blob: e7f2cfa6a2571afe6338368367a3aa269043daf1 [file] [log] [blame]
Linus Torvalds1da177e2005-04-16 15:20:36 -07001/*
2 * kernel/sched.c
3 *
4 * Kernel scheduler and related syscalls
5 *
6 * Copyright (C) 1991-2002 Linus Torvalds
7 *
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
11 * by Andrea Arcangeli
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
Ingo Molnarc31f2e82007-07-09 18:52:01 +020019 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
22 * by Peter Williams
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
Ingo Molnarb9131762008-01-25 21:08:19 +010025 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
Linus Torvalds1da177e2005-04-16 15:20:36 -070027 */
28
29#include <linux/mm.h>
30#include <linux/module.h>
31#include <linux/nmi.h>
32#include <linux/init.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020033#include <linux/uaccess.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070034#include <linux/highmem.h>
35#include <linux/smp_lock.h>
36#include <asm/mmu_context.h>
37#include <linux/interrupt.h>
Randy.Dunlapc59ede72006-01-11 12:17:46 -080038#include <linux/capability.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070039#include <linux/completion.h>
40#include <linux/kernel_stat.h>
Ingo Molnar9a11b49a2006-07-03 00:24:33 -070041#include <linux/debug_locks.h>
Ingo Molnarcdd6c482009-09-21 12:02:48 +020042#include <linux/perf_event.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070043#include <linux/security.h>
44#include <linux/notifier.h>
45#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080046#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080047#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070048#include <linux/blkdev.h>
49#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070050#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070051#include <linux/smp.h>
52#include <linux/threads.h>
53#include <linux/timer.h>
54#include <linux/rcupdate.h>
55#include <linux/cpu.h>
56#include <linux/cpuset.h>
57#include <linux/percpu.h>
58#include <linux/kthread.h>
Alexey Dobriyanb5aadf72008-10-06 13:23:43 +040059#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070060#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020061#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070062#include <linux/syscalls.h>
63#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070064#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080065#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070066#include <linux/delayacct.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020067#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020068#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010069#include <linux/hrtimer.h>
Reynes Philippe30914a52008-03-17 16:19:05 -070070#include <linux/tick.h>
Peter Zijlstraf00b45c2008-04-19 19:45:00 +020071#include <linux/debugfs.h>
72#include <linux/ctype.h>
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +020073#include <linux/ftrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070074
Eric Dumazet5517d862007-05-08 00:32:57 -070075#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020076#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070077
Gregory Haskins6e0534f2008-05-12 21:21:01 +020078#include "sched_cpupri.h"
79
Steven Rostedta8d154b2009-04-10 09:36:00 -040080#define CREATE_TRACE_POINTS
Steven Rostedtad8d75f2009-04-14 19:39:12 -040081#include <trace/events/sched.h>
Steven Rostedta8d154b2009-04-10 09:36:00 -040082
Linus Torvalds1da177e2005-04-16 15:20:36 -070083/*
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
86 * and back.
87 */
88#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
91
92/*
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
96 */
97#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
100
101/*
Ingo Molnard7876a02008-01-25 21:08:19 +0100102 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -0700103 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100104#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700105
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200106#define NICE_0_LOAD SCHED_LOAD_SCALE
107#define NICE_0_SHIFT SCHED_LOAD_SHIFT
108
Linus Torvalds1da177e2005-04-16 15:20:36 -0700109/*
110 * These are the 'tuning knobs' of the scheduler:
111 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700113 * Timeslices get refilled after they expire.
114 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700115#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700116
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200117/*
118 * single value that denotes runtime == period, ie unlimited time.
119 */
120#define RUNTIME_INF ((u64)~0ULL)
121
Ingo Molnare05606d2007-07-09 18:51:59 +0200122static inline int rt_policy(int policy)
123{
Roel Kluin3f33a7c2008-05-13 23:44:11 +0200124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
Ingo Molnare05606d2007-07-09 18:51:59 +0200125 return 1;
126 return 0;
127}
128
129static inline int task_has_rt_policy(struct task_struct *p)
130{
131 return rt_policy(p->policy);
132}
133
Linus Torvalds1da177e2005-04-16 15:20:36 -0700134/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200135 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700136 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200137struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
140};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700141
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200142struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
145 ktime_t rt_period;
146 u64 rt_runtime;
147 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200148};
149
150static struct rt_bandwidth def_rt_bandwidth;
151
152static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
153
154static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
155{
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
158 ktime_t now;
159 int overrun;
160 int idle = 0;
161
162 for (;;) {
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
165
166 if (!overrun)
167 break;
168
169 idle = do_sched_rt_period_timer(rt_b, overrun);
170 }
171
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
173}
174
175static
176void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
177{
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
180
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200181 spin_lock_init(&rt_b->rt_runtime_lock);
182
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200186}
187
Krzysztof Heltc8bfff62008-09-05 23:46:19 +0200188static inline int rt_bandwidth_enabled(void)
189{
190 return sysctl_sched_rt_runtime >= 0;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200191}
192
193static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
194{
195 ktime_t now;
196
Hiroshi Shimamotocac64d02009-02-25 09:59:26 -0800197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200198 return;
199
200 if (hrtimer_active(&rt_b->rt_period_timer))
201 return;
202
203 spin_lock(&rt_b->rt_runtime_lock);
204 for (;;) {
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100205 unsigned long delta;
206 ktime_t soft, hard;
207
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200208 if (hrtimer_active(&rt_b->rt_period_timer))
209 break;
210
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100213
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +0530218 HRTIMER_MODE_ABS_PINNED, 0);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200219 }
220 spin_unlock(&rt_b->rt_runtime_lock);
221}
222
223#ifdef CONFIG_RT_GROUP_SCHED
224static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
225{
226 hrtimer_cancel(&rt_b->rt_period_timer);
227}
228#endif
229
Heiko Carstens712555e2008-04-28 11:33:07 +0200230/*
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
233 */
234static DEFINE_MUTEX(sched_domains_mutex);
235
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100236#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200237
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700238#include <linux/cgroup.h>
239
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200240struct cfs_rq;
241
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100242static LIST_HEAD(task_groups);
243
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200244/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200245struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100246#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700247 struct cgroup_subsys_state css;
248#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100249
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530250#ifdef CONFIG_USER_SCHED
251 uid_t uid;
252#endif
253
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100254#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100260#endif
261
262#ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
265
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200266 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100267#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100268
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100269 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100270 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200271
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200275};
276
Dhaval Giani354d60c2008-04-19 19:44:59 +0200277#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200278
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530279/* Helper function to pass uid information to create_sched_user() */
280void set_tg_uid(struct user_struct *user)
281{
282 user->tg->uid = user->uid;
283}
284
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200285/*
286 * Root task group.
Anirban Sinha84e9dab2009-08-28 22:40:43 -0700287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200289 */
290struct task_group root_task_group;
291
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100292#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200293/* Default task group's sched entity on each cpu */
294static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295/* Default task group's cfs_rq on each cpu */
Linus Torvaldsada3fa12009-09-15 09:39:44 -0700296static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq, init_tg_cfs_rq);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200297#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100298
299#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100300static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
Tejun Heob9bf3122009-06-24 15:13:47 +0900301static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq, init_rt_rq);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200302#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200303#else /* !CONFIG_USER_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200304#define root_task_group init_task_group
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200305#endif /* CONFIG_USER_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100306
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100307/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100308 * a task group's cpu shares.
309 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100310static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100311
Cyrill Gorcunove9036b32009-10-26 22:24:14 +0300312#ifdef CONFIG_FAIR_GROUP_SCHED
313
Peter Zijlstra57310a92009-03-09 13:56:21 +0100314#ifdef CONFIG_SMP
315static int root_task_group_empty(void)
316{
317 return list_empty(&root_task_group.children);
318}
319#endif
320
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100321#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100322# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200323#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100324# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200325#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200326
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800327/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800328 * A weight of 0 or 1 can cause arithmetics problems.
329 * A weight of a cfs_rq is the sum of weights of which entities
330 * are queued on this cfs_rq, so a weight of a entity should not be
331 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800332 * (The default weight is 1024 - so there's no practical
333 * limitation from this.)
334 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200335#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800336#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200337
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100338static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100339#endif
340
341/* Default task group.
342 * Every task in system belong to this group at bootup.
343 */
Mike Travis434d53b2008-04-04 18:11:04 -0700344struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200345
346/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200347static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200348{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200349 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200350
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100351#ifdef CONFIG_USER_SCHED
David Howellsc69e8d92008-11-14 10:39:19 +1100352 rcu_read_lock();
353 tg = __task_cred(p)->user->tg;
354 rcu_read_unlock();
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100355#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700356 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
357 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200358#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100359 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200360#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200361 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200362}
363
364/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100365static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200366{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100367#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100368 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
369 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100370#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100371
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100372#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100373 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
374 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100375#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200376}
377
378#else
379
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100380static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Peter Zijlstra83378262008-06-27 13:41:37 +0200381static inline struct task_group *task_group(struct task_struct *p)
382{
383 return NULL;
384}
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200385
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100386#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200387
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200388/* CFS-related fields in a runqueue */
389struct cfs_rq {
390 struct load_weight load;
391 unsigned long nr_running;
392
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200393 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200394 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200395
396 struct rb_root tasks_timeline;
397 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200398
399 struct list_head tasks;
400 struct list_head *balance_iterator;
401
402 /*
403 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200404 * It is set to NULL otherwise (i.e when none are currently running).
405 */
Peter Zijlstra47932412008-11-04 21:25:09 +0100406 struct sched_entity *curr, *next, *last;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200407
Peter Zijlstra5ac5c4d2008-11-10 10:46:32 +0100408 unsigned int nr_spread_over;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200409
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200410#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200411 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
412
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100413 /*
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
417 *
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
420 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100421 struct list_head leaf_cfs_rq_list;
422 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200423
424#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200425 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200426 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200427 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200428 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200429
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200430 /*
431 * h_load = weight * f(tg)
432 *
433 * Where f(tg) is the recursive weight fraction assigned to
434 * this group.
435 */
436 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200437
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200438 /*
439 * this cpu's part of tg->shares
440 */
441 unsigned long shares;
Peter Zijlstraf1d239f2008-06-27 13:41:38 +0200442
443 /*
444 * load.weight at the time we set shares
445 */
446 unsigned long rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200447#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200448#endif
449};
450
451/* Real-Time classes' related field in a runqueue: */
452struct rt_rq {
453 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100454 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100455#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -0500456 struct {
457 int curr; /* highest queued rt task prio */
Gregory Haskins398a1532009-01-14 09:10:04 -0500458#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500459 int next; /* next highest */
Gregory Haskins398a1532009-01-14 09:10:04 -0500460#endif
Gregory Haskinse864c492008-12-29 09:39:49 -0500461 } highest_prio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100462#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100463#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100464 unsigned long rt_nr_migratory;
Peter Zijlstraa1ba4d82009-04-01 18:40:15 +0200465 unsigned long rt_nr_total;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100466 int overloaded;
Gregory Haskins917b6272008-12-29 09:39:53 -0500467 struct plist_head pushable_tasks;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100468#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100469 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100470 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200471 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100472 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200473 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100474
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100475#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100476 unsigned long rt_nr_boosted;
477
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100478 struct rq *rq;
479 struct list_head leaf_rt_rq_list;
480 struct task_group *tg;
481 struct sched_rt_entity *rt_se;
482#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200483};
484
Gregory Haskins57d885f2008-01-25 21:08:18 +0100485#ifdef CONFIG_SMP
486
487/*
488 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100489 * variables. Each exclusive cpuset essentially defines an island domain by
490 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100491 * exclusive cpuset is created, we also create and attach a new root-domain
492 * object.
493 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100494 */
495struct root_domain {
496 atomic_t refcount;
Rusty Russellc6c49272008-11-25 02:35:05 +1030497 cpumask_var_t span;
498 cpumask_var_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100499
Ingo Molnar0eab9142008-01-25 21:08:19 +0100500 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100501 * The "RT overload" flag: it gets set if a CPU has more than
502 * one runnable RT task.
503 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030504 cpumask_var_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100505 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200506#ifdef CONFIG_SMP
507 struct cpupri cpupri;
508#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100509};
510
Gregory Haskinsdc938522008-01-25 21:08:26 +0100511/*
512 * By default the system creates a single root-domain with all cpus as
513 * members (mimicking the global state we have today).
514 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100515static struct root_domain def_root_domain;
516
517#endif
518
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200519/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700520 * This is the main, per-CPU runqueue data structure.
521 *
522 * Locking rule: those places that want to lock multiple runqueues
523 * (such as the load balancing or the thread migration code), lock
524 * acquire operations must be ordered by ascending &runqueue.
525 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700526struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200527 /* runqueue lock: */
528 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700529
530 /*
531 * nr_running and cpu_load should be in the same cacheline because
532 * remote CPUs use both these fields when doing load calculation.
533 */
534 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200535 #define CPU_LOAD_IDX_MAX 5
536 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700537#ifdef CONFIG_NO_HZ
538 unsigned char in_nohz_recently;
539#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200540 /* capture load from *all* tasks on this cpu: */
541 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200542 unsigned long nr_load_updates;
543 u64 nr_switches;
544
545 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100546 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100547
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200548#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200549 /* list of leaf cfs_rq on this cpu: */
550 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100551#endif
552#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100553 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700554#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555
556 /*
557 * This is part of a global counter where only the total sum
558 * over all CPUs matters. A task can increase this counter on
559 * one CPU and if it got migrated afterwards it may decrease
560 * it on another CPU. Always updated under the runqueue lock:
561 */
562 unsigned long nr_uninterruptible;
563
Ingo Molnar36c8b582006-07-03 00:25:41 -0700564 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800565 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700566 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200567
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200568 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200569
Linus Torvalds1da177e2005-04-16 15:20:36 -0700570 atomic_t nr_iowait;
571
572#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100573 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700574 struct sched_domain *sd;
575
Henrik Austada0a522c2009-02-13 20:35:45 +0100576 unsigned char idle_at_tick;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700577 /* For active balancing */
Gregory Haskins3f029d32009-07-29 11:08:47 -0400578 int post_schedule;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700579 int active_balance;
580 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200581 /* cpu of this runqueue: */
582 int cpu;
Gregory Haskins1f11eb62008-06-04 15:04:05 -0400583 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700584
Peter Zijlstraa8a51d52008-06-27 13:41:26 +0200585 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586
Ingo Molnar36c8b582006-07-03 00:25:41 -0700587 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700588 struct list_head migration_queue;
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200589
590 u64 rt_avg;
591 u64 age_stamp;
Mike Galbraith1b9508f2009-11-04 17:53:50 +0100592 u64 idle_stamp;
593 u64 avg_idle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594#endif
595
Thomas Gleixnerdce48a82009-04-11 10:43:41 +0200596 /* calc_load related fields */
597 unsigned long calc_load_update;
598 long calc_load_active;
599
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100600#ifdef CONFIG_SCHED_HRTICK
Peter Zijlstra31656512008-07-18 18:01:23 +0200601#ifdef CONFIG_SMP
602 int hrtick_csd_pending;
603 struct call_single_data hrtick_csd;
604#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100605 struct hrtimer hrtick_timer;
606#endif
607
Linus Torvalds1da177e2005-04-16 15:20:36 -0700608#ifdef CONFIG_SCHEDSTATS
609 /* latency stats */
610 struct sched_info rq_sched_info;
Ken Chen9c2c4802008-12-16 23:41:22 -0800611 unsigned long long rq_cpu_time;
612 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700613
614 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200615 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700616
617 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200618 unsigned int sched_switch;
619 unsigned int sched_count;
620 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700621
622 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200623 unsigned int ttwu_count;
624 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200625
626 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200627 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700628#endif
629};
630
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700631static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700632
Peter Zijlstra7d478722009-09-14 19:55:44 +0200633static inline
634void check_preempt_curr(struct rq *rq, struct task_struct *p, int flags)
Ingo Molnardd41f592007-07-09 18:51:59 +0200635{
Peter Zijlstra7d478722009-09-14 19:55:44 +0200636 rq->curr->sched_class->check_preempt_curr(rq, p, flags);
Ingo Molnardd41f592007-07-09 18:51:59 +0200637}
638
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700639static inline int cpu_of(struct rq *rq)
640{
641#ifdef CONFIG_SMP
642 return rq->cpu;
643#else
644 return 0;
645#endif
646}
647
Ingo Molnar20d315d2007-07-09 18:51:58 +0200648/*
Nick Piggin674311d2005-06-25 14:57:27 -0700649 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700650 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700651 *
652 * The domain tree of any CPU may only be accessed from within
653 * preempt-disabled sections.
654 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700655#define for_each_domain(cpu, __sd) \
656 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700657
658#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
659#define this_rq() (&__get_cpu_var(runqueues))
660#define task_rq(p) cpu_rq(task_cpu(p))
661#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
Hitoshi Mitake54d35f22009-06-29 14:44:57 +0900662#define raw_rq() (&__raw_get_cpu_var(runqueues))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700663
Ingo Molnaraa9c4c02008-12-17 14:10:57 +0100664inline void update_rq_clock(struct rq *rq)
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200665{
666 rq->clock = sched_clock_cpu(cpu_of(rq));
667}
668
Ingo Molnare436d802007-07-19 21:28:35 +0200669/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200670 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
671 */
672#ifdef CONFIG_SCHED_DEBUG
673# define const_debug __read_mostly
674#else
675# define const_debug static const
676#endif
677
Ingo Molnar017730c2008-05-12 21:20:52 +0200678/**
679 * runqueue_is_locked
Randy Dunlape17b38b2009-10-11 19:12:00 -0700680 * @cpu: the processor in question.
Ingo Molnar017730c2008-05-12 21:20:52 +0200681 *
682 * Returns true if the current cpu runqueue is locked.
683 * This interface allows printk to be called with the runqueue lock
684 * held and know whether or not it is OK to wake up the klogd.
685 */
Andrew Morton89f19f02009-09-19 11:55:44 -0700686int runqueue_is_locked(int cpu)
Ingo Molnar017730c2008-05-12 21:20:52 +0200687{
Andrew Morton89f19f02009-09-19 11:55:44 -0700688 return spin_is_locked(&cpu_rq(cpu)->lock);
Ingo Molnar017730c2008-05-12 21:20:52 +0200689}
690
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200691/*
692 * Debugging: various feature bits
693 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200694
695#define SCHED_FEAT(name, enabled) \
696 __SCHED_FEAT_##name ,
697
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200698enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200699#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200700};
701
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200702#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200703
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200704#define SCHED_FEAT(name, enabled) \
705 (1UL << __SCHED_FEAT_##name) * enabled |
706
707const_debug unsigned int sysctl_sched_features =
708#include "sched_features.h"
709 0;
710
711#undef SCHED_FEAT
712
713#ifdef CONFIG_SCHED_DEBUG
714#define SCHED_FEAT(name, enabled) \
715 #name ,
716
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700717static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200718#include "sched_features.h"
719 NULL
720};
721
722#undef SCHED_FEAT
723
Li Zefan34f3a812008-10-30 15:23:32 +0800724static int sched_feat_show(struct seq_file *m, void *v)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200725{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200726 int i;
727
728 for (i = 0; sched_feat_names[i]; i++) {
Li Zefan34f3a812008-10-30 15:23:32 +0800729 if (!(sysctl_sched_features & (1UL << i)))
730 seq_puts(m, "NO_");
731 seq_printf(m, "%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200732 }
Li Zefan34f3a812008-10-30 15:23:32 +0800733 seq_puts(m, "\n");
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200734
Li Zefan34f3a812008-10-30 15:23:32 +0800735 return 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200736}
737
738static ssize_t
739sched_feat_write(struct file *filp, const char __user *ubuf,
740 size_t cnt, loff_t *ppos)
741{
742 char buf[64];
743 char *cmp = buf;
744 int neg = 0;
745 int i;
746
747 if (cnt > 63)
748 cnt = 63;
749
750 if (copy_from_user(&buf, ubuf, cnt))
751 return -EFAULT;
752
753 buf[cnt] = 0;
754
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200755 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200756 neg = 1;
757 cmp += 3;
758 }
759
760 for (i = 0; sched_feat_names[i]; i++) {
761 int len = strlen(sched_feat_names[i]);
762
763 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 if (neg)
765 sysctl_sched_features &= ~(1UL << i);
766 else
767 sysctl_sched_features |= (1UL << i);
768 break;
769 }
770 }
771
772 if (!sched_feat_names[i])
773 return -EINVAL;
774
Jan Blunck42994722009-11-20 17:40:37 +0100775 *ppos += cnt;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200776
777 return cnt;
778}
779
Li Zefan34f3a812008-10-30 15:23:32 +0800780static int sched_feat_open(struct inode *inode, struct file *filp)
781{
782 return single_open(filp, sched_feat_show, NULL);
783}
784
Alexey Dobriyan828c0952009-10-01 15:43:56 -0700785static const struct file_operations sched_feat_fops = {
Li Zefan34f3a812008-10-30 15:23:32 +0800786 .open = sched_feat_open,
787 .write = sched_feat_write,
788 .read = seq_read,
789 .llseek = seq_lseek,
790 .release = single_release,
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200791};
792
793static __init int sched_init_debug(void)
794{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200795 debugfs_create_file("sched_features", 0644, NULL, NULL,
796 &sched_feat_fops);
797
798 return 0;
799}
800late_initcall(sched_init_debug);
801
802#endif
803
804#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200805
806/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100807 * Number of tasks to iterate in a single balance run.
808 * Limited because this is done with IRQs disabled.
809 */
810const_debug unsigned int sysctl_sched_nr_migrate = 32;
811
812/*
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200813 * ratelimit for updating the group shares.
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200814 * default: 0.25ms
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200815 */
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200816unsigned int sysctl_sched_shares_ratelimit = 250000;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200817
818/*
Peter Zijlstraffda12a2008-10-17 19:27:02 +0200819 * Inject some fuzzyness into changing the per-cpu group shares
820 * this avoids remote rq-locks at the expense of fairness.
821 * default: 4
822 */
823unsigned int sysctl_sched_shares_thresh = 4;
824
825/*
Peter Zijlstrae9e92502009-09-01 10:34:37 +0200826 * period over which we average the RT time consumption, measured
827 * in ms.
828 *
829 * default: 1s
830 */
831const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
832
833/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100834 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100835 * default: 1s
836 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100837unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100838
Ingo Molnar6892b752008-02-13 14:02:36 +0100839static __read_mostly int scheduler_running;
840
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100841/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100842 * part of the period that we allow rt tasks to run in us.
843 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100844 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100845int sysctl_sched_rt_runtime = 950000;
846
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200847static inline u64 global_rt_period(void)
848{
849 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
850}
851
852static inline u64 global_rt_runtime(void)
853{
roel kluine26873b2008-07-22 16:51:15 -0400854 if (sysctl_sched_rt_runtime < 0)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200855 return RUNTIME_INF;
856
857 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
858}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100859
Linus Torvalds1da177e2005-04-16 15:20:36 -0700860#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700861# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700862#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700863#ifndef finish_arch_switch
864# define finish_arch_switch(prev) do { } while (0)
865#endif
866
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100867static inline int task_current(struct rq *rq, struct task_struct *p)
868{
869 return rq->curr == p;
870}
871
Nick Piggin4866cde2005-06-25 14:57:23 -0700872#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700873static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700874{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100875 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700876}
877
Ingo Molnar70b97a72006-07-03 00:25:42 -0700878static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700879{
880}
881
Ingo Molnar70b97a72006-07-03 00:25:42 -0700882static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700883{
Ingo Molnarda04c032005-09-13 11:17:59 +0200884#ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq->lock.owner = current;
887#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700888 /*
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
891 * prev into current:
892 */
893 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
894
Nick Piggin4866cde2005-06-25 14:57:23 -0700895 spin_unlock_irq(&rq->lock);
896}
897
898#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700899static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700900{
901#ifdef CONFIG_SMP
902 return p->oncpu;
903#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100904 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700905#endif
906}
907
Ingo Molnar70b97a72006-07-03 00:25:42 -0700908static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700909{
910#ifdef CONFIG_SMP
911 /*
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
914 * here.
915 */
916 next->oncpu = 1;
917#endif
918#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq->lock);
920#else
921 spin_unlock(&rq->lock);
922#endif
923}
924
Ingo Molnar70b97a72006-07-03 00:25:42 -0700925static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700926{
927#ifdef CONFIG_SMP
928 /*
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
931 * finished.
932 */
933 smp_wmb();
934 prev->oncpu = 0;
935#endif
936#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 local_irq_enable();
938#endif
939}
940#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700941
942/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
945 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700946static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700947 __acquires(rq->lock)
948{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200949 for (;;) {
950 struct rq *rq = task_rq(p);
951 spin_lock(&rq->lock);
952 if (likely(rq == task_rq(p)))
953 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700954 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700955 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700956}
957
958/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700959 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100960 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961 * explicitly disabling preemption.
962 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700963static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700964 __acquires(rq->lock)
965{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700966 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700967
Andi Kleen3a5c3592007-10-15 17:00:14 +0200968 for (;;) {
969 local_irq_save(*flags);
970 rq = task_rq(p);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
973 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700974 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700975 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700976}
977
Oleg Nesterovad474ca2008-11-10 15:39:30 +0100978void task_rq_unlock_wait(struct task_struct *p)
979{
980 struct rq *rq = task_rq(p);
981
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq->lock);
984}
985
Alexey Dobriyana9957442007-10-15 17:00:13 +0200986static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700987 __releases(rq->lock)
988{
989 spin_unlock(&rq->lock);
990}
991
Ingo Molnar70b97a72006-07-03 00:25:42 -0700992static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 __releases(rq->lock)
994{
995 spin_unlock_irqrestore(&rq->lock, *flags);
996}
997
Linus Torvalds1da177e2005-04-16 15:20:36 -0700998/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800999 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001000 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001001static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001002 __acquires(rq->lock)
1003{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001004 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005
1006 local_irq_disable();
1007 rq = this_rq();
1008 spin_lock(&rq->lock);
1009
1010 return rq;
1011}
1012
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001013#ifdef CONFIG_SCHED_HRTICK
1014/*
1015 * Use HR-timers to deliver accurate preemption points.
1016 *
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1019 * reschedule event.
1020 *
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1022 * rq->lock.
1023 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001024
1025/*
1026 * Use hrtick when:
1027 * - enabled by features
1028 * - hrtimer is actually high res
1029 */
1030static inline int hrtick_enabled(struct rq *rq)
1031{
1032 if (!sched_feat(HRTICK))
1033 return 0;
Ingo Molnarba420592008-07-20 11:02:06 +02001034 if (!cpu_active(cpu_of(rq)))
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001035 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001036 return hrtimer_is_hres_active(&rq->hrtick_timer);
1037}
1038
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001039static void hrtick_clear(struct rq *rq)
1040{
1041 if (hrtimer_active(&rq->hrtick_timer))
1042 hrtimer_cancel(&rq->hrtick_timer);
1043}
1044
1045/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1048 */
1049static enum hrtimer_restart hrtick(struct hrtimer *timer)
1050{
1051 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1052
1053 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1054
1055 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001056 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001057 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1058 spin_unlock(&rq->lock);
1059
1060 return HRTIMER_NORESTART;
1061}
1062
Rabin Vincent95e904c2008-05-11 05:55:33 +05301063#ifdef CONFIG_SMP
Peter Zijlstra31656512008-07-18 18:01:23 +02001064/*
1065 * called from hardirq (IPI) context
1066 */
1067static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001068{
Peter Zijlstra31656512008-07-18 18:01:23 +02001069 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001070
Peter Zijlstra31656512008-07-18 18:01:23 +02001071 spin_lock(&rq->lock);
1072 hrtimer_restart(&rq->hrtick_timer);
1073 rq->hrtick_csd_pending = 0;
1074 spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001075}
1076
Peter Zijlstra31656512008-07-18 18:01:23 +02001077/*
1078 * Called to set the hrtick timer state.
1079 *
1080 * called with rq->lock held and irqs disabled
1081 */
1082static void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001083{
Peter Zijlstra31656512008-07-18 18:01:23 +02001084 struct hrtimer *timer = &rq->hrtick_timer;
1085 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001086
Arjan van de Vencc584b22008-09-01 15:02:30 -07001087 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +02001088
1089 if (rq == this_rq()) {
1090 hrtimer_restart(timer);
1091 } else if (!rq->hrtick_csd_pending) {
Peter Zijlstra6e275632009-02-25 13:59:48 +01001092 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001093 rq->hrtick_csd_pending = 1;
1094 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001095}
1096
1097static int
1098hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1099{
1100 int cpu = (int)(long)hcpu;
1101
1102 switch (action) {
1103 case CPU_UP_CANCELED:
1104 case CPU_UP_CANCELED_FROZEN:
1105 case CPU_DOWN_PREPARE:
1106 case CPU_DOWN_PREPARE_FROZEN:
1107 case CPU_DEAD:
1108 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +02001109 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001110 return NOTIFY_OK;
1111 }
1112
1113 return NOTIFY_DONE;
1114}
1115
Rakib Mullickfa748202008-09-22 14:55:45 -07001116static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001117{
1118 hotcpu_notifier(hotplug_hrtick, 0);
1119}
Peter Zijlstra31656512008-07-18 18:01:23 +02001120#else
1121/*
1122 * Called to set the hrtick timer state.
1123 *
1124 * called with rq->lock held and irqs disabled
1125 */
1126static void hrtick_start(struct rq *rq, u64 delay)
1127{
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +01001128 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +05301129 HRTIMER_MODE_REL_PINNED, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001130}
1131
Andrew Morton006c75f2008-09-22 14:55:46 -07001132static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +02001133{
1134}
Rabin Vincent95e904c2008-05-11 05:55:33 +05301135#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001136
1137static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001138{
Peter Zijlstra31656512008-07-18 18:01:23 +02001139#ifdef CONFIG_SMP
1140 rq->hrtick_csd_pending = 0;
1141
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1145#endif
1146
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001149}
Andrew Morton006c75f2008-09-22 14:55:46 -07001150#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001151static inline void hrtick_clear(struct rq *rq)
1152{
1153}
1154
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001155static inline void init_rq_hrtick(struct rq *rq)
1156{
1157}
1158
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001159static inline void init_hrtick(void)
1160{
1161}
Andrew Morton006c75f2008-09-22 14:55:46 -07001162#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001163
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001164/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001165 * resched_task - mark a task 'to be rescheduled now'.
1166 *
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1169 * the target CPU.
1170 */
1171#ifdef CONFIG_SMP
1172
1173#ifndef tsk_is_polling
1174#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1175#endif
1176
Peter Zijlstra31656512008-07-18 18:01:23 +02001177static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001178{
1179 int cpu;
1180
1181 assert_spin_locked(&task_rq(p)->lock);
1182
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001183 if (test_tsk_need_resched(p))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001184 return;
1185
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001186 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001187
1188 cpu = task_cpu(p);
1189 if (cpu == smp_processor_id())
1190 return;
1191
1192 /* NEED_RESCHED must be visible before we test polling */
1193 smp_mb();
1194 if (!tsk_is_polling(p))
1195 smp_send_reschedule(cpu);
1196}
1197
1198static void resched_cpu(int cpu)
1199{
1200 struct rq *rq = cpu_rq(cpu);
1201 unsigned long flags;
1202
1203 if (!spin_trylock_irqsave(&rq->lock, flags))
1204 return;
1205 resched_task(cpu_curr(cpu));
1206 spin_unlock_irqrestore(&rq->lock, flags);
1207}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001208
1209#ifdef CONFIG_NO_HZ
1210/*
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1219 */
1220void wake_up_idle_cpu(int cpu)
1221{
1222 struct rq *rq = cpu_rq(cpu);
1223
1224 if (cpu == smp_processor_id())
1225 return;
1226
1227 /*
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1233 */
1234 if (rq->curr != rq->idle)
1235 return;
1236
1237 /*
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1241 */
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001242 set_tsk_need_resched(rq->idle);
Thomas Gleixner06d83082008-03-22 09:20:24 +01001243
1244 /* NEED_RESCHED must be visible before we test polling */
1245 smp_mb();
1246 if (!tsk_is_polling(rq->idle))
1247 smp_send_reschedule(cpu);
1248}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001249#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001250
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001251static u64 sched_avg_period(void)
1252{
1253 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1254}
1255
1256static void sched_avg_update(struct rq *rq)
1257{
1258 s64 period = sched_avg_period();
1259
1260 while ((s64)(rq->clock - rq->age_stamp) > period) {
1261 rq->age_stamp += period;
1262 rq->rt_avg /= 2;
1263 }
1264}
1265
1266static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1267{
1268 rq->rt_avg += rt_delta;
1269 sched_avg_update(rq);
1270}
1271
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001272#else /* !CONFIG_SMP */
Peter Zijlstra31656512008-07-18 18:01:23 +02001273static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001274{
1275 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001276 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001277}
Peter Zijlstrae9e92502009-09-01 10:34:37 +02001278
1279static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1280{
1281}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001282#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001283
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001284#if BITS_PER_LONG == 32
1285# define WMULT_CONST (~0UL)
1286#else
1287# define WMULT_CONST (1UL << 32)
1288#endif
1289
1290#define WMULT_SHIFT 32
1291
Ingo Molnar194081e2007-08-09 11:16:51 +02001292/*
1293 * Shift right and round:
1294 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001295#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001296
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001297/*
1298 * delta *= weight / lw
1299 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001300static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001301calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1302 struct load_weight *lw)
1303{
1304 u64 tmp;
1305
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001306 if (!lw->inv_weight) {
1307 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1308 lw->inv_weight = 1;
1309 else
1310 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1311 / (lw->weight+1);
1312 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001313
1314 tmp = (u64)delta_exec * weight;
1315 /*
1316 * Check whether we'd overflow the 64-bit multiplication:
1317 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001318 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001319 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001320 WMULT_SHIFT/2);
1321 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001322 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001323
Ingo Molnarecf691d2007-08-02 17:41:40 +02001324 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001325}
1326
Ingo Molnar10919852007-10-15 17:00:04 +02001327static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001328{
1329 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001330 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001331}
1332
Ingo Molnar10919852007-10-15 17:00:04 +02001333static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001334{
1335 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001336 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001337}
1338
Linus Torvalds1da177e2005-04-16 15:20:36 -07001339/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001340 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1341 * of tasks with abnormal "nice" values across CPUs the contribution that
1342 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001343 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001344 * scaled version of the new time slice allocation that they receive on time
1345 * slice expiry etc.
1346 */
1347
Peter Zijlstracce7ade2009-01-15 14:53:37 +01001348#define WEIGHT_IDLEPRIO 3
1349#define WMULT_IDLEPRIO 1431655765
Ingo Molnardd41f592007-07-09 18:51:59 +02001350
1351/*
1352 * Nice levels are multiplicative, with a gentle 10% change for every
1353 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1354 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1355 * that remained on nice 0.
1356 *
1357 * The "10% effect" is relative and cumulative: from _any_ nice level,
1358 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001359 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1360 * If a task goes up by ~10% and another task goes down by ~10% then
1361 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001362 */
1363static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001364 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1365 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1366 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1367 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1368 /* 0 */ 1024, 820, 655, 526, 423,
1369 /* 5 */ 335, 272, 215, 172, 137,
1370 /* 10 */ 110, 87, 70, 56, 45,
1371 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001372};
1373
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001374/*
1375 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1376 *
1377 * In cases where the weight does not change often, we can use the
1378 * precalculated inverse to speed up arithmetics by turning divisions
1379 * into multiplications:
1380 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001381static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001382 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1383 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1384 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1385 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1386 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1387 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1388 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1389 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001390};
Peter Williams2dd73a42006-06-27 02:54:34 -07001391
Ingo Molnardd41f592007-07-09 18:51:59 +02001392static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1393
1394/*
1395 * runqueue iterator, to support SMP load-balancing between different
1396 * scheduling classes, without having to expose their internal data
1397 * structures to the load-balancing proper:
1398 */
1399struct rq_iterator {
1400 void *arg;
1401 struct task_struct *(*start)(void *);
1402 struct task_struct *(*next)(void *);
1403};
1404
Peter Williamse1d14842007-10-24 18:23:51 +02001405#ifdef CONFIG_SMP
1406static unsigned long
1407balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1408 unsigned long max_load_move, struct sched_domain *sd,
1409 enum cpu_idle_type idle, int *all_pinned,
1410 int *this_best_prio, struct rq_iterator *iterator);
1411
1412static int
1413iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1414 struct sched_domain *sd, enum cpu_idle_type idle,
1415 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001416#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001417
Bharata B Raoef12fef2009-03-31 10:02:22 +05301418/* Time spent by the tasks of the cpu accounting group executing in ... */
1419enum cpuacct_stat_index {
1420 CPUACCT_STAT_USER, /* ... user mode */
1421 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1422
1423 CPUACCT_STAT_NSTATS,
1424};
1425
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001426#ifdef CONFIG_CGROUP_CPUACCT
1427static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
Bharata B Raoef12fef2009-03-31 10:02:22 +05301428static void cpuacct_update_stats(struct task_struct *tsk,
1429 enum cpuacct_stat_index idx, cputime_t val);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001430#else
1431static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
Bharata B Raoef12fef2009-03-31 10:02:22 +05301432static inline void cpuacct_update_stats(struct task_struct *tsk,
1433 enum cpuacct_stat_index idx, cputime_t val) {}
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001434#endif
1435
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001436static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1437{
1438 update_load_add(&rq->load, load);
1439}
1440
1441static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1442{
1443 update_load_sub(&rq->load, load);
1444}
1445
Ingo Molnar7940ca32008-08-19 13:40:47 +02001446#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001447typedef int (*tg_visitor)(struct task_group *, void *);
1448
1449/*
1450 * Iterate the full tree, calling @down when first entering a node and @up when
1451 * leaving it for the final time.
1452 */
1453static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1454{
1455 struct task_group *parent, *child;
1456 int ret;
1457
1458 rcu_read_lock();
1459 parent = &root_task_group;
1460down:
1461 ret = (*down)(parent, data);
1462 if (ret)
1463 goto out_unlock;
1464 list_for_each_entry_rcu(child, &parent->children, siblings) {
1465 parent = child;
1466 goto down;
1467
1468up:
1469 continue;
1470 }
1471 ret = (*up)(parent, data);
1472 if (ret)
1473 goto out_unlock;
1474
1475 child = parent;
1476 parent = parent->parent;
1477 if (parent)
1478 goto up;
1479out_unlock:
1480 rcu_read_unlock();
1481
1482 return ret;
1483}
1484
1485static int tg_nop(struct task_group *tg, void *data)
1486{
1487 return 0;
1488}
1489#endif
1490
Gregory Haskinse7693a32008-01-25 21:08:09 +01001491#ifdef CONFIG_SMP
Peter Zijlstraf5f08f32009-09-10 13:35:28 +02001492/* Used instead of source_load when we know the type == 0 */
1493static unsigned long weighted_cpuload(const int cpu)
1494{
1495 return cpu_rq(cpu)->load.weight;
1496}
1497
1498/*
1499 * Return a low guess at the load of a migration-source cpu weighted
1500 * according to the scheduling class and "nice" value.
1501 *
1502 * We want to under-estimate the load of migration sources, to
1503 * balance conservatively.
1504 */
1505static unsigned long source_load(int cpu, int type)
1506{
1507 struct rq *rq = cpu_rq(cpu);
1508 unsigned long total = weighted_cpuload(cpu);
1509
1510 if (type == 0 || !sched_feat(LB_BIAS))
1511 return total;
1512
1513 return min(rq->cpu_load[type-1], total);
1514}
1515
1516/*
1517 * Return a high guess at the load of a migration-target cpu weighted
1518 * according to the scheduling class and "nice" value.
1519 */
1520static unsigned long target_load(int cpu, int type)
1521{
1522 struct rq *rq = cpu_rq(cpu);
1523 unsigned long total = weighted_cpuload(cpu);
1524
1525 if (type == 0 || !sched_feat(LB_BIAS))
1526 return total;
1527
1528 return max(rq->cpu_load[type-1], total);
1529}
1530
Peter Zijlstraae154be2009-09-10 14:40:57 +02001531static struct sched_group *group_of(int cpu)
1532{
1533 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
1534
1535 if (!sd)
1536 return NULL;
1537
1538 return sd->groups;
1539}
1540
1541static unsigned long power_of(int cpu)
1542{
1543 struct sched_group *group = group_of(cpu);
1544
1545 if (!group)
1546 return SCHED_LOAD_SCALE;
1547
1548 return group->cpu_power;
1549}
1550
Gregory Haskinse7693a32008-01-25 21:08:09 +01001551static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001552
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001553static unsigned long cpu_avg_load_per_task(int cpu)
1554{
1555 struct rq *rq = cpu_rq(cpu);
Ingo Molnaraf6d5962008-11-29 20:45:15 +01001556 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001557
Steven Rostedt4cd42622008-11-26 21:04:24 -05001558 if (nr_running)
1559 rq->avg_load_per_task = rq->load.weight / nr_running;
Balbir Singha2d47772008-11-12 16:19:00 +05301560 else
1561 rq->avg_load_per_task = 0;
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001562
1563 return rq->avg_load_per_task;
1564}
1565
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001566#ifdef CONFIG_FAIR_GROUP_SCHED
1567
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001568static __read_mostly unsigned long *update_shares_data;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001569
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001570static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1571
1572/*
1573 * Calculate and set the cpu's group shares.
1574 */
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001575static void update_group_shares_cpu(struct task_group *tg, int cpu,
1576 unsigned long sd_shares,
1577 unsigned long sd_rq_weight,
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001578 unsigned long *usd_rq_weight)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001579{
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001580 unsigned long shares, rq_weight;
Peter Zijlstraa5004272009-07-27 14:04:49 +02001581 int boost = 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001582
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001583 rq_weight = usd_rq_weight[cpu];
Peter Zijlstraa5004272009-07-27 14:04:49 +02001584 if (!rq_weight) {
1585 boost = 1;
1586 rq_weight = NICE_0_LOAD;
1587 }
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001588
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001589 /*
Peter Zijlstraa8af7242009-08-21 13:58:54 +02001590 * \Sum_j shares_j * rq_weight_i
1591 * shares_i = -----------------------------
1592 * \Sum_j rq_weight_j
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001593 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001594 shares = (sd_shares * rq_weight) / sd_rq_weight;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001595 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001596
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001597 if (abs(shares - tg->se[cpu]->load.weight) >
1598 sysctl_sched_shares_thresh) {
1599 struct rq *rq = cpu_rq(cpu);
1600 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001601
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001602 spin_lock_irqsave(&rq->lock, flags);
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001603 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
Peter Zijlstraa5004272009-07-27 14:04:49 +02001604 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001605 __set_se_shares(tg->se[cpu], shares);
1606 spin_unlock_irqrestore(&rq->lock, flags);
1607 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001608}
1609
1610/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001611 * Re-compute the task group their per cpu shares over the given domain.
1612 * This needs to be done in a bottom-up fashion because the rq weight of a
1613 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001614 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001615static int tg_shares_up(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001616{
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001617 unsigned long weight, rq_weight = 0, shares = 0;
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001618 unsigned long *usd_rq_weight;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001619 struct sched_domain *sd = data;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001620 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001621 int i;
1622
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001623 if (!tg->se[0])
1624 return 0;
1625
1626 local_irq_save(flags);
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001627 usd_rq_weight = per_cpu_ptr(update_shares_data, smp_processor_id());
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001628
Rusty Russell758b2cd2008-11-25 02:35:04 +10301629 for_each_cpu(i, sched_domain_span(sd)) {
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001630 weight = tg->cfs_rq[i]->load.weight;
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001631 usd_rq_weight[i] = weight;
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001632
Ken Chenec4e0e22008-11-18 22:41:57 -08001633 /*
1634 * If there are currently no tasks on the cpu pretend there
1635 * is one of average load so that when a new task gets to
1636 * run here it will not get delayed by group starvation.
1637 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001638 if (!weight)
1639 weight = NICE_0_LOAD;
1640
Ken Chenec4e0e22008-11-18 22:41:57 -08001641 rq_weight += weight;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001642 shares += tg->cfs_rq[i]->shares;
1643 }
1644
1645 if ((!shares && rq_weight) || shares > tg->shares)
1646 shares = tg->shares;
1647
1648 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1649 shares = tg->shares;
1650
Rusty Russell758b2cd2008-11-25 02:35:04 +10301651 for_each_cpu(i, sched_domain_span(sd))
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09001652 update_group_shares_cpu(tg, i, shares, rq_weight, usd_rq_weight);
Peter Zijlstra34d76c42009-08-27 13:08:56 +02001653
1654 local_irq_restore(flags);
Peter Zijlstraeb755802008-08-19 12:33:05 +02001655
1656 return 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001657}
1658
1659/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001660 * Compute the cpu's hierarchical load factor for each task group.
1661 * This needs to be done in a top-down fashion because the load of a child
1662 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001663 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001664static int tg_load_down(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001665{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001666 unsigned long load;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001667 long cpu = (long)data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001668
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001669 if (!tg->parent) {
1670 load = cpu_rq(cpu)->load.weight;
1671 } else {
1672 load = tg->parent->cfs_rq[cpu]->h_load;
1673 load *= tg->cfs_rq[cpu]->shares;
1674 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1675 }
1676
1677 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001678
Peter Zijlstraeb755802008-08-19 12:33:05 +02001679 return 0;
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001680}
1681
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001682static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001683{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001684 s64 elapsed;
1685 u64 now;
1686
1687 if (root_task_group_empty())
1688 return;
1689
1690 now = cpu_clock(raw_smp_processor_id());
1691 elapsed = now - sd->last_update;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001692
1693 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1694 sd->last_update = now;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001695 walk_tg_tree(tg_nop, tg_shares_up, sd);
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001696 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001697}
1698
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001699static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1700{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001701 if (root_task_group_empty())
1702 return;
1703
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001704 spin_unlock(&rq->lock);
1705 update_shares(sd);
1706 spin_lock(&rq->lock);
1707}
1708
Peter Zijlstraeb755802008-08-19 12:33:05 +02001709static void update_h_load(long cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001710{
Peter Zijlstrae7097152009-06-03 15:41:20 +02001711 if (root_task_group_empty())
1712 return;
1713
Peter Zijlstraeb755802008-08-19 12:33:05 +02001714 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001715}
1716
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001717#else
1718
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001719static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001720{
1721}
1722
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001723static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1724{
1725}
1726
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001727#endif
1728
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001729#ifdef CONFIG_PREEMPT
1730
Peter Zijlstrab78bb862009-09-15 14:23:18 +02001731static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1732
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001733/*
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001734 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1735 * way at the expense of forcing extra atomic operations in all
1736 * invocations. This assures that the double_lock is acquired using the
1737 * same underlying policy as the spinlock_t on this architecture, which
1738 * reduces latency compared to the unfair variant below. However, it
1739 * also adds more overhead and therefore may reduce throughput.
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001740 */
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001741static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1742 __releases(this_rq->lock)
1743 __acquires(busiest->lock)
1744 __acquires(this_rq->lock)
1745{
1746 spin_unlock(&this_rq->lock);
1747 double_rq_lock(this_rq, busiest);
1748
1749 return 1;
1750}
1751
1752#else
1753/*
1754 * Unfair double_lock_balance: Optimizes throughput at the expense of
1755 * latency by eliminating extra atomic operations when the locks are
1756 * already in proper order on entry. This favors lower cpu-ids and will
1757 * grant the double lock to lower cpus over higher ids under contention,
1758 * regardless of entry order into the function.
1759 */
1760static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001761 __releases(this_rq->lock)
1762 __acquires(busiest->lock)
1763 __acquires(this_rq->lock)
1764{
1765 int ret = 0;
1766
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001767 if (unlikely(!spin_trylock(&busiest->lock))) {
1768 if (busiest < this_rq) {
1769 spin_unlock(&this_rq->lock);
1770 spin_lock(&busiest->lock);
1771 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1772 ret = 1;
1773 } else
1774 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1775 }
1776 return ret;
1777}
1778
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001779#endif /* CONFIG_PREEMPT */
1780
1781/*
1782 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1783 */
1784static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1785{
1786 if (unlikely(!irqs_disabled())) {
1787 /* printk() doesn't work good under rq->lock */
1788 spin_unlock(&this_rq->lock);
1789 BUG_ON(1);
1790 }
1791
1792 return _double_lock_balance(this_rq, busiest);
1793}
1794
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001795static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1796 __releases(busiest->lock)
1797{
1798 spin_unlock(&busiest->lock);
1799 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1800}
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001801#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001802
1803#ifdef CONFIG_FAIR_GROUP_SCHED
1804static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1805{
Vegard Nossum30432092008-06-27 21:35:50 +02001806#ifdef CONFIG_SMP
Ingo Molnar34e83e82008-06-27 15:42:36 +02001807 cfs_rq->shares = shares;
1808#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001809}
1810#endif
1811
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02001812static void calc_load_account_active(struct rq *this_rq);
1813
Ingo Molnardd41f592007-07-09 18:51:59 +02001814#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001815#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001816#include "sched_fair.c"
1817#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001818#ifdef CONFIG_SCHED_DEBUG
1819# include "sched_debug.c"
1820#endif
1821
1822#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb62008-06-04 15:04:05 -04001823#define for_each_class(class) \
1824 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001825
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001826static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001827{
1828 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001829}
1830
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001831static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001832{
1833 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001834}
1835
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001836static void set_load_weight(struct task_struct *p)
1837{
1838 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001839 p->se.load.weight = prio_to_weight[0] * 2;
1840 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1841 return;
1842 }
1843
1844 /*
1845 * SCHED_IDLE tasks get minimal weight:
1846 */
1847 if (p->policy == SCHED_IDLE) {
1848 p->se.load.weight = WEIGHT_IDLEPRIO;
1849 p->se.load.inv_weight = WMULT_IDLEPRIO;
1850 return;
1851 }
1852
1853 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1854 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001855}
1856
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001857static void update_avg(u64 *avg, u64 sample)
1858{
1859 s64 diff = sample - *avg;
1860 *avg += diff >> 3;
1861}
1862
Ingo Molnar8159f872007-08-09 11:16:49 +02001863static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001864{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001865 if (wakeup)
1866 p->se.start_runtime = p->se.sum_exec_runtime;
1867
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001868 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001869 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001870 p->se.on_rq = 1;
1871}
1872
Ingo Molnar69be72c2007-08-09 11:16:49 +02001873static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001874{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001875 if (sleep) {
1876 if (p->se.last_wakeup) {
1877 update_avg(&p->se.avg_overlap,
1878 p->se.sum_exec_runtime - p->se.last_wakeup);
1879 p->se.last_wakeup = 0;
1880 } else {
1881 update_avg(&p->se.avg_wakeup,
1882 sysctl_sched_wakeup_granularity);
1883 }
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001884 }
1885
Ankita Garg46ac22b2008-07-01 14:30:06 +05301886 sched_info_dequeued(p);
Ingo Molnarf02231e2007-08-09 11:16:48 +02001887 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001888 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001889}
1890
1891/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001892 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001893 */
Ingo Molnar14531182007-07-09 18:51:59 +02001894static inline int __normal_prio(struct task_struct *p)
1895{
Ingo Molnardd41f592007-07-09 18:51:59 +02001896 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001897}
1898
1899/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001900 * Calculate the expected normal priority: i.e. priority
1901 * without taking RT-inheritance into account. Might be
1902 * boosted by interactivity modifiers. Changes upon fork,
1903 * setprio syscalls, and whenever the interactivity
1904 * estimator recalculates.
1905 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001906static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001907{
1908 int prio;
1909
Ingo Molnare05606d2007-07-09 18:51:59 +02001910 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001911 prio = MAX_RT_PRIO-1 - p->rt_priority;
1912 else
1913 prio = __normal_prio(p);
1914 return prio;
1915}
1916
1917/*
1918 * Calculate the current priority, i.e. the priority
1919 * taken into account by the scheduler. This value might
1920 * be boosted by RT tasks, or might be boosted by
1921 * interactivity modifiers. Will be RT if the task got
1922 * RT-boosted. If not then it returns p->normal_prio.
1923 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001924static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001925{
1926 p->normal_prio = normal_prio(p);
1927 /*
1928 * If we are RT tasks or we were boosted to RT priority,
1929 * keep the priority unchanged. Otherwise, update priority
1930 * to the normal priority:
1931 */
1932 if (!rt_prio(p->prio))
1933 return p->normal_prio;
1934 return p->prio;
1935}
1936
1937/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001938 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001939 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001940static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001941{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001942 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001943 rq->nr_uninterruptible--;
1944
Ingo Molnar8159f872007-08-09 11:16:49 +02001945 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001946 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001947}
1948
1949/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001950 * deactivate_task - remove a task from the runqueue.
1951 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001952static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001953{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001954 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001955 rq->nr_uninterruptible++;
1956
Ingo Molnar69be72c2007-08-09 11:16:49 +02001957 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001958 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001959}
1960
Linus Torvalds1da177e2005-04-16 15:20:36 -07001961/**
1962 * task_curr - is this task currently executing on a CPU?
1963 * @p: the task in question.
1964 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001965inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001966{
1967 return cpu_curr(task_cpu(p)) == p;
1968}
1969
Ingo Molnardd41f592007-07-09 18:51:59 +02001970static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1971{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001972 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001973#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001974 /*
1975 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1976 * successfuly executed on another CPU. We must ensure that updates of
1977 * per-task data have been completed by this moment.
1978 */
1979 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001980 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001981#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001982}
1983
Steven Rostedtcb469842008-01-25 21:08:22 +01001984static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1985 const struct sched_class *prev_class,
1986 int oldprio, int running)
1987{
1988 if (prev_class != p->sched_class) {
1989 if (prev_class->switched_from)
1990 prev_class->switched_from(rq, p, running);
1991 p->sched_class->switched_to(rq, p, running);
1992 } else
1993 p->sched_class->prio_changed(rq, p, oldprio, running);
1994}
1995
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01001996/**
1997 * kthread_bind - bind a just-created kthread to a cpu.
Randy Dunlap968c8642009-11-06 15:31:08 -08001998 * @p: thread created by kthread_create().
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01001999 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2000 *
2001 * Description: This function is equivalent to set_cpus_allowed(),
2002 * except that @cpu doesn't need to be online, and the thread must be
2003 * stopped (i.e., just returned from kthread_create()).
2004 *
2005 * Function lives here instead of kthread.c because it messes with
2006 * scheduler internals which require locking.
2007 */
2008void kthread_bind(struct task_struct *p, unsigned int cpu)
2009{
2010 struct rq *rq = cpu_rq(cpu);
2011 unsigned long flags;
2012
2013 /* Must have done schedule() in kthread() before we set_task_cpu */
2014 if (!wait_task_inactive(p, TASK_UNINTERRUPTIBLE)) {
2015 WARN_ON(1);
2016 return;
2017 }
2018
2019 spin_lock_irqsave(&rq->lock, flags);
Mike Galbraith055a0082009-11-12 11:07:44 +01002020 update_rq_clock(rq);
Mike Galbraithb84ff7d2009-10-29 11:48:30 +01002021 set_task_cpu(p, cpu);
2022 p->cpus_allowed = cpumask_of_cpu(cpu);
2023 p->rt.nr_cpus_allowed = 1;
2024 p->flags |= PF_THREAD_BOUND;
2025 spin_unlock_irqrestore(&rq->lock, flags);
2026}
2027EXPORT_SYMBOL(kthread_bind);
2028
Linus Torvalds1da177e2005-04-16 15:20:36 -07002029#ifdef CONFIG_SMP
Ingo Molnarcc367732007-10-15 17:00:18 +02002030/*
2031 * Is this task likely cache-hot:
2032 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002033static int
Ingo Molnarcc367732007-10-15 17:00:18 +02002034task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2035{
2036 s64 delta;
2037
Ingo Molnarf540a602008-03-15 17:10:34 +01002038 /*
2039 * Buddy candidates are cache hot:
2040 */
Mike Galbraithf685cea2009-10-23 23:09:22 +02002041 if (sched_feat(CACHE_HOT_BUDDY) && this_rq()->nr_running &&
Peter Zijlstra47932412008-11-04 21:25:09 +01002042 (&p->se == cfs_rq_of(&p->se)->next ||
2043 &p->se == cfs_rq_of(&p->se)->last))
Ingo Molnarf540a602008-03-15 17:10:34 +01002044 return 1;
2045
Ingo Molnarcc367732007-10-15 17:00:18 +02002046 if (p->sched_class != &fair_sched_class)
2047 return 0;
2048
Ingo Molnar6bc16652007-10-15 17:00:18 +02002049 if (sysctl_sched_migration_cost == -1)
2050 return 1;
2051 if (sysctl_sched_migration_cost == 0)
2052 return 0;
2053
Ingo Molnarcc367732007-10-15 17:00:18 +02002054 delta = now - p->se.exec_start;
2055
2056 return delta < (s64)sysctl_sched_migration_cost;
2057}
2058
2059
Ingo Molnardd41f592007-07-09 18:51:59 +02002060void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02002061{
Ingo Molnardd41f592007-07-09 18:51:59 +02002062 int old_cpu = task_cpu(p);
2063 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002064 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2065 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02002066 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002067
2068 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002069
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08002070 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01002071
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002072#ifdef CONFIG_SCHEDSTATS
2073 if (p->se.wait_start)
2074 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02002075 if (p->se.sleep_start)
2076 p->se.sleep_start -= clock_offset;
2077 if (p->se.block_start)
2078 p->se.block_start -= clock_offset;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002079#endif
Ingo Molnarcc367732007-10-15 17:00:18 +02002080 if (old_cpu != new_cpu) {
Ingo Molnar6c594c22008-12-14 12:34:15 +01002081 p->se.nr_migrations++;
2082#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002083 if (task_hot(p, old_rq->clock, NULL))
2084 schedstat_inc(p, se.nr_forced2_migrations);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002085#endif
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002086 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS,
Peter Zijlstrae5289d42009-06-19 13:22:51 +02002087 1, 1, NULL, 0);
Ingo Molnar6c594c22008-12-14 12:34:15 +01002088 }
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02002089 p->se.vruntime -= old_cfsrq->min_vruntime -
2090 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02002091
2092 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02002093}
2094
Ingo Molnar70b97a72006-07-03 00:25:42 -07002095struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002096 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002097
Ingo Molnar36c8b582006-07-03 00:25:41 -07002098 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002099 int dest_cpu;
2100
Linus Torvalds1da177e2005-04-16 15:20:36 -07002101 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002102};
Linus Torvalds1da177e2005-04-16 15:20:36 -07002103
2104/*
2105 * The task's runqueue lock must be held.
2106 * Returns true if you have to wait for migration thread.
2107 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002108static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002109migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002110{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002111 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002112
2113 /*
2114 * If the task is not on a runqueue (and not running), then
2115 * it is sufficient to simply update the task's cpu field.
2116 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002117 if (!p->se.on_rq && !task_running(rq, p)) {
Mike Galbraith055a0082009-11-12 11:07:44 +01002118 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002119 set_task_cpu(p, dest_cpu);
2120 return 0;
2121 }
2122
2123 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002124 req->task = p;
2125 req->dest_cpu = dest_cpu;
2126 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002127
Linus Torvalds1da177e2005-04-16 15:20:36 -07002128 return 1;
2129}
2130
2131/*
Markus Metzgera26b89f2009-04-03 16:43:34 +02002132 * wait_task_context_switch - wait for a thread to complete at least one
2133 * context switch.
2134 *
2135 * @p must not be current.
2136 */
2137void wait_task_context_switch(struct task_struct *p)
2138{
2139 unsigned long nvcsw, nivcsw, flags;
2140 int running;
2141 struct rq *rq;
2142
2143 nvcsw = p->nvcsw;
2144 nivcsw = p->nivcsw;
2145 for (;;) {
2146 /*
2147 * The runqueue is assigned before the actual context
2148 * switch. We need to take the runqueue lock.
2149 *
2150 * We could check initially without the lock but it is
2151 * very likely that we need to take the lock in every
2152 * iteration.
2153 */
2154 rq = task_rq_lock(p, &flags);
2155 running = task_running(rq, p);
2156 task_rq_unlock(rq, &flags);
2157
2158 if (likely(!running))
2159 break;
2160 /*
2161 * The switch count is incremented before the actual
2162 * context switch. We thus wait for two switches to be
2163 * sure at least one completed.
2164 */
2165 if ((p->nvcsw - nvcsw) > 1)
2166 break;
2167 if ((p->nivcsw - nivcsw) > 1)
2168 break;
2169
2170 cpu_relax();
2171 }
2172}
2173
2174/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002175 * wait_task_inactive - wait for a thread to unschedule.
2176 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07002177 * If @match_state is nonzero, it's the @p->state value just checked and
2178 * not expected to change. If it changes, i.e. @p might have woken up,
2179 * then return zero. When we succeed in waiting for @p to be off its CPU,
2180 * we return a positive number (its total switch count). If a second call
2181 * a short while later returns the same number, the caller can be sure that
2182 * @p has remained unscheduled the whole time.
2183 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002184 * The caller must ensure that the task *will* unschedule sometime soon,
2185 * else this function might spin for a *long* time. This function can't
2186 * be called with interrupts off, or it may introduce deadlock with
2187 * smp_call_function() if an IPI is sent by the same process we are
2188 * waiting to become inactive.
2189 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002190unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002191{
2192 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002193 int running, on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002194 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002195 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002196
Andi Kleen3a5c3592007-10-15 17:00:14 +02002197 for (;;) {
2198 /*
2199 * We do the initial early heuristics without holding
2200 * any task-queue locks at all. We'll only try to get
2201 * the runqueue lock when things look like they will
2202 * work out!
2203 */
2204 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002205
Andi Kleen3a5c3592007-10-15 17:00:14 +02002206 /*
2207 * If the task is actively running on another CPU
2208 * still, just relax and busy-wait without holding
2209 * any locks.
2210 *
2211 * NOTE! Since we don't hold any locks, it's not
2212 * even sure that "rq" stays as the right runqueue!
2213 * But we don't care, since "task_running()" will
2214 * return false if the runqueue has changed and p
2215 * is actually now running somewhere else!
2216 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002217 while (task_running(rq, p)) {
2218 if (match_state && unlikely(p->state != match_state))
2219 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02002220 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07002221 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002222
Andi Kleen3a5c3592007-10-15 17:00:14 +02002223 /*
2224 * Ok, time to look more closely! We need the rq
2225 * lock now, to be *sure*. If we're wrong, we'll
2226 * just go back and repeat.
2227 */
2228 rq = task_rq_lock(p, &flags);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002229 trace_sched_wait_task(rq, p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02002230 running = task_running(rq, p);
2231 on_rq = p->se.on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002232 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07002233 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07002234 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Andi Kleen3a5c3592007-10-15 17:00:14 +02002235 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002236
Andi Kleen3a5c3592007-10-15 17:00:14 +02002237 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07002238 * If it changed from the expected state, bail out now.
2239 */
2240 if (unlikely(!ncsw))
2241 break;
2242
2243 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02002244 * Was it really running after all now that we
2245 * checked with the proper locks actually held?
2246 *
2247 * Oops. Go back and try again..
2248 */
2249 if (unlikely(running)) {
2250 cpu_relax();
2251 continue;
2252 }
2253
2254 /*
2255 * It's not enough that it's not actively running,
2256 * it must be off the runqueue _entirely_, and not
2257 * preempted!
2258 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00002259 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02002260 * running right now), it's preempted, and we should
2261 * yield - it could be a while.
2262 */
2263 if (unlikely(on_rq)) {
2264 schedule_timeout_uninterruptible(1);
2265 continue;
2266 }
2267
2268 /*
2269 * Ahh, all good. It wasn't running, and it wasn't
2270 * runnable, which means that it will never become
2271 * running in the future either. We're all done!
2272 */
2273 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002274 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07002275
2276 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002277}
2278
2279/***
2280 * kick_process - kick a running thread to enter/exit the kernel
2281 * @p: the to-be-kicked thread
2282 *
2283 * Cause a process which is running on another CPU to enter
2284 * kernel-mode, without any delay. (to get signals handled.)
2285 *
2286 * NOTE: this function doesnt have to take the runqueue lock,
2287 * because all it wants to ensure is that the remote task enters
2288 * the kernel. If the IPI races and the task has been migrated
2289 * to another CPU then no harm is done and the purpose has been
2290 * achieved as well.
2291 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002292void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002293{
2294 int cpu;
2295
2296 preempt_disable();
2297 cpu = task_cpu(p);
2298 if ((cpu != smp_processor_id()) && task_curr(p))
2299 smp_send_reschedule(cpu);
2300 preempt_enable();
2301}
Rusty Russellb43e3522009-06-12 22:27:00 -06002302EXPORT_SYMBOL_GPL(kick_process);
Nick Piggin476d1392005-06-25 14:57:29 -07002303#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002304
Thomas Gleixner0793a612008-12-04 20:12:29 +01002305/**
2306 * task_oncpu_function_call - call a function on the cpu on which a task runs
2307 * @p: the task to evaluate
2308 * @func: the function to be called
2309 * @info: the function call argument
2310 *
2311 * Calls the function @func when the task is currently running. This might
2312 * be on the current CPU, which just calls the function directly
2313 */
2314void task_oncpu_function_call(struct task_struct *p,
2315 void (*func) (void *info), void *info)
2316{
2317 int cpu;
2318
2319 preempt_disable();
2320 cpu = task_cpu(p);
2321 if (task_curr(p))
2322 smp_call_function_single(cpu, func, info, 1);
2323 preempt_enable();
2324}
2325
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326/***
2327 * try_to_wake_up - wake up a thread
2328 * @p: the to-be-woken-up thread
2329 * @state: the mask of task states that can be woken
2330 * @sync: do a synchronous wakeup?
2331 *
2332 * Put it on the run-queue if it's not already there. The "current"
2333 * thread is always on the run-queue (except when the actual
2334 * re-schedule is in progress), and as such you're allowed to do
2335 * the simpler "current->state = TASK_RUNNING" to mark yourself
2336 * runnable without the overhead of this.
2337 *
2338 * returns failure only if the task is already active.
2339 */
Peter Zijlstra7d478722009-09-14 19:55:44 +02002340static int try_to_wake_up(struct task_struct *p, unsigned int state,
2341 int wake_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002342{
Ingo Molnarcc367732007-10-15 17:00:18 +02002343 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002344 unsigned long flags;
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002345 struct rq *rq, *orig_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002346
Ingo Molnarb85d0662008-03-16 20:03:22 +01002347 if (!sched_feat(SYNC_WAKEUPS))
Peter Zijlstra7d478722009-09-14 19:55:44 +02002348 wake_flags &= ~WF_SYNC;
Ingo Molnarb85d0662008-03-16 20:03:22 +01002349
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002350 this_cpu = get_cpu();
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002351
Linus Torvalds04e2f172008-02-23 18:05:03 -08002352 smp_wmb();
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002353 rq = orig_rq = task_rq_lock(p, &flags);
Mike Galbraith03e89e42008-12-16 08:45:30 +01002354 update_rq_clock(rq);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002355 if (!(p->state & state))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002356 goto out;
2357
Ingo Molnardd41f592007-07-09 18:51:59 +02002358 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002359 goto out_running;
2360
2361 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002362 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002363
2364#ifdef CONFIG_SMP
2365 if (unlikely(task_running(rq, p)))
2366 goto out_activate;
2367
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002368 /*
2369 * In order to handle concurrent wakeups and release the rq->lock
2370 * we put the task in TASK_WAKING state.
Ingo Molnareb240732009-09-16 21:09:13 +02002371 *
2372 * First fix up the nr_uninterruptible count:
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002373 */
Ingo Molnareb240732009-09-16 21:09:13 +02002374 if (task_contributes_to_load(p))
2375 rq->nr_uninterruptible--;
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002376 p->state = TASK_WAKING;
2377 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378
Peter Zijlstra7d478722009-09-14 19:55:44 +02002379 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, wake_flags);
Mike Galbraith055a0082009-11-12 11:07:44 +01002380 if (cpu != orig_cpu) {
2381 local_irq_save(flags);
2382 rq = cpu_rq(cpu);
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002383 update_rq_clock(rq);
Mike Galbraith055a0082009-11-12 11:07:44 +01002384 set_task_cpu(p, cpu);
2385 local_irq_restore(flags);
2386 }
2387 rq = task_rq_lock(p, &flags);
Mike Galbraithf5dc3752009-10-09 08:35:03 +02002388
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002389 WARN_ON(p->state != TASK_WAKING);
2390 cpu = task_cpu(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002391
Gregory Haskinse7693a32008-01-25 21:08:09 +01002392#ifdef CONFIG_SCHEDSTATS
2393 schedstat_inc(rq, ttwu_count);
2394 if (cpu == this_cpu)
2395 schedstat_inc(rq, ttwu_local);
2396 else {
2397 struct sched_domain *sd;
2398 for_each_domain(this_cpu, sd) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302399 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Gregory Haskinse7693a32008-01-25 21:08:09 +01002400 schedstat_inc(sd, ttwu_wake_remote);
2401 break;
2402 }
2403 }
2404 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002405#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002406
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407out_activate:
2408#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002409 schedstat_inc(p, se.nr_wakeups);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002410 if (wake_flags & WF_SYNC)
Ingo Molnarcc367732007-10-15 17:00:18 +02002411 schedstat_inc(p, se.nr_wakeups_sync);
2412 if (orig_cpu != cpu)
2413 schedstat_inc(p, se.nr_wakeups_migrate);
2414 if (cpu == this_cpu)
2415 schedstat_inc(p, se.nr_wakeups_local);
2416 else
2417 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnardd41f592007-07-09 18:51:59 +02002418 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419 success = 1;
2420
Peter Zijlstra831451a2009-01-14 12:39:18 +01002421 /*
2422 * Only attribute actual wakeups done by this task.
2423 */
2424 if (!in_interrupt()) {
2425 struct sched_entity *se = &current->se;
2426 u64 sample = se->sum_exec_runtime;
2427
2428 if (se->last_wakeup)
2429 sample -= se->last_wakeup;
2430 else
2431 sample -= se->start_runtime;
2432 update_avg(&se->avg_wakeup, sample);
2433
2434 se->last_wakeup = se->sum_exec_runtime;
2435 }
2436
Linus Torvalds1da177e2005-04-16 15:20:36 -07002437out_running:
Peter Zijlstra468a15b2008-12-16 08:07:03 +01002438 trace_sched_wakeup(rq, p, success);
Peter Zijlstra7d478722009-09-14 19:55:44 +02002439 check_preempt_curr(rq, p, wake_flags);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002440
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002442#ifdef CONFIG_SMP
2443 if (p->sched_class->task_wake_up)
2444 p->sched_class->task_wake_up(rq, p);
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01002445
2446 if (unlikely(rq->idle_stamp)) {
2447 u64 delta = rq->clock - rq->idle_stamp;
2448 u64 max = 2*sysctl_sched_migration_cost;
2449
2450 if (delta > max)
2451 rq->avg_idle = max;
2452 else
2453 update_avg(&rq->avg_idle, delta);
2454 rq->idle_stamp = 0;
2455 }
Steven Rostedt9a897c52008-01-25 21:08:22 +01002456#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002457out:
2458 task_rq_unlock(rq, &flags);
Peter Zijlstrae9c84312009-09-15 14:43:03 +02002459 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002460
2461 return success;
2462}
2463
David Howells50fa6102009-04-28 15:01:38 +01002464/**
2465 * wake_up_process - Wake up a specific process
2466 * @p: The process to be woken up.
2467 *
2468 * Attempt to wake up the nominated process and move it to the set of runnable
2469 * processes. Returns 1 if the process was woken up, 0 if it was already
2470 * running.
2471 *
2472 * It may be assumed that this function implies a write memory barrier before
2473 * changing the task state if and only if any tasks are woken up.
2474 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002475int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002476{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002477 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002478}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002479EXPORT_SYMBOL(wake_up_process);
2480
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002481int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002482{
2483 return try_to_wake_up(p, state, 0);
2484}
2485
Linus Torvalds1da177e2005-04-16 15:20:36 -07002486/*
2487 * Perform scheduler related setup for a newly forked process p.
2488 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002489 *
2490 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002491 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002492static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002493{
Ingo Molnardd41f592007-07-09 18:51:59 +02002494 p->se.exec_start = 0;
2495 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002496 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002497 p->se.nr_migrations = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002498 p->se.last_wakeup = 0;
2499 p->se.avg_overlap = 0;
Peter Zijlstra831451a2009-01-14 12:39:18 +01002500 p->se.start_runtime = 0;
2501 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02002502 p->se.avg_running = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002503
2504#ifdef CONFIG_SCHEDSTATS
Lucas De Marchi77935272009-07-09 13:57:20 +02002505 p->se.wait_start = 0;
2506 p->se.wait_max = 0;
2507 p->se.wait_count = 0;
2508 p->se.wait_sum = 0;
2509
2510 p->se.sleep_start = 0;
2511 p->se.sleep_max = 0;
2512 p->se.sum_sleep_runtime = 0;
2513
2514 p->se.block_start = 0;
2515 p->se.block_max = 0;
2516 p->se.exec_max = 0;
2517 p->se.slice_max = 0;
2518
2519 p->se.nr_migrations_cold = 0;
2520 p->se.nr_failed_migrations_affine = 0;
2521 p->se.nr_failed_migrations_running = 0;
2522 p->se.nr_failed_migrations_hot = 0;
2523 p->se.nr_forced_migrations = 0;
2524 p->se.nr_forced2_migrations = 0;
2525
2526 p->se.nr_wakeups = 0;
2527 p->se.nr_wakeups_sync = 0;
2528 p->se.nr_wakeups_migrate = 0;
2529 p->se.nr_wakeups_local = 0;
2530 p->se.nr_wakeups_remote = 0;
2531 p->se.nr_wakeups_affine = 0;
2532 p->se.nr_wakeups_affine_attempts = 0;
2533 p->se.nr_wakeups_passive = 0;
2534 p->se.nr_wakeups_idle = 0;
2535
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002536#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002537
Peter Zijlstrafa717062008-01-25 21:08:27 +01002538 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002539 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002540 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002541
Avi Kivitye107be32007-07-26 13:40:43 +02002542#ifdef CONFIG_PREEMPT_NOTIFIERS
2543 INIT_HLIST_HEAD(&p->preempt_notifiers);
2544#endif
2545
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 /*
2547 * We mark the process as running here, but have not actually
2548 * inserted it onto the runqueue yet. This guarantees that
2549 * nobody will actually run it, and a signal or other external
2550 * event cannot wake it up and insert it on the runqueue either.
2551 */
2552 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002553}
2554
2555/*
2556 * fork()/clone()-time setup:
2557 */
2558void sched_fork(struct task_struct *p, int clone_flags)
2559{
2560 int cpu = get_cpu();
Mike Galbraith055a0082009-11-12 11:07:44 +01002561 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002562
2563 __sched_fork(p);
2564
Ingo Molnarb29739f2006-06-27 02:54:51 -07002565 /*
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002566 * Revert to default priority/policy on fork if requested.
2567 */
2568 if (unlikely(p->sched_reset_on_fork)) {
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002569 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR) {
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002570 p->policy = SCHED_NORMAL;
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002571 p->normal_prio = p->static_prio;
2572 }
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002573
Mike Galbraith6c697bdf2009-06-17 10:48:02 +02002574 if (PRIO_TO_NICE(p->static_prio) < 0) {
2575 p->static_prio = NICE_TO_PRIO(0);
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002576 p->normal_prio = p->static_prio;
Mike Galbraith6c697bdf2009-06-17 10:48:02 +02002577 set_load_weight(p);
2578 }
2579
Mike Galbraithb9dc29e2009-06-17 10:46:01 +02002580 /*
2581 * We don't need the reset flag anymore after the fork. It has
2582 * fulfilled its duty:
2583 */
2584 p->sched_reset_on_fork = 0;
2585 }
Lennart Poetteringca94c442009-06-15 17:17:47 +02002586
Peter Williamsf83f9ac2009-09-24 06:47:10 +00002587 /*
2588 * Make sure we do not leak PI boosting priority to the child.
2589 */
2590 p->prio = current->normal_prio;
2591
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002592 if (!rt_prio(p->prio))
2593 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002594
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002595#ifdef CONFIG_SMP
2596 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2597#endif
Mike Galbraith055a0082009-11-12 11:07:44 +01002598 local_irq_save(flags);
2599 update_rq_clock(cpu_rq(cpu));
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002600 set_task_cpu(p, cpu);
Mike Galbraith055a0082009-11-12 11:07:44 +01002601 local_irq_restore(flags);
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02002602
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002603#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002604 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002605 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002606#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002607#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002608 p->oncpu = 0;
2609#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002611 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002612 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002613#endif
Gregory Haskins917b6272008-12-29 09:39:53 -05002614 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2615
Nick Piggin476d1392005-06-25 14:57:29 -07002616 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002617}
2618
2619/*
2620 * wake_up_new_task - wake up a newly created task for the first time.
2621 *
2622 * This function will do some initial scheduler statistics housekeeping
2623 * that must be done for every newly created context, then puts the task
2624 * on the runqueue and wakes it.
2625 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002626void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002627{
2628 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002629 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630
2631 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002632 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002633 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002634
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002635 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002636 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002638 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002639 * Let the scheduling class do new task startup
2640 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002641 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002642 p->sched_class->task_new(rq, p);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002643 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002644 }
Ingo Molnarc71dd422008-12-19 01:09:51 +01002645 trace_sched_wakeup_new(rq, p, 1);
Peter Zijlstraa7558e02009-09-14 20:02:34 +02002646 check_preempt_curr(rq, p, WF_FORK);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002647#ifdef CONFIG_SMP
2648 if (p->sched_class->task_wake_up)
2649 p->sched_class->task_wake_up(rq, p);
2650#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002651 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652}
2653
Avi Kivitye107be32007-07-26 13:40:43 +02002654#ifdef CONFIG_PREEMPT_NOTIFIERS
2655
2656/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00002657 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07002658 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002659 */
2660void preempt_notifier_register(struct preempt_notifier *notifier)
2661{
2662 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2663}
2664EXPORT_SYMBOL_GPL(preempt_notifier_register);
2665
2666/**
2667 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002668 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002669 *
2670 * This is safe to call from within a preemption notifier.
2671 */
2672void preempt_notifier_unregister(struct preempt_notifier *notifier)
2673{
2674 hlist_del(&notifier->link);
2675}
2676EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2677
2678static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2679{
2680 struct preempt_notifier *notifier;
2681 struct hlist_node *node;
2682
2683 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2684 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2685}
2686
2687static void
2688fire_sched_out_preempt_notifiers(struct task_struct *curr,
2689 struct task_struct *next)
2690{
2691 struct preempt_notifier *notifier;
2692 struct hlist_node *node;
2693
2694 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2695 notifier->ops->sched_out(notifier, next);
2696}
2697
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002698#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002699
2700static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2701{
2702}
2703
2704static void
2705fire_sched_out_preempt_notifiers(struct task_struct *curr,
2706 struct task_struct *next)
2707{
2708}
2709
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002710#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002711
Linus Torvalds1da177e2005-04-16 15:20:36 -07002712/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002713 * prepare_task_switch - prepare to switch tasks
2714 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002715 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002716 * @next: the task we are going to switch to.
2717 *
2718 * This is called with the rq lock held and interrupts off. It must
2719 * be paired with a subsequent finish_task_switch after the context
2720 * switch.
2721 *
2722 * prepare_task_switch sets up locking and calls architecture specific
2723 * hooks.
2724 */
Avi Kivitye107be32007-07-26 13:40:43 +02002725static inline void
2726prepare_task_switch(struct rq *rq, struct task_struct *prev,
2727 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002728{
Avi Kivitye107be32007-07-26 13:40:43 +02002729 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002730 prepare_lock_switch(rq, next);
2731 prepare_arch_switch(next);
2732}
2733
2734/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002735 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002736 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002737 * @prev: the thread we just switched away from.
2738 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002739 * finish_task_switch must be called after the context switch, paired
2740 * with a prepare_task_switch call before the context switch.
2741 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2742 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002743 *
2744 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002745 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002746 * with the lock held can cause deadlocks; see schedule() for
2747 * details.)
2748 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002749static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002750 __releases(rq->lock)
2751{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002752 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002753 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002754
2755 rq->prev_mm = NULL;
2756
2757 /*
2758 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002759 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002760 * schedule one last time. The schedule call will never return, and
2761 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002762 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002763 * still held, otherwise prev could be scheduled on another cpu, die
2764 * there before we look at prev->state, and then the reference would
2765 * be dropped twice.
2766 * Manfred Spraul <manfred@colorfullife.com>
2767 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002768 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002769 finish_arch_switch(prev);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02002770 perf_event_task_sched_in(current, cpu_of(rq));
Nick Piggin4866cde2005-06-25 14:57:23 -07002771 finish_lock_switch(rq, prev);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002772
Avi Kivitye107be32007-07-26 13:40:43 +02002773 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 if (mm)
2775 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002776 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002777 /*
2778 * Remove function-return probe instances associated with this
2779 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002780 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002781 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002783 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002784}
2785
Gregory Haskins3f029d32009-07-29 11:08:47 -04002786#ifdef CONFIG_SMP
2787
2788/* assumes rq->lock is held */
2789static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2790{
2791 if (prev->sched_class->pre_schedule)
2792 prev->sched_class->pre_schedule(rq, prev);
2793}
2794
2795/* rq->lock is NOT held, but preemption is disabled */
2796static inline void post_schedule(struct rq *rq)
2797{
2798 if (rq->post_schedule) {
2799 unsigned long flags;
2800
2801 spin_lock_irqsave(&rq->lock, flags);
2802 if (rq->curr->sched_class->post_schedule)
2803 rq->curr->sched_class->post_schedule(rq);
2804 spin_unlock_irqrestore(&rq->lock, flags);
2805
2806 rq->post_schedule = 0;
2807 }
2808}
2809
2810#else
2811
2812static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2813{
2814}
2815
2816static inline void post_schedule(struct rq *rq)
2817{
2818}
2819
2820#endif
2821
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822/**
2823 * schedule_tail - first thing a freshly forked thread must call.
2824 * @prev: the thread we just switched away from.
2825 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002826asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002827 __releases(rq->lock)
2828{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002829 struct rq *rq = this_rq();
2830
Nick Piggin4866cde2005-06-25 14:57:23 -07002831 finish_task_switch(rq, prev);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002832
Gregory Haskins3f029d32009-07-29 11:08:47 -04002833 /*
2834 * FIXME: do we need to worry about rq being invalidated by the
2835 * task_switch?
2836 */
2837 post_schedule(rq);
Steven Rostedtda19ab52009-07-29 00:21:22 -04002838
Nick Piggin4866cde2005-06-25 14:57:23 -07002839#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2840 /* In this case, finish_task_switch does not reenable preemption */
2841 preempt_enable();
2842#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002844 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845}
2846
2847/*
2848 * context_switch - switch to the new MM and the new
2849 * thread's register state.
2850 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002851static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002852context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002853 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002854{
Ingo Molnardd41f592007-07-09 18:51:59 +02002855 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002856
Avi Kivitye107be32007-07-26 13:40:43 +02002857 prepare_task_switch(rq, prev, next);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002858 trace_sched_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002859 mm = next->mm;
2860 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002861 /*
2862 * For paravirt, this is coupled with an exit in switch_to to
2863 * combine the page table reload and the switch backend into
2864 * one hypercall.
2865 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08002866 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01002867
Tim Blechmann710390d2009-11-24 11:55:27 +01002868 if (likely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002869 next->active_mm = oldmm;
2870 atomic_inc(&oldmm->mm_count);
2871 enter_lazy_tlb(oldmm, next);
2872 } else
2873 switch_mm(oldmm, mm, next);
2874
Tim Blechmann710390d2009-11-24 11:55:27 +01002875 if (likely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877 rq->prev_mm = oldmm;
2878 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002879 /*
2880 * Since the runqueue lock will be released by the next
2881 * task (which is an invalid locking op but in the case
2882 * of the scheduler it's an obvious special-case), so we
2883 * do an early lockdep release here:
2884 */
2885#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002886 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002887#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888
2889 /* Here we just switch the register state and the stack. */
2890 switch_to(prev, next, prev);
2891
Ingo Molnardd41f592007-07-09 18:51:59 +02002892 barrier();
2893 /*
2894 * this_rq must be evaluated again because prev may have moved
2895 * CPUs since it called schedule(), thus the 'rq' on its stack
2896 * frame will be invalid.
2897 */
2898 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002899}
2900
2901/*
2902 * nr_running, nr_uninterruptible and nr_context_switches:
2903 *
2904 * externally visible scheduler statistics: current number of runnable
2905 * threads, current number of uninterruptible-sleeping threads, total
2906 * number of context switches performed since bootup.
2907 */
2908unsigned long nr_running(void)
2909{
2910 unsigned long i, sum = 0;
2911
2912 for_each_online_cpu(i)
2913 sum += cpu_rq(i)->nr_running;
2914
2915 return sum;
2916}
2917
2918unsigned long nr_uninterruptible(void)
2919{
2920 unsigned long i, sum = 0;
2921
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002922 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002923 sum += cpu_rq(i)->nr_uninterruptible;
2924
2925 /*
2926 * Since we read the counters lockless, it might be slightly
2927 * inaccurate. Do not allow it to go below zero though:
2928 */
2929 if (unlikely((long)sum < 0))
2930 sum = 0;
2931
2932 return sum;
2933}
2934
2935unsigned long long nr_context_switches(void)
2936{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002937 int i;
2938 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002939
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002940 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002941 sum += cpu_rq(i)->nr_switches;
2942
2943 return sum;
2944}
2945
2946unsigned long nr_iowait(void)
2947{
2948 unsigned long i, sum = 0;
2949
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002950 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002951 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2952
2953 return sum;
2954}
2955
Arjan van de Ven69d25872009-09-21 17:04:08 -07002956unsigned long nr_iowait_cpu(void)
2957{
2958 struct rq *this = this_rq();
2959 return atomic_read(&this->nr_iowait);
2960}
2961
2962unsigned long this_cpu_load(void)
2963{
2964 struct rq *this = this_rq();
2965 return this->cpu_load[0];
2966}
2967
2968
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002969/* Variables and functions for calc_load */
2970static atomic_long_t calc_load_tasks;
2971static unsigned long calc_load_update;
2972unsigned long avenrun[3];
2973EXPORT_SYMBOL(avenrun);
2974
Thomas Gleixner2d024942009-05-02 20:08:52 +02002975/**
2976 * get_avenrun - get the load average array
2977 * @loads: pointer to dest load array
2978 * @offset: offset to add
2979 * @shift: shift count to shift the result left
2980 *
2981 * These values are estimates at best, so no need for locking.
2982 */
2983void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2984{
2985 loads[0] = (avenrun[0] + offset) << shift;
2986 loads[1] = (avenrun[1] + offset) << shift;
2987 loads[2] = (avenrun[2] + offset) << shift;
2988}
2989
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002990static unsigned long
2991calc_load(unsigned long load, unsigned long exp, unsigned long active)
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002992{
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002993 load *= exp;
2994 load += active * (FIXED_1 - exp);
2995 return load >> FSHIFT;
2996}
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002997
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002998/*
2999 * calc_load - update the avenrun load estimates 10 ticks after the
3000 * CPUs have updated calc_load_tasks.
3001 */
3002void calc_global_load(void)
3003{
3004 unsigned long upd = calc_load_update + 10;
3005 long active;
3006
3007 if (time_before(jiffies, upd))
3008 return;
3009
3010 active = atomic_long_read(&calc_load_tasks);
3011 active = active > 0 ? active * FIXED_1 : 0;
3012
3013 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3014 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3015 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3016
3017 calc_load_update += LOAD_FREQ;
3018}
3019
3020/*
3021 * Either called from update_cpu_load() or from a cpu going idle
3022 */
3023static void calc_load_account_active(struct rq *this_rq)
3024{
3025 long nr_active, delta;
3026
3027 nr_active = this_rq->nr_running;
3028 nr_active += (long) this_rq->nr_uninterruptible;
3029
3030 if (nr_active != this_rq->calc_load_active) {
3031 delta = nr_active - this_rq->calc_load_active;
3032 this_rq->calc_load_active = nr_active;
3033 atomic_long_add(delta, &calc_load_tasks);
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003034 }
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003035}
3036
Linus Torvalds1da177e2005-04-16 15:20:36 -07003037/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003038 * Update rq->cpu_load[] statistics. This function is usually called every
3039 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07003040 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003041static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003042{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02003043 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02003044 int i, scale;
3045
3046 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02003047
3048 /* Update our load: */
3049 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3050 unsigned long old_load, new_load;
3051
3052 /* scale is effectively 1 << i now, and >> i divides by scale */
3053
3054 old_load = this_rq->cpu_load[i];
3055 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02003056 /*
3057 * Round up the averaging division if load is increasing. This
3058 * prevents us from getting stuck on 9 if the load is 10, for
3059 * example.
3060 */
3061 if (new_load > old_load)
3062 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003063 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3064 }
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02003065
3066 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3067 this_rq->calc_load_update += LOAD_FREQ;
3068 calc_load_account_active(this_rq);
3069 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003070}
3071
Ingo Molnardd41f592007-07-09 18:51:59 +02003072#ifdef CONFIG_SMP
3073
Ingo Molnar48f24c42006-07-03 00:25:40 -07003074/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 * double_rq_lock - safely lock two runqueues
3076 *
3077 * Note this does not disable interrupts like task_rq_lock,
3078 * you need to do so manually before calling.
3079 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003080static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003081 __acquires(rq1->lock)
3082 __acquires(rq2->lock)
3083{
Kirill Korotaev054b9102006-12-10 02:20:11 -08003084 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 if (rq1 == rq2) {
3086 spin_lock(&rq1->lock);
3087 __acquire(rq2->lock); /* Fake it out ;) */
3088 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003089 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090 spin_lock(&rq1->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003091 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003092 } else {
3093 spin_lock(&rq2->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003094 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003095 }
3096 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003097 update_rq_clock(rq1);
3098 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003099}
3100
3101/*
3102 * double_rq_unlock - safely unlock two runqueues
3103 *
3104 * Note this does not restore interrupts like task_rq_unlock,
3105 * you need to do so manually after calling.
3106 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003107static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003108 __releases(rq1->lock)
3109 __releases(rq2->lock)
3110{
3111 spin_unlock(&rq1->lock);
3112 if (rq1 != rq2)
3113 spin_unlock(&rq2->lock);
3114 else
3115 __release(rq2->lock);
3116}
3117
3118/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003119 * If dest_cpu is allowed for this process, migrate the task to it.
3120 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003121 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07003122 * the cpu_allowed mask is restored.
3123 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003124static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003126 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003127 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003128 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003129
3130 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e2008-11-25 02:35:14 +10303131 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
Max Krasnyanskye761b772008-07-15 04:43:49 -07003132 || unlikely(!cpu_active(dest_cpu)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003133 goto out;
3134
3135 /* force the process onto the specified CPU */
3136 if (migrate_task(p, dest_cpu, &req)) {
3137 /* Need to wait for migration thread (might exit: take ref). */
3138 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003139
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 get_task_struct(mt);
3141 task_rq_unlock(rq, &flags);
3142 wake_up_process(mt);
3143 put_task_struct(mt);
3144 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07003145
Linus Torvalds1da177e2005-04-16 15:20:36 -07003146 return;
3147 }
3148out:
3149 task_rq_unlock(rq, &flags);
3150}
3151
3152/*
Nick Piggin476d1392005-06-25 14:57:29 -07003153 * sched_exec - execve() is a valuable balancing opportunity, because at
3154 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003155 */
3156void sched_exec(void)
3157{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 int new_cpu, this_cpu = get_cpu();
Peter Zijlstra5f3edc12009-09-10 13:42:00 +02003159 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003161 if (new_cpu != this_cpu)
3162 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163}
3164
3165/*
3166 * pull_task - move a task from a remote runqueue to the local runqueue.
3167 * Both runqueues must be locked.
3168 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003169static void pull_task(struct rq *src_rq, struct task_struct *p,
3170 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003171{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003172 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003173 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003174 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003175 /*
3176 * Note that idle threads have a prio of MAX_PRIO, for this test
3177 * to be always true for them.
3178 */
Peter Zijlstra15afe092008-09-20 23:38:02 +02003179 check_preempt_curr(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003180}
3181
3182/*
3183 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3184 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08003185static
Ingo Molnar70b97a72006-07-03 00:25:42 -07003186int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003187 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003188 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189{
Luis Henriques708dc512009-03-16 19:59:02 +00003190 int tsk_cache_hot = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003191 /*
3192 * We do not migrate tasks that are:
3193 * 1) running (obviously), or
3194 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3195 * 3) are cache-hot on their current CPU.
3196 */
Rusty Russell96f874e2008-11-25 02:35:14 +10303197 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003198 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003199 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003200 }
Nick Piggin81026792005-06-25 14:57:07 -07003201 *all_pinned = 0;
3202
Ingo Molnarcc367732007-10-15 17:00:18 +02003203 if (task_running(rq, p)) {
3204 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07003205 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003206 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003207
Ingo Molnarda84d962007-10-15 17:00:18 +02003208 /*
3209 * Aggressive migration if:
3210 * 1) task is cache cold, or
3211 * 2) too many balance attempts have failed.
3212 */
3213
Luis Henriques708dc512009-03-16 19:59:02 +00003214 tsk_cache_hot = task_hot(p, rq->clock, sd);
3215 if (!tsk_cache_hot ||
3216 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003217#ifdef CONFIG_SCHEDSTATS
Luis Henriques708dc512009-03-16 19:59:02 +00003218 if (tsk_cache_hot) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003219 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02003220 schedstat_inc(p, se.nr_forced_migrations);
3221 }
Ingo Molnarda84d962007-10-15 17:00:18 +02003222#endif
3223 return 1;
3224 }
3225
Luis Henriques708dc512009-03-16 19:59:02 +00003226 if (tsk_cache_hot) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003227 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02003228 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003229 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003230 return 1;
3231}
3232
Peter Williamse1d14842007-10-24 18:23:51 +02003233static unsigned long
3234balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3235 unsigned long max_load_move, struct sched_domain *sd,
3236 enum cpu_idle_type idle, int *all_pinned,
3237 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003238{
Peter Zijlstra051c6762008-06-27 13:41:31 +02003239 int loops = 0, pulled = 0, pinned = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02003240 struct task_struct *p;
3241 long rem_load_move = max_load_move;
3242
Peter Williamse1d14842007-10-24 18:23:51 +02003243 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02003244 goto out;
3245
3246 pinned = 1;
3247
3248 /*
3249 * Start the load-balancing iterator:
3250 */
3251 p = iterator->start(iterator->arg);
3252next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003253 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02003254 goto out;
Peter Zijlstra051c6762008-06-27 13:41:31 +02003255
3256 if ((p->se.load.weight >> 1) > rem_load_move ||
Ingo Molnardd41f592007-07-09 18:51:59 +02003257 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003258 p = iterator->next(iterator->arg);
3259 goto next;
3260 }
3261
3262 pull_task(busiest, p, this_rq, this_cpu);
3263 pulled++;
3264 rem_load_move -= p->se.load.weight;
3265
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003266#ifdef CONFIG_PREEMPT
3267 /*
3268 * NEWIDLE balancing is a source of latency, so preemptible kernels
3269 * will stop after the first task is pulled to minimize the critical
3270 * section.
3271 */
3272 if (idle == CPU_NEWLY_IDLE)
3273 goto out;
3274#endif
3275
Ingo Molnardd41f592007-07-09 18:51:59 +02003276 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003277 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003278 */
Peter Williamse1d14842007-10-24 18:23:51 +02003279 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003280 if (p->prio < *this_best_prio)
3281 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003282 p = iterator->next(iterator->arg);
3283 goto next;
3284 }
3285out:
3286 /*
Peter Williamse1d14842007-10-24 18:23:51 +02003287 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02003288 * so we can safely collect pull_task() stats here rather than
3289 * inside pull_task().
3290 */
3291 schedstat_add(sd, lb_gained[idle], pulled);
3292
3293 if (all_pinned)
3294 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003295
3296 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003297}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003298
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299/*
Peter Williams43010652007-08-09 11:16:46 +02003300 * move_tasks tries to move up to max_load_move weighted load from busiest to
3301 * this_rq, as part of a balancing operation within domain "sd".
3302 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 *
3304 * Called with both runqueues locked.
3305 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003306static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003307 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003308 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003309 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003310{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003311 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003312 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003313 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003314
Ingo Molnardd41f592007-07-09 18:51:59 +02003315 do {
Peter Williams43010652007-08-09 11:16:46 +02003316 total_load_moved +=
3317 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003318 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003319 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003320 class = class->next;
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003321
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003322#ifdef CONFIG_PREEMPT
3323 /*
3324 * NEWIDLE balancing is a source of latency, so preemptible
3325 * kernels will stop after the first task is pulled to minimize
3326 * the critical section.
3327 */
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003328 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3329 break;
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003330#endif
Peter Williams43010652007-08-09 11:16:46 +02003331 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003332
Peter Williams43010652007-08-09 11:16:46 +02003333 return total_load_moved > 0;
3334}
3335
Peter Williamse1d14842007-10-24 18:23:51 +02003336static int
3337iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3338 struct sched_domain *sd, enum cpu_idle_type idle,
3339 struct rq_iterator *iterator)
3340{
3341 struct task_struct *p = iterator->start(iterator->arg);
3342 int pinned = 0;
3343
3344 while (p) {
3345 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3346 pull_task(busiest, p, this_rq, this_cpu);
3347 /*
3348 * Right now, this is only the second place pull_task()
3349 * is called, so we can safely collect pull_task()
3350 * stats here rather than inside pull_task().
3351 */
3352 schedstat_inc(sd, lb_gained[idle]);
3353
3354 return 1;
3355 }
3356 p = iterator->next(iterator->arg);
3357 }
3358
3359 return 0;
3360}
3361
Peter Williams43010652007-08-09 11:16:46 +02003362/*
3363 * move_one_task tries to move exactly one task from busiest to this_rq, as
3364 * part of active balancing operations within "domain".
3365 * Returns 1 if successful and 0 otherwise.
3366 *
3367 * Called with both runqueues locked.
3368 */
3369static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3370 struct sched_domain *sd, enum cpu_idle_type idle)
3371{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003372 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003373
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003374 for_each_class(class) {
Peter Williamse1d14842007-10-24 18:23:51 +02003375 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003376 return 1;
Hiroshi Shimamotocde7e5ca2009-08-18 13:01:01 +09003377 }
Peter Williams43010652007-08-09 11:16:46 +02003378
3379 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003380}
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303381/********** Helpers for find_busiest_group ************************/
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003382/*
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303383 * sd_lb_stats - Structure to store the statistics of a sched_domain
3384 * during load balancing.
3385 */
3386struct sd_lb_stats {
3387 struct sched_group *busiest; /* Busiest group in this sd */
3388 struct sched_group *this; /* Local group in this sd */
3389 unsigned long total_load; /* Total load of all groups in sd */
3390 unsigned long total_pwr; /* Total power of all groups in sd */
3391 unsigned long avg_load; /* Average load across all groups in sd */
3392
3393 /** Statistics of this group */
3394 unsigned long this_load;
3395 unsigned long this_load_per_task;
3396 unsigned long this_nr_running;
3397
3398 /* Statistics of the busiest group */
3399 unsigned long max_load;
3400 unsigned long busiest_load_per_task;
3401 unsigned long busiest_nr_running;
3402
3403 int group_imb; /* Is there imbalance in this sd */
3404#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3405 int power_savings_balance; /* Is powersave balance needed for this sd */
3406 struct sched_group *group_min; /* Least loaded group in sd */
3407 struct sched_group *group_leader; /* Group which relieves group_min */
3408 unsigned long min_load_per_task; /* load_per_task in group_min */
3409 unsigned long leader_nr_running; /* Nr running of group_leader */
3410 unsigned long min_nr_running; /* Nr running of group_min */
3411#endif
3412};
Linus Torvalds1da177e2005-04-16 15:20:36 -07003413
3414/*
Gautham R Shenoy381be782009-03-25 14:43:46 +05303415 * sg_lb_stats - stats of a sched_group required for load_balancing
3416 */
3417struct sg_lb_stats {
3418 unsigned long avg_load; /*Avg load across the CPUs of the group */
3419 unsigned long group_load; /* Total load over the CPUs of the group */
3420 unsigned long sum_nr_running; /* Nr tasks running in the group */
3421 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3422 unsigned long group_capacity;
3423 int group_imb; /* Is there an imbalance in the group ? */
3424};
3425
3426/**
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303427 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3428 * @group: The group whose first cpu is to be returned.
3429 */
3430static inline unsigned int group_first_cpu(struct sched_group *group)
3431{
3432 return cpumask_first(sched_group_cpus(group));
3433}
3434
3435/**
3436 * get_sd_load_idx - Obtain the load index for a given sched domain.
3437 * @sd: The sched_domain whose load_idx is to be obtained.
3438 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3439 */
3440static inline int get_sd_load_idx(struct sched_domain *sd,
3441 enum cpu_idle_type idle)
3442{
3443 int load_idx;
3444
3445 switch (idle) {
3446 case CPU_NOT_IDLE:
3447 load_idx = sd->busy_idx;
3448 break;
3449
3450 case CPU_NEWLY_IDLE:
3451 load_idx = sd->newidle_idx;
3452 break;
3453 default:
3454 load_idx = sd->idle_idx;
3455 break;
3456 }
3457
3458 return load_idx;
3459}
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303460
3461
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303462#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3463/**
3464 * init_sd_power_savings_stats - Initialize power savings statistics for
3465 * the given sched_domain, during load balancing.
3466 *
3467 * @sd: Sched domain whose power-savings statistics are to be initialized.
3468 * @sds: Variable containing the statistics for sd.
3469 * @idle: Idle status of the CPU at which we're performing load-balancing.
3470 */
3471static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3472 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3473{
3474 /*
3475 * Busy processors will not participate in power savings
3476 * balance.
3477 */
3478 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3479 sds->power_savings_balance = 0;
3480 else {
3481 sds->power_savings_balance = 1;
3482 sds->min_nr_running = ULONG_MAX;
3483 sds->leader_nr_running = 0;
3484 }
3485}
3486
3487/**
3488 * update_sd_power_savings_stats - Update the power saving stats for a
3489 * sched_domain while performing load balancing.
3490 *
3491 * @group: sched_group belonging to the sched_domain under consideration.
3492 * @sds: Variable containing the statistics of the sched_domain
3493 * @local_group: Does group contain the CPU for which we're performing
3494 * load balancing ?
3495 * @sgs: Variable containing the statistics of the group.
3496 */
3497static inline void update_sd_power_savings_stats(struct sched_group *group,
3498 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3499{
3500
3501 if (!sds->power_savings_balance)
3502 return;
3503
3504 /*
3505 * If the local group is idle or completely loaded
3506 * no need to do power savings balance at this domain
3507 */
3508 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3509 !sds->this_nr_running))
3510 sds->power_savings_balance = 0;
3511
3512 /*
3513 * If a group is already running at full capacity or idle,
3514 * don't include that group in power savings calculations
3515 */
3516 if (!sds->power_savings_balance ||
3517 sgs->sum_nr_running >= sgs->group_capacity ||
3518 !sgs->sum_nr_running)
3519 return;
3520
3521 /*
3522 * Calculate the group which has the least non-idle load.
3523 * This is the group from where we need to pick up the load
3524 * for saving power
3525 */
3526 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3527 (sgs->sum_nr_running == sds->min_nr_running &&
3528 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3529 sds->group_min = group;
3530 sds->min_nr_running = sgs->sum_nr_running;
3531 sds->min_load_per_task = sgs->sum_weighted_load /
3532 sgs->sum_nr_running;
3533 }
3534
3535 /*
3536 * Calculate the group which is almost near its
3537 * capacity but still has some space to pick up some load
3538 * from other group and save more power
3539 */
Gautham R Shenoyd899a782009-09-02 16:59:10 +05303540 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303541 return;
3542
3543 if (sgs->sum_nr_running > sds->leader_nr_running ||
3544 (sgs->sum_nr_running == sds->leader_nr_running &&
3545 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3546 sds->group_leader = group;
3547 sds->leader_nr_running = sgs->sum_nr_running;
3548 }
3549}
3550
3551/**
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003552 * check_power_save_busiest_group - see if there is potential for some power-savings balance
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303553 * @sds: Variable containing the statistics of the sched_domain
3554 * under consideration.
3555 * @this_cpu: Cpu at which we're currently performing load-balancing.
3556 * @imbalance: Variable to store the imbalance.
3557 *
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003558 * Description:
3559 * Check if we have potential to perform some power-savings balance.
3560 * If yes, set the busiest group to be the least loaded group in the
3561 * sched_domain, so that it's CPUs can be put to idle.
3562 *
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303563 * Returns 1 if there is potential to perform power-savings balance.
3564 * Else returns 0.
3565 */
3566static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3567 int this_cpu, unsigned long *imbalance)
3568{
3569 if (!sds->power_savings_balance)
3570 return 0;
3571
3572 if (sds->this != sds->group_leader ||
3573 sds->group_leader == sds->group_min)
3574 return 0;
3575
3576 *imbalance = sds->min_load_per_task;
3577 sds->busiest = sds->group_min;
3578
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303579 return 1;
3580
3581}
3582#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3583static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3584 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3585{
3586 return;
3587}
3588
3589static inline void update_sd_power_savings_stats(struct sched_group *group,
3590 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3591{
3592 return;
3593}
3594
3595static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3596 int this_cpu, unsigned long *imbalance)
3597{
3598 return 0;
3599}
3600#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3601
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003602
3603unsigned long default_scale_freq_power(struct sched_domain *sd, int cpu)
3604{
3605 return SCHED_LOAD_SCALE;
3606}
3607
3608unsigned long __weak arch_scale_freq_power(struct sched_domain *sd, int cpu)
3609{
3610 return default_scale_freq_power(sd, cpu);
3611}
3612
3613unsigned long default_scale_smt_power(struct sched_domain *sd, int cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003614{
3615 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3616 unsigned long smt_gain = sd->smt_gain;
3617
3618 smt_gain /= weight;
3619
3620 return smt_gain;
3621}
3622
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003623unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3624{
3625 return default_scale_smt_power(sd, cpu);
3626}
3627
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003628unsigned long scale_rt_power(int cpu)
3629{
3630 struct rq *rq = cpu_rq(cpu);
3631 u64 total, available;
3632
3633 sched_avg_update(rq);
3634
3635 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3636 available = total - rq->rt_avg;
3637
3638 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3639 total = SCHED_LOAD_SCALE;
3640
3641 total >>= SCHED_LOAD_SHIFT;
3642
3643 return div_u64(available, total);
3644}
3645
Peter Zijlstraab292302009-09-01 10:34:36 +02003646static void update_cpu_power(struct sched_domain *sd, int cpu)
3647{
3648 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3649 unsigned long power = SCHED_LOAD_SCALE;
3650 struct sched_group *sdg = sd->groups;
Peter Zijlstraab292302009-09-01 10:34:36 +02003651
Peter Zijlstra8e6598a2009-09-03 13:20:03 +02003652 if (sched_feat(ARCH_POWER))
3653 power *= arch_scale_freq_power(sd, cpu);
3654 else
3655 power *= default_scale_freq_power(sd, cpu);
3656
Peter Zijlstrad6a59aa2009-09-02 13:28:02 +02003657 power >>= SCHED_LOAD_SHIFT;
Peter Zijlstraab292302009-09-01 10:34:36 +02003658
3659 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
Peter Zijlstra8e6598a2009-09-03 13:20:03 +02003660 if (sched_feat(ARCH_POWER))
3661 power *= arch_scale_smt_power(sd, cpu);
3662 else
3663 power *= default_scale_smt_power(sd, cpu);
3664
Peter Zijlstraab292302009-09-01 10:34:36 +02003665 power >>= SCHED_LOAD_SHIFT;
3666 }
3667
Peter Zijlstrae9e92502009-09-01 10:34:37 +02003668 power *= scale_rt_power(cpu);
3669 power >>= SCHED_LOAD_SHIFT;
3670
3671 if (!power)
3672 power = 1;
Peter Zijlstraab292302009-09-01 10:34:36 +02003673
Peter Zijlstra18a38852009-09-01 10:34:39 +02003674 sdg->cpu_power = power;
Peter Zijlstraab292302009-09-01 10:34:36 +02003675}
3676
3677static void update_group_power(struct sched_domain *sd, int cpu)
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003678{
3679 struct sched_domain *child = sd->child;
3680 struct sched_group *group, *sdg = sd->groups;
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003681 unsigned long power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003682
3683 if (!child) {
Peter Zijlstraab292302009-09-01 10:34:36 +02003684 update_cpu_power(sd, cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003685 return;
3686 }
3687
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003688 power = 0;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003689
3690 group = child->groups;
3691 do {
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003692 power += group->cpu_power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003693 group = group->next;
3694 } while (group != child->groups);
Ingo Molnard7ea17a2009-09-04 11:49:25 +02003695
3696 sdg->cpu_power = power;
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003697}
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303698
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303699/**
3700 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
Randy Dunlape17b38b2009-10-11 19:12:00 -07003701 * @sd: The sched_domain whose statistics are to be updated.
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303702 * @group: sched_group whose statistics are to be updated.
3703 * @this_cpu: Cpu for which load balance is currently performed.
3704 * @idle: Idle status of this_cpu
3705 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3706 * @sd_idle: Idle status of the sched_domain containing group.
3707 * @local_group: Does group contain this_cpu.
3708 * @cpus: Set of cpus considered for load balancing.
3709 * @balance: Should we balance.
3710 * @sgs: variable to hold the statistics for this group.
3711 */
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003712static inline void update_sg_lb_stats(struct sched_domain *sd,
3713 struct sched_group *group, int this_cpu,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303714 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3715 int local_group, const struct cpumask *cpus,
3716 int *balance, struct sg_lb_stats *sgs)
3717{
3718 unsigned long load, max_cpu_load, min_cpu_load;
3719 int i;
3720 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3721 unsigned long sum_avg_load_per_task;
3722 unsigned long avg_load_per_task;
3723
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003724 if (local_group) {
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303725 balance_cpu = group_first_cpu(group);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003726 if (balance_cpu == this_cpu)
Peter Zijlstraab292302009-09-01 10:34:36 +02003727 update_group_power(sd, this_cpu);
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003728 }
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303729
3730 /* Tally up the load of all CPUs in the group */
3731 sum_avg_load_per_task = avg_load_per_task = 0;
3732 max_cpu_load = 0;
3733 min_cpu_load = ~0UL;
3734
3735 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3736 struct rq *rq = cpu_rq(i);
3737
3738 if (*sd_idle && rq->nr_running)
3739 *sd_idle = 0;
3740
3741 /* Bias balancing toward cpus of our domain */
3742 if (local_group) {
3743 if (idle_cpu(i) && !first_idle_cpu) {
3744 first_idle_cpu = 1;
3745 balance_cpu = i;
3746 }
3747
3748 load = target_load(i, load_idx);
3749 } else {
3750 load = source_load(i, load_idx);
3751 if (load > max_cpu_load)
3752 max_cpu_load = load;
3753 if (min_cpu_load > load)
3754 min_cpu_load = load;
3755 }
3756
3757 sgs->group_load += load;
3758 sgs->sum_nr_running += rq->nr_running;
3759 sgs->sum_weighted_load += weighted_cpuload(i);
3760
3761 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3762 }
3763
3764 /*
3765 * First idle cpu or the first cpu(busiest) in this sched group
3766 * is eligible for doing load balancing at this and above
3767 * domains. In the newly idle case, we will allow all the cpu's
3768 * to do the newly idle load balance.
3769 */
3770 if (idle != CPU_NEWLY_IDLE && local_group &&
3771 balance_cpu != this_cpu && balance) {
3772 *balance = 0;
3773 return;
3774 }
3775
3776 /* Adjust by relative CPU power of the group */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003777 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303778
3779
3780 /*
3781 * Consider the group unbalanced when the imbalance is larger
3782 * than the average weight of two tasks.
3783 *
3784 * APZ: with cgroup the avg task weight can vary wildly and
3785 * might not be a suitable number - should we keep a
3786 * normalized nr_running number somewhere that negates
3787 * the hierarchy?
3788 */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003789 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3790 group->cpu_power;
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303791
3792 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3793 sgs->group_imb = 1;
3794
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003795 sgs->group_capacity =
Peter Zijlstra18a38852009-09-01 10:34:39 +02003796 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303797}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003798
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303799/**
3800 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3801 * @sd: sched_domain whose statistics are to be updated.
3802 * @this_cpu: Cpu for which load balance is currently performed.
3803 * @idle: Idle status of this_cpu
3804 * @sd_idle: Idle status of the sched_domain containing group.
3805 * @cpus: Set of cpus considered for load balancing.
3806 * @balance: Should we balance.
3807 * @sds: variable to hold the statistics for this sched_domain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003808 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303809static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3810 enum cpu_idle_type idle, int *sd_idle,
3811 const struct cpumask *cpus, int *balance,
3812 struct sd_lb_stats *sds)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003813{
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003814 struct sched_domain *child = sd->child;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303815 struct sched_group *group = sd->groups;
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303816 struct sg_lb_stats sgs;
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003817 int load_idx, prefer_sibling = 0;
3818
3819 if (child && child->flags & SD_PREFER_SIBLING)
3820 prefer_sibling = 1;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303821
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303822 init_sd_power_savings_stats(sd, sds, idle);
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303823 load_idx = get_sd_load_idx(sd, idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003824
3825 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003826 int local_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003827
Rusty Russell758b2cd2008-11-25 02:35:04 +10303828 local_group = cpumask_test_cpu(this_cpu,
3829 sched_group_cpus(group));
Gautham R Shenoy381be782009-03-25 14:43:46 +05303830 memset(&sgs, 0, sizeof(sgs));
Peter Zijlstracc9fba72009-09-01 10:34:34 +02003831 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303832 local_group, cpus, balance, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003833
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303834 if (local_group && balance && !(*balance))
3835 return;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003836
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303837 sds->total_load += sgs.group_load;
Peter Zijlstra18a38852009-09-01 10:34:39 +02003838 sds->total_pwr += group->cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003839
Peter Zijlstrab5d978e2009-09-01 10:34:33 +02003840 /*
3841 * In case the child domain prefers tasks go to siblings
3842 * first, lower the group capacity to one so that we'll try
3843 * and move all the excess tasks away.
3844 */
3845 if (prefer_sibling)
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02003846 sgs.group_capacity = min(sgs.group_capacity, 1UL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003847
Linus Torvalds1da177e2005-04-16 15:20:36 -07003848 if (local_group) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303849 sds->this_load = sgs.avg_load;
3850 sds->this = group;
3851 sds->this_nr_running = sgs.sum_nr_running;
3852 sds->this_load_per_task = sgs.sum_weighted_load;
3853 } else if (sgs.avg_load > sds->max_load &&
Gautham R Shenoy381be782009-03-25 14:43:46 +05303854 (sgs.sum_nr_running > sgs.group_capacity ||
3855 sgs.group_imb)) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303856 sds->max_load = sgs.avg_load;
3857 sds->busiest = group;
3858 sds->busiest_nr_running = sgs.sum_nr_running;
3859 sds->busiest_load_per_task = sgs.sum_weighted_load;
3860 sds->group_imb = sgs.group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003861 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003862
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303863 update_sd_power_savings_stats(group, sds, local_group, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003864 group = group->next;
3865 } while (group != sd->groups);
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303866}
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303867
3868/**
3869 * fix_small_imbalance - Calculate the minor imbalance that exists
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303870 * amongst the groups of a sched_domain, during
3871 * load balancing.
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303872 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3873 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3874 * @imbalance: Variable to store the imbalance.
3875 */
3876static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3877 int this_cpu, unsigned long *imbalance)
3878{
3879 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3880 unsigned int imbn = 2;
3881
3882 if (sds->this_nr_running) {
3883 sds->this_load_per_task /= sds->this_nr_running;
3884 if (sds->busiest_load_per_task >
3885 sds->this_load_per_task)
3886 imbn = 1;
3887 } else
3888 sds->this_load_per_task =
3889 cpu_avg_load_per_task(this_cpu);
3890
3891 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3892 sds->busiest_load_per_task * imbn) {
3893 *imbalance = sds->busiest_load_per_task;
3894 return;
3895 }
3896
3897 /*
3898 * OK, we don't have enough imbalance to justify moving tasks,
3899 * however we may be able to increase total CPU power used by
3900 * moving them.
3901 */
3902
Peter Zijlstra18a38852009-09-01 10:34:39 +02003903 pwr_now += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303904 min(sds->busiest_load_per_task, sds->max_load);
Peter Zijlstra18a38852009-09-01 10:34:39 +02003905 pwr_now += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303906 min(sds->this_load_per_task, sds->this_load);
3907 pwr_now /= SCHED_LOAD_SCALE;
3908
3909 /* Amount of load we'd subtract */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003910 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3911 sds->busiest->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303912 if (sds->max_load > tmp)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003913 pwr_move += sds->busiest->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303914 min(sds->busiest_load_per_task, sds->max_load - tmp);
3915
3916 /* Amount of load we'd add */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003917 if (sds->max_load * sds->busiest->cpu_power <
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303918 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
Peter Zijlstra18a38852009-09-01 10:34:39 +02003919 tmp = (sds->max_load * sds->busiest->cpu_power) /
3920 sds->this->cpu_power;
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303921 else
Peter Zijlstra18a38852009-09-01 10:34:39 +02003922 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3923 sds->this->cpu_power;
3924 pwr_move += sds->this->cpu_power *
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303925 min(sds->this_load_per_task, sds->this_load + tmp);
3926 pwr_move /= SCHED_LOAD_SCALE;
3927
3928 /* Move if we gain throughput */
3929 if (pwr_move > pwr_now)
3930 *imbalance = sds->busiest_load_per_task;
3931}
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303932
3933/**
3934 * calculate_imbalance - Calculate the amount of imbalance present within the
3935 * groups of a given sched_domain during load balance.
3936 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3937 * @this_cpu: Cpu for which currently load balance is being performed.
3938 * @imbalance: The variable to store the imbalance.
3939 */
3940static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3941 unsigned long *imbalance)
3942{
3943 unsigned long max_pull;
3944 /*
3945 * In the presence of smp nice balancing, certain scenarios can have
3946 * max load less than avg load(as we skip the groups at or below
3947 * its cpu_power, while calculating max_load..)
3948 */
3949 if (sds->max_load < sds->avg_load) {
3950 *imbalance = 0;
3951 return fix_small_imbalance(sds, this_cpu, imbalance);
3952 }
3953
3954 /* Don't want to pull so many tasks that a group would go idle */
3955 max_pull = min(sds->max_load - sds->avg_load,
3956 sds->max_load - sds->busiest_load_per_task);
3957
3958 /* How much load to actually move to equalise the imbalance */
Peter Zijlstra18a38852009-09-01 10:34:39 +02003959 *imbalance = min(max_pull * sds->busiest->cpu_power,
3960 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303961 / SCHED_LOAD_SCALE;
3962
3963 /*
3964 * if *imbalance is less than the average load per runnable task
3965 * there is no gaurantee that any tasks will be moved so we'll have
3966 * a think about bumping its value to force at least one task to be
3967 * moved
3968 */
3969 if (*imbalance < sds->busiest_load_per_task)
3970 return fix_small_imbalance(sds, this_cpu, imbalance);
3971
3972}
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303973/******* find_busiest_group() helpers end here *********************/
3974
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303975/**
3976 * find_busiest_group - Returns the busiest group within the sched_domain
3977 * if there is an imbalance. If there isn't an imbalance, and
3978 * the user has opted for power-savings, it returns a group whose
3979 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3980 * such a group exists.
3981 *
3982 * Also calculates the amount of weighted load which should be moved
3983 * to restore balance.
3984 *
3985 * @sd: The sched_domain whose busiest group is to be returned.
3986 * @this_cpu: The cpu for which load balancing is currently being performed.
3987 * @imbalance: Variable which stores amount of weighted load which should
3988 * be moved to restore balance/put a group to idle.
3989 * @idle: The idle status of this_cpu.
3990 * @sd_idle: The idleness of sd
3991 * @cpus: The set of CPUs under consideration for load-balancing.
3992 * @balance: Pointer to a variable indicating if this_cpu
3993 * is the appropriate cpu to perform load balancing at this_level.
3994 *
3995 * Returns: - the busiest group if imbalance exists.
3996 * - If no imbalance and user has opted for power-savings balance,
3997 * return the least loaded group whose CPUs can be
3998 * put to idle by rebalancing its tasks onto our group.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003999 */
4000static struct sched_group *
4001find_busiest_group(struct sched_domain *sd, int this_cpu,
4002 unsigned long *imbalance, enum cpu_idle_type idle,
4003 int *sd_idle, const struct cpumask *cpus, int *balance)
4004{
Gautham R Shenoy37abe192009-03-25 14:44:01 +05304005 struct sd_lb_stats sds;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006
Gautham R Shenoy37abe192009-03-25 14:44:01 +05304007 memset(&sds, 0, sizeof(sds));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004008
Gautham R Shenoy37abe192009-03-25 14:44:01 +05304009 /*
4010 * Compute the various statistics relavent for load balancing at
4011 * this level.
4012 */
4013 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4014 balance, &sds);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004015
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304016 /* Cases where imbalance does not exist from POV of this_cpu */
4017 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4018 * at this level.
4019 * 2) There is no busy sibling group to pull from.
4020 * 3) This group is the busiest group.
4021 * 4) This group is more busy than the avg busieness at this
4022 * sched_domain.
4023 * 5) The imbalance is within the specified limit.
4024 * 6) Any rebalance would lead to ping-pong
4025 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05304026 if (balance && !(*balance))
4027 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004028
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304029 if (!sds.busiest || sds.busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004030 goto out_balanced;
4031
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304032 if (sds.this_load >= sds.max_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004033 goto out_balanced;
4034
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304035 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05304037 if (sds.this_load >= sds.avg_load)
4038 goto out_balanced;
4039
4040 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004041 goto out_balanced;
4042
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304043 sds.busiest_load_per_task /= sds.busiest_nr_running;
4044 if (sds.group_imb)
4045 sds.busiest_load_per_task =
4046 min(sds.busiest_load_per_task, sds.avg_load);
Ken Chen908a7c12007-10-17 16:55:11 +02004047
Linus Torvalds1da177e2005-04-16 15:20:36 -07004048 /*
4049 * We're trying to get all the cpus to the average_load, so we don't
4050 * want to push ourselves above the average load, nor do we wish to
4051 * reduce the max loaded cpu below the average load, as either of these
4052 * actions would just result in more rebalancing later, and ping-pong
4053 * tasks around. Thus we look for the minimum possible imbalance.
4054 * Negative imbalances (*we* are more loaded than anyone else) will
4055 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004056 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07004057 * appear as very large values with unsigned longs.
4058 */
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304059 if (sds.max_load <= sds.busiest_load_per_task)
Peter Williams2dd73a42006-06-27 02:54:34 -07004060 goto out_balanced;
4061
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05304062 /* Looks like there is an imbalance. Compute it */
4063 calculate_imbalance(&sds, this_cpu, imbalance);
Gautham R Shenoy222d6562009-03-25 14:43:56 +05304064 return sds.busiest;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004065
4066out_balanced:
Gautham R Shenoyc071df12009-03-25 14:44:22 +05304067 /*
4068 * There is no obvious imbalance. But check if we can do some balancing
4069 * to save power.
4070 */
4071 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4072 return sds.busiest;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004073ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 *imbalance = 0;
4075 return NULL;
4076}
4077
4078/*
4079 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4080 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004081static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004082find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Rusty Russell96f874e2008-11-25 02:35:14 +10304083 unsigned long imbalance, const struct cpumask *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084{
Ingo Molnar70b97a72006-07-03 00:25:42 -07004085 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07004086 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004087 int i;
4088
Rusty Russell758b2cd2008-11-25 02:35:04 +10304089 for_each_cpu(i, sched_group_cpus(group)) {
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004090 unsigned long power = power_of(i);
4091 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
Ingo Molnardd41f592007-07-09 18:51:59 +02004092 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004093
Rusty Russell96f874e2008-11-25 02:35:14 +10304094 if (!cpumask_test_cpu(i, cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004095 continue;
4096
Ingo Molnar48f24c42006-07-03 00:25:40 -07004097 rq = cpu_rq(i);
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004098 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4099 wl /= power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100
Peter Zijlstrabdb94aa2009-09-01 10:34:38 +02004101 if (capacity && rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07004102 continue;
4103
Ingo Molnardd41f592007-07-09 18:51:59 +02004104 if (wl > max_load) {
4105 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004106 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004107 }
4108 }
4109
4110 return busiest;
4111}
4112
4113/*
Nick Piggin77391d72005-06-25 14:57:30 -07004114 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4115 * so long as it is large enough.
4116 */
4117#define MAX_PINNED_INTERVAL 512
4118
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304119/* Working cpumask for load_balance and load_balance_newidle. */
4120static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4121
Nick Piggin77391d72005-06-25 14:57:30 -07004122/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004123 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4124 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004125 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004126static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004127 struct sched_domain *sd, enum cpu_idle_type idle,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304128 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004129{
Peter Williams43010652007-08-09 11:16:46 +02004130 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004131 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004132 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004133 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004134 unsigned long flags;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304135 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004136
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01004137 cpumask_copy(cpus, cpu_online_mask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004138
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004139 /*
4140 * When power savings policy is enabled for the parent domain, idle
4141 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02004142 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004143 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004144 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004145 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004146 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004147 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004148
Ingo Molnar2d723762007-10-15 17:00:12 +02004149 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004151redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004152 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004153 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07004154 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004155
Chen, Kenneth W06066712006-12-10 02:20:35 -08004156 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004157 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004158
Linus Torvalds1da177e2005-04-16 15:20:36 -07004159 if (!group) {
4160 schedstat_inc(sd, lb_nobusyg[idle]);
4161 goto out_balanced;
4162 }
4163
Mike Travis7c16ec52008-04-04 18:11:11 -07004164 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004165 if (!busiest) {
4166 schedstat_inc(sd, lb_nobusyq[idle]);
4167 goto out_balanced;
4168 }
4169
Nick Piggindb935db2005-06-25 14:57:11 -07004170 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004171
4172 schedstat_add(sd, lb_imbalance[idle], imbalance);
4173
Peter Williams43010652007-08-09 11:16:46 +02004174 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004175 if (busiest->nr_running > 1) {
4176 /*
4177 * Attempt to move tasks. If find_busiest_group has found
4178 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02004179 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07004180 * correctly treated as an imbalance.
4181 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004182 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07004183 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02004184 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07004185 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07004186 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004187 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07004188
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004189 /*
4190 * some other cpu did the load balance for us.
4191 */
Peter Williams43010652007-08-09 11:16:46 +02004192 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004193 resched_cpu(this_cpu);
4194
Nick Piggin81026792005-06-25 14:57:07 -07004195 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004196 if (unlikely(all_pinned)) {
Rusty Russell96f874e2008-11-25 02:35:14 +10304197 cpumask_clear_cpu(cpu_of(busiest), cpus);
4198 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004199 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07004200 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004201 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004202 }
Nick Piggin81026792005-06-25 14:57:07 -07004203
Peter Williams43010652007-08-09 11:16:46 +02004204 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004205 schedstat_inc(sd, lb_failed[idle]);
4206 sd->nr_balance_failed++;
4207
4208 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004209
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004210 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004211
4212 /* don't kick the migration_thread, if the curr
4213 * task on busiest cpu can't be moved to this_cpu
4214 */
Rusty Russell96f874e2008-11-25 02:35:14 +10304215 if (!cpumask_test_cpu(this_cpu,
4216 &busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004217 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004218 all_pinned = 1;
4219 goto out_one_pinned;
4220 }
4221
Linus Torvalds1da177e2005-04-16 15:20:36 -07004222 if (!busiest->active_balance) {
4223 busiest->active_balance = 1;
4224 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07004225 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004226 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004227 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07004228 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004229 wake_up_process(busiest->migration_thread);
4230
4231 /*
4232 * We've kicked active balancing, reset the failure
4233 * counter.
4234 */
Nick Piggin39507452005-06-25 14:57:09 -07004235 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004236 }
Nick Piggin81026792005-06-25 14:57:07 -07004237 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004238 sd->nr_balance_failed = 0;
4239
Nick Piggin81026792005-06-25 14:57:07 -07004240 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004241 /* We were unbalanced, so reset the balancing interval */
4242 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07004243 } else {
4244 /*
4245 * If we've begun active balancing, start to back off. This
4246 * case may not be covered by the all_pinned logic if there
4247 * is only 1 task on the busy runqueue (because we don't call
4248 * move_tasks).
4249 */
4250 if (sd->balance_interval < sd->max_interval)
4251 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004252 }
4253
Peter Williams43010652007-08-09 11:16:46 +02004254 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004255 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004256 ld_moved = -1;
4257
4258 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004259
4260out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004261 schedstat_inc(sd, lb_balanced[idle]);
4262
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004263 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004264
4265out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004266 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07004267 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4268 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004269 sd->balance_interval *= 2;
4270
Ingo Molnar48f24c42006-07-03 00:25:40 -07004271 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004272 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004273 ld_moved = -1;
4274 else
4275 ld_moved = 0;
4276out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004277 if (ld_moved)
4278 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004279 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004280}
4281
4282/*
4283 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4284 * tasks if there is an imbalance.
4285 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004286 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07004287 * this_rq is locked.
4288 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07004289static int
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304290load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291{
4292 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004293 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004294 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02004295 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07004296 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004297 int all_pinned = 0;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304298 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004299
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01004300 cpumask_copy(cpus, cpu_online_mask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004301
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004302 /*
4303 * When power savings policy is enabled for the parent domain, idle
4304 * sibling can pick up load irrespective of busy siblings. In this case,
4305 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004306 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004307 */
4308 if (sd->flags & SD_SHARE_CPUPOWER &&
4309 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004310 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004311
Ingo Molnar2d723762007-10-15 17:00:12 +02004312 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004313redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004314 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004315 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07004316 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004317 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004318 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004319 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320 }
4321
Mike Travis7c16ec52008-04-04 18:11:11 -07004322 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07004323 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004324 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004325 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326 }
4327
Nick Piggindb935db2005-06-25 14:57:11 -07004328 BUG_ON(busiest == this_rq);
4329
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004330 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004331
Peter Williams43010652007-08-09 11:16:46 +02004332 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004333 if (busiest->nr_running > 1) {
4334 /* Attempt to move tasks */
4335 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004336 /* this_rq->clock is already updated */
4337 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02004338 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004339 imbalance, sd, CPU_NEWLY_IDLE,
4340 &all_pinned);
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004341 double_unlock_balance(this_rq, busiest);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004342
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004343 if (unlikely(all_pinned)) {
Rusty Russell96f874e2008-11-25 02:35:14 +10304344 cpumask_clear_cpu(cpu_of(busiest), cpus);
4345 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004346 goto redo;
4347 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004348 }
4349
Peter Williams43010652007-08-09 11:16:46 +02004350 if (!ld_moved) {
Vaidyanathan Srinivasan36dffab2008-12-20 10:06:38 +05304351 int active_balance = 0;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304352
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004353 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004354 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4355 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004356 return -1;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304357
4358 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4359 return -1;
4360
4361 if (sd->nr_balance_failed++ < 2)
4362 return -1;
4363
4364 /*
4365 * The only task running in a non-idle cpu can be moved to this
4366 * cpu in an attempt to completely freeup the other CPU
4367 * package. The same method used to move task in load_balance()
4368 * have been extended for load_balance_newidle() to speedup
4369 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4370 *
4371 * The package power saving logic comes from
4372 * find_busiest_group(). If there are no imbalance, then
4373 * f_b_g() will return NULL. However when sched_mc={1,2} then
4374 * f_b_g() will select a group from which a running task may be
4375 * pulled to this cpu in order to make the other package idle.
4376 * If there is no opportunity to make a package idle and if
4377 * there are no imbalance, then f_b_g() will return NULL and no
4378 * action will be taken in load_balance_newidle().
4379 *
4380 * Under normal task pull operation due to imbalance, there
4381 * will be more than one task in the source run queue and
4382 * move_tasks() will succeed. ld_moved will be true and this
4383 * active balance code will not be triggered.
4384 */
4385
4386 /* Lock busiest in correct order while this_rq is held */
4387 double_lock_balance(this_rq, busiest);
4388
4389 /*
4390 * don't kick the migration_thread, if the curr
4391 * task on busiest cpu can't be moved to this_cpu
4392 */
Mike Travis6ca09df2008-12-31 18:08:45 -08004393 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304394 double_unlock_balance(this_rq, busiest);
4395 all_pinned = 1;
4396 return ld_moved;
4397 }
4398
4399 if (!busiest->active_balance) {
4400 busiest->active_balance = 1;
4401 busiest->push_cpu = this_cpu;
4402 active_balance = 1;
4403 }
4404
4405 double_unlock_balance(this_rq, busiest);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004406 /*
4407 * Should not call ttwu while holding a rq->lock
4408 */
4409 spin_unlock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304410 if (active_balance)
4411 wake_up_process(busiest->migration_thread);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004412 spin_lock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304413
Nick Piggin5969fe02005-09-10 00:26:19 -07004414 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004415 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004416
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004417 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02004418 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004419
4420out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004421 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004422 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004423 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004424 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004425 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004426
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004427 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004428}
4429
4430/*
4431 * idle_balance is called by schedule() if this_cpu is about to become
4432 * idle. Attempts to pull tasks from other CPUs.
4433 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004434static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004435{
4436 struct sched_domain *sd;
Vaidyanathan Srinivasanefbe0272008-12-08 20:52:49 +05304437 int pulled_task = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02004438 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004439
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004440 this_rq->idle_stamp = this_rq->clock;
4441
4442 if (this_rq->avg_idle < sysctl_sched_migration_cost)
4443 return;
4444
Linus Torvalds1da177e2005-04-16 15:20:36 -07004445 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004446 unsigned long interval;
4447
4448 if (!(sd->flags & SD_LOAD_BALANCE))
4449 continue;
4450
4451 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07004452 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07004453 pulled_task = load_balance_newidle(this_cpu, this_rq,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304454 sd);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004455
4456 interval = msecs_to_jiffies(sd->balance_interval);
4457 if (time_after(next_balance, sd->last_balance + interval))
4458 next_balance = sd->last_balance + interval;
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004459 if (pulled_task) {
4460 this_rq->idle_stamp = 0;
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004461 break;
Mike Galbraith1b9508f2009-11-04 17:53:50 +01004462 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004463 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004464 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004465 /*
4466 * We are going idle. next_balance may be set based on
4467 * a busy processor. So reset next_balance.
4468 */
4469 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02004470 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004471}
4472
4473/*
4474 * active_load_balance is run by migration threads. It pushes running tasks
4475 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4476 * running on each physical CPU where possible, and avoids physical /
4477 * logical imbalances.
4478 *
4479 * Called with busiest_rq locked.
4480 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004481static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004482{
Nick Piggin39507452005-06-25 14:57:09 -07004483 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004484 struct sched_domain *sd;
4485 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07004486
Ingo Molnar48f24c42006-07-03 00:25:40 -07004487 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07004488 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07004489 return;
4490
4491 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004492
4493 /*
Nick Piggin39507452005-06-25 14:57:09 -07004494 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004495 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07004496 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497 */
Nick Piggin39507452005-06-25 14:57:09 -07004498 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004499
Nick Piggin39507452005-06-25 14:57:09 -07004500 /* move a task from busiest_rq to target_rq */
4501 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004502 update_rq_clock(busiest_rq);
4503 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004504
Nick Piggin39507452005-06-25 14:57:09 -07004505 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004506 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07004507 if ((sd->flags & SD_LOAD_BALANCE) &&
Rusty Russell758b2cd2008-11-25 02:35:04 +10304508 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
Nick Piggin39507452005-06-25 14:57:09 -07004509 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004510 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004511
Ingo Molnar48f24c42006-07-03 00:25:40 -07004512 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004513 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514
Peter Williams43010652007-08-09 11:16:46 +02004515 if (move_one_task(target_rq, target_cpu, busiest_rq,
4516 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07004517 schedstat_inc(sd, alb_pushed);
4518 else
4519 schedstat_inc(sd, alb_failed);
4520 }
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004521 double_unlock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004522}
4523
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004524#ifdef CONFIG_NO_HZ
4525static struct {
4526 atomic_t load_balancer;
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304527 cpumask_var_t cpu_mask;
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304528 cpumask_var_t ilb_grp_nohz_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004529} nohz ____cacheline_aligned = {
4530 .load_balancer = ATOMIC_INIT(-1),
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004531};
4532
Arun R Bharadwajeea08f32009-04-16 12:16:41 +05304533int get_nohz_load_balancer(void)
4534{
4535 return atomic_read(&nohz.load_balancer);
4536}
4537
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304538#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4539/**
4540 * lowest_flag_domain - Return lowest sched_domain containing flag.
4541 * @cpu: The cpu whose lowest level of sched domain is to
4542 * be returned.
4543 * @flag: The flag to check for the lowest sched_domain
4544 * for the given cpu.
4545 *
4546 * Returns the lowest sched_domain of a cpu which contains the given flag.
4547 */
4548static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4549{
4550 struct sched_domain *sd;
4551
4552 for_each_domain(cpu, sd)
4553 if (sd && (sd->flags & flag))
4554 break;
4555
4556 return sd;
4557}
4558
4559/**
4560 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4561 * @cpu: The cpu whose domains we're iterating over.
4562 * @sd: variable holding the value of the power_savings_sd
4563 * for cpu.
4564 * @flag: The flag to filter the sched_domains to be iterated.
4565 *
4566 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4567 * set, starting from the lowest sched_domain to the highest.
4568 */
4569#define for_each_flag_domain(cpu, sd, flag) \
4570 for (sd = lowest_flag_domain(cpu, flag); \
4571 (sd && (sd->flags & flag)); sd = sd->parent)
4572
4573/**
4574 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4575 * @ilb_group: group to be checked for semi-idleness
4576 *
4577 * Returns: 1 if the group is semi-idle. 0 otherwise.
4578 *
4579 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4580 * and atleast one non-idle CPU. This helper function checks if the given
4581 * sched_group is semi-idle or not.
4582 */
4583static inline int is_semi_idle_group(struct sched_group *ilb_group)
4584{
4585 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4586 sched_group_cpus(ilb_group));
4587
4588 /*
4589 * A sched_group is semi-idle when it has atleast one busy cpu
4590 * and atleast one idle cpu.
4591 */
4592 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4593 return 0;
4594
4595 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4596 return 0;
4597
4598 return 1;
4599}
4600/**
4601 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4602 * @cpu: The cpu which is nominating a new idle_load_balancer.
4603 *
4604 * Returns: Returns the id of the idle load balancer if it exists,
4605 * Else, returns >= nr_cpu_ids.
4606 *
4607 * This algorithm picks the idle load balancer such that it belongs to a
4608 * semi-idle powersavings sched_domain. The idea is to try and avoid
4609 * completely idle packages/cores just for the purpose of idle load balancing
4610 * when there are other idle cpu's which are better suited for that job.
4611 */
4612static int find_new_ilb(int cpu)
4613{
4614 struct sched_domain *sd;
4615 struct sched_group *ilb_group;
4616
4617 /*
4618 * Have idle load balancer selection from semi-idle packages only
4619 * when power-aware load balancing is enabled
4620 */
4621 if (!(sched_smt_power_savings || sched_mc_power_savings))
4622 goto out_done;
4623
4624 /*
4625 * Optimize for the case when we have no idle CPUs or only one
4626 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4627 */
4628 if (cpumask_weight(nohz.cpu_mask) < 2)
4629 goto out_done;
4630
4631 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4632 ilb_group = sd->groups;
4633
4634 do {
4635 if (is_semi_idle_group(ilb_group))
4636 return cpumask_first(nohz.ilb_grp_nohz_mask);
4637
4638 ilb_group = ilb_group->next;
4639
4640 } while (ilb_group != sd->groups);
4641 }
4642
4643out_done:
4644 return cpumask_first(nohz.cpu_mask);
4645}
4646#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4647static inline int find_new_ilb(int call_cpu)
4648{
Gautham R Shenoy6e29ec52009-04-21 08:40:49 +05304649 return cpumask_first(nohz.cpu_mask);
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304650}
4651#endif
4652
Christoph Lameter7835b982006-12-10 02:20:22 -08004653/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004654 * This routine will try to nominate the ilb (idle load balancing)
4655 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4656 * load balancing on behalf of all those cpus. If all the cpus in the system
4657 * go into this tickless mode, then there will be no ilb owner (as there is
4658 * no need for one) and all the cpus will sleep till the next wakeup event
4659 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08004660 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004661 * For the ilb owner, tick is not stopped. And this tick will be used
4662 * for idle load balancing. ilb owner will still be part of
4663 * nohz.cpu_mask..
4664 *
4665 * While stopping the tick, this cpu will become the ilb owner if there
4666 * is no other owner. And will be the owner till that cpu becomes busy
4667 * or if all cpus in the system stop their ticks at which point
4668 * there is no need for ilb owner.
4669 *
4670 * When the ilb owner becomes busy, it nominates another owner, during the
4671 * next busy scheduler_tick()
4672 */
4673int select_nohz_load_balancer(int stop_tick)
4674{
4675 int cpu = smp_processor_id();
4676
4677 if (stop_tick) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004678 cpu_rq(cpu)->in_nohz_recently = 1;
4679
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004680 if (!cpu_active(cpu)) {
4681 if (atomic_read(&nohz.load_balancer) != cpu)
4682 return 0;
4683
4684 /*
4685 * If we are going offline and still the leader,
4686 * give up!
4687 */
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004688 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4689 BUG();
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004690
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004691 return 0;
4692 }
4693
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004694 cpumask_set_cpu(cpu, nohz.cpu_mask);
4695
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004696 /* time for ilb owner also to sleep */
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304697 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004698 if (atomic_read(&nohz.load_balancer) == cpu)
4699 atomic_set(&nohz.load_balancer, -1);
4700 return 0;
4701 }
4702
4703 if (atomic_read(&nohz.load_balancer) == -1) {
4704 /* make me the ilb owner */
4705 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4706 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304707 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4708 int new_ilb;
4709
4710 if (!(sched_smt_power_savings ||
4711 sched_mc_power_savings))
4712 return 1;
4713 /*
4714 * Check to see if there is a more power-efficient
4715 * ilb.
4716 */
4717 new_ilb = find_new_ilb(cpu);
4718 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4719 atomic_set(&nohz.load_balancer, -1);
4720 resched_cpu(new_ilb);
4721 return 0;
4722 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004723 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304724 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004725 } else {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304726 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004727 return 0;
4728
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304729 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004730
4731 if (atomic_read(&nohz.load_balancer) == cpu)
4732 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4733 BUG();
4734 }
4735 return 0;
4736}
4737#endif
4738
4739static DEFINE_SPINLOCK(balancing);
4740
4741/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004742 * It checks each scheduling domain to see if it is due to be balanced,
4743 * and initiates a balancing operation if so.
4744 *
4745 * Balancing parameters are set up in arch_init_sched_domains.
4746 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004747static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08004748{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004749 int balance = 1;
4750 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08004751 unsigned long interval;
4752 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004753 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08004754 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004755 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004756 int need_serialize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004757
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004758 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004759 if (!(sd->flags & SD_LOAD_BALANCE))
4760 continue;
4761
4762 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004763 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004764 interval *= sd->busy_factor;
4765
4766 /* scale ms to jiffies */
4767 interval = msecs_to_jiffies(interval);
4768 if (unlikely(!interval))
4769 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02004770 if (interval > HZ*NR_CPUS/10)
4771 interval = HZ*NR_CPUS/10;
4772
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004773 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004774
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004775 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08004776 if (!spin_trylock(&balancing))
4777 goto out;
4778 }
4779
Christoph Lameterc9819f42006-12-10 02:20:25 -08004780 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304781 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004782 /*
4783 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07004784 * longer idle, or one of our SMT siblings is
4785 * not idle.
4786 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004787 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004788 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004789 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004791 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08004792 spin_unlock(&balancing);
4793out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02004794 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08004795 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004796 update_next_balance = 1;
4797 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004798
4799 /*
4800 * Stop the load balance at this level. There is another
4801 * CPU in our sched group which is doing load balancing more
4802 * actively.
4803 */
4804 if (!balance)
4805 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004806 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02004807
4808 /*
4809 * next_balance will be updated only when there is a need.
4810 * When the cpu is attached to null domain for ex, it will not be
4811 * updated.
4812 */
4813 if (likely(update_next_balance))
4814 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004815}
4816
4817/*
4818 * run_rebalance_domains is triggered when needed from the scheduler tick.
4819 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4820 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4821 */
4822static void run_rebalance_domains(struct softirq_action *h)
4823{
Ingo Molnardd41f592007-07-09 18:51:59 +02004824 int this_cpu = smp_processor_id();
4825 struct rq *this_rq = cpu_rq(this_cpu);
4826 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4827 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004828
Ingo Molnardd41f592007-07-09 18:51:59 +02004829 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004830
4831#ifdef CONFIG_NO_HZ
4832 /*
4833 * If this cpu is the owner for idle load balancing, then do the
4834 * balancing on behalf of the other idle cpus whose ticks are
4835 * stopped.
4836 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004837 if (this_rq->idle_at_tick &&
4838 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004839 struct rq *rq;
4840 int balance_cpu;
4841
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304842 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4843 if (balance_cpu == this_cpu)
4844 continue;
4845
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004846 /*
4847 * If this cpu gets work to do, stop the load balancing
4848 * work being done for other cpus. Next load
4849 * balancing owner will pick it up.
4850 */
4851 if (need_resched())
4852 break;
4853
Oleg Nesterovde0cf892007-08-12 18:08:19 +02004854 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004855
4856 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004857 if (time_after(this_rq->next_balance, rq->next_balance))
4858 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004859 }
4860 }
4861#endif
4862}
4863
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004864static inline int on_null_domain(int cpu)
4865{
4866 return !rcu_dereference(cpu_rq(cpu)->sd);
4867}
4868
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004869/*
4870 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4871 *
4872 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4873 * idle load balancing owner or decide to stop the periodic load balancing,
4874 * if the whole system is idle.
4875 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004876static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004877{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004878#ifdef CONFIG_NO_HZ
4879 /*
4880 * If we were in the nohz mode recently and busy at the current
4881 * scheduler tick, then check if we need to nominate new idle
4882 * load balancer.
4883 */
4884 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4885 rq->in_nohz_recently = 0;
4886
4887 if (atomic_read(&nohz.load_balancer) == cpu) {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304888 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004889 atomic_set(&nohz.load_balancer, -1);
4890 }
4891
4892 if (atomic_read(&nohz.load_balancer) == -1) {
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304893 int ilb = find_new_ilb(cpu);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004894
Mike Travis434d53b2008-04-04 18:11:04 -07004895 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004896 resched_cpu(ilb);
4897 }
4898 }
4899
4900 /*
4901 * If this cpu is idle and doing idle load balancing for all the
4902 * cpus with ticks stopped, is it time for that to stop?
4903 */
4904 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304905 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004906 resched_cpu(cpu);
4907 return;
4908 }
4909
4910 /*
4911 * If this cpu is idle and the idle load balancing is done by
4912 * someone else, then no need raise the SCHED_SOFTIRQ
4913 */
4914 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304915 cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004916 return;
4917#endif
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004918 /* Don't need to rebalance while attached to NULL domain */
4919 if (time_after_eq(jiffies, rq->next_balance) &&
4920 likely(!on_null_domain(cpu)))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004921 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004922}
Ingo Molnardd41f592007-07-09 18:51:59 +02004923
4924#else /* CONFIG_SMP */
4925
Linus Torvalds1da177e2005-04-16 15:20:36 -07004926/*
4927 * on UP we do not need to balance between CPUs:
4928 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004929static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004930{
4931}
Ingo Molnardd41f592007-07-09 18:51:59 +02004932
Linus Torvalds1da177e2005-04-16 15:20:36 -07004933#endif
4934
Linus Torvalds1da177e2005-04-16 15:20:36 -07004935DEFINE_PER_CPU(struct kernel_stat, kstat);
4936
4937EXPORT_PER_CPU_SYMBOL(kstat);
4938
4939/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004940 * Return any ns on the sched_clock that have not yet been accounted in
Frank Mayharf06febc2008-09-12 09:54:39 -07004941 * @p in case that task is currently running.
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004942 *
4943 * Called with task_rq_lock() held on @rq.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004944 */
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004945static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4946{
4947 u64 ns = 0;
4948
4949 if (task_current(rq, p)) {
4950 update_rq_clock(rq);
4951 ns = rq->clock - p->se.exec_start;
4952 if ((s64)ns < 0)
4953 ns = 0;
4954 }
4955
4956 return ns;
4957}
4958
Frank Mayharbb34d922008-09-12 09:54:39 -07004959unsigned long long task_delta_exec(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004960{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004961 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004962 struct rq *rq;
Frank Mayharbb34d922008-09-12 09:54:39 -07004963 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004964
Ingo Molnar41b86e92007-07-09 18:51:58 +02004965 rq = task_rq_lock(p, &flags);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004966 ns = do_task_delta_exec(p, rq);
4967 task_rq_unlock(rq, &flags);
Ingo Molnar15084872008-09-30 08:28:17 +02004968
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004969 return ns;
4970}
Frank Mayharf06febc2008-09-12 09:54:39 -07004971
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004972/*
4973 * Return accounted runtime for the task.
4974 * In case the task is currently running, return the runtime plus current's
4975 * pending runtime that have not been accounted yet.
4976 */
4977unsigned long long task_sched_runtime(struct task_struct *p)
4978{
4979 unsigned long flags;
4980 struct rq *rq;
4981 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004982
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004983 rq = task_rq_lock(p, &flags);
4984 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4985 task_rq_unlock(rq, &flags);
4986
4987 return ns;
4988}
4989
4990/*
4991 * Return sum_exec_runtime for the thread group.
4992 * In case the task is currently running, return the sum plus current's
4993 * pending runtime that have not been accounted yet.
4994 *
4995 * Note that the thread group might have other running tasks as well,
4996 * so the return value not includes other pending runtime that other
4997 * running tasks might have.
4998 */
4999unsigned long long thread_group_sched_runtime(struct task_struct *p)
5000{
5001 struct task_cputime totals;
5002 unsigned long flags;
5003 struct rq *rq;
5004 u64 ns;
5005
5006 rq = task_rq_lock(p, &flags);
5007 thread_group_cputime(p, &totals);
5008 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005009 task_rq_unlock(rq, &flags);
5010
5011 return ns;
5012}
5013
5014/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005015 * Account user cpu time to a process.
5016 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07005017 * @cputime: the cpu time spent in user space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005018 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07005019 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005020void account_user_time(struct task_struct *p, cputime_t cputime,
5021 cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005022{
5023 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5024 cputime64_t tmp;
5025
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005026 /* Add user time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005027 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005028 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005029 account_group_user_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005030
5031 /* Add user time to cpustat. */
5032 tmp = cputime_to_cputime64(cputime);
5033 if (TASK_NICE(p) > 0)
5034 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5035 else
5036 cpustat->user = cputime64_add(cpustat->user, tmp);
Bharata B Raoef12fef2009-03-31 10:02:22 +05305037
5038 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
Jonathan Lim49b5cf32008-07-25 01:48:40 -07005039 /* Account for user time used */
5040 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005041}
5042
5043/*
Laurent Vivier94886b82007-10-15 17:00:19 +02005044 * Account guest cpu time to a process.
5045 * @p: the process that the cpu time gets accounted to
5046 * @cputime: the cpu time spent in virtual machine since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005047 * @cputime_scaled: cputime scaled by cpu frequency
Laurent Vivier94886b82007-10-15 17:00:19 +02005048 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005049static void account_guest_time(struct task_struct *p, cputime_t cputime,
5050 cputime_t cputime_scaled)
Laurent Vivier94886b82007-10-15 17:00:19 +02005051{
5052 cputime64_t tmp;
5053 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5054
5055 tmp = cputime_to_cputime64(cputime);
5056
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005057 /* Add guest time to process. */
Laurent Vivier94886b82007-10-15 17:00:19 +02005058 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005059 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005060 account_group_user_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02005061 p->gtime = cputime_add(p->gtime, cputime);
5062
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005063 /* Add guest time to cpustat. */
Ryota Ozakice0e7b22009-10-24 01:20:10 +09005064 if (TASK_NICE(p) > 0) {
5065 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5066 cpustat->guest_nice = cputime64_add(cpustat->guest_nice, tmp);
5067 } else {
5068 cpustat->user = cputime64_add(cpustat->user, tmp);
5069 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5070 }
Laurent Vivier94886b82007-10-15 17:00:19 +02005071}
5072
5073/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005074 * Account system cpu time to a process.
5075 * @p: the process that the cpu time gets accounted to
5076 * @hardirq_offset: the offset to subtract from hardirq_count()
5077 * @cputime: the cpu time spent in kernel space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005078 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07005079 */
5080void account_system_time(struct task_struct *p, int hardirq_offset,
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005081 cputime_t cputime, cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005082{
5083 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005084 cputime64_t tmp;
5085
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005086 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005087 account_guest_time(p, cputime, cputime_scaled);
Harvey Harrison983ed7a2008-04-24 18:17:55 -07005088 return;
5089 }
Laurent Vivier94886b82007-10-15 17:00:19 +02005090
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005091 /* Add system time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005092 p->stime = cputime_add(p->stime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01005093 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07005094 account_group_system_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005095
5096 /* Add system time to cpustat. */
5097 tmp = cputime_to_cputime64(cputime);
5098 if (hardirq_count() - hardirq_offset)
5099 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5100 else if (softirq_count())
5101 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005102 else
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005103 cpustat->system = cputime64_add(cpustat->system, tmp);
5104
Bharata B Raoef12fef2009-03-31 10:02:22 +05305105 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5106
Linus Torvalds1da177e2005-04-16 15:20:36 -07005107 /* Account for system time used */
5108 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005109}
5110
5111/*
5112 * Account for involuntary wait time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005113 * @steal: the cpu time spent in involuntary wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005114 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005115void account_steal_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005116{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005117 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005118 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5119
5120 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005121}
5122
Christoph Lameter7835b982006-12-10 02:20:22 -08005123/*
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005124 * Account for idle time.
5125 * @cputime: the cpu time spent in idle wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005127void account_idle_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005128{
5129 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005130 cputime64_t cputime64 = cputime_to_cputime64(cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131 struct rq *rq = this_rq();
5132
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005133 if (atomic_read(&rq->nr_iowait) > 0)
5134 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5135 else
5136 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
Christoph Lameter7835b982006-12-10 02:20:22 -08005137}
5138
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005139#ifndef CONFIG_VIRT_CPU_ACCOUNTING
5140
5141/*
5142 * Account a single tick of cpu time.
5143 * @p: the process that the cpu time gets accounted to
5144 * @user_tick: indicates if the tick is a user or a system tick
5145 */
5146void account_process_tick(struct task_struct *p, int user_tick)
5147{
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005148 cputime_t one_jiffy_scaled = cputime_to_scaled(cputime_one_jiffy);
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005149 struct rq *rq = this_rq();
5150
5151 if (user_tick)
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005152 account_user_time(p, cputime_one_jiffy, one_jiffy_scaled);
Eric Dumazetf5f293a2009-04-29 14:44:49 +02005153 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005154 account_system_time(p, HARDIRQ_OFFSET, cputime_one_jiffy,
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005155 one_jiffy_scaled);
5156 else
Stanislaw Gruszkaa42548a2009-07-29 12:15:29 +02005157 account_idle_time(cputime_one_jiffy);
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005158}
5159
5160/*
5161 * Account multiple ticks of steal time.
5162 * @p: the process from which the cpu time has been stolen
5163 * @ticks: number of stolen ticks
5164 */
5165void account_steal_ticks(unsigned long ticks)
5166{
5167 account_steal_time(jiffies_to_cputime(ticks));
5168}
5169
5170/*
5171 * Account multiple ticks of idle time.
5172 * @ticks: number of stolen ticks
5173 */
5174void account_idle_ticks(unsigned long ticks)
5175{
5176 account_idle_time(jiffies_to_cputime(ticks));
5177}
5178
5179#endif
5180
Christoph Lameter7835b982006-12-10 02:20:22 -08005181/*
Balbir Singh49048622008-09-05 18:12:23 +02005182 * Use precise platform statistics if available:
5183 */
5184#ifdef CONFIG_VIRT_CPU_ACCOUNTING
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005185void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005186{
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005187 *ut = p->utime;
5188 *st = p->stime;
Balbir Singh49048622008-09-05 18:12:23 +02005189}
5190
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005191void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005192{
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005193 struct task_cputime cputime;
5194
5195 thread_group_cputime(p, &cputime);
5196
5197 *ut = cputime.utime;
5198 *st = cputime.stime;
Balbir Singh49048622008-09-05 18:12:23 +02005199}
5200#else
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005201
5202#ifndef nsecs_to_cputime
Hidetoshi Setob7b20df92009-11-26 14:49:27 +09005203# define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005204#endif
5205
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005206void task_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
Balbir Singh49048622008-09-05 18:12:23 +02005207{
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005208 cputime_t rtime, utime = p->utime, total = cputime_add(utime, p->stime);
Balbir Singh49048622008-09-05 18:12:23 +02005209
5210 /*
5211 * Use CFS's precise accounting:
5212 */
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005213 rtime = nsecs_to_cputime(p->se.sum_exec_runtime);
Balbir Singh49048622008-09-05 18:12:23 +02005214
5215 if (total) {
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005216 u64 temp;
Balbir Singh49048622008-09-05 18:12:23 +02005217
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005218 temp = (u64)(rtime * utime);
Balbir Singh49048622008-09-05 18:12:23 +02005219 do_div(temp, total);
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005220 utime = (cputime_t)temp;
5221 } else
5222 utime = rtime;
Balbir Singh49048622008-09-05 18:12:23 +02005223
5224 /*
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005225 * Compare with previous values, to keep monotonicity:
Balbir Singh49048622008-09-05 18:12:23 +02005226 */
Hidetoshi Seto761b1d22009-11-12 13:33:45 +09005227 p->prev_utime = max(p->prev_utime, utime);
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005228 p->prev_stime = max(p->prev_stime, cputime_sub(rtime, p->prev_utime));
Balbir Singh49048622008-09-05 18:12:23 +02005229
Hidetoshi Setod99ca3b2009-12-02 17:26:47 +09005230 *ut = p->prev_utime;
5231 *st = p->prev_stime;
Hidetoshi Setod180c5b2009-11-26 14:48:30 +09005232}
Balbir Singh49048622008-09-05 18:12:23 +02005233
Hidetoshi Seto0cf55e12009-12-02 17:28:07 +09005234/*
5235 * Must be called with siglock held.
5236 */
5237void thread_group_times(struct task_struct *p, cputime_t *ut, cputime_t *st)
5238{
5239 struct signal_struct *sig = p->signal;
5240 struct task_cputime cputime;
5241 cputime_t rtime, utime, total;
5242
5243 thread_group_cputime(p, &cputime);
5244
5245 total = cputime_add(cputime.utime, cputime.stime);
5246 rtime = nsecs_to_cputime(cputime.sum_exec_runtime);
5247
5248 if (total) {
5249 u64 temp;
5250
5251 temp = (u64)(rtime * cputime.utime);
5252 do_div(temp, total);
5253 utime = (cputime_t)temp;
5254 } else
5255 utime = rtime;
5256
5257 sig->prev_utime = max(sig->prev_utime, utime);
5258 sig->prev_stime = max(sig->prev_stime,
5259 cputime_sub(rtime, sig->prev_utime));
5260
5261 *ut = sig->prev_utime;
5262 *st = sig->prev_stime;
Balbir Singh49048622008-09-05 18:12:23 +02005263}
5264#endif
5265
Balbir Singh49048622008-09-05 18:12:23 +02005266/*
Christoph Lameter7835b982006-12-10 02:20:22 -08005267 * This function gets called by the timer code, with HZ frequency.
5268 * We call it with interrupts disabled.
5269 *
5270 * It also gets called by the fork code, when changing the parent's
5271 * timeslices.
5272 */
5273void scheduler_tick(void)
5274{
Christoph Lameter7835b982006-12-10 02:20:22 -08005275 int cpu = smp_processor_id();
5276 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005277 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005278
5279 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08005280
Ingo Molnardd41f592007-07-09 18:51:59 +02005281 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005282 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02005283 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01005284 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02005285 spin_unlock(&rq->lock);
5286
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005287 perf_event_task_tick(curr, cpu);
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02005288
Christoph Lametere418e1c2006-12-10 02:20:23 -08005289#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02005290 rq->idle_at_tick = idle_cpu(cpu);
5291 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08005292#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293}
5294
Lai Jiangshan132380a2009-04-02 14:18:25 +08005295notrace unsigned long get_parent_ip(unsigned long addr)
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005296{
5297 if (in_lock_functions(addr)) {
5298 addr = CALLER_ADDR2;
5299 if (in_lock_functions(addr))
5300 addr = CALLER_ADDR3;
5301 }
5302 return addr;
5303}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005304
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05005305#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5306 defined(CONFIG_PREEMPT_TRACER))
5307
Srinivasa Ds43627582008-02-23 15:24:04 -08005308void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005310#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005311 /*
5312 * Underflow?
5313 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005314 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5315 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005316#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005317 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005318#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005319 /*
5320 * Spinlock count overflowing soon?
5321 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005322 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5323 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005324#endif
5325 if (preempt_count() == val)
5326 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005327}
5328EXPORT_SYMBOL(add_preempt_count);
5329
Srinivasa Ds43627582008-02-23 15:24:04 -08005330void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005331{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005332#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005333 /*
5334 * Underflow?
5335 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01005336 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005337 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338 /*
5339 * Is the spinlock portion underflowing?
5340 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005341 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5342 !(preempt_count() & PREEMPT_MASK)))
5343 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005344#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005345
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005346 if (preempt_count() == val)
5347 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005348 preempt_count() -= val;
5349}
5350EXPORT_SYMBOL(sub_preempt_count);
5351
5352#endif
5353
5354/*
Ingo Molnardd41f592007-07-09 18:51:59 +02005355 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005356 */
Ingo Molnardd41f592007-07-09 18:51:59 +02005357static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005358{
Satyam Sharma838225b2007-10-24 18:23:50 +02005359 struct pt_regs *regs = get_irq_regs();
5360
5361 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5362 prev->comm, prev->pid, preempt_count());
5363
Ingo Molnardd41f592007-07-09 18:51:59 +02005364 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07005365 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02005366 if (irqs_disabled())
5367 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02005368
5369 if (regs)
5370 show_regs(regs);
5371 else
5372 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02005373}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005374
Ingo Molnardd41f592007-07-09 18:51:59 +02005375/*
5376 * Various schedule()-time debugging checks and statistics:
5377 */
5378static inline void schedule_debug(struct task_struct *prev)
5379{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005380 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005381 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07005382 * schedule() atomically, we ignore that path for now.
5383 * Otherwise, whine if we are scheduling when we should not be.
5384 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02005385 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02005386 __schedule_bug(prev);
5387
Linus Torvalds1da177e2005-04-16 15:20:36 -07005388 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5389
Ingo Molnar2d723762007-10-15 17:00:12 +02005390 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005391#ifdef CONFIG_SCHEDSTATS
5392 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02005393 schedstat_inc(this_rq(), bkl_count);
5394 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005395 }
5396#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005397}
5398
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005399static void put_prev_task(struct rq *rq, struct task_struct *p)
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005400{
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005401 u64 runtime = p->se.sum_exec_runtime - p->se.prev_sum_exec_runtime;
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005402
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005403 update_avg(&p->se.avg_running, runtime);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005404
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005405 if (p->state == TASK_RUNNING) {
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005406 /*
5407 * In order to avoid avg_overlap growing stale when we are
5408 * indeed overlapping and hence not getting put to sleep, grow
5409 * the avg_overlap on preemption.
5410 *
5411 * We use the average preemption runtime because that
5412 * correlates to the amount of cache footprint a task can
5413 * build up.
5414 */
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005415 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5416 update_avg(&p->se.avg_overlap, runtime);
5417 } else {
5418 update_avg(&p->se.avg_running, 0);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005419 }
Peter Zijlstraad4b78b2009-09-16 12:31:31 +02005420 p->sched_class->put_prev_task(rq, p);
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005421}
5422
Ingo Molnardd41f592007-07-09 18:51:59 +02005423/*
5424 * Pick up the highest-prio task:
5425 */
5426static inline struct task_struct *
Wang Chenb67802e2009-03-02 13:55:26 +08005427pick_next_task(struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02005428{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02005429 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02005430 struct task_struct *p;
5431
5432 /*
5433 * Optimization: we know that if all tasks are in
5434 * the fair class we can call that function directly:
5435 */
5436 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005437 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005438 if (likely(p))
5439 return p;
5440 }
5441
5442 class = sched_class_highest;
5443 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005444 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005445 if (p)
5446 return p;
5447 /*
5448 * Will never be NULL as the idle class always
5449 * returns a non-NULL p:
5450 */
5451 class = class->next;
5452 }
5453}
5454
5455/*
5456 * schedule() is the main scheduler function.
5457 */
Peter Zijlstraff743342009-03-13 12:21:26 +01005458asmlinkage void __sched schedule(void)
Ingo Molnardd41f592007-07-09 18:51:59 +02005459{
5460 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08005461 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02005462 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02005463 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02005464
Peter Zijlstraff743342009-03-13 12:21:26 +01005465need_resched:
5466 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02005467 cpu = smp_processor_id();
5468 rq = cpu_rq(cpu);
Paul E. McKenneyd6714c22009-08-22 13:56:46 -07005469 rcu_sched_qs(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005470 prev = rq->curr;
5471 switch_count = &prev->nivcsw;
5472
Linus Torvalds1da177e2005-04-16 15:20:36 -07005473 release_kernel_lock(prev);
5474need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475
Ingo Molnardd41f592007-07-09 18:51:59 +02005476 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005477
Peter Zijlstra31656512008-07-18 18:01:23 +02005478 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02005479 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005480
Peter Zijlstra8cd162c2008-10-15 20:37:23 +02005481 spin_lock_irq(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005482 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02005483 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005484
Ingo Molnardd41f592007-07-09 18:51:59 +02005485 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04005486 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02005487 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04005488 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005489 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02005490 switch_count = &prev->nvcsw;
5491 }
5492
Gregory Haskins3f029d32009-07-29 11:08:47 -04005493 pre_schedule(rq, prev);
Steven Rostedtf65eda42008-01-25 21:08:07 +01005494
Ingo Molnardd41f592007-07-09 18:51:59 +02005495 if (unlikely(!rq->nr_running))
5496 idle_balance(cpu, rq);
5497
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005498 put_prev_task(rq, prev);
Wang Chenb67802e2009-03-02 13:55:26 +08005499 next = pick_next_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005500
Linus Torvalds1da177e2005-04-16 15:20:36 -07005501 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01005502 sched_info_switch(prev, next);
Ingo Molnarcdd6c482009-09-21 12:02:48 +02005503 perf_event_task_sched_out(prev, next, cpu);
David Simner673a90a2008-04-29 10:08:59 +01005504
Linus Torvalds1da177e2005-04-16 15:20:36 -07005505 rq->nr_switches++;
5506 rq->curr = next;
5507 ++*switch_count;
5508
Ingo Molnardd41f592007-07-09 18:51:59 +02005509 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005510 /*
5511 * the context switch might have flipped the stack from under
5512 * us, hence refresh the local variables.
5513 */
5514 cpu = smp_processor_id();
5515 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005516 } else
5517 spin_unlock_irq(&rq->lock);
5518
Gregory Haskins3f029d32009-07-29 11:08:47 -04005519 post_schedule(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005520
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005521 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005522 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005523
Linus Torvalds1da177e2005-04-16 15:20:36 -07005524 preempt_enable_no_resched();
Peter Zijlstraff743342009-03-13 12:21:26 +01005525 if (need_resched())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005526 goto need_resched;
5527}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528EXPORT_SYMBOL(schedule);
5529
Frederic Weisbeckerc08f7822009-12-02 20:49:17 +01005530#ifdef CONFIG_MUTEX_SPIN_ON_OWNER
Peter Zijlstra0d66bf62009-01-12 14:01:47 +01005531/*
5532 * Look out! "owner" is an entirely speculative pointer
5533 * access and not reliable.
5534 */
5535int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5536{
5537 unsigned int cpu;
5538 struct rq *rq;
5539
5540 if (!sched_feat(OWNER_SPIN))
5541 return 0;
5542
5543#ifdef CONFIG_DEBUG_PAGEALLOC
5544 /*
5545 * Need to access the cpu field knowing that
5546 * DEBUG_PAGEALLOC could have unmapped it if
5547 * the mutex owner just released it and exited.
5548 */
5549 if (probe_kernel_address(&owner->cpu, cpu))
5550 goto out;
5551#else
5552 cpu = owner->cpu;
5553#endif
5554
5555 /*
5556 * Even if the access succeeded (likely case),
5557 * the cpu field may no longer be valid.
5558 */
5559 if (cpu >= nr_cpumask_bits)
5560 goto out;
5561
5562 /*
5563 * We need to validate that we can do a
5564 * get_cpu() and that we have the percpu area.
5565 */
5566 if (!cpu_online(cpu))
5567 goto out;
5568
5569 rq = cpu_rq(cpu);
5570
5571 for (;;) {
5572 /*
5573 * Owner changed, break to re-assess state.
5574 */
5575 if (lock->owner != owner)
5576 break;
5577
5578 /*
5579 * Is that owner really running on that cpu?
5580 */
5581 if (task_thread_info(rq->curr) != owner || need_resched())
5582 return 0;
5583
5584 cpu_relax();
5585 }
5586out:
5587 return 1;
5588}
5589#endif
5590
Linus Torvalds1da177e2005-04-16 15:20:36 -07005591#ifdef CONFIG_PREEMPT
5592/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005593 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005594 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07005595 * occur there and call schedule directly.
5596 */
5597asmlinkage void __sched preempt_schedule(void)
5598{
5599 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005600
Linus Torvalds1da177e2005-04-16 15:20:36 -07005601 /*
5602 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005603 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07005605 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606 return;
5607
Andi Kleen3a5c3592007-10-15 17:00:14 +02005608 do {
5609 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005610 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005611 sub_preempt_count(PREEMPT_ACTIVE);
5612
5613 /*
5614 * Check again in case we missed a preemption opportunity
5615 * between schedule and now.
5616 */
5617 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005618 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005619}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005620EXPORT_SYMBOL(preempt_schedule);
5621
5622/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005623 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07005624 * off of irq context.
5625 * Note, that this is called and return with irqs disabled. This will
5626 * protect us against recursive calling from irq.
5627 */
5628asmlinkage void __sched preempt_schedule_irq(void)
5629{
5630 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005631
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005632 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005633 BUG_ON(ti->preempt_count || !irqs_disabled());
5634
Andi Kleen3a5c3592007-10-15 17:00:14 +02005635 do {
5636 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005637 local_irq_enable();
5638 schedule();
5639 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005640 sub_preempt_count(PREEMPT_ACTIVE);
5641
5642 /*
5643 * Check again in case we missed a preemption opportunity
5644 * between schedule and now.
5645 */
5646 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005647 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005648}
5649
5650#endif /* CONFIG_PREEMPT */
5651
Peter Zijlstra63859d42009-09-15 19:14:42 +02005652int default_wake_function(wait_queue_t *curr, unsigned mode, int wake_flags,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005653 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005654{
Peter Zijlstra63859d42009-09-15 19:14:42 +02005655 return try_to_wake_up(curr->private, mode, wake_flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005656}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005657EXPORT_SYMBOL(default_wake_function);
5658
5659/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005660 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5661 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07005662 * number) then we wake all the non-exclusive tasks and one exclusive task.
5663 *
5664 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005665 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07005666 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5667 */
Johannes Weiner78ddb082009-04-14 16:53:05 +02005668static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
Peter Zijlstra63859d42009-09-15 19:14:42 +02005669 int nr_exclusive, int wake_flags, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005670{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005671 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005672
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005673 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005674 unsigned flags = curr->flags;
5675
Peter Zijlstra63859d42009-09-15 19:14:42 +02005676 if (curr->func(curr, mode, wake_flags, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07005677 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005678 break;
5679 }
5680}
5681
5682/**
5683 * __wake_up - wake up threads blocked on a waitqueue.
5684 * @q: the waitqueue
5685 * @mode: which threads
5686 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07005687 * @key: is directly passed to the wakeup function
David Howells50fa6102009-04-28 15:01:38 +01005688 *
5689 * It may be assumed that this function implies a write memory barrier before
5690 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005691 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005692void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005693 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005694{
5695 unsigned long flags;
5696
5697 spin_lock_irqsave(&q->lock, flags);
5698 __wake_up_common(q, mode, nr_exclusive, 0, key);
5699 spin_unlock_irqrestore(&q->lock, flags);
5700}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005701EXPORT_SYMBOL(__wake_up);
5702
5703/*
5704 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5705 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005706void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005707{
5708 __wake_up_common(q, mode, 1, 0, NULL);
5709}
5710
Davide Libenzi4ede8162009-03-31 15:24:20 -07005711void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5712{
5713 __wake_up_common(q, mode, 1, 0, key);
5714}
5715
Linus Torvalds1da177e2005-04-16 15:20:36 -07005716/**
Davide Libenzi4ede8162009-03-31 15:24:20 -07005717 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005718 * @q: the waitqueue
5719 * @mode: which threads
5720 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Davide Libenzi4ede8162009-03-31 15:24:20 -07005721 * @key: opaque value to be passed to wakeup targets
Linus Torvalds1da177e2005-04-16 15:20:36 -07005722 *
5723 * The sync wakeup differs that the waker knows that it will schedule
5724 * away soon, so while the target thread will be woken up, it will not
5725 * be migrated to another CPU - ie. the two threads are 'synchronized'
5726 * with each other. This can prevent needless bouncing between CPUs.
5727 *
5728 * On UP it can prevent extra preemption.
David Howells50fa6102009-04-28 15:01:38 +01005729 *
5730 * It may be assumed that this function implies a write memory barrier before
5731 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732 */
Davide Libenzi4ede8162009-03-31 15:24:20 -07005733void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5734 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005735{
5736 unsigned long flags;
Peter Zijlstra7d478722009-09-14 19:55:44 +02005737 int wake_flags = WF_SYNC;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005738
5739 if (unlikely(!q))
5740 return;
5741
5742 if (unlikely(!nr_exclusive))
Peter Zijlstra7d478722009-09-14 19:55:44 +02005743 wake_flags = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005744
5745 spin_lock_irqsave(&q->lock, flags);
Peter Zijlstra7d478722009-09-14 19:55:44 +02005746 __wake_up_common(q, mode, nr_exclusive, wake_flags, key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005747 spin_unlock_irqrestore(&q->lock, flags);
5748}
Davide Libenzi4ede8162009-03-31 15:24:20 -07005749EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5750
5751/*
5752 * __wake_up_sync - see __wake_up_sync_key()
5753 */
5754void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5755{
5756 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5757}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005758EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5759
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005760/**
5761 * complete: - signals a single thread waiting on this completion
5762 * @x: holds the state of this particular completion
5763 *
5764 * This will wake up a single thread waiting on this completion. Threads will be
5765 * awakened in the same order in which they were queued.
5766 *
5767 * See also complete_all(), wait_for_completion() and related routines.
David Howells50fa6102009-04-28 15:01:38 +01005768 *
5769 * It may be assumed that this function implies a write memory barrier before
5770 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005771 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005772void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005773{
5774 unsigned long flags;
5775
5776 spin_lock_irqsave(&x->wait.lock, flags);
5777 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005778 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005779 spin_unlock_irqrestore(&x->wait.lock, flags);
5780}
5781EXPORT_SYMBOL(complete);
5782
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005783/**
5784 * complete_all: - signals all threads waiting on this completion
5785 * @x: holds the state of this particular completion
5786 *
5787 * This will wake up all threads waiting on this particular completion event.
David Howells50fa6102009-04-28 15:01:38 +01005788 *
5789 * It may be assumed that this function implies a write memory barrier before
5790 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005791 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005792void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005793{
5794 unsigned long flags;
5795
5796 spin_lock_irqsave(&x->wait.lock, flags);
5797 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005798 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005799 spin_unlock_irqrestore(&x->wait.lock, flags);
5800}
5801EXPORT_SYMBOL(complete_all);
5802
Andi Kleen8cbbe862007-10-15 17:00:14 +02005803static inline long __sched
5804do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005805{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806 if (!x->done) {
5807 DECLARE_WAITQUEUE(wait, current);
5808
5809 wait.flags |= WQ_FLAG_EXCLUSIVE;
5810 __add_wait_queue_tail(&x->wait, &wait);
5811 do {
Oleg Nesterov94d3d822008-08-20 16:54:41 -07005812 if (signal_pending_state(state, current)) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04005813 timeout = -ERESTARTSYS;
5814 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005815 }
5816 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005817 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005818 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005819 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005820 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005821 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005822 if (!x->done)
5823 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005824 }
5825 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04005826 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005827}
5828
5829static long __sched
5830wait_for_common(struct completion *x, long timeout, int state)
5831{
5832 might_sleep();
5833
5834 spin_lock_irq(&x->wait.lock);
5835 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005836 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005837 return timeout;
5838}
5839
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005840/**
5841 * wait_for_completion: - waits for completion of a task
5842 * @x: holds the state of this particular completion
5843 *
5844 * This waits to be signaled for completion of a specific task. It is NOT
5845 * interruptible and there is no timeout.
5846 *
5847 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5848 * and interrupt capability. Also see complete().
5849 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005850void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02005851{
5852 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005853}
5854EXPORT_SYMBOL(wait_for_completion);
5855
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005856/**
5857 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5858 * @x: holds the state of this particular completion
5859 * @timeout: timeout value in jiffies
5860 *
5861 * This waits for either a completion of a specific task to be signaled or for a
5862 * specified timeout to expire. The timeout is in jiffies. It is not
5863 * interruptible.
5864 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005865unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005866wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5867{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005868 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005869}
5870EXPORT_SYMBOL(wait_for_completion_timeout);
5871
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005872/**
5873 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5874 * @x: holds the state of this particular completion
5875 *
5876 * This waits for completion of a specific task to be signaled. It is
5877 * interruptible.
5878 */
Andi Kleen8cbbe862007-10-15 17:00:14 +02005879int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005880{
Andi Kleen51e97992007-10-18 21:32:55 +02005881 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5882 if (t == -ERESTARTSYS)
5883 return t;
5884 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005885}
5886EXPORT_SYMBOL(wait_for_completion_interruptible);
5887
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005888/**
5889 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5890 * @x: holds the state of this particular completion
5891 * @timeout: timeout value in jiffies
5892 *
5893 * This waits for either a completion of a specific task to be signaled or for a
5894 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5895 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005896unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005897wait_for_completion_interruptible_timeout(struct completion *x,
5898 unsigned long timeout)
5899{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005900 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005901}
5902EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5903
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005904/**
5905 * wait_for_completion_killable: - waits for completion of a task (killable)
5906 * @x: holds the state of this particular completion
5907 *
5908 * This waits to be signaled for completion of a specific task. It can be
5909 * interrupted by a kill signal.
5910 */
Matthew Wilcox009e5772007-12-06 12:29:54 -05005911int __sched wait_for_completion_killable(struct completion *x)
5912{
5913 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5914 if (t == -ERESTARTSYS)
5915 return t;
5916 return 0;
5917}
5918EXPORT_SYMBOL(wait_for_completion_killable);
5919
Dave Chinnerbe4de352008-08-15 00:40:44 -07005920/**
5921 * try_wait_for_completion - try to decrement a completion without blocking
5922 * @x: completion structure
5923 *
5924 * Returns: 0 if a decrement cannot be done without blocking
5925 * 1 if a decrement succeeded.
5926 *
5927 * If a completion is being used as a counting completion,
5928 * attempt to decrement the counter without blocking. This
5929 * enables us to avoid waiting if the resource the completion
5930 * is protecting is not available.
5931 */
5932bool try_wait_for_completion(struct completion *x)
5933{
5934 int ret = 1;
5935
5936 spin_lock_irq(&x->wait.lock);
5937 if (!x->done)
5938 ret = 0;
5939 else
5940 x->done--;
5941 spin_unlock_irq(&x->wait.lock);
5942 return ret;
5943}
5944EXPORT_SYMBOL(try_wait_for_completion);
5945
5946/**
5947 * completion_done - Test to see if a completion has any waiters
5948 * @x: completion structure
5949 *
5950 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5951 * 1 if there are no waiters.
5952 *
5953 */
5954bool completion_done(struct completion *x)
5955{
5956 int ret = 1;
5957
5958 spin_lock_irq(&x->wait.lock);
5959 if (!x->done)
5960 ret = 0;
5961 spin_unlock_irq(&x->wait.lock);
5962 return ret;
5963}
5964EXPORT_SYMBOL(completion_done);
5965
Andi Kleen8cbbe862007-10-15 17:00:14 +02005966static long __sched
5967sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02005968{
5969 unsigned long flags;
5970 wait_queue_t wait;
5971
5972 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005973
Andi Kleen8cbbe862007-10-15 17:00:14 +02005974 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005975
Andi Kleen8cbbe862007-10-15 17:00:14 +02005976 spin_lock_irqsave(&q->lock, flags);
5977 __add_wait_queue(q, &wait);
5978 spin_unlock(&q->lock);
5979 timeout = schedule_timeout(timeout);
5980 spin_lock_irq(&q->lock);
5981 __remove_wait_queue(q, &wait);
5982 spin_unlock_irqrestore(&q->lock, flags);
5983
5984 return timeout;
5985}
5986
5987void __sched interruptible_sleep_on(wait_queue_head_t *q)
5988{
5989 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005990}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005991EXPORT_SYMBOL(interruptible_sleep_on);
5992
Ingo Molnar0fec1712007-07-09 18:52:01 +02005993long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005994interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005995{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005996 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005997}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005998EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5999
Ingo Molnar0fec1712007-07-09 18:52:01 +02006000void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006001{
Andi Kleen8cbbe862007-10-15 17:00:14 +02006002 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006003}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006004EXPORT_SYMBOL(sleep_on);
6005
Ingo Molnar0fec1712007-07-09 18:52:01 +02006006long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006007{
Andi Kleen8cbbe862007-10-15 17:00:14 +02006008 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006009}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006010EXPORT_SYMBOL(sleep_on_timeout);
6011
Ingo Molnarb29739f2006-06-27 02:54:51 -07006012#ifdef CONFIG_RT_MUTEXES
6013
6014/*
6015 * rt_mutex_setprio - set the current priority of a task
6016 * @p: task
6017 * @prio: prio value (kernel-internal form)
6018 *
6019 * This function changes the 'effective' priority of a task. It does
6020 * not touch ->normal_prio like __setscheduler().
6021 *
6022 * Used by the rt_mutex code to implement priority inheritance logic.
6023 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006024void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07006025{
6026 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006027 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006028 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01006029 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006030
6031 BUG_ON(prio < 0 || prio > MAX_PRIO);
6032
6033 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006034 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006035
Andrew Mortond5f9f942007-05-08 20:27:06 -07006036 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006037 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006038 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006039 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02006040 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006041 if (running)
6042 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02006043
6044 if (rt_prio(prio))
6045 p->sched_class = &rt_sched_class;
6046 else
6047 p->sched_class = &fair_sched_class;
6048
Ingo Molnarb29739f2006-06-27 02:54:51 -07006049 p->prio = prio;
6050
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006051 if (running)
6052 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006053 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02006054 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006055
6056 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006057 }
6058 task_rq_unlock(rq, &flags);
6059}
6060
6061#endif
6062
Ingo Molnar36c8b582006-07-03 00:25:41 -07006063void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006064{
Ingo Molnardd41f592007-07-09 18:51:59 +02006065 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006066 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006067 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006068
6069 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6070 return;
6071 /*
6072 * We have to be careful, if called from sys_setpriority(),
6073 * the task might be in the middle of scheduling on another CPU.
6074 */
6075 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006076 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006077 /*
6078 * The RT priorities are set via sched_setscheduler(), but we still
6079 * allow the 'normal' nice value to be set - but as expected
6080 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02006081 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006082 */
Ingo Molnare05606d2007-07-09 18:51:59 +02006083 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006084 p->static_prio = NICE_TO_PRIO(nice);
6085 goto out_unlock;
6086 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006087 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02006088 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02006089 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006090
Linus Torvalds1da177e2005-04-16 15:20:36 -07006091 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07006092 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07006093 old_prio = p->prio;
6094 p->prio = effective_prio(p);
6095 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006096
Ingo Molnardd41f592007-07-09 18:51:59 +02006097 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02006098 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006099 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07006100 * If the task increased its priority or is running and
6101 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006102 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07006103 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006104 resched_task(rq->curr);
6105 }
6106out_unlock:
6107 task_rq_unlock(rq, &flags);
6108}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006109EXPORT_SYMBOL(set_user_nice);
6110
Matt Mackalle43379f2005-05-01 08:59:00 -07006111/*
6112 * can_nice - check if a task can reduce its nice value
6113 * @p: task
6114 * @nice: nice value
6115 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006116int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07006117{
Matt Mackall024f4742005-08-18 11:24:19 -07006118 /* convert nice value [19,-20] to rlimit style value [1,40] */
6119 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006120
Matt Mackalle43379f2005-05-01 08:59:00 -07006121 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6122 capable(CAP_SYS_NICE));
6123}
6124
Linus Torvalds1da177e2005-04-16 15:20:36 -07006125#ifdef __ARCH_WANT_SYS_NICE
6126
6127/*
6128 * sys_nice - change the priority of the current process.
6129 * @increment: priority increment
6130 *
6131 * sys_setpriority is a more generic, but much slower function that
6132 * does similar things.
6133 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006134SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006135{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006136 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006137
6138 /*
6139 * Setpriority might change our priority at the same moment.
6140 * We don't have to worry. Conceptually one call occurs first
6141 * and we have a single winner.
6142 */
Matt Mackalle43379f2005-05-01 08:59:00 -07006143 if (increment < -40)
6144 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006145 if (increment > 40)
6146 increment = 40;
6147
Américo Wang2b8f8362009-02-16 18:54:21 +08006148 nice = TASK_NICE(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006149 if (nice < -20)
6150 nice = -20;
6151 if (nice > 19)
6152 nice = 19;
6153
Matt Mackalle43379f2005-05-01 08:59:00 -07006154 if (increment < 0 && !can_nice(current, nice))
6155 return -EPERM;
6156
Linus Torvalds1da177e2005-04-16 15:20:36 -07006157 retval = security_task_setnice(current, nice);
6158 if (retval)
6159 return retval;
6160
6161 set_user_nice(current, nice);
6162 return 0;
6163}
6164
6165#endif
6166
6167/**
6168 * task_prio - return the priority value of a given task.
6169 * @p: the task in question.
6170 *
6171 * This is the priority value as seen by users in /proc.
6172 * RT tasks are offset by -200. Normal tasks are centered
6173 * around 0, value goes from -16 to +15.
6174 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006175int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006176{
6177 return p->prio - MAX_RT_PRIO;
6178}
6179
6180/**
6181 * task_nice - return the nice value of a given task.
6182 * @p: the task in question.
6183 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006184int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006185{
6186 return TASK_NICE(p);
6187}
Pavel Roskin150d8be2008-03-05 16:56:37 -05006188EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006189
6190/**
6191 * idle_cpu - is a given cpu idle currently?
6192 * @cpu: the processor in question.
6193 */
6194int idle_cpu(int cpu)
6195{
6196 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6197}
6198
Linus Torvalds1da177e2005-04-16 15:20:36 -07006199/**
6200 * idle_task - return the idle task for a given cpu.
6201 * @cpu: the processor in question.
6202 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006203struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006204{
6205 return cpu_rq(cpu)->idle;
6206}
6207
6208/**
6209 * find_process_by_pid - find a process with a matching PID value.
6210 * @pid: the pid in question.
6211 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02006212static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006213{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07006214 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006215}
6216
6217/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02006218static void
6219__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006220{
Ingo Molnardd41f592007-07-09 18:51:59 +02006221 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006222
Linus Torvalds1da177e2005-04-16 15:20:36 -07006223 p->policy = policy;
6224 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006225 p->normal_prio = normal_prio(p);
6226 /* we are holding p->pi_lock already */
6227 p->prio = rt_mutex_getprio(p);
Peter Zijlstraffd44db2009-11-10 20:12:01 +01006228 if (rt_prio(p->prio))
6229 p->sched_class = &rt_sched_class;
6230 else
6231 p->sched_class = &fair_sched_class;
Peter Williams2dd73a42006-06-27 02:54:34 -07006232 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006233}
6234
David Howellsc69e8d92008-11-14 10:39:19 +11006235/*
6236 * check the target process has a UID that matches the current process's
6237 */
6238static bool check_same_owner(struct task_struct *p)
6239{
6240 const struct cred *cred = current_cred(), *pcred;
6241 bool match;
6242
6243 rcu_read_lock();
6244 pcred = __task_cred(p);
6245 match = (cred->euid == pcred->euid ||
6246 cred->euid == pcred->uid);
6247 rcu_read_unlock();
6248 return match;
6249}
6250
Rusty Russell961ccdd2008-06-23 13:55:38 +10006251static int __sched_setscheduler(struct task_struct *p, int policy,
6252 struct sched_param *param, bool user)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006253{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006254 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006255 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01006256 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006257 struct rq *rq;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006258 int reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006259
Steven Rostedt66e53932006-06-27 02:54:44 -07006260 /* may grab non-irq protected spin_locks */
6261 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07006262recheck:
6263 /* double check policy once rq lock held */
Lennart Poetteringca94c442009-06-15 17:17:47 +02006264 if (policy < 0) {
6265 reset_on_fork = p->sched_reset_on_fork;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006266 policy = oldpolicy = p->policy;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006267 } else {
6268 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6269 policy &= ~SCHED_RESET_ON_FORK;
6270
6271 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6272 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6273 policy != SCHED_IDLE)
6274 return -EINVAL;
6275 }
6276
Linus Torvalds1da177e2005-04-16 15:20:36 -07006277 /*
6278 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02006279 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6280 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006281 */
6282 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006283 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04006284 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006285 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02006286 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006287 return -EINVAL;
6288
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006289 /*
6290 * Allow unprivileged RT tasks to decrease priority:
6291 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10006292 if (user && !capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02006293 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006294 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006295
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006296 if (!lock_task_sighand(p, &flags))
6297 return -ESRCH;
6298 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6299 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006300
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006301 /* can't set/change the rt policy */
6302 if (policy != p->policy && !rlim_rtprio)
6303 return -EPERM;
6304
6305 /* can't increase priority */
6306 if (param->sched_priority > p->rt_priority &&
6307 param->sched_priority > rlim_rtprio)
6308 return -EPERM;
6309 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006310 /*
6311 * Like positive nice levels, dont allow tasks to
6312 * move out of SCHED_IDLE either:
6313 */
6314 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6315 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006316
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006317 /* can't change other user's priorities */
David Howellsc69e8d92008-11-14 10:39:19 +11006318 if (!check_same_owner(p))
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006319 return -EPERM;
Lennart Poetteringca94c442009-06-15 17:17:47 +02006320
6321 /* Normal users shall not reset the sched_reset_on_fork flag */
6322 if (p->sched_reset_on_fork && !reset_on_fork)
6323 return -EPERM;
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006324 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006325
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006326 if (user) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006327#ifdef CONFIG_RT_GROUP_SCHED
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006328 /*
6329 * Do not allow realtime tasks into groups that have no runtime
6330 * assigned.
6331 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02006332 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6333 task_group(p)->rt_bandwidth.rt_runtime == 0)
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006334 return -EPERM;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006335#endif
6336
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006337 retval = security_task_setscheduler(p, policy, param);
6338 if (retval)
6339 return retval;
6340 }
6341
Linus Torvalds1da177e2005-04-16 15:20:36 -07006342 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07006343 * make sure no PI-waiters arrive (or leave) while we are
6344 * changing the priority of the task:
6345 */
6346 spin_lock_irqsave(&p->pi_lock, flags);
6347 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07006348 * To be able to change p->policy safely, the apropriate
6349 * runqueue lock must be held.
6350 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07006351 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006352 /* recheck policy now with rq lock held */
6353 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6354 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006355 __task_rq_unlock(rq);
6356 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006357 goto recheck;
6358 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02006359 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006360 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006361 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006362 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006363 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006364 if (running)
6365 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02006366
Lennart Poetteringca94c442009-06-15 17:17:47 +02006367 p->sched_reset_on_fork = reset_on_fork;
6368
Linus Torvalds1da177e2005-04-16 15:20:36 -07006369 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006370 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02006371
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006372 if (running)
6373 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006374 if (on_rq) {
6375 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006376
6377 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006378 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07006379 __task_rq_unlock(rq);
6380 spin_unlock_irqrestore(&p->pi_lock, flags);
6381
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07006382 rt_mutex_adjust_pi(p);
6383
Linus Torvalds1da177e2005-04-16 15:20:36 -07006384 return 0;
6385}
Rusty Russell961ccdd2008-06-23 13:55:38 +10006386
6387/**
6388 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6389 * @p: the task in question.
6390 * @policy: new policy.
6391 * @param: structure containing the new RT priority.
6392 *
6393 * NOTE that the task may be already dead.
6394 */
6395int sched_setscheduler(struct task_struct *p, int policy,
6396 struct sched_param *param)
6397{
6398 return __sched_setscheduler(p, policy, param, true);
6399}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006400EXPORT_SYMBOL_GPL(sched_setscheduler);
6401
Rusty Russell961ccdd2008-06-23 13:55:38 +10006402/**
6403 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6404 * @p: the task in question.
6405 * @policy: new policy.
6406 * @param: structure containing the new RT priority.
6407 *
6408 * Just like sched_setscheduler, only don't bother checking if the
6409 * current context has permission. For example, this is needed in
6410 * stop_machine(): we create temporary high priority worker threads,
6411 * but our caller might not have that capability.
6412 */
6413int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6414 struct sched_param *param)
6415{
6416 return __sched_setscheduler(p, policy, param, false);
6417}
6418
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006419static int
6420do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006421{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006422 struct sched_param lparam;
6423 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006424 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006425
6426 if (!param || pid < 0)
6427 return -EINVAL;
6428 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6429 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006430
6431 rcu_read_lock();
6432 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006433 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006434 if (p != NULL)
6435 retval = sched_setscheduler(p, policy, &lparam);
6436 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07006437
Linus Torvalds1da177e2005-04-16 15:20:36 -07006438 return retval;
6439}
6440
6441/**
6442 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6443 * @pid: the pid in question.
6444 * @policy: new policy.
6445 * @param: structure containing the new RT priority.
6446 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006447SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6448 struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006449{
Jason Baronc21761f2006-01-18 17:43:03 -08006450 /* negative values for policy are not valid */
6451 if (policy < 0)
6452 return -EINVAL;
6453
Linus Torvalds1da177e2005-04-16 15:20:36 -07006454 return do_sched_setscheduler(pid, policy, param);
6455}
6456
6457/**
6458 * sys_sched_setparam - set/change the RT priority of a thread
6459 * @pid: the pid in question.
6460 * @param: structure containing the new RT priority.
6461 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006462SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006463{
6464 return do_sched_setscheduler(pid, -1, param);
6465}
6466
6467/**
6468 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6469 * @pid: the pid in question.
6470 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006471SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006472{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006473 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006474 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006475
6476 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006477 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006478
6479 retval = -ESRCH;
6480 read_lock(&tasklist_lock);
6481 p = find_process_by_pid(pid);
6482 if (p) {
6483 retval = security_task_getscheduler(p);
6484 if (!retval)
Lennart Poetteringca94c442009-06-15 17:17:47 +02006485 retval = p->policy
6486 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006487 }
6488 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006489 return retval;
6490}
6491
6492/**
Lennart Poetteringca94c442009-06-15 17:17:47 +02006493 * sys_sched_getparam - get the RT priority of a thread
Linus Torvalds1da177e2005-04-16 15:20:36 -07006494 * @pid: the pid in question.
6495 * @param: structure containing the RT priority.
6496 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006497SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006498{
6499 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006500 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006501 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006502
6503 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006504 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006505
6506 read_lock(&tasklist_lock);
6507 p = find_process_by_pid(pid);
6508 retval = -ESRCH;
6509 if (!p)
6510 goto out_unlock;
6511
6512 retval = security_task_getscheduler(p);
6513 if (retval)
6514 goto out_unlock;
6515
6516 lp.sched_priority = p->rt_priority;
6517 read_unlock(&tasklist_lock);
6518
6519 /*
6520 * This one might sleep, we cannot do it with a spinlock held ...
6521 */
6522 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6523
Linus Torvalds1da177e2005-04-16 15:20:36 -07006524 return retval;
6525
6526out_unlock:
6527 read_unlock(&tasklist_lock);
6528 return retval;
6529}
6530
Rusty Russell96f874e2008-11-25 02:35:14 +10306531long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006532{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306533 cpumask_var_t cpus_allowed, new_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006534 struct task_struct *p;
6535 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006536
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006537 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006538 read_lock(&tasklist_lock);
6539
6540 p = find_process_by_pid(pid);
6541 if (!p) {
6542 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006543 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006544 return -ESRCH;
6545 }
6546
6547 /*
6548 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006549 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07006550 * usage count and then drop tasklist_lock.
6551 */
6552 get_task_struct(p);
6553 read_unlock(&tasklist_lock);
6554
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306555 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6556 retval = -ENOMEM;
6557 goto out_put_task;
6558 }
6559 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6560 retval = -ENOMEM;
6561 goto out_free_cpus_allowed;
6562 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006563 retval = -EPERM;
David Howellsc69e8d92008-11-14 10:39:19 +11006564 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006565 goto out_unlock;
6566
David Quigleye7834f82006-06-23 02:03:59 -07006567 retval = security_task_setscheduler(p, 0, NULL);
6568 if (retval)
6569 goto out_unlock;
6570
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306571 cpuset_cpus_allowed(p, cpus_allowed);
6572 cpumask_and(new_mask, in_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006573 again:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306574 retval = set_cpus_allowed_ptr(p, new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006575
Paul Menage8707d8b2007-10-18 23:40:22 -07006576 if (!retval) {
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306577 cpuset_cpus_allowed(p, cpus_allowed);
6578 if (!cpumask_subset(new_mask, cpus_allowed)) {
Paul Menage8707d8b2007-10-18 23:40:22 -07006579 /*
6580 * We must have raced with a concurrent cpuset
6581 * update. Just reset the cpus_allowed to the
6582 * cpuset's cpus_allowed
6583 */
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306584 cpumask_copy(new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006585 goto again;
6586 }
6587 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006588out_unlock:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306589 free_cpumask_var(new_mask);
6590out_free_cpus_allowed:
6591 free_cpumask_var(cpus_allowed);
6592out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006593 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006594 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006595 return retval;
6596}
6597
6598static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e2008-11-25 02:35:14 +10306599 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006600{
Rusty Russell96f874e2008-11-25 02:35:14 +10306601 if (len < cpumask_size())
6602 cpumask_clear(new_mask);
6603 else if (len > cpumask_size())
6604 len = cpumask_size();
6605
Linus Torvalds1da177e2005-04-16 15:20:36 -07006606 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6607}
6608
6609/**
6610 * sys_sched_setaffinity - set the cpu affinity of a process
6611 * @pid: pid of the process
6612 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6613 * @user_mask_ptr: user-space pointer to the new cpu mask
6614 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006615SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6616 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006617{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306618 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619 int retval;
6620
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306621 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6622 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006623
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306624 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6625 if (retval == 0)
6626 retval = sched_setaffinity(pid, new_mask);
6627 free_cpumask_var(new_mask);
6628 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006629}
6630
Rusty Russell96f874e2008-11-25 02:35:14 +10306631long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006632{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006633 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006634 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006635
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006636 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006637 read_lock(&tasklist_lock);
6638
6639 retval = -ESRCH;
6640 p = find_process_by_pid(pid);
6641 if (!p)
6642 goto out_unlock;
6643
David Quigleye7834f82006-06-23 02:03:59 -07006644 retval = security_task_getscheduler(p);
6645 if (retval)
6646 goto out_unlock;
6647
Rusty Russell96f874e2008-11-25 02:35:14 +10306648 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006649
6650out_unlock:
6651 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006652 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006653
Ulrich Drepper9531b622007-08-09 11:16:46 +02006654 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006655}
6656
6657/**
6658 * sys_sched_getaffinity - get the cpu affinity of a process
6659 * @pid: pid of the process
6660 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6661 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6662 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006663SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6664 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006665{
6666 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10306667 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006668
Rusty Russellf17c8602008-11-25 02:35:11 +10306669 if (len < cpumask_size())
Linus Torvalds1da177e2005-04-16 15:20:36 -07006670 return -EINVAL;
6671
Rusty Russellf17c8602008-11-25 02:35:11 +10306672 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6673 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006674
Rusty Russellf17c8602008-11-25 02:35:11 +10306675 ret = sched_getaffinity(pid, mask);
6676 if (ret == 0) {
6677 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6678 ret = -EFAULT;
6679 else
6680 ret = cpumask_size();
6681 }
6682 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006683
Rusty Russellf17c8602008-11-25 02:35:11 +10306684 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006685}
6686
6687/**
6688 * sys_sched_yield - yield the current processor to other threads.
6689 *
Ingo Molnardd41f592007-07-09 18:51:59 +02006690 * This function yields the current CPU to other tasks. If there are no
6691 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006692 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006693SYSCALL_DEFINE0(sched_yield)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006694{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006695 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006696
Ingo Molnar2d723762007-10-15 17:00:12 +02006697 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02006698 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006699
6700 /*
6701 * Since we are going to call schedule() anyway, there's
6702 * no need to preempt or enable interrupts:
6703 */
6704 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07006705 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006706 _raw_spin_unlock(&rq->lock);
6707 preempt_enable_no_resched();
6708
6709 schedule();
6710
6711 return 0;
6712}
6713
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006714static inline int should_resched(void)
6715{
6716 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6717}
6718
Andrew Mortone7b38402006-06-30 01:56:00 -07006719static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006720{
Frederic Weisbeckere7aaaa62009-07-16 15:44:29 +02006721 add_preempt_count(PREEMPT_ACTIVE);
6722 schedule();
6723 sub_preempt_count(PREEMPT_ACTIVE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006724}
6725
Herbert Xu02b67cc32008-01-25 21:08:28 +01006726int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006727{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006728 if (should_resched()) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006729 __cond_resched();
6730 return 1;
6731 }
6732 return 0;
6733}
Herbert Xu02b67cc32008-01-25 21:08:28 +01006734EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006735
6736/*
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006737 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
Linus Torvalds1da177e2005-04-16 15:20:36 -07006738 * call schedule, and on return reacquire the lock.
6739 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006740 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07006741 * operations here to prevent schedule() from being called twice (once via
6742 * spin_unlock(), once by hand).
6743 */
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006744int __cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006745{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006746 int resched = should_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07006747 int ret = 0;
6748
Peter Zijlstraf607c662009-07-20 19:16:29 +02006749 lockdep_assert_held(lock);
6750
Nick Piggin95c354f2008-01-30 13:31:20 +01006751 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006752 spin_unlock(lock);
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006753 if (resched)
Nick Piggin95c354f2008-01-30 13:31:20 +01006754 __cond_resched();
6755 else
6756 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07006757 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006758 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006759 }
Jan Kara6df3cec2005-06-13 15:52:32 -07006760 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006761}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006762EXPORT_SYMBOL(__cond_resched_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006763
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006764int __sched __cond_resched_softirq(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006765{
6766 BUG_ON(!in_softirq());
6767
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006768 if (should_resched()) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07006769 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006770 __cond_resched();
6771 local_bh_disable();
6772 return 1;
6773 }
6774 return 0;
6775}
Frederic Weisbecker613afbf2009-07-16 15:44:29 +02006776EXPORT_SYMBOL(__cond_resched_softirq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006777
Linus Torvalds1da177e2005-04-16 15:20:36 -07006778/**
6779 * yield - yield the current processor to other threads.
6780 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08006781 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07006782 * thread runnable and calls sys_sched_yield().
6783 */
6784void __sched yield(void)
6785{
6786 set_current_state(TASK_RUNNING);
6787 sys_sched_yield();
6788}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006789EXPORT_SYMBOL(yield);
6790
6791/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006792 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07006793 * that process accounting knows that this is a task in IO wait state.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006794 */
6795void __sched io_schedule(void)
6796{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006797 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006798
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006799 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006800 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006801 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006802 schedule();
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006803 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006804 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006805 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006806}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006807EXPORT_SYMBOL(io_schedule);
6808
6809long __sched io_schedule_timeout(long timeout)
6810{
Hitoshi Mitake54d35f22009-06-29 14:44:57 +09006811 struct rq *rq = raw_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006812 long ret;
6813
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006814 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006815 atomic_inc(&rq->nr_iowait);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006816 current->in_iowait = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006817 ret = schedule_timeout(timeout);
Arjan van de Ven8f0dfc32009-07-20 11:26:58 -07006818 current->in_iowait = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006819 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006820 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006821 return ret;
6822}
6823
6824/**
6825 * sys_sched_get_priority_max - return maximum RT priority.
6826 * @policy: scheduling class.
6827 *
6828 * this syscall returns the maximum rt_priority that can be used
6829 * by a given scheduling class.
6830 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006831SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006832{
6833 int ret = -EINVAL;
6834
6835 switch (policy) {
6836 case SCHED_FIFO:
6837 case SCHED_RR:
6838 ret = MAX_USER_RT_PRIO-1;
6839 break;
6840 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006841 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006842 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006843 ret = 0;
6844 break;
6845 }
6846 return ret;
6847}
6848
6849/**
6850 * sys_sched_get_priority_min - return minimum RT priority.
6851 * @policy: scheduling class.
6852 *
6853 * this syscall returns the minimum rt_priority that can be used
6854 * by a given scheduling class.
6855 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006856SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006857{
6858 int ret = -EINVAL;
6859
6860 switch (policy) {
6861 case SCHED_FIFO:
6862 case SCHED_RR:
6863 ret = 1;
6864 break;
6865 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006866 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006867 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006868 ret = 0;
6869 }
6870 return ret;
6871}
6872
6873/**
6874 * sys_sched_rr_get_interval - return the default timeslice of a process.
6875 * @pid: pid of the process.
6876 * @interval: userspace pointer to the timeslice value.
6877 *
6878 * this syscall writes the default timeslice value of a given process
6879 * into the user-space timespec buffer. A value of '0' means infinity.
6880 */
Heiko Carstens17da2bd2009-01-14 14:14:10 +01006881SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Heiko Carstens754fe8d2009-01-14 14:14:09 +01006882 struct timespec __user *, interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006883{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006884 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006885 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006886 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006887 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006888
6889 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006890 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006891
6892 retval = -ESRCH;
6893 read_lock(&tasklist_lock);
6894 p = find_process_by_pid(pid);
6895 if (!p)
6896 goto out_unlock;
6897
6898 retval = security_task_getscheduler(p);
6899 if (retval)
6900 goto out_unlock;
6901
Peter Williams0d721ce2009-09-21 01:31:53 +00006902 time_slice = p->sched_class->get_rr_interval(p);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006903
Linus Torvalds1da177e2005-04-16 15:20:36 -07006904 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006905 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006906 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006907 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006908
Linus Torvalds1da177e2005-04-16 15:20:36 -07006909out_unlock:
6910 read_unlock(&tasklist_lock);
6911 return retval;
6912}
6913
Steven Rostedt7c731e02008-05-12 21:20:41 +02006914static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006915
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006916void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006917{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006918 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006919 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006920
Linus Torvalds1da177e2005-04-16 15:20:36 -07006921 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006922 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07006923 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02006924#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07006925 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006926 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006927 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006928 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006929#else
6930 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006931 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006932 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006933 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006934#endif
6935#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05006936 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006937#endif
David Rientjesaa47b7e2009-05-04 01:38:05 -07006938 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6939 task_pid_nr(p), task_pid_nr(p->real_parent),
6940 (unsigned long)task_thread_info(p)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006941
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01006942 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006943}
6944
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006945void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006946{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006947 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006948
Ingo Molnar4bd77322007-07-11 21:21:47 +02006949#if BITS_PER_LONG == 32
6950 printk(KERN_INFO
6951 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006952#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02006953 printk(KERN_INFO
6954 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006955#endif
6956 read_lock(&tasklist_lock);
6957 do_each_thread(g, p) {
6958 /*
6959 * reset the NMI-timeout, listing all files on a slow
6960 * console might take alot of time:
6961 */
6962 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07006963 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006964 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006965 } while_each_thread(g, p);
6966
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07006967 touch_all_softlockup_watchdogs();
6968
Ingo Molnardd41f592007-07-09 18:51:59 +02006969#ifdef CONFIG_SCHED_DEBUG
6970 sysrq_sched_debug_show();
6971#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006972 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006973 /*
6974 * Only show locks if all tasks are dumped:
6975 */
Shmulik Ladkani93335a22009-11-25 15:23:41 +02006976 if (!state_filter)
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006977 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006978}
6979
Ingo Molnar1df21052007-07-09 18:51:58 +02006980void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6981{
Ingo Molnardd41f592007-07-09 18:51:59 +02006982 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02006983}
6984
Ingo Molnarf340c0d2005-06-28 16:40:42 +02006985/**
6986 * init_idle - set up an idle thread for a given CPU
6987 * @idle: task in question
6988 * @cpu: cpu the idle task belongs to
6989 *
6990 * NOTE: this function does not set the idle thread's NEED_RESCHED
6991 * flag, to make booting more robust.
6992 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07006993void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006994{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006995 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006996 unsigned long flags;
6997
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01006998 spin_lock_irqsave(&rq->lock, flags);
6999
Ingo Molnardd41f592007-07-09 18:51:59 +02007000 __sched_fork(idle);
7001 idle->se.exec_start = sched_clock();
7002
Ingo Molnarb29739f2006-06-27 02:54:51 -07007003 idle->prio = idle->normal_prio = MAX_PRIO;
Rusty Russell96f874e2008-11-25 02:35:14 +10307004 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
Ingo Molnardd41f592007-07-09 18:51:59 +02007005 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007006
Linus Torvalds1da177e2005-04-16 15:20:36 -07007007 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07007008#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7009 idle->oncpu = 1;
7010#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007011 spin_unlock_irqrestore(&rq->lock, flags);
7012
7013 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07007014#if defined(CONFIG_PREEMPT)
7015 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7016#else
Al Viroa1261f52005-11-13 16:06:55 -08007017 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07007018#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02007019 /*
7020 * The idle tasks have their own, simple scheduling class:
7021 */
7022 idle->sched_class = &idle_sched_class;
Frederic Weisbeckerfb526072008-11-25 21:07:04 +01007023 ftrace_graph_init_task(idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007024}
7025
7026/*
7027 * In a system that switches off the HZ timer nohz_cpu_mask
7028 * indicates which cpus entered this state. This is used
7029 * in the rcu update to wait only for active cpus. For system
7030 * which do not switch off the HZ timer nohz_cpu_mask should
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10307031 * always be CPU_BITS_NONE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007032 */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10307033cpumask_var_t nohz_cpu_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007034
Ingo Molnar19978ca2007-11-09 22:39:38 +01007035/*
7036 * Increase the granularity value when there are more CPUs,
7037 * because with more CPUs the 'effective latency' as visible
7038 * to users decreases. But the relationship is not linear,
7039 * so pick a second-best guess by going with the log2 of the
7040 * number of CPUs.
7041 *
7042 * This idea comes from the SD scheduler of Con Kolivas:
7043 */
7044static inline void sched_init_granularity(void)
7045{
7046 unsigned int factor = 1 + ilog2(num_online_cpus());
7047 const unsigned long limit = 200000000;
7048
7049 sysctl_sched_min_granularity *= factor;
7050 if (sysctl_sched_min_granularity > limit)
7051 sysctl_sched_min_granularity = limit;
7052
7053 sysctl_sched_latency *= factor;
7054 if (sysctl_sched_latency > limit)
7055 sysctl_sched_latency = limit;
7056
7057 sysctl_sched_wakeup_granularity *= factor;
Peter Zijlstra55cd5342008-08-04 08:54:26 +02007058
7059 sysctl_sched_shares_ratelimit *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01007060}
7061
Linus Torvalds1da177e2005-04-16 15:20:36 -07007062#ifdef CONFIG_SMP
7063/*
7064 * This is how migration works:
7065 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07007066 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07007067 * runqueue and wake up that CPU's migration thread.
7068 * 2) we down() the locked semaphore => thread blocks.
7069 * 3) migration thread wakes up (implicitly it forces the migrated
7070 * thread off the CPU)
7071 * 4) it gets the migration request and checks whether the migrated
7072 * task is still in the wrong runqueue.
7073 * 5) if it's in the wrong runqueue then the migration thread removes
7074 * it and puts it into the right queue.
7075 * 6) migration thread up()s the semaphore.
7076 * 7) we wake up and the migration is done.
7077 */
7078
7079/*
7080 * Change a given task's CPU affinity. Migrate the thread to a
7081 * proper CPU and schedule it away if the CPU it's executing on
7082 * is removed from the allowed bitmask.
7083 *
7084 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007085 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07007086 * call is not atomic; no spinlocks may be held.
7087 */
Rusty Russell96f874e2008-11-25 02:35:14 +10307088int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007089{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007090 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007091 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007092 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007093 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007094
7095 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e2008-11-25 02:35:14 +10307096 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007097 ret = -EINVAL;
7098 goto out;
7099 }
7100
David Rientjes9985b0b2008-06-05 12:57:11 -07007101 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
Rusty Russell96f874e2008-11-25 02:35:14 +10307102 !cpumask_equal(&p->cpus_allowed, new_mask))) {
David Rientjes9985b0b2008-06-05 12:57:11 -07007103 ret = -EINVAL;
7104 goto out;
7105 }
7106
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01007107 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007108 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01007109 else {
Rusty Russell96f874e2008-11-25 02:35:14 +10307110 cpumask_copy(&p->cpus_allowed, new_mask);
7111 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01007112 }
7113
Linus Torvalds1da177e2005-04-16 15:20:36 -07007114 /* Can the task run on the task's current CPU? If so, we're done */
Rusty Russell96f874e2008-11-25 02:35:14 +10307115 if (cpumask_test_cpu(task_cpu(p), new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007116 goto out;
7117
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307118 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007119 /* Need help from migration thread: drop lock and wait. */
Peter Zijlstra693525e2009-07-21 13:56:38 +02007120 struct task_struct *mt = rq->migration_thread;
7121
7122 get_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007123 task_rq_unlock(rq, &flags);
7124 wake_up_process(rq->migration_thread);
Peter Zijlstra693525e2009-07-21 13:56:38 +02007125 put_task_struct(mt);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007126 wait_for_completion(&req.done);
7127 tlb_migrate_finish(p->mm);
7128 return 0;
7129 }
7130out:
7131 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007132
Linus Torvalds1da177e2005-04-16 15:20:36 -07007133 return ret;
7134}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07007135EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007136
7137/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007138 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07007139 * this because either it can't run here any more (set_cpus_allowed()
7140 * away from this CPU, or CPU going down), or because we're
7141 * attempting to rebalance this task on exec (sched_exec).
7142 *
7143 * So we race with normal scheduler movements, but that's OK, as long
7144 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07007145 *
7146 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07007148static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007149{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007150 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02007151 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007152
Max Krasnyanskye761b772008-07-15 04:43:49 -07007153 if (unlikely(!cpu_active(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07007154 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007155
7156 rq_src = cpu_rq(src_cpu);
7157 rq_dest = cpu_rq(dest_cpu);
7158
7159 double_rq_lock(rq_src, rq_dest);
7160 /* Already moved. */
7161 if (task_cpu(p) != src_cpu)
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007162 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007163 /* Affinity changed (again). */
Rusty Russell96f874e2008-11-25 02:35:14 +10307164 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007165 goto fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007166
Ingo Molnardd41f592007-07-09 18:51:59 +02007167 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007168 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007169 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007170
Linus Torvalds1da177e2005-04-16 15:20:36 -07007171 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007172 if (on_rq) {
7173 activate_task(rq_dest, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02007174 check_preempt_curr(rq_dest, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007175 }
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007176done:
Kirill Korotaevefc30812006-06-27 02:54:32 -07007177 ret = 1;
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007178fail:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007179 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07007180 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007181}
7182
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007183#define RCU_MIGRATION_IDLE 0
7184#define RCU_MIGRATION_NEED_QS 1
7185#define RCU_MIGRATION_GOT_QS 2
7186#define RCU_MIGRATION_MUST_SYNC 3
7187
Linus Torvalds1da177e2005-04-16 15:20:36 -07007188/*
7189 * migration_thread - this is a highprio system thread that performs
7190 * thread migration by bumping thread off CPU then 'pushing' onto
7191 * another runqueue.
7192 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07007193static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007194{
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007195 int badcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007196 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007197 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007198
7199 rq = cpu_rq(cpu);
7200 BUG_ON(rq->migration_thread != current);
7201
7202 set_current_state(TASK_INTERRUPTIBLE);
7203 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007204 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007205 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007206
Linus Torvalds1da177e2005-04-16 15:20:36 -07007207 spin_lock_irq(&rq->lock);
7208
7209 if (cpu_is_offline(cpu)) {
7210 spin_unlock_irq(&rq->lock);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007211 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007212 }
7213
7214 if (rq->active_balance) {
7215 active_load_balance(rq, cpu);
7216 rq->active_balance = 0;
7217 }
7218
7219 head = &rq->migration_queue;
7220
7221 if (list_empty(head)) {
7222 spin_unlock_irq(&rq->lock);
7223 schedule();
7224 set_current_state(TASK_INTERRUPTIBLE);
7225 continue;
7226 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07007227 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007228 list_del_init(head->next);
7229
Paul E. McKenney03b042b2009-06-25 09:08:16 -07007230 if (req->task != NULL) {
7231 spin_unlock(&rq->lock);
7232 __migrate_task(req->task, cpu, req->dest_cpu);
7233 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7234 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7235 spin_unlock(&rq->lock);
7236 } else {
7237 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7238 spin_unlock(&rq->lock);
7239 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7240 }
Nick Piggin674311d2005-06-25 14:57:27 -07007241 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007242
7243 complete(&req->done);
7244 }
7245 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007246
Linus Torvalds1da177e2005-04-16 15:20:36 -07007247 return 0;
7248}
7249
7250#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007251
7252static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7253{
7254 int ret;
7255
7256 local_irq_disable();
7257 ret = __migrate_task(p, src_cpu, dest_cpu);
7258 local_irq_enable();
7259 return ret;
7260}
7261
Kirill Korotaev054b9102006-12-10 02:20:11 -08007262/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007263 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08007264 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007265static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007266{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007267 int dest_cpu;
Mike Travis6ca09df2008-12-31 18:08:45 -08007268 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007269
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307270again:
7271 /* Look for allowed, online CPU in same node. */
7272 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7273 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7274 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007275
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307276 /* Any allowed, online CPU? */
7277 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7278 if (dest_cpu < nr_cpu_ids)
7279 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007280
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307281 /* No more Mr. Nice Guy. */
7282 if (dest_cpu >= nr_cpu_ids) {
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307283 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7284 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07007285
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307286 /*
7287 * Don't tell them about moving exiting tasks or
7288 * kernel threads (both mm NULL), since they never
7289 * leave kernel.
7290 */
7291 if (p->mm && printk_ratelimit()) {
7292 printk(KERN_INFO "process %d (%s) no "
7293 "longer affine to cpu%d\n",
7294 task_pid_nr(p), p->comm, dead_cpu);
Andi Kleen3a5c3592007-10-15 17:00:14 +02007295 }
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307296 }
7297
7298move:
7299 /* It can have affinity changed while we were choosing. */
7300 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7301 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007302}
7303
7304/*
7305 * While a dead CPU has no uninterruptible tasks queued at this point,
7306 * it might still have a nonzero ->nr_uninterruptible counter, because
7307 * for performance reasons the counter is not stricly tracking tasks to
7308 * their home CPUs. So we just add the counter to another CPU's counter,
7309 * to keep the global sum constant after CPU-down:
7310 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07007311static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007312{
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307313 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007314 unsigned long flags;
7315
7316 local_irq_save(flags);
7317 double_rq_lock(rq_src, rq_dest);
7318 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7319 rq_src->nr_uninterruptible = 0;
7320 double_rq_unlock(rq_src, rq_dest);
7321 local_irq_restore(flags);
7322}
7323
7324/* Run through task list and migrate tasks from the dead cpu. */
7325static void migrate_live_tasks(int src_cpu)
7326{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007327 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007328
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007329 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007330
Ingo Molnar48f24c42006-07-03 00:25:40 -07007331 do_each_thread(t, p) {
7332 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007333 continue;
7334
Ingo Molnar48f24c42006-07-03 00:25:40 -07007335 if (task_cpu(p) == src_cpu)
7336 move_task_off_dead_cpu(src_cpu, p);
7337 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007338
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007339 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007340}
7341
Ingo Molnardd41f592007-07-09 18:51:59 +02007342/*
7343 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007344 * It does so by boosting its priority to highest possible.
7345 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007346 */
7347void sched_idle_next(void)
7348{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007349 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07007350 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007351 struct task_struct *p = rq->idle;
7352 unsigned long flags;
7353
7354 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007355 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007356
Ingo Molnar48f24c42006-07-03 00:25:40 -07007357 /*
7358 * Strictly not necessary since rest of the CPUs are stopped by now
7359 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007360 */
7361 spin_lock_irqsave(&rq->lock, flags);
7362
Ingo Molnardd41f592007-07-09 18:51:59 +02007363 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007364
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007365 update_rq_clock(rq);
7366 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007367
7368 spin_unlock_irqrestore(&rq->lock, flags);
7369}
7370
Ingo Molnar48f24c42006-07-03 00:25:40 -07007371/*
7372 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07007373 * offline.
7374 */
7375void idle_task_exit(void)
7376{
7377 struct mm_struct *mm = current->active_mm;
7378
7379 BUG_ON(cpu_online(smp_processor_id()));
7380
7381 if (mm != &init_mm)
7382 switch_mm(mm, &init_mm, current);
7383 mmdrop(mm);
7384}
7385
Kirill Korotaev054b9102006-12-10 02:20:11 -08007386/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007387static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007388{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007389 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007390
7391 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07007392 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007393
7394 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07007395 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007396
Ingo Molnar48f24c42006-07-03 00:25:40 -07007397 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007398
7399 /*
7400 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007401 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07007402 * fine.
7403 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007404 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007405 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007406 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007407
Ingo Molnar48f24c42006-07-03 00:25:40 -07007408 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007409}
7410
7411/* release_task() removes task from tasklist, so we won't find dead tasks. */
7412static void migrate_dead_tasks(unsigned int dead_cpu)
7413{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007414 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007415 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007416
Ingo Molnardd41f592007-07-09 18:51:59 +02007417 for ( ; ; ) {
7418 if (!rq->nr_running)
7419 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02007420 update_rq_clock(rq);
Wang Chenb67802e2009-03-02 13:55:26 +08007421 next = pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02007422 if (!next)
7423 break;
Dmitry Adamushko79c53792008-06-29 00:16:56 +02007424 next->sched_class->put_prev_task(rq, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02007425 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02007426
Linus Torvalds1da177e2005-04-16 15:20:36 -07007427 }
7428}
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007429
7430/*
7431 * remove the tasks which were accounted by rq from calc_load_tasks.
7432 */
7433static void calc_global_load_remove(struct rq *rq)
7434{
7435 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
Thomas Gleixnera468d382009-07-17 14:15:46 +02007436 rq->calc_load_active = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007437}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007438#endif /* CONFIG_HOTPLUG_CPU */
7439
Nick Piggine692ab52007-07-26 13:40:43 +02007440#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7441
7442static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007443 {
7444 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007445 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007446 },
Eric W. Biederman56992302009-11-05 15:38:40 -08007447 {}
Nick Piggine692ab52007-07-26 13:40:43 +02007448};
7449
7450static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007451 {
7452 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007453 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007454 .child = sd_ctl_dir,
7455 },
Eric W. Biederman56992302009-11-05 15:38:40 -08007456 {}
Nick Piggine692ab52007-07-26 13:40:43 +02007457};
7458
7459static struct ctl_table *sd_alloc_ctl_entry(int n)
7460{
7461 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02007462 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02007463
Nick Piggine692ab52007-07-26 13:40:43 +02007464 return entry;
7465}
7466
Milton Miller6382bc92007-10-15 17:00:19 +02007467static void sd_free_ctl_entry(struct ctl_table **tablep)
7468{
Milton Millercd7900762007-10-17 16:55:11 +02007469 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02007470
Milton Millercd7900762007-10-17 16:55:11 +02007471 /*
7472 * In the intermediate directories, both the child directory and
7473 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007474 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02007475 * static strings and all have proc handlers.
7476 */
7477 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02007478 if (entry->child)
7479 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02007480 if (entry->proc_handler == NULL)
7481 kfree(entry->procname);
7482 }
Milton Miller6382bc92007-10-15 17:00:19 +02007483
7484 kfree(*tablep);
7485 *tablep = NULL;
7486}
7487
Nick Piggine692ab52007-07-26 13:40:43 +02007488static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02007489set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02007490 const char *procname, void *data, int maxlen,
7491 mode_t mode, proc_handler *proc_handler)
7492{
Nick Piggine692ab52007-07-26 13:40:43 +02007493 entry->procname = procname;
7494 entry->data = data;
7495 entry->maxlen = maxlen;
7496 entry->mode = mode;
7497 entry->proc_handler = proc_handler;
7498}
7499
7500static struct ctl_table *
7501sd_alloc_ctl_domain_table(struct sched_domain *sd)
7502{
Ingo Molnara5d8c342008-10-09 11:35:51 +02007503 struct ctl_table *table = sd_alloc_ctl_entry(13);
Nick Piggine692ab52007-07-26 13:40:43 +02007504
Milton Millerad1cdc12007-10-15 17:00:19 +02007505 if (table == NULL)
7506 return NULL;
7507
Alexey Dobriyane0361852007-08-09 11:16:46 +02007508 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007509 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007510 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007511 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007512 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007513 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007514 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007515 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007516 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007517 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007518 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007519 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007520 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007521 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007522 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02007523 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007524 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02007525 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007526 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02007527 &sd->cache_nice_tries,
7528 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007529 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02007530 sizeof(int), 0644, proc_dointvec_minmax);
Ingo Molnara5d8c342008-10-09 11:35:51 +02007531 set_table_entry(&table[11], "name", sd->name,
7532 CORENAME_MAX_SIZE, 0444, proc_dostring);
7533 /* &table[12] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02007534
7535 return table;
7536}
7537
Ingo Molnar9a4e7152007-11-28 15:52:56 +01007538static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02007539{
7540 struct ctl_table *entry, *table;
7541 struct sched_domain *sd;
7542 int domain_num = 0, i;
7543 char buf[32];
7544
7545 for_each_domain(cpu, sd)
7546 domain_num++;
7547 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02007548 if (table == NULL)
7549 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02007550
7551 i = 0;
7552 for_each_domain(cpu, sd) {
7553 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007554 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007555 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007556 entry->child = sd_alloc_ctl_domain_table(sd);
7557 entry++;
7558 i++;
7559 }
7560 return table;
7561}
7562
7563static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02007564static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007565{
7566 int i, cpu_num = num_online_cpus();
7567 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7568 char buf[32];
7569
Milton Miller73785472007-10-24 18:23:48 +02007570 WARN_ON(sd_ctl_dir[0].child);
7571 sd_ctl_dir[0].child = entry;
7572
Milton Millerad1cdc12007-10-15 17:00:19 +02007573 if (entry == NULL)
7574 return;
7575
Milton Miller97b6ea72007-10-15 17:00:19 +02007576 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02007577 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007578 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007579 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007580 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02007581 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02007582 }
Milton Miller73785472007-10-24 18:23:48 +02007583
7584 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02007585 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7586}
Milton Miller6382bc92007-10-15 17:00:19 +02007587
Milton Miller73785472007-10-24 18:23:48 +02007588/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02007589static void unregister_sched_domain_sysctl(void)
7590{
Milton Miller73785472007-10-24 18:23:48 +02007591 if (sd_sysctl_header)
7592 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02007593 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007594 if (sd_ctl_dir[0].child)
7595 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02007596}
Nick Piggine692ab52007-07-26 13:40:43 +02007597#else
Milton Miller6382bc92007-10-15 17:00:19 +02007598static void register_sched_domain_sysctl(void)
7599{
7600}
7601static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007602{
7603}
7604#endif
7605
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007606static void set_rq_online(struct rq *rq)
7607{
7608 if (!rq->online) {
7609 const struct sched_class *class;
7610
Rusty Russellc6c49272008-11-25 02:35:05 +10307611 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007612 rq->online = 1;
7613
7614 for_each_class(class) {
7615 if (class->rq_online)
7616 class->rq_online(rq);
7617 }
7618 }
7619}
7620
7621static void set_rq_offline(struct rq *rq)
7622{
7623 if (rq->online) {
7624 const struct sched_class *class;
7625
7626 for_each_class(class) {
7627 if (class->rq_offline)
7628 class->rq_offline(rq);
7629 }
7630
Rusty Russellc6c49272008-11-25 02:35:05 +10307631 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007632 rq->online = 0;
7633 }
7634}
7635
Linus Torvalds1da177e2005-04-16 15:20:36 -07007636/*
7637 * migration_call - callback that gets triggered when a CPU is added.
7638 * Here we can start up the necessary migration thread for the new CPU.
7639 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007640static int __cpuinit
7641migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007642{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007643 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007644 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007645 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007646 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007647
7648 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07007649
Linus Torvalds1da177e2005-04-16 15:20:36 -07007650 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007651 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02007652 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007653 if (IS_ERR(p))
7654 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007655 kthread_bind(p, cpu);
7656 /* Must be high prio: stop_machine expects to yield to it. */
7657 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02007658 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007659 task_rq_unlock(rq, &flags);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007660 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007661 cpu_rq(cpu)->migration_thread = p;
Thomas Gleixnera468d382009-07-17 14:15:46 +02007662 rq->calc_load_update = calc_load_update;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007663 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007664
Linus Torvalds1da177e2005-04-16 15:20:36 -07007665 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007666 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007667 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007668 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007669
7670 /* Update our root-domain */
7671 rq = cpu_rq(cpu);
7672 spin_lock_irqsave(&rq->lock, flags);
7673 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307674 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007675
7676 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007677 }
7678 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007679 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007680
Linus Torvalds1da177e2005-04-16 15:20:36 -07007681#ifdef CONFIG_HOTPLUG_CPU
7682 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007683 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07007684 if (!cpu_rq(cpu)->migration_thread)
7685 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007686 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08007687 kthread_bind(cpu_rq(cpu)->migration_thread,
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307688 cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007689 kthread_stop(cpu_rq(cpu)->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007690 put_task_struct(cpu_rq(cpu)->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007691 cpu_rq(cpu)->migration_thread = NULL;
7692 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007693
Linus Torvalds1da177e2005-04-16 15:20:36 -07007694 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007695 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07007696 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007697 migrate_live_tasks(cpu);
7698 rq = cpu_rq(cpu);
7699 kthread_stop(rq->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007700 put_task_struct(rq->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007701 rq->migration_thread = NULL;
7702 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007703 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02007704 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007705 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007706 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02007707 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7708 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007709 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007710 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07007711 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007712 migrate_nr_uninterruptible(rq);
7713 BUG_ON(rq->nr_running != 0);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007714 calc_global_load_remove(rq);
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007715 /*
7716 * No need to migrate the tasks: it was best-effort if
7717 * they didn't take sched_hotcpu_mutex. Just wake up
7718 * the requestors.
7719 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007720 spin_lock_irq(&rq->lock);
7721 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007722 struct migration_req *req;
7723
Linus Torvalds1da177e2005-04-16 15:20:36 -07007724 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07007725 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007726 list_del_init(&req->list);
Brian King9a2bd242008-12-09 08:47:00 -06007727 spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007728 complete(&req->done);
Brian King9a2bd242008-12-09 08:47:00 -06007729 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007730 }
7731 spin_unlock_irq(&rq->lock);
7732 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007733
Gregory Haskins08f503b2008-03-10 17:59:11 -04007734 case CPU_DYING:
7735 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01007736 /* Update our root-domain */
7737 rq = cpu_rq(cpu);
7738 spin_lock_irqsave(&rq->lock, flags);
7739 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307740 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007741 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007742 }
7743 spin_unlock_irqrestore(&rq->lock, flags);
7744 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007745#endif
7746 }
7747 return NOTIFY_OK;
7748}
7749
Paul Mackerrasf38b0822009-06-02 21:05:16 +10007750/*
7751 * Register at high priority so that task migration (migrate_all_tasks)
7752 * happens before everything else. This has to be lower priority than
Ingo Molnarcdd6c482009-09-21 12:02:48 +02007753 * the notifier in the perf_event subsystem, though.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007754 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07007755static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007756 .notifier_call = migration_call,
7757 .priority = 10
7758};
7759
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007760static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007761{
7762 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07007763 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007764
7765 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07007766 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7767 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007768 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7769 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007770
Thomas Gleixnera004cd42009-07-21 09:54:05 +02007771 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007772}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007773early_initcall(migration_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007774#endif
7775
7776#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07007777
Ingo Molnar3e9830d2007-10-15 17:00:13 +02007778#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007779
Mike Travisf6630112009-11-17 18:22:15 -06007780static __read_mostly int sched_domain_debug_enabled;
7781
7782static int __init sched_domain_debug_setup(char *str)
7783{
7784 sched_domain_debug_enabled = 1;
7785
7786 return 0;
7787}
7788early_param("sched_debug", sched_domain_debug_setup);
7789
Mike Travis7c16ec52008-04-04 18:11:11 -07007790static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
Rusty Russell96f874e2008-11-25 02:35:14 +10307791 struct cpumask *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007792{
7793 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07007794 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007795
Rusty Russell968ea6d2008-12-13 21:55:51 +10307796 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
Rusty Russell96f874e2008-11-25 02:35:14 +10307797 cpumask_clear(groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007798
7799 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7800
7801 if (!(sd->flags & SD_LOAD_BALANCE)) {
7802 printk("does not load-balance\n");
7803 if (sd->parent)
7804 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7805 " has parent");
7806 return -1;
7807 }
7808
Li Zefaneefd7962008-11-04 16:15:37 +08007809 printk(KERN_CONT "span %s level %s\n", str, sd->name);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007810
Rusty Russell758b2cd2008-11-25 02:35:04 +10307811 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007812 printk(KERN_ERR "ERROR: domain->span does not contain "
7813 "CPU%d\n", cpu);
7814 }
Rusty Russell758b2cd2008-11-25 02:35:04 +10307815 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007816 printk(KERN_ERR "ERROR: domain->groups does not contain"
7817 " CPU%d\n", cpu);
7818 }
7819
7820 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7821 do {
7822 if (!group) {
7823 printk("\n");
7824 printk(KERN_ERR "ERROR: group is NULL\n");
7825 break;
7826 }
7827
Peter Zijlstra18a38852009-09-01 10:34:39 +02007828 if (!group->cpu_power) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007829 printk(KERN_CONT "\n");
7830 printk(KERN_ERR "ERROR: domain->cpu_power not "
7831 "set\n");
7832 break;
7833 }
7834
Rusty Russell758b2cd2008-11-25 02:35:04 +10307835 if (!cpumask_weight(sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007836 printk(KERN_CONT "\n");
7837 printk(KERN_ERR "ERROR: empty group\n");
7838 break;
7839 }
7840
Rusty Russell758b2cd2008-11-25 02:35:04 +10307841 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007842 printk(KERN_CONT "\n");
7843 printk(KERN_ERR "ERROR: repeated CPUs\n");
7844 break;
7845 }
7846
Rusty Russell758b2cd2008-11-25 02:35:04 +10307847 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007848
Rusty Russell968ea6d2008-12-13 21:55:51 +10307849 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307850
7851 printk(KERN_CONT " %s", str);
Peter Zijlstra18a38852009-09-01 10:34:39 +02007852 if (group->cpu_power != SCHED_LOAD_SCALE) {
7853 printk(KERN_CONT " (cpu_power = %d)",
7854 group->cpu_power);
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307855 }
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007856
7857 group = group->next;
7858 } while (group != sd->groups);
7859 printk(KERN_CONT "\n");
7860
Rusty Russell758b2cd2008-11-25 02:35:04 +10307861 if (!cpumask_equal(sched_domain_span(sd), groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007862 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7863
Rusty Russell758b2cd2008-11-25 02:35:04 +10307864 if (sd->parent &&
7865 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007866 printk(KERN_ERR "ERROR: parent span is not a superset "
7867 "of domain->span\n");
7868 return 0;
7869}
7870
Linus Torvalds1da177e2005-04-16 15:20:36 -07007871static void sched_domain_debug(struct sched_domain *sd, int cpu)
7872{
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307873 cpumask_var_t groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007874 int level = 0;
7875
Mike Travisf6630112009-11-17 18:22:15 -06007876 if (!sched_domain_debug_enabled)
7877 return;
7878
Nick Piggin41c7ce92005-06-25 14:57:24 -07007879 if (!sd) {
7880 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7881 return;
7882 }
7883
Linus Torvalds1da177e2005-04-16 15:20:36 -07007884 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7885
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307886 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007887 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7888 return;
7889 }
7890
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007891 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007892 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007893 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007894 level++;
7895 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08007896 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007897 break;
7898 }
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307899 free_cpumask_var(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007900}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007901#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007902# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007903#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007904
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007905static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007906{
Rusty Russell758b2cd2008-11-25 02:35:04 +10307907 if (cpumask_weight(sched_domain_span(sd)) == 1)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007908 return 1;
7909
7910 /* Following flags need at least 2 groups */
7911 if (sd->flags & (SD_LOAD_BALANCE |
7912 SD_BALANCE_NEWIDLE |
7913 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007914 SD_BALANCE_EXEC |
7915 SD_SHARE_CPUPOWER |
7916 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007917 if (sd->groups != sd->groups->next)
7918 return 0;
7919 }
7920
7921 /* Following flags don't use groups */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02007922 if (sd->flags & (SD_WAKE_AFFINE))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007923 return 0;
7924
7925 return 1;
7926}
7927
Ingo Molnar48f24c42006-07-03 00:25:40 -07007928static int
7929sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007930{
7931 unsigned long cflags = sd->flags, pflags = parent->flags;
7932
7933 if (sd_degenerate(parent))
7934 return 1;
7935
Rusty Russell758b2cd2008-11-25 02:35:04 +10307936 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007937 return 0;
7938
Suresh Siddha245af2c2005-06-25 14:57:25 -07007939 /* Flags needing groups don't count if only 1 group in parent */
7940 if (parent->groups == parent->groups->next) {
7941 pflags &= ~(SD_LOAD_BALANCE |
7942 SD_BALANCE_NEWIDLE |
7943 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007944 SD_BALANCE_EXEC |
7945 SD_SHARE_CPUPOWER |
7946 SD_SHARE_PKG_RESOURCES);
Ken Chen54364992008-12-07 18:47:37 -08007947 if (nr_node_ids == 1)
7948 pflags &= ~SD_SERIALIZE;
Suresh Siddha245af2c2005-06-25 14:57:25 -07007949 }
7950 if (~cflags & pflags)
7951 return 0;
7952
7953 return 1;
7954}
7955
Rusty Russellc6c49272008-11-25 02:35:05 +10307956static void free_rootdomain(struct root_domain *rd)
7957{
Peter Zijlstra047106a2009-11-16 10:28:09 +01007958 synchronize_sched();
7959
Rusty Russell68e74562008-11-25 02:35:13 +10307960 cpupri_cleanup(&rd->cpupri);
7961
Rusty Russellc6c49272008-11-25 02:35:05 +10307962 free_cpumask_var(rd->rto_mask);
7963 free_cpumask_var(rd->online);
7964 free_cpumask_var(rd->span);
7965 kfree(rd);
7966}
7967
Gregory Haskins57d885f2008-01-25 21:08:18 +01007968static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7969{
Ingo Molnara0490fa2009-02-12 11:35:40 +01007970 struct root_domain *old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007971 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007972
7973 spin_lock_irqsave(&rq->lock, flags);
7974
7975 if (rq->rd) {
Ingo Molnara0490fa2009-02-12 11:35:40 +01007976 old_rd = rq->rd;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007977
Rusty Russellc6c49272008-11-25 02:35:05 +10307978 if (cpumask_test_cpu(rq->cpu, old_rd->online))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007979 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007980
Rusty Russellc6c49272008-11-25 02:35:05 +10307981 cpumask_clear_cpu(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007982
Ingo Molnara0490fa2009-02-12 11:35:40 +01007983 /*
7984 * If we dont want to free the old_rt yet then
7985 * set old_rd to NULL to skip the freeing later
7986 * in this function:
7987 */
7988 if (!atomic_dec_and_test(&old_rd->refcount))
7989 old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007990 }
7991
7992 atomic_inc(&rd->refcount);
7993 rq->rd = rd;
7994
Rusty Russellc6c49272008-11-25 02:35:05 +10307995 cpumask_set_cpu(rq->cpu, rd->span);
Gregory Haskins00aec932009-07-30 10:57:23 -04007996 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007997 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007998
7999 spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnara0490fa2009-02-12 11:35:40 +01008000
8001 if (old_rd)
8002 free_rootdomain(old_rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01008003}
8004
Li Zefanfd5e1b52009-06-15 13:34:19 +08008005static int init_rootdomain(struct root_domain *rd, bool bootmem)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008006{
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008007 gfp_t gfp = GFP_KERNEL;
8008
Gregory Haskins57d885f2008-01-25 21:08:18 +01008009 memset(rd, 0, sizeof(*rd));
8010
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008011 if (bootmem)
8012 gfp = GFP_NOWAIT;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02008013
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008014 if (!alloc_cpumask_var(&rd->span, gfp))
Li Zefan0c910d22009-01-06 17:39:06 +08008015 goto out;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008016 if (!alloc_cpumask_var(&rd->online, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10308017 goto free_span;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03008018 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10308019 goto free_online;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02008020
Pekka Enberg0fb53022009-06-11 08:41:22 +03008021 if (cpupri_init(&rd->cpupri, bootmem) != 0)
Rusty Russell68e74562008-11-25 02:35:13 +10308022 goto free_rto_mask;
Rusty Russellc6c49272008-11-25 02:35:05 +10308023 return 0;
8024
Rusty Russell68e74562008-11-25 02:35:13 +10308025free_rto_mask:
8026 free_cpumask_var(rd->rto_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10308027free_online:
8028 free_cpumask_var(rd->online);
8029free_span:
8030 free_cpumask_var(rd->span);
Li Zefan0c910d22009-01-06 17:39:06 +08008031out:
Rusty Russellc6c49272008-11-25 02:35:05 +10308032 return -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008033}
8034
8035static void init_defrootdomain(void)
8036{
Rusty Russellc6c49272008-11-25 02:35:05 +10308037 init_rootdomain(&def_root_domain, true);
8038
Gregory Haskins57d885f2008-01-25 21:08:18 +01008039 atomic_set(&def_root_domain.refcount, 1);
8040}
8041
Gregory Haskinsdc938522008-01-25 21:08:26 +01008042static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008043{
8044 struct root_domain *rd;
8045
8046 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8047 if (!rd)
8048 return NULL;
8049
Rusty Russellc6c49272008-11-25 02:35:05 +10308050 if (init_rootdomain(rd, false) != 0) {
8051 kfree(rd);
8052 return NULL;
8053 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01008054
8055 return rd;
8056}
8057
Linus Torvalds1da177e2005-04-16 15:20:36 -07008058/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01008059 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07008060 * hold the hotplug lock.
8061 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01008062static void
8063cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008064{
Ingo Molnar70b97a72006-07-03 00:25:42 -07008065 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07008066 struct sched_domain *tmp;
8067
8068 /* Remove the sched domains which do not contribute to scheduling. */
Li Zefanf29c9b12008-11-06 09:45:16 +08008069 for (tmp = sd; tmp; ) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008070 struct sched_domain *parent = tmp->parent;
8071 if (!parent)
8072 break;
Li Zefanf29c9b12008-11-06 09:45:16 +08008073
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008074 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008075 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008076 if (parent->parent)
8077 parent->parent->child = tmp;
Li Zefanf29c9b12008-11-06 09:45:16 +08008078 } else
8079 tmp = tmp->parent;
Suresh Siddha245af2c2005-06-25 14:57:25 -07008080 }
8081
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008082 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07008083 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008084 if (sd)
8085 sd->child = NULL;
8086 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008087
8088 sched_domain_debug(sd, cpu);
8089
Gregory Haskins57d885f2008-01-25 21:08:18 +01008090 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07008091 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008092}
8093
8094/* cpus with isolated domains */
Rusty Russelldcc30a32008-11-25 02:35:12 +10308095static cpumask_var_t cpu_isolated_map;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008096
8097/* Setup the mask of cpus configured for isolated domains */
8098static int __init isolated_cpu_setup(char *str)
8099{
Rusty Russellbdddd292009-12-02 14:09:16 +10308100 alloc_bootmem_cpumask_var(&cpu_isolated_map);
Rusty Russell968ea6d2008-12-13 21:55:51 +10308101 cpulist_parse(str, cpu_isolated_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008102 return 1;
8103}
8104
Ingo Molnar8927f492007-10-15 17:00:13 +02008105__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008106
8107/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008108 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8109 * to a function which identifies what group(along with sched group) a CPU
Rusty Russell96f874e2008-11-25 02:35:14 +10308110 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8111 * (due to the fact that we keep track of groups covered with a struct cpumask).
Linus Torvalds1da177e2005-04-16 15:20:36 -07008112 *
8113 * init_sched_build_groups will build a circular linked list of the groups
8114 * covered by the given span, and will set each group's ->cpumask correctly,
8115 * and ->cpu_power to 0.
8116 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008117static void
Rusty Russell96f874e2008-11-25 02:35:14 +10308118init_sched_build_groups(const struct cpumask *span,
8119 const struct cpumask *cpu_map,
8120 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07008121 struct sched_group **sg,
Rusty Russell96f874e2008-11-25 02:35:14 +10308122 struct cpumask *tmpmask),
8123 struct cpumask *covered, struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008124{
8125 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008126 int i;
8127
Rusty Russell96f874e2008-11-25 02:35:14 +10308128 cpumask_clear(covered);
Mike Travis7c16ec52008-04-04 18:11:11 -07008129
Rusty Russellabcd0832008-11-25 02:35:02 +10308130 for_each_cpu(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008131 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07008132 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008133 int j;
8134
Rusty Russell758b2cd2008-11-25 02:35:04 +10308135 if (cpumask_test_cpu(i, covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008136 continue;
8137
Rusty Russell758b2cd2008-11-25 02:35:04 +10308138 cpumask_clear(sched_group_cpus(sg));
Peter Zijlstra18a38852009-09-01 10:34:39 +02008139 sg->cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008140
Rusty Russellabcd0832008-11-25 02:35:02 +10308141 for_each_cpu(j, span) {
Mike Travis7c16ec52008-04-04 18:11:11 -07008142 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008143 continue;
8144
Rusty Russell96f874e2008-11-25 02:35:14 +10308145 cpumask_set_cpu(j, covered);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308146 cpumask_set_cpu(j, sched_group_cpus(sg));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008147 }
8148 if (!first)
8149 first = sg;
8150 if (last)
8151 last->next = sg;
8152 last = sg;
8153 }
8154 last->next = first;
8155}
8156
John Hawkes9c1cfda2005-09-06 15:18:14 -07008157#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07008158
John Hawkes9c1cfda2005-09-06 15:18:14 -07008159#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08008160
John Hawkes9c1cfda2005-09-06 15:18:14 -07008161/**
8162 * find_next_best_node - find the next node to include in a sched_domain
8163 * @node: node whose sched_domain we're building
8164 * @used_nodes: nodes already in the sched_domain
8165 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008166 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07008167 * finds the closest node not already in the @used_nodes map.
8168 *
8169 * Should use nodemask_t.
8170 */
Mike Travisc5f59f02008-04-04 18:11:10 -07008171static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008172{
8173 int i, n, val, min_val, best_node = 0;
8174
8175 min_val = INT_MAX;
8176
Mike Travis076ac2a2008-05-12 21:21:12 +02008177 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008178 /* Start at @node */
Mike Travis076ac2a2008-05-12 21:21:12 +02008179 n = (node + i) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008180
8181 if (!nr_cpus_node(n))
8182 continue;
8183
8184 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07008185 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008186 continue;
8187
8188 /* Simple min distance search */
8189 val = node_distance(node, n);
8190
8191 if (val < min_val) {
8192 min_val = val;
8193 best_node = n;
8194 }
8195 }
8196
Mike Travisc5f59f02008-04-04 18:11:10 -07008197 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008198 return best_node;
8199}
8200
8201/**
8202 * sched_domain_node_span - get a cpumask for a node's sched_domain
8203 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07008204 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07008205 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008206 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07008207 * should be one that prevents unnecessary balancing, but also spreads tasks
8208 * out optimally.
8209 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308210static void sched_domain_node_span(int node, struct cpumask *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008211{
Mike Travisc5f59f02008-04-04 18:11:10 -07008212 nodemask_t used_nodes;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008213 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008214
Mike Travis6ca09df2008-12-31 18:08:45 -08008215 cpumask_clear(span);
Mike Travisc5f59f02008-04-04 18:11:10 -07008216 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008217
Mike Travis6ca09df2008-12-31 18:08:45 -08008218 cpumask_or(span, span, cpumask_of_node(node));
Mike Travisc5f59f02008-04-04 18:11:10 -07008219 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008220
8221 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07008222 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008223
Mike Travis6ca09df2008-12-31 18:08:45 -08008224 cpumask_or(span, span, cpumask_of_node(next_node));
John Hawkes9c1cfda2005-09-06 15:18:14 -07008225 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008226}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008227#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07008228
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008229int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008230
John Hawkes9c1cfda2005-09-06 15:18:14 -07008231/*
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308232 * The cpus mask in sched_group and sched_domain hangs off the end.
Ingo Molnar4200efd2009-05-19 09:22:19 +02008233 *
8234 * ( See the the comments in include/linux/sched.h:struct sched_group
8235 * and struct sched_domain. )
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308236 */
8237struct static_sched_group {
8238 struct sched_group sg;
8239 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8240};
8241
8242struct static_sched_domain {
8243 struct sched_domain sd;
8244 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8245};
8246
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008247struct s_data {
8248#ifdef CONFIG_NUMA
8249 int sd_allnodes;
8250 cpumask_var_t domainspan;
8251 cpumask_var_t covered;
8252 cpumask_var_t notcovered;
8253#endif
8254 cpumask_var_t nodemask;
8255 cpumask_var_t this_sibling_map;
8256 cpumask_var_t this_core_map;
8257 cpumask_var_t send_covered;
8258 cpumask_var_t tmpmask;
8259 struct sched_group **sched_group_nodes;
8260 struct root_domain *rd;
8261};
8262
Andreas Herrmann2109b992009-08-18 12:53:00 +02008263enum s_alloc {
8264 sa_sched_groups = 0,
8265 sa_rootdomain,
8266 sa_tmpmask,
8267 sa_send_covered,
8268 sa_this_core_map,
8269 sa_this_sibling_map,
8270 sa_nodemask,
8271 sa_sched_group_nodes,
8272#ifdef CONFIG_NUMA
8273 sa_notcovered,
8274 sa_covered,
8275 sa_domainspan,
8276#endif
8277 sa_none,
8278};
8279
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308280/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07008281 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07008282 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008283#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308284static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8285static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008286
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008287static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308288cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8289 struct sched_group **sg, struct cpumask *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008290{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008291 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308292 *sg = &per_cpu(sched_group_cpus, cpu).sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008293 return cpu;
8294}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008295#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008296
Ingo Molnar48f24c42006-07-03 00:25:40 -07008297/*
8298 * multi-core sched-domains:
8299 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008300#ifdef CONFIG_SCHED_MC
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308301static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8302static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008303#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008304
8305#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008306static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308307cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8308 struct sched_group **sg, struct cpumask *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008309{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008310 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07008311
Rusty Russellc69fc562009-03-13 14:49:46 +10308312 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308313 group = cpumask_first(mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008314 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308315 *sg = &per_cpu(sched_group_core, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008316 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008317}
8318#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008319static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308320cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8321 struct sched_group **sg, struct cpumask *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008322{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008323 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308324 *sg = &per_cpu(sched_group_core, cpu).sg;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008325 return cpu;
8326}
8327#endif
8328
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308329static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8330static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008331
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008332static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308333cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8334 struct sched_group **sg, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008335{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008336 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008337#ifdef CONFIG_SCHED_MC
Mike Travis6ca09df2008-12-31 18:08:45 -08008338 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308339 group = cpumask_first(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008340#elif defined(CONFIG_SCHED_SMT)
Rusty Russellc69fc562009-03-13 14:49:46 +10308341 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308342 group = cpumask_first(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008343#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008344 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008345#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008346 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308347 *sg = &per_cpu(sched_group_phys, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008348 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008349}
8350
8351#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07008352/*
8353 * The init_sched_build_groups can't handle what we want to do with node
8354 * groups, so roll our own. Now each node has its own list of groups which
8355 * gets dynamically allocated.
8356 */
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008357static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07008358static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008359
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008360static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308361static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008362
Rusty Russell96f874e2008-11-25 02:35:14 +10308363static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8364 struct sched_group **sg,
8365 struct cpumask *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008366{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008367 int group;
8368
Mike Travis6ca09df2008-12-31 18:08:45 -08008369 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308370 group = cpumask_first(nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008371
8372 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308373 *sg = &per_cpu(sched_group_allnodes, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008374 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008375}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008376
Siddha, Suresh B08069032006-03-27 01:15:23 -08008377static void init_numa_sched_groups_power(struct sched_group *group_head)
8378{
8379 struct sched_group *sg = group_head;
8380 int j;
8381
8382 if (!sg)
8383 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008384 do {
Rusty Russell758b2cd2008-11-25 02:35:04 +10308385 for_each_cpu(j, sched_group_cpus(sg)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008386 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008387
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308388 sd = &per_cpu(phys_domains, j).sd;
Miao Xie13318a72009-04-15 09:59:10 +08008389 if (j != group_first_cpu(sd->groups)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008390 /*
8391 * Only add "power" once for each
8392 * physical package.
8393 */
8394 continue;
8395 }
8396
Peter Zijlstra18a38852009-09-01 10:34:39 +02008397 sg->cpu_power += sd->groups->cpu_power;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008398 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02008399 sg = sg->next;
8400 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08008401}
Andreas Herrmann0601a882009-08-18 13:01:11 +02008402
8403static int build_numa_sched_groups(struct s_data *d,
8404 const struct cpumask *cpu_map, int num)
8405{
8406 struct sched_domain *sd;
8407 struct sched_group *sg, *prev;
8408 int n, j;
8409
8410 cpumask_clear(d->covered);
8411 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8412 if (cpumask_empty(d->nodemask)) {
8413 d->sched_group_nodes[num] = NULL;
8414 goto out;
8415 }
8416
8417 sched_domain_node_span(num, d->domainspan);
8418 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8419
8420 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8421 GFP_KERNEL, num);
8422 if (!sg) {
8423 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8424 num);
8425 return -ENOMEM;
8426 }
8427 d->sched_group_nodes[num] = sg;
8428
8429 for_each_cpu(j, d->nodemask) {
8430 sd = &per_cpu(node_domains, j).sd;
8431 sd->groups = sg;
8432 }
8433
Peter Zijlstra18a38852009-09-01 10:34:39 +02008434 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008435 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8436 sg->next = sg;
8437 cpumask_or(d->covered, d->covered, d->nodemask);
8438
8439 prev = sg;
8440 for (j = 0; j < nr_node_ids; j++) {
8441 n = (num + j) % nr_node_ids;
8442 cpumask_complement(d->notcovered, d->covered);
8443 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8444 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8445 if (cpumask_empty(d->tmpmask))
8446 break;
8447 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8448 if (cpumask_empty(d->tmpmask))
8449 continue;
8450 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8451 GFP_KERNEL, num);
8452 if (!sg) {
8453 printk(KERN_WARNING
8454 "Can not alloc domain group for node %d\n", j);
8455 return -ENOMEM;
8456 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008457 sg->cpu_power = 0;
Andreas Herrmann0601a882009-08-18 13:01:11 +02008458 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8459 sg->next = prev->next;
8460 cpumask_or(d->covered, d->covered, d->tmpmask);
8461 prev->next = sg;
8462 prev = sg;
8463 }
8464out:
8465 return 0;
8466}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008467#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008468
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008469#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008470/* Free memory allocated for various sched_group structures */
Rusty Russell96f874e2008-11-25 02:35:14 +10308471static void free_sched_groups(const struct cpumask *cpu_map,
8472 struct cpumask *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008473{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008474 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008475
Rusty Russellabcd0832008-11-25 02:35:02 +10308476 for_each_cpu(cpu, cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008477 struct sched_group **sched_group_nodes
8478 = sched_group_nodes_bycpu[cpu];
8479
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008480 if (!sched_group_nodes)
8481 continue;
8482
Mike Travis076ac2a2008-05-12 21:21:12 +02008483 for (i = 0; i < nr_node_ids; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008484 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8485
Mike Travis6ca09df2008-12-31 18:08:45 -08008486 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308487 if (cpumask_empty(nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008488 continue;
8489
8490 if (sg == NULL)
8491 continue;
8492 sg = sg->next;
8493next_sg:
8494 oldsg = sg;
8495 sg = sg->next;
8496 kfree(oldsg);
8497 if (oldsg != sched_group_nodes[i])
8498 goto next_sg;
8499 }
8500 kfree(sched_group_nodes);
8501 sched_group_nodes_bycpu[cpu] = NULL;
8502 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008503}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008504#else /* !CONFIG_NUMA */
Rusty Russell96f874e2008-11-25 02:35:14 +10308505static void free_sched_groups(const struct cpumask *cpu_map,
8506 struct cpumask *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008507{
8508}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008509#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008510
Linus Torvalds1da177e2005-04-16 15:20:36 -07008511/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008512 * Initialize sched groups cpu_power.
8513 *
8514 * cpu_power indicates the capacity of sched group, which is used while
8515 * distributing the load between different sched groups in a sched domain.
8516 * Typically cpu_power for all the groups in a sched domain will be same unless
8517 * there are asymmetries in the topology. If there are asymmetries, group
8518 * having more cpu_power will pickup more load compared to the group having
8519 * less cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008520 */
8521static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8522{
8523 struct sched_domain *child;
8524 struct sched_group *group;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008525 long power;
8526 int weight;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008527
8528 WARN_ON(!sd || !sd->groups);
8529
Miao Xie13318a72009-04-15 09:59:10 +08008530 if (cpu != group_first_cpu(sd->groups))
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008531 return;
8532
8533 child = sd->child;
8534
Peter Zijlstra18a38852009-09-01 10:34:39 +02008535 sd->groups->cpu_power = 0;
Eric Dumazet5517d862007-05-08 00:32:57 -07008536
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008537 if (!child) {
8538 power = SCHED_LOAD_SCALE;
8539 weight = cpumask_weight(sched_domain_span(sd));
8540 /*
8541 * SMT siblings share the power of a single core.
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008542 * Usually multiple threads get a better yield out of
8543 * that one core than a single thread would have,
8544 * reflect that in sd->smt_gain.
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008545 */
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008546 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8547 power *= sd->smt_gain;
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008548 power /= weight;
Peter Zijlstraa52bfd732009-09-01 10:34:35 +02008549 power >>= SCHED_LOAD_SHIFT;
8550 }
Peter Zijlstra18a38852009-09-01 10:34:39 +02008551 sd->groups->cpu_power += power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008552 return;
8553 }
8554
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008555 /*
Peter Zijlstraf93e65c2009-09-01 10:34:32 +02008556 * Add cpu_power of each child group to this groups cpu_power.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008557 */
8558 group = child->groups;
8559 do {
Peter Zijlstra18a38852009-09-01 10:34:39 +02008560 sd->groups->cpu_power += group->cpu_power;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008561 group = group->next;
8562 } while (group != child->groups);
8563}
8564
8565/*
Mike Travis7c16ec52008-04-04 18:11:11 -07008566 * Initializers for schedule domains
8567 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8568 */
8569
Ingo Molnara5d8c342008-10-09 11:35:51 +02008570#ifdef CONFIG_SCHED_DEBUG
8571# define SD_INIT_NAME(sd, type) sd->name = #type
8572#else
8573# define SD_INIT_NAME(sd, type) do { } while (0)
8574#endif
8575
Mike Travis7c16ec52008-04-04 18:11:11 -07008576#define SD_INIT(sd, type) sd_init_##type(sd)
Ingo Molnara5d8c342008-10-09 11:35:51 +02008577
Mike Travis7c16ec52008-04-04 18:11:11 -07008578#define SD_INIT_FUNC(type) \
8579static noinline void sd_init_##type(struct sched_domain *sd) \
8580{ \
8581 memset(sd, 0, sizeof(*sd)); \
8582 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008583 sd->level = SD_LV_##type; \
Ingo Molnara5d8c342008-10-09 11:35:51 +02008584 SD_INIT_NAME(sd, type); \
Mike Travis7c16ec52008-04-04 18:11:11 -07008585}
8586
8587SD_INIT_FUNC(CPU)
8588#ifdef CONFIG_NUMA
8589 SD_INIT_FUNC(ALLNODES)
8590 SD_INIT_FUNC(NODE)
8591#endif
8592#ifdef CONFIG_SCHED_SMT
8593 SD_INIT_FUNC(SIBLING)
8594#endif
8595#ifdef CONFIG_SCHED_MC
8596 SD_INIT_FUNC(MC)
8597#endif
8598
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008599static int default_relax_domain_level = -1;
8600
8601static int __init setup_relax_domain_level(char *str)
8602{
Li Zefan30e0e172008-05-13 10:27:17 +08008603 unsigned long val;
8604
8605 val = simple_strtoul(str, NULL, 0);
8606 if (val < SD_LV_MAX)
8607 default_relax_domain_level = val;
8608
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008609 return 1;
8610}
8611__setup("relax_domain_level=", setup_relax_domain_level);
8612
8613static void set_domain_attribute(struct sched_domain *sd,
8614 struct sched_domain_attr *attr)
8615{
8616 int request;
8617
8618 if (!attr || attr->relax_domain_level < 0) {
8619 if (default_relax_domain_level < 0)
8620 return;
8621 else
8622 request = default_relax_domain_level;
8623 } else
8624 request = attr->relax_domain_level;
8625 if (request < sd->level) {
8626 /* turn off idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008627 sd->flags &= ~(SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008628 } else {
8629 /* turn on idle balance on this domain */
Peter Zijlstrac88d5912009-09-10 13:50:02 +02008630 sd->flags |= (SD_BALANCE_WAKE|SD_BALANCE_NEWIDLE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008631 }
8632}
8633
Andreas Herrmann2109b992009-08-18 12:53:00 +02008634static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8635 const struct cpumask *cpu_map)
8636{
8637 switch (what) {
8638 case sa_sched_groups:
8639 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8640 d->sched_group_nodes = NULL;
8641 case sa_rootdomain:
8642 free_rootdomain(d->rd); /* fall through */
8643 case sa_tmpmask:
8644 free_cpumask_var(d->tmpmask); /* fall through */
8645 case sa_send_covered:
8646 free_cpumask_var(d->send_covered); /* fall through */
8647 case sa_this_core_map:
8648 free_cpumask_var(d->this_core_map); /* fall through */
8649 case sa_this_sibling_map:
8650 free_cpumask_var(d->this_sibling_map); /* fall through */
8651 case sa_nodemask:
8652 free_cpumask_var(d->nodemask); /* fall through */
8653 case sa_sched_group_nodes:
8654#ifdef CONFIG_NUMA
8655 kfree(d->sched_group_nodes); /* fall through */
8656 case sa_notcovered:
8657 free_cpumask_var(d->notcovered); /* fall through */
8658 case sa_covered:
8659 free_cpumask_var(d->covered); /* fall through */
8660 case sa_domainspan:
8661 free_cpumask_var(d->domainspan); /* fall through */
8662#endif
8663 case sa_none:
8664 break;
8665 }
8666}
8667
8668static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8669 const struct cpumask *cpu_map)
8670{
8671#ifdef CONFIG_NUMA
8672 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8673 return sa_none;
8674 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8675 return sa_domainspan;
8676 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8677 return sa_covered;
8678 /* Allocate the per-node list of sched groups */
8679 d->sched_group_nodes = kcalloc(nr_node_ids,
8680 sizeof(struct sched_group *), GFP_KERNEL);
8681 if (!d->sched_group_nodes) {
8682 printk(KERN_WARNING "Can not alloc sched group node list\n");
8683 return sa_notcovered;
8684 }
8685 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8686#endif
8687 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8688 return sa_sched_group_nodes;
8689 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8690 return sa_nodemask;
8691 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8692 return sa_this_sibling_map;
8693 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8694 return sa_this_core_map;
8695 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8696 return sa_send_covered;
8697 d->rd = alloc_rootdomain();
8698 if (!d->rd) {
8699 printk(KERN_WARNING "Cannot alloc root domain\n");
8700 return sa_tmpmask;
8701 }
8702 return sa_rootdomain;
8703}
8704
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008705static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8706 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8707{
8708 struct sched_domain *sd = NULL;
8709#ifdef CONFIG_NUMA
8710 struct sched_domain *parent;
8711
8712 d->sd_allnodes = 0;
8713 if (cpumask_weight(cpu_map) >
8714 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8715 sd = &per_cpu(allnodes_domains, i).sd;
8716 SD_INIT(sd, ALLNODES);
8717 set_domain_attribute(sd, attr);
8718 cpumask_copy(sched_domain_span(sd), cpu_map);
8719 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8720 d->sd_allnodes = 1;
8721 }
8722 parent = sd;
8723
8724 sd = &per_cpu(node_domains, i).sd;
8725 SD_INIT(sd, NODE);
8726 set_domain_attribute(sd, attr);
8727 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8728 sd->parent = parent;
8729 if (parent)
8730 parent->child = sd;
8731 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8732#endif
8733 return sd;
8734}
8735
Andreas Herrmann87cce662009-08-18 12:54:55 +02008736static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8737 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8738 struct sched_domain *parent, int i)
8739{
8740 struct sched_domain *sd;
8741 sd = &per_cpu(phys_domains, i).sd;
8742 SD_INIT(sd, CPU);
8743 set_domain_attribute(sd, attr);
8744 cpumask_copy(sched_domain_span(sd), d->nodemask);
8745 sd->parent = parent;
8746 if (parent)
8747 parent->child = sd;
8748 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8749 return sd;
8750}
8751
Andreas Herrmann410c4082009-08-18 12:56:14 +02008752static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8753 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8754 struct sched_domain *parent, int i)
8755{
8756 struct sched_domain *sd = parent;
8757#ifdef CONFIG_SCHED_MC
8758 sd = &per_cpu(core_domains, i).sd;
8759 SD_INIT(sd, MC);
8760 set_domain_attribute(sd, attr);
8761 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8762 sd->parent = parent;
8763 parent->child = sd;
8764 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8765#endif
8766 return sd;
8767}
8768
Andreas Herrmannd8173532009-08-18 12:57:03 +02008769static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8770 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8771 struct sched_domain *parent, int i)
8772{
8773 struct sched_domain *sd = parent;
8774#ifdef CONFIG_SCHED_SMT
8775 sd = &per_cpu(cpu_domains, i).sd;
8776 SD_INIT(sd, SIBLING);
8777 set_domain_attribute(sd, attr);
8778 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8779 sd->parent = parent;
8780 parent->child = sd;
8781 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8782#endif
8783 return sd;
8784}
8785
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008786static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8787 const struct cpumask *cpu_map, int cpu)
8788{
8789 switch (l) {
8790#ifdef CONFIG_SCHED_SMT
8791 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8792 cpumask_and(d->this_sibling_map, cpu_map,
8793 topology_thread_cpumask(cpu));
8794 if (cpu == cpumask_first(d->this_sibling_map))
8795 init_sched_build_groups(d->this_sibling_map, cpu_map,
8796 &cpu_to_cpu_group,
8797 d->send_covered, d->tmpmask);
8798 break;
8799#endif
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008800#ifdef CONFIG_SCHED_MC
8801 case SD_LV_MC: /* set up multi-core groups */
8802 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8803 if (cpu == cpumask_first(d->this_core_map))
8804 init_sched_build_groups(d->this_core_map, cpu_map,
8805 &cpu_to_core_group,
8806 d->send_covered, d->tmpmask);
8807 break;
8808#endif
Andreas Herrmann86548092009-08-18 12:59:28 +02008809 case SD_LV_CPU: /* set up physical groups */
8810 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8811 if (!cpumask_empty(d->nodemask))
8812 init_sched_build_groups(d->nodemask, cpu_map,
8813 &cpu_to_phys_group,
8814 d->send_covered, d->tmpmask);
8815 break;
Andreas Herrmannde616e32009-08-18 13:00:13 +02008816#ifdef CONFIG_NUMA
8817 case SD_LV_ALLNODES:
8818 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8819 d->send_covered, d->tmpmask);
8820 break;
8821#endif
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008822 default:
8823 break;
8824 }
8825}
8826
Mike Travis7c16ec52008-04-04 18:11:11 -07008827/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008828 * Build sched domains for a given set of cpus and attach the sched domains
8829 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07008830 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308831static int __build_sched_domains(const struct cpumask *cpu_map,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008832 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008833{
Andreas Herrmann2109b992009-08-18 12:53:00 +02008834 enum s_alloc alloc_state = sa_none;
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008835 struct s_data d;
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008836 struct sched_domain *sd;
Andreas Herrmann2109b992009-08-18 12:53:00 +02008837 int i;
John Hawkesd1b55132005-09-06 15:18:14 -07008838#ifdef CONFIG_NUMA
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008839 d.sd_allnodes = 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308840#endif
8841
Andreas Herrmann2109b992009-08-18 12:53:00 +02008842 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8843 if (alloc_state != sa_rootdomain)
8844 goto error;
8845 alloc_state = sa_sched_groups;
Mike Travis7c16ec52008-04-04 18:11:11 -07008846
Linus Torvalds1da177e2005-04-16 15:20:36 -07008847 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008848 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008849 */
Rusty Russellabcd0832008-11-25 02:35:02 +10308850 for_each_cpu(i, cpu_map) {
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008851 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8852 cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008853
Andreas Herrmann7f4588f2009-08-18 12:54:06 +02008854 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
Andreas Herrmann87cce662009-08-18 12:54:55 +02008855 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmann410c4082009-08-18 12:56:14 +02008856 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
Andreas Herrmannd8173532009-08-18 12:57:03 +02008857 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008858 }
8859
Rusty Russellabcd0832008-11-25 02:35:02 +10308860 for_each_cpu(i, cpu_map) {
Andreas Herrmann0e8e85c2009-08-18 12:57:51 +02008861 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
Andreas Herrmanna2af04c2009-08-18 12:58:38 +02008862 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008863 }
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008864
Linus Torvalds1da177e2005-04-16 15:20:36 -07008865 /* Set up physical groups */
Andreas Herrmann86548092009-08-18 12:59:28 +02008866 for (i = 0; i < nr_node_ids; i++)
8867 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008868
8869#ifdef CONFIG_NUMA
8870 /* Set up node groups */
Andreas Herrmannde616e32009-08-18 13:00:13 +02008871 if (d.sd_allnodes)
8872 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008873
Andreas Herrmann0601a882009-08-18 13:01:11 +02008874 for (i = 0; i < nr_node_ids; i++)
8875 if (build_numa_sched_groups(&d, cpu_map, i))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008876 goto error;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008877#endif
8878
8879 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008880#ifdef CONFIG_SCHED_SMT
Rusty Russellabcd0832008-11-25 02:35:02 +10308881 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008882 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008883 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008884 }
8885#endif
8886#ifdef CONFIG_SCHED_MC
Rusty Russellabcd0832008-11-25 02:35:02 +10308887 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008888 sd = &per_cpu(core_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008889 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008890 }
8891#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008892
Rusty Russellabcd0832008-11-25 02:35:02 +10308893 for_each_cpu(i, cpu_map) {
Andreas Herrmann294b0c92009-08-18 13:02:29 +02008894 sd = &per_cpu(phys_domains, i).sd;
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008895 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008896 }
8897
John Hawkes9c1cfda2005-09-06 15:18:14 -07008898#ifdef CONFIG_NUMA
Mike Travis076ac2a2008-05-12 21:21:12 +02008899 for (i = 0; i < nr_node_ids; i++)
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008900 init_numa_sched_groups_power(d.sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008901
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008902 if (d.sd_allnodes) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008903 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07008904
Rusty Russell96f874e2008-11-25 02:35:14 +10308905 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008906 d.tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07008907 init_numa_sched_groups_power(sg);
8908 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008909#endif
8910
Linus Torvalds1da177e2005-04-16 15:20:36 -07008911 /* Attach the domains */
Rusty Russellabcd0832008-11-25 02:35:02 +10308912 for_each_cpu(i, cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008913#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308914 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008915#elif defined(CONFIG_SCHED_MC)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308916 sd = &per_cpu(core_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008917#else
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308918 sd = &per_cpu(phys_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008919#endif
Andreas Herrmann49a02c52009-08-18 12:51:52 +02008920 cpu_attach_domain(sd, d.rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008921 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008922
Andreas Herrmann2109b992009-08-18 12:53:00 +02008923 d.sched_group_nodes = NULL; /* don't free this we still need it */
8924 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8925 return 0;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308926
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008927error:
Andreas Herrmann2109b992009-08-18 12:53:00 +02008928 __free_domain_allocs(&d, alloc_state, cpu_map);
8929 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008930}
Paul Jackson029190c2007-10-18 23:40:20 -07008931
Rusty Russell96f874e2008-11-25 02:35:14 +10308932static int build_sched_domains(const struct cpumask *cpu_map)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008933{
8934 return __build_sched_domains(cpu_map, NULL);
8935}
8936
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308937static cpumask_var_t *doms_cur; /* current sched domains */
Paul Jackson029190c2007-10-18 23:40:20 -07008938static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02008939static struct sched_domain_attr *dattr_cur;
8940 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07008941
8942/*
8943 * Special case: If a kmalloc of a doms_cur partition (array of
Rusty Russell42128232008-11-25 02:35:12 +10308944 * cpumask) fails, then fallback to a single sched domain,
8945 * as determined by the single cpumask fallback_doms.
Paul Jackson029190c2007-10-18 23:40:20 -07008946 */
Rusty Russell42128232008-11-25 02:35:12 +10308947static cpumask_var_t fallback_doms;
Paul Jackson029190c2007-10-18 23:40:20 -07008948
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008949/*
8950 * arch_update_cpu_topology lets virtualized architectures update the
8951 * cpu core maps. It is supposed to return 1 if the topology changed
8952 * or 0 if it stayed the same.
8953 */
8954int __attribute__((weak)) arch_update_cpu_topology(void)
Heiko Carstens22e52b02008-03-12 18:31:59 +01008955{
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008956 return 0;
Heiko Carstens22e52b02008-03-12 18:31:59 +01008957}
8958
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308959cpumask_var_t *alloc_sched_domains(unsigned int ndoms)
8960{
8961 int i;
8962 cpumask_var_t *doms;
8963
8964 doms = kmalloc(sizeof(*doms) * ndoms, GFP_KERNEL);
8965 if (!doms)
8966 return NULL;
8967 for (i = 0; i < ndoms; i++) {
8968 if (!alloc_cpumask_var(&doms[i], GFP_KERNEL)) {
8969 free_sched_domains(doms, i);
8970 return NULL;
8971 }
8972 }
8973 return doms;
8974}
8975
8976void free_sched_domains(cpumask_var_t doms[], unsigned int ndoms)
8977{
8978 unsigned int i;
8979 for (i = 0; i < ndoms; i++)
8980 free_cpumask_var(doms[i]);
8981 kfree(doms);
8982}
8983
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008984/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008985 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07008986 * For now this just excludes isolated cpus, but could be used to
8987 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008988 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308989static int arch_init_sched_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008990{
Milton Miller73785472007-10-24 18:23:48 +02008991 int err;
8992
Heiko Carstens22e52b02008-03-12 18:31:59 +01008993 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07008994 ndoms_cur = 1;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308995 doms_cur = alloc_sched_domains(ndoms_cur);
Paul Jackson029190c2007-10-18 23:40:20 -07008996 if (!doms_cur)
Rusty Russellacc3f5d2009-11-03 14:53:40 +10308997 doms_cur = &fallback_doms;
8998 cpumask_andnot(doms_cur[0], cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008999 dattr_cur = NULL;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309000 err = build_sched_domains(doms_cur[0]);
Milton Miller6382bc92007-10-15 17:00:19 +02009001 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02009002
9003 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009004}
9005
Rusty Russell96f874e2008-11-25 02:35:14 +10309006static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9007 struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009008{
Mike Travis7c16ec52008-04-04 18:11:11 -07009009 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07009010}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009011
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009012/*
9013 * Detach sched domains from a group of cpus specified in cpu_map
9014 * These cpus will now be attached to the NULL domain
9015 */
Rusty Russell96f874e2008-11-25 02:35:14 +10309016static void detach_destroy_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009017{
Rusty Russell96f874e2008-11-25 02:35:14 +10309018 /* Save because hotplug lock held. */
9019 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009020 int i;
9021
Rusty Russellabcd0832008-11-25 02:35:02 +10309022 for_each_cpu(i, cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01009023 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009024 synchronize_sched();
Rusty Russell96f874e2008-11-25 02:35:14 +10309025 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07009026}
9027
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009028/* handle null as "default" */
9029static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9030 struct sched_domain_attr *new, int idx_new)
9031{
9032 struct sched_domain_attr tmp;
9033
9034 /* fast path */
9035 if (!new && !cur)
9036 return 1;
9037
9038 tmp = SD_ATTR_INIT;
9039 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9040 new ? (new + idx_new) : &tmp,
9041 sizeof(struct sched_domain_attr));
9042}
9043
Paul Jackson029190c2007-10-18 23:40:20 -07009044/*
9045 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009046 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07009047 * doms_new[] to the current sched domain partitioning, doms_cur[].
9048 * It destroys each deleted domain and builds each new domain.
9049 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309050 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009051 * The masks don't intersect (don't overlap.) We should setup one
9052 * sched domain for each mask. CPUs not in any of the cpumasks will
9053 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07009054 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9055 * it as it is.
9056 *
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309057 * The passed in 'doms_new' should be allocated using
9058 * alloc_sched_domains. This routine takes ownership of it and will
9059 * free_sched_domains it when done with it. If the caller failed the
9060 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9061 * and partition_sched_domains() will fallback to the single partition
9062 * 'fallback_doms', it also forces the domains to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07009063 *
Rusty Russell96f874e2008-11-25 02:35:14 +10309064 * If doms_new == NULL it will be replaced with cpu_online_mask.
Li Zefan700018e2008-11-18 14:02:03 +08009065 * ndoms_new == 0 is a special case for destroying existing domains,
9066 * and it will not create the default domain.
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009067 *
Paul Jackson029190c2007-10-18 23:40:20 -07009068 * Call with hotplug lock held
9069 */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309070void partition_sched_domains(int ndoms_new, cpumask_var_t doms_new[],
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009071 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07009072{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009073 int i, j, n;
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009074 int new_topology;
Paul Jackson029190c2007-10-18 23:40:20 -07009075
Heiko Carstens712555e2008-04-28 11:33:07 +02009076 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01009077
Milton Miller73785472007-10-24 18:23:48 +02009078 /* always unregister in case we don't destroy any domains */
9079 unregister_sched_domain_sysctl();
9080
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009081 /* Let architecture update cpu core mappings. */
9082 new_topology = arch_update_cpu_topology();
9083
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009084 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07009085
9086 /* Destroy deleted domains */
9087 for (i = 0; i < ndoms_cur; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009088 for (j = 0; j < n && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309089 if (cpumask_equal(doms_cur[i], doms_new[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009090 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07009091 goto match1;
9092 }
9093 /* no match - a current sched domain not in new doms_new[] */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309094 detach_destroy_domains(doms_cur[i]);
Paul Jackson029190c2007-10-18 23:40:20 -07009095match1:
9096 ;
9097 }
9098
Max Krasnyanskye761b772008-07-15 04:43:49 -07009099 if (doms_new == NULL) {
9100 ndoms_cur = 0;
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309101 doms_new = &fallback_doms;
9102 cpumask_andnot(doms_new[0], cpu_online_mask, cpu_isolated_map);
Li Zefanfaa2f982008-11-04 16:20:23 +08009103 WARN_ON_ONCE(dattr_new);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009104 }
9105
Paul Jackson029190c2007-10-18 23:40:20 -07009106 /* Build new domains */
9107 for (i = 0; i < ndoms_new; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01009108 for (j = 0; j < ndoms_cur && !new_topology; j++) {
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309109 if (cpumask_equal(doms_new[i], doms_cur[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009110 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07009111 goto match2;
9112 }
9113 /* no match - add a new doms_new */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309114 __build_sched_domains(doms_new[i],
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009115 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07009116match2:
9117 ;
9118 }
9119
9120 /* Remember the new sched domains */
Rusty Russellacc3f5d2009-11-03 14:53:40 +10309121 if (doms_cur != &fallback_doms)
9122 free_sched_domains(doms_cur, ndoms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009123 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07009124 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09009125 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07009126 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02009127
9128 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01009129
Heiko Carstens712555e2008-04-28 11:33:07 +02009130 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07009131}
9132
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009133#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Li Zefanc70f22d2009-01-05 19:07:50 +08009134static void arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009135{
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009136 get_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009137
9138 /* Destroy domains first to force the rebuild */
9139 partition_sched_domains(0, NULL, NULL);
9140
Max Krasnyanskye761b772008-07-15 04:43:49 -07009141 rebuild_sched_domains();
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009142 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009143}
9144
9145static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9146{
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309147 unsigned int level = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009148
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309149 if (sscanf(buf, "%u", &level) != 1)
9150 return -EINVAL;
9151
9152 /*
9153 * level is always be positive so don't check for
9154 * level < POWERSAVINGS_BALANCE_NONE which is 0
9155 * What happens on 0 or 1 byte write,
9156 * need to check for count as well?
9157 */
9158
9159 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009160 return -EINVAL;
9161
9162 if (smt)
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309163 sched_smt_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009164 else
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05309165 sched_mc_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009166
Li Zefanc70f22d2009-01-05 19:07:50 +08009167 arch_reinit_sched_domains();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009168
Li Zefanc70f22d2009-01-05 19:07:50 +08009169 return count;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009170}
9171
Adrian Bunk6707de002007-08-12 18:08:19 +02009172#ifdef CONFIG_SCHED_MC
Andi Kleenf718cd42008-07-29 22:33:52 -07009173static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9174 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009175{
9176 return sprintf(page, "%u\n", sched_mc_power_savings);
9177}
Andi Kleenf718cd42008-07-29 22:33:52 -07009178static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
Adrian Bunk6707de002007-08-12 18:08:19 +02009179 const char *buf, size_t count)
9180{
9181 return sched_power_savings_store(buf, count, 0);
9182}
Andi Kleenf718cd42008-07-29 22:33:52 -07009183static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9184 sched_mc_power_savings_show,
9185 sched_mc_power_savings_store);
Adrian Bunk6707de002007-08-12 18:08:19 +02009186#endif
9187
9188#ifdef CONFIG_SCHED_SMT
Andi Kleenf718cd42008-07-29 22:33:52 -07009189static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9190 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02009191{
9192 return sprintf(page, "%u\n", sched_smt_power_savings);
9193}
Andi Kleenf718cd42008-07-29 22:33:52 -07009194static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
Adrian Bunk6707de002007-08-12 18:08:19 +02009195 const char *buf, size_t count)
9196{
9197 return sched_power_savings_store(buf, count, 1);
9198}
Andi Kleenf718cd42008-07-29 22:33:52 -07009199static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9200 sched_smt_power_savings_show,
Adrian Bunk6707de002007-08-12 18:08:19 +02009201 sched_smt_power_savings_store);
9202#endif
9203
Li Zefan39aac642009-01-05 19:18:02 +08009204int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009205{
9206 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07009207
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009208#ifdef CONFIG_SCHED_SMT
9209 if (smt_capable())
9210 err = sysfs_create_file(&cls->kset.kobj,
9211 &attr_sched_smt_power_savings.attr);
9212#endif
9213#ifdef CONFIG_SCHED_MC
9214 if (!err && mc_capable())
9215 err = sysfs_create_file(&cls->kset.kobj,
9216 &attr_sched_mc_power_savings.attr);
9217#endif
9218 return err;
9219}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009220#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07009221
Max Krasnyanskye761b772008-07-15 04:43:49 -07009222#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009223/*
Max Krasnyanskye761b772008-07-15 04:43:49 -07009224 * Add online and remove offline CPUs from the scheduler domains.
9225 * When cpusets are enabled they take over this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07009226 */
9227static int update_sched_domains(struct notifier_block *nfb,
9228 unsigned long action, void *hcpu)
9229{
Max Krasnyanskye761b772008-07-15 04:43:49 -07009230 switch (action) {
9231 case CPU_ONLINE:
9232 case CPU_ONLINE_FROZEN:
9233 case CPU_DEAD:
9234 case CPU_DEAD_FROZEN:
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07009235 partition_sched_domains(1, NULL, NULL);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009236 return NOTIFY_OK;
9237
9238 default:
9239 return NOTIFY_DONE;
9240 }
9241}
9242#endif
9243
9244static int update_runtime(struct notifier_block *nfb,
9245 unsigned long action, void *hcpu)
9246{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009247 int cpu = (int)(long)hcpu;
9248
Linus Torvalds1da177e2005-04-16 15:20:36 -07009249 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07009250 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009251 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009252 disable_runtime(cpu_rq(cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07009253 return NOTIFY_OK;
9254
Linus Torvalds1da177e2005-04-16 15:20:36 -07009255 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009256 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07009257 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07009258 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02009259 enable_runtime(cpu_rq(cpu));
Max Krasnyanskye761b772008-07-15 04:43:49 -07009260 return NOTIFY_OK;
9261
Linus Torvalds1da177e2005-04-16 15:20:36 -07009262 default:
9263 return NOTIFY_DONE;
9264 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009265}
Linus Torvalds1da177e2005-04-16 15:20:36 -07009266
9267void __init sched_init_smp(void)
9268{
Rusty Russelldcc30a32008-11-25 02:35:12 +10309269 cpumask_var_t non_isolated_cpus;
9270
9271 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
Yong Zhangcb5fd132009-09-14 20:20:16 +08009272 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
Nick Piggin5c1e1762006-10-03 01:14:04 -07009273
Mike Travis434d53b2008-04-04 18:11:04 -07009274#if defined(CONFIG_NUMA)
9275 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9276 GFP_KERNEL);
9277 BUG_ON(sched_group_nodes_bycpu == NULL);
9278#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009279 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02009280 mutex_lock(&sched_domains_mutex);
Rusty Russelldcc30a32008-11-25 02:35:12 +10309281 arch_init_sched_domains(cpu_online_mask);
9282 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9283 if (cpumask_empty(non_isolated_cpus))
9284 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02009285 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009286 put_online_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009287
9288#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009289 /* XXX: Theoretical race here - CPU may be hotplugged now */
9290 hotcpu_notifier(update_sched_domains, 0);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009291#endif
9292
9293 /* RT runtime code needs to handle some hotplug events */
9294 hotcpu_notifier(update_runtime, 0);
9295
Peter Zijlstrab328ca12008-04-29 10:02:46 +02009296 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07009297
9298 /* Move init over to a non-isolated CPU */
Rusty Russelldcc30a32008-11-25 02:35:12 +10309299 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07009300 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01009301 sched_init_granularity();
Rusty Russelldcc30a32008-11-25 02:35:12 +10309302 free_cpumask_var(non_isolated_cpus);
Rusty Russell42128232008-11-25 02:35:12 +10309303
Rusty Russell0e3900e2008-11-25 02:35:13 +10309304 init_sched_rt_class();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009305}
9306#else
9307void __init sched_init_smp(void)
9308{
Ingo Molnar19978ca2007-11-09 22:39:38 +01009309 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009310}
9311#endif /* CONFIG_SMP */
9312
Arun R Bharadwajcd1bb942009-04-16 12:15:34 +05309313const_debug unsigned int sysctl_timer_migration = 1;
9314
Linus Torvalds1da177e2005-04-16 15:20:36 -07009315int in_sched_functions(unsigned long addr)
9316{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009317 return in_lock_functions(addr) ||
9318 (addr >= (unsigned long)__sched_text_start
9319 && addr < (unsigned long)__sched_text_end);
9320}
9321
Alexey Dobriyana9957442007-10-15 17:00:13 +02009322static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02009323{
9324 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02009325 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02009326#ifdef CONFIG_FAIR_GROUP_SCHED
9327 cfs_rq->rq = rq;
9328#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02009329 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02009330}
9331
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009332static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9333{
9334 struct rt_prio_array *array;
9335 int i;
9336
9337 array = &rt_rq->active;
9338 for (i = 0; i < MAX_RT_PRIO; i++) {
9339 INIT_LIST_HEAD(array->queue + i);
9340 __clear_bit(i, array->bitmap);
9341 }
9342 /* delimiter for bitsearch: */
9343 __set_bit(MAX_RT_PRIO, array->bitmap);
9344
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009345#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -05009346 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins398a1532009-01-14 09:10:04 -05009347#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -05009348 rt_rq->highest_prio.next = MAX_RT_PRIO;
Peter Zijlstra48d5e252008-01-25 21:08:31 +01009349#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009350#endif
9351#ifdef CONFIG_SMP
9352 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009353 rt_rq->overloaded = 0;
Fabio Checconic20b08e2009-06-15 20:56:38 +02009354 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009355#endif
9356
9357 rt_rq->rt_time = 0;
9358 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009359 rt_rq->rt_runtime = 0;
9360 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009361
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009362#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01009363 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009364 rt_rq->rq = rq;
9365#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009366}
9367
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009368#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009369static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9370 struct sched_entity *se, int cpu, int add,
9371 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009372{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009373 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009374 tg->cfs_rq[cpu] = cfs_rq;
9375 init_cfs_rq(cfs_rq, rq);
9376 cfs_rq->tg = tg;
9377 if (add)
9378 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9379
9380 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009381 /* se could be NULL for init_task_group */
9382 if (!se)
9383 return;
9384
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009385 if (!parent)
9386 se->cfs_rq = &rq->cfs;
9387 else
9388 se->cfs_rq = parent->my_q;
9389
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009390 se->my_q = cfs_rq;
9391 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009392 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009393 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009394}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009395#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009396
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009397#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009398static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9399 struct sched_rt_entity *rt_se, int cpu, int add,
9400 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009401{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009402 struct rq *rq = cpu_rq(cpu);
9403
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009404 tg->rt_rq[cpu] = rt_rq;
9405 init_rt_rq(rt_rq, rq);
9406 rt_rq->tg = tg;
9407 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009408 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009409 if (add)
9410 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9411
9412 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009413 if (!rt_se)
9414 return;
9415
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009416 if (!parent)
9417 rt_se->rt_rq = &rq->rt;
9418 else
9419 rt_se->rt_rq = parent->my_q;
9420
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009421 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009422 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009423 INIT_LIST_HEAD(&rt_se->run_list);
9424}
9425#endif
9426
Linus Torvalds1da177e2005-04-16 15:20:36 -07009427void __init sched_init(void)
9428{
Ingo Molnardd41f592007-07-09 18:51:59 +02009429 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07009430 unsigned long alloc_size = 0, ptr;
9431
9432#ifdef CONFIG_FAIR_GROUP_SCHED
9433 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9434#endif
9435#ifdef CONFIG_RT_GROUP_SCHED
9436 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9437#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009438#ifdef CONFIG_USER_SCHED
9439 alloc_size *= 2;
9440#endif
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309441#ifdef CONFIG_CPUMASK_OFFSTACK
Rusty Russell8c083f02009-03-19 15:22:20 +10309442 alloc_size += num_possible_cpus() * cpumask_size();
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309443#endif
Mike Travis434d53b2008-04-04 18:11:04 -07009444 if (alloc_size) {
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03009445 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07009446
9447#ifdef CONFIG_FAIR_GROUP_SCHED
9448 init_task_group.se = (struct sched_entity **)ptr;
9449 ptr += nr_cpu_ids * sizeof(void **);
9450
9451 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9452 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009453
9454#ifdef CONFIG_USER_SCHED
9455 root_task_group.se = (struct sched_entity **)ptr;
9456 ptr += nr_cpu_ids * sizeof(void **);
9457
9458 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9459 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009460#endif /* CONFIG_USER_SCHED */
9461#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009462#ifdef CONFIG_RT_GROUP_SCHED
9463 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9464 ptr += nr_cpu_ids * sizeof(void **);
9465
9466 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009467 ptr += nr_cpu_ids * sizeof(void **);
9468
9469#ifdef CONFIG_USER_SCHED
9470 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9471 ptr += nr_cpu_ids * sizeof(void **);
9472
9473 root_task_group.rt_rq = (struct rt_rq **)ptr;
9474 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009475#endif /* CONFIG_USER_SCHED */
9476#endif /* CONFIG_RT_GROUP_SCHED */
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309477#ifdef CONFIG_CPUMASK_OFFSTACK
9478 for_each_possible_cpu(i) {
9479 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9480 ptr += cpumask_size();
9481 }
9482#endif /* CONFIG_CPUMASK_OFFSTACK */
Mike Travis434d53b2008-04-04 18:11:04 -07009483 }
Ingo Molnardd41f592007-07-09 18:51:59 +02009484
Gregory Haskins57d885f2008-01-25 21:08:18 +01009485#ifdef CONFIG_SMP
9486 init_defrootdomain();
9487#endif
9488
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009489 init_rt_bandwidth(&def_rt_bandwidth,
9490 global_rt_period(), global_rt_runtime());
9491
9492#ifdef CONFIG_RT_GROUP_SCHED
9493 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9494 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009495#ifdef CONFIG_USER_SCHED
9496 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9497 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009498#endif /* CONFIG_USER_SCHED */
9499#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009500
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009501#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009502 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009503 INIT_LIST_HEAD(&init_task_group.children);
9504
9505#ifdef CONFIG_USER_SCHED
9506 INIT_LIST_HEAD(&root_task_group.children);
9507 init_task_group.parent = &root_task_group;
9508 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009509#endif /* CONFIG_USER_SCHED */
9510#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009511
Jiri Kosina4a6cc4b2009-10-29 00:26:00 +09009512#if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9513 update_shares_data = __alloc_percpu(nr_cpu_ids * sizeof(unsigned long),
9514 __alignof__(unsigned long));
9515#endif
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08009516 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07009517 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009518
9519 rq = cpu_rq(i);
9520 spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07009521 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009522 rq->calc_load_active = 0;
9523 rq->calc_load_update = jiffies + LOAD_FREQ;
Ingo Molnardd41f592007-07-09 18:51:59 +02009524 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009525 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009526#ifdef CONFIG_FAIR_GROUP_SCHED
9527 init_task_group.shares = init_task_group_load;
9528 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009529#ifdef CONFIG_CGROUP_SCHED
9530 /*
9531 * How much cpu bandwidth does init_task_group get?
9532 *
9533 * In case of task-groups formed thr' the cgroup filesystem, it
9534 * gets 100% of the cpu resources in the system. This overall
9535 * system cpu resource is divided among the tasks of
9536 * init_task_group and its child task-groups in a fair manner,
9537 * based on each entity's (task or task-group's) weight
9538 * (se->load.weight).
9539 *
9540 * In other words, if init_task_group has 10 tasks of weight
9541 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9542 * then A0's share of the cpu resource is:
9543 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009544 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02009545 *
9546 * We achieve this by letting init_task_group's tasks sit
9547 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9548 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009549 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009550#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009551 root_task_group.shares = NICE_0_LOAD;
9552 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009553 /*
9554 * In case of task-groups formed thr' the user id of tasks,
9555 * init_task_group represents tasks belonging to root user.
9556 * Hence it forms a sibling of all subsequent groups formed.
9557 * In this case, init_task_group gets only a fraction of overall
9558 * system cpu resource, based on the weight assigned to root
9559 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9560 * by letting tasks of init_task_group sit in a separate cfs_rq
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009561 * (init_tg_cfs_rq) and having one entity represent this group of
Dhaval Giani354d60c2008-04-19 19:44:59 +02009562 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9563 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009564 init_tg_cfs_entry(&init_task_group,
Anirban Sinha84e9dab2009-08-28 22:40:43 -07009565 &per_cpu(init_tg_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009566 &per_cpu(init_sched_entity, i), i, 1,
9567 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009568
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009569#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02009570#endif /* CONFIG_FAIR_GROUP_SCHED */
9571
9572 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009573#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009574 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009575#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009576 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009577#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009578 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009579 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009580 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009581 &per_cpu(init_sched_rt_entity, i), i, 1,
9582 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009583#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009584#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009585
Ingo Molnardd41f592007-07-09 18:51:59 +02009586 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9587 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009588#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07009589 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01009590 rq->rd = NULL;
Gregory Haskins3f029d32009-07-29 11:08:47 -04009591 rq->post_schedule = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009592 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009593 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009594 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07009595 rq->cpu = i;
Gregory Haskins1f11eb62008-06-04 15:04:05 -04009596 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009597 rq->migration_thread = NULL;
Mike Galbraitheae0c9d2009-11-10 03:50:02 +01009598 rq->idle_stamp = 0;
9599 rq->avg_idle = 2*sysctl_sched_migration_cost;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009600 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01009601 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009602#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009603 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009604 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009605 }
9606
Peter Williams2dd73a42006-06-27 02:54:34 -07009607 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009608
Avi Kivitye107be32007-07-26 13:40:43 +02009609#ifdef CONFIG_PREEMPT_NOTIFIERS
9610 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9611#endif
9612
Christoph Lameterc9819f42006-12-10 02:20:25 -08009613#ifdef CONFIG_SMP
Carlos R. Mafra962cf362008-05-15 11:15:37 -03009614 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
Christoph Lameterc9819f42006-12-10 02:20:25 -08009615#endif
9616
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009617#ifdef CONFIG_RT_MUTEXES
9618 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9619#endif
9620
Linus Torvalds1da177e2005-04-16 15:20:36 -07009621 /*
9622 * The boot idle thread does lazy MMU switching as well:
9623 */
9624 atomic_inc(&init_mm.mm_count);
9625 enter_lazy_tlb(&init_mm, current);
9626
9627 /*
9628 * Make us the idle thread. Technically, schedule() should not be
9629 * called from this thread, however somewhere below it might be,
9630 * but because we are the idle thread, we just pick up running again
9631 * when this runqueue becomes "idle".
9632 */
9633 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009634
9635 calc_load_update = jiffies + LOAD_FREQ;
9636
Ingo Molnardd41f592007-07-09 18:51:59 +02009637 /*
9638 * During early bootup we pretend to be a normal task:
9639 */
9640 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01009641
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309642 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
Rusty Russell49557e62009-11-02 20:37:20 +10309643 zalloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309644#ifdef CONFIG_SMP
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309645#ifdef CONFIG_NO_HZ
Rusty Russell49557e62009-11-02 20:37:20 +10309646 zalloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009647 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309648#endif
Rusty Russellbdddd292009-12-02 14:09:16 +10309649 /* May be allocated at isolcpus cmdline parse time */
9650 if (cpu_isolated_map == NULL)
9651 zalloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309652#endif /* SMP */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309653
Ingo Molnarcdd6c482009-09-21 12:02:48 +02009654 perf_event_init();
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009655
Ingo Molnar6892b752008-02-13 14:02:36 +01009656 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009657}
9658
9659#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009660static inline int preempt_count_equals(int preempt_offset)
9661{
9662 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9663
9664 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9665}
9666
9667void __might_sleep(char *file, int line, int preempt_offset)
Linus Torvalds1da177e2005-04-16 15:20:36 -07009668{
Ingo Molnar48f24c42006-07-03 00:25:40 -07009669#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07009670 static unsigned long prev_jiffy; /* ratelimiting */
9671
Frederic Weisbeckere4aafea2009-07-16 15:44:29 +02009672 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9673 system_state != SYSTEM_RUNNING || oops_in_progress)
Ingo Molnaraef745f2008-08-28 11:34:43 +02009674 return;
9675 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9676 return;
9677 prev_jiffy = jiffies;
9678
9679 printk(KERN_ERR
9680 "BUG: sleeping function called from invalid context at %s:%d\n",
9681 file, line);
9682 printk(KERN_ERR
9683 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9684 in_atomic(), irqs_disabled(),
9685 current->pid, current->comm);
9686
9687 debug_show_held_locks(current);
9688 if (irqs_disabled())
9689 print_irqtrace_events(current);
9690 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009691#endif
9692}
9693EXPORT_SYMBOL(__might_sleep);
9694#endif
9695
9696#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009697static void normalize_task(struct rq *rq, struct task_struct *p)
9698{
9699 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02009700
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009701 update_rq_clock(rq);
9702 on_rq = p->se.on_rq;
9703 if (on_rq)
9704 deactivate_task(rq, p, 0);
9705 __setscheduler(rq, p, SCHED_NORMAL, 0);
9706 if (on_rq) {
9707 activate_task(rq, p, 0);
9708 resched_task(rq->curr);
9709 }
9710}
9711
Linus Torvalds1da177e2005-04-16 15:20:36 -07009712void normalize_rt_tasks(void)
9713{
Ingo Molnara0f98a12007-06-17 18:37:45 +02009714 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009715 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07009716 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009717
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009718 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009719 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02009720 /*
9721 * Only normalize user tasks:
9722 */
9723 if (!p->mm)
9724 continue;
9725
Ingo Molnardd41f592007-07-09 18:51:59 +02009726 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009727#ifdef CONFIG_SCHEDSTATS
9728 p->se.wait_start = 0;
9729 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009730 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009731#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02009732
9733 if (!rt_task(p)) {
9734 /*
9735 * Renice negative nice level userspace
9736 * tasks back to 0:
9737 */
9738 if (TASK_NICE(p) < 0 && p->mm)
9739 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009740 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02009741 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009742
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009743 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07009744 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009745
Ingo Molnar178be792007-10-15 17:00:18 +02009746 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009747
Ingo Molnarb29739f2006-06-27 02:54:51 -07009748 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009749 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009750 } while_each_thread(g, p);
9751
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009752 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009753}
9754
9755#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07009756
9757#ifdef CONFIG_IA64
9758/*
9759 * These functions are only useful for the IA64 MCA handling.
9760 *
9761 * They can only be called when the whole system has been
9762 * stopped - every CPU needs to be quiescent, and no scheduling
9763 * activity can take place. Using them for anything else would
9764 * be a serious bug, and as a result, they aren't even visible
9765 * under any other configuration.
9766 */
9767
9768/**
9769 * curr_task - return the current task for a given cpu.
9770 * @cpu: the processor in question.
9771 *
9772 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9773 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009774struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009775{
9776 return cpu_curr(cpu);
9777}
9778
9779/**
9780 * set_curr_task - set the current task for a given cpu.
9781 * @cpu: the processor in question.
9782 * @p: the task pointer to set.
9783 *
9784 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009785 * are serviced on a separate stack. It allows the architecture to switch the
9786 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07009787 * must be called with all CPU's synchronized, and interrupts disabled, the
9788 * and caller must save the original value of the current task (see
9789 * curr_task() above) and restore that value before reenabling interrupts and
9790 * re-starting the system.
9791 *
9792 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9793 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009794void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009795{
9796 cpu_curr(cpu) = p;
9797}
9798
9799#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009800
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009801#ifdef CONFIG_FAIR_GROUP_SCHED
9802static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009803{
9804 int i;
9805
9806 for_each_possible_cpu(i) {
9807 if (tg->cfs_rq)
9808 kfree(tg->cfs_rq[i]);
9809 if (tg->se)
9810 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009811 }
9812
9813 kfree(tg->cfs_rq);
9814 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009815}
9816
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009817static
9818int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009819{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009820 struct cfs_rq *cfs_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009821 struct sched_entity *se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009822 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009823 int i;
9824
Mike Travis434d53b2008-04-04 18:11:04 -07009825 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009826 if (!tg->cfs_rq)
9827 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009828 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009829 if (!tg->se)
9830 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009831
9832 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009833
9834 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009835 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009836
Li Zefaneab17222008-10-29 17:03:22 +08009837 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9838 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009839 if (!cfs_rq)
9840 goto err;
9841
Li Zefaneab17222008-10-29 17:03:22 +08009842 se = kzalloc_node(sizeof(struct sched_entity),
9843 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009844 if (!se)
9845 goto err;
9846
Li Zefaneab17222008-10-29 17:03:22 +08009847 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009848 }
9849
9850 return 1;
9851
9852 err:
9853 return 0;
9854}
9855
9856static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9857{
9858 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9859 &cpu_rq(cpu)->leaf_cfs_rq_list);
9860}
9861
9862static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9863{
9864 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9865}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009866#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009867static inline void free_fair_sched_group(struct task_group *tg)
9868{
9869}
9870
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009871static inline
9872int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009873{
9874 return 1;
9875}
9876
9877static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9878{
9879}
9880
9881static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9882{
9883}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009884#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009885
9886#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009887static void free_rt_sched_group(struct task_group *tg)
9888{
9889 int i;
9890
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009891 destroy_rt_bandwidth(&tg->rt_bandwidth);
9892
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009893 for_each_possible_cpu(i) {
9894 if (tg->rt_rq)
9895 kfree(tg->rt_rq[i]);
9896 if (tg->rt_se)
9897 kfree(tg->rt_se[i]);
9898 }
9899
9900 kfree(tg->rt_rq);
9901 kfree(tg->rt_se);
9902}
9903
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009904static
9905int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009906{
9907 struct rt_rq *rt_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009908 struct sched_rt_entity *rt_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009909 struct rq *rq;
9910 int i;
9911
Mike Travis434d53b2008-04-04 18:11:04 -07009912 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009913 if (!tg->rt_rq)
9914 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009915 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009916 if (!tg->rt_se)
9917 goto err;
9918
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009919 init_rt_bandwidth(&tg->rt_bandwidth,
9920 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009921
9922 for_each_possible_cpu(i) {
9923 rq = cpu_rq(i);
9924
Li Zefaneab17222008-10-29 17:03:22 +08009925 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9926 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009927 if (!rt_rq)
9928 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009929
Li Zefaneab17222008-10-29 17:03:22 +08009930 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9931 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009932 if (!rt_se)
9933 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009934
Li Zefaneab17222008-10-29 17:03:22 +08009935 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009936 }
9937
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009938 return 1;
9939
9940 err:
9941 return 0;
9942}
9943
9944static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9945{
9946 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9947 &cpu_rq(cpu)->leaf_rt_rq_list);
9948}
9949
9950static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9951{
9952 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9953}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009954#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009955static inline void free_rt_sched_group(struct task_group *tg)
9956{
9957}
9958
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009959static inline
9960int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009961{
9962 return 1;
9963}
9964
9965static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9966{
9967}
9968
9969static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9970{
9971}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009972#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009973
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009974#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009975static void free_sched_group(struct task_group *tg)
9976{
9977 free_fair_sched_group(tg);
9978 free_rt_sched_group(tg);
9979 kfree(tg);
9980}
9981
9982/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009983struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009984{
9985 struct task_group *tg;
9986 unsigned long flags;
9987 int i;
9988
9989 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9990 if (!tg)
9991 return ERR_PTR(-ENOMEM);
9992
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009993 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009994 goto err;
9995
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009996 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009997 goto err;
9998
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009999 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010000 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010001 register_fair_sched_group(tg, i);
10002 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010003 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010004 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010005
10006 WARN_ON(!parent); /* root should already exist */
10007
10008 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010009 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +080010010 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010011 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010012
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010013 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010014
10015err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010016 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010017 return ERR_PTR(-ENOMEM);
10018}
10019
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010020/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010021static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010022{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010023 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010024 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010025}
10026
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010027/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010028void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010029{
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010030 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010031 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010032
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010033 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010034 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010035 unregister_fair_sched_group(tg, i);
10036 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010037 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010038 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010039 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010040 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010041
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010042 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010043 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010044}
10045
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010046/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +020010047 * The caller of this function should have put the task in its new group
10048 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10049 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010050 */
10051void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010052{
10053 int on_rq, running;
10054 unsigned long flags;
10055 struct rq *rq;
10056
10057 rq = task_rq_lock(tsk, &flags);
10058
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010059 update_rq_clock(rq);
10060
Dmitry Adamushko051a1d12007-12-18 15:21:13 +010010061 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010062 on_rq = tsk->se.on_rq;
10063
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010064 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010065 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010066 if (unlikely(running))
10067 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010068
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010069 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010070
Peter Zijlstra810b3812008-02-29 15:21:01 -050010071#ifdef CONFIG_FAIR_GROUP_SCHED
10072 if (tsk->sched_class->moved_group)
10073 tsk->sched_class->moved_group(tsk);
10074#endif
10075
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -070010076 if (unlikely(running))
10077 tsk->sched_class->set_curr_task(rq);
10078 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +020010079 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010080
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010081 task_rq_unlock(rq, &flags);
10082}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010083#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010084
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010085#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010086static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010087{
10088 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010089 int on_rq;
10090
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010091 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010092 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010093 dequeue_entity(cfs_rq, se, 0);
10094
10095 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +020010096 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010097
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010098 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010099 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010100}
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010101
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010102static void set_se_shares(struct sched_entity *se, unsigned long shares)
10103{
10104 struct cfs_rq *cfs_rq = se->cfs_rq;
10105 struct rq *rq = cfs_rq->rq;
10106 unsigned long flags;
10107
10108 spin_lock_irqsave(&rq->lock, flags);
10109 __set_se_shares(se, shares);
10110 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010111}
10112
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010113static DEFINE_MUTEX(shares_mutex);
10114
Ingo Molnar4cf86d72007-10-15 17:00:14 +020010115int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010116{
10117 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010118 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +010010119
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010120 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010121 * We can't change the weight of the root cgroup.
10122 */
10123 if (!tg->se[0])
10124 return -EINVAL;
10125
Peter Zijlstra18d95a22008-04-19 19:45:00 +020010126 if (shares < MIN_SHARES)
10127 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010128 else if (shares > MAX_SHARES)
10129 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +010010130
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010131 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010132 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010133 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010134
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010135 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010136 for_each_possible_cpu(i)
10137 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010138 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010139 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010140
10141 /* wait for any ongoing reference to this group to finish */
10142 synchronize_sched();
10143
10144 /*
10145 * Now we are free to modify the group's share on each cpu
10146 * w/o tripping rebalance_share or load_balance_fair.
10147 */
10148 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010149 for_each_possible_cpu(i) {
10150 /*
10151 * force a rebalance
10152 */
10153 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +080010154 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +020010155 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +010010156
10157 /*
10158 * Enable load balance activity on this group, by inserting it back on
10159 * each cpu's rq->leaf_cfs_rq_list.
10160 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010161 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +010010162 for_each_possible_cpu(i)
10163 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +020010164 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010165 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010166done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +010010167 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +020010168 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +020010169}
10170
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010171unsigned long sched_group_shares(struct task_group *tg)
10172{
10173 return tg->shares;
10174}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010175#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +020010176
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010177#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010178/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010179 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010180 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010181static DEFINE_MUTEX(rt_constraints_mutex);
10182
10183static unsigned long to_ratio(u64 period, u64 runtime)
10184{
10185 if (runtime == RUNTIME_INF)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010186 return 1ULL << 20;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010187
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010188 return div64_u64(runtime << 20, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010189}
10190
Dhaval Giani521f1a242008-02-28 15:21:56 +053010191/* Must be called with tasklist_lock held */
10192static inline int tg_has_rt_tasks(struct task_group *tg)
10193{
10194 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010195
Dhaval Giani521f1a242008-02-28 15:21:56 +053010196 do_each_thread(g, p) {
10197 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10198 return 1;
10199 } while_each_thread(g, p);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010200
Dhaval Giani521f1a242008-02-28 15:21:56 +053010201 return 0;
10202}
10203
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010204struct rt_schedulable_data {
10205 struct task_group *tg;
10206 u64 rt_period;
10207 u64 rt_runtime;
10208};
10209
10210static int tg_schedulable(struct task_group *tg, void *data)
10211{
10212 struct rt_schedulable_data *d = data;
10213 struct task_group *child;
10214 unsigned long total, sum = 0;
10215 u64 period, runtime;
10216
10217 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10218 runtime = tg->rt_bandwidth.rt_runtime;
10219
10220 if (tg == d->tg) {
10221 period = d->rt_period;
10222 runtime = d->rt_runtime;
10223 }
10224
Peter Zijlstra98a48262009-01-14 10:56:32 +010010225#ifdef CONFIG_USER_SCHED
10226 if (tg == &root_task_group) {
10227 period = global_rt_period();
10228 runtime = global_rt_runtime();
10229 }
10230#endif
10231
Peter Zijlstra4653f802008-09-23 15:33:44 +020010232 /*
10233 * Cannot have more runtime than the period.
10234 */
10235 if (runtime > period && runtime != RUNTIME_INF)
10236 return -EINVAL;
10237
10238 /*
10239 * Ensure we don't starve existing RT tasks.
10240 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010241 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10242 return -EBUSY;
10243
10244 total = to_ratio(period, runtime);
10245
Peter Zijlstra4653f802008-09-23 15:33:44 +020010246 /*
10247 * Nobody can have more than the global setting allows.
10248 */
10249 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10250 return -EINVAL;
10251
10252 /*
10253 * The sum of our children's runtime should not exceed our own.
10254 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010255 list_for_each_entry_rcu(child, &tg->children, siblings) {
10256 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10257 runtime = child->rt_bandwidth.rt_runtime;
10258
10259 if (child == d->tg) {
10260 period = d->rt_period;
10261 runtime = d->rt_runtime;
10262 }
10263
10264 sum += to_ratio(period, runtime);
10265 }
10266
10267 if (sum > total)
10268 return -EINVAL;
10269
10270 return 0;
10271}
10272
10273static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10274{
10275 struct rt_schedulable_data data = {
10276 .tg = tg,
10277 .rt_period = period,
10278 .rt_runtime = runtime,
10279 };
10280
10281 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10282}
10283
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010284static int tg_set_bandwidth(struct task_group *tg,
10285 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010286{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010287 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010288
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010289 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +053010290 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010291 err = __rt_schedulable(tg, rt_period, rt_runtime);
10292 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +053010293 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010294
10295 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010296 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10297 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010298
10299 for_each_possible_cpu(i) {
10300 struct rt_rq *rt_rq = tg->rt_rq[i];
10301
10302 spin_lock(&rt_rq->rt_runtime_lock);
10303 rt_rq->rt_runtime = rt_runtime;
10304 spin_unlock(&rt_rq->rt_runtime_lock);
10305 }
10306 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010307 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +053010308 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010309 mutex_unlock(&rt_constraints_mutex);
10310
10311 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010312}
10313
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010314int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10315{
10316 u64 rt_runtime, rt_period;
10317
10318 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10319 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10320 if (rt_runtime_us < 0)
10321 rt_runtime = RUNTIME_INF;
10322
10323 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10324}
10325
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010326long sched_group_rt_runtime(struct task_group *tg)
10327{
10328 u64 rt_runtime_us;
10329
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010330 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010331 return -1;
10332
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010333 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010334 do_div(rt_runtime_us, NSEC_PER_USEC);
10335 return rt_runtime_us;
10336}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010337
10338int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10339{
10340 u64 rt_runtime, rt_period;
10341
10342 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10343 rt_runtime = tg->rt_bandwidth.rt_runtime;
10344
Raistlin619b0482008-06-26 18:54:09 +020010345 if (rt_period == 0)
10346 return -EINVAL;
10347
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010348 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10349}
10350
10351long sched_group_rt_period(struct task_group *tg)
10352{
10353 u64 rt_period_us;
10354
10355 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10356 do_div(rt_period_us, NSEC_PER_USEC);
10357 return rt_period_us;
10358}
10359
10360static int sched_rt_global_constraints(void)
10361{
Peter Zijlstra4653f802008-09-23 15:33:44 +020010362 u64 runtime, period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010363 int ret = 0;
10364
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010365 if (sysctl_sched_rt_period <= 0)
10366 return -EINVAL;
10367
Peter Zijlstra4653f802008-09-23 15:33:44 +020010368 runtime = global_rt_runtime();
10369 period = global_rt_period();
10370
10371 /*
10372 * Sanity check on the sysctl variables.
10373 */
10374 if (runtime > period && runtime != RUNTIME_INF)
10375 return -EINVAL;
Peter Zijlstra10b612f2008-06-19 14:22:27 +020010376
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010377 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010378 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +020010379 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010380 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010381 mutex_unlock(&rt_constraints_mutex);
10382
10383 return ret;
10384}
Dhaval Giani54e99122009-02-27 15:13:54 +053010385
10386int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10387{
10388 /* Don't accept realtime tasks when there is no way for them to run */
10389 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10390 return 0;
10391
10392 return 1;
10393}
10394
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010395#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010396static int sched_rt_global_constraints(void)
10397{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010398 unsigned long flags;
10399 int i;
10400
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010401 if (sysctl_sched_rt_period <= 0)
10402 return -EINVAL;
10403
Peter Zijlstra60aa6052009-05-05 17:50:21 +020010404 /*
10405 * There's always some RT tasks in the root group
10406 * -- migration, kstopmachine etc..
10407 */
10408 if (sysctl_sched_rt_runtime == 0)
10409 return -EBUSY;
10410
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010411 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10412 for_each_possible_cpu(i) {
10413 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10414
10415 spin_lock(&rt_rq->rt_runtime_lock);
10416 rt_rq->rt_runtime = global_rt_runtime();
10417 spin_unlock(&rt_rq->rt_runtime_lock);
10418 }
10419 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10420
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010421 return 0;
10422}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010423#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010424
10425int sched_rt_handler(struct ctl_table *table, int write,
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070010426 void __user *buffer, size_t *lenp,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010427 loff_t *ppos)
10428{
10429 int ret;
10430 int old_period, old_runtime;
10431 static DEFINE_MUTEX(mutex);
10432
10433 mutex_lock(&mutex);
10434 old_period = sysctl_sched_rt_period;
10435 old_runtime = sysctl_sched_rt_runtime;
10436
Alexey Dobriyan8d65af72009-09-23 15:57:19 -070010437 ret = proc_dointvec(table, write, buffer, lenp, ppos);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010438
10439 if (!ret && write) {
10440 ret = sched_rt_global_constraints();
10441 if (ret) {
10442 sysctl_sched_rt_period = old_period;
10443 sysctl_sched_rt_runtime = old_runtime;
10444 } else {
10445 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10446 def_rt_bandwidth.rt_period =
10447 ns_to_ktime(global_rt_period());
10448 }
10449 }
10450 mutex_unlock(&mutex);
10451
10452 return ret;
10453}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010454
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010455#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010456
10457/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +020010458static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010459{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010460 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10461 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010462}
10463
10464static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +020010465cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010466{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010467 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010468
Paul Menage2b01dfe2007-10-24 18:23:50 +020010469 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010470 /* This is early initialization for the top cgroup */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010471 return &init_task_group.css;
10472 }
10473
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010474 parent = cgroup_tg(cgrp->parent);
10475 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010476 if (IS_ERR(tg))
10477 return ERR_PTR(-ENOMEM);
10478
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010479 return &tg->css;
10480}
10481
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010482static void
10483cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010484{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010485 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010486
10487 sched_destroy_group(tg);
10488}
10489
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010490static int
Ben Blumbe367d02009-09-23 15:56:31 -070010491cpu_cgroup_can_attach_task(struct cgroup *cgrp, struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010492{
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010493#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Giani54e99122009-02-27 15:13:54 +053010494 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010495 return -EINVAL;
10496#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010497 /* We don't support RT-tasks being in separate groups */
10498 if (tsk->sched_class != &fair_sched_class)
10499 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010500#endif
Ben Blumbe367d02009-09-23 15:56:31 -070010501 return 0;
10502}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010503
Ben Blumbe367d02009-09-23 15:56:31 -070010504static int
10505cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10506 struct task_struct *tsk, bool threadgroup)
10507{
10508 int retval = cpu_cgroup_can_attach_task(cgrp, tsk);
10509 if (retval)
10510 return retval;
10511 if (threadgroup) {
10512 struct task_struct *c;
10513 rcu_read_lock();
10514 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10515 retval = cpu_cgroup_can_attach_task(cgrp, c);
10516 if (retval) {
10517 rcu_read_unlock();
10518 return retval;
10519 }
10520 }
10521 rcu_read_unlock();
10522 }
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010523 return 0;
10524}
10525
10526static void
Paul Menage2b01dfe2007-10-24 18:23:50 +020010527cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Ben Blumbe367d02009-09-23 15:56:31 -070010528 struct cgroup *old_cont, struct task_struct *tsk,
10529 bool threadgroup)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010530{
10531 sched_move_task(tsk);
Ben Blumbe367d02009-09-23 15:56:31 -070010532 if (threadgroup) {
10533 struct task_struct *c;
10534 rcu_read_lock();
10535 list_for_each_entry_rcu(c, &tsk->thread_group, thread_group) {
10536 sched_move_task(c);
10537 }
10538 rcu_read_unlock();
10539 }
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010540}
10541
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010542#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -070010543static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +020010544 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010545{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010546 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010547}
10548
Paul Menagef4c753b2008-04-29 00:59:56 -070010549static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010550{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010551 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010552
10553 return (u64) tg->shares;
10554}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010555#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010556
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010557#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -070010558static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -070010559 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010560{
Paul Menage06ecb272008-04-29 01:00:06 -070010561 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010562}
10563
Paul Menage06ecb272008-04-29 01:00:06 -070010564static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010565{
Paul Menage06ecb272008-04-29 01:00:06 -070010566 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010567}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010568
10569static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10570 u64 rt_period_us)
10571{
10572 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10573}
10574
10575static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10576{
10577 return sched_group_rt_period(cgroup_tg(cgrp));
10578}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010579#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010580
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010581static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010582#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010583 {
10584 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -070010585 .read_u64 = cpu_shares_read_u64,
10586 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010587 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010588#endif
10589#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010590 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010591 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -070010592 .read_s64 = cpu_rt_runtime_read,
10593 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010594 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010595 {
10596 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -070010597 .read_u64 = cpu_rt_period_read_uint,
10598 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010599 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010600#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010601};
10602
10603static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10604{
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010605 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010606}
10607
10608struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +010010609 .name = "cpu",
10610 .create = cpu_cgroup_create,
10611 .destroy = cpu_cgroup_destroy,
10612 .can_attach = cpu_cgroup_can_attach,
10613 .attach = cpu_cgroup_attach,
10614 .populate = cpu_cgroup_populate,
10615 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010616 .early_init = 1,
10617};
10618
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010619#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010620
10621#ifdef CONFIG_CGROUP_CPUACCT
10622
10623/*
10624 * CPU accounting code for task groups.
10625 *
10626 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10627 * (balbir@in.ibm.com).
10628 */
10629
Bharata B Rao934352f2008-11-10 20:41:13 +053010630/* track cpu usage of a group of tasks and its child groups */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010631struct cpuacct {
10632 struct cgroup_subsys_state css;
10633 /* cpuusage holds pointer to a u64-type object on every cpu */
10634 u64 *cpuusage;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010635 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
Bharata B Rao934352f2008-11-10 20:41:13 +053010636 struct cpuacct *parent;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010637};
10638
10639struct cgroup_subsys cpuacct_subsys;
10640
10641/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010642static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010643{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010644 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010645 struct cpuacct, css);
10646}
10647
10648/* return cpu accounting group to which this task belongs */
10649static inline struct cpuacct *task_ca(struct task_struct *tsk)
10650{
10651 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10652 struct cpuacct, css);
10653}
10654
10655/* create a new cpu accounting group */
10656static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +053010657 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010658{
10659 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010660 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010661
10662 if (!ca)
Bharata B Raoef12fef2009-03-31 10:02:22 +053010663 goto out;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010664
10665 ca->cpuusage = alloc_percpu(u64);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010666 if (!ca->cpuusage)
10667 goto out_free_ca;
10668
10669 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10670 if (percpu_counter_init(&ca->cpustat[i], 0))
10671 goto out_free_counters;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010672
Bharata B Rao934352f2008-11-10 20:41:13 +053010673 if (cgrp->parent)
10674 ca->parent = cgroup_ca(cgrp->parent);
10675
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010676 return &ca->css;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010677
10678out_free_counters:
10679 while (--i >= 0)
10680 percpu_counter_destroy(&ca->cpustat[i]);
10681 free_percpu(ca->cpuusage);
10682out_free_ca:
10683 kfree(ca);
10684out:
10685 return ERR_PTR(-ENOMEM);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010686}
10687
10688/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010689static void
Dhaval Giani32cd7562008-02-29 10:02:43 +053010690cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010691{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010692 struct cpuacct *ca = cgroup_ca(cgrp);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010693 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010694
Bharata B Raoef12fef2009-03-31 10:02:22 +053010695 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10696 percpu_counter_destroy(&ca->cpustat[i]);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010697 free_percpu(ca->cpuusage);
10698 kfree(ca);
10699}
10700
Ken Chen720f5492008-12-15 22:02:01 -080010701static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10702{
Rusty Russellb36128c2009-02-20 16:29:08 +090010703 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010704 u64 data;
10705
10706#ifndef CONFIG_64BIT
10707 /*
10708 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10709 */
10710 spin_lock_irq(&cpu_rq(cpu)->lock);
10711 data = *cpuusage;
10712 spin_unlock_irq(&cpu_rq(cpu)->lock);
10713#else
10714 data = *cpuusage;
10715#endif
10716
10717 return data;
10718}
10719
10720static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10721{
Rusty Russellb36128c2009-02-20 16:29:08 +090010722 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010723
10724#ifndef CONFIG_64BIT
10725 /*
10726 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10727 */
10728 spin_lock_irq(&cpu_rq(cpu)->lock);
10729 *cpuusage = val;
10730 spin_unlock_irq(&cpu_rq(cpu)->lock);
10731#else
10732 *cpuusage = val;
10733#endif
10734}
10735
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010736/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010737static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010738{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010739 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010740 u64 totalcpuusage = 0;
10741 int i;
10742
Ken Chen720f5492008-12-15 22:02:01 -080010743 for_each_present_cpu(i)
10744 totalcpuusage += cpuacct_cpuusage_read(ca, i);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010745
10746 return totalcpuusage;
10747}
10748
Dhaval Giani0297b802008-02-29 10:02:44 +053010749static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10750 u64 reset)
10751{
10752 struct cpuacct *ca = cgroup_ca(cgrp);
10753 int err = 0;
10754 int i;
10755
10756 if (reset) {
10757 err = -EINVAL;
10758 goto out;
10759 }
10760
Ken Chen720f5492008-12-15 22:02:01 -080010761 for_each_present_cpu(i)
10762 cpuacct_cpuusage_write(ca, i, 0);
Dhaval Giani0297b802008-02-29 10:02:44 +053010763
Dhaval Giani0297b802008-02-29 10:02:44 +053010764out:
10765 return err;
10766}
10767
Ken Chene9515c32008-12-15 22:04:15 -080010768static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10769 struct seq_file *m)
10770{
10771 struct cpuacct *ca = cgroup_ca(cgroup);
10772 u64 percpu;
10773 int i;
10774
10775 for_each_present_cpu(i) {
10776 percpu = cpuacct_cpuusage_read(ca, i);
10777 seq_printf(m, "%llu ", (unsigned long long) percpu);
10778 }
10779 seq_printf(m, "\n");
10780 return 0;
10781}
10782
Bharata B Raoef12fef2009-03-31 10:02:22 +053010783static const char *cpuacct_stat_desc[] = {
10784 [CPUACCT_STAT_USER] = "user",
10785 [CPUACCT_STAT_SYSTEM] = "system",
10786};
10787
10788static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10789 struct cgroup_map_cb *cb)
10790{
10791 struct cpuacct *ca = cgroup_ca(cgrp);
10792 int i;
10793
10794 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10795 s64 val = percpu_counter_read(&ca->cpustat[i]);
10796 val = cputime64_to_clock_t(val);
10797 cb->fill(cb, cpuacct_stat_desc[i], val);
10798 }
10799 return 0;
10800}
10801
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010802static struct cftype files[] = {
10803 {
10804 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -070010805 .read_u64 = cpuusage_read,
10806 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010807 },
Ken Chene9515c32008-12-15 22:04:15 -080010808 {
10809 .name = "usage_percpu",
10810 .read_seq_string = cpuacct_percpu_seq_read,
10811 },
Bharata B Raoef12fef2009-03-31 10:02:22 +053010812 {
10813 .name = "stat",
10814 .read_map = cpuacct_stats_show,
10815 },
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010816};
10817
Dhaval Giani32cd7562008-02-29 10:02:43 +053010818static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010819{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010820 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010821}
10822
10823/*
10824 * charge this task's execution time to its accounting group.
10825 *
10826 * called with rq->lock held.
10827 */
10828static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10829{
10830 struct cpuacct *ca;
Bharata B Rao934352f2008-11-10 20:41:13 +053010831 int cpu;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010832
Li Zefanc40c6f82009-02-26 15:40:15 +080010833 if (unlikely(!cpuacct_subsys.active))
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010834 return;
10835
Bharata B Rao934352f2008-11-10 20:41:13 +053010836 cpu = task_cpu(tsk);
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010837
10838 rcu_read_lock();
10839
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010840 ca = task_ca(tsk);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010841
Bharata B Rao934352f2008-11-10 20:41:13 +053010842 for (; ca; ca = ca->parent) {
Rusty Russellb36128c2009-02-20 16:29:08 +090010843 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010844 *cpuusage += cputime;
10845 }
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010846
10847 rcu_read_unlock();
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010848}
10849
Bharata B Raoef12fef2009-03-31 10:02:22 +053010850/*
10851 * Charge the system/user time to the task's accounting group.
10852 */
10853static void cpuacct_update_stats(struct task_struct *tsk,
10854 enum cpuacct_stat_index idx, cputime_t val)
10855{
10856 struct cpuacct *ca;
10857
10858 if (unlikely(!cpuacct_subsys.active))
10859 return;
10860
10861 rcu_read_lock();
10862 ca = task_ca(tsk);
10863
10864 do {
10865 percpu_counter_add(&ca->cpustat[idx], val);
10866 ca = ca->parent;
10867 } while (ca);
10868 rcu_read_unlock();
10869}
10870
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010871struct cgroup_subsys cpuacct_subsys = {
10872 .name = "cpuacct",
10873 .create = cpuacct_create,
10874 .destroy = cpuacct_destroy,
10875 .populate = cpuacct_populate,
10876 .subsys_id = cpuacct_subsys_id,
10877};
10878#endif /* CONFIG_CGROUP_CPUACCT */
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010879
10880#ifndef CONFIG_SMP
10881
10882int rcu_expedited_torture_stats(char *page)
10883{
10884 return 0;
10885}
10886EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10887
10888void synchronize_sched_expedited(void)
10889{
10890}
10891EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10892
10893#else /* #ifndef CONFIG_SMP */
10894
10895static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10896static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10897
10898#define RCU_EXPEDITED_STATE_POST -2
10899#define RCU_EXPEDITED_STATE_IDLE -1
10900
10901static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10902
10903int rcu_expedited_torture_stats(char *page)
10904{
10905 int cnt = 0;
10906 int cpu;
10907
10908 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10909 for_each_online_cpu(cpu) {
10910 cnt += sprintf(&page[cnt], " %d:%d",
10911 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10912 }
10913 cnt += sprintf(&page[cnt], "\n");
10914 return cnt;
10915}
10916EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10917
10918static long synchronize_sched_expedited_count;
10919
10920/*
10921 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10922 * approach to force grace period to end quickly. This consumes
10923 * significant time on all CPUs, and is thus not recommended for
10924 * any sort of common-case code.
10925 *
10926 * Note that it is illegal to call this function while holding any
10927 * lock that is acquired by a CPU-hotplug notifier. Failing to
10928 * observe this restriction will result in deadlock.
10929 */
10930void synchronize_sched_expedited(void)
10931{
10932 int cpu;
10933 unsigned long flags;
10934 bool need_full_sync = 0;
10935 struct rq *rq;
10936 struct migration_req *req;
10937 long snap;
10938 int trycount = 0;
10939
10940 smp_mb(); /* ensure prior mod happens before capturing snap. */
10941 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10942 get_online_cpus();
10943 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10944 put_online_cpus();
10945 if (trycount++ < 10)
10946 udelay(trycount * num_online_cpus());
10947 else {
10948 synchronize_sched();
10949 return;
10950 }
10951 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10952 smp_mb(); /* ensure test happens before caller kfree */
10953 return;
10954 }
10955 get_online_cpus();
10956 }
10957 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10958 for_each_online_cpu(cpu) {
10959 rq = cpu_rq(cpu);
10960 req = &per_cpu(rcu_migration_req, cpu);
10961 init_completion(&req->done);
10962 req->task = NULL;
10963 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10964 spin_lock_irqsave(&rq->lock, flags);
10965 list_add(&req->list, &rq->migration_queue);
10966 spin_unlock_irqrestore(&rq->lock, flags);
10967 wake_up_process(rq->migration_thread);
10968 }
10969 for_each_online_cpu(cpu) {
10970 rcu_expedited_state = cpu;
10971 req = &per_cpu(rcu_migration_req, cpu);
10972 rq = cpu_rq(cpu);
10973 wait_for_completion(&req->done);
10974 spin_lock_irqsave(&rq->lock, flags);
10975 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10976 need_full_sync = 1;
10977 req->dest_cpu = RCU_MIGRATION_IDLE;
10978 spin_unlock_irqrestore(&rq->lock, flags);
10979 }
10980 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
Paul E. McKenney956539b2009-11-10 13:37:20 -080010981 synchronize_sched_expedited_count++;
Paul E. McKenney03b042b2009-06-25 09:08:16 -070010982 mutex_unlock(&rcu_sched_expedited_mutex);
10983 put_online_cpus();
10984 if (need_full_sync)
10985 synchronize_sched();
10986}
10987EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10988
10989#endif /* #else #ifndef CONFIG_SMP */