blob: 01f55ada3598a8e46dd25f7953660d69fc5258ee [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 Molnar0d905bc2009-05-04 19:13:30 +020042#include <linux/perf_counter.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>
Eric Dumazet5517d862007-05-08 00:32:57 -070067#include <linux/reciprocal_div.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020068#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020069#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010070#include <linux/hrtimer.h>
Reynes Philippe30914a52008-03-17 16:19:05 -070071#include <linux/tick.h>
Peter Zijlstraf00b45c2008-04-19 19:45:00 +020072#include <linux/debugfs.h>
73#include <linux/ctype.h>
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +020074#include <linux/ftrace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070075
Eric Dumazet5517d862007-05-08 00:32:57 -070076#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020077#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070078
Gregory Haskins6e0534f2008-05-12 21:21:01 +020079#include "sched_cpupri.h"
80
Steven Rostedta8d154b2009-04-10 09:36:00 -040081#define CREATE_TRACE_POINTS
Steven Rostedtad8d75f2009-04-14 19:39:12 -040082#include <trace/events/sched.h>
Steven Rostedta8d154b2009-04-10 09:36:00 -040083
Linus Torvalds1da177e2005-04-16 15:20:36 -070084/*
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 * and back.
88 */
89#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92
93/*
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
97 */
98#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101
102/*
Ingo Molnard7876a02008-01-25 21:08:19 +0100103 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100105#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700106
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200107#define NICE_0_LOAD SCHED_LOAD_SCALE
108#define NICE_0_SHIFT SCHED_LOAD_SHIFT
109
Linus Torvalds1da177e2005-04-16 15:20:36 -0700110/*
111 * These are the 'tuning knobs' of the scheduler:
112 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114 * Timeslices get refilled after they expire.
115 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700116#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700117
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200118/*
119 * single value that denotes runtime == period, ie unlimited time.
120 */
121#define RUNTIME_INF ((u64)~0ULL)
122
Eric Dumazet5517d862007-05-08 00:32:57 -0700123#ifdef CONFIG_SMP
Steven Noonanfd2ab302009-01-11 01:04:22 -0800124
125static void double_rq_lock(struct rq *rq1, struct rq *rq2);
126
Eric Dumazet5517d862007-05-08 00:32:57 -0700127/*
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
130 */
131static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
132{
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
134}
135
136/*
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
139 */
140static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
141{
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
144}
145#endif
146
Ingo Molnare05606d2007-07-09 18:51:59 +0200147static inline int rt_policy(int policy)
148{
Roel Kluin3f33a7c2008-05-13 23:44:11 +0200149 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
Ingo Molnare05606d2007-07-09 18:51:59 +0200150 return 1;
151 return 0;
152}
153
154static inline int task_has_rt_policy(struct task_struct *p)
155{
156 return rt_policy(p->policy);
157}
158
Linus Torvalds1da177e2005-04-16 15:20:36 -0700159/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200160 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200162struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
165};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700166
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200167struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
170 ktime_t rt_period;
171 u64 rt_runtime;
172 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200173};
174
175static struct rt_bandwidth def_rt_bandwidth;
176
177static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
178
179static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
180{
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
183 ktime_t now;
184 int overrun;
185 int idle = 0;
186
187 for (;;) {
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
190
191 if (!overrun)
192 break;
193
194 idle = do_sched_rt_period_timer(rt_b, overrun);
195 }
196
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
198}
199
200static
201void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
202{
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
205
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200206 spin_lock_init(&rt_b->rt_runtime_lock);
207
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200211}
212
Krzysztof Heltc8bfff62008-09-05 23:46:19 +0200213static inline int rt_bandwidth_enabled(void)
214{
215 return sysctl_sched_rt_runtime >= 0;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200216}
217
218static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
219{
220 ktime_t now;
221
Hiroshi Shimamotocac64d02009-02-25 09:59:26 -0800222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200223 return;
224
225 if (hrtimer_active(&rt_b->rt_period_timer))
226 return;
227
228 spin_lock(&rt_b->rt_runtime_lock);
229 for (;;) {
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100230 unsigned long delta;
231 ktime_t soft, hard;
232
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200233 if (hrtimer_active(&rt_b->rt_period_timer))
234 break;
235
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +0100238
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +0530243 HRTIMER_MODE_ABS_PINNED, 0);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200244 }
245 spin_unlock(&rt_b->rt_runtime_lock);
246}
247
248#ifdef CONFIG_RT_GROUP_SCHED
249static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
250{
251 hrtimer_cancel(&rt_b->rt_period_timer);
252}
253#endif
254
Heiko Carstens712555e2008-04-28 11:33:07 +0200255/*
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
258 */
259static DEFINE_MUTEX(sched_domains_mutex);
260
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100261#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200262
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700263#include <linux/cgroup.h>
264
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200265struct cfs_rq;
266
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100267static LIST_HEAD(task_groups);
268
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200269/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200270struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100271#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700272 struct cgroup_subsys_state css;
273#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100274
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530275#ifdef CONFIG_USER_SCHED
276 uid_t uid;
277#endif
278
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100279#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100285#endif
286
287#ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
290
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200291 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100292#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100293
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100294 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100295 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200296
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200300};
301
Dhaval Giani354d60c2008-04-19 19:44:59 +0200302#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200303
Arun R Bharadwaj6c415b92008-12-01 20:49:05 +0530304/* Helper function to pass uid information to create_sched_user() */
305void set_tg_uid(struct user_struct *user)
306{
307 user->tg->uid = user->uid;
308}
309
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200310/*
311 * Root task group.
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
314 */
315struct task_group root_task_group;
316
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100317#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200318/* Default task group's sched entity on each cpu */
319static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320/* Default task group's cfs_rq on each cpu */
321static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200322#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100323
324#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100325static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200327#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200328#else /* !CONFIG_USER_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200329#define root_task_group init_task_group
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200330#endif /* CONFIG_USER_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100331
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100332/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100333 * a task group's cpu shares.
334 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100335static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100336
Peter Zijlstra57310a92009-03-09 13:56:21 +0100337#ifdef CONFIG_SMP
338static int root_task_group_empty(void)
339{
340 return list_empty(&root_task_group.children);
341}
342#endif
343
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100344#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100345#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100346# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200347#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100348# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200349#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200350
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800351/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
358 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200359#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800360#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200361
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100362static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100363#endif
364
365/* Default task group.
366 * Every task in system belong to this group at bootup.
367 */
Mike Travis434d53b2008-04-04 18:11:04 -0700368struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200369
370/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200371static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200372{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200373 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200374
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100375#ifdef CONFIG_USER_SCHED
David Howellsc69e8d92008-11-14 10:39:19 +1100376 rcu_read_lock();
377 tg = __task_cred(p)->user->tg;
378 rcu_read_unlock();
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100379#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200382#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100383 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200384#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200385 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200386}
387
388/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100389static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200390{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100391#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100394#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100395
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100396#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100399#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200400}
401
402#else
403
Peter Zijlstra57310a92009-03-09 13:56:21 +0100404#ifdef CONFIG_SMP
405static int root_task_group_empty(void)
406{
407 return 1;
408}
409#endif
410
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100411static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Peter Zijlstra83378262008-06-27 13:41:37 +0200412static inline struct task_group *task_group(struct task_struct *p)
413{
414 return NULL;
415}
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200416
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100417#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200418
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200419/* CFS-related fields in a runqueue */
420struct cfs_rq {
421 struct load_weight load;
422 unsigned long nr_running;
423
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200424 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200425 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200426
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200429
430 struct list_head tasks;
431 struct list_head *balance_iterator;
432
433 /*
434 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200435 * It is set to NULL otherwise (i.e when none are currently running).
436 */
Peter Zijlstra47932412008-11-04 21:25:09 +0100437 struct sched_entity *curr, *next, *last;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200438
Peter Zijlstra5ac5c4d2008-11-10 10:46:32 +0100439 unsigned int nr_spread_over;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200440
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200441#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
443
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100444 /*
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
448 *
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
451 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200454
455#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200456 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200457 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200458 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200459 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200460
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200461 /*
462 * h_load = weight * f(tg)
463 *
464 * Where f(tg) is the recursive weight fraction assigned to
465 * this group.
466 */
467 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200468
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200469 /*
470 * this cpu's part of tg->shares
471 */
472 unsigned long shares;
Peter Zijlstraf1d239f2008-06-27 13:41:38 +0200473
474 /*
475 * load.weight at the time we set shares
476 */
477 unsigned long rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200478#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200479#endif
480};
481
482/* Real-Time classes' related field in a runqueue: */
483struct rt_rq {
484 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100485 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100486#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -0500487 struct {
488 int curr; /* highest queued rt task prio */
Gregory Haskins398a1532009-01-14 09:10:04 -0500489#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -0500490 int next; /* next highest */
Gregory Haskins398a1532009-01-14 09:10:04 -0500491#endif
Gregory Haskinse864c492008-12-29 09:39:49 -0500492 } highest_prio;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100493#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100494#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100495 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100496 int overloaded;
Gregory Haskins917b6272008-12-29 09:39:53 -0500497 struct plist_head pushable_tasks;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100498#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100499 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100500 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200501 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100502 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200503 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100504
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100505#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100506 unsigned long rt_nr_boosted;
507
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100508 struct rq *rq;
509 struct list_head leaf_rt_rq_list;
510 struct task_group *tg;
511 struct sched_rt_entity *rt_se;
512#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200513};
514
Gregory Haskins57d885f2008-01-25 21:08:18 +0100515#ifdef CONFIG_SMP
516
517/*
518 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100519 * variables. Each exclusive cpuset essentially defines an island domain by
520 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100521 * exclusive cpuset is created, we also create and attach a new root-domain
522 * object.
523 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100524 */
525struct root_domain {
526 atomic_t refcount;
Rusty Russellc6c49272008-11-25 02:35:05 +1030527 cpumask_var_t span;
528 cpumask_var_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100529
Ingo Molnar0eab9142008-01-25 21:08:19 +0100530 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100531 * The "RT overload" flag: it gets set if a CPU has more than
532 * one runnable RT task.
533 */
Rusty Russellc6c49272008-11-25 02:35:05 +1030534 cpumask_var_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100535 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200536#ifdef CONFIG_SMP
537 struct cpupri cpupri;
538#endif
Vaidyanathan Srinivasan7a09b1a2008-12-18 23:26:22 +0530539#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
540 /*
541 * Preferred wake up cpu nominated by sched_mc balance that will be
542 * used when most cpus are idle in the system indicating overall very
543 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
544 */
545 unsigned int sched_mc_preferred_wakeup_cpu;
546#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100547};
548
Gregory Haskinsdc938522008-01-25 21:08:26 +0100549/*
550 * By default the system creates a single root-domain with all cpus as
551 * members (mimicking the global state we have today).
552 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100553static struct root_domain def_root_domain;
554
555#endif
556
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200557/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700558 * This is the main, per-CPU runqueue data structure.
559 *
560 * Locking rule: those places that want to lock multiple runqueues
561 * (such as the load balancing or the thread migration code), lock
562 * acquire operations must be ordered by ascending &runqueue.
563 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700564struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200565 /* runqueue lock: */
566 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700567
568 /*
569 * nr_running and cpu_load should be in the same cacheline because
570 * remote CPUs use both these fields when doing load calculation.
571 */
572 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200573 #define CPU_LOAD_IDX_MAX 5
574 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700575#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200576 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700577 unsigned char in_nohz_recently;
578#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200579 /* capture load from *all* tasks on this cpu: */
580 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200581 unsigned long nr_load_updates;
582 u64 nr_switches;
Paul Mackerras23a185c2009-02-09 22:42:47 +1100583 u64 nr_migrations_in;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200584
585 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100586 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100587
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200588#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200589 /* list of leaf cfs_rq on this cpu: */
590 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100591#endif
592#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100593 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700594#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596 /*
597 * This is part of a global counter where only the total sum
598 * over all CPUs matters. A task can increase this counter on
599 * one CPU and if it got migrated afterwards it may decrease
600 * it on another CPU. Always updated under the runqueue lock:
601 */
602 unsigned long nr_uninterruptible;
603
Ingo Molnar36c8b582006-07-03 00:25:41 -0700604 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800605 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700606 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200607
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200608 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200609
Linus Torvalds1da177e2005-04-16 15:20:36 -0700610 atomic_t nr_iowait;
611
612#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100613 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700614 struct sched_domain *sd;
615
Henrik Austada0a522c2009-02-13 20:35:45 +0100616 unsigned char idle_at_tick;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700617 /* For active balancing */
618 int active_balance;
619 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200620 /* cpu of this runqueue: */
621 int cpu;
Gregory Haskins1f11eb62008-06-04 15:04:05 -0400622 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700623
Peter Zijlstraa8a51d52008-06-27 13:41:26 +0200624 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700625
Ingo Molnar36c8b582006-07-03 00:25:41 -0700626 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700627 struct list_head migration_queue;
628#endif
629
Thomas Gleixnerdce48a82009-04-11 10:43:41 +0200630 /* calc_load related fields */
631 unsigned long calc_load_update;
632 long calc_load_active;
633
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100634#ifdef CONFIG_SCHED_HRTICK
Peter Zijlstra31656512008-07-18 18:01:23 +0200635#ifdef CONFIG_SMP
636 int hrtick_csd_pending;
637 struct call_single_data hrtick_csd;
638#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100639 struct hrtimer hrtick_timer;
640#endif
641
Linus Torvalds1da177e2005-04-16 15:20:36 -0700642#ifdef CONFIG_SCHEDSTATS
643 /* latency stats */
644 struct sched_info rq_sched_info;
Ken Chen9c2c4802008-12-16 23:41:22 -0800645 unsigned long long rq_cpu_time;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700647
648 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200649 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700650
651 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200652 unsigned int sched_switch;
653 unsigned int sched_count;
654 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700655
656 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200657 unsigned int ttwu_count;
658 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200659
660 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200661 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700662#endif
663};
664
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700665static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700666
Peter Zijlstra15afe092008-09-20 23:38:02 +0200667static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
Ingo Molnardd41f592007-07-09 18:51:59 +0200668{
Peter Zijlstra15afe092008-09-20 23:38:02 +0200669 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
Ingo Molnardd41f592007-07-09 18:51:59 +0200670}
671
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700672static inline int cpu_of(struct rq *rq)
673{
674#ifdef CONFIG_SMP
675 return rq->cpu;
676#else
677 return 0;
678#endif
679}
680
Ingo Molnar20d315d2007-07-09 18:51:58 +0200681/*
Nick Piggin674311d2005-06-25 14:57:27 -0700682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700683 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700684 *
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
687 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700688#define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700690
691#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692#define this_rq() (&__get_cpu_var(runqueues))
693#define task_rq(p) cpu_rq(task_cpu(p))
694#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
695
Ingo Molnaraa9c4c02008-12-17 14:10:57 +0100696inline void update_rq_clock(struct rq *rq)
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200697{
698 rq->clock = sched_clock_cpu(cpu_of(rq));
699}
700
Ingo Molnare436d802007-07-19 21:28:35 +0200701/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200702 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
703 */
704#ifdef CONFIG_SCHED_DEBUG
705# define const_debug __read_mostly
706#else
707# define const_debug static const
708#endif
709
Ingo Molnar017730c2008-05-12 21:20:52 +0200710/**
711 * runqueue_is_locked
712 *
713 * Returns true if the current cpu runqueue is locked.
714 * This interface allows printk to be called with the runqueue lock
715 * held and know whether or not it is OK to wake up the klogd.
716 */
717int runqueue_is_locked(void)
718{
719 int cpu = get_cpu();
720 struct rq *rq = cpu_rq(cpu);
721 int ret;
722
723 ret = spin_is_locked(&rq->lock);
724 put_cpu();
725 return ret;
726}
727
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200728/*
729 * Debugging: various feature bits
730 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200731
732#define SCHED_FEAT(name, enabled) \
733 __SCHED_FEAT_##name ,
734
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200735enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200736#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200737};
738
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200739#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200740
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200741#define SCHED_FEAT(name, enabled) \
742 (1UL << __SCHED_FEAT_##name) * enabled |
743
744const_debug unsigned int sysctl_sched_features =
745#include "sched_features.h"
746 0;
747
748#undef SCHED_FEAT
749
750#ifdef CONFIG_SCHED_DEBUG
751#define SCHED_FEAT(name, enabled) \
752 #name ,
753
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700754static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200755#include "sched_features.h"
756 NULL
757};
758
759#undef SCHED_FEAT
760
Li Zefan34f3a812008-10-30 15:23:32 +0800761static int sched_feat_show(struct seq_file *m, void *v)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200762{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200763 int i;
764
765 for (i = 0; sched_feat_names[i]; i++) {
Li Zefan34f3a812008-10-30 15:23:32 +0800766 if (!(sysctl_sched_features & (1UL << i)))
767 seq_puts(m, "NO_");
768 seq_printf(m, "%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200769 }
Li Zefan34f3a812008-10-30 15:23:32 +0800770 seq_puts(m, "\n");
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200771
Li Zefan34f3a812008-10-30 15:23:32 +0800772 return 0;
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200773}
774
775static ssize_t
776sched_feat_write(struct file *filp, const char __user *ubuf,
777 size_t cnt, loff_t *ppos)
778{
779 char buf[64];
780 char *cmp = buf;
781 int neg = 0;
782 int i;
783
784 if (cnt > 63)
785 cnt = 63;
786
787 if (copy_from_user(&buf, ubuf, cnt))
788 return -EFAULT;
789
790 buf[cnt] = 0;
791
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200792 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200793 neg = 1;
794 cmp += 3;
795 }
796
797 for (i = 0; sched_feat_names[i]; i++) {
798 int len = strlen(sched_feat_names[i]);
799
800 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
801 if (neg)
802 sysctl_sched_features &= ~(1UL << i);
803 else
804 sysctl_sched_features |= (1UL << i);
805 break;
806 }
807 }
808
809 if (!sched_feat_names[i])
810 return -EINVAL;
811
812 filp->f_pos += cnt;
813
814 return cnt;
815}
816
Li Zefan34f3a812008-10-30 15:23:32 +0800817static int sched_feat_open(struct inode *inode, struct file *filp)
818{
819 return single_open(filp, sched_feat_show, NULL);
820}
821
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200822static struct file_operations sched_feat_fops = {
Li Zefan34f3a812008-10-30 15:23:32 +0800823 .open = sched_feat_open,
824 .write = sched_feat_write,
825 .read = seq_read,
826 .llseek = seq_lseek,
827 .release = single_release,
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200828};
829
830static __init int sched_init_debug(void)
831{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200832 debugfs_create_file("sched_features", 0644, NULL, NULL,
833 &sched_feat_fops);
834
835 return 0;
836}
837late_initcall(sched_init_debug);
838
839#endif
840
841#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200842
843/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100844 * Number of tasks to iterate in a single balance run.
845 * Limited because this is done with IRQs disabled.
846 */
847const_debug unsigned int sysctl_sched_nr_migrate = 32;
848
849/*
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200850 * ratelimit for updating the group shares.
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200851 * default: 0.25ms
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200852 */
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200853unsigned int sysctl_sched_shares_ratelimit = 250000;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200854
855/*
Peter Zijlstraffda12a2008-10-17 19:27:02 +0200856 * Inject some fuzzyness into changing the per-cpu group shares
857 * this avoids remote rq-locks at the expense of fairness.
858 * default: 4
859 */
860unsigned int sysctl_sched_shares_thresh = 4;
861
862/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100863 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100864 * default: 1s
865 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100866unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100867
Ingo Molnar6892b752008-02-13 14:02:36 +0100868static __read_mostly int scheduler_running;
869
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100870/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100871 * part of the period that we allow rt tasks to run in us.
872 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100873 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100874int sysctl_sched_rt_runtime = 950000;
875
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200876static inline u64 global_rt_period(void)
877{
878 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
879}
880
881static inline u64 global_rt_runtime(void)
882{
roel kluine26873b2008-07-22 16:51:15 -0400883 if (sysctl_sched_rt_runtime < 0)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200884 return RUNTIME_INF;
885
886 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
887}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100888
Linus Torvalds1da177e2005-04-16 15:20:36 -0700889#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700890# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700891#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700892#ifndef finish_arch_switch
893# define finish_arch_switch(prev) do { } while (0)
894#endif
895
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100896static inline int task_current(struct rq *rq, struct task_struct *p)
897{
898 return rq->curr == p;
899}
900
Nick Piggin4866cde2005-06-25 14:57:23 -0700901#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700902static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700903{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100904 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700905}
906
Ingo Molnar70b97a72006-07-03 00:25:42 -0700907static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700908{
909}
910
Ingo Molnar70b97a72006-07-03 00:25:42 -0700911static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700912{
Ingo Molnarda04c032005-09-13 11:17:59 +0200913#ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq->lock.owner = current;
916#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700917 /*
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
920 * prev into current:
921 */
922 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
923
Nick Piggin4866cde2005-06-25 14:57:23 -0700924 spin_unlock_irq(&rq->lock);
925}
926
927#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700928static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700929{
930#ifdef CONFIG_SMP
931 return p->oncpu;
932#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100933 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700934#endif
935}
936
Ingo Molnar70b97a72006-07-03 00:25:42 -0700937static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700938{
939#ifdef CONFIG_SMP
940 /*
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
943 * here.
944 */
945 next->oncpu = 1;
946#endif
947#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq->lock);
949#else
950 spin_unlock(&rq->lock);
951#endif
952}
953
Ingo Molnar70b97a72006-07-03 00:25:42 -0700954static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700955{
956#ifdef CONFIG_SMP
957 /*
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
960 * finished.
961 */
962 smp_wmb();
963 prev->oncpu = 0;
964#endif
965#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
966 local_irq_enable();
967#endif
968}
969#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970
971/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
974 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700975static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700976 __acquires(rq->lock)
977{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200978 for (;;) {
979 struct rq *rq = task_rq(p);
980 spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
982 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700983 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700984 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700985}
986
987/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100989 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700990 * explicitly disabling preemption.
991 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700992static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700993 __acquires(rq->lock)
994{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700995 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700996
Andi Kleen3a5c3592007-10-15 17:00:14 +0200997 for (;;) {
998 local_irq_save(*flags);
999 rq = task_rq(p);
1000 spin_lock(&rq->lock);
1001 if (likely(rq == task_rq(p)))
1002 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001003 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001004 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07001005}
1006
Oleg Nesterovad474ca2008-11-10 15:39:30 +01001007void task_rq_unlock_wait(struct task_struct *p)
1008{
1009 struct rq *rq = task_rq(p);
1010
1011 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1012 spin_unlock_wait(&rq->lock);
1013}
1014
Alexey Dobriyana9957442007-10-15 17:00:13 +02001015static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001016 __releases(rq->lock)
1017{
1018 spin_unlock(&rq->lock);
1019}
1020
Ingo Molnar70b97a72006-07-03 00:25:42 -07001021static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001022 __releases(rq->lock)
1023{
1024 spin_unlock_irqrestore(&rq->lock, *flags);
1025}
1026
Linus Torvalds1da177e2005-04-16 15:20:36 -07001027/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -08001028 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001029 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02001030static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001031 __acquires(rq->lock)
1032{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001033 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001034
1035 local_irq_disable();
1036 rq = this_rq();
1037 spin_lock(&rq->lock);
1038
1039 return rq;
1040}
1041
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001042#ifdef CONFIG_SCHED_HRTICK
1043/*
1044 * Use HR-timers to deliver accurate preemption points.
1045 *
1046 * Its all a bit involved since we cannot program an hrt while holding the
1047 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1048 * reschedule event.
1049 *
1050 * When we get rescheduled we reprogram the hrtick_timer outside of the
1051 * rq->lock.
1052 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001053
1054/*
1055 * Use hrtick when:
1056 * - enabled by features
1057 * - hrtimer is actually high res
1058 */
1059static inline int hrtick_enabled(struct rq *rq)
1060{
1061 if (!sched_feat(HRTICK))
1062 return 0;
Ingo Molnarba420592008-07-20 11:02:06 +02001063 if (!cpu_active(cpu_of(rq)))
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001064 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001065 return hrtimer_is_hres_active(&rq->hrtick_timer);
1066}
1067
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001068static void hrtick_clear(struct rq *rq)
1069{
1070 if (hrtimer_active(&rq->hrtick_timer))
1071 hrtimer_cancel(&rq->hrtick_timer);
1072}
1073
1074/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001075 * High-resolution timer tick.
1076 * Runs from hardirq context with interrupts disabled.
1077 */
1078static enum hrtimer_restart hrtick(struct hrtimer *timer)
1079{
1080 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1081
1082 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1083
1084 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001085 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001086 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1087 spin_unlock(&rq->lock);
1088
1089 return HRTIMER_NORESTART;
1090}
1091
Rabin Vincent95e904c2008-05-11 05:55:33 +05301092#ifdef CONFIG_SMP
Peter Zijlstra31656512008-07-18 18:01:23 +02001093/*
1094 * called from hardirq (IPI) context
1095 */
1096static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001097{
Peter Zijlstra31656512008-07-18 18:01:23 +02001098 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001099
Peter Zijlstra31656512008-07-18 18:01:23 +02001100 spin_lock(&rq->lock);
1101 hrtimer_restart(&rq->hrtick_timer);
1102 rq->hrtick_csd_pending = 0;
1103 spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001104}
1105
Peter Zijlstra31656512008-07-18 18:01:23 +02001106/*
1107 * Called to set the hrtick timer state.
1108 *
1109 * called with rq->lock held and irqs disabled
1110 */
1111static void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001112{
Peter Zijlstra31656512008-07-18 18:01:23 +02001113 struct hrtimer *timer = &rq->hrtick_timer;
1114 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001115
Arjan van de Vencc584b22008-09-01 15:02:30 -07001116 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +02001117
1118 if (rq == this_rq()) {
1119 hrtimer_restart(timer);
1120 } else if (!rq->hrtick_csd_pending) {
Peter Zijlstra6e275632009-02-25 13:59:48 +01001121 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001122 rq->hrtick_csd_pending = 1;
1123 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001124}
1125
1126static int
1127hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1128{
1129 int cpu = (int)(long)hcpu;
1130
1131 switch (action) {
1132 case CPU_UP_CANCELED:
1133 case CPU_UP_CANCELED_FROZEN:
1134 case CPU_DOWN_PREPARE:
1135 case CPU_DOWN_PREPARE_FROZEN:
1136 case CPU_DEAD:
1137 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +02001138 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001139 return NOTIFY_OK;
1140 }
1141
1142 return NOTIFY_DONE;
1143}
1144
Rakib Mullickfa748202008-09-22 14:55:45 -07001145static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001146{
1147 hotcpu_notifier(hotplug_hrtick, 0);
1148}
Peter Zijlstra31656512008-07-18 18:01:23 +02001149#else
1150/*
1151 * Called to set the hrtick timer state.
1152 *
1153 * called with rq->lock held and irqs disabled
1154 */
1155static void hrtick_start(struct rq *rq, u64 delay)
1156{
Peter Zijlstra7f1e2ca2009-03-13 12:21:27 +01001157 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
Arun R Bharadwaj5c333862009-04-16 12:14:37 +05301158 HRTIMER_MODE_REL_PINNED, 0);
Peter Zijlstra31656512008-07-18 18:01:23 +02001159}
1160
Andrew Morton006c75f2008-09-22 14:55:46 -07001161static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +02001162{
1163}
Rabin Vincent95e904c2008-05-11 05:55:33 +05301164#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001165
1166static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001167{
Peter Zijlstra31656512008-07-18 18:01:23 +02001168#ifdef CONFIG_SMP
1169 rq->hrtick_csd_pending = 0;
1170
1171 rq->hrtick_csd.flags = 0;
1172 rq->hrtick_csd.func = __hrtick_start;
1173 rq->hrtick_csd.info = rq;
1174#endif
1175
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001176 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1177 rq->hrtick_timer.function = hrtick;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001178}
Andrew Morton006c75f2008-09-22 14:55:46 -07001179#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001180static inline void hrtick_clear(struct rq *rq)
1181{
1182}
1183
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001184static inline void init_rq_hrtick(struct rq *rq)
1185{
1186}
1187
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001188static inline void init_hrtick(void)
1189{
1190}
Andrew Morton006c75f2008-09-22 14:55:46 -07001191#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001192
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001193/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001194 * resched_task - mark a task 'to be rescheduled now'.
1195 *
1196 * On UP this means the setting of the need_resched flag, on SMP it
1197 * might also involve a cross-CPU call to trigger the scheduler on
1198 * the target CPU.
1199 */
1200#ifdef CONFIG_SMP
1201
1202#ifndef tsk_is_polling
1203#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1204#endif
1205
Peter Zijlstra31656512008-07-18 18:01:23 +02001206static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001207{
1208 int cpu;
1209
1210 assert_spin_locked(&task_rq(p)->lock);
1211
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001212 if (test_tsk_need_resched(p))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001213 return;
1214
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001215 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001216
1217 cpu = task_cpu(p);
1218 if (cpu == smp_processor_id())
1219 return;
1220
1221 /* NEED_RESCHED must be visible before we test polling */
1222 smp_mb();
1223 if (!tsk_is_polling(p))
1224 smp_send_reschedule(cpu);
1225}
1226
1227static void resched_cpu(int cpu)
1228{
1229 struct rq *rq = cpu_rq(cpu);
1230 unsigned long flags;
1231
1232 if (!spin_trylock_irqsave(&rq->lock, flags))
1233 return;
1234 resched_task(cpu_curr(cpu));
1235 spin_unlock_irqrestore(&rq->lock, flags);
1236}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001237
1238#ifdef CONFIG_NO_HZ
1239/*
1240 * When add_timer_on() enqueues a timer into the timer wheel of an
1241 * idle CPU then this timer might expire before the next timer event
1242 * which is scheduled to wake up that CPU. In case of a completely
1243 * idle system the next event might even be infinite time into the
1244 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1245 * leaves the inner idle loop so the newly added timer is taken into
1246 * account when the CPU goes back to idle and evaluates the timer
1247 * wheel for the next timer event.
1248 */
1249void wake_up_idle_cpu(int cpu)
1250{
1251 struct rq *rq = cpu_rq(cpu);
1252
1253 if (cpu == smp_processor_id())
1254 return;
1255
1256 /*
1257 * This is safe, as this function is called with the timer
1258 * wheel base lock of (cpu) held. When the CPU is on the way
1259 * to idle and has not yet set rq->curr to idle then it will
1260 * be serialized on the timer wheel base lock and take the new
1261 * timer into account automatically.
1262 */
1263 if (rq->curr != rq->idle)
1264 return;
1265
1266 /*
1267 * We can set TIF_RESCHED on the idle task of the other CPU
1268 * lockless. The worst case is that the other CPU runs the
1269 * idle task through an additional NOOP schedule()
1270 */
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08001271 set_tsk_need_resched(rq->idle);
Thomas Gleixner06d83082008-03-22 09:20:24 +01001272
1273 /* NEED_RESCHED must be visible before we test polling */
1274 smp_mb();
1275 if (!tsk_is_polling(rq->idle))
1276 smp_send_reschedule(cpu);
1277}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001278#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001279
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001280#else /* !CONFIG_SMP */
Peter Zijlstra31656512008-07-18 18:01:23 +02001281static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001282{
1283 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001284 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001285}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001286#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001287
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001288#if BITS_PER_LONG == 32
1289# define WMULT_CONST (~0UL)
1290#else
1291# define WMULT_CONST (1UL << 32)
1292#endif
1293
1294#define WMULT_SHIFT 32
1295
Ingo Molnar194081e2007-08-09 11:16:51 +02001296/*
1297 * Shift right and round:
1298 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001299#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001300
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001301/*
1302 * delta *= weight / lw
1303 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001304static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001305calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1306 struct load_weight *lw)
1307{
1308 u64 tmp;
1309
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001310 if (!lw->inv_weight) {
1311 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1312 lw->inv_weight = 1;
1313 else
1314 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1315 / (lw->weight+1);
1316 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001317
1318 tmp = (u64)delta_exec * weight;
1319 /*
1320 * Check whether we'd overflow the 64-bit multiplication:
1321 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001322 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001323 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001324 WMULT_SHIFT/2);
1325 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001326 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001327
Ingo Molnarecf691d2007-08-02 17:41:40 +02001328 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001329}
1330
Ingo Molnar10919852007-10-15 17:00:04 +02001331static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001332{
1333 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001334 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001335}
1336
Ingo Molnar10919852007-10-15 17:00:04 +02001337static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001338{
1339 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001340 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001341}
1342
Linus Torvalds1da177e2005-04-16 15:20:36 -07001343/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001348 * scaled version of the new time slice allocation that they receive on time
1349 * slice expiry etc.
1350 */
1351
Peter Zijlstracce7ade2009-01-15 14:53:37 +01001352#define WEIGHT_IDLEPRIO 3
1353#define WMULT_IDLEPRIO 1431655765
Ingo Molnardd41f592007-07-09 18:51:59 +02001354
1355/*
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1360 *
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001366 */
1367static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001376};
1377
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001378/*
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 *
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1384 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001385static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001394};
Peter Williams2dd73a42006-06-27 02:54:34 -07001395
Ingo Molnardd41f592007-07-09 18:51:59 +02001396static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1397
1398/*
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1402 */
1403struct rq_iterator {
1404 void *arg;
1405 struct task_struct *(*start)(void *);
1406 struct task_struct *(*next)(void *);
1407};
1408
Peter Williamse1d14842007-10-24 18:23:51 +02001409#ifdef CONFIG_SMP
1410static unsigned long
1411balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1412 unsigned long max_load_move, struct sched_domain *sd,
1413 enum cpu_idle_type idle, int *all_pinned,
1414 int *this_best_prio, struct rq_iterator *iterator);
1415
1416static int
1417iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1418 struct sched_domain *sd, enum cpu_idle_type idle,
1419 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001420#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001421
Bharata B Raoef12fef2009-03-31 10:02:22 +05301422/* Time spent by the tasks of the cpu accounting group executing in ... */
1423enum cpuacct_stat_index {
1424 CPUACCT_STAT_USER, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1426
1427 CPUACCT_STAT_NSTATS,
1428};
1429
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001430#ifdef CONFIG_CGROUP_CPUACCT
1431static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
Bharata B Raoef12fef2009-03-31 10:02:22 +05301432static 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#else
1435static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
Bharata B Raoef12fef2009-03-31 10:02:22 +05301436static inline void cpuacct_update_stats(struct task_struct *tsk,
1437 enum cpuacct_stat_index idx, cputime_t val) {}
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001438#endif
1439
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001440static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1441{
1442 update_load_add(&rq->load, load);
1443}
1444
1445static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1446{
1447 update_load_sub(&rq->load, load);
1448}
1449
Ingo Molnar7940ca32008-08-19 13:40:47 +02001450#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001451typedef int (*tg_visitor)(struct task_group *, void *);
1452
1453/*
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1456 */
1457static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1458{
1459 struct task_group *parent, *child;
1460 int ret;
1461
1462 rcu_read_lock();
1463 parent = &root_task_group;
1464down:
1465 ret = (*down)(parent, data);
1466 if (ret)
1467 goto out_unlock;
1468 list_for_each_entry_rcu(child, &parent->children, siblings) {
1469 parent = child;
1470 goto down;
1471
1472up:
1473 continue;
1474 }
1475 ret = (*up)(parent, data);
1476 if (ret)
1477 goto out_unlock;
1478
1479 child = parent;
1480 parent = parent->parent;
1481 if (parent)
1482 goto up;
1483out_unlock:
1484 rcu_read_unlock();
1485
1486 return ret;
1487}
1488
1489static int tg_nop(struct task_group *tg, void *data)
1490{
1491 return 0;
1492}
1493#endif
1494
Gregory Haskinse7693a32008-01-25 21:08:09 +01001495#ifdef CONFIG_SMP
1496static unsigned long source_load(int cpu, int type);
1497static unsigned long target_load(int cpu, int type);
Gregory Haskinse7693a32008-01-25 21:08:09 +01001498static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001499
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001500static unsigned long cpu_avg_load_per_task(int cpu)
1501{
1502 struct rq *rq = cpu_rq(cpu);
Ingo Molnaraf6d5962008-11-29 20:45:15 +01001503 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001504
Steven Rostedt4cd42622008-11-26 21:04:24 -05001505 if (nr_running)
1506 rq->avg_load_per_task = rq->load.weight / nr_running;
Balbir Singha2d47772008-11-12 16:19:00 +05301507 else
1508 rq->avg_load_per_task = 0;
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001509
1510 return rq->avg_load_per_task;
1511}
1512
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001513#ifdef CONFIG_FAIR_GROUP_SCHED
1514
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001515static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1516
1517/*
1518 * Calculate and set the cpu's group shares.
1519 */
1520static void
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001521update_group_shares_cpu(struct task_group *tg, int cpu,
1522 unsigned long sd_shares, unsigned long sd_rq_weight)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001523{
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001524 unsigned long shares;
1525 unsigned long rq_weight;
1526
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001527 if (!tg->se[cpu])
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001528 return;
1529
Ken Chenec4e0e22008-11-18 22:41:57 -08001530 rq_weight = tg->cfs_rq[cpu]->rq_weight;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001531
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001532 /*
1533 * \Sum shares * rq_weight
1534 * shares = -----------------------
1535 * \Sum rq_weight
1536 *
1537 */
Ken Chenec4e0e22008-11-18 22:41:57 -08001538 shares = (sd_shares * rq_weight) / sd_rq_weight;
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001539 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001540
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001541 if (abs(shares - tg->se[cpu]->load.weight) >
1542 sysctl_sched_shares_thresh) {
1543 struct rq *rq = cpu_rq(cpu);
1544 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001545
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001546 spin_lock_irqsave(&rq->lock, flags);
Ken Chenec4e0e22008-11-18 22:41:57 -08001547 tg->cfs_rq[cpu]->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001548
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001549 __set_se_shares(tg->se[cpu], shares);
1550 spin_unlock_irqrestore(&rq->lock, flags);
1551 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001552}
1553
1554/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001555 * Re-compute the task group their per cpu shares over the given domain.
1556 * This needs to be done in a bottom-up fashion because the rq weight of a
1557 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001558 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001559static int tg_shares_up(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001560{
Ken Chenec4e0e22008-11-18 22:41:57 -08001561 unsigned long weight, rq_weight = 0;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001562 unsigned long shares = 0;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001563 struct sched_domain *sd = data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001564 int i;
1565
Rusty Russell758b2cd2008-11-25 02:35:04 +10301566 for_each_cpu(i, sched_domain_span(sd)) {
Ken Chenec4e0e22008-11-18 22:41:57 -08001567 /*
1568 * If there are currently no tasks on the cpu pretend there
1569 * is one of average load so that when a new task gets to
1570 * run here it will not get delayed by group starvation.
1571 */
1572 weight = tg->cfs_rq[i]->load.weight;
1573 if (!weight)
1574 weight = NICE_0_LOAD;
1575
1576 tg->cfs_rq[i]->rq_weight = weight;
1577 rq_weight += weight;
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001578 shares += tg->cfs_rq[i]->shares;
1579 }
1580
1581 if ((!shares && rq_weight) || shares > tg->shares)
1582 shares = tg->shares;
1583
1584 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1585 shares = tg->shares;
1586
Rusty Russell758b2cd2008-11-25 02:35:04 +10301587 for_each_cpu(i, sched_domain_span(sd))
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001588 update_group_shares_cpu(tg, i, shares, rq_weight);
Peter Zijlstraeb755802008-08-19 12:33:05 +02001589
1590 return 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001591}
1592
1593/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001594 * Compute the cpu's hierarchical load factor for each task group.
1595 * This needs to be done in a top-down fashion because the load of a child
1596 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001597 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001598static int tg_load_down(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001599{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001600 unsigned long load;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001601 long cpu = (long)data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001602
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001603 if (!tg->parent) {
1604 load = cpu_rq(cpu)->load.weight;
1605 } else {
1606 load = tg->parent->cfs_rq[cpu]->h_load;
1607 load *= tg->cfs_rq[cpu]->shares;
1608 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1609 }
1610
1611 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001612
Peter Zijlstraeb755802008-08-19 12:33:05 +02001613 return 0;
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001614}
1615
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001616static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001617{
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001618 u64 now = cpu_clock(raw_smp_processor_id());
1619 s64 elapsed = now - sd->last_update;
1620
1621 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1622 sd->last_update = now;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001623 walk_tg_tree(tg_nop, tg_shares_up, sd);
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001624 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001625}
1626
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001627static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1628{
1629 spin_unlock(&rq->lock);
1630 update_shares(sd);
1631 spin_lock(&rq->lock);
1632}
1633
Peter Zijlstraeb755802008-08-19 12:33:05 +02001634static void update_h_load(long cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001635{
Peter Zijlstraeb755802008-08-19 12:33:05 +02001636 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001637}
1638
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001639#else
1640
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001641static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001642{
1643}
1644
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001645static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1646{
1647}
1648
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001649#endif
1650
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001651#ifdef CONFIG_PREEMPT
1652
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001653/*
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001654 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1655 * way at the expense of forcing extra atomic operations in all
1656 * invocations. This assures that the double_lock is acquired using the
1657 * same underlying policy as the spinlock_t on this architecture, which
1658 * reduces latency compared to the unfair variant below. However, it
1659 * also adds more overhead and therefore may reduce throughput.
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001660 */
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001661static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1662 __releases(this_rq->lock)
1663 __acquires(busiest->lock)
1664 __acquires(this_rq->lock)
1665{
1666 spin_unlock(&this_rq->lock);
1667 double_rq_lock(this_rq, busiest);
1668
1669 return 1;
1670}
1671
1672#else
1673/*
1674 * Unfair double_lock_balance: Optimizes throughput at the expense of
1675 * latency by eliminating extra atomic operations when the locks are
1676 * already in proper order on entry. This favors lower cpu-ids and will
1677 * grant the double lock to lower cpus over higher ids under contention,
1678 * regardless of entry order into the function.
1679 */
1680static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001681 __releases(this_rq->lock)
1682 __acquires(busiest->lock)
1683 __acquires(this_rq->lock)
1684{
1685 int ret = 0;
1686
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001687 if (unlikely(!spin_trylock(&busiest->lock))) {
1688 if (busiest < this_rq) {
1689 spin_unlock(&this_rq->lock);
1690 spin_lock(&busiest->lock);
1691 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1692 ret = 1;
1693 } else
1694 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1695 }
1696 return ret;
1697}
1698
Gregory Haskins8f45e2b2008-12-29 09:39:51 -05001699#endif /* CONFIG_PREEMPT */
1700
1701/*
1702 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1703 */
1704static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1705{
1706 if (unlikely(!irqs_disabled())) {
1707 /* printk() doesn't work good under rq->lock */
1708 spin_unlock(&this_rq->lock);
1709 BUG_ON(1);
1710 }
1711
1712 return _double_lock_balance(this_rq, busiest);
1713}
1714
Alexey Dobriyan70574a92008-11-28 22:08:00 +03001715static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1716 __releases(busiest->lock)
1717{
1718 spin_unlock(&busiest->lock);
1719 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1720}
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001721#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001722
1723#ifdef CONFIG_FAIR_GROUP_SCHED
1724static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1725{
Vegard Nossum30432092008-06-27 21:35:50 +02001726#ifdef CONFIG_SMP
Ingo Molnar34e83e82008-06-27 15:42:36 +02001727 cfs_rq->shares = shares;
1728#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001729}
1730#endif
1731
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02001732static void calc_load_account_active(struct rq *this_rq);
1733
Ingo Molnardd41f592007-07-09 18:51:59 +02001734#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001735#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001736#include "sched_fair.c"
1737#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001738#ifdef CONFIG_SCHED_DEBUG
1739# include "sched_debug.c"
1740#endif
1741
1742#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb62008-06-04 15:04:05 -04001743#define for_each_class(class) \
1744 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001745
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001746static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001747{
1748 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001749}
1750
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001751static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001752{
1753 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001754}
1755
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001756static void set_load_weight(struct task_struct *p)
1757{
1758 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001759 p->se.load.weight = prio_to_weight[0] * 2;
1760 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1761 return;
1762 }
1763
1764 /*
1765 * SCHED_IDLE tasks get minimal weight:
1766 */
1767 if (p->policy == SCHED_IDLE) {
1768 p->se.load.weight = WEIGHT_IDLEPRIO;
1769 p->se.load.inv_weight = WMULT_IDLEPRIO;
1770 return;
1771 }
1772
1773 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1774 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001775}
1776
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001777static void update_avg(u64 *avg, u64 sample)
1778{
1779 s64 diff = sample - *avg;
1780 *avg += diff >> 3;
1781}
1782
Ingo Molnar8159f872007-08-09 11:16:49 +02001783static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001784{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001785 if (wakeup)
1786 p->se.start_runtime = p->se.sum_exec_runtime;
1787
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001788 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001789 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001790 p->se.on_rq = 1;
1791}
1792
Ingo Molnar69be72c2007-08-09 11:16:49 +02001793static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001794{
Peter Zijlstra831451a2009-01-14 12:39:18 +01001795 if (sleep) {
1796 if (p->se.last_wakeup) {
1797 update_avg(&p->se.avg_overlap,
1798 p->se.sum_exec_runtime - p->se.last_wakeup);
1799 p->se.last_wakeup = 0;
1800 } else {
1801 update_avg(&p->se.avg_wakeup,
1802 sysctl_sched_wakeup_granularity);
1803 }
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001804 }
1805
Ankita Garg46ac22b2008-07-01 14:30:06 +05301806 sched_info_dequeued(p);
Ingo Molnarf02231e2007-08-09 11:16:48 +02001807 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001808 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001809}
1810
1811/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001812 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001813 */
Ingo Molnar14531182007-07-09 18:51:59 +02001814static inline int __normal_prio(struct task_struct *p)
1815{
Ingo Molnardd41f592007-07-09 18:51:59 +02001816 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001817}
1818
1819/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001820 * Calculate the expected normal priority: i.e. priority
1821 * without taking RT-inheritance into account. Might be
1822 * boosted by interactivity modifiers. Changes upon fork,
1823 * setprio syscalls, and whenever the interactivity
1824 * estimator recalculates.
1825 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001826static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001827{
1828 int prio;
1829
Ingo Molnare05606d2007-07-09 18:51:59 +02001830 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001831 prio = MAX_RT_PRIO-1 - p->rt_priority;
1832 else
1833 prio = __normal_prio(p);
1834 return prio;
1835}
1836
1837/*
1838 * Calculate the current priority, i.e. the priority
1839 * taken into account by the scheduler. This value might
1840 * be boosted by RT tasks, or might be boosted by
1841 * interactivity modifiers. Will be RT if the task got
1842 * RT-boosted. If not then it returns p->normal_prio.
1843 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001844static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001845{
1846 p->normal_prio = normal_prio(p);
1847 /*
1848 * If we are RT tasks or we were boosted to RT priority,
1849 * keep the priority unchanged. Otherwise, update priority
1850 * to the normal priority:
1851 */
1852 if (!rt_prio(p->prio))
1853 return p->normal_prio;
1854 return p->prio;
1855}
1856
1857/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001858 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001859 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001860static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001861{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001862 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001863 rq->nr_uninterruptible--;
1864
Ingo Molnar8159f872007-08-09 11:16:49 +02001865 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001866 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001867}
1868
1869/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001870 * deactivate_task - remove a task from the runqueue.
1871 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001872static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001873{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001874 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001875 rq->nr_uninterruptible++;
1876
Ingo Molnar69be72c2007-08-09 11:16:49 +02001877 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001878 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001879}
1880
Linus Torvalds1da177e2005-04-16 15:20:36 -07001881/**
1882 * task_curr - is this task currently executing on a CPU?
1883 * @p: the task in question.
1884 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001885inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886{
1887 return cpu_curr(task_cpu(p)) == p;
1888}
1889
Ingo Molnardd41f592007-07-09 18:51:59 +02001890static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1891{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001892 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001893#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001894 /*
1895 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1896 * successfuly executed on another CPU. We must ensure that updates of
1897 * per-task data have been completed by this moment.
1898 */
1899 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001900 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001901#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001902}
1903
Steven Rostedtcb469842008-01-25 21:08:22 +01001904static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1905 const struct sched_class *prev_class,
1906 int oldprio, int running)
1907{
1908 if (prev_class != p->sched_class) {
1909 if (prev_class->switched_from)
1910 prev_class->switched_from(rq, p, running);
1911 p->sched_class->switched_to(rq, p, running);
1912 } else
1913 p->sched_class->prio_changed(rq, p, oldprio, running);
1914}
1915
Linus Torvalds1da177e2005-04-16 15:20:36 -07001916#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001917
Thomas Gleixnere958b362008-06-04 23:22:32 +02001918/* Used instead of source_load when we know the type == 0 */
1919static unsigned long weighted_cpuload(const int cpu)
1920{
1921 return cpu_rq(cpu)->load.weight;
1922}
1923
Ingo Molnarcc367732007-10-15 17:00:18 +02001924/*
1925 * Is this task likely cache-hot:
1926 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001927static int
Ingo Molnarcc367732007-10-15 17:00:18 +02001928task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1929{
1930 s64 delta;
1931
Ingo Molnarf540a602008-03-15 17:10:34 +01001932 /*
1933 * Buddy candidates are cache hot:
1934 */
Peter Zijlstra47932412008-11-04 21:25:09 +01001935 if (sched_feat(CACHE_HOT_BUDDY) &&
1936 (&p->se == cfs_rq_of(&p->se)->next ||
1937 &p->se == cfs_rq_of(&p->se)->last))
Ingo Molnarf540a602008-03-15 17:10:34 +01001938 return 1;
1939
Ingo Molnarcc367732007-10-15 17:00:18 +02001940 if (p->sched_class != &fair_sched_class)
1941 return 0;
1942
Ingo Molnar6bc16652007-10-15 17:00:18 +02001943 if (sysctl_sched_migration_cost == -1)
1944 return 1;
1945 if (sysctl_sched_migration_cost == 0)
1946 return 0;
1947
Ingo Molnarcc367732007-10-15 17:00:18 +02001948 delta = now - p->se.exec_start;
1949
1950 return delta < (s64)sysctl_sched_migration_cost;
1951}
1952
1953
Ingo Molnardd41f592007-07-09 18:51:59 +02001954void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001955{
Ingo Molnardd41f592007-07-09 18:51:59 +02001956 int old_cpu = task_cpu(p);
1957 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001958 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1959 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001960 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001961
1962 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001963
Mathieu Desnoyersde1d7282009-05-05 16:49:59 +08001964 trace_sched_migrate_task(p, new_cpu);
Peter Zijlstracbc34ed2008-12-10 08:08:22 +01001965
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001966#ifdef CONFIG_SCHEDSTATS
1967 if (p->se.wait_start)
1968 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001969 if (p->se.sleep_start)
1970 p->se.sleep_start -= clock_offset;
1971 if (p->se.block_start)
1972 p->se.block_start -= clock_offset;
Ingo Molnar6c594c22008-12-14 12:34:15 +01001973#endif
Ingo Molnarcc367732007-10-15 17:00:18 +02001974 if (old_cpu != new_cpu) {
Ingo Molnar6c594c22008-12-14 12:34:15 +01001975 p->se.nr_migrations++;
Paul Mackerras23a185c2009-02-09 22:42:47 +11001976 new_rq->nr_migrations_in++;
Ingo Molnar6c594c22008-12-14 12:34:15 +01001977#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02001978 if (task_hot(p, old_rq->clock, NULL))
1979 schedstat_inc(p, se.nr_forced2_migrations);
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001980#endif
Peter Zijlstrae5289d42009-06-19 13:22:51 +02001981 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
1982 1, 1, NULL, 0);
Ingo Molnar6c594c22008-12-14 12:34:15 +01001983 }
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001984 p->se.vruntime -= old_cfsrq->min_vruntime -
1985 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001986
1987 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001988}
1989
Ingo Molnar70b97a72006-07-03 00:25:42 -07001990struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001991 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992
Ingo Molnar36c8b582006-07-03 00:25:41 -07001993 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001994 int dest_cpu;
1995
Linus Torvalds1da177e2005-04-16 15:20:36 -07001996 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001997};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001998
1999/*
2000 * The task's runqueue lock must be held.
2001 * Returns true if you have to wait for migration thread.
2002 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002003static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07002004migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002005{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002006 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002007
2008 /*
2009 * If the task is not on a runqueue (and not running), then
2010 * it is sufficient to simply update the task's cpu field.
2011 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002012 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002013 set_task_cpu(p, dest_cpu);
2014 return 0;
2015 }
2016
2017 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002018 req->task = p;
2019 req->dest_cpu = dest_cpu;
2020 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07002021
Linus Torvalds1da177e2005-04-16 15:20:36 -07002022 return 1;
2023}
2024
2025/*
Markus Metzgera26b89f2009-04-03 16:43:34 +02002026 * wait_task_context_switch - wait for a thread to complete at least one
2027 * context switch.
2028 *
2029 * @p must not be current.
2030 */
2031void wait_task_context_switch(struct task_struct *p)
2032{
2033 unsigned long nvcsw, nivcsw, flags;
2034 int running;
2035 struct rq *rq;
2036
2037 nvcsw = p->nvcsw;
2038 nivcsw = p->nivcsw;
2039 for (;;) {
2040 /*
2041 * The runqueue is assigned before the actual context
2042 * switch. We need to take the runqueue lock.
2043 *
2044 * We could check initially without the lock but it is
2045 * very likely that we need to take the lock in every
2046 * iteration.
2047 */
2048 rq = task_rq_lock(p, &flags);
2049 running = task_running(rq, p);
2050 task_rq_unlock(rq, &flags);
2051
2052 if (likely(!running))
2053 break;
2054 /*
2055 * The switch count is incremented before the actual
2056 * context switch. We thus wait for two switches to be
2057 * sure at least one completed.
2058 */
2059 if ((p->nvcsw - nvcsw) > 1)
2060 break;
2061 if ((p->nivcsw - nivcsw) > 1)
2062 break;
2063
2064 cpu_relax();
2065 }
2066}
2067
2068/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002069 * wait_task_inactive - wait for a thread to unschedule.
2070 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07002071 * If @match_state is nonzero, it's the @p->state value just checked and
2072 * not expected to change. If it changes, i.e. @p might have woken up,
2073 * then return zero. When we succeed in waiting for @p to be off its CPU,
2074 * we return a positive number (its total switch count). If a second call
2075 * a short while later returns the same number, the caller can be sure that
2076 * @p has remained unscheduled the whole time.
2077 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07002078 * The caller must ensure that the task *will* unschedule sometime soon,
2079 * else this function might spin for a *long* time. This function can't
2080 * be called with interrupts off, or it may introduce deadlock with
2081 * smp_call_function() if an IPI is sent by the same process we are
2082 * waiting to become inactive.
2083 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002084unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002085{
2086 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002087 int running, on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002088 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002089 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002090
Andi Kleen3a5c3592007-10-15 17:00:14 +02002091 for (;;) {
2092 /*
2093 * We do the initial early heuristics without holding
2094 * any task-queue locks at all. We'll only try to get
2095 * the runqueue lock when things look like they will
2096 * work out!
2097 */
2098 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002099
Andi Kleen3a5c3592007-10-15 17:00:14 +02002100 /*
2101 * If the task is actively running on another CPU
2102 * still, just relax and busy-wait without holding
2103 * any locks.
2104 *
2105 * NOTE! Since we don't hold any locks, it's not
2106 * even sure that "rq" stays as the right runqueue!
2107 * But we don't care, since "task_running()" will
2108 * return false if the runqueue has changed and p
2109 * is actually now running somewhere else!
2110 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07002111 while (task_running(rq, p)) {
2112 if (match_state && unlikely(p->state != match_state))
2113 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02002114 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07002115 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002116
Andi Kleen3a5c3592007-10-15 17:00:14 +02002117 /*
2118 * Ok, time to look more closely! We need the rq
2119 * lock now, to be *sure*. If we're wrong, we'll
2120 * just go back and repeat.
2121 */
2122 rq = task_rq_lock(p, &flags);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002123 trace_sched_wait_task(rq, p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02002124 running = task_running(rq, p);
2125 on_rq = p->se.on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07002126 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07002127 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07002128 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Andi Kleen3a5c3592007-10-15 17:00:14 +02002129 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07002130
Andi Kleen3a5c3592007-10-15 17:00:14 +02002131 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07002132 * If it changed from the expected state, bail out now.
2133 */
2134 if (unlikely(!ncsw))
2135 break;
2136
2137 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02002138 * Was it really running after all now that we
2139 * checked with the proper locks actually held?
2140 *
2141 * Oops. Go back and try again..
2142 */
2143 if (unlikely(running)) {
2144 cpu_relax();
2145 continue;
2146 }
2147
2148 /*
2149 * It's not enough that it's not actively running,
2150 * it must be off the runqueue _entirely_, and not
2151 * preempted!
2152 *
Luis Henriques80dd99b2009-03-16 19:58:09 +00002153 * So if it was still runnable (but just not actively
Andi Kleen3a5c3592007-10-15 17:00:14 +02002154 * running right now), it's preempted, and we should
2155 * yield - it could be a while.
2156 */
2157 if (unlikely(on_rq)) {
2158 schedule_timeout_uninterruptible(1);
2159 continue;
2160 }
2161
2162 /*
2163 * Ahh, all good. It wasn't running, and it wasn't
2164 * runnable, which means that it will never become
2165 * running in the future either. We're all done!
2166 */
2167 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002168 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07002169
2170 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002171}
2172
2173/***
2174 * kick_process - kick a running thread to enter/exit the kernel
2175 * @p: the to-be-kicked thread
2176 *
2177 * Cause a process which is running on another CPU to enter
2178 * kernel-mode, without any delay. (to get signals handled.)
2179 *
2180 * NOTE: this function doesnt have to take the runqueue lock,
2181 * because all it wants to ensure is that the remote task enters
2182 * the kernel. If the IPI races and the task has been migrated
2183 * to another CPU then no harm is done and the purpose has been
2184 * achieved as well.
2185 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002186void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002187{
2188 int cpu;
2189
2190 preempt_disable();
2191 cpu = task_cpu(p);
2192 if ((cpu != smp_processor_id()) && task_curr(p))
2193 smp_send_reschedule(cpu);
2194 preempt_enable();
2195}
Rusty Russellb43e3522009-06-12 22:27:00 -06002196EXPORT_SYMBOL_GPL(kick_process);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002197
2198/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002199 * Return a low guess at the load of a migration-source cpu weighted
2200 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201 *
2202 * We want to under-estimate the load of migration sources, to
2203 * balance conservatively.
2204 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002205static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002206{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002207 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002208 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002209
Peter Zijlstra93b75212008-06-27 13:41:33 +02002210 if (type == 0 || !sched_feat(LB_BIAS))
Ingo Molnardd41f592007-07-09 18:51:59 +02002211 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002212
Ingo Molnardd41f592007-07-09 18:51:59 +02002213 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002214}
2215
2216/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002217 * Return a high guess at the load of a migration-target cpu weighted
2218 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002220static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002221{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002222 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002223 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002224
Peter Zijlstra93b75212008-06-27 13:41:33 +02002225 if (type == 0 || !sched_feat(LB_BIAS))
Ingo Molnardd41f592007-07-09 18:51:59 +02002226 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002227
Ingo Molnardd41f592007-07-09 18:51:59 +02002228 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07002229}
2230
2231/*
Nick Piggin147cbb42005-06-25 14:57:19 -07002232 * find_idlest_group finds and returns the least busy CPU group within the
2233 * domain.
2234 */
2235static struct sched_group *
2236find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2237{
2238 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2239 unsigned long min_load = ULONG_MAX, this_load = 0;
2240 int load_idx = sd->forkexec_idx;
2241 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2242
2243 do {
2244 unsigned long load, avg_load;
2245 int local_group;
2246 int i;
2247
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002248 /* Skip over this group if it has no CPUs allowed */
Rusty Russell758b2cd2008-11-25 02:35:04 +10302249 if (!cpumask_intersects(sched_group_cpus(group),
2250 &p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02002251 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002252
Rusty Russell758b2cd2008-11-25 02:35:04 +10302253 local_group = cpumask_test_cpu(this_cpu,
2254 sched_group_cpus(group));
Nick Piggin147cbb42005-06-25 14:57:19 -07002255
2256 /* Tally up the load of all CPUs in the group */
2257 avg_load = 0;
2258
Rusty Russell758b2cd2008-11-25 02:35:04 +10302259 for_each_cpu(i, sched_group_cpus(group)) {
Nick Piggin147cbb42005-06-25 14:57:19 -07002260 /* Bias balancing toward cpus of our domain */
2261 if (local_group)
2262 load = source_load(i, load_idx);
2263 else
2264 load = target_load(i, load_idx);
2265
2266 avg_load += load;
2267 }
2268
2269 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002270 avg_load = sg_div_cpu_power(group,
2271 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07002272
2273 if (local_group) {
2274 this_load = avg_load;
2275 this = group;
2276 } else if (avg_load < min_load) {
2277 min_load = avg_load;
2278 idlest = group;
2279 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02002280 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07002281
2282 if (!idlest || 100*this_load < imbalance*min_load)
2283 return NULL;
2284 return idlest;
2285}
2286
2287/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07002288 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07002289 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002290static int
Rusty Russell758b2cd2008-11-25 02:35:04 +10302291find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
Nick Piggin147cbb42005-06-25 14:57:19 -07002292{
2293 unsigned long load, min_load = ULONG_MAX;
2294 int idlest = -1;
2295 int i;
2296
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002297 /* Traverse only the allowed CPUs */
Rusty Russell758b2cd2008-11-25 02:35:04 +10302298 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002299 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07002300
2301 if (load < min_load || (load == min_load && i == this_cpu)) {
2302 min_load = load;
2303 idlest = i;
2304 }
2305 }
2306
2307 return idlest;
2308}
2309
Nick Piggin476d1392005-06-25 14:57:29 -07002310/*
2311 * sched_balance_self: balance the current task (running on cpu) in domains
2312 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2313 * SD_BALANCE_EXEC.
2314 *
2315 * Balance, ie. select the least loaded group.
2316 *
2317 * Returns the target CPU number, or the same CPU if no balancing is needed.
2318 *
2319 * preempt must be disabled.
2320 */
2321static int sched_balance_self(int cpu, int flag)
2322{
2323 struct task_struct *t = current;
2324 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07002325
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002326 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02002327 /*
2328 * If power savings logic is enabled for a domain, stop there.
2329 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002330 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2331 break;
Nick Piggin476d1392005-06-25 14:57:29 -07002332 if (tmp->flags & flag)
2333 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002334 }
Nick Piggin476d1392005-06-25 14:57:29 -07002335
Peter Zijlstra039a1c42008-06-27 13:41:25 +02002336 if (sd)
2337 update_shares(sd);
2338
Nick Piggin476d1392005-06-25 14:57:29 -07002339 while (sd) {
Nick Piggin476d1392005-06-25 14:57:29 -07002340 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002341 int new_cpu, weight;
2342
2343 if (!(sd->flags & flag)) {
2344 sd = sd->child;
2345 continue;
2346 }
Nick Piggin476d1392005-06-25 14:57:29 -07002347
Nick Piggin476d1392005-06-25 14:57:29 -07002348 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002349 if (!group) {
2350 sd = sd->child;
2351 continue;
2352 }
Nick Piggin476d1392005-06-25 14:57:29 -07002353
Rusty Russell758b2cd2008-11-25 02:35:04 +10302354 new_cpu = find_idlest_cpu(group, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002355 if (new_cpu == -1 || new_cpu == cpu) {
2356 /* Now try balancing at a lower domain level of cpu */
2357 sd = sd->child;
2358 continue;
2359 }
Nick Piggin476d1392005-06-25 14:57:29 -07002360
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002361 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07002362 cpu = new_cpu;
Rusty Russell758b2cd2008-11-25 02:35:04 +10302363 weight = cpumask_weight(sched_domain_span(sd));
Nick Piggin476d1392005-06-25 14:57:29 -07002364 sd = NULL;
Nick Piggin476d1392005-06-25 14:57:29 -07002365 for_each_domain(cpu, tmp) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302366 if (weight <= cpumask_weight(sched_domain_span(tmp)))
Nick Piggin476d1392005-06-25 14:57:29 -07002367 break;
2368 if (tmp->flags & flag)
2369 sd = tmp;
2370 }
2371 /* while loop will break here if sd == NULL */
2372 }
2373
2374 return cpu;
2375}
2376
2377#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002378
Thomas Gleixner0793a612008-12-04 20:12:29 +01002379/**
2380 * task_oncpu_function_call - call a function on the cpu on which a task runs
2381 * @p: the task to evaluate
2382 * @func: the function to be called
2383 * @info: the function call argument
2384 *
2385 * Calls the function @func when the task is currently running. This might
2386 * be on the current CPU, which just calls the function directly
2387 */
2388void task_oncpu_function_call(struct task_struct *p,
2389 void (*func) (void *info), void *info)
2390{
2391 int cpu;
2392
2393 preempt_disable();
2394 cpu = task_cpu(p);
2395 if (task_curr(p))
2396 smp_call_function_single(cpu, func, info, 1);
2397 preempt_enable();
2398}
2399
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400/***
2401 * try_to_wake_up - wake up a thread
2402 * @p: the to-be-woken-up thread
2403 * @state: the mask of task states that can be woken
2404 * @sync: do a synchronous wakeup?
2405 *
2406 * Put it on the run-queue if it's not already there. The "current"
2407 * thread is always on the run-queue (except when the actual
2408 * re-schedule is in progress), and as such you're allowed to do
2409 * the simpler "current->state = TASK_RUNNING" to mark yourself
2410 * runnable without the overhead of this.
2411 *
2412 * returns failure only if the task is already active.
2413 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002414static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002415{
Ingo Molnarcc367732007-10-15 17:00:18 +02002416 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002417 unsigned long flags;
2418 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002419 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002420
Ingo Molnarb85d0662008-03-16 20:03:22 +01002421 if (!sched_feat(SYNC_WAKEUPS))
2422 sync = 0;
2423
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002424#ifdef CONFIG_SMP
Peter Zijlstra57310a92009-03-09 13:56:21 +01002425 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002426 struct sched_domain *sd;
2427
2428 this_cpu = raw_smp_processor_id();
2429 cpu = task_cpu(p);
2430
2431 for_each_domain(this_cpu, sd) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302432 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002433 update_shares(sd);
2434 break;
2435 }
2436 }
2437 }
2438#endif
2439
Linus Torvalds04e2f172008-02-23 18:05:03 -08002440 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002441 rq = task_rq_lock(p, &flags);
Mike Galbraith03e89e42008-12-16 08:45:30 +01002442 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002443 old_state = p->state;
2444 if (!(old_state & state))
2445 goto out;
2446
Ingo Molnardd41f592007-07-09 18:51:59 +02002447 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002448 goto out_running;
2449
2450 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002451 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002452 this_cpu = smp_processor_id();
2453
2454#ifdef CONFIG_SMP
2455 if (unlikely(task_running(rq, p)))
2456 goto out_activate;
2457
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01002458 cpu = p->sched_class->select_task_rq(p, sync);
2459 if (cpu != orig_cpu) {
2460 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002461 task_rq_unlock(rq, &flags);
2462 /* might preempt at this point */
2463 rq = task_rq_lock(p, &flags);
2464 old_state = p->state;
2465 if (!(old_state & state))
2466 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02002467 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002468 goto out_running;
2469
2470 this_cpu = smp_processor_id();
2471 cpu = task_cpu(p);
2472 }
2473
Gregory Haskinse7693a32008-01-25 21:08:09 +01002474#ifdef CONFIG_SCHEDSTATS
2475 schedstat_inc(rq, ttwu_count);
2476 if (cpu == this_cpu)
2477 schedstat_inc(rq, ttwu_local);
2478 else {
2479 struct sched_domain *sd;
2480 for_each_domain(this_cpu, sd) {
Rusty Russell758b2cd2008-11-25 02:35:04 +10302481 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Gregory Haskinse7693a32008-01-25 21:08:09 +01002482 schedstat_inc(sd, ttwu_wake_remote);
2483 break;
2484 }
2485 }
2486 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002487#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002488
Linus Torvalds1da177e2005-04-16 15:20:36 -07002489out_activate:
2490#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002491 schedstat_inc(p, se.nr_wakeups);
2492 if (sync)
2493 schedstat_inc(p, se.nr_wakeups_sync);
2494 if (orig_cpu != cpu)
2495 schedstat_inc(p, se.nr_wakeups_migrate);
2496 if (cpu == this_cpu)
2497 schedstat_inc(p, se.nr_wakeups_local);
2498 else
2499 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnardd41f592007-07-09 18:51:59 +02002500 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002501 success = 1;
2502
Peter Zijlstra831451a2009-01-14 12:39:18 +01002503 /*
2504 * Only attribute actual wakeups done by this task.
2505 */
2506 if (!in_interrupt()) {
2507 struct sched_entity *se = &current->se;
2508 u64 sample = se->sum_exec_runtime;
2509
2510 if (se->last_wakeup)
2511 sample -= se->last_wakeup;
2512 else
2513 sample -= se->start_runtime;
2514 update_avg(&se->avg_wakeup, sample);
2515
2516 se->last_wakeup = se->sum_exec_runtime;
2517 }
2518
Linus Torvalds1da177e2005-04-16 15:20:36 -07002519out_running:
Peter Zijlstra468a15b2008-12-16 08:07:03 +01002520 trace_sched_wakeup(rq, p, success);
Peter Zijlstra15afe092008-09-20 23:38:02 +02002521 check_preempt_curr(rq, p, sync);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002522
Linus Torvalds1da177e2005-04-16 15:20:36 -07002523 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002524#ifdef CONFIG_SMP
2525 if (p->sched_class->task_wake_up)
2526 p->sched_class->task_wake_up(rq, p);
2527#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002528out:
2529 task_rq_unlock(rq, &flags);
2530
2531 return success;
2532}
2533
David Howells50fa6102009-04-28 15:01:38 +01002534/**
2535 * wake_up_process - Wake up a specific process
2536 * @p: The process to be woken up.
2537 *
2538 * Attempt to wake up the nominated process and move it to the set of runnable
2539 * processes. Returns 1 if the process was woken up, 0 if it was already
2540 * running.
2541 *
2542 * It may be assumed that this function implies a write memory barrier before
2543 * changing the task state if and only if any tasks are woken up.
2544 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002545int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002547 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002549EXPORT_SYMBOL(wake_up_process);
2550
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002551int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002552{
2553 return try_to_wake_up(p, state, 0);
2554}
2555
Linus Torvalds1da177e2005-04-16 15:20:36 -07002556/*
2557 * Perform scheduler related setup for a newly forked process p.
2558 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002559 *
2560 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002561 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002562static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002563{
Ingo Molnardd41f592007-07-09 18:51:59 +02002564 p->se.exec_start = 0;
2565 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002566 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar6c594c22008-12-14 12:34:15 +01002567 p->se.nr_migrations = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002568 p->se.last_wakeup = 0;
2569 p->se.avg_overlap = 0;
Peter Zijlstra831451a2009-01-14 12:39:18 +01002570 p->se.start_runtime = 0;
2571 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002572
2573#ifdef CONFIG_SCHEDSTATS
2574 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002575 p->se.sum_sleep_runtime = 0;
2576 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002577 p->se.block_start = 0;
2578 p->se.sleep_max = 0;
2579 p->se.block_max = 0;
2580 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02002581 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002582 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002583#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002584
Peter Zijlstrafa717062008-01-25 21:08:27 +01002585 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002586 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002587 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002588
Avi Kivitye107be32007-07-26 13:40:43 +02002589#ifdef CONFIG_PREEMPT_NOTIFIERS
2590 INIT_HLIST_HEAD(&p->preempt_notifiers);
2591#endif
2592
Linus Torvalds1da177e2005-04-16 15:20:36 -07002593 /*
2594 * We mark the process as running here, but have not actually
2595 * inserted it onto the runqueue yet. This guarantees that
2596 * nobody will actually run it, and a signal or other external
2597 * event cannot wake it up and insert it on the runqueue either.
2598 */
2599 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002600}
2601
2602/*
2603 * fork()/clone()-time setup:
2604 */
2605void sched_fork(struct task_struct *p, int clone_flags)
2606{
2607 int cpu = get_cpu();
2608
2609 __sched_fork(p);
2610
2611#ifdef CONFIG_SMP
2612 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2613#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02002614 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07002615
2616 /*
2617 * Make sure we do not leak PI boosting priority to the child:
2618 */
2619 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002620 if (!rt_prio(p->prio))
2621 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002622
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002623#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002624 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002625 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002626#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002627#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002628 p->oncpu = 0;
2629#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002631 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002632 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002633#endif
Gregory Haskins917b6272008-12-29 09:39:53 -05002634 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2635
Nick Piggin476d1392005-06-25 14:57:29 -07002636 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002637}
2638
2639/*
2640 * wake_up_new_task - wake up a newly created task for the first time.
2641 *
2642 * This function will do some initial scheduler statistics housekeeping
2643 * that must be done for every newly created context, then puts the task
2644 * on the runqueue and wakes it.
2645 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002646void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002647{
2648 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002649 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002650
2651 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002652 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002653 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002654
2655 p->prio = effective_prio(p);
2656
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002657 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002658 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002659 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002660 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002661 * Let the scheduling class do new task startup
2662 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002663 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002664 p->sched_class->task_new(rq, p);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002665 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002666 }
Ingo Molnarc71dd422008-12-19 01:09:51 +01002667 trace_sched_wakeup_new(rq, p, 1);
Peter Zijlstra15afe092008-09-20 23:38:02 +02002668 check_preempt_curr(rq, p, 0);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002669#ifdef CONFIG_SMP
2670 if (p->sched_class->task_wake_up)
2671 p->sched_class->task_wake_up(rq, p);
2672#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002673 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002674}
2675
Avi Kivitye107be32007-07-26 13:40:43 +02002676#ifdef CONFIG_PREEMPT_NOTIFIERS
2677
2678/**
Luis Henriques80dd99b2009-03-16 19:58:09 +00002679 * preempt_notifier_register - tell me when current is being preempted & rescheduled
Randy Dunlap421cee22007-07-31 00:37:50 -07002680 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002681 */
2682void preempt_notifier_register(struct preempt_notifier *notifier)
2683{
2684 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2685}
2686EXPORT_SYMBOL_GPL(preempt_notifier_register);
2687
2688/**
2689 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002690 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002691 *
2692 * This is safe to call from within a preemption notifier.
2693 */
2694void preempt_notifier_unregister(struct preempt_notifier *notifier)
2695{
2696 hlist_del(&notifier->link);
2697}
2698EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2699
2700static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2701{
2702 struct preempt_notifier *notifier;
2703 struct hlist_node *node;
2704
2705 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2706 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2707}
2708
2709static void
2710fire_sched_out_preempt_notifiers(struct task_struct *curr,
2711 struct task_struct *next)
2712{
2713 struct preempt_notifier *notifier;
2714 struct hlist_node *node;
2715
2716 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2717 notifier->ops->sched_out(notifier, next);
2718}
2719
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002720#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002721
2722static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2723{
2724}
2725
2726static void
2727fire_sched_out_preempt_notifiers(struct task_struct *curr,
2728 struct task_struct *next)
2729{
2730}
2731
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002732#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002733
Linus Torvalds1da177e2005-04-16 15:20:36 -07002734/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002735 * prepare_task_switch - prepare to switch tasks
2736 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002737 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002738 * @next: the task we are going to switch to.
2739 *
2740 * This is called with the rq lock held and interrupts off. It must
2741 * be paired with a subsequent finish_task_switch after the context
2742 * switch.
2743 *
2744 * prepare_task_switch sets up locking and calls architecture specific
2745 * hooks.
2746 */
Avi Kivitye107be32007-07-26 13:40:43 +02002747static inline void
2748prepare_task_switch(struct rq *rq, struct task_struct *prev,
2749 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002750{
Avi Kivitye107be32007-07-26 13:40:43 +02002751 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002752 prepare_lock_switch(rq, next);
2753 prepare_arch_switch(next);
2754}
2755
2756/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002757 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002758 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002759 * @prev: the thread we just switched away from.
2760 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002761 * finish_task_switch must be called after the context switch, paired
2762 * with a prepare_task_switch call before the context switch.
2763 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2764 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002765 *
2766 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002767 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 * with the lock held can cause deadlocks; see schedule() for
2769 * details.)
2770 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002771static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002772 __releases(rq->lock)
2773{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002774 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002775 long prev_state;
Gregory Haskins967fc042008-12-29 09:39:52 -05002776#ifdef CONFIG_SMP
2777 int post_schedule = 0;
2778
2779 if (current->sched_class->needs_post_schedule)
2780 post_schedule = current->sched_class->needs_post_schedule(rq);
2781#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782
2783 rq->prev_mm = NULL;
2784
2785 /*
2786 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002787 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002788 * schedule one last time. The schedule call will never return, and
2789 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002790 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 * still held, otherwise prev could be scheduled on another cpu, die
2792 * there before we look at prev->state, and then the reference would
2793 * be dropped twice.
2794 * Manfred Spraul <manfred@colorfullife.com>
2795 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002796 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002797 finish_arch_switch(prev);
Thomas Gleixner0793a612008-12-04 20:12:29 +01002798 perf_counter_task_sched_in(current, cpu_of(rq));
Nick Piggin4866cde2005-06-25 14:57:23 -07002799 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002800#ifdef CONFIG_SMP
Gregory Haskins967fc042008-12-29 09:39:52 -05002801 if (post_schedule)
Steven Rostedt9a897c52008-01-25 21:08:22 +01002802 current->sched_class->post_schedule(rq);
2803#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002804
Avi Kivitye107be32007-07-26 13:40:43 +02002805 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002806 if (mm)
2807 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002808 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002809 /*
2810 * Remove function-return probe instances associated with this
2811 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002812 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002813 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002814 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816}
2817
2818/**
2819 * schedule_tail - first thing a freshly forked thread must call.
2820 * @prev: the thread we just switched away from.
2821 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002822asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002823 __releases(rq->lock)
2824{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002825 struct rq *rq = this_rq();
2826
Nick Piggin4866cde2005-06-25 14:57:23 -07002827 finish_task_switch(rq, prev);
2828#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2829 /* In this case, finish_task_switch does not reenable preemption */
2830 preempt_enable();
2831#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002832 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002833 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002834}
2835
2836/*
2837 * context_switch - switch to the new MM and the new
2838 * thread's register state.
2839 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002840static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002841context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002842 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002843{
Ingo Molnardd41f592007-07-09 18:51:59 +02002844 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002845
Avi Kivitye107be32007-07-26 13:40:43 +02002846 prepare_task_switch(rq, prev, next);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002847 trace_sched_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002848 mm = next->mm;
2849 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002850 /*
2851 * For paravirt, this is coupled with an exit in switch_to to
2852 * combine the page table reload and the switch backend into
2853 * one hypercall.
2854 */
Jeremy Fitzhardinge224101e2009-02-18 11:18:57 -08002855 arch_start_context_switch(prev);
Zachary Amsden9226d122007-02-13 13:26:21 +01002856
Ingo Molnardd41f592007-07-09 18:51:59 +02002857 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 next->active_mm = oldmm;
2859 atomic_inc(&oldmm->mm_count);
2860 enter_lazy_tlb(oldmm, next);
2861 } else
2862 switch_mm(oldmm, mm, next);
2863
Ingo Molnardd41f592007-07-09 18:51:59 +02002864 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002865 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002866 rq->prev_mm = oldmm;
2867 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002868 /*
2869 * Since the runqueue lock will be released by the next
2870 * task (which is an invalid locking op but in the case
2871 * of the scheduler it's an obvious special-case), so we
2872 * do an early lockdep release here:
2873 */
2874#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002875 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002876#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002877
2878 /* Here we just switch the register state and the stack. */
2879 switch_to(prev, next, prev);
2880
Ingo Molnardd41f592007-07-09 18:51:59 +02002881 barrier();
2882 /*
2883 * this_rq must be evaluated again because prev may have moved
2884 * CPUs since it called schedule(), thus the 'rq' on its stack
2885 * frame will be invalid.
2886 */
2887 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002888}
2889
2890/*
2891 * nr_running, nr_uninterruptible and nr_context_switches:
2892 *
2893 * externally visible scheduler statistics: current number of runnable
2894 * threads, current number of uninterruptible-sleeping threads, total
2895 * number of context switches performed since bootup.
2896 */
2897unsigned long nr_running(void)
2898{
2899 unsigned long i, sum = 0;
2900
2901 for_each_online_cpu(i)
2902 sum += cpu_rq(i)->nr_running;
2903
2904 return sum;
2905}
2906
2907unsigned long nr_uninterruptible(void)
2908{
2909 unsigned long i, sum = 0;
2910
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002911 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002912 sum += cpu_rq(i)->nr_uninterruptible;
2913
2914 /*
2915 * Since we read the counters lockless, it might be slightly
2916 * inaccurate. Do not allow it to go below zero though:
2917 */
2918 if (unlikely((long)sum < 0))
2919 sum = 0;
2920
2921 return sum;
2922}
2923
2924unsigned long long nr_context_switches(void)
2925{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002926 int i;
2927 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002928
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002929 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002930 sum += cpu_rq(i)->nr_switches;
2931
2932 return sum;
2933}
2934
2935unsigned long nr_iowait(void)
2936{
2937 unsigned long i, sum = 0;
2938
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002939 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002940 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2941
2942 return sum;
2943}
2944
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002945/* Variables and functions for calc_load */
2946static atomic_long_t calc_load_tasks;
2947static unsigned long calc_load_update;
2948unsigned long avenrun[3];
2949EXPORT_SYMBOL(avenrun);
2950
Thomas Gleixner2d024942009-05-02 20:08:52 +02002951/**
2952 * get_avenrun - get the load average array
2953 * @loads: pointer to dest load array
2954 * @offset: offset to add
2955 * @shift: shift count to shift the result left
2956 *
2957 * These values are estimates at best, so no need for locking.
2958 */
2959void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2960{
2961 loads[0] = (avenrun[0] + offset) << shift;
2962 loads[1] = (avenrun[1] + offset) << shift;
2963 loads[2] = (avenrun[2] + offset) << shift;
2964}
2965
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002966static unsigned long
2967calc_load(unsigned long load, unsigned long exp, unsigned long active)
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002968{
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002969 load *= exp;
2970 load += active * (FIXED_1 - exp);
2971 return load >> FSHIFT;
2972}
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002973
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02002974/*
2975 * calc_load - update the avenrun load estimates 10 ticks after the
2976 * CPUs have updated calc_load_tasks.
2977 */
2978void calc_global_load(void)
2979{
2980 unsigned long upd = calc_load_update + 10;
2981 long active;
2982
2983 if (time_before(jiffies, upd))
2984 return;
2985
2986 active = atomic_long_read(&calc_load_tasks);
2987 active = active > 0 ? active * FIXED_1 : 0;
2988
2989 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2990 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2991 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2992
2993 calc_load_update += LOAD_FREQ;
2994}
2995
2996/*
2997 * Either called from update_cpu_load() or from a cpu going idle
2998 */
2999static void calc_load_account_active(struct rq *this_rq)
3000{
3001 long nr_active, delta;
3002
3003 nr_active = this_rq->nr_running;
3004 nr_active += (long) this_rq->nr_uninterruptible;
3005
3006 if (nr_active != this_rq->calc_load_active) {
3007 delta = nr_active - this_rq->calc_load_active;
3008 this_rq->calc_load_active = nr_active;
3009 atomic_long_add(delta, &calc_load_tasks);
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003010 }
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08003011}
3012
Linus Torvalds1da177e2005-04-16 15:20:36 -07003013/*
Paul Mackerras23a185c2009-02-09 22:42:47 +11003014 * Externally visible per-cpu scheduler statistics:
Paul Mackerras23a185c2009-02-09 22:42:47 +11003015 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3016 */
Paul Mackerras23a185c2009-02-09 22:42:47 +11003017u64 cpu_nr_migrations(int cpu)
3018{
3019 return cpu_rq(cpu)->nr_migrations_in;
3020}
3021
3022/*
Ingo Molnardd41f592007-07-09 18:51:59 +02003023 * Update rq->cpu_load[] statistics. This function is usually called every
3024 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07003025 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003026static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003027{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02003028 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02003029 int i, scale;
3030
3031 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02003032
3033 /* Update our load: */
3034 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3035 unsigned long old_load, new_load;
3036
3037 /* scale is effectively 1 << i now, and >> i divides by scale */
3038
3039 old_load = this_rq->cpu_load[i];
3040 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02003041 /*
3042 * Round up the averaging division if load is increasing. This
3043 * prevents us from getting stuck on 9 if the load is 10, for
3044 * example.
3045 */
3046 if (new_load > old_load)
3047 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003048 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3049 }
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02003050
3051 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3052 this_rq->calc_load_update += LOAD_FREQ;
3053 calc_load_account_active(this_rq);
3054 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003055}
3056
Ingo Molnardd41f592007-07-09 18:51:59 +02003057#ifdef CONFIG_SMP
3058
Ingo Molnar48f24c42006-07-03 00:25:40 -07003059/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003060 * double_rq_lock - safely lock two runqueues
3061 *
3062 * Note this does not disable interrupts like task_rq_lock,
3063 * you need to do so manually before calling.
3064 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003065static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003066 __acquires(rq1->lock)
3067 __acquires(rq2->lock)
3068{
Kirill Korotaev054b9102006-12-10 02:20:11 -08003069 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07003070 if (rq1 == rq2) {
3071 spin_lock(&rq1->lock);
3072 __acquire(rq2->lock); /* Fake it out ;) */
3073 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003074 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003075 spin_lock(&rq1->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003076 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003077 } else {
3078 spin_lock(&rq2->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02003079 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003080 }
3081 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003082 update_rq_clock(rq1);
3083 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003084}
3085
3086/*
3087 * double_rq_unlock - safely unlock two runqueues
3088 *
3089 * Note this does not restore interrupts like task_rq_unlock,
3090 * you need to do so manually after calling.
3091 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003092static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003093 __releases(rq1->lock)
3094 __releases(rq2->lock)
3095{
3096 spin_unlock(&rq1->lock);
3097 if (rq1 != rq2)
3098 spin_unlock(&rq2->lock);
3099 else
3100 __release(rq2->lock);
3101}
3102
3103/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003104 * If dest_cpu is allowed for this process, migrate the task to it.
3105 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003106 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07003107 * the cpu_allowed mask is restored.
3108 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07003109static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003110{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003111 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003112 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003113 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
3115 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e2008-11-25 02:35:14 +10303116 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
Max Krasnyanskye761b772008-07-15 04:43:49 -07003117 || unlikely(!cpu_active(dest_cpu)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003118 goto out;
3119
3120 /* force the process onto the specified CPU */
3121 if (migrate_task(p, dest_cpu, &req)) {
3122 /* Need to wait for migration thread (might exit: take ref). */
3123 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07003124
Linus Torvalds1da177e2005-04-16 15:20:36 -07003125 get_task_struct(mt);
3126 task_rq_unlock(rq, &flags);
3127 wake_up_process(mt);
3128 put_task_struct(mt);
3129 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07003130
Linus Torvalds1da177e2005-04-16 15:20:36 -07003131 return;
3132 }
3133out:
3134 task_rq_unlock(rq, &flags);
3135}
3136
3137/*
Nick Piggin476d1392005-06-25 14:57:29 -07003138 * sched_exec - execve() is a valuable balancing opportunity, because at
3139 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003140 */
3141void sched_exec(void)
3142{
Linus Torvalds1da177e2005-04-16 15:20:36 -07003143 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003144 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003145 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07003146 if (new_cpu != this_cpu)
3147 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148}
3149
3150/*
3151 * pull_task - move a task from a remote runqueue to the local runqueue.
3152 * Both runqueues must be locked.
3153 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003154static void pull_task(struct rq *src_rq, struct task_struct *p,
3155 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003156{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02003157 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003158 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003159 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003160 /*
3161 * Note that idle threads have a prio of MAX_PRIO, for this test
3162 * to be always true for them.
3163 */
Peter Zijlstra15afe092008-09-20 23:38:02 +02003164 check_preempt_curr(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003165}
3166
3167/*
3168 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3169 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08003170static
Ingo Molnar70b97a72006-07-03 00:25:42 -07003171int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003172 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07003173 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003174{
Luis Henriques708dc512009-03-16 19:59:02 +00003175 int tsk_cache_hot = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003176 /*
3177 * We do not migrate tasks that are:
3178 * 1) running (obviously), or
3179 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3180 * 3) are cache-hot on their current CPU.
3181 */
Rusty Russell96f874e2008-11-25 02:35:14 +10303182 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003183 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003184 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003185 }
Nick Piggin81026792005-06-25 14:57:07 -07003186 *all_pinned = 0;
3187
Ingo Molnarcc367732007-10-15 17:00:18 +02003188 if (task_running(rq, p)) {
3189 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07003190 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003191 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003192
Ingo Molnarda84d962007-10-15 17:00:18 +02003193 /*
3194 * Aggressive migration if:
3195 * 1) task is cache cold, or
3196 * 2) too many balance attempts have failed.
3197 */
3198
Luis Henriques708dc512009-03-16 19:59:02 +00003199 tsk_cache_hot = task_hot(p, rq->clock, sd);
3200 if (!tsk_cache_hot ||
3201 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003202#ifdef CONFIG_SCHEDSTATS
Luis Henriques708dc512009-03-16 19:59:02 +00003203 if (tsk_cache_hot) {
Ingo Molnarda84d962007-10-15 17:00:18 +02003204 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02003205 schedstat_inc(p, se.nr_forced_migrations);
3206 }
Ingo Molnarda84d962007-10-15 17:00:18 +02003207#endif
3208 return 1;
3209 }
3210
Luis Henriques708dc512009-03-16 19:59:02 +00003211 if (tsk_cache_hot) {
Ingo Molnarcc367732007-10-15 17:00:18 +02003212 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02003213 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02003214 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 return 1;
3216}
3217
Peter Williamse1d14842007-10-24 18:23:51 +02003218static unsigned long
3219balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3220 unsigned long max_load_move, struct sched_domain *sd,
3221 enum cpu_idle_type idle, int *all_pinned,
3222 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02003223{
Peter Zijlstra051c6762008-06-27 13:41:31 +02003224 int loops = 0, pulled = 0, pinned = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02003225 struct task_struct *p;
3226 long rem_load_move = max_load_move;
3227
Peter Williamse1d14842007-10-24 18:23:51 +02003228 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02003229 goto out;
3230
3231 pinned = 1;
3232
3233 /*
3234 * Start the load-balancing iterator:
3235 */
3236 p = iterator->start(iterator->arg);
3237next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003238 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02003239 goto out;
Peter Zijlstra051c6762008-06-27 13:41:31 +02003240
3241 if ((p->se.load.weight >> 1) > rem_load_move ||
Ingo Molnardd41f592007-07-09 18:51:59 +02003242 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003243 p = iterator->next(iterator->arg);
3244 goto next;
3245 }
3246
3247 pull_task(busiest, p, this_rq, this_cpu);
3248 pulled++;
3249 rem_load_move -= p->se.load.weight;
3250
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003251#ifdef CONFIG_PREEMPT
3252 /*
3253 * NEWIDLE balancing is a source of latency, so preemptible kernels
3254 * will stop after the first task is pulled to minimize the critical
3255 * section.
3256 */
3257 if (idle == CPU_NEWLY_IDLE)
3258 goto out;
3259#endif
3260
Ingo Molnardd41f592007-07-09 18:51:59 +02003261 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01003262 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003263 */
Peter Williamse1d14842007-10-24 18:23:51 +02003264 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003265 if (p->prio < *this_best_prio)
3266 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02003267 p = iterator->next(iterator->arg);
3268 goto next;
3269 }
3270out:
3271 /*
Peter Williamse1d14842007-10-24 18:23:51 +02003272 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02003273 * so we can safely collect pull_task() stats here rather than
3274 * inside pull_task().
3275 */
3276 schedstat_add(sd, lb_gained[idle], pulled);
3277
3278 if (all_pinned)
3279 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003280
3281 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003282}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003283
Linus Torvalds1da177e2005-04-16 15:20:36 -07003284/*
Peter Williams43010652007-08-09 11:16:46 +02003285 * move_tasks tries to move up to max_load_move weighted load from busiest to
3286 * this_rq, as part of a balancing operation within domain "sd".
3287 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003288 *
3289 * Called with both runqueues locked.
3290 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003291static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003292 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003293 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003294 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003295{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003296 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003297 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003298 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003299
Ingo Molnardd41f592007-07-09 18:51:59 +02003300 do {
Peter Williams43010652007-08-09 11:16:46 +02003301 total_load_moved +=
3302 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003303 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003304 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003305 class = class->next;
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003306
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003307#ifdef CONFIG_PREEMPT
3308 /*
3309 * NEWIDLE balancing is a source of latency, so preemptible
3310 * kernels will stop after the first task is pulled to minimize
3311 * the critical section.
3312 */
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003313 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3314 break;
Gregory Haskins7e96fa52008-12-29 09:39:50 -05003315#endif
Peter Williams43010652007-08-09 11:16:46 +02003316 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003317
Peter Williams43010652007-08-09 11:16:46 +02003318 return total_load_moved > 0;
3319}
3320
Peter Williamse1d14842007-10-24 18:23:51 +02003321static int
3322iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3323 struct sched_domain *sd, enum cpu_idle_type idle,
3324 struct rq_iterator *iterator)
3325{
3326 struct task_struct *p = iterator->start(iterator->arg);
3327 int pinned = 0;
3328
3329 while (p) {
3330 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3331 pull_task(busiest, p, this_rq, this_cpu);
3332 /*
3333 * Right now, this is only the second place pull_task()
3334 * is called, so we can safely collect pull_task()
3335 * stats here rather than inside pull_task().
3336 */
3337 schedstat_inc(sd, lb_gained[idle]);
3338
3339 return 1;
3340 }
3341 p = iterator->next(iterator->arg);
3342 }
3343
3344 return 0;
3345}
3346
Peter Williams43010652007-08-09 11:16:46 +02003347/*
3348 * move_one_task tries to move exactly one task from busiest to this_rq, as
3349 * part of active balancing operations within "domain".
3350 * Returns 1 if successful and 0 otherwise.
3351 *
3352 * Called with both runqueues locked.
3353 */
3354static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3355 struct sched_domain *sd, enum cpu_idle_type idle)
3356{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003357 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003358
3359 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02003360 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003361 return 1;
3362
3363 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003364}
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303365/********** Helpers for find_busiest_group ************************/
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003366/*
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303367 * sd_lb_stats - Structure to store the statistics of a sched_domain
3368 * during load balancing.
3369 */
3370struct sd_lb_stats {
3371 struct sched_group *busiest; /* Busiest group in this sd */
3372 struct sched_group *this; /* Local group in this sd */
3373 unsigned long total_load; /* Total load of all groups in sd */
3374 unsigned long total_pwr; /* Total power of all groups in sd */
3375 unsigned long avg_load; /* Average load across all groups in sd */
3376
3377 /** Statistics of this group */
3378 unsigned long this_load;
3379 unsigned long this_load_per_task;
3380 unsigned long this_nr_running;
3381
3382 /* Statistics of the busiest group */
3383 unsigned long max_load;
3384 unsigned long busiest_load_per_task;
3385 unsigned long busiest_nr_running;
3386
3387 int group_imb; /* Is there imbalance in this sd */
3388#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3389 int power_savings_balance; /* Is powersave balance needed for this sd */
3390 struct sched_group *group_min; /* Least loaded group in sd */
3391 struct sched_group *group_leader; /* Group which relieves group_min */
3392 unsigned long min_load_per_task; /* load_per_task in group_min */
3393 unsigned long leader_nr_running; /* Nr running of group_leader */
3394 unsigned long min_nr_running; /* Nr running of group_min */
3395#endif
3396};
Linus Torvalds1da177e2005-04-16 15:20:36 -07003397
3398/*
Gautham R Shenoy381be782009-03-25 14:43:46 +05303399 * sg_lb_stats - stats of a sched_group required for load_balancing
3400 */
3401struct sg_lb_stats {
3402 unsigned long avg_load; /*Avg load across the CPUs of the group */
3403 unsigned long group_load; /* Total load over the CPUs of the group */
3404 unsigned long sum_nr_running; /* Nr tasks running in the group */
3405 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3406 unsigned long group_capacity;
3407 int group_imb; /* Is there an imbalance in the group ? */
3408};
3409
3410/**
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303411 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3412 * @group: The group whose first cpu is to be returned.
3413 */
3414static inline unsigned int group_first_cpu(struct sched_group *group)
3415{
3416 return cpumask_first(sched_group_cpus(group));
3417}
3418
3419/**
3420 * get_sd_load_idx - Obtain the load index for a given sched domain.
3421 * @sd: The sched_domain whose load_idx is to be obtained.
3422 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3423 */
3424static inline int get_sd_load_idx(struct sched_domain *sd,
3425 enum cpu_idle_type idle)
3426{
3427 int load_idx;
3428
3429 switch (idle) {
3430 case CPU_NOT_IDLE:
3431 load_idx = sd->busy_idx;
3432 break;
3433
3434 case CPU_NEWLY_IDLE:
3435 load_idx = sd->newidle_idx;
3436 break;
3437 default:
3438 load_idx = sd->idle_idx;
3439 break;
3440 }
3441
3442 return load_idx;
3443}
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303444
3445
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303446#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3447/**
3448 * init_sd_power_savings_stats - Initialize power savings statistics for
3449 * the given sched_domain, during load balancing.
3450 *
3451 * @sd: Sched domain whose power-savings statistics are to be initialized.
3452 * @sds: Variable containing the statistics for sd.
3453 * @idle: Idle status of the CPU at which we're performing load-balancing.
3454 */
3455static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3456 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3457{
3458 /*
3459 * Busy processors will not participate in power savings
3460 * balance.
3461 */
3462 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3463 sds->power_savings_balance = 0;
3464 else {
3465 sds->power_savings_balance = 1;
3466 sds->min_nr_running = ULONG_MAX;
3467 sds->leader_nr_running = 0;
3468 }
3469}
3470
3471/**
3472 * update_sd_power_savings_stats - Update the power saving stats for a
3473 * sched_domain while performing load balancing.
3474 *
3475 * @group: sched_group belonging to the sched_domain under consideration.
3476 * @sds: Variable containing the statistics of the sched_domain
3477 * @local_group: Does group contain the CPU for which we're performing
3478 * load balancing ?
3479 * @sgs: Variable containing the statistics of the group.
3480 */
3481static inline void update_sd_power_savings_stats(struct sched_group *group,
3482 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3483{
3484
3485 if (!sds->power_savings_balance)
3486 return;
3487
3488 /*
3489 * If the local group is idle or completely loaded
3490 * no need to do power savings balance at this domain
3491 */
3492 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3493 !sds->this_nr_running))
3494 sds->power_savings_balance = 0;
3495
3496 /*
3497 * If a group is already running at full capacity or idle,
3498 * don't include that group in power savings calculations
3499 */
3500 if (!sds->power_savings_balance ||
3501 sgs->sum_nr_running >= sgs->group_capacity ||
3502 !sgs->sum_nr_running)
3503 return;
3504
3505 /*
3506 * Calculate the group which has the least non-idle load.
3507 * This is the group from where we need to pick up the load
3508 * for saving power
3509 */
3510 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3511 (sgs->sum_nr_running == sds->min_nr_running &&
3512 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3513 sds->group_min = group;
3514 sds->min_nr_running = sgs->sum_nr_running;
3515 sds->min_load_per_task = sgs->sum_weighted_load /
3516 sgs->sum_nr_running;
3517 }
3518
3519 /*
3520 * Calculate the group which is almost near its
3521 * capacity but still has some space to pick up some load
3522 * from other group and save more power
3523 */
3524 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3525 return;
3526
3527 if (sgs->sum_nr_running > sds->leader_nr_running ||
3528 (sgs->sum_nr_running == sds->leader_nr_running &&
3529 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3530 sds->group_leader = group;
3531 sds->leader_nr_running = sgs->sum_nr_running;
3532 }
3533}
3534
3535/**
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003536 * check_power_save_busiest_group - see if there is potential for some power-savings balance
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303537 * @sds: Variable containing the statistics of the sched_domain
3538 * under consideration.
3539 * @this_cpu: Cpu at which we're currently performing load-balancing.
3540 * @imbalance: Variable to store the imbalance.
3541 *
Randy Dunlapd5ac5372009-03-28 21:52:47 -07003542 * Description:
3543 * Check if we have potential to perform some power-savings balance.
3544 * If yes, set the busiest group to be the least loaded group in the
3545 * sched_domain, so that it's CPUs can be put to idle.
3546 *
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303547 * Returns 1 if there is potential to perform power-savings balance.
3548 * Else returns 0.
3549 */
3550static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3551 int this_cpu, unsigned long *imbalance)
3552{
3553 if (!sds->power_savings_balance)
3554 return 0;
3555
3556 if (sds->this != sds->group_leader ||
3557 sds->group_leader == sds->group_min)
3558 return 0;
3559
3560 *imbalance = sds->min_load_per_task;
3561 sds->busiest = sds->group_min;
3562
3563 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3564 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3565 group_first_cpu(sds->group_leader);
3566 }
3567
3568 return 1;
3569
3570}
3571#else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3572static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3573 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3574{
3575 return;
3576}
3577
3578static inline void update_sd_power_savings_stats(struct sched_group *group,
3579 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3580{
3581 return;
3582}
3583
3584static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3585 int this_cpu, unsigned long *imbalance)
3586{
3587 return 0;
3588}
3589#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3590
3591
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303592/**
3593 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3594 * @group: sched_group whose statistics are to be updated.
3595 * @this_cpu: Cpu for which load balance is currently performed.
3596 * @idle: Idle status of this_cpu
3597 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3598 * @sd_idle: Idle status of the sched_domain containing group.
3599 * @local_group: Does group contain this_cpu.
3600 * @cpus: Set of cpus considered for load balancing.
3601 * @balance: Should we balance.
3602 * @sgs: variable to hold the statistics for this group.
3603 */
3604static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3605 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3606 int local_group, const struct cpumask *cpus,
3607 int *balance, struct sg_lb_stats *sgs)
3608{
3609 unsigned long load, max_cpu_load, min_cpu_load;
3610 int i;
3611 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3612 unsigned long sum_avg_load_per_task;
3613 unsigned long avg_load_per_task;
3614
3615 if (local_group)
3616 balance_cpu = group_first_cpu(group);
3617
3618 /* Tally up the load of all CPUs in the group */
3619 sum_avg_load_per_task = avg_load_per_task = 0;
3620 max_cpu_load = 0;
3621 min_cpu_load = ~0UL;
3622
3623 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3624 struct rq *rq = cpu_rq(i);
3625
3626 if (*sd_idle && rq->nr_running)
3627 *sd_idle = 0;
3628
3629 /* Bias balancing toward cpus of our domain */
3630 if (local_group) {
3631 if (idle_cpu(i) && !first_idle_cpu) {
3632 first_idle_cpu = 1;
3633 balance_cpu = i;
3634 }
3635
3636 load = target_load(i, load_idx);
3637 } else {
3638 load = source_load(i, load_idx);
3639 if (load > max_cpu_load)
3640 max_cpu_load = load;
3641 if (min_cpu_load > load)
3642 min_cpu_load = load;
3643 }
3644
3645 sgs->group_load += load;
3646 sgs->sum_nr_running += rq->nr_running;
3647 sgs->sum_weighted_load += weighted_cpuload(i);
3648
3649 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3650 }
3651
3652 /*
3653 * First idle cpu or the first cpu(busiest) in this sched group
3654 * is eligible for doing load balancing at this and above
3655 * domains. In the newly idle case, we will allow all the cpu's
3656 * to do the newly idle load balance.
3657 */
3658 if (idle != CPU_NEWLY_IDLE && local_group &&
3659 balance_cpu != this_cpu && balance) {
3660 *balance = 0;
3661 return;
3662 }
3663
3664 /* Adjust by relative CPU power of the group */
3665 sgs->avg_load = sg_div_cpu_power(group,
3666 sgs->group_load * SCHED_LOAD_SCALE);
3667
3668
3669 /*
3670 * Consider the group unbalanced when the imbalance is larger
3671 * than the average weight of two tasks.
3672 *
3673 * APZ: with cgroup the avg task weight can vary wildly and
3674 * might not be a suitable number - should we keep a
3675 * normalized nr_running number somewhere that negates
3676 * the hierarchy?
3677 */
3678 avg_load_per_task = sg_div_cpu_power(group,
3679 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3680
3681 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3682 sgs->group_imb = 1;
3683
3684 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3685
3686}
Linus Torvalds1da177e2005-04-16 15:20:36 -07003687
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303688/**
3689 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3690 * @sd: sched_domain whose statistics are to be updated.
3691 * @this_cpu: Cpu for which load balance is currently performed.
3692 * @idle: Idle status of this_cpu
3693 * @sd_idle: Idle status of the sched_domain containing group.
3694 * @cpus: Set of cpus considered for load balancing.
3695 * @balance: Should we balance.
3696 * @sds: variable to hold the statistics for this sched_domain.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003697 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303698static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3699 enum cpu_idle_type idle, int *sd_idle,
3700 const struct cpumask *cpus, int *balance,
3701 struct sd_lb_stats *sds)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003702{
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303703 struct sched_group *group = sd->groups;
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303704 struct sg_lb_stats sgs;
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303705 int load_idx;
3706
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303707 init_sd_power_savings_stats(sd, sds, idle);
Gautham R Shenoy67bb6c02009-03-25 14:43:35 +05303708 load_idx = get_sd_load_idx(sd, idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003709
3710 do {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003711 int local_group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003712
Rusty Russell758b2cd2008-11-25 02:35:04 +10303713 local_group = cpumask_test_cpu(this_cpu,
3714 sched_group_cpus(group));
Gautham R Shenoy381be782009-03-25 14:43:46 +05303715 memset(&sgs, 0, sizeof(sgs));
Gautham R Shenoy1f8c5532009-03-25 14:43:51 +05303716 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3717 local_group, cpus, balance, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003718
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303719 if (local_group && balance && !(*balance))
3720 return;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003721
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303722 sds->total_load += sgs.group_load;
3723 sds->total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003724
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725 if (local_group) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303726 sds->this_load = sgs.avg_load;
3727 sds->this = group;
3728 sds->this_nr_running = sgs.sum_nr_running;
3729 sds->this_load_per_task = sgs.sum_weighted_load;
3730 } else if (sgs.avg_load > sds->max_load &&
Gautham R Shenoy381be782009-03-25 14:43:46 +05303731 (sgs.sum_nr_running > sgs.group_capacity ||
3732 sgs.group_imb)) {
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303733 sds->max_load = sgs.avg_load;
3734 sds->busiest = group;
3735 sds->busiest_nr_running = sgs.sum_nr_running;
3736 sds->busiest_load_per_task = sgs.sum_weighted_load;
3737 sds->group_imb = sgs.group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003738 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003739
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303740 update_sd_power_savings_stats(group, sds, local_group, &sgs);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003741 group = group->next;
3742 } while (group != sd->groups);
3743
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303744}
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303745
3746/**
3747 * fix_small_imbalance - Calculate the minor imbalance that exists
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303748 * amongst the groups of a sched_domain, during
3749 * load balancing.
Gautham R Shenoy2e6f44a2009-03-25 14:44:06 +05303750 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3751 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3752 * @imbalance: Variable to store the imbalance.
3753 */
3754static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3755 int this_cpu, unsigned long *imbalance)
3756{
3757 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3758 unsigned int imbn = 2;
3759
3760 if (sds->this_nr_running) {
3761 sds->this_load_per_task /= sds->this_nr_running;
3762 if (sds->busiest_load_per_task >
3763 sds->this_load_per_task)
3764 imbn = 1;
3765 } else
3766 sds->this_load_per_task =
3767 cpu_avg_load_per_task(this_cpu);
3768
3769 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3770 sds->busiest_load_per_task * imbn) {
3771 *imbalance = sds->busiest_load_per_task;
3772 return;
3773 }
3774
3775 /*
3776 * OK, we don't have enough imbalance to justify moving tasks,
3777 * however we may be able to increase total CPU power used by
3778 * moving them.
3779 */
3780
3781 pwr_now += sds->busiest->__cpu_power *
3782 min(sds->busiest_load_per_task, sds->max_load);
3783 pwr_now += sds->this->__cpu_power *
3784 min(sds->this_load_per_task, sds->this_load);
3785 pwr_now /= SCHED_LOAD_SCALE;
3786
3787 /* Amount of load we'd subtract */
3788 tmp = sg_div_cpu_power(sds->busiest,
3789 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3790 if (sds->max_load > tmp)
3791 pwr_move += sds->busiest->__cpu_power *
3792 min(sds->busiest_load_per_task, sds->max_load - tmp);
3793
3794 /* Amount of load we'd add */
3795 if (sds->max_load * sds->busiest->__cpu_power <
3796 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3797 tmp = sg_div_cpu_power(sds->this,
3798 sds->max_load * sds->busiest->__cpu_power);
3799 else
3800 tmp = sg_div_cpu_power(sds->this,
3801 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3802 pwr_move += sds->this->__cpu_power *
3803 min(sds->this_load_per_task, sds->this_load + tmp);
3804 pwr_move /= SCHED_LOAD_SCALE;
3805
3806 /* Move if we gain throughput */
3807 if (pwr_move > pwr_now)
3808 *imbalance = sds->busiest_load_per_task;
3809}
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303810
3811/**
3812 * calculate_imbalance - Calculate the amount of imbalance present within the
3813 * groups of a given sched_domain during load balance.
3814 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3815 * @this_cpu: Cpu for which currently load balance is being performed.
3816 * @imbalance: The variable to store the imbalance.
3817 */
3818static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3819 unsigned long *imbalance)
3820{
3821 unsigned long max_pull;
3822 /*
3823 * In the presence of smp nice balancing, certain scenarios can have
3824 * max load less than avg load(as we skip the groups at or below
3825 * its cpu_power, while calculating max_load..)
3826 */
3827 if (sds->max_load < sds->avg_load) {
3828 *imbalance = 0;
3829 return fix_small_imbalance(sds, this_cpu, imbalance);
3830 }
3831
3832 /* Don't want to pull so many tasks that a group would go idle */
3833 max_pull = min(sds->max_load - sds->avg_load,
3834 sds->max_load - sds->busiest_load_per_task);
3835
3836 /* How much load to actually move to equalise the imbalance */
3837 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3838 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3839 / SCHED_LOAD_SCALE;
3840
3841 /*
3842 * if *imbalance is less than the average load per runnable task
3843 * there is no gaurantee that any tasks will be moved so we'll have
3844 * a think about bumping its value to force at least one task to be
3845 * moved
3846 */
3847 if (*imbalance < sds->busiest_load_per_task)
3848 return fix_small_imbalance(sds, this_cpu, imbalance);
3849
3850}
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303851/******* find_busiest_group() helpers end here *********************/
3852
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303853/**
3854 * find_busiest_group - Returns the busiest group within the sched_domain
3855 * if there is an imbalance. If there isn't an imbalance, and
3856 * the user has opted for power-savings, it returns a group whose
3857 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3858 * such a group exists.
3859 *
3860 * Also calculates the amount of weighted load which should be moved
3861 * to restore balance.
3862 *
3863 * @sd: The sched_domain whose busiest group is to be returned.
3864 * @this_cpu: The cpu for which load balancing is currently being performed.
3865 * @imbalance: Variable which stores amount of weighted load which should
3866 * be moved to restore balance/put a group to idle.
3867 * @idle: The idle status of this_cpu.
3868 * @sd_idle: The idleness of sd
3869 * @cpus: The set of CPUs under consideration for load-balancing.
3870 * @balance: Pointer to a variable indicating if this_cpu
3871 * is the appropriate cpu to perform load balancing at this_level.
3872 *
3873 * Returns: - the busiest group if imbalance exists.
3874 * - If no imbalance and user has opted for power-savings balance,
3875 * return the least loaded group whose CPUs can be
3876 * put to idle by rebalancing its tasks onto our group.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003877 */
3878static struct sched_group *
3879find_busiest_group(struct sched_domain *sd, int this_cpu,
3880 unsigned long *imbalance, enum cpu_idle_type idle,
3881 int *sd_idle, const struct cpumask *cpus, int *balance)
3882{
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303883 struct sd_lb_stats sds;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003884
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303885 memset(&sds, 0, sizeof(sds));
Linus Torvalds1da177e2005-04-16 15:20:36 -07003886
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303887 /*
3888 * Compute the various statistics relavent for load balancing at
3889 * this level.
3890 */
3891 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3892 balance, &sds);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003893
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303894 /* Cases where imbalance does not exist from POV of this_cpu */
3895 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3896 * at this level.
3897 * 2) There is no busy sibling group to pull from.
3898 * 3) This group is the busiest group.
3899 * 4) This group is more busy than the avg busieness at this
3900 * sched_domain.
3901 * 5) The imbalance is within the specified limit.
3902 * 6) Any rebalance would lead to ping-pong
3903 */
Gautham R Shenoy37abe192009-03-25 14:44:01 +05303904 if (balance && !(*balance))
3905 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003906
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303907 if (!sds.busiest || sds.busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003908 goto out_balanced;
3909
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303910 if (sds.this_load >= sds.max_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003911 goto out_balanced;
3912
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303913 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003914
Gautham R Shenoyb7bb4c92009-03-25 14:44:27 +05303915 if (sds.this_load >= sds.avg_load)
3916 goto out_balanced;
3917
3918 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003919 goto out_balanced;
3920
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303921 sds.busiest_load_per_task /= sds.busiest_nr_running;
3922 if (sds.group_imb)
3923 sds.busiest_load_per_task =
3924 min(sds.busiest_load_per_task, sds.avg_load);
Ken Chen908a7c12007-10-17 16:55:11 +02003925
Linus Torvalds1da177e2005-04-16 15:20:36 -07003926 /*
3927 * We're trying to get all the cpus to the average_load, so we don't
3928 * want to push ourselves above the average load, nor do we wish to
3929 * reduce the max loaded cpu below the average load, as either of these
3930 * actions would just result in more rebalancing later, and ping-pong
3931 * tasks around. Thus we look for the minimum possible imbalance.
3932 * Negative imbalances (*we* are more loaded than anyone else) will
3933 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003934 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003935 * appear as very large values with unsigned longs.
3936 */
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303937 if (sds.max_load <= sds.busiest_load_per_task)
Peter Williams2dd73a42006-06-27 02:54:34 -07003938 goto out_balanced;
3939
Gautham R Shenoydbc523a2009-03-25 14:44:12 +05303940 /* Looks like there is an imbalance. Compute it */
3941 calculate_imbalance(&sds, this_cpu, imbalance);
Gautham R Shenoy222d6562009-03-25 14:43:56 +05303942 return sds.busiest;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003943
3944out_balanced:
Gautham R Shenoyc071df12009-03-25 14:44:22 +05303945 /*
3946 * There is no obvious imbalance. But check if we can do some balancing
3947 * to save power.
3948 */
3949 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3950 return sds.busiest;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003951ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003952 *imbalance = 0;
3953 return NULL;
3954}
3955
3956/*
3957 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3958 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003959static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003960find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Rusty Russell96f874e2008-11-25 02:35:14 +10303961 unsigned long imbalance, const struct cpumask *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003962{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003963 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07003964 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003965 int i;
3966
Rusty Russell758b2cd2008-11-25 02:35:04 +10303967 for_each_cpu(i, sched_group_cpus(group)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003968 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003969
Rusty Russell96f874e2008-11-25 02:35:14 +10303970 if (!cpumask_test_cpu(i, cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003971 continue;
3972
Ingo Molnar48f24c42006-07-03 00:25:40 -07003973 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02003974 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003975
Ingo Molnardd41f592007-07-09 18:51:59 +02003976 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07003977 continue;
3978
Ingo Molnardd41f592007-07-09 18:51:59 +02003979 if (wl > max_load) {
3980 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003981 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003982 }
3983 }
3984
3985 return busiest;
3986}
3987
3988/*
Nick Piggin77391d72005-06-25 14:57:30 -07003989 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3990 * so long as it is large enough.
3991 */
3992#define MAX_PINNED_INTERVAL 512
3993
Rusty Russelldf7c8e82009-03-19 15:22:20 +10303994/* Working cpumask for load_balance and load_balance_newidle. */
3995static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3996
Nick Piggin77391d72005-06-25 14:57:30 -07003997/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003998 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3999 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004000 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004001static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004002 struct sched_domain *sd, enum cpu_idle_type idle,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304003 int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004004{
Peter Williams43010652007-08-09 11:16:46 +02004005 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004006 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004007 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004008 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004009 unsigned long flags;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304010 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Nick Piggin5969fe02005-09-10 00:26:19 -07004011
Rusty Russell96f874e2008-11-25 02:35:14 +10304012 cpumask_setall(cpus);
Mike Travis7c16ec52008-04-04 18:11:11 -07004013
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004014 /*
4015 * When power savings policy is enabled for the parent domain, idle
4016 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02004017 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004018 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004019 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004020 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004021 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004022 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004023
Ingo Molnar2d723762007-10-15 17:00:12 +02004024 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004025
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004026redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004027 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004028 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07004029 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004030
Chen, Kenneth W06066712006-12-10 02:20:35 -08004031 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004032 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004033
Linus Torvalds1da177e2005-04-16 15:20:36 -07004034 if (!group) {
4035 schedstat_inc(sd, lb_nobusyg[idle]);
4036 goto out_balanced;
4037 }
4038
Mike Travis7c16ec52008-04-04 18:11:11 -07004039 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040 if (!busiest) {
4041 schedstat_inc(sd, lb_nobusyq[idle]);
4042 goto out_balanced;
4043 }
4044
Nick Piggindb935db2005-06-25 14:57:11 -07004045 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004046
4047 schedstat_add(sd, lb_imbalance[idle], imbalance);
4048
Peter Williams43010652007-08-09 11:16:46 +02004049 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004050 if (busiest->nr_running > 1) {
4051 /*
4052 * Attempt to move tasks. If find_busiest_group has found
4053 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02004054 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07004055 * correctly treated as an imbalance.
4056 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004057 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07004058 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02004059 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07004060 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07004061 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004062 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07004063
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004064 /*
4065 * some other cpu did the load balance for us.
4066 */
Peter Williams43010652007-08-09 11:16:46 +02004067 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004068 resched_cpu(this_cpu);
4069
Nick Piggin81026792005-06-25 14:57:07 -07004070 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004071 if (unlikely(all_pinned)) {
Rusty Russell96f874e2008-11-25 02:35:14 +10304072 cpumask_clear_cpu(cpu_of(busiest), cpus);
4073 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004074 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07004075 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004076 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004077 }
Nick Piggin81026792005-06-25 14:57:07 -07004078
Peter Williams43010652007-08-09 11:16:46 +02004079 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080 schedstat_inc(sd, lb_failed[idle]);
4081 sd->nr_balance_failed++;
4082
4083 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004084
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004085 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004086
4087 /* don't kick the migration_thread, if the curr
4088 * task on busiest cpu can't be moved to this_cpu
4089 */
Rusty Russell96f874e2008-11-25 02:35:14 +10304090 if (!cpumask_test_cpu(this_cpu,
4091 &busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004092 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004093 all_pinned = 1;
4094 goto out_one_pinned;
4095 }
4096
Linus Torvalds1da177e2005-04-16 15:20:36 -07004097 if (!busiest->active_balance) {
4098 busiest->active_balance = 1;
4099 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07004100 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004101 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08004102 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07004103 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004104 wake_up_process(busiest->migration_thread);
4105
4106 /*
4107 * We've kicked active balancing, reset the failure
4108 * counter.
4109 */
Nick Piggin39507452005-06-25 14:57:09 -07004110 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004111 }
Nick Piggin81026792005-06-25 14:57:07 -07004112 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004113 sd->nr_balance_failed = 0;
4114
Nick Piggin81026792005-06-25 14:57:07 -07004115 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004116 /* We were unbalanced, so reset the balancing interval */
4117 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07004118 } else {
4119 /*
4120 * If we've begun active balancing, start to back off. This
4121 * case may not be covered by the all_pinned logic if there
4122 * is only 1 task on the busy runqueue (because we don't call
4123 * move_tasks).
4124 */
4125 if (sd->balance_interval < sd->max_interval)
4126 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004127 }
4128
Peter Williams43010652007-08-09 11:16:46 +02004129 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004130 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004131 ld_moved = -1;
4132
4133 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004134
4135out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004136 schedstat_inc(sd, lb_balanced[idle]);
4137
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004138 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004139
4140out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004141 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07004142 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4143 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004144 sd->balance_interval *= 2;
4145
Ingo Molnar48f24c42006-07-03 00:25:40 -07004146 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004147 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004148 ld_moved = -1;
4149 else
4150 ld_moved = 0;
4151out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02004152 if (ld_moved)
4153 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004154 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004155}
4156
4157/*
4158 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4159 * tasks if there is an imbalance.
4160 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004161 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162 * this_rq is locked.
4163 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07004164static int
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304165load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004166{
4167 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004168 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004169 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02004170 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07004171 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004172 int all_pinned = 0;
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304173 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
Mike Travis7c16ec52008-04-04 18:11:11 -07004174
Rusty Russell96f874e2008-11-25 02:35:14 +10304175 cpumask_setall(cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07004176
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004177 /*
4178 * When power savings policy is enabled for the parent domain, idle
4179 * sibling can pick up load irrespective of busy siblings. In this case,
4180 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004181 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004182 */
4183 if (sd->flags & SD_SHARE_CPUPOWER &&
4184 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004185 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004186
Ingo Molnar2d723762007-10-15 17:00:12 +02004187 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004188redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004189 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004190 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07004191 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004192 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004193 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004194 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004195 }
4196
Mike Travis7c16ec52008-04-04 18:11:11 -07004197 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07004198 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004199 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004200 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004201 }
4202
Nick Piggindb935db2005-06-25 14:57:11 -07004203 BUG_ON(busiest == this_rq);
4204
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004205 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004206
Peter Williams43010652007-08-09 11:16:46 +02004207 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004208 if (busiest->nr_running > 1) {
4209 /* Attempt to move tasks */
4210 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004211 /* this_rq->clock is already updated */
4212 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02004213 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004214 imbalance, sd, CPU_NEWLY_IDLE,
4215 &all_pinned);
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004216 double_unlock_balance(this_rq, busiest);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004217
Suresh Siddha969bb4e2007-07-19 21:28:35 +02004218 if (unlikely(all_pinned)) {
Rusty Russell96f874e2008-11-25 02:35:14 +10304219 cpumask_clear_cpu(cpu_of(busiest), cpus);
4220 if (!cpumask_empty(cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07004221 goto redo;
4222 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07004223 }
4224
Peter Williams43010652007-08-09 11:16:46 +02004225 if (!ld_moved) {
Vaidyanathan Srinivasan36dffab2008-12-20 10:06:38 +05304226 int active_balance = 0;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304227
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004228 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004229 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4230 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004231 return -1;
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304232
4233 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4234 return -1;
4235
4236 if (sd->nr_balance_failed++ < 2)
4237 return -1;
4238
4239 /*
4240 * The only task running in a non-idle cpu can be moved to this
4241 * cpu in an attempt to completely freeup the other CPU
4242 * package. The same method used to move task in load_balance()
4243 * have been extended for load_balance_newidle() to speedup
4244 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4245 *
4246 * The package power saving logic comes from
4247 * find_busiest_group(). If there are no imbalance, then
4248 * f_b_g() will return NULL. However when sched_mc={1,2} then
4249 * f_b_g() will select a group from which a running task may be
4250 * pulled to this cpu in order to make the other package idle.
4251 * If there is no opportunity to make a package idle and if
4252 * there are no imbalance, then f_b_g() will return NULL and no
4253 * action will be taken in load_balance_newidle().
4254 *
4255 * Under normal task pull operation due to imbalance, there
4256 * will be more than one task in the source run queue and
4257 * move_tasks() will succeed. ld_moved will be true and this
4258 * active balance code will not be triggered.
4259 */
4260
4261 /* Lock busiest in correct order while this_rq is held */
4262 double_lock_balance(this_rq, busiest);
4263
4264 /*
4265 * don't kick the migration_thread, if the curr
4266 * task on busiest cpu can't be moved to this_cpu
4267 */
Mike Travis6ca09df2008-12-31 18:08:45 -08004268 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304269 double_unlock_balance(this_rq, busiest);
4270 all_pinned = 1;
4271 return ld_moved;
4272 }
4273
4274 if (!busiest->active_balance) {
4275 busiest->active_balance = 1;
4276 busiest->push_cpu = this_cpu;
4277 active_balance = 1;
4278 }
4279
4280 double_unlock_balance(this_rq, busiest);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004281 /*
4282 * Should not call ttwu while holding a rq->lock
4283 */
4284 spin_unlock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304285 if (active_balance)
4286 wake_up_process(busiest->migration_thread);
Peter Zijlstrada8d5082009-01-07 15:28:57 +01004287 spin_lock(&this_rq->lock);
Vaidyanathan Srinivasanad273b32008-12-18 23:26:36 +05304288
Nick Piggin5969fe02005-09-10 00:26:19 -07004289 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004290 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004291
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02004292 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02004293 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004294
4295out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004296 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07004297 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07004298 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07004299 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004300 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004301
Nick Piggin16cfb1c2005-06-25 14:57:08 -07004302 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004303}
4304
4305/*
4306 * idle_balance is called by schedule() if this_cpu is about to become
4307 * idle. Attempts to pull tasks from other CPUs.
4308 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004309static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310{
4311 struct sched_domain *sd;
Vaidyanathan Srinivasanefbe0272008-12-08 20:52:49 +05304312 int pulled_task = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02004313 unsigned long next_balance = jiffies + HZ;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004314
4315 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004316 unsigned long interval;
4317
4318 if (!(sd->flags & SD_LOAD_BALANCE))
4319 continue;
4320
4321 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07004322 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07004323 pulled_task = load_balance_newidle(this_cpu, this_rq,
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304324 sd);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07004325
4326 interval = msecs_to_jiffies(sd->balance_interval);
4327 if (time_after(next_balance, sd->last_balance + interval))
4328 next_balance = sd->last_balance + interval;
4329 if (pulled_task)
4330 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004332 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004333 /*
4334 * We are going idle. next_balance may be set based on
4335 * a busy processor. So reset next_balance.
4336 */
4337 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02004338 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004339}
4340
4341/*
4342 * active_load_balance is run by migration threads. It pushes running tasks
4343 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4344 * running on each physical CPU where possible, and avoids physical /
4345 * logical imbalances.
4346 *
4347 * Called with busiest_rq locked.
4348 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004349static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004350{
Nick Piggin39507452005-06-25 14:57:09 -07004351 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004352 struct sched_domain *sd;
4353 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07004354
Ingo Molnar48f24c42006-07-03 00:25:40 -07004355 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07004356 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07004357 return;
4358
4359 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004360
4361 /*
Nick Piggin39507452005-06-25 14:57:09 -07004362 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004363 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07004364 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004365 */
Nick Piggin39507452005-06-25 14:57:09 -07004366 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367
Nick Piggin39507452005-06-25 14:57:09 -07004368 /* move a task from busiest_rq to target_rq */
4369 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02004370 update_rq_clock(busiest_rq);
4371 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004372
Nick Piggin39507452005-06-25 14:57:09 -07004373 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004374 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07004375 if ((sd->flags & SD_LOAD_BALANCE) &&
Rusty Russell758b2cd2008-11-25 02:35:04 +10304376 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
Nick Piggin39507452005-06-25 14:57:09 -07004377 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07004378 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07004379
Ingo Molnar48f24c42006-07-03 00:25:40 -07004380 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004381 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004382
Peter Williams43010652007-08-09 11:16:46 +02004383 if (move_one_task(target_rq, target_cpu, busiest_rq,
4384 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07004385 schedstat_inc(sd, alb_pushed);
4386 else
4387 schedstat_inc(sd, alb_failed);
4388 }
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02004389 double_unlock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004390}
4391
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004392#ifdef CONFIG_NO_HZ
4393static struct {
4394 atomic_t load_balancer;
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304395 cpumask_var_t cpu_mask;
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304396 cpumask_var_t ilb_grp_nohz_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004397} nohz ____cacheline_aligned = {
4398 .load_balancer = ATOMIC_INIT(-1),
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004399};
4400
Arun R Bharadwajeea08f32009-04-16 12:16:41 +05304401int get_nohz_load_balancer(void)
4402{
4403 return atomic_read(&nohz.load_balancer);
4404}
4405
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304406#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4407/**
4408 * lowest_flag_domain - Return lowest sched_domain containing flag.
4409 * @cpu: The cpu whose lowest level of sched domain is to
4410 * be returned.
4411 * @flag: The flag to check for the lowest sched_domain
4412 * for the given cpu.
4413 *
4414 * Returns the lowest sched_domain of a cpu which contains the given flag.
4415 */
4416static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4417{
4418 struct sched_domain *sd;
4419
4420 for_each_domain(cpu, sd)
4421 if (sd && (sd->flags & flag))
4422 break;
4423
4424 return sd;
4425}
4426
4427/**
4428 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4429 * @cpu: The cpu whose domains we're iterating over.
4430 * @sd: variable holding the value of the power_savings_sd
4431 * for cpu.
4432 * @flag: The flag to filter the sched_domains to be iterated.
4433 *
4434 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4435 * set, starting from the lowest sched_domain to the highest.
4436 */
4437#define for_each_flag_domain(cpu, sd, flag) \
4438 for (sd = lowest_flag_domain(cpu, flag); \
4439 (sd && (sd->flags & flag)); sd = sd->parent)
4440
4441/**
4442 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4443 * @ilb_group: group to be checked for semi-idleness
4444 *
4445 * Returns: 1 if the group is semi-idle. 0 otherwise.
4446 *
4447 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4448 * and atleast one non-idle CPU. This helper function checks if the given
4449 * sched_group is semi-idle or not.
4450 */
4451static inline int is_semi_idle_group(struct sched_group *ilb_group)
4452{
4453 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4454 sched_group_cpus(ilb_group));
4455
4456 /*
4457 * A sched_group is semi-idle when it has atleast one busy cpu
4458 * and atleast one idle cpu.
4459 */
4460 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4461 return 0;
4462
4463 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4464 return 0;
4465
4466 return 1;
4467}
4468/**
4469 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4470 * @cpu: The cpu which is nominating a new idle_load_balancer.
4471 *
4472 * Returns: Returns the id of the idle load balancer if it exists,
4473 * Else, returns >= nr_cpu_ids.
4474 *
4475 * This algorithm picks the idle load balancer such that it belongs to a
4476 * semi-idle powersavings sched_domain. The idea is to try and avoid
4477 * completely idle packages/cores just for the purpose of idle load balancing
4478 * when there are other idle cpu's which are better suited for that job.
4479 */
4480static int find_new_ilb(int cpu)
4481{
4482 struct sched_domain *sd;
4483 struct sched_group *ilb_group;
4484
4485 /*
4486 * Have idle load balancer selection from semi-idle packages only
4487 * when power-aware load balancing is enabled
4488 */
4489 if (!(sched_smt_power_savings || sched_mc_power_savings))
4490 goto out_done;
4491
4492 /*
4493 * Optimize for the case when we have no idle CPUs or only one
4494 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4495 */
4496 if (cpumask_weight(nohz.cpu_mask) < 2)
4497 goto out_done;
4498
4499 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4500 ilb_group = sd->groups;
4501
4502 do {
4503 if (is_semi_idle_group(ilb_group))
4504 return cpumask_first(nohz.ilb_grp_nohz_mask);
4505
4506 ilb_group = ilb_group->next;
4507
4508 } while (ilb_group != sd->groups);
4509 }
4510
4511out_done:
4512 return cpumask_first(nohz.cpu_mask);
4513}
4514#else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4515static inline int find_new_ilb(int call_cpu)
4516{
Gautham R Shenoy6e29ec52009-04-21 08:40:49 +05304517 return cpumask_first(nohz.cpu_mask);
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304518}
4519#endif
4520
Christoph Lameter7835b982006-12-10 02:20:22 -08004521/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004522 * This routine will try to nominate the ilb (idle load balancing)
4523 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4524 * load balancing on behalf of all those cpus. If all the cpus in the system
4525 * go into this tickless mode, then there will be no ilb owner (as there is
4526 * no need for one) and all the cpus will sleep till the next wakeup event
4527 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08004528 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004529 * For the ilb owner, tick is not stopped. And this tick will be used
4530 * for idle load balancing. ilb owner will still be part of
4531 * nohz.cpu_mask..
4532 *
4533 * While stopping the tick, this cpu will become the ilb owner if there
4534 * is no other owner. And will be the owner till that cpu becomes busy
4535 * or if all cpus in the system stop their ticks at which point
4536 * there is no need for ilb owner.
4537 *
4538 * When the ilb owner becomes busy, it nominates another owner, during the
4539 * next busy scheduler_tick()
4540 */
4541int select_nohz_load_balancer(int stop_tick)
4542{
4543 int cpu = smp_processor_id();
4544
4545 if (stop_tick) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004546 cpu_rq(cpu)->in_nohz_recently = 1;
4547
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004548 if (!cpu_active(cpu)) {
4549 if (atomic_read(&nohz.load_balancer) != cpu)
4550 return 0;
4551
4552 /*
4553 * If we are going offline and still the leader,
4554 * give up!
4555 */
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004556 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4557 BUG();
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004558
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004559 return 0;
4560 }
4561
Suresh Siddha483b4ee2009-02-04 11:59:44 -08004562 cpumask_set_cpu(cpu, nohz.cpu_mask);
4563
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004564 /* time for ilb owner also to sleep */
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304565 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004566 if (atomic_read(&nohz.load_balancer) == cpu)
4567 atomic_set(&nohz.load_balancer, -1);
4568 return 0;
4569 }
4570
4571 if (atomic_read(&nohz.load_balancer) == -1) {
4572 /* make me the ilb owner */
4573 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4574 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304575 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4576 int new_ilb;
4577
4578 if (!(sched_smt_power_savings ||
4579 sched_mc_power_savings))
4580 return 1;
4581 /*
4582 * Check to see if there is a more power-efficient
4583 * ilb.
4584 */
4585 new_ilb = find_new_ilb(cpu);
4586 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4587 atomic_set(&nohz.load_balancer, -1);
4588 resched_cpu(new_ilb);
4589 return 0;
4590 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004591 return 1;
Gautham R Shenoye790fb02009-04-14 10:25:35 +05304592 }
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004593 } else {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304594 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004595 return 0;
4596
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304597 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004598
4599 if (atomic_read(&nohz.load_balancer) == cpu)
4600 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4601 BUG();
4602 }
4603 return 0;
4604}
4605#endif
4606
4607static DEFINE_SPINLOCK(balancing);
4608
4609/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004610 * It checks each scheduling domain to see if it is due to be balanced,
4611 * and initiates a balancing operation if so.
4612 *
4613 * Balancing parameters are set up in arch_init_sched_domains.
4614 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02004615static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08004616{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004617 int balance = 1;
4618 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08004619 unsigned long interval;
4620 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004621 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08004622 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004623 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004624 int need_serialize;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004625
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004626 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004627 if (!(sd->flags & SD_LOAD_BALANCE))
4628 continue;
4629
4630 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004631 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004632 interval *= sd->busy_factor;
4633
4634 /* scale ms to jiffies */
4635 interval = msecs_to_jiffies(interval);
4636 if (unlikely(!interval))
4637 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02004638 if (interval > HZ*NR_CPUS/10)
4639 interval = HZ*NR_CPUS/10;
4640
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004641 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004642
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004643 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08004644 if (!spin_trylock(&balancing))
4645 goto out;
4646 }
4647
Christoph Lameterc9819f42006-12-10 02:20:25 -08004648 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Rusty Russelldf7c8e82009-03-19 15:22:20 +10304649 if (load_balance(cpu, rq, sd, idle, &balance)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07004650 /*
4651 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07004652 * longer idle, or one of our SMT siblings is
4653 * not idle.
4654 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02004655 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004656 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08004657 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004658 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02004659 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08004660 spin_unlock(&balancing);
4661out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02004662 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08004663 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02004664 update_next_balance = 1;
4665 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08004666
4667 /*
4668 * Stop the load balance at this level. There is another
4669 * CPU in our sched group which is doing load balancing more
4670 * actively.
4671 */
4672 if (!balance)
4673 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004674 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02004675
4676 /*
4677 * next_balance will be updated only when there is a need.
4678 * When the cpu is attached to null domain for ex, it will not be
4679 * updated.
4680 */
4681 if (likely(update_next_balance))
4682 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004683}
4684
4685/*
4686 * run_rebalance_domains is triggered when needed from the scheduler tick.
4687 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4688 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4689 */
4690static void run_rebalance_domains(struct softirq_action *h)
4691{
Ingo Molnardd41f592007-07-09 18:51:59 +02004692 int this_cpu = smp_processor_id();
4693 struct rq *this_rq = cpu_rq(this_cpu);
4694 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4695 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004696
Ingo Molnardd41f592007-07-09 18:51:59 +02004697 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004698
4699#ifdef CONFIG_NO_HZ
4700 /*
4701 * If this cpu is the owner for idle load balancing, then do the
4702 * balancing on behalf of the other idle cpus whose ticks are
4703 * stopped.
4704 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004705 if (this_rq->idle_at_tick &&
4706 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004707 struct rq *rq;
4708 int balance_cpu;
4709
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304710 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4711 if (balance_cpu == this_cpu)
4712 continue;
4713
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004714 /*
4715 * If this cpu gets work to do, stop the load balancing
4716 * work being done for other cpus. Next load
4717 * balancing owner will pick it up.
4718 */
4719 if (need_resched())
4720 break;
4721
Oleg Nesterovde0cf892007-08-12 18:08:19 +02004722 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004723
4724 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004725 if (time_after(this_rq->next_balance, rq->next_balance))
4726 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004727 }
4728 }
4729#endif
4730}
4731
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004732static inline int on_null_domain(int cpu)
4733{
4734 return !rcu_dereference(cpu_rq(cpu)->sd);
4735}
4736
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004737/*
4738 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4739 *
4740 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4741 * idle load balancing owner or decide to stop the periodic load balancing,
4742 * if the whole system is idle.
4743 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004744static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004745{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004746#ifdef CONFIG_NO_HZ
4747 /*
4748 * If we were in the nohz mode recently and busy at the current
4749 * scheduler tick, then check if we need to nominate new idle
4750 * load balancer.
4751 */
4752 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4753 rq->in_nohz_recently = 0;
4754
4755 if (atomic_read(&nohz.load_balancer) == cpu) {
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304756 cpumask_clear_cpu(cpu, nohz.cpu_mask);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004757 atomic_set(&nohz.load_balancer, -1);
4758 }
4759
4760 if (atomic_read(&nohz.load_balancer) == -1) {
Gautham R Shenoyf711f602009-04-14 10:25:30 +05304761 int ilb = find_new_ilb(cpu);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004762
Mike Travis434d53b2008-04-04 18:11:04 -07004763 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004764 resched_cpu(ilb);
4765 }
4766 }
4767
4768 /*
4769 * If this cpu is idle and doing idle load balancing for all the
4770 * cpus with ticks stopped, is it time for that to stop?
4771 */
4772 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304773 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004774 resched_cpu(cpu);
4775 return;
4776 }
4777
4778 /*
4779 * If this cpu is idle and the idle load balancing is done by
4780 * someone else, then no need raise the SCHED_SOFTIRQ
4781 */
4782 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
Rusty Russell7d1e6a92008-11-25 02:35:09 +10304783 cpumask_test_cpu(cpu, nohz.cpu_mask))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004784 return;
4785#endif
Frederic Weisbecker8a0be9e2009-03-05 01:27:02 +01004786 /* Don't need to rebalance while attached to NULL domain */
4787 if (time_after_eq(jiffies, rq->next_balance) &&
4788 likely(!on_null_domain(cpu)))
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004789 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004790}
Ingo Molnardd41f592007-07-09 18:51:59 +02004791
4792#else /* CONFIG_SMP */
4793
Linus Torvalds1da177e2005-04-16 15:20:36 -07004794/*
4795 * on UP we do not need to balance between CPUs:
4796 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004797static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004798{
4799}
Ingo Molnardd41f592007-07-09 18:51:59 +02004800
Linus Torvalds1da177e2005-04-16 15:20:36 -07004801#endif
4802
Linus Torvalds1da177e2005-04-16 15:20:36 -07004803DEFINE_PER_CPU(struct kernel_stat, kstat);
4804
4805EXPORT_PER_CPU_SYMBOL(kstat);
4806
4807/*
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004808 * Return any ns on the sched_clock that have not yet been accounted in
Frank Mayharf06febc2008-09-12 09:54:39 -07004809 * @p in case that task is currently running.
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004810 *
4811 * Called with task_rq_lock() held on @rq.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004812 */
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004813static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4814{
4815 u64 ns = 0;
4816
4817 if (task_current(rq, p)) {
4818 update_rq_clock(rq);
4819 ns = rq->clock - p->se.exec_start;
4820 if ((s64)ns < 0)
4821 ns = 0;
4822 }
4823
4824 return ns;
4825}
4826
Frank Mayharbb34d922008-09-12 09:54:39 -07004827unsigned long long task_delta_exec(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004828{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004829 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004830 struct rq *rq;
Frank Mayharbb34d922008-09-12 09:54:39 -07004831 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004832
Ingo Molnar41b86e92007-07-09 18:51:58 +02004833 rq = task_rq_lock(p, &flags);
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004834 ns = do_task_delta_exec(p, rq);
4835 task_rq_unlock(rq, &flags);
Ingo Molnar15084872008-09-30 08:28:17 +02004836
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004837 return ns;
4838}
Frank Mayharf06febc2008-09-12 09:54:39 -07004839
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004840/*
4841 * Return accounted runtime for the task.
4842 * In case the task is currently running, return the runtime plus current's
4843 * pending runtime that have not been accounted yet.
4844 */
4845unsigned long long task_sched_runtime(struct task_struct *p)
4846{
4847 unsigned long flags;
4848 struct rq *rq;
4849 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004850
Hidetoshi Setoc5f8d992009-03-31 16:56:03 +09004851 rq = task_rq_lock(p, &flags);
4852 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4853 task_rq_unlock(rq, &flags);
4854
4855 return ns;
4856}
4857
4858/*
4859 * Return sum_exec_runtime for the thread group.
4860 * In case the task is currently running, return the sum plus current's
4861 * pending runtime that have not been accounted yet.
4862 *
4863 * Note that the thread group might have other running tasks as well,
4864 * so the return value not includes other pending runtime that other
4865 * running tasks might have.
4866 */
4867unsigned long long thread_group_sched_runtime(struct task_struct *p)
4868{
4869 struct task_cputime totals;
4870 unsigned long flags;
4871 struct rq *rq;
4872 u64 ns;
4873
4874 rq = task_rq_lock(p, &flags);
4875 thread_group_cputime(p, &totals);
4876 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004877 task_rq_unlock(rq, &flags);
4878
4879 return ns;
4880}
4881
4882/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004883 * Account user cpu time to a process.
4884 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004885 * @cputime: the cpu time spent in user space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004886 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07004887 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004888void account_user_time(struct task_struct *p, cputime_t cputime,
4889 cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004890{
4891 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4892 cputime64_t tmp;
4893
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004894 /* Add user time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004895 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004896 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07004897 account_group_user_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004898
4899 /* Add user time to cpustat. */
4900 tmp = cputime_to_cputime64(cputime);
4901 if (TASK_NICE(p) > 0)
4902 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4903 else
4904 cpustat->user = cputime64_add(cpustat->user, tmp);
Bharata B Raoef12fef2009-03-31 10:02:22 +05304905
4906 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
Jonathan Lim49b5cf32008-07-25 01:48:40 -07004907 /* Account for user time used */
4908 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004909}
4910
4911/*
Laurent Vivier94886b82007-10-15 17:00:19 +02004912 * Account guest cpu time to a process.
4913 * @p: the process that the cpu time gets accounted to
4914 * @cputime: the cpu time spent in virtual machine since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004915 * @cputime_scaled: cputime scaled by cpu frequency
Laurent Vivier94886b82007-10-15 17:00:19 +02004916 */
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004917static void account_guest_time(struct task_struct *p, cputime_t cputime,
4918 cputime_t cputime_scaled)
Laurent Vivier94886b82007-10-15 17:00:19 +02004919{
4920 cputime64_t tmp;
4921 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4922
4923 tmp = cputime_to_cputime64(cputime);
4924
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004925 /* Add guest time to process. */
Laurent Vivier94886b82007-10-15 17:00:19 +02004926 p->utime = cputime_add(p->utime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004927 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07004928 account_group_user_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02004929 p->gtime = cputime_add(p->gtime, cputime);
4930
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004931 /* Add guest time to cpustat. */
Laurent Vivier94886b82007-10-15 17:00:19 +02004932 cpustat->user = cputime64_add(cpustat->user, tmp);
4933 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4934}
4935
4936/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004937 * Account system cpu time to a process.
4938 * @p: the process that the cpu time gets accounted to
4939 * @hardirq_offset: the offset to subtract from hardirq_count()
4940 * @cputime: the cpu time spent in kernel space since the last update
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004941 * @cputime_scaled: cputime scaled by cpu frequency
Linus Torvalds1da177e2005-04-16 15:20:36 -07004942 */
4943void account_system_time(struct task_struct *p, int hardirq_offset,
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004944 cputime_t cputime, cputime_t cputime_scaled)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004945{
4946 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004947 cputime64_t tmp;
4948
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004949 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004950 account_guest_time(p, cputime, cputime_scaled);
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004951 return;
4952 }
Laurent Vivier94886b82007-10-15 17:00:19 +02004953
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004954 /* Add system time to process. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004955 p->stime = cputime_add(p->stime, cputime);
Martin Schwidefsky457533a2008-12-31 15:11:37 +01004956 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
Frank Mayharf06febc2008-09-12 09:54:39 -07004957 account_group_system_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004958
4959 /* Add system time to cpustat. */
4960 tmp = cputime_to_cputime64(cputime);
4961 if (hardirq_count() - hardirq_offset)
4962 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4963 else if (softirq_count())
4964 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004965 else
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004966 cpustat->system = cputime64_add(cpustat->system, tmp);
4967
Bharata B Raoef12fef2009-03-31 10:02:22 +05304968 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4969
Linus Torvalds1da177e2005-04-16 15:20:36 -07004970 /* Account for system time used */
4971 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004972}
4973
4974/*
4975 * Account for involuntary wait time.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004976 * @steal: the cpu time spent in involuntary wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07004977 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004978void account_steal_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004979{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004980 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004981 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4982
4983 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004984}
4985
Christoph Lameter7835b982006-12-10 02:20:22 -08004986/*
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004987 * Account for idle time.
4988 * @cputime: the cpu time spent in idle wait
Linus Torvalds1da177e2005-04-16 15:20:36 -07004989 */
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004990void account_idle_time(cputime_t cputime)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004991{
4992 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004993 cputime64_t cputime64 = cputime_to_cputime64(cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004994 struct rq *rq = this_rq();
4995
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01004996 if (atomic_read(&rq->nr_iowait) > 0)
4997 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4998 else
4999 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
Christoph Lameter7835b982006-12-10 02:20:22 -08005000}
5001
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005002#ifndef CONFIG_VIRT_CPU_ACCOUNTING
5003
5004/*
5005 * Account a single tick of cpu time.
5006 * @p: the process that the cpu time gets accounted to
5007 * @user_tick: indicates if the tick is a user or a system tick
5008 */
5009void account_process_tick(struct task_struct *p, int user_tick)
5010{
5011 cputime_t one_jiffy = jiffies_to_cputime(1);
5012 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5013 struct rq *rq = this_rq();
5014
5015 if (user_tick)
5016 account_user_time(p, one_jiffy, one_jiffy_scaled);
Eric Dumazetf5f293a2009-04-29 14:44:49 +02005017 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
Martin Schwidefsky79741dd2008-12-31 15:11:38 +01005018 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5019 one_jiffy_scaled);
5020 else
5021 account_idle_time(one_jiffy);
5022}
5023
5024/*
5025 * Account multiple ticks of steal time.
5026 * @p: the process from which the cpu time has been stolen
5027 * @ticks: number of stolen ticks
5028 */
5029void account_steal_ticks(unsigned long ticks)
5030{
5031 account_steal_time(jiffies_to_cputime(ticks));
5032}
5033
5034/*
5035 * Account multiple ticks of idle time.
5036 * @ticks: number of stolen ticks
5037 */
5038void account_idle_ticks(unsigned long ticks)
5039{
5040 account_idle_time(jiffies_to_cputime(ticks));
5041}
5042
5043#endif
5044
Christoph Lameter7835b982006-12-10 02:20:22 -08005045/*
Balbir Singh49048622008-09-05 18:12:23 +02005046 * Use precise platform statistics if available:
5047 */
5048#ifdef CONFIG_VIRT_CPU_ACCOUNTING
5049cputime_t task_utime(struct task_struct *p)
5050{
5051 return p->utime;
5052}
5053
5054cputime_t task_stime(struct task_struct *p)
5055{
5056 return p->stime;
5057}
5058#else
5059cputime_t task_utime(struct task_struct *p)
5060{
5061 clock_t utime = cputime_to_clock_t(p->utime),
5062 total = utime + cputime_to_clock_t(p->stime);
5063 u64 temp;
5064
5065 /*
5066 * Use CFS's precise accounting:
5067 */
5068 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5069
5070 if (total) {
5071 temp *= utime;
5072 do_div(temp, total);
5073 }
5074 utime = (clock_t)temp;
5075
5076 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5077 return p->prev_utime;
5078}
5079
5080cputime_t task_stime(struct task_struct *p)
5081{
5082 clock_t stime;
5083
5084 /*
5085 * Use CFS's precise accounting. (we subtract utime from
5086 * the total, to make sure the total observed by userspace
5087 * grows monotonically - apps rely on that):
5088 */
5089 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5090 cputime_to_clock_t(task_utime(p));
5091
5092 if (stime >= 0)
5093 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5094
5095 return p->prev_stime;
5096}
5097#endif
5098
5099inline cputime_t task_gtime(struct task_struct *p)
5100{
5101 return p->gtime;
5102}
5103
5104/*
Christoph Lameter7835b982006-12-10 02:20:22 -08005105 * This function gets called by the timer code, with HZ frequency.
5106 * We call it with interrupts disabled.
5107 *
5108 * It also gets called by the fork code, when changing the parent's
5109 * timeslices.
5110 */
5111void scheduler_tick(void)
5112{
Christoph Lameter7835b982006-12-10 02:20:22 -08005113 int cpu = smp_processor_id();
5114 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005115 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005116
5117 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08005118
Ingo Molnardd41f592007-07-09 18:51:59 +02005119 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005120 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02005121 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01005122 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02005123 spin_unlock(&rq->lock);
5124
Peter Zijlstrae220d2d2009-05-23 18:28:55 +02005125 perf_counter_task_tick(curr, cpu);
5126
Christoph Lametere418e1c2006-12-10 02:20:23 -08005127#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02005128 rq->idle_at_tick = idle_cpu(cpu);
5129 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08005130#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131}
5132
Lai Jiangshan132380a2009-04-02 14:18:25 +08005133notrace unsigned long get_parent_ip(unsigned long addr)
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005134{
5135 if (in_lock_functions(addr)) {
5136 addr = CALLER_ADDR2;
5137 if (in_lock_functions(addr))
5138 addr = CALLER_ADDR3;
5139 }
5140 return addr;
5141}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005142
Steven Rostedt7e49fcc2009-01-22 19:01:40 -05005143#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5144 defined(CONFIG_PREEMPT_TRACER))
5145
Srinivasa Ds43627582008-02-23 15:24:04 -08005146void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005147{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005148#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005149 /*
5150 * Underflow?
5151 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005152 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5153 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005154#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005155 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005156#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005157 /*
5158 * Spinlock count overflowing soon?
5159 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08005160 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5161 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005162#endif
5163 if (preempt_count() == val)
5164 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005165}
5166EXPORT_SYMBOL(add_preempt_count);
5167
Srinivasa Ds43627582008-02-23 15:24:04 -08005168void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005169{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005170#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07005171 /*
5172 * Underflow?
5173 */
Ingo Molnar01e3eb82009-01-12 13:00:50 +01005174 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005175 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005176 /*
5177 * Is the spinlock portion underflowing?
5178 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005179 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5180 !(preempt_count() & PREEMPT_MASK)))
5181 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005182#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07005183
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02005184 if (preempt_count() == val)
5185 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005186 preempt_count() -= val;
5187}
5188EXPORT_SYMBOL(sub_preempt_count);
5189
5190#endif
5191
5192/*
Ingo Molnardd41f592007-07-09 18:51:59 +02005193 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005194 */
Ingo Molnardd41f592007-07-09 18:51:59 +02005195static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005196{
Satyam Sharma838225b2007-10-24 18:23:50 +02005197 struct pt_regs *regs = get_irq_regs();
5198
5199 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5200 prev->comm, prev->pid, preempt_count());
5201
Ingo Molnardd41f592007-07-09 18:51:59 +02005202 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07005203 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02005204 if (irqs_disabled())
5205 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02005206
5207 if (regs)
5208 show_regs(regs);
5209 else
5210 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02005211}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005212
Ingo Molnardd41f592007-07-09 18:51:59 +02005213/*
5214 * Various schedule()-time debugging checks and statistics:
5215 */
5216static inline void schedule_debug(struct task_struct *prev)
5217{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005218 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005219 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07005220 * schedule() atomically, we ignore that path for now.
5221 * Otherwise, whine if we are scheduling when we should not be.
5222 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02005223 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02005224 __schedule_bug(prev);
5225
Linus Torvalds1da177e2005-04-16 15:20:36 -07005226 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5227
Ingo Molnar2d723762007-10-15 17:00:12 +02005228 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005229#ifdef CONFIG_SCHEDSTATS
5230 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02005231 schedstat_inc(this_rq(), bkl_count);
5232 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02005233 }
5234#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005235}
5236
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005237static void put_prev_task(struct rq *rq, struct task_struct *prev)
5238{
5239 if (prev->state == TASK_RUNNING) {
5240 u64 runtime = prev->se.sum_exec_runtime;
5241
5242 runtime -= prev->se.prev_sum_exec_runtime;
5243 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5244
5245 /*
5246 * In order to avoid avg_overlap growing stale when we are
5247 * indeed overlapping and hence not getting put to sleep, grow
5248 * the avg_overlap on preemption.
5249 *
5250 * We use the average preemption runtime because that
5251 * correlates to the amount of cache footprint a task can
5252 * build up.
5253 */
5254 update_avg(&prev->se.avg_overlap, runtime);
5255 }
5256 prev->sched_class->put_prev_task(rq, prev);
5257}
5258
Ingo Molnardd41f592007-07-09 18:51:59 +02005259/*
5260 * Pick up the highest-prio task:
5261 */
5262static inline struct task_struct *
Wang Chenb67802e2009-03-02 13:55:26 +08005263pick_next_task(struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02005264{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02005265 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02005266 struct task_struct *p;
5267
5268 /*
5269 * Optimization: we know that if all tasks are in
5270 * the fair class we can call that function directly:
5271 */
5272 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005273 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005274 if (likely(p))
5275 return p;
5276 }
5277
5278 class = sched_class_highest;
5279 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02005280 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005281 if (p)
5282 return p;
5283 /*
5284 * Will never be NULL as the idle class always
5285 * returns a non-NULL p:
5286 */
5287 class = class->next;
5288 }
5289}
5290
5291/*
5292 * schedule() is the main scheduler function.
5293 */
Peter Zijlstraff743342009-03-13 12:21:26 +01005294asmlinkage void __sched schedule(void)
Ingo Molnardd41f592007-07-09 18:51:59 +02005295{
5296 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08005297 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02005298 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02005299 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02005300
Peter Zijlstraff743342009-03-13 12:21:26 +01005301need_resched:
5302 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02005303 cpu = smp_processor_id();
5304 rq = cpu_rq(cpu);
5305 rcu_qsctr_inc(cpu);
5306 prev = rq->curr;
5307 switch_count = &prev->nivcsw;
5308
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309 release_kernel_lock(prev);
5310need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005311
Ingo Molnardd41f592007-07-09 18:51:59 +02005312 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005313
Peter Zijlstra31656512008-07-18 18:01:23 +02005314 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02005315 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005316
Peter Zijlstra8cd162c2008-10-15 20:37:23 +02005317 spin_lock_irq(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02005318 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02005319 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005320
Ingo Molnardd41f592007-07-09 18:51:59 +02005321 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04005322 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02005323 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04005324 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005325 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02005326 switch_count = &prev->nvcsw;
5327 }
5328
Steven Rostedt9a897c52008-01-25 21:08:22 +01005329#ifdef CONFIG_SMP
5330 if (prev->sched_class->pre_schedule)
5331 prev->sched_class->pre_schedule(rq, prev);
5332#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01005333
Ingo Molnardd41f592007-07-09 18:51:59 +02005334 if (unlikely(!rq->nr_running))
5335 idle_balance(cpu, rq);
5336
Mike Galbraithdf1c99d2009-03-10 19:08:11 +01005337 put_prev_task(rq, prev);
Wang Chenb67802e2009-03-02 13:55:26 +08005338 next = pick_next_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005339
Linus Torvalds1da177e2005-04-16 15:20:36 -07005340 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01005341 sched_info_switch(prev, next);
Paul Mackerras564c2b22009-05-22 14:27:22 +10005342 perf_counter_task_sched_out(prev, next, cpu);
David Simner673a90a2008-04-29 10:08:59 +01005343
Linus Torvalds1da177e2005-04-16 15:20:36 -07005344 rq->nr_switches++;
5345 rq->curr = next;
5346 ++*switch_count;
5347
Ingo Molnardd41f592007-07-09 18:51:59 +02005348 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005349 /*
5350 * the context switch might have flipped the stack from under
5351 * us, hence refresh the local variables.
5352 */
5353 cpu = smp_processor_id();
5354 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005355 } else
5356 spin_unlock_irq(&rq->lock);
5357
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005358 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005359 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01005360
Linus Torvalds1da177e2005-04-16 15:20:36 -07005361 preempt_enable_no_resched();
Peter Zijlstraff743342009-03-13 12:21:26 +01005362 if (need_resched())
Linus Torvalds1da177e2005-04-16 15:20:36 -07005363 goto need_resched;
5364}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005365EXPORT_SYMBOL(schedule);
5366
Peter Zijlstra0d66bf62009-01-12 14:01:47 +01005367#ifdef CONFIG_SMP
5368/*
5369 * Look out! "owner" is an entirely speculative pointer
5370 * access and not reliable.
5371 */
5372int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5373{
5374 unsigned int cpu;
5375 struct rq *rq;
5376
5377 if (!sched_feat(OWNER_SPIN))
5378 return 0;
5379
5380#ifdef CONFIG_DEBUG_PAGEALLOC
5381 /*
5382 * Need to access the cpu field knowing that
5383 * DEBUG_PAGEALLOC could have unmapped it if
5384 * the mutex owner just released it and exited.
5385 */
5386 if (probe_kernel_address(&owner->cpu, cpu))
5387 goto out;
5388#else
5389 cpu = owner->cpu;
5390#endif
5391
5392 /*
5393 * Even if the access succeeded (likely case),
5394 * the cpu field may no longer be valid.
5395 */
5396 if (cpu >= nr_cpumask_bits)
5397 goto out;
5398
5399 /*
5400 * We need to validate that we can do a
5401 * get_cpu() and that we have the percpu area.
5402 */
5403 if (!cpu_online(cpu))
5404 goto out;
5405
5406 rq = cpu_rq(cpu);
5407
5408 for (;;) {
5409 /*
5410 * Owner changed, break to re-assess state.
5411 */
5412 if (lock->owner != owner)
5413 break;
5414
5415 /*
5416 * Is that owner really running on that cpu?
5417 */
5418 if (task_thread_info(rq->curr) != owner || need_resched())
5419 return 0;
5420
5421 cpu_relax();
5422 }
5423out:
5424 return 1;
5425}
5426#endif
5427
Linus Torvalds1da177e2005-04-16 15:20:36 -07005428#ifdef CONFIG_PREEMPT
5429/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005430 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005431 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07005432 * occur there and call schedule directly.
5433 */
5434asmlinkage void __sched preempt_schedule(void)
5435{
5436 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005437
Linus Torvalds1da177e2005-04-16 15:20:36 -07005438 /*
5439 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005440 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07005441 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07005442 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005443 return;
5444
Andi Kleen3a5c3592007-10-15 17:00:14 +02005445 do {
5446 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005447 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005448 sub_preempt_count(PREEMPT_ACTIVE);
5449
5450 /*
5451 * Check again in case we missed a preemption opportunity
5452 * between schedule and now.
5453 */
5454 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005455 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005456}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005457EXPORT_SYMBOL(preempt_schedule);
5458
5459/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005460 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07005461 * off of irq context.
5462 * Note, that this is called and return with irqs disabled. This will
5463 * protect us against recursive calling from irq.
5464 */
5465asmlinkage void __sched preempt_schedule_irq(void)
5466{
5467 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01005468
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005469 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005470 BUG_ON(ti->preempt_count || !irqs_disabled());
5471
Andi Kleen3a5c3592007-10-15 17:00:14 +02005472 do {
5473 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02005474 local_irq_enable();
5475 schedule();
5476 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02005477 sub_preempt_count(PREEMPT_ACTIVE);
5478
5479 /*
5480 * Check again in case we missed a preemption opportunity
5481 * between schedule and now.
5482 */
5483 barrier();
Lai Jiangshan5ed0cec2009-03-06 19:40:20 +08005484 } while (need_resched());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005485}
5486
5487#endif /* CONFIG_PREEMPT */
5488
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005489int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5490 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005491{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005492 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005493}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005494EXPORT_SYMBOL(default_wake_function);
5495
5496/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005497 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5498 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07005499 * number) then we wake all the non-exclusive tasks and one exclusive task.
5500 *
5501 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005502 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07005503 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5504 */
Johannes Weiner78ddb082009-04-14 16:53:05 +02005505static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
Johannes Weiner777c6c52009-02-04 15:12:14 -08005506 int nr_exclusive, int sync, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005507{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005508 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005509
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02005510 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07005511 unsigned flags = curr->flags;
5512
Linus Torvalds1da177e2005-04-16 15:20:36 -07005513 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07005514 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005515 break;
5516 }
5517}
5518
5519/**
5520 * __wake_up - wake up threads blocked on a waitqueue.
5521 * @q: the waitqueue
5522 * @mode: which threads
5523 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07005524 * @key: is directly passed to the wakeup function
David Howells50fa6102009-04-28 15:01:38 +01005525 *
5526 * It may be assumed that this function implies a write memory barrier before
5527 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005528 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005529void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005530 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005531{
5532 unsigned long flags;
5533
5534 spin_lock_irqsave(&q->lock, flags);
5535 __wake_up_common(q, mode, nr_exclusive, 0, key);
5536 spin_unlock_irqrestore(&q->lock, flags);
5537}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005538EXPORT_SYMBOL(__wake_up);
5539
5540/*
5541 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5542 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08005543void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005544{
5545 __wake_up_common(q, mode, 1, 0, NULL);
5546}
5547
Davide Libenzi4ede8162009-03-31 15:24:20 -07005548void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5549{
5550 __wake_up_common(q, mode, 1, 0, key);
5551}
5552
Linus Torvalds1da177e2005-04-16 15:20:36 -07005553/**
Davide Libenzi4ede8162009-03-31 15:24:20 -07005554 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005555 * @q: the waitqueue
5556 * @mode: which threads
5557 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Davide Libenzi4ede8162009-03-31 15:24:20 -07005558 * @key: opaque value to be passed to wakeup targets
Linus Torvalds1da177e2005-04-16 15:20:36 -07005559 *
5560 * The sync wakeup differs that the waker knows that it will schedule
5561 * away soon, so while the target thread will be woken up, it will not
5562 * be migrated to another CPU - ie. the two threads are 'synchronized'
5563 * with each other. This can prevent needless bouncing between CPUs.
5564 *
5565 * On UP it can prevent extra preemption.
David Howells50fa6102009-04-28 15:01:38 +01005566 *
5567 * It may be assumed that this function implies a write memory barrier before
5568 * changing the task state if and only if any tasks are woken up.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005569 */
Davide Libenzi4ede8162009-03-31 15:24:20 -07005570void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5571 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005572{
5573 unsigned long flags;
5574 int sync = 1;
5575
5576 if (unlikely(!q))
5577 return;
5578
5579 if (unlikely(!nr_exclusive))
5580 sync = 0;
5581
5582 spin_lock_irqsave(&q->lock, flags);
Davide Libenzi4ede8162009-03-31 15:24:20 -07005583 __wake_up_common(q, mode, nr_exclusive, sync, key);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005584 spin_unlock_irqrestore(&q->lock, flags);
5585}
Davide Libenzi4ede8162009-03-31 15:24:20 -07005586EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5587
5588/*
5589 * __wake_up_sync - see __wake_up_sync_key()
5590 */
5591void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5592{
5593 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5594}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005595EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5596
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005597/**
5598 * complete: - signals a single thread waiting on this completion
5599 * @x: holds the state of this particular completion
5600 *
5601 * This will wake up a single thread waiting on this completion. Threads will be
5602 * awakened in the same order in which they were queued.
5603 *
5604 * See also complete_all(), wait_for_completion() and related routines.
David Howells50fa6102009-04-28 15:01:38 +01005605 *
5606 * It may be assumed that this function implies a write memory barrier before
5607 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005608 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005609void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005610{
5611 unsigned long flags;
5612
5613 spin_lock_irqsave(&x->wait.lock, flags);
5614 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005615 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005616 spin_unlock_irqrestore(&x->wait.lock, flags);
5617}
5618EXPORT_SYMBOL(complete);
5619
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005620/**
5621 * complete_all: - signals all threads waiting on this completion
5622 * @x: holds the state of this particular completion
5623 *
5624 * This will wake up all threads waiting on this particular completion event.
David Howells50fa6102009-04-28 15:01:38 +01005625 *
5626 * It may be assumed that this function implies a write memory barrier before
5627 * changing the task state if and only if any tasks are woken up.
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005628 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005629void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005630{
5631 unsigned long flags;
5632
5633 spin_lock_irqsave(&x->wait.lock, flags);
5634 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05005635 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005636 spin_unlock_irqrestore(&x->wait.lock, flags);
5637}
5638EXPORT_SYMBOL(complete_all);
5639
Andi Kleen8cbbe862007-10-15 17:00:14 +02005640static inline long __sched
5641do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005642{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005643 if (!x->done) {
5644 DECLARE_WAITQUEUE(wait, current);
5645
5646 wait.flags |= WQ_FLAG_EXCLUSIVE;
5647 __add_wait_queue_tail(&x->wait, &wait);
5648 do {
Oleg Nesterov94d3d822008-08-20 16:54:41 -07005649 if (signal_pending_state(state, current)) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04005650 timeout = -ERESTARTSYS;
5651 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005652 }
5653 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005654 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005655 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005656 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005657 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005658 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04005659 if (!x->done)
5660 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661 }
5662 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04005663 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02005664}
5665
5666static long __sched
5667wait_for_common(struct completion *x, long timeout, int state)
5668{
5669 might_sleep();
5670
5671 spin_lock_irq(&x->wait.lock);
5672 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005673 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02005674 return timeout;
5675}
5676
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005677/**
5678 * wait_for_completion: - waits for completion of a task
5679 * @x: holds the state of this particular completion
5680 *
5681 * This waits to be signaled for completion of a specific task. It is NOT
5682 * interruptible and there is no timeout.
5683 *
5684 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5685 * and interrupt capability. Also see complete().
5686 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005687void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02005688{
5689 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005690}
5691EXPORT_SYMBOL(wait_for_completion);
5692
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005693/**
5694 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5695 * @x: holds the state of this particular completion
5696 * @timeout: timeout value in jiffies
5697 *
5698 * This waits for either a completion of a specific task to be signaled or for a
5699 * specified timeout to expire. The timeout is in jiffies. It is not
5700 * interruptible.
5701 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005702unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005703wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5704{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005705 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005706}
5707EXPORT_SYMBOL(wait_for_completion_timeout);
5708
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005709/**
5710 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5711 * @x: holds the state of this particular completion
5712 *
5713 * This waits for completion of a specific task to be signaled. It is
5714 * interruptible.
5715 */
Andi Kleen8cbbe862007-10-15 17:00:14 +02005716int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005717{
Andi Kleen51e97992007-10-18 21:32:55 +02005718 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5719 if (t == -ERESTARTSYS)
5720 return t;
5721 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005722}
5723EXPORT_SYMBOL(wait_for_completion_interruptible);
5724
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005725/**
5726 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5727 * @x: holds the state of this particular completion
5728 * @timeout: timeout value in jiffies
5729 *
5730 * This waits for either a completion of a specific task to be signaled or for a
5731 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5732 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02005733unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07005734wait_for_completion_interruptible_timeout(struct completion *x,
5735 unsigned long timeout)
5736{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005737 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005738}
5739EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5740
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02005741/**
5742 * wait_for_completion_killable: - waits for completion of a task (killable)
5743 * @x: holds the state of this particular completion
5744 *
5745 * This waits to be signaled for completion of a specific task. It can be
5746 * interrupted by a kill signal.
5747 */
Matthew Wilcox009e5772007-12-06 12:29:54 -05005748int __sched wait_for_completion_killable(struct completion *x)
5749{
5750 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5751 if (t == -ERESTARTSYS)
5752 return t;
5753 return 0;
5754}
5755EXPORT_SYMBOL(wait_for_completion_killable);
5756
Dave Chinnerbe4de352008-08-15 00:40:44 -07005757/**
5758 * try_wait_for_completion - try to decrement a completion without blocking
5759 * @x: completion structure
5760 *
5761 * Returns: 0 if a decrement cannot be done without blocking
5762 * 1 if a decrement succeeded.
5763 *
5764 * If a completion is being used as a counting completion,
5765 * attempt to decrement the counter without blocking. This
5766 * enables us to avoid waiting if the resource the completion
5767 * is protecting is not available.
5768 */
5769bool try_wait_for_completion(struct completion *x)
5770{
5771 int ret = 1;
5772
5773 spin_lock_irq(&x->wait.lock);
5774 if (!x->done)
5775 ret = 0;
5776 else
5777 x->done--;
5778 spin_unlock_irq(&x->wait.lock);
5779 return ret;
5780}
5781EXPORT_SYMBOL(try_wait_for_completion);
5782
5783/**
5784 * completion_done - Test to see if a completion has any waiters
5785 * @x: completion structure
5786 *
5787 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5788 * 1 if there are no waiters.
5789 *
5790 */
5791bool completion_done(struct completion *x)
5792{
5793 int ret = 1;
5794
5795 spin_lock_irq(&x->wait.lock);
5796 if (!x->done)
5797 ret = 0;
5798 spin_unlock_irq(&x->wait.lock);
5799 return ret;
5800}
5801EXPORT_SYMBOL(completion_done);
5802
Andi Kleen8cbbe862007-10-15 17:00:14 +02005803static long __sched
5804sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02005805{
5806 unsigned long flags;
5807 wait_queue_t wait;
5808
5809 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005810
Andi Kleen8cbbe862007-10-15 17:00:14 +02005811 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005812
Andi Kleen8cbbe862007-10-15 17:00:14 +02005813 spin_lock_irqsave(&q->lock, flags);
5814 __add_wait_queue(q, &wait);
5815 spin_unlock(&q->lock);
5816 timeout = schedule_timeout(timeout);
5817 spin_lock_irq(&q->lock);
5818 __remove_wait_queue(q, &wait);
5819 spin_unlock_irqrestore(&q->lock, flags);
5820
5821 return timeout;
5822}
5823
5824void __sched interruptible_sleep_on(wait_queue_head_t *q)
5825{
5826 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005827}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005828EXPORT_SYMBOL(interruptible_sleep_on);
5829
Ingo Molnar0fec1712007-07-09 18:52:01 +02005830long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005831interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005832{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005833 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005834}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005835EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5836
Ingo Molnar0fec1712007-07-09 18:52:01 +02005837void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005838{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005839 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005840}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005841EXPORT_SYMBOL(sleep_on);
5842
Ingo Molnar0fec1712007-07-09 18:52:01 +02005843long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005844{
Andi Kleen8cbbe862007-10-15 17:00:14 +02005845 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005846}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005847EXPORT_SYMBOL(sleep_on_timeout);
5848
Ingo Molnarb29739f2006-06-27 02:54:51 -07005849#ifdef CONFIG_RT_MUTEXES
5850
5851/*
5852 * rt_mutex_setprio - set the current priority of a task
5853 * @p: task
5854 * @prio: prio value (kernel-internal form)
5855 *
5856 * This function changes the 'effective' priority of a task. It does
5857 * not touch ->normal_prio like __setscheduler().
5858 *
5859 * Used by the rt_mutex code to implement priority inheritance logic.
5860 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005861void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07005862{
5863 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02005864 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005865 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01005866 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005867
5868 BUG_ON(prio < 0 || prio > MAX_PRIO);
5869
5870 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005871 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07005872
Andrew Mortond5f9f942007-05-08 20:27:06 -07005873 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02005874 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01005875 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005876 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02005877 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005878 if (running)
5879 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02005880
5881 if (rt_prio(prio))
5882 p->sched_class = &rt_sched_class;
5883 else
5884 p->sched_class = &fair_sched_class;
5885
Ingo Molnarb29739f2006-06-27 02:54:51 -07005886 p->prio = prio;
5887
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005888 if (running)
5889 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005890 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02005891 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01005892
5893 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07005894 }
5895 task_rq_unlock(rq, &flags);
5896}
5897
5898#endif
5899
Ingo Molnar36c8b582006-07-03 00:25:41 -07005900void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005901{
Ingo Molnardd41f592007-07-09 18:51:59 +02005902 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005903 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005904 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005905
5906 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5907 return;
5908 /*
5909 * We have to be careful, if called from sys_setpriority(),
5910 * the task might be in the middle of scheduling on another CPU.
5911 */
5912 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02005913 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005914 /*
5915 * The RT priorities are set via sched_setscheduler(), but we still
5916 * allow the 'normal' nice value to be set - but as expected
5917 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02005918 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005919 */
Ingo Molnare05606d2007-07-09 18:51:59 +02005920 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005921 p->static_prio = NICE_TO_PRIO(nice);
5922 goto out_unlock;
5923 }
Ingo Molnardd41f592007-07-09 18:51:59 +02005924 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02005925 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02005926 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005927
Linus Torvalds1da177e2005-04-16 15:20:36 -07005928 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07005929 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07005930 old_prio = p->prio;
5931 p->prio = effective_prio(p);
5932 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005933
Ingo Molnardd41f592007-07-09 18:51:59 +02005934 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02005935 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005936 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07005937 * If the task increased its priority or is running and
5938 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005939 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07005940 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005941 resched_task(rq->curr);
5942 }
5943out_unlock:
5944 task_rq_unlock(rq, &flags);
5945}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005946EXPORT_SYMBOL(set_user_nice);
5947
Matt Mackalle43379f2005-05-01 08:59:00 -07005948/*
5949 * can_nice - check if a task can reduce its nice value
5950 * @p: task
5951 * @nice: nice value
5952 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005953int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07005954{
Matt Mackall024f4742005-08-18 11:24:19 -07005955 /* convert nice value [19,-20] to rlimit style value [1,40] */
5956 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005957
Matt Mackalle43379f2005-05-01 08:59:00 -07005958 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5959 capable(CAP_SYS_NICE));
5960}
5961
Linus Torvalds1da177e2005-04-16 15:20:36 -07005962#ifdef __ARCH_WANT_SYS_NICE
5963
5964/*
5965 * sys_nice - change the priority of the current process.
5966 * @increment: priority increment
5967 *
5968 * sys_setpriority is a more generic, but much slower function that
5969 * does similar things.
5970 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01005971SYSCALL_DEFINE1(nice, int, increment)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005972{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005973 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005974
5975 /*
5976 * Setpriority might change our priority at the same moment.
5977 * We don't have to worry. Conceptually one call occurs first
5978 * and we have a single winner.
5979 */
Matt Mackalle43379f2005-05-01 08:59:00 -07005980 if (increment < -40)
5981 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005982 if (increment > 40)
5983 increment = 40;
5984
Américo Wang2b8f8362009-02-16 18:54:21 +08005985 nice = TASK_NICE(current) + increment;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005986 if (nice < -20)
5987 nice = -20;
5988 if (nice > 19)
5989 nice = 19;
5990
Matt Mackalle43379f2005-05-01 08:59:00 -07005991 if (increment < 0 && !can_nice(current, nice))
5992 return -EPERM;
5993
Linus Torvalds1da177e2005-04-16 15:20:36 -07005994 retval = security_task_setnice(current, nice);
5995 if (retval)
5996 return retval;
5997
5998 set_user_nice(current, nice);
5999 return 0;
6000}
6001
6002#endif
6003
6004/**
6005 * task_prio - return the priority value of a given task.
6006 * @p: the task in question.
6007 *
6008 * This is the priority value as seen by users in /proc.
6009 * RT tasks are offset by -200. Normal tasks are centered
6010 * around 0, value goes from -16 to +15.
6011 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006012int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006013{
6014 return p->prio - MAX_RT_PRIO;
6015}
6016
6017/**
6018 * task_nice - return the nice value of a given task.
6019 * @p: the task in question.
6020 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006021int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006022{
6023 return TASK_NICE(p);
6024}
Pavel Roskin150d8be2008-03-05 16:56:37 -05006025EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006026
6027/**
6028 * idle_cpu - is a given cpu idle currently?
6029 * @cpu: the processor in question.
6030 */
6031int idle_cpu(int cpu)
6032{
6033 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6034}
6035
Linus Torvalds1da177e2005-04-16 15:20:36 -07006036/**
6037 * idle_task - return the idle task for a given cpu.
6038 * @cpu: the processor in question.
6039 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006040struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006041{
6042 return cpu_rq(cpu)->idle;
6043}
6044
6045/**
6046 * find_process_by_pid - find a process with a matching PID value.
6047 * @pid: the pid in question.
6048 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02006049static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006050{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07006051 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006052}
6053
6054/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02006055static void
6056__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006057{
Ingo Molnardd41f592007-07-09 18:51:59 +02006058 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006059
Linus Torvalds1da177e2005-04-16 15:20:36 -07006060 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02006061 switch (p->policy) {
6062 case SCHED_NORMAL:
6063 case SCHED_BATCH:
6064 case SCHED_IDLE:
6065 p->sched_class = &fair_sched_class;
6066 break;
6067 case SCHED_FIFO:
6068 case SCHED_RR:
6069 p->sched_class = &rt_sched_class;
6070 break;
6071 }
6072
Linus Torvalds1da177e2005-04-16 15:20:36 -07006073 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006074 p->normal_prio = normal_prio(p);
6075 /* we are holding p->pi_lock already */
6076 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07006077 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006078}
6079
David Howellsc69e8d92008-11-14 10:39:19 +11006080/*
6081 * check the target process has a UID that matches the current process's
6082 */
6083static bool check_same_owner(struct task_struct *p)
6084{
6085 const struct cred *cred = current_cred(), *pcred;
6086 bool match;
6087
6088 rcu_read_lock();
6089 pcred = __task_cred(p);
6090 match = (cred->euid == pcred->euid ||
6091 cred->euid == pcred->uid);
6092 rcu_read_unlock();
6093 return match;
6094}
6095
Rusty Russell961ccdd2008-06-23 13:55:38 +10006096static int __sched_setscheduler(struct task_struct *p, int policy,
6097 struct sched_param *param, bool user)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006098{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02006099 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006100 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01006101 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006102 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006103
Steven Rostedt66e53932006-06-27 02:54:44 -07006104 /* may grab non-irq protected spin_locks */
6105 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07006106recheck:
6107 /* double check policy once rq lock held */
6108 if (policy < 0)
6109 policy = oldpolicy = p->policy;
6110 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02006111 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6112 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08006113 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006114 /*
6115 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02006116 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6117 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006118 */
6119 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006120 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04006121 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006122 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02006123 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006124 return -EINVAL;
6125
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006126 /*
6127 * Allow unprivileged RT tasks to decrease priority:
6128 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10006129 if (user && !capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02006130 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006131 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006132
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006133 if (!lock_task_sighand(p, &flags))
6134 return -ESRCH;
6135 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6136 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006137
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006138 /* can't set/change the rt policy */
6139 if (policy != p->policy && !rlim_rtprio)
6140 return -EPERM;
6141
6142 /* can't increase priority */
6143 if (param->sched_priority > p->rt_priority &&
6144 param->sched_priority > rlim_rtprio)
6145 return -EPERM;
6146 }
Ingo Molnardd41f592007-07-09 18:51:59 +02006147 /*
6148 * Like positive nice levels, dont allow tasks to
6149 * move out of SCHED_IDLE either:
6150 */
6151 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6152 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07006153
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006154 /* can't change other user's priorities */
David Howellsc69e8d92008-11-14 10:39:19 +11006155 if (!check_same_owner(p))
Olivier Croquette37e4ab32005-06-25 14:57:32 -07006156 return -EPERM;
6157 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006158
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006159 if (user) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006160#ifdef CONFIG_RT_GROUP_SCHED
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006161 /*
6162 * Do not allow realtime tasks into groups that have no runtime
6163 * assigned.
6164 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02006165 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6166 task_group(p)->rt_bandwidth.rt_runtime == 0)
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006167 return -EPERM;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01006168#endif
6169
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07006170 retval = security_task_setscheduler(p, policy, param);
6171 if (retval)
6172 return retval;
6173 }
6174
Linus Torvalds1da177e2005-04-16 15:20:36 -07006175 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07006176 * make sure no PI-waiters arrive (or leave) while we are
6177 * changing the priority of the task:
6178 */
6179 spin_lock_irqsave(&p->pi_lock, flags);
6180 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07006181 * To be able to change p->policy safely, the apropriate
6182 * runqueue lock must be held.
6183 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07006184 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006185 /* recheck policy now with rq lock held */
6186 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6187 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07006188 __task_rq_unlock(rq);
6189 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006190 goto recheck;
6191 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02006192 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006193 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01006194 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006195 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006196 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006197 if (running)
6198 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02006199
Linus Torvalds1da177e2005-04-16 15:20:36 -07006200 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02006201 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02006202
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07006203 if (running)
6204 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02006205 if (on_rq) {
6206 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01006207
6208 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006209 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07006210 __task_rq_unlock(rq);
6211 spin_unlock_irqrestore(&p->pi_lock, flags);
6212
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07006213 rt_mutex_adjust_pi(p);
6214
Linus Torvalds1da177e2005-04-16 15:20:36 -07006215 return 0;
6216}
Rusty Russell961ccdd2008-06-23 13:55:38 +10006217
6218/**
6219 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6220 * @p: the task in question.
6221 * @policy: new policy.
6222 * @param: structure containing the new RT priority.
6223 *
6224 * NOTE that the task may be already dead.
6225 */
6226int sched_setscheduler(struct task_struct *p, int policy,
6227 struct sched_param *param)
6228{
6229 return __sched_setscheduler(p, policy, param, true);
6230}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006231EXPORT_SYMBOL_GPL(sched_setscheduler);
6232
Rusty Russell961ccdd2008-06-23 13:55:38 +10006233/**
6234 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6235 * @p: the task in question.
6236 * @policy: new policy.
6237 * @param: structure containing the new RT priority.
6238 *
6239 * Just like sched_setscheduler, only don't bother checking if the
6240 * current context has permission. For example, this is needed in
6241 * stop_machine(): we create temporary high priority worker threads,
6242 * but our caller might not have that capability.
6243 */
6244int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6245 struct sched_param *param)
6246{
6247 return __sched_setscheduler(p, policy, param, false);
6248}
6249
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006250static int
6251do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006252{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006253 struct sched_param lparam;
6254 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006255 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006256
6257 if (!param || pid < 0)
6258 return -EINVAL;
6259 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6260 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006261
6262 rcu_read_lock();
6263 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006264 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07006265 if (p != NULL)
6266 retval = sched_setscheduler(p, policy, &lparam);
6267 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07006268
Linus Torvalds1da177e2005-04-16 15:20:36 -07006269 return retval;
6270}
6271
6272/**
6273 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6274 * @pid: the pid in question.
6275 * @policy: new policy.
6276 * @param: structure containing the new RT priority.
6277 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006278SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6279 struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006280{
Jason Baronc21761f2006-01-18 17:43:03 -08006281 /* negative values for policy are not valid */
6282 if (policy < 0)
6283 return -EINVAL;
6284
Linus Torvalds1da177e2005-04-16 15:20:36 -07006285 return do_sched_setscheduler(pid, policy, param);
6286}
6287
6288/**
6289 * sys_sched_setparam - set/change the RT priority of a thread
6290 * @pid: the pid in question.
6291 * @param: structure containing the new RT priority.
6292 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006293SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006294{
6295 return do_sched_setscheduler(pid, -1, param);
6296}
6297
6298/**
6299 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6300 * @pid: the pid in question.
6301 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006302SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006303{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006304 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006305 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006306
6307 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006308 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006309
6310 retval = -ESRCH;
6311 read_lock(&tasklist_lock);
6312 p = find_process_by_pid(pid);
6313 if (p) {
6314 retval = security_task_getscheduler(p);
6315 if (!retval)
6316 retval = p->policy;
6317 }
6318 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006319 return retval;
6320}
6321
6322/**
6323 * sys_sched_getscheduler - get the RT priority of a thread
6324 * @pid: the pid in question.
6325 * @param: structure containing the RT priority.
6326 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006327SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006328{
6329 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006330 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006331 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006332
6333 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006334 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006335
6336 read_lock(&tasklist_lock);
6337 p = find_process_by_pid(pid);
6338 retval = -ESRCH;
6339 if (!p)
6340 goto out_unlock;
6341
6342 retval = security_task_getscheduler(p);
6343 if (retval)
6344 goto out_unlock;
6345
6346 lp.sched_priority = p->rt_priority;
6347 read_unlock(&tasklist_lock);
6348
6349 /*
6350 * This one might sleep, we cannot do it with a spinlock held ...
6351 */
6352 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6353
Linus Torvalds1da177e2005-04-16 15:20:36 -07006354 return retval;
6355
6356out_unlock:
6357 read_unlock(&tasklist_lock);
6358 return retval;
6359}
6360
Rusty Russell96f874e2008-11-25 02:35:14 +10306361long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006362{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306363 cpumask_var_t cpus_allowed, new_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006364 struct task_struct *p;
6365 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006366
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006367 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006368 read_lock(&tasklist_lock);
6369
6370 p = find_process_by_pid(pid);
6371 if (!p) {
6372 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006373 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006374 return -ESRCH;
6375 }
6376
6377 /*
6378 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006379 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07006380 * usage count and then drop tasklist_lock.
6381 */
6382 get_task_struct(p);
6383 read_unlock(&tasklist_lock);
6384
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306385 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6386 retval = -ENOMEM;
6387 goto out_put_task;
6388 }
6389 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6390 retval = -ENOMEM;
6391 goto out_free_cpus_allowed;
6392 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006393 retval = -EPERM;
David Howellsc69e8d92008-11-14 10:39:19 +11006394 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006395 goto out_unlock;
6396
David Quigleye7834f82006-06-23 02:03:59 -07006397 retval = security_task_setscheduler(p, 0, NULL);
6398 if (retval)
6399 goto out_unlock;
6400
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306401 cpuset_cpus_allowed(p, cpus_allowed);
6402 cpumask_and(new_mask, in_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006403 again:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306404 retval = set_cpus_allowed_ptr(p, new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006405
Paul Menage8707d8b2007-10-18 23:40:22 -07006406 if (!retval) {
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306407 cpuset_cpus_allowed(p, cpus_allowed);
6408 if (!cpumask_subset(new_mask, cpus_allowed)) {
Paul Menage8707d8b2007-10-18 23:40:22 -07006409 /*
6410 * We must have raced with a concurrent cpuset
6411 * update. Just reset the cpus_allowed to the
6412 * cpuset's cpus_allowed
6413 */
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306414 cpumask_copy(new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07006415 goto again;
6416 }
6417 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006418out_unlock:
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306419 free_cpumask_var(new_mask);
6420out_free_cpus_allowed:
6421 free_cpumask_var(cpus_allowed);
6422out_put_task:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006423 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006424 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006425 return retval;
6426}
6427
6428static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
Rusty Russell96f874e2008-11-25 02:35:14 +10306429 struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006430{
Rusty Russell96f874e2008-11-25 02:35:14 +10306431 if (len < cpumask_size())
6432 cpumask_clear(new_mask);
6433 else if (len > cpumask_size())
6434 len = cpumask_size();
6435
Linus Torvalds1da177e2005-04-16 15:20:36 -07006436 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6437}
6438
6439/**
6440 * sys_sched_setaffinity - set the cpu affinity of a process
6441 * @pid: pid of the process
6442 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6443 * @user_mask_ptr: user-space pointer to the new cpu mask
6444 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006445SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6446 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006447{
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306448 cpumask_var_t new_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006449 int retval;
6450
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306451 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6452 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006453
Rusty Russell5a16f3d2008-11-25 02:35:11 +10306454 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6455 if (retval == 0)
6456 retval = sched_setaffinity(pid, new_mask);
6457 free_cpumask_var(new_mask);
6458 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006459}
6460
Rusty Russell96f874e2008-11-25 02:35:14 +10306461long sched_getaffinity(pid_t pid, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006462{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006463 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006464 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006465
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006466 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006467 read_lock(&tasklist_lock);
6468
6469 retval = -ESRCH;
6470 p = find_process_by_pid(pid);
6471 if (!p)
6472 goto out_unlock;
6473
David Quigleye7834f82006-06-23 02:03:59 -07006474 retval = security_task_getscheduler(p);
6475 if (retval)
6476 goto out_unlock;
6477
Rusty Russell96f874e2008-11-25 02:35:14 +10306478 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006479
6480out_unlock:
6481 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01006482 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006483
Ulrich Drepper9531b622007-08-09 11:16:46 +02006484 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006485}
6486
6487/**
6488 * sys_sched_getaffinity - get the cpu affinity of a process
6489 * @pid: pid of the process
6490 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6491 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6492 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006493SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6494 unsigned long __user *, user_mask_ptr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006495{
6496 int ret;
Rusty Russellf17c8602008-11-25 02:35:11 +10306497 cpumask_var_t mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006498
Rusty Russellf17c8602008-11-25 02:35:11 +10306499 if (len < cpumask_size())
Linus Torvalds1da177e2005-04-16 15:20:36 -07006500 return -EINVAL;
6501
Rusty Russellf17c8602008-11-25 02:35:11 +10306502 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6503 return -ENOMEM;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006504
Rusty Russellf17c8602008-11-25 02:35:11 +10306505 ret = sched_getaffinity(pid, mask);
6506 if (ret == 0) {
6507 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6508 ret = -EFAULT;
6509 else
6510 ret = cpumask_size();
6511 }
6512 free_cpumask_var(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006513
Rusty Russellf17c8602008-11-25 02:35:11 +10306514 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006515}
6516
6517/**
6518 * sys_sched_yield - yield the current processor to other threads.
6519 *
Ingo Molnardd41f592007-07-09 18:51:59 +02006520 * This function yields the current CPU to other tasks. If there are no
6521 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006522 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006523SYSCALL_DEFINE0(sched_yield)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006524{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006525 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006526
Ingo Molnar2d723762007-10-15 17:00:12 +02006527 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02006528 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006529
6530 /*
6531 * Since we are going to call schedule() anyway, there's
6532 * no need to preempt or enable interrupts:
6533 */
6534 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07006535 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006536 _raw_spin_unlock(&rq->lock);
6537 preempt_enable_no_resched();
6538
6539 schedule();
6540
6541 return 0;
6542}
6543
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006544static inline int should_resched(void)
6545{
6546 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6547}
6548
Andrew Mortone7b38402006-06-30 01:56:00 -07006549static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006550{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07006551#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6552 __might_sleep(__FILE__, __LINE__);
6553#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07006554 /*
6555 * The BKS might be reacquired before we have dropped
6556 * PREEMPT_ACTIVE, which could trigger a second
6557 * cond_resched() call.
6558 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006559 do {
6560 add_preempt_count(PREEMPT_ACTIVE);
6561 schedule();
6562 sub_preempt_count(PREEMPT_ACTIVE);
6563 } while (need_resched());
6564}
6565
Herbert Xu02b67cc32008-01-25 21:08:28 +01006566int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006567{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006568 if (should_resched()) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006569 __cond_resched();
6570 return 1;
6571 }
6572 return 0;
6573}
Herbert Xu02b67cc32008-01-25 21:08:28 +01006574EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006575
6576/*
6577 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6578 * call schedule, and on return reacquire the lock.
6579 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006580 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07006581 * operations here to prevent schedule() from being called twice (once via
6582 * spin_unlock(), once by hand).
6583 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006584int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006585{
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006586 int resched = should_resched();
Jan Kara6df3cec2005-06-13 15:52:32 -07006587 int ret = 0;
6588
Nick Piggin95c354f2008-01-30 13:31:20 +01006589 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006590 spin_unlock(lock);
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006591 if (resched)
Nick Piggin95c354f2008-01-30 13:31:20 +01006592 __cond_resched();
6593 else
6594 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07006595 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006596 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006597 }
Jan Kara6df3cec2005-06-13 15:52:32 -07006598 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006599}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006600EXPORT_SYMBOL(cond_resched_lock);
6601
6602int __sched cond_resched_softirq(void)
6603{
6604 BUG_ON(!in_softirq());
6605
Peter Zijlstrad86ee482009-07-10 14:57:57 +02006606 if (should_resched()) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07006607 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006608 __cond_resched();
6609 local_bh_disable();
6610 return 1;
6611 }
6612 return 0;
6613}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006614EXPORT_SYMBOL(cond_resched_softirq);
6615
Linus Torvalds1da177e2005-04-16 15:20:36 -07006616/**
6617 * yield - yield the current processor to other threads.
6618 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08006619 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07006620 * thread runnable and calls sys_sched_yield().
6621 */
6622void __sched yield(void)
6623{
6624 set_current_state(TASK_RUNNING);
6625 sys_sched_yield();
6626}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006627EXPORT_SYMBOL(yield);
6628
6629/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006630 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07006631 * that process accounting knows that this is a task in IO wait state.
6632 *
6633 * But don't do that if it is a deliberate, throttling IO wait (this task
6634 * has set its backing_dev_info: the queue against which it should throttle)
6635 */
6636void __sched io_schedule(void)
6637{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006638 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006639
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006640 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006641 atomic_inc(&rq->nr_iowait);
6642 schedule();
6643 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006644 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006645}
Linus Torvalds1da177e2005-04-16 15:20:36 -07006646EXPORT_SYMBOL(io_schedule);
6647
6648long __sched io_schedule_timeout(long timeout)
6649{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006650 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006651 long ret;
6652
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006653 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006654 atomic_inc(&rq->nr_iowait);
6655 ret = schedule_timeout(timeout);
6656 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07006657 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006658 return ret;
6659}
6660
6661/**
6662 * sys_sched_get_priority_max - return maximum RT priority.
6663 * @policy: scheduling class.
6664 *
6665 * this syscall returns the maximum rt_priority that can be used
6666 * by a given scheduling class.
6667 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006668SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006669{
6670 int ret = -EINVAL;
6671
6672 switch (policy) {
6673 case SCHED_FIFO:
6674 case SCHED_RR:
6675 ret = MAX_USER_RT_PRIO-1;
6676 break;
6677 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006678 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006679 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006680 ret = 0;
6681 break;
6682 }
6683 return ret;
6684}
6685
6686/**
6687 * sys_sched_get_priority_min - return minimum RT priority.
6688 * @policy: scheduling class.
6689 *
6690 * this syscall returns the minimum rt_priority that can be used
6691 * by a given scheduling class.
6692 */
Heiko Carstens5add95d2009-01-14 14:14:08 +01006693SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006694{
6695 int ret = -EINVAL;
6696
6697 switch (policy) {
6698 case SCHED_FIFO:
6699 case SCHED_RR:
6700 ret = 1;
6701 break;
6702 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08006703 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02006704 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006705 ret = 0;
6706 }
6707 return ret;
6708}
6709
6710/**
6711 * sys_sched_rr_get_interval - return the default timeslice of a process.
6712 * @pid: pid of the process.
6713 * @interval: userspace pointer to the timeslice value.
6714 *
6715 * this syscall writes the default timeslice value of a given process
6716 * into the user-space timespec buffer. A value of '0' means infinity.
6717 */
Heiko Carstens17da2bd2009-01-14 14:14:10 +01006718SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
Heiko Carstens754fe8d2009-01-14 14:14:09 +01006719 struct timespec __user *, interval)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006720{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006721 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006722 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006723 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006724 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006725
6726 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006727 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006728
6729 retval = -ESRCH;
6730 read_lock(&tasklist_lock);
6731 p = find_process_by_pid(pid);
6732 if (!p)
6733 goto out_unlock;
6734
6735 retval = security_task_getscheduler(p);
6736 if (retval)
6737 goto out_unlock;
6738
Ingo Molnar77034932007-12-04 17:04:39 +01006739 /*
6740 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6741 * tasks that are on an otherwise idle runqueue:
6742 */
6743 time_slice = 0;
6744 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006745 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08006746 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006747 struct sched_entity *se = &p->se;
6748 unsigned long flags;
6749 struct rq *rq;
6750
6751 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01006752 if (rq->cfs.load.weight)
6753 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006754 task_rq_unlock(rq, &flags);
6755 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006756 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02006757 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006758 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006759 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006760
Linus Torvalds1da177e2005-04-16 15:20:36 -07006761out_unlock:
6762 read_unlock(&tasklist_lock);
6763 return retval;
6764}
6765
Steven Rostedt7c731e02008-05-12 21:20:41 +02006766static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006767
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006768void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006769{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006770 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07006771 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006772
Linus Torvalds1da177e2005-04-16 15:20:36 -07006773 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006774 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07006775 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02006776#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07006777 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006778 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006779 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006780 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006781#else
6782 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006783 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006784 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02006785 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006786#endif
6787#ifdef CONFIG_DEBUG_STACK_USAGE
Eric Sandeen7c9f8862008-04-22 16:38:23 -05006788 free = stack_not_used(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006789#endif
David Rientjesaa47b7e2009-05-04 01:38:05 -07006790 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6791 task_pid_nr(p), task_pid_nr(p->real_parent),
6792 (unsigned long)task_thread_info(p)->flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006793
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01006794 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006795}
6796
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006797void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006798{
Ingo Molnar36c8b582006-07-03 00:25:41 -07006799 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006800
Ingo Molnar4bd77322007-07-11 21:21:47 +02006801#if BITS_PER_LONG == 32
6802 printk(KERN_INFO
6803 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006804#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02006805 printk(KERN_INFO
6806 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07006807#endif
6808 read_lock(&tasklist_lock);
6809 do_each_thread(g, p) {
6810 /*
6811 * reset the NMI-timeout, listing all files on a slow
6812 * console might take alot of time:
6813 */
6814 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07006815 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01006816 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006817 } while_each_thread(g, p);
6818
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07006819 touch_all_softlockup_watchdogs();
6820
Ingo Molnardd41f592007-07-09 18:51:59 +02006821#ifdef CONFIG_SCHED_DEBUG
6822 sysrq_sched_debug_show();
6823#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006824 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08006825 /*
6826 * Only show locks if all tasks are dumped:
6827 */
6828 if (state_filter == -1)
6829 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006830}
6831
Ingo Molnar1df21052007-07-09 18:51:58 +02006832void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6833{
Ingo Molnardd41f592007-07-09 18:51:59 +02006834 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02006835}
6836
Ingo Molnarf340c0d2005-06-28 16:40:42 +02006837/**
6838 * init_idle - set up an idle thread for a given CPU
6839 * @idle: task in question
6840 * @cpu: cpu the idle task belongs to
6841 *
6842 * NOTE: this function does not set the idle thread's NEED_RESCHED
6843 * flag, to make booting more robust.
6844 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07006845void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006846{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006847 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006848 unsigned long flags;
6849
Ingo Molnar5cbd54e2008-11-12 20:05:50 +01006850 spin_lock_irqsave(&rq->lock, flags);
6851
Ingo Molnardd41f592007-07-09 18:51:59 +02006852 __sched_fork(idle);
6853 idle->se.exec_start = sched_clock();
6854
Ingo Molnarb29739f2006-06-27 02:54:51 -07006855 idle->prio = idle->normal_prio = MAX_PRIO;
Rusty Russell96f874e2008-11-25 02:35:14 +10306856 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
Ingo Molnardd41f592007-07-09 18:51:59 +02006857 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006858
Linus Torvalds1da177e2005-04-16 15:20:36 -07006859 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07006860#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6861 idle->oncpu = 1;
6862#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07006863 spin_unlock_irqrestore(&rq->lock, flags);
6864
6865 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07006866#if defined(CONFIG_PREEMPT)
6867 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6868#else
Al Viroa1261f52005-11-13 16:06:55 -08006869 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07006870#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02006871 /*
6872 * The idle tasks have their own, simple scheduling class:
6873 */
6874 idle->sched_class = &idle_sched_class;
Frederic Weisbeckerfb526072008-11-25 21:07:04 +01006875 ftrace_graph_init_task(idle);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006876}
6877
6878/*
6879 * In a system that switches off the HZ timer nohz_cpu_mask
6880 * indicates which cpus entered this state. This is used
6881 * in the rcu update to wait only for active cpus. For system
6882 * which do not switch off the HZ timer nohz_cpu_mask should
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10306883 * always be CPU_BITS_NONE.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006884 */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10306885cpumask_var_t nohz_cpu_mask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006886
Ingo Molnar19978ca2007-11-09 22:39:38 +01006887/*
6888 * Increase the granularity value when there are more CPUs,
6889 * because with more CPUs the 'effective latency' as visible
6890 * to users decreases. But the relationship is not linear,
6891 * so pick a second-best guess by going with the log2 of the
6892 * number of CPUs.
6893 *
6894 * This idea comes from the SD scheduler of Con Kolivas:
6895 */
6896static inline void sched_init_granularity(void)
6897{
6898 unsigned int factor = 1 + ilog2(num_online_cpus());
6899 const unsigned long limit = 200000000;
6900
6901 sysctl_sched_min_granularity *= factor;
6902 if (sysctl_sched_min_granularity > limit)
6903 sysctl_sched_min_granularity = limit;
6904
6905 sysctl_sched_latency *= factor;
6906 if (sysctl_sched_latency > limit)
6907 sysctl_sched_latency = limit;
6908
6909 sysctl_sched_wakeup_granularity *= factor;
Peter Zijlstra55cd5342008-08-04 08:54:26 +02006910
6911 sysctl_sched_shares_ratelimit *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01006912}
6913
Linus Torvalds1da177e2005-04-16 15:20:36 -07006914#ifdef CONFIG_SMP
6915/*
6916 * This is how migration works:
6917 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07006918 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07006919 * runqueue and wake up that CPU's migration thread.
6920 * 2) we down() the locked semaphore => thread blocks.
6921 * 3) migration thread wakes up (implicitly it forces the migrated
6922 * thread off the CPU)
6923 * 4) it gets the migration request and checks whether the migrated
6924 * task is still in the wrong runqueue.
6925 * 5) if it's in the wrong runqueue then the migration thread removes
6926 * it and puts it into the right queue.
6927 * 6) migration thread up()s the semaphore.
6928 * 7) we wake up and the migration is done.
6929 */
6930
6931/*
6932 * Change a given task's CPU affinity. Migrate the thread to a
6933 * proper CPU and schedule it away if the CPU it's executing on
6934 * is removed from the allowed bitmask.
6935 *
6936 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006937 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07006938 * call is not atomic; no spinlocks may be held.
6939 */
Rusty Russell96f874e2008-11-25 02:35:14 +10306940int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006941{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006942 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006943 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006944 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006945 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006946
6947 rq = task_rq_lock(p, &flags);
Rusty Russell96f874e2008-11-25 02:35:14 +10306948 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006949 ret = -EINVAL;
6950 goto out;
6951 }
6952
David Rientjes9985b0b2008-06-05 12:57:11 -07006953 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
Rusty Russell96f874e2008-11-25 02:35:14 +10306954 !cpumask_equal(&p->cpus_allowed, new_mask))) {
David Rientjes9985b0b2008-06-05 12:57:11 -07006955 ret = -EINVAL;
6956 goto out;
6957 }
6958
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01006959 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07006960 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01006961 else {
Rusty Russell96f874e2008-11-25 02:35:14 +10306962 cpumask_copy(&p->cpus_allowed, new_mask);
6963 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01006964 }
6965
Linus Torvalds1da177e2005-04-16 15:20:36 -07006966 /* Can the task run on the task's current CPU? If so, we're done */
Rusty Russell96f874e2008-11-25 02:35:14 +10306967 if (cpumask_test_cpu(task_cpu(p), new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006968 goto out;
6969
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10306970 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006971 /* Need help from migration thread: drop lock and wait. */
6972 task_rq_unlock(rq, &flags);
6973 wake_up_process(rq->migration_thread);
6974 wait_for_completion(&req.done);
6975 tlb_migrate_finish(p->mm);
6976 return 0;
6977 }
6978out:
6979 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006980
Linus Torvalds1da177e2005-04-16 15:20:36 -07006981 return ret;
6982}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07006983EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006984
6985/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006986 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07006987 * this because either it can't run here any more (set_cpus_allowed()
6988 * away from this CPU, or CPU going down), or because we're
6989 * attempting to rebalance this task on exec (sched_exec).
6990 *
6991 * So we race with normal scheduler movements, but that's OK, as long
6992 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07006993 *
6994 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006995 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07006996static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006997{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006998 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02006999 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007000
Max Krasnyanskye761b772008-07-15 04:43:49 -07007001 if (unlikely(!cpu_active(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07007002 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007003
7004 rq_src = cpu_rq(src_cpu);
7005 rq_dest = cpu_rq(dest_cpu);
7006
7007 double_rq_lock(rq_src, rq_dest);
7008 /* Already moved. */
7009 if (task_cpu(p) != src_cpu)
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007010 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007011 /* Affinity changed (again). */
Rusty Russell96f874e2008-11-25 02:35:14 +10307012 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007013 goto fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007014
Ingo Molnardd41f592007-07-09 18:51:59 +02007015 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007016 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007017 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02007018
Linus Torvalds1da177e2005-04-16 15:20:36 -07007019 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007020 if (on_rq) {
7021 activate_task(rq_dest, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02007022 check_preempt_curr(rq_dest, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007023 }
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007024done:
Kirill Korotaevefc30812006-06-27 02:54:32 -07007025 ret = 1;
Linus Torvaldsb1e38732008-07-10 11:25:03 -07007026fail:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007027 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07007028 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007029}
7030
7031/*
7032 * migration_thread - this is a highprio system thread that performs
7033 * thread migration by bumping thread off CPU then 'pushing' onto
7034 * another runqueue.
7035 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07007036static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007037{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007038 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007039 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007040
7041 rq = cpu_rq(cpu);
7042 BUG_ON(rq->migration_thread != current);
7043
7044 set_current_state(TASK_INTERRUPTIBLE);
7045 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007046 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007047 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007048
Linus Torvalds1da177e2005-04-16 15:20:36 -07007049 spin_lock_irq(&rq->lock);
7050
7051 if (cpu_is_offline(cpu)) {
7052 spin_unlock_irq(&rq->lock);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007053 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007054 }
7055
7056 if (rq->active_balance) {
7057 active_load_balance(rq, cpu);
7058 rq->active_balance = 0;
7059 }
7060
7061 head = &rq->migration_queue;
7062
7063 if (list_empty(head)) {
7064 spin_unlock_irq(&rq->lock);
7065 schedule();
7066 set_current_state(TASK_INTERRUPTIBLE);
7067 continue;
7068 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07007069 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007070 list_del_init(head->next);
7071
Nick Piggin674311d2005-06-25 14:57:27 -07007072 spin_unlock(&rq->lock);
7073 __migrate_task(req->task, cpu, req->dest_cpu);
7074 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007075
7076 complete(&req->done);
7077 }
7078 __set_current_state(TASK_RUNNING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007079
Linus Torvalds1da177e2005-04-16 15:20:36 -07007080 return 0;
7081}
7082
7083#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007084
7085static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7086{
7087 int ret;
7088
7089 local_irq_disable();
7090 ret = __migrate_task(p, src_cpu, dest_cpu);
7091 local_irq_enable();
7092 return ret;
7093}
7094
Kirill Korotaev054b9102006-12-10 02:20:11 -08007095/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007096 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08007097 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007098static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007099{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007100 int dest_cpu;
Mike Travis6ca09df2008-12-31 18:08:45 -08007101 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007102
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307103again:
7104 /* Look for allowed, online CPU in same node. */
7105 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7106 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7107 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007108
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307109 /* Any allowed, online CPU? */
7110 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7111 if (dest_cpu < nr_cpu_ids)
7112 goto move;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007113
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307114 /* No more Mr. Nice Guy. */
7115 if (dest_cpu >= nr_cpu_ids) {
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307116 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7117 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
Mike Travisf9a86fc2008-04-04 18:11:07 -07007118
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307119 /*
7120 * Don't tell them about moving exiting tasks or
7121 * kernel threads (both mm NULL), since they never
7122 * leave kernel.
7123 */
7124 if (p->mm && printk_ratelimit()) {
7125 printk(KERN_INFO "process %d (%s) no "
7126 "longer affine to cpu%d\n",
7127 task_pid_nr(p), p->comm, dead_cpu);
Andi Kleen3a5c3592007-10-15 17:00:14 +02007128 }
Rusty Russelle76bd8d2008-11-25 02:35:11 +10307129 }
7130
7131move:
7132 /* It can have affinity changed while we were choosing. */
7133 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7134 goto again;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007135}
7136
7137/*
7138 * While a dead CPU has no uninterruptible tasks queued at this point,
7139 * it might still have a nonzero ->nr_uninterruptible counter, because
7140 * for performance reasons the counter is not stricly tracking tasks to
7141 * their home CPUs. So we just add the counter to another CPU's counter,
7142 * to keep the global sum constant after CPU-down:
7143 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07007144static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007145{
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307146 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007147 unsigned long flags;
7148
7149 local_irq_save(flags);
7150 double_rq_lock(rq_src, rq_dest);
7151 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7152 rq_src->nr_uninterruptible = 0;
7153 double_rq_unlock(rq_src, rq_dest);
7154 local_irq_restore(flags);
7155}
7156
7157/* Run through task list and migrate tasks from the dead cpu. */
7158static void migrate_live_tasks(int src_cpu)
7159{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007160 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007161
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007162 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007163
Ingo Molnar48f24c42006-07-03 00:25:40 -07007164 do_each_thread(t, p) {
7165 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007166 continue;
7167
Ingo Molnar48f24c42006-07-03 00:25:40 -07007168 if (task_cpu(p) == src_cpu)
7169 move_task_off_dead_cpu(src_cpu, p);
7170 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007171
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007172 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007173}
7174
Ingo Molnardd41f592007-07-09 18:51:59 +02007175/*
7176 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007177 * It does so by boosting its priority to highest possible.
7178 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007179 */
7180void sched_idle_next(void)
7181{
Ingo Molnar48f24c42006-07-03 00:25:40 -07007182 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07007183 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007184 struct task_struct *p = rq->idle;
7185 unsigned long flags;
7186
7187 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007188 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007189
Ingo Molnar48f24c42006-07-03 00:25:40 -07007190 /*
7191 * Strictly not necessary since rest of the CPUs are stopped by now
7192 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007193 */
7194 spin_lock_irqsave(&rq->lock, flags);
7195
Ingo Molnardd41f592007-07-09 18:51:59 +02007196 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007197
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01007198 update_rq_clock(rq);
7199 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007200
7201 spin_unlock_irqrestore(&rq->lock, flags);
7202}
7203
Ingo Molnar48f24c42006-07-03 00:25:40 -07007204/*
7205 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07007206 * offline.
7207 */
7208void idle_task_exit(void)
7209{
7210 struct mm_struct *mm = current->active_mm;
7211
7212 BUG_ON(cpu_online(smp_processor_id()));
7213
7214 if (mm != &init_mm)
7215 switch_mm(mm, &init_mm, current);
7216 mmdrop(mm);
7217}
7218
Kirill Korotaev054b9102006-12-10 02:20:11 -08007219/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07007220static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007221{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007222 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007223
7224 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07007225 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007226
7227 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07007228 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007229
Ingo Molnar48f24c42006-07-03 00:25:40 -07007230 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007231
7232 /*
7233 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007234 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07007235 * fine.
7236 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007237 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007238 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07007239 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007240
Ingo Molnar48f24c42006-07-03 00:25:40 -07007241 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007242}
7243
7244/* release_task() removes task from tasklist, so we won't find dead tasks. */
7245static void migrate_dead_tasks(unsigned int dead_cpu)
7246{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007247 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02007248 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007249
Ingo Molnardd41f592007-07-09 18:51:59 +02007250 for ( ; ; ) {
7251 if (!rq->nr_running)
7252 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02007253 update_rq_clock(rq);
Wang Chenb67802e2009-03-02 13:55:26 +08007254 next = pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02007255 if (!next)
7256 break;
Dmitry Adamushko79c53792008-06-29 00:16:56 +02007257 next->sched_class->put_prev_task(rq, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02007258 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02007259
Linus Torvalds1da177e2005-04-16 15:20:36 -07007260 }
7261}
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007262
7263/*
7264 * remove the tasks which were accounted by rq from calc_load_tasks.
7265 */
7266static void calc_global_load_remove(struct rq *rq)
7267{
7268 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7269}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007270#endif /* CONFIG_HOTPLUG_CPU */
7271
Nick Piggine692ab52007-07-26 13:40:43 +02007272#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7273
7274static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007275 {
7276 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007277 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007278 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01007279 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02007280};
7281
7282static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02007283 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007284 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007285 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007286 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02007287 .child = sd_ctl_dir,
7288 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01007289 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02007290};
7291
7292static struct ctl_table *sd_alloc_ctl_entry(int n)
7293{
7294 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02007295 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02007296
Nick Piggine692ab52007-07-26 13:40:43 +02007297 return entry;
7298}
7299
Milton Miller6382bc92007-10-15 17:00:19 +02007300static void sd_free_ctl_entry(struct ctl_table **tablep)
7301{
Milton Millercd7900762007-10-17 16:55:11 +02007302 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02007303
Milton Millercd7900762007-10-17 16:55:11 +02007304 /*
7305 * In the intermediate directories, both the child directory and
7306 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007307 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02007308 * static strings and all have proc handlers.
7309 */
7310 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02007311 if (entry->child)
7312 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02007313 if (entry->proc_handler == NULL)
7314 kfree(entry->procname);
7315 }
Milton Miller6382bc92007-10-15 17:00:19 +02007316
7317 kfree(*tablep);
7318 *tablep = NULL;
7319}
7320
Nick Piggine692ab52007-07-26 13:40:43 +02007321static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02007322set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02007323 const char *procname, void *data, int maxlen,
7324 mode_t mode, proc_handler *proc_handler)
7325{
Nick Piggine692ab52007-07-26 13:40:43 +02007326 entry->procname = procname;
7327 entry->data = data;
7328 entry->maxlen = maxlen;
7329 entry->mode = mode;
7330 entry->proc_handler = proc_handler;
7331}
7332
7333static struct ctl_table *
7334sd_alloc_ctl_domain_table(struct sched_domain *sd)
7335{
Ingo Molnara5d8c342008-10-09 11:35:51 +02007336 struct ctl_table *table = sd_alloc_ctl_entry(13);
Nick Piggine692ab52007-07-26 13:40:43 +02007337
Milton Millerad1cdc12007-10-15 17:00:19 +02007338 if (table == NULL)
7339 return NULL;
7340
Alexey Dobriyane0361852007-08-09 11:16:46 +02007341 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007342 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007343 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02007344 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007345 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007346 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007347 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007348 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007349 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007350 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007351 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007352 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007353 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02007354 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007355 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02007356 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02007357 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02007358 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007359 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02007360 &sd->cache_nice_tries,
7361 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02007362 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02007363 sizeof(int), 0644, proc_dointvec_minmax);
Ingo Molnara5d8c342008-10-09 11:35:51 +02007364 set_table_entry(&table[11], "name", sd->name,
7365 CORENAME_MAX_SIZE, 0444, proc_dostring);
7366 /* &table[12] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02007367
7368 return table;
7369}
7370
Ingo Molnar9a4e7152007-11-28 15:52:56 +01007371static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02007372{
7373 struct ctl_table *entry, *table;
7374 struct sched_domain *sd;
7375 int domain_num = 0, i;
7376 char buf[32];
7377
7378 for_each_domain(cpu, sd)
7379 domain_num++;
7380 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02007381 if (table == NULL)
7382 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02007383
7384 i = 0;
7385 for_each_domain(cpu, sd) {
7386 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007387 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007388 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007389 entry->child = sd_alloc_ctl_domain_table(sd);
7390 entry++;
7391 i++;
7392 }
7393 return table;
7394}
7395
7396static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02007397static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007398{
7399 int i, cpu_num = num_online_cpus();
7400 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7401 char buf[32];
7402
Milton Miller73785472007-10-24 18:23:48 +02007403 WARN_ON(sd_ctl_dir[0].child);
7404 sd_ctl_dir[0].child = entry;
7405
Milton Millerad1cdc12007-10-15 17:00:19 +02007406 if (entry == NULL)
7407 return;
7408
Milton Miller97b6ea72007-10-15 17:00:19 +02007409 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02007410 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02007411 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02007412 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02007413 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02007414 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02007415 }
Milton Miller73785472007-10-24 18:23:48 +02007416
7417 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02007418 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7419}
Milton Miller6382bc92007-10-15 17:00:19 +02007420
Milton Miller73785472007-10-24 18:23:48 +02007421/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02007422static void unregister_sched_domain_sysctl(void)
7423{
Milton Miller73785472007-10-24 18:23:48 +02007424 if (sd_sysctl_header)
7425 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02007426 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007427 if (sd_ctl_dir[0].child)
7428 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02007429}
Nick Piggine692ab52007-07-26 13:40:43 +02007430#else
Milton Miller6382bc92007-10-15 17:00:19 +02007431static void register_sched_domain_sysctl(void)
7432{
7433}
7434static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02007435{
7436}
7437#endif
7438
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007439static void set_rq_online(struct rq *rq)
7440{
7441 if (!rq->online) {
7442 const struct sched_class *class;
7443
Rusty Russellc6c49272008-11-25 02:35:05 +10307444 cpumask_set_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007445 rq->online = 1;
7446
7447 for_each_class(class) {
7448 if (class->rq_online)
7449 class->rq_online(rq);
7450 }
7451 }
7452}
7453
7454static void set_rq_offline(struct rq *rq)
7455{
7456 if (rq->online) {
7457 const struct sched_class *class;
7458
7459 for_each_class(class) {
7460 if (class->rq_offline)
7461 class->rq_offline(rq);
7462 }
7463
Rusty Russellc6c49272008-11-25 02:35:05 +10307464 cpumask_clear_cpu(rq->cpu, rq->rd->online);
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007465 rq->online = 0;
7466 }
7467}
7468
Linus Torvalds1da177e2005-04-16 15:20:36 -07007469/*
7470 * migration_call - callback that gets triggered when a CPU is added.
7471 * Here we can start up the necessary migration thread for the new CPU.
7472 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007473static int __cpuinit
7474migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007475{
Linus Torvalds1da177e2005-04-16 15:20:36 -07007476 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007477 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007478 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07007479 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007480
7481 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07007482
Linus Torvalds1da177e2005-04-16 15:20:36 -07007483 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007484 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02007485 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007486 if (IS_ERR(p))
7487 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007488 kthread_bind(p, cpu);
7489 /* Must be high prio: stop_machine expects to yield to it. */
7490 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02007491 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007492 task_rq_unlock(rq, &flags);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007493 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007494 cpu_rq(cpu)->migration_thread = p;
7495 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007496
Linus Torvalds1da177e2005-04-16 15:20:36 -07007497 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007498 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02007499 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007500 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007501
7502 /* Update our root-domain */
7503 rq = cpu_rq(cpu);
7504 spin_lock_irqsave(&rq->lock, flags);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007505 rq->calc_load_update = calc_load_update;
7506 rq->calc_load_active = 0;
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007507 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307508 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007509
7510 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04007511 }
7512 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007513 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007514
Linus Torvalds1da177e2005-04-16 15:20:36 -07007515#ifdef CONFIG_HOTPLUG_CPU
7516 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007517 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07007518 if (!cpu_rq(cpu)->migration_thread)
7519 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007520 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08007521 kthread_bind(cpu_rq(cpu)->migration_thread,
Rusty Russell1e5ce4f2008-11-25 02:35:03 +10307522 cpumask_any(cpu_online_mask));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007523 kthread_stop(cpu_rq(cpu)->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007524 put_task_struct(cpu_rq(cpu)->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007525 cpu_rq(cpu)->migration_thread = NULL;
7526 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007527
Linus Torvalds1da177e2005-04-16 15:20:36 -07007528 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007529 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07007530 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007531 migrate_live_tasks(cpu);
7532 rq = cpu_rq(cpu);
7533 kthread_stop(rq->migration_thread);
Oleg Nesterov371cbb32009-06-17 16:27:45 -07007534 put_task_struct(rq->migration_thread);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007535 rq->migration_thread = NULL;
7536 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007537 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02007538 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02007539 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007540 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02007541 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7542 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007543 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07007544 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07007545 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07007546 migrate_nr_uninterruptible(rq);
7547 BUG_ON(rq->nr_running != 0);
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02007548 calc_global_load_remove(rq);
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007549 /*
7550 * No need to migrate the tasks: it was best-effort if
7551 * they didn't take sched_hotcpu_mutex. Just wake up
7552 * the requestors.
7553 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007554 spin_lock_irq(&rq->lock);
7555 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07007556 struct migration_req *req;
7557
Linus Torvalds1da177e2005-04-16 15:20:36 -07007558 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07007559 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007560 list_del_init(&req->list);
Brian King9a2bd242008-12-09 08:47:00 -06007561 spin_unlock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007562 complete(&req->done);
Brian King9a2bd242008-12-09 08:47:00 -06007563 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007564 }
7565 spin_unlock_irq(&rq->lock);
7566 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007567
Gregory Haskins08f503b2008-03-10 17:59:11 -04007568 case CPU_DYING:
7569 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01007570 /* Update our root-domain */
7571 rq = cpu_rq(cpu);
7572 spin_lock_irqsave(&rq->lock, flags);
7573 if (rq->rd) {
Rusty Russellc6c49272008-11-25 02:35:05 +10307574 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007575 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007576 }
7577 spin_unlock_irqrestore(&rq->lock, flags);
7578 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007579#endif
7580 }
7581 return NOTIFY_OK;
7582}
7583
Paul Mackerrasf38b0822009-06-02 21:05:16 +10007584/*
7585 * Register at high priority so that task migration (migrate_all_tasks)
7586 * happens before everything else. This has to be lower priority than
7587 * the notifier in the perf_counter subsystem, though.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007588 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07007589static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007590 .notifier_call = migration_call,
7591 .priority = 10
7592};
7593
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007594static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007595{
7596 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07007597 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007598
7599 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07007600 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7601 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007602 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7603 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007604
7605 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007606}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07007607early_initcall(migration_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007608#endif
7609
7610#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07007611
Ingo Molnar3e9830d2007-10-15 17:00:13 +02007612#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007613
Mike Travis7c16ec52008-04-04 18:11:11 -07007614static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
Rusty Russell96f874e2008-11-25 02:35:14 +10307615 struct cpumask *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007616{
7617 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07007618 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007619
Rusty Russell968ea6d2008-12-13 21:55:51 +10307620 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
Rusty Russell96f874e2008-11-25 02:35:14 +10307621 cpumask_clear(groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007622
7623 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7624
7625 if (!(sd->flags & SD_LOAD_BALANCE)) {
7626 printk("does not load-balance\n");
7627 if (sd->parent)
7628 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7629 " has parent");
7630 return -1;
7631 }
7632
Li Zefaneefd7962008-11-04 16:15:37 +08007633 printk(KERN_CONT "span %s level %s\n", str, sd->name);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007634
Rusty Russell758b2cd2008-11-25 02:35:04 +10307635 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007636 printk(KERN_ERR "ERROR: domain->span does not contain "
7637 "CPU%d\n", cpu);
7638 }
Rusty Russell758b2cd2008-11-25 02:35:04 +10307639 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007640 printk(KERN_ERR "ERROR: domain->groups does not contain"
7641 " CPU%d\n", cpu);
7642 }
7643
7644 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7645 do {
7646 if (!group) {
7647 printk("\n");
7648 printk(KERN_ERR "ERROR: group is NULL\n");
7649 break;
7650 }
7651
7652 if (!group->__cpu_power) {
7653 printk(KERN_CONT "\n");
7654 printk(KERN_ERR "ERROR: domain->cpu_power not "
7655 "set\n");
7656 break;
7657 }
7658
Rusty Russell758b2cd2008-11-25 02:35:04 +10307659 if (!cpumask_weight(sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007660 printk(KERN_CONT "\n");
7661 printk(KERN_ERR "ERROR: empty group\n");
7662 break;
7663 }
7664
Rusty Russell758b2cd2008-11-25 02:35:04 +10307665 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007666 printk(KERN_CONT "\n");
7667 printk(KERN_ERR "ERROR: repeated CPUs\n");
7668 break;
7669 }
7670
Rusty Russell758b2cd2008-11-25 02:35:04 +10307671 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007672
Rusty Russell968ea6d2008-12-13 21:55:51 +10307673 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
Gautham R Shenoy381512c2009-04-14 09:09:36 +05307674
7675 printk(KERN_CONT " %s", str);
7676 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7677 printk(KERN_CONT " (__cpu_power = %d)",
7678 group->__cpu_power);
7679 }
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007680
7681 group = group->next;
7682 } while (group != sd->groups);
7683 printk(KERN_CONT "\n");
7684
Rusty Russell758b2cd2008-11-25 02:35:04 +10307685 if (!cpumask_equal(sched_domain_span(sd), groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007686 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7687
Rusty Russell758b2cd2008-11-25 02:35:04 +10307688 if (sd->parent &&
7689 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007690 printk(KERN_ERR "ERROR: parent span is not a superset "
7691 "of domain->span\n");
7692 return 0;
7693}
7694
Linus Torvalds1da177e2005-04-16 15:20:36 -07007695static void sched_domain_debug(struct sched_domain *sd, int cpu)
7696{
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307697 cpumask_var_t groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007698 int level = 0;
7699
Nick Piggin41c7ce92005-06-25 14:57:24 -07007700 if (!sd) {
7701 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7702 return;
7703 }
7704
Linus Torvalds1da177e2005-04-16 15:20:36 -07007705 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7706
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307707 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007708 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7709 return;
7710 }
7711
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007712 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007713 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007714 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007715 level++;
7716 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08007717 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02007718 break;
7719 }
Rusty Russelld5dd3db2008-11-25 02:35:12 +10307720 free_cpumask_var(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007721}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007722#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07007723# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007724#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007725
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007726static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007727{
Rusty Russell758b2cd2008-11-25 02:35:04 +10307728 if (cpumask_weight(sched_domain_span(sd)) == 1)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007729 return 1;
7730
7731 /* Following flags need at least 2 groups */
7732 if (sd->flags & (SD_LOAD_BALANCE |
7733 SD_BALANCE_NEWIDLE |
7734 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007735 SD_BALANCE_EXEC |
7736 SD_SHARE_CPUPOWER |
7737 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007738 if (sd->groups != sd->groups->next)
7739 return 0;
7740 }
7741
7742 /* Following flags don't use groups */
7743 if (sd->flags & (SD_WAKE_IDLE |
7744 SD_WAKE_AFFINE |
7745 SD_WAKE_BALANCE))
7746 return 0;
7747
7748 return 1;
7749}
7750
Ingo Molnar48f24c42006-07-03 00:25:40 -07007751static int
7752sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07007753{
7754 unsigned long cflags = sd->flags, pflags = parent->flags;
7755
7756 if (sd_degenerate(parent))
7757 return 1;
7758
Rusty Russell758b2cd2008-11-25 02:35:04 +10307759 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
Suresh Siddha245af2c2005-06-25 14:57:25 -07007760 return 0;
7761
7762 /* Does parent contain flags not in child? */
7763 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7764 if (cflags & SD_WAKE_AFFINE)
7765 pflags &= ~SD_WAKE_BALANCE;
7766 /* Flags needing groups don't count if only 1 group in parent */
7767 if (parent->groups == parent->groups->next) {
7768 pflags &= ~(SD_LOAD_BALANCE |
7769 SD_BALANCE_NEWIDLE |
7770 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007771 SD_BALANCE_EXEC |
7772 SD_SHARE_CPUPOWER |
7773 SD_SHARE_PKG_RESOURCES);
Ken Chen54364992008-12-07 18:47:37 -08007774 if (nr_node_ids == 1)
7775 pflags &= ~SD_SERIALIZE;
Suresh Siddha245af2c2005-06-25 14:57:25 -07007776 }
7777 if (~cflags & pflags)
7778 return 0;
7779
7780 return 1;
7781}
7782
Rusty Russellc6c49272008-11-25 02:35:05 +10307783static void free_rootdomain(struct root_domain *rd)
7784{
Rusty Russell68e74562008-11-25 02:35:13 +10307785 cpupri_cleanup(&rd->cpupri);
7786
Rusty Russellc6c49272008-11-25 02:35:05 +10307787 free_cpumask_var(rd->rto_mask);
7788 free_cpumask_var(rd->online);
7789 free_cpumask_var(rd->span);
7790 kfree(rd);
7791}
7792
Gregory Haskins57d885f2008-01-25 21:08:18 +01007793static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7794{
Ingo Molnara0490fa2009-02-12 11:35:40 +01007795 struct root_domain *old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007796 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007797
7798 spin_lock_irqsave(&rq->lock, flags);
7799
7800 if (rq->rd) {
Ingo Molnara0490fa2009-02-12 11:35:40 +01007801 old_rd = rq->rd;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007802
Rusty Russellc6c49272008-11-25 02:35:05 +10307803 if (cpumask_test_cpu(rq->cpu, old_rd->online))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007804 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007805
Rusty Russellc6c49272008-11-25 02:35:05 +10307806 cpumask_clear_cpu(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01007807
Ingo Molnara0490fa2009-02-12 11:35:40 +01007808 /*
7809 * If we dont want to free the old_rt yet then
7810 * set old_rd to NULL to skip the freeing later
7811 * in this function:
7812 */
7813 if (!atomic_dec_and_test(&old_rd->refcount))
7814 old_rd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007815 }
7816
7817 atomic_inc(&rd->refcount);
7818 rq->rd = rd;
7819
Rusty Russellc6c49272008-11-25 02:35:05 +10307820 cpumask_set_cpu(rq->cpu, rd->span);
7821 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04007822 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007823
7824 spin_unlock_irqrestore(&rq->lock, flags);
Ingo Molnara0490fa2009-02-12 11:35:40 +01007825
7826 if (old_rd)
7827 free_rootdomain(old_rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01007828}
7829
Li Zefanfd5e1b52009-06-15 13:34:19 +08007830static int init_rootdomain(struct root_domain *rd, bool bootmem)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007831{
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007832 gfp_t gfp = GFP_KERNEL;
7833
Gregory Haskins57d885f2008-01-25 21:08:18 +01007834 memset(rd, 0, sizeof(*rd));
7835
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007836 if (bootmem)
7837 gfp = GFP_NOWAIT;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02007838
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007839 if (!alloc_cpumask_var(&rd->span, gfp))
Li Zefan0c910d22009-01-06 17:39:06 +08007840 goto out;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007841 if (!alloc_cpumask_var(&rd->online, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10307842 goto free_span;
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03007843 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
Rusty Russellc6c49272008-11-25 02:35:05 +10307844 goto free_online;
Gregory Haskins6e0534f2008-05-12 21:21:01 +02007845
Pekka Enberg0fb53022009-06-11 08:41:22 +03007846 if (cpupri_init(&rd->cpupri, bootmem) != 0)
Rusty Russell68e74562008-11-25 02:35:13 +10307847 goto free_rto_mask;
Rusty Russellc6c49272008-11-25 02:35:05 +10307848 return 0;
7849
Rusty Russell68e74562008-11-25 02:35:13 +10307850free_rto_mask:
7851 free_cpumask_var(rd->rto_mask);
Rusty Russellc6c49272008-11-25 02:35:05 +10307852free_online:
7853 free_cpumask_var(rd->online);
7854free_span:
7855 free_cpumask_var(rd->span);
Li Zefan0c910d22009-01-06 17:39:06 +08007856out:
Rusty Russellc6c49272008-11-25 02:35:05 +10307857 return -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007858}
7859
7860static void init_defrootdomain(void)
7861{
Rusty Russellc6c49272008-11-25 02:35:05 +10307862 init_rootdomain(&def_root_domain, true);
7863
Gregory Haskins57d885f2008-01-25 21:08:18 +01007864 atomic_set(&def_root_domain.refcount, 1);
7865}
7866
Gregory Haskinsdc938522008-01-25 21:08:26 +01007867static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007868{
7869 struct root_domain *rd;
7870
7871 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7872 if (!rd)
7873 return NULL;
7874
Rusty Russellc6c49272008-11-25 02:35:05 +10307875 if (init_rootdomain(rd, false) != 0) {
7876 kfree(rd);
7877 return NULL;
7878 }
Gregory Haskins57d885f2008-01-25 21:08:18 +01007879
7880 return rd;
7881}
7882
Linus Torvalds1da177e2005-04-16 15:20:36 -07007883/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01007884 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07007885 * hold the hotplug lock.
7886 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01007887static void
7888cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007889{
Ingo Molnar70b97a72006-07-03 00:25:42 -07007890 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07007891 struct sched_domain *tmp;
7892
7893 /* Remove the sched domains which do not contribute to scheduling. */
Li Zefanf29c9b12008-11-06 09:45:16 +08007894 for (tmp = sd; tmp; ) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007895 struct sched_domain *parent = tmp->parent;
7896 if (!parent)
7897 break;
Li Zefanf29c9b12008-11-06 09:45:16 +08007898
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007899 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007900 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007901 if (parent->parent)
7902 parent->parent->child = tmp;
Li Zefanf29c9b12008-11-06 09:45:16 +08007903 } else
7904 tmp = tmp->parent;
Suresh Siddha245af2c2005-06-25 14:57:25 -07007905 }
7906
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007907 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07007908 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007909 if (sd)
7910 sd->child = NULL;
7911 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007912
7913 sched_domain_debug(sd, cpu);
7914
Gregory Haskins57d885f2008-01-25 21:08:18 +01007915 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07007916 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007917}
7918
7919/* cpus with isolated domains */
Rusty Russelldcc30a32008-11-25 02:35:12 +10307920static cpumask_var_t cpu_isolated_map;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007921
7922/* Setup the mask of cpus configured for isolated domains */
7923static int __init isolated_cpu_setup(char *str)
7924{
Rusty Russell968ea6d2008-12-13 21:55:51 +10307925 cpulist_parse(str, cpu_isolated_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007926 return 1;
7927}
7928
Ingo Molnar8927f492007-10-15 17:00:13 +02007929__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007930
7931/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007932 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7933 * to a function which identifies what group(along with sched group) a CPU
Rusty Russell96f874e2008-11-25 02:35:14 +10307934 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7935 * (due to the fact that we keep track of groups covered with a struct cpumask).
Linus Torvalds1da177e2005-04-16 15:20:36 -07007936 *
7937 * init_sched_build_groups will build a circular linked list of the groups
7938 * covered by the given span, and will set each group's ->cpumask correctly,
7939 * and ->cpu_power to 0.
7940 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007941static void
Rusty Russell96f874e2008-11-25 02:35:14 +10307942init_sched_build_groups(const struct cpumask *span,
7943 const struct cpumask *cpu_map,
7944 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007945 struct sched_group **sg,
Rusty Russell96f874e2008-11-25 02:35:14 +10307946 struct cpumask *tmpmask),
7947 struct cpumask *covered, struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007948{
7949 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007950 int i;
7951
Rusty Russell96f874e2008-11-25 02:35:14 +10307952 cpumask_clear(covered);
Mike Travis7c16ec52008-04-04 18:11:11 -07007953
Rusty Russellabcd0832008-11-25 02:35:02 +10307954 for_each_cpu(i, span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007955 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007956 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007957 int j;
7958
Rusty Russell758b2cd2008-11-25 02:35:04 +10307959 if (cpumask_test_cpu(i, covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007960 continue;
7961
Rusty Russell758b2cd2008-11-25 02:35:04 +10307962 cpumask_clear(sched_group_cpus(sg));
Eric Dumazet5517d862007-05-08 00:32:57 -07007963 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007964
Rusty Russellabcd0832008-11-25 02:35:02 +10307965 for_each_cpu(j, span) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007966 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007967 continue;
7968
Rusty Russell96f874e2008-11-25 02:35:14 +10307969 cpumask_set_cpu(j, covered);
Rusty Russell758b2cd2008-11-25 02:35:04 +10307970 cpumask_set_cpu(j, sched_group_cpus(sg));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007971 }
7972 if (!first)
7973 first = sg;
7974 if (last)
7975 last->next = sg;
7976 last = sg;
7977 }
7978 last->next = first;
7979}
7980
John Hawkes9c1cfda2005-09-06 15:18:14 -07007981#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07007982
John Hawkes9c1cfda2005-09-06 15:18:14 -07007983#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08007984
John Hawkes9c1cfda2005-09-06 15:18:14 -07007985/**
7986 * find_next_best_node - find the next node to include in a sched_domain
7987 * @node: node whose sched_domain we're building
7988 * @used_nodes: nodes already in the sched_domain
7989 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007990 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07007991 * finds the closest node not already in the @used_nodes map.
7992 *
7993 * Should use nodemask_t.
7994 */
Mike Travisc5f59f02008-04-04 18:11:10 -07007995static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07007996{
7997 int i, n, val, min_val, best_node = 0;
7998
7999 min_val = INT_MAX;
8000
Mike Travis076ac2a2008-05-12 21:21:12 +02008001 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008002 /* Start at @node */
Mike Travis076ac2a2008-05-12 21:21:12 +02008003 n = (node + i) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008004
8005 if (!nr_cpus_node(n))
8006 continue;
8007
8008 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07008009 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008010 continue;
8011
8012 /* Simple min distance search */
8013 val = node_distance(node, n);
8014
8015 if (val < min_val) {
8016 min_val = val;
8017 best_node = n;
8018 }
8019 }
8020
Mike Travisc5f59f02008-04-04 18:11:10 -07008021 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008022 return best_node;
8023}
8024
8025/**
8026 * sched_domain_node_span - get a cpumask for a node's sched_domain
8027 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07008028 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07008029 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008030 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07008031 * should be one that prevents unnecessary balancing, but also spreads tasks
8032 * out optimally.
8033 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308034static void sched_domain_node_span(int node, struct cpumask *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07008035{
Mike Travisc5f59f02008-04-04 18:11:10 -07008036 nodemask_t used_nodes;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008037 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008038
Mike Travis6ca09df2008-12-31 18:08:45 -08008039 cpumask_clear(span);
Mike Travisc5f59f02008-04-04 18:11:10 -07008040 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008041
Mike Travis6ca09df2008-12-31 18:08:45 -08008042 cpumask_or(span, span, cpumask_of_node(node));
Mike Travisc5f59f02008-04-04 18:11:10 -07008043 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008044
8045 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07008046 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008047
Mike Travis6ca09df2008-12-31 18:08:45 -08008048 cpumask_or(span, span, cpumask_of_node(next_node));
John Hawkes9c1cfda2005-09-06 15:18:14 -07008049 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008050}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008051#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07008052
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008053int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008054
John Hawkes9c1cfda2005-09-06 15:18:14 -07008055/*
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308056 * The cpus mask in sched_group and sched_domain hangs off the end.
Ingo Molnar4200efd2009-05-19 09:22:19 +02008057 *
8058 * ( See the the comments in include/linux/sched.h:struct sched_group
8059 * and struct sched_domain. )
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308060 */
8061struct static_sched_group {
8062 struct sched_group sg;
8063 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8064};
8065
8066struct static_sched_domain {
8067 struct sched_domain sd;
8068 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8069};
8070
8071/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07008072 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07008073 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008074#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308075static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8076static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008077
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008078static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308079cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8080 struct sched_group **sg, struct cpumask *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008081{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008082 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308083 *sg = &per_cpu(sched_group_cpus, cpu).sg;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008084 return cpu;
8085}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008086#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008087
Ingo Molnar48f24c42006-07-03 00:25:40 -07008088/*
8089 * multi-core sched-domains:
8090 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008091#ifdef CONFIG_SCHED_MC
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308092static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8093static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008094#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008095
8096#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008097static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308098cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8099 struct sched_group **sg, struct cpumask *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008100{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008101 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07008102
Rusty Russellc69fc562009-03-13 14:49:46 +10308103 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308104 group = cpumask_first(mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008105 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308106 *sg = &per_cpu(sched_group_core, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008107 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008108}
8109#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008110static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308111cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8112 struct sched_group **sg, struct cpumask *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008113{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008114 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308115 *sg = &per_cpu(sched_group_core, cpu).sg;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008116 return cpu;
8117}
8118#endif
8119
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308120static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8121static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07008122
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008123static int
Rusty Russell96f874e2008-11-25 02:35:14 +10308124cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8125 struct sched_group **sg, struct cpumask *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008126{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008127 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008128#ifdef CONFIG_SCHED_MC
Mike Travis6ca09df2008-12-31 18:08:45 -08008129 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308130 group = cpumask_first(mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008131#elif defined(CONFIG_SCHED_SMT)
Rusty Russellc69fc562009-03-13 14:49:46 +10308132 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308133 group = cpumask_first(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008134#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008135 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008136#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008137 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308138 *sg = &per_cpu(sched_group_phys, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008139 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008140}
8141
8142#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07008143/*
8144 * The init_sched_build_groups can't handle what we want to do with node
8145 * groups, so roll our own. Now each node has its own list of groups which
8146 * gets dynamically allocated.
8147 */
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008148static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07008149static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008150
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008151static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308152static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008153
Rusty Russell96f874e2008-11-25 02:35:14 +10308154static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8155 struct sched_group **sg,
8156 struct cpumask *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008157{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008158 int group;
8159
Mike Travis6ca09df2008-12-31 18:08:45 -08008160 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308161 group = cpumask_first(nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008162
8163 if (sg)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308164 *sg = &per_cpu(sched_group_allnodes, group).sg;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008165 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008166}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008167
Siddha, Suresh B08069032006-03-27 01:15:23 -08008168static void init_numa_sched_groups_power(struct sched_group *group_head)
8169{
8170 struct sched_group *sg = group_head;
8171 int j;
8172
8173 if (!sg)
8174 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02008175 do {
Rusty Russell758b2cd2008-11-25 02:35:04 +10308176 for_each_cpu(j, sched_group_cpus(sg)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008177 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08008178
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308179 sd = &per_cpu(phys_domains, j).sd;
Miao Xie13318a72009-04-15 09:59:10 +08008180 if (j != group_first_cpu(sd->groups)) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02008181 /*
8182 * Only add "power" once for each
8183 * physical package.
8184 */
8185 continue;
8186 }
8187
8188 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08008189 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02008190 sg = sg->next;
8191 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08008192}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008193#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07008194
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008195#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008196/* Free memory allocated for various sched_group structures */
Rusty Russell96f874e2008-11-25 02:35:14 +10308197static void free_sched_groups(const struct cpumask *cpu_map,
8198 struct cpumask *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008199{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008200 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008201
Rusty Russellabcd0832008-11-25 02:35:02 +10308202 for_each_cpu(cpu, cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008203 struct sched_group **sched_group_nodes
8204 = sched_group_nodes_bycpu[cpu];
8205
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008206 if (!sched_group_nodes)
8207 continue;
8208
Mike Travis076ac2a2008-05-12 21:21:12 +02008209 for (i = 0; i < nr_node_ids; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008210 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8211
Mike Travis6ca09df2008-12-31 18:08:45 -08008212 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308213 if (cpumask_empty(nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008214 continue;
8215
8216 if (sg == NULL)
8217 continue;
8218 sg = sg->next;
8219next_sg:
8220 oldsg = sg;
8221 sg = sg->next;
8222 kfree(oldsg);
8223 if (oldsg != sched_group_nodes[i])
8224 goto next_sg;
8225 }
8226 kfree(sched_group_nodes);
8227 sched_group_nodes_bycpu[cpu] = NULL;
8228 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008229}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008230#else /* !CONFIG_NUMA */
Rusty Russell96f874e2008-11-25 02:35:14 +10308231static void free_sched_groups(const struct cpumask *cpu_map,
8232 struct cpumask *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008233{
8234}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008235#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008236
Linus Torvalds1da177e2005-04-16 15:20:36 -07008237/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008238 * Initialize sched groups cpu_power.
8239 *
8240 * cpu_power indicates the capacity of sched group, which is used while
8241 * distributing the load between different sched groups in a sched domain.
8242 * Typically cpu_power for all the groups in a sched domain will be same unless
8243 * there are asymmetries in the topology. If there are asymmetries, group
8244 * having more cpu_power will pickup more load compared to the group having
8245 * less cpu_power.
8246 *
8247 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8248 * the maximum number of tasks a group can handle in the presence of other idle
8249 * or lightly loaded groups in the same sched domain.
8250 */
8251static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8252{
8253 struct sched_domain *child;
8254 struct sched_group *group;
8255
8256 WARN_ON(!sd || !sd->groups);
8257
Miao Xie13318a72009-04-15 09:59:10 +08008258 if (cpu != group_first_cpu(sd->groups))
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008259 return;
8260
8261 child = sd->child;
8262
Eric Dumazet5517d862007-05-08 00:32:57 -07008263 sd->groups->__cpu_power = 0;
8264
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008265 /*
8266 * For perf policy, if the groups in child domain share resources
8267 * (for example cores sharing some portions of the cache hierarchy
8268 * or SMT), then set this domain groups cpu_power such that each group
8269 * can handle only one task, when there are other idle groups in the
8270 * same sched domain.
8271 */
8272 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8273 (child->flags &
8274 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07008275 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008276 return;
8277 }
8278
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008279 /*
8280 * add cpu_power of each child group to this groups cpu_power
8281 */
8282 group = child->groups;
8283 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07008284 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008285 group = group->next;
8286 } while (group != child->groups);
8287}
8288
8289/*
Mike Travis7c16ec52008-04-04 18:11:11 -07008290 * Initializers for schedule domains
8291 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8292 */
8293
Ingo Molnara5d8c342008-10-09 11:35:51 +02008294#ifdef CONFIG_SCHED_DEBUG
8295# define SD_INIT_NAME(sd, type) sd->name = #type
8296#else
8297# define SD_INIT_NAME(sd, type) do { } while (0)
8298#endif
8299
Mike Travis7c16ec52008-04-04 18:11:11 -07008300#define SD_INIT(sd, type) sd_init_##type(sd)
Ingo Molnara5d8c342008-10-09 11:35:51 +02008301
Mike Travis7c16ec52008-04-04 18:11:11 -07008302#define SD_INIT_FUNC(type) \
8303static noinline void sd_init_##type(struct sched_domain *sd) \
8304{ \
8305 memset(sd, 0, sizeof(*sd)); \
8306 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008307 sd->level = SD_LV_##type; \
Ingo Molnara5d8c342008-10-09 11:35:51 +02008308 SD_INIT_NAME(sd, type); \
Mike Travis7c16ec52008-04-04 18:11:11 -07008309}
8310
8311SD_INIT_FUNC(CPU)
8312#ifdef CONFIG_NUMA
8313 SD_INIT_FUNC(ALLNODES)
8314 SD_INIT_FUNC(NODE)
8315#endif
8316#ifdef CONFIG_SCHED_SMT
8317 SD_INIT_FUNC(SIBLING)
8318#endif
8319#ifdef CONFIG_SCHED_MC
8320 SD_INIT_FUNC(MC)
8321#endif
8322
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008323static int default_relax_domain_level = -1;
8324
8325static int __init setup_relax_domain_level(char *str)
8326{
Li Zefan30e0e172008-05-13 10:27:17 +08008327 unsigned long val;
8328
8329 val = simple_strtoul(str, NULL, 0);
8330 if (val < SD_LV_MAX)
8331 default_relax_domain_level = val;
8332
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008333 return 1;
8334}
8335__setup("relax_domain_level=", setup_relax_domain_level);
8336
8337static void set_domain_attribute(struct sched_domain *sd,
8338 struct sched_domain_attr *attr)
8339{
8340 int request;
8341
8342 if (!attr || attr->relax_domain_level < 0) {
8343 if (default_relax_domain_level < 0)
8344 return;
8345 else
8346 request = default_relax_domain_level;
8347 } else
8348 request = attr->relax_domain_level;
8349 if (request < sd->level) {
8350 /* turn off idle balance on this domain */
8351 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8352 } else {
8353 /* turn on idle balance on this domain */
8354 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8355 }
8356}
8357
Mike Travis7c16ec52008-04-04 18:11:11 -07008358/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008359 * Build sched domains for a given set of cpus and attach the sched domains
8360 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07008361 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308362static int __build_sched_domains(const struct cpumask *cpu_map,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008363 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008364{
Rusty Russell3404c8d2008-11-25 02:35:03 +10308365 int i, err = -ENOMEM;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008366 struct root_domain *rd;
Rusty Russell3404c8d2008-11-25 02:35:03 +10308367 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8368 tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07008369#ifdef CONFIG_NUMA
Rusty Russell3404c8d2008-11-25 02:35:03 +10308370 cpumask_var_t domainspan, covered, notcovered;
John Hawkesd1b55132005-09-06 15:18:14 -07008371 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008372 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07008373
Rusty Russell3404c8d2008-11-25 02:35:03 +10308374 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8375 goto out;
8376 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8377 goto free_domainspan;
8378 if (!alloc_cpumask_var(&notcovered, GFP_KERNEL))
8379 goto free_covered;
8380#endif
8381
8382 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8383 goto free_notcovered;
8384 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8385 goto free_nodemask;
8386 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8387 goto free_this_sibling_map;
8388 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8389 goto free_this_core_map;
8390 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8391 goto free_send_covered;
8392
8393#ifdef CONFIG_NUMA
John Hawkesd1b55132005-09-06 15:18:14 -07008394 /*
8395 * Allocate the per-node list of sched groups
8396 */
Mike Travis076ac2a2008-05-12 21:21:12 +02008397 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008398 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07008399 if (!sched_group_nodes) {
8400 printk(KERN_WARNING "Can not alloc sched group node list\n");
Rusty Russell3404c8d2008-11-25 02:35:03 +10308401 goto free_tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07008402 }
John Hawkesd1b55132005-09-06 15:18:14 -07008403#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008404
Gregory Haskinsdc938522008-01-25 21:08:26 +01008405 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01008406 if (!rd) {
8407 printk(KERN_WARNING "Cannot alloc root domain\n");
Rusty Russell3404c8d2008-11-25 02:35:03 +10308408 goto free_sched_groups;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008409 }
8410
Mike Travis7c16ec52008-04-04 18:11:11 -07008411#ifdef CONFIG_NUMA
Rusty Russell96f874e2008-11-25 02:35:14 +10308412 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
Mike Travis7c16ec52008-04-04 18:11:11 -07008413#endif
8414
Linus Torvalds1da177e2005-04-16 15:20:36 -07008415 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008416 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008417 */
Rusty Russellabcd0832008-11-25 02:35:02 +10308418 for_each_cpu(i, cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008419 struct sched_domain *sd = NULL, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008420
Mike Travis6ca09df2008-12-31 18:08:45 -08008421 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008422
8423#ifdef CONFIG_NUMA
Rusty Russell96f874e2008-11-25 02:35:14 +10308424 if (cpumask_weight(cpu_map) >
8425 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008426 sd = &per_cpu(allnodes_domains, i).sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008427 SD_INIT(sd, ALLNODES);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008428 set_domain_attribute(sd, attr);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308429 cpumask_copy(sched_domain_span(sd), cpu_map);
Mike Travis7c16ec52008-04-04 18:11:11 -07008430 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008431 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008432 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008433 } else
8434 p = NULL;
8435
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008436 sd = &per_cpu(node_domains, i).sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008437 SD_INIT(sd, NODE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008438 set_domain_attribute(sd, attr);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308439 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
John Hawkes9c1cfda2005-09-06 15:18:14 -07008440 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008441 if (p)
8442 p->child = sd;
Rusty Russell758b2cd2008-11-25 02:35:04 +10308443 cpumask_and(sched_domain_span(sd),
8444 sched_domain_span(sd), cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008445#endif
8446
8447 p = sd;
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308448 sd = &per_cpu(phys_domains, i).sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008449 SD_INIT(sd, CPU);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008450 set_domain_attribute(sd, attr);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308451 cpumask_copy(sched_domain_span(sd), nodemask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008452 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008453 if (p)
8454 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008455 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008456
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008457#ifdef CONFIG_SCHED_MC
8458 p = sd;
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308459 sd = &per_cpu(core_domains, i).sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008460 SD_INIT(sd, MC);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008461 set_domain_attribute(sd, attr);
Mike Travis6ca09df2008-12-31 18:08:45 -08008462 cpumask_and(sched_domain_span(sd), cpu_map,
8463 cpu_coregroup_mask(i));
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008464 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008465 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008466 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008467#endif
8468
Linus Torvalds1da177e2005-04-16 15:20:36 -07008469#ifdef CONFIG_SCHED_SMT
8470 p = sd;
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308471 sd = &per_cpu(cpu_domains, i).sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008472 SD_INIT(sd, SIBLING);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008473 set_domain_attribute(sd, attr);
Rusty Russell758b2cd2008-11-25 02:35:04 +10308474 cpumask_and(sched_domain_span(sd),
Rusty Russellc69fc562009-03-13 14:49:46 +10308475 topology_thread_cpumask(i), cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008476 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07008477 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07008478 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008479#endif
8480 }
8481
8482#ifdef CONFIG_SCHED_SMT
8483 /* Set up CPU (sibling) groups */
Rusty Russellabcd0832008-11-25 02:35:02 +10308484 for_each_cpu(i, cpu_map) {
Rusty Russell96f874e2008-11-25 02:35:14 +10308485 cpumask_and(this_sibling_map,
Rusty Russellc69fc562009-03-13 14:49:46 +10308486 topology_thread_cpumask(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308487 if (i != cpumask_first(this_sibling_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008488 continue;
8489
Ingo Molnardd41f592007-07-09 18:51:59 +02008490 init_sched_build_groups(this_sibling_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07008491 &cpu_to_cpu_group,
8492 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008493 }
8494#endif
8495
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008496#ifdef CONFIG_SCHED_MC
8497 /* Set up multi-core groups */
Rusty Russellabcd0832008-11-25 02:35:02 +10308498 for_each_cpu(i, cpu_map) {
Mike Travis6ca09df2008-12-31 18:08:45 -08008499 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308500 if (i != cpumask_first(this_core_map))
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008501 continue;
Mike Travis7c16ec52008-04-04 18:11:11 -07008502
Ingo Molnardd41f592007-07-09 18:51:59 +02008503 init_sched_build_groups(this_core_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07008504 &cpu_to_core_group,
8505 send_covered, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008506 }
8507#endif
8508
Linus Torvalds1da177e2005-04-16 15:20:36 -07008509 /* Set up physical groups */
Mike Travis076ac2a2008-05-12 21:21:12 +02008510 for (i = 0; i < nr_node_ids; i++) {
Mike Travis6ca09df2008-12-31 18:08:45 -08008511 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308512 if (cpumask_empty(nodemask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07008513 continue;
8514
Mike Travis7c16ec52008-04-04 18:11:11 -07008515 init_sched_build_groups(nodemask, cpu_map,
8516 &cpu_to_phys_group,
8517 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008518 }
8519
8520#ifdef CONFIG_NUMA
8521 /* Set up node groups */
Mike Travis7c16ec52008-04-04 18:11:11 -07008522 if (sd_allnodes) {
Mike Travis7c16ec52008-04-04 18:11:11 -07008523 init_sched_build_groups(cpu_map, cpu_map,
8524 &cpu_to_allnodes_group,
8525 send_covered, tmpmask);
8526 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008527
Mike Travis076ac2a2008-05-12 21:21:12 +02008528 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008529 /* Set up node groups */
8530 struct sched_group *sg, *prev;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008531 int j;
8532
Rusty Russell96f874e2008-11-25 02:35:14 +10308533 cpumask_clear(covered);
Mike Travis6ca09df2008-12-31 18:08:45 -08008534 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
Rusty Russell96f874e2008-11-25 02:35:14 +10308535 if (cpumask_empty(nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07008536 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008537 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07008538 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008539
Mike Travis4bdbaad32008-04-15 16:35:52 -07008540 sched_domain_node_span(i, domainspan);
Rusty Russell96f874e2008-11-25 02:35:14 +10308541 cpumask_and(domainspan, domainspan, cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008542
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308543 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8544 GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008545 if (!sg) {
8546 printk(KERN_WARNING "Can not alloc domain group for "
8547 "node %d\n", i);
8548 goto error;
8549 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008550 sched_group_nodes[i] = sg;
Rusty Russellabcd0832008-11-25 02:35:02 +10308551 for_each_cpu(j, nodemask) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07008552 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02008553
Rusty Russell62ea9ce2009-01-11 01:04:16 +01008554 sd = &per_cpu(node_domains, j).sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008555 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008556 }
Eric Dumazet5517d862007-05-08 00:32:57 -07008557 sg->__cpu_power = 0;
Rusty Russell758b2cd2008-11-25 02:35:04 +10308558 cpumask_copy(sched_group_cpus(sg), nodemask);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008559 sg->next = sg;
Rusty Russell96f874e2008-11-25 02:35:14 +10308560 cpumask_or(covered, covered, nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008561 prev = sg;
8562
Mike Travis076ac2a2008-05-12 21:21:12 +02008563 for (j = 0; j < nr_node_ids; j++) {
Mike Travis076ac2a2008-05-12 21:21:12 +02008564 int n = (i + j) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008565
Rusty Russell96f874e2008-11-25 02:35:14 +10308566 cpumask_complement(notcovered, covered);
8567 cpumask_and(tmpmask, notcovered, cpu_map);
8568 cpumask_and(tmpmask, tmpmask, domainspan);
8569 if (cpumask_empty(tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008570 break;
8571
Mike Travis6ca09df2008-12-31 18:08:45 -08008572 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
Rusty Russell96f874e2008-11-25 02:35:14 +10308573 if (cpumask_empty(tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07008574 continue;
8575
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308576 sg = kmalloc_node(sizeof(struct sched_group) +
8577 cpumask_size(),
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07008578 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008579 if (!sg) {
8580 printk(KERN_WARNING
8581 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008582 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07008583 }
Eric Dumazet5517d862007-05-08 00:32:57 -07008584 sg->__cpu_power = 0;
Rusty Russell758b2cd2008-11-25 02:35:04 +10308585 cpumask_copy(sched_group_cpus(sg), tmpmask);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008586 sg->next = prev->next;
Rusty Russell96f874e2008-11-25 02:35:14 +10308587 cpumask_or(covered, covered, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008588 prev->next = sg;
8589 prev = sg;
8590 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008591 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008592#endif
8593
8594 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008595#ifdef CONFIG_SCHED_SMT
Rusty Russellabcd0832008-11-25 02:35:02 +10308596 for_each_cpu(i, cpu_map) {
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308597 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02008598
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008599 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008600 }
8601#endif
8602#ifdef CONFIG_SCHED_MC
Rusty Russellabcd0832008-11-25 02:35:02 +10308603 for_each_cpu(i, cpu_map) {
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308604 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02008605
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008606 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008607 }
8608#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008609
Rusty Russellabcd0832008-11-25 02:35:02 +10308610 for_each_cpu(i, cpu_map) {
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308611 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02008612
Siddha, Suresh B89c47102006-10-03 01:14:09 -07008613 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008614 }
8615
John Hawkes9c1cfda2005-09-06 15:18:14 -07008616#ifdef CONFIG_NUMA
Mike Travis076ac2a2008-05-12 21:21:12 +02008617 for (i = 0; i < nr_node_ids; i++)
Siddha, Suresh B08069032006-03-27 01:15:23 -08008618 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008619
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08008620 if (sd_allnodes) {
8621 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07008622
Rusty Russell96f874e2008-11-25 02:35:14 +10308623 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
Mike Travis7c16ec52008-04-04 18:11:11 -07008624 tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07008625 init_numa_sched_groups_power(sg);
8626 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07008627#endif
8628
Linus Torvalds1da177e2005-04-16 15:20:36 -07008629 /* Attach the domains */
Rusty Russellabcd0832008-11-25 02:35:02 +10308630 for_each_cpu(i, cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008631 struct sched_domain *sd;
8632#ifdef CONFIG_SCHED_SMT
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308633 sd = &per_cpu(cpu_domains, i).sd;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08008634#elif defined(CONFIG_SCHED_MC)
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308635 sd = &per_cpu(core_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008636#else
Rusty Russell6c99e9a2008-11-25 02:35:04 +10308637 sd = &per_cpu(phys_domains, i).sd;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008638#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01008639 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008640 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008641
Rusty Russell3404c8d2008-11-25 02:35:03 +10308642 err = 0;
8643
8644free_tmpmask:
8645 free_cpumask_var(tmpmask);
8646free_send_covered:
8647 free_cpumask_var(send_covered);
8648free_this_core_map:
8649 free_cpumask_var(this_core_map);
8650free_this_sibling_map:
8651 free_cpumask_var(this_sibling_map);
8652free_nodemask:
8653 free_cpumask_var(nodemask);
8654free_notcovered:
8655#ifdef CONFIG_NUMA
8656 free_cpumask_var(notcovered);
8657free_covered:
8658 free_cpumask_var(covered);
8659free_domainspan:
8660 free_cpumask_var(domainspan);
8661out:
8662#endif
8663 return err;
8664
8665free_sched_groups:
8666#ifdef CONFIG_NUMA
8667 kfree(sched_group_nodes);
8668#endif
8669 goto free_tmpmask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008670
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008671#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07008672error:
Mike Travis7c16ec52008-04-04 18:11:11 -07008673 free_sched_groups(cpu_map, tmpmask);
Rusty Russellc6c49272008-11-25 02:35:05 +10308674 free_rootdomain(rd);
Rusty Russell3404c8d2008-11-25 02:35:03 +10308675 goto free_tmpmask;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07008676#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008677}
Paul Jackson029190c2007-10-18 23:40:20 -07008678
Rusty Russell96f874e2008-11-25 02:35:14 +10308679static int build_sched_domains(const struct cpumask *cpu_map)
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008680{
8681 return __build_sched_domains(cpu_map, NULL);
8682}
8683
Rusty Russell96f874e2008-11-25 02:35:14 +10308684static struct cpumask *doms_cur; /* current sched domains */
Paul Jackson029190c2007-10-18 23:40:20 -07008685static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02008686static struct sched_domain_attr *dattr_cur;
8687 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07008688
8689/*
8690 * Special case: If a kmalloc of a doms_cur partition (array of
Rusty Russell42128232008-11-25 02:35:12 +10308691 * cpumask) fails, then fallback to a single sched domain,
8692 * as determined by the single cpumask fallback_doms.
Paul Jackson029190c2007-10-18 23:40:20 -07008693 */
Rusty Russell42128232008-11-25 02:35:12 +10308694static cpumask_var_t fallback_doms;
Paul Jackson029190c2007-10-18 23:40:20 -07008695
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008696/*
8697 * arch_update_cpu_topology lets virtualized architectures update the
8698 * cpu core maps. It is supposed to return 1 if the topology changed
8699 * or 0 if it stayed the same.
8700 */
8701int __attribute__((weak)) arch_update_cpu_topology(void)
Heiko Carstens22e52b02008-03-12 18:31:59 +01008702{
Heiko Carstensee79d1b2008-12-09 18:49:50 +01008703 return 0;
Heiko Carstens22e52b02008-03-12 18:31:59 +01008704}
8705
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008706/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008707 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07008708 * For now this just excludes isolated cpus, but could be used to
8709 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008710 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308711static int arch_init_sched_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008712{
Milton Miller73785472007-10-24 18:23:48 +02008713 int err;
8714
Heiko Carstens22e52b02008-03-12 18:31:59 +01008715 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07008716 ndoms_cur = 1;
Rusty Russell96f874e2008-11-25 02:35:14 +10308717 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
Paul Jackson029190c2007-10-18 23:40:20 -07008718 if (!doms_cur)
Rusty Russell42128232008-11-25 02:35:12 +10308719 doms_cur = fallback_doms;
Rusty Russelldcc30a32008-11-25 02:35:12 +10308720 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008721 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02008722 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02008723 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02008724
8725 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008726}
8727
Rusty Russell96f874e2008-11-25 02:35:14 +10308728static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8729 struct cpumask *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07008730{
Mike Travis7c16ec52008-04-04 18:11:11 -07008731 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07008732}
Linus Torvalds1da177e2005-04-16 15:20:36 -07008733
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008734/*
8735 * Detach sched domains from a group of cpus specified in cpu_map
8736 * These cpus will now be attached to the NULL domain
8737 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308738static void detach_destroy_domains(const struct cpumask *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008739{
Rusty Russell96f874e2008-11-25 02:35:14 +10308740 /* Save because hotplug lock held. */
8741 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008742 int i;
8743
Rusty Russellabcd0832008-11-25 02:35:02 +10308744 for_each_cpu(i, cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01008745 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008746 synchronize_sched();
Rusty Russell96f874e2008-11-25 02:35:14 +10308747 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07008748}
8749
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008750/* handle null as "default" */
8751static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8752 struct sched_domain_attr *new, int idx_new)
8753{
8754 struct sched_domain_attr tmp;
8755
8756 /* fast path */
8757 if (!new && !cur)
8758 return 1;
8759
8760 tmp = SD_ATTR_INIT;
8761 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8762 new ? (new + idx_new) : &tmp,
8763 sizeof(struct sched_domain_attr));
8764}
8765
Paul Jackson029190c2007-10-18 23:40:20 -07008766/*
8767 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008768 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07008769 * doms_new[] to the current sched domain partitioning, doms_cur[].
8770 * It destroys each deleted domain and builds each new domain.
8771 *
Rusty Russell96f874e2008-11-25 02:35:14 +10308772 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008773 * The masks don't intersect (don't overlap.) We should setup one
8774 * sched domain for each mask. CPUs not in any of the cpumasks will
8775 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07008776 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8777 * it as it is.
8778 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008779 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8780 * ownership of it and will kfree it when done with it. If the caller
Li Zefan700018e2008-11-18 14:02:03 +08008781 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8782 * ndoms_new == 1, and partition_sched_domains() will fallback to
8783 * the single partition 'fallback_doms', it also forces the domains
8784 * to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07008785 *
Rusty Russell96f874e2008-11-25 02:35:14 +10308786 * If doms_new == NULL it will be replaced with cpu_online_mask.
Li Zefan700018e2008-11-18 14:02:03 +08008787 * ndoms_new == 0 is a special case for destroying existing domains,
8788 * and it will not create the default domain.
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008789 *
Paul Jackson029190c2007-10-18 23:40:20 -07008790 * Call with hotplug lock held
8791 */
Rusty Russell96f874e2008-11-25 02:35:14 +10308792/* FIXME: Change to struct cpumask *doms_new[] */
8793void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008794 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07008795{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008796 int i, j, n;
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008797 int new_topology;
Paul Jackson029190c2007-10-18 23:40:20 -07008798
Heiko Carstens712555e2008-04-28 11:33:07 +02008799 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01008800
Milton Miller73785472007-10-24 18:23:48 +02008801 /* always unregister in case we don't destroy any domains */
8802 unregister_sched_domain_sysctl();
8803
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008804 /* Let architecture update cpu core mappings. */
8805 new_topology = arch_update_cpu_topology();
8806
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008807 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07008808
8809 /* Destroy deleted domains */
8810 for (i = 0; i < ndoms_cur; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008811 for (j = 0; j < n && !new_topology; j++) {
Rusty Russell96f874e2008-11-25 02:35:14 +10308812 if (cpumask_equal(&doms_cur[i], &doms_new[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008813 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07008814 goto match1;
8815 }
8816 /* no match - a current sched domain not in new doms_new[] */
8817 detach_destroy_domains(doms_cur + i);
8818match1:
8819 ;
8820 }
8821
Max Krasnyanskye761b772008-07-15 04:43:49 -07008822 if (doms_new == NULL) {
8823 ndoms_cur = 0;
Rusty Russell42128232008-11-25 02:35:12 +10308824 doms_new = fallback_doms;
Rusty Russelldcc30a32008-11-25 02:35:12 +10308825 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
Li Zefanfaa2f982008-11-04 16:20:23 +08008826 WARN_ON_ONCE(dattr_new);
Max Krasnyanskye761b772008-07-15 04:43:49 -07008827 }
8828
Paul Jackson029190c2007-10-18 23:40:20 -07008829 /* Build new domains */
8830 for (i = 0; i < ndoms_new; i++) {
Heiko Carstensd65bd5e2008-12-09 18:49:51 +01008831 for (j = 0; j < ndoms_cur && !new_topology; j++) {
Rusty Russell96f874e2008-11-25 02:35:14 +10308832 if (cpumask_equal(&doms_new[i], &doms_cur[j])
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008833 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07008834 goto match2;
8835 }
8836 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008837 __build_sched_domains(doms_new + i,
8838 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07008839match2:
8840 ;
8841 }
8842
8843 /* Remember the new sched domains */
Rusty Russell42128232008-11-25 02:35:12 +10308844 if (doms_cur != fallback_doms)
Paul Jackson029190c2007-10-18 23:40:20 -07008845 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008846 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07008847 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09008848 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07008849 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02008850
8851 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01008852
Heiko Carstens712555e2008-04-28 11:33:07 +02008853 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07008854}
8855
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008856#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Li Zefanc70f22d2009-01-05 19:07:50 +08008857static void arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008858{
Gautham R Shenoy95402b32008-01-25 21:08:02 +01008859 get_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008860
8861 /* Destroy domains first to force the rebuild */
8862 partition_sched_domains(0, NULL, NULL);
8863
Max Krasnyanskye761b772008-07-15 04:43:49 -07008864 rebuild_sched_domains();
Gautham R Shenoy95402b32008-01-25 21:08:02 +01008865 put_online_cpus();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008866}
8867
8868static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8869{
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05308870 unsigned int level = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008871
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05308872 if (sscanf(buf, "%u", &level) != 1)
8873 return -EINVAL;
8874
8875 /*
8876 * level is always be positive so don't check for
8877 * level < POWERSAVINGS_BALANCE_NONE which is 0
8878 * What happens on 0 or 1 byte write,
8879 * need to check for count as well?
8880 */
8881
8882 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008883 return -EINVAL;
8884
8885 if (smt)
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05308886 sched_smt_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008887 else
Gautham R Shenoyafb8a9b2008-12-18 23:26:09 +05308888 sched_mc_power_savings = level;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008889
Li Zefanc70f22d2009-01-05 19:07:50 +08008890 arch_reinit_sched_domains();
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008891
Li Zefanc70f22d2009-01-05 19:07:50 +08008892 return count;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008893}
8894
Adrian Bunk6707de002007-08-12 18:08:19 +02008895#ifdef CONFIG_SCHED_MC
Andi Kleenf718cd42008-07-29 22:33:52 -07008896static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8897 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02008898{
8899 return sprintf(page, "%u\n", sched_mc_power_savings);
8900}
Andi Kleenf718cd42008-07-29 22:33:52 -07008901static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
Adrian Bunk6707de002007-08-12 18:08:19 +02008902 const char *buf, size_t count)
8903{
8904 return sched_power_savings_store(buf, count, 0);
8905}
Andi Kleenf718cd42008-07-29 22:33:52 -07008906static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8907 sched_mc_power_savings_show,
8908 sched_mc_power_savings_store);
Adrian Bunk6707de002007-08-12 18:08:19 +02008909#endif
8910
8911#ifdef CONFIG_SCHED_SMT
Andi Kleenf718cd42008-07-29 22:33:52 -07008912static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8913 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02008914{
8915 return sprintf(page, "%u\n", sched_smt_power_savings);
8916}
Andi Kleenf718cd42008-07-29 22:33:52 -07008917static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
Adrian Bunk6707de002007-08-12 18:08:19 +02008918 const char *buf, size_t count)
8919{
8920 return sched_power_savings_store(buf, count, 1);
8921}
Andi Kleenf718cd42008-07-29 22:33:52 -07008922static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8923 sched_smt_power_savings_show,
Adrian Bunk6707de002007-08-12 18:08:19 +02008924 sched_smt_power_savings_store);
8925#endif
8926
Li Zefan39aac642009-01-05 19:18:02 +08008927int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008928{
8929 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07008930
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008931#ifdef CONFIG_SCHED_SMT
8932 if (smt_capable())
8933 err = sysfs_create_file(&cls->kset.kobj,
8934 &attr_sched_smt_power_savings.attr);
8935#endif
8936#ifdef CONFIG_SCHED_MC
8937 if (!err && mc_capable())
8938 err = sysfs_create_file(&cls->kset.kobj,
8939 &attr_sched_mc_power_savings.attr);
8940#endif
8941 return err;
8942}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008943#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07008944
Max Krasnyanskye761b772008-07-15 04:43:49 -07008945#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07008946/*
Max Krasnyanskye761b772008-07-15 04:43:49 -07008947 * Add online and remove offline CPUs from the scheduler domains.
8948 * When cpusets are enabled they take over this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07008949 */
8950static int update_sched_domains(struct notifier_block *nfb,
8951 unsigned long action, void *hcpu)
8952{
Max Krasnyanskye761b772008-07-15 04:43:49 -07008953 switch (action) {
8954 case CPU_ONLINE:
8955 case CPU_ONLINE_FROZEN:
8956 case CPU_DEAD:
8957 case CPU_DEAD_FROZEN:
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07008958 partition_sched_domains(1, NULL, NULL);
Max Krasnyanskye761b772008-07-15 04:43:49 -07008959 return NOTIFY_OK;
8960
8961 default:
8962 return NOTIFY_DONE;
8963 }
8964}
8965#endif
8966
8967static int update_runtime(struct notifier_block *nfb,
8968 unsigned long action, void *hcpu)
8969{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02008970 int cpu = (int)(long)hcpu;
8971
Linus Torvalds1da177e2005-04-16 15:20:36 -07008972 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07008973 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07008974 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02008975 disable_runtime(cpu_rq(cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07008976 return NOTIFY_OK;
8977
Linus Torvalds1da177e2005-04-16 15:20:36 -07008978 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07008979 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07008980 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07008981 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02008982 enable_runtime(cpu_rq(cpu));
Max Krasnyanskye761b772008-07-15 04:43:49 -07008983 return NOTIFY_OK;
8984
Linus Torvalds1da177e2005-04-16 15:20:36 -07008985 default:
8986 return NOTIFY_DONE;
8987 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008988}
Linus Torvalds1da177e2005-04-16 15:20:36 -07008989
8990void __init sched_init_smp(void)
8991{
Rusty Russelldcc30a32008-11-25 02:35:12 +10308992 cpumask_var_t non_isolated_cpus;
8993
8994 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
Nick Piggin5c1e1762006-10-03 01:14:04 -07008995
Mike Travis434d53b2008-04-04 18:11:04 -07008996#if defined(CONFIG_NUMA)
8997 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8998 GFP_KERNEL);
8999 BUG_ON(sched_group_nodes_bycpu == NULL);
9000#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009001 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02009002 mutex_lock(&sched_domains_mutex);
Rusty Russelldcc30a32008-11-25 02:35:12 +10309003 arch_init_sched_domains(cpu_online_mask);
9004 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9005 if (cpumask_empty(non_isolated_cpus))
9006 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02009007 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01009008 put_online_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07009009
9010#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07009011 /* XXX: Theoretical race here - CPU may be hotplugged now */
9012 hotcpu_notifier(update_sched_domains, 0);
Max Krasnyanskye761b772008-07-15 04:43:49 -07009013#endif
9014
9015 /* RT runtime code needs to handle some hotplug events */
9016 hotcpu_notifier(update_runtime, 0);
9017
Peter Zijlstrab328ca12008-04-29 10:02:46 +02009018 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07009019
9020 /* Move init over to a non-isolated CPU */
Rusty Russelldcc30a32008-11-25 02:35:12 +10309021 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07009022 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01009023 sched_init_granularity();
Rusty Russelldcc30a32008-11-25 02:35:12 +10309024 free_cpumask_var(non_isolated_cpus);
Rusty Russell42128232008-11-25 02:35:12 +10309025
9026 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
Rusty Russell0e3900e2008-11-25 02:35:13 +10309027 init_sched_rt_class();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009028}
9029#else
9030void __init sched_init_smp(void)
9031{
Ingo Molnar19978ca2007-11-09 22:39:38 +01009032 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009033}
9034#endif /* CONFIG_SMP */
9035
Arun R Bharadwajcd1bb942009-04-16 12:15:34 +05309036const_debug unsigned int sysctl_timer_migration = 1;
9037
Linus Torvalds1da177e2005-04-16 15:20:36 -07009038int in_sched_functions(unsigned long addr)
9039{
Linus Torvalds1da177e2005-04-16 15:20:36 -07009040 return in_lock_functions(addr) ||
9041 (addr >= (unsigned long)__sched_text_start
9042 && addr < (unsigned long)__sched_text_end);
9043}
9044
Alexey Dobriyana9957442007-10-15 17:00:13 +02009045static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02009046{
9047 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02009048 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02009049#ifdef CONFIG_FAIR_GROUP_SCHED
9050 cfs_rq->rq = rq;
9051#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02009052 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02009053}
9054
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009055static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9056{
9057 struct rt_prio_array *array;
9058 int i;
9059
9060 array = &rt_rq->active;
9061 for (i = 0; i < MAX_RT_PRIO; i++) {
9062 INIT_LIST_HEAD(array->queue + i);
9063 __clear_bit(i, array->bitmap);
9064 }
9065 /* delimiter for bitsearch: */
9066 __set_bit(MAX_RT_PRIO, array->bitmap);
9067
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009068#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Gregory Haskinse864c492008-12-29 09:39:49 -05009069 rt_rq->highest_prio.curr = MAX_RT_PRIO;
Gregory Haskins398a1532009-01-14 09:10:04 -05009070#ifdef CONFIG_SMP
Gregory Haskinse864c492008-12-29 09:39:49 -05009071 rt_rq->highest_prio.next = MAX_RT_PRIO;
Peter Zijlstra48d5e252008-01-25 21:08:31 +01009072#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009073#endif
9074#ifdef CONFIG_SMP
9075 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009076 rt_rq->overloaded = 0;
Gregory Haskins917b6272008-12-29 09:39:53 -05009077 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009078#endif
9079
9080 rt_rq->rt_time = 0;
9081 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009082 rt_rq->rt_runtime = 0;
9083 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009084
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009085#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01009086 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009087 rt_rq->rq = rq;
9088#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009089}
9090
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009091#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009092static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9093 struct sched_entity *se, int cpu, int add,
9094 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009095{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009096 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009097 tg->cfs_rq[cpu] = cfs_rq;
9098 init_cfs_rq(cfs_rq, rq);
9099 cfs_rq->tg = tg;
9100 if (add)
9101 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9102
9103 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009104 /* se could be NULL for init_task_group */
9105 if (!se)
9106 return;
9107
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009108 if (!parent)
9109 se->cfs_rq = &rq->cfs;
9110 else
9111 se->cfs_rq = parent->my_q;
9112
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009113 se->my_q = cfs_rq;
9114 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009115 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009116 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009117}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009118#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009119
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009120#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009121static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9122 struct sched_rt_entity *rt_se, int cpu, int add,
9123 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009124{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009125 struct rq *rq = cpu_rq(cpu);
9126
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009127 tg->rt_rq[cpu] = rt_rq;
9128 init_rt_rq(rt_rq, rq);
9129 rt_rq->tg = tg;
9130 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009131 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009132 if (add)
9133 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9134
9135 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02009136 if (!rt_se)
9137 return;
9138
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009139 if (!parent)
9140 rt_se->rt_rq = &rq->rt;
9141 else
9142 rt_se->rt_rq = parent->my_q;
9143
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009144 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009145 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009146 INIT_LIST_HEAD(&rt_se->run_list);
9147}
9148#endif
9149
Linus Torvalds1da177e2005-04-16 15:20:36 -07009150void __init sched_init(void)
9151{
Ingo Molnardd41f592007-07-09 18:51:59 +02009152 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07009153 unsigned long alloc_size = 0, ptr;
9154
9155#ifdef CONFIG_FAIR_GROUP_SCHED
9156 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9157#endif
9158#ifdef CONFIG_RT_GROUP_SCHED
9159 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9160#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009161#ifdef CONFIG_USER_SCHED
9162 alloc_size *= 2;
9163#endif
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309164#ifdef CONFIG_CPUMASK_OFFSTACK
Rusty Russell8c083f02009-03-19 15:22:20 +10309165 alloc_size += num_possible_cpus() * cpumask_size();
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309166#endif
Mike Travis434d53b2008-04-04 18:11:04 -07009167 /*
9168 * As sched_init() is called before page_alloc is setup,
9169 * we use alloc_bootmem().
9170 */
9171 if (alloc_size) {
Pekka Enberg36b7b6d2009-06-10 23:42:36 +03009172 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
Mike Travis434d53b2008-04-04 18:11:04 -07009173
9174#ifdef CONFIG_FAIR_GROUP_SCHED
9175 init_task_group.se = (struct sched_entity **)ptr;
9176 ptr += nr_cpu_ids * sizeof(void **);
9177
9178 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9179 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009180
9181#ifdef CONFIG_USER_SCHED
9182 root_task_group.se = (struct sched_entity **)ptr;
9183 ptr += nr_cpu_ids * sizeof(void **);
9184
9185 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9186 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009187#endif /* CONFIG_USER_SCHED */
9188#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07009189#ifdef CONFIG_RT_GROUP_SCHED
9190 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9191 ptr += nr_cpu_ids * sizeof(void **);
9192
9193 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009194 ptr += nr_cpu_ids * sizeof(void **);
9195
9196#ifdef CONFIG_USER_SCHED
9197 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9198 ptr += nr_cpu_ids * sizeof(void **);
9199
9200 root_task_group.rt_rq = (struct rt_rq **)ptr;
9201 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009202#endif /* CONFIG_USER_SCHED */
9203#endif /* CONFIG_RT_GROUP_SCHED */
Rusty Russelldf7c8e82009-03-19 15:22:20 +10309204#ifdef CONFIG_CPUMASK_OFFSTACK
9205 for_each_possible_cpu(i) {
9206 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9207 ptr += cpumask_size();
9208 }
9209#endif /* CONFIG_CPUMASK_OFFSTACK */
Mike Travis434d53b2008-04-04 18:11:04 -07009210 }
Ingo Molnardd41f592007-07-09 18:51:59 +02009211
Gregory Haskins57d885f2008-01-25 21:08:18 +01009212#ifdef CONFIG_SMP
9213 init_defrootdomain();
9214#endif
9215
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009216 init_rt_bandwidth(&def_rt_bandwidth,
9217 global_rt_period(), global_rt_runtime());
9218
9219#ifdef CONFIG_RT_GROUP_SCHED
9220 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9221 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009222#ifdef CONFIG_USER_SCHED
9223 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9224 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009225#endif /* CONFIG_USER_SCHED */
9226#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009227
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009228#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009229 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009230 INIT_LIST_HEAD(&init_task_group.children);
9231
9232#ifdef CONFIG_USER_SCHED
9233 INIT_LIST_HEAD(&root_task_group.children);
9234 init_task_group.parent = &root_task_group;
9235 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009236#endif /* CONFIG_USER_SCHED */
9237#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009238
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08009239 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07009240 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009241
9242 rq = cpu_rq(i);
9243 spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07009244 rq->nr_running = 0;
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009245 rq->calc_load_active = 0;
9246 rq->calc_load_update = jiffies + LOAD_FREQ;
Ingo Molnardd41f592007-07-09 18:51:59 +02009247 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01009248 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009249#ifdef CONFIG_FAIR_GROUP_SCHED
9250 init_task_group.shares = init_task_group_load;
9251 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009252#ifdef CONFIG_CGROUP_SCHED
9253 /*
9254 * How much cpu bandwidth does init_task_group get?
9255 *
9256 * In case of task-groups formed thr' the cgroup filesystem, it
9257 * gets 100% of the cpu resources in the system. This overall
9258 * system cpu resource is divided among the tasks of
9259 * init_task_group and its child task-groups in a fair manner,
9260 * based on each entity's (task or task-group's) weight
9261 * (se->load.weight).
9262 *
9263 * In other words, if init_task_group has 10 tasks of weight
9264 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9265 * then A0's share of the cpu resource is:
9266 *
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009267 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
Dhaval Giani354d60c2008-04-19 19:44:59 +02009268 *
9269 * We achieve this by letting init_task_group's tasks sit
9270 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9271 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009272 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009273#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009274 root_task_group.shares = NICE_0_LOAD;
9275 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009276 /*
9277 * In case of task-groups formed thr' the user id of tasks,
9278 * init_task_group represents tasks belonging to root user.
9279 * Hence it forms a sibling of all subsequent groups formed.
9280 * In this case, init_task_group gets only a fraction of overall
9281 * system cpu resource, based on the weight assigned to root
9282 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9283 * by letting tasks of init_task_group sit in a separate cfs_rq
9284 * (init_cfs_rq) and having one entity represent this group of
9285 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9286 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009287 init_tg_cfs_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009288 &per_cpu(init_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009289 &per_cpu(init_sched_entity, i), i, 1,
9290 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009291
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009292#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02009293#endif /* CONFIG_FAIR_GROUP_SCHED */
9294
9295 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009296#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009297 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009298#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009299 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009300#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009301 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009302 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009303 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02009304 &per_cpu(init_sched_rt_entity, i), i, 1,
9305 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02009306#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009307#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07009308
Ingo Molnardd41f592007-07-09 18:51:59 +02009309 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9310 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009311#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07009312 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01009313 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009314 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009315 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009316 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07009317 rq->cpu = i;
Gregory Haskins1f11eb62008-06-04 15:04:05 -04009318 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009319 rq->migration_thread = NULL;
9320 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01009321 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009322#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01009323 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009324 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009325 }
9326
Peter Williams2dd73a42006-06-27 02:54:34 -07009327 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009328
Avi Kivitye107be32007-07-26 13:40:43 +02009329#ifdef CONFIG_PREEMPT_NOTIFIERS
9330 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9331#endif
9332
Christoph Lameterc9819f42006-12-10 02:20:25 -08009333#ifdef CONFIG_SMP
Carlos R. Mafra962cf362008-05-15 11:15:37 -03009334 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
Christoph Lameterc9819f42006-12-10 02:20:25 -08009335#endif
9336
Heiko Carstensb50f60c2006-07-30 03:03:52 -07009337#ifdef CONFIG_RT_MUTEXES
9338 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9339#endif
9340
Linus Torvalds1da177e2005-04-16 15:20:36 -07009341 /*
9342 * The boot idle thread does lazy MMU switching as well:
9343 */
9344 atomic_inc(&init_mm.mm_count);
9345 enter_lazy_tlb(&init_mm, current);
9346
9347 /*
9348 * Make us the idle thread. Technically, schedule() should not be
9349 * called from this thread, however somewhere below it might be,
9350 * but because we are the idle thread, we just pick up running again
9351 * when this runqueue becomes "idle".
9352 */
9353 init_idle(current, smp_processor_id());
Thomas Gleixnerdce48a82009-04-11 10:43:41 +02009354
9355 calc_load_update = jiffies + LOAD_FREQ;
9356
Ingo Molnardd41f592007-07-09 18:51:59 +02009357 /*
9358 * During early bootup we pretend to be a normal task:
9359 */
9360 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01009361
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309362 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009363 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309364#ifdef CONFIG_SMP
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309365#ifdef CONFIG_NO_HZ
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009366 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9367 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
Rusty Russell7d1e6a92008-11-25 02:35:09 +10309368#endif
Pekka Enberg4bdddf82009-06-11 08:35:27 +03009369 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
Rusty Russellbf4d83f2008-11-25 09:57:51 +10309370#endif /* SMP */
Rusty Russell6a7b3dc2008-11-25 02:35:04 +10309371
Ingo Molnar0d905bc2009-05-04 19:13:30 +02009372 perf_counter_init();
9373
Ingo Molnar6892b752008-02-13 14:02:36 +01009374 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009375}
9376
9377#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9378void __might_sleep(char *file, int line)
9379{
Ingo Molnar48f24c42006-07-03 00:25:40 -07009380#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07009381 static unsigned long prev_jiffy; /* ratelimiting */
9382
Ingo Molnaraef745f2008-08-28 11:34:43 +02009383 if ((!in_atomic() && !irqs_disabled()) ||
9384 system_state != SYSTEM_RUNNING || oops_in_progress)
9385 return;
9386 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9387 return;
9388 prev_jiffy = jiffies;
9389
9390 printk(KERN_ERR
9391 "BUG: sleeping function called from invalid context at %s:%d\n",
9392 file, line);
9393 printk(KERN_ERR
9394 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9395 in_atomic(), irqs_disabled(),
9396 current->pid, current->comm);
9397
9398 debug_show_held_locks(current);
9399 if (irqs_disabled())
9400 print_irqtrace_events(current);
9401 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07009402#endif
9403}
9404EXPORT_SYMBOL(__might_sleep);
9405#endif
9406
9407#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009408static void normalize_task(struct rq *rq, struct task_struct *p)
9409{
9410 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02009411
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009412 update_rq_clock(rq);
9413 on_rq = p->se.on_rq;
9414 if (on_rq)
9415 deactivate_task(rq, p, 0);
9416 __setscheduler(rq, p, SCHED_NORMAL, 0);
9417 if (on_rq) {
9418 activate_task(rq, p, 0);
9419 resched_task(rq->curr);
9420 }
9421}
9422
Linus Torvalds1da177e2005-04-16 15:20:36 -07009423void normalize_rt_tasks(void)
9424{
Ingo Molnara0f98a12007-06-17 18:37:45 +02009425 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009426 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07009427 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07009428
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009429 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009430 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02009431 /*
9432 * Only normalize user tasks:
9433 */
9434 if (!p->mm)
9435 continue;
9436
Ingo Molnardd41f592007-07-09 18:51:59 +02009437 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009438#ifdef CONFIG_SCHEDSTATS
9439 p->se.wait_start = 0;
9440 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02009441 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02009442#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02009443
9444 if (!rt_task(p)) {
9445 /*
9446 * Renice negative nice level userspace
9447 * tasks back to 0:
9448 */
9449 if (TASK_NICE(p) < 0 && p->mm)
9450 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009451 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02009452 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07009453
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009454 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07009455 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009456
Ingo Molnar178be792007-10-15 17:00:18 +02009457 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02009458
Ingo Molnarb29739f2006-06-27 02:54:51 -07009459 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009460 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02009461 } while_each_thread(g, p);
9462
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01009463 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07009464}
9465
9466#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07009467
9468#ifdef CONFIG_IA64
9469/*
9470 * These functions are only useful for the IA64 MCA handling.
9471 *
9472 * They can only be called when the whole system has been
9473 * stopped - every CPU needs to be quiescent, and no scheduling
9474 * activity can take place. Using them for anything else would
9475 * be a serious bug, and as a result, they aren't even visible
9476 * under any other configuration.
9477 */
9478
9479/**
9480 * curr_task - return the current task for a given cpu.
9481 * @cpu: the processor in question.
9482 *
9483 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9484 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009485struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009486{
9487 return cpu_curr(cpu);
9488}
9489
9490/**
9491 * set_curr_task - set the current task for a given cpu.
9492 * @cpu: the processor in question.
9493 * @p: the task pointer to set.
9494 *
9495 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009496 * are serviced on a separate stack. It allows the architecture to switch the
9497 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07009498 * must be called with all CPU's synchronized, and interrupts disabled, the
9499 * and caller must save the original value of the current task (see
9500 * curr_task() above) and restore that value before reenabling interrupts and
9501 * re-starting the system.
9502 *
9503 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9504 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07009505void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07009506{
9507 cpu_curr(cpu) = p;
9508}
9509
9510#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009511
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009512#ifdef CONFIG_FAIR_GROUP_SCHED
9513static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009514{
9515 int i;
9516
9517 for_each_possible_cpu(i) {
9518 if (tg->cfs_rq)
9519 kfree(tg->cfs_rq[i]);
9520 if (tg->se)
9521 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009522 }
9523
9524 kfree(tg->cfs_rq);
9525 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009526}
9527
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009528static
9529int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009530{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009531 struct cfs_rq *cfs_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009532 struct sched_entity *se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009533 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009534 int i;
9535
Mike Travis434d53b2008-04-04 18:11:04 -07009536 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009537 if (!tg->cfs_rq)
9538 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009539 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009540 if (!tg->se)
9541 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009542
9543 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009544
9545 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009546 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009547
Li Zefaneab17222008-10-29 17:03:22 +08009548 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9549 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009550 if (!cfs_rq)
9551 goto err;
9552
Li Zefaneab17222008-10-29 17:03:22 +08009553 se = kzalloc_node(sizeof(struct sched_entity),
9554 GFP_KERNEL, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009555 if (!se)
9556 goto err;
9557
Li Zefaneab17222008-10-29 17:03:22 +08009558 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009559 }
9560
9561 return 1;
9562
9563 err:
9564 return 0;
9565}
9566
9567static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9568{
9569 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9570 &cpu_rq(cpu)->leaf_cfs_rq_list);
9571}
9572
9573static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9574{
9575 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9576}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009577#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009578static inline void free_fair_sched_group(struct task_group *tg)
9579{
9580}
9581
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009582static inline
9583int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009584{
9585 return 1;
9586}
9587
9588static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9589{
9590}
9591
9592static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9593{
9594}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009595#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009596
9597#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009598static void free_rt_sched_group(struct task_group *tg)
9599{
9600 int i;
9601
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009602 destroy_rt_bandwidth(&tg->rt_bandwidth);
9603
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009604 for_each_possible_cpu(i) {
9605 if (tg->rt_rq)
9606 kfree(tg->rt_rq[i]);
9607 if (tg->rt_se)
9608 kfree(tg->rt_se[i]);
9609 }
9610
9611 kfree(tg->rt_rq);
9612 kfree(tg->rt_se);
9613}
9614
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009615static
9616int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009617{
9618 struct rt_rq *rt_rq;
Li Zefaneab17222008-10-29 17:03:22 +08009619 struct sched_rt_entity *rt_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009620 struct rq *rq;
9621 int i;
9622
Mike Travis434d53b2008-04-04 18:11:04 -07009623 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009624 if (!tg->rt_rq)
9625 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07009626 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009627 if (!tg->rt_se)
9628 goto err;
9629
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009630 init_rt_bandwidth(&tg->rt_bandwidth,
9631 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009632
9633 for_each_possible_cpu(i) {
9634 rq = cpu_rq(i);
9635
Li Zefaneab17222008-10-29 17:03:22 +08009636 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9637 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009638 if (!rt_rq)
9639 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009640
Li Zefaneab17222008-10-29 17:03:22 +08009641 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9642 GFP_KERNEL, cpu_to_node(i));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009643 if (!rt_se)
9644 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009645
Li Zefaneab17222008-10-29 17:03:22 +08009646 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009647 }
9648
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009649 return 1;
9650
9651 err:
9652 return 0;
9653}
9654
9655static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9656{
9657 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9658 &cpu_rq(cpu)->leaf_rt_rq_list);
9659}
9660
9661static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9662{
9663 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9664}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009665#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009666static inline void free_rt_sched_group(struct task_group *tg)
9667{
9668}
9669
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009670static inline
9671int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009672{
9673 return 1;
9674}
9675
9676static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9677{
9678}
9679
9680static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9681{
9682}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009683#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009684
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009685#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009686static void free_sched_group(struct task_group *tg)
9687{
9688 free_fair_sched_group(tg);
9689 free_rt_sched_group(tg);
9690 kfree(tg);
9691}
9692
9693/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009694struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009695{
9696 struct task_group *tg;
9697 unsigned long flags;
9698 int i;
9699
9700 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9701 if (!tg)
9702 return ERR_PTR(-ENOMEM);
9703
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009704 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009705 goto err;
9706
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009707 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009708 goto err;
9709
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009710 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009711 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009712 register_fair_sched_group(tg, i);
9713 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009714 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009715 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009716
9717 WARN_ON(!parent); /* root should already exist */
9718
9719 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009720 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +08009721 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009722 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009723
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009724 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009725
9726err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009727 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009728 return ERR_PTR(-ENOMEM);
9729}
9730
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009731/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009732static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009733{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009734 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009735 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009736}
9737
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009738/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02009739void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009740{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009741 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009742 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009743
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009744 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009745 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009746 unregister_fair_sched_group(tg, i);
9747 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009748 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009749 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009750 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009751 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009752
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009753 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009754 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009755}
9756
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009757/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02009758 * The caller of this function should have put the task in its new group
9759 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9760 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009761 */
9762void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009763{
9764 int on_rq, running;
9765 unsigned long flags;
9766 struct rq *rq;
9767
9768 rq = task_rq_lock(tsk, &flags);
9769
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009770 update_rq_clock(rq);
9771
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01009772 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009773 on_rq = tsk->se.on_rq;
9774
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009775 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009776 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009777 if (unlikely(running))
9778 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009779
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009780 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009781
Peter Zijlstra810b3812008-02-29 15:21:01 -05009782#ifdef CONFIG_FAIR_GROUP_SCHED
9783 if (tsk->sched_class->moved_group)
9784 tsk->sched_class->moved_group(tsk);
9785#endif
9786
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07009787 if (unlikely(running))
9788 tsk->sched_class->set_curr_task(rq);
9789 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02009790 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009791
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009792 task_rq_unlock(rq, &flags);
9793}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009794#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009795
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009796#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009797static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009798{
9799 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009800 int on_rq;
9801
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009802 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009803 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009804 dequeue_entity(cfs_rq, se, 0);
9805
9806 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02009807 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009808
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009809 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009810 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009811}
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009812
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009813static void set_se_shares(struct sched_entity *se, unsigned long shares)
9814{
9815 struct cfs_rq *cfs_rq = se->cfs_rq;
9816 struct rq *rq = cfs_rq->rq;
9817 unsigned long flags;
9818
9819 spin_lock_irqsave(&rq->lock, flags);
9820 __set_se_shares(se, shares);
9821 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009822}
9823
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009824static DEFINE_MUTEX(shares_mutex);
9825
Ingo Molnar4cf86d72007-10-15 17:00:14 +02009826int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009827{
9828 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009829 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +01009830
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009831 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009832 * We can't change the weight of the root cgroup.
9833 */
9834 if (!tg->se[0])
9835 return -EINVAL;
9836
Peter Zijlstra18d95a22008-04-19 19:45:00 +02009837 if (shares < MIN_SHARES)
9838 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08009839 else if (shares > MAX_SHARES)
9840 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01009841
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009842 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009843 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02009844 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009845
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009846 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009847 for_each_possible_cpu(i)
9848 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009849 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009850 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01009851
9852 /* wait for any ongoing reference to this group to finish */
9853 synchronize_sched();
9854
9855 /*
9856 * Now we are free to modify the group's share on each cpu
9857 * w/o tripping rebalance_share or load_balance_fair.
9858 */
9859 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009860 for_each_possible_cpu(i) {
9861 /*
9862 * force a rebalance
9863 */
9864 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +08009865 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02009866 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01009867
9868 /*
9869 * Enable load balance activity on this group, by inserting it back on
9870 * each cpu's rq->leaf_cfs_rq_list.
9871 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009872 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01009873 for_each_possible_cpu(i)
9874 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02009875 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009876 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +02009877done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +01009878 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02009879 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02009880}
9881
Dhaval Giani5cb350b2007-10-15 17:00:14 +02009882unsigned long sched_group_shares(struct task_group *tg)
9883{
9884 return tg->shares;
9885}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009886#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +02009887
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009888#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009889/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009890 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009891 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009892static DEFINE_MUTEX(rt_constraints_mutex);
9893
9894static unsigned long to_ratio(u64 period, u64 runtime)
9895{
9896 if (runtime == RUNTIME_INF)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009897 return 1ULL << 20;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009898
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009899 return div64_u64(runtime << 20, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009900}
9901
Dhaval Giani521f1a242008-02-28 15:21:56 +05309902/* Must be called with tasklist_lock held */
9903static inline int tg_has_rt_tasks(struct task_group *tg)
9904{
9905 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009906
Dhaval Giani521f1a242008-02-28 15:21:56 +05309907 do_each_thread(g, p) {
9908 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9909 return 1;
9910 } while_each_thread(g, p);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009911
Dhaval Giani521f1a242008-02-28 15:21:56 +05309912 return 0;
9913}
9914
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009915struct rt_schedulable_data {
9916 struct task_group *tg;
9917 u64 rt_period;
9918 u64 rt_runtime;
9919};
9920
9921static int tg_schedulable(struct task_group *tg, void *data)
9922{
9923 struct rt_schedulable_data *d = data;
9924 struct task_group *child;
9925 unsigned long total, sum = 0;
9926 u64 period, runtime;
9927
9928 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9929 runtime = tg->rt_bandwidth.rt_runtime;
9930
9931 if (tg == d->tg) {
9932 period = d->rt_period;
9933 runtime = d->rt_runtime;
9934 }
9935
Peter Zijlstra98a48262009-01-14 10:56:32 +01009936#ifdef CONFIG_USER_SCHED
9937 if (tg == &root_task_group) {
9938 period = global_rt_period();
9939 runtime = global_rt_runtime();
9940 }
9941#endif
9942
Peter Zijlstra4653f802008-09-23 15:33:44 +02009943 /*
9944 * Cannot have more runtime than the period.
9945 */
9946 if (runtime > period && runtime != RUNTIME_INF)
9947 return -EINVAL;
9948
9949 /*
9950 * Ensure we don't starve existing RT tasks.
9951 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009952 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9953 return -EBUSY;
9954
9955 total = to_ratio(period, runtime);
9956
Peter Zijlstra4653f802008-09-23 15:33:44 +02009957 /*
9958 * Nobody can have more than the global setting allows.
9959 */
9960 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9961 return -EINVAL;
9962
9963 /*
9964 * The sum of our children's runtime should not exceed our own.
9965 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009966 list_for_each_entry_rcu(child, &tg->children, siblings) {
9967 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9968 runtime = child->rt_bandwidth.rt_runtime;
9969
9970 if (child == d->tg) {
9971 period = d->rt_period;
9972 runtime = d->rt_runtime;
9973 }
9974
9975 sum += to_ratio(period, runtime);
9976 }
9977
9978 if (sum > total)
9979 return -EINVAL;
9980
9981 return 0;
9982}
9983
9984static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9985{
9986 struct rt_schedulable_data data = {
9987 .tg = tg,
9988 .rt_period = period,
9989 .rt_runtime = runtime,
9990 };
9991
9992 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9993}
9994
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009995static int tg_set_bandwidth(struct task_group *tg,
9996 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009997{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009998 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009999
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010000 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +053010001 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010002 err = __rt_schedulable(tg, rt_period, rt_runtime);
10003 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +053010004 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010005
10006 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010007 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10008 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010009
10010 for_each_possible_cpu(i) {
10011 struct rt_rq *rt_rq = tg->rt_rq[i];
10012
10013 spin_lock(&rt_rq->rt_runtime_lock);
10014 rt_rq->rt_runtime = rt_runtime;
10015 spin_unlock(&rt_rq->rt_runtime_lock);
10016 }
10017 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010018 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +053010019 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010020 mutex_unlock(&rt_constraints_mutex);
10021
10022 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010023}
10024
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010025int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10026{
10027 u64 rt_runtime, rt_period;
10028
10029 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10030 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10031 if (rt_runtime_us < 0)
10032 rt_runtime = RUNTIME_INF;
10033
10034 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10035}
10036
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010037long sched_group_rt_runtime(struct task_group *tg)
10038{
10039 u64 rt_runtime_us;
10040
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010041 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010042 return -1;
10043
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010044 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010045 do_div(rt_runtime_us, NSEC_PER_USEC);
10046 return rt_runtime_us;
10047}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010048
10049int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10050{
10051 u64 rt_runtime, rt_period;
10052
10053 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10054 rt_runtime = tg->rt_bandwidth.rt_runtime;
10055
Raistlin619b0482008-06-26 18:54:09 +020010056 if (rt_period == 0)
10057 return -EINVAL;
10058
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010059 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10060}
10061
10062long sched_group_rt_period(struct task_group *tg)
10063{
10064 u64 rt_period_us;
10065
10066 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10067 do_div(rt_period_us, NSEC_PER_USEC);
10068 return rt_period_us;
10069}
10070
10071static int sched_rt_global_constraints(void)
10072{
Peter Zijlstra4653f802008-09-23 15:33:44 +020010073 u64 runtime, period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010074 int ret = 0;
10075
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010076 if (sysctl_sched_rt_period <= 0)
10077 return -EINVAL;
10078
Peter Zijlstra4653f802008-09-23 15:33:44 +020010079 runtime = global_rt_runtime();
10080 period = global_rt_period();
10081
10082 /*
10083 * Sanity check on the sysctl variables.
10084 */
10085 if (runtime > period && runtime != RUNTIME_INF)
10086 return -EINVAL;
Peter Zijlstra10b612f2008-06-19 14:22:27 +020010087
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010088 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010089 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +020010090 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +020010091 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010092 mutex_unlock(&rt_constraints_mutex);
10093
10094 return ret;
10095}
Dhaval Giani54e99122009-02-27 15:13:54 +053010096
10097int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10098{
10099 /* Don't accept realtime tasks when there is no way for them to run */
10100 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10101 return 0;
10102
10103 return 1;
10104}
10105
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010106#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010107static int sched_rt_global_constraints(void)
10108{
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010109 unsigned long flags;
10110 int i;
10111
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -070010112 if (sysctl_sched_rt_period <= 0)
10113 return -EINVAL;
10114
Peter Zijlstra60aa6052009-05-05 17:50:21 +020010115 /*
10116 * There's always some RT tasks in the root group
10117 * -- migration, kstopmachine etc..
10118 */
10119 if (sysctl_sched_rt_runtime == 0)
10120 return -EBUSY;
10121
Peter Zijlstraac086bc2008-04-19 19:44:58 +020010122 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10123 for_each_possible_cpu(i) {
10124 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10125
10126 spin_lock(&rt_rq->rt_runtime_lock);
10127 rt_rq->rt_runtime = global_rt_runtime();
10128 spin_unlock(&rt_rq->rt_runtime_lock);
10129 }
10130 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10131
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010132 return 0;
10133}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010134#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010135
10136int sched_rt_handler(struct ctl_table *table, int write,
10137 struct file *filp, void __user *buffer, size_t *lenp,
10138 loff_t *ppos)
10139{
10140 int ret;
10141 int old_period, old_runtime;
10142 static DEFINE_MUTEX(mutex);
10143
10144 mutex_lock(&mutex);
10145 old_period = sysctl_sched_rt_period;
10146 old_runtime = sysctl_sched_rt_runtime;
10147
10148 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10149
10150 if (!ret && write) {
10151 ret = sched_rt_global_constraints();
10152 if (ret) {
10153 sysctl_sched_rt_period = old_period;
10154 sysctl_sched_rt_runtime = old_runtime;
10155 } else {
10156 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10157 def_rt_bandwidth.rt_period =
10158 ns_to_ktime(global_rt_period());
10159 }
10160 }
10161 mutex_unlock(&mutex);
10162
10163 return ret;
10164}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010165
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010166#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010167
10168/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +020010169static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010170{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010171 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10172 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010173}
10174
10175static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +020010176cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010177{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010178 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010179
Paul Menage2b01dfe2007-10-24 18:23:50 +020010180 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010181 /* This is early initialization for the top cgroup */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010182 return &init_task_group.css;
10183 }
10184
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +020010185 parent = cgroup_tg(cgrp->parent);
10186 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010187 if (IS_ERR(tg))
10188 return ERR_PTR(-ENOMEM);
10189
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010190 return &tg->css;
10191}
10192
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010193static void
10194cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010195{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010196 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010197
10198 sched_destroy_group(tg);
10199}
10200
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010201static int
10202cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10203 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010204{
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010205#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Giani54e99122009-02-27 15:13:54 +053010206 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010207 return -EINVAL;
10208#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010209 /* We don't support RT-tasks being in separate groups */
10210 if (tsk->sched_class != &fair_sched_class)
10211 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +010010212#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010213
10214 return 0;
10215}
10216
10217static void
Paul Menage2b01dfe2007-10-24 18:23:50 +020010218cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010219 struct cgroup *old_cont, struct task_struct *tsk)
10220{
10221 sched_move_task(tsk);
10222}
10223
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010224#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -070010225static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +020010226 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010227{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010228 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010229}
10230
Paul Menagef4c753b2008-04-29 00:59:56 -070010231static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010232{
Paul Menage2b01dfe2007-10-24 18:23:50 +020010233 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010234
10235 return (u64) tg->shares;
10236}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010237#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010238
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010239#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -070010240static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -070010241 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010242{
Paul Menage06ecb272008-04-29 01:00:06 -070010243 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010244}
10245
Paul Menage06ecb272008-04-29 01:00:06 -070010246static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010247{
Paul Menage06ecb272008-04-29 01:00:06 -070010248 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010249}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010250
10251static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10252 u64 rt_period_us)
10253{
10254 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10255}
10256
10257static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10258{
10259 return sched_group_rt_period(cgroup_tg(cgrp));
10260}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +020010261#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010262
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010263static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010264#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010265 {
10266 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -070010267 .read_u64 = cpu_shares_read_u64,
10268 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010269 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010270#endif
10271#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010272 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +010010273 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -070010274 .read_s64 = cpu_rt_runtime_read,
10275 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +010010276 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010277 {
10278 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -070010279 .read_u64 = cpu_rt_period_read_uint,
10280 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +020010281 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010282#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010283};
10284
10285static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10286{
Paul Menagefe5c7cc2007-10-29 21:18:11 +010010287 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010288}
10289
10290struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +010010291 .name = "cpu",
10292 .create = cpu_cgroup_create,
10293 .destroy = cpu_cgroup_destroy,
10294 .can_attach = cpu_cgroup_can_attach,
10295 .attach = cpu_cgroup_attach,
10296 .populate = cpu_cgroup_populate,
10297 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -070010298 .early_init = 1,
10299};
10300
Peter Zijlstra052f1dc2008-02-13 15:45:40 +010010301#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010302
10303#ifdef CONFIG_CGROUP_CPUACCT
10304
10305/*
10306 * CPU accounting code for task groups.
10307 *
10308 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10309 * (balbir@in.ibm.com).
10310 */
10311
Bharata B Rao934352f2008-11-10 20:41:13 +053010312/* track cpu usage of a group of tasks and its child groups */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010313struct cpuacct {
10314 struct cgroup_subsys_state css;
10315 /* cpuusage holds pointer to a u64-type object on every cpu */
10316 u64 *cpuusage;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010317 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
Bharata B Rao934352f2008-11-10 20:41:13 +053010318 struct cpuacct *parent;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010319};
10320
10321struct cgroup_subsys cpuacct_subsys;
10322
10323/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010324static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010325{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010326 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010327 struct cpuacct, css);
10328}
10329
10330/* return cpu accounting group to which this task belongs */
10331static inline struct cpuacct *task_ca(struct task_struct *tsk)
10332{
10333 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10334 struct cpuacct, css);
10335}
10336
10337/* create a new cpu accounting group */
10338static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +053010339 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010340{
10341 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010342 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010343
10344 if (!ca)
Bharata B Raoef12fef2009-03-31 10:02:22 +053010345 goto out;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010346
10347 ca->cpuusage = alloc_percpu(u64);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010348 if (!ca->cpuusage)
10349 goto out_free_ca;
10350
10351 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10352 if (percpu_counter_init(&ca->cpustat[i], 0))
10353 goto out_free_counters;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010354
Bharata B Rao934352f2008-11-10 20:41:13 +053010355 if (cgrp->parent)
10356 ca->parent = cgroup_ca(cgrp->parent);
10357
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010358 return &ca->css;
Bharata B Raoef12fef2009-03-31 10:02:22 +053010359
10360out_free_counters:
10361 while (--i >= 0)
10362 percpu_counter_destroy(&ca->cpustat[i]);
10363 free_percpu(ca->cpuusage);
10364out_free_ca:
10365 kfree(ca);
10366out:
10367 return ERR_PTR(-ENOMEM);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010368}
10369
10370/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +010010371static void
Dhaval Giani32cd7562008-02-29 10:02:43 +053010372cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010373{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010374 struct cpuacct *ca = cgroup_ca(cgrp);
Bharata B Raoef12fef2009-03-31 10:02:22 +053010375 int i;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010376
Bharata B Raoef12fef2009-03-31 10:02:22 +053010377 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10378 percpu_counter_destroy(&ca->cpustat[i]);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010379 free_percpu(ca->cpuusage);
10380 kfree(ca);
10381}
10382
Ken Chen720f5492008-12-15 22:02:01 -080010383static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10384{
Rusty Russellb36128c2009-02-20 16:29:08 +090010385 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010386 u64 data;
10387
10388#ifndef CONFIG_64BIT
10389 /*
10390 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10391 */
10392 spin_lock_irq(&cpu_rq(cpu)->lock);
10393 data = *cpuusage;
10394 spin_unlock_irq(&cpu_rq(cpu)->lock);
10395#else
10396 data = *cpuusage;
10397#endif
10398
10399 return data;
10400}
10401
10402static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10403{
Rusty Russellb36128c2009-02-20 16:29:08 +090010404 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Ken Chen720f5492008-12-15 22:02:01 -080010405
10406#ifndef CONFIG_64BIT
10407 /*
10408 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10409 */
10410 spin_lock_irq(&cpu_rq(cpu)->lock);
10411 *cpuusage = val;
10412 spin_unlock_irq(&cpu_rq(cpu)->lock);
10413#else
10414 *cpuusage = val;
10415#endif
10416}
10417
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010418/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +053010419static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010420{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010421 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010422 u64 totalcpuusage = 0;
10423 int i;
10424
Ken Chen720f5492008-12-15 22:02:01 -080010425 for_each_present_cpu(i)
10426 totalcpuusage += cpuacct_cpuusage_read(ca, i);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010427
10428 return totalcpuusage;
10429}
10430
Dhaval Giani0297b802008-02-29 10:02:44 +053010431static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10432 u64 reset)
10433{
10434 struct cpuacct *ca = cgroup_ca(cgrp);
10435 int err = 0;
10436 int i;
10437
10438 if (reset) {
10439 err = -EINVAL;
10440 goto out;
10441 }
10442
Ken Chen720f5492008-12-15 22:02:01 -080010443 for_each_present_cpu(i)
10444 cpuacct_cpuusage_write(ca, i, 0);
Dhaval Giani0297b802008-02-29 10:02:44 +053010445
Dhaval Giani0297b802008-02-29 10:02:44 +053010446out:
10447 return err;
10448}
10449
Ken Chene9515c32008-12-15 22:04:15 -080010450static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10451 struct seq_file *m)
10452{
10453 struct cpuacct *ca = cgroup_ca(cgroup);
10454 u64 percpu;
10455 int i;
10456
10457 for_each_present_cpu(i) {
10458 percpu = cpuacct_cpuusage_read(ca, i);
10459 seq_printf(m, "%llu ", (unsigned long long) percpu);
10460 }
10461 seq_printf(m, "\n");
10462 return 0;
10463}
10464
Bharata B Raoef12fef2009-03-31 10:02:22 +053010465static const char *cpuacct_stat_desc[] = {
10466 [CPUACCT_STAT_USER] = "user",
10467 [CPUACCT_STAT_SYSTEM] = "system",
10468};
10469
10470static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10471 struct cgroup_map_cb *cb)
10472{
10473 struct cpuacct *ca = cgroup_ca(cgrp);
10474 int i;
10475
10476 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10477 s64 val = percpu_counter_read(&ca->cpustat[i]);
10478 val = cputime64_to_clock_t(val);
10479 cb->fill(cb, cpuacct_stat_desc[i], val);
10480 }
10481 return 0;
10482}
10483
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010484static struct cftype files[] = {
10485 {
10486 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -070010487 .read_u64 = cpuusage_read,
10488 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010489 },
Ken Chene9515c32008-12-15 22:04:15 -080010490 {
10491 .name = "usage_percpu",
10492 .read_seq_string = cpuacct_percpu_seq_read,
10493 },
Bharata B Raoef12fef2009-03-31 10:02:22 +053010494 {
10495 .name = "stat",
10496 .read_map = cpuacct_stats_show,
10497 },
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010498};
10499
Dhaval Giani32cd7562008-02-29 10:02:43 +053010500static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010501{
Dhaval Giani32cd7562008-02-29 10:02:43 +053010502 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010503}
10504
10505/*
10506 * charge this task's execution time to its accounting group.
10507 *
10508 * called with rq->lock held.
10509 */
10510static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10511{
10512 struct cpuacct *ca;
Bharata B Rao934352f2008-11-10 20:41:13 +053010513 int cpu;
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010514
Li Zefanc40c6f82009-02-26 15:40:15 +080010515 if (unlikely(!cpuacct_subsys.active))
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010516 return;
10517
Bharata B Rao934352f2008-11-10 20:41:13 +053010518 cpu = task_cpu(tsk);
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010519
10520 rcu_read_lock();
10521
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010522 ca = task_ca(tsk);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010523
Bharata B Rao934352f2008-11-10 20:41:13 +053010524 for (; ca; ca = ca->parent) {
Rusty Russellb36128c2009-02-20 16:29:08 +090010525 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010526 *cpuusage += cputime;
10527 }
Bharata B Raoa18b83b2009-03-23 10:02:53 +053010528
10529 rcu_read_unlock();
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010530}
10531
Bharata B Raoef12fef2009-03-31 10:02:22 +053010532/*
10533 * Charge the system/user time to the task's accounting group.
10534 */
10535static void cpuacct_update_stats(struct task_struct *tsk,
10536 enum cpuacct_stat_index idx, cputime_t val)
10537{
10538 struct cpuacct *ca;
10539
10540 if (unlikely(!cpuacct_subsys.active))
10541 return;
10542
10543 rcu_read_lock();
10544 ca = task_ca(tsk);
10545
10546 do {
10547 percpu_counter_add(&ca->cpustat[idx], val);
10548 ca = ca->parent;
10549 } while (ca);
10550 rcu_read_unlock();
10551}
10552
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +010010553struct cgroup_subsys cpuacct_subsys = {
10554 .name = "cpuacct",
10555 .create = cpuacct_create,
10556 .destroy = cpuacct_destroy,
10557 .populate = cpuacct_populate,
10558 .subsys_id = cpuacct_subsys_id,
10559};
10560#endif /* CONFIG_CGROUP_CPUACCT */