blob: e8819bc6f462c18761c11861b4808fd28223b431 [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>
Linus Torvalds1da177e2005-04-16 15:20:36 -070042#include <linux/security.h>
43#include <linux/notifier.h>
44#include <linux/profile.h>
Nigel Cunningham7dfb7102006-12-06 20:34:23 -080045#include <linux/freezer.h>
akpm@osdl.org198e2f12006-01-12 01:05:30 -080046#include <linux/vmalloc.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070047#include <linux/blkdev.h>
48#include <linux/delay.h>
Pavel Emelyanovb4888932007-10-18 23:40:14 -070049#include <linux/pid_namespace.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070050#include <linux/smp.h>
51#include <linux/threads.h>
52#include <linux/timer.h>
53#include <linux/rcupdate.h>
54#include <linux/cpu.h>
55#include <linux/cpuset.h>
56#include <linux/percpu.h>
57#include <linux/kthread.h>
Alexey Dobriyanb5aadf72008-10-06 13:23:43 +040058#include <linux/proc_fs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070059#include <linux/seq_file.h>
Nick Piggine692ab52007-07-26 13:40:43 +020060#include <linux/sysctl.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070061#include <linux/syscalls.h>
62#include <linux/times.h>
Jay Lan8f0ab512006-09-30 23:28:59 -070063#include <linux/tsacct_kern.h>
bibo maoc6fd91f2006-03-26 01:38:20 -080064#include <linux/kprobes.h>
Shailabh Nagar0ff92242006-07-14 00:24:37 -070065#include <linux/delayacct.h>
Eric Dumazet5517d862007-05-08 00:32:57 -070066#include <linux/reciprocal_div.h>
Ingo Molnardff06c12007-07-09 18:52:00 +020067#include <linux/unistd.h>
Jens Axboef5ff8422007-09-21 09:19:54 +020068#include <linux/pagemap.h>
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +010069#include <linux/hrtimer.h>
Reynes Philippe30914a52008-03-17 16:19:05 -070070#include <linux/tick.h>
Mike Travis434d53b2008-04-04 18:11:04 -070071#include <linux/bootmem.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>
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -040075#include <trace/sched.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070076
Eric Dumazet5517d862007-05-08 00:32:57 -070077#include <asm/tlb.h>
Satyam Sharma838225b2007-10-24 18:23:50 +020078#include <asm/irq_regs.h>
Linus Torvalds1da177e2005-04-16 15:20:36 -070079
Gregory Haskins6e0534f2008-05-12 21:21:01 +020080#include "sched_cpupri.h"
81
Linus Torvalds1da177e2005-04-16 15:20:36 -070082/*
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
85 * and back.
86 */
87#define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88#define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89#define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
90
91/*
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
95 */
96#define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97#define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98#define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
99
100/*
Ingo Molnard7876a02008-01-25 21:08:19 +0100101 * Helpers for converting nanosecond timing to jiffy resolution
Linus Torvalds1da177e2005-04-16 15:20:36 -0700102 */
Eric Dumazetd6322fa2007-11-09 22:39:38 +0100103#define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
Linus Torvalds1da177e2005-04-16 15:20:36 -0700104
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200105#define NICE_0_LOAD SCHED_LOAD_SCALE
106#define NICE_0_SHIFT SCHED_LOAD_SHIFT
107
Linus Torvalds1da177e2005-04-16 15:20:36 -0700108/*
109 * These are the 'tuning knobs' of the scheduler:
110 *
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +0200111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
Linus Torvalds1da177e2005-04-16 15:20:36 -0700112 * Timeslices get refilled after they expire.
113 */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700114#define DEF_TIMESLICE (100 * HZ / 1000)
Peter Williams2dd73a42006-06-27 02:54:34 -0700115
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200116/*
117 * single value that denotes runtime == period, ie unlimited time.
118 */
119#define RUNTIME_INF ((u64)~0ULL)
120
Eric Dumazet5517d862007-05-08 00:32:57 -0700121#ifdef CONFIG_SMP
122/*
123 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
124 * Since cpu_power is a 'constant', we can use a reciprocal divide.
125 */
126static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
127{
128 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129}
130
131/*
132 * Each time a sched group cpu_power is changed,
133 * we must compute its reciprocal value
134 */
135static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
136{
137 sg->__cpu_power += val;
138 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139}
140#endif
141
Ingo Molnare05606d2007-07-09 18:51:59 +0200142static inline int rt_policy(int policy)
143{
Roel Kluin3f33a7c2008-05-13 23:44:11 +0200144 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
Ingo Molnare05606d2007-07-09 18:51:59 +0200145 return 1;
146 return 0;
147}
148
149static inline int task_has_rt_policy(struct task_struct *p)
150{
151 return rt_policy(p->policy);
152}
153
Linus Torvalds1da177e2005-04-16 15:20:36 -0700154/*
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200155 * This is the priority-queue data structure of the RT scheduling class:
Linus Torvalds1da177e2005-04-16 15:20:36 -0700156 */
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200157struct rt_prio_array {
158 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
159 struct list_head queue[MAX_RT_PRIO];
160};
Linus Torvalds1da177e2005-04-16 15:20:36 -0700161
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200162struct rt_bandwidth {
Ingo Molnarea736ed2008-03-25 13:51:45 +0100163 /* nests inside the rq lock: */
164 spinlock_t rt_runtime_lock;
165 ktime_t rt_period;
166 u64 rt_runtime;
167 struct hrtimer rt_period_timer;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200168};
169
170static struct rt_bandwidth def_rt_bandwidth;
171
172static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
173
174static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
175{
176 struct rt_bandwidth *rt_b =
177 container_of(timer, struct rt_bandwidth, rt_period_timer);
178 ktime_t now;
179 int overrun;
180 int idle = 0;
181
182 for (;;) {
183 now = hrtimer_cb_get_time(timer);
184 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
185
186 if (!overrun)
187 break;
188
189 idle = do_sched_rt_period_timer(rt_b, overrun);
190 }
191
192 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193}
194
195static
196void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
197{
198 rt_b->rt_period = ns_to_ktime(period);
199 rt_b->rt_runtime = runtime;
200
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200201 spin_lock_init(&rt_b->rt_runtime_lock);
202
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200203 hrtimer_init(&rt_b->rt_period_timer,
204 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
205 rt_b->rt_period_timer.function = sched_rt_period_timer;
Thomas Gleixnerccc7dad2008-09-29 15:47:42 +0200206 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200207}
208
Krzysztof Heltc8bfff62008-09-05 23:46:19 +0200209static inline int rt_bandwidth_enabled(void)
210{
211 return sysctl_sched_rt_runtime >= 0;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200212}
213
214static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
215{
216 ktime_t now;
217
Peter Zijlstra0b148fa2008-08-19 12:33:04 +0200218 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200219 return;
220
221 if (hrtimer_active(&rt_b->rt_period_timer))
222 return;
223
224 spin_lock(&rt_b->rt_runtime_lock);
225 for (;;) {
226 if (hrtimer_active(&rt_b->rt_period_timer))
227 break;
228
229 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
230 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
Arjan van de Vencc584b22008-09-01 15:02:30 -0700231 hrtimer_start_expires(&rt_b->rt_period_timer,
232 HRTIMER_MODE_ABS);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200233 }
234 spin_unlock(&rt_b->rt_runtime_lock);
235}
236
237#ifdef CONFIG_RT_GROUP_SCHED
238static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
239{
240 hrtimer_cancel(&rt_b->rt_period_timer);
241}
242#endif
243
Heiko Carstens712555e2008-04-28 11:33:07 +0200244/*
245 * sched_domains_mutex serializes calls to arch_init_sched_domains,
246 * detach_destroy_domains and partition_sched_domains.
247 */
248static DEFINE_MUTEX(sched_domains_mutex);
249
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100250#ifdef CONFIG_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200251
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700252#include <linux/cgroup.h>
253
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200254struct cfs_rq;
255
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100256static LIST_HEAD(task_groups);
257
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200258/* task group related information */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200259struct task_group {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100260#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700261 struct cgroup_subsys_state css;
262#endif
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100263
264#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200265 /* schedulable entities of this group on each cpu */
266 struct sched_entity **se;
267 /* runqueue "owned" by this group on each cpu */
268 struct cfs_rq **cfs_rq;
269 unsigned long shares;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100270#endif
271
272#ifdef CONFIG_RT_GROUP_SCHED
273 struct sched_rt_entity **rt_se;
274 struct rt_rq **rt_rq;
275
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200276 struct rt_bandwidth rt_bandwidth;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100277#endif
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +0100278
Srivatsa Vaddagiriae8393e2007-10-29 21:18:11 +0100279 struct rcu_head rcu;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100280 struct list_head list;
Peter Zijlstraf473aa52008-04-19 19:45:00 +0200281
282 struct task_group *parent;
283 struct list_head siblings;
284 struct list_head children;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200285};
286
Dhaval Giani354d60c2008-04-19 19:44:59 +0200287#ifdef CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200288
289/*
290 * Root task group.
291 * Every UID task group (including init_task_group aka UID-0) will
292 * be a child to this group.
293 */
294struct task_group root_task_group;
295
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100296#ifdef CONFIG_FAIR_GROUP_SCHED
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200297/* Default task group's sched entity on each cpu */
298static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
299/* Default task group's cfs_rq on each cpu */
300static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200301#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100302
303#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100304static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
305static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200306#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200307#else /* !CONFIG_USER_SCHED */
Peter Zijlstraeff766a2008-04-19 19:45:00 +0200308#define root_task_group init_task_group
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +0200309#endif /* CONFIG_USER_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100310
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100311/* task_group_lock serializes add/remove of task groups and also changes to
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100312 * a task group's cpu shares.
313 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +0100314static DEFINE_SPINLOCK(task_group_lock);
Srivatsa Vaddagiriec2c5072008-01-25 21:07:59 +0100315
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100316#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100317#ifdef CONFIG_USER_SCHED
Ingo Molnar0eab9142008-01-25 21:08:19 +0100318# define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200319#else /* !CONFIG_USER_SCHED */
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100320# define INIT_TASK_GROUP_LOAD NICE_0_LOAD
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +0200321#endif /* CONFIG_USER_SCHED */
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200322
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800323/*
Lai Jiangshan2e084782008-06-12 16:42:58 +0800324 * A weight of 0 or 1 can cause arithmetics problems.
325 * A weight of a cfs_rq is the sum of weights of which entities
326 * are queued on this cfs_rq, so a weight of a entity should not be
327 * too large, so as the shares value of a task group.
Miao Xiecb4ad1f2008-04-28 12:54:56 +0800328 * (The default weight is 1024 - so there's no practical
329 * limitation from this.)
330 */
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200331#define MIN_SHARES 2
Lai Jiangshan2e084782008-06-12 16:42:58 +0800332#define MAX_SHARES (1UL << 18)
Peter Zijlstra18d95a22008-04-19 19:45:00 +0200333
Srivatsa Vaddagiri93f992c2008-01-25 21:07:59 +0100334static int init_task_group_load = INIT_TASK_GROUP_LOAD;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100335#endif
336
337/* Default task group.
338 * Every task in system belong to this group at bootup.
339 */
Mike Travis434d53b2008-04-04 18:11:04 -0700340struct task_group init_task_group;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200341
342/* return group to which a task belongs */
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200343static inline struct task_group *task_group(struct task_struct *p)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200344{
Ingo Molnar4cf86d72007-10-15 17:00:14 +0200345 struct task_group *tg;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200346
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100347#ifdef CONFIG_USER_SCHED
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200348 tg = p->user->tg;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100349#elif defined(CONFIG_CGROUP_SCHED)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -0700350 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
351 struct task_group, css);
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200352#else
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100353 tg = &init_task_group;
Srivatsa Vaddagiri24e377a2007-10-15 17:00:09 +0200354#endif
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +0200355 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200356}
357
358/* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100359static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200360{
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100361#ifdef CONFIG_FAIR_GROUP_SCHED
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +0100362 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
363 p->se.parent = task_group(p)->se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100364#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100365
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100366#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100367 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
368 p->rt.parent = task_group(p)->rt_se[cpu];
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100369#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200370}
371
372#else
373
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100374static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
Peter Zijlstra83378262008-06-27 13:41:37 +0200375static inline struct task_group *task_group(struct task_struct *p)
376{
377 return NULL;
378}
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200379
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100380#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +0200381
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200382/* CFS-related fields in a runqueue */
383struct cfs_rq {
384 struct load_weight load;
385 unsigned long nr_running;
386
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200387 u64 exec_clock;
Ingo Molnare9acbff2007-10-15 17:00:04 +0200388 u64 min_vruntime;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200389
390 struct rb_root tasks_timeline;
391 struct rb_node *rb_leftmost;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +0200392
393 struct list_head tasks;
394 struct list_head *balance_iterator;
395
396 /*
397 * 'curr' points to currently running entity on this cfs_rq.
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200398 * It is set to NULL otherwise (i.e when none are currently running).
399 */
Peter Zijlstraaa2ac252008-03-14 21:12:12 +0100400 struct sched_entity *curr, *next;
Peter Zijlstraddc97292007-10-15 17:00:10 +0200401
402 unsigned long nr_spread_over;
403
Ingo Molnar62160e3f2007-10-15 17:00:03 +0200404#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200405 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
406
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100407 /*
408 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200409 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
410 * (like users, containers etc.)
411 *
412 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
413 * list is used during load balance.
414 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100415 struct list_head leaf_cfs_rq_list;
416 struct task_group *tg; /* group that "owns" this runqueue */
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200417
418#ifdef CONFIG_SMP
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200419 /*
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200420 * the part of load.weight contributed by tasks
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200421 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200422 unsigned long task_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200423
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200424 /*
425 * h_load = weight * f(tg)
426 *
427 * Where f(tg) is the recursive weight fraction assigned to
428 * this group.
429 */
430 unsigned long h_load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200431
Peter Zijlstrac8cba852008-06-27 13:41:23 +0200432 /*
433 * this cpu's part of tg->shares
434 */
435 unsigned long shares;
Peter Zijlstraf1d239f2008-06-27 13:41:38 +0200436
437 /*
438 * load.weight at the time we set shares
439 */
440 unsigned long rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +0200441#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200442#endif
443};
444
445/* Real-Time classes' related field in a runqueue: */
446struct rt_rq {
447 struct rt_prio_array active;
Steven Rostedt63489e42008-01-25 21:08:03 +0100448 unsigned long rt_nr_running;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100449#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100450 int highest_prio; /* highest queued rt task prio */
451#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100452#ifdef CONFIG_SMP
Gregory Haskins73fe6aa2008-01-25 21:08:07 +0100453 unsigned long rt_nr_migratory;
Gregory Haskinsa22d7fc2008-01-25 21:08:12 +0100454 int overloaded;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100455#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100456 int rt_throttled;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100457 u64 rt_time;
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200458 u64 rt_runtime;
Ingo Molnarea736ed2008-03-25 13:51:45 +0100459 /* Nests inside the rq lock: */
Peter Zijlstraac086bc2008-04-19 19:44:58 +0200460 spinlock_t rt_runtime_lock;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100461
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100462#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +0100463 unsigned long rt_nr_boosted;
464
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100465 struct rq *rq;
466 struct list_head leaf_rt_rq_list;
467 struct task_group *tg;
468 struct sched_rt_entity *rt_se;
469#endif
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200470};
471
Gregory Haskins57d885f2008-01-25 21:08:18 +0100472#ifdef CONFIG_SMP
473
474/*
475 * We add the notion of a root-domain which will be used to define per-domain
Ingo Molnar0eab9142008-01-25 21:08:19 +0100476 * variables. Each exclusive cpuset essentially defines an island domain by
477 * fully partitioning the member cpus from any other cpuset. Whenever a new
Gregory Haskins57d885f2008-01-25 21:08:18 +0100478 * exclusive cpuset is created, we also create and attach a new root-domain
479 * object.
480 *
Gregory Haskins57d885f2008-01-25 21:08:18 +0100481 */
482struct root_domain {
483 atomic_t refcount;
484 cpumask_t span;
485 cpumask_t online;
Gregory Haskins637f5082008-01-25 21:08:18 +0100486
Ingo Molnar0eab9142008-01-25 21:08:19 +0100487 /*
Gregory Haskins637f5082008-01-25 21:08:18 +0100488 * The "RT overload" flag: it gets set if a CPU has more than
489 * one runnable RT task.
490 */
491 cpumask_t rto_mask;
Ingo Molnar0eab9142008-01-25 21:08:19 +0100492 atomic_t rto_count;
Gregory Haskins6e0534f2008-05-12 21:21:01 +0200493#ifdef CONFIG_SMP
494 struct cpupri cpupri;
495#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +0100496};
497
Gregory Haskinsdc938522008-01-25 21:08:26 +0100498/*
499 * By default the system creates a single root-domain with all cpus as
500 * members (mimicking the global state we have today).
501 */
Gregory Haskins57d885f2008-01-25 21:08:18 +0100502static struct root_domain def_root_domain;
503
504#endif
505
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200506/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700507 * This is the main, per-CPU runqueue data structure.
508 *
509 * Locking rule: those places that want to lock multiple runqueues
510 * (such as the load balancing or the thread migration code), lock
511 * acquire operations must be ordered by ascending &runqueue.
512 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700513struct rq {
Ingo Molnard8016492007-10-18 21:32:55 +0200514 /* runqueue lock: */
515 spinlock_t lock;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700516
517 /*
518 * nr_running and cpu_load should be in the same cacheline because
519 * remote CPUs use both these fields when doing load calculation.
520 */
521 unsigned long nr_running;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200522 #define CPU_LOAD_IDX_MAX 5
523 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
Siddha, Suresh Bbdecea32007-05-08 00:32:48 -0700524 unsigned char idle_at_tick;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700525#ifdef CONFIG_NO_HZ
Guillaume Chazarain15934a32008-04-19 19:44:57 +0200526 unsigned long last_tick_seen;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -0700527 unsigned char in_nohz_recently;
528#endif
Ingo Molnard8016492007-10-18 21:32:55 +0200529 /* capture load from *all* tasks on this cpu: */
530 struct load_weight load;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200531 unsigned long nr_load_updates;
532 u64 nr_switches;
533
534 struct cfs_rq cfs;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100535 struct rt_rq rt;
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100536
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200537#ifdef CONFIG_FAIR_GROUP_SCHED
Ingo Molnard8016492007-10-18 21:32:55 +0200538 /* list of leaf cfs_rq on this cpu: */
539 struct list_head leaf_cfs_rq_list;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +0100540#endif
541#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +0100542 struct list_head leaf_rt_rq_list;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700543#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -0700544
545 /*
546 * This is part of a global counter where only the total sum
547 * over all CPUs matters. A task can increase this counter on
548 * one CPU and if it got migrated afterwards it may decrease
549 * it on another CPU. Always updated under the runqueue lock:
550 */
551 unsigned long nr_uninterruptible;
552
Ingo Molnar36c8b582006-07-03 00:25:41 -0700553 struct task_struct *curr, *idle;
Christoph Lameterc9819f42006-12-10 02:20:25 -0800554 unsigned long next_balance;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700555 struct mm_struct *prev_mm;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200556
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200557 u64 clock;
Ingo Molnar6aa645e2007-07-09 18:51:58 +0200558
Linus Torvalds1da177e2005-04-16 15:20:36 -0700559 atomic_t nr_iowait;
560
561#ifdef CONFIG_SMP
Ingo Molnar0eab9142008-01-25 21:08:19 +0100562 struct root_domain *rd;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700563 struct sched_domain *sd;
564
565 /* For active balancing */
566 int active_balance;
567 int push_cpu;
Ingo Molnard8016492007-10-18 21:32:55 +0200568 /* cpu of this runqueue: */
569 int cpu;
Gregory Haskins1f11eb62008-06-04 15:04:05 -0400570 int online;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700571
Peter Zijlstraa8a51d52008-06-27 13:41:26 +0200572 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700573
Ingo Molnar36c8b582006-07-03 00:25:41 -0700574 struct task_struct *migration_thread;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700575 struct list_head migration_queue;
576#endif
577
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100578#ifdef CONFIG_SCHED_HRTICK
Peter Zijlstra31656512008-07-18 18:01:23 +0200579#ifdef CONFIG_SMP
580 int hrtick_csd_pending;
581 struct call_single_data hrtick_csd;
582#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100583 struct hrtimer hrtick_timer;
584#endif
585
Linus Torvalds1da177e2005-04-16 15:20:36 -0700586#ifdef CONFIG_SCHEDSTATS
587 /* latency stats */
588 struct sched_info rq_sched_info;
589
590 /* sys_sched_yield() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200591 unsigned int yld_exp_empty;
592 unsigned int yld_act_empty;
593 unsigned int yld_both_empty;
594 unsigned int yld_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700595
596 /* schedule() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200597 unsigned int sched_switch;
598 unsigned int sched_count;
599 unsigned int sched_goidle;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700600
601 /* try_to_wake_up() stats */
Ken Chen480b9432007-10-18 21:32:56 +0200602 unsigned int ttwu_count;
603 unsigned int ttwu_local;
Ingo Molnarb8efb562007-10-15 17:00:10 +0200604
605 /* BKL stats */
Ken Chen480b9432007-10-18 21:32:56 +0200606 unsigned int bkl_count;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700607#endif
608};
609
Fenghua Yuf34e3b62007-07-19 01:48:13 -0700610static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700611
Peter Zijlstra15afe092008-09-20 23:38:02 +0200612static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
Ingo Molnardd41f592007-07-09 18:51:59 +0200613{
Peter Zijlstra15afe092008-09-20 23:38:02 +0200614 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
Ingo Molnardd41f592007-07-09 18:51:59 +0200615}
616
Christoph Lameter0a2966b2006-09-25 23:30:51 -0700617static inline int cpu_of(struct rq *rq)
618{
619#ifdef CONFIG_SMP
620 return rq->cpu;
621#else
622 return 0;
623#endif
624}
625
Ingo Molnar20d315d2007-07-09 18:51:58 +0200626/*
Nick Piggin674311d2005-06-25 14:57:27 -0700627 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -0700628 * See detach_destroy_domains: synchronize_sched for details.
Nick Piggin674311d2005-06-25 14:57:27 -0700629 *
630 * The domain tree of any CPU may only be accessed from within
631 * preempt-disabled sections.
632 */
Ingo Molnar48f24c42006-07-03 00:25:40 -0700633#define for_each_domain(cpu, __sd) \
634 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700635
636#define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
637#define this_rq() (&__get_cpu_var(runqueues))
638#define task_rq(p) cpu_rq(task_cpu(p))
639#define cpu_curr(cpu) (cpu_rq(cpu)->curr)
640
Peter Zijlstra3e51f332008-05-03 18:29:28 +0200641static inline void update_rq_clock(struct rq *rq)
642{
643 rq->clock = sched_clock_cpu(cpu_of(rq));
644}
645
Ingo Molnare436d802007-07-19 21:28:35 +0200646/*
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200647 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
648 */
649#ifdef CONFIG_SCHED_DEBUG
650# define const_debug __read_mostly
651#else
652# define const_debug static const
653#endif
654
Ingo Molnar017730c2008-05-12 21:20:52 +0200655/**
656 * runqueue_is_locked
657 *
658 * Returns true if the current cpu runqueue is locked.
659 * This interface allows printk to be called with the runqueue lock
660 * held and know whether or not it is OK to wake up the klogd.
661 */
662int runqueue_is_locked(void)
663{
664 int cpu = get_cpu();
665 struct rq *rq = cpu_rq(cpu);
666 int ret;
667
668 ret = spin_is_locked(&rq->lock);
669 put_cpu();
670 return ret;
671}
672
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200673/*
674 * Debugging: various feature bits
675 */
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200676
677#define SCHED_FEAT(name, enabled) \
678 __SCHED_FEAT_##name ,
679
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200680enum {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200681#include "sched_features.h"
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200682};
683
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200684#undef SCHED_FEAT
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200685
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200686#define SCHED_FEAT(name, enabled) \
687 (1UL << __SCHED_FEAT_##name) * enabled |
688
689const_debug unsigned int sysctl_sched_features =
690#include "sched_features.h"
691 0;
692
693#undef SCHED_FEAT
694
695#ifdef CONFIG_SCHED_DEBUG
696#define SCHED_FEAT(name, enabled) \
697 #name ,
698
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700699static __read_mostly char *sched_feat_names[] = {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200700#include "sched_features.h"
701 NULL
702};
703
704#undef SCHED_FEAT
705
Harvey Harrison983ed7a2008-04-24 18:17:55 -0700706static int sched_feat_open(struct inode *inode, struct file *filp)
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200707{
708 filp->private_data = inode->i_private;
709 return 0;
710}
711
712static ssize_t
713sched_feat_read(struct file *filp, char __user *ubuf,
714 size_t cnt, loff_t *ppos)
715{
716 char *buf;
717 int r = 0;
718 int len = 0;
719 int i;
720
721 for (i = 0; sched_feat_names[i]; i++) {
722 len += strlen(sched_feat_names[i]);
723 len += 4;
724 }
725
726 buf = kmalloc(len + 2, GFP_KERNEL);
727 if (!buf)
728 return -ENOMEM;
729
730 for (i = 0; sched_feat_names[i]; i++) {
731 if (sysctl_sched_features & (1UL << i))
732 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
733 else
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200734 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200735 }
736
737 r += sprintf(buf + r, "\n");
738 WARN_ON(r >= len + 2);
739
740 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
741
742 kfree(buf);
743
744 return r;
745}
746
747static ssize_t
748sched_feat_write(struct file *filp, const char __user *ubuf,
749 size_t cnt, loff_t *ppos)
750{
751 char buf[64];
752 char *cmp = buf;
753 int neg = 0;
754 int i;
755
756 if (cnt > 63)
757 cnt = 63;
758
759 if (copy_from_user(&buf, ubuf, cnt))
760 return -EFAULT;
761
762 buf[cnt] = 0;
763
Ingo Molnarc24b7c52008-04-18 10:55:34 +0200764 if (strncmp(buf, "NO_", 3) == 0) {
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200765 neg = 1;
766 cmp += 3;
767 }
768
769 for (i = 0; sched_feat_names[i]; i++) {
770 int len = strlen(sched_feat_names[i]);
771
772 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
773 if (neg)
774 sysctl_sched_features &= ~(1UL << i);
775 else
776 sysctl_sched_features |= (1UL << i);
777 break;
778 }
779 }
780
781 if (!sched_feat_names[i])
782 return -EINVAL;
783
784 filp->f_pos += cnt;
785
786 return cnt;
787}
788
789static struct file_operations sched_feat_fops = {
790 .open = sched_feat_open,
791 .read = sched_feat_read,
792 .write = sched_feat_write,
793};
794
795static __init int sched_init_debug(void)
796{
Peter Zijlstraf00b45c2008-04-19 19:45:00 +0200797 debugfs_create_file("sched_features", 0644, NULL, NULL,
798 &sched_feat_fops);
799
800 return 0;
801}
802late_initcall(sched_init_debug);
803
804#endif
805
806#define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
Ingo Molnarbf5c91b2007-10-15 17:00:04 +0200807
808/*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +0100809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
811 */
812const_debug unsigned int sysctl_sched_nr_migrate = 32;
813
814/*
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200815 * ratelimit for updating the group shares.
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200816 * default: 0.25ms
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200817 */
Peter Zijlstra55cd5342008-08-04 08:54:26 +0200818unsigned int sysctl_sched_shares_ratelimit = 250000;
Peter Zijlstra2398f2c2008-06-27 13:41:35 +0200819
820/*
Peter Zijlstraffda12a2008-10-17 19:27:02 +0200821 * Inject some fuzzyness into changing the per-cpu group shares
822 * this avoids remote rq-locks at the expense of fairness.
823 * default: 4
824 */
825unsigned int sysctl_sched_shares_thresh = 4;
826
827/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100828 * period over which we measure -rt task cpu usage in us.
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100829 * default: 1s
830 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100831unsigned int sysctl_sched_rt_period = 1000000;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100832
Ingo Molnar6892b752008-02-13 14:02:36 +0100833static __read_mostly int scheduler_running;
834
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100835/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100836 * part of the period that we allow rt tasks to run in us.
837 * default: 0.95s
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100838 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +0100839int sysctl_sched_rt_runtime = 950000;
840
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200841static inline u64 global_rt_period(void)
842{
843 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
844}
845
846static inline u64 global_rt_runtime(void)
847{
roel kluine26873b2008-07-22 16:51:15 -0400848 if (sysctl_sched_rt_runtime < 0)
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +0200849 return RUNTIME_INF;
850
851 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
852}
Peter Zijlstrafa85ae22008-01-25 21:08:29 +0100853
Linus Torvalds1da177e2005-04-16 15:20:36 -0700854#ifndef prepare_arch_switch
Nick Piggin4866cde2005-06-25 14:57:23 -0700855# define prepare_arch_switch(next) do { } while (0)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700856#endif
Nick Piggin4866cde2005-06-25 14:57:23 -0700857#ifndef finish_arch_switch
858# define finish_arch_switch(prev) do { } while (0)
859#endif
860
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100861static inline int task_current(struct rq *rq, struct task_struct *p)
862{
863 return rq->curr == p;
864}
865
Nick Piggin4866cde2005-06-25 14:57:23 -0700866#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar70b97a72006-07-03 00:25:42 -0700867static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700868{
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100869 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700870}
871
Ingo Molnar70b97a72006-07-03 00:25:42 -0700872static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700873{
874}
875
Ingo Molnar70b97a72006-07-03 00:25:42 -0700876static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700877{
Ingo Molnarda04c032005-09-13 11:17:59 +0200878#ifdef CONFIG_DEBUG_SPINLOCK
879 /* this is a valid case when another task releases the spinlock */
880 rq->lock.owner = current;
881#endif
Ingo Molnar8a25d5d2006-07-03 00:24:54 -0700882 /*
883 * If we are tracking spinlock dependencies then we have to
884 * fix up the runqueue lock - which gets 'carried over' from
885 * prev into current:
886 */
887 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
888
Nick Piggin4866cde2005-06-25 14:57:23 -0700889 spin_unlock_irq(&rq->lock);
890}
891
892#else /* __ARCH_WANT_UNLOCKED_CTXSW */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700893static inline int task_running(struct rq *rq, struct task_struct *p)
Nick Piggin4866cde2005-06-25 14:57:23 -0700894{
895#ifdef CONFIG_SMP
896 return p->oncpu;
897#else
Dmitry Adamushko051a1d12007-12-18 15:21:13 +0100898 return task_current(rq, p);
Nick Piggin4866cde2005-06-25 14:57:23 -0700899#endif
900}
901
Ingo Molnar70b97a72006-07-03 00:25:42 -0700902static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -0700903{
904#ifdef CONFIG_SMP
905 /*
906 * We can optimise this out completely for !SMP, because the
907 * SMP rebalancing from interrupt is the only thing that cares
908 * here.
909 */
910 next->oncpu = 1;
911#endif
912#ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
913 spin_unlock_irq(&rq->lock);
914#else
915 spin_unlock(&rq->lock);
916#endif
917}
918
Ingo Molnar70b97a72006-07-03 00:25:42 -0700919static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
Nick Piggin4866cde2005-06-25 14:57:23 -0700920{
921#ifdef CONFIG_SMP
922 /*
923 * After ->oncpu is cleared, the task can be moved to a different CPU.
924 * We must ensure this doesn't happen until the switch is completely
925 * finished.
926 */
927 smp_wmb();
928 prev->oncpu = 0;
929#endif
930#ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
931 local_irq_enable();
932#endif
933}
934#endif /* __ARCH_WANT_UNLOCKED_CTXSW */
Linus Torvalds1da177e2005-04-16 15:20:36 -0700935
936/*
Ingo Molnarb29739f2006-06-27 02:54:51 -0700937 * __task_rq_lock - lock the runqueue a given task resides on.
938 * Must be called interrupts disabled.
939 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700940static inline struct rq *__task_rq_lock(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700941 __acquires(rq->lock)
942{
Andi Kleen3a5c3592007-10-15 17:00:14 +0200943 for (;;) {
944 struct rq *rq = task_rq(p);
945 spin_lock(&rq->lock);
946 if (likely(rq == task_rq(p)))
947 return rq;
Ingo Molnarb29739f2006-06-27 02:54:51 -0700948 spin_unlock(&rq->lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -0700949 }
Ingo Molnarb29739f2006-06-27 02:54:51 -0700950}
951
952/*
Linus Torvalds1da177e2005-04-16 15:20:36 -0700953 * task_rq_lock - lock the runqueue a given task resides on and disable
Ingo Molnar41a2d6c2007-12-05 15:46:09 +0100954 * interrupts. Note the ordering: we can safely lookup the task_rq without
Linus Torvalds1da177e2005-04-16 15:20:36 -0700955 * explicitly disabling preemption.
956 */
Ingo Molnar70b97a72006-07-03 00:25:42 -0700957static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700958 __acquires(rq->lock)
959{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700960 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700961
Andi Kleen3a5c3592007-10-15 17:00:14 +0200962 for (;;) {
963 local_irq_save(*flags);
964 rq = task_rq(p);
965 spin_lock(&rq->lock);
966 if (likely(rq == task_rq(p)))
967 return rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700968 spin_unlock_irqrestore(&rq->lock, *flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -0700969 }
Linus Torvalds1da177e2005-04-16 15:20:36 -0700970}
971
Alexey Dobriyana9957442007-10-15 17:00:13 +0200972static void __task_rq_unlock(struct rq *rq)
Ingo Molnarb29739f2006-06-27 02:54:51 -0700973 __releases(rq->lock)
974{
975 spin_unlock(&rq->lock);
976}
977
Ingo Molnar70b97a72006-07-03 00:25:42 -0700978static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700979 __releases(rq->lock)
980{
981 spin_unlock_irqrestore(&rq->lock, *flags);
982}
983
Linus Torvalds1da177e2005-04-16 15:20:36 -0700984/*
Robert P. J. Daycc2a73b2006-12-10 02:20:00 -0800985 * this_rq_lock - lock this runqueue and disable interrupts.
Linus Torvalds1da177e2005-04-16 15:20:36 -0700986 */
Alexey Dobriyana9957442007-10-15 17:00:13 +0200987static struct rq *this_rq_lock(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -0700988 __acquires(rq->lock)
989{
Ingo Molnar70b97a72006-07-03 00:25:42 -0700990 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -0700991
992 local_irq_disable();
993 rq = this_rq();
994 spin_lock(&rq->lock);
995
996 return rq;
997}
998
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +0100999#ifdef CONFIG_SCHED_HRTICK
1000/*
1001 * Use HR-timers to deliver accurate preemption points.
1002 *
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * reschedule event.
1006 *
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1008 * rq->lock.
1009 */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001010
1011/*
1012 * Use hrtick when:
1013 * - enabled by features
1014 * - hrtimer is actually high res
1015 */
1016static inline int hrtick_enabled(struct rq *rq)
1017{
1018 if (!sched_feat(HRTICK))
1019 return 0;
Ingo Molnarba420592008-07-20 11:02:06 +02001020 if (!cpu_active(cpu_of(rq)))
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001021 return 0;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001022 return hrtimer_is_hres_active(&rq->hrtick_timer);
1023}
1024
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001025static void hrtick_clear(struct rq *rq)
1026{
1027 if (hrtimer_active(&rq->hrtick_timer))
1028 hrtimer_cancel(&rq->hrtick_timer);
1029}
1030
1031/*
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1034 */
1035static enum hrtimer_restart hrtick(struct hrtimer *timer)
1036{
1037 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1038
1039 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1040
1041 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02001042 update_rq_clock(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001043 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1044 spin_unlock(&rq->lock);
1045
1046 return HRTIMER_NORESTART;
1047}
1048
Rabin Vincent95e904c2008-05-11 05:55:33 +05301049#ifdef CONFIG_SMP
Peter Zijlstra31656512008-07-18 18:01:23 +02001050/*
1051 * called from hardirq (IPI) context
1052 */
1053static void __hrtick_start(void *arg)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001054{
Peter Zijlstra31656512008-07-18 18:01:23 +02001055 struct rq *rq = arg;
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001056
Peter Zijlstra31656512008-07-18 18:01:23 +02001057 spin_lock(&rq->lock);
1058 hrtimer_restart(&rq->hrtick_timer);
1059 rq->hrtick_csd_pending = 0;
1060 spin_unlock(&rq->lock);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001061}
1062
Peter Zijlstra31656512008-07-18 18:01:23 +02001063/*
1064 * Called to set the hrtick timer state.
1065 *
1066 * called with rq->lock held and irqs disabled
1067 */
1068static void hrtick_start(struct rq *rq, u64 delay)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001069{
Peter Zijlstra31656512008-07-18 18:01:23 +02001070 struct hrtimer *timer = &rq->hrtick_timer;
1071 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001072
Arjan van de Vencc584b22008-09-01 15:02:30 -07001073 hrtimer_set_expires(timer, time);
Peter Zijlstra31656512008-07-18 18:01:23 +02001074
1075 if (rq == this_rq()) {
1076 hrtimer_restart(timer);
1077 } else if (!rq->hrtick_csd_pending) {
1078 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1079 rq->hrtick_csd_pending = 1;
1080 }
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001081}
1082
1083static int
1084hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1085{
1086 int cpu = (int)(long)hcpu;
1087
1088 switch (action) {
1089 case CPU_UP_CANCELED:
1090 case CPU_UP_CANCELED_FROZEN:
1091 case CPU_DOWN_PREPARE:
1092 case CPU_DOWN_PREPARE_FROZEN:
1093 case CPU_DEAD:
1094 case CPU_DEAD_FROZEN:
Peter Zijlstra31656512008-07-18 18:01:23 +02001095 hrtick_clear(cpu_rq(cpu));
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001096 return NOTIFY_OK;
1097 }
1098
1099 return NOTIFY_DONE;
1100}
1101
Rakib Mullickfa748202008-09-22 14:55:45 -07001102static __init void init_hrtick(void)
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001103{
1104 hotcpu_notifier(hotplug_hrtick, 0);
1105}
Peter Zijlstra31656512008-07-18 18:01:23 +02001106#else
1107/*
1108 * Called to set the hrtick timer state.
1109 *
1110 * called with rq->lock held and irqs disabled
1111 */
1112static void hrtick_start(struct rq *rq, u64 delay)
1113{
1114 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1115}
1116
Andrew Morton006c75f2008-09-22 14:55:46 -07001117static inline void init_hrtick(void)
Peter Zijlstra31656512008-07-18 18:01:23 +02001118{
1119}
Rabin Vincent95e904c2008-05-11 05:55:33 +05301120#endif /* CONFIG_SMP */
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001121
1122static void init_rq_hrtick(struct rq *rq)
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001123{
Peter Zijlstra31656512008-07-18 18:01:23 +02001124#ifdef CONFIG_SMP
1125 rq->hrtick_csd_pending = 0;
1126
1127 rq->hrtick_csd.flags = 0;
1128 rq->hrtick_csd.func = __hrtick_start;
1129 rq->hrtick_csd.info = rq;
1130#endif
1131
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001132 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1133 rq->hrtick_timer.function = hrtick;
Thomas Gleixnerccc7dad2008-09-29 15:47:42 +02001134 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001135}
Andrew Morton006c75f2008-09-22 14:55:46 -07001136#else /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001137static inline void hrtick_clear(struct rq *rq)
1138{
1139}
1140
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001141static inline void init_rq_hrtick(struct rq *rq)
1142{
1143}
1144
Peter Zijlstrab328ca12008-04-29 10:02:46 +02001145static inline void init_hrtick(void)
1146{
1147}
Andrew Morton006c75f2008-09-22 14:55:46 -07001148#endif /* CONFIG_SCHED_HRTICK */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01001149
Ingo Molnar1b9f19c2007-07-09 18:51:59 +02001150/*
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001151 * resched_task - mark a task 'to be rescheduled now'.
1152 *
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1155 * the target CPU.
1156 */
1157#ifdef CONFIG_SMP
1158
1159#ifndef tsk_is_polling
1160#define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161#endif
1162
Peter Zijlstra31656512008-07-18 18:01:23 +02001163static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001164{
1165 int cpu;
1166
1167 assert_spin_locked(&task_rq(p)->lock);
1168
Peter Zijlstra31656512008-07-18 18:01:23 +02001169 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001170 return;
1171
Peter Zijlstra31656512008-07-18 18:01:23 +02001172 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001173
1174 cpu = task_cpu(p);
1175 if (cpu == smp_processor_id())
1176 return;
1177
1178 /* NEED_RESCHED must be visible before we test polling */
1179 smp_mb();
1180 if (!tsk_is_polling(p))
1181 smp_send_reschedule(cpu);
1182}
1183
1184static void resched_cpu(int cpu)
1185{
1186 struct rq *rq = cpu_rq(cpu);
1187 unsigned long flags;
1188
1189 if (!spin_trylock_irqsave(&rq->lock, flags))
1190 return;
1191 resched_task(cpu_curr(cpu));
1192 spin_unlock_irqrestore(&rq->lock, flags);
1193}
Thomas Gleixner06d83082008-03-22 09:20:24 +01001194
1195#ifdef CONFIG_NO_HZ
1196/*
1197 * When add_timer_on() enqueues a timer into the timer wheel of an
1198 * idle CPU then this timer might expire before the next timer event
1199 * which is scheduled to wake up that CPU. In case of a completely
1200 * idle system the next event might even be infinite time into the
1201 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1202 * leaves the inner idle loop so the newly added timer is taken into
1203 * account when the CPU goes back to idle and evaluates the timer
1204 * wheel for the next timer event.
1205 */
1206void wake_up_idle_cpu(int cpu)
1207{
1208 struct rq *rq = cpu_rq(cpu);
1209
1210 if (cpu == smp_processor_id())
1211 return;
1212
1213 /*
1214 * This is safe, as this function is called with the timer
1215 * wheel base lock of (cpu) held. When the CPU is on the way
1216 * to idle and has not yet set rq->curr to idle then it will
1217 * be serialized on the timer wheel base lock and take the new
1218 * timer into account automatically.
1219 */
1220 if (rq->curr != rq->idle)
1221 return;
1222
1223 /*
1224 * We can set TIF_RESCHED on the idle task of the other CPU
1225 * lockless. The worst case is that the other CPU runs the
1226 * idle task through an additional NOOP schedule()
1227 */
1228 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1229
1230 /* NEED_RESCHED must be visible before we test polling */
1231 smp_mb();
1232 if (!tsk_is_polling(rq->idle))
1233 smp_send_reschedule(cpu);
1234}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001235#endif /* CONFIG_NO_HZ */
Thomas Gleixner06d83082008-03-22 09:20:24 +01001236
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001237#else /* !CONFIG_SMP */
Peter Zijlstra31656512008-07-18 18:01:23 +02001238static void resched_task(struct task_struct *p)
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001239{
1240 assert_spin_locked(&task_rq(p)->lock);
Peter Zijlstra31656512008-07-18 18:01:23 +02001241 set_tsk_need_resched(p);
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001242}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02001243#endif /* CONFIG_SMP */
Ingo Molnarc24d20d2007-07-09 18:51:59 +02001244
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001245#if BITS_PER_LONG == 32
1246# define WMULT_CONST (~0UL)
1247#else
1248# define WMULT_CONST (1UL << 32)
1249#endif
1250
1251#define WMULT_SHIFT 32
1252
Ingo Molnar194081e2007-08-09 11:16:51 +02001253/*
1254 * Shift right and round:
1255 */
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001256#define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
Ingo Molnar194081e2007-08-09 11:16:51 +02001257
Peter Zijlstraa7be37a2008-06-27 13:41:11 +02001258/*
1259 * delta *= weight / lw
1260 */
Ingo Molnarcb1c4fc2007-08-02 17:41:40 +02001261static unsigned long
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001262calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1263 struct load_weight *lw)
1264{
1265 u64 tmp;
1266
Lai Jiangshan7a232e02008-06-12 16:43:07 +08001267 if (!lw->inv_weight) {
1268 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1269 lw->inv_weight = 1;
1270 else
1271 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1272 / (lw->weight+1);
1273 }
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001274
1275 tmp = (u64)delta_exec * weight;
1276 /*
1277 * Check whether we'd overflow the 64-bit multiplication:
1278 */
Ingo Molnar194081e2007-08-09 11:16:51 +02001279 if (unlikely(tmp > WMULT_CONST))
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001280 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
Ingo Molnar194081e2007-08-09 11:16:51 +02001281 WMULT_SHIFT/2);
1282 else
Ingo Molnarcf2ab462007-09-05 14:32:49 +02001283 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001284
Ingo Molnarecf691d2007-08-02 17:41:40 +02001285 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001286}
1287
Ingo Molnar10919852007-10-15 17:00:04 +02001288static inline void update_load_add(struct load_weight *lw, unsigned long inc)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001289{
1290 lw->weight += inc;
Ingo Molnare89996a2008-03-14 23:48:28 +01001291 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001292}
1293
Ingo Molnar10919852007-10-15 17:00:04 +02001294static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001295{
1296 lw->weight -= dec;
Ingo Molnare89996a2008-03-14 23:48:28 +01001297 lw->inv_weight = 0;
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001298}
1299
Linus Torvalds1da177e2005-04-16 15:20:36 -07001300/*
Peter Williams2dd73a42006-06-27 02:54:34 -07001301 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1302 * of tasks with abnormal "nice" values across CPUs the contribution that
1303 * each task makes to its run queue's load is weighted according to its
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01001304 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
Peter Williams2dd73a42006-06-27 02:54:34 -07001305 * scaled version of the new time slice allocation that they receive on time
1306 * slice expiry etc.
1307 */
1308
Ingo Molnardd41f592007-07-09 18:51:59 +02001309#define WEIGHT_IDLEPRIO 2
1310#define WMULT_IDLEPRIO (1 << 31)
1311
1312/*
1313 * Nice levels are multiplicative, with a gentle 10% change for every
1314 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1315 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1316 * that remained on nice 0.
1317 *
1318 * The "10% effect" is relative and cumulative: from _any_ nice level,
1319 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
Ingo Molnarf9153ee2007-07-16 09:46:30 +02001320 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1321 * If a task goes up by ~10% and another task goes down by ~10% then
1322 * the relative distance between them is ~25%.)
Ingo Molnardd41f592007-07-09 18:51:59 +02001323 */
1324static const int prio_to_weight[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001325 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1326 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1327 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1328 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1329 /* 0 */ 1024, 820, 655, 526, 423,
1330 /* 5 */ 335, 272, 215, 172, 137,
1331 /* 10 */ 110, 87, 70, 56, 45,
1332 /* 15 */ 36, 29, 23, 18, 15,
Ingo Molnardd41f592007-07-09 18:51:59 +02001333};
1334
Ingo Molnar5714d2d2007-07-16 09:46:31 +02001335/*
1336 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1337 *
1338 * In cases where the weight does not change often, we can use the
1339 * precalculated inverse to speed up arithmetics by turning divisions
1340 * into multiplications:
1341 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001342static const u32 prio_to_wmult[40] = {
Ingo Molnar254753d2007-08-09 11:16:51 +02001343 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1344 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1345 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1346 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1347 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1348 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1349 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1350 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
Ingo Molnardd41f592007-07-09 18:51:59 +02001351};
Peter Williams2dd73a42006-06-27 02:54:34 -07001352
Ingo Molnardd41f592007-07-09 18:51:59 +02001353static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1354
1355/*
1356 * runqueue iterator, to support SMP load-balancing between different
1357 * scheduling classes, without having to expose their internal data
1358 * structures to the load-balancing proper:
1359 */
1360struct rq_iterator {
1361 void *arg;
1362 struct task_struct *(*start)(void *);
1363 struct task_struct *(*next)(void *);
1364};
1365
Peter Williamse1d14842007-10-24 18:23:51 +02001366#ifdef CONFIG_SMP
1367static unsigned long
1368balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1369 unsigned long max_load_move, struct sched_domain *sd,
1370 enum cpu_idle_type idle, int *all_pinned,
1371 int *this_best_prio, struct rq_iterator *iterator);
1372
1373static int
1374iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1375 struct sched_domain *sd, enum cpu_idle_type idle,
1376 struct rq_iterator *iterator);
Peter Williamse1d14842007-10-24 18:23:51 +02001377#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02001378
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01001379#ifdef CONFIG_CGROUP_CPUACCT
1380static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1381#else
1382static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1383#endif
1384
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001385static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1386{
1387 update_load_add(&rq->load, load);
1388}
1389
1390static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1391{
1392 update_load_sub(&rq->load, load);
1393}
1394
Ingo Molnar7940ca32008-08-19 13:40:47 +02001395#if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
Peter Zijlstraeb755802008-08-19 12:33:05 +02001396typedef int (*tg_visitor)(struct task_group *, void *);
1397
1398/*
1399 * Iterate the full tree, calling @down when first entering a node and @up when
1400 * leaving it for the final time.
1401 */
1402static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1403{
1404 struct task_group *parent, *child;
1405 int ret;
1406
1407 rcu_read_lock();
1408 parent = &root_task_group;
1409down:
1410 ret = (*down)(parent, data);
1411 if (ret)
1412 goto out_unlock;
1413 list_for_each_entry_rcu(child, &parent->children, siblings) {
1414 parent = child;
1415 goto down;
1416
1417up:
1418 continue;
1419 }
1420 ret = (*up)(parent, data);
1421 if (ret)
1422 goto out_unlock;
1423
1424 child = parent;
1425 parent = parent->parent;
1426 if (parent)
1427 goto up;
1428out_unlock:
1429 rcu_read_unlock();
1430
1431 return ret;
1432}
1433
1434static int tg_nop(struct task_group *tg, void *data)
1435{
1436 return 0;
1437}
1438#endif
1439
Gregory Haskinse7693a32008-01-25 21:08:09 +01001440#ifdef CONFIG_SMP
1441static unsigned long source_load(int cpu, int type);
1442static unsigned long target_load(int cpu, int type);
Gregory Haskinse7693a32008-01-25 21:08:09 +01001443static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001444
Peter Zijlstraa8a51d52008-06-27 13:41:26 +02001445static unsigned long cpu_avg_load_per_task(int cpu)
1446{
1447 struct rq *rq = cpu_rq(cpu);
1448
1449 if (rq->nr_running)
1450 rq->avg_load_per_task = rq->load.weight / rq->nr_running;
1451
1452 return rq->avg_load_per_task;
1453}
1454
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001455#ifdef CONFIG_FAIR_GROUP_SCHED
1456
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001457static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1458
1459/*
1460 * Calculate and set the cpu's group shares.
1461 */
1462static void
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001463update_group_shares_cpu(struct task_group *tg, int cpu,
1464 unsigned long sd_shares, unsigned long sd_rq_weight)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001465{
1466 int boost = 0;
1467 unsigned long shares;
1468 unsigned long rq_weight;
1469
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001470 if (!tg->se[cpu])
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001471 return;
1472
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001473 rq_weight = tg->cfs_rq[cpu]->load.weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001474
1475 /*
1476 * If there are currently no tasks on the cpu pretend there is one of
1477 * average load so that when a new task gets to run here it will not
1478 * get delayed by group starvation.
1479 */
1480 if (!rq_weight) {
1481 boost = 1;
1482 rq_weight = NICE_0_LOAD;
1483 }
1484
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001485 if (unlikely(rq_weight > sd_rq_weight))
1486 rq_weight = sd_rq_weight;
1487
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001488 /*
1489 * \Sum shares * rq_weight
1490 * shares = -----------------------
1491 * \Sum rq_weight
1492 *
1493 */
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001494 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001495 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001496
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001497 if (abs(shares - tg->se[cpu]->load.weight) >
1498 sysctl_sched_shares_thresh) {
1499 struct rq *rq = cpu_rq(cpu);
1500 unsigned long flags;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001501
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001502 spin_lock_irqsave(&rq->lock, flags);
1503 /*
1504 * record the actual number of shares, not the boosted amount.
1505 */
1506 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1507 tg->cfs_rq[cpu]->rq_weight = rq_weight;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001508
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001509 __set_se_shares(tg->se[cpu], shares);
1510 spin_unlock_irqrestore(&rq->lock, flags);
1511 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001512}
1513
1514/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001515 * Re-compute the task group their per cpu shares over the given domain.
1516 * This needs to be done in a bottom-up fashion because the rq weight of a
1517 * parent group depends on the shares of its child groups.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001518 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001519static int tg_shares_up(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001520{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001521 unsigned long rq_weight = 0;
1522 unsigned long shares = 0;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001523 struct sched_domain *sd = data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001524 int i;
1525
1526 for_each_cpu_mask(i, sd->span) {
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001527 rq_weight += tg->cfs_rq[i]->load.weight;
1528 shares += tg->cfs_rq[i]->shares;
1529 }
1530
1531 if ((!shares && rq_weight) || shares > tg->shares)
1532 shares = tg->shares;
1533
1534 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1535 shares = tg->shares;
1536
Peter Zijlstracd809172008-06-27 13:41:34 +02001537 if (!rq_weight)
1538 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1539
Peter Zijlstraffda12a2008-10-17 19:27:02 +02001540 for_each_cpu_mask(i, sd->span)
1541 update_group_shares_cpu(tg, i, shares, rq_weight);
Peter Zijlstraeb755802008-08-19 12:33:05 +02001542
1543 return 0;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001544}
1545
1546/*
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001547 * Compute the cpu's hierarchical load factor for each task group.
1548 * This needs to be done in a top-down fashion because the load of a child
1549 * group is a fraction of its parents load.
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001550 */
Peter Zijlstraeb755802008-08-19 12:33:05 +02001551static int tg_load_down(struct task_group *tg, void *data)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001552{
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001553 unsigned long load;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001554 long cpu = (long)data;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001555
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001556 if (!tg->parent) {
1557 load = cpu_rq(cpu)->load.weight;
1558 } else {
1559 load = tg->parent->cfs_rq[cpu]->h_load;
1560 load *= tg->cfs_rq[cpu]->shares;
1561 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1562 }
1563
1564 tg->cfs_rq[cpu]->h_load = load;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001565
Peter Zijlstraeb755802008-08-19 12:33:05 +02001566 return 0;
Peter Zijlstra4d8d5952008-06-27 13:41:19 +02001567}
1568
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001569static void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001570{
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001571 u64 now = cpu_clock(raw_smp_processor_id());
1572 s64 elapsed = now - sd->last_update;
1573
1574 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1575 sd->last_update = now;
Peter Zijlstraeb755802008-08-19 12:33:05 +02001576 walk_tg_tree(tg_nop, tg_shares_up, sd);
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02001577 }
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001578}
1579
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001580static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1581{
1582 spin_unlock(&rq->lock);
1583 update_shares(sd);
1584 spin_lock(&rq->lock);
1585}
1586
Peter Zijlstraeb755802008-08-19 12:33:05 +02001587static void update_h_load(long cpu)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001588{
Peter Zijlstraeb755802008-08-19 12:33:05 +02001589 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001590}
1591
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001592#else
1593
Peter Zijlstrac8cba852008-06-27 13:41:23 +02001594static inline void update_shares(struct sched_domain *sd)
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001595{
1596}
1597
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02001598static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1599{
1600}
1601
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001602#endif
1603
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001604#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001605
1606#ifdef CONFIG_FAIR_GROUP_SCHED
1607static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1608{
Vegard Nossum30432092008-06-27 21:35:50 +02001609#ifdef CONFIG_SMP
Ingo Molnar34e83e82008-06-27 15:42:36 +02001610 cfs_rq->shares = shares;
1611#endif
Peter Zijlstra18d95a22008-04-19 19:45:00 +02001612}
1613#endif
1614
Ingo Molnardd41f592007-07-09 18:51:59 +02001615#include "sched_stats.h"
Ingo Molnardd41f592007-07-09 18:51:59 +02001616#include "sched_idletask.c"
Ingo Molnar5522d5d2007-10-15 17:00:12 +02001617#include "sched_fair.c"
1618#include "sched_rt.c"
Ingo Molnardd41f592007-07-09 18:51:59 +02001619#ifdef CONFIG_SCHED_DEBUG
1620# include "sched_debug.c"
1621#endif
1622
1623#define sched_class_highest (&rt_sched_class)
Gregory Haskins1f11eb62008-06-04 15:04:05 -04001624#define for_each_class(class) \
1625 for (class = sched_class_highest; class; class = class->next)
Ingo Molnardd41f592007-07-09 18:51:59 +02001626
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001627static void inc_nr_running(struct rq *rq)
Ingo Molnar6363ca52008-05-29 11:28:57 +02001628{
1629 rq->nr_running++;
Ingo Molnar6363ca52008-05-29 11:28:57 +02001630}
1631
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001632static void dec_nr_running(struct rq *rq)
Ingo Molnar9c217242007-08-02 17:41:40 +02001633{
1634 rq->nr_running--;
Ingo Molnar9c217242007-08-02 17:41:40 +02001635}
1636
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001637static void set_load_weight(struct task_struct *p)
1638{
1639 if (task_has_rt_policy(p)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02001640 p->se.load.weight = prio_to_weight[0] * 2;
1641 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1642 return;
1643 }
1644
1645 /*
1646 * SCHED_IDLE tasks get minimal weight:
1647 */
1648 if (p->policy == SCHED_IDLE) {
1649 p->se.load.weight = WEIGHT_IDLEPRIO;
1650 p->se.load.inv_weight = WMULT_IDLEPRIO;
1651 return;
1652 }
1653
1654 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1655 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
Ingo Molnar45bf76d2007-07-09 18:51:59 +02001656}
1657
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001658static void update_avg(u64 *avg, u64 sample)
1659{
1660 s64 diff = sample - *avg;
1661 *avg += diff >> 3;
1662}
1663
Ingo Molnar8159f872007-08-09 11:16:49 +02001664static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001665{
1666 sched_info_queued(p);
Ingo Molnarfd390f62007-08-09 11:16:48 +02001667 p->sched_class->enqueue_task(rq, p, wakeup);
Ingo Molnardd41f592007-07-09 18:51:59 +02001668 p->se.on_rq = 1;
1669}
1670
Ingo Molnar69be72c2007-08-09 11:16:49 +02001671static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
Ingo Molnardd41f592007-07-09 18:51:59 +02001672{
Gregory Haskins2087a1a2008-06-27 14:30:00 -06001673 if (sleep && p->se.last_wakeup) {
1674 update_avg(&p->se.avg_overlap,
1675 p->se.sum_exec_runtime - p->se.last_wakeup);
1676 p->se.last_wakeup = 0;
1677 }
1678
Ankita Garg46ac22b2008-07-01 14:30:06 +05301679 sched_info_dequeued(p);
Ingo Molnarf02231e2007-08-09 11:16:48 +02001680 p->sched_class->dequeue_task(rq, p, sleep);
Ingo Molnardd41f592007-07-09 18:51:59 +02001681 p->se.on_rq = 0;
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001682}
1683
1684/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001685 * __normal_prio - return the priority that is based on the static prio
Ingo Molnar71f8bd42007-07-09 18:51:59 +02001686 */
Ingo Molnar14531182007-07-09 18:51:59 +02001687static inline int __normal_prio(struct task_struct *p)
1688{
Ingo Molnardd41f592007-07-09 18:51:59 +02001689 return p->static_prio;
Ingo Molnar14531182007-07-09 18:51:59 +02001690}
1691
1692/*
Ingo Molnarb29739f2006-06-27 02:54:51 -07001693 * Calculate the expected normal priority: i.e. priority
1694 * without taking RT-inheritance into account. Might be
1695 * boosted by interactivity modifiers. Changes upon fork,
1696 * setprio syscalls, and whenever the interactivity
1697 * estimator recalculates.
1698 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001699static inline int normal_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001700{
1701 int prio;
1702
Ingo Molnare05606d2007-07-09 18:51:59 +02001703 if (task_has_rt_policy(p))
Ingo Molnarb29739f2006-06-27 02:54:51 -07001704 prio = MAX_RT_PRIO-1 - p->rt_priority;
1705 else
1706 prio = __normal_prio(p);
1707 return prio;
1708}
1709
1710/*
1711 * Calculate the current priority, i.e. the priority
1712 * taken into account by the scheduler. This value might
1713 * be boosted by RT tasks, or might be boosted by
1714 * interactivity modifiers. Will be RT if the task got
1715 * RT-boosted. If not then it returns p->normal_prio.
1716 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001717static int effective_prio(struct task_struct *p)
Ingo Molnarb29739f2006-06-27 02:54:51 -07001718{
1719 p->normal_prio = normal_prio(p);
1720 /*
1721 * If we are RT tasks or we were boosted to RT priority,
1722 * keep the priority unchanged. Otherwise, update priority
1723 * to the normal priority:
1724 */
1725 if (!rt_prio(p->prio))
1726 return p->normal_prio;
1727 return p->prio;
1728}
1729
1730/*
Ingo Molnardd41f592007-07-09 18:51:59 +02001731 * activate_task - move a task to the runqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07001732 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001733static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001734{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001735 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001736 rq->nr_uninterruptible--;
1737
Ingo Molnar8159f872007-08-09 11:16:49 +02001738 enqueue_task(rq, p, wakeup);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001739 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001740}
1741
1742/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07001743 * deactivate_task - remove a task from the runqueue.
1744 */
Ingo Molnar2e1cb742007-08-09 11:16:49 +02001745static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001746{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05001747 if (task_contributes_to_load(p))
Ingo Molnardd41f592007-07-09 18:51:59 +02001748 rq->nr_uninterruptible++;
1749
Ingo Molnar69be72c2007-08-09 11:16:49 +02001750 dequeue_task(rq, p, sleep);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02001751 dec_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001752}
1753
Linus Torvalds1da177e2005-04-16 15:20:36 -07001754/**
1755 * task_curr - is this task currently executing on a CPU?
1756 * @p: the task in question.
1757 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001758inline int task_curr(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001759{
1760 return cpu_curr(task_cpu(p)) == p;
1761}
1762
Ingo Molnardd41f592007-07-09 18:51:59 +02001763static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1764{
Peter Zijlstra6f505b12008-01-25 21:08:30 +01001765 set_task_rq(p, cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02001766#ifdef CONFIG_SMP
Dmitry Adamushkoce96b5a2007-11-15 20:57:40 +01001767 /*
1768 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1769 * successfuly executed on another CPU. We must ensure that updates of
1770 * per-task data have been completed by this moment.
1771 */
1772 smp_wmb();
Ingo Molnardd41f592007-07-09 18:51:59 +02001773 task_thread_info(p)->cpu = cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02001774#endif
Peter Williams2dd73a42006-06-27 02:54:34 -07001775}
1776
Steven Rostedtcb469842008-01-25 21:08:22 +01001777static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1778 const struct sched_class *prev_class,
1779 int oldprio, int running)
1780{
1781 if (prev_class != p->sched_class) {
1782 if (prev_class->switched_from)
1783 prev_class->switched_from(rq, p, running);
1784 p->sched_class->switched_to(rq, p, running);
1785 } else
1786 p->sched_class->prio_changed(rq, p, oldprio, running);
1787}
1788
Linus Torvalds1da177e2005-04-16 15:20:36 -07001789#ifdef CONFIG_SMP
Ingo Molnarc65cc872007-07-09 18:51:58 +02001790
Thomas Gleixnere958b362008-06-04 23:22:32 +02001791/* Used instead of source_load when we know the type == 0 */
1792static unsigned long weighted_cpuload(const int cpu)
1793{
1794 return cpu_rq(cpu)->load.weight;
1795}
1796
Ingo Molnarcc367732007-10-15 17:00:18 +02001797/*
1798 * Is this task likely cache-hot:
1799 */
Gregory Haskinse7693a32008-01-25 21:08:09 +01001800static int
Ingo Molnarcc367732007-10-15 17:00:18 +02001801task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1802{
1803 s64 delta;
1804
Ingo Molnarf540a602008-03-15 17:10:34 +01001805 /*
1806 * Buddy candidates are cache hot:
1807 */
Ingo Molnard25ce4c2008-03-17 09:36:53 +01001808 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
Ingo Molnarf540a602008-03-15 17:10:34 +01001809 return 1;
1810
Ingo Molnarcc367732007-10-15 17:00:18 +02001811 if (p->sched_class != &fair_sched_class)
1812 return 0;
1813
Ingo Molnar6bc16652007-10-15 17:00:18 +02001814 if (sysctl_sched_migration_cost == -1)
1815 return 1;
1816 if (sysctl_sched_migration_cost == 0)
1817 return 0;
1818
Ingo Molnarcc367732007-10-15 17:00:18 +02001819 delta = now - p->se.exec_start;
1820
1821 return delta < (s64)sysctl_sched_migration_cost;
1822}
1823
1824
Ingo Molnardd41f592007-07-09 18:51:59 +02001825void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
Ingo Molnarc65cc872007-07-09 18:51:58 +02001826{
Ingo Molnardd41f592007-07-09 18:51:59 +02001827 int old_cpu = task_cpu(p);
1828 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001829 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1830 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
Ingo Molnarbbdba7c2007-10-15 17:00:06 +02001831 u64 clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001832
1833 clock_offset = old_rq->clock - new_rq->clock;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001834
1835#ifdef CONFIG_SCHEDSTATS
1836 if (p->se.wait_start)
1837 p->se.wait_start -= clock_offset;
Ingo Molnardd41f592007-07-09 18:51:59 +02001838 if (p->se.sleep_start)
1839 p->se.sleep_start -= clock_offset;
1840 if (p->se.block_start)
1841 p->se.block_start -= clock_offset;
Ingo Molnarcc367732007-10-15 17:00:18 +02001842 if (old_cpu != new_cpu) {
1843 schedstat_inc(p, se.nr_migrations);
1844 if (task_hot(p, old_rq->clock, NULL))
1845 schedstat_inc(p, se.nr_forced2_migrations);
1846 }
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02001847#endif
Srivatsa Vaddagiri2830cf82007-10-15 17:00:12 +02001848 p->se.vruntime -= old_cfsrq->min_vruntime -
1849 new_cfsrq->min_vruntime;
Ingo Molnardd41f592007-07-09 18:51:59 +02001850
1851 __set_task_cpu(p, new_cpu);
Ingo Molnarc65cc872007-07-09 18:51:58 +02001852}
1853
Ingo Molnar70b97a72006-07-03 00:25:42 -07001854struct migration_req {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001855 struct list_head list;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001856
Ingo Molnar36c8b582006-07-03 00:25:41 -07001857 struct task_struct *task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001858 int dest_cpu;
1859
Linus Torvalds1da177e2005-04-16 15:20:36 -07001860 struct completion done;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001861};
Linus Torvalds1da177e2005-04-16 15:20:36 -07001862
1863/*
1864 * The task's runqueue lock must be held.
1865 * Returns true if you have to wait for migration thread.
1866 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07001867static int
Ingo Molnar70b97a72006-07-03 00:25:42 -07001868migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001869{
Ingo Molnar70b97a72006-07-03 00:25:42 -07001870 struct rq *rq = task_rq(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001871
1872 /*
1873 * If the task is not on a runqueue (and not running), then
1874 * it is sufficient to simply update the task's cpu field.
1875 */
Ingo Molnardd41f592007-07-09 18:51:59 +02001876 if (!p->se.on_rq && !task_running(rq, p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07001877 set_task_cpu(p, dest_cpu);
1878 return 0;
1879 }
1880
1881 init_completion(&req->done);
Linus Torvalds1da177e2005-04-16 15:20:36 -07001882 req->task = p;
1883 req->dest_cpu = dest_cpu;
1884 list_add(&req->list, &rq->migration_queue);
Ingo Molnar48f24c42006-07-03 00:25:40 -07001885
Linus Torvalds1da177e2005-04-16 15:20:36 -07001886 return 1;
1887}
1888
1889/*
1890 * wait_task_inactive - wait for a thread to unschedule.
1891 *
Roland McGrath85ba2d82008-07-25 19:45:58 -07001892 * If @match_state is nonzero, it's the @p->state value just checked and
1893 * not expected to change. If it changes, i.e. @p might have woken up,
1894 * then return zero. When we succeed in waiting for @p to be off its CPU,
1895 * we return a positive number (its total switch count). If a second call
1896 * a short while later returns the same number, the caller can be sure that
1897 * @p has remained unscheduled the whole time.
1898 *
Linus Torvalds1da177e2005-04-16 15:20:36 -07001899 * The caller must ensure that the task *will* unschedule sometime soon,
1900 * else this function might spin for a *long* time. This function can't
1901 * be called with interrupts off, or it may introduce deadlock with
1902 * smp_call_function() if an IPI is sent by the same process we are
1903 * waiting to become inactive.
1904 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07001905unsigned long wait_task_inactive(struct task_struct *p, long match_state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07001906{
1907 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02001908 int running, on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07001909 unsigned long ncsw;
Ingo Molnar70b97a72006-07-03 00:25:42 -07001910 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001911
Andi Kleen3a5c3592007-10-15 17:00:14 +02001912 for (;;) {
1913 /*
1914 * We do the initial early heuristics without holding
1915 * any task-queue locks at all. We'll only try to get
1916 * the runqueue lock when things look like they will
1917 * work out!
1918 */
1919 rq = task_rq(p);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001920
Andi Kleen3a5c3592007-10-15 17:00:14 +02001921 /*
1922 * If the task is actively running on another CPU
1923 * still, just relax and busy-wait without holding
1924 * any locks.
1925 *
1926 * NOTE! Since we don't hold any locks, it's not
1927 * even sure that "rq" stays as the right runqueue!
1928 * But we don't care, since "task_running()" will
1929 * return false if the runqueue has changed and p
1930 * is actually now running somewhere else!
1931 */
Roland McGrath85ba2d82008-07-25 19:45:58 -07001932 while (task_running(rq, p)) {
1933 if (match_state && unlikely(p->state != match_state))
1934 return 0;
Andi Kleen3a5c3592007-10-15 17:00:14 +02001935 cpu_relax();
Roland McGrath85ba2d82008-07-25 19:45:58 -07001936 }
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001937
Andi Kleen3a5c3592007-10-15 17:00:14 +02001938 /*
1939 * Ok, time to look more closely! We need the rq
1940 * lock now, to be *sure*. If we're wrong, we'll
1941 * just go back and repeat.
1942 */
1943 rq = task_rq_lock(p, &flags);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04001944 trace_sched_wait_task(rq, p);
Andi Kleen3a5c3592007-10-15 17:00:14 +02001945 running = task_running(rq, p);
1946 on_rq = p->se.on_rq;
Roland McGrath85ba2d82008-07-25 19:45:58 -07001947 ncsw = 0;
Oleg Nesterovf31e11d2008-08-20 16:54:44 -07001948 if (!match_state || p->state == match_state)
Oleg Nesterov93dcf552008-08-20 16:54:44 -07001949 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
Andi Kleen3a5c3592007-10-15 17:00:14 +02001950 task_rq_unlock(rq, &flags);
Linus Torvaldsfa490cf2007-06-18 09:34:40 -07001951
Andi Kleen3a5c3592007-10-15 17:00:14 +02001952 /*
Roland McGrath85ba2d82008-07-25 19:45:58 -07001953 * If it changed from the expected state, bail out now.
1954 */
1955 if (unlikely(!ncsw))
1956 break;
1957
1958 /*
Andi Kleen3a5c3592007-10-15 17:00:14 +02001959 * Was it really running after all now that we
1960 * checked with the proper locks actually held?
1961 *
1962 * Oops. Go back and try again..
1963 */
1964 if (unlikely(running)) {
1965 cpu_relax();
1966 continue;
1967 }
1968
1969 /*
1970 * It's not enough that it's not actively running,
1971 * it must be off the runqueue _entirely_, and not
1972 * preempted!
1973 *
1974 * So if it wa still runnable (but just not actively
1975 * running right now), it's preempted, and we should
1976 * yield - it could be a while.
1977 */
1978 if (unlikely(on_rq)) {
1979 schedule_timeout_uninterruptible(1);
1980 continue;
1981 }
1982
1983 /*
1984 * Ahh, all good. It wasn't running, and it wasn't
1985 * runnable, which means that it will never become
1986 * running in the future either. We're all done!
1987 */
1988 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001989 }
Roland McGrath85ba2d82008-07-25 19:45:58 -07001990
1991 return ncsw;
Linus Torvalds1da177e2005-04-16 15:20:36 -07001992}
1993
1994/***
1995 * kick_process - kick a running thread to enter/exit the kernel
1996 * @p: the to-be-kicked thread
1997 *
1998 * Cause a process which is running on another CPU to enter
1999 * kernel-mode, without any delay. (to get signals handled.)
2000 *
2001 * NOTE: this function doesnt have to take the runqueue lock,
2002 * because all it wants to ensure is that the remote task enters
2003 * the kernel. If the IPI races and the task has been migrated
2004 * to another CPU then no harm is done and the purpose has been
2005 * achieved as well.
2006 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002007void kick_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002008{
2009 int cpu;
2010
2011 preempt_disable();
2012 cpu = task_cpu(p);
2013 if ((cpu != smp_processor_id()) && task_curr(p))
2014 smp_send_reschedule(cpu);
2015 preempt_enable();
2016}
2017
2018/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002019 * Return a low guess at the load of a migration-source cpu weighted
2020 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002021 *
2022 * We want to under-estimate the load of migration sources, to
2023 * balance conservatively.
2024 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002025static unsigned long source_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002026{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002027 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002028 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002029
Peter Zijlstra93b75212008-06-27 13:41:33 +02002030 if (type == 0 || !sched_feat(LB_BIAS))
Ingo Molnardd41f592007-07-09 18:51:59 +02002031 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002032
Ingo Molnardd41f592007-07-09 18:51:59 +02002033 return min(rq->cpu_load[type-1], total);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002034}
2035
2036/*
Peter Williams2dd73a42006-06-27 02:54:34 -07002037 * Return a high guess at the load of a migration-target cpu weighted
2038 * according to the scheduling class and "nice" value.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002039 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002040static unsigned long target_load(int cpu, int type)
Con Kolivasb9104722005-11-08 21:38:55 -08002041{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002042 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002043 unsigned long total = weighted_cpuload(cpu);
Nick Piggina2000572006-02-10 01:51:02 -08002044
Peter Zijlstra93b75212008-06-27 13:41:33 +02002045 if (type == 0 || !sched_feat(LB_BIAS))
Ingo Molnardd41f592007-07-09 18:51:59 +02002046 return total;
Peter Williams2dd73a42006-06-27 02:54:34 -07002047
Ingo Molnardd41f592007-07-09 18:51:59 +02002048 return max(rq->cpu_load[type-1], total);
Peter Williams2dd73a42006-06-27 02:54:34 -07002049}
2050
2051/*
Nick Piggin147cbb42005-06-25 14:57:19 -07002052 * find_idlest_group finds and returns the least busy CPU group within the
2053 * domain.
2054 */
2055static struct sched_group *
2056find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2057{
2058 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2059 unsigned long min_load = ULONG_MAX, this_load = 0;
2060 int load_idx = sd->forkexec_idx;
2061 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2062
2063 do {
2064 unsigned long load, avg_load;
2065 int local_group;
2066 int i;
2067
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002068 /* Skip over this group if it has no CPUs allowed */
2069 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
Andi Kleen3a5c3592007-10-15 17:00:14 +02002070 continue;
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002071
Nick Piggin147cbb42005-06-25 14:57:19 -07002072 local_group = cpu_isset(this_cpu, group->cpumask);
Nick Piggin147cbb42005-06-25 14:57:19 -07002073
2074 /* Tally up the load of all CPUs in the group */
2075 avg_load = 0;
2076
Mike Travis363ab6f2008-05-12 21:21:13 +02002077 for_each_cpu_mask_nr(i, group->cpumask) {
Nick Piggin147cbb42005-06-25 14:57:19 -07002078 /* Bias balancing toward cpus of our domain */
2079 if (local_group)
2080 load = source_load(i, load_idx);
2081 else
2082 load = target_load(i, load_idx);
2083
2084 avg_load += load;
2085 }
2086
2087 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07002088 avg_load = sg_div_cpu_power(group,
2089 avg_load * SCHED_LOAD_SCALE);
Nick Piggin147cbb42005-06-25 14:57:19 -07002090
2091 if (local_group) {
2092 this_load = avg_load;
2093 this = group;
2094 } else if (avg_load < min_load) {
2095 min_load = avg_load;
2096 idlest = group;
2097 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02002098 } while (group = group->next, group != sd->groups);
Nick Piggin147cbb42005-06-25 14:57:19 -07002099
2100 if (!idlest || 100*this_load < imbalance*min_load)
2101 return NULL;
2102 return idlest;
2103}
2104
2105/*
Satoru Takeuchi0feaece2006-10-03 01:14:10 -07002106 * find_idlest_cpu - find the idlest cpu among the cpus in group.
Nick Piggin147cbb42005-06-25 14:57:19 -07002107 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002108static int
Mike Travis7c16ec52008-04-04 18:11:11 -07002109find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2110 cpumask_t *tmp)
Nick Piggin147cbb42005-06-25 14:57:19 -07002111{
2112 unsigned long load, min_load = ULONG_MAX;
2113 int idlest = -1;
2114 int i;
2115
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002116 /* Traverse only the allowed CPUs */
Mike Travis7c16ec52008-04-04 18:11:11 -07002117 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
M.Baris Demirayda5a5522005-09-10 00:26:09 -07002118
Mike Travis363ab6f2008-05-12 21:21:13 +02002119 for_each_cpu_mask_nr(i, *tmp) {
Peter Williams2dd73a42006-06-27 02:54:34 -07002120 load = weighted_cpuload(i);
Nick Piggin147cbb42005-06-25 14:57:19 -07002121
2122 if (load < min_load || (load == min_load && i == this_cpu)) {
2123 min_load = load;
2124 idlest = i;
2125 }
2126 }
2127
2128 return idlest;
2129}
2130
Nick Piggin476d1392005-06-25 14:57:29 -07002131/*
2132 * sched_balance_self: balance the current task (running on cpu) in domains
2133 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2134 * SD_BALANCE_EXEC.
2135 *
2136 * Balance, ie. select the least loaded group.
2137 *
2138 * Returns the target CPU number, or the same CPU if no balancing is needed.
2139 *
2140 * preempt must be disabled.
2141 */
2142static int sched_balance_self(int cpu, int flag)
2143{
2144 struct task_struct *t = current;
2145 struct sched_domain *tmp, *sd = NULL;
Nick Piggin147cbb42005-06-25 14:57:19 -07002146
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002147 for_each_domain(cpu, tmp) {
Ingo Molnar9761eea2007-07-09 18:52:00 +02002148 /*
2149 * If power savings logic is enabled for a domain, stop there.
2150 */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07002151 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2152 break;
Nick Piggin476d1392005-06-25 14:57:29 -07002153 if (tmp->flags & flag)
2154 sd = tmp;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002155 }
Nick Piggin476d1392005-06-25 14:57:29 -07002156
Peter Zijlstra039a1c42008-06-27 13:41:25 +02002157 if (sd)
2158 update_shares(sd);
2159
Nick Piggin476d1392005-06-25 14:57:29 -07002160 while (sd) {
Mike Travis7c16ec52008-04-04 18:11:11 -07002161 cpumask_t span, tmpmask;
Nick Piggin476d1392005-06-25 14:57:29 -07002162 struct sched_group *group;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002163 int new_cpu, weight;
2164
2165 if (!(sd->flags & flag)) {
2166 sd = sd->child;
2167 continue;
2168 }
Nick Piggin476d1392005-06-25 14:57:29 -07002169
2170 span = sd->span;
2171 group = find_idlest_group(sd, t, cpu);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002172 if (!group) {
2173 sd = sd->child;
2174 continue;
2175 }
Nick Piggin476d1392005-06-25 14:57:29 -07002176
Mike Travis7c16ec52008-04-04 18:11:11 -07002177 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002178 if (new_cpu == -1 || new_cpu == cpu) {
2179 /* Now try balancing at a lower domain level of cpu */
2180 sd = sd->child;
2181 continue;
2182 }
Nick Piggin476d1392005-06-25 14:57:29 -07002183
Siddha, Suresh B1a848872006-10-03 01:14:08 -07002184 /* Now try balancing at a lower domain level of new_cpu */
Nick Piggin476d1392005-06-25 14:57:29 -07002185 cpu = new_cpu;
Nick Piggin476d1392005-06-25 14:57:29 -07002186 sd = NULL;
2187 weight = cpus_weight(span);
2188 for_each_domain(cpu, tmp) {
2189 if (weight <= cpus_weight(tmp->span))
2190 break;
2191 if (tmp->flags & flag)
2192 sd = tmp;
2193 }
2194 /* while loop will break here if sd == NULL */
2195 }
2196
2197 return cpu;
2198}
2199
2200#endif /* CONFIG_SMP */
Linus Torvalds1da177e2005-04-16 15:20:36 -07002201
Linus Torvalds1da177e2005-04-16 15:20:36 -07002202/***
2203 * try_to_wake_up - wake up a thread
2204 * @p: the to-be-woken-up thread
2205 * @state: the mask of task states that can be woken
2206 * @sync: do a synchronous wakeup?
2207 *
2208 * Put it on the run-queue if it's not already there. The "current"
2209 * thread is always on the run-queue (except when the actual
2210 * re-schedule is in progress), and as such you're allowed to do
2211 * the simpler "current->state = TASK_RUNNING" to mark yourself
2212 * runnable without the overhead of this.
2213 *
2214 * returns failure only if the task is already active.
2215 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002216static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002217{
Ingo Molnarcc367732007-10-15 17:00:18 +02002218 int cpu, orig_cpu, this_cpu, success = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002219 unsigned long flags;
2220 long old_state;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002221 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002222
Ingo Molnarb85d0662008-03-16 20:03:22 +01002223 if (!sched_feat(SYNC_WAKEUPS))
2224 sync = 0;
2225
Peter Zijlstra2398f2c2008-06-27 13:41:35 +02002226#ifdef CONFIG_SMP
2227 if (sched_feat(LB_WAKEUP_UPDATE)) {
2228 struct sched_domain *sd;
2229
2230 this_cpu = raw_smp_processor_id();
2231 cpu = task_cpu(p);
2232
2233 for_each_domain(this_cpu, sd) {
2234 if (cpu_isset(cpu, sd->span)) {
2235 update_shares(sd);
2236 break;
2237 }
2238 }
2239 }
2240#endif
2241
Linus Torvalds04e2f172008-02-23 18:05:03 -08002242 smp_wmb();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002243 rq = task_rq_lock(p, &flags);
2244 old_state = p->state;
2245 if (!(old_state & state))
2246 goto out;
2247
Ingo Molnardd41f592007-07-09 18:51:59 +02002248 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002249 goto out_running;
2250
2251 cpu = task_cpu(p);
Ingo Molnarcc367732007-10-15 17:00:18 +02002252 orig_cpu = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002253 this_cpu = smp_processor_id();
2254
2255#ifdef CONFIG_SMP
2256 if (unlikely(task_running(rq, p)))
2257 goto out_activate;
2258
Dmitry Adamushko5d2f5a62008-01-25 21:08:21 +01002259 cpu = p->sched_class->select_task_rq(p, sync);
2260 if (cpu != orig_cpu) {
2261 set_task_cpu(p, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002262 task_rq_unlock(rq, &flags);
2263 /* might preempt at this point */
2264 rq = task_rq_lock(p, &flags);
2265 old_state = p->state;
2266 if (!(old_state & state))
2267 goto out;
Ingo Molnardd41f592007-07-09 18:51:59 +02002268 if (p->se.on_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002269 goto out_running;
2270
2271 this_cpu = smp_processor_id();
2272 cpu = task_cpu(p);
2273 }
2274
Gregory Haskinse7693a32008-01-25 21:08:09 +01002275#ifdef CONFIG_SCHEDSTATS
2276 schedstat_inc(rq, ttwu_count);
2277 if (cpu == this_cpu)
2278 schedstat_inc(rq, ttwu_local);
2279 else {
2280 struct sched_domain *sd;
2281 for_each_domain(this_cpu, sd) {
2282 if (cpu_isset(cpu, sd->span)) {
2283 schedstat_inc(sd, ttwu_wake_remote);
2284 break;
2285 }
2286 }
2287 }
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002288#endif /* CONFIG_SCHEDSTATS */
Gregory Haskinse7693a32008-01-25 21:08:09 +01002289
Linus Torvalds1da177e2005-04-16 15:20:36 -07002290out_activate:
2291#endif /* CONFIG_SMP */
Ingo Molnarcc367732007-10-15 17:00:18 +02002292 schedstat_inc(p, se.nr_wakeups);
2293 if (sync)
2294 schedstat_inc(p, se.nr_wakeups_sync);
2295 if (orig_cpu != cpu)
2296 schedstat_inc(p, se.nr_wakeups_migrate);
2297 if (cpu == this_cpu)
2298 schedstat_inc(p, se.nr_wakeups_local);
2299 else
2300 schedstat_inc(p, se.nr_wakeups_remote);
Ingo Molnar2daa3572007-08-09 11:16:51 +02002301 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02002302 activate_task(rq, p, 1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002303 success = 1;
2304
2305out_running:
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002306 trace_sched_wakeup(rq, p);
Peter Zijlstra15afe092008-09-20 23:38:02 +02002307 check_preempt_curr(rq, p, sync);
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002308
Linus Torvalds1da177e2005-04-16 15:20:36 -07002309 p->state = TASK_RUNNING;
Steven Rostedt9a897c52008-01-25 21:08:22 +01002310#ifdef CONFIG_SMP
2311 if (p->sched_class->task_wake_up)
2312 p->sched_class->task_wake_up(rq, p);
2313#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002314out:
Gregory Haskins2087a1a2008-06-27 14:30:00 -06002315 current->se.last_wakeup = current->se.sum_exec_runtime;
2316
Linus Torvalds1da177e2005-04-16 15:20:36 -07002317 task_rq_unlock(rq, &flags);
2318
2319 return success;
2320}
2321
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002322int wake_up_process(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002323{
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05002324 return try_to_wake_up(p, TASK_ALL, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002325}
Linus Torvalds1da177e2005-04-16 15:20:36 -07002326EXPORT_SYMBOL(wake_up_process);
2327
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002328int wake_up_state(struct task_struct *p, unsigned int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002329{
2330 return try_to_wake_up(p, state, 0);
2331}
2332
Linus Torvalds1da177e2005-04-16 15:20:36 -07002333/*
2334 * Perform scheduler related setup for a newly forked process p.
2335 * p is forked by current.
Ingo Molnardd41f592007-07-09 18:51:59 +02002336 *
2337 * __sched_fork() is basic setup used by init_idle() too:
Linus Torvalds1da177e2005-04-16 15:20:36 -07002338 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002339static void __sched_fork(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002340{
Ingo Molnardd41f592007-07-09 18:51:59 +02002341 p->se.exec_start = 0;
2342 p->se.sum_exec_runtime = 0;
Ingo Molnarf6cf8912007-08-28 12:53:24 +02002343 p->se.prev_sum_exec_runtime = 0;
Ingo Molnar4ae7d5c2008-03-19 01:42:00 +01002344 p->se.last_wakeup = 0;
2345 p->se.avg_overlap = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002346
2347#ifdef CONFIG_SCHEDSTATS
2348 p->se.wait_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002349 p->se.sum_sleep_runtime = 0;
2350 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002351 p->se.block_start = 0;
2352 p->se.sleep_max = 0;
2353 p->se.block_max = 0;
2354 p->se.exec_max = 0;
Ingo Molnareba1ed42007-10-15 17:00:02 +02002355 p->se.slice_max = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002356 p->se.wait_max = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02002357#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002358
Peter Zijlstrafa717062008-01-25 21:08:27 +01002359 INIT_LIST_HEAD(&p->rt.run_list);
Ingo Molnardd41f592007-07-09 18:51:59 +02002360 p->se.on_rq = 0;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02002361 INIT_LIST_HEAD(&p->se.group_node);
Nick Piggin476d1392005-06-25 14:57:29 -07002362
Avi Kivitye107be32007-07-26 13:40:43 +02002363#ifdef CONFIG_PREEMPT_NOTIFIERS
2364 INIT_HLIST_HEAD(&p->preempt_notifiers);
2365#endif
2366
Linus Torvalds1da177e2005-04-16 15:20:36 -07002367 /*
2368 * We mark the process as running here, but have not actually
2369 * inserted it onto the runqueue yet. This guarantees that
2370 * nobody will actually run it, and a signal or other external
2371 * event cannot wake it up and insert it on the runqueue either.
2372 */
2373 p->state = TASK_RUNNING;
Ingo Molnardd41f592007-07-09 18:51:59 +02002374}
2375
2376/*
2377 * fork()/clone()-time setup:
2378 */
2379void sched_fork(struct task_struct *p, int clone_flags)
2380{
2381 int cpu = get_cpu();
2382
2383 __sched_fork(p);
2384
2385#ifdef CONFIG_SMP
2386 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2387#endif
Ingo Molnar02e4bac2007-10-15 17:00:11 +02002388 set_task_cpu(p, cpu);
Ingo Molnarb29739f2006-06-27 02:54:51 -07002389
2390 /*
2391 * Make sure we do not leak PI boosting priority to the child:
2392 */
2393 p->prio = current->normal_prio;
Hiroshi Shimamoto2ddbf952007-10-15 17:00:11 +02002394 if (!rt_prio(p->prio))
2395 p->sched_class = &fair_sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07002396
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002397#if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
Ingo Molnardd41f592007-07-09 18:51:59 +02002398 if (likely(sched_info_on()))
Chandra Seetharaman52f17b62006-07-14 00:24:38 -07002399 memset(&p->sched_info, 0, sizeof(p->sched_info));
Linus Torvalds1da177e2005-04-16 15:20:36 -07002400#endif
Chen, Kenneth Wd6077cb2006-02-14 13:53:10 -08002401#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
Nick Piggin4866cde2005-06-25 14:57:23 -07002402 p->oncpu = 0;
2403#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002404#ifdef CONFIG_PREEMPT
Nick Piggin4866cde2005-06-25 14:57:23 -07002405 /* Want to start with kernel preemption disabled. */
Al Viroa1261f52005-11-13 16:06:55 -08002406 task_thread_info(p)->preempt_count = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002407#endif
Nick Piggin476d1392005-06-25 14:57:29 -07002408 put_cpu();
Linus Torvalds1da177e2005-04-16 15:20:36 -07002409}
2410
2411/*
2412 * wake_up_new_task - wake up a newly created task for the first time.
2413 *
2414 * This function will do some initial scheduler statistics housekeeping
2415 * that must be done for every newly created context, then puts the task
2416 * on the runqueue and wakes it.
2417 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08002418void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002419{
2420 unsigned long flags;
Ingo Molnardd41f592007-07-09 18:51:59 +02002421 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002422
2423 rq = task_rq_lock(p, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002424 BUG_ON(p->state != TASK_RUNNING);
Ingo Molnara8e504d2007-08-09 11:16:47 +02002425 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002426
2427 p->prio = effective_prio(p);
2428
Srivatsa Vaddagirib9dca1e2007-10-17 16:55:11 +02002429 if (!p->sched_class->task_new || !current->se.on_rq) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002430 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002431 } else {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002432 /*
Ingo Molnardd41f592007-07-09 18:51:59 +02002433 * Let the scheduling class do new task startup
2434 * management (if any):
Linus Torvalds1da177e2005-04-16 15:20:36 -07002435 */
Ingo Molnaree0827d2007-08-09 11:16:49 +02002436 p->sched_class->task_new(rq, p);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02002437 inc_nr_running(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002438 }
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002439 trace_sched_wakeup_new(rq, p);
Peter Zijlstra15afe092008-09-20 23:38:02 +02002440 check_preempt_curr(rq, p, 0);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002441#ifdef CONFIG_SMP
2442 if (p->sched_class->task_wake_up)
2443 p->sched_class->task_wake_up(rq, p);
2444#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02002445 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002446}
2447
Avi Kivitye107be32007-07-26 13:40:43 +02002448#ifdef CONFIG_PREEMPT_NOTIFIERS
2449
2450/**
Randy Dunlap421cee22007-07-31 00:37:50 -07002451 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2452 * @notifier: notifier struct to register
Avi Kivitye107be32007-07-26 13:40:43 +02002453 */
2454void preempt_notifier_register(struct preempt_notifier *notifier)
2455{
2456 hlist_add_head(&notifier->link, &current->preempt_notifiers);
2457}
2458EXPORT_SYMBOL_GPL(preempt_notifier_register);
2459
2460/**
2461 * preempt_notifier_unregister - no longer interested in preemption notifications
Randy Dunlap421cee22007-07-31 00:37:50 -07002462 * @notifier: notifier struct to unregister
Avi Kivitye107be32007-07-26 13:40:43 +02002463 *
2464 * This is safe to call from within a preemption notifier.
2465 */
2466void preempt_notifier_unregister(struct preempt_notifier *notifier)
2467{
2468 hlist_del(&notifier->link);
2469}
2470EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2471
2472static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2473{
2474 struct preempt_notifier *notifier;
2475 struct hlist_node *node;
2476
2477 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2478 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2479}
2480
2481static void
2482fire_sched_out_preempt_notifiers(struct task_struct *curr,
2483 struct task_struct *next)
2484{
2485 struct preempt_notifier *notifier;
2486 struct hlist_node *node;
2487
2488 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2489 notifier->ops->sched_out(notifier, next);
2490}
2491
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002492#else /* !CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002493
2494static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2495{
2496}
2497
2498static void
2499fire_sched_out_preempt_notifiers(struct task_struct *curr,
2500 struct task_struct *next)
2501{
2502}
2503
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02002504#endif /* CONFIG_PREEMPT_NOTIFIERS */
Avi Kivitye107be32007-07-26 13:40:43 +02002505
Linus Torvalds1da177e2005-04-16 15:20:36 -07002506/**
Nick Piggin4866cde2005-06-25 14:57:23 -07002507 * prepare_task_switch - prepare to switch tasks
2508 * @rq: the runqueue preparing to switch
Randy Dunlap421cee22007-07-31 00:37:50 -07002509 * @prev: the current task that is being switched out
Nick Piggin4866cde2005-06-25 14:57:23 -07002510 * @next: the task we are going to switch to.
2511 *
2512 * This is called with the rq lock held and interrupts off. It must
2513 * be paired with a subsequent finish_task_switch after the context
2514 * switch.
2515 *
2516 * prepare_task_switch sets up locking and calls architecture specific
2517 * hooks.
2518 */
Avi Kivitye107be32007-07-26 13:40:43 +02002519static inline void
2520prepare_task_switch(struct rq *rq, struct task_struct *prev,
2521 struct task_struct *next)
Nick Piggin4866cde2005-06-25 14:57:23 -07002522{
Avi Kivitye107be32007-07-26 13:40:43 +02002523 fire_sched_out_preempt_notifiers(prev, next);
Nick Piggin4866cde2005-06-25 14:57:23 -07002524 prepare_lock_switch(rq, next);
2525 prepare_arch_switch(next);
2526}
2527
2528/**
Linus Torvalds1da177e2005-04-16 15:20:36 -07002529 * finish_task_switch - clean up after a task-switch
Jeff Garzik344baba2005-09-07 01:15:17 -04002530 * @rq: runqueue associated with task-switch
Linus Torvalds1da177e2005-04-16 15:20:36 -07002531 * @prev: the thread we just switched away from.
2532 *
Nick Piggin4866cde2005-06-25 14:57:23 -07002533 * finish_task_switch must be called after the context switch, paired
2534 * with a prepare_task_switch call before the context switch.
2535 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2536 * and do any other architecture-specific cleanup actions.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002537 *
2538 * Note that we may have delayed dropping an mm in context_switch(). If
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002539 * so, we finish that here outside of the runqueue lock. (Doing it
Linus Torvalds1da177e2005-04-16 15:20:36 -07002540 * with the lock held can cause deadlocks; see schedule() for
2541 * details.)
2542 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02002543static void finish_task_switch(struct rq *rq, struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002544 __releases(rq->lock)
2545{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002546 struct mm_struct *mm = rq->prev_mm;
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002547 long prev_state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002548
2549 rq->prev_mm = NULL;
2550
2551 /*
2552 * A task struct has one reference for the use as "current".
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002553 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002554 * schedule one last time. The schedule call will never return, and
2555 * the scheduled task must drop that reference.
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002556 * The test for TASK_DEAD must occur while the runqueue locks are
Linus Torvalds1da177e2005-04-16 15:20:36 -07002557 * still held, otherwise prev could be scheduled on another cpu, die
2558 * there before we look at prev->state, and then the reference would
2559 * be dropped twice.
2560 * Manfred Spraul <manfred@colorfullife.com>
2561 */
Oleg Nesterov55a101f2006-09-29 02:01:10 -07002562 prev_state = prev->state;
Nick Piggin4866cde2005-06-25 14:57:23 -07002563 finish_arch_switch(prev);
2564 finish_lock_switch(rq, prev);
Steven Rostedt9a897c52008-01-25 21:08:22 +01002565#ifdef CONFIG_SMP
2566 if (current->sched_class->post_schedule)
2567 current->sched_class->post_schedule(rq);
2568#endif
Steven Rostedte8fa1362008-01-25 21:08:05 +01002569
Avi Kivitye107be32007-07-26 13:40:43 +02002570 fire_sched_in_preempt_notifiers(current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002571 if (mm)
2572 mmdrop(mm);
Oleg Nesterovc394cc92006-09-29 02:01:11 -07002573 if (unlikely(prev_state == TASK_DEAD)) {
bibo maoc6fd91f2006-03-26 01:38:20 -08002574 /*
2575 * Remove function-return probe instances associated with this
2576 * task and put them back on the free list.
Ingo Molnar9761eea2007-07-09 18:52:00 +02002577 */
bibo maoc6fd91f2006-03-26 01:38:20 -08002578 kprobe_flush_task(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002579 put_task_struct(prev);
bibo maoc6fd91f2006-03-26 01:38:20 -08002580 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002581}
2582
2583/**
2584 * schedule_tail - first thing a freshly forked thread must call.
2585 * @prev: the thread we just switched away from.
2586 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002587asmlinkage void schedule_tail(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002588 __releases(rq->lock)
2589{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002590 struct rq *rq = this_rq();
2591
Nick Piggin4866cde2005-06-25 14:57:23 -07002592 finish_task_switch(rq, prev);
2593#ifdef __ARCH_WANT_UNLOCKED_CTXSW
2594 /* In this case, finish_task_switch does not reenable preemption */
2595 preempt_enable();
2596#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002597 if (current->set_child_tid)
Pavel Emelyanovb4888932007-10-18 23:40:14 -07002598 put_user(task_pid_vnr(current), current->set_child_tid);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002599}
2600
2601/*
2602 * context_switch - switch to the new MM and the new
2603 * thread's register state.
2604 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002605static inline void
Ingo Molnar70b97a72006-07-03 00:25:42 -07002606context_switch(struct rq *rq, struct task_struct *prev,
Ingo Molnar36c8b582006-07-03 00:25:41 -07002607 struct task_struct *next)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002608{
Ingo Molnardd41f592007-07-09 18:51:59 +02002609 struct mm_struct *mm, *oldmm;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002610
Avi Kivitye107be32007-07-26 13:40:43 +02002611 prepare_task_switch(rq, prev, next);
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002612 trace_sched_switch(rq, prev, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02002613 mm = next->mm;
2614 oldmm = prev->active_mm;
Zachary Amsden9226d122007-02-13 13:26:21 +01002615 /*
2616 * For paravirt, this is coupled with an exit in switch_to to
2617 * combine the page table reload and the switch backend into
2618 * one hypercall.
2619 */
2620 arch_enter_lazy_cpu_mode();
2621
Ingo Molnardd41f592007-07-09 18:51:59 +02002622 if (unlikely(!mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002623 next->active_mm = oldmm;
2624 atomic_inc(&oldmm->mm_count);
2625 enter_lazy_tlb(oldmm, next);
2626 } else
2627 switch_mm(oldmm, mm, next);
2628
Ingo Molnardd41f592007-07-09 18:51:59 +02002629 if (unlikely(!prev->mm)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002630 prev->active_mm = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002631 rq->prev_mm = oldmm;
2632 }
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002633 /*
2634 * Since the runqueue lock will be released by the next
2635 * task (which is an invalid locking op but in the case
2636 * of the scheduler it's an obvious special-case), so we
2637 * do an early lockdep release here:
2638 */
2639#ifndef __ARCH_WANT_UNLOCKED_CTXSW
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07002640 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Ingo Molnar3a5f5e42006-07-14 00:24:27 -07002641#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07002642
2643 /* Here we just switch the register state and the stack. */
2644 switch_to(prev, next, prev);
2645
Ingo Molnardd41f592007-07-09 18:51:59 +02002646 barrier();
2647 /*
2648 * this_rq must be evaluated again because prev may have moved
2649 * CPUs since it called schedule(), thus the 'rq' on its stack
2650 * frame will be invalid.
2651 */
2652 finish_task_switch(this_rq(), prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002653}
2654
2655/*
2656 * nr_running, nr_uninterruptible and nr_context_switches:
2657 *
2658 * externally visible scheduler statistics: current number of runnable
2659 * threads, current number of uninterruptible-sleeping threads, total
2660 * number of context switches performed since bootup.
2661 */
2662unsigned long nr_running(void)
2663{
2664 unsigned long i, sum = 0;
2665
2666 for_each_online_cpu(i)
2667 sum += cpu_rq(i)->nr_running;
2668
2669 return sum;
2670}
2671
2672unsigned long nr_uninterruptible(void)
2673{
2674 unsigned long i, sum = 0;
2675
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002676 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002677 sum += cpu_rq(i)->nr_uninterruptible;
2678
2679 /*
2680 * Since we read the counters lockless, it might be slightly
2681 * inaccurate. Do not allow it to go below zero though:
2682 */
2683 if (unlikely((long)sum < 0))
2684 sum = 0;
2685
2686 return sum;
2687}
2688
2689unsigned long long nr_context_switches(void)
2690{
Steven Rostedtcc94abf2006-06-27 02:54:31 -07002691 int i;
2692 unsigned long long sum = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002693
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002694 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002695 sum += cpu_rq(i)->nr_switches;
2696
2697 return sum;
2698}
2699
2700unsigned long nr_iowait(void)
2701{
2702 unsigned long i, sum = 0;
2703
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08002704 for_each_possible_cpu(i)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002705 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2706
2707 return sum;
2708}
2709
Jack Steinerdb1b1fe2006-03-31 02:31:21 -08002710unsigned long nr_active(void)
2711{
2712 unsigned long i, running = 0, uninterruptible = 0;
2713
2714 for_each_online_cpu(i) {
2715 running += cpu_rq(i)->nr_running;
2716 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2717 }
2718
2719 if (unlikely((long)uninterruptible < 0))
2720 uninterruptible = 0;
2721
2722 return running + uninterruptible;
2723}
2724
Linus Torvalds1da177e2005-04-16 15:20:36 -07002725/*
Ingo Molnardd41f592007-07-09 18:51:59 +02002726 * Update rq->cpu_load[] statistics. This function is usually called every
2727 * scheduler tick (TICK_NSEC).
Ingo Molnar48f24c42006-07-03 00:25:40 -07002728 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002729static void update_cpu_load(struct rq *this_rq)
Ingo Molnar48f24c42006-07-03 00:25:40 -07002730{
Dmitry Adamushko495eca42007-10-15 17:00:06 +02002731 unsigned long this_load = this_rq->load.weight;
Ingo Molnardd41f592007-07-09 18:51:59 +02002732 int i, scale;
2733
2734 this_rq->nr_load_updates++;
Ingo Molnardd41f592007-07-09 18:51:59 +02002735
2736 /* Update our load: */
2737 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2738 unsigned long old_load, new_load;
2739
2740 /* scale is effectively 1 << i now, and >> i divides by scale */
2741
2742 old_load = this_rq->cpu_load[i];
2743 new_load = this_load;
Ingo Molnara25707f2007-10-15 17:00:03 +02002744 /*
2745 * Round up the averaging division if load is increasing. This
2746 * prevents us from getting stuck on 9 if the load is 10, for
2747 * example.
2748 */
2749 if (new_load > old_load)
2750 new_load += scale-1;
Ingo Molnardd41f592007-07-09 18:51:59 +02002751 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2752 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07002753}
2754
Ingo Molnardd41f592007-07-09 18:51:59 +02002755#ifdef CONFIG_SMP
2756
Ingo Molnar48f24c42006-07-03 00:25:40 -07002757/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002758 * double_rq_lock - safely lock two runqueues
2759 *
2760 * Note this does not disable interrupts like task_rq_lock,
2761 * you need to do so manually before calling.
2762 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002763static void double_rq_lock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002764 __acquires(rq1->lock)
2765 __acquires(rq2->lock)
2766{
Kirill Korotaev054b9102006-12-10 02:20:11 -08002767 BUG_ON(!irqs_disabled());
Linus Torvalds1da177e2005-04-16 15:20:36 -07002768 if (rq1 == rq2) {
2769 spin_lock(&rq1->lock);
2770 __acquire(rq2->lock); /* Fake it out ;) */
2771 } else {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002772 if (rq1 < rq2) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002773 spin_lock(&rq1->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02002774 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002775 } else {
2776 spin_lock(&rq2->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02002777 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002778 }
2779 }
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02002780 update_rq_clock(rq1);
2781 update_rq_clock(rq2);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002782}
2783
2784/*
2785 * double_rq_unlock - safely unlock two runqueues
2786 *
2787 * Note this does not restore interrupts like task_rq_unlock,
2788 * you need to do so manually after calling.
2789 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07002790static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002791 __releases(rq1->lock)
2792 __releases(rq2->lock)
2793{
2794 spin_unlock(&rq1->lock);
2795 if (rq1 != rq2)
2796 spin_unlock(&rq2->lock);
2797 else
2798 __release(rq2->lock);
2799}
2800
2801/*
2802 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2803 */
Steven Rostedte8fa1362008-01-25 21:08:05 +01002804static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002805 __releases(this_rq->lock)
2806 __acquires(busiest->lock)
2807 __acquires(this_rq->lock)
2808{
Steven Rostedte8fa1362008-01-25 21:08:05 +01002809 int ret = 0;
2810
Kirill Korotaev054b9102006-12-10 02:20:11 -08002811 if (unlikely(!irqs_disabled())) {
2812 /* printk() doesn't work good under rq->lock */
2813 spin_unlock(&this_rq->lock);
2814 BUG_ON(1);
2815 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002816 if (unlikely(!spin_trylock(&busiest->lock))) {
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07002817 if (busiest < this_rq) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07002818 spin_unlock(&this_rq->lock);
2819 spin_lock(&busiest->lock);
Peter Zijlstra5e710e32008-07-30 13:26:57 +02002820 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
Steven Rostedte8fa1362008-01-25 21:08:05 +01002821 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002822 } else
Peter Zijlstra5e710e32008-07-30 13:26:57 +02002823 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002824 }
Steven Rostedte8fa1362008-01-25 21:08:05 +01002825 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002826}
2827
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02002828static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2829 __releases(busiest->lock)
2830{
2831 spin_unlock(&busiest->lock);
2832 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2833}
2834
Linus Torvalds1da177e2005-04-16 15:20:36 -07002835/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07002836 * If dest_cpu is allowed for this process, migrate the task to it.
2837 * This is accomplished by forcing the cpu_allowed mask to only
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01002838 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
Linus Torvalds1da177e2005-04-16 15:20:36 -07002839 * the cpu_allowed mask is restored.
2840 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07002841static void sched_migrate_task(struct task_struct *p, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002842{
Ingo Molnar70b97a72006-07-03 00:25:42 -07002843 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002844 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07002845 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07002846
2847 rq = task_rq_lock(p, &flags);
2848 if (!cpu_isset(dest_cpu, p->cpus_allowed)
Max Krasnyanskye761b772008-07-15 04:43:49 -07002849 || unlikely(!cpu_active(dest_cpu)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07002850 goto out;
2851
Mathieu Desnoyers0a16b602008-07-18 12:16:17 -04002852 trace_sched_migrate_task(rq, p, dest_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002853 /* force the process onto the specified CPU */
2854 if (migrate_task(p, dest_cpu, &req)) {
2855 /* Need to wait for migration thread (might exit: take ref). */
2856 struct task_struct *mt = rq->migration_thread;
Ingo Molnar36c8b582006-07-03 00:25:41 -07002857
Linus Torvalds1da177e2005-04-16 15:20:36 -07002858 get_task_struct(mt);
2859 task_rq_unlock(rq, &flags);
2860 wake_up_process(mt);
2861 put_task_struct(mt);
2862 wait_for_completion(&req.done);
Ingo Molnar36c8b582006-07-03 00:25:41 -07002863
Linus Torvalds1da177e2005-04-16 15:20:36 -07002864 return;
2865 }
2866out:
2867 task_rq_unlock(rq, &flags);
2868}
2869
2870/*
Nick Piggin476d1392005-06-25 14:57:29 -07002871 * sched_exec - execve() is a valuable balancing opportunity, because at
2872 * this point the task has the smallest effective memory and cache footprint.
Linus Torvalds1da177e2005-04-16 15:20:36 -07002873 */
2874void sched_exec(void)
2875{
Linus Torvalds1da177e2005-04-16 15:20:36 -07002876 int new_cpu, this_cpu = get_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002877 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002878 put_cpu();
Nick Piggin476d1392005-06-25 14:57:29 -07002879 if (new_cpu != this_cpu)
2880 sched_migrate_task(current, new_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002881}
2882
2883/*
2884 * pull_task - move a task from a remote runqueue to the local runqueue.
2885 * Both runqueues must be locked.
2886 */
Ingo Molnardd41f592007-07-09 18:51:59 +02002887static void pull_task(struct rq *src_rq, struct task_struct *p,
2888 struct rq *this_rq, int this_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002889{
Ingo Molnar2e1cb742007-08-09 11:16:49 +02002890 deactivate_task(src_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002891 set_task_cpu(p, this_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02002892 activate_task(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002893 /*
2894 * Note that idle threads have a prio of MAX_PRIO, for this test
2895 * to be always true for them.
2896 */
Peter Zijlstra15afe092008-09-20 23:38:02 +02002897 check_preempt_curr(this_rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002898}
2899
2900/*
2901 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2902 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08002903static
Ingo Molnar70b97a72006-07-03 00:25:42 -07002904int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02002905 struct sched_domain *sd, enum cpu_idle_type idle,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07002906 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07002907{
2908 /*
2909 * We do not migrate tasks that are:
2910 * 1) running (obviously), or
2911 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2912 * 3) are cache-hot on their current CPU.
2913 */
Ingo Molnarcc367732007-10-15 17:00:18 +02002914 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2915 schedstat_inc(p, se.nr_failed_migrations_affine);
Linus Torvalds1da177e2005-04-16 15:20:36 -07002916 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002917 }
Nick Piggin81026792005-06-25 14:57:07 -07002918 *all_pinned = 0;
2919
Ingo Molnarcc367732007-10-15 17:00:18 +02002920 if (task_running(rq, p)) {
2921 schedstat_inc(p, se.nr_failed_migrations_running);
Nick Piggin81026792005-06-25 14:57:07 -07002922 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002923 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002924
Ingo Molnarda84d962007-10-15 17:00:18 +02002925 /*
2926 * Aggressive migration if:
2927 * 1) task is cache cold, or
2928 * 2) too many balance attempts have failed.
2929 */
2930
Ingo Molnar6bc16652007-10-15 17:00:18 +02002931 if (!task_hot(p, rq->clock, sd) ||
2932 sd->nr_balance_failed > sd->cache_nice_tries) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002933#ifdef CONFIG_SCHEDSTATS
Ingo Molnarcc367732007-10-15 17:00:18 +02002934 if (task_hot(p, rq->clock, sd)) {
Ingo Molnarda84d962007-10-15 17:00:18 +02002935 schedstat_inc(sd, lb_hot_gained[idle]);
Ingo Molnarcc367732007-10-15 17:00:18 +02002936 schedstat_inc(p, se.nr_forced_migrations);
2937 }
Ingo Molnarda84d962007-10-15 17:00:18 +02002938#endif
2939 return 1;
2940 }
2941
Ingo Molnarcc367732007-10-15 17:00:18 +02002942 if (task_hot(p, rq->clock, sd)) {
2943 schedstat_inc(p, se.nr_failed_migrations_hot);
Ingo Molnarda84d962007-10-15 17:00:18 +02002944 return 0;
Ingo Molnarcc367732007-10-15 17:00:18 +02002945 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07002946 return 1;
2947}
2948
Peter Williamse1d14842007-10-24 18:23:51 +02002949static unsigned long
2950balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2951 unsigned long max_load_move, struct sched_domain *sd,
2952 enum cpu_idle_type idle, int *all_pinned,
2953 int *this_best_prio, struct rq_iterator *iterator)
Ingo Molnardd41f592007-07-09 18:51:59 +02002954{
Peter Zijlstra051c6762008-06-27 13:41:31 +02002955 int loops = 0, pulled = 0, pinned = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02002956 struct task_struct *p;
2957 long rem_load_move = max_load_move;
2958
Peter Williamse1d14842007-10-24 18:23:51 +02002959 if (max_load_move == 0)
Ingo Molnardd41f592007-07-09 18:51:59 +02002960 goto out;
2961
2962 pinned = 1;
2963
2964 /*
2965 * Start the load-balancing iterator:
2966 */
2967 p = iterator->start(iterator->arg);
2968next:
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002969 if (!p || loops++ > sysctl_sched_nr_migrate)
Ingo Molnardd41f592007-07-09 18:51:59 +02002970 goto out;
Peter Zijlstra051c6762008-06-27 13:41:31 +02002971
2972 if ((p->se.load.weight >> 1) > rem_load_move ||
Ingo Molnardd41f592007-07-09 18:51:59 +02002973 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
Ingo Molnardd41f592007-07-09 18:51:59 +02002974 p = iterator->next(iterator->arg);
2975 goto next;
2976 }
2977
2978 pull_task(busiest, p, this_rq, this_cpu);
2979 pulled++;
2980 rem_load_move -= p->se.load.weight;
2981
2982 /*
Peter Zijlstrab82d9fd2007-11-09 22:39:39 +01002983 * We only want to steal up to the prescribed amount of weighted load.
Ingo Molnardd41f592007-07-09 18:51:59 +02002984 */
Peter Williamse1d14842007-10-24 18:23:51 +02002985 if (rem_load_move > 0) {
Peter Williamsa4ac01c2007-08-09 11:16:46 +02002986 if (p->prio < *this_best_prio)
2987 *this_best_prio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02002988 p = iterator->next(iterator->arg);
2989 goto next;
2990 }
2991out:
2992 /*
Peter Williamse1d14842007-10-24 18:23:51 +02002993 * Right now, this is one of only two places pull_task() is called,
Ingo Molnardd41f592007-07-09 18:51:59 +02002994 * so we can safely collect pull_task() stats here rather than
2995 * inside pull_task().
2996 */
2997 schedstat_add(sd, lb_gained[idle], pulled);
2998
2999 if (all_pinned)
3000 *all_pinned = pinned;
Peter Williamse1d14842007-10-24 18:23:51 +02003001
3002 return max_load_move - rem_load_move;
Ingo Molnardd41f592007-07-09 18:51:59 +02003003}
Ingo Molnar48f24c42006-07-03 00:25:40 -07003004
Linus Torvalds1da177e2005-04-16 15:20:36 -07003005/*
Peter Williams43010652007-08-09 11:16:46 +02003006 * move_tasks tries to move up to max_load_move weighted load from busiest to
3007 * this_rq, as part of a balancing operation within domain "sd".
3008 * Returns 1 if successful and 0 otherwise.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003009 *
3010 * Called with both runqueues locked.
3011 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003012static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
Peter Williams43010652007-08-09 11:16:46 +02003013 unsigned long max_load_move,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003014 struct sched_domain *sd, enum cpu_idle_type idle,
Peter Williams2dd73a42006-06-27 02:54:34 -07003015 int *all_pinned)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003016{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003017 const struct sched_class *class = sched_class_highest;
Peter Williams43010652007-08-09 11:16:46 +02003018 unsigned long total_load_moved = 0;
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003019 int this_best_prio = this_rq->curr->prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003020
Ingo Molnardd41f592007-07-09 18:51:59 +02003021 do {
Peter Williams43010652007-08-09 11:16:46 +02003022 total_load_moved +=
3023 class->load_balance(this_rq, this_cpu, busiest,
Peter Williamse1d14842007-10-24 18:23:51 +02003024 max_load_move - total_load_moved,
Peter Williamsa4ac01c2007-08-09 11:16:46 +02003025 sd, idle, all_pinned, &this_best_prio);
Ingo Molnardd41f592007-07-09 18:51:59 +02003026 class = class->next;
Gregory Haskinsc4acb2c2008-06-27 14:29:55 -06003027
3028 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3029 break;
3030
Peter Williams43010652007-08-09 11:16:46 +02003031 } while (class && max_load_move > total_load_moved);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003032
Peter Williams43010652007-08-09 11:16:46 +02003033 return total_load_moved > 0;
3034}
3035
Peter Williamse1d14842007-10-24 18:23:51 +02003036static int
3037iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3038 struct sched_domain *sd, enum cpu_idle_type idle,
3039 struct rq_iterator *iterator)
3040{
3041 struct task_struct *p = iterator->start(iterator->arg);
3042 int pinned = 0;
3043
3044 while (p) {
3045 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3046 pull_task(busiest, p, this_rq, this_cpu);
3047 /*
3048 * Right now, this is only the second place pull_task()
3049 * is called, so we can safely collect pull_task()
3050 * stats here rather than inside pull_task().
3051 */
3052 schedstat_inc(sd, lb_gained[idle]);
3053
3054 return 1;
3055 }
3056 p = iterator->next(iterator->arg);
3057 }
3058
3059 return 0;
3060}
3061
Peter Williams43010652007-08-09 11:16:46 +02003062/*
3063 * move_one_task tries to move exactly one task from busiest to this_rq, as
3064 * part of active balancing operations within "domain".
3065 * Returns 1 if successful and 0 otherwise.
3066 *
3067 * Called with both runqueues locked.
3068 */
3069static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3070 struct sched_domain *sd, enum cpu_idle_type idle)
3071{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02003072 const struct sched_class *class;
Peter Williams43010652007-08-09 11:16:46 +02003073
3074 for (class = sched_class_highest; class; class = class->next)
Peter Williamse1d14842007-10-24 18:23:51 +02003075 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
Peter Williams43010652007-08-09 11:16:46 +02003076 return 1;
3077
3078 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003079}
3080
3081/*
3082 * find_busiest_group finds and returns the busiest CPU group within the
Ingo Molnar48f24c42006-07-03 00:25:40 -07003083 * domain. It calculates and returns the amount of weighted load which
3084 * should be moved to restore balance via the imbalance parameter.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003085 */
3086static struct sched_group *
3087find_busiest_group(struct sched_domain *sd, int this_cpu,
Ingo Molnardd41f592007-07-09 18:51:59 +02003088 unsigned long *imbalance, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003089 int *sd_idle, const cpumask_t *cpus, int *balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003090{
3091 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3092 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003093 unsigned long max_pull;
Peter Williams2dd73a42006-06-27 02:54:34 -07003094 unsigned long busiest_load_per_task, busiest_nr_running;
3095 unsigned long this_load_per_task, this_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003096 int load_idx, group_imb = 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003097#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3098 int power_savings_balance = 1;
3099 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3100 unsigned long min_nr_running = ULONG_MAX;
3101 struct sched_group *group_min = NULL, *group_leader = NULL;
3102#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003103
3104 max_load = this_load = total_load = total_pwr = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003105 busiest_load_per_task = busiest_nr_running = 0;
3106 this_load_per_task = this_nr_running = 0;
Peter Zijlstra408ed062008-06-27 13:41:28 +02003107
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003108 if (idle == CPU_NOT_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003109 load_idx = sd->busy_idx;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003110 else if (idle == CPU_NEWLY_IDLE)
Nick Piggin78979862005-06-25 14:57:13 -07003111 load_idx = sd->newidle_idx;
3112 else
3113 load_idx = sd->idle_idx;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003114
3115 do {
Ken Chen908a7c12007-10-17 16:55:11 +02003116 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003117 int local_group;
3118 int i;
Ken Chen908a7c12007-10-17 16:55:11 +02003119 int __group_imb = 0;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003120 unsigned int balance_cpu = -1, first_idle_cpu = 0;
Peter Williams2dd73a42006-06-27 02:54:34 -07003121 unsigned long sum_nr_running, sum_weighted_load;
Peter Zijlstra408ed062008-06-27 13:41:28 +02003122 unsigned long sum_avg_load_per_task;
3123 unsigned long avg_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003124
3125 local_group = cpu_isset(this_cpu, group->cpumask);
3126
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003127 if (local_group)
3128 balance_cpu = first_cpu(group->cpumask);
3129
Linus Torvalds1da177e2005-04-16 15:20:36 -07003130 /* Tally up the load of all CPUs in the group */
Peter Williams2dd73a42006-06-27 02:54:34 -07003131 sum_weighted_load = sum_nr_running = avg_load = 0;
Peter Zijlstra408ed062008-06-27 13:41:28 +02003132 sum_avg_load_per_task = avg_load_per_task = 0;
3133
Ken Chen908a7c12007-10-17 16:55:11 +02003134 max_cpu_load = 0;
3135 min_cpu_load = ~0UL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003136
Mike Travis363ab6f2008-05-12 21:21:13 +02003137 for_each_cpu_mask_nr(i, group->cpumask) {
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003138 struct rq *rq;
3139
3140 if (!cpu_isset(i, *cpus))
3141 continue;
3142
3143 rq = cpu_rq(i);
Peter Williams2dd73a42006-06-27 02:54:34 -07003144
Suresh Siddha9439aab2007-07-19 21:28:35 +02003145 if (*sd_idle && rq->nr_running)
Nick Piggin5969fe02005-09-10 00:26:19 -07003146 *sd_idle = 0;
3147
Linus Torvalds1da177e2005-04-16 15:20:36 -07003148 /* Bias balancing toward cpus of our domain */
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003149 if (local_group) {
3150 if (idle_cpu(i) && !first_idle_cpu) {
3151 first_idle_cpu = 1;
3152 balance_cpu = i;
3153 }
3154
Nick Piggina2000572006-02-10 01:51:02 -08003155 load = target_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003156 } else {
Nick Piggina2000572006-02-10 01:51:02 -08003157 load = source_load(i, load_idx);
Ken Chen908a7c12007-10-17 16:55:11 +02003158 if (load > max_cpu_load)
3159 max_cpu_load = load;
3160 if (min_cpu_load > load)
3161 min_cpu_load = load;
3162 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003163
3164 avg_load += load;
Peter Williams2dd73a42006-06-27 02:54:34 -07003165 sum_nr_running += rq->nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003166 sum_weighted_load += weighted_cpuload(i);
Peter Zijlstra408ed062008-06-27 13:41:28 +02003167
3168 sum_avg_load_per_task += cpu_avg_load_per_task(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003169 }
3170
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003171 /*
3172 * First idle cpu or the first cpu(busiest) in this sched group
3173 * is eligible for doing load balancing at this and above
Suresh Siddha9439aab2007-07-19 21:28:35 +02003174 * domains. In the newly idle case, we will allow all the cpu's
3175 * to do the newly idle load balance.
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003176 */
Suresh Siddha9439aab2007-07-19 21:28:35 +02003177 if (idle != CPU_NEWLY_IDLE && local_group &&
3178 balance_cpu != this_cpu && balance) {
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003179 *balance = 0;
3180 goto ret;
3181 }
3182
Linus Torvalds1da177e2005-04-16 15:20:36 -07003183 total_load += avg_load;
Eric Dumazet5517d862007-05-08 00:32:57 -07003184 total_pwr += group->__cpu_power;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003185
3186 /* Adjust by relative CPU power of the group */
Eric Dumazet5517d862007-05-08 00:32:57 -07003187 avg_load = sg_div_cpu_power(group,
3188 avg_load * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003189
Peter Zijlstra408ed062008-06-27 13:41:28 +02003190
3191 /*
3192 * Consider the group unbalanced when the imbalance is larger
3193 * than the average weight of two tasks.
3194 *
3195 * APZ: with cgroup the avg task weight can vary wildly and
3196 * might not be a suitable number - should we keep a
3197 * normalized nr_running number somewhere that negates
3198 * the hierarchy?
3199 */
3200 avg_load_per_task = sg_div_cpu_power(group,
3201 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3202
3203 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
Ken Chen908a7c12007-10-17 16:55:11 +02003204 __group_imb = 1;
3205
Eric Dumazet5517d862007-05-08 00:32:57 -07003206 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003207
Linus Torvalds1da177e2005-04-16 15:20:36 -07003208 if (local_group) {
3209 this_load = avg_load;
3210 this = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003211 this_nr_running = sum_nr_running;
3212 this_load_per_task = sum_weighted_load;
3213 } else if (avg_load > max_load &&
Ken Chen908a7c12007-10-17 16:55:11 +02003214 (sum_nr_running > group_capacity || __group_imb)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003215 max_load = avg_load;
3216 busiest = group;
Peter Williams2dd73a42006-06-27 02:54:34 -07003217 busiest_nr_running = sum_nr_running;
3218 busiest_load_per_task = sum_weighted_load;
Ken Chen908a7c12007-10-17 16:55:11 +02003219 group_imb = __group_imb;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003220 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003221
3222#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3223 /*
3224 * Busy processors will not participate in power savings
3225 * balance.
3226 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003227 if (idle == CPU_NOT_IDLE ||
3228 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3229 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003230
3231 /*
3232 * If the local group is idle or completely loaded
3233 * no need to do power savings balance at this domain
3234 */
3235 if (local_group && (this_nr_running >= group_capacity ||
3236 !this_nr_running))
3237 power_savings_balance = 0;
3238
Ingo Molnardd41f592007-07-09 18:51:59 +02003239 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003240 * If a group is already running at full capacity or idle,
3241 * don't include that group in power savings calculations
Ingo Molnardd41f592007-07-09 18:51:59 +02003242 */
3243 if (!power_savings_balance || sum_nr_running >= group_capacity
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003244 || !sum_nr_running)
Ingo Molnardd41f592007-07-09 18:51:59 +02003245 goto group_next;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003246
Ingo Molnardd41f592007-07-09 18:51:59 +02003247 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003248 * Calculate the group which has the least non-idle load.
Ingo Molnardd41f592007-07-09 18:51:59 +02003249 * This is the group from where we need to pick up the load
3250 * for saving power
3251 */
3252 if ((sum_nr_running < min_nr_running) ||
3253 (sum_nr_running == min_nr_running &&
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003254 first_cpu(group->cpumask) <
3255 first_cpu(group_min->cpumask))) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003256 group_min = group;
3257 min_nr_running = sum_nr_running;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003258 min_load_per_task = sum_weighted_load /
3259 sum_nr_running;
Ingo Molnardd41f592007-07-09 18:51:59 +02003260 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003261
Ingo Molnardd41f592007-07-09 18:51:59 +02003262 /*
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003263 * Calculate the group which is almost near its
Ingo Molnardd41f592007-07-09 18:51:59 +02003264 * capacity but still has some space to pick up some load
3265 * from other group and save more power
3266 */
3267 if (sum_nr_running <= group_capacity - 1) {
3268 if (sum_nr_running > leader_nr_running ||
3269 (sum_nr_running == leader_nr_running &&
3270 first_cpu(group->cpumask) >
3271 first_cpu(group_leader->cpumask))) {
3272 group_leader = group;
3273 leader_nr_running = sum_nr_running;
3274 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07003275 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003276group_next:
3277#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07003278 group = group->next;
3279 } while (group != sd->groups);
3280
Peter Williams2dd73a42006-06-27 02:54:34 -07003281 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003282 goto out_balanced;
3283
3284 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3285
3286 if (this_load >= avg_load ||
3287 100*max_load <= sd->imbalance_pct*this_load)
3288 goto out_balanced;
3289
Peter Williams2dd73a42006-06-27 02:54:34 -07003290 busiest_load_per_task /= busiest_nr_running;
Ken Chen908a7c12007-10-17 16:55:11 +02003291 if (group_imb)
3292 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3293
Linus Torvalds1da177e2005-04-16 15:20:36 -07003294 /*
3295 * We're trying to get all the cpus to the average_load, so we don't
3296 * want to push ourselves above the average load, nor do we wish to
3297 * reduce the max loaded cpu below the average load, as either of these
3298 * actions would just result in more rebalancing later, and ping-pong
3299 * tasks around. Thus we look for the minimum possible imbalance.
3300 * Negative imbalances (*we* are more loaded than anyone else) will
3301 * be counted as no imbalance for these purposes -- we can't fix that
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003302 * by pulling tasks to us. Be careful of negative numbers as they'll
Linus Torvalds1da177e2005-04-16 15:20:36 -07003303 * appear as very large values with unsigned longs.
3304 */
Peter Williams2dd73a42006-06-27 02:54:34 -07003305 if (max_load <= busiest_load_per_task)
3306 goto out_balanced;
3307
3308 /*
3309 * In the presence of smp nice balancing, certain scenarios can have
3310 * max load less than avg load(as we skip the groups at or below
3311 * its cpu_power, while calculating max_load..)
3312 */
3313 if (max_load < avg_load) {
3314 *imbalance = 0;
3315 goto small_imbalance;
3316 }
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003317
3318 /* Don't want to pull so many tasks that a group would go idle */
Peter Williams2dd73a42006-06-27 02:54:34 -07003319 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
Siddha, Suresh B0c117f12005-09-10 00:26:21 -07003320
Linus Torvalds1da177e2005-04-16 15:20:36 -07003321 /* How much load to actually move to equalise the imbalance */
Eric Dumazet5517d862007-05-08 00:32:57 -07003322 *imbalance = min(max_pull * busiest->__cpu_power,
3323 (avg_load - this_load) * this->__cpu_power)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003324 / SCHED_LOAD_SCALE;
3325
Peter Williams2dd73a42006-06-27 02:54:34 -07003326 /*
3327 * if *imbalance is less than the average load per runnable task
3328 * there is no gaurantee that any tasks will be moved so we'll have
3329 * a think about bumping its value to force at least one task to be
3330 * moved
3331 */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003332 if (*imbalance < busiest_load_per_task) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07003333 unsigned long tmp, pwr_now, pwr_move;
Peter Williams2dd73a42006-06-27 02:54:34 -07003334 unsigned int imbn;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003335
Peter Williams2dd73a42006-06-27 02:54:34 -07003336small_imbalance:
3337 pwr_move = pwr_now = 0;
3338 imbn = 2;
3339 if (this_nr_running) {
3340 this_load_per_task /= this_nr_running;
3341 if (busiest_load_per_task > this_load_per_task)
3342 imbn = 1;
3343 } else
Peter Zijlstra408ed062008-06-27 13:41:28 +02003344 this_load_per_task = cpu_avg_load_per_task(this_cpu);
Peter Williams2dd73a42006-06-27 02:54:34 -07003345
Peter Zijlstra01c8c572008-10-24 11:06:12 +02003346 if (max_load - this_load + busiest_load_per_task >=
Ingo Molnardd41f592007-07-09 18:51:59 +02003347 busiest_load_per_task * imbn) {
Peter Williams2dd73a42006-06-27 02:54:34 -07003348 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003349 return busiest;
3350 }
3351
3352 /*
3353 * OK, we don't have enough imbalance to justify moving tasks,
3354 * however we may be able to increase total CPU power used by
3355 * moving them.
3356 */
3357
Eric Dumazet5517d862007-05-08 00:32:57 -07003358 pwr_now += busiest->__cpu_power *
3359 min(busiest_load_per_task, max_load);
3360 pwr_now += this->__cpu_power *
3361 min(this_load_per_task, this_load);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003362 pwr_now /= SCHED_LOAD_SCALE;
3363
3364 /* Amount of load we'd subtract */
Eric Dumazet5517d862007-05-08 00:32:57 -07003365 tmp = sg_div_cpu_power(busiest,
3366 busiest_load_per_task * SCHED_LOAD_SCALE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003367 if (max_load > tmp)
Eric Dumazet5517d862007-05-08 00:32:57 -07003368 pwr_move += busiest->__cpu_power *
Peter Williams2dd73a42006-06-27 02:54:34 -07003369 min(busiest_load_per_task, max_load - tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003370
3371 /* Amount of load we'd add */
Eric Dumazet5517d862007-05-08 00:32:57 -07003372 if (max_load * busiest->__cpu_power <
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08003373 busiest_load_per_task * SCHED_LOAD_SCALE)
Eric Dumazet5517d862007-05-08 00:32:57 -07003374 tmp = sg_div_cpu_power(this,
3375 max_load * busiest->__cpu_power);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003376 else
Eric Dumazet5517d862007-05-08 00:32:57 -07003377 tmp = sg_div_cpu_power(this,
3378 busiest_load_per_task * SCHED_LOAD_SCALE);
3379 pwr_move += this->__cpu_power *
3380 min(this_load_per_task, this_load + tmp);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003381 pwr_move /= SCHED_LOAD_SCALE;
3382
3383 /* Move if we gain throughput */
Suresh Siddha7fd0d2d2007-09-05 14:32:48 +02003384 if (pwr_move > pwr_now)
3385 *imbalance = busiest_load_per_task;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003386 }
3387
Linus Torvalds1da177e2005-04-16 15:20:36 -07003388 return busiest;
3389
3390out_balanced:
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003391#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003392 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003393 goto ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003394
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003395 if (this == group_leader && group_leader != group_min) {
3396 *imbalance = min_load_per_task;
3397 return group_min;
3398 }
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07003399#endif
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003400ret:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003401 *imbalance = 0;
3402 return NULL;
3403}
3404
3405/*
3406 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3407 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003408static struct rq *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003409find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003410 unsigned long imbalance, const cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003411{
Ingo Molnar70b97a72006-07-03 00:25:42 -07003412 struct rq *busiest = NULL, *rq;
Peter Williams2dd73a42006-06-27 02:54:34 -07003413 unsigned long max_load = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003414 int i;
3415
Mike Travis363ab6f2008-05-12 21:21:13 +02003416 for_each_cpu_mask_nr(i, group->cpumask) {
Ingo Molnardd41f592007-07-09 18:51:59 +02003417 unsigned long wl;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003418
3419 if (!cpu_isset(i, *cpus))
3420 continue;
3421
Ingo Molnar48f24c42006-07-03 00:25:40 -07003422 rq = cpu_rq(i);
Ingo Molnardd41f592007-07-09 18:51:59 +02003423 wl = weighted_cpuload(i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003424
Ingo Molnardd41f592007-07-09 18:51:59 +02003425 if (rq->nr_running == 1 && wl > imbalance)
Peter Williams2dd73a42006-06-27 02:54:34 -07003426 continue;
3427
Ingo Molnardd41f592007-07-09 18:51:59 +02003428 if (wl > max_load) {
3429 max_load = wl;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003430 busiest = rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003431 }
3432 }
3433
3434 return busiest;
3435}
3436
3437/*
Nick Piggin77391d72005-06-25 14:57:30 -07003438 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3439 * so long as it is large enough.
3440 */
3441#define MAX_PINNED_INTERVAL 512
3442
3443/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07003444 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3445 * tasks if there is an imbalance.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003446 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003447static int load_balance(int this_cpu, struct rq *this_rq,
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003448 struct sched_domain *sd, enum cpu_idle_type idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003449 int *balance, cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003450{
Peter Williams43010652007-08-09 11:16:46 +02003451 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003452 struct sched_group *group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003453 unsigned long imbalance;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003454 struct rq *busiest;
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003455 unsigned long flags;
Nick Piggin5969fe02005-09-10 00:26:19 -07003456
Mike Travis7c16ec52008-04-04 18:11:11 -07003457 cpus_setall(*cpus);
3458
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003459 /*
3460 * When power savings policy is enabled for the parent domain, idle
3461 * sibling can pick up load irrespective of busy siblings. In this case,
Ingo Molnardd41f592007-07-09 18:51:59 +02003462 * let the state of idle sibling percolate up as CPU_IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003463 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003464 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003465 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003466 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003467 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003468
Ingo Molnar2d723762007-10-15 17:00:12 +02003469 schedstat_inc(sd, lb_count[idle]);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003470
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003471redo:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02003472 update_shares(sd);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003473 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
Mike Travis7c16ec52008-04-04 18:11:11 -07003474 cpus, balance);
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003475
Chen, Kenneth W06066712006-12-10 02:20:35 -08003476 if (*balance == 0)
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003477 goto out_balanced;
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003478
Linus Torvalds1da177e2005-04-16 15:20:36 -07003479 if (!group) {
3480 schedstat_inc(sd, lb_nobusyg[idle]);
3481 goto out_balanced;
3482 }
3483
Mike Travis7c16ec52008-04-04 18:11:11 -07003484 busiest = find_busiest_queue(group, idle, imbalance, cpus);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003485 if (!busiest) {
3486 schedstat_inc(sd, lb_nobusyq[idle]);
3487 goto out_balanced;
3488 }
3489
Nick Piggindb935db2005-06-25 14:57:11 -07003490 BUG_ON(busiest == this_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003491
3492 schedstat_add(sd, lb_imbalance[idle], imbalance);
3493
Peter Williams43010652007-08-09 11:16:46 +02003494 ld_moved = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003495 if (busiest->nr_running > 1) {
3496 /*
3497 * Attempt to move tasks. If find_busiest_group has found
3498 * an imbalance but busiest->nr_running <= 1, the group is
Peter Williams43010652007-08-09 11:16:46 +02003499 * still unbalanced. ld_moved simply stays zero, so it is
Linus Torvalds1da177e2005-04-16 15:20:36 -07003500 * correctly treated as an imbalance.
3501 */
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003502 local_irq_save(flags);
Nick Piggine17224b2005-09-10 00:26:18 -07003503 double_rq_lock(this_rq, busiest);
Peter Williams43010652007-08-09 11:16:46 +02003504 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Ingo Molnar48f24c42006-07-03 00:25:40 -07003505 imbalance, sd, idle, &all_pinned);
Nick Piggine17224b2005-09-10 00:26:18 -07003506 double_rq_unlock(this_rq, busiest);
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003507 local_irq_restore(flags);
Nick Piggin81026792005-06-25 14:57:07 -07003508
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003509 /*
3510 * some other cpu did the load balance for us.
3511 */
Peter Williams43010652007-08-09 11:16:46 +02003512 if (ld_moved && this_cpu != smp_processor_id())
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003513 resched_cpu(this_cpu);
3514
Nick Piggin81026792005-06-25 14:57:07 -07003515 /* All tasks on this runqueue were pinned by CPU affinity */
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003516 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003517 cpu_clear(cpu_of(busiest), *cpus);
3518 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003519 goto redo;
Nick Piggin81026792005-06-25 14:57:07 -07003520 goto out_balanced;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003521 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003522 }
Nick Piggin81026792005-06-25 14:57:07 -07003523
Peter Williams43010652007-08-09 11:16:46 +02003524 if (!ld_moved) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003525 schedstat_inc(sd, lb_failed[idle]);
3526 sd->nr_balance_failed++;
3527
3528 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003529
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003530 spin_lock_irqsave(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003531
3532 /* don't kick the migration_thread, if the curr
3533 * task on busiest cpu can't be moved to this_cpu
3534 */
3535 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003536 spin_unlock_irqrestore(&busiest->lock, flags);
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003537 all_pinned = 1;
3538 goto out_one_pinned;
3539 }
3540
Linus Torvalds1da177e2005-04-16 15:20:36 -07003541 if (!busiest->active_balance) {
3542 busiest->active_balance = 1;
3543 busiest->push_cpu = this_cpu;
Nick Piggin81026792005-06-25 14:57:07 -07003544 active_balance = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003545 }
Christoph Lameterfe2eea32006-12-10 02:20:21 -08003546 spin_unlock_irqrestore(&busiest->lock, flags);
Nick Piggin81026792005-06-25 14:57:07 -07003547 if (active_balance)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003548 wake_up_process(busiest->migration_thread);
3549
3550 /*
3551 * We've kicked active balancing, reset the failure
3552 * counter.
3553 */
Nick Piggin39507452005-06-25 14:57:09 -07003554 sd->nr_balance_failed = sd->cache_nice_tries+1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003555 }
Nick Piggin81026792005-06-25 14:57:07 -07003556 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07003557 sd->nr_balance_failed = 0;
3558
Nick Piggin81026792005-06-25 14:57:07 -07003559 if (likely(!active_balance)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003560 /* We were unbalanced, so reset the balancing interval */
3561 sd->balance_interval = sd->min_interval;
Nick Piggin81026792005-06-25 14:57:07 -07003562 } else {
3563 /*
3564 * If we've begun active balancing, start to back off. This
3565 * case may not be covered by the all_pinned logic if there
3566 * is only 1 task on the busy runqueue (because we don't call
3567 * move_tasks).
3568 */
3569 if (sd->balance_interval < sd->max_interval)
3570 sd->balance_interval *= 2;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003571 }
3572
Peter Williams43010652007-08-09 11:16:46 +02003573 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003574 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003575 ld_moved = -1;
3576
3577 goto out;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003578
3579out_balanced:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003580 schedstat_inc(sd, lb_balanced[idle]);
3581
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003582 sd->nr_balance_failed = 0;
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003583
3584out_one_pinned:
Linus Torvalds1da177e2005-04-16 15:20:36 -07003585 /* tune up the balancing interval */
Nick Piggin77391d72005-06-25 14:57:30 -07003586 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3587 (sd->balance_interval < sd->max_interval))
Linus Torvalds1da177e2005-04-16 15:20:36 -07003588 sd->balance_interval *= 2;
3589
Ingo Molnar48f24c42006-07-03 00:25:40 -07003590 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003591 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003592 ld_moved = -1;
3593 else
3594 ld_moved = 0;
3595out:
Peter Zijlstrac8cba852008-06-27 13:41:23 +02003596 if (ld_moved)
3597 update_shares(sd);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02003598 return ld_moved;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003599}
3600
3601/*
3602 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3603 * tasks if there is an imbalance.
3604 *
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003605 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
Linus Torvalds1da177e2005-04-16 15:20:36 -07003606 * this_rq is locked.
3607 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07003608static int
Mike Travis7c16ec52008-04-04 18:11:11 -07003609load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3610 cpumask_t *cpus)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003611{
3612 struct sched_group *group;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003613 struct rq *busiest = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003614 unsigned long imbalance;
Peter Williams43010652007-08-09 11:16:46 +02003615 int ld_moved = 0;
Nick Piggin5969fe02005-09-10 00:26:19 -07003616 int sd_idle = 0;
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003617 int all_pinned = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07003618
3619 cpus_setall(*cpus);
Nick Piggin5969fe02005-09-10 00:26:19 -07003620
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003621 /*
3622 * When power savings policy is enabled for the parent domain, idle
3623 * sibling can pick up load irrespective of busy siblings. In this case,
3624 * let the state of idle sibling percolate up as IDLE, instead of
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003625 * portraying it as CPU_NOT_IDLE.
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003626 */
3627 if (sd->flags & SD_SHARE_CPUPOWER &&
3628 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003629 sd_idle = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003630
Ingo Molnar2d723762007-10-15 17:00:12 +02003631 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003632redo:
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02003633 update_shares_locked(this_rq, sd);
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003634 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
Mike Travis7c16ec52008-04-04 18:11:11 -07003635 &sd_idle, cpus, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003636 if (!group) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003637 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003638 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003639 }
3640
Mike Travis7c16ec52008-04-04 18:11:11 -07003641 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
Nick Piggindb935db2005-06-25 14:57:11 -07003642 if (!busiest) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003643 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003644 goto out_balanced;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003645 }
3646
Nick Piggindb935db2005-06-25 14:57:11 -07003647 BUG_ON(busiest == this_rq);
3648
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003649 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003650
Peter Williams43010652007-08-09 11:16:46 +02003651 ld_moved = 0;
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003652 if (busiest->nr_running > 1) {
3653 /* Attempt to move tasks */
3654 double_lock_balance(this_rq, busiest);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003655 /* this_rq->clock is already updated */
3656 update_rq_clock(busiest);
Peter Williams43010652007-08-09 11:16:46 +02003657 ld_moved = move_tasks(this_rq, this_cpu, busiest,
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003658 imbalance, sd, CPU_NEWLY_IDLE,
3659 &all_pinned);
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02003660 double_unlock_balance(this_rq, busiest);
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003661
Suresh Siddha969bb4e2007-07-19 21:28:35 +02003662 if (unlikely(all_pinned)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003663 cpu_clear(cpu_of(busiest), *cpus);
3664 if (!cpus_empty(*cpus))
Christoph Lameter0a2966b2006-09-25 23:30:51 -07003665 goto redo;
3666 }
Nick Piggind6d5cfa2005-09-10 00:26:16 -07003667 }
3668
Peter Williams43010652007-08-09 11:16:46 +02003669 if (!ld_moved) {
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003670 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003671 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3672 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003673 return -1;
3674 } else
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003675 sd->nr_balance_failed = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003676
Peter Zijlstra3e5459b2008-06-27 13:41:24 +02003677 update_shares_locked(this_rq, sd);
Peter Williams43010652007-08-09 11:16:46 +02003678 return ld_moved;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003679
3680out_balanced:
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003681 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
Ingo Molnar48f24c42006-07-03 00:25:40 -07003682 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
Siddha, Suresh B89c47102006-10-03 01:14:09 -07003683 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
Nick Piggin5969fe02005-09-10 00:26:19 -07003684 return -1;
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003685 sd->nr_balance_failed = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07003686
Nick Piggin16cfb1c2005-06-25 14:57:08 -07003687 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003688}
3689
3690/*
3691 * idle_balance is called by schedule() if this_cpu is about to become
3692 * idle. Attempts to pull tasks from other CPUs.
3693 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003694static void idle_balance(int this_cpu, struct rq *this_rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003695{
3696 struct sched_domain *sd;
Ingo Molnardd41f592007-07-09 18:51:59 +02003697 int pulled_task = -1;
3698 unsigned long next_balance = jiffies + HZ;
Mike Travis7c16ec52008-04-04 18:11:11 -07003699 cpumask_t tmpmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003700
3701 for_each_domain(this_cpu, sd) {
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003702 unsigned long interval;
3703
3704 if (!(sd->flags & SD_LOAD_BALANCE))
3705 continue;
3706
3707 if (sd->flags & SD_BALANCE_NEWIDLE)
Ingo Molnar48f24c42006-07-03 00:25:40 -07003708 /* If we've pulled tasks over stop searching: */
Mike Travis7c16ec52008-04-04 18:11:11 -07003709 pulled_task = load_balance_newidle(this_cpu, this_rq,
3710 sd, &tmpmask);
Christoph Lameter92c4ca52007-06-23 17:16:33 -07003711
3712 interval = msecs_to_jiffies(sd->balance_interval);
3713 if (time_after(next_balance, sd->last_balance + interval))
3714 next_balance = sd->last_balance + interval;
3715 if (pulled_task)
3716 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003717 }
Ingo Molnardd41f592007-07-09 18:51:59 +02003718 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003719 /*
3720 * We are going idle. next_balance may be set based on
3721 * a busy processor. So reset next_balance.
3722 */
3723 this_rq->next_balance = next_balance;
Ingo Molnardd41f592007-07-09 18:51:59 +02003724 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003725}
3726
3727/*
3728 * active_load_balance is run by migration threads. It pushes running tasks
3729 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3730 * running on each physical CPU where possible, and avoids physical /
3731 * logical imbalances.
3732 *
3733 * Called with busiest_rq locked.
3734 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07003735static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003736{
Nick Piggin39507452005-06-25 14:57:09 -07003737 int target_cpu = busiest_rq->push_cpu;
Ingo Molnar70b97a72006-07-03 00:25:42 -07003738 struct sched_domain *sd;
3739 struct rq *target_rq;
Nick Piggin39507452005-06-25 14:57:09 -07003740
Ingo Molnar48f24c42006-07-03 00:25:40 -07003741 /* Is there any task to move? */
Nick Piggin39507452005-06-25 14:57:09 -07003742 if (busiest_rq->nr_running <= 1)
Nick Piggin39507452005-06-25 14:57:09 -07003743 return;
3744
3745 target_rq = cpu_rq(target_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003746
3747 /*
Nick Piggin39507452005-06-25 14:57:09 -07003748 * This condition is "impossible", if it occurs
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003749 * we need to fix it. Originally reported by
Nick Piggin39507452005-06-25 14:57:09 -07003750 * Bjorn Helgaas on a 128-cpu setup.
Linus Torvalds1da177e2005-04-16 15:20:36 -07003751 */
Nick Piggin39507452005-06-25 14:57:09 -07003752 BUG_ON(busiest_rq == target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003753
Nick Piggin39507452005-06-25 14:57:09 -07003754 /* move a task from busiest_rq to target_rq */
3755 double_lock_balance(busiest_rq, target_rq);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02003756 update_rq_clock(busiest_rq);
3757 update_rq_clock(target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003758
Nick Piggin39507452005-06-25 14:57:09 -07003759 /* Search for an sd spanning us and the target CPU. */
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003760 for_each_domain(target_cpu, sd) {
Nick Piggin39507452005-06-25 14:57:09 -07003761 if ((sd->flags & SD_LOAD_BALANCE) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07003762 cpu_isset(busiest_cpu, sd->span))
Nick Piggin39507452005-06-25 14:57:09 -07003763 break;
Chen, Kenneth Wc96d1452006-06-27 02:54:28 -07003764 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07003765
Ingo Molnar48f24c42006-07-03 00:25:40 -07003766 if (likely(sd)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02003767 schedstat_inc(sd, alb_count);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003768
Peter Williams43010652007-08-09 11:16:46 +02003769 if (move_one_task(target_rq, target_cpu, busiest_rq,
3770 sd, CPU_IDLE))
Ingo Molnar48f24c42006-07-03 00:25:40 -07003771 schedstat_inc(sd, alb_pushed);
3772 else
3773 schedstat_inc(sd, alb_failed);
3774 }
Peter Zijlstra1b12bbc2008-08-11 09:30:22 +02003775 double_unlock_balance(busiest_rq, target_rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07003776}
3777
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003778#ifdef CONFIG_NO_HZ
3779static struct {
3780 atomic_t load_balancer;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01003781 cpumask_t cpu_mask;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003782} nohz ____cacheline_aligned = {
3783 .load_balancer = ATOMIC_INIT(-1),
3784 .cpu_mask = CPU_MASK_NONE,
3785};
3786
Christoph Lameter7835b982006-12-10 02:20:22 -08003787/*
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003788 * This routine will try to nominate the ilb (idle load balancing)
3789 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3790 * load balancing on behalf of all those cpus. If all the cpus in the system
3791 * go into this tickless mode, then there will be no ilb owner (as there is
3792 * no need for one) and all the cpus will sleep till the next wakeup event
3793 * arrives...
Christoph Lameter7835b982006-12-10 02:20:22 -08003794 *
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003795 * For the ilb owner, tick is not stopped. And this tick will be used
3796 * for idle load balancing. ilb owner will still be part of
3797 * nohz.cpu_mask..
3798 *
3799 * While stopping the tick, this cpu will become the ilb owner if there
3800 * is no other owner. And will be the owner till that cpu becomes busy
3801 * or if all cpus in the system stop their ticks at which point
3802 * there is no need for ilb owner.
3803 *
3804 * When the ilb owner becomes busy, it nominates another owner, during the
3805 * next busy scheduler_tick()
3806 */
3807int select_nohz_load_balancer(int stop_tick)
3808{
3809 int cpu = smp_processor_id();
3810
3811 if (stop_tick) {
3812 cpu_set(cpu, nohz.cpu_mask);
3813 cpu_rq(cpu)->in_nohz_recently = 1;
3814
3815 /*
3816 * If we are going offline and still the leader, give up!
3817 */
Max Krasnyanskye761b772008-07-15 04:43:49 -07003818 if (!cpu_active(cpu) &&
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003819 atomic_read(&nohz.load_balancer) == cpu) {
3820 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3821 BUG();
3822 return 0;
3823 }
3824
3825 /* time for ilb owner also to sleep */
3826 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3827 if (atomic_read(&nohz.load_balancer) == cpu)
3828 atomic_set(&nohz.load_balancer, -1);
3829 return 0;
3830 }
3831
3832 if (atomic_read(&nohz.load_balancer) == -1) {
3833 /* make me the ilb owner */
3834 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3835 return 1;
3836 } else if (atomic_read(&nohz.load_balancer) == cpu)
3837 return 1;
3838 } else {
3839 if (!cpu_isset(cpu, nohz.cpu_mask))
3840 return 0;
3841
3842 cpu_clear(cpu, nohz.cpu_mask);
3843
3844 if (atomic_read(&nohz.load_balancer) == cpu)
3845 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3846 BUG();
3847 }
3848 return 0;
3849}
3850#endif
3851
3852static DEFINE_SPINLOCK(balancing);
3853
3854/*
Christoph Lameter7835b982006-12-10 02:20:22 -08003855 * It checks each scheduling domain to see if it is due to be balanced,
3856 * and initiates a balancing operation if so.
3857 *
3858 * Balancing parameters are set up in arch_init_sched_domains.
3859 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02003860static void rebalance_domains(int cpu, enum cpu_idle_type idle)
Christoph Lameter7835b982006-12-10 02:20:22 -08003861{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003862 int balance = 1;
3863 struct rq *rq = cpu_rq(cpu);
Christoph Lameter7835b982006-12-10 02:20:22 -08003864 unsigned long interval;
3865 struct sched_domain *sd;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003866 /* Earliest time when we have to do rebalance again */
Christoph Lameterc9819f42006-12-10 02:20:25 -08003867 unsigned long next_balance = jiffies + 60*HZ;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003868 int update_next_balance = 0;
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003869 int need_serialize;
Mike Travis7c16ec52008-04-04 18:11:11 -07003870 cpumask_t tmp;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003871
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003872 for_each_domain(cpu, sd) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07003873 if (!(sd->flags & SD_LOAD_BALANCE))
3874 continue;
3875
3876 interval = sd->balance_interval;
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003877 if (idle != CPU_IDLE)
Linus Torvalds1da177e2005-04-16 15:20:36 -07003878 interval *= sd->busy_factor;
3879
3880 /* scale ms to jiffies */
3881 interval = msecs_to_jiffies(interval);
3882 if (unlikely(!interval))
3883 interval = 1;
Ingo Molnardd41f592007-07-09 18:51:59 +02003884 if (interval > HZ*NR_CPUS/10)
3885 interval = HZ*NR_CPUS/10;
3886
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003887 need_serialize = sd->flags & SD_SERIALIZE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003888
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003889 if (need_serialize) {
Christoph Lameter08c183f2006-12-10 02:20:29 -08003890 if (!spin_trylock(&balancing))
3891 goto out;
3892 }
3893
Christoph Lameterc9819f42006-12-10 02:20:25 -08003894 if (time_after_eq(jiffies, sd->last_balance + interval)) {
Mike Travis7c16ec52008-04-04 18:11:11 -07003895 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
Siddha, Suresh Bfa3b6dd2005-09-10 00:26:21 -07003896 /*
3897 * We've pulled tasks over so either we're no
Nick Piggin5969fe02005-09-10 00:26:19 -07003898 * longer idle, or one of our SMT siblings is
3899 * not idle.
3900 */
Ingo Molnard15bcfd2007-07-09 18:51:57 +02003901 idle = CPU_NOT_IDLE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003902 }
Christoph Lameter1bd77f22006-12-10 02:20:27 -08003903 sd->last_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003904 }
Dmitry Adamushkod07355f2008-05-12 21:21:15 +02003905 if (need_serialize)
Christoph Lameter08c183f2006-12-10 02:20:29 -08003906 spin_unlock(&balancing);
3907out:
Suresh Siddhaf549da82007-08-23 15:18:02 +02003908 if (time_after(next_balance, sd->last_balance + interval)) {
Christoph Lameterc9819f42006-12-10 02:20:25 -08003909 next_balance = sd->last_balance + interval;
Suresh Siddhaf549da82007-08-23 15:18:02 +02003910 update_next_balance = 1;
3911 }
Siddha, Suresh B783609c2006-12-10 02:20:33 -08003912
3913 /*
3914 * Stop the load balance at this level. There is another
3915 * CPU in our sched group which is doing load balancing more
3916 * actively.
3917 */
3918 if (!balance)
3919 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07003920 }
Suresh Siddhaf549da82007-08-23 15:18:02 +02003921
3922 /*
3923 * next_balance will be updated only when there is a need.
3924 * When the cpu is attached to null domain for ex, it will not be
3925 * updated.
3926 */
3927 if (likely(update_next_balance))
3928 rq->next_balance = next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003929}
3930
3931/*
3932 * run_rebalance_domains is triggered when needed from the scheduler tick.
3933 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3934 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3935 */
3936static void run_rebalance_domains(struct softirq_action *h)
3937{
Ingo Molnardd41f592007-07-09 18:51:59 +02003938 int this_cpu = smp_processor_id();
3939 struct rq *this_rq = cpu_rq(this_cpu);
3940 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3941 CPU_IDLE : CPU_NOT_IDLE;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003942
Ingo Molnardd41f592007-07-09 18:51:59 +02003943 rebalance_domains(this_cpu, idle);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003944
3945#ifdef CONFIG_NO_HZ
3946 /*
3947 * If this cpu is the owner for idle load balancing, then do the
3948 * balancing on behalf of the other idle cpus whose ticks are
3949 * stopped.
3950 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003951 if (this_rq->idle_at_tick &&
3952 atomic_read(&nohz.load_balancer) == this_cpu) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003953 cpumask_t cpus = nohz.cpu_mask;
3954 struct rq *rq;
3955 int balance_cpu;
3956
Ingo Molnardd41f592007-07-09 18:51:59 +02003957 cpu_clear(this_cpu, cpus);
Mike Travis363ab6f2008-05-12 21:21:13 +02003958 for_each_cpu_mask_nr(balance_cpu, cpus) {
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003959 /*
3960 * If this cpu gets work to do, stop the load balancing
3961 * work being done for other cpus. Next load
3962 * balancing owner will pick it up.
3963 */
3964 if (need_resched())
3965 break;
3966
Oleg Nesterovde0cf892007-08-12 18:08:19 +02003967 rebalance_domains(balance_cpu, CPU_IDLE);
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003968
3969 rq = cpu_rq(balance_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02003970 if (time_after(this_rq->next_balance, rq->next_balance))
3971 this_rq->next_balance = rq->next_balance;
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003972 }
3973 }
3974#endif
3975}
3976
3977/*
3978 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3979 *
3980 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3981 * idle load balancing owner or decide to stop the periodic load balancing,
3982 * if the whole system is idle.
3983 */
Ingo Molnardd41f592007-07-09 18:51:59 +02003984static inline void trigger_load_balance(struct rq *rq, int cpu)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003985{
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07003986#ifdef CONFIG_NO_HZ
3987 /*
3988 * If we were in the nohz mode recently and busy at the current
3989 * scheduler tick, then check if we need to nominate new idle
3990 * load balancer.
3991 */
3992 if (rq->in_nohz_recently && !rq->idle_at_tick) {
3993 rq->in_nohz_recently = 0;
3994
3995 if (atomic_read(&nohz.load_balancer) == cpu) {
3996 cpu_clear(cpu, nohz.cpu_mask);
3997 atomic_set(&nohz.load_balancer, -1);
3998 }
3999
4000 if (atomic_read(&nohz.load_balancer) == -1) {
4001 /*
4002 * simple selection for now: Nominate the
4003 * first cpu in the nohz list to be the next
4004 * ilb owner.
4005 *
4006 * TBD: Traverse the sched domains and nominate
4007 * the nearest cpu in the nohz.cpu_mask.
4008 */
4009 int ilb = first_cpu(nohz.cpu_mask);
4010
Mike Travis434d53b2008-04-04 18:11:04 -07004011 if (ilb < nr_cpu_ids)
Siddha, Suresh B46cb4b72007-05-08 00:32:51 -07004012 resched_cpu(ilb);
4013 }
4014 }
4015
4016 /*
4017 * If this cpu is idle and doing idle load balancing for all the
4018 * cpus with ticks stopped, is it time for that to stop?
4019 */
4020 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4021 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4022 resched_cpu(cpu);
4023 return;
4024 }
4025
4026 /*
4027 * If this cpu is idle and the idle load balancing is done by
4028 * someone else, then no need raise the SCHED_SOFTIRQ
4029 */
4030 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4031 cpu_isset(cpu, nohz.cpu_mask))
4032 return;
4033#endif
4034 if (time_after_eq(jiffies, rq->next_balance))
4035 raise_softirq(SCHED_SOFTIRQ);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004036}
Ingo Molnardd41f592007-07-09 18:51:59 +02004037
4038#else /* CONFIG_SMP */
4039
Linus Torvalds1da177e2005-04-16 15:20:36 -07004040/*
4041 * on UP we do not need to balance between CPUs:
4042 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07004043static inline void idle_balance(int cpu, struct rq *rq)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004044{
4045}
Ingo Molnardd41f592007-07-09 18:51:59 +02004046
Linus Torvalds1da177e2005-04-16 15:20:36 -07004047#endif
4048
Linus Torvalds1da177e2005-04-16 15:20:36 -07004049DEFINE_PER_CPU(struct kernel_stat, kstat);
4050
4051EXPORT_PER_CPU_SYMBOL(kstat);
4052
4053/*
Frank Mayharf06febc2008-09-12 09:54:39 -07004054 * Return any ns on the sched_clock that have not yet been banked in
4055 * @p in case that task is currently running.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004056 */
Frank Mayharbb34d922008-09-12 09:54:39 -07004057unsigned long long task_delta_exec(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004058{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004059 unsigned long flags;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004060 struct rq *rq;
Frank Mayharbb34d922008-09-12 09:54:39 -07004061 u64 ns = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07004062
Ingo Molnar41b86e92007-07-09 18:51:58 +02004063 rq = task_rq_lock(p, &flags);
Ingo Molnar15084872008-09-30 08:28:17 +02004064
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004065 if (task_current(rq, p)) {
Frank Mayharf06febc2008-09-12 09:54:39 -07004066 u64 delta_exec;
4067
Ingo Molnara8e504d2007-08-09 11:16:47 +02004068 update_rq_clock(rq);
4069 delta_exec = rq->clock - p->se.exec_start;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004070 if ((s64)delta_exec > 0)
Frank Mayharbb34d922008-09-12 09:54:39 -07004071 ns = delta_exec;
Ingo Molnar41b86e92007-07-09 18:51:58 +02004072 }
Ingo Molnar48f24c42006-07-03 00:25:40 -07004073
Linus Torvalds1da177e2005-04-16 15:20:36 -07004074 task_rq_unlock(rq, &flags);
4075
4076 return ns;
4077}
4078
4079/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004080 * Account user cpu time to a process.
4081 * @p: the process that the cpu time gets accounted to
Linus Torvalds1da177e2005-04-16 15:20:36 -07004082 * @cputime: the cpu time spent in user space since the last update
4083 */
4084void account_user_time(struct task_struct *p, cputime_t cputime)
4085{
4086 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4087 cputime64_t tmp;
4088
4089 p->utime = cputime_add(p->utime, cputime);
Frank Mayharf06febc2008-09-12 09:54:39 -07004090 account_group_user_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004091
4092 /* Add user time to cpustat. */
4093 tmp = cputime_to_cputime64(cputime);
4094 if (TASK_NICE(p) > 0)
4095 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4096 else
4097 cpustat->user = cputime64_add(cpustat->user, tmp);
Jonathan Lim49b5cf32008-07-25 01:48:40 -07004098 /* Account for user time used */
4099 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004100}
4101
4102/*
Laurent Vivier94886b82007-10-15 17:00:19 +02004103 * Account guest cpu time to a process.
4104 * @p: the process that the cpu time gets accounted to
4105 * @cputime: the cpu time spent in virtual machine since the last update
4106 */
Adrian Bunkf7402e02007-10-29 21:18:10 +01004107static void account_guest_time(struct task_struct *p, cputime_t cputime)
Laurent Vivier94886b82007-10-15 17:00:19 +02004108{
4109 cputime64_t tmp;
4110 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4111
4112 tmp = cputime_to_cputime64(cputime);
4113
4114 p->utime = cputime_add(p->utime, cputime);
Frank Mayharf06febc2008-09-12 09:54:39 -07004115 account_group_user_time(p, cputime);
Laurent Vivier94886b82007-10-15 17:00:19 +02004116 p->gtime = cputime_add(p->gtime, cputime);
4117
4118 cpustat->user = cputime64_add(cpustat->user, tmp);
4119 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4120}
4121
4122/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004123 * Account scaled user cpu time to a process.
4124 * @p: the process that the cpu time gets accounted to
4125 * @cputime: the cpu time spent in user space since the last update
4126 */
4127void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4128{
4129 p->utimescaled = cputime_add(p->utimescaled, cputime);
4130}
4131
4132/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004133 * Account system cpu time to a process.
4134 * @p: the process that the cpu time gets accounted to
4135 * @hardirq_offset: the offset to subtract from hardirq_count()
4136 * @cputime: the cpu time spent in kernel space since the last update
4137 */
4138void account_system_time(struct task_struct *p, int hardirq_offset,
4139 cputime_t cputime)
4140{
4141 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004142 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004143 cputime64_t tmp;
4144
Harvey Harrison983ed7a2008-04-24 18:17:55 -07004145 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4146 account_guest_time(p, cputime);
4147 return;
4148 }
Laurent Vivier94886b82007-10-15 17:00:19 +02004149
Linus Torvalds1da177e2005-04-16 15:20:36 -07004150 p->stime = cputime_add(p->stime, cputime);
Frank Mayharf06febc2008-09-12 09:54:39 -07004151 account_group_system_time(p, cputime);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004152
4153 /* Add system time to cpustat. */
4154 tmp = cputime_to_cputime64(cputime);
4155 if (hardirq_count() - hardirq_offset)
4156 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4157 else if (softirq_count())
4158 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004159 else if (p != rq->idle)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004160 cpustat->system = cputime64_add(cpustat->system, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004161 else if (atomic_read(&rq->nr_iowait) > 0)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004162 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4163 else
4164 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4165 /* Account for system time used */
4166 acct_update_integrals(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004167}
4168
4169/*
Michael Neulingc66f08b2007-10-18 03:06:34 -07004170 * Account scaled system cpu time to a process.
4171 * @p: the process that the cpu time gets accounted to
4172 * @hardirq_offset: the offset to subtract from hardirq_count()
4173 * @cputime: the cpu time spent in kernel space since the last update
4174 */
4175void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4176{
4177 p->stimescaled = cputime_add(p->stimescaled, cputime);
4178}
4179
4180/*
Linus Torvalds1da177e2005-04-16 15:20:36 -07004181 * Account for involuntary wait time.
4182 * @p: the process from which the cpu time has been stolen
4183 * @steal: the cpu time spent in involuntary wait
4184 */
4185void account_steal_time(struct task_struct *p, cputime_t steal)
4186{
4187 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4188 cputime64_t tmp = cputime_to_cputime64(steal);
Ingo Molnar70b97a72006-07-03 00:25:42 -07004189 struct rq *rq = this_rq();
Linus Torvalds1da177e2005-04-16 15:20:36 -07004190
4191 if (p == rq->idle) {
4192 p->stime = cputime_add(p->stime, steal);
Frank Mayharf06febc2008-09-12 09:54:39 -07004193 account_group_system_time(p, steal);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004194 if (atomic_read(&rq->nr_iowait) > 0)
4195 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4196 else
4197 cpustat->idle = cputime64_add(cpustat->idle, tmp);
Andrew Mortoncfb52852007-11-14 16:59:45 -08004198 } else
Linus Torvalds1da177e2005-04-16 15:20:36 -07004199 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4200}
4201
Christoph Lameter7835b982006-12-10 02:20:22 -08004202/*
Balbir Singh49048622008-09-05 18:12:23 +02004203 * Use precise platform statistics if available:
4204 */
4205#ifdef CONFIG_VIRT_CPU_ACCOUNTING
4206cputime_t task_utime(struct task_struct *p)
4207{
4208 return p->utime;
4209}
4210
4211cputime_t task_stime(struct task_struct *p)
4212{
4213 return p->stime;
4214}
4215#else
4216cputime_t task_utime(struct task_struct *p)
4217{
4218 clock_t utime = cputime_to_clock_t(p->utime),
4219 total = utime + cputime_to_clock_t(p->stime);
4220 u64 temp;
4221
4222 /*
4223 * Use CFS's precise accounting:
4224 */
4225 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4226
4227 if (total) {
4228 temp *= utime;
4229 do_div(temp, total);
4230 }
4231 utime = (clock_t)temp;
4232
4233 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4234 return p->prev_utime;
4235}
4236
4237cputime_t task_stime(struct task_struct *p)
4238{
4239 clock_t stime;
4240
4241 /*
4242 * Use CFS's precise accounting. (we subtract utime from
4243 * the total, to make sure the total observed by userspace
4244 * grows monotonically - apps rely on that):
4245 */
4246 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4247 cputime_to_clock_t(task_utime(p));
4248
4249 if (stime >= 0)
4250 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4251
4252 return p->prev_stime;
4253}
4254#endif
4255
4256inline cputime_t task_gtime(struct task_struct *p)
4257{
4258 return p->gtime;
4259}
4260
4261/*
Christoph Lameter7835b982006-12-10 02:20:22 -08004262 * This function gets called by the timer code, with HZ frequency.
4263 * We call it with interrupts disabled.
4264 *
4265 * It also gets called by the fork code, when changing the parent's
4266 * timeslices.
4267 */
4268void scheduler_tick(void)
4269{
Christoph Lameter7835b982006-12-10 02:20:22 -08004270 int cpu = smp_processor_id();
4271 struct rq *rq = cpu_rq(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02004272 struct task_struct *curr = rq->curr;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004273
4274 sched_clock_tick();
Christoph Lameter7835b982006-12-10 02:20:22 -08004275
Ingo Molnardd41f592007-07-09 18:51:59 +02004276 spin_lock(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004277 update_rq_clock(rq);
Ingo Molnarf1a438d2007-08-09 11:16:45 +02004278 update_cpu_load(rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01004279 curr->sched_class->task_tick(rq, curr, 0);
Ingo Molnardd41f592007-07-09 18:51:59 +02004280 spin_unlock(&rq->lock);
4281
Christoph Lametere418e1c2006-12-10 02:20:23 -08004282#ifdef CONFIG_SMP
Ingo Molnardd41f592007-07-09 18:51:59 +02004283 rq->idle_at_tick = idle_cpu(cpu);
4284 trigger_load_balance(rq, cpu);
Christoph Lametere418e1c2006-12-10 02:20:23 -08004285#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004286}
4287
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004288#if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4289 defined(CONFIG_PREEMPT_TRACER))
4290
4291static inline unsigned long get_parent_ip(unsigned long addr)
4292{
4293 if (in_lock_functions(addr)) {
4294 addr = CALLER_ADDR2;
4295 if (in_lock_functions(addr))
4296 addr = CALLER_ADDR3;
4297 }
4298 return addr;
4299}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004300
Srinivasa Ds43627582008-02-23 15:24:04 -08004301void __kprobes add_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004302{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004303#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004304 /*
4305 * Underflow?
4306 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004307 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4308 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004309#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07004310 preempt_count() += val;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004311#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004312 /*
4313 * Spinlock count overflowing soon?
4314 */
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08004315 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4316 PREEMPT_MASK - 10);
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004317#endif
4318 if (preempt_count() == val)
4319 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004320}
4321EXPORT_SYMBOL(add_preempt_count);
4322
Srinivasa Ds43627582008-02-23 15:24:04 -08004323void __kprobes sub_preempt_count(int val)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004324{
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004325#ifdef CONFIG_DEBUG_PREEMPT
Linus Torvalds1da177e2005-04-16 15:20:36 -07004326 /*
4327 * Underflow?
4328 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004329 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4330 return;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004331 /*
4332 * Is the spinlock portion underflowing?
4333 */
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004334 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4335 !(preempt_count() & PREEMPT_MASK)))
4336 return;
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004337#endif
Ingo Molnar9a11b49a2006-07-03 00:24:33 -07004338
Steven Rostedt6cd8a4b2008-05-12 21:20:42 +02004339 if (preempt_count() == val)
4340 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004341 preempt_count() -= val;
4342}
4343EXPORT_SYMBOL(sub_preempt_count);
4344
4345#endif
4346
4347/*
Ingo Molnardd41f592007-07-09 18:51:59 +02004348 * Print scheduling while atomic bug:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004349 */
Ingo Molnardd41f592007-07-09 18:51:59 +02004350static noinline void __schedule_bug(struct task_struct *prev)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004351{
Satyam Sharma838225b2007-10-24 18:23:50 +02004352 struct pt_regs *regs = get_irq_regs();
4353
4354 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4355 prev->comm, prev->pid, preempt_count());
4356
Ingo Molnardd41f592007-07-09 18:51:59 +02004357 debug_show_held_locks(prev);
Arjan van de Vene21f5b12008-05-23 09:05:58 -07004358 print_modules();
Ingo Molnardd41f592007-07-09 18:51:59 +02004359 if (irqs_disabled())
4360 print_irqtrace_events(prev);
Satyam Sharma838225b2007-10-24 18:23:50 +02004361
4362 if (regs)
4363 show_regs(regs);
4364 else
4365 dump_stack();
Ingo Molnardd41f592007-07-09 18:51:59 +02004366}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004367
Ingo Molnardd41f592007-07-09 18:51:59 +02004368/*
4369 * Various schedule()-time debugging checks and statistics:
4370 */
4371static inline void schedule_debug(struct task_struct *prev)
4372{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004373 /*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004374 * Test if we are atomic. Since do_exit() needs to call into
Linus Torvalds1da177e2005-04-16 15:20:36 -07004375 * schedule() atomically, we ignore that path for now.
4376 * Otherwise, whine if we are scheduling when we should not be.
4377 */
Roel Kluin3f33a7c2008-05-13 23:44:11 +02004378 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
Ingo Molnardd41f592007-07-09 18:51:59 +02004379 __schedule_bug(prev);
4380
Linus Torvalds1da177e2005-04-16 15:20:36 -07004381 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4382
Ingo Molnar2d723762007-10-15 17:00:12 +02004383 schedstat_inc(this_rq(), sched_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004384#ifdef CONFIG_SCHEDSTATS
4385 if (unlikely(prev->lock_depth >= 0)) {
Ingo Molnar2d723762007-10-15 17:00:12 +02004386 schedstat_inc(this_rq(), bkl_count);
4387 schedstat_inc(prev, sched_info.bkl_count);
Ingo Molnarb8efb562007-10-15 17:00:10 +02004388 }
4389#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02004390}
4391
4392/*
4393 * Pick up the highest-prio task:
4394 */
4395static inline struct task_struct *
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004396pick_next_task(struct rq *rq, struct task_struct *prev)
Ingo Molnardd41f592007-07-09 18:51:59 +02004397{
Ingo Molnar5522d5d2007-10-15 17:00:12 +02004398 const struct sched_class *class;
Ingo Molnardd41f592007-07-09 18:51:59 +02004399 struct task_struct *p;
4400
4401 /*
4402 * Optimization: we know that if all tasks are in
4403 * the fair class we can call that function directly:
4404 */
4405 if (likely(rq->nr_running == rq->cfs.nr_running)) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004406 p = fair_sched_class.pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004407 if (likely(p))
4408 return p;
4409 }
4410
4411 class = sched_class_highest;
4412 for ( ; ; ) {
Ingo Molnarfb8d4722007-08-09 11:16:48 +02004413 p = class->pick_next_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004414 if (p)
4415 return p;
4416 /*
4417 * Will never be NULL as the idle class always
4418 * returns a non-NULL p:
4419 */
4420 class = class->next;
4421 }
4422}
4423
4424/*
4425 * schedule() is the main scheduler function.
4426 */
4427asmlinkage void __sched schedule(void)
4428{
4429 struct task_struct *prev, *next;
Harvey Harrison67ca7bd2008-02-15 09:56:36 -08004430 unsigned long *switch_count;
Ingo Molnardd41f592007-07-09 18:51:59 +02004431 struct rq *rq;
Peter Zijlstra31656512008-07-18 18:01:23 +02004432 int cpu;
Ingo Molnardd41f592007-07-09 18:51:59 +02004433
Linus Torvalds1da177e2005-04-16 15:20:36 -07004434need_resched:
4435 preempt_disable();
Ingo Molnardd41f592007-07-09 18:51:59 +02004436 cpu = smp_processor_id();
4437 rq = cpu_rq(cpu);
4438 rcu_qsctr_inc(cpu);
4439 prev = rq->curr;
4440 switch_count = &prev->nivcsw;
4441
Linus Torvalds1da177e2005-04-16 15:20:36 -07004442 release_kernel_lock(prev);
4443need_resched_nonpreemptible:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004444
Ingo Molnardd41f592007-07-09 18:51:59 +02004445 schedule_debug(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004446
Peter Zijlstra31656512008-07-18 18:01:23 +02004447 if (sched_feat(HRTICK))
Mike Galbraithf333fdc2008-05-12 21:20:55 +02004448 hrtick_clear(rq);
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004449
Peter Zijlstra8cd162c2008-10-15 20:37:23 +02004450 spin_lock_irq(&rq->lock);
Peter Zijlstra3e51f332008-05-03 18:29:28 +02004451 update_rq_clock(rq);
Ingo Molnar1e819952007-10-15 17:00:13 +02004452 clear_tsk_need_resched(prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004453
Ingo Molnardd41f592007-07-09 18:51:59 +02004454 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
Oleg Nesterov16882c12008-06-08 21:20:41 +04004455 if (unlikely(signal_pending_state(prev->state, prev)))
Ingo Molnardd41f592007-07-09 18:51:59 +02004456 prev->state = TASK_RUNNING;
Oleg Nesterov16882c12008-06-08 21:20:41 +04004457 else
Ingo Molnar2e1cb742007-08-09 11:16:49 +02004458 deactivate_task(rq, prev, 1);
Ingo Molnardd41f592007-07-09 18:51:59 +02004459 switch_count = &prev->nvcsw;
4460 }
4461
Steven Rostedt9a897c52008-01-25 21:08:22 +01004462#ifdef CONFIG_SMP
4463 if (prev->sched_class->pre_schedule)
4464 prev->sched_class->pre_schedule(rq, prev);
4465#endif
Steven Rostedtf65eda42008-01-25 21:08:07 +01004466
Ingo Molnardd41f592007-07-09 18:51:59 +02004467 if (unlikely(!rq->nr_running))
4468 idle_balance(cpu, rq);
4469
Ingo Molnar31ee5292007-08-09 11:16:49 +02004470 prev->sched_class->put_prev_task(rq, prev);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02004471 next = pick_next_task(rq, prev);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004472
Linus Torvalds1da177e2005-04-16 15:20:36 -07004473 if (likely(prev != next)) {
David Simner673a90a2008-04-29 10:08:59 +01004474 sched_info_switch(prev, next);
4475
Linus Torvalds1da177e2005-04-16 15:20:36 -07004476 rq->nr_switches++;
4477 rq->curr = next;
4478 ++*switch_count;
4479
Ingo Molnardd41f592007-07-09 18:51:59 +02004480 context_switch(rq, prev, next); /* unlocks the rq */
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004481 /*
4482 * the context switch might have flipped the stack from under
4483 * us, hence refresh the local variables.
4484 */
4485 cpu = smp_processor_id();
4486 rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004487 } else
4488 spin_unlock_irq(&rq->lock);
4489
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004490 if (unlikely(reacquire_kernel_lock(current) < 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004491 goto need_resched_nonpreemptible;
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01004492
Linus Torvalds1da177e2005-04-16 15:20:36 -07004493 preempt_enable_no_resched();
4494 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4495 goto need_resched;
4496}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004497EXPORT_SYMBOL(schedule);
4498
4499#ifdef CONFIG_PREEMPT
4500/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004501 * this is the entry point to schedule() from in-kernel preemption
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004502 * off of preempt_enable. Kernel preemptions off return from interrupt
Linus Torvalds1da177e2005-04-16 15:20:36 -07004503 * occur there and call schedule directly.
4504 */
4505asmlinkage void __sched preempt_schedule(void)
4506{
4507 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004508
Linus Torvalds1da177e2005-04-16 15:20:36 -07004509 /*
4510 * If there is a non-zero preempt_count or interrupts are disabled,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004511 * we do not want to preempt the current task. Just return..
Linus Torvalds1da177e2005-04-16 15:20:36 -07004512 */
Nick Pigginbeed33a2006-10-11 01:21:52 -07004513 if (likely(ti->preempt_count || irqs_disabled()))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004514 return;
4515
Andi Kleen3a5c3592007-10-15 17:00:14 +02004516 do {
4517 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004518 schedule();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004519 sub_preempt_count(PREEMPT_ACTIVE);
4520
4521 /*
4522 * Check again in case we missed a preemption opportunity
4523 * between schedule and now.
4524 */
4525 barrier();
4526 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004527}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004528EXPORT_SYMBOL(preempt_schedule);
4529
4530/*
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004531 * this is the entry point to schedule() from kernel preemption
Linus Torvalds1da177e2005-04-16 15:20:36 -07004532 * off of irq context.
4533 * Note, that this is called and return with irqs disabled. This will
4534 * protect us against recursive calling from irq.
4535 */
4536asmlinkage void __sched preempt_schedule_irq(void)
4537{
4538 struct thread_info *ti = current_thread_info();
Ingo Molnar6478d882008-01-25 21:08:33 +01004539
Andreas Mohr2ed6e342006-07-10 04:43:52 -07004540 /* Catch callers which need to be fixed */
Linus Torvalds1da177e2005-04-16 15:20:36 -07004541 BUG_ON(ti->preempt_count || !irqs_disabled());
4542
Andi Kleen3a5c3592007-10-15 17:00:14 +02004543 do {
4544 add_preempt_count(PREEMPT_ACTIVE);
Andi Kleen3a5c3592007-10-15 17:00:14 +02004545 local_irq_enable();
4546 schedule();
4547 local_irq_disable();
Andi Kleen3a5c3592007-10-15 17:00:14 +02004548 sub_preempt_count(PREEMPT_ACTIVE);
4549
4550 /*
4551 * Check again in case we missed a preemption opportunity
4552 * between schedule and now.
4553 */
4554 barrier();
4555 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
Linus Torvalds1da177e2005-04-16 15:20:36 -07004556}
4557
4558#endif /* CONFIG_PREEMPT */
4559
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004560int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4561 void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004562{
Ingo Molnar48f24c42006-07-03 00:25:40 -07004563 return try_to_wake_up(curr->private, mode, sync);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004564}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004565EXPORT_SYMBOL(default_wake_function);
4566
4567/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004568 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4569 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
Linus Torvalds1da177e2005-04-16 15:20:36 -07004570 * number) then we wake all the non-exclusive tasks and one exclusive task.
4571 *
4572 * There are circumstances in which we can try to wake a task which has already
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01004573 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
Linus Torvalds1da177e2005-04-16 15:20:36 -07004574 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4575 */
4576static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4577 int nr_exclusive, int sync, void *key)
4578{
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004579 wait_queue_t *curr, *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004580
Matthias Kaehlcke2e458742007-10-15 17:00:02 +02004581 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
Ingo Molnar48f24c42006-07-03 00:25:40 -07004582 unsigned flags = curr->flags;
4583
Linus Torvalds1da177e2005-04-16 15:20:36 -07004584 if (curr->func(curr, mode, sync, key) &&
Ingo Molnar48f24c42006-07-03 00:25:40 -07004585 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004586 break;
4587 }
4588}
4589
4590/**
4591 * __wake_up - wake up threads blocked on a waitqueue.
4592 * @q: the waitqueue
4593 * @mode: which threads
4594 * @nr_exclusive: how many wake-one or wake-many threads to wake up
Martin Waitz67be2dd2005-05-01 08:59:26 -07004595 * @key: is directly passed to the wakeup function
Linus Torvalds1da177e2005-04-16 15:20:36 -07004596 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004597void __wake_up(wait_queue_head_t *q, unsigned int mode,
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004598 int nr_exclusive, void *key)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004599{
4600 unsigned long flags;
4601
4602 spin_lock_irqsave(&q->lock, flags);
4603 __wake_up_common(q, mode, nr_exclusive, 0, key);
4604 spin_unlock_irqrestore(&q->lock, flags);
4605}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004606EXPORT_SYMBOL(__wake_up);
4607
4608/*
4609 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4610 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004611void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004612{
4613 __wake_up_common(q, mode, 1, 0, NULL);
4614}
4615
4616/**
Martin Waitz67be2dd2005-05-01 08:59:26 -07004617 * __wake_up_sync - wake up threads blocked on a waitqueue.
Linus Torvalds1da177e2005-04-16 15:20:36 -07004618 * @q: the waitqueue
4619 * @mode: which threads
4620 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4621 *
4622 * The sync wakeup differs that the waker knows that it will schedule
4623 * away soon, so while the target thread will be woken up, it will not
4624 * be migrated to another CPU - ie. the two threads are 'synchronized'
4625 * with each other. This can prevent needless bouncing between CPUs.
4626 *
4627 * On UP it can prevent extra preemption.
4628 */
Harvey Harrison7ad5b3a2008-02-08 04:19:53 -08004629void
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004630__wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004631{
4632 unsigned long flags;
4633 int sync = 1;
4634
4635 if (unlikely(!q))
4636 return;
4637
4638 if (unlikely(!nr_exclusive))
4639 sync = 0;
4640
4641 spin_lock_irqsave(&q->lock, flags);
4642 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4643 spin_unlock_irqrestore(&q->lock, flags);
4644}
4645EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4646
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004647/**
4648 * complete: - signals a single thread waiting on this completion
4649 * @x: holds the state of this particular completion
4650 *
4651 * This will wake up a single thread waiting on this completion. Threads will be
4652 * awakened in the same order in which they were queued.
4653 *
4654 * See also complete_all(), wait_for_completion() and related routines.
4655 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02004656void complete(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004657{
4658 unsigned long flags;
4659
4660 spin_lock_irqsave(&x->wait.lock, flags);
4661 x->done++;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004662 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004663 spin_unlock_irqrestore(&x->wait.lock, flags);
4664}
4665EXPORT_SYMBOL(complete);
4666
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004667/**
4668 * complete_all: - signals all threads waiting on this completion
4669 * @x: holds the state of this particular completion
4670 *
4671 * This will wake up all threads waiting on this particular completion event.
4672 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02004673void complete_all(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004674{
4675 unsigned long flags;
4676
4677 spin_lock_irqsave(&x->wait.lock, flags);
4678 x->done += UINT_MAX/2;
Matthew Wilcoxd9514f62007-12-06 11:07:07 -05004679 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004680 spin_unlock_irqrestore(&x->wait.lock, flags);
4681}
4682EXPORT_SYMBOL(complete_all);
4683
Andi Kleen8cbbe862007-10-15 17:00:14 +02004684static inline long __sched
4685do_wait_for_common(struct completion *x, long timeout, int state)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004686{
Linus Torvalds1da177e2005-04-16 15:20:36 -07004687 if (!x->done) {
4688 DECLARE_WAITQUEUE(wait, current);
4689
4690 wait.flags |= WQ_FLAG_EXCLUSIVE;
4691 __add_wait_queue_tail(&x->wait, &wait);
4692 do {
Oleg Nesterov94d3d822008-08-20 16:54:41 -07004693 if (signal_pending_state(state, current)) {
Oleg Nesterovea71a542008-06-20 18:32:20 +04004694 timeout = -ERESTARTSYS;
4695 break;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004696 }
4697 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004698 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004699 timeout = schedule_timeout(timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004700 spin_lock_irq(&x->wait.lock);
Oleg Nesterovea71a542008-06-20 18:32:20 +04004701 } while (!x->done && timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004702 __remove_wait_queue(&x->wait, &wait);
Oleg Nesterovea71a542008-06-20 18:32:20 +04004703 if (!x->done)
4704 return timeout;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004705 }
4706 x->done--;
Oleg Nesterovea71a542008-06-20 18:32:20 +04004707 return timeout ?: 1;
Andi Kleen8cbbe862007-10-15 17:00:14 +02004708}
4709
4710static long __sched
4711wait_for_common(struct completion *x, long timeout, int state)
4712{
4713 might_sleep();
4714
4715 spin_lock_irq(&x->wait.lock);
4716 timeout = do_wait_for_common(x, timeout, state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004717 spin_unlock_irq(&x->wait.lock);
Andi Kleen8cbbe862007-10-15 17:00:14 +02004718 return timeout;
4719}
4720
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004721/**
4722 * wait_for_completion: - waits for completion of a task
4723 * @x: holds the state of this particular completion
4724 *
4725 * This waits to be signaled for completion of a specific task. It is NOT
4726 * interruptible and there is no timeout.
4727 *
4728 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4729 * and interrupt capability. Also see complete().
4730 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02004731void __sched wait_for_completion(struct completion *x)
Andi Kleen8cbbe862007-10-15 17:00:14 +02004732{
4733 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004734}
4735EXPORT_SYMBOL(wait_for_completion);
4736
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004737/**
4738 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4739 * @x: holds the state of this particular completion
4740 * @timeout: timeout value in jiffies
4741 *
4742 * This waits for either a completion of a specific task to be signaled or for a
4743 * specified timeout to expire. The timeout is in jiffies. It is not
4744 * interruptible.
4745 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02004746unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004747wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4748{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004749 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004750}
4751EXPORT_SYMBOL(wait_for_completion_timeout);
4752
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004753/**
4754 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4755 * @x: holds the state of this particular completion
4756 *
4757 * This waits for completion of a specific task to be signaled. It is
4758 * interruptible.
4759 */
Andi Kleen8cbbe862007-10-15 17:00:14 +02004760int __sched wait_for_completion_interruptible(struct completion *x)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004761{
Andi Kleen51e97992007-10-18 21:32:55 +02004762 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4763 if (t == -ERESTARTSYS)
4764 return t;
4765 return 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004766}
4767EXPORT_SYMBOL(wait_for_completion_interruptible);
4768
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004769/**
4770 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4771 * @x: holds the state of this particular completion
4772 * @timeout: timeout value in jiffies
4773 *
4774 * This waits for either a completion of a specific task to be signaled or for a
4775 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4776 */
Ingo Molnarb15136e2007-10-24 18:23:48 +02004777unsigned long __sched
Linus Torvalds1da177e2005-04-16 15:20:36 -07004778wait_for_completion_interruptible_timeout(struct completion *x,
4779 unsigned long timeout)
4780{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004781 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004782}
4783EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4784
Kevin Diggs65eb3dc2008-08-26 10:26:54 +02004785/**
4786 * wait_for_completion_killable: - waits for completion of a task (killable)
4787 * @x: holds the state of this particular completion
4788 *
4789 * This waits to be signaled for completion of a specific task. It can be
4790 * interrupted by a kill signal.
4791 */
Matthew Wilcox009e5772007-12-06 12:29:54 -05004792int __sched wait_for_completion_killable(struct completion *x)
4793{
4794 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4795 if (t == -ERESTARTSYS)
4796 return t;
4797 return 0;
4798}
4799EXPORT_SYMBOL(wait_for_completion_killable);
4800
Dave Chinnerbe4de352008-08-15 00:40:44 -07004801/**
4802 * try_wait_for_completion - try to decrement a completion without blocking
4803 * @x: completion structure
4804 *
4805 * Returns: 0 if a decrement cannot be done without blocking
4806 * 1 if a decrement succeeded.
4807 *
4808 * If a completion is being used as a counting completion,
4809 * attempt to decrement the counter without blocking. This
4810 * enables us to avoid waiting if the resource the completion
4811 * is protecting is not available.
4812 */
4813bool try_wait_for_completion(struct completion *x)
4814{
4815 int ret = 1;
4816
4817 spin_lock_irq(&x->wait.lock);
4818 if (!x->done)
4819 ret = 0;
4820 else
4821 x->done--;
4822 spin_unlock_irq(&x->wait.lock);
4823 return ret;
4824}
4825EXPORT_SYMBOL(try_wait_for_completion);
4826
4827/**
4828 * completion_done - Test to see if a completion has any waiters
4829 * @x: completion structure
4830 *
4831 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4832 * 1 if there are no waiters.
4833 *
4834 */
4835bool completion_done(struct completion *x)
4836{
4837 int ret = 1;
4838
4839 spin_lock_irq(&x->wait.lock);
4840 if (!x->done)
4841 ret = 0;
4842 spin_unlock_irq(&x->wait.lock);
4843 return ret;
4844}
4845EXPORT_SYMBOL(completion_done);
4846
Andi Kleen8cbbe862007-10-15 17:00:14 +02004847static long __sched
4848sleep_on_common(wait_queue_head_t *q, int state, long timeout)
Ingo Molnar0fec1712007-07-09 18:52:01 +02004849{
4850 unsigned long flags;
4851 wait_queue_t wait;
4852
4853 init_waitqueue_entry(&wait, current);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004854
Andi Kleen8cbbe862007-10-15 17:00:14 +02004855 __set_current_state(state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004856
Andi Kleen8cbbe862007-10-15 17:00:14 +02004857 spin_lock_irqsave(&q->lock, flags);
4858 __add_wait_queue(q, &wait);
4859 spin_unlock(&q->lock);
4860 timeout = schedule_timeout(timeout);
4861 spin_lock_irq(&q->lock);
4862 __remove_wait_queue(q, &wait);
4863 spin_unlock_irqrestore(&q->lock, flags);
4864
4865 return timeout;
4866}
4867
4868void __sched interruptible_sleep_on(wait_queue_head_t *q)
4869{
4870 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004871}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004872EXPORT_SYMBOL(interruptible_sleep_on);
4873
Ingo Molnar0fec1712007-07-09 18:52:01 +02004874long __sched
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07004875interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004876{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004877 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004878}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004879EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4880
Ingo Molnar0fec1712007-07-09 18:52:01 +02004881void __sched sleep_on(wait_queue_head_t *q)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004882{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004883 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004884}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004885EXPORT_SYMBOL(sleep_on);
4886
Ingo Molnar0fec1712007-07-09 18:52:01 +02004887long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004888{
Andi Kleen8cbbe862007-10-15 17:00:14 +02004889 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004890}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004891EXPORT_SYMBOL(sleep_on_timeout);
4892
Ingo Molnarb29739f2006-06-27 02:54:51 -07004893#ifdef CONFIG_RT_MUTEXES
4894
4895/*
4896 * rt_mutex_setprio - set the current priority of a task
4897 * @p: task
4898 * @prio: prio value (kernel-internal form)
4899 *
4900 * This function changes the 'effective' priority of a task. It does
4901 * not touch ->normal_prio like __setscheduler().
4902 *
4903 * Used by the rt_mutex code to implement priority inheritance logic.
4904 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004905void rt_mutex_setprio(struct task_struct *p, int prio)
Ingo Molnarb29739f2006-06-27 02:54:51 -07004906{
4907 unsigned long flags;
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02004908 int oldprio, on_rq, running;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004909 struct rq *rq;
Steven Rostedtcb469842008-01-25 21:08:22 +01004910 const struct sched_class *prev_class = p->sched_class;
Ingo Molnarb29739f2006-06-27 02:54:51 -07004911
4912 BUG_ON(prio < 0 || prio > MAX_PRIO);
4913
4914 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004915 update_rq_clock(rq);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004916
Andrew Mortond5f9f942007-05-08 20:27:06 -07004917 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02004918 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01004919 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004920 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004921 dequeue_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004922 if (running)
4923 p->sched_class->put_prev_task(rq, p);
Ingo Molnardd41f592007-07-09 18:51:59 +02004924
4925 if (rt_prio(prio))
4926 p->sched_class = &rt_sched_class;
4927 else
4928 p->sched_class = &fair_sched_class;
4929
Ingo Molnarb29739f2006-06-27 02:54:51 -07004930 p->prio = prio;
4931
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07004932 if (running)
4933 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02004934 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004935 enqueue_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01004936
4937 check_class_changed(rq, p, prev_class, oldprio, running);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004938 }
4939 task_rq_unlock(rq, &flags);
4940}
4941
4942#endif
4943
Ingo Molnar36c8b582006-07-03 00:25:41 -07004944void set_user_nice(struct task_struct *p, long nice)
Linus Torvalds1da177e2005-04-16 15:20:36 -07004945{
Ingo Molnardd41f592007-07-09 18:51:59 +02004946 int old_prio, delta, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004947 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07004948 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004949
4950 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4951 return;
4952 /*
4953 * We have to be careful, if called from sys_setpriority(),
4954 * the task might be in the middle of scheduling on another CPU.
4955 */
4956 rq = task_rq_lock(p, &flags);
Ingo Molnara8e504d2007-08-09 11:16:47 +02004957 update_rq_clock(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004958 /*
4959 * The RT priorities are set via sched_setscheduler(), but we still
4960 * allow the 'normal' nice value to be set - but as expected
4961 * it wont have any effect on scheduling until the task is
Ingo Molnardd41f592007-07-09 18:51:59 +02004962 * SCHED_FIFO/SCHED_RR:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004963 */
Ingo Molnare05606d2007-07-09 18:51:59 +02004964 if (task_has_rt_policy(p)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07004965 p->static_prio = NICE_TO_PRIO(nice);
4966 goto out_unlock;
4967 }
Ingo Molnardd41f592007-07-09 18:51:59 +02004968 on_rq = p->se.on_rq;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02004969 if (on_rq)
Ingo Molnar69be72c2007-08-09 11:16:49 +02004970 dequeue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004971
Linus Torvalds1da177e2005-04-16 15:20:36 -07004972 p->static_prio = NICE_TO_PRIO(nice);
Peter Williams2dd73a42006-06-27 02:54:34 -07004973 set_load_weight(p);
Ingo Molnarb29739f2006-06-27 02:54:51 -07004974 old_prio = p->prio;
4975 p->prio = effective_prio(p);
4976 delta = p->prio - old_prio;
Linus Torvalds1da177e2005-04-16 15:20:36 -07004977
Ingo Molnardd41f592007-07-09 18:51:59 +02004978 if (on_rq) {
Ingo Molnar8159f872007-08-09 11:16:49 +02004979 enqueue_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07004980 /*
Andrew Mortond5f9f942007-05-08 20:27:06 -07004981 * If the task increased its priority or is running and
4982 * lowered its priority, then reschedule its CPU:
Linus Torvalds1da177e2005-04-16 15:20:36 -07004983 */
Andrew Mortond5f9f942007-05-08 20:27:06 -07004984 if (delta < 0 || (delta > 0 && task_running(rq, p)))
Linus Torvalds1da177e2005-04-16 15:20:36 -07004985 resched_task(rq->curr);
4986 }
4987out_unlock:
4988 task_rq_unlock(rq, &flags);
4989}
Linus Torvalds1da177e2005-04-16 15:20:36 -07004990EXPORT_SYMBOL(set_user_nice);
4991
Matt Mackalle43379f2005-05-01 08:59:00 -07004992/*
4993 * can_nice - check if a task can reduce its nice value
4994 * @p: task
4995 * @nice: nice value
4996 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07004997int can_nice(const struct task_struct *p, const int nice)
Matt Mackalle43379f2005-05-01 08:59:00 -07004998{
Matt Mackall024f4742005-08-18 11:24:19 -07004999 /* convert nice value [19,-20] to rlimit style value [1,40] */
5000 int nice_rlim = 20 - nice;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005001
Matt Mackalle43379f2005-05-01 08:59:00 -07005002 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5003 capable(CAP_SYS_NICE));
5004}
5005
Linus Torvalds1da177e2005-04-16 15:20:36 -07005006#ifdef __ARCH_WANT_SYS_NICE
5007
5008/*
5009 * sys_nice - change the priority of the current process.
5010 * @increment: priority increment
5011 *
5012 * sys_setpriority is a more generic, but much slower function that
5013 * does similar things.
5014 */
5015asmlinkage long sys_nice(int increment)
5016{
Ingo Molnar48f24c42006-07-03 00:25:40 -07005017 long nice, retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005018
5019 /*
5020 * Setpriority might change our priority at the same moment.
5021 * We don't have to worry. Conceptually one call occurs first
5022 * and we have a single winner.
5023 */
Matt Mackalle43379f2005-05-01 08:59:00 -07005024 if (increment < -40)
5025 increment = -40;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005026 if (increment > 40)
5027 increment = 40;
5028
5029 nice = PRIO_TO_NICE(current->static_prio) + increment;
5030 if (nice < -20)
5031 nice = -20;
5032 if (nice > 19)
5033 nice = 19;
5034
Matt Mackalle43379f2005-05-01 08:59:00 -07005035 if (increment < 0 && !can_nice(current, nice))
5036 return -EPERM;
5037
Linus Torvalds1da177e2005-04-16 15:20:36 -07005038 retval = security_task_setnice(current, nice);
5039 if (retval)
5040 return retval;
5041
5042 set_user_nice(current, nice);
5043 return 0;
5044}
5045
5046#endif
5047
5048/**
5049 * task_prio - return the priority value of a given task.
5050 * @p: the task in question.
5051 *
5052 * This is the priority value as seen by users in /proc.
5053 * RT tasks are offset by -200. Normal tasks are centered
5054 * around 0, value goes from -16 to +15.
5055 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005056int task_prio(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005057{
5058 return p->prio - MAX_RT_PRIO;
5059}
5060
5061/**
5062 * task_nice - return the nice value of a given task.
5063 * @p: the task in question.
5064 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005065int task_nice(const struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005066{
5067 return TASK_NICE(p);
5068}
Pavel Roskin150d8be2008-03-05 16:56:37 -05005069EXPORT_SYMBOL(task_nice);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005070
5071/**
5072 * idle_cpu - is a given cpu idle currently?
5073 * @cpu: the processor in question.
5074 */
5075int idle_cpu(int cpu)
5076{
5077 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5078}
5079
Linus Torvalds1da177e2005-04-16 15:20:36 -07005080/**
5081 * idle_task - return the idle task for a given cpu.
5082 * @cpu: the processor in question.
5083 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07005084struct task_struct *idle_task(int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005085{
5086 return cpu_rq(cpu)->idle;
5087}
5088
5089/**
5090 * find_process_by_pid - find a process with a matching PID value.
5091 * @pid: the pid in question.
5092 */
Alexey Dobriyana9957442007-10-15 17:00:13 +02005093static struct task_struct *find_process_by_pid(pid_t pid)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005094{
Pavel Emelyanov228ebcb2007-10-18 23:40:16 -07005095 return pid ? find_task_by_vpid(pid) : current;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005096}
5097
5098/* Actually do priority change: must hold rq lock. */
Ingo Molnardd41f592007-07-09 18:51:59 +02005099static void
5100__setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005101{
Ingo Molnardd41f592007-07-09 18:51:59 +02005102 BUG_ON(p->se.on_rq);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005103
Linus Torvalds1da177e2005-04-16 15:20:36 -07005104 p->policy = policy;
Ingo Molnardd41f592007-07-09 18:51:59 +02005105 switch (p->policy) {
5106 case SCHED_NORMAL:
5107 case SCHED_BATCH:
5108 case SCHED_IDLE:
5109 p->sched_class = &fair_sched_class;
5110 break;
5111 case SCHED_FIFO:
5112 case SCHED_RR:
5113 p->sched_class = &rt_sched_class;
5114 break;
5115 }
5116
Linus Torvalds1da177e2005-04-16 15:20:36 -07005117 p->rt_priority = prio;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005118 p->normal_prio = normal_prio(p);
5119 /* we are holding p->pi_lock already */
5120 p->prio = rt_mutex_getprio(p);
Peter Williams2dd73a42006-06-27 02:54:34 -07005121 set_load_weight(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005122}
5123
Rusty Russell961ccdd2008-06-23 13:55:38 +10005124static int __sched_setscheduler(struct task_struct *p, int policy,
5125 struct sched_param *param, bool user)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005126{
Srivatsa Vaddagiri83b699e2007-10-15 17:00:08 +02005127 int retval, oldprio, oldpolicy = -1, on_rq, running;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005128 unsigned long flags;
Steven Rostedtcb469842008-01-25 21:08:22 +01005129 const struct sched_class *prev_class = p->sched_class;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005130 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005131
Steven Rostedt66e53932006-06-27 02:54:44 -07005132 /* may grab non-irq protected spin_locks */
5133 BUG_ON(in_interrupt());
Linus Torvalds1da177e2005-04-16 15:20:36 -07005134recheck:
5135 /* double check policy once rq lock held */
5136 if (policy < 0)
5137 policy = oldpolicy = p->policy;
5138 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
Ingo Molnardd41f592007-07-09 18:51:59 +02005139 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5140 policy != SCHED_IDLE)
Ingo Molnarb0a94992006-01-14 13:20:41 -08005141 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005142 /*
5143 * Valid priorities for SCHED_FIFO and SCHED_RR are
Ingo Molnardd41f592007-07-09 18:51:59 +02005144 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5145 * SCHED_BATCH and SCHED_IDLE is 0.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005146 */
5147 if (param->sched_priority < 0 ||
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005148 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
Steven Rostedtd46523e2005-07-25 16:28:39 -04005149 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005150 return -EINVAL;
Ingo Molnare05606d2007-07-09 18:51:59 +02005151 if (rt_policy(policy) != (param->sched_priority != 0))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005152 return -EINVAL;
5153
Olivier Croquette37e4ab32005-06-25 14:57:32 -07005154 /*
5155 * Allow unprivileged RT tasks to decrease priority:
5156 */
Rusty Russell961ccdd2008-06-23 13:55:38 +10005157 if (user && !capable(CAP_SYS_NICE)) {
Ingo Molnare05606d2007-07-09 18:51:59 +02005158 if (rt_policy(policy)) {
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005159 unsigned long rlim_rtprio;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005160
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005161 if (!lock_task_sighand(p, &flags))
5162 return -ESRCH;
5163 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5164 unlock_task_sighand(p, &flags);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005165
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005166 /* can't set/change the rt policy */
5167 if (policy != p->policy && !rlim_rtprio)
5168 return -EPERM;
5169
5170 /* can't increase priority */
5171 if (param->sched_priority > p->rt_priority &&
5172 param->sched_priority > rlim_rtprio)
5173 return -EPERM;
5174 }
Ingo Molnardd41f592007-07-09 18:51:59 +02005175 /*
5176 * Like positive nice levels, dont allow tasks to
5177 * move out of SCHED_IDLE either:
5178 */
5179 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5180 return -EPERM;
Oleg Nesterov8dc3e902006-09-29 02:00:50 -07005181
Olivier Croquette37e4ab32005-06-25 14:57:32 -07005182 /* can't change other user's priorities */
5183 if ((current->euid != p->euid) &&
5184 (current->euid != p->uid))
5185 return -EPERM;
5186 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005187
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07005188 if (user) {
Peter Zijlstrab68aa232008-02-13 15:45:40 +01005189#ifdef CONFIG_RT_GROUP_SCHED
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07005190 /*
5191 * Do not allow realtime tasks into groups that have no runtime
5192 * assigned.
5193 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02005194 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5195 task_group(p)->rt_bandwidth.rt_runtime == 0)
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07005196 return -EPERM;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01005197#endif
5198
Jeremy Fitzhardinge725aad22008-08-03 09:33:03 -07005199 retval = security_task_setscheduler(p, policy, param);
5200 if (retval)
5201 return retval;
5202 }
5203
Linus Torvalds1da177e2005-04-16 15:20:36 -07005204 /*
Ingo Molnarb29739f2006-06-27 02:54:51 -07005205 * make sure no PI-waiters arrive (or leave) while we are
5206 * changing the priority of the task:
5207 */
5208 spin_lock_irqsave(&p->pi_lock, flags);
5209 /*
Linus Torvalds1da177e2005-04-16 15:20:36 -07005210 * To be able to change p->policy safely, the apropriate
5211 * runqueue lock must be held.
5212 */
Ingo Molnarb29739f2006-06-27 02:54:51 -07005213 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005214 /* recheck policy now with rq lock held */
5215 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5216 policy = oldpolicy = -1;
Ingo Molnarb29739f2006-06-27 02:54:51 -07005217 __task_rq_unlock(rq);
5218 spin_unlock_irqrestore(&p->pi_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005219 goto recheck;
5220 }
Ingo Molnar2daa3572007-08-09 11:16:51 +02005221 update_rq_clock(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005222 on_rq = p->se.on_rq;
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01005223 running = task_current(rq, p);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005224 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02005225 deactivate_task(rq, p, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005226 if (running)
5227 p->sched_class->put_prev_task(rq, p);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02005228
Linus Torvalds1da177e2005-04-16 15:20:36 -07005229 oldprio = p->prio;
Ingo Molnardd41f592007-07-09 18:51:59 +02005230 __setscheduler(rq, p, policy, param->sched_priority);
Dmitry Adamushkof6b53202007-10-15 17:00:08 +02005231
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07005232 if (running)
5233 p->sched_class->set_curr_task(rq);
Ingo Molnardd41f592007-07-09 18:51:59 +02005234 if (on_rq) {
5235 activate_task(rq, p, 0);
Steven Rostedtcb469842008-01-25 21:08:22 +01005236
5237 check_class_changed(rq, p, prev_class, oldprio, running);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005238 }
Ingo Molnarb29739f2006-06-27 02:54:51 -07005239 __task_rq_unlock(rq);
5240 spin_unlock_irqrestore(&p->pi_lock, flags);
5241
Thomas Gleixner95e02ca2006-06-27 02:55:02 -07005242 rt_mutex_adjust_pi(p);
5243
Linus Torvalds1da177e2005-04-16 15:20:36 -07005244 return 0;
5245}
Rusty Russell961ccdd2008-06-23 13:55:38 +10005246
5247/**
5248 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5249 * @p: the task in question.
5250 * @policy: new policy.
5251 * @param: structure containing the new RT priority.
5252 *
5253 * NOTE that the task may be already dead.
5254 */
5255int sched_setscheduler(struct task_struct *p, int policy,
5256 struct sched_param *param)
5257{
5258 return __sched_setscheduler(p, policy, param, true);
5259}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005260EXPORT_SYMBOL_GPL(sched_setscheduler);
5261
Rusty Russell961ccdd2008-06-23 13:55:38 +10005262/**
5263 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5264 * @p: the task in question.
5265 * @policy: new policy.
5266 * @param: structure containing the new RT priority.
5267 *
5268 * Just like sched_setscheduler, only don't bother checking if the
5269 * current context has permission. For example, this is needed in
5270 * stop_machine(): we create temporary high priority worker threads,
5271 * but our caller might not have that capability.
5272 */
5273int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5274 struct sched_param *param)
5275{
5276 return __sched_setscheduler(p, policy, param, false);
5277}
5278
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005279static int
5280do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005281{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005282 struct sched_param lparam;
5283 struct task_struct *p;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005284 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005285
5286 if (!param || pid < 0)
5287 return -EINVAL;
5288 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5289 return -EFAULT;
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005290
5291 rcu_read_lock();
5292 retval = -ESRCH;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005293 p = find_process_by_pid(pid);
Oleg Nesterov5fe1d752006-09-29 02:00:48 -07005294 if (p != NULL)
5295 retval = sched_setscheduler(p, policy, &lparam);
5296 rcu_read_unlock();
Ingo Molnar36c8b582006-07-03 00:25:41 -07005297
Linus Torvalds1da177e2005-04-16 15:20:36 -07005298 return retval;
5299}
5300
5301/**
5302 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5303 * @pid: the pid in question.
5304 * @policy: new policy.
5305 * @param: structure containing the new RT priority.
5306 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005307asmlinkage long
5308sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005309{
Jason Baronc21761f2006-01-18 17:43:03 -08005310 /* negative values for policy are not valid */
5311 if (policy < 0)
5312 return -EINVAL;
5313
Linus Torvalds1da177e2005-04-16 15:20:36 -07005314 return do_sched_setscheduler(pid, policy, param);
5315}
5316
5317/**
5318 * sys_sched_setparam - set/change the RT priority of a thread
5319 * @pid: the pid in question.
5320 * @param: structure containing the new RT priority.
5321 */
5322asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5323{
5324 return do_sched_setscheduler(pid, -1, param);
5325}
5326
5327/**
5328 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5329 * @pid: the pid in question.
5330 */
5331asmlinkage long sys_sched_getscheduler(pid_t pid)
5332{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005333 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005334 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005335
5336 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005337 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005338
5339 retval = -ESRCH;
5340 read_lock(&tasklist_lock);
5341 p = find_process_by_pid(pid);
5342 if (p) {
5343 retval = security_task_getscheduler(p);
5344 if (!retval)
5345 retval = p->policy;
5346 }
5347 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005348 return retval;
5349}
5350
5351/**
5352 * sys_sched_getscheduler - get the RT priority of a thread
5353 * @pid: the pid in question.
5354 * @param: structure containing the RT priority.
5355 */
5356asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5357{
5358 struct sched_param lp;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005359 struct task_struct *p;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005360 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005361
5362 if (!param || pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005363 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005364
5365 read_lock(&tasklist_lock);
5366 p = find_process_by_pid(pid);
5367 retval = -ESRCH;
5368 if (!p)
5369 goto out_unlock;
5370
5371 retval = security_task_getscheduler(p);
5372 if (retval)
5373 goto out_unlock;
5374
5375 lp.sched_priority = p->rt_priority;
5376 read_unlock(&tasklist_lock);
5377
5378 /*
5379 * This one might sleep, we cannot do it with a spinlock held ...
5380 */
5381 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5382
Linus Torvalds1da177e2005-04-16 15:20:36 -07005383 return retval;
5384
5385out_unlock:
5386 read_unlock(&tasklist_lock);
5387 return retval;
5388}
5389
Mike Travisb53e9212008-04-04 18:11:08 -07005390long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005391{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005392 cpumask_t cpus_allowed;
Mike Travisb53e9212008-04-04 18:11:08 -07005393 cpumask_t new_mask = *in_mask;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005394 struct task_struct *p;
5395 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005396
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005397 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005398 read_lock(&tasklist_lock);
5399
5400 p = find_process_by_pid(pid);
5401 if (!p) {
5402 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005403 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005404 return -ESRCH;
5405 }
5406
5407 /*
5408 * It is not safe to call set_cpus_allowed with the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005409 * tasklist_lock held. We will bump the task_struct's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005410 * usage count and then drop tasklist_lock.
5411 */
5412 get_task_struct(p);
5413 read_unlock(&tasklist_lock);
5414
5415 retval = -EPERM;
5416 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5417 !capable(CAP_SYS_NICE))
5418 goto out_unlock;
5419
David Quigleye7834f82006-06-23 02:03:59 -07005420 retval = security_task_setscheduler(p, 0, NULL);
5421 if (retval)
5422 goto out_unlock;
5423
Mike Travisf9a86fc2008-04-04 18:11:07 -07005424 cpuset_cpus_allowed(p, &cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005425 cpus_and(new_mask, new_mask, cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005426 again:
Mike Travis7c16ec52008-04-04 18:11:11 -07005427 retval = set_cpus_allowed_ptr(p, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005428
Paul Menage8707d8b2007-10-18 23:40:22 -07005429 if (!retval) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07005430 cpuset_cpus_allowed(p, &cpus_allowed);
Paul Menage8707d8b2007-10-18 23:40:22 -07005431 if (!cpus_subset(new_mask, cpus_allowed)) {
5432 /*
5433 * We must have raced with a concurrent cpuset
5434 * update. Just reset the cpus_allowed to the
5435 * cpuset's cpus_allowed
5436 */
5437 new_mask = cpus_allowed;
5438 goto again;
5439 }
5440 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005441out_unlock:
5442 put_task_struct(p);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005443 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005444 return retval;
5445}
5446
5447static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5448 cpumask_t *new_mask)
5449{
5450 if (len < sizeof(cpumask_t)) {
5451 memset(new_mask, 0, sizeof(cpumask_t));
5452 } else if (len > sizeof(cpumask_t)) {
5453 len = sizeof(cpumask_t);
5454 }
5455 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5456}
5457
5458/**
5459 * sys_sched_setaffinity - set the cpu affinity of a process
5460 * @pid: pid of the process
5461 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5462 * @user_mask_ptr: user-space pointer to the new cpu mask
5463 */
5464asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5465 unsigned long __user *user_mask_ptr)
5466{
5467 cpumask_t new_mask;
5468 int retval;
5469
5470 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5471 if (retval)
5472 return retval;
5473
Mike Travisb53e9212008-04-04 18:11:08 -07005474 return sched_setaffinity(pid, &new_mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005475}
5476
Linus Torvalds1da177e2005-04-16 15:20:36 -07005477long sched_getaffinity(pid_t pid, cpumask_t *mask)
5478{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005479 struct task_struct *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005480 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005481
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005482 get_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005483 read_lock(&tasklist_lock);
5484
5485 retval = -ESRCH;
5486 p = find_process_by_pid(pid);
5487 if (!p)
5488 goto out_unlock;
5489
David Quigleye7834f82006-06-23 02:03:59 -07005490 retval = security_task_getscheduler(p);
5491 if (retval)
5492 goto out_unlock;
5493
Jack Steiner2f7016d2006-02-01 03:05:18 -08005494 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005495
5496out_unlock:
5497 read_unlock(&tasklist_lock);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01005498 put_online_cpus();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005499
Ulrich Drepper9531b622007-08-09 11:16:46 +02005500 return retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005501}
5502
5503/**
5504 * sys_sched_getaffinity - get the cpu affinity of a process
5505 * @pid: pid of the process
5506 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5507 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5508 */
5509asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5510 unsigned long __user *user_mask_ptr)
5511{
5512 int ret;
5513 cpumask_t mask;
5514
5515 if (len < sizeof(cpumask_t))
5516 return -EINVAL;
5517
5518 ret = sched_getaffinity(pid, &mask);
5519 if (ret < 0)
5520 return ret;
5521
5522 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5523 return -EFAULT;
5524
5525 return sizeof(cpumask_t);
5526}
5527
5528/**
5529 * sys_sched_yield - yield the current processor to other threads.
5530 *
Ingo Molnardd41f592007-07-09 18:51:59 +02005531 * This function yields the current CPU to other tasks. If there are no
5532 * other threads running on this CPU then this function will return.
Linus Torvalds1da177e2005-04-16 15:20:36 -07005533 */
5534asmlinkage long sys_sched_yield(void)
5535{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005536 struct rq *rq = this_rq_lock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005537
Ingo Molnar2d723762007-10-15 17:00:12 +02005538 schedstat_inc(rq, yld_count);
Dmitry Adamushko4530d7a2007-10-15 17:00:08 +02005539 current->sched_class->yield_task(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005540
5541 /*
5542 * Since we are going to call schedule() anyway, there's
5543 * no need to preempt or enable interrupts:
5544 */
5545 __release(rq->lock);
Ingo Molnar8a25d5d2006-07-03 00:24:54 -07005546 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005547 _raw_spin_unlock(&rq->lock);
5548 preempt_enable_no_resched();
5549
5550 schedule();
5551
5552 return 0;
5553}
5554
Andrew Mortone7b38402006-06-30 01:56:00 -07005555static void __cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005556{
Ingo Molnar8e0a43d2006-06-23 02:05:23 -07005557#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5558 __might_sleep(__FILE__, __LINE__);
5559#endif
Ingo Molnar5bbcfd92005-07-07 17:57:04 -07005560 /*
5561 * The BKS might be reacquired before we have dropped
5562 * PREEMPT_ACTIVE, which could trigger a second
5563 * cond_resched() call.
5564 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07005565 do {
5566 add_preempt_count(PREEMPT_ACTIVE);
5567 schedule();
5568 sub_preempt_count(PREEMPT_ACTIVE);
5569 } while (need_resched());
5570}
5571
Herbert Xu02b67cc32008-01-25 21:08:28 +01005572int __sched _cond_resched(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005573{
Ingo Molnar94142322006-12-29 16:48:13 -08005574 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5575 system_state == SYSTEM_RUNNING) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005576 __cond_resched();
5577 return 1;
5578 }
5579 return 0;
5580}
Herbert Xu02b67cc32008-01-25 21:08:28 +01005581EXPORT_SYMBOL(_cond_resched);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005582
5583/*
5584 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5585 * call schedule, and on return reacquire the lock.
5586 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005587 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
Linus Torvalds1da177e2005-04-16 15:20:36 -07005588 * operations here to prevent schedule() from being called twice (once via
5589 * spin_unlock(), once by hand).
5590 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07005591int cond_resched_lock(spinlock_t *lock)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005592{
Nick Piggin95c354f2008-01-30 13:31:20 +01005593 int resched = need_resched() && system_state == SYSTEM_RUNNING;
Jan Kara6df3cec2005-06-13 15:52:32 -07005594 int ret = 0;
5595
Nick Piggin95c354f2008-01-30 13:31:20 +01005596 if (spin_needbreak(lock) || resched) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005597 spin_unlock(lock);
Nick Piggin95c354f2008-01-30 13:31:20 +01005598 if (resched && need_resched())
5599 __cond_resched();
5600 else
5601 cpu_relax();
Jan Kara6df3cec2005-06-13 15:52:32 -07005602 ret = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005603 spin_lock(lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005604 }
Jan Kara6df3cec2005-06-13 15:52:32 -07005605 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005606}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005607EXPORT_SYMBOL(cond_resched_lock);
5608
5609int __sched cond_resched_softirq(void)
5610{
5611 BUG_ON(!in_softirq());
5612
Ingo Molnar94142322006-12-29 16:48:13 -08005613 if (need_resched() && system_state == SYSTEM_RUNNING) {
Thomas Gleixner98d825672007-05-23 13:58:18 -07005614 local_bh_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005615 __cond_resched();
5616 local_bh_disable();
5617 return 1;
5618 }
5619 return 0;
5620}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005621EXPORT_SYMBOL(cond_resched_softirq);
5622
Linus Torvalds1da177e2005-04-16 15:20:36 -07005623/**
5624 * yield - yield the current processor to other threads.
5625 *
Robert P. J. Day72fd4a32007-02-10 01:45:59 -08005626 * This is a shortcut for kernel-space yielding - it marks the
Linus Torvalds1da177e2005-04-16 15:20:36 -07005627 * thread runnable and calls sys_sched_yield().
5628 */
5629void __sched yield(void)
5630{
5631 set_current_state(TASK_RUNNING);
5632 sys_sched_yield();
5633}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005634EXPORT_SYMBOL(yield);
5635
5636/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005637 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
Linus Torvalds1da177e2005-04-16 15:20:36 -07005638 * that process accounting knows that this is a task in IO wait state.
5639 *
5640 * But don't do that if it is a deliberate, throttling IO wait (this task
5641 * has set its backing_dev_info: the queue against which it should throttle)
5642 */
5643void __sched io_schedule(void)
5644{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005645 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005646
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005647 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005648 atomic_inc(&rq->nr_iowait);
5649 schedule();
5650 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005651 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005652}
Linus Torvalds1da177e2005-04-16 15:20:36 -07005653EXPORT_SYMBOL(io_schedule);
5654
5655long __sched io_schedule_timeout(long timeout)
5656{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005657 struct rq *rq = &__raw_get_cpu_var(runqueues);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005658 long ret;
5659
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005660 delayacct_blkio_start();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005661 atomic_inc(&rq->nr_iowait);
5662 ret = schedule_timeout(timeout);
5663 atomic_dec(&rq->nr_iowait);
Shailabh Nagar0ff92242006-07-14 00:24:37 -07005664 delayacct_blkio_end();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005665 return ret;
5666}
5667
5668/**
5669 * sys_sched_get_priority_max - return maximum RT priority.
5670 * @policy: scheduling class.
5671 *
5672 * this syscall returns the maximum rt_priority that can be used
5673 * by a given scheduling class.
5674 */
5675asmlinkage long sys_sched_get_priority_max(int policy)
5676{
5677 int ret = -EINVAL;
5678
5679 switch (policy) {
5680 case SCHED_FIFO:
5681 case SCHED_RR:
5682 ret = MAX_USER_RT_PRIO-1;
5683 break;
5684 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005685 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005686 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005687 ret = 0;
5688 break;
5689 }
5690 return ret;
5691}
5692
5693/**
5694 * sys_sched_get_priority_min - return minimum RT priority.
5695 * @policy: scheduling class.
5696 *
5697 * this syscall returns the minimum rt_priority that can be used
5698 * by a given scheduling class.
5699 */
5700asmlinkage long sys_sched_get_priority_min(int policy)
5701{
5702 int ret = -EINVAL;
5703
5704 switch (policy) {
5705 case SCHED_FIFO:
5706 case SCHED_RR:
5707 ret = 1;
5708 break;
5709 case SCHED_NORMAL:
Ingo Molnarb0a94992006-01-14 13:20:41 -08005710 case SCHED_BATCH:
Ingo Molnardd41f592007-07-09 18:51:59 +02005711 case SCHED_IDLE:
Linus Torvalds1da177e2005-04-16 15:20:36 -07005712 ret = 0;
5713 }
5714 return ret;
5715}
5716
5717/**
5718 * sys_sched_rr_get_interval - return the default timeslice of a process.
5719 * @pid: pid of the process.
5720 * @interval: userspace pointer to the timeslice value.
5721 *
5722 * this syscall writes the default timeslice value of a given process
5723 * into the user-space timespec buffer. A value of '0' means infinity.
5724 */
5725asmlinkage
5726long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5727{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005728 struct task_struct *p;
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005729 unsigned int time_slice;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005730 int retval;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005731 struct timespec t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005732
5733 if (pid < 0)
Andi Kleen3a5c3592007-10-15 17:00:14 +02005734 return -EINVAL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005735
5736 retval = -ESRCH;
5737 read_lock(&tasklist_lock);
5738 p = find_process_by_pid(pid);
5739 if (!p)
5740 goto out_unlock;
5741
5742 retval = security_task_getscheduler(p);
5743 if (retval)
5744 goto out_unlock;
5745
Ingo Molnar77034932007-12-04 17:04:39 +01005746 /*
5747 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5748 * tasks that are on an otherwise idle runqueue:
5749 */
5750 time_slice = 0;
5751 if (p->policy == SCHED_RR) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005752 time_slice = DEF_TIMESLICE;
Miao Xie1868f952008-03-07 09:35:06 +08005753 } else if (p->policy != SCHED_FIFO) {
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005754 struct sched_entity *se = &p->se;
5755 unsigned long flags;
5756 struct rq *rq;
5757
5758 rq = task_rq_lock(p, &flags);
Ingo Molnar77034932007-12-04 17:04:39 +01005759 if (rq->cfs.load.weight)
5760 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005761 task_rq_unlock(rq, &flags);
5762 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07005763 read_unlock(&tasklist_lock);
Dmitry Adamushkoa4ec24b2007-10-15 17:00:13 +02005764 jiffies_to_timespec(time_slice, &t);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005765 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005766 return retval;
Andi Kleen3a5c3592007-10-15 17:00:14 +02005767
Linus Torvalds1da177e2005-04-16 15:20:36 -07005768out_unlock:
5769 read_unlock(&tasklist_lock);
5770 return retval;
5771}
5772
Steven Rostedt7c731e02008-05-12 21:20:41 +02005773static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005774
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005775void sched_show_task(struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005776{
Linus Torvalds1da177e2005-04-16 15:20:36 -07005777 unsigned long free = 0;
Ingo Molnar36c8b582006-07-03 00:25:41 -07005778 unsigned state;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005779
Linus Torvalds1da177e2005-04-16 15:20:36 -07005780 state = p->state ? __ffs(p->state) + 1 : 0;
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005781 printk(KERN_INFO "%-13.13s %c", p->comm,
Andreas Mohr2ed6e342006-07-10 04:43:52 -07005782 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
Ingo Molnar4bd77322007-07-11 21:21:47 +02005783#if BITS_PER_LONG == 32
Linus Torvalds1da177e2005-04-16 15:20:36 -07005784 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005785 printk(KERN_CONT " running ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005786 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005787 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005788#else
5789 if (state == TASK_RUNNING)
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005790 printk(KERN_CONT " running task ");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005791 else
Ingo Molnarcc4ea792007-10-18 21:32:56 +02005792 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005793#endif
5794#ifdef CONFIG_DEBUG_STACK_USAGE
5795 {
Al Viro10ebffd2005-11-13 16:06:56 -08005796 unsigned long *n = end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005797 while (!*n)
5798 n++;
Al Viro10ebffd2005-11-13 16:06:56 -08005799 free = (unsigned long)n - (unsigned long)end_of_stack(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005800 }
5801#endif
Pavel Emelyanovba25f9d2007-10-18 23:40:40 -07005802 printk(KERN_CONT "%5lu %5d %6d\n", free,
Roland McGrathfcfd50a2008-01-09 00:03:23 -08005803 task_pid_nr(p), task_pid_nr(p->real_parent));
Linus Torvalds1da177e2005-04-16 15:20:36 -07005804
Nick Piggin5fb5e6d2008-01-25 21:08:34 +01005805 show_stack(p, NULL);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005806}
5807
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005808void show_state_filter(unsigned long state_filter)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005809{
Ingo Molnar36c8b582006-07-03 00:25:41 -07005810 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005811
Ingo Molnar4bd77322007-07-11 21:21:47 +02005812#if BITS_PER_LONG == 32
5813 printk(KERN_INFO
5814 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005815#else
Ingo Molnar4bd77322007-07-11 21:21:47 +02005816 printk(KERN_INFO
5817 " task PC stack pid father\n");
Linus Torvalds1da177e2005-04-16 15:20:36 -07005818#endif
5819 read_lock(&tasklist_lock);
5820 do_each_thread(g, p) {
5821 /*
5822 * reset the NMI-timeout, listing all files on a slow
5823 * console might take alot of time:
5824 */
5825 touch_nmi_watchdog();
Ingo Molnar39bc89f2007-04-25 20:50:03 -07005826 if (!state_filter || (p->state & state_filter))
Ingo Molnar82a1fcb2008-01-25 21:08:02 +01005827 sched_show_task(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005828 } while_each_thread(g, p);
5829
Jeremy Fitzhardinge04c91672007-05-08 00:28:05 -07005830 touch_all_softlockup_watchdogs();
5831
Ingo Molnardd41f592007-07-09 18:51:59 +02005832#ifdef CONFIG_SCHED_DEBUG
5833 sysrq_sched_debug_show();
5834#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005835 read_unlock(&tasklist_lock);
Ingo Molnare59e2ae2006-12-06 20:35:59 -08005836 /*
5837 * Only show locks if all tasks are dumped:
5838 */
5839 if (state_filter == -1)
5840 debug_show_all_locks();
Linus Torvalds1da177e2005-04-16 15:20:36 -07005841}
5842
Ingo Molnar1df21052007-07-09 18:51:58 +02005843void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5844{
Ingo Molnardd41f592007-07-09 18:51:59 +02005845 idle->sched_class = &idle_sched_class;
Ingo Molnar1df21052007-07-09 18:51:58 +02005846}
5847
Ingo Molnarf340c0d2005-06-28 16:40:42 +02005848/**
5849 * init_idle - set up an idle thread for a given CPU
5850 * @idle: task in question
5851 * @cpu: cpu the idle task belongs to
5852 *
5853 * NOTE: this function does not set the idle thread's NEED_RESCHED
5854 * flag, to make booting more robust.
5855 */
Nick Piggin5c1e1762006-10-03 01:14:04 -07005856void __cpuinit init_idle(struct task_struct *idle, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005857{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005858 struct rq *rq = cpu_rq(cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005859 unsigned long flags;
5860
Ingo Molnardd41f592007-07-09 18:51:59 +02005861 __sched_fork(idle);
5862 idle->se.exec_start = sched_clock();
5863
Ingo Molnarb29739f2006-06-27 02:54:51 -07005864 idle->prio = idle->normal_prio = MAX_PRIO;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005865 idle->cpus_allowed = cpumask_of_cpu(cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02005866 __set_task_cpu(idle, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005867
5868 spin_lock_irqsave(&rq->lock, flags);
5869 rq->curr = rq->idle = idle;
Nick Piggin4866cde2005-06-25 14:57:23 -07005870#if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5871 idle->oncpu = 1;
5872#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07005873 spin_unlock_irqrestore(&rq->lock, flags);
5874
5875 /* Set the preempt count _outside_ the spinlocks! */
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005876#if defined(CONFIG_PREEMPT)
5877 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5878#else
Al Viroa1261f52005-11-13 16:06:55 -08005879 task_thread_info(idle)->preempt_count = 0;
Linus Torvalds8e3e0762008-05-10 20:58:02 -07005880#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02005881 /*
5882 * The idle tasks have their own, simple scheduling class:
5883 */
5884 idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005885}
5886
5887/*
5888 * In a system that switches off the HZ timer nohz_cpu_mask
5889 * indicates which cpus entered this state. This is used
5890 * in the rcu update to wait only for active cpus. For system
5891 * which do not switch off the HZ timer nohz_cpu_mask should
5892 * always be CPU_MASK_NONE.
5893 */
5894cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5895
Ingo Molnar19978ca2007-11-09 22:39:38 +01005896/*
5897 * Increase the granularity value when there are more CPUs,
5898 * because with more CPUs the 'effective latency' as visible
5899 * to users decreases. But the relationship is not linear,
5900 * so pick a second-best guess by going with the log2 of the
5901 * number of CPUs.
5902 *
5903 * This idea comes from the SD scheduler of Con Kolivas:
5904 */
5905static inline void sched_init_granularity(void)
5906{
5907 unsigned int factor = 1 + ilog2(num_online_cpus());
5908 const unsigned long limit = 200000000;
5909
5910 sysctl_sched_min_granularity *= factor;
5911 if (sysctl_sched_min_granularity > limit)
5912 sysctl_sched_min_granularity = limit;
5913
5914 sysctl_sched_latency *= factor;
5915 if (sysctl_sched_latency > limit)
5916 sysctl_sched_latency = limit;
5917
5918 sysctl_sched_wakeup_granularity *= factor;
Peter Zijlstra55cd5342008-08-04 08:54:26 +02005919
5920 sysctl_sched_shares_ratelimit *= factor;
Ingo Molnar19978ca2007-11-09 22:39:38 +01005921}
5922
Linus Torvalds1da177e2005-04-16 15:20:36 -07005923#ifdef CONFIG_SMP
5924/*
5925 * This is how migration works:
5926 *
Ingo Molnar70b97a72006-07-03 00:25:42 -07005927 * 1) we queue a struct migration_req structure in the source CPU's
Linus Torvalds1da177e2005-04-16 15:20:36 -07005928 * runqueue and wake up that CPU's migration thread.
5929 * 2) we down() the locked semaphore => thread blocks.
5930 * 3) migration thread wakes up (implicitly it forces the migrated
5931 * thread off the CPU)
5932 * 4) it gets the migration request and checks whether the migrated
5933 * task is still in the wrong runqueue.
5934 * 5) if it's in the wrong runqueue then the migration thread removes
5935 * it and puts it into the right queue.
5936 * 6) migration thread up()s the semaphore.
5937 * 7) we wake up and the migration is done.
5938 */
5939
5940/*
5941 * Change a given task's CPU affinity. Migrate the thread to a
5942 * proper CPU and schedule it away if the CPU it's executing on
5943 * is removed from the allowed bitmask.
5944 *
5945 * NOTE: the caller must have a valid reference to the task, the
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005946 * task must not exit() & deallocate itself prematurely. The
Linus Torvalds1da177e2005-04-16 15:20:36 -07005947 * call is not atomic; no spinlocks may be held.
5948 */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005949int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07005950{
Ingo Molnar70b97a72006-07-03 00:25:42 -07005951 struct migration_req req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005952 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07005953 struct rq *rq;
Ingo Molnar48f24c42006-07-03 00:25:40 -07005954 int ret = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07005955
5956 rq = task_rq_lock(p, &flags);
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005957 if (!cpus_intersects(*new_mask, cpu_online_map)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005958 ret = -EINVAL;
5959 goto out;
5960 }
5961
David Rientjes9985b0b2008-06-05 12:57:11 -07005962 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5963 !cpus_equal(p->cpus_allowed, *new_mask))) {
5964 ret = -EINVAL;
5965 goto out;
5966 }
5967
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005968 if (p->sched_class->set_cpus_allowed)
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005969 p->sched_class->set_cpus_allowed(p, new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005970 else {
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005971 p->cpus_allowed = *new_mask;
5972 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
Gregory Haskins73fe6aa2008-01-25 21:08:07 +01005973 }
5974
Linus Torvalds1da177e2005-04-16 15:20:36 -07005975 /* Can the task run on the task's current CPU? If so, we're done */
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005976 if (cpu_isset(task_cpu(p), *new_mask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07005977 goto out;
5978
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005979 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07005980 /* Need help from migration thread: drop lock and wait. */
5981 task_rq_unlock(rq, &flags);
5982 wake_up_process(rq->migration_thread);
5983 wait_for_completion(&req.done);
5984 tlb_migrate_finish(p->mm);
5985 return 0;
5986 }
5987out:
5988 task_rq_unlock(rq, &flags);
Ingo Molnar48f24c42006-07-03 00:25:40 -07005989
Linus Torvalds1da177e2005-04-16 15:20:36 -07005990 return ret;
5991}
Mike Traviscd8ba7c2008-03-26 14:23:49 -07005992EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
Linus Torvalds1da177e2005-04-16 15:20:36 -07005993
5994/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01005995 * Move (not current) task off this cpu, onto dest cpu. We're doing
Linus Torvalds1da177e2005-04-16 15:20:36 -07005996 * this because either it can't run here any more (set_cpus_allowed()
5997 * away from this CPU, or CPU going down), or because we're
5998 * attempting to rebalance this task on exec (sched_exec).
5999 *
6000 * So we race with normal scheduler movements, but that's OK, as long
6001 * as the task is no longer on this CPU.
Kirill Korotaevefc30812006-06-27 02:54:32 -07006002 *
6003 * Returns non-zero if task was successfully migrated.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006004 */
Kirill Korotaevefc30812006-06-27 02:54:32 -07006005static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006006{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006007 struct rq *rq_dest, *rq_src;
Ingo Molnardd41f592007-07-09 18:51:59 +02006008 int ret = 0, on_rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006009
Max Krasnyanskye761b772008-07-15 04:43:49 -07006010 if (unlikely(!cpu_active(dest_cpu)))
Kirill Korotaevefc30812006-06-27 02:54:32 -07006011 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006012
6013 rq_src = cpu_rq(src_cpu);
6014 rq_dest = cpu_rq(dest_cpu);
6015
6016 double_rq_lock(rq_src, rq_dest);
6017 /* Already moved. */
6018 if (task_cpu(p) != src_cpu)
Linus Torvaldsb1e38732008-07-10 11:25:03 -07006019 goto done;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006020 /* Affinity changed (again). */
6021 if (!cpu_isset(dest_cpu, p->cpus_allowed))
Linus Torvaldsb1e38732008-07-10 11:25:03 -07006022 goto fail;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006023
Ingo Molnardd41f592007-07-09 18:51:59 +02006024 on_rq = p->se.on_rq;
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02006025 if (on_rq)
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006026 deactivate_task(rq_src, p, 0);
Ingo Molnar6e82a3b2007-08-09 11:16:51 +02006027
Linus Torvalds1da177e2005-04-16 15:20:36 -07006028 set_task_cpu(p, dest_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02006029 if (on_rq) {
6030 activate_task(rq_dest, p, 0);
Peter Zijlstra15afe092008-09-20 23:38:02 +02006031 check_preempt_curr(rq_dest, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006032 }
Linus Torvaldsb1e38732008-07-10 11:25:03 -07006033done:
Kirill Korotaevefc30812006-06-27 02:54:32 -07006034 ret = 1;
Linus Torvaldsb1e38732008-07-10 11:25:03 -07006035fail:
Linus Torvalds1da177e2005-04-16 15:20:36 -07006036 double_rq_unlock(rq_src, rq_dest);
Kirill Korotaevefc30812006-06-27 02:54:32 -07006037 return ret;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006038}
6039
6040/*
6041 * migration_thread - this is a highprio system thread that performs
6042 * thread migration by bumping thread off CPU then 'pushing' onto
6043 * another runqueue.
6044 */
Ingo Molnar95cdf3b2005-09-10 00:26:11 -07006045static int migration_thread(void *data)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006046{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006047 int cpu = (long)data;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006048 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006049
6050 rq = cpu_rq(cpu);
6051 BUG_ON(rq->migration_thread != current);
6052
6053 set_current_state(TASK_INTERRUPTIBLE);
6054 while (!kthread_should_stop()) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006055 struct migration_req *req;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006056 struct list_head *head;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006057
Linus Torvalds1da177e2005-04-16 15:20:36 -07006058 spin_lock_irq(&rq->lock);
6059
6060 if (cpu_is_offline(cpu)) {
6061 spin_unlock_irq(&rq->lock);
6062 goto wait_to_die;
6063 }
6064
6065 if (rq->active_balance) {
6066 active_load_balance(rq, cpu);
6067 rq->active_balance = 0;
6068 }
6069
6070 head = &rq->migration_queue;
6071
6072 if (list_empty(head)) {
6073 spin_unlock_irq(&rq->lock);
6074 schedule();
6075 set_current_state(TASK_INTERRUPTIBLE);
6076 continue;
6077 }
Ingo Molnar70b97a72006-07-03 00:25:42 -07006078 req = list_entry(head->next, struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006079 list_del_init(head->next);
6080
Nick Piggin674311d2005-06-25 14:57:27 -07006081 spin_unlock(&rq->lock);
6082 __migrate_task(req->task, cpu, req->dest_cpu);
6083 local_irq_enable();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006084
6085 complete(&req->done);
6086 }
6087 __set_current_state(TASK_RUNNING);
6088 return 0;
6089
6090wait_to_die:
6091 /* Wait for kthread_stop */
6092 set_current_state(TASK_INTERRUPTIBLE);
6093 while (!kthread_should_stop()) {
6094 schedule();
6095 set_current_state(TASK_INTERRUPTIBLE);
6096 }
6097 __set_current_state(TASK_RUNNING);
6098 return 0;
6099}
6100
6101#ifdef CONFIG_HOTPLUG_CPU
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006102
6103static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6104{
6105 int ret;
6106
6107 local_irq_disable();
6108 ret = __migrate_task(p, src_cpu, dest_cpu);
6109 local_irq_enable();
6110 return ret;
6111}
6112
Kirill Korotaev054b9102006-12-10 02:20:11 -08006113/*
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006114 * Figure out where task on dead CPU should go, use force if necessary.
Kirill Korotaev054b9102006-12-10 02:20:11 -08006115 * NOTE: interrupts should be disabled by the caller
6116 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006117static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006118{
Kirill Korotaevefc30812006-06-27 02:54:32 -07006119 unsigned long flags;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006120 cpumask_t mask;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006121 struct rq *rq;
6122 int dest_cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006123
Andi Kleen3a5c3592007-10-15 17:00:14 +02006124 do {
6125 /* On same node? */
6126 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6127 cpus_and(mask, mask, p->cpus_allowed);
6128 dest_cpu = any_online_cpu(mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006129
Andi Kleen3a5c3592007-10-15 17:00:14 +02006130 /* On any allowed CPU? */
Mike Travis434d53b2008-04-04 18:11:04 -07006131 if (dest_cpu >= nr_cpu_ids)
Andi Kleen3a5c3592007-10-15 17:00:14 +02006132 dest_cpu = any_online_cpu(p->cpus_allowed);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006133
Andi Kleen3a5c3592007-10-15 17:00:14 +02006134 /* No more Mr. Nice Guy. */
Mike Travis434d53b2008-04-04 18:11:04 -07006135 if (dest_cpu >= nr_cpu_ids) {
Mike Travisf9a86fc2008-04-04 18:11:07 -07006136 cpumask_t cpus_allowed;
6137
6138 cpuset_cpus_allowed_locked(p, &cpus_allowed);
Cliff Wickman470fd642007-10-18 23:40:46 -07006139 /*
6140 * Try to stay on the same cpuset, where the
6141 * current cpuset may be a subset of all cpus.
6142 * The cpuset_cpus_allowed_locked() variant of
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006143 * cpuset_cpus_allowed() will not block. It must be
Cliff Wickman470fd642007-10-18 23:40:46 -07006144 * called within calls to cpuset_lock/cpuset_unlock.
6145 */
Andi Kleen3a5c3592007-10-15 17:00:14 +02006146 rq = task_rq_lock(p, &flags);
Cliff Wickman470fd642007-10-18 23:40:46 -07006147 p->cpus_allowed = cpus_allowed;
Andi Kleen3a5c3592007-10-15 17:00:14 +02006148 dest_cpu = any_online_cpu(p->cpus_allowed);
6149 task_rq_unlock(rq, &flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006150
Andi Kleen3a5c3592007-10-15 17:00:14 +02006151 /*
6152 * Don't tell them about moving exiting tasks or
6153 * kernel threads (both mm NULL), since they never
6154 * leave kernel.
6155 */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006156 if (p->mm && printk_ratelimit()) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02006157 printk(KERN_INFO "process %d (%s) no "
6158 "longer affine to cpu%d\n",
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006159 task_pid_nr(p), p->comm, dead_cpu);
6160 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02006161 }
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006162 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006163}
6164
6165/*
6166 * While a dead CPU has no uninterruptible tasks queued at this point,
6167 * it might still have a nonzero ->nr_uninterruptible counter, because
6168 * for performance reasons the counter is not stricly tracking tasks to
6169 * their home CPUs. So we just add the counter to another CPU's counter,
6170 * to keep the global sum constant after CPU-down:
6171 */
Ingo Molnar70b97a72006-07-03 00:25:42 -07006172static void migrate_nr_uninterruptible(struct rq *rq_src)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006173{
Mike Travis7c16ec52008-04-04 18:11:11 -07006174 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006175 unsigned long flags;
6176
6177 local_irq_save(flags);
6178 double_rq_lock(rq_src, rq_dest);
6179 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6180 rq_src->nr_uninterruptible = 0;
6181 double_rq_unlock(rq_src, rq_dest);
6182 local_irq_restore(flags);
6183}
6184
6185/* Run through task list and migrate tasks from the dead cpu. */
6186static void migrate_live_tasks(int src_cpu)
6187{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006188 struct task_struct *p, *t;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006189
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006190 read_lock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006191
Ingo Molnar48f24c42006-07-03 00:25:40 -07006192 do_each_thread(t, p) {
6193 if (p == current)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006194 continue;
6195
Ingo Molnar48f24c42006-07-03 00:25:40 -07006196 if (task_cpu(p) == src_cpu)
6197 move_task_off_dead_cpu(src_cpu, p);
6198 } while_each_thread(t, p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006199
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006200 read_unlock(&tasklist_lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006201}
6202
Ingo Molnardd41f592007-07-09 18:51:59 +02006203/*
6204 * Schedules idle task to be the next runnable task on current CPU.
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01006205 * It does so by boosting its priority to highest possible.
6206 * Used by CPU offline code.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006207 */
6208void sched_idle_next(void)
6209{
Ingo Molnar48f24c42006-07-03 00:25:40 -07006210 int this_cpu = smp_processor_id();
Ingo Molnar70b97a72006-07-03 00:25:42 -07006211 struct rq *rq = cpu_rq(this_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006212 struct task_struct *p = rq->idle;
6213 unsigned long flags;
6214
6215 /* cpu has to be offline */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006216 BUG_ON(cpu_online(this_cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006217
Ingo Molnar48f24c42006-07-03 00:25:40 -07006218 /*
6219 * Strictly not necessary since rest of the CPUs are stopped by now
6220 * and interrupts disabled on the current cpu.
Linus Torvalds1da177e2005-04-16 15:20:36 -07006221 */
6222 spin_lock_irqsave(&rq->lock, flags);
6223
Ingo Molnardd41f592007-07-09 18:51:59 +02006224 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006225
Dmitry Adamushko94bc9a72007-11-15 20:57:40 +01006226 update_rq_clock(rq);
6227 activate_task(rq, p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006228
6229 spin_unlock_irqrestore(&rq->lock, flags);
6230}
6231
Ingo Molnar48f24c42006-07-03 00:25:40 -07006232/*
6233 * Ensures that the idle task is using init_mm right before its cpu goes
Linus Torvalds1da177e2005-04-16 15:20:36 -07006234 * offline.
6235 */
6236void idle_task_exit(void)
6237{
6238 struct mm_struct *mm = current->active_mm;
6239
6240 BUG_ON(cpu_online(smp_processor_id()));
6241
6242 if (mm != &init_mm)
6243 switch_mm(mm, &init_mm, current);
6244 mmdrop(mm);
6245}
6246
Kirill Korotaev054b9102006-12-10 02:20:11 -08006247/* called under rq->lock with disabled interrupts */
Ingo Molnar36c8b582006-07-03 00:25:41 -07006248static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006249{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006250 struct rq *rq = cpu_rq(dead_cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006251
6252 /* Must be exiting, otherwise would be on tasklist. */
Eugene Teo270f7222007-10-18 23:40:38 -07006253 BUG_ON(!p->exit_state);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006254
6255 /* Cannot have done final schedule yet: would have vanished. */
Oleg Nesterovc394cc92006-09-29 02:01:11 -07006256 BUG_ON(p->state == TASK_DEAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006257
Ingo Molnar48f24c42006-07-03 00:25:40 -07006258 get_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006259
6260 /*
6261 * Drop lock around migration; if someone else moves it,
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006262 * that's OK. No task can be added to this CPU, so iteration is
Linus Torvalds1da177e2005-04-16 15:20:36 -07006263 * fine.
6264 */
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006265 spin_unlock_irq(&rq->lock);
Ingo Molnar48f24c42006-07-03 00:25:40 -07006266 move_task_off_dead_cpu(dead_cpu, p);
Oleg Nesterovf7b4cdd2007-10-16 23:30:56 -07006267 spin_lock_irq(&rq->lock);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006268
Ingo Molnar48f24c42006-07-03 00:25:40 -07006269 put_task_struct(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006270}
6271
6272/* release_task() removes task from tasklist, so we won't find dead tasks. */
6273static void migrate_dead_tasks(unsigned int dead_cpu)
6274{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006275 struct rq *rq = cpu_rq(dead_cpu);
Ingo Molnardd41f592007-07-09 18:51:59 +02006276 struct task_struct *next;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006277
Ingo Molnardd41f592007-07-09 18:51:59 +02006278 for ( ; ; ) {
6279 if (!rq->nr_running)
6280 break;
Ingo Molnara8e504d2007-08-09 11:16:47 +02006281 update_rq_clock(rq);
Ingo Molnarff95f3d2007-08-09 11:16:49 +02006282 next = pick_next_task(rq, rq->curr);
Ingo Molnardd41f592007-07-09 18:51:59 +02006283 if (!next)
6284 break;
Dmitry Adamushko79c53792008-06-29 00:16:56 +02006285 next->sched_class->put_prev_task(rq, next);
Ingo Molnardd41f592007-07-09 18:51:59 +02006286 migrate_dead(dead_cpu, next);
Nick Piggine692ab52007-07-26 13:40:43 +02006287
Linus Torvalds1da177e2005-04-16 15:20:36 -07006288 }
6289}
6290#endif /* CONFIG_HOTPLUG_CPU */
6291
Nick Piggine692ab52007-07-26 13:40:43 +02006292#if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6293
6294static struct ctl_table sd_ctl_dir[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006295 {
6296 .procname = "sched_domain",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006297 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006298 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006299 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006300};
6301
6302static struct ctl_table sd_ctl_root[] = {
Alexey Dobriyane0361852007-08-09 11:16:46 +02006303 {
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006304 .ctl_name = CTL_KERN,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006305 .procname = "kernel",
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006306 .mode = 0555,
Alexey Dobriyane0361852007-08-09 11:16:46 +02006307 .child = sd_ctl_dir,
6308 },
Ingo Molnar38605ca2007-10-29 21:18:11 +01006309 {0, },
Nick Piggine692ab52007-07-26 13:40:43 +02006310};
6311
6312static struct ctl_table *sd_alloc_ctl_entry(int n)
6313{
6314 struct ctl_table *entry =
Milton Miller5cf9f062007-10-15 17:00:19 +02006315 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
Nick Piggine692ab52007-07-26 13:40:43 +02006316
Nick Piggine692ab52007-07-26 13:40:43 +02006317 return entry;
6318}
6319
Milton Miller6382bc92007-10-15 17:00:19 +02006320static void sd_free_ctl_entry(struct ctl_table **tablep)
6321{
Milton Millercd7900762007-10-17 16:55:11 +02006322 struct ctl_table *entry;
Milton Miller6382bc92007-10-15 17:00:19 +02006323
Milton Millercd7900762007-10-17 16:55:11 +02006324 /*
6325 * In the intermediate directories, both the child directory and
6326 * procname are dynamically allocated and could fail but the mode
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006327 * will always be set. In the lowest directory the names are
Milton Millercd7900762007-10-17 16:55:11 +02006328 * static strings and all have proc handlers.
6329 */
6330 for (entry = *tablep; entry->mode; entry++) {
Milton Miller6382bc92007-10-15 17:00:19 +02006331 if (entry->child)
6332 sd_free_ctl_entry(&entry->child);
Milton Millercd7900762007-10-17 16:55:11 +02006333 if (entry->proc_handler == NULL)
6334 kfree(entry->procname);
6335 }
Milton Miller6382bc92007-10-15 17:00:19 +02006336
6337 kfree(*tablep);
6338 *tablep = NULL;
6339}
6340
Nick Piggine692ab52007-07-26 13:40:43 +02006341static void
Alexey Dobriyane0361852007-08-09 11:16:46 +02006342set_table_entry(struct ctl_table *entry,
Nick Piggine692ab52007-07-26 13:40:43 +02006343 const char *procname, void *data, int maxlen,
6344 mode_t mode, proc_handler *proc_handler)
6345{
Nick Piggine692ab52007-07-26 13:40:43 +02006346 entry->procname = procname;
6347 entry->data = data;
6348 entry->maxlen = maxlen;
6349 entry->mode = mode;
6350 entry->proc_handler = proc_handler;
6351}
6352
6353static struct ctl_table *
6354sd_alloc_ctl_domain_table(struct sched_domain *sd)
6355{
Ingo Molnara5d8c342008-10-09 11:35:51 +02006356 struct ctl_table *table = sd_alloc_ctl_entry(13);
Nick Piggine692ab52007-07-26 13:40:43 +02006357
Milton Millerad1cdc12007-10-15 17:00:19 +02006358 if (table == NULL)
6359 return NULL;
6360
Alexey Dobriyane0361852007-08-09 11:16:46 +02006361 set_table_entry(&table[0], "min_interval", &sd->min_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006362 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006363 set_table_entry(&table[1], "max_interval", &sd->max_interval,
Nick Piggine692ab52007-07-26 13:40:43 +02006364 sizeof(long), 0644, proc_doulongvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006365 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006366 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006367 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006368 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006369 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006370 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006371 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006372 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006373 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
Nick Piggine692ab52007-07-26 13:40:43 +02006374 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006375 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
Nick Piggine692ab52007-07-26 13:40:43 +02006376 sizeof(int), 0644, proc_dointvec_minmax);
Alexey Dobriyane0361852007-08-09 11:16:46 +02006377 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
Nick Piggine692ab52007-07-26 13:40:43 +02006378 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006379 set_table_entry(&table[9], "cache_nice_tries",
Nick Piggine692ab52007-07-26 13:40:43 +02006380 &sd->cache_nice_tries,
6381 sizeof(int), 0644, proc_dointvec_minmax);
Zou Nan haiace8b3d2007-10-15 17:00:14 +02006382 set_table_entry(&table[10], "flags", &sd->flags,
Nick Piggine692ab52007-07-26 13:40:43 +02006383 sizeof(int), 0644, proc_dointvec_minmax);
Ingo Molnara5d8c342008-10-09 11:35:51 +02006384 set_table_entry(&table[11], "name", sd->name,
6385 CORENAME_MAX_SIZE, 0444, proc_dostring);
6386 /* &table[12] is terminator */
Nick Piggine692ab52007-07-26 13:40:43 +02006387
6388 return table;
6389}
6390
Ingo Molnar9a4e7152007-11-28 15:52:56 +01006391static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
Nick Piggine692ab52007-07-26 13:40:43 +02006392{
6393 struct ctl_table *entry, *table;
6394 struct sched_domain *sd;
6395 int domain_num = 0, i;
6396 char buf[32];
6397
6398 for_each_domain(cpu, sd)
6399 domain_num++;
6400 entry = table = sd_alloc_ctl_entry(domain_num + 1);
Milton Millerad1cdc12007-10-15 17:00:19 +02006401 if (table == NULL)
6402 return NULL;
Nick Piggine692ab52007-07-26 13:40:43 +02006403
6404 i = 0;
6405 for_each_domain(cpu, sd) {
6406 snprintf(buf, 32, "domain%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006407 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006408 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006409 entry->child = sd_alloc_ctl_domain_table(sd);
6410 entry++;
6411 i++;
6412 }
6413 return table;
6414}
6415
6416static struct ctl_table_header *sd_sysctl_header;
Milton Miller6382bc92007-10-15 17:00:19 +02006417static void register_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006418{
6419 int i, cpu_num = num_online_cpus();
6420 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6421 char buf[32];
6422
Milton Miller73785472007-10-24 18:23:48 +02006423 WARN_ON(sd_ctl_dir[0].child);
6424 sd_ctl_dir[0].child = entry;
6425
Milton Millerad1cdc12007-10-15 17:00:19 +02006426 if (entry == NULL)
6427 return;
6428
Milton Miller97b6ea72007-10-15 17:00:19 +02006429 for_each_online_cpu(i) {
Nick Piggine692ab52007-07-26 13:40:43 +02006430 snprintf(buf, 32, "cpu%d", i);
Nick Piggine692ab52007-07-26 13:40:43 +02006431 entry->procname = kstrdup(buf, GFP_KERNEL);
Eric W. Biedermanc57baf12007-08-23 15:18:02 +02006432 entry->mode = 0555;
Nick Piggine692ab52007-07-26 13:40:43 +02006433 entry->child = sd_alloc_ctl_cpu_table(i);
Milton Miller97b6ea72007-10-15 17:00:19 +02006434 entry++;
Nick Piggine692ab52007-07-26 13:40:43 +02006435 }
Milton Miller73785472007-10-24 18:23:48 +02006436
6437 WARN_ON(sd_sysctl_header);
Nick Piggine692ab52007-07-26 13:40:43 +02006438 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6439}
Milton Miller6382bc92007-10-15 17:00:19 +02006440
Milton Miller73785472007-10-24 18:23:48 +02006441/* may be called multiple times per register */
Milton Miller6382bc92007-10-15 17:00:19 +02006442static void unregister_sched_domain_sysctl(void)
6443{
Milton Miller73785472007-10-24 18:23:48 +02006444 if (sd_sysctl_header)
6445 unregister_sysctl_table(sd_sysctl_header);
Milton Miller6382bc92007-10-15 17:00:19 +02006446 sd_sysctl_header = NULL;
Milton Miller73785472007-10-24 18:23:48 +02006447 if (sd_ctl_dir[0].child)
6448 sd_free_ctl_entry(&sd_ctl_dir[0].child);
Milton Miller6382bc92007-10-15 17:00:19 +02006449}
Nick Piggine692ab52007-07-26 13:40:43 +02006450#else
Milton Miller6382bc92007-10-15 17:00:19 +02006451static void register_sched_domain_sysctl(void)
6452{
6453}
6454static void unregister_sched_domain_sysctl(void)
Nick Piggine692ab52007-07-26 13:40:43 +02006455{
6456}
6457#endif
6458
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006459static void set_rq_online(struct rq *rq)
6460{
6461 if (!rq->online) {
6462 const struct sched_class *class;
6463
6464 cpu_set(rq->cpu, rq->rd->online);
6465 rq->online = 1;
6466
6467 for_each_class(class) {
6468 if (class->rq_online)
6469 class->rq_online(rq);
6470 }
6471 }
6472}
6473
6474static void set_rq_offline(struct rq *rq)
6475{
6476 if (rq->online) {
6477 const struct sched_class *class;
6478
6479 for_each_class(class) {
6480 if (class->rq_offline)
6481 class->rq_offline(rq);
6482 }
6483
6484 cpu_clear(rq->cpu, rq->rd->online);
6485 rq->online = 0;
6486 }
6487}
6488
Linus Torvalds1da177e2005-04-16 15:20:36 -07006489/*
6490 * migration_call - callback that gets triggered when a CPU is added.
6491 * Here we can start up the necessary migration thread for the new CPU.
6492 */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006493static int __cpuinit
6494migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006495{
Linus Torvalds1da177e2005-04-16 15:20:36 -07006496 struct task_struct *p;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006497 int cpu = (long)hcpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006498 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07006499 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006500
6501 switch (action) {
Gautham R Shenoy5be93612007-05-09 02:34:04 -07006502
Linus Torvalds1da177e2005-04-16 15:20:36 -07006503 case CPU_UP_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006504 case CPU_UP_PREPARE_FROZEN:
Ingo Molnardd41f592007-07-09 18:51:59 +02006505 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006506 if (IS_ERR(p))
6507 return NOTIFY_BAD;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006508 kthread_bind(p, cpu);
6509 /* Must be high prio: stop_machine expects to yield to it. */
6510 rq = task_rq_lock(p, &flags);
Ingo Molnardd41f592007-07-09 18:51:59 +02006511 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006512 task_rq_unlock(rq, &flags);
6513 cpu_rq(cpu)->migration_thread = p;
6514 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006515
Linus Torvalds1da177e2005-04-16 15:20:36 -07006516 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006517 case CPU_ONLINE_FROZEN:
Robert P. J. Day3a4fa0a2007-10-19 23:10:43 +02006518 /* Strictly unnecessary, as first user will wake it. */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006519 wake_up_process(cpu_rq(cpu)->migration_thread);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006520
6521 /* Update our root-domain */
6522 rq = cpu_rq(cpu);
6523 spin_lock_irqsave(&rq->lock, flags);
6524 if (rq->rd) {
6525 BUG_ON(!cpu_isset(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006526
6527 set_rq_online(rq);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006528 }
6529 spin_unlock_irqrestore(&rq->lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006530 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006531
Linus Torvalds1da177e2005-04-16 15:20:36 -07006532#ifdef CONFIG_HOTPLUG_CPU
6533 case CPU_UP_CANCELED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006534 case CPU_UP_CANCELED_FROZEN:
Heiko Carstensfc75cdf2006-06-25 05:49:10 -07006535 if (!cpu_rq(cpu)->migration_thread)
6536 break;
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006537 /* Unbind it from offline cpu so it can run. Fall thru. */
Heiko Carstensa4c4af72005-11-07 00:58:38 -08006538 kthread_bind(cpu_rq(cpu)->migration_thread,
6539 any_online_cpu(cpu_online_map));
Linus Torvalds1da177e2005-04-16 15:20:36 -07006540 kthread_stop(cpu_rq(cpu)->migration_thread);
6541 cpu_rq(cpu)->migration_thread = NULL;
6542 break;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006543
Linus Torvalds1da177e2005-04-16 15:20:36 -07006544 case CPU_DEAD:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07006545 case CPU_DEAD_FROZEN:
Cliff Wickman470fd642007-10-18 23:40:46 -07006546 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006547 migrate_live_tasks(cpu);
6548 rq = cpu_rq(cpu);
6549 kthread_stop(rq->migration_thread);
6550 rq->migration_thread = NULL;
6551 /* Idle task back to normal (off runqueue, low prio) */
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006552 spin_lock_irq(&rq->lock);
Ingo Molnara8e504d2007-08-09 11:16:47 +02006553 update_rq_clock(rq);
Ingo Molnar2e1cb742007-08-09 11:16:49 +02006554 deactivate_task(rq, rq->idle, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006555 rq->idle->static_prio = MAX_PRIO;
Ingo Molnardd41f592007-07-09 18:51:59 +02006556 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6557 rq->idle->sched_class = &idle_sched_class;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006558 migrate_dead_tasks(cpu);
Oleg Nesterovd2da2722007-10-16 23:30:56 -07006559 spin_unlock_irq(&rq->lock);
Cliff Wickman470fd642007-10-18 23:40:46 -07006560 cpuset_unlock();
Linus Torvalds1da177e2005-04-16 15:20:36 -07006561 migrate_nr_uninterruptible(rq);
6562 BUG_ON(rq->nr_running != 0);
6563
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006564 /*
6565 * No need to migrate the tasks: it was best-effort if
6566 * they didn't take sched_hotcpu_mutex. Just wake up
6567 * the requestors.
6568 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006569 spin_lock_irq(&rq->lock);
6570 while (!list_empty(&rq->migration_queue)) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07006571 struct migration_req *req;
6572
Linus Torvalds1da177e2005-04-16 15:20:36 -07006573 req = list_entry(rq->migration_queue.next,
Ingo Molnar70b97a72006-07-03 00:25:42 -07006574 struct migration_req, list);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006575 list_del_init(&req->list);
6576 complete(&req->done);
6577 }
6578 spin_unlock_irq(&rq->lock);
6579 break;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006580
Gregory Haskins08f503b2008-03-10 17:59:11 -04006581 case CPU_DYING:
6582 case CPU_DYING_FROZEN:
Gregory Haskins57d885f2008-01-25 21:08:18 +01006583 /* Update our root-domain */
6584 rq = cpu_rq(cpu);
6585 spin_lock_irqsave(&rq->lock, flags);
6586 if (rq->rd) {
6587 BUG_ON(!cpu_isset(cpu, rq->rd->span));
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006588 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006589 }
6590 spin_unlock_irqrestore(&rq->lock, flags);
6591 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006592#endif
6593 }
6594 return NOTIFY_OK;
6595}
6596
6597/* Register at highest priority so that task migration (migrate_all_tasks)
6598 * happens before everything else.
6599 */
Chandra Seetharaman26c21432006-06-27 02:54:10 -07006600static struct notifier_block __cpuinitdata migration_notifier = {
Linus Torvalds1da177e2005-04-16 15:20:36 -07006601 .notifier_call = migration_call,
6602 .priority = 10
6603};
6604
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07006605static int __init migration_init(void)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006606{
6607 void *cpu = (void *)(long)smp_processor_id();
Akinobu Mita07dccf32006-09-29 02:00:22 -07006608 int err;
Ingo Molnar48f24c42006-07-03 00:25:40 -07006609
6610 /* Start one for the boot CPU: */
Akinobu Mita07dccf32006-09-29 02:00:22 -07006611 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6612 BUG_ON(err == NOTIFY_BAD);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006613 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6614 register_cpu_notifier(&migration_notifier);
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07006615
6616 return err;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006617}
Eduard - Gabriel Munteanu7babe8d2008-07-25 19:45:11 -07006618early_initcall(migration_init);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006619#endif
6620
6621#ifdef CONFIG_SMP
Christoph Lameter476f3532007-05-06 14:48:58 -07006622
Ingo Molnar3e9830d2007-10-15 17:00:13 +02006623#ifdef CONFIG_SCHED_DEBUG
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006624
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306625static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6626{
6627 switch (lvl) {
6628 case SD_LV_NONE:
6629 return "NONE";
6630 case SD_LV_SIBLING:
6631 return "SIBLING";
6632 case SD_LV_MC:
6633 return "MC";
6634 case SD_LV_CPU:
6635 return "CPU";
6636 case SD_LV_NODE:
6637 return "NODE";
6638 case SD_LV_ALLNODES:
6639 return "ALLNODES";
6640 case SD_LV_MAX:
6641 return "MAX";
6642
6643 }
6644 return "MAX";
6645}
6646
Mike Travis7c16ec52008-04-04 18:11:11 -07006647static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6648 cpumask_t *groupmask)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006649{
6650 struct sched_group *group = sd->groups;
Mike Travis434d53b2008-04-04 18:11:04 -07006651 char str[256];
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006652
Mike Travis434d53b2008-04-04 18:11:04 -07006653 cpulist_scnprintf(str, sizeof(str), sd->span);
Mike Travis7c16ec52008-04-04 18:11:11 -07006654 cpus_clear(*groupmask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006655
6656 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6657
6658 if (!(sd->flags & SD_LOAD_BALANCE)) {
6659 printk("does not load-balance\n");
6660 if (sd->parent)
6661 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6662 " has parent");
6663 return -1;
6664 }
6665
Gautham R Shenoy099f98c2008-05-29 20:56:32 +05306666 printk(KERN_CONT "span %s level %s\n",
6667 str, sd_level_to_string(sd->level));
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006668
6669 if (!cpu_isset(cpu, sd->span)) {
6670 printk(KERN_ERR "ERROR: domain->span does not contain "
6671 "CPU%d\n", cpu);
6672 }
6673 if (!cpu_isset(cpu, group->cpumask)) {
6674 printk(KERN_ERR "ERROR: domain->groups does not contain"
6675 " CPU%d\n", cpu);
6676 }
6677
6678 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6679 do {
6680 if (!group) {
6681 printk("\n");
6682 printk(KERN_ERR "ERROR: group is NULL\n");
6683 break;
6684 }
6685
6686 if (!group->__cpu_power) {
6687 printk(KERN_CONT "\n");
6688 printk(KERN_ERR "ERROR: domain->cpu_power not "
6689 "set\n");
6690 break;
6691 }
6692
6693 if (!cpus_weight(group->cpumask)) {
6694 printk(KERN_CONT "\n");
6695 printk(KERN_ERR "ERROR: empty group\n");
6696 break;
6697 }
6698
Mike Travis7c16ec52008-04-04 18:11:11 -07006699 if (cpus_intersects(*groupmask, group->cpumask)) {
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006700 printk(KERN_CONT "\n");
6701 printk(KERN_ERR "ERROR: repeated CPUs\n");
6702 break;
6703 }
6704
Mike Travis7c16ec52008-04-04 18:11:11 -07006705 cpus_or(*groupmask, *groupmask, group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006706
Mike Travis434d53b2008-04-04 18:11:04 -07006707 cpulist_scnprintf(str, sizeof(str), group->cpumask);
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006708 printk(KERN_CONT " %s", str);
6709
6710 group = group->next;
6711 } while (group != sd->groups);
6712 printk(KERN_CONT "\n");
6713
Mike Travis7c16ec52008-04-04 18:11:11 -07006714 if (!cpus_equal(sd->span, *groupmask))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006715 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6716
Mike Travis7c16ec52008-04-04 18:11:11 -07006717 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006718 printk(KERN_ERR "ERROR: parent span is not a superset "
6719 "of domain->span\n");
6720 return 0;
6721}
6722
Linus Torvalds1da177e2005-04-16 15:20:36 -07006723static void sched_domain_debug(struct sched_domain *sd, int cpu)
6724{
Mike Travis7c16ec52008-04-04 18:11:11 -07006725 cpumask_t *groupmask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006726 int level = 0;
6727
Nick Piggin41c7ce92005-06-25 14:57:24 -07006728 if (!sd) {
6729 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6730 return;
6731 }
6732
Linus Torvalds1da177e2005-04-16 15:20:36 -07006733 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6734
Mike Travis7c16ec52008-04-04 18:11:11 -07006735 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6736 if (!groupmask) {
6737 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6738 return;
6739 }
6740
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006741 for (;;) {
Mike Travis7c16ec52008-04-04 18:11:11 -07006742 if (sched_domain_debug_one(sd, cpu, level, groupmask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006743 break;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006744 level++;
6745 sd = sd->parent;
Miguel Ojeda Sandonis33859f72006-12-10 02:20:38 -08006746 if (!sd)
Ingo Molnar4dcf6af2007-10-24 18:23:48 +02006747 break;
6748 }
Mike Travis7c16ec52008-04-04 18:11:11 -07006749 kfree(groupmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006750}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006751#else /* !CONFIG_SCHED_DEBUG */
Ingo Molnar48f24c42006-07-03 00:25:40 -07006752# define sched_domain_debug(sd, cpu) do { } while (0)
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02006753#endif /* CONFIG_SCHED_DEBUG */
Linus Torvalds1da177e2005-04-16 15:20:36 -07006754
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07006755static int sd_degenerate(struct sched_domain *sd)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006756{
6757 if (cpus_weight(sd->span) == 1)
6758 return 1;
6759
6760 /* Following flags need at least 2 groups */
6761 if (sd->flags & (SD_LOAD_BALANCE |
6762 SD_BALANCE_NEWIDLE |
6763 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006764 SD_BALANCE_EXEC |
6765 SD_SHARE_CPUPOWER |
6766 SD_SHARE_PKG_RESOURCES)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006767 if (sd->groups != sd->groups->next)
6768 return 0;
6769 }
6770
6771 /* Following flags don't use groups */
6772 if (sd->flags & (SD_WAKE_IDLE |
6773 SD_WAKE_AFFINE |
6774 SD_WAKE_BALANCE))
6775 return 0;
6776
6777 return 1;
6778}
6779
Ingo Molnar48f24c42006-07-03 00:25:40 -07006780static int
6781sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
Suresh Siddha245af2c2005-06-25 14:57:25 -07006782{
6783 unsigned long cflags = sd->flags, pflags = parent->flags;
6784
6785 if (sd_degenerate(parent))
6786 return 1;
6787
6788 if (!cpus_equal(sd->span, parent->span))
6789 return 0;
6790
6791 /* Does parent contain flags not in child? */
6792 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6793 if (cflags & SD_WAKE_AFFINE)
6794 pflags &= ~SD_WAKE_BALANCE;
6795 /* Flags needing groups don't count if only 1 group in parent */
6796 if (parent->groups == parent->groups->next) {
6797 pflags &= ~(SD_LOAD_BALANCE |
6798 SD_BALANCE_NEWIDLE |
6799 SD_BALANCE_FORK |
Siddha, Suresh B89c47102006-10-03 01:14:09 -07006800 SD_BALANCE_EXEC |
6801 SD_SHARE_CPUPOWER |
6802 SD_SHARE_PKG_RESOURCES);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006803 }
6804 if (~cflags & pflags)
6805 return 0;
6806
6807 return 1;
6808}
6809
Gregory Haskins57d885f2008-01-25 21:08:18 +01006810static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6811{
6812 unsigned long flags;
Gregory Haskins57d885f2008-01-25 21:08:18 +01006813
6814 spin_lock_irqsave(&rq->lock, flags);
6815
6816 if (rq->rd) {
6817 struct root_domain *old_rd = rq->rd;
6818
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006819 if (cpu_isset(rq->cpu, old_rd->online))
6820 set_rq_offline(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006821
Gregory Haskinsdc938522008-01-25 21:08:26 +01006822 cpu_clear(rq->cpu, old_rd->span);
Gregory Haskinsdc938522008-01-25 21:08:26 +01006823
Gregory Haskins57d885f2008-01-25 21:08:18 +01006824 if (atomic_dec_and_test(&old_rd->refcount))
6825 kfree(old_rd);
6826 }
6827
6828 atomic_inc(&rd->refcount);
6829 rq->rd = rd;
6830
Gregory Haskinsdc938522008-01-25 21:08:26 +01006831 cpu_set(rq->cpu, rd->span);
Gregory Haskins1f94ef52008-03-10 16:52:41 -04006832 if (cpu_isset(rq->cpu, cpu_online_map))
Gregory Haskins1f11eb62008-06-04 15:04:05 -04006833 set_rq_online(rq);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006834
6835 spin_unlock_irqrestore(&rq->lock, flags);
6836}
6837
Gregory Haskinsdc938522008-01-25 21:08:26 +01006838static void init_rootdomain(struct root_domain *rd)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006839{
6840 memset(rd, 0, sizeof(*rd));
6841
Gregory Haskinsdc938522008-01-25 21:08:26 +01006842 cpus_clear(rd->span);
6843 cpus_clear(rd->online);
Gregory Haskins6e0534f2008-05-12 21:21:01 +02006844
6845 cpupri_init(&rd->cpupri);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006846}
6847
6848static void init_defrootdomain(void)
6849{
Gregory Haskinsdc938522008-01-25 21:08:26 +01006850 init_rootdomain(&def_root_domain);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006851 atomic_set(&def_root_domain.refcount, 1);
6852}
6853
Gregory Haskinsdc938522008-01-25 21:08:26 +01006854static struct root_domain *alloc_rootdomain(void)
Gregory Haskins57d885f2008-01-25 21:08:18 +01006855{
6856 struct root_domain *rd;
6857
6858 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6859 if (!rd)
6860 return NULL;
6861
Gregory Haskinsdc938522008-01-25 21:08:26 +01006862 init_rootdomain(rd);
Gregory Haskins57d885f2008-01-25 21:08:18 +01006863
6864 return rd;
6865}
6866
Linus Torvalds1da177e2005-04-16 15:20:36 -07006867/*
Ingo Molnar0eab9142008-01-25 21:08:19 +01006868 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
Linus Torvalds1da177e2005-04-16 15:20:36 -07006869 * hold the hotplug lock.
6870 */
Ingo Molnar0eab9142008-01-25 21:08:19 +01006871static void
6872cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006873{
Ingo Molnar70b97a72006-07-03 00:25:42 -07006874 struct rq *rq = cpu_rq(cpu);
Suresh Siddha245af2c2005-06-25 14:57:25 -07006875 struct sched_domain *tmp;
6876
6877 /* Remove the sched domains which do not contribute to scheduling. */
6878 for (tmp = sd; tmp; tmp = tmp->parent) {
6879 struct sched_domain *parent = tmp->parent;
6880 if (!parent)
6881 break;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006882 if (sd_parent_degenerate(tmp, parent)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006883 tmp->parent = parent->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006884 if (parent->parent)
6885 parent->parent->child = tmp;
6886 }
Suresh Siddha245af2c2005-06-25 14:57:25 -07006887 }
6888
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006889 if (sd && sd_degenerate(sd)) {
Suresh Siddha245af2c2005-06-25 14:57:25 -07006890 sd = sd->parent;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07006891 if (sd)
6892 sd->child = NULL;
6893 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07006894
6895 sched_domain_debug(sd, cpu);
6896
Gregory Haskins57d885f2008-01-25 21:08:18 +01006897 rq_attach_root(rq, rd);
Nick Piggin674311d2005-06-25 14:57:27 -07006898 rcu_assign_pointer(rq->sd, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006899}
6900
6901/* cpus with isolated domains */
Tim Chen67af63a2006-12-22 01:07:50 -08006902static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006903
6904/* Setup the mask of cpus configured for isolated domains */
6905static int __init isolated_cpu_setup(char *str)
6906{
Mike Travis13b40c12008-07-01 10:32:50 -07006907 static int __initdata ints[NR_CPUS];
6908 int i;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006909
6910 str = get_options(str, ARRAY_SIZE(ints), ints);
6911 cpus_clear(cpu_isolated_map);
6912 for (i = 1; i <= ints[0]; i++)
6913 if (ints[i] < NR_CPUS)
6914 cpu_set(ints[i], cpu_isolated_map);
6915 return 1;
6916}
6917
Ingo Molnar8927f492007-10-15 17:00:13 +02006918__setup("isolcpus=", isolated_cpu_setup);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006919
6920/*
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006921 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6922 * to a function which identifies what group(along with sched group) a CPU
6923 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6924 * (due to the fact that we keep track of groups covered with a cpumask_t).
Linus Torvalds1da177e2005-04-16 15:20:36 -07006925 *
6926 * init_sched_build_groups will build a circular linked list of the groups
6927 * covered by the given span, and will set each group's ->cpumask correctly,
6928 * and ->cpu_power to 0.
6929 */
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07006930static void
Mike Travis7c16ec52008-04-04 18:11:11 -07006931init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006932 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07006933 struct sched_group **sg,
6934 cpumask_t *tmpmask),
6935 cpumask_t *covered, cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006936{
6937 struct sched_group *first = NULL, *last = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006938 int i;
6939
Mike Travis7c16ec52008-04-04 18:11:11 -07006940 cpus_clear(*covered);
6941
Mike Travis363ab6f2008-05-12 21:21:13 +02006942 for_each_cpu_mask_nr(i, *span) {
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08006943 struct sched_group *sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07006944 int group = group_fn(i, cpu_map, &sg, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006945 int j;
6946
Mike Travis7c16ec52008-04-04 18:11:11 -07006947 if (cpu_isset(i, *covered))
Linus Torvalds1da177e2005-04-16 15:20:36 -07006948 continue;
6949
Mike Travis7c16ec52008-04-04 18:11:11 -07006950 cpus_clear(sg->cpumask);
Eric Dumazet5517d862007-05-08 00:32:57 -07006951 sg->__cpu_power = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07006952
Mike Travis363ab6f2008-05-12 21:21:13 +02006953 for_each_cpu_mask_nr(j, *span) {
Mike Travis7c16ec52008-04-04 18:11:11 -07006954 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
Linus Torvalds1da177e2005-04-16 15:20:36 -07006955 continue;
6956
Mike Travis7c16ec52008-04-04 18:11:11 -07006957 cpu_set(j, *covered);
Linus Torvalds1da177e2005-04-16 15:20:36 -07006958 cpu_set(j, sg->cpumask);
6959 }
6960 if (!first)
6961 first = sg;
6962 if (last)
6963 last->next = sg;
6964 last = sg;
6965 }
6966 last->next = first;
6967}
6968
John Hawkes9c1cfda2005-09-06 15:18:14 -07006969#define SD_NODES_PER_DOMAIN 16
Linus Torvalds1da177e2005-04-16 15:20:36 -07006970
John Hawkes9c1cfda2005-09-06 15:18:14 -07006971#ifdef CONFIG_NUMA
akpm@osdl.org198e2f12006-01-12 01:05:30 -08006972
John Hawkes9c1cfda2005-09-06 15:18:14 -07006973/**
6974 * find_next_best_node - find the next node to include in a sched_domain
6975 * @node: node whose sched_domain we're building
6976 * @used_nodes: nodes already in the sched_domain
6977 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01006978 * Find the next node to include in a given scheduling domain. Simply
John Hawkes9c1cfda2005-09-06 15:18:14 -07006979 * finds the closest node not already in the @used_nodes map.
6980 *
6981 * Should use nodemask_t.
6982 */
Mike Travisc5f59f02008-04-04 18:11:10 -07006983static int find_next_best_node(int node, nodemask_t *used_nodes)
John Hawkes9c1cfda2005-09-06 15:18:14 -07006984{
6985 int i, n, val, min_val, best_node = 0;
6986
6987 min_val = INT_MAX;
6988
Mike Travis076ac2a2008-05-12 21:21:12 +02006989 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07006990 /* Start at @node */
Mike Travis076ac2a2008-05-12 21:21:12 +02006991 n = (node + i) % nr_node_ids;
John Hawkes9c1cfda2005-09-06 15:18:14 -07006992
6993 if (!nr_cpus_node(n))
6994 continue;
6995
6996 /* Skip already used nodes */
Mike Travisc5f59f02008-04-04 18:11:10 -07006997 if (node_isset(n, *used_nodes))
John Hawkes9c1cfda2005-09-06 15:18:14 -07006998 continue;
6999
7000 /* Simple min distance search */
7001 val = node_distance(node, n);
7002
7003 if (val < min_val) {
7004 min_val = val;
7005 best_node = n;
7006 }
7007 }
7008
Mike Travisc5f59f02008-04-04 18:11:10 -07007009 node_set(best_node, *used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007010 return best_node;
7011}
7012
7013/**
7014 * sched_domain_node_span - get a cpumask for a node's sched_domain
7015 * @node: node whose cpumask we're constructing
Randy Dunlap73486722008-04-22 10:07:22 -07007016 * @span: resulting cpumask
John Hawkes9c1cfda2005-09-06 15:18:14 -07007017 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007018 * Given a node, construct a good cpumask for its sched_domain to span. It
John Hawkes9c1cfda2005-09-06 15:18:14 -07007019 * should be one that prevents unnecessary balancing, but also spreads tasks
7020 * out optimally.
7021 */
Mike Travis4bdbaad32008-04-15 16:35:52 -07007022static void sched_domain_node_span(int node, cpumask_t *span)
John Hawkes9c1cfda2005-09-06 15:18:14 -07007023{
Mike Travisc5f59f02008-04-04 18:11:10 -07007024 nodemask_t used_nodes;
Mike Travisc5f59f02008-04-04 18:11:10 -07007025 node_to_cpumask_ptr(nodemask, node);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007026 int i;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007027
Mike Travis4bdbaad32008-04-15 16:35:52 -07007028 cpus_clear(*span);
Mike Travisc5f59f02008-04-04 18:11:10 -07007029 nodes_clear(used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007030
Mike Travis4bdbaad32008-04-15 16:35:52 -07007031 cpus_or(*span, *span, *nodemask);
Mike Travisc5f59f02008-04-04 18:11:10 -07007032 node_set(node, used_nodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007033
7034 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
Mike Travisc5f59f02008-04-04 18:11:10 -07007035 int next_node = find_next_best_node(node, &used_nodes);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007036
Mike Travisc5f59f02008-04-04 18:11:10 -07007037 node_to_cpumask_ptr_next(nodemask, next_node);
Mike Travis4bdbaad32008-04-15 16:35:52 -07007038 cpus_or(*span, *span, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007039 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007040}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007041#endif /* CONFIG_NUMA */
John Hawkes9c1cfda2005-09-06 15:18:14 -07007042
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007043int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007044
John Hawkes9c1cfda2005-09-06 15:18:14 -07007045/*
Ingo Molnar48f24c42006-07-03 00:25:40 -07007046 * SMT sched-domains:
John Hawkes9c1cfda2005-09-06 15:18:14 -07007047 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007048#ifdef CONFIG_SCHED_SMT
7049static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007050static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007051
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007052static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007053cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7054 cpumask_t *unused)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007055{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007056 if (sg)
7057 *sg = &per_cpu(sched_group_cpus, cpu);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007058 return cpu;
7059}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007060#endif /* CONFIG_SCHED_SMT */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007061
Ingo Molnar48f24c42006-07-03 00:25:40 -07007062/*
7063 * multi-core sched-domains:
7064 */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007065#ifdef CONFIG_SCHED_MC
7066static DEFINE_PER_CPU(struct sched_domain, core_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007067static DEFINE_PER_CPU(struct sched_group, sched_group_core);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007068#endif /* CONFIG_SCHED_MC */
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007069
7070#if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007071static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007072cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7073 cpumask_t *mask)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007074{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007075 int group;
Mike Travis7c16ec52008-04-04 18:11:11 -07007076
7077 *mask = per_cpu(cpu_sibling_map, cpu);
7078 cpus_and(*mask, *mask, *cpu_map);
7079 group = first_cpu(*mask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007080 if (sg)
7081 *sg = &per_cpu(sched_group_core, group);
7082 return group;
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007083}
7084#elif defined(CONFIG_SCHED_MC)
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007085static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007086cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7087 cpumask_t *unused)
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007088{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007089 if (sg)
7090 *sg = &per_cpu(sched_group_core, cpu);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007091 return cpu;
7092}
7093#endif
7094
Linus Torvalds1da177e2005-04-16 15:20:36 -07007095static DEFINE_PER_CPU(struct sched_domain, phys_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007096static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
Ingo Molnar48f24c42006-07-03 00:25:40 -07007097
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007098static int
Mike Travis7c16ec52008-04-04 18:11:11 -07007099cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7100 cpumask_t *mask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007101{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007102 int group;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007103#ifdef CONFIG_SCHED_MC
Mike Travis7c16ec52008-04-04 18:11:11 -07007104 *mask = cpu_coregroup_map(cpu);
7105 cpus_and(*mask, *mask, *cpu_map);
7106 group = first_cpu(*mask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007107#elif defined(CONFIG_SCHED_SMT)
Mike Travis7c16ec52008-04-04 18:11:11 -07007108 *mask = per_cpu(cpu_sibling_map, cpu);
7109 cpus_and(*mask, *mask, *cpu_map);
7110 group = first_cpu(*mask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007111#else
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007112 group = cpu;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007113#endif
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007114 if (sg)
7115 *sg = &per_cpu(sched_group_phys, group);
7116 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007117}
7118
7119#ifdef CONFIG_NUMA
John Hawkes9c1cfda2005-09-06 15:18:14 -07007120/*
7121 * The init_sched_build_groups can't handle what we want to do with node
7122 * groups, so roll our own. Now each node has its own list of groups which
7123 * gets dynamically allocated.
7124 */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007125static DEFINE_PER_CPU(struct sched_domain, node_domains);
Mike Travis434d53b2008-04-04 18:11:04 -07007126static struct sched_group ***sched_group_nodes_bycpu;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007127
7128static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007129static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007130
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007131static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007132 struct sched_group **sg, cpumask_t *nodemask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007133{
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007134 int group;
7135
Mike Travis7c16ec52008-04-04 18:11:11 -07007136 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7137 cpus_and(*nodemask, *nodemask, *cpu_map);
7138 group = first_cpu(*nodemask);
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007139
7140 if (sg)
7141 *sg = &per_cpu(sched_group_allnodes, group);
7142 return group;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007143}
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007144
Siddha, Suresh B08069032006-03-27 01:15:23 -08007145static void init_numa_sched_groups_power(struct sched_group *group_head)
7146{
7147 struct sched_group *sg = group_head;
7148 int j;
7149
7150 if (!sg)
7151 return;
Andi Kleen3a5c3592007-10-15 17:00:14 +02007152 do {
Mike Travis363ab6f2008-05-12 21:21:13 +02007153 for_each_cpu_mask_nr(j, sg->cpumask) {
Andi Kleen3a5c3592007-10-15 17:00:14 +02007154 struct sched_domain *sd;
Siddha, Suresh B08069032006-03-27 01:15:23 -08007155
Andi Kleen3a5c3592007-10-15 17:00:14 +02007156 sd = &per_cpu(phys_domains, j);
7157 if (j != first_cpu(sd->groups->cpumask)) {
7158 /*
7159 * Only add "power" once for each
7160 * physical package.
7161 */
7162 continue;
7163 }
7164
7165 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
Siddha, Suresh B08069032006-03-27 01:15:23 -08007166 }
Andi Kleen3a5c3592007-10-15 17:00:14 +02007167 sg = sg->next;
7168 } while (sg != group_head);
Siddha, Suresh B08069032006-03-27 01:15:23 -08007169}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007170#endif /* CONFIG_NUMA */
Linus Torvalds1da177e2005-04-16 15:20:36 -07007171
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007172#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007173/* Free memory allocated for various sched_group structures */
Mike Travis7c16ec52008-04-04 18:11:11 -07007174static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007175{
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007176 int cpu, i;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007177
Mike Travis363ab6f2008-05-12 21:21:13 +02007178 for_each_cpu_mask_nr(cpu, *cpu_map) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007179 struct sched_group **sched_group_nodes
7180 = sched_group_nodes_bycpu[cpu];
7181
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007182 if (!sched_group_nodes)
7183 continue;
7184
Mike Travis076ac2a2008-05-12 21:21:12 +02007185 for (i = 0; i < nr_node_ids; i++) {
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007186 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7187
Mike Travis7c16ec52008-04-04 18:11:11 -07007188 *nodemask = node_to_cpumask(i);
7189 cpus_and(*nodemask, *nodemask, *cpu_map);
7190 if (cpus_empty(*nodemask))
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007191 continue;
7192
7193 if (sg == NULL)
7194 continue;
7195 sg = sg->next;
7196next_sg:
7197 oldsg = sg;
7198 sg = sg->next;
7199 kfree(oldsg);
7200 if (oldsg != sched_group_nodes[i])
7201 goto next_sg;
7202 }
7203 kfree(sched_group_nodes);
7204 sched_group_nodes_bycpu[cpu] = NULL;
7205 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007206}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007207#else /* !CONFIG_NUMA */
Mike Travis7c16ec52008-04-04 18:11:11 -07007208static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007209{
7210}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007211#endif /* CONFIG_NUMA */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007212
Linus Torvalds1da177e2005-04-16 15:20:36 -07007213/*
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007214 * Initialize sched groups cpu_power.
7215 *
7216 * cpu_power indicates the capacity of sched group, which is used while
7217 * distributing the load between different sched groups in a sched domain.
7218 * Typically cpu_power for all the groups in a sched domain will be same unless
7219 * there are asymmetries in the topology. If there are asymmetries, group
7220 * having more cpu_power will pickup more load compared to the group having
7221 * less cpu_power.
7222 *
7223 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7224 * the maximum number of tasks a group can handle in the presence of other idle
7225 * or lightly loaded groups in the same sched domain.
7226 */
7227static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7228{
7229 struct sched_domain *child;
7230 struct sched_group *group;
7231
7232 WARN_ON(!sd || !sd->groups);
7233
7234 if (cpu != first_cpu(sd->groups->cpumask))
7235 return;
7236
7237 child = sd->child;
7238
Eric Dumazet5517d862007-05-08 00:32:57 -07007239 sd->groups->__cpu_power = 0;
7240
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007241 /*
7242 * For perf policy, if the groups in child domain share resources
7243 * (for example cores sharing some portions of the cache hierarchy
7244 * or SMT), then set this domain groups cpu_power such that each group
7245 * can handle only one task, when there are other idle groups in the
7246 * same sched domain.
7247 */
7248 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7249 (child->flags &
7250 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
Eric Dumazet5517d862007-05-08 00:32:57 -07007251 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007252 return;
7253 }
7254
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007255 /*
7256 * add cpu_power of each child group to this groups cpu_power
7257 */
7258 group = child->groups;
7259 do {
Eric Dumazet5517d862007-05-08 00:32:57 -07007260 sg_inc_cpu_power(sd->groups, group->__cpu_power);
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007261 group = group->next;
7262 } while (group != child->groups);
7263}
7264
7265/*
Mike Travis7c16ec52008-04-04 18:11:11 -07007266 * Initializers for schedule domains
7267 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7268 */
7269
Ingo Molnara5d8c342008-10-09 11:35:51 +02007270#ifdef CONFIG_SCHED_DEBUG
7271# define SD_INIT_NAME(sd, type) sd->name = #type
7272#else
7273# define SD_INIT_NAME(sd, type) do { } while (0)
7274#endif
7275
Mike Travis7c16ec52008-04-04 18:11:11 -07007276#define SD_INIT(sd, type) sd_init_##type(sd)
Ingo Molnara5d8c342008-10-09 11:35:51 +02007277
Mike Travis7c16ec52008-04-04 18:11:11 -07007278#define SD_INIT_FUNC(type) \
7279static noinline void sd_init_##type(struct sched_domain *sd) \
7280{ \
7281 memset(sd, 0, sizeof(*sd)); \
7282 *sd = SD_##type##_INIT; \
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007283 sd->level = SD_LV_##type; \
Ingo Molnara5d8c342008-10-09 11:35:51 +02007284 SD_INIT_NAME(sd, type); \
Mike Travis7c16ec52008-04-04 18:11:11 -07007285}
7286
7287SD_INIT_FUNC(CPU)
7288#ifdef CONFIG_NUMA
7289 SD_INIT_FUNC(ALLNODES)
7290 SD_INIT_FUNC(NODE)
7291#endif
7292#ifdef CONFIG_SCHED_SMT
7293 SD_INIT_FUNC(SIBLING)
7294#endif
7295#ifdef CONFIG_SCHED_MC
7296 SD_INIT_FUNC(MC)
7297#endif
7298
7299/*
7300 * To minimize stack usage kmalloc room for cpumasks and share the
7301 * space as the usage in build_sched_domains() dictates. Used only
7302 * if the amount of space is significant.
7303 */
7304struct allmasks {
7305 cpumask_t tmpmask; /* make this one first */
7306 union {
7307 cpumask_t nodemask;
7308 cpumask_t this_sibling_map;
7309 cpumask_t this_core_map;
7310 };
7311 cpumask_t send_covered;
7312
7313#ifdef CONFIG_NUMA
7314 cpumask_t domainspan;
7315 cpumask_t covered;
7316 cpumask_t notcovered;
7317#endif
7318};
7319
7320#if NR_CPUS > 128
7321#define SCHED_CPUMASK_ALLOC 1
7322#define SCHED_CPUMASK_FREE(v) kfree(v)
7323#define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7324#else
7325#define SCHED_CPUMASK_ALLOC 0
7326#define SCHED_CPUMASK_FREE(v)
7327#define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7328#endif
7329
7330#define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7331 ((unsigned long)(a) + offsetof(struct allmasks, v))
7332
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007333static int default_relax_domain_level = -1;
7334
7335static int __init setup_relax_domain_level(char *str)
7336{
Li Zefan30e0e172008-05-13 10:27:17 +08007337 unsigned long val;
7338
7339 val = simple_strtoul(str, NULL, 0);
7340 if (val < SD_LV_MAX)
7341 default_relax_domain_level = val;
7342
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007343 return 1;
7344}
7345__setup("relax_domain_level=", setup_relax_domain_level);
7346
7347static void set_domain_attribute(struct sched_domain *sd,
7348 struct sched_domain_attr *attr)
7349{
7350 int request;
7351
7352 if (!attr || attr->relax_domain_level < 0) {
7353 if (default_relax_domain_level < 0)
7354 return;
7355 else
7356 request = default_relax_domain_level;
7357 } else
7358 request = attr->relax_domain_level;
7359 if (request < sd->level) {
7360 /* turn off idle balance on this domain */
7361 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7362 } else {
7363 /* turn on idle balance on this domain */
7364 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7365 }
7366}
7367
Mike Travis7c16ec52008-04-04 18:11:11 -07007368/*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007369 * Build sched domains for a given set of cpus and attach the sched domains
7370 * to the individual cpus
Linus Torvalds1da177e2005-04-16 15:20:36 -07007371 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007372static int __build_sched_domains(const cpumask_t *cpu_map,
7373 struct sched_domain_attr *attr)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007374{
7375 int i;
Gregory Haskins57d885f2008-01-25 21:08:18 +01007376 struct root_domain *rd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007377 SCHED_CPUMASK_DECLARE(allmasks);
7378 cpumask_t *tmpmask;
John Hawkesd1b55132005-09-06 15:18:14 -07007379#ifdef CONFIG_NUMA
7380 struct sched_group **sched_group_nodes = NULL;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007381 int sd_allnodes = 0;
John Hawkesd1b55132005-09-06 15:18:14 -07007382
7383 /*
7384 * Allocate the per-node list of sched groups
7385 */
Mike Travis076ac2a2008-05-12 21:21:12 +02007386 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007387 GFP_KERNEL);
John Hawkesd1b55132005-09-06 15:18:14 -07007388 if (!sched_group_nodes) {
7389 printk(KERN_WARNING "Can not alloc sched group node list\n");
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007390 return -ENOMEM;
John Hawkesd1b55132005-09-06 15:18:14 -07007391 }
John Hawkesd1b55132005-09-06 15:18:14 -07007392#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007393
Gregory Haskinsdc938522008-01-25 21:08:26 +01007394 rd = alloc_rootdomain();
Gregory Haskins57d885f2008-01-25 21:08:18 +01007395 if (!rd) {
7396 printk(KERN_WARNING "Cannot alloc root domain\n");
Mike Travis7c16ec52008-04-04 18:11:11 -07007397#ifdef CONFIG_NUMA
7398 kfree(sched_group_nodes);
7399#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007400 return -ENOMEM;
7401 }
7402
Mike Travis7c16ec52008-04-04 18:11:11 -07007403#if SCHED_CPUMASK_ALLOC
7404 /* get space for all scratch cpumask variables */
7405 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7406 if (!allmasks) {
7407 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7408 kfree(rd);
7409#ifdef CONFIG_NUMA
7410 kfree(sched_group_nodes);
7411#endif
7412 return -ENOMEM;
7413 }
7414#endif
7415 tmpmask = (cpumask_t *)allmasks;
7416
7417
7418#ifdef CONFIG_NUMA
7419 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7420#endif
7421
Linus Torvalds1da177e2005-04-16 15:20:36 -07007422 /*
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007423 * Set up domains for cpus specified by the cpu_map.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007424 */
Mike Travis363ab6f2008-05-12 21:21:13 +02007425 for_each_cpu_mask_nr(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007426 struct sched_domain *sd = NULL, *p;
Mike Travis7c16ec52008-04-04 18:11:11 -07007427 SCHED_CPUMASK_VAR(nodemask, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007428
Mike Travis7c16ec52008-04-04 18:11:11 -07007429 *nodemask = node_to_cpumask(cpu_to_node(i));
7430 cpus_and(*nodemask, *nodemask, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007431
7432#ifdef CONFIG_NUMA
Ingo Molnardd41f592007-07-09 18:51:59 +02007433 if (cpus_weight(*cpu_map) >
Mike Travis7c16ec52008-04-04 18:11:11 -07007434 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007435 sd = &per_cpu(allnodes_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007436 SD_INIT(sd, ALLNODES);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007437 set_domain_attribute(sd, attr);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007438 sd->span = *cpu_map;
Mike Travis7c16ec52008-04-04 18:11:11 -07007439 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007440 p = sd;
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007441 sd_allnodes = 1;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007442 } else
7443 p = NULL;
7444
Linus Torvalds1da177e2005-04-16 15:20:36 -07007445 sd = &per_cpu(node_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007446 SD_INIT(sd, NODE);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007447 set_domain_attribute(sd, attr);
Mike Travis4bdbaad32008-04-15 16:35:52 -07007448 sched_domain_node_span(cpu_to_node(i), &sd->span);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007449 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007450 if (p)
7451 p->child = sd;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007452 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007453#endif
7454
7455 p = sd;
7456 sd = &per_cpu(phys_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007457 SD_INIT(sd, CPU);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007458 set_domain_attribute(sd, attr);
Mike Travis7c16ec52008-04-04 18:11:11 -07007459 sd->span = *nodemask;
Linus Torvalds1da177e2005-04-16 15:20:36 -07007460 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007461 if (p)
7462 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007463 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007464
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007465#ifdef CONFIG_SCHED_MC
7466 p = sd;
7467 sd = &per_cpu(core_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007468 SD_INIT(sd, MC);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007469 set_domain_attribute(sd, attr);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007470 sd->span = cpu_coregroup_map(i);
7471 cpus_and(sd->span, sd->span, *cpu_map);
7472 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007473 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007474 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007475#endif
7476
Linus Torvalds1da177e2005-04-16 15:20:36 -07007477#ifdef CONFIG_SCHED_SMT
7478 p = sd;
7479 sd = &per_cpu(cpu_domains, i);
Mike Travis7c16ec52008-04-04 18:11:11 -07007480 SD_INIT(sd, SIBLING);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007481 set_domain_attribute(sd, attr);
Mike Travisd5a74302007-10-16 01:24:05 -07007482 sd->span = per_cpu(cpu_sibling_map, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007483 cpus_and(sd->span, sd->span, *cpu_map);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007484 sd->parent = p;
Siddha, Suresh B1a848872006-10-03 01:14:08 -07007485 p->child = sd;
Mike Travis7c16ec52008-04-04 18:11:11 -07007486 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007487#endif
7488 }
7489
7490#ifdef CONFIG_SCHED_SMT
7491 /* Set up CPU (sibling) groups */
Mike Travis363ab6f2008-05-12 21:21:13 +02007492 for_each_cpu_mask_nr(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007493 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7494 SCHED_CPUMASK_VAR(send_covered, allmasks);
7495
7496 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7497 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7498 if (i != first_cpu(*this_sibling_map))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007499 continue;
7500
Ingo Molnardd41f592007-07-09 18:51:59 +02007501 init_sched_build_groups(this_sibling_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007502 &cpu_to_cpu_group,
7503 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007504 }
7505#endif
7506
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007507#ifdef CONFIG_SCHED_MC
7508 /* Set up multi-core groups */
Mike Travis363ab6f2008-05-12 21:21:13 +02007509 for_each_cpu_mask_nr(i, *cpu_map) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007510 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7511 SCHED_CPUMASK_VAR(send_covered, allmasks);
7512
7513 *this_core_map = cpu_coregroup_map(i);
7514 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7515 if (i != first_cpu(*this_core_map))
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007516 continue;
Mike Travis7c16ec52008-04-04 18:11:11 -07007517
Ingo Molnardd41f592007-07-09 18:51:59 +02007518 init_sched_build_groups(this_core_map, cpu_map,
Mike Travis7c16ec52008-04-04 18:11:11 -07007519 &cpu_to_core_group,
7520 send_covered, tmpmask);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007521 }
7522#endif
7523
Linus Torvalds1da177e2005-04-16 15:20:36 -07007524 /* Set up physical groups */
Mike Travis076ac2a2008-05-12 21:21:12 +02007525 for (i = 0; i < nr_node_ids; i++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007526 SCHED_CPUMASK_VAR(nodemask, allmasks);
7527 SCHED_CPUMASK_VAR(send_covered, allmasks);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007528
Mike Travis7c16ec52008-04-04 18:11:11 -07007529 *nodemask = node_to_cpumask(i);
7530 cpus_and(*nodemask, *nodemask, *cpu_map);
7531 if (cpus_empty(*nodemask))
Linus Torvalds1da177e2005-04-16 15:20:36 -07007532 continue;
7533
Mike Travis7c16ec52008-04-04 18:11:11 -07007534 init_sched_build_groups(nodemask, cpu_map,
7535 &cpu_to_phys_group,
7536 send_covered, tmpmask);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007537 }
7538
7539#ifdef CONFIG_NUMA
7540 /* Set up node groups */
Mike Travis7c16ec52008-04-04 18:11:11 -07007541 if (sd_allnodes) {
7542 SCHED_CPUMASK_VAR(send_covered, allmasks);
7543
7544 init_sched_build_groups(cpu_map, cpu_map,
7545 &cpu_to_allnodes_group,
7546 send_covered, tmpmask);
7547 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007548
Mike Travis076ac2a2008-05-12 21:21:12 +02007549 for (i = 0; i < nr_node_ids; i++) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007550 /* Set up node groups */
7551 struct sched_group *sg, *prev;
Mike Travis7c16ec52008-04-04 18:11:11 -07007552 SCHED_CPUMASK_VAR(nodemask, allmasks);
7553 SCHED_CPUMASK_VAR(domainspan, allmasks);
7554 SCHED_CPUMASK_VAR(covered, allmasks);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007555 int j;
7556
Mike Travis7c16ec52008-04-04 18:11:11 -07007557 *nodemask = node_to_cpumask(i);
7558 cpus_clear(*covered);
7559
7560 cpus_and(*nodemask, *nodemask, *cpu_map);
7561 if (cpus_empty(*nodemask)) {
John Hawkesd1b55132005-09-06 15:18:14 -07007562 sched_group_nodes[i] = NULL;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007563 continue;
John Hawkesd1b55132005-09-06 15:18:14 -07007564 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007565
Mike Travis4bdbaad32008-04-15 16:35:52 -07007566 sched_domain_node_span(i, domainspan);
Mike Travis7c16ec52008-04-04 18:11:11 -07007567 cpus_and(*domainspan, *domainspan, *cpu_map);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007568
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007569 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007570 if (!sg) {
7571 printk(KERN_WARNING "Can not alloc domain group for "
7572 "node %d\n", i);
7573 goto error;
7574 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007575 sched_group_nodes[i] = sg;
Mike Travis363ab6f2008-05-12 21:21:13 +02007576 for_each_cpu_mask_nr(j, *nodemask) {
John Hawkes9c1cfda2005-09-06 15:18:14 -07007577 struct sched_domain *sd;
Ingo Molnar9761eea2007-07-09 18:52:00 +02007578
John Hawkes9c1cfda2005-09-06 15:18:14 -07007579 sd = &per_cpu(node_domains, j);
7580 sd->groups = sg;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007581 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007582 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007583 sg->cpumask = *nodemask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007584 sg->next = sg;
Mike Travis7c16ec52008-04-04 18:11:11 -07007585 cpus_or(*covered, *covered, *nodemask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007586 prev = sg;
7587
Mike Travis076ac2a2008-05-12 21:21:12 +02007588 for (j = 0; j < nr_node_ids; j++) {
Mike Travis7c16ec52008-04-04 18:11:11 -07007589 SCHED_CPUMASK_VAR(notcovered, allmasks);
Mike Travis076ac2a2008-05-12 21:21:12 +02007590 int n = (i + j) % nr_node_ids;
Mike Travisc5f59f02008-04-04 18:11:10 -07007591 node_to_cpumask_ptr(pnodemask, n);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007592
Mike Travis7c16ec52008-04-04 18:11:11 -07007593 cpus_complement(*notcovered, *covered);
7594 cpus_and(*tmpmask, *notcovered, *cpu_map);
7595 cpus_and(*tmpmask, *tmpmask, *domainspan);
7596 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007597 break;
7598
Mike Travis7c16ec52008-04-04 18:11:11 -07007599 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7600 if (cpus_empty(*tmpmask))
John Hawkes9c1cfda2005-09-06 15:18:14 -07007601 continue;
7602
Srivatsa Vaddagiri15f0b672006-06-27 02:54:40 -07007603 sg = kmalloc_node(sizeof(struct sched_group),
7604 GFP_KERNEL, i);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007605 if (!sg) {
7606 printk(KERN_WARNING
7607 "Can not alloc domain group for node %d\n", j);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007608 goto error;
John Hawkes9c1cfda2005-09-06 15:18:14 -07007609 }
Eric Dumazet5517d862007-05-08 00:32:57 -07007610 sg->__cpu_power = 0;
Mike Travis7c16ec52008-04-04 18:11:11 -07007611 sg->cpumask = *tmpmask;
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007612 sg->next = prev->next;
Mike Travis7c16ec52008-04-04 18:11:11 -07007613 cpus_or(*covered, *covered, *tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007614 prev->next = sg;
7615 prev = sg;
7616 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007617 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007618#endif
7619
7620 /* Calculate CPU power for physical packages and nodes */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007621#ifdef CONFIG_SCHED_SMT
Mike Travis363ab6f2008-05-12 21:21:13 +02007622 for_each_cpu_mask_nr(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007623 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7624
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007625 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007626 }
7627#endif
7628#ifdef CONFIG_SCHED_MC
Mike Travis363ab6f2008-05-12 21:21:13 +02007629 for_each_cpu_mask_nr(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007630 struct sched_domain *sd = &per_cpu(core_domains, i);
7631
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007632 init_sched_groups_power(i, sd);
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007633 }
7634#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007635
Mike Travis363ab6f2008-05-12 21:21:13 +02007636 for_each_cpu_mask_nr(i, *cpu_map) {
Ingo Molnardd41f592007-07-09 18:51:59 +02007637 struct sched_domain *sd = &per_cpu(phys_domains, i);
7638
Siddha, Suresh B89c47102006-10-03 01:14:09 -07007639 init_sched_groups_power(i, sd);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007640 }
7641
John Hawkes9c1cfda2005-09-06 15:18:14 -07007642#ifdef CONFIG_NUMA
Mike Travis076ac2a2008-05-12 21:21:12 +02007643 for (i = 0; i < nr_node_ids; i++)
Siddha, Suresh B08069032006-03-27 01:15:23 -08007644 init_numa_sched_groups_power(sched_group_nodes[i]);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007645
Siddha, Suresh B6711cab2006-12-10 02:20:07 -08007646 if (sd_allnodes) {
7647 struct sched_group *sg;
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007648
Mike Travis7c16ec52008-04-04 18:11:11 -07007649 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7650 tmpmask);
Siddha, Suresh Bf712c0c2006-07-30 03:02:59 -07007651 init_numa_sched_groups_power(sg);
7652 }
John Hawkes9c1cfda2005-09-06 15:18:14 -07007653#endif
7654
Linus Torvalds1da177e2005-04-16 15:20:36 -07007655 /* Attach the domains */
Mike Travis363ab6f2008-05-12 21:21:13 +02007656 for_each_cpu_mask_nr(i, *cpu_map) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007657 struct sched_domain *sd;
7658#ifdef CONFIG_SCHED_SMT
7659 sd = &per_cpu(cpu_domains, i);
Siddha, Suresh B1e9f28f2006-03-27 01:15:22 -08007660#elif defined(CONFIG_SCHED_MC)
7661 sd = &per_cpu(core_domains, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007662#else
7663 sd = &per_cpu(phys_domains, i);
7664#endif
Gregory Haskins57d885f2008-01-25 21:08:18 +01007665 cpu_attach_domain(sd, rd, i);
Linus Torvalds1da177e2005-04-16 15:20:36 -07007666 }
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007667
Mike Travis7c16ec52008-04-04 18:11:11 -07007668 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007669 return 0;
7670
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007671#ifdef CONFIG_NUMA
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007672error:
Mike Travis7c16ec52008-04-04 18:11:11 -07007673 free_sched_groups(cpu_map, tmpmask);
7674 SCHED_CPUMASK_FREE((void *)allmasks);
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007675 return -ENOMEM;
Siddha, Suresh Ba6160582006-10-03 01:14:06 -07007676#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07007677}
Paul Jackson029190c2007-10-18 23:40:20 -07007678
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007679static int build_sched_domains(const cpumask_t *cpu_map)
7680{
7681 return __build_sched_domains(cpu_map, NULL);
7682}
7683
Paul Jackson029190c2007-10-18 23:40:20 -07007684static cpumask_t *doms_cur; /* current sched domains */
7685static int ndoms_cur; /* number of sched domains in 'doms_cur' */
Ingo Molnar4285f5942008-05-16 17:47:14 +02007686static struct sched_domain_attr *dattr_cur;
7687 /* attribues of custom domains in 'doms_cur' */
Paul Jackson029190c2007-10-18 23:40:20 -07007688
7689/*
7690 * Special case: If a kmalloc of a doms_cur partition (array of
7691 * cpumask_t) fails, then fallback to a single sched domain,
7692 * as determined by the single cpumask_t fallback_doms.
7693 */
7694static cpumask_t fallback_doms;
7695
Heiko Carstens22e52b02008-03-12 18:31:59 +01007696void __attribute__((weak)) arch_update_cpu_topology(void)
7697{
7698}
7699
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007700/*
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007701 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
Paul Jackson029190c2007-10-18 23:40:20 -07007702 * For now this just excludes isolated cpus, but could be used to
7703 * exclude other special cases in the future.
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007704 */
Srivatsa Vaddagiri51888ca2006-06-27 02:54:38 -07007705static int arch_init_sched_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007706{
Milton Miller73785472007-10-24 18:23:48 +02007707 int err;
7708
Heiko Carstens22e52b02008-03-12 18:31:59 +01007709 arch_update_cpu_topology();
Paul Jackson029190c2007-10-18 23:40:20 -07007710 ndoms_cur = 1;
7711 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7712 if (!doms_cur)
7713 doms_cur = &fallback_doms;
7714 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007715 dattr_cur = NULL;
Milton Miller73785472007-10-24 18:23:48 +02007716 err = build_sched_domains(doms_cur);
Milton Miller6382bc92007-10-15 17:00:19 +02007717 register_sched_domain_sysctl();
Milton Miller73785472007-10-24 18:23:48 +02007718
7719 return err;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007720}
7721
Mike Travis7c16ec52008-04-04 18:11:11 -07007722static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7723 cpumask_t *tmpmask)
Linus Torvalds1da177e2005-04-16 15:20:36 -07007724{
Mike Travis7c16ec52008-04-04 18:11:11 -07007725 free_sched_groups(cpu_map, tmpmask);
John Hawkes9c1cfda2005-09-06 15:18:14 -07007726}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007727
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007728/*
7729 * Detach sched domains from a group of cpus specified in cpu_map
7730 * These cpus will now be attached to the NULL domain
7731 */
Arjan van de Ven858119e2006-01-14 13:20:43 -08007732static void detach_destroy_domains(const cpumask_t *cpu_map)
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007733{
Mike Travis7c16ec52008-04-04 18:11:11 -07007734 cpumask_t tmpmask;
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007735 int i;
7736
Milton Miller6382bc92007-10-15 17:00:19 +02007737 unregister_sched_domain_sysctl();
7738
Mike Travis363ab6f2008-05-12 21:21:13 +02007739 for_each_cpu_mask_nr(i, *cpu_map)
Gregory Haskins57d885f2008-01-25 21:08:18 +01007740 cpu_attach_domain(NULL, &def_root_domain, i);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007741 synchronize_sched();
Mike Travis7c16ec52008-04-04 18:11:11 -07007742 arch_destroy_sched_domains(cpu_map, &tmpmask);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007743}
7744
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007745/* handle null as "default" */
7746static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7747 struct sched_domain_attr *new, int idx_new)
7748{
7749 struct sched_domain_attr tmp;
7750
7751 /* fast path */
7752 if (!new && !cur)
7753 return 1;
7754
7755 tmp = SD_ATTR_INIT;
7756 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7757 new ? (new + idx_new) : &tmp,
7758 sizeof(struct sched_domain_attr));
7759}
7760
Paul Jackson029190c2007-10-18 23:40:20 -07007761/*
7762 * Partition sched domains as specified by the 'ndoms_new'
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007763 * cpumasks in the array doms_new[] of cpumasks. This compares
Paul Jackson029190c2007-10-18 23:40:20 -07007764 * doms_new[] to the current sched domain partitioning, doms_cur[].
7765 * It destroys each deleted domain and builds each new domain.
7766 *
7767 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007768 * The masks don't intersect (don't overlap.) We should setup one
7769 * sched domain for each mask. CPUs not in any of the cpumasks will
7770 * not be load balanced. If the same cpumask appears both in the
Paul Jackson029190c2007-10-18 23:40:20 -07007771 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7772 * it as it is.
7773 *
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01007774 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7775 * ownership of it and will kfree it when done with it. If the caller
Paul Jackson029190c2007-10-18 23:40:20 -07007776 * failed the kmalloc call, then it can pass in doms_new == NULL,
7777 * and partition_sched_domains() will fallback to the single partition
Max Krasnyanskye761b772008-07-15 04:43:49 -07007778 * 'fallback_doms', it also forces the domains to be rebuilt.
Paul Jackson029190c2007-10-18 23:40:20 -07007779 *
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007780 * If doms_new==NULL it will be replaced with cpu_online_map.
7781 * ndoms_new==0 is a special case for destroying existing domains.
7782 * It will not create the default domain.
7783 *
Paul Jackson029190c2007-10-18 23:40:20 -07007784 * Call with hotplug lock held
7785 */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007786void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7787 struct sched_domain_attr *dattr_new)
Paul Jackson029190c2007-10-18 23:40:20 -07007788{
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007789 int i, j, n;
Paul Jackson029190c2007-10-18 23:40:20 -07007790
Heiko Carstens712555e2008-04-28 11:33:07 +02007791 mutex_lock(&sched_domains_mutex);
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007792
Milton Miller73785472007-10-24 18:23:48 +02007793 /* always unregister in case we don't destroy any domains */
7794 unregister_sched_domain_sysctl();
7795
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007796 n = doms_new ? ndoms_new : 0;
Paul Jackson029190c2007-10-18 23:40:20 -07007797
7798 /* Destroy deleted domains */
7799 for (i = 0; i < ndoms_cur; i++) {
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007800 for (j = 0; j < n; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007801 if (cpus_equal(doms_cur[i], doms_new[j])
7802 && dattrs_equal(dattr_cur, i, dattr_new, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007803 goto match1;
7804 }
7805 /* no match - a current sched domain not in new doms_new[] */
7806 detach_destroy_domains(doms_cur + i);
7807match1:
7808 ;
7809 }
7810
Max Krasnyanskye761b772008-07-15 04:43:49 -07007811 if (doms_new == NULL) {
7812 ndoms_cur = 0;
Max Krasnyanskye761b772008-07-15 04:43:49 -07007813 doms_new = &fallback_doms;
7814 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7815 dattr_new = NULL;
7816 }
7817
Paul Jackson029190c2007-10-18 23:40:20 -07007818 /* Build new domains */
7819 for (i = 0; i < ndoms_new; i++) {
7820 for (j = 0; j < ndoms_cur; j++) {
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007821 if (cpus_equal(doms_new[i], doms_cur[j])
7822 && dattrs_equal(dattr_new, i, dattr_cur, j))
Paul Jackson029190c2007-10-18 23:40:20 -07007823 goto match2;
7824 }
7825 /* no match - add a new doms_new */
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007826 __build_sched_domains(doms_new + i,
7827 dattr_new ? dattr_new + i : NULL);
Paul Jackson029190c2007-10-18 23:40:20 -07007828match2:
7829 ;
7830 }
7831
7832 /* Remember the new sched domains */
7833 if (doms_cur != &fallback_doms)
7834 kfree(doms_cur);
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007835 kfree(dattr_cur); /* kfree(NULL) is safe */
Paul Jackson029190c2007-10-18 23:40:20 -07007836 doms_cur = doms_new;
Hidetoshi Seto1d3504f2008-04-15 14:04:23 +09007837 dattr_cur = dattr_new;
Paul Jackson029190c2007-10-18 23:40:20 -07007838 ndoms_cur = ndoms_new;
Milton Miller73785472007-10-24 18:23:48 +02007839
7840 register_sched_domain_sysctl();
Srivatsa Vaddagiria1835612008-01-25 21:08:00 +01007841
Heiko Carstens712555e2008-04-28 11:33:07 +02007842 mutex_unlock(&sched_domains_mutex);
Paul Jackson029190c2007-10-18 23:40:20 -07007843}
7844
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007845#if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
Heiko Carstens9aefd0a2008-03-12 18:31:58 +01007846int arch_reinit_sched_domains(void)
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007847{
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007848 get_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007849
7850 /* Destroy domains first to force the rebuild */
7851 partition_sched_domains(0, NULL, NULL);
7852
Max Krasnyanskye761b772008-07-15 04:43:49 -07007853 rebuild_sched_domains();
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007854 put_online_cpus();
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007855
Max Krasnyanskye761b772008-07-15 04:43:49 -07007856 return 0;
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007857}
7858
7859static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7860{
7861 int ret;
7862
7863 if (buf[0] != '0' && buf[0] != '1')
7864 return -EINVAL;
7865
7866 if (smt)
7867 sched_smt_power_savings = (buf[0] == '1');
7868 else
7869 sched_mc_power_savings = (buf[0] == '1');
7870
7871 ret = arch_reinit_sched_domains();
7872
7873 return ret ? ret : count;
7874}
7875
Adrian Bunk6707de002007-08-12 18:08:19 +02007876#ifdef CONFIG_SCHED_MC
Andi Kleenf718cd42008-07-29 22:33:52 -07007877static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7878 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02007879{
7880 return sprintf(page, "%u\n", sched_mc_power_savings);
7881}
Andi Kleenf718cd42008-07-29 22:33:52 -07007882static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
Adrian Bunk6707de002007-08-12 18:08:19 +02007883 const char *buf, size_t count)
7884{
7885 return sched_power_savings_store(buf, count, 0);
7886}
Andi Kleenf718cd42008-07-29 22:33:52 -07007887static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7888 sched_mc_power_savings_show,
7889 sched_mc_power_savings_store);
Adrian Bunk6707de002007-08-12 18:08:19 +02007890#endif
7891
7892#ifdef CONFIG_SCHED_SMT
Andi Kleenf718cd42008-07-29 22:33:52 -07007893static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7894 char *page)
Adrian Bunk6707de002007-08-12 18:08:19 +02007895{
7896 return sprintf(page, "%u\n", sched_smt_power_savings);
7897}
Andi Kleenf718cd42008-07-29 22:33:52 -07007898static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
Adrian Bunk6707de002007-08-12 18:08:19 +02007899 const char *buf, size_t count)
7900{
7901 return sched_power_savings_store(buf, count, 1);
7902}
Andi Kleenf718cd42008-07-29 22:33:52 -07007903static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7904 sched_smt_power_savings_show,
Adrian Bunk6707de002007-08-12 18:08:19 +02007905 sched_smt_power_savings_store);
7906#endif
7907
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007908int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7909{
7910 int err = 0;
Ingo Molnar48f24c42006-07-03 00:25:40 -07007911
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007912#ifdef CONFIG_SCHED_SMT
7913 if (smt_capable())
7914 err = sysfs_create_file(&cls->kset.kobj,
7915 &attr_sched_smt_power_savings.attr);
7916#endif
7917#ifdef CONFIG_SCHED_MC
7918 if (!err && mc_capable())
7919 err = sysfs_create_file(&cls->kset.kobj,
7920 &attr_sched_mc_power_savings.attr);
7921#endif
7922 return err;
7923}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02007924#endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
Siddha, Suresh B5c45bf22006-06-27 02:54:42 -07007925
Max Krasnyanskye761b772008-07-15 04:43:49 -07007926#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07007927/*
Max Krasnyanskye761b772008-07-15 04:43:49 -07007928 * Add online and remove offline CPUs from the scheduler domains.
7929 * When cpusets are enabled they take over this function.
Linus Torvalds1da177e2005-04-16 15:20:36 -07007930 */
7931static int update_sched_domains(struct notifier_block *nfb,
7932 unsigned long action, void *hcpu)
7933{
Max Krasnyanskye761b772008-07-15 04:43:49 -07007934 switch (action) {
7935 case CPU_ONLINE:
7936 case CPU_ONLINE_FROZEN:
7937 case CPU_DEAD:
7938 case CPU_DEAD_FROZEN:
Max Krasnyanskydfb512e2008-08-29 13:11:41 -07007939 partition_sched_domains(1, NULL, NULL);
Max Krasnyanskye761b772008-07-15 04:43:49 -07007940 return NOTIFY_OK;
7941
7942 default:
7943 return NOTIFY_DONE;
7944 }
7945}
7946#endif
7947
7948static int update_runtime(struct notifier_block *nfb,
7949 unsigned long action, void *hcpu)
7950{
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007951 int cpu = (int)(long)hcpu;
7952
Linus Torvalds1da177e2005-04-16 15:20:36 -07007953 switch (action) {
Linus Torvalds1da177e2005-04-16 15:20:36 -07007954 case CPU_DOWN_PREPARE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007955 case CPU_DOWN_PREPARE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007956 disable_runtime(cpu_rq(cpu));
Linus Torvalds1da177e2005-04-16 15:20:36 -07007957 return NOTIFY_OK;
7958
Linus Torvalds1da177e2005-04-16 15:20:36 -07007959 case CPU_DOWN_FAILED:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007960 case CPU_DOWN_FAILED_FROZEN:
Linus Torvalds1da177e2005-04-16 15:20:36 -07007961 case CPU_ONLINE:
Rafael J. Wysocki8bb78442007-05-09 02:35:10 -07007962 case CPU_ONLINE_FROZEN:
Peter Zijlstra7def2be2008-06-05 14:49:58 +02007963 enable_runtime(cpu_rq(cpu));
Max Krasnyanskye761b772008-07-15 04:43:49 -07007964 return NOTIFY_OK;
7965
Linus Torvalds1da177e2005-04-16 15:20:36 -07007966 default:
7967 return NOTIFY_DONE;
7968 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07007969}
Linus Torvalds1da177e2005-04-16 15:20:36 -07007970
7971void __init sched_init_smp(void)
7972{
Nick Piggin5c1e1762006-10-03 01:14:04 -07007973 cpumask_t non_isolated_cpus;
7974
Mike Travis434d53b2008-04-04 18:11:04 -07007975#if defined(CONFIG_NUMA)
7976 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7977 GFP_KERNEL);
7978 BUG_ON(sched_group_nodes_bycpu == NULL);
7979#endif
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007980 get_online_cpus();
Heiko Carstens712555e2008-04-28 11:33:07 +02007981 mutex_lock(&sched_domains_mutex);
Dinakar Guniguntala1a20ff22005-06-25 14:57:33 -07007982 arch_init_sched_domains(&cpu_online_map);
Nathan Lynche5e56732007-01-10 23:15:28 -08007983 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
Nick Piggin5c1e1762006-10-03 01:14:04 -07007984 if (cpus_empty(non_isolated_cpus))
7985 cpu_set(smp_processor_id(), non_isolated_cpus);
Heiko Carstens712555e2008-04-28 11:33:07 +02007986 mutex_unlock(&sched_domains_mutex);
Gautham R Shenoy95402b32008-01-25 21:08:02 +01007987 put_online_cpus();
Max Krasnyanskye761b772008-07-15 04:43:49 -07007988
7989#ifndef CONFIG_CPUSETS
Linus Torvalds1da177e2005-04-16 15:20:36 -07007990 /* XXX: Theoretical race here - CPU may be hotplugged now */
7991 hotcpu_notifier(update_sched_domains, 0);
Max Krasnyanskye761b772008-07-15 04:43:49 -07007992#endif
7993
7994 /* RT runtime code needs to handle some hotplug events */
7995 hotcpu_notifier(update_runtime, 0);
7996
Peter Zijlstrab328ca12008-04-29 10:02:46 +02007997 init_hrtick();
Nick Piggin5c1e1762006-10-03 01:14:04 -07007998
7999 /* Move init over to a non-isolated CPU */
Mike Travis7c16ec52008-04-04 18:11:11 -07008000 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
Nick Piggin5c1e1762006-10-03 01:14:04 -07008001 BUG();
Ingo Molnar19978ca2007-11-09 22:39:38 +01008002 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008003}
8004#else
8005void __init sched_init_smp(void)
8006{
Ingo Molnar19978ca2007-11-09 22:39:38 +01008007 sched_init_granularity();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008008}
8009#endif /* CONFIG_SMP */
8010
8011int in_sched_functions(unsigned long addr)
8012{
Linus Torvalds1da177e2005-04-16 15:20:36 -07008013 return in_lock_functions(addr) ||
8014 (addr >= (unsigned long)__sched_text_start
8015 && addr < (unsigned long)__sched_text_end);
8016}
8017
Alexey Dobriyana9957442007-10-15 17:00:13 +02008018static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
Ingo Molnardd41f592007-07-09 18:51:59 +02008019{
8020 cfs_rq->tasks_timeline = RB_ROOT;
Peter Zijlstra4a55bd52008-04-19 19:45:00 +02008021 INIT_LIST_HEAD(&cfs_rq->tasks);
Ingo Molnardd41f592007-07-09 18:51:59 +02008022#ifdef CONFIG_FAIR_GROUP_SCHED
8023 cfs_rq->rq = rq;
8024#endif
Peter Zijlstra67e9fb22007-10-15 17:00:10 +02008025 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
Ingo Molnardd41f592007-07-09 18:51:59 +02008026}
8027
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008028static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8029{
8030 struct rt_prio_array *array;
8031 int i;
8032
8033 array = &rt_rq->active;
8034 for (i = 0; i < MAX_RT_PRIO; i++) {
8035 INIT_LIST_HEAD(array->queue + i);
8036 __clear_bit(i, array->bitmap);
8037 }
8038 /* delimiter for bitsearch: */
8039 __set_bit(MAX_RT_PRIO, array->bitmap);
8040
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008041#if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
Peter Zijlstra48d5e252008-01-25 21:08:31 +01008042 rt_rq->highest_prio = MAX_RT_PRIO;
8043#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008044#ifdef CONFIG_SMP
8045 rt_rq->rt_nr_migratory = 0;
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008046 rt_rq->overloaded = 0;
8047#endif
8048
8049 rt_rq->rt_time = 0;
8050 rt_rq->rt_throttled = 0;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008051 rt_rq->rt_runtime = 0;
8052 spin_lock_init(&rt_rq->rt_runtime_lock);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008053
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008054#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra23b0fdf2008-02-13 15:45:39 +01008055 rt_rq->rt_nr_boosted = 0;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008056 rt_rq->rq = rq;
8057#endif
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008058}
8059
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008060#ifdef CONFIG_FAIR_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008061static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8062 struct sched_entity *se, int cpu, int add,
8063 struct sched_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008064{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008065 struct rq *rq = cpu_rq(cpu);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008066 tg->cfs_rq[cpu] = cfs_rq;
8067 init_cfs_rq(cfs_rq, rq);
8068 cfs_rq->tg = tg;
8069 if (add)
8070 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8071
8072 tg->se[cpu] = se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02008073 /* se could be NULL for init_task_group */
8074 if (!se)
8075 return;
8076
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008077 if (!parent)
8078 se->cfs_rq = &rq->cfs;
8079 else
8080 se->cfs_rq = parent->my_q;
8081
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008082 se->my_q = cfs_rq;
8083 se->load.weight = tg->shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02008084 se->load.inv_weight = 0;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008085 se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008086}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008087#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008088
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008089#ifdef CONFIG_RT_GROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008090static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8091 struct sched_rt_entity *rt_se, int cpu, int add,
8092 struct sched_rt_entity *parent)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008093{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008094 struct rq *rq = cpu_rq(cpu);
8095
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008096 tg->rt_rq[cpu] = rt_rq;
8097 init_rt_rq(rt_rq, rq);
8098 rt_rq->tg = tg;
8099 rt_rq->rt_se = rt_se;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008100 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008101 if (add)
8102 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8103
8104 tg->rt_se[cpu] = rt_se;
Dhaval Giani354d60c2008-04-19 19:44:59 +02008105 if (!rt_se)
8106 return;
8107
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008108 if (!parent)
8109 rt_se->rt_rq = &rq->rt;
8110 else
8111 rt_se->rt_rq = parent->my_q;
8112
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008113 rt_se->my_q = rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008114 rt_se->parent = parent;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008115 INIT_LIST_HEAD(&rt_se->run_list);
8116}
8117#endif
8118
Linus Torvalds1da177e2005-04-16 15:20:36 -07008119void __init sched_init(void)
8120{
Ingo Molnardd41f592007-07-09 18:51:59 +02008121 int i, j;
Mike Travis434d53b2008-04-04 18:11:04 -07008122 unsigned long alloc_size = 0, ptr;
8123
8124#ifdef CONFIG_FAIR_GROUP_SCHED
8125 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8126#endif
8127#ifdef CONFIG_RT_GROUP_SCHED
8128 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8129#endif
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008130#ifdef CONFIG_USER_SCHED
8131 alloc_size *= 2;
8132#endif
Mike Travis434d53b2008-04-04 18:11:04 -07008133 /*
8134 * As sched_init() is called before page_alloc is setup,
8135 * we use alloc_bootmem().
8136 */
8137 if (alloc_size) {
David Miller5a9d3222008-04-24 20:46:20 -07008138 ptr = (unsigned long)alloc_bootmem(alloc_size);
Mike Travis434d53b2008-04-04 18:11:04 -07008139
8140#ifdef CONFIG_FAIR_GROUP_SCHED
8141 init_task_group.se = (struct sched_entity **)ptr;
8142 ptr += nr_cpu_ids * sizeof(void **);
8143
8144 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8145 ptr += nr_cpu_ids * sizeof(void **);
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008146
8147#ifdef CONFIG_USER_SCHED
8148 root_task_group.se = (struct sched_entity **)ptr;
8149 ptr += nr_cpu_ids * sizeof(void **);
8150
8151 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8152 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008153#endif /* CONFIG_USER_SCHED */
8154#endif /* CONFIG_FAIR_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07008155#ifdef CONFIG_RT_GROUP_SCHED
8156 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8157 ptr += nr_cpu_ids * sizeof(void **);
8158
8159 init_task_group.rt_rq = (struct rt_rq **)ptr;
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008160 ptr += nr_cpu_ids * sizeof(void **);
8161
8162#ifdef CONFIG_USER_SCHED
8163 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8164 ptr += nr_cpu_ids * sizeof(void **);
8165
8166 root_task_group.rt_rq = (struct rt_rq **)ptr;
8167 ptr += nr_cpu_ids * sizeof(void **);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008168#endif /* CONFIG_USER_SCHED */
8169#endif /* CONFIG_RT_GROUP_SCHED */
Mike Travis434d53b2008-04-04 18:11:04 -07008170 }
Ingo Molnardd41f592007-07-09 18:51:59 +02008171
Gregory Haskins57d885f2008-01-25 21:08:18 +01008172#ifdef CONFIG_SMP
8173 init_defrootdomain();
8174#endif
8175
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008176 init_rt_bandwidth(&def_rt_bandwidth,
8177 global_rt_period(), global_rt_runtime());
8178
8179#ifdef CONFIG_RT_GROUP_SCHED
8180 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8181 global_rt_period(), global_rt_runtime());
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008182#ifdef CONFIG_USER_SCHED
8183 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8184 global_rt_period(), RUNTIME_INF);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008185#endif /* CONFIG_USER_SCHED */
8186#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008187
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008188#ifdef CONFIG_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008189 list_add(&init_task_group.list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008190 INIT_LIST_HEAD(&init_task_group.children);
8191
8192#ifdef CONFIG_USER_SCHED
8193 INIT_LIST_HEAD(&root_task_group.children);
8194 init_task_group.parent = &root_task_group;
8195 list_add(&init_task_group.siblings, &root_task_group.children);
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008196#endif /* CONFIG_USER_SCHED */
8197#endif /* CONFIG_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008198
KAMEZAWA Hiroyuki0a945022006-03-28 01:56:37 -08008199 for_each_possible_cpu(i) {
Ingo Molnar70b97a72006-07-03 00:25:42 -07008200 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008201
8202 rq = cpu_rq(i);
8203 spin_lock_init(&rq->lock);
Nick Piggin78979862005-06-25 14:57:13 -07008204 rq->nr_running = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008205 init_cfs_rq(&rq->cfs, rq);
Peter Zijlstrafa85ae22008-01-25 21:08:29 +01008206 init_rt_rq(&rq->rt, rq);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008207#ifdef CONFIG_FAIR_GROUP_SCHED
8208 init_task_group.shares = init_task_group_load;
8209 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008210#ifdef CONFIG_CGROUP_SCHED
8211 /*
8212 * How much cpu bandwidth does init_task_group get?
8213 *
8214 * In case of task-groups formed thr' the cgroup filesystem, it
8215 * gets 100% of the cpu resources in the system. This overall
8216 * system cpu resource is divided among the tasks of
8217 * init_task_group and its child task-groups in a fair manner,
8218 * based on each entity's (task or task-group's) weight
8219 * (se->load.weight).
8220 *
8221 * In other words, if init_task_group has 10 tasks of weight
8222 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8223 * then A0's share of the cpu resource is:
8224 *
8225 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8226 *
8227 * We achieve this by letting init_task_group's tasks sit
8228 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8229 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008230 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008231#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008232 root_task_group.shares = NICE_0_LOAD;
8233 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008234 /*
8235 * In case of task-groups formed thr' the user id of tasks,
8236 * init_task_group represents tasks belonging to root user.
8237 * Hence it forms a sibling of all subsequent groups formed.
8238 * In this case, init_task_group gets only a fraction of overall
8239 * system cpu resource, based on the weight assigned to root
8240 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8241 * by letting tasks of init_task_group sit in a separate cfs_rq
8242 * (init_cfs_rq) and having one entity represent this group of
8243 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8244 */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008245 init_tg_cfs_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008246 &per_cpu(init_cfs_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008247 &per_cpu(init_sched_entity, i), i, 1,
8248 root_task_group.se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008249
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008250#endif
Dhaval Giani354d60c2008-04-19 19:44:59 +02008251#endif /* CONFIG_FAIR_GROUP_SCHED */
8252
8253 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008254#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008255 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008256#ifdef CONFIG_CGROUP_SCHED
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008257 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008258#elif defined CONFIG_USER_SCHED
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008259 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008260 init_tg_rt_entry(&init_task_group,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008261 &per_cpu(init_rt_rq, i),
Peter Zijlstraeff766a2008-04-19 19:45:00 +02008262 &per_cpu(init_sched_rt_entity, i), i, 1,
8263 root_task_group.rt_se[i]);
Dhaval Giani354d60c2008-04-19 19:44:59 +02008264#endif
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008265#endif
Linus Torvalds1da177e2005-04-16 15:20:36 -07008266
Ingo Molnardd41f592007-07-09 18:51:59 +02008267 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8268 rq->cpu_load[j] = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008269#ifdef CONFIG_SMP
Nick Piggin41c7ce92005-06-25 14:57:24 -07008270 rq->sd = NULL;
Gregory Haskins57d885f2008-01-25 21:08:18 +01008271 rq->rd = NULL;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008272 rq->active_balance = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008273 rq->next_balance = jiffies;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008274 rq->push_cpu = 0;
Christoph Lameter0a2966b2006-09-25 23:30:51 -07008275 rq->cpu = i;
Gregory Haskins1f11eb62008-06-04 15:04:05 -04008276 rq->online = 0;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008277 rq->migration_thread = NULL;
8278 INIT_LIST_HEAD(&rq->migration_queue);
Gregory Haskinsdc938522008-01-25 21:08:26 +01008279 rq_attach_root(rq, &def_root_domain);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008280#endif
Peter Zijlstra8f4d37e2008-01-25 21:08:29 +01008281 init_rq_hrtick(rq);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008282 atomic_set(&rq->nr_iowait, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008283 }
8284
Peter Williams2dd73a42006-06-27 02:54:34 -07008285 set_load_weight(&init_task);
Heiko Carstensb50f60c2006-07-30 03:03:52 -07008286
Avi Kivitye107be32007-07-26 13:40:43 +02008287#ifdef CONFIG_PREEMPT_NOTIFIERS
8288 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8289#endif
8290
Christoph Lameterc9819f42006-12-10 02:20:25 -08008291#ifdef CONFIG_SMP
Carlos R. Mafra962cf362008-05-15 11:15:37 -03008292 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
Christoph Lameterc9819f42006-12-10 02:20:25 -08008293#endif
8294
Heiko Carstensb50f60c2006-07-30 03:03:52 -07008295#ifdef CONFIG_RT_MUTEXES
8296 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8297#endif
8298
Linus Torvalds1da177e2005-04-16 15:20:36 -07008299 /*
8300 * The boot idle thread does lazy MMU switching as well:
8301 */
8302 atomic_inc(&init_mm.mm_count);
8303 enter_lazy_tlb(&init_mm, current);
8304
8305 /*
8306 * Make us the idle thread. Technically, schedule() should not be
8307 * called from this thread, however somewhere below it might be,
8308 * but because we are the idle thread, we just pick up running again
8309 * when this runqueue becomes "idle".
8310 */
8311 init_idle(current, smp_processor_id());
Ingo Molnardd41f592007-07-09 18:51:59 +02008312 /*
8313 * During early bootup we pretend to be a normal task:
8314 */
8315 current->sched_class = &fair_sched_class;
Ingo Molnar6892b752008-02-13 14:02:36 +01008316
8317 scheduler_running = 1;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008318}
8319
8320#ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8321void __might_sleep(char *file, int line)
8322{
Ingo Molnar48f24c42006-07-03 00:25:40 -07008323#ifdef in_atomic
Linus Torvalds1da177e2005-04-16 15:20:36 -07008324 static unsigned long prev_jiffy; /* ratelimiting */
8325
Ingo Molnaraef745f2008-08-28 11:34:43 +02008326 if ((!in_atomic() && !irqs_disabled()) ||
8327 system_state != SYSTEM_RUNNING || oops_in_progress)
8328 return;
8329 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8330 return;
8331 prev_jiffy = jiffies;
8332
8333 printk(KERN_ERR
8334 "BUG: sleeping function called from invalid context at %s:%d\n",
8335 file, line);
8336 printk(KERN_ERR
8337 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8338 in_atomic(), irqs_disabled(),
8339 current->pid, current->comm);
8340
8341 debug_show_held_locks(current);
8342 if (irqs_disabled())
8343 print_irqtrace_events(current);
8344 dump_stack();
Linus Torvalds1da177e2005-04-16 15:20:36 -07008345#endif
8346}
8347EXPORT_SYMBOL(__might_sleep);
8348#endif
8349
8350#ifdef CONFIG_MAGIC_SYSRQ
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008351static void normalize_task(struct rq *rq, struct task_struct *p)
8352{
8353 int on_rq;
Peter Zijlstra3e51f332008-05-03 18:29:28 +02008354
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008355 update_rq_clock(rq);
8356 on_rq = p->se.on_rq;
8357 if (on_rq)
8358 deactivate_task(rq, p, 0);
8359 __setscheduler(rq, p, SCHED_NORMAL, 0);
8360 if (on_rq) {
8361 activate_task(rq, p, 0);
8362 resched_task(rq->curr);
8363 }
8364}
8365
Linus Torvalds1da177e2005-04-16 15:20:36 -07008366void normalize_rt_tasks(void)
8367{
Ingo Molnara0f98a12007-06-17 18:37:45 +02008368 struct task_struct *g, *p;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008369 unsigned long flags;
Ingo Molnar70b97a72006-07-03 00:25:42 -07008370 struct rq *rq;
Linus Torvalds1da177e2005-04-16 15:20:36 -07008371
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008372 read_lock_irqsave(&tasklist_lock, flags);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008373 do_each_thread(g, p) {
Ingo Molnar178be792007-10-15 17:00:18 +02008374 /*
8375 * Only normalize user tasks:
8376 */
8377 if (!p->mm)
8378 continue;
8379
Ingo Molnardd41f592007-07-09 18:51:59 +02008380 p->se.exec_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008381#ifdef CONFIG_SCHEDSTATS
8382 p->se.wait_start = 0;
8383 p->se.sleep_start = 0;
Ingo Molnardd41f592007-07-09 18:51:59 +02008384 p->se.block_start = 0;
Ingo Molnar6cfb0d52007-08-02 17:41:40 +02008385#endif
Ingo Molnardd41f592007-07-09 18:51:59 +02008386
8387 if (!rt_task(p)) {
8388 /*
8389 * Renice negative nice level userspace
8390 * tasks back to 0:
8391 */
8392 if (TASK_NICE(p) < 0 && p->mm)
8393 set_user_nice(p, 0);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008394 continue;
Ingo Molnardd41f592007-07-09 18:51:59 +02008395 }
Linus Torvalds1da177e2005-04-16 15:20:36 -07008396
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008397 spin_lock(&p->pi_lock);
Ingo Molnarb29739f2006-06-27 02:54:51 -07008398 rq = __task_rq_lock(p);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008399
Ingo Molnar178be792007-10-15 17:00:18 +02008400 normalize_task(rq, p);
Andi Kleen3a5e4dc2007-10-15 17:00:15 +02008401
Ingo Molnarb29739f2006-06-27 02:54:51 -07008402 __task_rq_unlock(rq);
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008403 spin_unlock(&p->pi_lock);
Ingo Molnara0f98a12007-06-17 18:37:45 +02008404 } while_each_thread(g, p);
8405
Peter Zijlstra4cf5d772008-02-13 15:45:39 +01008406 read_unlock_irqrestore(&tasklist_lock, flags);
Linus Torvalds1da177e2005-04-16 15:20:36 -07008407}
8408
8409#endif /* CONFIG_MAGIC_SYSRQ */
Linus Torvalds1df5c102005-09-12 07:59:21 -07008410
8411#ifdef CONFIG_IA64
8412/*
8413 * These functions are only useful for the IA64 MCA handling.
8414 *
8415 * They can only be called when the whole system has been
8416 * stopped - every CPU needs to be quiescent, and no scheduling
8417 * activity can take place. Using them for anything else would
8418 * be a serious bug, and as a result, they aren't even visible
8419 * under any other configuration.
8420 */
8421
8422/**
8423 * curr_task - return the current task for a given cpu.
8424 * @cpu: the processor in question.
8425 *
8426 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8427 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008428struct task_struct *curr_task(int cpu)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008429{
8430 return cpu_curr(cpu);
8431}
8432
8433/**
8434 * set_curr_task - set the current task for a given cpu.
8435 * @cpu: the processor in question.
8436 * @p: the task pointer to set.
8437 *
8438 * Description: This function must only be used when non-maskable interrupts
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01008439 * are serviced on a separate stack. It allows the architecture to switch the
8440 * notion of the current task on a cpu in a non-blocking manner. This function
Linus Torvalds1df5c102005-09-12 07:59:21 -07008441 * must be called with all CPU's synchronized, and interrupts disabled, the
8442 * and caller must save the original value of the current task (see
8443 * curr_task() above) and restore that value before reenabling interrupts and
8444 * re-starting the system.
8445 *
8446 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8447 */
Ingo Molnar36c8b582006-07-03 00:25:41 -07008448void set_curr_task(int cpu, struct task_struct *p)
Linus Torvalds1df5c102005-09-12 07:59:21 -07008449{
8450 cpu_curr(cpu) = p;
8451}
8452
8453#endif
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008454
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008455#ifdef CONFIG_FAIR_GROUP_SCHED
8456static void free_fair_sched_group(struct task_group *tg)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008457{
8458 int i;
8459
8460 for_each_possible_cpu(i) {
8461 if (tg->cfs_rq)
8462 kfree(tg->cfs_rq[i]);
8463 if (tg->se)
8464 kfree(tg->se[i]);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008465 }
8466
8467 kfree(tg->cfs_rq);
8468 kfree(tg->se);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008469}
8470
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008471static
8472int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008473{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008474 struct cfs_rq *cfs_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008475 struct sched_entity *se, *parent_se;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008476 struct rq *rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008477 int i;
8478
Mike Travis434d53b2008-04-04 18:11:04 -07008479 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008480 if (!tg->cfs_rq)
8481 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008482 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008483 if (!tg->se)
8484 goto err;
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008485
8486 tg->shares = NICE_0_LOAD;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008487
8488 for_each_possible_cpu(i) {
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008489 rq = cpu_rq(i);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008490
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008491 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8492 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008493 if (!cfs_rq)
8494 goto err;
8495
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008496 se = kmalloc_node(sizeof(struct sched_entity),
8497 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008498 if (!se)
8499 goto err;
8500
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008501 parent_se = parent ? parent->se[i] : NULL;
8502 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008503 }
8504
8505 return 1;
8506
8507 err:
8508 return 0;
8509}
8510
8511static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8512{
8513 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8514 &cpu_rq(cpu)->leaf_cfs_rq_list);
8515}
8516
8517static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8518{
8519 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8520}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008521#else /* !CONFG_FAIR_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008522static inline void free_fair_sched_group(struct task_group *tg)
8523{
8524}
8525
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008526static inline
8527int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008528{
8529 return 1;
8530}
8531
8532static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8533{
8534}
8535
8536static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8537{
8538}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008539#endif /* CONFIG_FAIR_GROUP_SCHED */
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008540
8541#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008542static void free_rt_sched_group(struct task_group *tg)
8543{
8544 int i;
8545
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008546 destroy_rt_bandwidth(&tg->rt_bandwidth);
8547
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008548 for_each_possible_cpu(i) {
8549 if (tg->rt_rq)
8550 kfree(tg->rt_rq[i]);
8551 if (tg->rt_se)
8552 kfree(tg->rt_se[i]);
8553 }
8554
8555 kfree(tg->rt_rq);
8556 kfree(tg->rt_se);
8557}
8558
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008559static
8560int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008561{
8562 struct rt_rq *rt_rq;
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008563 struct sched_rt_entity *rt_se, *parent_se;
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008564 struct rq *rq;
8565 int i;
8566
Mike Travis434d53b2008-04-04 18:11:04 -07008567 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008568 if (!tg->rt_rq)
8569 goto err;
Mike Travis434d53b2008-04-04 18:11:04 -07008570 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008571 if (!tg->rt_se)
8572 goto err;
8573
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008574 init_rt_bandwidth(&tg->rt_bandwidth,
8575 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008576
8577 for_each_possible_cpu(i) {
8578 rq = cpu_rq(i);
8579
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008580 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8581 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8582 if (!rt_rq)
8583 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008584
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008585 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8586 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8587 if (!rt_se)
8588 goto err;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008589
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008590 parent_se = parent ? parent->rt_se[i] : NULL;
8591 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008592 }
8593
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008594 return 1;
8595
8596 err:
8597 return 0;
8598}
8599
8600static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8601{
8602 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8603 &cpu_rq(cpu)->leaf_rt_rq_list);
8604}
8605
8606static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8607{
8608 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8609}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008610#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008611static inline void free_rt_sched_group(struct task_group *tg)
8612{
8613}
8614
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008615static inline
8616int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008617{
8618 return 1;
8619}
8620
8621static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8622{
8623}
8624
8625static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8626{
8627}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008628#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008629
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008630#ifdef CONFIG_GROUP_SCHED
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008631static void free_sched_group(struct task_group *tg)
8632{
8633 free_fair_sched_group(tg);
8634 free_rt_sched_group(tg);
8635 kfree(tg);
8636}
8637
8638/* allocate runqueue etc for a new task group */
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008639struct task_group *sched_create_group(struct task_group *parent)
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008640{
8641 struct task_group *tg;
8642 unsigned long flags;
8643 int i;
8644
8645 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8646 if (!tg)
8647 return ERR_PTR(-ENOMEM);
8648
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008649 if (!alloc_fair_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008650 goto err;
8651
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008652 if (!alloc_rt_sched_group(tg, parent))
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008653 goto err;
8654
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008655 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008656 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008657 register_fair_sched_group(tg, i);
8658 register_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008659 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008660 list_add_rcu(&tg->list, &task_groups);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008661
8662 WARN_ON(!parent); /* root should already exist */
8663
8664 tg->parent = parent;
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008665 INIT_LIST_HEAD(&tg->children);
Zhang, Yanmin09f27242030-08-14 15:56:40 +08008666 list_add_rcu(&tg->siblings, &parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008667 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008668
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008669 return tg;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008670
8671err:
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008672 free_sched_group(tg);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008673 return ERR_PTR(-ENOMEM);
8674}
8675
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008676/* rcu callback to free various structures associated with a task group */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008677static void free_sched_group_rcu(struct rcu_head *rhp)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008678{
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008679 /* now it should be safe to free those cfs_rqs */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008680 free_sched_group(container_of(rhp, struct task_group, rcu));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008681}
8682
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008683/* Destroy runqueue etc associated with a task group */
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008684void sched_destroy_group(struct task_group *tg)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008685{
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008686 unsigned long flags;
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008687 int i;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008688
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008689 spin_lock_irqsave(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008690 for_each_possible_cpu(i) {
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008691 unregister_fair_sched_group(tg, i);
8692 unregister_rt_sched_group(tg, i);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008693 }
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008694 list_del_rcu(&tg->list);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008695 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008696 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008697
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008698 /* wait for possible concurrent references to cfs_rqs complete */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008699 call_rcu(&tg->rcu, free_sched_group_rcu);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008700}
8701
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008702/* change task's runqueue when it moves between groups.
Ingo Molnar3a252012007-10-15 17:00:12 +02008703 * The caller of this function should have put the task in its new group
8704 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8705 * reflect its new group.
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008706 */
8707void sched_move_task(struct task_struct *tsk)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008708{
8709 int on_rq, running;
8710 unsigned long flags;
8711 struct rq *rq;
8712
8713 rq = task_rq_lock(tsk, &flags);
8714
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008715 update_rq_clock(rq);
8716
Dmitry Adamushko051a1d12007-12-18 15:21:13 +01008717 running = task_current(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008718 on_rq = tsk->se.on_rq;
8719
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008720 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008721 dequeue_task(rq, tsk, 0);
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008722 if (unlikely(running))
8723 tsk->sched_class->put_prev_task(rq, tsk);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008724
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008725 set_task_rq(tsk, task_cpu(tsk));
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008726
Peter Zijlstra810b3812008-02-29 15:21:01 -05008727#ifdef CONFIG_FAIR_GROUP_SCHED
8728 if (tsk->sched_class->moved_group)
8729 tsk->sched_class->moved_group(tsk);
8730#endif
8731
Hiroshi Shimamoto0e1f3482008-03-10 11:01:20 -07008732 if (unlikely(running))
8733 tsk->sched_class->set_curr_task(rq);
8734 if (on_rq)
Dmitry Adamushko7074bad2007-10-15 17:00:07 +02008735 enqueue_task(rq, tsk, 0);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008736
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008737 task_rq_unlock(rq, &flags);
8738}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02008739#endif /* CONFIG_GROUP_SCHED */
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008740
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008741#ifdef CONFIG_FAIR_GROUP_SCHED
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008742static void __set_se_shares(struct sched_entity *se, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008743{
8744 struct cfs_rq *cfs_rq = se->cfs_rq;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008745 int on_rq;
8746
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008747 on_rq = se->on_rq;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008748 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008749 dequeue_entity(cfs_rq, se, 0);
8750
8751 se->load.weight = shares;
Peter Zijlstrae05510d2008-05-05 23:56:17 +02008752 se->load.inv_weight = 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008753
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008754 if (on_rq)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008755 enqueue_entity(cfs_rq, se, 0);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008756}
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008757
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008758static void set_se_shares(struct sched_entity *se, unsigned long shares)
8759{
8760 struct cfs_rq *cfs_rq = se->cfs_rq;
8761 struct rq *rq = cfs_rq->rq;
8762 unsigned long flags;
8763
8764 spin_lock_irqsave(&rq->lock, flags);
8765 __set_se_shares(se, shares);
8766 spin_unlock_irqrestore(&rq->lock, flags);
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008767}
8768
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008769static DEFINE_MUTEX(shares_mutex);
8770
Ingo Molnar4cf86d72007-10-15 17:00:14 +02008771int sched_group_set_shares(struct task_group *tg, unsigned long shares)
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008772{
8773 int i;
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008774 unsigned long flags;
Ingo Molnarc61935f2008-01-22 11:24:58 +01008775
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008776 /*
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02008777 * We can't change the weight of the root cgroup.
8778 */
8779 if (!tg->se[0])
8780 return -EINVAL;
8781
Peter Zijlstra18d95a22008-04-19 19:45:00 +02008782 if (shares < MIN_SHARES)
8783 shares = MIN_SHARES;
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008784 else if (shares > MAX_SHARES)
8785 shares = MAX_SHARES;
Peter Zijlstra62fb1852008-02-25 17:34:02 +01008786
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008787 mutex_lock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008788 if (tg->shares == shares)
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008789 goto done;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008790
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008791 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008792 for_each_possible_cpu(i)
8793 unregister_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008794 list_del_rcu(&tg->siblings);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008795 spin_unlock_irqrestore(&task_group_lock, flags);
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008796
8797 /* wait for any ongoing reference to this group to finish */
8798 synchronize_sched();
8799
8800 /*
8801 * Now we are free to modify the group's share on each cpu
8802 * w/o tripping rebalance_share or load_balance_fair.
8803 */
8804 tg->shares = shares;
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008805 for_each_possible_cpu(i) {
8806 /*
8807 * force a rebalance
8808 */
8809 cfs_rq_set_shares(tg->cfs_rq[i], 0);
Miao Xiecb4ad1f2008-04-28 12:54:56 +08008810 set_se_shares(tg->se[i], shares);
Peter Zijlstrac09595f2008-06-27 13:41:14 +02008811 }
Srivatsa Vaddagiri6b2d7702008-01-25 21:08:00 +01008812
8813 /*
8814 * Enable load balance activity on this group, by inserting it back on
8815 * each cpu's rq->leaf_cfs_rq_list.
8816 */
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008817 spin_lock_irqsave(&task_group_lock, flags);
Peter Zijlstrabccbe082008-02-13 15:45:40 +01008818 for_each_possible_cpu(i)
8819 register_fair_sched_group(tg, i);
Peter Zijlstraf473aa52008-04-19 19:45:00 +02008820 list_add_rcu(&tg->siblings, &tg->parent->children);
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008821 spin_unlock_irqrestore(&task_group_lock, flags);
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008822done:
Peter Zijlstra8ed36992008-02-13 15:45:39 +01008823 mutex_unlock(&shares_mutex);
Srivatsa Vaddagiri9b5b7752007-10-15 17:00:09 +02008824 return 0;
Srivatsa Vaddagiri29f59db2007-10-15 17:00:07 +02008825}
8826
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008827unsigned long sched_group_shares(struct task_group *tg)
8828{
8829 return tg->shares;
8830}
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008831#endif
Dhaval Giani5cb350b2007-10-15 17:00:14 +02008832
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01008833#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008834/*
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008835 * Ensure that the real time constraints are schedulable.
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008836 */
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008837static DEFINE_MUTEX(rt_constraints_mutex);
8838
8839static unsigned long to_ratio(u64 period, u64 runtime)
8840{
8841 if (runtime == RUNTIME_INF)
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008842 return 1ULL << 20;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008843
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008844 return div64_u64(runtime << 20, period);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008845}
8846
Dhaval Giani521f1a242008-02-28 15:21:56 +05308847/* Must be called with tasklist_lock held */
8848static inline int tg_has_rt_tasks(struct task_group *tg)
8849{
8850 struct task_struct *g, *p;
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008851
Dhaval Giani521f1a242008-02-28 15:21:56 +05308852 do_each_thread(g, p) {
8853 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8854 return 1;
8855 } while_each_thread(g, p);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008856
Dhaval Giani521f1a242008-02-28 15:21:56 +05308857 return 0;
8858}
8859
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008860struct rt_schedulable_data {
8861 struct task_group *tg;
8862 u64 rt_period;
8863 u64 rt_runtime;
8864};
8865
8866static int tg_schedulable(struct task_group *tg, void *data)
8867{
8868 struct rt_schedulable_data *d = data;
8869 struct task_group *child;
8870 unsigned long total, sum = 0;
8871 u64 period, runtime;
8872
8873 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8874 runtime = tg->rt_bandwidth.rt_runtime;
8875
8876 if (tg == d->tg) {
8877 period = d->rt_period;
8878 runtime = d->rt_runtime;
8879 }
8880
Peter Zijlstra4653f802008-09-23 15:33:44 +02008881 /*
8882 * Cannot have more runtime than the period.
8883 */
8884 if (runtime > period && runtime != RUNTIME_INF)
8885 return -EINVAL;
8886
8887 /*
8888 * Ensure we don't starve existing RT tasks.
8889 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008890 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8891 return -EBUSY;
8892
8893 total = to_ratio(period, runtime);
8894
Peter Zijlstra4653f802008-09-23 15:33:44 +02008895 /*
8896 * Nobody can have more than the global setting allows.
8897 */
8898 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8899 return -EINVAL;
8900
8901 /*
8902 * The sum of our children's runtime should not exceed our own.
8903 */
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008904 list_for_each_entry_rcu(child, &tg->children, siblings) {
8905 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8906 runtime = child->rt_bandwidth.rt_runtime;
8907
8908 if (child == d->tg) {
8909 period = d->rt_period;
8910 runtime = d->rt_runtime;
8911 }
8912
8913 sum += to_ratio(period, runtime);
8914 }
8915
8916 if (sum > total)
8917 return -EINVAL;
8918
8919 return 0;
8920}
8921
8922static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8923{
8924 struct rt_schedulable_data data = {
8925 .tg = tg,
8926 .rt_period = period,
8927 .rt_runtime = runtime,
8928 };
8929
8930 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8931}
8932
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008933static int tg_set_bandwidth(struct task_group *tg,
8934 u64 rt_period, u64 rt_runtime)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008935{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008936 int i, err = 0;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008937
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008938 mutex_lock(&rt_constraints_mutex);
Dhaval Giani521f1a242008-02-28 15:21:56 +05308939 read_lock(&tasklist_lock);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02008940 err = __rt_schedulable(tg, rt_period, rt_runtime);
8941 if (err)
Dhaval Giani521f1a242008-02-28 15:21:56 +05308942 goto unlock;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008943
8944 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008945 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8946 tg->rt_bandwidth.rt_runtime = rt_runtime;
Peter Zijlstraac086bc2008-04-19 19:44:58 +02008947
8948 for_each_possible_cpu(i) {
8949 struct rt_rq *rt_rq = tg->rt_rq[i];
8950
8951 spin_lock(&rt_rq->rt_runtime_lock);
8952 rt_rq->rt_runtime = rt_runtime;
8953 spin_unlock(&rt_rq->rt_runtime_lock);
8954 }
8955 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008956 unlock:
Dhaval Giani521f1a242008-02-28 15:21:56 +05308957 read_unlock(&tasklist_lock);
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008958 mutex_unlock(&rt_constraints_mutex);
8959
8960 return err;
Peter Zijlstra6f505b12008-01-25 21:08:30 +01008961}
8962
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008963int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8964{
8965 u64 rt_runtime, rt_period;
8966
8967 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8968 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8969 if (rt_runtime_us < 0)
8970 rt_runtime = RUNTIME_INF;
8971
8972 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8973}
8974
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008975long sched_group_rt_runtime(struct task_group *tg)
8976{
8977 u64 rt_runtime_us;
8978
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008979 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008980 return -1;
8981
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008982 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01008983 do_div(rt_runtime_us, NSEC_PER_USEC);
8984 return rt_runtime_us;
8985}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008986
8987int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8988{
8989 u64 rt_runtime, rt_period;
8990
8991 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8992 rt_runtime = tg->rt_bandwidth.rt_runtime;
8993
Raistlin619b0482008-06-26 18:54:09 +02008994 if (rt_period == 0)
8995 return -EINVAL;
8996
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02008997 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8998}
8999
9000long sched_group_rt_period(struct task_group *tg)
9001{
9002 u64 rt_period_us;
9003
9004 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9005 do_div(rt_period_us, NSEC_PER_USEC);
9006 return rt_period_us;
9007}
9008
9009static int sched_rt_global_constraints(void)
9010{
Peter Zijlstra4653f802008-09-23 15:33:44 +02009011 u64 runtime, period;
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009012 int ret = 0;
9013
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -07009014 if (sysctl_sched_rt_period <= 0)
9015 return -EINVAL;
9016
Peter Zijlstra4653f802008-09-23 15:33:44 +02009017 runtime = global_rt_runtime();
9018 period = global_rt_period();
9019
9020 /*
9021 * Sanity check on the sysctl variables.
9022 */
9023 if (runtime > period && runtime != RUNTIME_INF)
9024 return -EINVAL;
Peter Zijlstra10b612f2008-06-19 14:22:27 +02009025
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009026 mutex_lock(&rt_constraints_mutex);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009027 read_lock(&tasklist_lock);
Peter Zijlstra4653f802008-09-23 15:33:44 +02009028 ret = __rt_schedulable(NULL, 0, 0);
Peter Zijlstra9a7e0b12008-08-19 12:33:06 +02009029 read_unlock(&tasklist_lock);
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009030 mutex_unlock(&rt_constraints_mutex);
9031
9032 return ret;
9033}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009034#else /* !CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009035static int sched_rt_global_constraints(void)
9036{
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009037 unsigned long flags;
9038 int i;
9039
Hiroshi Shimamotoec5d4982008-09-10 17:00:19 -07009040 if (sysctl_sched_rt_period <= 0)
9041 return -EINVAL;
9042
Peter Zijlstraac086bc2008-04-19 19:44:58 +02009043 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9044 for_each_possible_cpu(i) {
9045 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9046
9047 spin_lock(&rt_rq->rt_runtime_lock);
9048 rt_rq->rt_runtime = global_rt_runtime();
9049 spin_unlock(&rt_rq->rt_runtime_lock);
9050 }
9051 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9052
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009053 return 0;
9054}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009055#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009056
9057int sched_rt_handler(struct ctl_table *table, int write,
9058 struct file *filp, void __user *buffer, size_t *lenp,
9059 loff_t *ppos)
9060{
9061 int ret;
9062 int old_period, old_runtime;
9063 static DEFINE_MUTEX(mutex);
9064
9065 mutex_lock(&mutex);
9066 old_period = sysctl_sched_rt_period;
9067 old_runtime = sysctl_sched_rt_runtime;
9068
9069 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9070
9071 if (!ret && write) {
9072 ret = sched_rt_global_constraints();
9073 if (ret) {
9074 sysctl_sched_rt_period = old_period;
9075 sysctl_sched_rt_runtime = old_runtime;
9076 } else {
9077 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9078 def_rt_bandwidth.rt_period =
9079 ns_to_ktime(global_rt_period());
9080 }
9081 }
9082 mutex_unlock(&mutex);
9083
9084 return ret;
9085}
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009086
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009087#ifdef CONFIG_CGROUP_SCHED
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009088
9089/* return corresponding task_group object of a cgroup */
Paul Menage2b01dfe2007-10-24 18:23:50 +02009090static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009091{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009092 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9093 struct task_group, css);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009094}
9095
9096static struct cgroup_subsys_state *
Paul Menage2b01dfe2007-10-24 18:23:50 +02009097cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009098{
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009099 struct task_group *tg, *parent;
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009100
Paul Menage2b01dfe2007-10-24 18:23:50 +02009101 if (!cgrp->parent) {
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009102 /* This is early initialization for the top cgroup */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009103 return &init_task_group.css;
9104 }
9105
Dhaval Gianiec7dc8a2008-04-19 19:44:59 +02009106 parent = cgroup_tg(cgrp->parent);
9107 tg = sched_create_group(parent);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009108 if (IS_ERR(tg))
9109 return ERR_PTR(-ENOMEM);
9110
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009111 return &tg->css;
9112}
9113
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009114static void
9115cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009116{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009117 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009118
9119 sched_destroy_group(tg);
9120}
9121
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009122static int
9123cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9124 struct task_struct *tsk)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009125{
Peter Zijlstrab68aa232008-02-13 15:45:40 +01009126#ifdef CONFIG_RT_GROUP_SCHED
9127 /* Don't accept realtime tasks when there is no way for them to run */
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009128 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
Peter Zijlstrab68aa232008-02-13 15:45:40 +01009129 return -EINVAL;
9130#else
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009131 /* We don't support RT-tasks being in separate groups */
9132 if (tsk->sched_class != &fair_sched_class)
9133 return -EINVAL;
Peter Zijlstrab68aa232008-02-13 15:45:40 +01009134#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009135
9136 return 0;
9137}
9138
9139static void
Paul Menage2b01dfe2007-10-24 18:23:50 +02009140cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009141 struct cgroup *old_cont, struct task_struct *tsk)
9142{
9143 sched_move_task(tsk);
9144}
9145
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009146#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagef4c753b2008-04-29 00:59:56 -07009147static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
Paul Menage2b01dfe2007-10-24 18:23:50 +02009148 u64 shareval)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009149{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009150 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009151}
9152
Paul Menagef4c753b2008-04-29 00:59:56 -07009153static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009154{
Paul Menage2b01dfe2007-10-24 18:23:50 +02009155 struct task_group *tg = cgroup_tg(cgrp);
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009156
9157 return (u64) tg->shares;
9158}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009159#endif /* CONFIG_FAIR_GROUP_SCHED */
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009160
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009161#ifdef CONFIG_RT_GROUP_SCHED
Mirco Tischler0c708142008-05-14 16:05:46 -07009162static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
Paul Menage06ecb272008-04-29 01:00:06 -07009163 s64 val)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009164{
Paul Menage06ecb272008-04-29 01:00:06 -07009165 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009166}
9167
Paul Menage06ecb272008-04-29 01:00:06 -07009168static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009169{
Paul Menage06ecb272008-04-29 01:00:06 -07009170 return sched_group_rt_runtime(cgroup_tg(cgrp));
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009171}
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009172
9173static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9174 u64 rt_period_us)
9175{
9176 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9177}
9178
9179static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9180{
9181 return sched_group_rt_period(cgroup_tg(cgrp));
9182}
Dhaval Giani6d6bc0a2008-05-30 14:23:45 +02009183#endif /* CONFIG_RT_GROUP_SCHED */
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009184
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009185static struct cftype cpu_files[] = {
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009186#ifdef CONFIG_FAIR_GROUP_SCHED
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009187 {
9188 .name = "shares",
Paul Menagef4c753b2008-04-29 00:59:56 -07009189 .read_u64 = cpu_shares_read_u64,
9190 .write_u64 = cpu_shares_write_u64,
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009191 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009192#endif
9193#ifdef CONFIG_RT_GROUP_SCHED
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009194 {
Peter Zijlstra9f0c1e52008-02-13 15:45:39 +01009195 .name = "rt_runtime_us",
Paul Menage06ecb272008-04-29 01:00:06 -07009196 .read_s64 = cpu_rt_runtime_read,
9197 .write_s64 = cpu_rt_runtime_write,
Peter Zijlstra6f505b12008-01-25 21:08:30 +01009198 },
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009199 {
9200 .name = "rt_period_us",
Paul Menagef4c753b2008-04-29 00:59:56 -07009201 .read_u64 = cpu_rt_period_read_uint,
9202 .write_u64 = cpu_rt_period_write_uint,
Peter Zijlstrad0b27fa2008-04-19 19:44:57 +02009203 },
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009204#endif
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009205};
9206
9207static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9208{
Paul Menagefe5c7cc2007-10-29 21:18:11 +01009209 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009210}
9211
9212struct cgroup_subsys cpu_cgroup_subsys = {
Ingo Molnar38605ca2007-10-29 21:18:11 +01009213 .name = "cpu",
9214 .create = cpu_cgroup_create,
9215 .destroy = cpu_cgroup_destroy,
9216 .can_attach = cpu_cgroup_can_attach,
9217 .attach = cpu_cgroup_attach,
9218 .populate = cpu_cgroup_populate,
9219 .subsys_id = cpu_cgroup_subsys_id,
Srivatsa Vaddagiri68318b82007-10-18 23:41:03 -07009220 .early_init = 1,
9221};
9222
Peter Zijlstra052f1dc2008-02-13 15:45:40 +01009223#endif /* CONFIG_CGROUP_SCHED */
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009224
9225#ifdef CONFIG_CGROUP_CPUACCT
9226
9227/*
9228 * CPU accounting code for task groups.
9229 *
9230 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9231 * (balbir@in.ibm.com).
9232 */
9233
9234/* track cpu usage of a group of tasks */
9235struct cpuacct {
9236 struct cgroup_subsys_state css;
9237 /* cpuusage holds pointer to a u64-type object on every cpu */
9238 u64 *cpuusage;
9239};
9240
9241struct cgroup_subsys cpuacct_subsys;
9242
9243/* return cpu accounting group corresponding to this container */
Dhaval Giani32cd7562008-02-29 10:02:43 +05309244static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009245{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309246 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009247 struct cpuacct, css);
9248}
9249
9250/* return cpu accounting group to which this task belongs */
9251static inline struct cpuacct *task_ca(struct task_struct *tsk)
9252{
9253 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9254 struct cpuacct, css);
9255}
9256
9257/* create a new cpu accounting group */
9258static struct cgroup_subsys_state *cpuacct_create(
Dhaval Giani32cd7562008-02-29 10:02:43 +05309259 struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009260{
9261 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9262
9263 if (!ca)
9264 return ERR_PTR(-ENOMEM);
9265
9266 ca->cpuusage = alloc_percpu(u64);
9267 if (!ca->cpuusage) {
9268 kfree(ca);
9269 return ERR_PTR(-ENOMEM);
9270 }
9271
9272 return &ca->css;
9273}
9274
9275/* destroy an existing cpu accounting group */
Ingo Molnar41a2d6c2007-12-05 15:46:09 +01009276static void
Dhaval Giani32cd7562008-02-29 10:02:43 +05309277cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009278{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309279 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009280
9281 free_percpu(ca->cpuusage);
9282 kfree(ca);
9283}
9284
9285/* return total cpu usage (in nanoseconds) of a group */
Dhaval Giani32cd7562008-02-29 10:02:43 +05309286static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009287{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309288 struct cpuacct *ca = cgroup_ca(cgrp);
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009289 u64 totalcpuusage = 0;
9290 int i;
9291
9292 for_each_possible_cpu(i) {
9293 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9294
9295 /*
9296 * Take rq->lock to make 64-bit addition safe on 32-bit
9297 * platforms.
9298 */
9299 spin_lock_irq(&cpu_rq(i)->lock);
9300 totalcpuusage += *cpuusage;
9301 spin_unlock_irq(&cpu_rq(i)->lock);
9302 }
9303
9304 return totalcpuusage;
9305}
9306
Dhaval Giani0297b802008-02-29 10:02:44 +05309307static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9308 u64 reset)
9309{
9310 struct cpuacct *ca = cgroup_ca(cgrp);
9311 int err = 0;
9312 int i;
9313
9314 if (reset) {
9315 err = -EINVAL;
9316 goto out;
9317 }
9318
9319 for_each_possible_cpu(i) {
9320 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9321
9322 spin_lock_irq(&cpu_rq(i)->lock);
9323 *cpuusage = 0;
9324 spin_unlock_irq(&cpu_rq(i)->lock);
9325 }
9326out:
9327 return err;
9328}
9329
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009330static struct cftype files[] = {
9331 {
9332 .name = "usage",
Paul Menagef4c753b2008-04-29 00:59:56 -07009333 .read_u64 = cpuusage_read,
9334 .write_u64 = cpuusage_write,
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009335 },
9336};
9337
Dhaval Giani32cd7562008-02-29 10:02:43 +05309338static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009339{
Dhaval Giani32cd7562008-02-29 10:02:43 +05309340 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
Srivatsa Vaddagirid842de82007-12-02 20:04:49 +01009341}
9342
9343/*
9344 * charge this task's execution time to its accounting group.
9345 *
9346 * called with rq->lock held.
9347 */
9348static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9349{
9350 struct cpuacct *ca;
9351
9352 if (!cpuacct_subsys.active)
9353 return;
9354
9355 ca = task_ca(tsk);
9356 if (ca) {
9357 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9358
9359 *cpuusage += cputime;
9360 }
9361}
9362
9363struct cgroup_subsys cpuacct_subsys = {
9364 .name = "cpuacct",
9365 .create = cpuacct_create,
9366 .destroy = cpuacct_destroy,
9367 .populate = cpuacct_populate,
9368 .subsys_id = cpuacct_subsys_id,
9369};
9370#endif /* CONFIG_CGROUP_CPUACCT */