4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
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
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
19 * 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
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #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>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
78 * Scheduler clock - returns current time in nanosec units.
79 * This is default implementation.
80 * Architectures and sub-architectures can override this.
82 unsigned long long __attribute__((weak)) sched_clock(void)
84 return (unsigned long long)jiffies * (NSEC_PER_SEC / HZ);
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
133 return reciprocal_divide(load, sg->reciprocal_cpu_power);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
142 sg->__cpu_power += val;
143 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
147 static inline int rt_policy(int policy)
149 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
154 static inline int task_has_rt_policy(struct task_struct *p)
156 return rt_policy(p->policy);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array {
163 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
164 struct list_head queue[MAX_RT_PRIO];
167 struct rt_bandwidth {
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock;
172 struct hrtimer rt_period_timer;
175 static struct rt_bandwidth def_rt_bandwidth;
177 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
179 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
181 struct rt_bandwidth *rt_b =
182 container_of(timer, struct rt_bandwidth, rt_period_timer);
188 now = hrtimer_cb_get_time(timer);
189 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
194 idle = do_sched_rt_period_timer(rt_b, overrun);
197 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
201 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
203 rt_b->rt_period = ns_to_ktime(period);
204 rt_b->rt_runtime = runtime;
206 spin_lock_init(&rt_b->rt_runtime_lock);
208 hrtimer_init(&rt_b->rt_period_timer,
209 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
210 rt_b->rt_period_timer.function = sched_rt_period_timer;
211 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
214 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
218 if (rt_b->rt_runtime == RUNTIME_INF)
221 if (hrtimer_active(&rt_b->rt_period_timer))
224 spin_lock(&rt_b->rt_runtime_lock);
226 if (hrtimer_active(&rt_b->rt_period_timer))
229 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
230 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
231 hrtimer_start(&rt_b->rt_period_timer,
232 rt_b->rt_period_timer.expires,
235 spin_unlock(&rt_b->rt_runtime_lock);
238 #ifdef CONFIG_RT_GROUP_SCHED
239 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
241 hrtimer_cancel(&rt_b->rt_period_timer);
245 #ifdef CONFIG_GROUP_SCHED
247 #include <linux/cgroup.h>
251 static LIST_HEAD(task_groups);
253 /* task group related information */
255 #ifdef CONFIG_CGROUP_SCHED
256 struct cgroup_subsys_state css;
259 #ifdef CONFIG_FAIR_GROUP_SCHED
260 /* schedulable entities of this group on each cpu */
261 struct sched_entity **se;
262 /* runqueue "owned" by this group on each cpu */
263 struct cfs_rq **cfs_rq;
264 unsigned long shares;
267 #ifdef CONFIG_RT_GROUP_SCHED
268 struct sched_rt_entity **rt_se;
269 struct rt_rq **rt_rq;
271 struct rt_bandwidth rt_bandwidth;
275 struct list_head list;
277 struct task_group *parent;
278 struct list_head siblings;
279 struct list_head children;
282 #ifdef CONFIG_USER_SCHED
286 * Every UID task group (including init_task_group aka UID-0) will
287 * be a child to this group.
289 struct task_group root_task_group;
291 #ifdef CONFIG_FAIR_GROUP_SCHED
292 /* Default task group's sched entity on each cpu */
293 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
294 /* Default task group's cfs_rq on each cpu */
295 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
298 #ifdef CONFIG_RT_GROUP_SCHED
299 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
300 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
303 #define root_task_group init_task_group
306 /* task_group_lock serializes add/remove of task groups and also changes to
307 * a task group's cpu shares.
309 static DEFINE_SPINLOCK(task_group_lock);
311 /* doms_cur_mutex serializes access to doms_cur[] array */
312 static DEFINE_MUTEX(doms_cur_mutex);
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 #ifdef CONFIG_USER_SCHED
316 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
318 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
323 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
326 /* Default task group.
327 * Every task in system belong to this group at bootup.
329 struct task_group init_task_group;
331 /* return group to which a task belongs */
332 static inline struct task_group *task_group(struct task_struct *p)
334 struct task_group *tg;
336 #ifdef CONFIG_USER_SCHED
338 #elif defined(CONFIG_CGROUP_SCHED)
339 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
340 struct task_group, css);
342 tg = &init_task_group;
347 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
348 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
350 #ifdef CONFIG_FAIR_GROUP_SCHED
351 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
352 p->se.parent = task_group(p)->se[cpu];
355 #ifdef CONFIG_RT_GROUP_SCHED
356 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
357 p->rt.parent = task_group(p)->rt_se[cpu];
361 static inline void lock_doms_cur(void)
363 mutex_lock(&doms_cur_mutex);
366 static inline void unlock_doms_cur(void)
368 mutex_unlock(&doms_cur_mutex);
373 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
374 static inline void lock_doms_cur(void) { }
375 static inline void unlock_doms_cur(void) { }
377 #endif /* CONFIG_GROUP_SCHED */
379 /* CFS-related fields in a runqueue */
381 struct load_weight load;
382 unsigned long nr_running;
387 struct rb_root tasks_timeline;
388 struct rb_node *rb_leftmost;
390 struct list_head tasks;
391 struct list_head *balance_iterator;
394 * 'curr' points to currently running entity on this cfs_rq.
395 * It is set to NULL otherwise (i.e when none are currently running).
397 struct sched_entity *curr, *next;
399 unsigned long nr_spread_over;
401 #ifdef CONFIG_FAIR_GROUP_SCHED
402 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
405 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
406 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
407 * (like users, containers etc.)
409 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
410 * list is used during load balance.
412 struct list_head leaf_cfs_rq_list;
413 struct task_group *tg; /* group that "owns" this runqueue */
416 unsigned long task_weight;
417 unsigned long shares;
419 * We need space to build a sched_domain wide view of the full task
420 * group tree, in order to avoid depending on dynamic memory allocation
421 * during the load balancing we place this in the per cpu task group
422 * hierarchy. This limits the load balancing to one instance per cpu,
423 * but more should not be needed anyway.
425 struct aggregate_struct {
427 * load = weight(cpus) * f(tg)
429 * Where f(tg) is the recursive weight fraction assigned to
435 * part of the group weight distributed to this span.
437 unsigned long shares;
440 * The sum of all runqueue weights within this span.
442 unsigned long rq_weight;
445 * Weight contributed by tasks; this is the part we can
446 * influence by moving tasks around.
448 unsigned long task_weight;
454 /* Real-Time classes' related field in a runqueue: */
456 struct rt_prio_array active;
457 unsigned long rt_nr_running;
458 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
459 int highest_prio; /* highest queued rt task prio */
462 unsigned long rt_nr_migratory;
468 /* Nests inside the rq lock: */
469 spinlock_t rt_runtime_lock;
471 #ifdef CONFIG_RT_GROUP_SCHED
472 unsigned long rt_nr_boosted;
475 struct list_head leaf_rt_rq_list;
476 struct task_group *tg;
477 struct sched_rt_entity *rt_se;
484 * We add the notion of a root-domain which will be used to define per-domain
485 * variables. Each exclusive cpuset essentially defines an island domain by
486 * fully partitioning the member cpus from any other cpuset. Whenever a new
487 * exclusive cpuset is created, we also create and attach a new root-domain
497 * The "RT overload" flag: it gets set if a CPU has more than
498 * one runnable RT task.
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
508 static struct root_domain def_root_domain;
513 * This is the main, per-CPU runqueue data structure.
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
527 unsigned long nr_running;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
530 unsigned char idle_at_tick;
532 unsigned long last_tick_seen;
533 unsigned char in_nohz_recently;
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load;
537 unsigned long nr_load_updates;
543 #ifdef CONFIG_FAIR_GROUP_SCHED
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list;
547 #ifdef CONFIG_RT_GROUP_SCHED
548 struct list_head leaf_rt_rq_list;
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
557 unsigned long nr_uninterruptible;
559 struct task_struct *curr, *idle;
560 unsigned long next_balance;
561 struct mm_struct *prev_mm;
563 u64 clock, prev_clock_raw;
566 unsigned int clock_warps, clock_overflows, clock_underflows;
568 unsigned int clock_deep_idle_events;
574 struct root_domain *rd;
575 struct sched_domain *sd;
577 /* For active balancing */
580 /* cpu of this runqueue: */
583 struct task_struct *migration_thread;
584 struct list_head migration_queue;
587 #ifdef CONFIG_SCHED_HRTICK
588 unsigned long hrtick_flags;
589 ktime_t hrtick_expire;
590 struct hrtimer hrtick_timer;
593 #ifdef CONFIG_SCHEDSTATS
595 struct sched_info rq_sched_info;
597 /* sys_sched_yield() stats */
598 unsigned int yld_exp_empty;
599 unsigned int yld_act_empty;
600 unsigned int yld_both_empty;
601 unsigned int yld_count;
603 /* schedule() stats */
604 unsigned int sched_switch;
605 unsigned int sched_count;
606 unsigned int sched_goidle;
608 /* try_to_wake_up() stats */
609 unsigned int ttwu_count;
610 unsigned int ttwu_local;
613 unsigned int bkl_count;
615 struct lock_class_key rq_lock_key;
618 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
620 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
622 rq->curr->sched_class->check_preempt_curr(rq, p);
625 static inline int cpu_of(struct rq *rq)
635 static inline bool nohz_on(int cpu)
637 return tick_get_tick_sched(cpu)->nohz_mode != NOHZ_MODE_INACTIVE;
640 static inline u64 max_skipped_ticks(struct rq *rq)
642 return nohz_on(cpu_of(rq)) ? jiffies - rq->last_tick_seen + 2 : 1;
645 static inline void update_last_tick_seen(struct rq *rq)
647 rq->last_tick_seen = jiffies;
650 static inline u64 max_skipped_ticks(struct rq *rq)
655 static inline void update_last_tick_seen(struct rq *rq)
661 * Update the per-runqueue clock, as finegrained as the platform can give
662 * us, but without assuming monotonicity, etc.:
664 static void __update_rq_clock(struct rq *rq)
666 u64 prev_raw = rq->prev_clock_raw;
667 u64 now = sched_clock();
668 s64 delta = now - prev_raw;
669 u64 clock = rq->clock;
671 #ifdef CONFIG_SCHED_DEBUG
672 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
675 * Protect against sched_clock() occasionally going backwards:
677 if (unlikely(delta < 0)) {
682 * Catch too large forward jumps too:
684 u64 max_jump = max_skipped_ticks(rq) * TICK_NSEC;
685 u64 max_time = rq->tick_timestamp + max_jump;
687 if (unlikely(clock + delta > max_time)) {
688 if (clock < max_time)
692 rq->clock_overflows++;
694 if (unlikely(delta > rq->clock_max_delta))
695 rq->clock_max_delta = delta;
700 rq->prev_clock_raw = now;
704 static void update_rq_clock(struct rq *rq)
706 if (likely(smp_processor_id() == cpu_of(rq)))
707 __update_rq_clock(rq);
711 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
712 * See detach_destroy_domains: synchronize_sched for details.
714 * The domain tree of any CPU may only be accessed from within
715 * preempt-disabled sections.
717 #define for_each_domain(cpu, __sd) \
718 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
720 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
721 #define this_rq() (&__get_cpu_var(runqueues))
722 #define task_rq(p) cpu_rq(task_cpu(p))
723 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
726 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
728 #ifdef CONFIG_SCHED_DEBUG
729 # define const_debug __read_mostly
731 # define const_debug static const
735 * Debugging: various feature bits
738 #define SCHED_FEAT(name, enabled) \
739 __SCHED_FEAT_##name ,
742 #include "sched_features.h"
747 #define SCHED_FEAT(name, enabled) \
748 (1UL << __SCHED_FEAT_##name) * enabled |
750 const_debug unsigned int sysctl_sched_features =
751 #include "sched_features.h"
756 #ifdef CONFIG_SCHED_DEBUG
757 #define SCHED_FEAT(name, enabled) \
760 __read_mostly char *sched_feat_names[] = {
761 #include "sched_features.h"
767 int sched_feat_open(struct inode *inode, struct file *filp)
769 filp->private_data = inode->i_private;
774 sched_feat_read(struct file *filp, char __user *ubuf,
775 size_t cnt, loff_t *ppos)
782 for (i = 0; sched_feat_names[i]; i++) {
783 len += strlen(sched_feat_names[i]);
787 buf = kmalloc(len + 2, GFP_KERNEL);
791 for (i = 0; sched_feat_names[i]; i++) {
792 if (sysctl_sched_features & (1UL << i))
793 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
795 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
798 r += sprintf(buf + r, "\n");
799 WARN_ON(r >= len + 2);
801 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
809 sched_feat_write(struct file *filp, const char __user *ubuf,
810 size_t cnt, loff_t *ppos)
820 if (copy_from_user(&buf, ubuf, cnt))
825 if (strncmp(buf, "NO_", 3) == 0) {
830 for (i = 0; sched_feat_names[i]; i++) {
831 int len = strlen(sched_feat_names[i]);
833 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
835 sysctl_sched_features &= ~(1UL << i);
837 sysctl_sched_features |= (1UL << i);
842 if (!sched_feat_names[i])
850 static struct file_operations sched_feat_fops = {
851 .open = sched_feat_open,
852 .read = sched_feat_read,
853 .write = sched_feat_write,
856 static __init int sched_init_debug(void)
858 debugfs_create_file("sched_features", 0644, NULL, NULL,
863 late_initcall(sched_init_debug);
867 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
870 * Number of tasks to iterate in a single balance run.
871 * Limited because this is done with IRQs disabled.
873 const_debug unsigned int sysctl_sched_nr_migrate = 32;
876 * period over which we measure -rt task cpu usage in us.
879 unsigned int sysctl_sched_rt_period = 1000000;
881 static __read_mostly int scheduler_running;
884 * part of the period that we allow rt tasks to run in us.
887 int sysctl_sched_rt_runtime = 950000;
889 static inline u64 global_rt_period(void)
891 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
894 static inline u64 global_rt_runtime(void)
896 if (sysctl_sched_rt_period < 0)
899 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
902 static const unsigned long long time_sync_thresh = 100000;
904 static DEFINE_PER_CPU(unsigned long long, time_offset);
905 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
908 * Global lock which we take every now and then to synchronize
909 * the CPUs time. This method is not warp-safe, but it's good
910 * enough to synchronize slowly diverging time sources and thus
911 * it's good enough for tracing:
913 static DEFINE_SPINLOCK(time_sync_lock);
914 static unsigned long long prev_global_time;
916 static unsigned long long __sync_cpu_clock(cycles_t time, int cpu)
920 spin_lock_irqsave(&time_sync_lock, flags);
922 if (time < prev_global_time) {
923 per_cpu(time_offset, cpu) += prev_global_time - time;
924 time = prev_global_time;
926 prev_global_time = time;
929 spin_unlock_irqrestore(&time_sync_lock, flags);
934 static unsigned long long __cpu_clock(int cpu)
936 unsigned long long now;
941 * Only call sched_clock() if the scheduler has already been
942 * initialized (some code might call cpu_clock() very early):
944 if (unlikely(!scheduler_running))
947 local_irq_save(flags);
951 local_irq_restore(flags);
957 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
958 * clock constructed from sched_clock():
960 unsigned long long cpu_clock(int cpu)
962 unsigned long long prev_cpu_time, time, delta_time;
964 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
965 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
966 delta_time = time-prev_cpu_time;
968 if (unlikely(delta_time > time_sync_thresh))
969 time = __sync_cpu_clock(time, cpu);
973 EXPORT_SYMBOL_GPL(cpu_clock);
975 #ifndef prepare_arch_switch
976 # define prepare_arch_switch(next) do { } while (0)
978 #ifndef finish_arch_switch
979 # define finish_arch_switch(prev) do { } while (0)
982 static inline int task_current(struct rq *rq, struct task_struct *p)
984 return rq->curr == p;
987 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
988 static inline int task_running(struct rq *rq, struct task_struct *p)
990 return task_current(rq, p);
993 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
997 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
999 #ifdef CONFIG_DEBUG_SPINLOCK
1000 /* this is a valid case when another task releases the spinlock */
1001 rq->lock.owner = current;
1004 * If we are tracking spinlock dependencies then we have to
1005 * fix up the runqueue lock - which gets 'carried over' from
1006 * prev into current:
1008 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
1010 spin_unlock_irq(&rq->lock);
1013 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1014 static inline int task_running(struct rq *rq, struct task_struct *p)
1019 return task_current(rq, p);
1023 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
1027 * We can optimise this out completely for !SMP, because the
1028 * SMP rebalancing from interrupt is the only thing that cares
1033 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1034 spin_unlock_irq(&rq->lock);
1036 spin_unlock(&rq->lock);
1040 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
1044 * After ->oncpu is cleared, the task can be moved to a different CPU.
1045 * We must ensure this doesn't happen until the switch is completely
1051 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1055 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1058 * __task_rq_lock - lock the runqueue a given task resides on.
1059 * Must be called interrupts disabled.
1061 static inline struct rq *__task_rq_lock(struct task_struct *p)
1062 __acquires(rq->lock)
1065 struct rq *rq = task_rq(p);
1066 spin_lock(&rq->lock);
1067 if (likely(rq == task_rq(p)))
1069 spin_unlock(&rq->lock);
1074 * task_rq_lock - lock the runqueue a given task resides on and disable
1075 * interrupts. Note the ordering: we can safely lookup the task_rq without
1076 * explicitly disabling preemption.
1078 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1079 __acquires(rq->lock)
1084 local_irq_save(*flags);
1086 spin_lock(&rq->lock);
1087 if (likely(rq == task_rq(p)))
1089 spin_unlock_irqrestore(&rq->lock, *flags);
1093 static void __task_rq_unlock(struct rq *rq)
1094 __releases(rq->lock)
1096 spin_unlock(&rq->lock);
1099 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1100 __releases(rq->lock)
1102 spin_unlock_irqrestore(&rq->lock, *flags);
1106 * this_rq_lock - lock this runqueue and disable interrupts.
1108 static struct rq *this_rq_lock(void)
1109 __acquires(rq->lock)
1113 local_irq_disable();
1115 spin_lock(&rq->lock);
1121 * We are going deep-idle (irqs are disabled):
1123 void sched_clock_idle_sleep_event(void)
1125 struct rq *rq = cpu_rq(smp_processor_id());
1127 WARN_ON(!irqs_disabled());
1128 spin_lock(&rq->lock);
1129 __update_rq_clock(rq);
1130 spin_unlock(&rq->lock);
1131 rq->clock_deep_idle_events++;
1133 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event);
1136 * We just idled delta nanoseconds (called with irqs disabled):
1138 void sched_clock_idle_wakeup_event(u64 delta_ns)
1140 struct rq *rq = cpu_rq(smp_processor_id());
1141 u64 now = sched_clock();
1143 WARN_ON(!irqs_disabled());
1144 rq->idle_clock += delta_ns;
1146 * Override the previous timestamp and ignore all
1147 * sched_clock() deltas that occured while we idled,
1148 * and use the PM-provided delta_ns to advance the
1151 spin_lock(&rq->lock);
1152 rq->prev_clock_raw = now;
1153 rq->clock += delta_ns;
1154 spin_unlock(&rq->lock);
1155 touch_softlockup_watchdog();
1157 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event);
1159 static void __resched_task(struct task_struct *p, int tif_bit);
1161 static inline void resched_task(struct task_struct *p)
1163 __resched_task(p, TIF_NEED_RESCHED);
1166 #ifdef CONFIG_SCHED_HRTICK
1168 * Use HR-timers to deliver accurate preemption points.
1170 * Its all a bit involved since we cannot program an hrt while holding the
1171 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1174 * When we get rescheduled we reprogram the hrtick_timer outside of the
1177 static inline void resched_hrt(struct task_struct *p)
1179 __resched_task(p, TIF_HRTICK_RESCHED);
1182 static inline void resched_rq(struct rq *rq)
1184 unsigned long flags;
1186 spin_lock_irqsave(&rq->lock, flags);
1187 resched_task(rq->curr);
1188 spin_unlock_irqrestore(&rq->lock, flags);
1192 HRTICK_SET, /* re-programm hrtick_timer */
1193 HRTICK_RESET, /* not a new slice */
1198 * - enabled by features
1199 * - hrtimer is actually high res
1201 static inline int hrtick_enabled(struct rq *rq)
1203 if (!sched_feat(HRTICK))
1205 return hrtimer_is_hres_active(&rq->hrtick_timer);
1209 * Called to set the hrtick timer state.
1211 * called with rq->lock held and irqs disabled
1213 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1215 assert_spin_locked(&rq->lock);
1218 * preempt at: now + delay
1221 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1223 * indicate we need to program the timer
1225 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1227 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1230 * New slices are called from the schedule path and don't need a
1231 * forced reschedule.
1234 resched_hrt(rq->curr);
1237 static void hrtick_clear(struct rq *rq)
1239 if (hrtimer_active(&rq->hrtick_timer))
1240 hrtimer_cancel(&rq->hrtick_timer);
1244 * Update the timer from the possible pending state.
1246 static void hrtick_set(struct rq *rq)
1250 unsigned long flags;
1252 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1254 spin_lock_irqsave(&rq->lock, flags);
1255 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1256 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1257 time = rq->hrtick_expire;
1258 clear_thread_flag(TIF_HRTICK_RESCHED);
1259 spin_unlock_irqrestore(&rq->lock, flags);
1262 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1263 if (reset && !hrtimer_active(&rq->hrtick_timer))
1270 * High-resolution timer tick.
1271 * Runs from hardirq context with interrupts disabled.
1273 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1275 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1277 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1279 spin_lock(&rq->lock);
1280 __update_rq_clock(rq);
1281 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1282 spin_unlock(&rq->lock);
1284 return HRTIMER_NORESTART;
1287 static inline void init_rq_hrtick(struct rq *rq)
1289 rq->hrtick_flags = 0;
1290 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1291 rq->hrtick_timer.function = hrtick;
1292 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1295 void hrtick_resched(void)
1298 unsigned long flags;
1300 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1303 local_irq_save(flags);
1304 rq = cpu_rq(smp_processor_id());
1306 local_irq_restore(flags);
1309 static inline void hrtick_clear(struct rq *rq)
1313 static inline void hrtick_set(struct rq *rq)
1317 static inline void init_rq_hrtick(struct rq *rq)
1321 void hrtick_resched(void)
1327 * resched_task - mark a task 'to be rescheduled now'.
1329 * On UP this means the setting of the need_resched flag, on SMP it
1330 * might also involve a cross-CPU call to trigger the scheduler on
1335 #ifndef tsk_is_polling
1336 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1339 static void __resched_task(struct task_struct *p, int tif_bit)
1343 assert_spin_locked(&task_rq(p)->lock);
1345 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1348 set_tsk_thread_flag(p, tif_bit);
1351 if (cpu == smp_processor_id())
1354 /* NEED_RESCHED must be visible before we test polling */
1356 if (!tsk_is_polling(p))
1357 smp_send_reschedule(cpu);
1360 static void resched_cpu(int cpu)
1362 struct rq *rq = cpu_rq(cpu);
1363 unsigned long flags;
1365 if (!spin_trylock_irqsave(&rq->lock, flags))
1367 resched_task(cpu_curr(cpu));
1368 spin_unlock_irqrestore(&rq->lock, flags);
1373 * When add_timer_on() enqueues a timer into the timer wheel of an
1374 * idle CPU then this timer might expire before the next timer event
1375 * which is scheduled to wake up that CPU. In case of a completely
1376 * idle system the next event might even be infinite time into the
1377 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1378 * leaves the inner idle loop so the newly added timer is taken into
1379 * account when the CPU goes back to idle and evaluates the timer
1380 * wheel for the next timer event.
1382 void wake_up_idle_cpu(int cpu)
1384 struct rq *rq = cpu_rq(cpu);
1386 if (cpu == smp_processor_id())
1390 * This is safe, as this function is called with the timer
1391 * wheel base lock of (cpu) held. When the CPU is on the way
1392 * to idle and has not yet set rq->curr to idle then it will
1393 * be serialized on the timer wheel base lock and take the new
1394 * timer into account automatically.
1396 if (rq->curr != rq->idle)
1400 * We can set TIF_RESCHED on the idle task of the other CPU
1401 * lockless. The worst case is that the other CPU runs the
1402 * idle task through an additional NOOP schedule()
1404 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1406 /* NEED_RESCHED must be visible before we test polling */
1408 if (!tsk_is_polling(rq->idle))
1409 smp_send_reschedule(cpu);
1414 static void __resched_task(struct task_struct *p, int tif_bit)
1416 assert_spin_locked(&task_rq(p)->lock);
1417 set_tsk_thread_flag(p, tif_bit);
1421 #if BITS_PER_LONG == 32
1422 # define WMULT_CONST (~0UL)
1424 # define WMULT_CONST (1UL << 32)
1427 #define WMULT_SHIFT 32
1430 * Shift right and round:
1432 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1435 * delta *= weight / lw
1437 static unsigned long
1438 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1439 struct load_weight *lw)
1443 if (!lw->inv_weight)
1444 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1446 tmp = (u64)delta_exec * weight;
1448 * Check whether we'd overflow the 64-bit multiplication:
1450 if (unlikely(tmp > WMULT_CONST))
1451 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1454 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1456 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1459 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1465 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1472 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1473 * of tasks with abnormal "nice" values across CPUs the contribution that
1474 * each task makes to its run queue's load is weighted according to its
1475 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1476 * scaled version of the new time slice allocation that they receive on time
1480 #define WEIGHT_IDLEPRIO 2
1481 #define WMULT_IDLEPRIO (1 << 31)
1484 * Nice levels are multiplicative, with a gentle 10% change for every
1485 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1486 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1487 * that remained on nice 0.
1489 * The "10% effect" is relative and cumulative: from _any_ nice level,
1490 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1491 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1492 * If a task goes up by ~10% and another task goes down by ~10% then
1493 * the relative distance between them is ~25%.)
1495 static const int prio_to_weight[40] = {
1496 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1497 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1498 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1499 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1500 /* 0 */ 1024, 820, 655, 526, 423,
1501 /* 5 */ 335, 272, 215, 172, 137,
1502 /* 10 */ 110, 87, 70, 56, 45,
1503 /* 15 */ 36, 29, 23, 18, 15,
1507 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1509 * In cases where the weight does not change often, we can use the
1510 * precalculated inverse to speed up arithmetics by turning divisions
1511 * into multiplications:
1513 static const u32 prio_to_wmult[40] = {
1514 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1515 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1516 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1517 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1518 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1519 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1520 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1521 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1524 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1527 * runqueue iterator, to support SMP load-balancing between different
1528 * scheduling classes, without having to expose their internal data
1529 * structures to the load-balancing proper:
1531 struct rq_iterator {
1533 struct task_struct *(*start)(void *);
1534 struct task_struct *(*next)(void *);
1538 static unsigned long
1539 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1540 unsigned long max_load_move, struct sched_domain *sd,
1541 enum cpu_idle_type idle, int *all_pinned,
1542 int *this_best_prio, struct rq_iterator *iterator);
1545 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1546 struct sched_domain *sd, enum cpu_idle_type idle,
1547 struct rq_iterator *iterator);
1550 #ifdef CONFIG_CGROUP_CPUACCT
1551 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1553 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1556 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1558 update_load_add(&rq->load, load);
1561 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1563 update_load_sub(&rq->load, load);
1567 static unsigned long source_load(int cpu, int type);
1568 static unsigned long target_load(int cpu, int type);
1569 static unsigned long cpu_avg_load_per_task(int cpu);
1570 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1572 #ifdef CONFIG_FAIR_GROUP_SCHED
1575 * Group load balancing.
1577 * We calculate a few balance domain wide aggregate numbers; load and weight.
1578 * Given the pictures below, and assuming each item has equal weight:
1589 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1590 * which equals 1/9-th of the total load.
1593 * The weight of this group on the selected cpus.
1596 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1600 * Part of the rq_weight contributed by tasks; all groups except B would
1604 static inline struct aggregate_struct *
1605 aggregate(struct task_group *tg, struct sched_domain *sd)
1607 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1610 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1613 * Iterate the full tree, calling @down when first entering a node and @up when
1614 * leaving it for the final time.
1617 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1618 struct sched_domain *sd)
1620 struct task_group *parent, *child;
1623 parent = &root_task_group;
1625 (*down)(parent, sd);
1626 list_for_each_entry_rcu(child, &parent->children, siblings) {
1636 parent = parent->parent;
1643 * Calculate the aggregate runqueue weight.
1646 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1648 unsigned long rq_weight = 0;
1649 unsigned long task_weight = 0;
1652 for_each_cpu_mask(i, sd->span) {
1653 rq_weight += tg->cfs_rq[i]->load.weight;
1654 task_weight += tg->cfs_rq[i]->task_weight;
1657 aggregate(tg, sd)->rq_weight = rq_weight;
1658 aggregate(tg, sd)->task_weight = task_weight;
1662 * Compute the weight of this group on the given cpus.
1665 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1667 unsigned long shares = 0;
1670 for_each_cpu_mask(i, sd->span)
1671 shares += tg->cfs_rq[i]->shares;
1673 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1674 shares = tg->shares;
1676 aggregate(tg, sd)->shares = shares;
1680 * Compute the load fraction assigned to this group, relies on the aggregate
1681 * weight and this group's parent's load, i.e. top-down.
1684 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1692 for_each_cpu_mask(i, sd->span)
1693 load += cpu_rq(i)->load.weight;
1696 load = aggregate(tg->parent, sd)->load;
1699 * shares is our weight in the parent's rq so
1700 * shares/parent->rq_weight gives our fraction of the load
1702 load *= aggregate(tg, sd)->shares;
1703 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1706 aggregate(tg, sd)->load = load;
1709 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1712 * Calculate and set the cpu's group shares.
1715 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1719 unsigned long shares;
1720 unsigned long rq_weight;
1725 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1728 * If there are currently no tasks on the cpu pretend there is one of
1729 * average load so that when a new task gets to run here it will not
1730 * get delayed by group starvation.
1734 rq_weight = NICE_0_LOAD;
1738 * \Sum shares * rq_weight
1739 * shares = -----------------------
1743 shares = aggregate(tg, sd)->shares * rq_weight;
1744 shares /= aggregate(tg, sd)->rq_weight + 1;
1747 * record the actual number of shares, not the boosted amount.
1749 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1751 if (shares < MIN_SHARES)
1752 shares = MIN_SHARES;
1754 __set_se_shares(tg->se[tcpu], shares);
1758 * Re-adjust the weights on the cpu the task came from and on the cpu the
1762 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1765 unsigned long shares;
1767 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1769 __update_group_shares_cpu(tg, sd, scpu);
1770 __update_group_shares_cpu(tg, sd, dcpu);
1773 * ensure we never loose shares due to rounding errors in the
1774 * above redistribution.
1776 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1778 tg->cfs_rq[dcpu]->shares += shares;
1782 * Because changing a group's shares changes the weight of the super-group
1783 * we need to walk up the tree and change all shares until we hit the root.
1786 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1790 __move_group_shares(tg, sd, scpu, dcpu);
1796 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1798 unsigned long shares = aggregate(tg, sd)->shares;
1801 for_each_cpu_mask(i, sd->span) {
1802 struct rq *rq = cpu_rq(i);
1803 unsigned long flags;
1805 spin_lock_irqsave(&rq->lock, flags);
1806 __update_group_shares_cpu(tg, sd, i);
1807 spin_unlock_irqrestore(&rq->lock, flags);
1810 aggregate_group_shares(tg, sd);
1813 * ensure we never loose shares due to rounding errors in the
1814 * above redistribution.
1816 shares -= aggregate(tg, sd)->shares;
1818 tg->cfs_rq[sd->first_cpu]->shares += shares;
1819 aggregate(tg, sd)->shares += shares;
1824 * Calculate the accumulative weight and recursive load of each task group
1825 * while walking down the tree.
1828 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1830 aggregate_group_weight(tg, sd);
1831 aggregate_group_shares(tg, sd);
1832 aggregate_group_load(tg, sd);
1836 * Rebalance the cpu shares while walking back up the tree.
1839 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1841 aggregate_group_set_shares(tg, sd);
1844 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1846 static void __init init_aggregate(void)
1850 for_each_possible_cpu(i)
1851 spin_lock_init(&per_cpu(aggregate_lock, i));
1854 static int get_aggregate(struct sched_domain *sd)
1856 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1859 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1863 static void put_aggregate(struct sched_domain *sd)
1865 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1868 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1870 cfs_rq->shares = shares;
1875 static inline void init_aggregate(void)
1879 static inline int get_aggregate(struct sched_domain *sd)
1884 static inline void put_aggregate(struct sched_domain *sd)
1889 #else /* CONFIG_SMP */
1891 #ifdef CONFIG_FAIR_GROUP_SCHED
1892 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1897 #endif /* CONFIG_SMP */
1899 #include "sched_stats.h"
1900 #include "sched_idletask.c"
1901 #include "sched_fair.c"
1902 #include "sched_rt.c"
1903 #ifdef CONFIG_SCHED_DEBUG
1904 # include "sched_debug.c"
1907 #define sched_class_highest (&rt_sched_class)
1909 static void inc_nr_running(struct rq *rq)
1914 static void dec_nr_running(struct rq *rq)
1919 static void set_load_weight(struct task_struct *p)
1921 if (task_has_rt_policy(p)) {
1922 p->se.load.weight = prio_to_weight[0] * 2;
1923 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1928 * SCHED_IDLE tasks get minimal weight:
1930 if (p->policy == SCHED_IDLE) {
1931 p->se.load.weight = WEIGHT_IDLEPRIO;
1932 p->se.load.inv_weight = WMULT_IDLEPRIO;
1936 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1937 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1940 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1942 sched_info_queued(p);
1943 p->sched_class->enqueue_task(rq, p, wakeup);
1947 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1949 p->sched_class->dequeue_task(rq, p, sleep);
1954 * __normal_prio - return the priority that is based on the static prio
1956 static inline int __normal_prio(struct task_struct *p)
1958 return p->static_prio;
1962 * Calculate the expected normal priority: i.e. priority
1963 * without taking RT-inheritance into account. Might be
1964 * boosted by interactivity modifiers. Changes upon fork,
1965 * setprio syscalls, and whenever the interactivity
1966 * estimator recalculates.
1968 static inline int normal_prio(struct task_struct *p)
1972 if (task_has_rt_policy(p))
1973 prio = MAX_RT_PRIO-1 - p->rt_priority;
1975 prio = __normal_prio(p);
1980 * Calculate the current priority, i.e. the priority
1981 * taken into account by the scheduler. This value might
1982 * be boosted by RT tasks, or might be boosted by
1983 * interactivity modifiers. Will be RT if the task got
1984 * RT-boosted. If not then it returns p->normal_prio.
1986 static int effective_prio(struct task_struct *p)
1988 p->normal_prio = normal_prio(p);
1990 * If we are RT tasks or we were boosted to RT priority,
1991 * keep the priority unchanged. Otherwise, update priority
1992 * to the normal priority:
1994 if (!rt_prio(p->prio))
1995 return p->normal_prio;
2000 * activate_task - move a task to the runqueue.
2002 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
2004 if (task_contributes_to_load(p))
2005 rq->nr_uninterruptible--;
2007 enqueue_task(rq, p, wakeup);
2012 * deactivate_task - remove a task from the runqueue.
2014 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
2016 if (task_contributes_to_load(p))
2017 rq->nr_uninterruptible++;
2019 dequeue_task(rq, p, sleep);
2024 * task_curr - is this task currently executing on a CPU?
2025 * @p: the task in question.
2027 inline int task_curr(const struct task_struct *p)
2029 return cpu_curr(task_cpu(p)) == p;
2032 /* Used instead of source_load when we know the type == 0 */
2033 unsigned long weighted_cpuload(const int cpu)
2035 return cpu_rq(cpu)->load.weight;
2038 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
2040 set_task_rq(p, cpu);
2043 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2044 * successfuly executed on another CPU. We must ensure that updates of
2045 * per-task data have been completed by this moment.
2048 task_thread_info(p)->cpu = cpu;
2052 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2053 const struct sched_class *prev_class,
2054 int oldprio, int running)
2056 if (prev_class != p->sched_class) {
2057 if (prev_class->switched_from)
2058 prev_class->switched_from(rq, p, running);
2059 p->sched_class->switched_to(rq, p, running);
2061 p->sched_class->prio_changed(rq, p, oldprio, running);
2067 * Is this task likely cache-hot:
2070 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2075 * Buddy candidates are cache hot:
2077 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2080 if (p->sched_class != &fair_sched_class)
2083 if (sysctl_sched_migration_cost == -1)
2085 if (sysctl_sched_migration_cost == 0)
2088 delta = now - p->se.exec_start;
2090 return delta < (s64)sysctl_sched_migration_cost;
2094 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2096 int old_cpu = task_cpu(p);
2097 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2098 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2099 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2102 clock_offset = old_rq->clock - new_rq->clock;
2104 #ifdef CONFIG_SCHEDSTATS
2105 if (p->se.wait_start)
2106 p->se.wait_start -= clock_offset;
2107 if (p->se.sleep_start)
2108 p->se.sleep_start -= clock_offset;
2109 if (p->se.block_start)
2110 p->se.block_start -= clock_offset;
2111 if (old_cpu != new_cpu) {
2112 schedstat_inc(p, se.nr_migrations);
2113 if (task_hot(p, old_rq->clock, NULL))
2114 schedstat_inc(p, se.nr_forced2_migrations);
2117 p->se.vruntime -= old_cfsrq->min_vruntime -
2118 new_cfsrq->min_vruntime;
2120 __set_task_cpu(p, new_cpu);
2123 struct migration_req {
2124 struct list_head list;
2126 struct task_struct *task;
2129 struct completion done;
2133 * The task's runqueue lock must be held.
2134 * Returns true if you have to wait for migration thread.
2137 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2139 struct rq *rq = task_rq(p);
2142 * If the task is not on a runqueue (and not running), then
2143 * it is sufficient to simply update the task's cpu field.
2145 if (!p->se.on_rq && !task_running(rq, p)) {
2146 set_task_cpu(p, dest_cpu);
2150 init_completion(&req->done);
2152 req->dest_cpu = dest_cpu;
2153 list_add(&req->list, &rq->migration_queue);
2159 * wait_task_inactive - wait for a thread to unschedule.
2161 * The caller must ensure that the task *will* unschedule sometime soon,
2162 * else this function might spin for a *long* time. This function can't
2163 * be called with interrupts off, or it may introduce deadlock with
2164 * smp_call_function() if an IPI is sent by the same process we are
2165 * waiting to become inactive.
2167 void wait_task_inactive(struct task_struct *p)
2169 unsigned long flags;
2175 * We do the initial early heuristics without holding
2176 * any task-queue locks at all. We'll only try to get
2177 * the runqueue lock when things look like they will
2183 * If the task is actively running on another CPU
2184 * still, just relax and busy-wait without holding
2187 * NOTE! Since we don't hold any locks, it's not
2188 * even sure that "rq" stays as the right runqueue!
2189 * But we don't care, since "task_running()" will
2190 * return false if the runqueue has changed and p
2191 * is actually now running somewhere else!
2193 while (task_running(rq, p))
2197 * Ok, time to look more closely! We need the rq
2198 * lock now, to be *sure*. If we're wrong, we'll
2199 * just go back and repeat.
2201 rq = task_rq_lock(p, &flags);
2202 running = task_running(rq, p);
2203 on_rq = p->se.on_rq;
2204 task_rq_unlock(rq, &flags);
2207 * Was it really running after all now that we
2208 * checked with the proper locks actually held?
2210 * Oops. Go back and try again..
2212 if (unlikely(running)) {
2218 * It's not enough that it's not actively running,
2219 * it must be off the runqueue _entirely_, and not
2222 * So if it wa still runnable (but just not actively
2223 * running right now), it's preempted, and we should
2224 * yield - it could be a while.
2226 if (unlikely(on_rq)) {
2227 schedule_timeout_uninterruptible(1);
2232 * Ahh, all good. It wasn't running, and it wasn't
2233 * runnable, which means that it will never become
2234 * running in the future either. We're all done!
2241 * kick_process - kick a running thread to enter/exit the kernel
2242 * @p: the to-be-kicked thread
2244 * Cause a process which is running on another CPU to enter
2245 * kernel-mode, without any delay. (to get signals handled.)
2247 * NOTE: this function doesnt have to take the runqueue lock,
2248 * because all it wants to ensure is that the remote task enters
2249 * the kernel. If the IPI races and the task has been migrated
2250 * to another CPU then no harm is done and the purpose has been
2253 void kick_process(struct task_struct *p)
2259 if ((cpu != smp_processor_id()) && task_curr(p))
2260 smp_send_reschedule(cpu);
2265 * Return a low guess at the load of a migration-source cpu weighted
2266 * according to the scheduling class and "nice" value.
2268 * We want to under-estimate the load of migration sources, to
2269 * balance conservatively.
2271 static unsigned long source_load(int cpu, int type)
2273 struct rq *rq = cpu_rq(cpu);
2274 unsigned long total = weighted_cpuload(cpu);
2279 return min(rq->cpu_load[type-1], total);
2283 * Return a high guess at the load of a migration-target cpu weighted
2284 * according to the scheduling class and "nice" value.
2286 static unsigned long target_load(int cpu, int type)
2288 struct rq *rq = cpu_rq(cpu);
2289 unsigned long total = weighted_cpuload(cpu);
2294 return max(rq->cpu_load[type-1], total);
2298 * Return the average load per task on the cpu's run queue
2300 static unsigned long cpu_avg_load_per_task(int cpu)
2302 struct rq *rq = cpu_rq(cpu);
2303 unsigned long total = weighted_cpuload(cpu);
2304 unsigned long n = rq->nr_running;
2306 return n ? total / n : SCHED_LOAD_SCALE;
2310 * find_idlest_group finds and returns the least busy CPU group within the
2313 static struct sched_group *
2314 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2316 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2317 unsigned long min_load = ULONG_MAX, this_load = 0;
2318 int load_idx = sd->forkexec_idx;
2319 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2322 unsigned long load, avg_load;
2326 /* Skip over this group if it has no CPUs allowed */
2327 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2330 local_group = cpu_isset(this_cpu, group->cpumask);
2332 /* Tally up the load of all CPUs in the group */
2335 for_each_cpu_mask(i, group->cpumask) {
2336 /* Bias balancing toward cpus of our domain */
2338 load = source_load(i, load_idx);
2340 load = target_load(i, load_idx);
2345 /* Adjust by relative CPU power of the group */
2346 avg_load = sg_div_cpu_power(group,
2347 avg_load * SCHED_LOAD_SCALE);
2350 this_load = avg_load;
2352 } else if (avg_load < min_load) {
2353 min_load = avg_load;
2356 } while (group = group->next, group != sd->groups);
2358 if (!idlest || 100*this_load < imbalance*min_load)
2364 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2367 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2370 unsigned long load, min_load = ULONG_MAX;
2374 /* Traverse only the allowed CPUs */
2375 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2377 for_each_cpu_mask(i, *tmp) {
2378 load = weighted_cpuload(i);
2380 if (load < min_load || (load == min_load && i == this_cpu)) {
2390 * sched_balance_self: balance the current task (running on cpu) in domains
2391 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2394 * Balance, ie. select the least loaded group.
2396 * Returns the target CPU number, or the same CPU if no balancing is needed.
2398 * preempt must be disabled.
2400 static int sched_balance_self(int cpu, int flag)
2402 struct task_struct *t = current;
2403 struct sched_domain *tmp, *sd = NULL;
2405 for_each_domain(cpu, tmp) {
2407 * If power savings logic is enabled for a domain, stop there.
2409 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2411 if (tmp->flags & flag)
2416 cpumask_t span, tmpmask;
2417 struct sched_group *group;
2418 int new_cpu, weight;
2420 if (!(sd->flags & flag)) {
2426 group = find_idlest_group(sd, t, cpu);
2432 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2433 if (new_cpu == -1 || new_cpu == cpu) {
2434 /* Now try balancing at a lower domain level of cpu */
2439 /* Now try balancing at a lower domain level of new_cpu */
2442 weight = cpus_weight(span);
2443 for_each_domain(cpu, tmp) {
2444 if (weight <= cpus_weight(tmp->span))
2446 if (tmp->flags & flag)
2449 /* while loop will break here if sd == NULL */
2455 #endif /* CONFIG_SMP */
2458 * try_to_wake_up - wake up a thread
2459 * @p: the to-be-woken-up thread
2460 * @state: the mask of task states that can be woken
2461 * @sync: do a synchronous wakeup?
2463 * Put it on the run-queue if it's not already there. The "current"
2464 * thread is always on the run-queue (except when the actual
2465 * re-schedule is in progress), and as such you're allowed to do
2466 * the simpler "current->state = TASK_RUNNING" to mark yourself
2467 * runnable without the overhead of this.
2469 * returns failure only if the task is already active.
2471 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2473 int cpu, orig_cpu, this_cpu, success = 0;
2474 unsigned long flags;
2478 if (!sched_feat(SYNC_WAKEUPS))
2482 rq = task_rq_lock(p, &flags);
2483 old_state = p->state;
2484 if (!(old_state & state))
2492 this_cpu = smp_processor_id();
2495 if (unlikely(task_running(rq, p)))
2498 cpu = p->sched_class->select_task_rq(p, sync);
2499 if (cpu != orig_cpu) {
2500 set_task_cpu(p, cpu);
2501 task_rq_unlock(rq, &flags);
2502 /* might preempt at this point */
2503 rq = task_rq_lock(p, &flags);
2504 old_state = p->state;
2505 if (!(old_state & state))
2510 this_cpu = smp_processor_id();
2514 #ifdef CONFIG_SCHEDSTATS
2515 schedstat_inc(rq, ttwu_count);
2516 if (cpu == this_cpu)
2517 schedstat_inc(rq, ttwu_local);
2519 struct sched_domain *sd;
2520 for_each_domain(this_cpu, sd) {
2521 if (cpu_isset(cpu, sd->span)) {
2522 schedstat_inc(sd, ttwu_wake_remote);
2530 #endif /* CONFIG_SMP */
2531 schedstat_inc(p, se.nr_wakeups);
2533 schedstat_inc(p, se.nr_wakeups_sync);
2534 if (orig_cpu != cpu)
2535 schedstat_inc(p, se.nr_wakeups_migrate);
2536 if (cpu == this_cpu)
2537 schedstat_inc(p, se.nr_wakeups_local);
2539 schedstat_inc(p, se.nr_wakeups_remote);
2540 update_rq_clock(rq);
2541 activate_task(rq, p, 1);
2545 check_preempt_curr(rq, p);
2547 p->state = TASK_RUNNING;
2549 if (p->sched_class->task_wake_up)
2550 p->sched_class->task_wake_up(rq, p);
2553 task_rq_unlock(rq, &flags);
2558 int wake_up_process(struct task_struct *p)
2560 return try_to_wake_up(p, TASK_ALL, 0);
2562 EXPORT_SYMBOL(wake_up_process);
2564 int wake_up_state(struct task_struct *p, unsigned int state)
2566 return try_to_wake_up(p, state, 0);
2570 * Perform scheduler related setup for a newly forked process p.
2571 * p is forked by current.
2573 * __sched_fork() is basic setup used by init_idle() too:
2575 static void __sched_fork(struct task_struct *p)
2577 p->se.exec_start = 0;
2578 p->se.sum_exec_runtime = 0;
2579 p->se.prev_sum_exec_runtime = 0;
2580 p->se.last_wakeup = 0;
2581 p->se.avg_overlap = 0;
2583 #ifdef CONFIG_SCHEDSTATS
2584 p->se.wait_start = 0;
2585 p->se.sum_sleep_runtime = 0;
2586 p->se.sleep_start = 0;
2587 p->se.block_start = 0;
2588 p->se.sleep_max = 0;
2589 p->se.block_max = 0;
2591 p->se.slice_max = 0;
2595 INIT_LIST_HEAD(&p->rt.run_list);
2597 INIT_LIST_HEAD(&p->se.group_node);
2599 #ifdef CONFIG_PREEMPT_NOTIFIERS
2600 INIT_HLIST_HEAD(&p->preempt_notifiers);
2604 * We mark the process as running here, but have not actually
2605 * inserted it onto the runqueue yet. This guarantees that
2606 * nobody will actually run it, and a signal or other external
2607 * event cannot wake it up and insert it on the runqueue either.
2609 p->state = TASK_RUNNING;
2613 * fork()/clone()-time setup:
2615 void sched_fork(struct task_struct *p, int clone_flags)
2617 int cpu = get_cpu();
2622 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2624 set_task_cpu(p, cpu);
2627 * Make sure we do not leak PI boosting priority to the child:
2629 p->prio = current->normal_prio;
2630 if (!rt_prio(p->prio))
2631 p->sched_class = &fair_sched_class;
2633 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2634 if (likely(sched_info_on()))
2635 memset(&p->sched_info, 0, sizeof(p->sched_info));
2637 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2640 #ifdef CONFIG_PREEMPT
2641 /* Want to start with kernel preemption disabled. */
2642 task_thread_info(p)->preempt_count = 1;
2648 * wake_up_new_task - wake up a newly created task for the first time.
2650 * This function will do some initial scheduler statistics housekeeping
2651 * that must be done for every newly created context, then puts the task
2652 * on the runqueue and wakes it.
2654 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2656 unsigned long flags;
2659 rq = task_rq_lock(p, &flags);
2660 BUG_ON(p->state != TASK_RUNNING);
2661 update_rq_clock(rq);
2663 p->prio = effective_prio(p);
2665 if (!p->sched_class->task_new || !current->se.on_rq) {
2666 activate_task(rq, p, 0);
2669 * Let the scheduling class do new task startup
2670 * management (if any):
2672 p->sched_class->task_new(rq, p);
2675 check_preempt_curr(rq, p);
2677 if (p->sched_class->task_wake_up)
2678 p->sched_class->task_wake_up(rq, p);
2680 task_rq_unlock(rq, &flags);
2683 #ifdef CONFIG_PREEMPT_NOTIFIERS
2686 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2687 * @notifier: notifier struct to register
2689 void preempt_notifier_register(struct preempt_notifier *notifier)
2691 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2693 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2696 * preempt_notifier_unregister - no longer interested in preemption notifications
2697 * @notifier: notifier struct to unregister
2699 * This is safe to call from within a preemption notifier.
2701 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2703 hlist_del(¬ifier->link);
2705 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2707 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2709 struct preempt_notifier *notifier;
2710 struct hlist_node *node;
2712 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2713 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2717 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2718 struct task_struct *next)
2720 struct preempt_notifier *notifier;
2721 struct hlist_node *node;
2723 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2724 notifier->ops->sched_out(notifier, next);
2729 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2734 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2735 struct task_struct *next)
2742 * prepare_task_switch - prepare to switch tasks
2743 * @rq: the runqueue preparing to switch
2744 * @prev: the current task that is being switched out
2745 * @next: the task we are going to switch to.
2747 * This is called with the rq lock held and interrupts off. It must
2748 * be paired with a subsequent finish_task_switch after the context
2751 * prepare_task_switch sets up locking and calls architecture specific
2755 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2756 struct task_struct *next)
2758 fire_sched_out_preempt_notifiers(prev, next);
2759 prepare_lock_switch(rq, next);
2760 prepare_arch_switch(next);
2764 * finish_task_switch - clean up after a task-switch
2765 * @rq: runqueue associated with task-switch
2766 * @prev: the thread we just switched away from.
2768 * finish_task_switch must be called after the context switch, paired
2769 * with a prepare_task_switch call before the context switch.
2770 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2771 * and do any other architecture-specific cleanup actions.
2773 * Note that we may have delayed dropping an mm in context_switch(). If
2774 * so, we finish that here outside of the runqueue lock. (Doing it
2775 * with the lock held can cause deadlocks; see schedule() for
2778 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2779 __releases(rq->lock)
2781 struct mm_struct *mm = rq->prev_mm;
2787 * A task struct has one reference for the use as "current".
2788 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2789 * schedule one last time. The schedule call will never return, and
2790 * the scheduled task must drop that reference.
2791 * The test for TASK_DEAD must occur while the runqueue locks are
2792 * still held, otherwise prev could be scheduled on another cpu, die
2793 * there before we look at prev->state, and then the reference would
2795 * Manfred Spraul <manfred@colorfullife.com>
2797 prev_state = prev->state;
2798 finish_arch_switch(prev);
2799 finish_lock_switch(rq, prev);
2801 if (current->sched_class->post_schedule)
2802 current->sched_class->post_schedule(rq);
2805 fire_sched_in_preempt_notifiers(current);
2808 if (unlikely(prev_state == TASK_DEAD)) {
2810 * Remove function-return probe instances associated with this
2811 * task and put them back on the free list.
2813 kprobe_flush_task(prev);
2814 put_task_struct(prev);
2819 * schedule_tail - first thing a freshly forked thread must call.
2820 * @prev: the thread we just switched away from.
2822 asmlinkage void schedule_tail(struct task_struct *prev)
2823 __releases(rq->lock)
2825 struct rq *rq = this_rq();
2827 finish_task_switch(rq, prev);
2828 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2829 /* In this case, finish_task_switch does not reenable preemption */
2832 if (current->set_child_tid)
2833 put_user(task_pid_vnr(current), current->set_child_tid);
2837 * context_switch - switch to the new MM and the new
2838 * thread's register state.
2841 context_switch(struct rq *rq, struct task_struct *prev,
2842 struct task_struct *next)
2844 struct mm_struct *mm, *oldmm;
2846 prepare_task_switch(rq, prev, next);
2848 oldmm = prev->active_mm;
2850 * For paravirt, this is coupled with an exit in switch_to to
2851 * combine the page table reload and the switch backend into
2854 arch_enter_lazy_cpu_mode();
2856 if (unlikely(!mm)) {
2857 next->active_mm = oldmm;
2858 atomic_inc(&oldmm->mm_count);
2859 enter_lazy_tlb(oldmm, next);
2861 switch_mm(oldmm, mm, next);
2863 if (unlikely(!prev->mm)) {
2864 prev->active_mm = NULL;
2865 rq->prev_mm = oldmm;
2868 * Since the runqueue lock will be released by the next
2869 * task (which is an invalid locking op but in the case
2870 * of the scheduler it's an obvious special-case), so we
2871 * do an early lockdep release here:
2873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2874 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2877 /* Here we just switch the register state and the stack. */
2878 switch_to(prev, next, prev);
2882 * this_rq must be evaluated again because prev may have moved
2883 * CPUs since it called schedule(), thus the 'rq' on its stack
2884 * frame will be invalid.
2886 finish_task_switch(this_rq(), prev);
2890 * nr_running, nr_uninterruptible and nr_context_switches:
2892 * externally visible scheduler statistics: current number of runnable
2893 * threads, current number of uninterruptible-sleeping threads, total
2894 * number of context switches performed since bootup.
2896 unsigned long nr_running(void)
2898 unsigned long i, sum = 0;
2900 for_each_online_cpu(i)
2901 sum += cpu_rq(i)->nr_running;
2906 unsigned long nr_uninterruptible(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += cpu_rq(i)->nr_uninterruptible;
2914 * Since we read the counters lockless, it might be slightly
2915 * inaccurate. Do not allow it to go below zero though:
2917 if (unlikely((long)sum < 0))
2923 unsigned long long nr_context_switches(void)
2926 unsigned long long sum = 0;
2928 for_each_possible_cpu(i)
2929 sum += cpu_rq(i)->nr_switches;
2934 unsigned long nr_iowait(void)
2936 unsigned long i, sum = 0;
2938 for_each_possible_cpu(i)
2939 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2944 unsigned long nr_active(void)
2946 unsigned long i, running = 0, uninterruptible = 0;
2948 for_each_online_cpu(i) {
2949 running += cpu_rq(i)->nr_running;
2950 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2953 if (unlikely((long)uninterruptible < 0))
2954 uninterruptible = 0;
2956 return running + uninterruptible;
2960 * Update rq->cpu_load[] statistics. This function is usually called every
2961 * scheduler tick (TICK_NSEC).
2963 static void update_cpu_load(struct rq *this_rq)
2965 unsigned long this_load = this_rq->load.weight;
2968 this_rq->nr_load_updates++;
2970 /* Update our load: */
2971 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2972 unsigned long old_load, new_load;
2974 /* scale is effectively 1 << i now, and >> i divides by scale */
2976 old_load = this_rq->cpu_load[i];
2977 new_load = this_load;
2979 * Round up the averaging division if load is increasing. This
2980 * prevents us from getting stuck on 9 if the load is 10, for
2983 if (new_load > old_load)
2984 new_load += scale-1;
2985 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2992 * double_rq_lock - safely lock two runqueues
2994 * Note this does not disable interrupts like task_rq_lock,
2995 * you need to do so manually before calling.
2997 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2998 __acquires(rq1->lock)
2999 __acquires(rq2->lock)
3001 BUG_ON(!irqs_disabled());
3003 spin_lock(&rq1->lock);
3004 __acquire(rq2->lock); /* Fake it out ;) */
3007 spin_lock(&rq1->lock);
3008 spin_lock(&rq2->lock);
3010 spin_lock(&rq2->lock);
3011 spin_lock(&rq1->lock);
3014 update_rq_clock(rq1);
3015 update_rq_clock(rq2);
3019 * double_rq_unlock - safely unlock two runqueues
3021 * Note this does not restore interrupts like task_rq_unlock,
3022 * you need to do so manually after calling.
3024 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3025 __releases(rq1->lock)
3026 __releases(rq2->lock)
3028 spin_unlock(&rq1->lock);
3030 spin_unlock(&rq2->lock);
3032 __release(rq2->lock);
3036 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3038 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
3039 __releases(this_rq->lock)
3040 __acquires(busiest->lock)
3041 __acquires(this_rq->lock)
3045 if (unlikely(!irqs_disabled())) {
3046 /* printk() doesn't work good under rq->lock */
3047 spin_unlock(&this_rq->lock);
3050 if (unlikely(!spin_trylock(&busiest->lock))) {
3051 if (busiest < this_rq) {
3052 spin_unlock(&this_rq->lock);
3053 spin_lock(&busiest->lock);
3054 spin_lock(&this_rq->lock);
3057 spin_lock(&busiest->lock);
3063 * If dest_cpu is allowed for this process, migrate the task to it.
3064 * This is accomplished by forcing the cpu_allowed mask to only
3065 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3066 * the cpu_allowed mask is restored.
3068 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3070 struct migration_req req;
3071 unsigned long flags;
3074 rq = task_rq_lock(p, &flags);
3075 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3076 || unlikely(cpu_is_offline(dest_cpu)))
3079 /* force the process onto the specified CPU */
3080 if (migrate_task(p, dest_cpu, &req)) {
3081 /* Need to wait for migration thread (might exit: take ref). */
3082 struct task_struct *mt = rq->migration_thread;
3084 get_task_struct(mt);
3085 task_rq_unlock(rq, &flags);
3086 wake_up_process(mt);
3087 put_task_struct(mt);
3088 wait_for_completion(&req.done);
3093 task_rq_unlock(rq, &flags);
3097 * sched_exec - execve() is a valuable balancing opportunity, because at
3098 * this point the task has the smallest effective memory and cache footprint.
3100 void sched_exec(void)
3102 int new_cpu, this_cpu = get_cpu();
3103 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3105 if (new_cpu != this_cpu)
3106 sched_migrate_task(current, new_cpu);
3110 * pull_task - move a task from a remote runqueue to the local runqueue.
3111 * Both runqueues must be locked.
3113 static void pull_task(struct rq *src_rq, struct task_struct *p,
3114 struct rq *this_rq, int this_cpu)
3116 deactivate_task(src_rq, p, 0);
3117 set_task_cpu(p, this_cpu);
3118 activate_task(this_rq, p, 0);
3120 * Note that idle threads have a prio of MAX_PRIO, for this test
3121 * to be always true for them.
3123 check_preempt_curr(this_rq, p);
3127 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3130 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3131 struct sched_domain *sd, enum cpu_idle_type idle,
3135 * We do not migrate tasks that are:
3136 * 1) running (obviously), or
3137 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3138 * 3) are cache-hot on their current CPU.
3140 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3141 schedstat_inc(p, se.nr_failed_migrations_affine);
3146 if (task_running(rq, p)) {
3147 schedstat_inc(p, se.nr_failed_migrations_running);
3152 * Aggressive migration if:
3153 * 1) task is cache cold, or
3154 * 2) too many balance attempts have failed.
3157 if (!task_hot(p, rq->clock, sd) ||
3158 sd->nr_balance_failed > sd->cache_nice_tries) {
3159 #ifdef CONFIG_SCHEDSTATS
3160 if (task_hot(p, rq->clock, sd)) {
3161 schedstat_inc(sd, lb_hot_gained[idle]);
3162 schedstat_inc(p, se.nr_forced_migrations);
3168 if (task_hot(p, rq->clock, sd)) {
3169 schedstat_inc(p, se.nr_failed_migrations_hot);
3175 static unsigned long
3176 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3177 unsigned long max_load_move, struct sched_domain *sd,
3178 enum cpu_idle_type idle, int *all_pinned,
3179 int *this_best_prio, struct rq_iterator *iterator)
3181 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3182 struct task_struct *p;
3183 long rem_load_move = max_load_move;
3185 if (max_load_move == 0)
3191 * Start the load-balancing iterator:
3193 p = iterator->start(iterator->arg);
3195 if (!p || loops++ > sysctl_sched_nr_migrate)
3198 * To help distribute high priority tasks across CPUs we don't
3199 * skip a task if it will be the highest priority task (i.e. smallest
3200 * prio value) on its new queue regardless of its load weight
3202 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3203 SCHED_LOAD_SCALE_FUZZ;
3204 if ((skip_for_load && p->prio >= *this_best_prio) ||
3205 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3206 p = iterator->next(iterator->arg);
3210 pull_task(busiest, p, this_rq, this_cpu);
3212 rem_load_move -= p->se.load.weight;
3215 * We only want to steal up to the prescribed amount of weighted load.
3217 if (rem_load_move > 0) {
3218 if (p->prio < *this_best_prio)
3219 *this_best_prio = p->prio;
3220 p = iterator->next(iterator->arg);
3225 * Right now, this is one of only two places pull_task() is called,
3226 * so we can safely collect pull_task() stats here rather than
3227 * inside pull_task().
3229 schedstat_add(sd, lb_gained[idle], pulled);
3232 *all_pinned = pinned;
3234 return max_load_move - rem_load_move;
3238 * move_tasks tries to move up to max_load_move weighted load from busiest to
3239 * this_rq, as part of a balancing operation within domain "sd".
3240 * Returns 1 if successful and 0 otherwise.
3242 * Called with both runqueues locked.
3244 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3245 unsigned long max_load_move,
3246 struct sched_domain *sd, enum cpu_idle_type idle,
3249 const struct sched_class *class = sched_class_highest;
3250 unsigned long total_load_moved = 0;
3251 int this_best_prio = this_rq->curr->prio;
3255 class->load_balance(this_rq, this_cpu, busiest,
3256 max_load_move - total_load_moved,
3257 sd, idle, all_pinned, &this_best_prio);
3258 class = class->next;
3259 } while (class && max_load_move > total_load_moved);
3261 return total_load_moved > 0;
3265 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3266 struct sched_domain *sd, enum cpu_idle_type idle,
3267 struct rq_iterator *iterator)
3269 struct task_struct *p = iterator->start(iterator->arg);
3273 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3274 pull_task(busiest, p, this_rq, this_cpu);
3276 * Right now, this is only the second place pull_task()
3277 * is called, so we can safely collect pull_task()
3278 * stats here rather than inside pull_task().
3280 schedstat_inc(sd, lb_gained[idle]);
3284 p = iterator->next(iterator->arg);
3291 * move_one_task tries to move exactly one task from busiest to this_rq, as
3292 * part of active balancing operations within "domain".
3293 * Returns 1 if successful and 0 otherwise.
3295 * Called with both runqueues locked.
3297 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3298 struct sched_domain *sd, enum cpu_idle_type idle)
3300 const struct sched_class *class;
3302 for (class = sched_class_highest; class; class = class->next)
3303 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3310 * find_busiest_group finds and returns the busiest CPU group within the
3311 * domain. It calculates and returns the amount of weighted load which
3312 * should be moved to restore balance via the imbalance parameter.
3314 static struct sched_group *
3315 find_busiest_group(struct sched_domain *sd, int this_cpu,
3316 unsigned long *imbalance, enum cpu_idle_type idle,
3317 int *sd_idle, const cpumask_t *cpus, int *balance)
3319 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3320 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3321 unsigned long max_pull;
3322 unsigned long busiest_load_per_task, busiest_nr_running;
3323 unsigned long this_load_per_task, this_nr_running;
3324 int load_idx, group_imb = 0;
3325 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3326 int power_savings_balance = 1;
3327 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3328 unsigned long min_nr_running = ULONG_MAX;
3329 struct sched_group *group_min = NULL, *group_leader = NULL;
3332 max_load = this_load = total_load = total_pwr = 0;
3333 busiest_load_per_task = busiest_nr_running = 0;
3334 this_load_per_task = this_nr_running = 0;
3335 if (idle == CPU_NOT_IDLE)
3336 load_idx = sd->busy_idx;
3337 else if (idle == CPU_NEWLY_IDLE)
3338 load_idx = sd->newidle_idx;
3340 load_idx = sd->idle_idx;
3343 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3346 int __group_imb = 0;
3347 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3348 unsigned long sum_nr_running, sum_weighted_load;
3350 local_group = cpu_isset(this_cpu, group->cpumask);
3353 balance_cpu = first_cpu(group->cpumask);
3355 /* Tally up the load of all CPUs in the group */
3356 sum_weighted_load = sum_nr_running = avg_load = 0;
3358 min_cpu_load = ~0UL;
3360 for_each_cpu_mask(i, group->cpumask) {
3363 if (!cpu_isset(i, *cpus))
3368 if (*sd_idle && rq->nr_running)
3371 /* Bias balancing toward cpus of our domain */
3373 if (idle_cpu(i) && !first_idle_cpu) {
3378 load = target_load(i, load_idx);
3380 load = source_load(i, load_idx);
3381 if (load > max_cpu_load)
3382 max_cpu_load = load;
3383 if (min_cpu_load > load)
3384 min_cpu_load = load;
3388 sum_nr_running += rq->nr_running;
3389 sum_weighted_load += weighted_cpuload(i);
3393 * First idle cpu or the first cpu(busiest) in this sched group
3394 * is eligible for doing load balancing at this and above
3395 * domains. In the newly idle case, we will allow all the cpu's
3396 * to do the newly idle load balance.
3398 if (idle != CPU_NEWLY_IDLE && local_group &&
3399 balance_cpu != this_cpu && balance) {
3404 total_load += avg_load;
3405 total_pwr += group->__cpu_power;
3407 /* Adjust by relative CPU power of the group */
3408 avg_load = sg_div_cpu_power(group,
3409 avg_load * SCHED_LOAD_SCALE);
3411 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3414 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3417 this_load = avg_load;
3419 this_nr_running = sum_nr_running;
3420 this_load_per_task = sum_weighted_load;
3421 } else if (avg_load > max_load &&
3422 (sum_nr_running > group_capacity || __group_imb)) {
3423 max_load = avg_load;
3425 busiest_nr_running = sum_nr_running;
3426 busiest_load_per_task = sum_weighted_load;
3427 group_imb = __group_imb;
3430 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3432 * Busy processors will not participate in power savings
3435 if (idle == CPU_NOT_IDLE ||
3436 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3440 * If the local group is idle or completely loaded
3441 * no need to do power savings balance at this domain
3443 if (local_group && (this_nr_running >= group_capacity ||
3445 power_savings_balance = 0;
3448 * If a group is already running at full capacity or idle,
3449 * don't include that group in power savings calculations
3451 if (!power_savings_balance || sum_nr_running >= group_capacity
3456 * Calculate the group which has the least non-idle load.
3457 * This is the group from where we need to pick up the load
3460 if ((sum_nr_running < min_nr_running) ||
3461 (sum_nr_running == min_nr_running &&
3462 first_cpu(group->cpumask) <
3463 first_cpu(group_min->cpumask))) {
3465 min_nr_running = sum_nr_running;
3466 min_load_per_task = sum_weighted_load /
3471 * Calculate the group which is almost near its
3472 * capacity but still has some space to pick up some load
3473 * from other group and save more power
3475 if (sum_nr_running <= group_capacity - 1) {
3476 if (sum_nr_running > leader_nr_running ||
3477 (sum_nr_running == leader_nr_running &&
3478 first_cpu(group->cpumask) >
3479 first_cpu(group_leader->cpumask))) {
3480 group_leader = group;
3481 leader_nr_running = sum_nr_running;
3486 group = group->next;
3487 } while (group != sd->groups);
3489 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3492 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3494 if (this_load >= avg_load ||
3495 100*max_load <= sd->imbalance_pct*this_load)
3498 busiest_load_per_task /= busiest_nr_running;
3500 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3503 * We're trying to get all the cpus to the average_load, so we don't
3504 * want to push ourselves above the average load, nor do we wish to
3505 * reduce the max loaded cpu below the average load, as either of these
3506 * actions would just result in more rebalancing later, and ping-pong
3507 * tasks around. Thus we look for the minimum possible imbalance.
3508 * Negative imbalances (*we* are more loaded than anyone else) will
3509 * be counted as no imbalance for these purposes -- we can't fix that
3510 * by pulling tasks to us. Be careful of negative numbers as they'll
3511 * appear as very large values with unsigned longs.
3513 if (max_load <= busiest_load_per_task)
3517 * In the presence of smp nice balancing, certain scenarios can have
3518 * max load less than avg load(as we skip the groups at or below
3519 * its cpu_power, while calculating max_load..)
3521 if (max_load < avg_load) {
3523 goto small_imbalance;
3526 /* Don't want to pull so many tasks that a group would go idle */
3527 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3529 /* How much load to actually move to equalise the imbalance */
3530 *imbalance = min(max_pull * busiest->__cpu_power,
3531 (avg_load - this_load) * this->__cpu_power)
3535 * if *imbalance is less than the average load per runnable task
3536 * there is no gaurantee that any tasks will be moved so we'll have
3537 * a think about bumping its value to force at least one task to be
3540 if (*imbalance < busiest_load_per_task) {
3541 unsigned long tmp, pwr_now, pwr_move;
3545 pwr_move = pwr_now = 0;
3547 if (this_nr_running) {
3548 this_load_per_task /= this_nr_running;
3549 if (busiest_load_per_task > this_load_per_task)
3552 this_load_per_task = SCHED_LOAD_SCALE;
3554 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3555 busiest_load_per_task * imbn) {
3556 *imbalance = busiest_load_per_task;
3561 * OK, we don't have enough imbalance to justify moving tasks,
3562 * however we may be able to increase total CPU power used by
3566 pwr_now += busiest->__cpu_power *
3567 min(busiest_load_per_task, max_load);
3568 pwr_now += this->__cpu_power *
3569 min(this_load_per_task, this_load);
3570 pwr_now /= SCHED_LOAD_SCALE;
3572 /* Amount of load we'd subtract */
3573 tmp = sg_div_cpu_power(busiest,
3574 busiest_load_per_task * SCHED_LOAD_SCALE);
3576 pwr_move += busiest->__cpu_power *
3577 min(busiest_load_per_task, max_load - tmp);
3579 /* Amount of load we'd add */
3580 if (max_load * busiest->__cpu_power <
3581 busiest_load_per_task * SCHED_LOAD_SCALE)
3582 tmp = sg_div_cpu_power(this,
3583 max_load * busiest->__cpu_power);
3585 tmp = sg_div_cpu_power(this,
3586 busiest_load_per_task * SCHED_LOAD_SCALE);
3587 pwr_move += this->__cpu_power *
3588 min(this_load_per_task, this_load + tmp);
3589 pwr_move /= SCHED_LOAD_SCALE;
3591 /* Move if we gain throughput */
3592 if (pwr_move > pwr_now)
3593 *imbalance = busiest_load_per_task;
3599 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3600 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3603 if (this == group_leader && group_leader != group_min) {
3604 *imbalance = min_load_per_task;
3614 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3617 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3618 unsigned long imbalance, const cpumask_t *cpus)
3620 struct rq *busiest = NULL, *rq;
3621 unsigned long max_load = 0;
3624 for_each_cpu_mask(i, group->cpumask) {
3627 if (!cpu_isset(i, *cpus))
3631 wl = weighted_cpuload(i);
3633 if (rq->nr_running == 1 && wl > imbalance)
3636 if (wl > max_load) {
3646 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3647 * so long as it is large enough.
3649 #define MAX_PINNED_INTERVAL 512
3652 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3653 * tasks if there is an imbalance.
3655 static int load_balance(int this_cpu, struct rq *this_rq,
3656 struct sched_domain *sd, enum cpu_idle_type idle,
3657 int *balance, cpumask_t *cpus)
3659 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3660 struct sched_group *group;
3661 unsigned long imbalance;
3663 unsigned long flags;
3664 int unlock_aggregate;
3668 unlock_aggregate = get_aggregate(sd);
3671 * When power savings policy is enabled for the parent domain, idle
3672 * sibling can pick up load irrespective of busy siblings. In this case,
3673 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3674 * portraying it as CPU_NOT_IDLE.
3676 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3677 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3680 schedstat_inc(sd, lb_count[idle]);
3683 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3690 schedstat_inc(sd, lb_nobusyg[idle]);
3694 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3696 schedstat_inc(sd, lb_nobusyq[idle]);
3700 BUG_ON(busiest == this_rq);
3702 schedstat_add(sd, lb_imbalance[idle], imbalance);
3705 if (busiest->nr_running > 1) {
3707 * Attempt to move tasks. If find_busiest_group has found
3708 * an imbalance but busiest->nr_running <= 1, the group is
3709 * still unbalanced. ld_moved simply stays zero, so it is
3710 * correctly treated as an imbalance.
3712 local_irq_save(flags);
3713 double_rq_lock(this_rq, busiest);
3714 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3715 imbalance, sd, idle, &all_pinned);
3716 double_rq_unlock(this_rq, busiest);
3717 local_irq_restore(flags);
3720 * some other cpu did the load balance for us.
3722 if (ld_moved && this_cpu != smp_processor_id())
3723 resched_cpu(this_cpu);
3725 /* All tasks on this runqueue were pinned by CPU affinity */
3726 if (unlikely(all_pinned)) {
3727 cpu_clear(cpu_of(busiest), *cpus);
3728 if (!cpus_empty(*cpus))
3735 schedstat_inc(sd, lb_failed[idle]);
3736 sd->nr_balance_failed++;
3738 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3740 spin_lock_irqsave(&busiest->lock, flags);
3742 /* don't kick the migration_thread, if the curr
3743 * task on busiest cpu can't be moved to this_cpu
3745 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3746 spin_unlock_irqrestore(&busiest->lock, flags);
3748 goto out_one_pinned;
3751 if (!busiest->active_balance) {
3752 busiest->active_balance = 1;
3753 busiest->push_cpu = this_cpu;
3756 spin_unlock_irqrestore(&busiest->lock, flags);
3758 wake_up_process(busiest->migration_thread);
3761 * We've kicked active balancing, reset the failure
3764 sd->nr_balance_failed = sd->cache_nice_tries+1;
3767 sd->nr_balance_failed = 0;
3769 if (likely(!active_balance)) {
3770 /* We were unbalanced, so reset the balancing interval */
3771 sd->balance_interval = sd->min_interval;
3774 * If we've begun active balancing, start to back off. This
3775 * case may not be covered by the all_pinned logic if there
3776 * is only 1 task on the busy runqueue (because we don't call
3779 if (sd->balance_interval < sd->max_interval)
3780 sd->balance_interval *= 2;
3783 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3784 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3790 schedstat_inc(sd, lb_balanced[idle]);
3792 sd->nr_balance_failed = 0;
3795 /* tune up the balancing interval */
3796 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3797 (sd->balance_interval < sd->max_interval))
3798 sd->balance_interval *= 2;
3800 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3801 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3806 if (unlock_aggregate)
3812 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3813 * tasks if there is an imbalance.
3815 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3816 * this_rq is locked.
3819 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3822 struct sched_group *group;
3823 struct rq *busiest = NULL;
3824 unsigned long imbalance;
3832 * When power savings policy is enabled for the parent domain, idle
3833 * sibling can pick up load irrespective of busy siblings. In this case,
3834 * let the state of idle sibling percolate up as IDLE, instead of
3835 * portraying it as CPU_NOT_IDLE.
3837 if (sd->flags & SD_SHARE_CPUPOWER &&
3838 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3841 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3843 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3844 &sd_idle, cpus, NULL);
3846 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3850 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3852 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3856 BUG_ON(busiest == this_rq);
3858 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3861 if (busiest->nr_running > 1) {
3862 /* Attempt to move tasks */
3863 double_lock_balance(this_rq, busiest);
3864 /* this_rq->clock is already updated */
3865 update_rq_clock(busiest);
3866 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3867 imbalance, sd, CPU_NEWLY_IDLE,
3869 spin_unlock(&busiest->lock);
3871 if (unlikely(all_pinned)) {
3872 cpu_clear(cpu_of(busiest), *cpus);
3873 if (!cpus_empty(*cpus))
3879 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3880 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3881 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3884 sd->nr_balance_failed = 0;
3889 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3890 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3891 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3893 sd->nr_balance_failed = 0;
3899 * idle_balance is called by schedule() if this_cpu is about to become
3900 * idle. Attempts to pull tasks from other CPUs.
3902 static void idle_balance(int this_cpu, struct rq *this_rq)
3904 struct sched_domain *sd;
3905 int pulled_task = -1;
3906 unsigned long next_balance = jiffies + HZ;
3909 for_each_domain(this_cpu, sd) {
3910 unsigned long interval;
3912 if (!(sd->flags & SD_LOAD_BALANCE))
3915 if (sd->flags & SD_BALANCE_NEWIDLE)
3916 /* If we've pulled tasks over stop searching: */
3917 pulled_task = load_balance_newidle(this_cpu, this_rq,
3920 interval = msecs_to_jiffies(sd->balance_interval);
3921 if (time_after(next_balance, sd->last_balance + interval))
3922 next_balance = sd->last_balance + interval;
3926 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3928 * We are going idle. next_balance may be set based on
3929 * a busy processor. So reset next_balance.
3931 this_rq->next_balance = next_balance;
3936 * active_load_balance is run by migration threads. It pushes running tasks
3937 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3938 * running on each physical CPU where possible, and avoids physical /
3939 * logical imbalances.
3941 * Called with busiest_rq locked.
3943 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3945 int target_cpu = busiest_rq->push_cpu;
3946 struct sched_domain *sd;
3947 struct rq *target_rq;
3949 /* Is there any task to move? */
3950 if (busiest_rq->nr_running <= 1)
3953 target_rq = cpu_rq(target_cpu);
3956 * This condition is "impossible", if it occurs
3957 * we need to fix it. Originally reported by
3958 * Bjorn Helgaas on a 128-cpu setup.
3960 BUG_ON(busiest_rq == target_rq);
3962 /* move a task from busiest_rq to target_rq */
3963 double_lock_balance(busiest_rq, target_rq);
3964 update_rq_clock(busiest_rq);
3965 update_rq_clock(target_rq);
3967 /* Search for an sd spanning us and the target CPU. */
3968 for_each_domain(target_cpu, sd) {
3969 if ((sd->flags & SD_LOAD_BALANCE) &&
3970 cpu_isset(busiest_cpu, sd->span))
3975 schedstat_inc(sd, alb_count);
3977 if (move_one_task(target_rq, target_cpu, busiest_rq,
3979 schedstat_inc(sd, alb_pushed);
3981 schedstat_inc(sd, alb_failed);
3983 spin_unlock(&target_rq->lock);
3988 atomic_t load_balancer;
3990 } nohz ____cacheline_aligned = {
3991 .load_balancer = ATOMIC_INIT(-1),
3992 .cpu_mask = CPU_MASK_NONE,
3996 * This routine will try to nominate the ilb (idle load balancing)
3997 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3998 * load balancing on behalf of all those cpus. If all the cpus in the system
3999 * go into this tickless mode, then there will be no ilb owner (as there is
4000 * no need for one) and all the cpus will sleep till the next wakeup event
4003 * For the ilb owner, tick is not stopped. And this tick will be used
4004 * for idle load balancing. ilb owner will still be part of
4007 * While stopping the tick, this cpu will become the ilb owner if there
4008 * is no other owner. And will be the owner till that cpu becomes busy
4009 * or if all cpus in the system stop their ticks at which point
4010 * there is no need for ilb owner.
4012 * When the ilb owner becomes busy, it nominates another owner, during the
4013 * next busy scheduler_tick()
4015 int select_nohz_load_balancer(int stop_tick)
4017 int cpu = smp_processor_id();
4020 cpu_set(cpu, nohz.cpu_mask);
4021 cpu_rq(cpu)->in_nohz_recently = 1;
4024 * If we are going offline and still the leader, give up!
4026 if (cpu_is_offline(cpu) &&
4027 atomic_read(&nohz.load_balancer) == cpu) {
4028 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4033 /* time for ilb owner also to sleep */
4034 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4035 if (atomic_read(&nohz.load_balancer) == cpu)
4036 atomic_set(&nohz.load_balancer, -1);
4040 if (atomic_read(&nohz.load_balancer) == -1) {
4041 /* make me the ilb owner */
4042 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4044 } else if (atomic_read(&nohz.load_balancer) == cpu)
4047 if (!cpu_isset(cpu, nohz.cpu_mask))
4050 cpu_clear(cpu, nohz.cpu_mask);
4052 if (atomic_read(&nohz.load_balancer) == cpu)
4053 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4060 static DEFINE_SPINLOCK(balancing);
4063 * It checks each scheduling domain to see if it is due to be balanced,
4064 * and initiates a balancing operation if so.
4066 * Balancing parameters are set up in arch_init_sched_domains.
4068 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4071 struct rq *rq = cpu_rq(cpu);
4072 unsigned long interval;
4073 struct sched_domain *sd;
4074 /* Earliest time when we have to do rebalance again */
4075 unsigned long next_balance = jiffies + 60*HZ;
4076 int update_next_balance = 0;
4079 for_each_domain(cpu, sd) {
4080 if (!(sd->flags & SD_LOAD_BALANCE))
4083 interval = sd->balance_interval;
4084 if (idle != CPU_IDLE)
4085 interval *= sd->busy_factor;
4087 /* scale ms to jiffies */
4088 interval = msecs_to_jiffies(interval);
4089 if (unlikely(!interval))
4091 if (interval > HZ*NR_CPUS/10)
4092 interval = HZ*NR_CPUS/10;
4095 if (sd->flags & SD_SERIALIZE) {
4096 if (!spin_trylock(&balancing))
4100 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4101 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4103 * We've pulled tasks over so either we're no
4104 * longer idle, or one of our SMT siblings is
4107 idle = CPU_NOT_IDLE;
4109 sd->last_balance = jiffies;
4111 if (sd->flags & SD_SERIALIZE)
4112 spin_unlock(&balancing);
4114 if (time_after(next_balance, sd->last_balance + interval)) {
4115 next_balance = sd->last_balance + interval;
4116 update_next_balance = 1;
4120 * Stop the load balance at this level. There is another
4121 * CPU in our sched group which is doing load balancing more
4129 * next_balance will be updated only when there is a need.
4130 * When the cpu is attached to null domain for ex, it will not be
4133 if (likely(update_next_balance))
4134 rq->next_balance = next_balance;
4138 * run_rebalance_domains is triggered when needed from the scheduler tick.
4139 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4140 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4142 static void run_rebalance_domains(struct softirq_action *h)
4144 int this_cpu = smp_processor_id();
4145 struct rq *this_rq = cpu_rq(this_cpu);
4146 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4147 CPU_IDLE : CPU_NOT_IDLE;
4149 rebalance_domains(this_cpu, idle);
4153 * If this cpu is the owner for idle load balancing, then do the
4154 * balancing on behalf of the other idle cpus whose ticks are
4157 if (this_rq->idle_at_tick &&
4158 atomic_read(&nohz.load_balancer) == this_cpu) {
4159 cpumask_t cpus = nohz.cpu_mask;
4163 cpu_clear(this_cpu, cpus);
4164 for_each_cpu_mask(balance_cpu, cpus) {
4166 * If this cpu gets work to do, stop the load balancing
4167 * work being done for other cpus. Next load
4168 * balancing owner will pick it up.
4173 rebalance_domains(balance_cpu, CPU_IDLE);
4175 rq = cpu_rq(balance_cpu);
4176 if (time_after(this_rq->next_balance, rq->next_balance))
4177 this_rq->next_balance = rq->next_balance;
4184 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4186 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4187 * idle load balancing owner or decide to stop the periodic load balancing,
4188 * if the whole system is idle.
4190 static inline void trigger_load_balance(struct rq *rq, int cpu)
4194 * If we were in the nohz mode recently and busy at the current
4195 * scheduler tick, then check if we need to nominate new idle
4198 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4199 rq->in_nohz_recently = 0;
4201 if (atomic_read(&nohz.load_balancer) == cpu) {
4202 cpu_clear(cpu, nohz.cpu_mask);
4203 atomic_set(&nohz.load_balancer, -1);
4206 if (atomic_read(&nohz.load_balancer) == -1) {
4208 * simple selection for now: Nominate the
4209 * first cpu in the nohz list to be the next
4212 * TBD: Traverse the sched domains and nominate
4213 * the nearest cpu in the nohz.cpu_mask.
4215 int ilb = first_cpu(nohz.cpu_mask);
4217 if (ilb < nr_cpu_ids)
4223 * If this cpu is idle and doing idle load balancing for all the
4224 * cpus with ticks stopped, is it time for that to stop?
4226 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4227 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4233 * If this cpu is idle and the idle load balancing is done by
4234 * someone else, then no need raise the SCHED_SOFTIRQ
4236 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4237 cpu_isset(cpu, nohz.cpu_mask))
4240 if (time_after_eq(jiffies, rq->next_balance))
4241 raise_softirq(SCHED_SOFTIRQ);
4244 #else /* CONFIG_SMP */
4247 * on UP we do not need to balance between CPUs:
4249 static inline void idle_balance(int cpu, struct rq *rq)
4255 DEFINE_PER_CPU(struct kernel_stat, kstat);
4257 EXPORT_PER_CPU_SYMBOL(kstat);
4260 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4261 * that have not yet been banked in case the task is currently running.
4263 unsigned long long task_sched_runtime(struct task_struct *p)
4265 unsigned long flags;
4269 rq = task_rq_lock(p, &flags);
4270 ns = p->se.sum_exec_runtime;
4271 if (task_current(rq, p)) {
4272 update_rq_clock(rq);
4273 delta_exec = rq->clock - p->se.exec_start;
4274 if ((s64)delta_exec > 0)
4277 task_rq_unlock(rq, &flags);
4283 * Account user cpu time to a process.
4284 * @p: the process that the cpu time gets accounted to
4285 * @cputime: the cpu time spent in user space since the last update
4287 void account_user_time(struct task_struct *p, cputime_t cputime)
4289 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4292 p->utime = cputime_add(p->utime, cputime);
4294 /* Add user time to cpustat. */
4295 tmp = cputime_to_cputime64(cputime);
4296 if (TASK_NICE(p) > 0)
4297 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4299 cpustat->user = cputime64_add(cpustat->user, tmp);
4303 * Account guest cpu time to a process.
4304 * @p: the process that the cpu time gets accounted to
4305 * @cputime: the cpu time spent in virtual machine since the last update
4307 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4310 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4312 tmp = cputime_to_cputime64(cputime);
4314 p->utime = cputime_add(p->utime, cputime);
4315 p->gtime = cputime_add(p->gtime, cputime);
4317 cpustat->user = cputime64_add(cpustat->user, tmp);
4318 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4322 * Account scaled user cpu time to a process.
4323 * @p: the process that the cpu time gets accounted to
4324 * @cputime: the cpu time spent in user space since the last update
4326 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4328 p->utimescaled = cputime_add(p->utimescaled, cputime);
4332 * Account system cpu time to a process.
4333 * @p: the process that the cpu time gets accounted to
4334 * @hardirq_offset: the offset to subtract from hardirq_count()
4335 * @cputime: the cpu time spent in kernel space since the last update
4337 void account_system_time(struct task_struct *p, int hardirq_offset,
4340 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4341 struct rq *rq = this_rq();
4344 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0))
4345 return account_guest_time(p, cputime);
4347 p->stime = cputime_add(p->stime, cputime);
4349 /* Add system time to cpustat. */
4350 tmp = cputime_to_cputime64(cputime);
4351 if (hardirq_count() - hardirq_offset)
4352 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4353 else if (softirq_count())
4354 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4355 else if (p != rq->idle)
4356 cpustat->system = cputime64_add(cpustat->system, tmp);
4357 else if (atomic_read(&rq->nr_iowait) > 0)
4358 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4360 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4361 /* Account for system time used */
4362 acct_update_integrals(p);
4366 * Account scaled system cpu time to a process.
4367 * @p: the process that the cpu time gets accounted to
4368 * @hardirq_offset: the offset to subtract from hardirq_count()
4369 * @cputime: the cpu time spent in kernel space since the last update
4371 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4373 p->stimescaled = cputime_add(p->stimescaled, cputime);
4377 * Account for involuntary wait time.
4378 * @p: the process from which the cpu time has been stolen
4379 * @steal: the cpu time spent in involuntary wait
4381 void account_steal_time(struct task_struct *p, cputime_t steal)
4383 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4384 cputime64_t tmp = cputime_to_cputime64(steal);
4385 struct rq *rq = this_rq();
4387 if (p == rq->idle) {
4388 p->stime = cputime_add(p->stime, steal);
4389 if (atomic_read(&rq->nr_iowait) > 0)
4390 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4392 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4394 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4398 * This function gets called by the timer code, with HZ frequency.
4399 * We call it with interrupts disabled.
4401 * It also gets called by the fork code, when changing the parent's
4404 void scheduler_tick(void)
4406 int cpu = smp_processor_id();
4407 struct rq *rq = cpu_rq(cpu);
4408 struct task_struct *curr = rq->curr;
4409 u64 next_tick = rq->tick_timestamp + TICK_NSEC;
4411 spin_lock(&rq->lock);
4412 __update_rq_clock(rq);
4414 * Let rq->clock advance by at least TICK_NSEC:
4416 if (unlikely(rq->clock < next_tick)) {
4417 rq->clock = next_tick;
4418 rq->clock_underflows++;
4420 rq->tick_timestamp = rq->clock;
4421 update_last_tick_seen(rq);
4422 update_cpu_load(rq);
4423 curr->sched_class->task_tick(rq, curr, 0);
4424 spin_unlock(&rq->lock);
4427 rq->idle_at_tick = idle_cpu(cpu);
4428 trigger_load_balance(rq, cpu);
4432 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4434 void __kprobes add_preempt_count(int val)
4439 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4441 preempt_count() += val;
4443 * Spinlock count overflowing soon?
4445 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4448 EXPORT_SYMBOL(add_preempt_count);
4450 void __kprobes sub_preempt_count(int val)
4455 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4458 * Is the spinlock portion underflowing?
4460 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4461 !(preempt_count() & PREEMPT_MASK)))
4464 preempt_count() -= val;
4466 EXPORT_SYMBOL(sub_preempt_count);
4471 * Print scheduling while atomic bug:
4473 static noinline void __schedule_bug(struct task_struct *prev)
4475 struct pt_regs *regs = get_irq_regs();
4477 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4478 prev->comm, prev->pid, preempt_count());
4480 debug_show_held_locks(prev);
4481 if (irqs_disabled())
4482 print_irqtrace_events(prev);
4491 * Various schedule()-time debugging checks and statistics:
4493 static inline void schedule_debug(struct task_struct *prev)
4496 * Test if we are atomic. Since do_exit() needs to call into
4497 * schedule() atomically, we ignore that path for now.
4498 * Otherwise, whine if we are scheduling when we should not be.
4500 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4501 __schedule_bug(prev);
4503 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4505 schedstat_inc(this_rq(), sched_count);
4506 #ifdef CONFIG_SCHEDSTATS
4507 if (unlikely(prev->lock_depth >= 0)) {
4508 schedstat_inc(this_rq(), bkl_count);
4509 schedstat_inc(prev, sched_info.bkl_count);
4515 * Pick up the highest-prio task:
4517 static inline struct task_struct *
4518 pick_next_task(struct rq *rq, struct task_struct *prev)
4520 const struct sched_class *class;
4521 struct task_struct *p;
4524 * Optimization: we know that if all tasks are in
4525 * the fair class we can call that function directly:
4527 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4528 p = fair_sched_class.pick_next_task(rq);
4533 class = sched_class_highest;
4535 p = class->pick_next_task(rq);
4539 * Will never be NULL as the idle class always
4540 * returns a non-NULL p:
4542 class = class->next;
4547 * schedule() is the main scheduler function.
4549 asmlinkage void __sched schedule(void)
4551 struct task_struct *prev, *next;
4552 unsigned long *switch_count;
4558 cpu = smp_processor_id();
4562 switch_count = &prev->nivcsw;
4564 release_kernel_lock(prev);
4565 need_resched_nonpreemptible:
4567 schedule_debug(prev);
4572 * Do the rq-clock update outside the rq lock:
4574 local_irq_disable();
4575 __update_rq_clock(rq);
4576 spin_lock(&rq->lock);
4577 clear_tsk_need_resched(prev);
4579 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4580 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4581 signal_pending(prev))) {
4582 prev->state = TASK_RUNNING;
4584 deactivate_task(rq, prev, 1);
4586 switch_count = &prev->nvcsw;
4590 if (prev->sched_class->pre_schedule)
4591 prev->sched_class->pre_schedule(rq, prev);
4594 if (unlikely(!rq->nr_running))
4595 idle_balance(cpu, rq);
4597 prev->sched_class->put_prev_task(rq, prev);
4598 next = pick_next_task(rq, prev);
4600 sched_info_switch(prev, next);
4602 if (likely(prev != next)) {
4607 context_switch(rq, prev, next); /* unlocks the rq */
4609 * the context switch might have flipped the stack from under
4610 * us, hence refresh the local variables.
4612 cpu = smp_processor_id();
4615 spin_unlock_irq(&rq->lock);
4619 if (unlikely(reacquire_kernel_lock(current) < 0))
4620 goto need_resched_nonpreemptible;
4622 preempt_enable_no_resched();
4623 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4626 EXPORT_SYMBOL(schedule);
4628 #ifdef CONFIG_PREEMPT
4630 * this is the entry point to schedule() from in-kernel preemption
4631 * off of preempt_enable. Kernel preemptions off return from interrupt
4632 * occur there and call schedule directly.
4634 asmlinkage void __sched preempt_schedule(void)
4636 struct thread_info *ti = current_thread_info();
4637 struct task_struct *task = current;
4638 int saved_lock_depth;
4641 * If there is a non-zero preempt_count or interrupts are disabled,
4642 * we do not want to preempt the current task. Just return..
4644 if (likely(ti->preempt_count || irqs_disabled()))
4648 add_preempt_count(PREEMPT_ACTIVE);
4651 * We keep the big kernel semaphore locked, but we
4652 * clear ->lock_depth so that schedule() doesnt
4653 * auto-release the semaphore:
4655 saved_lock_depth = task->lock_depth;
4656 task->lock_depth = -1;
4658 task->lock_depth = saved_lock_depth;
4659 sub_preempt_count(PREEMPT_ACTIVE);
4662 * Check again in case we missed a preemption opportunity
4663 * between schedule and now.
4666 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4668 EXPORT_SYMBOL(preempt_schedule);
4671 * this is the entry point to schedule() from kernel preemption
4672 * off of irq context.
4673 * Note, that this is called and return with irqs disabled. This will
4674 * protect us against recursive calling from irq.
4676 asmlinkage void __sched preempt_schedule_irq(void)
4678 struct thread_info *ti = current_thread_info();
4679 struct task_struct *task = current;
4680 int saved_lock_depth;
4682 /* Catch callers which need to be fixed */
4683 BUG_ON(ti->preempt_count || !irqs_disabled());
4686 add_preempt_count(PREEMPT_ACTIVE);
4689 * We keep the big kernel semaphore locked, but we
4690 * clear ->lock_depth so that schedule() doesnt
4691 * auto-release the semaphore:
4693 saved_lock_depth = task->lock_depth;
4694 task->lock_depth = -1;
4697 local_irq_disable();
4698 task->lock_depth = saved_lock_depth;
4699 sub_preempt_count(PREEMPT_ACTIVE);
4702 * Check again in case we missed a preemption opportunity
4703 * between schedule and now.
4706 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4709 #endif /* CONFIG_PREEMPT */
4711 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4714 return try_to_wake_up(curr->private, mode, sync);
4716 EXPORT_SYMBOL(default_wake_function);
4719 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4720 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4721 * number) then we wake all the non-exclusive tasks and one exclusive task.
4723 * There are circumstances in which we can try to wake a task which has already
4724 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4725 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4727 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4728 int nr_exclusive, int sync, void *key)
4730 wait_queue_t *curr, *next;
4732 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4733 unsigned flags = curr->flags;
4735 if (curr->func(curr, mode, sync, key) &&
4736 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4742 * __wake_up - wake up threads blocked on a waitqueue.
4744 * @mode: which threads
4745 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4746 * @key: is directly passed to the wakeup function
4748 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4749 int nr_exclusive, void *key)
4751 unsigned long flags;
4753 spin_lock_irqsave(&q->lock, flags);
4754 __wake_up_common(q, mode, nr_exclusive, 0, key);
4755 spin_unlock_irqrestore(&q->lock, flags);
4757 EXPORT_SYMBOL(__wake_up);
4760 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4762 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4764 __wake_up_common(q, mode, 1, 0, NULL);
4768 * __wake_up_sync - wake up threads blocked on a waitqueue.
4770 * @mode: which threads
4771 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4773 * The sync wakeup differs that the waker knows that it will schedule
4774 * away soon, so while the target thread will be woken up, it will not
4775 * be migrated to another CPU - ie. the two threads are 'synchronized'
4776 * with each other. This can prevent needless bouncing between CPUs.
4778 * On UP it can prevent extra preemption.
4781 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4783 unsigned long flags;
4789 if (unlikely(!nr_exclusive))
4792 spin_lock_irqsave(&q->lock, flags);
4793 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4794 spin_unlock_irqrestore(&q->lock, flags);
4796 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4798 void complete(struct completion *x)
4800 unsigned long flags;
4802 spin_lock_irqsave(&x->wait.lock, flags);
4804 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4805 spin_unlock_irqrestore(&x->wait.lock, flags);
4807 EXPORT_SYMBOL(complete);
4809 void complete_all(struct completion *x)
4811 unsigned long flags;
4813 spin_lock_irqsave(&x->wait.lock, flags);
4814 x->done += UINT_MAX/2;
4815 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4816 spin_unlock_irqrestore(&x->wait.lock, flags);
4818 EXPORT_SYMBOL(complete_all);
4820 static inline long __sched
4821 do_wait_for_common(struct completion *x, long timeout, int state)
4824 DECLARE_WAITQUEUE(wait, current);
4826 wait.flags |= WQ_FLAG_EXCLUSIVE;
4827 __add_wait_queue_tail(&x->wait, &wait);
4829 if ((state == TASK_INTERRUPTIBLE &&
4830 signal_pending(current)) ||
4831 (state == TASK_KILLABLE &&
4832 fatal_signal_pending(current))) {
4833 __remove_wait_queue(&x->wait, &wait);
4834 return -ERESTARTSYS;
4836 __set_current_state(state);
4837 spin_unlock_irq(&x->wait.lock);
4838 timeout = schedule_timeout(timeout);
4839 spin_lock_irq(&x->wait.lock);
4841 __remove_wait_queue(&x->wait, &wait);
4845 __remove_wait_queue(&x->wait, &wait);
4852 wait_for_common(struct completion *x, long timeout, int state)
4856 spin_lock_irq(&x->wait.lock);
4857 timeout = do_wait_for_common(x, timeout, state);
4858 spin_unlock_irq(&x->wait.lock);
4862 void __sched wait_for_completion(struct completion *x)
4864 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4866 EXPORT_SYMBOL(wait_for_completion);
4868 unsigned long __sched
4869 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4871 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4873 EXPORT_SYMBOL(wait_for_completion_timeout);
4875 int __sched wait_for_completion_interruptible(struct completion *x)
4877 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4878 if (t == -ERESTARTSYS)
4882 EXPORT_SYMBOL(wait_for_completion_interruptible);
4884 unsigned long __sched
4885 wait_for_completion_interruptible_timeout(struct completion *x,
4886 unsigned long timeout)
4888 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4890 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4892 int __sched wait_for_completion_killable(struct completion *x)
4894 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4895 if (t == -ERESTARTSYS)
4899 EXPORT_SYMBOL(wait_for_completion_killable);
4902 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4904 unsigned long flags;
4907 init_waitqueue_entry(&wait, current);
4909 __set_current_state(state);
4911 spin_lock_irqsave(&q->lock, flags);
4912 __add_wait_queue(q, &wait);
4913 spin_unlock(&q->lock);
4914 timeout = schedule_timeout(timeout);
4915 spin_lock_irq(&q->lock);
4916 __remove_wait_queue(q, &wait);
4917 spin_unlock_irqrestore(&q->lock, flags);
4922 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4924 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4926 EXPORT_SYMBOL(interruptible_sleep_on);
4929 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4931 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4933 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4935 void __sched sleep_on(wait_queue_head_t *q)
4937 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4939 EXPORT_SYMBOL(sleep_on);
4941 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4943 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4945 EXPORT_SYMBOL(sleep_on_timeout);
4947 #ifdef CONFIG_RT_MUTEXES
4950 * rt_mutex_setprio - set the current priority of a task
4952 * @prio: prio value (kernel-internal form)
4954 * This function changes the 'effective' priority of a task. It does
4955 * not touch ->normal_prio like __setscheduler().
4957 * Used by the rt_mutex code to implement priority inheritance logic.
4959 void rt_mutex_setprio(struct task_struct *p, int prio)
4961 unsigned long flags;
4962 int oldprio, on_rq, running;
4964 const struct sched_class *prev_class = p->sched_class;
4966 BUG_ON(prio < 0 || prio > MAX_PRIO);
4968 rq = task_rq_lock(p, &flags);
4969 update_rq_clock(rq);
4972 on_rq = p->se.on_rq;
4973 running = task_current(rq, p);
4975 dequeue_task(rq, p, 0);
4977 p->sched_class->put_prev_task(rq, p);
4980 p->sched_class = &rt_sched_class;
4982 p->sched_class = &fair_sched_class;
4987 p->sched_class->set_curr_task(rq);
4989 enqueue_task(rq, p, 0);
4991 check_class_changed(rq, p, prev_class, oldprio, running);
4993 task_rq_unlock(rq, &flags);
4998 void set_user_nice(struct task_struct *p, long nice)
5000 int old_prio, delta, on_rq;
5001 unsigned long flags;
5004 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5007 * We have to be careful, if called from sys_setpriority(),
5008 * the task might be in the middle of scheduling on another CPU.
5010 rq = task_rq_lock(p, &flags);
5011 update_rq_clock(rq);
5013 * The RT priorities are set via sched_setscheduler(), but we still
5014 * allow the 'normal' nice value to be set - but as expected
5015 * it wont have any effect on scheduling until the task is
5016 * SCHED_FIFO/SCHED_RR:
5018 if (task_has_rt_policy(p)) {
5019 p->static_prio = NICE_TO_PRIO(nice);
5022 on_rq = p->se.on_rq;
5024 dequeue_task(rq, p, 0);
5026 p->static_prio = NICE_TO_PRIO(nice);
5029 p->prio = effective_prio(p);
5030 delta = p->prio - old_prio;
5033 enqueue_task(rq, p, 0);
5035 * If the task increased its priority or is running and
5036 * lowered its priority, then reschedule its CPU:
5038 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5039 resched_task(rq->curr);
5042 task_rq_unlock(rq, &flags);
5044 EXPORT_SYMBOL(set_user_nice);
5047 * can_nice - check if a task can reduce its nice value
5051 int can_nice(const struct task_struct *p, const int nice)
5053 /* convert nice value [19,-20] to rlimit style value [1,40] */
5054 int nice_rlim = 20 - nice;
5056 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5057 capable(CAP_SYS_NICE));
5060 #ifdef __ARCH_WANT_SYS_NICE
5063 * sys_nice - change the priority of the current process.
5064 * @increment: priority increment
5066 * sys_setpriority is a more generic, but much slower function that
5067 * does similar things.
5069 asmlinkage long sys_nice(int increment)
5074 * Setpriority might change our priority at the same moment.
5075 * We don't have to worry. Conceptually one call occurs first
5076 * and we have a single winner.
5078 if (increment < -40)
5083 nice = PRIO_TO_NICE(current->static_prio) + increment;
5089 if (increment < 0 && !can_nice(current, nice))
5092 retval = security_task_setnice(current, nice);
5096 set_user_nice(current, nice);
5103 * task_prio - return the priority value of a given task.
5104 * @p: the task in question.
5106 * This is the priority value as seen by users in /proc.
5107 * RT tasks are offset by -200. Normal tasks are centered
5108 * around 0, value goes from -16 to +15.
5110 int task_prio(const struct task_struct *p)
5112 return p->prio - MAX_RT_PRIO;
5116 * task_nice - return the nice value of a given task.
5117 * @p: the task in question.
5119 int task_nice(const struct task_struct *p)
5121 return TASK_NICE(p);
5123 EXPORT_SYMBOL(task_nice);
5126 * idle_cpu - is a given cpu idle currently?
5127 * @cpu: the processor in question.
5129 int idle_cpu(int cpu)
5131 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5135 * idle_task - return the idle task for a given cpu.
5136 * @cpu: the processor in question.
5138 struct task_struct *idle_task(int cpu)
5140 return cpu_rq(cpu)->idle;
5144 * find_process_by_pid - find a process with a matching PID value.
5145 * @pid: the pid in question.
5147 static struct task_struct *find_process_by_pid(pid_t pid)
5149 return pid ? find_task_by_vpid(pid) : current;
5152 /* Actually do priority change: must hold rq lock. */
5154 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5156 BUG_ON(p->se.on_rq);
5159 switch (p->policy) {
5163 p->sched_class = &fair_sched_class;
5167 p->sched_class = &rt_sched_class;
5171 p->rt_priority = prio;
5172 p->normal_prio = normal_prio(p);
5173 /* we are holding p->pi_lock already */
5174 p->prio = rt_mutex_getprio(p);
5179 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5180 * @p: the task in question.
5181 * @policy: new policy.
5182 * @param: structure containing the new RT priority.
5184 * NOTE that the task may be already dead.
5186 int sched_setscheduler(struct task_struct *p, int policy,
5187 struct sched_param *param)
5189 int retval, oldprio, oldpolicy = -1, on_rq, running;
5190 unsigned long flags;
5191 const struct sched_class *prev_class = p->sched_class;
5194 /* may grab non-irq protected spin_locks */
5195 BUG_ON(in_interrupt());
5197 /* double check policy once rq lock held */
5199 policy = oldpolicy = p->policy;
5200 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5201 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5202 policy != SCHED_IDLE)
5205 * Valid priorities for SCHED_FIFO and SCHED_RR are
5206 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5207 * SCHED_BATCH and SCHED_IDLE is 0.
5209 if (param->sched_priority < 0 ||
5210 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5211 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5213 if (rt_policy(policy) != (param->sched_priority != 0))
5217 * Allow unprivileged RT tasks to decrease priority:
5219 if (!capable(CAP_SYS_NICE)) {
5220 if (rt_policy(policy)) {
5221 unsigned long rlim_rtprio;
5223 if (!lock_task_sighand(p, &flags))
5225 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5226 unlock_task_sighand(p, &flags);
5228 /* can't set/change the rt policy */
5229 if (policy != p->policy && !rlim_rtprio)
5232 /* can't increase priority */
5233 if (param->sched_priority > p->rt_priority &&
5234 param->sched_priority > rlim_rtprio)
5238 * Like positive nice levels, dont allow tasks to
5239 * move out of SCHED_IDLE either:
5241 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5244 /* can't change other user's priorities */
5245 if ((current->euid != p->euid) &&
5246 (current->euid != p->uid))
5250 #ifdef CONFIG_RT_GROUP_SCHED
5252 * Do not allow realtime tasks into groups that have no runtime
5255 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5259 retval = security_task_setscheduler(p, policy, param);
5263 * make sure no PI-waiters arrive (or leave) while we are
5264 * changing the priority of the task:
5266 spin_lock_irqsave(&p->pi_lock, flags);
5268 * To be able to change p->policy safely, the apropriate
5269 * runqueue lock must be held.
5271 rq = __task_rq_lock(p);
5272 /* recheck policy now with rq lock held */
5273 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5274 policy = oldpolicy = -1;
5275 __task_rq_unlock(rq);
5276 spin_unlock_irqrestore(&p->pi_lock, flags);
5279 update_rq_clock(rq);
5280 on_rq = p->se.on_rq;
5281 running = task_current(rq, p);
5283 deactivate_task(rq, p, 0);
5285 p->sched_class->put_prev_task(rq, p);
5288 __setscheduler(rq, p, policy, param->sched_priority);
5291 p->sched_class->set_curr_task(rq);
5293 activate_task(rq, p, 0);
5295 check_class_changed(rq, p, prev_class, oldprio, running);
5297 __task_rq_unlock(rq);
5298 spin_unlock_irqrestore(&p->pi_lock, flags);
5300 rt_mutex_adjust_pi(p);
5304 EXPORT_SYMBOL_GPL(sched_setscheduler);
5307 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5309 struct sched_param lparam;
5310 struct task_struct *p;
5313 if (!param || pid < 0)
5315 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5320 p = find_process_by_pid(pid);
5322 retval = sched_setscheduler(p, policy, &lparam);
5329 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5330 * @pid: the pid in question.
5331 * @policy: new policy.
5332 * @param: structure containing the new RT priority.
5335 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5337 /* negative values for policy are not valid */
5341 return do_sched_setscheduler(pid, policy, param);
5345 * sys_sched_setparam - set/change the RT priority of a thread
5346 * @pid: the pid in question.
5347 * @param: structure containing the new RT priority.
5349 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5351 return do_sched_setscheduler(pid, -1, param);
5355 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5356 * @pid: the pid in question.
5358 asmlinkage long sys_sched_getscheduler(pid_t pid)
5360 struct task_struct *p;
5367 read_lock(&tasklist_lock);
5368 p = find_process_by_pid(pid);
5370 retval = security_task_getscheduler(p);
5374 read_unlock(&tasklist_lock);
5379 * sys_sched_getscheduler - get the RT priority of a thread
5380 * @pid: the pid in question.
5381 * @param: structure containing the RT priority.
5383 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5385 struct sched_param lp;
5386 struct task_struct *p;
5389 if (!param || pid < 0)
5392 read_lock(&tasklist_lock);
5393 p = find_process_by_pid(pid);
5398 retval = security_task_getscheduler(p);
5402 lp.sched_priority = p->rt_priority;
5403 read_unlock(&tasklist_lock);
5406 * This one might sleep, we cannot do it with a spinlock held ...
5408 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5413 read_unlock(&tasklist_lock);
5417 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5419 cpumask_t cpus_allowed;
5420 cpumask_t new_mask = *in_mask;
5421 struct task_struct *p;
5425 read_lock(&tasklist_lock);
5427 p = find_process_by_pid(pid);
5429 read_unlock(&tasklist_lock);
5435 * It is not safe to call set_cpus_allowed with the
5436 * tasklist_lock held. We will bump the task_struct's
5437 * usage count and then drop tasklist_lock.
5440 read_unlock(&tasklist_lock);
5443 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5444 !capable(CAP_SYS_NICE))
5447 retval = security_task_setscheduler(p, 0, NULL);
5451 cpuset_cpus_allowed(p, &cpus_allowed);
5452 cpus_and(new_mask, new_mask, cpus_allowed);
5454 retval = set_cpus_allowed_ptr(p, &new_mask);
5457 cpuset_cpus_allowed(p, &cpus_allowed);
5458 if (!cpus_subset(new_mask, cpus_allowed)) {
5460 * We must have raced with a concurrent cpuset
5461 * update. Just reset the cpus_allowed to the
5462 * cpuset's cpus_allowed
5464 new_mask = cpus_allowed;
5474 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5475 cpumask_t *new_mask)
5477 if (len < sizeof(cpumask_t)) {
5478 memset(new_mask, 0, sizeof(cpumask_t));
5479 } else if (len > sizeof(cpumask_t)) {
5480 len = sizeof(cpumask_t);
5482 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5486 * sys_sched_setaffinity - set the cpu affinity of a process
5487 * @pid: pid of the process
5488 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5489 * @user_mask_ptr: user-space pointer to the new cpu mask
5491 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5492 unsigned long __user *user_mask_ptr)
5497 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5501 return sched_setaffinity(pid, &new_mask);
5505 * Represents all cpu's present in the system
5506 * In systems capable of hotplug, this map could dynamically grow
5507 * as new cpu's are detected in the system via any platform specific
5508 * method, such as ACPI for e.g.
5511 cpumask_t cpu_present_map __read_mostly;
5512 EXPORT_SYMBOL(cpu_present_map);
5515 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5516 EXPORT_SYMBOL(cpu_online_map);
5518 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5519 EXPORT_SYMBOL(cpu_possible_map);
5522 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5524 struct task_struct *p;
5528 read_lock(&tasklist_lock);
5531 p = find_process_by_pid(pid);
5535 retval = security_task_getscheduler(p);
5539 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5542 read_unlock(&tasklist_lock);
5549 * sys_sched_getaffinity - get the cpu affinity of a process
5550 * @pid: pid of the process
5551 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5552 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5554 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5555 unsigned long __user *user_mask_ptr)
5560 if (len < sizeof(cpumask_t))
5563 ret = sched_getaffinity(pid, &mask);
5567 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5570 return sizeof(cpumask_t);
5574 * sys_sched_yield - yield the current processor to other threads.
5576 * This function yields the current CPU to other tasks. If there are no
5577 * other threads running on this CPU then this function will return.
5579 asmlinkage long sys_sched_yield(void)
5581 struct rq *rq = this_rq_lock();
5583 schedstat_inc(rq, yld_count);
5584 current->sched_class->yield_task(rq);
5587 * Since we are going to call schedule() anyway, there's
5588 * no need to preempt or enable interrupts:
5590 __release(rq->lock);
5591 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5592 _raw_spin_unlock(&rq->lock);
5593 preempt_enable_no_resched();
5600 static void __cond_resched(void)
5602 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5603 __might_sleep(__FILE__, __LINE__);
5606 * The BKS might be reacquired before we have dropped
5607 * PREEMPT_ACTIVE, which could trigger a second
5608 * cond_resched() call.
5611 add_preempt_count(PREEMPT_ACTIVE);
5613 sub_preempt_count(PREEMPT_ACTIVE);
5614 } while (need_resched());
5617 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5618 int __sched _cond_resched(void)
5620 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5621 system_state == SYSTEM_RUNNING) {
5627 EXPORT_SYMBOL(_cond_resched);
5631 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5632 * call schedule, and on return reacquire the lock.
5634 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5635 * operations here to prevent schedule() from being called twice (once via
5636 * spin_unlock(), once by hand).
5638 int cond_resched_lock(spinlock_t *lock)
5640 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5643 if (spin_needbreak(lock) || resched) {
5645 if (resched && need_resched())
5654 EXPORT_SYMBOL(cond_resched_lock);
5656 int __sched cond_resched_softirq(void)
5658 BUG_ON(!in_softirq());
5660 if (need_resched() && system_state == SYSTEM_RUNNING) {
5668 EXPORT_SYMBOL(cond_resched_softirq);
5671 * yield - yield the current processor to other threads.
5673 * This is a shortcut for kernel-space yielding - it marks the
5674 * thread runnable and calls sys_sched_yield().
5676 void __sched yield(void)
5678 set_current_state(TASK_RUNNING);
5681 EXPORT_SYMBOL(yield);
5684 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5685 * that process accounting knows that this is a task in IO wait state.
5687 * But don't do that if it is a deliberate, throttling IO wait (this task
5688 * has set its backing_dev_info: the queue against which it should throttle)
5690 void __sched io_schedule(void)
5692 struct rq *rq = &__raw_get_cpu_var(runqueues);
5694 delayacct_blkio_start();
5695 atomic_inc(&rq->nr_iowait);
5697 atomic_dec(&rq->nr_iowait);
5698 delayacct_blkio_end();
5700 EXPORT_SYMBOL(io_schedule);
5702 long __sched io_schedule_timeout(long timeout)
5704 struct rq *rq = &__raw_get_cpu_var(runqueues);
5707 delayacct_blkio_start();
5708 atomic_inc(&rq->nr_iowait);
5709 ret = schedule_timeout(timeout);
5710 atomic_dec(&rq->nr_iowait);
5711 delayacct_blkio_end();
5716 * sys_sched_get_priority_max - return maximum RT priority.
5717 * @policy: scheduling class.
5719 * this syscall returns the maximum rt_priority that can be used
5720 * by a given scheduling class.
5722 asmlinkage long sys_sched_get_priority_max(int policy)
5729 ret = MAX_USER_RT_PRIO-1;
5741 * sys_sched_get_priority_min - return minimum RT priority.
5742 * @policy: scheduling class.
5744 * this syscall returns the minimum rt_priority that can be used
5745 * by a given scheduling class.
5747 asmlinkage long sys_sched_get_priority_min(int policy)
5765 * sys_sched_rr_get_interval - return the default timeslice of a process.
5766 * @pid: pid of the process.
5767 * @interval: userspace pointer to the timeslice value.
5769 * this syscall writes the default timeslice value of a given process
5770 * into the user-space timespec buffer. A value of '0' means infinity.
5773 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5775 struct task_struct *p;
5776 unsigned int time_slice;
5784 read_lock(&tasklist_lock);
5785 p = find_process_by_pid(pid);
5789 retval = security_task_getscheduler(p);
5794 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5795 * tasks that are on an otherwise idle runqueue:
5798 if (p->policy == SCHED_RR) {
5799 time_slice = DEF_TIMESLICE;
5800 } else if (p->policy != SCHED_FIFO) {
5801 struct sched_entity *se = &p->se;
5802 unsigned long flags;
5805 rq = task_rq_lock(p, &flags);
5806 if (rq->cfs.load.weight)
5807 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5808 task_rq_unlock(rq, &flags);
5810 read_unlock(&tasklist_lock);
5811 jiffies_to_timespec(time_slice, &t);
5812 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5816 read_unlock(&tasklist_lock);
5820 static const char stat_nam[] = "RSDTtZX";
5822 void sched_show_task(struct task_struct *p)
5824 unsigned long free = 0;
5827 state = p->state ? __ffs(p->state) + 1 : 0;
5828 printk(KERN_INFO "%-13.13s %c", p->comm,
5829 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5830 #if BITS_PER_LONG == 32
5831 if (state == TASK_RUNNING)
5832 printk(KERN_CONT " running ");
5834 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5836 if (state == TASK_RUNNING)
5837 printk(KERN_CONT " running task ");
5839 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5841 #ifdef CONFIG_DEBUG_STACK_USAGE
5843 unsigned long *n = end_of_stack(p);
5846 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5849 printk(KERN_CONT "%5lu %5d %6d\n", free,
5850 task_pid_nr(p), task_pid_nr(p->real_parent));
5852 show_stack(p, NULL);
5855 void show_state_filter(unsigned long state_filter)
5857 struct task_struct *g, *p;
5859 #if BITS_PER_LONG == 32
5861 " task PC stack pid father\n");
5864 " task PC stack pid father\n");
5866 read_lock(&tasklist_lock);
5867 do_each_thread(g, p) {
5869 * reset the NMI-timeout, listing all files on a slow
5870 * console might take alot of time:
5872 touch_nmi_watchdog();
5873 if (!state_filter || (p->state & state_filter))
5875 } while_each_thread(g, p);
5877 touch_all_softlockup_watchdogs();
5879 #ifdef CONFIG_SCHED_DEBUG
5880 sysrq_sched_debug_show();
5882 read_unlock(&tasklist_lock);
5884 * Only show locks if all tasks are dumped:
5886 if (state_filter == -1)
5887 debug_show_all_locks();
5890 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5892 idle->sched_class = &idle_sched_class;
5896 * init_idle - set up an idle thread for a given CPU
5897 * @idle: task in question
5898 * @cpu: cpu the idle task belongs to
5900 * NOTE: this function does not set the idle thread's NEED_RESCHED
5901 * flag, to make booting more robust.
5903 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5905 struct rq *rq = cpu_rq(cpu);
5906 unsigned long flags;
5909 idle->se.exec_start = sched_clock();
5911 idle->prio = idle->normal_prio = MAX_PRIO;
5912 idle->cpus_allowed = cpumask_of_cpu(cpu);
5913 __set_task_cpu(idle, cpu);
5915 spin_lock_irqsave(&rq->lock, flags);
5916 rq->curr = rq->idle = idle;
5917 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5920 spin_unlock_irqrestore(&rq->lock, flags);
5922 /* Set the preempt count _outside_ the spinlocks! */
5923 task_thread_info(idle)->preempt_count = 0;
5926 * The idle tasks have their own, simple scheduling class:
5928 idle->sched_class = &idle_sched_class;
5932 * In a system that switches off the HZ timer nohz_cpu_mask
5933 * indicates which cpus entered this state. This is used
5934 * in the rcu update to wait only for active cpus. For system
5935 * which do not switch off the HZ timer nohz_cpu_mask should
5936 * always be CPU_MASK_NONE.
5938 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5941 * Increase the granularity value when there are more CPUs,
5942 * because with more CPUs the 'effective latency' as visible
5943 * to users decreases. But the relationship is not linear,
5944 * so pick a second-best guess by going with the log2 of the
5947 * This idea comes from the SD scheduler of Con Kolivas:
5949 static inline void sched_init_granularity(void)
5951 unsigned int factor = 1 + ilog2(num_online_cpus());
5952 const unsigned long limit = 200000000;
5954 sysctl_sched_min_granularity *= factor;
5955 if (sysctl_sched_min_granularity > limit)
5956 sysctl_sched_min_granularity = limit;
5958 sysctl_sched_latency *= factor;
5959 if (sysctl_sched_latency > limit)
5960 sysctl_sched_latency = limit;
5962 sysctl_sched_wakeup_granularity *= factor;
5967 * This is how migration works:
5969 * 1) we queue a struct migration_req structure in the source CPU's
5970 * runqueue and wake up that CPU's migration thread.
5971 * 2) we down() the locked semaphore => thread blocks.
5972 * 3) migration thread wakes up (implicitly it forces the migrated
5973 * thread off the CPU)
5974 * 4) it gets the migration request and checks whether the migrated
5975 * task is still in the wrong runqueue.
5976 * 5) if it's in the wrong runqueue then the migration thread removes
5977 * it and puts it into the right queue.
5978 * 6) migration thread up()s the semaphore.
5979 * 7) we wake up and the migration is done.
5983 * Change a given task's CPU affinity. Migrate the thread to a
5984 * proper CPU and schedule it away if the CPU it's executing on
5985 * is removed from the allowed bitmask.
5987 * NOTE: the caller must have a valid reference to the task, the
5988 * task must not exit() & deallocate itself prematurely. The
5989 * call is not atomic; no spinlocks may be held.
5991 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5993 struct migration_req req;
5994 unsigned long flags;
5998 rq = task_rq_lock(p, &flags);
5999 if (!cpus_intersects(*new_mask, cpu_online_map)) {
6004 if (p->sched_class->set_cpus_allowed)
6005 p->sched_class->set_cpus_allowed(p, new_mask);
6007 p->cpus_allowed = *new_mask;
6008 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
6011 /* Can the task run on the task's current CPU? If so, we're done */
6012 if (cpu_isset(task_cpu(p), *new_mask))
6015 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6016 /* Need help from migration thread: drop lock and wait. */
6017 task_rq_unlock(rq, &flags);
6018 wake_up_process(rq->migration_thread);
6019 wait_for_completion(&req.done);
6020 tlb_migrate_finish(p->mm);
6024 task_rq_unlock(rq, &flags);
6028 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6031 * Move (not current) task off this cpu, onto dest cpu. We're doing
6032 * this because either it can't run here any more (set_cpus_allowed()
6033 * away from this CPU, or CPU going down), or because we're
6034 * attempting to rebalance this task on exec (sched_exec).
6036 * So we race with normal scheduler movements, but that's OK, as long
6037 * as the task is no longer on this CPU.
6039 * Returns non-zero if task was successfully migrated.
6041 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6043 struct rq *rq_dest, *rq_src;
6046 if (unlikely(cpu_is_offline(dest_cpu)))
6049 rq_src = cpu_rq(src_cpu);
6050 rq_dest = cpu_rq(dest_cpu);
6052 double_rq_lock(rq_src, rq_dest);
6053 /* Already moved. */
6054 if (task_cpu(p) != src_cpu)
6056 /* Affinity changed (again). */
6057 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6060 on_rq = p->se.on_rq;
6062 deactivate_task(rq_src, p, 0);
6064 set_task_cpu(p, dest_cpu);
6066 activate_task(rq_dest, p, 0);
6067 check_preempt_curr(rq_dest, p);
6071 double_rq_unlock(rq_src, rq_dest);
6076 * migration_thread - this is a highprio system thread that performs
6077 * thread migration by bumping thread off CPU then 'pushing' onto
6080 static int migration_thread(void *data)
6082 int cpu = (long)data;
6086 BUG_ON(rq->migration_thread != current);
6088 set_current_state(TASK_INTERRUPTIBLE);
6089 while (!kthread_should_stop()) {
6090 struct migration_req *req;
6091 struct list_head *head;
6093 spin_lock_irq(&rq->lock);
6095 if (cpu_is_offline(cpu)) {
6096 spin_unlock_irq(&rq->lock);
6100 if (rq->active_balance) {
6101 active_load_balance(rq, cpu);
6102 rq->active_balance = 0;
6105 head = &rq->migration_queue;
6107 if (list_empty(head)) {
6108 spin_unlock_irq(&rq->lock);
6110 set_current_state(TASK_INTERRUPTIBLE);
6113 req = list_entry(head->next, struct migration_req, list);
6114 list_del_init(head->next);
6116 spin_unlock(&rq->lock);
6117 __migrate_task(req->task, cpu, req->dest_cpu);
6120 complete(&req->done);
6122 __set_current_state(TASK_RUNNING);
6126 /* Wait for kthread_stop */
6127 set_current_state(TASK_INTERRUPTIBLE);
6128 while (!kthread_should_stop()) {
6130 set_current_state(TASK_INTERRUPTIBLE);
6132 __set_current_state(TASK_RUNNING);
6136 #ifdef CONFIG_HOTPLUG_CPU
6138 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6142 local_irq_disable();
6143 ret = __migrate_task(p, src_cpu, dest_cpu);
6149 * Figure out where task on dead CPU should go, use force if necessary.
6150 * NOTE: interrupts should be disabled by the caller
6152 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6154 unsigned long flags;
6161 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6162 cpus_and(mask, mask, p->cpus_allowed);
6163 dest_cpu = any_online_cpu(mask);
6165 /* On any allowed CPU? */
6166 if (dest_cpu >= nr_cpu_ids)
6167 dest_cpu = any_online_cpu(p->cpus_allowed);
6169 /* No more Mr. Nice Guy. */
6170 if (dest_cpu >= nr_cpu_ids) {
6171 cpumask_t cpus_allowed;
6173 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6175 * Try to stay on the same cpuset, where the
6176 * current cpuset may be a subset of all cpus.
6177 * The cpuset_cpus_allowed_locked() variant of
6178 * cpuset_cpus_allowed() will not block. It must be
6179 * called within calls to cpuset_lock/cpuset_unlock.
6181 rq = task_rq_lock(p, &flags);
6182 p->cpus_allowed = cpus_allowed;
6183 dest_cpu = any_online_cpu(p->cpus_allowed);
6184 task_rq_unlock(rq, &flags);
6187 * Don't tell them about moving exiting tasks or
6188 * kernel threads (both mm NULL), since they never
6191 if (p->mm && printk_ratelimit()) {
6192 printk(KERN_INFO "process %d (%s) no "
6193 "longer affine to cpu%d\n",
6194 task_pid_nr(p), p->comm, dead_cpu);
6197 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6201 * While a dead CPU has no uninterruptible tasks queued at this point,
6202 * it might still have a nonzero ->nr_uninterruptible counter, because
6203 * for performance reasons the counter is not stricly tracking tasks to
6204 * their home CPUs. So we just add the counter to another CPU's counter,
6205 * to keep the global sum constant after CPU-down:
6207 static void migrate_nr_uninterruptible(struct rq *rq_src)
6209 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6210 unsigned long flags;
6212 local_irq_save(flags);
6213 double_rq_lock(rq_src, rq_dest);
6214 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6215 rq_src->nr_uninterruptible = 0;
6216 double_rq_unlock(rq_src, rq_dest);
6217 local_irq_restore(flags);
6220 /* Run through task list and migrate tasks from the dead cpu. */
6221 static void migrate_live_tasks(int src_cpu)
6223 struct task_struct *p, *t;
6225 read_lock(&tasklist_lock);
6227 do_each_thread(t, p) {
6231 if (task_cpu(p) == src_cpu)
6232 move_task_off_dead_cpu(src_cpu, p);
6233 } while_each_thread(t, p);
6235 read_unlock(&tasklist_lock);
6239 * Schedules idle task to be the next runnable task on current CPU.
6240 * It does so by boosting its priority to highest possible.
6241 * Used by CPU offline code.
6243 void sched_idle_next(void)
6245 int this_cpu = smp_processor_id();
6246 struct rq *rq = cpu_rq(this_cpu);
6247 struct task_struct *p = rq->idle;
6248 unsigned long flags;
6250 /* cpu has to be offline */
6251 BUG_ON(cpu_online(this_cpu));
6254 * Strictly not necessary since rest of the CPUs are stopped by now
6255 * and interrupts disabled on the current cpu.
6257 spin_lock_irqsave(&rq->lock, flags);
6259 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6261 update_rq_clock(rq);
6262 activate_task(rq, p, 0);
6264 spin_unlock_irqrestore(&rq->lock, flags);
6268 * Ensures that the idle task is using init_mm right before its cpu goes
6271 void idle_task_exit(void)
6273 struct mm_struct *mm = current->active_mm;
6275 BUG_ON(cpu_online(smp_processor_id()));
6278 switch_mm(mm, &init_mm, current);
6282 /* called under rq->lock with disabled interrupts */
6283 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6285 struct rq *rq = cpu_rq(dead_cpu);
6287 /* Must be exiting, otherwise would be on tasklist. */
6288 BUG_ON(!p->exit_state);
6290 /* Cannot have done final schedule yet: would have vanished. */
6291 BUG_ON(p->state == TASK_DEAD);
6296 * Drop lock around migration; if someone else moves it,
6297 * that's OK. No task can be added to this CPU, so iteration is
6300 spin_unlock_irq(&rq->lock);
6301 move_task_off_dead_cpu(dead_cpu, p);
6302 spin_lock_irq(&rq->lock);
6307 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6308 static void migrate_dead_tasks(unsigned int dead_cpu)
6310 struct rq *rq = cpu_rq(dead_cpu);
6311 struct task_struct *next;
6314 if (!rq->nr_running)
6316 update_rq_clock(rq);
6317 next = pick_next_task(rq, rq->curr);
6320 migrate_dead(dead_cpu, next);
6324 #endif /* CONFIG_HOTPLUG_CPU */
6326 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6328 static struct ctl_table sd_ctl_dir[] = {
6330 .procname = "sched_domain",
6336 static struct ctl_table sd_ctl_root[] = {
6338 .ctl_name = CTL_KERN,
6339 .procname = "kernel",
6341 .child = sd_ctl_dir,
6346 static struct ctl_table *sd_alloc_ctl_entry(int n)
6348 struct ctl_table *entry =
6349 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6354 static void sd_free_ctl_entry(struct ctl_table **tablep)
6356 struct ctl_table *entry;
6359 * In the intermediate directories, both the child directory and
6360 * procname are dynamically allocated and could fail but the mode
6361 * will always be set. In the lowest directory the names are
6362 * static strings and all have proc handlers.
6364 for (entry = *tablep; entry->mode; entry++) {
6366 sd_free_ctl_entry(&entry->child);
6367 if (entry->proc_handler == NULL)
6368 kfree(entry->procname);
6376 set_table_entry(struct ctl_table *entry,
6377 const char *procname, void *data, int maxlen,
6378 mode_t mode, proc_handler *proc_handler)
6380 entry->procname = procname;
6382 entry->maxlen = maxlen;
6384 entry->proc_handler = proc_handler;
6387 static struct ctl_table *
6388 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6390 struct ctl_table *table = sd_alloc_ctl_entry(12);
6395 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6396 sizeof(long), 0644, proc_doulongvec_minmax);
6397 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6398 sizeof(long), 0644, proc_doulongvec_minmax);
6399 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6400 sizeof(int), 0644, proc_dointvec_minmax);
6401 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6402 sizeof(int), 0644, proc_dointvec_minmax);
6403 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6404 sizeof(int), 0644, proc_dointvec_minmax);
6405 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6406 sizeof(int), 0644, proc_dointvec_minmax);
6407 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6408 sizeof(int), 0644, proc_dointvec_minmax);
6409 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6410 sizeof(int), 0644, proc_dointvec_minmax);
6411 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6412 sizeof(int), 0644, proc_dointvec_minmax);
6413 set_table_entry(&table[9], "cache_nice_tries",
6414 &sd->cache_nice_tries,
6415 sizeof(int), 0644, proc_dointvec_minmax);
6416 set_table_entry(&table[10], "flags", &sd->flags,
6417 sizeof(int), 0644, proc_dointvec_minmax);
6418 /* &table[11] is terminator */
6423 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6425 struct ctl_table *entry, *table;
6426 struct sched_domain *sd;
6427 int domain_num = 0, i;
6430 for_each_domain(cpu, sd)
6432 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6437 for_each_domain(cpu, sd) {
6438 snprintf(buf, 32, "domain%d", i);
6439 entry->procname = kstrdup(buf, GFP_KERNEL);
6441 entry->child = sd_alloc_ctl_domain_table(sd);
6448 static struct ctl_table_header *sd_sysctl_header;
6449 static void register_sched_domain_sysctl(void)
6451 int i, cpu_num = num_online_cpus();
6452 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6455 WARN_ON(sd_ctl_dir[0].child);
6456 sd_ctl_dir[0].child = entry;
6461 for_each_online_cpu(i) {
6462 snprintf(buf, 32, "cpu%d", i);
6463 entry->procname = kstrdup(buf, GFP_KERNEL);
6465 entry->child = sd_alloc_ctl_cpu_table(i);
6469 WARN_ON(sd_sysctl_header);
6470 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6473 /* may be called multiple times per register */
6474 static void unregister_sched_domain_sysctl(void)
6476 if (sd_sysctl_header)
6477 unregister_sysctl_table(sd_sysctl_header);
6478 sd_sysctl_header = NULL;
6479 if (sd_ctl_dir[0].child)
6480 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6483 static void register_sched_domain_sysctl(void)
6486 static void unregister_sched_domain_sysctl(void)
6492 * migration_call - callback that gets triggered when a CPU is added.
6493 * Here we can start up the necessary migration thread for the new CPU.
6495 static int __cpuinit
6496 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6498 struct task_struct *p;
6499 int cpu = (long)hcpu;
6500 unsigned long flags;
6505 case CPU_UP_PREPARE:
6506 case CPU_UP_PREPARE_FROZEN:
6507 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6510 kthread_bind(p, cpu);
6511 /* Must be high prio: stop_machine expects to yield to it. */
6512 rq = task_rq_lock(p, &flags);
6513 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6514 task_rq_unlock(rq, &flags);
6515 cpu_rq(cpu)->migration_thread = p;
6519 case CPU_ONLINE_FROZEN:
6520 /* Strictly unnecessary, as first user will wake it. */
6521 wake_up_process(cpu_rq(cpu)->migration_thread);
6523 /* Update our root-domain */
6525 spin_lock_irqsave(&rq->lock, flags);
6527 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6528 cpu_set(cpu, rq->rd->online);
6530 spin_unlock_irqrestore(&rq->lock, flags);
6533 #ifdef CONFIG_HOTPLUG_CPU
6534 case CPU_UP_CANCELED:
6535 case CPU_UP_CANCELED_FROZEN:
6536 if (!cpu_rq(cpu)->migration_thread)
6538 /* Unbind it from offline cpu so it can run. Fall thru. */
6539 kthread_bind(cpu_rq(cpu)->migration_thread,
6540 any_online_cpu(cpu_online_map));
6541 kthread_stop(cpu_rq(cpu)->migration_thread);
6542 cpu_rq(cpu)->migration_thread = NULL;
6546 case CPU_DEAD_FROZEN:
6547 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6548 migrate_live_tasks(cpu);
6550 kthread_stop(rq->migration_thread);
6551 rq->migration_thread = NULL;
6552 /* Idle task back to normal (off runqueue, low prio) */
6553 spin_lock_irq(&rq->lock);
6554 update_rq_clock(rq);
6555 deactivate_task(rq, rq->idle, 0);
6556 rq->idle->static_prio = MAX_PRIO;
6557 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6558 rq->idle->sched_class = &idle_sched_class;
6559 migrate_dead_tasks(cpu);
6560 spin_unlock_irq(&rq->lock);
6562 migrate_nr_uninterruptible(rq);
6563 BUG_ON(rq->nr_running != 0);
6566 * No need to migrate the tasks: it was best-effort if
6567 * they didn't take sched_hotcpu_mutex. Just wake up
6570 spin_lock_irq(&rq->lock);
6571 while (!list_empty(&rq->migration_queue)) {
6572 struct migration_req *req;
6574 req = list_entry(rq->migration_queue.next,
6575 struct migration_req, list);
6576 list_del_init(&req->list);
6577 complete(&req->done);
6579 spin_unlock_irq(&rq->lock);
6583 case CPU_DYING_FROZEN:
6584 /* Update our root-domain */
6586 spin_lock_irqsave(&rq->lock, flags);
6588 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6589 cpu_clear(cpu, rq->rd->online);
6591 spin_unlock_irqrestore(&rq->lock, flags);
6598 /* Register at highest priority so that task migration (migrate_all_tasks)
6599 * happens before everything else.
6601 static struct notifier_block __cpuinitdata migration_notifier = {
6602 .notifier_call = migration_call,
6606 void __init migration_init(void)
6608 void *cpu = (void *)(long)smp_processor_id();
6611 /* Start one for the boot CPU: */
6612 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6613 BUG_ON(err == NOTIFY_BAD);
6614 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6615 register_cpu_notifier(&migration_notifier);
6621 #ifdef CONFIG_SCHED_DEBUG
6623 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6624 cpumask_t *groupmask)
6626 struct sched_group *group = sd->groups;
6629 cpulist_scnprintf(str, sizeof(str), sd->span);
6630 cpus_clear(*groupmask);
6632 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6634 if (!(sd->flags & SD_LOAD_BALANCE)) {
6635 printk("does not load-balance\n");
6637 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6642 printk(KERN_CONT "span %s\n", str);
6644 if (!cpu_isset(cpu, sd->span)) {
6645 printk(KERN_ERR "ERROR: domain->span does not contain "
6648 if (!cpu_isset(cpu, group->cpumask)) {
6649 printk(KERN_ERR "ERROR: domain->groups does not contain"
6653 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6657 printk(KERN_ERR "ERROR: group is NULL\n");
6661 if (!group->__cpu_power) {
6662 printk(KERN_CONT "\n");
6663 printk(KERN_ERR "ERROR: domain->cpu_power not "
6668 if (!cpus_weight(group->cpumask)) {
6669 printk(KERN_CONT "\n");
6670 printk(KERN_ERR "ERROR: empty group\n");
6674 if (cpus_intersects(*groupmask, group->cpumask)) {
6675 printk(KERN_CONT "\n");
6676 printk(KERN_ERR "ERROR: repeated CPUs\n");
6680 cpus_or(*groupmask, *groupmask, group->cpumask);
6682 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6683 printk(KERN_CONT " %s", str);
6685 group = group->next;
6686 } while (group != sd->groups);
6687 printk(KERN_CONT "\n");
6689 if (!cpus_equal(sd->span, *groupmask))
6690 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6692 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6693 printk(KERN_ERR "ERROR: parent span is not a superset "
6694 "of domain->span\n");
6698 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6700 cpumask_t *groupmask;
6704 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6708 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6710 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6712 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6717 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6727 # define sched_domain_debug(sd, cpu) do { } while (0)
6730 static int sd_degenerate(struct sched_domain *sd)
6732 if (cpus_weight(sd->span) == 1)
6735 /* Following flags need at least 2 groups */
6736 if (sd->flags & (SD_LOAD_BALANCE |
6737 SD_BALANCE_NEWIDLE |
6741 SD_SHARE_PKG_RESOURCES)) {
6742 if (sd->groups != sd->groups->next)
6746 /* Following flags don't use groups */
6747 if (sd->flags & (SD_WAKE_IDLE |
6756 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6758 unsigned long cflags = sd->flags, pflags = parent->flags;
6760 if (sd_degenerate(parent))
6763 if (!cpus_equal(sd->span, parent->span))
6766 /* Does parent contain flags not in child? */
6767 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6768 if (cflags & SD_WAKE_AFFINE)
6769 pflags &= ~SD_WAKE_BALANCE;
6770 /* Flags needing groups don't count if only 1 group in parent */
6771 if (parent->groups == parent->groups->next) {
6772 pflags &= ~(SD_LOAD_BALANCE |
6773 SD_BALANCE_NEWIDLE |
6777 SD_SHARE_PKG_RESOURCES);
6779 if (~cflags & pflags)
6785 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6787 unsigned long flags;
6788 const struct sched_class *class;
6790 spin_lock_irqsave(&rq->lock, flags);
6793 struct root_domain *old_rd = rq->rd;
6795 for (class = sched_class_highest; class; class = class->next) {
6796 if (class->leave_domain)
6797 class->leave_domain(rq);
6800 cpu_clear(rq->cpu, old_rd->span);
6801 cpu_clear(rq->cpu, old_rd->online);
6803 if (atomic_dec_and_test(&old_rd->refcount))
6807 atomic_inc(&rd->refcount);
6810 cpu_set(rq->cpu, rd->span);
6811 if (cpu_isset(rq->cpu, cpu_online_map))
6812 cpu_set(rq->cpu, rd->online);
6814 for (class = sched_class_highest; class; class = class->next) {
6815 if (class->join_domain)
6816 class->join_domain(rq);
6819 spin_unlock_irqrestore(&rq->lock, flags);
6822 static void init_rootdomain(struct root_domain *rd)
6824 memset(rd, 0, sizeof(*rd));
6826 cpus_clear(rd->span);
6827 cpus_clear(rd->online);
6830 static void init_defrootdomain(void)
6832 init_rootdomain(&def_root_domain);
6833 atomic_set(&def_root_domain.refcount, 1);
6836 static struct root_domain *alloc_rootdomain(void)
6838 struct root_domain *rd;
6840 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6844 init_rootdomain(rd);
6850 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6851 * hold the hotplug lock.
6854 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6856 struct rq *rq = cpu_rq(cpu);
6857 struct sched_domain *tmp;
6859 /* Remove the sched domains which do not contribute to scheduling. */
6860 for (tmp = sd; tmp; tmp = tmp->parent) {
6861 struct sched_domain *parent = tmp->parent;
6864 if (sd_parent_degenerate(tmp, parent)) {
6865 tmp->parent = parent->parent;
6867 parent->parent->child = tmp;
6871 if (sd && sd_degenerate(sd)) {
6877 sched_domain_debug(sd, cpu);
6879 rq_attach_root(rq, rd);
6880 rcu_assign_pointer(rq->sd, sd);
6883 /* cpus with isolated domains */
6884 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6886 /* Setup the mask of cpus configured for isolated domains */
6887 static int __init isolated_cpu_setup(char *str)
6889 int ints[NR_CPUS], i;
6891 str = get_options(str, ARRAY_SIZE(ints), ints);
6892 cpus_clear(cpu_isolated_map);
6893 for (i = 1; i <= ints[0]; i++)
6894 if (ints[i] < NR_CPUS)
6895 cpu_set(ints[i], cpu_isolated_map);
6899 __setup("isolcpus=", isolated_cpu_setup);
6902 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6903 * to a function which identifies what group(along with sched group) a CPU
6904 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6905 * (due to the fact that we keep track of groups covered with a cpumask_t).
6907 * init_sched_build_groups will build a circular linked list of the groups
6908 * covered by the given span, and will set each group's ->cpumask correctly,
6909 * and ->cpu_power to 0.
6912 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6913 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6914 struct sched_group **sg,
6915 cpumask_t *tmpmask),
6916 cpumask_t *covered, cpumask_t *tmpmask)
6918 struct sched_group *first = NULL, *last = NULL;
6921 cpus_clear(*covered);
6923 for_each_cpu_mask(i, *span) {
6924 struct sched_group *sg;
6925 int group = group_fn(i, cpu_map, &sg, tmpmask);
6928 if (cpu_isset(i, *covered))
6931 cpus_clear(sg->cpumask);
6932 sg->__cpu_power = 0;
6934 for_each_cpu_mask(j, *span) {
6935 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6938 cpu_set(j, *covered);
6939 cpu_set(j, sg->cpumask);
6950 #define SD_NODES_PER_DOMAIN 16
6955 * find_next_best_node - find the next node to include in a sched_domain
6956 * @node: node whose sched_domain we're building
6957 * @used_nodes: nodes already in the sched_domain
6959 * Find the next node to include in a given scheduling domain. Simply
6960 * finds the closest node not already in the @used_nodes map.
6962 * Should use nodemask_t.
6964 static int find_next_best_node(int node, nodemask_t *used_nodes)
6966 int i, n, val, min_val, best_node = 0;
6970 for (i = 0; i < MAX_NUMNODES; i++) {
6971 /* Start at @node */
6972 n = (node + i) % MAX_NUMNODES;
6974 if (!nr_cpus_node(n))
6977 /* Skip already used nodes */
6978 if (node_isset(n, *used_nodes))
6981 /* Simple min distance search */
6982 val = node_distance(node, n);
6984 if (val < min_val) {
6990 node_set(best_node, *used_nodes);
6995 * sched_domain_node_span - get a cpumask for a node's sched_domain
6996 * @node: node whose cpumask we're constructing
6997 * @span: resulting cpumask
6999 * Given a node, construct a good cpumask for its sched_domain to span. It
7000 * should be one that prevents unnecessary balancing, but also spreads tasks
7003 static void sched_domain_node_span(int node, cpumask_t *span)
7005 nodemask_t used_nodes;
7006 node_to_cpumask_ptr(nodemask, node);
7010 nodes_clear(used_nodes);
7012 cpus_or(*span, *span, *nodemask);
7013 node_set(node, used_nodes);
7015 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7016 int next_node = find_next_best_node(node, &used_nodes);
7018 node_to_cpumask_ptr_next(nodemask, next_node);
7019 cpus_or(*span, *span, *nodemask);
7024 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7027 * SMT sched-domains:
7029 #ifdef CONFIG_SCHED_SMT
7030 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7031 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7034 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7038 *sg = &per_cpu(sched_group_cpus, cpu);
7044 * multi-core sched-domains:
7046 #ifdef CONFIG_SCHED_MC
7047 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7048 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7051 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7053 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7058 *mask = per_cpu(cpu_sibling_map, cpu);
7059 cpus_and(*mask, *mask, *cpu_map);
7060 group = first_cpu(*mask);
7062 *sg = &per_cpu(sched_group_core, group);
7065 #elif defined(CONFIG_SCHED_MC)
7067 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7071 *sg = &per_cpu(sched_group_core, cpu);
7076 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7077 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7080 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7084 #ifdef CONFIG_SCHED_MC
7085 *mask = cpu_coregroup_map(cpu);
7086 cpus_and(*mask, *mask, *cpu_map);
7087 group = first_cpu(*mask);
7088 #elif defined(CONFIG_SCHED_SMT)
7089 *mask = per_cpu(cpu_sibling_map, cpu);
7090 cpus_and(*mask, *mask, *cpu_map);
7091 group = first_cpu(*mask);
7096 *sg = &per_cpu(sched_group_phys, group);
7102 * The init_sched_build_groups can't handle what we want to do with node
7103 * groups, so roll our own. Now each node has its own list of groups which
7104 * gets dynamically allocated.
7106 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7107 static struct sched_group ***sched_group_nodes_bycpu;
7109 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7110 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7112 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7113 struct sched_group **sg, cpumask_t *nodemask)
7117 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7118 cpus_and(*nodemask, *nodemask, *cpu_map);
7119 group = first_cpu(*nodemask);
7122 *sg = &per_cpu(sched_group_allnodes, group);
7126 static void init_numa_sched_groups_power(struct sched_group *group_head)
7128 struct sched_group *sg = group_head;
7134 for_each_cpu_mask(j, sg->cpumask) {
7135 struct sched_domain *sd;
7137 sd = &per_cpu(phys_domains, j);
7138 if (j != first_cpu(sd->groups->cpumask)) {
7140 * Only add "power" once for each
7146 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7149 } while (sg != group_head);
7154 /* Free memory allocated for various sched_group structures */
7155 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7159 for_each_cpu_mask(cpu, *cpu_map) {
7160 struct sched_group **sched_group_nodes
7161 = sched_group_nodes_bycpu[cpu];
7163 if (!sched_group_nodes)
7166 for (i = 0; i < MAX_NUMNODES; i++) {
7167 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7169 *nodemask = node_to_cpumask(i);
7170 cpus_and(*nodemask, *nodemask, *cpu_map);
7171 if (cpus_empty(*nodemask))
7181 if (oldsg != sched_group_nodes[i])
7184 kfree(sched_group_nodes);
7185 sched_group_nodes_bycpu[cpu] = NULL;
7189 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7195 * Initialize sched groups cpu_power.
7197 * cpu_power indicates the capacity of sched group, which is used while
7198 * distributing the load between different sched groups in a sched domain.
7199 * Typically cpu_power for all the groups in a sched domain will be same unless
7200 * there are asymmetries in the topology. If there are asymmetries, group
7201 * having more cpu_power will pickup more load compared to the group having
7204 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7205 * the maximum number of tasks a group can handle in the presence of other idle
7206 * or lightly loaded groups in the same sched domain.
7208 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7210 struct sched_domain *child;
7211 struct sched_group *group;
7213 WARN_ON(!sd || !sd->groups);
7215 if (cpu != first_cpu(sd->groups->cpumask))
7220 sd->groups->__cpu_power = 0;
7223 * For perf policy, if the groups in child domain share resources
7224 * (for example cores sharing some portions of the cache hierarchy
7225 * or SMT), then set this domain groups cpu_power such that each group
7226 * can handle only one task, when there are other idle groups in the
7227 * same sched domain.
7229 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7231 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7232 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7237 * add cpu_power of each child group to this groups cpu_power
7239 group = child->groups;
7241 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7242 group = group->next;
7243 } while (group != child->groups);
7247 * Initializers for schedule domains
7248 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7251 #define SD_INIT(sd, type) sd_init_##type(sd)
7252 #define SD_INIT_FUNC(type) \
7253 static noinline void sd_init_##type(struct sched_domain *sd) \
7255 memset(sd, 0, sizeof(*sd)); \
7256 *sd = SD_##type##_INIT; \
7257 sd->level = SD_LV_##type; \
7262 SD_INIT_FUNC(ALLNODES)
7265 #ifdef CONFIG_SCHED_SMT
7266 SD_INIT_FUNC(SIBLING)
7268 #ifdef CONFIG_SCHED_MC
7273 * To minimize stack usage kmalloc room for cpumasks and share the
7274 * space as the usage in build_sched_domains() dictates. Used only
7275 * if the amount of space is significant.
7278 cpumask_t tmpmask; /* make this one first */
7281 cpumask_t this_sibling_map;
7282 cpumask_t this_core_map;
7284 cpumask_t send_covered;
7287 cpumask_t domainspan;
7289 cpumask_t notcovered;
7294 #define SCHED_CPUMASK_ALLOC 1
7295 #define SCHED_CPUMASK_FREE(v) kfree(v)
7296 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7298 #define SCHED_CPUMASK_ALLOC 0
7299 #define SCHED_CPUMASK_FREE(v)
7300 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7303 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7304 ((unsigned long)(a) + offsetof(struct allmasks, v))
7306 static int default_relax_domain_level = -1;
7308 static int __init setup_relax_domain_level(char *str)
7310 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7313 __setup("relax_domain_level=", setup_relax_domain_level);
7315 static void set_domain_attribute(struct sched_domain *sd,
7316 struct sched_domain_attr *attr)
7320 if (!attr || attr->relax_domain_level < 0) {
7321 if (default_relax_domain_level < 0)
7324 request = default_relax_domain_level;
7326 request = attr->relax_domain_level;
7327 if (request < sd->level) {
7328 /* turn off idle balance on this domain */
7329 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7331 /* turn on idle balance on this domain */
7332 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7337 * Build sched domains for a given set of cpus and attach the sched domains
7338 * to the individual cpus
7340 static int __build_sched_domains(const cpumask_t *cpu_map,
7341 struct sched_domain_attr *attr)
7344 struct root_domain *rd;
7345 SCHED_CPUMASK_DECLARE(allmasks);
7348 struct sched_group **sched_group_nodes = NULL;
7349 int sd_allnodes = 0;
7352 * Allocate the per-node list of sched groups
7354 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7356 if (!sched_group_nodes) {
7357 printk(KERN_WARNING "Can not alloc sched group node list\n");
7362 rd = alloc_rootdomain();
7364 printk(KERN_WARNING "Cannot alloc root domain\n");
7366 kfree(sched_group_nodes);
7371 #if SCHED_CPUMASK_ALLOC
7372 /* get space for all scratch cpumask variables */
7373 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7375 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7378 kfree(sched_group_nodes);
7383 tmpmask = (cpumask_t *)allmasks;
7387 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7391 * Set up domains for cpus specified by the cpu_map.
7393 for_each_cpu_mask(i, *cpu_map) {
7394 struct sched_domain *sd = NULL, *p;
7395 SCHED_CPUMASK_VAR(nodemask, allmasks);
7397 *nodemask = node_to_cpumask(cpu_to_node(i));
7398 cpus_and(*nodemask, *nodemask, *cpu_map);
7401 if (cpus_weight(*cpu_map) >
7402 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7403 sd = &per_cpu(allnodes_domains, i);
7404 SD_INIT(sd, ALLNODES);
7405 set_domain_attribute(sd, attr);
7406 sd->span = *cpu_map;
7407 sd->first_cpu = first_cpu(sd->span);
7408 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7414 sd = &per_cpu(node_domains, i);
7416 set_domain_attribute(sd, attr);
7417 sched_domain_node_span(cpu_to_node(i), &sd->span);
7418 sd->first_cpu = first_cpu(sd->span);
7422 cpus_and(sd->span, sd->span, *cpu_map);
7426 sd = &per_cpu(phys_domains, i);
7428 set_domain_attribute(sd, attr);
7429 sd->span = *nodemask;
7430 sd->first_cpu = first_cpu(sd->span);
7434 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7436 #ifdef CONFIG_SCHED_MC
7438 sd = &per_cpu(core_domains, i);
7440 set_domain_attribute(sd, attr);
7441 sd->span = cpu_coregroup_map(i);
7442 sd->first_cpu = first_cpu(sd->span);
7443 cpus_and(sd->span, sd->span, *cpu_map);
7446 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7449 #ifdef CONFIG_SCHED_SMT
7451 sd = &per_cpu(cpu_domains, i);
7452 SD_INIT(sd, SIBLING);
7453 set_domain_attribute(sd, attr);
7454 sd->span = per_cpu(cpu_sibling_map, i);
7455 sd->first_cpu = first_cpu(sd->span);
7456 cpus_and(sd->span, sd->span, *cpu_map);
7459 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7463 #ifdef CONFIG_SCHED_SMT
7464 /* Set up CPU (sibling) groups */
7465 for_each_cpu_mask(i, *cpu_map) {
7466 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7467 SCHED_CPUMASK_VAR(send_covered, allmasks);
7469 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7470 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7471 if (i != first_cpu(*this_sibling_map))
7474 init_sched_build_groups(this_sibling_map, cpu_map,
7476 send_covered, tmpmask);
7480 #ifdef CONFIG_SCHED_MC
7481 /* Set up multi-core groups */
7482 for_each_cpu_mask(i, *cpu_map) {
7483 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7484 SCHED_CPUMASK_VAR(send_covered, allmasks);
7486 *this_core_map = cpu_coregroup_map(i);
7487 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7488 if (i != first_cpu(*this_core_map))
7491 init_sched_build_groups(this_core_map, cpu_map,
7493 send_covered, tmpmask);
7497 /* Set up physical groups */
7498 for (i = 0; i < MAX_NUMNODES; i++) {
7499 SCHED_CPUMASK_VAR(nodemask, allmasks);
7500 SCHED_CPUMASK_VAR(send_covered, allmasks);
7502 *nodemask = node_to_cpumask(i);
7503 cpus_and(*nodemask, *nodemask, *cpu_map);
7504 if (cpus_empty(*nodemask))
7507 init_sched_build_groups(nodemask, cpu_map,
7509 send_covered, tmpmask);
7513 /* Set up node groups */
7515 SCHED_CPUMASK_VAR(send_covered, allmasks);
7517 init_sched_build_groups(cpu_map, cpu_map,
7518 &cpu_to_allnodes_group,
7519 send_covered, tmpmask);
7522 for (i = 0; i < MAX_NUMNODES; i++) {
7523 /* Set up node groups */
7524 struct sched_group *sg, *prev;
7525 SCHED_CPUMASK_VAR(nodemask, allmasks);
7526 SCHED_CPUMASK_VAR(domainspan, allmasks);
7527 SCHED_CPUMASK_VAR(covered, allmasks);
7530 *nodemask = node_to_cpumask(i);
7531 cpus_clear(*covered);
7533 cpus_and(*nodemask, *nodemask, *cpu_map);
7534 if (cpus_empty(*nodemask)) {
7535 sched_group_nodes[i] = NULL;
7539 sched_domain_node_span(i, domainspan);
7540 cpus_and(*domainspan, *domainspan, *cpu_map);
7542 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7544 printk(KERN_WARNING "Can not alloc domain group for "
7548 sched_group_nodes[i] = sg;
7549 for_each_cpu_mask(j, *nodemask) {
7550 struct sched_domain *sd;
7552 sd = &per_cpu(node_domains, j);
7555 sg->__cpu_power = 0;
7556 sg->cpumask = *nodemask;
7558 cpus_or(*covered, *covered, *nodemask);
7561 for (j = 0; j < MAX_NUMNODES; j++) {
7562 SCHED_CPUMASK_VAR(notcovered, allmasks);
7563 int n = (i + j) % MAX_NUMNODES;
7564 node_to_cpumask_ptr(pnodemask, n);
7566 cpus_complement(*notcovered, *covered);
7567 cpus_and(*tmpmask, *notcovered, *cpu_map);
7568 cpus_and(*tmpmask, *tmpmask, *domainspan);
7569 if (cpus_empty(*tmpmask))
7572 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7573 if (cpus_empty(*tmpmask))
7576 sg = kmalloc_node(sizeof(struct sched_group),
7580 "Can not alloc domain group for node %d\n", j);
7583 sg->__cpu_power = 0;
7584 sg->cpumask = *tmpmask;
7585 sg->next = prev->next;
7586 cpus_or(*covered, *covered, *tmpmask);
7593 /* Calculate CPU power for physical packages and nodes */
7594 #ifdef CONFIG_SCHED_SMT
7595 for_each_cpu_mask(i, *cpu_map) {
7596 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7598 init_sched_groups_power(i, sd);
7601 #ifdef CONFIG_SCHED_MC
7602 for_each_cpu_mask(i, *cpu_map) {
7603 struct sched_domain *sd = &per_cpu(core_domains, i);
7605 init_sched_groups_power(i, sd);
7609 for_each_cpu_mask(i, *cpu_map) {
7610 struct sched_domain *sd = &per_cpu(phys_domains, i);
7612 init_sched_groups_power(i, sd);
7616 for (i = 0; i < MAX_NUMNODES; i++)
7617 init_numa_sched_groups_power(sched_group_nodes[i]);
7620 struct sched_group *sg;
7622 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7624 init_numa_sched_groups_power(sg);
7628 /* Attach the domains */
7629 for_each_cpu_mask(i, *cpu_map) {
7630 struct sched_domain *sd;
7631 #ifdef CONFIG_SCHED_SMT
7632 sd = &per_cpu(cpu_domains, i);
7633 #elif defined(CONFIG_SCHED_MC)
7634 sd = &per_cpu(core_domains, i);
7636 sd = &per_cpu(phys_domains, i);
7638 cpu_attach_domain(sd, rd, i);
7641 SCHED_CPUMASK_FREE((void *)allmasks);
7646 free_sched_groups(cpu_map, tmpmask);
7647 SCHED_CPUMASK_FREE((void *)allmasks);
7652 static int build_sched_domains(const cpumask_t *cpu_map)
7654 return __build_sched_domains(cpu_map, NULL);
7657 static cpumask_t *doms_cur; /* current sched domains */
7658 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7659 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7663 * Special case: If a kmalloc of a doms_cur partition (array of
7664 * cpumask_t) fails, then fallback to a single sched domain,
7665 * as determined by the single cpumask_t fallback_doms.
7667 static cpumask_t fallback_doms;
7669 void __attribute__((weak)) arch_update_cpu_topology(void)
7674 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7675 * For now this just excludes isolated cpus, but could be used to
7676 * exclude other special cases in the future.
7678 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7682 arch_update_cpu_topology();
7684 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7686 doms_cur = &fallback_doms;
7687 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7689 err = build_sched_domains(doms_cur);
7690 register_sched_domain_sysctl();
7695 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7698 free_sched_groups(cpu_map, tmpmask);
7702 * Detach sched domains from a group of cpus specified in cpu_map
7703 * These cpus will now be attached to the NULL domain
7705 static void detach_destroy_domains(const cpumask_t *cpu_map)
7710 unregister_sched_domain_sysctl();
7712 for_each_cpu_mask(i, *cpu_map)
7713 cpu_attach_domain(NULL, &def_root_domain, i);
7714 synchronize_sched();
7715 arch_destroy_sched_domains(cpu_map, &tmpmask);
7718 /* handle null as "default" */
7719 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7720 struct sched_domain_attr *new, int idx_new)
7722 struct sched_domain_attr tmp;
7729 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7730 new ? (new + idx_new) : &tmp,
7731 sizeof(struct sched_domain_attr));
7735 * Partition sched domains as specified by the 'ndoms_new'
7736 * cpumasks in the array doms_new[] of cpumasks. This compares
7737 * doms_new[] to the current sched domain partitioning, doms_cur[].
7738 * It destroys each deleted domain and builds each new domain.
7740 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7741 * The masks don't intersect (don't overlap.) We should setup one
7742 * sched domain for each mask. CPUs not in any of the cpumasks will
7743 * not be load balanced. If the same cpumask appears both in the
7744 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7747 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7748 * ownership of it and will kfree it when done with it. If the caller
7749 * failed the kmalloc call, then it can pass in doms_new == NULL,
7750 * and partition_sched_domains() will fallback to the single partition
7753 * Call with hotplug lock held
7755 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7756 struct sched_domain_attr *dattr_new)
7762 /* always unregister in case we don't destroy any domains */
7763 unregister_sched_domain_sysctl();
7765 if (doms_new == NULL) {
7767 doms_new = &fallback_doms;
7768 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7772 /* Destroy deleted domains */
7773 for (i = 0; i < ndoms_cur; i++) {
7774 for (j = 0; j < ndoms_new; j++) {
7775 if (cpus_equal(doms_cur[i], doms_new[j])
7776 && dattrs_equal(dattr_cur, i, dattr_new, j))
7779 /* no match - a current sched domain not in new doms_new[] */
7780 detach_destroy_domains(doms_cur + i);
7785 /* Build new domains */
7786 for (i = 0; i < ndoms_new; i++) {
7787 for (j = 0; j < ndoms_cur; j++) {
7788 if (cpus_equal(doms_new[i], doms_cur[j])
7789 && dattrs_equal(dattr_new, i, dattr_cur, j))
7792 /* no match - add a new doms_new */
7793 __build_sched_domains(doms_new + i,
7794 dattr_new ? dattr_new + i : NULL);
7799 /* Remember the new sched domains */
7800 if (doms_cur != &fallback_doms)
7802 kfree(dattr_cur); /* kfree(NULL) is safe */
7803 doms_cur = doms_new;
7804 dattr_cur = dattr_new;
7805 ndoms_cur = ndoms_new;
7807 register_sched_domain_sysctl();
7812 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7813 int arch_reinit_sched_domains(void)
7818 detach_destroy_domains(&cpu_online_map);
7819 err = arch_init_sched_domains(&cpu_online_map);
7825 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7829 if (buf[0] != '0' && buf[0] != '1')
7833 sched_smt_power_savings = (buf[0] == '1');
7835 sched_mc_power_savings = (buf[0] == '1');
7837 ret = arch_reinit_sched_domains();
7839 return ret ? ret : count;
7842 #ifdef CONFIG_SCHED_MC
7843 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7845 return sprintf(page, "%u\n", sched_mc_power_savings);
7847 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7848 const char *buf, size_t count)
7850 return sched_power_savings_store(buf, count, 0);
7852 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7853 sched_mc_power_savings_store);
7856 #ifdef CONFIG_SCHED_SMT
7857 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7859 return sprintf(page, "%u\n", sched_smt_power_savings);
7861 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7862 const char *buf, size_t count)
7864 return sched_power_savings_store(buf, count, 1);
7866 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7867 sched_smt_power_savings_store);
7870 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7874 #ifdef CONFIG_SCHED_SMT
7876 err = sysfs_create_file(&cls->kset.kobj,
7877 &attr_sched_smt_power_savings.attr);
7879 #ifdef CONFIG_SCHED_MC
7880 if (!err && mc_capable())
7881 err = sysfs_create_file(&cls->kset.kobj,
7882 &attr_sched_mc_power_savings.attr);
7889 * Force a reinitialization of the sched domains hierarchy. The domains
7890 * and groups cannot be updated in place without racing with the balancing
7891 * code, so we temporarily attach all running cpus to the NULL domain
7892 * which will prevent rebalancing while the sched domains are recalculated.
7894 static int update_sched_domains(struct notifier_block *nfb,
7895 unsigned long action, void *hcpu)
7898 case CPU_UP_PREPARE:
7899 case CPU_UP_PREPARE_FROZEN:
7900 case CPU_DOWN_PREPARE:
7901 case CPU_DOWN_PREPARE_FROZEN:
7902 detach_destroy_domains(&cpu_online_map);
7905 case CPU_UP_CANCELED:
7906 case CPU_UP_CANCELED_FROZEN:
7907 case CPU_DOWN_FAILED:
7908 case CPU_DOWN_FAILED_FROZEN:
7910 case CPU_ONLINE_FROZEN:
7912 case CPU_DEAD_FROZEN:
7914 * Fall through and re-initialise the domains.
7921 /* The hotplug lock is already held by cpu_up/cpu_down */
7922 arch_init_sched_domains(&cpu_online_map);
7927 void __init sched_init_smp(void)
7929 cpumask_t non_isolated_cpus;
7931 #if defined(CONFIG_NUMA)
7932 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7934 BUG_ON(sched_group_nodes_bycpu == NULL);
7937 arch_init_sched_domains(&cpu_online_map);
7938 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7939 if (cpus_empty(non_isolated_cpus))
7940 cpu_set(smp_processor_id(), non_isolated_cpus);
7942 /* XXX: Theoretical race here - CPU may be hotplugged now */
7943 hotcpu_notifier(update_sched_domains, 0);
7945 /* Move init over to a non-isolated CPU */
7946 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7948 sched_init_granularity();
7951 void __init sched_init_smp(void)
7953 sched_init_granularity();
7955 #endif /* CONFIG_SMP */
7957 int in_sched_functions(unsigned long addr)
7959 return in_lock_functions(addr) ||
7960 (addr >= (unsigned long)__sched_text_start
7961 && addr < (unsigned long)__sched_text_end);
7964 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7966 cfs_rq->tasks_timeline = RB_ROOT;
7967 INIT_LIST_HEAD(&cfs_rq->tasks);
7968 #ifdef CONFIG_FAIR_GROUP_SCHED
7971 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7974 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7976 struct rt_prio_array *array;
7979 array = &rt_rq->active;
7980 for (i = 0; i < MAX_RT_PRIO; i++) {
7981 INIT_LIST_HEAD(array->queue + i);
7982 __clear_bit(i, array->bitmap);
7984 /* delimiter for bitsearch: */
7985 __set_bit(MAX_RT_PRIO, array->bitmap);
7987 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7988 rt_rq->highest_prio = MAX_RT_PRIO;
7991 rt_rq->rt_nr_migratory = 0;
7992 rt_rq->overloaded = 0;
7996 rt_rq->rt_throttled = 0;
7997 rt_rq->rt_runtime = 0;
7998 spin_lock_init(&rt_rq->rt_runtime_lock);
8000 #ifdef CONFIG_RT_GROUP_SCHED
8001 rt_rq->rt_nr_boosted = 0;
8006 #ifdef CONFIG_FAIR_GROUP_SCHED
8007 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8008 struct sched_entity *se, int cpu, int add,
8009 struct sched_entity *parent)
8011 struct rq *rq = cpu_rq(cpu);
8012 tg->cfs_rq[cpu] = cfs_rq;
8013 init_cfs_rq(cfs_rq, rq);
8016 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8019 /* se could be NULL for init_task_group */
8024 se->cfs_rq = &rq->cfs;
8026 se->cfs_rq = parent->my_q;
8029 se->load.weight = tg->shares;
8030 se->load.inv_weight = 0;
8031 se->parent = parent;
8035 #ifdef CONFIG_RT_GROUP_SCHED
8036 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8037 struct sched_rt_entity *rt_se, int cpu, int add,
8038 struct sched_rt_entity *parent)
8040 struct rq *rq = cpu_rq(cpu);
8042 tg->rt_rq[cpu] = rt_rq;
8043 init_rt_rq(rt_rq, rq);
8045 rt_rq->rt_se = rt_se;
8046 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8048 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8050 tg->rt_se[cpu] = rt_se;
8055 rt_se->rt_rq = &rq->rt;
8057 rt_se->rt_rq = parent->my_q;
8059 rt_se->rt_rq = &rq->rt;
8060 rt_se->my_q = rt_rq;
8061 rt_se->parent = parent;
8062 INIT_LIST_HEAD(&rt_se->run_list);
8066 void __init sched_init(void)
8069 unsigned long alloc_size = 0, ptr;
8071 #ifdef CONFIG_FAIR_GROUP_SCHED
8072 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8074 #ifdef CONFIG_RT_GROUP_SCHED
8075 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8077 #ifdef CONFIG_USER_SCHED
8081 * As sched_init() is called before page_alloc is setup,
8082 * we use alloc_bootmem().
8085 ptr = (unsigned long)alloc_bootmem(alloc_size);
8087 #ifdef CONFIG_FAIR_GROUP_SCHED
8088 init_task_group.se = (struct sched_entity **)ptr;
8089 ptr += nr_cpu_ids * sizeof(void **);
8091 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8092 ptr += nr_cpu_ids * sizeof(void **);
8094 #ifdef CONFIG_USER_SCHED
8095 root_task_group.se = (struct sched_entity **)ptr;
8096 ptr += nr_cpu_ids * sizeof(void **);
8098 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8099 ptr += nr_cpu_ids * sizeof(void **);
8102 #ifdef CONFIG_RT_GROUP_SCHED
8103 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8104 ptr += nr_cpu_ids * sizeof(void **);
8106 init_task_group.rt_rq = (struct rt_rq **)ptr;
8107 ptr += nr_cpu_ids * sizeof(void **);
8109 #ifdef CONFIG_USER_SCHED
8110 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8111 ptr += nr_cpu_ids * sizeof(void **);
8113 root_task_group.rt_rq = (struct rt_rq **)ptr;
8114 ptr += nr_cpu_ids * sizeof(void **);
8121 init_defrootdomain();
8124 init_rt_bandwidth(&def_rt_bandwidth,
8125 global_rt_period(), global_rt_runtime());
8127 #ifdef CONFIG_RT_GROUP_SCHED
8128 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8129 global_rt_period(), global_rt_runtime());
8130 #ifdef CONFIG_USER_SCHED
8131 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8132 global_rt_period(), RUNTIME_INF);
8136 #ifdef CONFIG_GROUP_SCHED
8137 list_add(&init_task_group.list, &task_groups);
8138 INIT_LIST_HEAD(&init_task_group.children);
8140 #ifdef CONFIG_USER_SCHED
8141 INIT_LIST_HEAD(&root_task_group.children);
8142 init_task_group.parent = &root_task_group;
8143 list_add(&init_task_group.siblings, &root_task_group.children);
8147 for_each_possible_cpu(i) {
8151 spin_lock_init(&rq->lock);
8152 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8155 update_last_tick_seen(rq);
8156 init_cfs_rq(&rq->cfs, rq);
8157 init_rt_rq(&rq->rt, rq);
8158 #ifdef CONFIG_FAIR_GROUP_SCHED
8159 init_task_group.shares = init_task_group_load;
8160 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8161 #ifdef CONFIG_CGROUP_SCHED
8163 * How much cpu bandwidth does init_task_group get?
8165 * In case of task-groups formed thr' the cgroup filesystem, it
8166 * gets 100% of the cpu resources in the system. This overall
8167 * system cpu resource is divided among the tasks of
8168 * init_task_group and its child task-groups in a fair manner,
8169 * based on each entity's (task or task-group's) weight
8170 * (se->load.weight).
8172 * In other words, if init_task_group has 10 tasks of weight
8173 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8174 * then A0's share of the cpu resource is:
8176 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8178 * We achieve this by letting init_task_group's tasks sit
8179 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8181 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8182 #elif defined CONFIG_USER_SCHED
8183 root_task_group.shares = NICE_0_LOAD;
8184 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8186 * In case of task-groups formed thr' the user id of tasks,
8187 * init_task_group represents tasks belonging to root user.
8188 * Hence it forms a sibling of all subsequent groups formed.
8189 * In this case, init_task_group gets only a fraction of overall
8190 * system cpu resource, based on the weight assigned to root
8191 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8192 * by letting tasks of init_task_group sit in a separate cfs_rq
8193 * (init_cfs_rq) and having one entity represent this group of
8194 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8196 init_tg_cfs_entry(&init_task_group,
8197 &per_cpu(init_cfs_rq, i),
8198 &per_cpu(init_sched_entity, i), i, 1,
8199 root_task_group.se[i]);
8202 #endif /* CONFIG_FAIR_GROUP_SCHED */
8204 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8205 #ifdef CONFIG_RT_GROUP_SCHED
8206 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8207 #ifdef CONFIG_CGROUP_SCHED
8208 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8209 #elif defined CONFIG_USER_SCHED
8210 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8211 init_tg_rt_entry(&init_task_group,
8212 &per_cpu(init_rt_rq, i),
8213 &per_cpu(init_sched_rt_entity, i), i, 1,
8214 root_task_group.rt_se[i]);
8218 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8219 rq->cpu_load[j] = 0;
8223 rq->active_balance = 0;
8224 rq->next_balance = jiffies;
8227 rq->migration_thread = NULL;
8228 INIT_LIST_HEAD(&rq->migration_queue);
8229 rq_attach_root(rq, &def_root_domain);
8232 atomic_set(&rq->nr_iowait, 0);
8235 set_load_weight(&init_task);
8237 #ifdef CONFIG_PREEMPT_NOTIFIERS
8238 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8242 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8245 #ifdef CONFIG_RT_MUTEXES
8246 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8250 * The boot idle thread does lazy MMU switching as well:
8252 atomic_inc(&init_mm.mm_count);
8253 enter_lazy_tlb(&init_mm, current);
8256 * Make us the idle thread. Technically, schedule() should not be
8257 * called from this thread, however somewhere below it might be,
8258 * but because we are the idle thread, we just pick up running again
8259 * when this runqueue becomes "idle".
8261 init_idle(current, smp_processor_id());
8263 * During early bootup we pretend to be a normal task:
8265 current->sched_class = &fair_sched_class;
8267 scheduler_running = 1;
8270 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8271 void __might_sleep(char *file, int line)
8274 static unsigned long prev_jiffy; /* ratelimiting */
8276 if ((in_atomic() || irqs_disabled()) &&
8277 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8278 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8280 prev_jiffy = jiffies;
8281 printk(KERN_ERR "BUG: sleeping function called from invalid"
8282 " context at %s:%d\n", file, line);
8283 printk("in_atomic():%d, irqs_disabled():%d\n",
8284 in_atomic(), irqs_disabled());
8285 debug_show_held_locks(current);
8286 if (irqs_disabled())
8287 print_irqtrace_events(current);
8292 EXPORT_SYMBOL(__might_sleep);
8295 #ifdef CONFIG_MAGIC_SYSRQ
8296 static void normalize_task(struct rq *rq, struct task_struct *p)
8299 update_rq_clock(rq);
8300 on_rq = p->se.on_rq;
8302 deactivate_task(rq, p, 0);
8303 __setscheduler(rq, p, SCHED_NORMAL, 0);
8305 activate_task(rq, p, 0);
8306 resched_task(rq->curr);
8310 void normalize_rt_tasks(void)
8312 struct task_struct *g, *p;
8313 unsigned long flags;
8316 read_lock_irqsave(&tasklist_lock, flags);
8317 do_each_thread(g, p) {
8319 * Only normalize user tasks:
8324 p->se.exec_start = 0;
8325 #ifdef CONFIG_SCHEDSTATS
8326 p->se.wait_start = 0;
8327 p->se.sleep_start = 0;
8328 p->se.block_start = 0;
8330 task_rq(p)->clock = 0;
8334 * Renice negative nice level userspace
8337 if (TASK_NICE(p) < 0 && p->mm)
8338 set_user_nice(p, 0);
8342 spin_lock(&p->pi_lock);
8343 rq = __task_rq_lock(p);
8345 normalize_task(rq, p);
8347 __task_rq_unlock(rq);
8348 spin_unlock(&p->pi_lock);
8349 } while_each_thread(g, p);
8351 read_unlock_irqrestore(&tasklist_lock, flags);
8354 #endif /* CONFIG_MAGIC_SYSRQ */
8358 * These functions are only useful for the IA64 MCA handling.
8360 * They can only be called when the whole system has been
8361 * stopped - every CPU needs to be quiescent, and no scheduling
8362 * activity can take place. Using them for anything else would
8363 * be a serious bug, and as a result, they aren't even visible
8364 * under any other configuration.
8368 * curr_task - return the current task for a given cpu.
8369 * @cpu: the processor in question.
8371 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8373 struct task_struct *curr_task(int cpu)
8375 return cpu_curr(cpu);
8379 * set_curr_task - set the current task for a given cpu.
8380 * @cpu: the processor in question.
8381 * @p: the task pointer to set.
8383 * Description: This function must only be used when non-maskable interrupts
8384 * are serviced on a separate stack. It allows the architecture to switch the
8385 * notion of the current task on a cpu in a non-blocking manner. This function
8386 * must be called with all CPU's synchronized, and interrupts disabled, the
8387 * and caller must save the original value of the current task (see
8388 * curr_task() above) and restore that value before reenabling interrupts and
8389 * re-starting the system.
8391 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8393 void set_curr_task(int cpu, struct task_struct *p)
8400 #ifdef CONFIG_FAIR_GROUP_SCHED
8401 static void free_fair_sched_group(struct task_group *tg)
8405 for_each_possible_cpu(i) {
8407 kfree(tg->cfs_rq[i]);
8417 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8419 struct cfs_rq *cfs_rq;
8420 struct sched_entity *se, *parent_se;
8424 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8427 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8431 tg->shares = NICE_0_LOAD;
8433 for_each_possible_cpu(i) {
8436 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8437 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8441 se = kmalloc_node(sizeof(struct sched_entity),
8442 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8446 parent_se = parent ? parent->se[i] : NULL;
8447 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8456 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8458 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8459 &cpu_rq(cpu)->leaf_cfs_rq_list);
8462 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8464 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8467 static inline void free_fair_sched_group(struct task_group *tg)
8472 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8477 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8481 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8486 #ifdef CONFIG_RT_GROUP_SCHED
8487 static void free_rt_sched_group(struct task_group *tg)
8491 destroy_rt_bandwidth(&tg->rt_bandwidth);
8493 for_each_possible_cpu(i) {
8495 kfree(tg->rt_rq[i]);
8497 kfree(tg->rt_se[i]);
8505 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8507 struct rt_rq *rt_rq;
8508 struct sched_rt_entity *rt_se, *parent_se;
8512 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8515 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8519 init_rt_bandwidth(&tg->rt_bandwidth,
8520 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8522 for_each_possible_cpu(i) {
8525 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8526 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8530 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8531 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8535 parent_se = parent ? parent->rt_se[i] : NULL;
8536 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8545 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8547 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8548 &cpu_rq(cpu)->leaf_rt_rq_list);
8551 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8553 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8556 static inline void free_rt_sched_group(struct task_group *tg)
8561 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8566 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8570 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8575 #ifdef CONFIG_GROUP_SCHED
8576 static void free_sched_group(struct task_group *tg)
8578 free_fair_sched_group(tg);
8579 free_rt_sched_group(tg);
8583 /* allocate runqueue etc for a new task group */
8584 struct task_group *sched_create_group(struct task_group *parent)
8586 struct task_group *tg;
8587 unsigned long flags;
8590 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8592 return ERR_PTR(-ENOMEM);
8594 if (!alloc_fair_sched_group(tg, parent))
8597 if (!alloc_rt_sched_group(tg, parent))
8600 spin_lock_irqsave(&task_group_lock, flags);
8601 for_each_possible_cpu(i) {
8602 register_fair_sched_group(tg, i);
8603 register_rt_sched_group(tg, i);
8605 list_add_rcu(&tg->list, &task_groups);
8607 WARN_ON(!parent); /* root should already exist */
8609 tg->parent = parent;
8610 list_add_rcu(&tg->siblings, &parent->children);
8611 INIT_LIST_HEAD(&tg->children);
8612 spin_unlock_irqrestore(&task_group_lock, flags);
8617 free_sched_group(tg);
8618 return ERR_PTR(-ENOMEM);
8621 /* rcu callback to free various structures associated with a task group */
8622 static void free_sched_group_rcu(struct rcu_head *rhp)
8624 /* now it should be safe to free those cfs_rqs */
8625 free_sched_group(container_of(rhp, struct task_group, rcu));
8628 /* Destroy runqueue etc associated with a task group */
8629 void sched_destroy_group(struct task_group *tg)
8631 unsigned long flags;
8634 spin_lock_irqsave(&task_group_lock, flags);
8635 for_each_possible_cpu(i) {
8636 unregister_fair_sched_group(tg, i);
8637 unregister_rt_sched_group(tg, i);
8639 list_del_rcu(&tg->list);
8640 list_del_rcu(&tg->siblings);
8641 spin_unlock_irqrestore(&task_group_lock, flags);
8643 /* wait for possible concurrent references to cfs_rqs complete */
8644 call_rcu(&tg->rcu, free_sched_group_rcu);
8647 /* change task's runqueue when it moves between groups.
8648 * The caller of this function should have put the task in its new group
8649 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8650 * reflect its new group.
8652 void sched_move_task(struct task_struct *tsk)
8655 unsigned long flags;
8658 rq = task_rq_lock(tsk, &flags);
8660 update_rq_clock(rq);
8662 running = task_current(rq, tsk);
8663 on_rq = tsk->se.on_rq;
8666 dequeue_task(rq, tsk, 0);
8667 if (unlikely(running))
8668 tsk->sched_class->put_prev_task(rq, tsk);
8670 set_task_rq(tsk, task_cpu(tsk));
8672 #ifdef CONFIG_FAIR_GROUP_SCHED
8673 if (tsk->sched_class->moved_group)
8674 tsk->sched_class->moved_group(tsk);
8677 if (unlikely(running))
8678 tsk->sched_class->set_curr_task(rq);
8680 enqueue_task(rq, tsk, 0);
8682 task_rq_unlock(rq, &flags);
8686 #ifdef CONFIG_FAIR_GROUP_SCHED
8687 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8689 struct cfs_rq *cfs_rq = se->cfs_rq;
8694 dequeue_entity(cfs_rq, se, 0);
8696 se->load.weight = shares;
8697 se->load.inv_weight = 0;
8700 enqueue_entity(cfs_rq, se, 0);
8703 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8705 struct cfs_rq *cfs_rq = se->cfs_rq;
8706 struct rq *rq = cfs_rq->rq;
8707 unsigned long flags;
8709 spin_lock_irqsave(&rq->lock, flags);
8710 __set_se_shares(se, shares);
8711 spin_unlock_irqrestore(&rq->lock, flags);
8714 static DEFINE_MUTEX(shares_mutex);
8716 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8719 unsigned long flags;
8722 * We can't change the weight of the root cgroup.
8728 * A weight of 0 or 1 can cause arithmetics problems.
8729 * (The default weight is 1024 - so there's no practical
8730 * limitation from this.)
8732 if (shares < MIN_SHARES)
8733 shares = MIN_SHARES;
8735 mutex_lock(&shares_mutex);
8736 if (tg->shares == shares)
8739 spin_lock_irqsave(&task_group_lock, flags);
8740 for_each_possible_cpu(i)
8741 unregister_fair_sched_group(tg, i);
8742 list_del_rcu(&tg->siblings);
8743 spin_unlock_irqrestore(&task_group_lock, flags);
8745 /* wait for any ongoing reference to this group to finish */
8746 synchronize_sched();
8749 * Now we are free to modify the group's share on each cpu
8750 * w/o tripping rebalance_share or load_balance_fair.
8752 tg->shares = shares;
8753 for_each_possible_cpu(i) {
8757 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8758 set_se_shares(tg->se[i], shares/nr_cpu_ids);
8762 * Enable load balance activity on this group, by inserting it back on
8763 * each cpu's rq->leaf_cfs_rq_list.
8765 spin_lock_irqsave(&task_group_lock, flags);
8766 for_each_possible_cpu(i)
8767 register_fair_sched_group(tg, i);
8768 list_add_rcu(&tg->siblings, &tg->parent->children);
8769 spin_unlock_irqrestore(&task_group_lock, flags);
8771 mutex_unlock(&shares_mutex);
8775 unsigned long sched_group_shares(struct task_group *tg)
8781 #ifdef CONFIG_RT_GROUP_SCHED
8783 * Ensure that the real time constraints are schedulable.
8785 static DEFINE_MUTEX(rt_constraints_mutex);
8787 static unsigned long to_ratio(u64 period, u64 runtime)
8789 if (runtime == RUNTIME_INF)
8792 return div64_u64(runtime << 16, period);
8795 #ifdef CONFIG_CGROUP_SCHED
8796 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8798 struct task_group *tgi, *parent = tg->parent;
8799 unsigned long total = 0;
8802 if (global_rt_period() < period)
8805 return to_ratio(period, runtime) <
8806 to_ratio(global_rt_period(), global_rt_runtime());
8809 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8813 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8817 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8818 tgi->rt_bandwidth.rt_runtime);
8822 return total + to_ratio(period, runtime) <
8823 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8824 parent->rt_bandwidth.rt_runtime);
8826 #elif defined CONFIG_USER_SCHED
8827 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8829 struct task_group *tgi;
8830 unsigned long total = 0;
8831 unsigned long global_ratio =
8832 to_ratio(global_rt_period(), global_rt_runtime());
8835 list_for_each_entry_rcu(tgi, &task_groups, list) {
8839 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8840 tgi->rt_bandwidth.rt_runtime);
8844 return total + to_ratio(period, runtime) < global_ratio;
8848 /* Must be called with tasklist_lock held */
8849 static inline int tg_has_rt_tasks(struct task_group *tg)
8851 struct task_struct *g, *p;
8852 do_each_thread(g, p) {
8853 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8855 } while_each_thread(g, p);
8859 static int tg_set_bandwidth(struct task_group *tg,
8860 u64 rt_period, u64 rt_runtime)
8864 mutex_lock(&rt_constraints_mutex);
8865 read_lock(&tasklist_lock);
8866 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8870 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8875 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8876 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8877 tg->rt_bandwidth.rt_runtime = rt_runtime;
8879 for_each_possible_cpu(i) {
8880 struct rt_rq *rt_rq = tg->rt_rq[i];
8882 spin_lock(&rt_rq->rt_runtime_lock);
8883 rt_rq->rt_runtime = rt_runtime;
8884 spin_unlock(&rt_rq->rt_runtime_lock);
8886 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8888 read_unlock(&tasklist_lock);
8889 mutex_unlock(&rt_constraints_mutex);
8894 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8896 u64 rt_runtime, rt_period;
8898 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8899 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8900 if (rt_runtime_us < 0)
8901 rt_runtime = RUNTIME_INF;
8903 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8906 long sched_group_rt_runtime(struct task_group *tg)
8910 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8913 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8914 do_div(rt_runtime_us, NSEC_PER_USEC);
8915 return rt_runtime_us;
8918 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8920 u64 rt_runtime, rt_period;
8922 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8923 rt_runtime = tg->rt_bandwidth.rt_runtime;
8925 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8928 long sched_group_rt_period(struct task_group *tg)
8932 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8933 do_div(rt_period_us, NSEC_PER_USEC);
8934 return rt_period_us;
8937 static int sched_rt_global_constraints(void)
8941 mutex_lock(&rt_constraints_mutex);
8942 if (!__rt_schedulable(NULL, 1, 0))
8944 mutex_unlock(&rt_constraints_mutex);
8949 static int sched_rt_global_constraints(void)
8951 unsigned long flags;
8954 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8955 for_each_possible_cpu(i) {
8956 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8958 spin_lock(&rt_rq->rt_runtime_lock);
8959 rt_rq->rt_runtime = global_rt_runtime();
8960 spin_unlock(&rt_rq->rt_runtime_lock);
8962 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8968 int sched_rt_handler(struct ctl_table *table, int write,
8969 struct file *filp, void __user *buffer, size_t *lenp,
8973 int old_period, old_runtime;
8974 static DEFINE_MUTEX(mutex);
8977 old_period = sysctl_sched_rt_period;
8978 old_runtime = sysctl_sched_rt_runtime;
8980 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8982 if (!ret && write) {
8983 ret = sched_rt_global_constraints();
8985 sysctl_sched_rt_period = old_period;
8986 sysctl_sched_rt_runtime = old_runtime;
8988 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8989 def_rt_bandwidth.rt_period =
8990 ns_to_ktime(global_rt_period());
8993 mutex_unlock(&mutex);
8998 #ifdef CONFIG_CGROUP_SCHED
9000 /* return corresponding task_group object of a cgroup */
9001 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9003 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9004 struct task_group, css);
9007 static struct cgroup_subsys_state *
9008 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9010 struct task_group *tg, *parent;
9012 if (!cgrp->parent) {
9013 /* This is early initialization for the top cgroup */
9014 init_task_group.css.cgroup = cgrp;
9015 return &init_task_group.css;
9018 parent = cgroup_tg(cgrp->parent);
9019 tg = sched_create_group(parent);
9021 return ERR_PTR(-ENOMEM);
9023 /* Bind the cgroup to task_group object we just created */
9024 tg->css.cgroup = cgrp;
9030 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9032 struct task_group *tg = cgroup_tg(cgrp);
9034 sched_destroy_group(tg);
9038 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9039 struct task_struct *tsk)
9041 #ifdef CONFIG_RT_GROUP_SCHED
9042 /* Don't accept realtime tasks when there is no way for them to run */
9043 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9046 /* We don't support RT-tasks being in separate groups */
9047 if (tsk->sched_class != &fair_sched_class)
9055 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9056 struct cgroup *old_cont, struct task_struct *tsk)
9058 sched_move_task(tsk);
9061 #ifdef CONFIG_FAIR_GROUP_SCHED
9062 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9065 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9068 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9070 struct task_group *tg = cgroup_tg(cgrp);
9072 return (u64) tg->shares;
9076 #ifdef CONFIG_RT_GROUP_SCHED
9077 static ssize_t cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9080 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9083 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9085 return sched_group_rt_runtime(cgroup_tg(cgrp));
9088 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9091 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9094 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9096 return sched_group_rt_period(cgroup_tg(cgrp));
9100 static struct cftype cpu_files[] = {
9101 #ifdef CONFIG_FAIR_GROUP_SCHED
9104 .read_u64 = cpu_shares_read_u64,
9105 .write_u64 = cpu_shares_write_u64,
9108 #ifdef CONFIG_RT_GROUP_SCHED
9110 .name = "rt_runtime_us",
9111 .read_s64 = cpu_rt_runtime_read,
9112 .write_s64 = cpu_rt_runtime_write,
9115 .name = "rt_period_us",
9116 .read_u64 = cpu_rt_period_read_uint,
9117 .write_u64 = cpu_rt_period_write_uint,
9122 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9124 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9127 struct cgroup_subsys cpu_cgroup_subsys = {
9129 .create = cpu_cgroup_create,
9130 .destroy = cpu_cgroup_destroy,
9131 .can_attach = cpu_cgroup_can_attach,
9132 .attach = cpu_cgroup_attach,
9133 .populate = cpu_cgroup_populate,
9134 .subsys_id = cpu_cgroup_subsys_id,
9138 #endif /* CONFIG_CGROUP_SCHED */
9140 #ifdef CONFIG_CGROUP_CPUACCT
9143 * CPU accounting code for task groups.
9145 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9146 * (balbir@in.ibm.com).
9149 /* track cpu usage of a group of tasks */
9151 struct cgroup_subsys_state css;
9152 /* cpuusage holds pointer to a u64-type object on every cpu */
9156 struct cgroup_subsys cpuacct_subsys;
9158 /* return cpu accounting group corresponding to this container */
9159 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9161 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9162 struct cpuacct, css);
9165 /* return cpu accounting group to which this task belongs */
9166 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9168 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9169 struct cpuacct, css);
9172 /* create a new cpu accounting group */
9173 static struct cgroup_subsys_state *cpuacct_create(
9174 struct cgroup_subsys *ss, struct cgroup *cgrp)
9176 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9179 return ERR_PTR(-ENOMEM);
9181 ca->cpuusage = alloc_percpu(u64);
9182 if (!ca->cpuusage) {
9184 return ERR_PTR(-ENOMEM);
9190 /* destroy an existing cpu accounting group */
9192 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9194 struct cpuacct *ca = cgroup_ca(cgrp);
9196 free_percpu(ca->cpuusage);
9200 /* return total cpu usage (in nanoseconds) of a group */
9201 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9203 struct cpuacct *ca = cgroup_ca(cgrp);
9204 u64 totalcpuusage = 0;
9207 for_each_possible_cpu(i) {
9208 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9211 * Take rq->lock to make 64-bit addition safe on 32-bit
9214 spin_lock_irq(&cpu_rq(i)->lock);
9215 totalcpuusage += *cpuusage;
9216 spin_unlock_irq(&cpu_rq(i)->lock);
9219 return totalcpuusage;
9222 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9225 struct cpuacct *ca = cgroup_ca(cgrp);
9234 for_each_possible_cpu(i) {
9235 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9237 spin_lock_irq(&cpu_rq(i)->lock);
9239 spin_unlock_irq(&cpu_rq(i)->lock);
9245 static struct cftype files[] = {
9248 .read_u64 = cpuusage_read,
9249 .write_u64 = cpuusage_write,
9253 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9255 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9259 * charge this task's execution time to its accounting group.
9261 * called with rq->lock held.
9263 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9267 if (!cpuacct_subsys.active)
9272 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9274 *cpuusage += cputime;
9278 struct cgroup_subsys cpuacct_subsys = {
9280 .create = cpuacct_create,
9281 .destroy = cpuacct_destroy,
9282 .populate = cpuacct_populate,
9283 .subsys_id = cpuacct_subsys_id,
9285 #endif /* CONFIG_CGROUP_CPUACCT */