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/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/reciprocal_div.h>
68 #include <linux/unistd.h>
69 #include <linux/pagemap.h>
70 #include <linux/hrtimer.h>
71 #include <linux/tick.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
125 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
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 || 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;
213 static inline int rt_bandwidth_enabled(void)
215 return sysctl_sched_rt_runtime >= 0;
218 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
222 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 spin_lock(&rt_b->rt_runtime_lock);
233 if (hrtimer_active(&rt_b->rt_period_timer))
236 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
237 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
240 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
241 delta = ktime_to_ns(ktime_sub(hard, soft));
242 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
243 HRTIMER_MODE_ABS_PINNED, 0);
245 spin_unlock(&rt_b->rt_runtime_lock);
248 #ifdef CONFIG_RT_GROUP_SCHED
249 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
251 hrtimer_cancel(&rt_b->rt_period_timer);
256 * sched_domains_mutex serializes calls to arch_init_sched_domains,
257 * detach_destroy_domains and partition_sched_domains.
259 static DEFINE_MUTEX(sched_domains_mutex);
261 #ifdef CONFIG_GROUP_SCHED
263 #include <linux/cgroup.h>
267 static LIST_HEAD(task_groups);
269 /* task group related information */
271 #ifdef CONFIG_CGROUP_SCHED
272 struct cgroup_subsys_state css;
275 #ifdef CONFIG_USER_SCHED
279 #ifdef CONFIG_FAIR_GROUP_SCHED
280 /* schedulable entities of this group on each cpu */
281 struct sched_entity **se;
282 /* runqueue "owned" by this group on each cpu */
283 struct cfs_rq **cfs_rq;
284 unsigned long shares;
287 #ifdef CONFIG_RT_GROUP_SCHED
288 struct sched_rt_entity **rt_se;
289 struct rt_rq **rt_rq;
291 struct rt_bandwidth rt_bandwidth;
295 struct list_head list;
297 struct task_group *parent;
298 struct list_head siblings;
299 struct list_head children;
302 #ifdef CONFIG_USER_SCHED
304 /* Helper function to pass uid information to create_sched_user() */
305 void set_tg_uid(struct user_struct *user)
307 user->tg->uid = user->uid;
312 * Every UID task group (including init_task_group aka UID-0) will
313 * be a child to this group.
315 struct task_group root_task_group;
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 /* Default task group's sched entity on each cpu */
319 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
320 /* Default task group's cfs_rq on each cpu */
321 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_FAIR_GROUP_SCHED */
324 #ifdef CONFIG_RT_GROUP_SCHED
325 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
326 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
327 #endif /* CONFIG_RT_GROUP_SCHED */
328 #else /* !CONFIG_USER_SCHED */
329 #define root_task_group init_task_group
330 #endif /* CONFIG_USER_SCHED */
332 /* task_group_lock serializes add/remove of task groups and also changes to
333 * a task group's cpu shares.
335 static DEFINE_SPINLOCK(task_group_lock);
338 static int root_task_group_empty(void)
340 return list_empty(&root_task_group.children);
344 #ifdef CONFIG_FAIR_GROUP_SCHED
345 #ifdef CONFIG_USER_SCHED
346 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
347 #else /* !CONFIG_USER_SCHED */
348 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
349 #endif /* CONFIG_USER_SCHED */
352 * A weight of 0 or 1 can cause arithmetics problems.
353 * A weight of a cfs_rq is the sum of weights of which entities
354 * are queued on this cfs_rq, so a weight of a entity should not be
355 * too large, so as the shares value of a task group.
356 * (The default weight is 1024 - so there's no practical
357 * limitation from this.)
360 #define MAX_SHARES (1UL << 18)
362 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
365 /* Default task group.
366 * Every task in system belong to this group at bootup.
368 struct task_group init_task_group;
370 /* return group to which a task belongs */
371 static inline struct task_group *task_group(struct task_struct *p)
373 struct task_group *tg;
375 #ifdef CONFIG_USER_SCHED
377 tg = __task_cred(p)->user->tg;
379 #elif defined(CONFIG_CGROUP_SCHED)
380 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
381 struct task_group, css);
383 tg = &init_task_group;
388 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
389 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
391 #ifdef CONFIG_FAIR_GROUP_SCHED
392 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
393 p->se.parent = task_group(p)->se[cpu];
396 #ifdef CONFIG_RT_GROUP_SCHED
397 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
398 p->rt.parent = task_group(p)->rt_se[cpu];
405 static int root_task_group_empty(void)
411 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
412 static inline struct task_group *task_group(struct task_struct *p)
417 #endif /* CONFIG_GROUP_SCHED */
419 /* CFS-related fields in a runqueue */
421 struct load_weight load;
422 unsigned long nr_running;
427 struct rb_root tasks_timeline;
428 struct rb_node *rb_leftmost;
430 struct list_head tasks;
431 struct list_head *balance_iterator;
434 * 'curr' points to currently running entity on this cfs_rq.
435 * It is set to NULL otherwise (i.e when none are currently running).
437 struct sched_entity *curr, *next, *last;
439 unsigned int nr_spread_over;
441 #ifdef CONFIG_FAIR_GROUP_SCHED
442 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
445 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
446 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
447 * (like users, containers etc.)
449 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
450 * list is used during load balance.
452 struct list_head leaf_cfs_rq_list;
453 struct task_group *tg; /* group that "owns" this runqueue */
457 * the part of load.weight contributed by tasks
459 unsigned long task_weight;
462 * h_load = weight * f(tg)
464 * Where f(tg) is the recursive weight fraction assigned to
467 unsigned long h_load;
470 * this cpu's part of tg->shares
472 unsigned long shares;
475 * load.weight at the time we set shares
477 unsigned long rq_weight;
482 /* Real-Time classes' related field in a runqueue: */
484 struct rt_prio_array active;
485 unsigned long rt_nr_running;
486 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
488 int curr; /* highest queued rt task prio */
490 int next; /* next highest */
495 unsigned long rt_nr_migratory;
496 unsigned long rt_nr_total;
498 struct plist_head pushable_tasks;
503 /* Nests inside the rq lock: */
504 spinlock_t rt_runtime_lock;
506 #ifdef CONFIG_RT_GROUP_SCHED
507 unsigned long rt_nr_boosted;
510 struct list_head leaf_rt_rq_list;
511 struct task_group *tg;
512 struct sched_rt_entity *rt_se;
519 * We add the notion of a root-domain which will be used to define per-domain
520 * variables. Each exclusive cpuset essentially defines an island domain by
521 * fully partitioning the member cpus from any other cpuset. Whenever a new
522 * exclusive cpuset is created, we also create and attach a new root-domain
529 cpumask_var_t online;
532 * The "RT overload" flag: it gets set if a CPU has more than
533 * one runnable RT task.
535 cpumask_var_t rto_mask;
538 struct cpupri cpupri;
540 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
542 * Preferred wake up cpu nominated by sched_mc balance that will be
543 * used when most cpus are idle in the system indicating overall very
544 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
546 unsigned int sched_mc_preferred_wakeup_cpu;
551 * By default the system creates a single root-domain with all cpus as
552 * members (mimicking the global state we have today).
554 static struct root_domain def_root_domain;
559 * This is the main, per-CPU runqueue data structure.
561 * Locking rule: those places that want to lock multiple runqueues
562 * (such as the load balancing or the thread migration code), lock
563 * acquire operations must be ordered by ascending &runqueue.
570 * nr_running and cpu_load should be in the same cacheline because
571 * remote CPUs use both these fields when doing load calculation.
573 unsigned long nr_running;
574 #define CPU_LOAD_IDX_MAX 5
575 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
577 unsigned long last_tick_seen;
578 unsigned char in_nohz_recently;
580 /* capture load from *all* tasks on this cpu: */
581 struct load_weight load;
582 unsigned long nr_load_updates;
584 u64 nr_migrations_in;
589 #ifdef CONFIG_FAIR_GROUP_SCHED
590 /* list of leaf cfs_rq on this cpu: */
591 struct list_head leaf_cfs_rq_list;
593 #ifdef CONFIG_RT_GROUP_SCHED
594 struct list_head leaf_rt_rq_list;
598 * This is part of a global counter where only the total sum
599 * over all CPUs matters. A task can increase this counter on
600 * one CPU and if it got migrated afterwards it may decrease
601 * it on another CPU. Always updated under the runqueue lock:
603 unsigned long nr_uninterruptible;
605 struct task_struct *curr, *idle;
606 unsigned long next_balance;
607 struct mm_struct *prev_mm;
614 struct root_domain *rd;
615 struct sched_domain *sd;
617 unsigned char idle_at_tick;
618 /* For active balancing */
622 /* cpu of this runqueue: */
626 unsigned long avg_load_per_task;
628 struct task_struct *migration_thread;
629 struct list_head migration_queue;
632 /* calc_load related fields */
633 unsigned long calc_load_update;
634 long calc_load_active;
636 #ifdef CONFIG_SCHED_HRTICK
638 int hrtick_csd_pending;
639 struct call_single_data hrtick_csd;
641 struct hrtimer hrtick_timer;
644 #ifdef CONFIG_SCHEDSTATS
646 struct sched_info rq_sched_info;
647 unsigned long long rq_cpu_time;
648 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
650 /* sys_sched_yield() stats */
651 unsigned int yld_count;
653 /* schedule() stats */
654 unsigned int sched_switch;
655 unsigned int sched_count;
656 unsigned int sched_goidle;
658 /* try_to_wake_up() stats */
659 unsigned int ttwu_count;
660 unsigned int ttwu_local;
663 unsigned int bkl_count;
667 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
669 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
671 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
674 static inline int cpu_of(struct rq *rq)
684 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
685 * See detach_destroy_domains: synchronize_sched for details.
687 * The domain tree of any CPU may only be accessed from within
688 * preempt-disabled sections.
690 #define for_each_domain(cpu, __sd) \
691 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
693 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
694 #define this_rq() (&__get_cpu_var(runqueues))
695 #define task_rq(p) cpu_rq(task_cpu(p))
696 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
697 #define raw_rq() (&__raw_get_cpu_var(runqueues))
699 inline void update_rq_clock(struct rq *rq)
701 rq->clock = sched_clock_cpu(cpu_of(rq));
705 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
707 #ifdef CONFIG_SCHED_DEBUG
708 # define const_debug __read_mostly
710 # define const_debug static const
716 * Returns true if the current cpu runqueue is locked.
717 * This interface allows printk to be called with the runqueue lock
718 * held and know whether or not it is OK to wake up the klogd.
720 int runqueue_is_locked(void)
723 struct rq *rq = cpu_rq(cpu);
726 ret = spin_is_locked(&rq->lock);
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug unsigned int sysctl_sched_features =
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly char *sched_feat_names[] = {
758 #include "sched_features.h"
764 static int sched_feat_show(struct seq_file *m, void *v)
768 for (i = 0; sched_feat_names[i]; i++) {
769 if (!(sysctl_sched_features & (1UL << i)))
771 seq_printf(m, "%s ", sched_feat_names[i]);
779 sched_feat_write(struct file *filp, const char __user *ubuf,
780 size_t cnt, loff_t *ppos)
790 if (copy_from_user(&buf, ubuf, cnt))
795 if (strncmp(buf, "NO_", 3) == 0) {
800 for (i = 0; sched_feat_names[i]; i++) {
801 int len = strlen(sched_feat_names[i]);
803 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
805 sysctl_sched_features &= ~(1UL << i);
807 sysctl_sched_features |= (1UL << i);
812 if (!sched_feat_names[i])
820 static int sched_feat_open(struct inode *inode, struct file *filp)
822 return single_open(filp, sched_feat_show, NULL);
825 static struct file_operations sched_feat_fops = {
826 .open = sched_feat_open,
827 .write = sched_feat_write,
830 .release = single_release,
833 static __init int sched_init_debug(void)
835 debugfs_create_file("sched_features", 0644, NULL, NULL,
840 late_initcall(sched_init_debug);
844 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
847 * Number of tasks to iterate in a single balance run.
848 * Limited because this is done with IRQs disabled.
850 const_debug unsigned int sysctl_sched_nr_migrate = 32;
853 * ratelimit for updating the group shares.
856 unsigned int sysctl_sched_shares_ratelimit = 250000;
859 * Inject some fuzzyness into changing the per-cpu group shares
860 * this avoids remote rq-locks at the expense of fairness.
863 unsigned int sysctl_sched_shares_thresh = 4;
866 * period over which we measure -rt task cpu usage in us.
869 unsigned int sysctl_sched_rt_period = 1000000;
871 static __read_mostly int scheduler_running;
874 * part of the period that we allow rt tasks to run in us.
877 int sysctl_sched_rt_runtime = 950000;
879 static inline u64 global_rt_period(void)
881 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
884 static inline u64 global_rt_runtime(void)
886 if (sysctl_sched_rt_runtime < 0)
889 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
892 #ifndef prepare_arch_switch
893 # define prepare_arch_switch(next) do { } while (0)
895 #ifndef finish_arch_switch
896 # define finish_arch_switch(prev) do { } while (0)
899 static inline int task_current(struct rq *rq, struct task_struct *p)
901 return rq->curr == p;
904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
905 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 #ifdef CONFIG_DEBUG_SPINLOCK
917 /* this is a valid case when another task releases the spinlock */
918 rq->lock.owner = current;
921 * If we are tracking spinlock dependencies then we have to
922 * fix up the runqueue lock - which gets 'carried over' from
925 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
927 spin_unlock_irq(&rq->lock);
930 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
931 static inline int task_running(struct rq *rq, struct task_struct *p)
936 return task_current(rq, p);
940 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
944 * We can optimise this out completely for !SMP, because the
945 * SMP rebalancing from interrupt is the only thing that cares
950 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
951 spin_unlock_irq(&rq->lock);
953 spin_unlock(&rq->lock);
957 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
961 * After ->oncpu is cleared, the task can be moved to a different CPU.
962 * We must ensure this doesn't happen until the switch is completely
968 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
972 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
975 * __task_rq_lock - lock the runqueue a given task resides on.
976 * Must be called interrupts disabled.
978 static inline struct rq *__task_rq_lock(struct task_struct *p)
982 struct rq *rq = task_rq(p);
983 spin_lock(&rq->lock);
984 if (likely(rq == task_rq(p)))
986 spin_unlock(&rq->lock);
991 * task_rq_lock - lock the runqueue a given task resides on and disable
992 * interrupts. Note the ordering: we can safely lookup the task_rq without
993 * explicitly disabling preemption.
995 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1001 local_irq_save(*flags);
1003 spin_lock(&rq->lock);
1004 if (likely(rq == task_rq(p)))
1006 spin_unlock_irqrestore(&rq->lock, *flags);
1010 void task_rq_unlock_wait(struct task_struct *p)
1012 struct rq *rq = task_rq(p);
1014 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1015 spin_unlock_wait(&rq->lock);
1018 static void __task_rq_unlock(struct rq *rq)
1019 __releases(rq->lock)
1021 spin_unlock(&rq->lock);
1024 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1025 __releases(rq->lock)
1027 spin_unlock_irqrestore(&rq->lock, *flags);
1031 * this_rq_lock - lock this runqueue and disable interrupts.
1033 static struct rq *this_rq_lock(void)
1034 __acquires(rq->lock)
1038 local_irq_disable();
1040 spin_lock(&rq->lock);
1045 #ifdef CONFIG_SCHED_HRTICK
1047 * Use HR-timers to deliver accurate preemption points.
1049 * Its all a bit involved since we cannot program an hrt while holding the
1050 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1053 * When we get rescheduled we reprogram the hrtick_timer outside of the
1059 * - enabled by features
1060 * - hrtimer is actually high res
1062 static inline int hrtick_enabled(struct rq *rq)
1064 if (!sched_feat(HRTICK))
1066 if (!cpu_active(cpu_of(rq)))
1068 return hrtimer_is_hres_active(&rq->hrtick_timer);
1071 static void hrtick_clear(struct rq *rq)
1073 if (hrtimer_active(&rq->hrtick_timer))
1074 hrtimer_cancel(&rq->hrtick_timer);
1078 * High-resolution timer tick.
1079 * Runs from hardirq context with interrupts disabled.
1081 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1083 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1085 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1087 spin_lock(&rq->lock);
1088 update_rq_clock(rq);
1089 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1090 spin_unlock(&rq->lock);
1092 return HRTIMER_NORESTART;
1097 * called from hardirq (IPI) context
1099 static void __hrtick_start(void *arg)
1101 struct rq *rq = arg;
1103 spin_lock(&rq->lock);
1104 hrtimer_restart(&rq->hrtick_timer);
1105 rq->hrtick_csd_pending = 0;
1106 spin_unlock(&rq->lock);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 struct hrtimer *timer = &rq->hrtick_timer;
1117 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1119 hrtimer_set_expires(timer, time);
1121 if (rq == this_rq()) {
1122 hrtimer_restart(timer);
1123 } else if (!rq->hrtick_csd_pending) {
1124 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1125 rq->hrtick_csd_pending = 1;
1130 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1132 int cpu = (int)(long)hcpu;
1135 case CPU_UP_CANCELED:
1136 case CPU_UP_CANCELED_FROZEN:
1137 case CPU_DOWN_PREPARE:
1138 case CPU_DOWN_PREPARE_FROZEN:
1140 case CPU_DEAD_FROZEN:
1141 hrtick_clear(cpu_rq(cpu));
1148 static __init void init_hrtick(void)
1150 hotcpu_notifier(hotplug_hrtick, 0);
1154 * Called to set the hrtick timer state.
1156 * called with rq->lock held and irqs disabled
1158 static void hrtick_start(struct rq *rq, u64 delay)
1160 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1161 HRTIMER_MODE_REL_PINNED, 0);
1164 static inline void init_hrtick(void)
1167 #endif /* CONFIG_SMP */
1169 static void init_rq_hrtick(struct rq *rq)
1172 rq->hrtick_csd_pending = 0;
1174 rq->hrtick_csd.flags = 0;
1175 rq->hrtick_csd.func = __hrtick_start;
1176 rq->hrtick_csd.info = rq;
1179 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1180 rq->hrtick_timer.function = hrtick;
1182 #else /* CONFIG_SCHED_HRTICK */
1183 static inline void hrtick_clear(struct rq *rq)
1187 static inline void init_rq_hrtick(struct rq *rq)
1191 static inline void init_hrtick(void)
1194 #endif /* CONFIG_SCHED_HRTICK */
1197 * resched_task - mark a task 'to be rescheduled now'.
1199 * On UP this means the setting of the need_resched flag, on SMP it
1200 * might also involve a cross-CPU call to trigger the scheduler on
1205 #ifndef tsk_is_polling
1206 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1209 static void resched_task(struct task_struct *p)
1213 assert_spin_locked(&task_rq(p)->lock);
1215 if (test_tsk_need_resched(p))
1218 set_tsk_need_resched(p);
1221 if (cpu == smp_processor_id())
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(p))
1227 smp_send_reschedule(cpu);
1230 static void resched_cpu(int cpu)
1232 struct rq *rq = cpu_rq(cpu);
1233 unsigned long flags;
1235 if (!spin_trylock_irqsave(&rq->lock, flags))
1237 resched_task(cpu_curr(cpu));
1238 spin_unlock_irqrestore(&rq->lock, flags);
1243 * When add_timer_on() enqueues a timer into the timer wheel of an
1244 * idle CPU then this timer might expire before the next timer event
1245 * which is scheduled to wake up that CPU. In case of a completely
1246 * idle system the next event might even be infinite time into the
1247 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1248 * leaves the inner idle loop so the newly added timer is taken into
1249 * account when the CPU goes back to idle and evaluates the timer
1250 * wheel for the next timer event.
1252 void wake_up_idle_cpu(int cpu)
1254 struct rq *rq = cpu_rq(cpu);
1256 if (cpu == smp_processor_id())
1260 * This is safe, as this function is called with the timer
1261 * wheel base lock of (cpu) held. When the CPU is on the way
1262 * to idle and has not yet set rq->curr to idle then it will
1263 * be serialized on the timer wheel base lock and take the new
1264 * timer into account automatically.
1266 if (rq->curr != rq->idle)
1270 * We can set TIF_RESCHED on the idle task of the other CPU
1271 * lockless. The worst case is that the other CPU runs the
1272 * idle task through an additional NOOP schedule()
1274 set_tsk_need_resched(rq->idle);
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(rq->idle))
1279 smp_send_reschedule(cpu);
1281 #endif /* CONFIG_NO_HZ */
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct *p)
1286 assert_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p);
1289 #endif /* CONFIG_SMP */
1291 #if BITS_PER_LONG == 32
1292 # define WMULT_CONST (~0UL)
1294 # define WMULT_CONST (1UL << 32)
1297 #define WMULT_SHIFT 32
1300 * Shift right and round:
1302 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305 * delta *= weight / lw
1307 static unsigned long
1308 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1309 struct load_weight *lw)
1313 if (!lw->inv_weight) {
1314 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1317 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1321 tmp = (u64)delta_exec * weight;
1323 * Check whether we'd overflow the 64-bit multiplication:
1325 if (unlikely(tmp > WMULT_CONST))
1326 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1329 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1331 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1334 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1340 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1348 * of tasks with abnormal "nice" values across CPUs the contribution that
1349 * each task makes to its run queue's load is weighted according to its
1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1351 * scaled version of the new time slice allocation that they receive on time
1355 #define WEIGHT_IDLEPRIO 3
1356 #define WMULT_IDLEPRIO 1431655765
1359 * Nice levels are multiplicative, with a gentle 10% change for every
1360 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1361 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1362 * that remained on nice 0.
1364 * The "10% effect" is relative and cumulative: from _any_ nice level,
1365 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1366 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1367 * If a task goes up by ~10% and another task goes down by ~10% then
1368 * the relative distance between them is ~25%.)
1370 static const int prio_to_weight[40] = {
1371 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1372 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1373 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1374 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1375 /* 0 */ 1024, 820, 655, 526, 423,
1376 /* 5 */ 335, 272, 215, 172, 137,
1377 /* 10 */ 110, 87, 70, 56, 45,
1378 /* 15 */ 36, 29, 23, 18, 15,
1382 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1384 * In cases where the weight does not change often, we can use the
1385 * precalculated inverse to speed up arithmetics by turning divisions
1386 * into multiplications:
1388 static const u32 prio_to_wmult[40] = {
1389 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1390 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1391 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1392 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1393 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1394 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1395 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1396 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1399 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1402 * runqueue iterator, to support SMP load-balancing between different
1403 * scheduling classes, without having to expose their internal data
1404 * structures to the load-balancing proper:
1406 struct rq_iterator {
1408 struct task_struct *(*start)(void *);
1409 struct task_struct *(*next)(void *);
1413 static unsigned long
1414 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 unsigned long max_load_move, struct sched_domain *sd,
1416 enum cpu_idle_type idle, int *all_pinned,
1417 int *this_best_prio, struct rq_iterator *iterator);
1420 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1421 struct sched_domain *sd, enum cpu_idle_type idle,
1422 struct rq_iterator *iterator);
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 static unsigned long source_load(int cpu, int type);
1500 static unsigned long target_load(int cpu, int type);
1501 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1503 static unsigned long cpu_avg_load_per_task(int cpu)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1509 rq->avg_load_per_task = rq->load.weight / nr_running;
1511 rq->avg_load_per_task = 0;
1513 return rq->avg_load_per_task;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 struct update_shares_data {
1519 unsigned long rq_weight[NR_CPUS];
1522 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1524 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1527 * Calculate and set the cpu's group shares.
1529 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1530 unsigned long sd_shares,
1531 unsigned long sd_rq_weight,
1532 struct update_shares_data *usd)
1534 unsigned long shares, rq_weight;
1537 rq_weight = usd->rq_weight[cpu];
1540 rq_weight = NICE_0_LOAD;
1544 * \Sum_j shares_j * rq_weight_i
1545 * shares_i = -----------------------------
1546 * \Sum_j rq_weight_j
1548 shares = (sd_shares * rq_weight) / sd_rq_weight;
1549 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1551 if (abs(shares - tg->se[cpu]->load.weight) >
1552 sysctl_sched_shares_thresh) {
1553 struct rq *rq = cpu_rq(cpu);
1554 unsigned long flags;
1556 spin_lock_irqsave(&rq->lock, flags);
1557 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1558 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1559 __set_se_shares(tg->se[cpu], shares);
1560 spin_unlock_irqrestore(&rq->lock, flags);
1565 * Re-compute the task group their per cpu shares over the given domain.
1566 * This needs to be done in a bottom-up fashion because the rq weight of a
1567 * parent group depends on the shares of its child groups.
1569 static int tg_shares_up(struct task_group *tg, void *data)
1571 unsigned long weight, rq_weight = 0, shares = 0;
1572 struct update_shares_data *usd;
1573 struct sched_domain *sd = data;
1574 unsigned long flags;
1580 local_irq_save(flags);
1581 usd = &__get_cpu_var(update_shares_data);
1583 for_each_cpu(i, sched_domain_span(sd)) {
1584 weight = tg->cfs_rq[i]->load.weight;
1585 usd->rq_weight[i] = weight;
1588 * If there are currently no tasks on the cpu pretend there
1589 * is one of average load so that when a new task gets to
1590 * run here it will not get delayed by group starvation.
1593 weight = NICE_0_LOAD;
1595 rq_weight += weight;
1596 shares += tg->cfs_rq[i]->shares;
1599 if ((!shares && rq_weight) || shares > tg->shares)
1600 shares = tg->shares;
1602 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1603 shares = tg->shares;
1605 for_each_cpu(i, sched_domain_span(sd))
1606 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1608 local_irq_restore(flags);
1614 * Compute the cpu's hierarchical load factor for each task group.
1615 * This needs to be done in a top-down fashion because the load of a child
1616 * group is a fraction of its parents load.
1618 static int tg_load_down(struct task_group *tg, void *data)
1621 long cpu = (long)data;
1624 load = cpu_rq(cpu)->load.weight;
1626 load = tg->parent->cfs_rq[cpu]->h_load;
1627 load *= tg->cfs_rq[cpu]->shares;
1628 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1631 tg->cfs_rq[cpu]->h_load = load;
1636 static void update_shares(struct sched_domain *sd)
1641 if (root_task_group_empty())
1644 now = cpu_clock(raw_smp_processor_id());
1645 elapsed = now - sd->last_update;
1647 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1648 sd->last_update = now;
1649 walk_tg_tree(tg_nop, tg_shares_up, sd);
1653 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1655 if (root_task_group_empty())
1658 spin_unlock(&rq->lock);
1660 spin_lock(&rq->lock);
1663 static void update_h_load(long cpu)
1665 if (root_task_group_empty())
1668 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1673 static inline void update_shares(struct sched_domain *sd)
1677 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1683 #ifdef CONFIG_PREEMPT
1686 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1687 * way at the expense of forcing extra atomic operations in all
1688 * invocations. This assures that the double_lock is acquired using the
1689 * same underlying policy as the spinlock_t on this architecture, which
1690 * reduces latency compared to the unfair variant below. However, it
1691 * also adds more overhead and therefore may reduce throughput.
1693 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1694 __releases(this_rq->lock)
1695 __acquires(busiest->lock)
1696 __acquires(this_rq->lock)
1698 spin_unlock(&this_rq->lock);
1699 double_rq_lock(this_rq, busiest);
1706 * Unfair double_lock_balance: Optimizes throughput at the expense of
1707 * latency by eliminating extra atomic operations when the locks are
1708 * already in proper order on entry. This favors lower cpu-ids and will
1709 * grant the double lock to lower cpus over higher ids under contention,
1710 * regardless of entry order into the function.
1712 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1713 __releases(this_rq->lock)
1714 __acquires(busiest->lock)
1715 __acquires(this_rq->lock)
1719 if (unlikely(!spin_trylock(&busiest->lock))) {
1720 if (busiest < this_rq) {
1721 spin_unlock(&this_rq->lock);
1722 spin_lock(&busiest->lock);
1723 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1726 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1731 #endif /* CONFIG_PREEMPT */
1734 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1736 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1738 if (unlikely(!irqs_disabled())) {
1739 /* printk() doesn't work good under rq->lock */
1740 spin_unlock(&this_rq->lock);
1744 return _double_lock_balance(this_rq, busiest);
1747 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1748 __releases(busiest->lock)
1750 spin_unlock(&busiest->lock);
1751 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1755 #ifdef CONFIG_FAIR_GROUP_SCHED
1756 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1759 cfs_rq->shares = shares;
1764 static void calc_load_account_active(struct rq *this_rq);
1766 #include "sched_stats.h"
1767 #include "sched_idletask.c"
1768 #include "sched_fair.c"
1769 #include "sched_rt.c"
1770 #ifdef CONFIG_SCHED_DEBUG
1771 # include "sched_debug.c"
1774 #define sched_class_highest (&rt_sched_class)
1775 #define for_each_class(class) \
1776 for (class = sched_class_highest; class; class = class->next)
1778 static void inc_nr_running(struct rq *rq)
1783 static void dec_nr_running(struct rq *rq)
1788 static void set_load_weight(struct task_struct *p)
1790 if (task_has_rt_policy(p)) {
1791 p->se.load.weight = prio_to_weight[0] * 2;
1792 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1797 * SCHED_IDLE tasks get minimal weight:
1799 if (p->policy == SCHED_IDLE) {
1800 p->se.load.weight = WEIGHT_IDLEPRIO;
1801 p->se.load.inv_weight = WMULT_IDLEPRIO;
1805 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1806 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1809 static void update_avg(u64 *avg, u64 sample)
1811 s64 diff = sample - *avg;
1815 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1818 p->se.start_runtime = p->se.sum_exec_runtime;
1820 sched_info_queued(p);
1821 p->sched_class->enqueue_task(rq, p, wakeup);
1825 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1828 if (p->se.last_wakeup) {
1829 update_avg(&p->se.avg_overlap,
1830 p->se.sum_exec_runtime - p->se.last_wakeup);
1831 p->se.last_wakeup = 0;
1833 update_avg(&p->se.avg_wakeup,
1834 sysctl_sched_wakeup_granularity);
1838 sched_info_dequeued(p);
1839 p->sched_class->dequeue_task(rq, p, sleep);
1844 * __normal_prio - return the priority that is based on the static prio
1846 static inline int __normal_prio(struct task_struct *p)
1848 return p->static_prio;
1852 * Calculate the expected normal priority: i.e. priority
1853 * without taking RT-inheritance into account. Might be
1854 * boosted by interactivity modifiers. Changes upon fork,
1855 * setprio syscalls, and whenever the interactivity
1856 * estimator recalculates.
1858 static inline int normal_prio(struct task_struct *p)
1862 if (task_has_rt_policy(p))
1863 prio = MAX_RT_PRIO-1 - p->rt_priority;
1865 prio = __normal_prio(p);
1870 * Calculate the current priority, i.e. the priority
1871 * taken into account by the scheduler. This value might
1872 * be boosted by RT tasks, or might be boosted by
1873 * interactivity modifiers. Will be RT if the task got
1874 * RT-boosted. If not then it returns p->normal_prio.
1876 static int effective_prio(struct task_struct *p)
1878 p->normal_prio = normal_prio(p);
1880 * If we are RT tasks or we were boosted to RT priority,
1881 * keep the priority unchanged. Otherwise, update priority
1882 * to the normal priority:
1884 if (!rt_prio(p->prio))
1885 return p->normal_prio;
1890 * activate_task - move a task to the runqueue.
1892 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1894 if (task_contributes_to_load(p))
1895 rq->nr_uninterruptible--;
1897 enqueue_task(rq, p, wakeup);
1902 * deactivate_task - remove a task from the runqueue.
1904 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1906 if (task_contributes_to_load(p))
1907 rq->nr_uninterruptible++;
1909 dequeue_task(rq, p, sleep);
1914 * task_curr - is this task currently executing on a CPU?
1915 * @p: the task in question.
1917 inline int task_curr(const struct task_struct *p)
1919 return cpu_curr(task_cpu(p)) == p;
1922 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1924 set_task_rq(p, cpu);
1927 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1928 * successfuly executed on another CPU. We must ensure that updates of
1929 * per-task data have been completed by this moment.
1932 task_thread_info(p)->cpu = cpu;
1936 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1937 const struct sched_class *prev_class,
1938 int oldprio, int running)
1940 if (prev_class != p->sched_class) {
1941 if (prev_class->switched_from)
1942 prev_class->switched_from(rq, p, running);
1943 p->sched_class->switched_to(rq, p, running);
1945 p->sched_class->prio_changed(rq, p, oldprio, running);
1950 /* Used instead of source_load when we know the type == 0 */
1951 static unsigned long weighted_cpuload(const int cpu)
1953 return cpu_rq(cpu)->load.weight;
1957 * Is this task likely cache-hot:
1960 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1965 * Buddy candidates are cache hot:
1967 if (sched_feat(CACHE_HOT_BUDDY) &&
1968 (&p->se == cfs_rq_of(&p->se)->next ||
1969 &p->se == cfs_rq_of(&p->se)->last))
1972 if (p->sched_class != &fair_sched_class)
1975 if (sysctl_sched_migration_cost == -1)
1977 if (sysctl_sched_migration_cost == 0)
1980 delta = now - p->se.exec_start;
1982 return delta < (s64)sysctl_sched_migration_cost;
1986 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1988 int old_cpu = task_cpu(p);
1989 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1990 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1991 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1994 clock_offset = old_rq->clock - new_rq->clock;
1996 trace_sched_migrate_task(p, new_cpu);
1998 #ifdef CONFIG_SCHEDSTATS
1999 if (p->se.wait_start)
2000 p->se.wait_start -= clock_offset;
2001 if (p->se.sleep_start)
2002 p->se.sleep_start -= clock_offset;
2003 if (p->se.block_start)
2004 p->se.block_start -= clock_offset;
2006 if (old_cpu != new_cpu) {
2007 p->se.nr_migrations++;
2008 new_rq->nr_migrations_in++;
2009 #ifdef CONFIG_SCHEDSTATS
2010 if (task_hot(p, old_rq->clock, NULL))
2011 schedstat_inc(p, se.nr_forced2_migrations);
2013 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2016 p->se.vruntime -= old_cfsrq->min_vruntime -
2017 new_cfsrq->min_vruntime;
2019 __set_task_cpu(p, new_cpu);
2022 struct migration_req {
2023 struct list_head list;
2025 struct task_struct *task;
2028 struct completion done;
2032 * The task's runqueue lock must be held.
2033 * Returns true if you have to wait for migration thread.
2036 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2038 struct rq *rq = task_rq(p);
2041 * If the task is not on a runqueue (and not running), then
2042 * it is sufficient to simply update the task's cpu field.
2044 if (!p->se.on_rq && !task_running(rq, p)) {
2045 set_task_cpu(p, dest_cpu);
2049 init_completion(&req->done);
2051 req->dest_cpu = dest_cpu;
2052 list_add(&req->list, &rq->migration_queue);
2058 * wait_task_context_switch - wait for a thread to complete at least one
2061 * @p must not be current.
2063 void wait_task_context_switch(struct task_struct *p)
2065 unsigned long nvcsw, nivcsw, flags;
2073 * The runqueue is assigned before the actual context
2074 * switch. We need to take the runqueue lock.
2076 * We could check initially without the lock but it is
2077 * very likely that we need to take the lock in every
2080 rq = task_rq_lock(p, &flags);
2081 running = task_running(rq, p);
2082 task_rq_unlock(rq, &flags);
2084 if (likely(!running))
2087 * The switch count is incremented before the actual
2088 * context switch. We thus wait for two switches to be
2089 * sure at least one completed.
2091 if ((p->nvcsw - nvcsw) > 1)
2093 if ((p->nivcsw - nivcsw) > 1)
2101 * wait_task_inactive - wait for a thread to unschedule.
2103 * If @match_state is nonzero, it's the @p->state value just checked and
2104 * not expected to change. If it changes, i.e. @p might have woken up,
2105 * then return zero. When we succeed in waiting for @p to be off its CPU,
2106 * we return a positive number (its total switch count). If a second call
2107 * a short while later returns the same number, the caller can be sure that
2108 * @p has remained unscheduled the whole time.
2110 * The caller must ensure that the task *will* unschedule sometime soon,
2111 * else this function might spin for a *long* time. This function can't
2112 * be called with interrupts off, or it may introduce deadlock with
2113 * smp_call_function() if an IPI is sent by the same process we are
2114 * waiting to become inactive.
2116 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2118 unsigned long flags;
2125 * We do the initial early heuristics without holding
2126 * any task-queue locks at all. We'll only try to get
2127 * the runqueue lock when things look like they will
2133 * If the task is actively running on another CPU
2134 * still, just relax and busy-wait without holding
2137 * NOTE! Since we don't hold any locks, it's not
2138 * even sure that "rq" stays as the right runqueue!
2139 * But we don't care, since "task_running()" will
2140 * return false if the runqueue has changed and p
2141 * is actually now running somewhere else!
2143 while (task_running(rq, p)) {
2144 if (match_state && unlikely(p->state != match_state))
2150 * Ok, time to look more closely! We need the rq
2151 * lock now, to be *sure*. If we're wrong, we'll
2152 * just go back and repeat.
2154 rq = task_rq_lock(p, &flags);
2155 trace_sched_wait_task(rq, p);
2156 running = task_running(rq, p);
2157 on_rq = p->se.on_rq;
2159 if (!match_state || p->state == match_state)
2160 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2161 task_rq_unlock(rq, &flags);
2164 * If it changed from the expected state, bail out now.
2166 if (unlikely(!ncsw))
2170 * Was it really running after all now that we
2171 * checked with the proper locks actually held?
2173 * Oops. Go back and try again..
2175 if (unlikely(running)) {
2181 * It's not enough that it's not actively running,
2182 * it must be off the runqueue _entirely_, and not
2185 * So if it was still runnable (but just not actively
2186 * running right now), it's preempted, and we should
2187 * yield - it could be a while.
2189 if (unlikely(on_rq)) {
2190 schedule_timeout_uninterruptible(1);
2195 * Ahh, all good. It wasn't running, and it wasn't
2196 * runnable, which means that it will never become
2197 * running in the future either. We're all done!
2206 * kick_process - kick a running thread to enter/exit the kernel
2207 * @p: the to-be-kicked thread
2209 * Cause a process which is running on another CPU to enter
2210 * kernel-mode, without any delay. (to get signals handled.)
2212 * NOTE: this function doesnt have to take the runqueue lock,
2213 * because all it wants to ensure is that the remote task enters
2214 * the kernel. If the IPI races and the task has been migrated
2215 * to another CPU then no harm is done and the purpose has been
2218 void kick_process(struct task_struct *p)
2224 if ((cpu != smp_processor_id()) && task_curr(p))
2225 smp_send_reschedule(cpu);
2228 EXPORT_SYMBOL_GPL(kick_process);
2231 * Return a low guess at the load of a migration-source cpu weighted
2232 * according to the scheduling class and "nice" value.
2234 * We want to under-estimate the load of migration sources, to
2235 * balance conservatively.
2237 static unsigned long source_load(int cpu, int type)
2239 struct rq *rq = cpu_rq(cpu);
2240 unsigned long total = weighted_cpuload(cpu);
2242 if (type == 0 || !sched_feat(LB_BIAS))
2245 return min(rq->cpu_load[type-1], total);
2249 * Return a high guess at the load of a migration-target cpu weighted
2250 * according to the scheduling class and "nice" value.
2252 static unsigned long target_load(int cpu, int type)
2254 struct rq *rq = cpu_rq(cpu);
2255 unsigned long total = weighted_cpuload(cpu);
2257 if (type == 0 || !sched_feat(LB_BIAS))
2260 return max(rq->cpu_load[type-1], total);
2264 * find_idlest_group finds and returns the least busy CPU group within the
2267 static struct sched_group *
2268 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2270 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2271 unsigned long min_load = ULONG_MAX, this_load = 0;
2272 int load_idx = sd->forkexec_idx;
2273 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2276 unsigned long load, avg_load;
2280 /* Skip over this group if it has no CPUs allowed */
2281 if (!cpumask_intersects(sched_group_cpus(group),
2285 local_group = cpumask_test_cpu(this_cpu,
2286 sched_group_cpus(group));
2288 /* Tally up the load of all CPUs in the group */
2291 for_each_cpu(i, sched_group_cpus(group)) {
2292 /* Bias balancing toward cpus of our domain */
2294 load = source_load(i, load_idx);
2296 load = target_load(i, load_idx);
2301 /* Adjust by relative CPU power of the group */
2302 avg_load = sg_div_cpu_power(group,
2303 avg_load * SCHED_LOAD_SCALE);
2306 this_load = avg_load;
2308 } else if (avg_load < min_load) {
2309 min_load = avg_load;
2312 } while (group = group->next, group != sd->groups);
2314 if (!idlest || 100*this_load < imbalance*min_load)
2320 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2323 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2325 unsigned long load, min_load = ULONG_MAX;
2329 /* Traverse only the allowed CPUs */
2330 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2331 load = weighted_cpuload(i);
2333 if (load < min_load || (load == min_load && i == this_cpu)) {
2343 * sched_balance_self: balance the current task (running on cpu) in domains
2344 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2347 * Balance, ie. select the least loaded group.
2349 * Returns the target CPU number, or the same CPU if no balancing is needed.
2351 * preempt must be disabled.
2353 static int sched_balance_self(int cpu, int flag)
2355 struct task_struct *t = current;
2356 struct sched_domain *tmp, *sd = NULL;
2358 for_each_domain(cpu, tmp) {
2360 * If power savings logic is enabled for a domain, stop there.
2362 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2364 if (tmp->flags & flag)
2372 struct sched_group *group;
2373 int new_cpu, weight;
2375 if (!(sd->flags & flag)) {
2380 group = find_idlest_group(sd, t, cpu);
2386 new_cpu = find_idlest_cpu(group, t, cpu);
2387 if (new_cpu == -1 || new_cpu == cpu) {
2388 /* Now try balancing at a lower domain level of cpu */
2393 /* Now try balancing at a lower domain level of new_cpu */
2395 weight = cpumask_weight(sched_domain_span(sd));
2397 for_each_domain(cpu, tmp) {
2398 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2400 if (tmp->flags & flag)
2403 /* while loop will break here if sd == NULL */
2409 #endif /* CONFIG_SMP */
2412 * task_oncpu_function_call - call a function on the cpu on which a task runs
2413 * @p: the task to evaluate
2414 * @func: the function to be called
2415 * @info: the function call argument
2417 * Calls the function @func when the task is currently running. This might
2418 * be on the current CPU, which just calls the function directly
2420 void task_oncpu_function_call(struct task_struct *p,
2421 void (*func) (void *info), void *info)
2428 smp_call_function_single(cpu, func, info, 1);
2433 * try_to_wake_up - wake up a thread
2434 * @p: the to-be-woken-up thread
2435 * @state: the mask of task states that can be woken
2436 * @sync: do a synchronous wakeup?
2438 * Put it on the run-queue if it's not already there. The "current"
2439 * thread is always on the run-queue (except when the actual
2440 * re-schedule is in progress), and as such you're allowed to do
2441 * the simpler "current->state = TASK_RUNNING" to mark yourself
2442 * runnable without the overhead of this.
2444 * returns failure only if the task is already active.
2446 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2448 int cpu, orig_cpu, this_cpu, success = 0;
2449 unsigned long flags;
2453 if (!sched_feat(SYNC_WAKEUPS))
2457 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2458 struct sched_domain *sd;
2460 this_cpu = raw_smp_processor_id();
2463 for_each_domain(this_cpu, sd) {
2464 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2473 rq = task_rq_lock(p, &flags);
2474 update_rq_clock(rq);
2475 old_state = p->state;
2476 if (!(old_state & state))
2484 this_cpu = smp_processor_id();
2487 if (unlikely(task_running(rq, p)))
2490 cpu = p->sched_class->select_task_rq(p, sync);
2491 if (cpu != orig_cpu) {
2492 set_task_cpu(p, cpu);
2493 task_rq_unlock(rq, &flags);
2494 /* might preempt at this point */
2495 rq = task_rq_lock(p, &flags);
2496 old_state = p->state;
2497 if (!(old_state & state))
2502 this_cpu = smp_processor_id();
2506 #ifdef CONFIG_SCHEDSTATS
2507 schedstat_inc(rq, ttwu_count);
2508 if (cpu == this_cpu)
2509 schedstat_inc(rq, ttwu_local);
2511 struct sched_domain *sd;
2512 for_each_domain(this_cpu, sd) {
2513 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2514 schedstat_inc(sd, ttwu_wake_remote);
2519 #endif /* CONFIG_SCHEDSTATS */
2522 #endif /* CONFIG_SMP */
2523 schedstat_inc(p, se.nr_wakeups);
2525 schedstat_inc(p, se.nr_wakeups_sync);
2526 if (orig_cpu != cpu)
2527 schedstat_inc(p, se.nr_wakeups_migrate);
2528 if (cpu == this_cpu)
2529 schedstat_inc(p, se.nr_wakeups_local);
2531 schedstat_inc(p, se.nr_wakeups_remote);
2532 activate_task(rq, p, 1);
2536 * Only attribute actual wakeups done by this task.
2538 if (!in_interrupt()) {
2539 struct sched_entity *se = ¤t->se;
2540 u64 sample = se->sum_exec_runtime;
2542 if (se->last_wakeup)
2543 sample -= se->last_wakeup;
2545 sample -= se->start_runtime;
2546 update_avg(&se->avg_wakeup, sample);
2548 se->last_wakeup = se->sum_exec_runtime;
2552 trace_sched_wakeup(rq, p, success);
2553 check_preempt_curr(rq, p, sync);
2555 p->state = TASK_RUNNING;
2557 if (p->sched_class->task_wake_up)
2558 p->sched_class->task_wake_up(rq, p);
2561 task_rq_unlock(rq, &flags);
2567 * wake_up_process - Wake up a specific process
2568 * @p: The process to be woken up.
2570 * Attempt to wake up the nominated process and move it to the set of runnable
2571 * processes. Returns 1 if the process was woken up, 0 if it was already
2574 * It may be assumed that this function implies a write memory barrier before
2575 * changing the task state if and only if any tasks are woken up.
2577 int wake_up_process(struct task_struct *p)
2579 return try_to_wake_up(p, TASK_ALL, 0);
2581 EXPORT_SYMBOL(wake_up_process);
2583 int wake_up_state(struct task_struct *p, unsigned int state)
2585 return try_to_wake_up(p, state, 0);
2589 * Perform scheduler related setup for a newly forked process p.
2590 * p is forked by current.
2592 * __sched_fork() is basic setup used by init_idle() too:
2594 static void __sched_fork(struct task_struct *p)
2596 p->se.exec_start = 0;
2597 p->se.sum_exec_runtime = 0;
2598 p->se.prev_sum_exec_runtime = 0;
2599 p->se.nr_migrations = 0;
2600 p->se.last_wakeup = 0;
2601 p->se.avg_overlap = 0;
2602 p->se.start_runtime = 0;
2603 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2605 #ifdef CONFIG_SCHEDSTATS
2606 p->se.wait_start = 0;
2608 p->se.wait_count = 0;
2611 p->se.sleep_start = 0;
2612 p->se.sleep_max = 0;
2613 p->se.sum_sleep_runtime = 0;
2615 p->se.block_start = 0;
2616 p->se.block_max = 0;
2618 p->se.slice_max = 0;
2620 p->se.nr_migrations_cold = 0;
2621 p->se.nr_failed_migrations_affine = 0;
2622 p->se.nr_failed_migrations_running = 0;
2623 p->se.nr_failed_migrations_hot = 0;
2624 p->se.nr_forced_migrations = 0;
2625 p->se.nr_forced2_migrations = 0;
2627 p->se.nr_wakeups = 0;
2628 p->se.nr_wakeups_sync = 0;
2629 p->se.nr_wakeups_migrate = 0;
2630 p->se.nr_wakeups_local = 0;
2631 p->se.nr_wakeups_remote = 0;
2632 p->se.nr_wakeups_affine = 0;
2633 p->se.nr_wakeups_affine_attempts = 0;
2634 p->se.nr_wakeups_passive = 0;
2635 p->se.nr_wakeups_idle = 0;
2639 INIT_LIST_HEAD(&p->rt.run_list);
2641 INIT_LIST_HEAD(&p->se.group_node);
2643 #ifdef CONFIG_PREEMPT_NOTIFIERS
2644 INIT_HLIST_HEAD(&p->preempt_notifiers);
2648 * We mark the process as running here, but have not actually
2649 * inserted it onto the runqueue yet. This guarantees that
2650 * nobody will actually run it, and a signal or other external
2651 * event cannot wake it up and insert it on the runqueue either.
2653 p->state = TASK_RUNNING;
2657 * fork()/clone()-time setup:
2659 void sched_fork(struct task_struct *p, int clone_flags)
2661 int cpu = get_cpu();
2666 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2668 set_task_cpu(p, cpu);
2671 * Make sure we do not leak PI boosting priority to the child.
2673 p->prio = current->normal_prio;
2676 * Revert to default priority/policy on fork if requested.
2678 if (unlikely(p->sched_reset_on_fork)) {
2679 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2680 p->policy = SCHED_NORMAL;
2682 if (p->normal_prio < DEFAULT_PRIO)
2683 p->prio = DEFAULT_PRIO;
2685 if (PRIO_TO_NICE(p->static_prio) < 0) {
2686 p->static_prio = NICE_TO_PRIO(0);
2691 * We don't need the reset flag anymore after the fork. It has
2692 * fulfilled its duty:
2694 p->sched_reset_on_fork = 0;
2697 if (!rt_prio(p->prio))
2698 p->sched_class = &fair_sched_class;
2700 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2701 if (likely(sched_info_on()))
2702 memset(&p->sched_info, 0, sizeof(p->sched_info));
2704 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2707 #ifdef CONFIG_PREEMPT
2708 /* Want to start with kernel preemption disabled. */
2709 task_thread_info(p)->preempt_count = 1;
2711 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2717 * wake_up_new_task - wake up a newly created task for the first time.
2719 * This function will do some initial scheduler statistics housekeeping
2720 * that must be done for every newly created context, then puts the task
2721 * on the runqueue and wakes it.
2723 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2725 unsigned long flags;
2728 rq = task_rq_lock(p, &flags);
2729 BUG_ON(p->state != TASK_RUNNING);
2730 update_rq_clock(rq);
2732 p->prio = effective_prio(p);
2734 if (!p->sched_class->task_new || !current->se.on_rq) {
2735 activate_task(rq, p, 0);
2738 * Let the scheduling class do new task startup
2739 * management (if any):
2741 p->sched_class->task_new(rq, p);
2744 trace_sched_wakeup_new(rq, p, 1);
2745 check_preempt_curr(rq, p, 0);
2747 if (p->sched_class->task_wake_up)
2748 p->sched_class->task_wake_up(rq, p);
2750 task_rq_unlock(rq, &flags);
2753 #ifdef CONFIG_PREEMPT_NOTIFIERS
2756 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2757 * @notifier: notifier struct to register
2759 void preempt_notifier_register(struct preempt_notifier *notifier)
2761 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2763 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2766 * preempt_notifier_unregister - no longer interested in preemption notifications
2767 * @notifier: notifier struct to unregister
2769 * This is safe to call from within a preemption notifier.
2771 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2773 hlist_del(¬ifier->link);
2775 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2777 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2779 struct preempt_notifier *notifier;
2780 struct hlist_node *node;
2782 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2783 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2787 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2788 struct task_struct *next)
2790 struct preempt_notifier *notifier;
2791 struct hlist_node *node;
2793 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2794 notifier->ops->sched_out(notifier, next);
2797 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2799 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2804 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2805 struct task_struct *next)
2809 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2812 * prepare_task_switch - prepare to switch tasks
2813 * @rq: the runqueue preparing to switch
2814 * @prev: the current task that is being switched out
2815 * @next: the task we are going to switch to.
2817 * This is called with the rq lock held and interrupts off. It must
2818 * be paired with a subsequent finish_task_switch after the context
2821 * prepare_task_switch sets up locking and calls architecture specific
2825 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2826 struct task_struct *next)
2828 fire_sched_out_preempt_notifiers(prev, next);
2829 prepare_lock_switch(rq, next);
2830 prepare_arch_switch(next);
2834 * finish_task_switch - clean up after a task-switch
2835 * @rq: runqueue associated with task-switch
2836 * @prev: the thread we just switched away from.
2838 * finish_task_switch must be called after the context switch, paired
2839 * with a prepare_task_switch call before the context switch.
2840 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2841 * and do any other architecture-specific cleanup actions.
2843 * Note that we may have delayed dropping an mm in context_switch(). If
2844 * so, we finish that here outside of the runqueue lock. (Doing it
2845 * with the lock held can cause deadlocks; see schedule() for
2848 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2849 __releases(rq->lock)
2851 struct mm_struct *mm = rq->prev_mm;
2857 * A task struct has one reference for the use as "current".
2858 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2859 * schedule one last time. The schedule call will never return, and
2860 * the scheduled task must drop that reference.
2861 * The test for TASK_DEAD must occur while the runqueue locks are
2862 * still held, otherwise prev could be scheduled on another cpu, die
2863 * there before we look at prev->state, and then the reference would
2865 * Manfred Spraul <manfred@colorfullife.com>
2867 prev_state = prev->state;
2868 finish_arch_switch(prev);
2869 perf_counter_task_sched_in(current, cpu_of(rq));
2870 finish_lock_switch(rq, prev);
2872 fire_sched_in_preempt_notifiers(current);
2875 if (unlikely(prev_state == TASK_DEAD)) {
2877 * Remove function-return probe instances associated with this
2878 * task and put them back on the free list.
2880 kprobe_flush_task(prev);
2881 put_task_struct(prev);
2887 /* assumes rq->lock is held */
2888 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2890 if (prev->sched_class->pre_schedule)
2891 prev->sched_class->pre_schedule(rq, prev);
2894 /* rq->lock is NOT held, but preemption is disabled */
2895 static inline void post_schedule(struct rq *rq)
2897 if (rq->post_schedule) {
2898 unsigned long flags;
2900 spin_lock_irqsave(&rq->lock, flags);
2901 if (rq->curr->sched_class->post_schedule)
2902 rq->curr->sched_class->post_schedule(rq);
2903 spin_unlock_irqrestore(&rq->lock, flags);
2905 rq->post_schedule = 0;
2911 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2915 static inline void post_schedule(struct rq *rq)
2922 * schedule_tail - first thing a freshly forked thread must call.
2923 * @prev: the thread we just switched away from.
2925 asmlinkage void schedule_tail(struct task_struct *prev)
2926 __releases(rq->lock)
2928 struct rq *rq = this_rq();
2930 finish_task_switch(rq, prev);
2933 * FIXME: do we need to worry about rq being invalidated by the
2938 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2939 /* In this case, finish_task_switch does not reenable preemption */
2942 if (current->set_child_tid)
2943 put_user(task_pid_vnr(current), current->set_child_tid);
2947 * context_switch - switch to the new MM and the new
2948 * thread's register state.
2951 context_switch(struct rq *rq, struct task_struct *prev,
2952 struct task_struct *next)
2954 struct mm_struct *mm, *oldmm;
2956 prepare_task_switch(rq, prev, next);
2957 trace_sched_switch(rq, prev, next);
2959 oldmm = prev->active_mm;
2961 * For paravirt, this is coupled with an exit in switch_to to
2962 * combine the page table reload and the switch backend into
2965 arch_start_context_switch(prev);
2967 if (unlikely(!mm)) {
2968 next->active_mm = oldmm;
2969 atomic_inc(&oldmm->mm_count);
2970 enter_lazy_tlb(oldmm, next);
2972 switch_mm(oldmm, mm, next);
2974 if (unlikely(!prev->mm)) {
2975 prev->active_mm = NULL;
2976 rq->prev_mm = oldmm;
2979 * Since the runqueue lock will be released by the next
2980 * task (which is an invalid locking op but in the case
2981 * of the scheduler it's an obvious special-case), so we
2982 * do an early lockdep release here:
2984 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2985 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2988 /* Here we just switch the register state and the stack. */
2989 switch_to(prev, next, prev);
2993 * this_rq must be evaluated again because prev may have moved
2994 * CPUs since it called schedule(), thus the 'rq' on its stack
2995 * frame will be invalid.
2997 finish_task_switch(this_rq(), prev);
3001 * nr_running, nr_uninterruptible and nr_context_switches:
3003 * externally visible scheduler statistics: current number of runnable
3004 * threads, current number of uninterruptible-sleeping threads, total
3005 * number of context switches performed since bootup.
3007 unsigned long nr_running(void)
3009 unsigned long i, sum = 0;
3011 for_each_online_cpu(i)
3012 sum += cpu_rq(i)->nr_running;
3017 unsigned long nr_uninterruptible(void)
3019 unsigned long i, sum = 0;
3021 for_each_possible_cpu(i)
3022 sum += cpu_rq(i)->nr_uninterruptible;
3025 * Since we read the counters lockless, it might be slightly
3026 * inaccurate. Do not allow it to go below zero though:
3028 if (unlikely((long)sum < 0))
3034 unsigned long long nr_context_switches(void)
3037 unsigned long long sum = 0;
3039 for_each_possible_cpu(i)
3040 sum += cpu_rq(i)->nr_switches;
3045 unsigned long nr_iowait(void)
3047 unsigned long i, sum = 0;
3049 for_each_possible_cpu(i)
3050 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3055 /* Variables and functions for calc_load */
3056 static atomic_long_t calc_load_tasks;
3057 static unsigned long calc_load_update;
3058 unsigned long avenrun[3];
3059 EXPORT_SYMBOL(avenrun);
3062 * get_avenrun - get the load average array
3063 * @loads: pointer to dest load array
3064 * @offset: offset to add
3065 * @shift: shift count to shift the result left
3067 * These values are estimates at best, so no need for locking.
3069 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3071 loads[0] = (avenrun[0] + offset) << shift;
3072 loads[1] = (avenrun[1] + offset) << shift;
3073 loads[2] = (avenrun[2] + offset) << shift;
3076 static unsigned long
3077 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3080 load += active * (FIXED_1 - exp);
3081 return load >> FSHIFT;
3085 * calc_load - update the avenrun load estimates 10 ticks after the
3086 * CPUs have updated calc_load_tasks.
3088 void calc_global_load(void)
3090 unsigned long upd = calc_load_update + 10;
3093 if (time_before(jiffies, upd))
3096 active = atomic_long_read(&calc_load_tasks);
3097 active = active > 0 ? active * FIXED_1 : 0;
3099 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3100 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3101 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3103 calc_load_update += LOAD_FREQ;
3107 * Either called from update_cpu_load() or from a cpu going idle
3109 static void calc_load_account_active(struct rq *this_rq)
3111 long nr_active, delta;
3113 nr_active = this_rq->nr_running;
3114 nr_active += (long) this_rq->nr_uninterruptible;
3116 if (nr_active != this_rq->calc_load_active) {
3117 delta = nr_active - this_rq->calc_load_active;
3118 this_rq->calc_load_active = nr_active;
3119 atomic_long_add(delta, &calc_load_tasks);
3124 * Externally visible per-cpu scheduler statistics:
3125 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3127 u64 cpu_nr_migrations(int cpu)
3129 return cpu_rq(cpu)->nr_migrations_in;
3133 * Update rq->cpu_load[] statistics. This function is usually called every
3134 * scheduler tick (TICK_NSEC).
3136 static void update_cpu_load(struct rq *this_rq)
3138 unsigned long this_load = this_rq->load.weight;
3141 this_rq->nr_load_updates++;
3143 /* Update our load: */
3144 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3145 unsigned long old_load, new_load;
3147 /* scale is effectively 1 << i now, and >> i divides by scale */
3149 old_load = this_rq->cpu_load[i];
3150 new_load = this_load;
3152 * Round up the averaging division if load is increasing. This
3153 * prevents us from getting stuck on 9 if the load is 10, for
3156 if (new_load > old_load)
3157 new_load += scale-1;
3158 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3161 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3162 this_rq->calc_load_update += LOAD_FREQ;
3163 calc_load_account_active(this_rq);
3170 * double_rq_lock - safely lock two runqueues
3172 * Note this does not disable interrupts like task_rq_lock,
3173 * you need to do so manually before calling.
3175 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3176 __acquires(rq1->lock)
3177 __acquires(rq2->lock)
3179 BUG_ON(!irqs_disabled());
3181 spin_lock(&rq1->lock);
3182 __acquire(rq2->lock); /* Fake it out ;) */
3185 spin_lock(&rq1->lock);
3186 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3188 spin_lock(&rq2->lock);
3189 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3192 update_rq_clock(rq1);
3193 update_rq_clock(rq2);
3197 * double_rq_unlock - safely unlock two runqueues
3199 * Note this does not restore interrupts like task_rq_unlock,
3200 * you need to do so manually after calling.
3202 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3203 __releases(rq1->lock)
3204 __releases(rq2->lock)
3206 spin_unlock(&rq1->lock);
3208 spin_unlock(&rq2->lock);
3210 __release(rq2->lock);
3214 * If dest_cpu is allowed for this process, migrate the task to it.
3215 * This is accomplished by forcing the cpu_allowed mask to only
3216 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3217 * the cpu_allowed mask is restored.
3219 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3221 struct migration_req req;
3222 unsigned long flags;
3225 rq = task_rq_lock(p, &flags);
3226 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3227 || unlikely(!cpu_active(dest_cpu)))
3230 /* force the process onto the specified CPU */
3231 if (migrate_task(p, dest_cpu, &req)) {
3232 /* Need to wait for migration thread (might exit: take ref). */
3233 struct task_struct *mt = rq->migration_thread;
3235 get_task_struct(mt);
3236 task_rq_unlock(rq, &flags);
3237 wake_up_process(mt);
3238 put_task_struct(mt);
3239 wait_for_completion(&req.done);
3244 task_rq_unlock(rq, &flags);
3248 * sched_exec - execve() is a valuable balancing opportunity, because at
3249 * this point the task has the smallest effective memory and cache footprint.
3251 void sched_exec(void)
3253 int new_cpu, this_cpu = get_cpu();
3254 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3256 if (new_cpu != this_cpu)
3257 sched_migrate_task(current, new_cpu);
3261 * pull_task - move a task from a remote runqueue to the local runqueue.
3262 * Both runqueues must be locked.
3264 static void pull_task(struct rq *src_rq, struct task_struct *p,
3265 struct rq *this_rq, int this_cpu)
3267 deactivate_task(src_rq, p, 0);
3268 set_task_cpu(p, this_cpu);
3269 activate_task(this_rq, p, 0);
3271 * Note that idle threads have a prio of MAX_PRIO, for this test
3272 * to be always true for them.
3274 check_preempt_curr(this_rq, p, 0);
3278 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3281 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3282 struct sched_domain *sd, enum cpu_idle_type idle,
3285 int tsk_cache_hot = 0;
3287 * We do not migrate tasks that are:
3288 * 1) running (obviously), or
3289 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3290 * 3) are cache-hot on their current CPU.
3292 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3293 schedstat_inc(p, se.nr_failed_migrations_affine);
3298 if (task_running(rq, p)) {
3299 schedstat_inc(p, se.nr_failed_migrations_running);
3304 * Aggressive migration if:
3305 * 1) task is cache cold, or
3306 * 2) too many balance attempts have failed.
3309 tsk_cache_hot = task_hot(p, rq->clock, sd);
3310 if (!tsk_cache_hot ||
3311 sd->nr_balance_failed > sd->cache_nice_tries) {
3312 #ifdef CONFIG_SCHEDSTATS
3313 if (tsk_cache_hot) {
3314 schedstat_inc(sd, lb_hot_gained[idle]);
3315 schedstat_inc(p, se.nr_forced_migrations);
3321 if (tsk_cache_hot) {
3322 schedstat_inc(p, se.nr_failed_migrations_hot);
3328 static unsigned long
3329 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3330 unsigned long max_load_move, struct sched_domain *sd,
3331 enum cpu_idle_type idle, int *all_pinned,
3332 int *this_best_prio, struct rq_iterator *iterator)
3334 int loops = 0, pulled = 0, pinned = 0;
3335 struct task_struct *p;
3336 long rem_load_move = max_load_move;
3338 if (max_load_move == 0)
3344 * Start the load-balancing iterator:
3346 p = iterator->start(iterator->arg);
3348 if (!p || loops++ > sysctl_sched_nr_migrate)
3351 if ((p->se.load.weight >> 1) > rem_load_move ||
3352 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3353 p = iterator->next(iterator->arg);
3357 pull_task(busiest, p, this_rq, this_cpu);
3359 rem_load_move -= p->se.load.weight;
3361 #ifdef CONFIG_PREEMPT
3363 * NEWIDLE balancing is a source of latency, so preemptible kernels
3364 * will stop after the first task is pulled to minimize the critical
3367 if (idle == CPU_NEWLY_IDLE)
3372 * We only want to steal up to the prescribed amount of weighted load.
3374 if (rem_load_move > 0) {
3375 if (p->prio < *this_best_prio)
3376 *this_best_prio = p->prio;
3377 p = iterator->next(iterator->arg);
3382 * Right now, this is one of only two places pull_task() is called,
3383 * so we can safely collect pull_task() stats here rather than
3384 * inside pull_task().
3386 schedstat_add(sd, lb_gained[idle], pulled);
3389 *all_pinned = pinned;
3391 return max_load_move - rem_load_move;
3395 * move_tasks tries to move up to max_load_move weighted load from busiest to
3396 * this_rq, as part of a balancing operation within domain "sd".
3397 * Returns 1 if successful and 0 otherwise.
3399 * Called with both runqueues locked.
3401 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3402 unsigned long max_load_move,
3403 struct sched_domain *sd, enum cpu_idle_type idle,
3406 const struct sched_class *class = sched_class_highest;
3407 unsigned long total_load_moved = 0;
3408 int this_best_prio = this_rq->curr->prio;
3412 class->load_balance(this_rq, this_cpu, busiest,
3413 max_load_move - total_load_moved,
3414 sd, idle, all_pinned, &this_best_prio);
3415 class = class->next;
3417 #ifdef CONFIG_PREEMPT
3419 * NEWIDLE balancing is a source of latency, so preemptible
3420 * kernels will stop after the first task is pulled to minimize
3421 * the critical section.
3423 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3426 } while (class && max_load_move > total_load_moved);
3428 return total_load_moved > 0;
3432 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3433 struct sched_domain *sd, enum cpu_idle_type idle,
3434 struct rq_iterator *iterator)
3436 struct task_struct *p = iterator->start(iterator->arg);
3440 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3441 pull_task(busiest, p, this_rq, this_cpu);
3443 * Right now, this is only the second place pull_task()
3444 * is called, so we can safely collect pull_task()
3445 * stats here rather than inside pull_task().
3447 schedstat_inc(sd, lb_gained[idle]);
3451 p = iterator->next(iterator->arg);
3458 * move_one_task tries to move exactly one task from busiest to this_rq, as
3459 * part of active balancing operations within "domain".
3460 * Returns 1 if successful and 0 otherwise.
3462 * Called with both runqueues locked.
3464 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3465 struct sched_domain *sd, enum cpu_idle_type idle)
3467 const struct sched_class *class;
3469 for_each_class(class) {
3470 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3476 /********** Helpers for find_busiest_group ************************/
3478 * sd_lb_stats - Structure to store the statistics of a sched_domain
3479 * during load balancing.
3481 struct sd_lb_stats {
3482 struct sched_group *busiest; /* Busiest group in this sd */
3483 struct sched_group *this; /* Local group in this sd */
3484 unsigned long total_load; /* Total load of all groups in sd */
3485 unsigned long total_pwr; /* Total power of all groups in sd */
3486 unsigned long avg_load; /* Average load across all groups in sd */
3488 /** Statistics of this group */
3489 unsigned long this_load;
3490 unsigned long this_load_per_task;
3491 unsigned long this_nr_running;
3493 /* Statistics of the busiest group */
3494 unsigned long max_load;
3495 unsigned long busiest_load_per_task;
3496 unsigned long busiest_nr_running;
3498 int group_imb; /* Is there imbalance in this sd */
3499 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3500 int power_savings_balance; /* Is powersave balance needed for this sd */
3501 struct sched_group *group_min; /* Least loaded group in sd */
3502 struct sched_group *group_leader; /* Group which relieves group_min */
3503 unsigned long min_load_per_task; /* load_per_task in group_min */
3504 unsigned long leader_nr_running; /* Nr running of group_leader */
3505 unsigned long min_nr_running; /* Nr running of group_min */
3510 * sg_lb_stats - stats of a sched_group required for load_balancing
3512 struct sg_lb_stats {
3513 unsigned long avg_load; /*Avg load across the CPUs of the group */
3514 unsigned long group_load; /* Total load over the CPUs of the group */
3515 unsigned long sum_nr_running; /* Nr tasks running in the group */
3516 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3517 unsigned long group_capacity;
3518 int group_imb; /* Is there an imbalance in the group ? */
3522 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3523 * @group: The group whose first cpu is to be returned.
3525 static inline unsigned int group_first_cpu(struct sched_group *group)
3527 return cpumask_first(sched_group_cpus(group));
3531 * get_sd_load_idx - Obtain the load index for a given sched domain.
3532 * @sd: The sched_domain whose load_idx is to be obtained.
3533 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3535 static inline int get_sd_load_idx(struct sched_domain *sd,
3536 enum cpu_idle_type idle)
3542 load_idx = sd->busy_idx;
3545 case CPU_NEWLY_IDLE:
3546 load_idx = sd->newidle_idx;
3549 load_idx = sd->idle_idx;
3557 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3559 * init_sd_power_savings_stats - Initialize power savings statistics for
3560 * the given sched_domain, during load balancing.
3562 * @sd: Sched domain whose power-savings statistics are to be initialized.
3563 * @sds: Variable containing the statistics for sd.
3564 * @idle: Idle status of the CPU at which we're performing load-balancing.
3566 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3567 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3570 * Busy processors will not participate in power savings
3573 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3574 sds->power_savings_balance = 0;
3576 sds->power_savings_balance = 1;
3577 sds->min_nr_running = ULONG_MAX;
3578 sds->leader_nr_running = 0;
3583 * update_sd_power_savings_stats - Update the power saving stats for a
3584 * sched_domain while performing load balancing.
3586 * @group: sched_group belonging to the sched_domain under consideration.
3587 * @sds: Variable containing the statistics of the sched_domain
3588 * @local_group: Does group contain the CPU for which we're performing
3590 * @sgs: Variable containing the statistics of the group.
3592 static inline void update_sd_power_savings_stats(struct sched_group *group,
3593 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3596 if (!sds->power_savings_balance)
3600 * If the local group is idle or completely loaded
3601 * no need to do power savings balance at this domain
3603 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3604 !sds->this_nr_running))
3605 sds->power_savings_balance = 0;
3608 * If a group is already running at full capacity or idle,
3609 * don't include that group in power savings calculations
3611 if (!sds->power_savings_balance ||
3612 sgs->sum_nr_running >= sgs->group_capacity ||
3613 !sgs->sum_nr_running)
3617 * Calculate the group which has the least non-idle load.
3618 * This is the group from where we need to pick up the load
3621 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3622 (sgs->sum_nr_running == sds->min_nr_running &&
3623 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3624 sds->group_min = group;
3625 sds->min_nr_running = sgs->sum_nr_running;
3626 sds->min_load_per_task = sgs->sum_weighted_load /
3627 sgs->sum_nr_running;
3631 * Calculate the group which is almost near its
3632 * capacity but still has some space to pick up some load
3633 * from other group and save more power
3635 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3638 if (sgs->sum_nr_running > sds->leader_nr_running ||
3639 (sgs->sum_nr_running == sds->leader_nr_running &&
3640 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3641 sds->group_leader = group;
3642 sds->leader_nr_running = sgs->sum_nr_running;
3647 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3648 * @sds: Variable containing the statistics of the sched_domain
3649 * under consideration.
3650 * @this_cpu: Cpu at which we're currently performing load-balancing.
3651 * @imbalance: Variable to store the imbalance.
3654 * Check if we have potential to perform some power-savings balance.
3655 * If yes, set the busiest group to be the least loaded group in the
3656 * sched_domain, so that it's CPUs can be put to idle.
3658 * Returns 1 if there is potential to perform power-savings balance.
3661 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3662 int this_cpu, unsigned long *imbalance)
3664 if (!sds->power_savings_balance)
3667 if (sds->this != sds->group_leader ||
3668 sds->group_leader == sds->group_min)
3671 *imbalance = sds->min_load_per_task;
3672 sds->busiest = sds->group_min;
3674 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3675 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3676 group_first_cpu(sds->group_leader);
3682 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3683 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3684 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3689 static inline void update_sd_power_savings_stats(struct sched_group *group,
3690 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3695 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3696 int this_cpu, unsigned long *imbalance)
3700 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3704 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3705 * @group: sched_group whose statistics are to be updated.
3706 * @this_cpu: Cpu for which load balance is currently performed.
3707 * @idle: Idle status of this_cpu
3708 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3709 * @sd_idle: Idle status of the sched_domain containing group.
3710 * @local_group: Does group contain this_cpu.
3711 * @cpus: Set of cpus considered for load balancing.
3712 * @balance: Should we balance.
3713 * @sgs: variable to hold the statistics for this group.
3715 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3716 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3717 int local_group, const struct cpumask *cpus,
3718 int *balance, struct sg_lb_stats *sgs)
3720 unsigned long load, max_cpu_load, min_cpu_load;
3722 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3723 unsigned long sum_avg_load_per_task;
3724 unsigned long avg_load_per_task;
3727 balance_cpu = group_first_cpu(group);
3729 /* Tally up the load of all CPUs in the group */
3730 sum_avg_load_per_task = avg_load_per_task = 0;
3732 min_cpu_load = ~0UL;
3734 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3735 struct rq *rq = cpu_rq(i);
3737 if (*sd_idle && rq->nr_running)
3740 /* Bias balancing toward cpus of our domain */
3742 if (idle_cpu(i) && !first_idle_cpu) {
3747 load = target_load(i, load_idx);
3749 load = source_load(i, load_idx);
3750 if (load > max_cpu_load)
3751 max_cpu_load = load;
3752 if (min_cpu_load > load)
3753 min_cpu_load = load;
3756 sgs->group_load += load;
3757 sgs->sum_nr_running += rq->nr_running;
3758 sgs->sum_weighted_load += weighted_cpuload(i);
3760 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3764 * First idle cpu or the first cpu(busiest) in this sched group
3765 * is eligible for doing load balancing at this and above
3766 * domains. In the newly idle case, we will allow all the cpu's
3767 * to do the newly idle load balance.
3769 if (idle != CPU_NEWLY_IDLE && local_group &&
3770 balance_cpu != this_cpu && balance) {
3775 /* Adjust by relative CPU power of the group */
3776 sgs->avg_load = sg_div_cpu_power(group,
3777 sgs->group_load * SCHED_LOAD_SCALE);
3781 * Consider the group unbalanced when the imbalance is larger
3782 * than the average weight of two tasks.
3784 * APZ: with cgroup the avg task weight can vary wildly and
3785 * might not be a suitable number - should we keep a
3786 * normalized nr_running number somewhere that negates
3789 avg_load_per_task = sg_div_cpu_power(group,
3790 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3792 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3795 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3800 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3801 * @sd: sched_domain whose statistics are to be updated.
3802 * @this_cpu: Cpu for which load balance is currently performed.
3803 * @idle: Idle status of this_cpu
3804 * @sd_idle: Idle status of the sched_domain containing group.
3805 * @cpus: Set of cpus considered for load balancing.
3806 * @balance: Should we balance.
3807 * @sds: variable to hold the statistics for this sched_domain.
3809 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3810 enum cpu_idle_type idle, int *sd_idle,
3811 const struct cpumask *cpus, int *balance,
3812 struct sd_lb_stats *sds)
3814 struct sched_group *group = sd->groups;
3815 struct sg_lb_stats sgs;
3818 init_sd_power_savings_stats(sd, sds, idle);
3819 load_idx = get_sd_load_idx(sd, idle);
3824 local_group = cpumask_test_cpu(this_cpu,
3825 sched_group_cpus(group));
3826 memset(&sgs, 0, sizeof(sgs));
3827 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3828 local_group, cpus, balance, &sgs);
3830 if (local_group && balance && !(*balance))
3833 sds->total_load += sgs.group_load;
3834 sds->total_pwr += group->__cpu_power;
3837 sds->this_load = sgs.avg_load;
3839 sds->this_nr_running = sgs.sum_nr_running;
3840 sds->this_load_per_task = sgs.sum_weighted_load;
3841 } else if (sgs.avg_load > sds->max_load &&
3842 (sgs.sum_nr_running > sgs.group_capacity ||
3844 sds->max_load = sgs.avg_load;
3845 sds->busiest = group;
3846 sds->busiest_nr_running = sgs.sum_nr_running;
3847 sds->busiest_load_per_task = sgs.sum_weighted_load;
3848 sds->group_imb = sgs.group_imb;
3851 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3852 group = group->next;
3853 } while (group != sd->groups);
3858 * fix_small_imbalance - Calculate the minor imbalance that exists
3859 * amongst the groups of a sched_domain, during
3861 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3862 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3863 * @imbalance: Variable to store the imbalance.
3865 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3866 int this_cpu, unsigned long *imbalance)
3868 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3869 unsigned int imbn = 2;
3871 if (sds->this_nr_running) {
3872 sds->this_load_per_task /= sds->this_nr_running;
3873 if (sds->busiest_load_per_task >
3874 sds->this_load_per_task)
3877 sds->this_load_per_task =
3878 cpu_avg_load_per_task(this_cpu);
3880 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3881 sds->busiest_load_per_task * imbn) {
3882 *imbalance = sds->busiest_load_per_task;
3887 * OK, we don't have enough imbalance to justify moving tasks,
3888 * however we may be able to increase total CPU power used by
3892 pwr_now += sds->busiest->__cpu_power *
3893 min(sds->busiest_load_per_task, sds->max_load);
3894 pwr_now += sds->this->__cpu_power *
3895 min(sds->this_load_per_task, sds->this_load);
3896 pwr_now /= SCHED_LOAD_SCALE;
3898 /* Amount of load we'd subtract */
3899 tmp = sg_div_cpu_power(sds->busiest,
3900 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3901 if (sds->max_load > tmp)
3902 pwr_move += sds->busiest->__cpu_power *
3903 min(sds->busiest_load_per_task, sds->max_load - tmp);
3905 /* Amount of load we'd add */
3906 if (sds->max_load * sds->busiest->__cpu_power <
3907 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3908 tmp = sg_div_cpu_power(sds->this,
3909 sds->max_load * sds->busiest->__cpu_power);
3911 tmp = sg_div_cpu_power(sds->this,
3912 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3913 pwr_move += sds->this->__cpu_power *
3914 min(sds->this_load_per_task, sds->this_load + tmp);
3915 pwr_move /= SCHED_LOAD_SCALE;
3917 /* Move if we gain throughput */
3918 if (pwr_move > pwr_now)
3919 *imbalance = sds->busiest_load_per_task;
3923 * calculate_imbalance - Calculate the amount of imbalance present within the
3924 * groups of a given sched_domain during load balance.
3925 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3926 * @this_cpu: Cpu for which currently load balance is being performed.
3927 * @imbalance: The variable to store the imbalance.
3929 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3930 unsigned long *imbalance)
3932 unsigned long max_pull;
3934 * In the presence of smp nice balancing, certain scenarios can have
3935 * max load less than avg load(as we skip the groups at or below
3936 * its cpu_power, while calculating max_load..)
3938 if (sds->max_load < sds->avg_load) {
3940 return fix_small_imbalance(sds, this_cpu, imbalance);
3943 /* Don't want to pull so many tasks that a group would go idle */
3944 max_pull = min(sds->max_load - sds->avg_load,
3945 sds->max_load - sds->busiest_load_per_task);
3947 /* How much load to actually move to equalise the imbalance */
3948 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3949 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3953 * if *imbalance is less than the average load per runnable task
3954 * there is no gaurantee that any tasks will be moved so we'll have
3955 * a think about bumping its value to force at least one task to be
3958 if (*imbalance < sds->busiest_load_per_task)
3959 return fix_small_imbalance(sds, this_cpu, imbalance);
3962 /******* find_busiest_group() helpers end here *********************/
3965 * find_busiest_group - Returns the busiest group within the sched_domain
3966 * if there is an imbalance. If there isn't an imbalance, and
3967 * the user has opted for power-savings, it returns a group whose
3968 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3969 * such a group exists.
3971 * Also calculates the amount of weighted load which should be moved
3972 * to restore balance.
3974 * @sd: The sched_domain whose busiest group is to be returned.
3975 * @this_cpu: The cpu for which load balancing is currently being performed.
3976 * @imbalance: Variable which stores amount of weighted load which should
3977 * be moved to restore balance/put a group to idle.
3978 * @idle: The idle status of this_cpu.
3979 * @sd_idle: The idleness of sd
3980 * @cpus: The set of CPUs under consideration for load-balancing.
3981 * @balance: Pointer to a variable indicating if this_cpu
3982 * is the appropriate cpu to perform load balancing at this_level.
3984 * Returns: - the busiest group if imbalance exists.
3985 * - If no imbalance and user has opted for power-savings balance,
3986 * return the least loaded group whose CPUs can be
3987 * put to idle by rebalancing its tasks onto our group.
3989 static struct sched_group *
3990 find_busiest_group(struct sched_domain *sd, int this_cpu,
3991 unsigned long *imbalance, enum cpu_idle_type idle,
3992 int *sd_idle, const struct cpumask *cpus, int *balance)
3994 struct sd_lb_stats sds;
3996 memset(&sds, 0, sizeof(sds));
3999 * Compute the various statistics relavent for load balancing at
4002 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4005 /* Cases where imbalance does not exist from POV of this_cpu */
4006 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4008 * 2) There is no busy sibling group to pull from.
4009 * 3) This group is the busiest group.
4010 * 4) This group is more busy than the avg busieness at this
4012 * 5) The imbalance is within the specified limit.
4013 * 6) Any rebalance would lead to ping-pong
4015 if (balance && !(*balance))
4018 if (!sds.busiest || sds.busiest_nr_running == 0)
4021 if (sds.this_load >= sds.max_load)
4024 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4026 if (sds.this_load >= sds.avg_load)
4029 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4032 sds.busiest_load_per_task /= sds.busiest_nr_running;
4034 sds.busiest_load_per_task =
4035 min(sds.busiest_load_per_task, sds.avg_load);
4038 * We're trying to get all the cpus to the average_load, so we don't
4039 * want to push ourselves above the average load, nor do we wish to
4040 * reduce the max loaded cpu below the average load, as either of these
4041 * actions would just result in more rebalancing later, and ping-pong
4042 * tasks around. Thus we look for the minimum possible imbalance.
4043 * Negative imbalances (*we* are more loaded than anyone else) will
4044 * be counted as no imbalance for these purposes -- we can't fix that
4045 * by pulling tasks to us. Be careful of negative numbers as they'll
4046 * appear as very large values with unsigned longs.
4048 if (sds.max_load <= sds.busiest_load_per_task)
4051 /* Looks like there is an imbalance. Compute it */
4052 calculate_imbalance(&sds, this_cpu, imbalance);
4057 * There is no obvious imbalance. But check if we can do some balancing
4060 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4068 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4071 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4072 unsigned long imbalance, const struct cpumask *cpus)
4074 struct rq *busiest = NULL, *rq;
4075 unsigned long max_load = 0;
4078 for_each_cpu(i, sched_group_cpus(group)) {
4081 if (!cpumask_test_cpu(i, cpus))
4085 wl = weighted_cpuload(i);
4087 if (rq->nr_running == 1 && wl > imbalance)
4090 if (wl > max_load) {
4100 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4101 * so long as it is large enough.
4103 #define MAX_PINNED_INTERVAL 512
4105 /* Working cpumask for load_balance and load_balance_newidle. */
4106 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4109 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4110 * tasks if there is an imbalance.
4112 static int load_balance(int this_cpu, struct rq *this_rq,
4113 struct sched_domain *sd, enum cpu_idle_type idle,
4116 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4117 struct sched_group *group;
4118 unsigned long imbalance;
4120 unsigned long flags;
4121 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4123 cpumask_setall(cpus);
4126 * When power savings policy is enabled for the parent domain, idle
4127 * sibling can pick up load irrespective of busy siblings. In this case,
4128 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4129 * portraying it as CPU_NOT_IDLE.
4131 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4132 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4135 schedstat_inc(sd, lb_count[idle]);
4139 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4146 schedstat_inc(sd, lb_nobusyg[idle]);
4150 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4152 schedstat_inc(sd, lb_nobusyq[idle]);
4156 BUG_ON(busiest == this_rq);
4158 schedstat_add(sd, lb_imbalance[idle], imbalance);
4161 if (busiest->nr_running > 1) {
4163 * Attempt to move tasks. If find_busiest_group has found
4164 * an imbalance but busiest->nr_running <= 1, the group is
4165 * still unbalanced. ld_moved simply stays zero, so it is
4166 * correctly treated as an imbalance.
4168 local_irq_save(flags);
4169 double_rq_lock(this_rq, busiest);
4170 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4171 imbalance, sd, idle, &all_pinned);
4172 double_rq_unlock(this_rq, busiest);
4173 local_irq_restore(flags);
4176 * some other cpu did the load balance for us.
4178 if (ld_moved && this_cpu != smp_processor_id())
4179 resched_cpu(this_cpu);
4181 /* All tasks on this runqueue were pinned by CPU affinity */
4182 if (unlikely(all_pinned)) {
4183 cpumask_clear_cpu(cpu_of(busiest), cpus);
4184 if (!cpumask_empty(cpus))
4191 schedstat_inc(sd, lb_failed[idle]);
4192 sd->nr_balance_failed++;
4194 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4196 spin_lock_irqsave(&busiest->lock, flags);
4198 /* don't kick the migration_thread, if the curr
4199 * task on busiest cpu can't be moved to this_cpu
4201 if (!cpumask_test_cpu(this_cpu,
4202 &busiest->curr->cpus_allowed)) {
4203 spin_unlock_irqrestore(&busiest->lock, flags);
4205 goto out_one_pinned;
4208 if (!busiest->active_balance) {
4209 busiest->active_balance = 1;
4210 busiest->push_cpu = this_cpu;
4213 spin_unlock_irqrestore(&busiest->lock, flags);
4215 wake_up_process(busiest->migration_thread);
4218 * We've kicked active balancing, reset the failure
4221 sd->nr_balance_failed = sd->cache_nice_tries+1;
4224 sd->nr_balance_failed = 0;
4226 if (likely(!active_balance)) {
4227 /* We were unbalanced, so reset the balancing interval */
4228 sd->balance_interval = sd->min_interval;
4231 * If we've begun active balancing, start to back off. This
4232 * case may not be covered by the all_pinned logic if there
4233 * is only 1 task on the busy runqueue (because we don't call
4236 if (sd->balance_interval < sd->max_interval)
4237 sd->balance_interval *= 2;
4240 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4241 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4247 schedstat_inc(sd, lb_balanced[idle]);
4249 sd->nr_balance_failed = 0;
4252 /* tune up the balancing interval */
4253 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4254 (sd->balance_interval < sd->max_interval))
4255 sd->balance_interval *= 2;
4257 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4258 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4269 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4270 * tasks if there is an imbalance.
4272 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4273 * this_rq is locked.
4276 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4278 struct sched_group *group;
4279 struct rq *busiest = NULL;
4280 unsigned long imbalance;
4284 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4286 cpumask_setall(cpus);
4289 * When power savings policy is enabled for the parent domain, idle
4290 * sibling can pick up load irrespective of busy siblings. In this case,
4291 * let the state of idle sibling percolate up as IDLE, instead of
4292 * portraying it as CPU_NOT_IDLE.
4294 if (sd->flags & SD_SHARE_CPUPOWER &&
4295 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4298 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4300 update_shares_locked(this_rq, sd);
4301 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4302 &sd_idle, cpus, NULL);
4304 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4308 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4310 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4314 BUG_ON(busiest == this_rq);
4316 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4319 if (busiest->nr_running > 1) {
4320 /* Attempt to move tasks */
4321 double_lock_balance(this_rq, busiest);
4322 /* this_rq->clock is already updated */
4323 update_rq_clock(busiest);
4324 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4325 imbalance, sd, CPU_NEWLY_IDLE,
4327 double_unlock_balance(this_rq, busiest);
4329 if (unlikely(all_pinned)) {
4330 cpumask_clear_cpu(cpu_of(busiest), cpus);
4331 if (!cpumask_empty(cpus))
4337 int active_balance = 0;
4339 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4340 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4341 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4344 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4347 if (sd->nr_balance_failed++ < 2)
4351 * The only task running in a non-idle cpu can be moved to this
4352 * cpu in an attempt to completely freeup the other CPU
4353 * package. The same method used to move task in load_balance()
4354 * have been extended for load_balance_newidle() to speedup
4355 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4357 * The package power saving logic comes from
4358 * find_busiest_group(). If there are no imbalance, then
4359 * f_b_g() will return NULL. However when sched_mc={1,2} then
4360 * f_b_g() will select a group from which a running task may be
4361 * pulled to this cpu in order to make the other package idle.
4362 * If there is no opportunity to make a package idle and if
4363 * there are no imbalance, then f_b_g() will return NULL and no
4364 * action will be taken in load_balance_newidle().
4366 * Under normal task pull operation due to imbalance, there
4367 * will be more than one task in the source run queue and
4368 * move_tasks() will succeed. ld_moved will be true and this
4369 * active balance code will not be triggered.
4372 /* Lock busiest in correct order while this_rq is held */
4373 double_lock_balance(this_rq, busiest);
4376 * don't kick the migration_thread, if the curr
4377 * task on busiest cpu can't be moved to this_cpu
4379 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4380 double_unlock_balance(this_rq, busiest);
4385 if (!busiest->active_balance) {
4386 busiest->active_balance = 1;
4387 busiest->push_cpu = this_cpu;
4391 double_unlock_balance(this_rq, busiest);
4393 * Should not call ttwu while holding a rq->lock
4395 spin_unlock(&this_rq->lock);
4397 wake_up_process(busiest->migration_thread);
4398 spin_lock(&this_rq->lock);
4401 sd->nr_balance_failed = 0;
4403 update_shares_locked(this_rq, sd);
4407 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4408 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4409 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4411 sd->nr_balance_failed = 0;
4417 * idle_balance is called by schedule() if this_cpu is about to become
4418 * idle. Attempts to pull tasks from other CPUs.
4420 static void idle_balance(int this_cpu, struct rq *this_rq)
4422 struct sched_domain *sd;
4423 int pulled_task = 0;
4424 unsigned long next_balance = jiffies + HZ;
4426 for_each_domain(this_cpu, sd) {
4427 unsigned long interval;
4429 if (!(sd->flags & SD_LOAD_BALANCE))
4432 if (sd->flags & SD_BALANCE_NEWIDLE)
4433 /* If we've pulled tasks over stop searching: */
4434 pulled_task = load_balance_newidle(this_cpu, this_rq,
4437 interval = msecs_to_jiffies(sd->balance_interval);
4438 if (time_after(next_balance, sd->last_balance + interval))
4439 next_balance = sd->last_balance + interval;
4443 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4445 * We are going idle. next_balance may be set based on
4446 * a busy processor. So reset next_balance.
4448 this_rq->next_balance = next_balance;
4453 * active_load_balance is run by migration threads. It pushes running tasks
4454 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4455 * running on each physical CPU where possible, and avoids physical /
4456 * logical imbalances.
4458 * Called with busiest_rq locked.
4460 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4462 int target_cpu = busiest_rq->push_cpu;
4463 struct sched_domain *sd;
4464 struct rq *target_rq;
4466 /* Is there any task to move? */
4467 if (busiest_rq->nr_running <= 1)
4470 target_rq = cpu_rq(target_cpu);
4473 * This condition is "impossible", if it occurs
4474 * we need to fix it. Originally reported by
4475 * Bjorn Helgaas on a 128-cpu setup.
4477 BUG_ON(busiest_rq == target_rq);
4479 /* move a task from busiest_rq to target_rq */
4480 double_lock_balance(busiest_rq, target_rq);
4481 update_rq_clock(busiest_rq);
4482 update_rq_clock(target_rq);
4484 /* Search for an sd spanning us and the target CPU. */
4485 for_each_domain(target_cpu, sd) {
4486 if ((sd->flags & SD_LOAD_BALANCE) &&
4487 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4492 schedstat_inc(sd, alb_count);
4494 if (move_one_task(target_rq, target_cpu, busiest_rq,
4496 schedstat_inc(sd, alb_pushed);
4498 schedstat_inc(sd, alb_failed);
4500 double_unlock_balance(busiest_rq, target_rq);
4505 atomic_t load_balancer;
4506 cpumask_var_t cpu_mask;
4507 cpumask_var_t ilb_grp_nohz_mask;
4508 } nohz ____cacheline_aligned = {
4509 .load_balancer = ATOMIC_INIT(-1),
4512 int get_nohz_load_balancer(void)
4514 return atomic_read(&nohz.load_balancer);
4517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4519 * lowest_flag_domain - Return lowest sched_domain containing flag.
4520 * @cpu: The cpu whose lowest level of sched domain is to
4522 * @flag: The flag to check for the lowest sched_domain
4523 * for the given cpu.
4525 * Returns the lowest sched_domain of a cpu which contains the given flag.
4527 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4529 struct sched_domain *sd;
4531 for_each_domain(cpu, sd)
4532 if (sd && (sd->flags & flag))
4539 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4540 * @cpu: The cpu whose domains we're iterating over.
4541 * @sd: variable holding the value of the power_savings_sd
4543 * @flag: The flag to filter the sched_domains to be iterated.
4545 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4546 * set, starting from the lowest sched_domain to the highest.
4548 #define for_each_flag_domain(cpu, sd, flag) \
4549 for (sd = lowest_flag_domain(cpu, flag); \
4550 (sd && (sd->flags & flag)); sd = sd->parent)
4553 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4554 * @ilb_group: group to be checked for semi-idleness
4556 * Returns: 1 if the group is semi-idle. 0 otherwise.
4558 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4559 * and atleast one non-idle CPU. This helper function checks if the given
4560 * sched_group is semi-idle or not.
4562 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4564 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4565 sched_group_cpus(ilb_group));
4568 * A sched_group is semi-idle when it has atleast one busy cpu
4569 * and atleast one idle cpu.
4571 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4574 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4580 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4581 * @cpu: The cpu which is nominating a new idle_load_balancer.
4583 * Returns: Returns the id of the idle load balancer if it exists,
4584 * Else, returns >= nr_cpu_ids.
4586 * This algorithm picks the idle load balancer such that it belongs to a
4587 * semi-idle powersavings sched_domain. The idea is to try and avoid
4588 * completely idle packages/cores just for the purpose of idle load balancing
4589 * when there are other idle cpu's which are better suited for that job.
4591 static int find_new_ilb(int cpu)
4593 struct sched_domain *sd;
4594 struct sched_group *ilb_group;
4597 * Have idle load balancer selection from semi-idle packages only
4598 * when power-aware load balancing is enabled
4600 if (!(sched_smt_power_savings || sched_mc_power_savings))
4604 * Optimize for the case when we have no idle CPUs or only one
4605 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4607 if (cpumask_weight(nohz.cpu_mask) < 2)
4610 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4611 ilb_group = sd->groups;
4614 if (is_semi_idle_group(ilb_group))
4615 return cpumask_first(nohz.ilb_grp_nohz_mask);
4617 ilb_group = ilb_group->next;
4619 } while (ilb_group != sd->groups);
4623 return cpumask_first(nohz.cpu_mask);
4625 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4626 static inline int find_new_ilb(int call_cpu)
4628 return cpumask_first(nohz.cpu_mask);
4633 * This routine will try to nominate the ilb (idle load balancing)
4634 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4635 * load balancing on behalf of all those cpus. If all the cpus in the system
4636 * go into this tickless mode, then there will be no ilb owner (as there is
4637 * no need for one) and all the cpus will sleep till the next wakeup event
4640 * For the ilb owner, tick is not stopped. And this tick will be used
4641 * for idle load balancing. ilb owner will still be part of
4644 * While stopping the tick, this cpu will become the ilb owner if there
4645 * is no other owner. And will be the owner till that cpu becomes busy
4646 * or if all cpus in the system stop their ticks at which point
4647 * there is no need for ilb owner.
4649 * When the ilb owner becomes busy, it nominates another owner, during the
4650 * next busy scheduler_tick()
4652 int select_nohz_load_balancer(int stop_tick)
4654 int cpu = smp_processor_id();
4657 cpu_rq(cpu)->in_nohz_recently = 1;
4659 if (!cpu_active(cpu)) {
4660 if (atomic_read(&nohz.load_balancer) != cpu)
4664 * If we are going offline and still the leader,
4667 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4673 cpumask_set_cpu(cpu, nohz.cpu_mask);
4675 /* time for ilb owner also to sleep */
4676 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4677 if (atomic_read(&nohz.load_balancer) == cpu)
4678 atomic_set(&nohz.load_balancer, -1);
4682 if (atomic_read(&nohz.load_balancer) == -1) {
4683 /* make me the ilb owner */
4684 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4686 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4689 if (!(sched_smt_power_savings ||
4690 sched_mc_power_savings))
4693 * Check to see if there is a more power-efficient
4696 new_ilb = find_new_ilb(cpu);
4697 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4698 atomic_set(&nohz.load_balancer, -1);
4699 resched_cpu(new_ilb);
4705 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4708 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4710 if (atomic_read(&nohz.load_balancer) == cpu)
4711 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4718 static DEFINE_SPINLOCK(balancing);
4721 * It checks each scheduling domain to see if it is due to be balanced,
4722 * and initiates a balancing operation if so.
4724 * Balancing parameters are set up in arch_init_sched_domains.
4726 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4729 struct rq *rq = cpu_rq(cpu);
4730 unsigned long interval;
4731 struct sched_domain *sd;
4732 /* Earliest time when we have to do rebalance again */
4733 unsigned long next_balance = jiffies + 60*HZ;
4734 int update_next_balance = 0;
4737 for_each_domain(cpu, sd) {
4738 if (!(sd->flags & SD_LOAD_BALANCE))
4741 interval = sd->balance_interval;
4742 if (idle != CPU_IDLE)
4743 interval *= sd->busy_factor;
4745 /* scale ms to jiffies */
4746 interval = msecs_to_jiffies(interval);
4747 if (unlikely(!interval))
4749 if (interval > HZ*NR_CPUS/10)
4750 interval = HZ*NR_CPUS/10;
4752 need_serialize = sd->flags & SD_SERIALIZE;
4754 if (need_serialize) {
4755 if (!spin_trylock(&balancing))
4759 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4760 if (load_balance(cpu, rq, sd, idle, &balance)) {
4762 * We've pulled tasks over so either we're no
4763 * longer idle, or one of our SMT siblings is
4766 idle = CPU_NOT_IDLE;
4768 sd->last_balance = jiffies;
4771 spin_unlock(&balancing);
4773 if (time_after(next_balance, sd->last_balance + interval)) {
4774 next_balance = sd->last_balance + interval;
4775 update_next_balance = 1;
4779 * Stop the load balance at this level. There is another
4780 * CPU in our sched group which is doing load balancing more
4788 * next_balance will be updated only when there is a need.
4789 * When the cpu is attached to null domain for ex, it will not be
4792 if (likely(update_next_balance))
4793 rq->next_balance = next_balance;
4797 * run_rebalance_domains is triggered when needed from the scheduler tick.
4798 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4799 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4801 static void run_rebalance_domains(struct softirq_action *h)
4803 int this_cpu = smp_processor_id();
4804 struct rq *this_rq = cpu_rq(this_cpu);
4805 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4806 CPU_IDLE : CPU_NOT_IDLE;
4808 rebalance_domains(this_cpu, idle);
4812 * If this cpu is the owner for idle load balancing, then do the
4813 * balancing on behalf of the other idle cpus whose ticks are
4816 if (this_rq->idle_at_tick &&
4817 atomic_read(&nohz.load_balancer) == this_cpu) {
4821 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4822 if (balance_cpu == this_cpu)
4826 * If this cpu gets work to do, stop the load balancing
4827 * work being done for other cpus. Next load
4828 * balancing owner will pick it up.
4833 rebalance_domains(balance_cpu, CPU_IDLE);
4835 rq = cpu_rq(balance_cpu);
4836 if (time_after(this_rq->next_balance, rq->next_balance))
4837 this_rq->next_balance = rq->next_balance;
4843 static inline int on_null_domain(int cpu)
4845 return !rcu_dereference(cpu_rq(cpu)->sd);
4849 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4851 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4852 * idle load balancing owner or decide to stop the periodic load balancing,
4853 * if the whole system is idle.
4855 static inline void trigger_load_balance(struct rq *rq, int cpu)
4859 * If we were in the nohz mode recently and busy at the current
4860 * scheduler tick, then check if we need to nominate new idle
4863 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4864 rq->in_nohz_recently = 0;
4866 if (atomic_read(&nohz.load_balancer) == cpu) {
4867 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4868 atomic_set(&nohz.load_balancer, -1);
4871 if (atomic_read(&nohz.load_balancer) == -1) {
4872 int ilb = find_new_ilb(cpu);
4874 if (ilb < nr_cpu_ids)
4880 * If this cpu is idle and doing idle load balancing for all the
4881 * cpus with ticks stopped, is it time for that to stop?
4883 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4884 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4890 * If this cpu is idle and the idle load balancing is done by
4891 * someone else, then no need raise the SCHED_SOFTIRQ
4893 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4894 cpumask_test_cpu(cpu, nohz.cpu_mask))
4897 /* Don't need to rebalance while attached to NULL domain */
4898 if (time_after_eq(jiffies, rq->next_balance) &&
4899 likely(!on_null_domain(cpu)))
4900 raise_softirq(SCHED_SOFTIRQ);
4903 #else /* CONFIG_SMP */
4906 * on UP we do not need to balance between CPUs:
4908 static inline void idle_balance(int cpu, struct rq *rq)
4914 DEFINE_PER_CPU(struct kernel_stat, kstat);
4916 EXPORT_PER_CPU_SYMBOL(kstat);
4919 * Return any ns on the sched_clock that have not yet been accounted in
4920 * @p in case that task is currently running.
4922 * Called with task_rq_lock() held on @rq.
4924 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4928 if (task_current(rq, p)) {
4929 update_rq_clock(rq);
4930 ns = rq->clock - p->se.exec_start;
4938 unsigned long long task_delta_exec(struct task_struct *p)
4940 unsigned long flags;
4944 rq = task_rq_lock(p, &flags);
4945 ns = do_task_delta_exec(p, rq);
4946 task_rq_unlock(rq, &flags);
4952 * Return accounted runtime for the task.
4953 * In case the task is currently running, return the runtime plus current's
4954 * pending runtime that have not been accounted yet.
4956 unsigned long long task_sched_runtime(struct task_struct *p)
4958 unsigned long flags;
4962 rq = task_rq_lock(p, &flags);
4963 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4964 task_rq_unlock(rq, &flags);
4970 * Return sum_exec_runtime for the thread group.
4971 * In case the task is currently running, return the sum plus current's
4972 * pending runtime that have not been accounted yet.
4974 * Note that the thread group might have other running tasks as well,
4975 * so the return value not includes other pending runtime that other
4976 * running tasks might have.
4978 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4980 struct task_cputime totals;
4981 unsigned long flags;
4985 rq = task_rq_lock(p, &flags);
4986 thread_group_cputime(p, &totals);
4987 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4988 task_rq_unlock(rq, &flags);
4994 * Account user cpu time to a process.
4995 * @p: the process that the cpu time gets accounted to
4996 * @cputime: the cpu time spent in user space since the last update
4997 * @cputime_scaled: cputime scaled by cpu frequency
4999 void account_user_time(struct task_struct *p, cputime_t cputime,
5000 cputime_t cputime_scaled)
5002 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5005 /* Add user time to process. */
5006 p->utime = cputime_add(p->utime, cputime);
5007 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5008 account_group_user_time(p, cputime);
5010 /* Add user time to cpustat. */
5011 tmp = cputime_to_cputime64(cputime);
5012 if (TASK_NICE(p) > 0)
5013 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5015 cpustat->user = cputime64_add(cpustat->user, tmp);
5017 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5018 /* Account for user time used */
5019 acct_update_integrals(p);
5023 * Account guest cpu time to a process.
5024 * @p: the process that the cpu time gets accounted to
5025 * @cputime: the cpu time spent in virtual machine since the last update
5026 * @cputime_scaled: cputime scaled by cpu frequency
5028 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5029 cputime_t cputime_scaled)
5032 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5034 tmp = cputime_to_cputime64(cputime);
5036 /* Add guest time to process. */
5037 p->utime = cputime_add(p->utime, cputime);
5038 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5039 account_group_user_time(p, cputime);
5040 p->gtime = cputime_add(p->gtime, cputime);
5042 /* Add guest time to cpustat. */
5043 cpustat->user = cputime64_add(cpustat->user, tmp);
5044 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5048 * Account system cpu time to a process.
5049 * @p: the process that the cpu time gets accounted to
5050 * @hardirq_offset: the offset to subtract from hardirq_count()
5051 * @cputime: the cpu time spent in kernel space since the last update
5052 * @cputime_scaled: cputime scaled by cpu frequency
5054 void account_system_time(struct task_struct *p, int hardirq_offset,
5055 cputime_t cputime, cputime_t cputime_scaled)
5057 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5060 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5061 account_guest_time(p, cputime, cputime_scaled);
5065 /* Add system time to process. */
5066 p->stime = cputime_add(p->stime, cputime);
5067 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5068 account_group_system_time(p, cputime);
5070 /* Add system time to cpustat. */
5071 tmp = cputime_to_cputime64(cputime);
5072 if (hardirq_count() - hardirq_offset)
5073 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5074 else if (softirq_count())
5075 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5077 cpustat->system = cputime64_add(cpustat->system, tmp);
5079 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5081 /* Account for system time used */
5082 acct_update_integrals(p);
5086 * Account for involuntary wait time.
5087 * @steal: the cpu time spent in involuntary wait
5089 void account_steal_time(cputime_t cputime)
5091 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5092 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5094 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5098 * Account for idle time.
5099 * @cputime: the cpu time spent in idle wait
5101 void account_idle_time(cputime_t cputime)
5103 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5104 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5105 struct rq *rq = this_rq();
5107 if (atomic_read(&rq->nr_iowait) > 0)
5108 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5110 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5113 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5116 * Account a single tick of cpu time.
5117 * @p: the process that the cpu time gets accounted to
5118 * @user_tick: indicates if the tick is a user or a system tick
5120 void account_process_tick(struct task_struct *p, int user_tick)
5122 cputime_t one_jiffy = jiffies_to_cputime(1);
5123 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5124 struct rq *rq = this_rq();
5127 account_user_time(p, one_jiffy, one_jiffy_scaled);
5128 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5129 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5132 account_idle_time(one_jiffy);
5136 * Account multiple ticks of steal time.
5137 * @p: the process from which the cpu time has been stolen
5138 * @ticks: number of stolen ticks
5140 void account_steal_ticks(unsigned long ticks)
5142 account_steal_time(jiffies_to_cputime(ticks));
5146 * Account multiple ticks of idle time.
5147 * @ticks: number of stolen ticks
5149 void account_idle_ticks(unsigned long ticks)
5151 account_idle_time(jiffies_to_cputime(ticks));
5157 * Use precise platform statistics if available:
5159 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5160 cputime_t task_utime(struct task_struct *p)
5165 cputime_t task_stime(struct task_struct *p)
5170 cputime_t task_utime(struct task_struct *p)
5172 clock_t utime = cputime_to_clock_t(p->utime),
5173 total = utime + cputime_to_clock_t(p->stime);
5177 * Use CFS's precise accounting:
5179 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5183 do_div(temp, total);
5185 utime = (clock_t)temp;
5187 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5188 return p->prev_utime;
5191 cputime_t task_stime(struct task_struct *p)
5196 * Use CFS's precise accounting. (we subtract utime from
5197 * the total, to make sure the total observed by userspace
5198 * grows monotonically - apps rely on that):
5200 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5201 cputime_to_clock_t(task_utime(p));
5204 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5206 return p->prev_stime;
5210 inline cputime_t task_gtime(struct task_struct *p)
5216 * This function gets called by the timer code, with HZ frequency.
5217 * We call it with interrupts disabled.
5219 * It also gets called by the fork code, when changing the parent's
5222 void scheduler_tick(void)
5224 int cpu = smp_processor_id();
5225 struct rq *rq = cpu_rq(cpu);
5226 struct task_struct *curr = rq->curr;
5230 spin_lock(&rq->lock);
5231 update_rq_clock(rq);
5232 update_cpu_load(rq);
5233 curr->sched_class->task_tick(rq, curr, 0);
5234 spin_unlock(&rq->lock);
5236 perf_counter_task_tick(curr, cpu);
5239 rq->idle_at_tick = idle_cpu(cpu);
5240 trigger_load_balance(rq, cpu);
5244 notrace unsigned long get_parent_ip(unsigned long addr)
5246 if (in_lock_functions(addr)) {
5247 addr = CALLER_ADDR2;
5248 if (in_lock_functions(addr))
5249 addr = CALLER_ADDR3;
5254 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5255 defined(CONFIG_PREEMPT_TRACER))
5257 void __kprobes add_preempt_count(int val)
5259 #ifdef CONFIG_DEBUG_PREEMPT
5263 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5266 preempt_count() += val;
5267 #ifdef CONFIG_DEBUG_PREEMPT
5269 * Spinlock count overflowing soon?
5271 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5274 if (preempt_count() == val)
5275 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5277 EXPORT_SYMBOL(add_preempt_count);
5279 void __kprobes sub_preempt_count(int val)
5281 #ifdef CONFIG_DEBUG_PREEMPT
5285 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5288 * Is the spinlock portion underflowing?
5290 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5291 !(preempt_count() & PREEMPT_MASK)))
5295 if (preempt_count() == val)
5296 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5297 preempt_count() -= val;
5299 EXPORT_SYMBOL(sub_preempt_count);
5304 * Print scheduling while atomic bug:
5306 static noinline void __schedule_bug(struct task_struct *prev)
5308 struct pt_regs *regs = get_irq_regs();
5310 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5311 prev->comm, prev->pid, preempt_count());
5313 debug_show_held_locks(prev);
5315 if (irqs_disabled())
5316 print_irqtrace_events(prev);
5325 * Various schedule()-time debugging checks and statistics:
5327 static inline void schedule_debug(struct task_struct *prev)
5330 * Test if we are atomic. Since do_exit() needs to call into
5331 * schedule() atomically, we ignore that path for now.
5332 * Otherwise, whine if we are scheduling when we should not be.
5334 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5335 __schedule_bug(prev);
5337 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5339 schedstat_inc(this_rq(), sched_count);
5340 #ifdef CONFIG_SCHEDSTATS
5341 if (unlikely(prev->lock_depth >= 0)) {
5342 schedstat_inc(this_rq(), bkl_count);
5343 schedstat_inc(prev, sched_info.bkl_count);
5348 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5350 if (prev->state == TASK_RUNNING) {
5351 u64 runtime = prev->se.sum_exec_runtime;
5353 runtime -= prev->se.prev_sum_exec_runtime;
5354 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5357 * In order to avoid avg_overlap growing stale when we are
5358 * indeed overlapping and hence not getting put to sleep, grow
5359 * the avg_overlap on preemption.
5361 * We use the average preemption runtime because that
5362 * correlates to the amount of cache footprint a task can
5365 update_avg(&prev->se.avg_overlap, runtime);
5367 prev->sched_class->put_prev_task(rq, prev);
5371 * Pick up the highest-prio task:
5373 static inline struct task_struct *
5374 pick_next_task(struct rq *rq)
5376 const struct sched_class *class;
5377 struct task_struct *p;
5380 * Optimization: we know that if all tasks are in
5381 * the fair class we can call that function directly:
5383 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5384 p = fair_sched_class.pick_next_task(rq);
5389 class = sched_class_highest;
5391 p = class->pick_next_task(rq);
5395 * Will never be NULL as the idle class always
5396 * returns a non-NULL p:
5398 class = class->next;
5403 * schedule() is the main scheduler function.
5405 asmlinkage void __sched schedule(void)
5407 struct task_struct *prev, *next;
5408 unsigned long *switch_count;
5414 cpu = smp_processor_id();
5418 switch_count = &prev->nivcsw;
5420 release_kernel_lock(prev);
5421 need_resched_nonpreemptible:
5423 schedule_debug(prev);
5425 if (sched_feat(HRTICK))
5428 spin_lock_irq(&rq->lock);
5429 update_rq_clock(rq);
5430 clear_tsk_need_resched(prev);
5432 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5433 if (unlikely(signal_pending_state(prev->state, prev)))
5434 prev->state = TASK_RUNNING;
5436 deactivate_task(rq, prev, 1);
5437 switch_count = &prev->nvcsw;
5440 pre_schedule(rq, prev);
5442 if (unlikely(!rq->nr_running))
5443 idle_balance(cpu, rq);
5445 put_prev_task(rq, prev);
5446 next = pick_next_task(rq);
5448 if (likely(prev != next)) {
5449 sched_info_switch(prev, next);
5450 perf_counter_task_sched_out(prev, next, cpu);
5456 context_switch(rq, prev, next); /* unlocks the rq */
5458 * the context switch might have flipped the stack from under
5459 * us, hence refresh the local variables.
5461 cpu = smp_processor_id();
5464 spin_unlock_irq(&rq->lock);
5468 if (unlikely(reacquire_kernel_lock(current) < 0))
5469 goto need_resched_nonpreemptible;
5471 preempt_enable_no_resched();
5475 EXPORT_SYMBOL(schedule);
5479 * Look out! "owner" is an entirely speculative pointer
5480 * access and not reliable.
5482 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5487 if (!sched_feat(OWNER_SPIN))
5490 #ifdef CONFIG_DEBUG_PAGEALLOC
5492 * Need to access the cpu field knowing that
5493 * DEBUG_PAGEALLOC could have unmapped it if
5494 * the mutex owner just released it and exited.
5496 if (probe_kernel_address(&owner->cpu, cpu))
5503 * Even if the access succeeded (likely case),
5504 * the cpu field may no longer be valid.
5506 if (cpu >= nr_cpumask_bits)
5510 * We need to validate that we can do a
5511 * get_cpu() and that we have the percpu area.
5513 if (!cpu_online(cpu))
5520 * Owner changed, break to re-assess state.
5522 if (lock->owner != owner)
5526 * Is that owner really running on that cpu?
5528 if (task_thread_info(rq->curr) != owner || need_resched())
5538 #ifdef CONFIG_PREEMPT
5540 * this is the entry point to schedule() from in-kernel preemption
5541 * off of preempt_enable. Kernel preemptions off return from interrupt
5542 * occur there and call schedule directly.
5544 asmlinkage void __sched preempt_schedule(void)
5546 struct thread_info *ti = current_thread_info();
5549 * If there is a non-zero preempt_count or interrupts are disabled,
5550 * we do not want to preempt the current task. Just return..
5552 if (likely(ti->preempt_count || irqs_disabled()))
5556 add_preempt_count(PREEMPT_ACTIVE);
5558 sub_preempt_count(PREEMPT_ACTIVE);
5561 * Check again in case we missed a preemption opportunity
5562 * between schedule and now.
5565 } while (need_resched());
5567 EXPORT_SYMBOL(preempt_schedule);
5570 * this is the entry point to schedule() from kernel preemption
5571 * off of irq context.
5572 * Note, that this is called and return with irqs disabled. This will
5573 * protect us against recursive calling from irq.
5575 asmlinkage void __sched preempt_schedule_irq(void)
5577 struct thread_info *ti = current_thread_info();
5579 /* Catch callers which need to be fixed */
5580 BUG_ON(ti->preempt_count || !irqs_disabled());
5583 add_preempt_count(PREEMPT_ACTIVE);
5586 local_irq_disable();
5587 sub_preempt_count(PREEMPT_ACTIVE);
5590 * Check again in case we missed a preemption opportunity
5591 * between schedule and now.
5594 } while (need_resched());
5597 #endif /* CONFIG_PREEMPT */
5599 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5602 return try_to_wake_up(curr->private, mode, sync);
5604 EXPORT_SYMBOL(default_wake_function);
5607 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5608 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5609 * number) then we wake all the non-exclusive tasks and one exclusive task.
5611 * There are circumstances in which we can try to wake a task which has already
5612 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5613 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5615 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5616 int nr_exclusive, int sync, void *key)
5618 wait_queue_t *curr, *next;
5620 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5621 unsigned flags = curr->flags;
5623 if (curr->func(curr, mode, sync, key) &&
5624 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5630 * __wake_up - wake up threads blocked on a waitqueue.
5632 * @mode: which threads
5633 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5634 * @key: is directly passed to the wakeup function
5636 * It may be assumed that this function implies a write memory barrier before
5637 * changing the task state if and only if any tasks are woken up.
5639 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5640 int nr_exclusive, void *key)
5642 unsigned long flags;
5644 spin_lock_irqsave(&q->lock, flags);
5645 __wake_up_common(q, mode, nr_exclusive, 0, key);
5646 spin_unlock_irqrestore(&q->lock, flags);
5648 EXPORT_SYMBOL(__wake_up);
5651 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5653 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5655 __wake_up_common(q, mode, 1, 0, NULL);
5658 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5660 __wake_up_common(q, mode, 1, 0, key);
5664 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5666 * @mode: which threads
5667 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5668 * @key: opaque value to be passed to wakeup targets
5670 * The sync wakeup differs that the waker knows that it will schedule
5671 * away soon, so while the target thread will be woken up, it will not
5672 * be migrated to another CPU - ie. the two threads are 'synchronized'
5673 * with each other. This can prevent needless bouncing between CPUs.
5675 * On UP it can prevent extra preemption.
5677 * It may be assumed that this function implies a write memory barrier before
5678 * changing the task state if and only if any tasks are woken up.
5680 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5681 int nr_exclusive, void *key)
5683 unsigned long flags;
5689 if (unlikely(!nr_exclusive))
5692 spin_lock_irqsave(&q->lock, flags);
5693 __wake_up_common(q, mode, nr_exclusive, sync, key);
5694 spin_unlock_irqrestore(&q->lock, flags);
5696 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5699 * __wake_up_sync - see __wake_up_sync_key()
5701 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5703 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5705 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5708 * complete: - signals a single thread waiting on this completion
5709 * @x: holds the state of this particular completion
5711 * This will wake up a single thread waiting on this completion. Threads will be
5712 * awakened in the same order in which they were queued.
5714 * See also complete_all(), wait_for_completion() and related routines.
5716 * It may be assumed that this function implies a write memory barrier before
5717 * changing the task state if and only if any tasks are woken up.
5719 void complete(struct completion *x)
5721 unsigned long flags;
5723 spin_lock_irqsave(&x->wait.lock, flags);
5725 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5726 spin_unlock_irqrestore(&x->wait.lock, flags);
5728 EXPORT_SYMBOL(complete);
5731 * complete_all: - signals all threads waiting on this completion
5732 * @x: holds the state of this particular completion
5734 * This will wake up all threads waiting on this particular completion event.
5736 * It may be assumed that this function implies a write memory barrier before
5737 * changing the task state if and only if any tasks are woken up.
5739 void complete_all(struct completion *x)
5741 unsigned long flags;
5743 spin_lock_irqsave(&x->wait.lock, flags);
5744 x->done += UINT_MAX/2;
5745 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5746 spin_unlock_irqrestore(&x->wait.lock, flags);
5748 EXPORT_SYMBOL(complete_all);
5750 static inline long __sched
5751 do_wait_for_common(struct completion *x, long timeout, int state)
5754 DECLARE_WAITQUEUE(wait, current);
5756 wait.flags |= WQ_FLAG_EXCLUSIVE;
5757 __add_wait_queue_tail(&x->wait, &wait);
5759 if (signal_pending_state(state, current)) {
5760 timeout = -ERESTARTSYS;
5763 __set_current_state(state);
5764 spin_unlock_irq(&x->wait.lock);
5765 timeout = schedule_timeout(timeout);
5766 spin_lock_irq(&x->wait.lock);
5767 } while (!x->done && timeout);
5768 __remove_wait_queue(&x->wait, &wait);
5773 return timeout ?: 1;
5777 wait_for_common(struct completion *x, long timeout, int state)
5781 spin_lock_irq(&x->wait.lock);
5782 timeout = do_wait_for_common(x, timeout, state);
5783 spin_unlock_irq(&x->wait.lock);
5788 * wait_for_completion: - waits for completion of a task
5789 * @x: holds the state of this particular completion
5791 * This waits to be signaled for completion of a specific task. It is NOT
5792 * interruptible and there is no timeout.
5794 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5795 * and interrupt capability. Also see complete().
5797 void __sched wait_for_completion(struct completion *x)
5799 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5801 EXPORT_SYMBOL(wait_for_completion);
5804 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5805 * @x: holds the state of this particular completion
5806 * @timeout: timeout value in jiffies
5808 * This waits for either a completion of a specific task to be signaled or for a
5809 * specified timeout to expire. The timeout is in jiffies. It is not
5812 unsigned long __sched
5813 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5815 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5817 EXPORT_SYMBOL(wait_for_completion_timeout);
5820 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5821 * @x: holds the state of this particular completion
5823 * This waits for completion of a specific task to be signaled. It is
5826 int __sched wait_for_completion_interruptible(struct completion *x)
5828 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5829 if (t == -ERESTARTSYS)
5833 EXPORT_SYMBOL(wait_for_completion_interruptible);
5836 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5837 * @x: holds the state of this particular completion
5838 * @timeout: timeout value in jiffies
5840 * This waits for either a completion of a specific task to be signaled or for a
5841 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5843 unsigned long __sched
5844 wait_for_completion_interruptible_timeout(struct completion *x,
5845 unsigned long timeout)
5847 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5849 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5852 * wait_for_completion_killable: - waits for completion of a task (killable)
5853 * @x: holds the state of this particular completion
5855 * This waits to be signaled for completion of a specific task. It can be
5856 * interrupted by a kill signal.
5858 int __sched wait_for_completion_killable(struct completion *x)
5860 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5861 if (t == -ERESTARTSYS)
5865 EXPORT_SYMBOL(wait_for_completion_killable);
5868 * try_wait_for_completion - try to decrement a completion without blocking
5869 * @x: completion structure
5871 * Returns: 0 if a decrement cannot be done without blocking
5872 * 1 if a decrement succeeded.
5874 * If a completion is being used as a counting completion,
5875 * attempt to decrement the counter without blocking. This
5876 * enables us to avoid waiting if the resource the completion
5877 * is protecting is not available.
5879 bool try_wait_for_completion(struct completion *x)
5883 spin_lock_irq(&x->wait.lock);
5888 spin_unlock_irq(&x->wait.lock);
5891 EXPORT_SYMBOL(try_wait_for_completion);
5894 * completion_done - Test to see if a completion has any waiters
5895 * @x: completion structure
5897 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5898 * 1 if there are no waiters.
5901 bool completion_done(struct completion *x)
5905 spin_lock_irq(&x->wait.lock);
5908 spin_unlock_irq(&x->wait.lock);
5911 EXPORT_SYMBOL(completion_done);
5914 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5916 unsigned long flags;
5919 init_waitqueue_entry(&wait, current);
5921 __set_current_state(state);
5923 spin_lock_irqsave(&q->lock, flags);
5924 __add_wait_queue(q, &wait);
5925 spin_unlock(&q->lock);
5926 timeout = schedule_timeout(timeout);
5927 spin_lock_irq(&q->lock);
5928 __remove_wait_queue(q, &wait);
5929 spin_unlock_irqrestore(&q->lock, flags);
5934 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5936 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5938 EXPORT_SYMBOL(interruptible_sleep_on);
5941 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5943 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5945 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5947 void __sched sleep_on(wait_queue_head_t *q)
5949 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5951 EXPORT_SYMBOL(sleep_on);
5953 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5955 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5957 EXPORT_SYMBOL(sleep_on_timeout);
5959 #ifdef CONFIG_RT_MUTEXES
5962 * rt_mutex_setprio - set the current priority of a task
5964 * @prio: prio value (kernel-internal form)
5966 * This function changes the 'effective' priority of a task. It does
5967 * not touch ->normal_prio like __setscheduler().
5969 * Used by the rt_mutex code to implement priority inheritance logic.
5971 void rt_mutex_setprio(struct task_struct *p, int prio)
5973 unsigned long flags;
5974 int oldprio, on_rq, running;
5976 const struct sched_class *prev_class = p->sched_class;
5978 BUG_ON(prio < 0 || prio > MAX_PRIO);
5980 rq = task_rq_lock(p, &flags);
5981 update_rq_clock(rq);
5984 on_rq = p->se.on_rq;
5985 running = task_current(rq, p);
5987 dequeue_task(rq, p, 0);
5989 p->sched_class->put_prev_task(rq, p);
5992 p->sched_class = &rt_sched_class;
5994 p->sched_class = &fair_sched_class;
5999 p->sched_class->set_curr_task(rq);
6001 enqueue_task(rq, p, 0);
6003 check_class_changed(rq, p, prev_class, oldprio, running);
6005 task_rq_unlock(rq, &flags);
6010 void set_user_nice(struct task_struct *p, long nice)
6012 int old_prio, delta, on_rq;
6013 unsigned long flags;
6016 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6019 * We have to be careful, if called from sys_setpriority(),
6020 * the task might be in the middle of scheduling on another CPU.
6022 rq = task_rq_lock(p, &flags);
6023 update_rq_clock(rq);
6025 * The RT priorities are set via sched_setscheduler(), but we still
6026 * allow the 'normal' nice value to be set - but as expected
6027 * it wont have any effect on scheduling until the task is
6028 * SCHED_FIFO/SCHED_RR:
6030 if (task_has_rt_policy(p)) {
6031 p->static_prio = NICE_TO_PRIO(nice);
6034 on_rq = p->se.on_rq;
6036 dequeue_task(rq, p, 0);
6038 p->static_prio = NICE_TO_PRIO(nice);
6041 p->prio = effective_prio(p);
6042 delta = p->prio - old_prio;
6045 enqueue_task(rq, p, 0);
6047 * If the task increased its priority or is running and
6048 * lowered its priority, then reschedule its CPU:
6050 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6051 resched_task(rq->curr);
6054 task_rq_unlock(rq, &flags);
6056 EXPORT_SYMBOL(set_user_nice);
6059 * can_nice - check if a task can reduce its nice value
6063 int can_nice(const struct task_struct *p, const int nice)
6065 /* convert nice value [19,-20] to rlimit style value [1,40] */
6066 int nice_rlim = 20 - nice;
6068 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6069 capable(CAP_SYS_NICE));
6072 #ifdef __ARCH_WANT_SYS_NICE
6075 * sys_nice - change the priority of the current process.
6076 * @increment: priority increment
6078 * sys_setpriority is a more generic, but much slower function that
6079 * does similar things.
6081 SYSCALL_DEFINE1(nice, int, increment)
6086 * Setpriority might change our priority at the same moment.
6087 * We don't have to worry. Conceptually one call occurs first
6088 * and we have a single winner.
6090 if (increment < -40)
6095 nice = TASK_NICE(current) + increment;
6101 if (increment < 0 && !can_nice(current, nice))
6104 retval = security_task_setnice(current, nice);
6108 set_user_nice(current, nice);
6115 * task_prio - return the priority value of a given task.
6116 * @p: the task in question.
6118 * This is the priority value as seen by users in /proc.
6119 * RT tasks are offset by -200. Normal tasks are centered
6120 * around 0, value goes from -16 to +15.
6122 int task_prio(const struct task_struct *p)
6124 return p->prio - MAX_RT_PRIO;
6128 * task_nice - return the nice value of a given task.
6129 * @p: the task in question.
6131 int task_nice(const struct task_struct *p)
6133 return TASK_NICE(p);
6135 EXPORT_SYMBOL(task_nice);
6138 * idle_cpu - is a given cpu idle currently?
6139 * @cpu: the processor in question.
6141 int idle_cpu(int cpu)
6143 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6147 * idle_task - return the idle task for a given cpu.
6148 * @cpu: the processor in question.
6150 struct task_struct *idle_task(int cpu)
6152 return cpu_rq(cpu)->idle;
6156 * find_process_by_pid - find a process with a matching PID value.
6157 * @pid: the pid in question.
6159 static struct task_struct *find_process_by_pid(pid_t pid)
6161 return pid ? find_task_by_vpid(pid) : current;
6164 /* Actually do priority change: must hold rq lock. */
6166 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6168 BUG_ON(p->se.on_rq);
6171 switch (p->policy) {
6175 p->sched_class = &fair_sched_class;
6179 p->sched_class = &rt_sched_class;
6183 p->rt_priority = prio;
6184 p->normal_prio = normal_prio(p);
6185 /* we are holding p->pi_lock already */
6186 p->prio = rt_mutex_getprio(p);
6191 * check the target process has a UID that matches the current process's
6193 static bool check_same_owner(struct task_struct *p)
6195 const struct cred *cred = current_cred(), *pcred;
6199 pcred = __task_cred(p);
6200 match = (cred->euid == pcred->euid ||
6201 cred->euid == pcred->uid);
6206 static int __sched_setscheduler(struct task_struct *p, int policy,
6207 struct sched_param *param, bool user)
6209 int retval, oldprio, oldpolicy = -1, on_rq, running;
6210 unsigned long flags;
6211 const struct sched_class *prev_class = p->sched_class;
6215 /* may grab non-irq protected spin_locks */
6216 BUG_ON(in_interrupt());
6218 /* double check policy once rq lock held */
6220 reset_on_fork = p->sched_reset_on_fork;
6221 policy = oldpolicy = p->policy;
6223 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6224 policy &= ~SCHED_RESET_ON_FORK;
6226 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6227 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6228 policy != SCHED_IDLE)
6233 * Valid priorities for SCHED_FIFO and SCHED_RR are
6234 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6235 * SCHED_BATCH and SCHED_IDLE is 0.
6237 if (param->sched_priority < 0 ||
6238 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6239 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6241 if (rt_policy(policy) != (param->sched_priority != 0))
6245 * Allow unprivileged RT tasks to decrease priority:
6247 if (user && !capable(CAP_SYS_NICE)) {
6248 if (rt_policy(policy)) {
6249 unsigned long rlim_rtprio;
6251 if (!lock_task_sighand(p, &flags))
6253 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6254 unlock_task_sighand(p, &flags);
6256 /* can't set/change the rt policy */
6257 if (policy != p->policy && !rlim_rtprio)
6260 /* can't increase priority */
6261 if (param->sched_priority > p->rt_priority &&
6262 param->sched_priority > rlim_rtprio)
6266 * Like positive nice levels, dont allow tasks to
6267 * move out of SCHED_IDLE either:
6269 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6272 /* can't change other user's priorities */
6273 if (!check_same_owner(p))
6276 /* Normal users shall not reset the sched_reset_on_fork flag */
6277 if (p->sched_reset_on_fork && !reset_on_fork)
6282 #ifdef CONFIG_RT_GROUP_SCHED
6284 * Do not allow realtime tasks into groups that have no runtime
6287 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6288 task_group(p)->rt_bandwidth.rt_runtime == 0)
6292 retval = security_task_setscheduler(p, policy, param);
6298 * make sure no PI-waiters arrive (or leave) while we are
6299 * changing the priority of the task:
6301 spin_lock_irqsave(&p->pi_lock, flags);
6303 * To be able to change p->policy safely, the apropriate
6304 * runqueue lock must be held.
6306 rq = __task_rq_lock(p);
6307 /* recheck policy now with rq lock held */
6308 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6309 policy = oldpolicy = -1;
6310 __task_rq_unlock(rq);
6311 spin_unlock_irqrestore(&p->pi_lock, flags);
6314 update_rq_clock(rq);
6315 on_rq = p->se.on_rq;
6316 running = task_current(rq, p);
6318 deactivate_task(rq, p, 0);
6320 p->sched_class->put_prev_task(rq, p);
6322 p->sched_reset_on_fork = reset_on_fork;
6325 __setscheduler(rq, p, policy, param->sched_priority);
6328 p->sched_class->set_curr_task(rq);
6330 activate_task(rq, p, 0);
6332 check_class_changed(rq, p, prev_class, oldprio, running);
6334 __task_rq_unlock(rq);
6335 spin_unlock_irqrestore(&p->pi_lock, flags);
6337 rt_mutex_adjust_pi(p);
6343 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6344 * @p: the task in question.
6345 * @policy: new policy.
6346 * @param: structure containing the new RT priority.
6348 * NOTE that the task may be already dead.
6350 int sched_setscheduler(struct task_struct *p, int policy,
6351 struct sched_param *param)
6353 return __sched_setscheduler(p, policy, param, true);
6355 EXPORT_SYMBOL_GPL(sched_setscheduler);
6358 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6359 * @p: the task in question.
6360 * @policy: new policy.
6361 * @param: structure containing the new RT priority.
6363 * Just like sched_setscheduler, only don't bother checking if the
6364 * current context has permission. For example, this is needed in
6365 * stop_machine(): we create temporary high priority worker threads,
6366 * but our caller might not have that capability.
6368 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6369 struct sched_param *param)
6371 return __sched_setscheduler(p, policy, param, false);
6375 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6377 struct sched_param lparam;
6378 struct task_struct *p;
6381 if (!param || pid < 0)
6383 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6388 p = find_process_by_pid(pid);
6390 retval = sched_setscheduler(p, policy, &lparam);
6397 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6398 * @pid: the pid in question.
6399 * @policy: new policy.
6400 * @param: structure containing the new RT priority.
6402 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6403 struct sched_param __user *, param)
6405 /* negative values for policy are not valid */
6409 return do_sched_setscheduler(pid, policy, param);
6413 * sys_sched_setparam - set/change the RT priority of a thread
6414 * @pid: the pid in question.
6415 * @param: structure containing the new RT priority.
6417 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6419 return do_sched_setscheduler(pid, -1, param);
6423 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6424 * @pid: the pid in question.
6426 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6428 struct task_struct *p;
6435 read_lock(&tasklist_lock);
6436 p = find_process_by_pid(pid);
6438 retval = security_task_getscheduler(p);
6441 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6443 read_unlock(&tasklist_lock);
6448 * sys_sched_getparam - get the RT priority of a thread
6449 * @pid: the pid in question.
6450 * @param: structure containing the RT priority.
6452 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6454 struct sched_param lp;
6455 struct task_struct *p;
6458 if (!param || pid < 0)
6461 read_lock(&tasklist_lock);
6462 p = find_process_by_pid(pid);
6467 retval = security_task_getscheduler(p);
6471 lp.sched_priority = p->rt_priority;
6472 read_unlock(&tasklist_lock);
6475 * This one might sleep, we cannot do it with a spinlock held ...
6477 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6482 read_unlock(&tasklist_lock);
6486 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6488 cpumask_var_t cpus_allowed, new_mask;
6489 struct task_struct *p;
6493 read_lock(&tasklist_lock);
6495 p = find_process_by_pid(pid);
6497 read_unlock(&tasklist_lock);
6503 * It is not safe to call set_cpus_allowed with the
6504 * tasklist_lock held. We will bump the task_struct's
6505 * usage count and then drop tasklist_lock.
6508 read_unlock(&tasklist_lock);
6510 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6514 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6516 goto out_free_cpus_allowed;
6519 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6522 retval = security_task_setscheduler(p, 0, NULL);
6526 cpuset_cpus_allowed(p, cpus_allowed);
6527 cpumask_and(new_mask, in_mask, cpus_allowed);
6529 retval = set_cpus_allowed_ptr(p, new_mask);
6532 cpuset_cpus_allowed(p, cpus_allowed);
6533 if (!cpumask_subset(new_mask, cpus_allowed)) {
6535 * We must have raced with a concurrent cpuset
6536 * update. Just reset the cpus_allowed to the
6537 * cpuset's cpus_allowed
6539 cpumask_copy(new_mask, cpus_allowed);
6544 free_cpumask_var(new_mask);
6545 out_free_cpus_allowed:
6546 free_cpumask_var(cpus_allowed);
6553 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6554 struct cpumask *new_mask)
6556 if (len < cpumask_size())
6557 cpumask_clear(new_mask);
6558 else if (len > cpumask_size())
6559 len = cpumask_size();
6561 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6565 * sys_sched_setaffinity - set the cpu affinity of a process
6566 * @pid: pid of the process
6567 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6568 * @user_mask_ptr: user-space pointer to the new cpu mask
6570 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6571 unsigned long __user *, user_mask_ptr)
6573 cpumask_var_t new_mask;
6576 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6579 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6581 retval = sched_setaffinity(pid, new_mask);
6582 free_cpumask_var(new_mask);
6586 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6588 struct task_struct *p;
6592 read_lock(&tasklist_lock);
6595 p = find_process_by_pid(pid);
6599 retval = security_task_getscheduler(p);
6603 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6606 read_unlock(&tasklist_lock);
6613 * sys_sched_getaffinity - get the cpu affinity of a process
6614 * @pid: pid of the process
6615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6616 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6618 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6619 unsigned long __user *, user_mask_ptr)
6624 if (len < cpumask_size())
6627 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6630 ret = sched_getaffinity(pid, mask);
6632 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6635 ret = cpumask_size();
6637 free_cpumask_var(mask);
6643 * sys_sched_yield - yield the current processor to other threads.
6645 * This function yields the current CPU to other tasks. If there are no
6646 * other threads running on this CPU then this function will return.
6648 SYSCALL_DEFINE0(sched_yield)
6650 struct rq *rq = this_rq_lock();
6652 schedstat_inc(rq, yld_count);
6653 current->sched_class->yield_task(rq);
6656 * Since we are going to call schedule() anyway, there's
6657 * no need to preempt or enable interrupts:
6659 __release(rq->lock);
6660 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6661 _raw_spin_unlock(&rq->lock);
6662 preempt_enable_no_resched();
6669 static inline int should_resched(void)
6671 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6674 static void __cond_resched(void)
6676 add_preempt_count(PREEMPT_ACTIVE);
6678 sub_preempt_count(PREEMPT_ACTIVE);
6681 int __sched _cond_resched(void)
6683 if (should_resched()) {
6689 EXPORT_SYMBOL(_cond_resched);
6692 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6693 * call schedule, and on return reacquire the lock.
6695 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6696 * operations here to prevent schedule() from being called twice (once via
6697 * spin_unlock(), once by hand).
6699 int __cond_resched_lock(spinlock_t *lock)
6701 int resched = should_resched();
6704 if (spin_needbreak(lock) || resched) {
6715 EXPORT_SYMBOL(__cond_resched_lock);
6717 int __sched __cond_resched_softirq(void)
6719 BUG_ON(!in_softirq());
6721 if (should_resched()) {
6729 EXPORT_SYMBOL(__cond_resched_softirq);
6732 * yield - yield the current processor to other threads.
6734 * This is a shortcut for kernel-space yielding - it marks the
6735 * thread runnable and calls sys_sched_yield().
6737 void __sched yield(void)
6739 set_current_state(TASK_RUNNING);
6742 EXPORT_SYMBOL(yield);
6745 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6746 * that process accounting knows that this is a task in IO wait state.
6748 * But don't do that if it is a deliberate, throttling IO wait (this task
6749 * has set its backing_dev_info: the queue against which it should throttle)
6751 void __sched io_schedule(void)
6753 struct rq *rq = raw_rq();
6755 delayacct_blkio_start();
6756 atomic_inc(&rq->nr_iowait);
6757 current->in_iowait = 1;
6759 current->in_iowait = 0;
6760 atomic_dec(&rq->nr_iowait);
6761 delayacct_blkio_end();
6763 EXPORT_SYMBOL(io_schedule);
6765 long __sched io_schedule_timeout(long timeout)
6767 struct rq *rq = raw_rq();
6770 delayacct_blkio_start();
6771 atomic_inc(&rq->nr_iowait);
6772 current->in_iowait = 1;
6773 ret = schedule_timeout(timeout);
6774 current->in_iowait = 0;
6775 atomic_dec(&rq->nr_iowait);
6776 delayacct_blkio_end();
6781 * sys_sched_get_priority_max - return maximum RT priority.
6782 * @policy: scheduling class.
6784 * this syscall returns the maximum rt_priority that can be used
6785 * by a given scheduling class.
6787 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6794 ret = MAX_USER_RT_PRIO-1;
6806 * sys_sched_get_priority_min - return minimum RT priority.
6807 * @policy: scheduling class.
6809 * this syscall returns the minimum rt_priority that can be used
6810 * by a given scheduling class.
6812 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6830 * sys_sched_rr_get_interval - return the default timeslice of a process.
6831 * @pid: pid of the process.
6832 * @interval: userspace pointer to the timeslice value.
6834 * this syscall writes the default timeslice value of a given process
6835 * into the user-space timespec buffer. A value of '0' means infinity.
6837 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6838 struct timespec __user *, interval)
6840 struct task_struct *p;
6841 unsigned int time_slice;
6849 read_lock(&tasklist_lock);
6850 p = find_process_by_pid(pid);
6854 retval = security_task_getscheduler(p);
6859 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6860 * tasks that are on an otherwise idle runqueue:
6863 if (p->policy == SCHED_RR) {
6864 time_slice = DEF_TIMESLICE;
6865 } else if (p->policy != SCHED_FIFO) {
6866 struct sched_entity *se = &p->se;
6867 unsigned long flags;
6870 rq = task_rq_lock(p, &flags);
6871 if (rq->cfs.load.weight)
6872 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6873 task_rq_unlock(rq, &flags);
6875 read_unlock(&tasklist_lock);
6876 jiffies_to_timespec(time_slice, &t);
6877 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6881 read_unlock(&tasklist_lock);
6885 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6887 void sched_show_task(struct task_struct *p)
6889 unsigned long free = 0;
6892 state = p->state ? __ffs(p->state) + 1 : 0;
6893 printk(KERN_INFO "%-13.13s %c", p->comm,
6894 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6895 #if BITS_PER_LONG == 32
6896 if (state == TASK_RUNNING)
6897 printk(KERN_CONT " running ");
6899 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6901 if (state == TASK_RUNNING)
6902 printk(KERN_CONT " running task ");
6904 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6906 #ifdef CONFIG_DEBUG_STACK_USAGE
6907 free = stack_not_used(p);
6909 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6910 task_pid_nr(p), task_pid_nr(p->real_parent),
6911 (unsigned long)task_thread_info(p)->flags);
6913 show_stack(p, NULL);
6916 void show_state_filter(unsigned long state_filter)
6918 struct task_struct *g, *p;
6920 #if BITS_PER_LONG == 32
6922 " task PC stack pid father\n");
6925 " task PC stack pid father\n");
6927 read_lock(&tasklist_lock);
6928 do_each_thread(g, p) {
6930 * reset the NMI-timeout, listing all files on a slow
6931 * console might take alot of time:
6933 touch_nmi_watchdog();
6934 if (!state_filter || (p->state & state_filter))
6936 } while_each_thread(g, p);
6938 touch_all_softlockup_watchdogs();
6940 #ifdef CONFIG_SCHED_DEBUG
6941 sysrq_sched_debug_show();
6943 read_unlock(&tasklist_lock);
6945 * Only show locks if all tasks are dumped:
6947 if (state_filter == -1)
6948 debug_show_all_locks();
6951 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6953 idle->sched_class = &idle_sched_class;
6957 * init_idle - set up an idle thread for a given CPU
6958 * @idle: task in question
6959 * @cpu: cpu the idle task belongs to
6961 * NOTE: this function does not set the idle thread's NEED_RESCHED
6962 * flag, to make booting more robust.
6964 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6966 struct rq *rq = cpu_rq(cpu);
6967 unsigned long flags;
6969 spin_lock_irqsave(&rq->lock, flags);
6972 idle->se.exec_start = sched_clock();
6974 idle->prio = idle->normal_prio = MAX_PRIO;
6975 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6976 __set_task_cpu(idle, cpu);
6978 rq->curr = rq->idle = idle;
6979 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6982 spin_unlock_irqrestore(&rq->lock, flags);
6984 /* Set the preempt count _outside_ the spinlocks! */
6985 #if defined(CONFIG_PREEMPT)
6986 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6988 task_thread_info(idle)->preempt_count = 0;
6991 * The idle tasks have their own, simple scheduling class:
6993 idle->sched_class = &idle_sched_class;
6994 ftrace_graph_init_task(idle);
6998 * In a system that switches off the HZ timer nohz_cpu_mask
6999 * indicates which cpus entered this state. This is used
7000 * in the rcu update to wait only for active cpus. For system
7001 * which do not switch off the HZ timer nohz_cpu_mask should
7002 * always be CPU_BITS_NONE.
7004 cpumask_var_t nohz_cpu_mask;
7007 * Increase the granularity value when there are more CPUs,
7008 * because with more CPUs the 'effective latency' as visible
7009 * to users decreases. But the relationship is not linear,
7010 * so pick a second-best guess by going with the log2 of the
7013 * This idea comes from the SD scheduler of Con Kolivas:
7015 static inline void sched_init_granularity(void)
7017 unsigned int factor = 1 + ilog2(num_online_cpus());
7018 const unsigned long limit = 200000000;
7020 sysctl_sched_min_granularity *= factor;
7021 if (sysctl_sched_min_granularity > limit)
7022 sysctl_sched_min_granularity = limit;
7024 sysctl_sched_latency *= factor;
7025 if (sysctl_sched_latency > limit)
7026 sysctl_sched_latency = limit;
7028 sysctl_sched_wakeup_granularity *= factor;
7030 sysctl_sched_shares_ratelimit *= factor;
7035 * This is how migration works:
7037 * 1) we queue a struct migration_req structure in the source CPU's
7038 * runqueue and wake up that CPU's migration thread.
7039 * 2) we down() the locked semaphore => thread blocks.
7040 * 3) migration thread wakes up (implicitly it forces the migrated
7041 * thread off the CPU)
7042 * 4) it gets the migration request and checks whether the migrated
7043 * task is still in the wrong runqueue.
7044 * 5) if it's in the wrong runqueue then the migration thread removes
7045 * it and puts it into the right queue.
7046 * 6) migration thread up()s the semaphore.
7047 * 7) we wake up and the migration is done.
7051 * Change a given task's CPU affinity. Migrate the thread to a
7052 * proper CPU and schedule it away if the CPU it's executing on
7053 * is removed from the allowed bitmask.
7055 * NOTE: the caller must have a valid reference to the task, the
7056 * task must not exit() & deallocate itself prematurely. The
7057 * call is not atomic; no spinlocks may be held.
7059 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7061 struct migration_req req;
7062 unsigned long flags;
7066 rq = task_rq_lock(p, &flags);
7067 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7072 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7073 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7078 if (p->sched_class->set_cpus_allowed)
7079 p->sched_class->set_cpus_allowed(p, new_mask);
7081 cpumask_copy(&p->cpus_allowed, new_mask);
7082 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7085 /* Can the task run on the task's current CPU? If so, we're done */
7086 if (cpumask_test_cpu(task_cpu(p), new_mask))
7089 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7090 /* Need help from migration thread: drop lock and wait. */
7091 struct task_struct *mt = rq->migration_thread;
7093 get_task_struct(mt);
7094 task_rq_unlock(rq, &flags);
7095 wake_up_process(rq->migration_thread);
7096 put_task_struct(mt);
7097 wait_for_completion(&req.done);
7098 tlb_migrate_finish(p->mm);
7102 task_rq_unlock(rq, &flags);
7106 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7109 * Move (not current) task off this cpu, onto dest cpu. We're doing
7110 * this because either it can't run here any more (set_cpus_allowed()
7111 * away from this CPU, or CPU going down), or because we're
7112 * attempting to rebalance this task on exec (sched_exec).
7114 * So we race with normal scheduler movements, but that's OK, as long
7115 * as the task is no longer on this CPU.
7117 * Returns non-zero if task was successfully migrated.
7119 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7121 struct rq *rq_dest, *rq_src;
7124 if (unlikely(!cpu_active(dest_cpu)))
7127 rq_src = cpu_rq(src_cpu);
7128 rq_dest = cpu_rq(dest_cpu);
7130 double_rq_lock(rq_src, rq_dest);
7131 /* Already moved. */
7132 if (task_cpu(p) != src_cpu)
7134 /* Affinity changed (again). */
7135 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7138 on_rq = p->se.on_rq;
7140 deactivate_task(rq_src, p, 0);
7142 set_task_cpu(p, dest_cpu);
7144 activate_task(rq_dest, p, 0);
7145 check_preempt_curr(rq_dest, p, 0);
7150 double_rq_unlock(rq_src, rq_dest);
7155 * migration_thread - this is a highprio system thread that performs
7156 * thread migration by bumping thread off CPU then 'pushing' onto
7159 static int migration_thread(void *data)
7161 int cpu = (long)data;
7165 BUG_ON(rq->migration_thread != current);
7167 set_current_state(TASK_INTERRUPTIBLE);
7168 while (!kthread_should_stop()) {
7169 struct migration_req *req;
7170 struct list_head *head;
7172 spin_lock_irq(&rq->lock);
7174 if (cpu_is_offline(cpu)) {
7175 spin_unlock_irq(&rq->lock);
7179 if (rq->active_balance) {
7180 active_load_balance(rq, cpu);
7181 rq->active_balance = 0;
7184 head = &rq->migration_queue;
7186 if (list_empty(head)) {
7187 spin_unlock_irq(&rq->lock);
7189 set_current_state(TASK_INTERRUPTIBLE);
7192 req = list_entry(head->next, struct migration_req, list);
7193 list_del_init(head->next);
7195 spin_unlock(&rq->lock);
7196 __migrate_task(req->task, cpu, req->dest_cpu);
7199 complete(&req->done);
7201 __set_current_state(TASK_RUNNING);
7206 #ifdef CONFIG_HOTPLUG_CPU
7208 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7212 local_irq_disable();
7213 ret = __migrate_task(p, src_cpu, dest_cpu);
7219 * Figure out where task on dead CPU should go, use force if necessary.
7221 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7224 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7227 /* Look for allowed, online CPU in same node. */
7228 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7229 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7232 /* Any allowed, online CPU? */
7233 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7234 if (dest_cpu < nr_cpu_ids)
7237 /* No more Mr. Nice Guy. */
7238 if (dest_cpu >= nr_cpu_ids) {
7239 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7240 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7243 * Don't tell them about moving exiting tasks or
7244 * kernel threads (both mm NULL), since they never
7247 if (p->mm && printk_ratelimit()) {
7248 printk(KERN_INFO "process %d (%s) no "
7249 "longer affine to cpu%d\n",
7250 task_pid_nr(p), p->comm, dead_cpu);
7255 /* It can have affinity changed while we were choosing. */
7256 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7261 * While a dead CPU has no uninterruptible tasks queued at this point,
7262 * it might still have a nonzero ->nr_uninterruptible counter, because
7263 * for performance reasons the counter is not stricly tracking tasks to
7264 * their home CPUs. So we just add the counter to another CPU's counter,
7265 * to keep the global sum constant after CPU-down:
7267 static void migrate_nr_uninterruptible(struct rq *rq_src)
7269 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7270 unsigned long flags;
7272 local_irq_save(flags);
7273 double_rq_lock(rq_src, rq_dest);
7274 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7275 rq_src->nr_uninterruptible = 0;
7276 double_rq_unlock(rq_src, rq_dest);
7277 local_irq_restore(flags);
7280 /* Run through task list and migrate tasks from the dead cpu. */
7281 static void migrate_live_tasks(int src_cpu)
7283 struct task_struct *p, *t;
7285 read_lock(&tasklist_lock);
7287 do_each_thread(t, p) {
7291 if (task_cpu(p) == src_cpu)
7292 move_task_off_dead_cpu(src_cpu, p);
7293 } while_each_thread(t, p);
7295 read_unlock(&tasklist_lock);
7299 * Schedules idle task to be the next runnable task on current CPU.
7300 * It does so by boosting its priority to highest possible.
7301 * Used by CPU offline code.
7303 void sched_idle_next(void)
7305 int this_cpu = smp_processor_id();
7306 struct rq *rq = cpu_rq(this_cpu);
7307 struct task_struct *p = rq->idle;
7308 unsigned long flags;
7310 /* cpu has to be offline */
7311 BUG_ON(cpu_online(this_cpu));
7314 * Strictly not necessary since rest of the CPUs are stopped by now
7315 * and interrupts disabled on the current cpu.
7317 spin_lock_irqsave(&rq->lock, flags);
7319 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7321 update_rq_clock(rq);
7322 activate_task(rq, p, 0);
7324 spin_unlock_irqrestore(&rq->lock, flags);
7328 * Ensures that the idle task is using init_mm right before its cpu goes
7331 void idle_task_exit(void)
7333 struct mm_struct *mm = current->active_mm;
7335 BUG_ON(cpu_online(smp_processor_id()));
7338 switch_mm(mm, &init_mm, current);
7342 /* called under rq->lock with disabled interrupts */
7343 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7345 struct rq *rq = cpu_rq(dead_cpu);
7347 /* Must be exiting, otherwise would be on tasklist. */
7348 BUG_ON(!p->exit_state);
7350 /* Cannot have done final schedule yet: would have vanished. */
7351 BUG_ON(p->state == TASK_DEAD);
7356 * Drop lock around migration; if someone else moves it,
7357 * that's OK. No task can be added to this CPU, so iteration is
7360 spin_unlock_irq(&rq->lock);
7361 move_task_off_dead_cpu(dead_cpu, p);
7362 spin_lock_irq(&rq->lock);
7367 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7368 static void migrate_dead_tasks(unsigned int dead_cpu)
7370 struct rq *rq = cpu_rq(dead_cpu);
7371 struct task_struct *next;
7374 if (!rq->nr_running)
7376 update_rq_clock(rq);
7377 next = pick_next_task(rq);
7380 next->sched_class->put_prev_task(rq, next);
7381 migrate_dead(dead_cpu, next);
7387 * remove the tasks which were accounted by rq from calc_load_tasks.
7389 static void calc_global_load_remove(struct rq *rq)
7391 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7392 rq->calc_load_active = 0;
7394 #endif /* CONFIG_HOTPLUG_CPU */
7396 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7398 static struct ctl_table sd_ctl_dir[] = {
7400 .procname = "sched_domain",
7406 static struct ctl_table sd_ctl_root[] = {
7408 .ctl_name = CTL_KERN,
7409 .procname = "kernel",
7411 .child = sd_ctl_dir,
7416 static struct ctl_table *sd_alloc_ctl_entry(int n)
7418 struct ctl_table *entry =
7419 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7424 static void sd_free_ctl_entry(struct ctl_table **tablep)
7426 struct ctl_table *entry;
7429 * In the intermediate directories, both the child directory and
7430 * procname are dynamically allocated and could fail but the mode
7431 * will always be set. In the lowest directory the names are
7432 * static strings and all have proc handlers.
7434 for (entry = *tablep; entry->mode; entry++) {
7436 sd_free_ctl_entry(&entry->child);
7437 if (entry->proc_handler == NULL)
7438 kfree(entry->procname);
7446 set_table_entry(struct ctl_table *entry,
7447 const char *procname, void *data, int maxlen,
7448 mode_t mode, proc_handler *proc_handler)
7450 entry->procname = procname;
7452 entry->maxlen = maxlen;
7454 entry->proc_handler = proc_handler;
7457 static struct ctl_table *
7458 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7460 struct ctl_table *table = sd_alloc_ctl_entry(13);
7465 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7466 sizeof(long), 0644, proc_doulongvec_minmax);
7467 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7468 sizeof(long), 0644, proc_doulongvec_minmax);
7469 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7470 sizeof(int), 0644, proc_dointvec_minmax);
7471 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7472 sizeof(int), 0644, proc_dointvec_minmax);
7473 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7474 sizeof(int), 0644, proc_dointvec_minmax);
7475 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7476 sizeof(int), 0644, proc_dointvec_minmax);
7477 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7478 sizeof(int), 0644, proc_dointvec_minmax);
7479 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7480 sizeof(int), 0644, proc_dointvec_minmax);
7481 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7482 sizeof(int), 0644, proc_dointvec_minmax);
7483 set_table_entry(&table[9], "cache_nice_tries",
7484 &sd->cache_nice_tries,
7485 sizeof(int), 0644, proc_dointvec_minmax);
7486 set_table_entry(&table[10], "flags", &sd->flags,
7487 sizeof(int), 0644, proc_dointvec_minmax);
7488 set_table_entry(&table[11], "name", sd->name,
7489 CORENAME_MAX_SIZE, 0444, proc_dostring);
7490 /* &table[12] is terminator */
7495 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7497 struct ctl_table *entry, *table;
7498 struct sched_domain *sd;
7499 int domain_num = 0, i;
7502 for_each_domain(cpu, sd)
7504 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7509 for_each_domain(cpu, sd) {
7510 snprintf(buf, 32, "domain%d", i);
7511 entry->procname = kstrdup(buf, GFP_KERNEL);
7513 entry->child = sd_alloc_ctl_domain_table(sd);
7520 static struct ctl_table_header *sd_sysctl_header;
7521 static void register_sched_domain_sysctl(void)
7523 int i, cpu_num = num_online_cpus();
7524 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7527 WARN_ON(sd_ctl_dir[0].child);
7528 sd_ctl_dir[0].child = entry;
7533 for_each_online_cpu(i) {
7534 snprintf(buf, 32, "cpu%d", i);
7535 entry->procname = kstrdup(buf, GFP_KERNEL);
7537 entry->child = sd_alloc_ctl_cpu_table(i);
7541 WARN_ON(sd_sysctl_header);
7542 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7545 /* may be called multiple times per register */
7546 static void unregister_sched_domain_sysctl(void)
7548 if (sd_sysctl_header)
7549 unregister_sysctl_table(sd_sysctl_header);
7550 sd_sysctl_header = NULL;
7551 if (sd_ctl_dir[0].child)
7552 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7555 static void register_sched_domain_sysctl(void)
7558 static void unregister_sched_domain_sysctl(void)
7563 static void set_rq_online(struct rq *rq)
7566 const struct sched_class *class;
7568 cpumask_set_cpu(rq->cpu, rq->rd->online);
7571 for_each_class(class) {
7572 if (class->rq_online)
7573 class->rq_online(rq);
7578 static void set_rq_offline(struct rq *rq)
7581 const struct sched_class *class;
7583 for_each_class(class) {
7584 if (class->rq_offline)
7585 class->rq_offline(rq);
7588 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7594 * migration_call - callback that gets triggered when a CPU is added.
7595 * Here we can start up the necessary migration thread for the new CPU.
7597 static int __cpuinit
7598 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7600 struct task_struct *p;
7601 int cpu = (long)hcpu;
7602 unsigned long flags;
7607 case CPU_UP_PREPARE:
7608 case CPU_UP_PREPARE_FROZEN:
7609 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7612 kthread_bind(p, cpu);
7613 /* Must be high prio: stop_machine expects to yield to it. */
7614 rq = task_rq_lock(p, &flags);
7615 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7616 task_rq_unlock(rq, &flags);
7618 cpu_rq(cpu)->migration_thread = p;
7619 rq->calc_load_update = calc_load_update;
7623 case CPU_ONLINE_FROZEN:
7624 /* Strictly unnecessary, as first user will wake it. */
7625 wake_up_process(cpu_rq(cpu)->migration_thread);
7627 /* Update our root-domain */
7629 spin_lock_irqsave(&rq->lock, flags);
7631 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7635 spin_unlock_irqrestore(&rq->lock, flags);
7638 #ifdef CONFIG_HOTPLUG_CPU
7639 case CPU_UP_CANCELED:
7640 case CPU_UP_CANCELED_FROZEN:
7641 if (!cpu_rq(cpu)->migration_thread)
7643 /* Unbind it from offline cpu so it can run. Fall thru. */
7644 kthread_bind(cpu_rq(cpu)->migration_thread,
7645 cpumask_any(cpu_online_mask));
7646 kthread_stop(cpu_rq(cpu)->migration_thread);
7647 put_task_struct(cpu_rq(cpu)->migration_thread);
7648 cpu_rq(cpu)->migration_thread = NULL;
7652 case CPU_DEAD_FROZEN:
7653 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7654 migrate_live_tasks(cpu);
7656 kthread_stop(rq->migration_thread);
7657 put_task_struct(rq->migration_thread);
7658 rq->migration_thread = NULL;
7659 /* Idle task back to normal (off runqueue, low prio) */
7660 spin_lock_irq(&rq->lock);
7661 update_rq_clock(rq);
7662 deactivate_task(rq, rq->idle, 0);
7663 rq->idle->static_prio = MAX_PRIO;
7664 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7665 rq->idle->sched_class = &idle_sched_class;
7666 migrate_dead_tasks(cpu);
7667 spin_unlock_irq(&rq->lock);
7669 migrate_nr_uninterruptible(rq);
7670 BUG_ON(rq->nr_running != 0);
7671 calc_global_load_remove(rq);
7673 * No need to migrate the tasks: it was best-effort if
7674 * they didn't take sched_hotcpu_mutex. Just wake up
7677 spin_lock_irq(&rq->lock);
7678 while (!list_empty(&rq->migration_queue)) {
7679 struct migration_req *req;
7681 req = list_entry(rq->migration_queue.next,
7682 struct migration_req, list);
7683 list_del_init(&req->list);
7684 spin_unlock_irq(&rq->lock);
7685 complete(&req->done);
7686 spin_lock_irq(&rq->lock);
7688 spin_unlock_irq(&rq->lock);
7692 case CPU_DYING_FROZEN:
7693 /* Update our root-domain */
7695 spin_lock_irqsave(&rq->lock, flags);
7697 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7700 spin_unlock_irqrestore(&rq->lock, flags);
7708 * Register at high priority so that task migration (migrate_all_tasks)
7709 * happens before everything else. This has to be lower priority than
7710 * the notifier in the perf_counter subsystem, though.
7712 static struct notifier_block __cpuinitdata migration_notifier = {
7713 .notifier_call = migration_call,
7717 static int __init migration_init(void)
7719 void *cpu = (void *)(long)smp_processor_id();
7722 /* Start one for the boot CPU: */
7723 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7724 BUG_ON(err == NOTIFY_BAD);
7725 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7726 register_cpu_notifier(&migration_notifier);
7730 early_initcall(migration_init);
7735 #ifdef CONFIG_SCHED_DEBUG
7737 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7738 struct cpumask *groupmask)
7740 struct sched_group *group = sd->groups;
7743 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7744 cpumask_clear(groupmask);
7746 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7748 if (!(sd->flags & SD_LOAD_BALANCE)) {
7749 printk("does not load-balance\n");
7751 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7756 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7758 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7759 printk(KERN_ERR "ERROR: domain->span does not contain "
7762 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7763 printk(KERN_ERR "ERROR: domain->groups does not contain"
7767 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7771 printk(KERN_ERR "ERROR: group is NULL\n");
7775 if (!group->__cpu_power) {
7776 printk(KERN_CONT "\n");
7777 printk(KERN_ERR "ERROR: domain->cpu_power not "
7782 if (!cpumask_weight(sched_group_cpus(group))) {
7783 printk(KERN_CONT "\n");
7784 printk(KERN_ERR "ERROR: empty group\n");
7788 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7789 printk(KERN_CONT "\n");
7790 printk(KERN_ERR "ERROR: repeated CPUs\n");
7794 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7796 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7798 printk(KERN_CONT " %s", str);
7799 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7800 printk(KERN_CONT " (__cpu_power = %d)",
7801 group->__cpu_power);
7804 group = group->next;
7805 } while (group != sd->groups);
7806 printk(KERN_CONT "\n");
7808 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7809 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7812 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7813 printk(KERN_ERR "ERROR: parent span is not a superset "
7814 "of domain->span\n");
7818 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7820 cpumask_var_t groupmask;
7824 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7828 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7830 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7831 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7836 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7843 free_cpumask_var(groupmask);
7845 #else /* !CONFIG_SCHED_DEBUG */
7846 # define sched_domain_debug(sd, cpu) do { } while (0)
7847 #endif /* CONFIG_SCHED_DEBUG */
7849 static int sd_degenerate(struct sched_domain *sd)
7851 if (cpumask_weight(sched_domain_span(sd)) == 1)
7854 /* Following flags need at least 2 groups */
7855 if (sd->flags & (SD_LOAD_BALANCE |
7856 SD_BALANCE_NEWIDLE |
7860 SD_SHARE_PKG_RESOURCES)) {
7861 if (sd->groups != sd->groups->next)
7865 /* Following flags don't use groups */
7866 if (sd->flags & (SD_WAKE_IDLE |
7875 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7877 unsigned long cflags = sd->flags, pflags = parent->flags;
7879 if (sd_degenerate(parent))
7882 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7885 /* Does parent contain flags not in child? */
7886 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7887 if (cflags & SD_WAKE_AFFINE)
7888 pflags &= ~SD_WAKE_BALANCE;
7889 /* Flags needing groups don't count if only 1 group in parent */
7890 if (parent->groups == parent->groups->next) {
7891 pflags &= ~(SD_LOAD_BALANCE |
7892 SD_BALANCE_NEWIDLE |
7896 SD_SHARE_PKG_RESOURCES);
7897 if (nr_node_ids == 1)
7898 pflags &= ~SD_SERIALIZE;
7900 if (~cflags & pflags)
7906 static void free_rootdomain(struct root_domain *rd)
7908 cpupri_cleanup(&rd->cpupri);
7910 free_cpumask_var(rd->rto_mask);
7911 free_cpumask_var(rd->online);
7912 free_cpumask_var(rd->span);
7916 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7918 struct root_domain *old_rd = NULL;
7919 unsigned long flags;
7921 spin_lock_irqsave(&rq->lock, flags);
7926 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7929 cpumask_clear_cpu(rq->cpu, old_rd->span);
7932 * If we dont want to free the old_rt yet then
7933 * set old_rd to NULL to skip the freeing later
7936 if (!atomic_dec_and_test(&old_rd->refcount))
7940 atomic_inc(&rd->refcount);
7943 cpumask_set_cpu(rq->cpu, rd->span);
7944 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7947 spin_unlock_irqrestore(&rq->lock, flags);
7950 free_rootdomain(old_rd);
7953 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7955 gfp_t gfp = GFP_KERNEL;
7957 memset(rd, 0, sizeof(*rd));
7962 if (!alloc_cpumask_var(&rd->span, gfp))
7964 if (!alloc_cpumask_var(&rd->online, gfp))
7966 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7969 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7974 free_cpumask_var(rd->rto_mask);
7976 free_cpumask_var(rd->online);
7978 free_cpumask_var(rd->span);
7983 static void init_defrootdomain(void)
7985 init_rootdomain(&def_root_domain, true);
7987 atomic_set(&def_root_domain.refcount, 1);
7990 static struct root_domain *alloc_rootdomain(void)
7992 struct root_domain *rd;
7994 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7998 if (init_rootdomain(rd, false) != 0) {
8007 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8008 * hold the hotplug lock.
8011 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8013 struct rq *rq = cpu_rq(cpu);
8014 struct sched_domain *tmp;
8016 /* Remove the sched domains which do not contribute to scheduling. */
8017 for (tmp = sd; tmp; ) {
8018 struct sched_domain *parent = tmp->parent;
8022 if (sd_parent_degenerate(tmp, parent)) {
8023 tmp->parent = parent->parent;
8025 parent->parent->child = tmp;
8030 if (sd && sd_degenerate(sd)) {
8036 sched_domain_debug(sd, cpu);
8038 rq_attach_root(rq, rd);
8039 rcu_assign_pointer(rq->sd, sd);
8042 /* cpus with isolated domains */
8043 static cpumask_var_t cpu_isolated_map;
8045 /* Setup the mask of cpus configured for isolated domains */
8046 static int __init isolated_cpu_setup(char *str)
8048 cpulist_parse(str, cpu_isolated_map);
8052 __setup("isolcpus=", isolated_cpu_setup);
8055 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8056 * to a function which identifies what group(along with sched group) a CPU
8057 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8058 * (due to the fact that we keep track of groups covered with a struct cpumask).
8060 * init_sched_build_groups will build a circular linked list of the groups
8061 * covered by the given span, and will set each group's ->cpumask correctly,
8062 * and ->cpu_power to 0.
8065 init_sched_build_groups(const struct cpumask *span,
8066 const struct cpumask *cpu_map,
8067 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8068 struct sched_group **sg,
8069 struct cpumask *tmpmask),
8070 struct cpumask *covered, struct cpumask *tmpmask)
8072 struct sched_group *first = NULL, *last = NULL;
8075 cpumask_clear(covered);
8077 for_each_cpu(i, span) {
8078 struct sched_group *sg;
8079 int group = group_fn(i, cpu_map, &sg, tmpmask);
8082 if (cpumask_test_cpu(i, covered))
8085 cpumask_clear(sched_group_cpus(sg));
8086 sg->__cpu_power = 0;
8088 for_each_cpu(j, span) {
8089 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8092 cpumask_set_cpu(j, covered);
8093 cpumask_set_cpu(j, sched_group_cpus(sg));
8104 #define SD_NODES_PER_DOMAIN 16
8109 * find_next_best_node - find the next node to include in a sched_domain
8110 * @node: node whose sched_domain we're building
8111 * @used_nodes: nodes already in the sched_domain
8113 * Find the next node to include in a given scheduling domain. Simply
8114 * finds the closest node not already in the @used_nodes map.
8116 * Should use nodemask_t.
8118 static int find_next_best_node(int node, nodemask_t *used_nodes)
8120 int i, n, val, min_val, best_node = 0;
8124 for (i = 0; i < nr_node_ids; i++) {
8125 /* Start at @node */
8126 n = (node + i) % nr_node_ids;
8128 if (!nr_cpus_node(n))
8131 /* Skip already used nodes */
8132 if (node_isset(n, *used_nodes))
8135 /* Simple min distance search */
8136 val = node_distance(node, n);
8138 if (val < min_val) {
8144 node_set(best_node, *used_nodes);
8149 * sched_domain_node_span - get a cpumask for a node's sched_domain
8150 * @node: node whose cpumask we're constructing
8151 * @span: resulting cpumask
8153 * Given a node, construct a good cpumask for its sched_domain to span. It
8154 * should be one that prevents unnecessary balancing, but also spreads tasks
8157 static void sched_domain_node_span(int node, struct cpumask *span)
8159 nodemask_t used_nodes;
8162 cpumask_clear(span);
8163 nodes_clear(used_nodes);
8165 cpumask_or(span, span, cpumask_of_node(node));
8166 node_set(node, used_nodes);
8168 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8169 int next_node = find_next_best_node(node, &used_nodes);
8171 cpumask_or(span, span, cpumask_of_node(next_node));
8174 #endif /* CONFIG_NUMA */
8176 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8179 * The cpus mask in sched_group and sched_domain hangs off the end.
8181 * ( See the the comments in include/linux/sched.h:struct sched_group
8182 * and struct sched_domain. )
8184 struct static_sched_group {
8185 struct sched_group sg;
8186 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8189 struct static_sched_domain {
8190 struct sched_domain sd;
8191 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8197 cpumask_var_t domainspan;
8198 cpumask_var_t covered;
8199 cpumask_var_t notcovered;
8201 cpumask_var_t nodemask;
8202 cpumask_var_t this_sibling_map;
8203 cpumask_var_t this_core_map;
8204 cpumask_var_t send_covered;
8205 cpumask_var_t tmpmask;
8206 struct sched_group **sched_group_nodes;
8207 struct root_domain *rd;
8211 sa_sched_groups = 0,
8216 sa_this_sibling_map,
8218 sa_sched_group_nodes,
8228 * SMT sched-domains:
8230 #ifdef CONFIG_SCHED_SMT
8231 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8232 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8235 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8236 struct sched_group **sg, struct cpumask *unused)
8239 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8242 #endif /* CONFIG_SCHED_SMT */
8245 * multi-core sched-domains:
8247 #ifdef CONFIG_SCHED_MC
8248 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8249 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8250 #endif /* CONFIG_SCHED_MC */
8252 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8254 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8255 struct sched_group **sg, struct cpumask *mask)
8259 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8260 group = cpumask_first(mask);
8262 *sg = &per_cpu(sched_group_core, group).sg;
8265 #elif defined(CONFIG_SCHED_MC)
8267 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8268 struct sched_group **sg, struct cpumask *unused)
8271 *sg = &per_cpu(sched_group_core, cpu).sg;
8276 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8277 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8280 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8281 struct sched_group **sg, struct cpumask *mask)
8284 #ifdef CONFIG_SCHED_MC
8285 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8286 group = cpumask_first(mask);
8287 #elif defined(CONFIG_SCHED_SMT)
8288 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8289 group = cpumask_first(mask);
8294 *sg = &per_cpu(sched_group_phys, group).sg;
8300 * The init_sched_build_groups can't handle what we want to do with node
8301 * groups, so roll our own. Now each node has its own list of groups which
8302 * gets dynamically allocated.
8304 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8305 static struct sched_group ***sched_group_nodes_bycpu;
8307 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8308 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8310 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8311 struct sched_group **sg,
8312 struct cpumask *nodemask)
8316 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8317 group = cpumask_first(nodemask);
8320 *sg = &per_cpu(sched_group_allnodes, group).sg;
8324 static void init_numa_sched_groups_power(struct sched_group *group_head)
8326 struct sched_group *sg = group_head;
8332 for_each_cpu(j, sched_group_cpus(sg)) {
8333 struct sched_domain *sd;
8335 sd = &per_cpu(phys_domains, j).sd;
8336 if (j != group_first_cpu(sd->groups)) {
8338 * Only add "power" once for each
8344 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8347 } while (sg != group_head);
8350 static int build_numa_sched_groups(struct s_data *d,
8351 const struct cpumask *cpu_map, int num)
8353 struct sched_domain *sd;
8354 struct sched_group *sg, *prev;
8357 cpumask_clear(d->covered);
8358 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8359 if (cpumask_empty(d->nodemask)) {
8360 d->sched_group_nodes[num] = NULL;
8364 sched_domain_node_span(num, d->domainspan);
8365 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8367 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8370 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8374 d->sched_group_nodes[num] = sg;
8376 for_each_cpu(j, d->nodemask) {
8377 sd = &per_cpu(node_domains, j).sd;
8381 sg->__cpu_power = 0;
8382 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8384 cpumask_or(d->covered, d->covered, d->nodemask);
8387 for (j = 0; j < nr_node_ids; j++) {
8388 n = (num + j) % nr_node_ids;
8389 cpumask_complement(d->notcovered, d->covered);
8390 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8391 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8392 if (cpumask_empty(d->tmpmask))
8394 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8395 if (cpumask_empty(d->tmpmask))
8397 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8401 "Can not alloc domain group for node %d\n", j);
8404 sg->__cpu_power = 0;
8405 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8406 sg->next = prev->next;
8407 cpumask_or(d->covered, d->covered, d->tmpmask);
8414 #endif /* CONFIG_NUMA */
8417 /* Free memory allocated for various sched_group structures */
8418 static void free_sched_groups(const struct cpumask *cpu_map,
8419 struct cpumask *nodemask)
8423 for_each_cpu(cpu, cpu_map) {
8424 struct sched_group **sched_group_nodes
8425 = sched_group_nodes_bycpu[cpu];
8427 if (!sched_group_nodes)
8430 for (i = 0; i < nr_node_ids; i++) {
8431 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8433 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8434 if (cpumask_empty(nodemask))
8444 if (oldsg != sched_group_nodes[i])
8447 kfree(sched_group_nodes);
8448 sched_group_nodes_bycpu[cpu] = NULL;
8451 #else /* !CONFIG_NUMA */
8452 static void free_sched_groups(const struct cpumask *cpu_map,
8453 struct cpumask *nodemask)
8456 #endif /* CONFIG_NUMA */
8459 * Initialize sched groups cpu_power.
8461 * cpu_power indicates the capacity of sched group, which is used while
8462 * distributing the load between different sched groups in a sched domain.
8463 * Typically cpu_power for all the groups in a sched domain will be same unless
8464 * there are asymmetries in the topology. If there are asymmetries, group
8465 * having more cpu_power will pickup more load compared to the group having
8468 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8469 * the maximum number of tasks a group can handle in the presence of other idle
8470 * or lightly loaded groups in the same sched domain.
8472 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8474 struct sched_domain *child;
8475 struct sched_group *group;
8477 WARN_ON(!sd || !sd->groups);
8479 if (cpu != group_first_cpu(sd->groups))
8484 sd->groups->__cpu_power = 0;
8487 * For perf policy, if the groups in child domain share resources
8488 * (for example cores sharing some portions of the cache hierarchy
8489 * or SMT), then set this domain groups cpu_power such that each group
8490 * can handle only one task, when there are other idle groups in the
8491 * same sched domain.
8493 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8495 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8496 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8501 * add cpu_power of each child group to this groups cpu_power
8503 group = child->groups;
8505 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8506 group = group->next;
8507 } while (group != child->groups);
8511 * Initializers for schedule domains
8512 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8515 #ifdef CONFIG_SCHED_DEBUG
8516 # define SD_INIT_NAME(sd, type) sd->name = #type
8518 # define SD_INIT_NAME(sd, type) do { } while (0)
8521 #define SD_INIT(sd, type) sd_init_##type(sd)
8523 #define SD_INIT_FUNC(type) \
8524 static noinline void sd_init_##type(struct sched_domain *sd) \
8526 memset(sd, 0, sizeof(*sd)); \
8527 *sd = SD_##type##_INIT; \
8528 sd->level = SD_LV_##type; \
8529 SD_INIT_NAME(sd, type); \
8534 SD_INIT_FUNC(ALLNODES)
8537 #ifdef CONFIG_SCHED_SMT
8538 SD_INIT_FUNC(SIBLING)
8540 #ifdef CONFIG_SCHED_MC
8544 static int default_relax_domain_level = -1;
8546 static int __init setup_relax_domain_level(char *str)
8550 val = simple_strtoul(str, NULL, 0);
8551 if (val < SD_LV_MAX)
8552 default_relax_domain_level = val;
8556 __setup("relax_domain_level=", setup_relax_domain_level);
8558 static void set_domain_attribute(struct sched_domain *sd,
8559 struct sched_domain_attr *attr)
8563 if (!attr || attr->relax_domain_level < 0) {
8564 if (default_relax_domain_level < 0)
8567 request = default_relax_domain_level;
8569 request = attr->relax_domain_level;
8570 if (request < sd->level) {
8571 /* turn off idle balance on this domain */
8572 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8574 /* turn on idle balance on this domain */
8575 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8579 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8580 const struct cpumask *cpu_map)
8583 case sa_sched_groups:
8584 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8585 d->sched_group_nodes = NULL;
8587 free_rootdomain(d->rd); /* fall through */
8589 free_cpumask_var(d->tmpmask); /* fall through */
8590 case sa_send_covered:
8591 free_cpumask_var(d->send_covered); /* fall through */
8592 case sa_this_core_map:
8593 free_cpumask_var(d->this_core_map); /* fall through */
8594 case sa_this_sibling_map:
8595 free_cpumask_var(d->this_sibling_map); /* fall through */
8597 free_cpumask_var(d->nodemask); /* fall through */
8598 case sa_sched_group_nodes:
8600 kfree(d->sched_group_nodes); /* fall through */
8602 free_cpumask_var(d->notcovered); /* fall through */
8604 free_cpumask_var(d->covered); /* fall through */
8606 free_cpumask_var(d->domainspan); /* fall through */
8613 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8614 const struct cpumask *cpu_map)
8617 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8619 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8620 return sa_domainspan;
8621 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8623 /* Allocate the per-node list of sched groups */
8624 d->sched_group_nodes = kcalloc(nr_node_ids,
8625 sizeof(struct sched_group *), GFP_KERNEL);
8626 if (!d->sched_group_nodes) {
8627 printk(KERN_WARNING "Can not alloc sched group node list\n");
8628 return sa_notcovered;
8630 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8632 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8633 return sa_sched_group_nodes;
8634 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8636 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8637 return sa_this_sibling_map;
8638 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8639 return sa_this_core_map;
8640 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8641 return sa_send_covered;
8642 d->rd = alloc_rootdomain();
8644 printk(KERN_WARNING "Cannot alloc root domain\n");
8647 return sa_rootdomain;
8650 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8651 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8653 struct sched_domain *sd = NULL;
8655 struct sched_domain *parent;
8658 if (cpumask_weight(cpu_map) >
8659 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8660 sd = &per_cpu(allnodes_domains, i).sd;
8661 SD_INIT(sd, ALLNODES);
8662 set_domain_attribute(sd, attr);
8663 cpumask_copy(sched_domain_span(sd), cpu_map);
8664 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8669 sd = &per_cpu(node_domains, i).sd;
8671 set_domain_attribute(sd, attr);
8672 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8673 sd->parent = parent;
8676 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8681 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8682 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8683 struct sched_domain *parent, int i)
8685 struct sched_domain *sd;
8686 sd = &per_cpu(phys_domains, i).sd;
8688 set_domain_attribute(sd, attr);
8689 cpumask_copy(sched_domain_span(sd), d->nodemask);
8690 sd->parent = parent;
8693 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8697 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8698 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8699 struct sched_domain *parent, int i)
8701 struct sched_domain *sd = parent;
8702 #ifdef CONFIG_SCHED_MC
8703 sd = &per_cpu(core_domains, i).sd;
8705 set_domain_attribute(sd, attr);
8706 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8707 sd->parent = parent;
8709 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8714 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8715 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8716 struct sched_domain *parent, int i)
8718 struct sched_domain *sd = parent;
8719 #ifdef CONFIG_SCHED_SMT
8720 sd = &per_cpu(cpu_domains, i).sd;
8721 SD_INIT(sd, SIBLING);
8722 set_domain_attribute(sd, attr);
8723 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8724 sd->parent = parent;
8726 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8731 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8732 const struct cpumask *cpu_map, int cpu)
8735 #ifdef CONFIG_SCHED_SMT
8736 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8737 cpumask_and(d->this_sibling_map, cpu_map,
8738 topology_thread_cpumask(cpu));
8739 if (cpu == cpumask_first(d->this_sibling_map))
8740 init_sched_build_groups(d->this_sibling_map, cpu_map,
8742 d->send_covered, d->tmpmask);
8745 #ifdef CONFIG_SCHED_MC
8746 case SD_LV_MC: /* set up multi-core groups */
8747 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8748 if (cpu == cpumask_first(d->this_core_map))
8749 init_sched_build_groups(d->this_core_map, cpu_map,
8751 d->send_covered, d->tmpmask);
8754 case SD_LV_CPU: /* set up physical groups */
8755 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8756 if (!cpumask_empty(d->nodemask))
8757 init_sched_build_groups(d->nodemask, cpu_map,
8759 d->send_covered, d->tmpmask);
8762 case SD_LV_ALLNODES:
8763 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8764 d->send_covered, d->tmpmask);
8773 * Build sched domains for a given set of cpus and attach the sched domains
8774 * to the individual cpus
8776 static int __build_sched_domains(const struct cpumask *cpu_map,
8777 struct sched_domain_attr *attr)
8779 enum s_alloc alloc_state = sa_none;
8781 struct sched_domain *sd;
8787 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8788 if (alloc_state != sa_rootdomain)
8790 alloc_state = sa_sched_groups;
8793 * Set up domains for cpus specified by the cpu_map.
8795 for_each_cpu(i, cpu_map) {
8796 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8799 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8800 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8801 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8802 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8805 for_each_cpu(i, cpu_map) {
8806 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8807 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8810 /* Set up physical groups */
8811 for (i = 0; i < nr_node_ids; i++)
8812 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8815 /* Set up node groups */
8817 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8819 for (i = 0; i < nr_node_ids; i++)
8820 if (build_numa_sched_groups(&d, cpu_map, i))
8824 /* Calculate CPU power for physical packages and nodes */
8825 #ifdef CONFIG_SCHED_SMT
8826 for_each_cpu(i, cpu_map) {
8827 sd = &per_cpu(cpu_domains, i).sd;
8828 init_sched_groups_power(i, sd);
8831 #ifdef CONFIG_SCHED_MC
8832 for_each_cpu(i, cpu_map) {
8833 sd = &per_cpu(core_domains, i).sd;
8834 init_sched_groups_power(i, sd);
8838 for_each_cpu(i, cpu_map) {
8839 sd = &per_cpu(phys_domains, i).sd;
8840 init_sched_groups_power(i, sd);
8844 for (i = 0; i < nr_node_ids; i++)
8845 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8847 if (d.sd_allnodes) {
8848 struct sched_group *sg;
8850 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8852 init_numa_sched_groups_power(sg);
8856 /* Attach the domains */
8857 for_each_cpu(i, cpu_map) {
8858 #ifdef CONFIG_SCHED_SMT
8859 sd = &per_cpu(cpu_domains, i).sd;
8860 #elif defined(CONFIG_SCHED_MC)
8861 sd = &per_cpu(core_domains, i).sd;
8863 sd = &per_cpu(phys_domains, i).sd;
8865 cpu_attach_domain(sd, d.rd, i);
8868 d.sched_group_nodes = NULL; /* don't free this we still need it */
8869 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8873 __free_domain_allocs(&d, alloc_state, cpu_map);
8877 static int build_sched_domains(const struct cpumask *cpu_map)
8879 return __build_sched_domains(cpu_map, NULL);
8882 static struct cpumask *doms_cur; /* current sched domains */
8883 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8884 static struct sched_domain_attr *dattr_cur;
8885 /* attribues of custom domains in 'doms_cur' */
8888 * Special case: If a kmalloc of a doms_cur partition (array of
8889 * cpumask) fails, then fallback to a single sched domain,
8890 * as determined by the single cpumask fallback_doms.
8892 static cpumask_var_t fallback_doms;
8895 * arch_update_cpu_topology lets virtualized architectures update the
8896 * cpu core maps. It is supposed to return 1 if the topology changed
8897 * or 0 if it stayed the same.
8899 int __attribute__((weak)) arch_update_cpu_topology(void)
8905 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8906 * For now this just excludes isolated cpus, but could be used to
8907 * exclude other special cases in the future.
8909 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8913 arch_update_cpu_topology();
8915 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8917 doms_cur = fallback_doms;
8918 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8920 err = build_sched_domains(doms_cur);
8921 register_sched_domain_sysctl();
8926 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8927 struct cpumask *tmpmask)
8929 free_sched_groups(cpu_map, tmpmask);
8933 * Detach sched domains from a group of cpus specified in cpu_map
8934 * These cpus will now be attached to the NULL domain
8936 static void detach_destroy_domains(const struct cpumask *cpu_map)
8938 /* Save because hotplug lock held. */
8939 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8942 for_each_cpu(i, cpu_map)
8943 cpu_attach_domain(NULL, &def_root_domain, i);
8944 synchronize_sched();
8945 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8948 /* handle null as "default" */
8949 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8950 struct sched_domain_attr *new, int idx_new)
8952 struct sched_domain_attr tmp;
8959 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8960 new ? (new + idx_new) : &tmp,
8961 sizeof(struct sched_domain_attr));
8965 * Partition sched domains as specified by the 'ndoms_new'
8966 * cpumasks in the array doms_new[] of cpumasks. This compares
8967 * doms_new[] to the current sched domain partitioning, doms_cur[].
8968 * It destroys each deleted domain and builds each new domain.
8970 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8971 * The masks don't intersect (don't overlap.) We should setup one
8972 * sched domain for each mask. CPUs not in any of the cpumasks will
8973 * not be load balanced. If the same cpumask appears both in the
8974 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8977 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8978 * ownership of it and will kfree it when done with it. If the caller
8979 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8980 * ndoms_new == 1, and partition_sched_domains() will fallback to
8981 * the single partition 'fallback_doms', it also forces the domains
8984 * If doms_new == NULL it will be replaced with cpu_online_mask.
8985 * ndoms_new == 0 is a special case for destroying existing domains,
8986 * and it will not create the default domain.
8988 * Call with hotplug lock held
8990 /* FIXME: Change to struct cpumask *doms_new[] */
8991 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8992 struct sched_domain_attr *dattr_new)
8997 mutex_lock(&sched_domains_mutex);
8999 /* always unregister in case we don't destroy any domains */
9000 unregister_sched_domain_sysctl();
9002 /* Let architecture update cpu core mappings. */
9003 new_topology = arch_update_cpu_topology();
9005 n = doms_new ? ndoms_new : 0;
9007 /* Destroy deleted domains */
9008 for (i = 0; i < ndoms_cur; i++) {
9009 for (j = 0; j < n && !new_topology; j++) {
9010 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9011 && dattrs_equal(dattr_cur, i, dattr_new, j))
9014 /* no match - a current sched domain not in new doms_new[] */
9015 detach_destroy_domains(doms_cur + i);
9020 if (doms_new == NULL) {
9022 doms_new = fallback_doms;
9023 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9024 WARN_ON_ONCE(dattr_new);
9027 /* Build new domains */
9028 for (i = 0; i < ndoms_new; i++) {
9029 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9030 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9031 && dattrs_equal(dattr_new, i, dattr_cur, j))
9034 /* no match - add a new doms_new */
9035 __build_sched_domains(doms_new + i,
9036 dattr_new ? dattr_new + i : NULL);
9041 /* Remember the new sched domains */
9042 if (doms_cur != fallback_doms)
9044 kfree(dattr_cur); /* kfree(NULL) is safe */
9045 doms_cur = doms_new;
9046 dattr_cur = dattr_new;
9047 ndoms_cur = ndoms_new;
9049 register_sched_domain_sysctl();
9051 mutex_unlock(&sched_domains_mutex);
9054 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9055 static void arch_reinit_sched_domains(void)
9059 /* Destroy domains first to force the rebuild */
9060 partition_sched_domains(0, NULL, NULL);
9062 rebuild_sched_domains();
9066 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9068 unsigned int level = 0;
9070 if (sscanf(buf, "%u", &level) != 1)
9074 * level is always be positive so don't check for
9075 * level < POWERSAVINGS_BALANCE_NONE which is 0
9076 * What happens on 0 or 1 byte write,
9077 * need to check for count as well?
9080 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9084 sched_smt_power_savings = level;
9086 sched_mc_power_savings = level;
9088 arch_reinit_sched_domains();
9093 #ifdef CONFIG_SCHED_MC
9094 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9097 return sprintf(page, "%u\n", sched_mc_power_savings);
9099 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9100 const char *buf, size_t count)
9102 return sched_power_savings_store(buf, count, 0);
9104 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9105 sched_mc_power_savings_show,
9106 sched_mc_power_savings_store);
9109 #ifdef CONFIG_SCHED_SMT
9110 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9113 return sprintf(page, "%u\n", sched_smt_power_savings);
9115 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9116 const char *buf, size_t count)
9118 return sched_power_savings_store(buf, count, 1);
9120 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9121 sched_smt_power_savings_show,
9122 sched_smt_power_savings_store);
9125 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9129 #ifdef CONFIG_SCHED_SMT
9131 err = sysfs_create_file(&cls->kset.kobj,
9132 &attr_sched_smt_power_savings.attr);
9134 #ifdef CONFIG_SCHED_MC
9135 if (!err && mc_capable())
9136 err = sysfs_create_file(&cls->kset.kobj,
9137 &attr_sched_mc_power_savings.attr);
9141 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9143 #ifndef CONFIG_CPUSETS
9145 * Add online and remove offline CPUs from the scheduler domains.
9146 * When cpusets are enabled they take over this function.
9148 static int update_sched_domains(struct notifier_block *nfb,
9149 unsigned long action, void *hcpu)
9153 case CPU_ONLINE_FROZEN:
9155 case CPU_DEAD_FROZEN:
9156 partition_sched_domains(1, NULL, NULL);
9165 static int update_runtime(struct notifier_block *nfb,
9166 unsigned long action, void *hcpu)
9168 int cpu = (int)(long)hcpu;
9171 case CPU_DOWN_PREPARE:
9172 case CPU_DOWN_PREPARE_FROZEN:
9173 disable_runtime(cpu_rq(cpu));
9176 case CPU_DOWN_FAILED:
9177 case CPU_DOWN_FAILED_FROZEN:
9179 case CPU_ONLINE_FROZEN:
9180 enable_runtime(cpu_rq(cpu));
9188 void __init sched_init_smp(void)
9190 cpumask_var_t non_isolated_cpus;
9192 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9194 #if defined(CONFIG_NUMA)
9195 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9197 BUG_ON(sched_group_nodes_bycpu == NULL);
9200 mutex_lock(&sched_domains_mutex);
9201 arch_init_sched_domains(cpu_online_mask);
9202 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9203 if (cpumask_empty(non_isolated_cpus))
9204 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9205 mutex_unlock(&sched_domains_mutex);
9208 #ifndef CONFIG_CPUSETS
9209 /* XXX: Theoretical race here - CPU may be hotplugged now */
9210 hotcpu_notifier(update_sched_domains, 0);
9213 /* RT runtime code needs to handle some hotplug events */
9214 hotcpu_notifier(update_runtime, 0);
9218 /* Move init over to a non-isolated CPU */
9219 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9221 sched_init_granularity();
9222 free_cpumask_var(non_isolated_cpus);
9224 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9225 init_sched_rt_class();
9228 void __init sched_init_smp(void)
9230 sched_init_granularity();
9232 #endif /* CONFIG_SMP */
9234 const_debug unsigned int sysctl_timer_migration = 1;
9236 int in_sched_functions(unsigned long addr)
9238 return in_lock_functions(addr) ||
9239 (addr >= (unsigned long)__sched_text_start
9240 && addr < (unsigned long)__sched_text_end);
9243 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9245 cfs_rq->tasks_timeline = RB_ROOT;
9246 INIT_LIST_HEAD(&cfs_rq->tasks);
9247 #ifdef CONFIG_FAIR_GROUP_SCHED
9250 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9253 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9255 struct rt_prio_array *array;
9258 array = &rt_rq->active;
9259 for (i = 0; i < MAX_RT_PRIO; i++) {
9260 INIT_LIST_HEAD(array->queue + i);
9261 __clear_bit(i, array->bitmap);
9263 /* delimiter for bitsearch: */
9264 __set_bit(MAX_RT_PRIO, array->bitmap);
9266 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9267 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9269 rt_rq->highest_prio.next = MAX_RT_PRIO;
9273 rt_rq->rt_nr_migratory = 0;
9274 rt_rq->overloaded = 0;
9275 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9279 rt_rq->rt_throttled = 0;
9280 rt_rq->rt_runtime = 0;
9281 spin_lock_init(&rt_rq->rt_runtime_lock);
9283 #ifdef CONFIG_RT_GROUP_SCHED
9284 rt_rq->rt_nr_boosted = 0;
9289 #ifdef CONFIG_FAIR_GROUP_SCHED
9290 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9291 struct sched_entity *se, int cpu, int add,
9292 struct sched_entity *parent)
9294 struct rq *rq = cpu_rq(cpu);
9295 tg->cfs_rq[cpu] = cfs_rq;
9296 init_cfs_rq(cfs_rq, rq);
9299 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9302 /* se could be NULL for init_task_group */
9307 se->cfs_rq = &rq->cfs;
9309 se->cfs_rq = parent->my_q;
9312 se->load.weight = tg->shares;
9313 se->load.inv_weight = 0;
9314 se->parent = parent;
9318 #ifdef CONFIG_RT_GROUP_SCHED
9319 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9320 struct sched_rt_entity *rt_se, int cpu, int add,
9321 struct sched_rt_entity *parent)
9323 struct rq *rq = cpu_rq(cpu);
9325 tg->rt_rq[cpu] = rt_rq;
9326 init_rt_rq(rt_rq, rq);
9328 rt_rq->rt_se = rt_se;
9329 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9331 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9333 tg->rt_se[cpu] = rt_se;
9338 rt_se->rt_rq = &rq->rt;
9340 rt_se->rt_rq = parent->my_q;
9342 rt_se->my_q = rt_rq;
9343 rt_se->parent = parent;
9344 INIT_LIST_HEAD(&rt_se->run_list);
9348 void __init sched_init(void)
9351 unsigned long alloc_size = 0, ptr;
9353 #ifdef CONFIG_FAIR_GROUP_SCHED
9354 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9356 #ifdef CONFIG_RT_GROUP_SCHED
9357 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9359 #ifdef CONFIG_USER_SCHED
9362 #ifdef CONFIG_CPUMASK_OFFSTACK
9363 alloc_size += num_possible_cpus() * cpumask_size();
9366 * As sched_init() is called before page_alloc is setup,
9367 * we use alloc_bootmem().
9370 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9372 #ifdef CONFIG_FAIR_GROUP_SCHED
9373 init_task_group.se = (struct sched_entity **)ptr;
9374 ptr += nr_cpu_ids * sizeof(void **);
9376 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9377 ptr += nr_cpu_ids * sizeof(void **);
9379 #ifdef CONFIG_USER_SCHED
9380 root_task_group.se = (struct sched_entity **)ptr;
9381 ptr += nr_cpu_ids * sizeof(void **);
9383 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9384 ptr += nr_cpu_ids * sizeof(void **);
9385 #endif /* CONFIG_USER_SCHED */
9386 #endif /* CONFIG_FAIR_GROUP_SCHED */
9387 #ifdef CONFIG_RT_GROUP_SCHED
9388 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9389 ptr += nr_cpu_ids * sizeof(void **);
9391 init_task_group.rt_rq = (struct rt_rq **)ptr;
9392 ptr += nr_cpu_ids * sizeof(void **);
9394 #ifdef CONFIG_USER_SCHED
9395 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9396 ptr += nr_cpu_ids * sizeof(void **);
9398 root_task_group.rt_rq = (struct rt_rq **)ptr;
9399 ptr += nr_cpu_ids * sizeof(void **);
9400 #endif /* CONFIG_USER_SCHED */
9401 #endif /* CONFIG_RT_GROUP_SCHED */
9402 #ifdef CONFIG_CPUMASK_OFFSTACK
9403 for_each_possible_cpu(i) {
9404 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9405 ptr += cpumask_size();
9407 #endif /* CONFIG_CPUMASK_OFFSTACK */
9411 init_defrootdomain();
9414 init_rt_bandwidth(&def_rt_bandwidth,
9415 global_rt_period(), global_rt_runtime());
9417 #ifdef CONFIG_RT_GROUP_SCHED
9418 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9419 global_rt_period(), global_rt_runtime());
9420 #ifdef CONFIG_USER_SCHED
9421 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9422 global_rt_period(), RUNTIME_INF);
9423 #endif /* CONFIG_USER_SCHED */
9424 #endif /* CONFIG_RT_GROUP_SCHED */
9426 #ifdef CONFIG_GROUP_SCHED
9427 list_add(&init_task_group.list, &task_groups);
9428 INIT_LIST_HEAD(&init_task_group.children);
9430 #ifdef CONFIG_USER_SCHED
9431 INIT_LIST_HEAD(&root_task_group.children);
9432 init_task_group.parent = &root_task_group;
9433 list_add(&init_task_group.siblings, &root_task_group.children);
9434 #endif /* CONFIG_USER_SCHED */
9435 #endif /* CONFIG_GROUP_SCHED */
9437 for_each_possible_cpu(i) {
9441 spin_lock_init(&rq->lock);
9443 rq->calc_load_active = 0;
9444 rq->calc_load_update = jiffies + LOAD_FREQ;
9445 init_cfs_rq(&rq->cfs, rq);
9446 init_rt_rq(&rq->rt, rq);
9447 #ifdef CONFIG_FAIR_GROUP_SCHED
9448 init_task_group.shares = init_task_group_load;
9449 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9450 #ifdef CONFIG_CGROUP_SCHED
9452 * How much cpu bandwidth does init_task_group get?
9454 * In case of task-groups formed thr' the cgroup filesystem, it
9455 * gets 100% of the cpu resources in the system. This overall
9456 * system cpu resource is divided among the tasks of
9457 * init_task_group and its child task-groups in a fair manner,
9458 * based on each entity's (task or task-group's) weight
9459 * (se->load.weight).
9461 * In other words, if init_task_group has 10 tasks of weight
9462 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9463 * then A0's share of the cpu resource is:
9465 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9467 * We achieve this by letting init_task_group's tasks sit
9468 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9470 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9471 #elif defined CONFIG_USER_SCHED
9472 root_task_group.shares = NICE_0_LOAD;
9473 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9475 * In case of task-groups formed thr' the user id of tasks,
9476 * init_task_group represents tasks belonging to root user.
9477 * Hence it forms a sibling of all subsequent groups formed.
9478 * In this case, init_task_group gets only a fraction of overall
9479 * system cpu resource, based on the weight assigned to root
9480 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9481 * by letting tasks of init_task_group sit in a separate cfs_rq
9482 * (init_tg_cfs_rq) and having one entity represent this group of
9483 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9485 init_tg_cfs_entry(&init_task_group,
9486 &per_cpu(init_tg_cfs_rq, i),
9487 &per_cpu(init_sched_entity, i), i, 1,
9488 root_task_group.se[i]);
9491 #endif /* CONFIG_FAIR_GROUP_SCHED */
9493 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9494 #ifdef CONFIG_RT_GROUP_SCHED
9495 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9496 #ifdef CONFIG_CGROUP_SCHED
9497 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9498 #elif defined CONFIG_USER_SCHED
9499 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9500 init_tg_rt_entry(&init_task_group,
9501 &per_cpu(init_rt_rq, i),
9502 &per_cpu(init_sched_rt_entity, i), i, 1,
9503 root_task_group.rt_se[i]);
9507 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9508 rq->cpu_load[j] = 0;
9512 rq->post_schedule = 0;
9513 rq->active_balance = 0;
9514 rq->next_balance = jiffies;
9518 rq->migration_thread = NULL;
9519 INIT_LIST_HEAD(&rq->migration_queue);
9520 rq_attach_root(rq, &def_root_domain);
9523 atomic_set(&rq->nr_iowait, 0);
9526 set_load_weight(&init_task);
9528 #ifdef CONFIG_PREEMPT_NOTIFIERS
9529 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9533 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9536 #ifdef CONFIG_RT_MUTEXES
9537 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9541 * The boot idle thread does lazy MMU switching as well:
9543 atomic_inc(&init_mm.mm_count);
9544 enter_lazy_tlb(&init_mm, current);
9547 * Make us the idle thread. Technically, schedule() should not be
9548 * called from this thread, however somewhere below it might be,
9549 * but because we are the idle thread, we just pick up running again
9550 * when this runqueue becomes "idle".
9552 init_idle(current, smp_processor_id());
9554 calc_load_update = jiffies + LOAD_FREQ;
9557 * During early bootup we pretend to be a normal task:
9559 current->sched_class = &fair_sched_class;
9561 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9562 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9565 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9566 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9568 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9571 perf_counter_init();
9573 scheduler_running = 1;
9576 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9577 static inline int preempt_count_equals(int preempt_offset)
9579 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9581 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9584 void __might_sleep(char *file, int line, int preempt_offset)
9587 static unsigned long prev_jiffy; /* ratelimiting */
9589 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9590 system_state != SYSTEM_RUNNING || oops_in_progress)
9592 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9594 prev_jiffy = jiffies;
9597 "BUG: sleeping function called from invalid context at %s:%d\n",
9600 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9601 in_atomic(), irqs_disabled(),
9602 current->pid, current->comm);
9604 debug_show_held_locks(current);
9605 if (irqs_disabled())
9606 print_irqtrace_events(current);
9610 EXPORT_SYMBOL(__might_sleep);
9613 #ifdef CONFIG_MAGIC_SYSRQ
9614 static void normalize_task(struct rq *rq, struct task_struct *p)
9618 update_rq_clock(rq);
9619 on_rq = p->se.on_rq;
9621 deactivate_task(rq, p, 0);
9622 __setscheduler(rq, p, SCHED_NORMAL, 0);
9624 activate_task(rq, p, 0);
9625 resched_task(rq->curr);
9629 void normalize_rt_tasks(void)
9631 struct task_struct *g, *p;
9632 unsigned long flags;
9635 read_lock_irqsave(&tasklist_lock, flags);
9636 do_each_thread(g, p) {
9638 * Only normalize user tasks:
9643 p->se.exec_start = 0;
9644 #ifdef CONFIG_SCHEDSTATS
9645 p->se.wait_start = 0;
9646 p->se.sleep_start = 0;
9647 p->se.block_start = 0;
9652 * Renice negative nice level userspace
9655 if (TASK_NICE(p) < 0 && p->mm)
9656 set_user_nice(p, 0);
9660 spin_lock(&p->pi_lock);
9661 rq = __task_rq_lock(p);
9663 normalize_task(rq, p);
9665 __task_rq_unlock(rq);
9666 spin_unlock(&p->pi_lock);
9667 } while_each_thread(g, p);
9669 read_unlock_irqrestore(&tasklist_lock, flags);
9672 #endif /* CONFIG_MAGIC_SYSRQ */
9676 * These functions are only useful for the IA64 MCA handling.
9678 * They can only be called when the whole system has been
9679 * stopped - every CPU needs to be quiescent, and no scheduling
9680 * activity can take place. Using them for anything else would
9681 * be a serious bug, and as a result, they aren't even visible
9682 * under any other configuration.
9686 * curr_task - return the current task for a given cpu.
9687 * @cpu: the processor in question.
9689 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9691 struct task_struct *curr_task(int cpu)
9693 return cpu_curr(cpu);
9697 * set_curr_task - set the current task for a given cpu.
9698 * @cpu: the processor in question.
9699 * @p: the task pointer to set.
9701 * Description: This function must only be used when non-maskable interrupts
9702 * are serviced on a separate stack. It allows the architecture to switch the
9703 * notion of the current task on a cpu in a non-blocking manner. This function
9704 * must be called with all CPU's synchronized, and interrupts disabled, the
9705 * and caller must save the original value of the current task (see
9706 * curr_task() above) and restore that value before reenabling interrupts and
9707 * re-starting the system.
9709 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9711 void set_curr_task(int cpu, struct task_struct *p)
9718 #ifdef CONFIG_FAIR_GROUP_SCHED
9719 static void free_fair_sched_group(struct task_group *tg)
9723 for_each_possible_cpu(i) {
9725 kfree(tg->cfs_rq[i]);
9735 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9737 struct cfs_rq *cfs_rq;
9738 struct sched_entity *se;
9742 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9745 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9749 tg->shares = NICE_0_LOAD;
9751 for_each_possible_cpu(i) {
9754 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9755 GFP_KERNEL, cpu_to_node(i));
9759 se = kzalloc_node(sizeof(struct sched_entity),
9760 GFP_KERNEL, cpu_to_node(i));
9764 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9773 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9775 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9776 &cpu_rq(cpu)->leaf_cfs_rq_list);
9779 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9781 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9783 #else /* !CONFG_FAIR_GROUP_SCHED */
9784 static inline void free_fair_sched_group(struct task_group *tg)
9789 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9794 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9798 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9801 #endif /* CONFIG_FAIR_GROUP_SCHED */
9803 #ifdef CONFIG_RT_GROUP_SCHED
9804 static void free_rt_sched_group(struct task_group *tg)
9808 destroy_rt_bandwidth(&tg->rt_bandwidth);
9810 for_each_possible_cpu(i) {
9812 kfree(tg->rt_rq[i]);
9814 kfree(tg->rt_se[i]);
9822 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9824 struct rt_rq *rt_rq;
9825 struct sched_rt_entity *rt_se;
9829 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9832 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9836 init_rt_bandwidth(&tg->rt_bandwidth,
9837 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9839 for_each_possible_cpu(i) {
9842 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9843 GFP_KERNEL, cpu_to_node(i));
9847 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9848 GFP_KERNEL, cpu_to_node(i));
9852 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9861 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9863 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9864 &cpu_rq(cpu)->leaf_rt_rq_list);
9867 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9869 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9871 #else /* !CONFIG_RT_GROUP_SCHED */
9872 static inline void free_rt_sched_group(struct task_group *tg)
9877 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9882 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9886 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9889 #endif /* CONFIG_RT_GROUP_SCHED */
9891 #ifdef CONFIG_GROUP_SCHED
9892 static void free_sched_group(struct task_group *tg)
9894 free_fair_sched_group(tg);
9895 free_rt_sched_group(tg);
9899 /* allocate runqueue etc for a new task group */
9900 struct task_group *sched_create_group(struct task_group *parent)
9902 struct task_group *tg;
9903 unsigned long flags;
9906 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9908 return ERR_PTR(-ENOMEM);
9910 if (!alloc_fair_sched_group(tg, parent))
9913 if (!alloc_rt_sched_group(tg, parent))
9916 spin_lock_irqsave(&task_group_lock, flags);
9917 for_each_possible_cpu(i) {
9918 register_fair_sched_group(tg, i);
9919 register_rt_sched_group(tg, i);
9921 list_add_rcu(&tg->list, &task_groups);
9923 WARN_ON(!parent); /* root should already exist */
9925 tg->parent = parent;
9926 INIT_LIST_HEAD(&tg->children);
9927 list_add_rcu(&tg->siblings, &parent->children);
9928 spin_unlock_irqrestore(&task_group_lock, flags);
9933 free_sched_group(tg);
9934 return ERR_PTR(-ENOMEM);
9937 /* rcu callback to free various structures associated with a task group */
9938 static void free_sched_group_rcu(struct rcu_head *rhp)
9940 /* now it should be safe to free those cfs_rqs */
9941 free_sched_group(container_of(rhp, struct task_group, rcu));
9944 /* Destroy runqueue etc associated with a task group */
9945 void sched_destroy_group(struct task_group *tg)
9947 unsigned long flags;
9950 spin_lock_irqsave(&task_group_lock, flags);
9951 for_each_possible_cpu(i) {
9952 unregister_fair_sched_group(tg, i);
9953 unregister_rt_sched_group(tg, i);
9955 list_del_rcu(&tg->list);
9956 list_del_rcu(&tg->siblings);
9957 spin_unlock_irqrestore(&task_group_lock, flags);
9959 /* wait for possible concurrent references to cfs_rqs complete */
9960 call_rcu(&tg->rcu, free_sched_group_rcu);
9963 /* change task's runqueue when it moves between groups.
9964 * The caller of this function should have put the task in its new group
9965 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9966 * reflect its new group.
9968 void sched_move_task(struct task_struct *tsk)
9971 unsigned long flags;
9974 rq = task_rq_lock(tsk, &flags);
9976 update_rq_clock(rq);
9978 running = task_current(rq, tsk);
9979 on_rq = tsk->se.on_rq;
9982 dequeue_task(rq, tsk, 0);
9983 if (unlikely(running))
9984 tsk->sched_class->put_prev_task(rq, tsk);
9986 set_task_rq(tsk, task_cpu(tsk));
9988 #ifdef CONFIG_FAIR_GROUP_SCHED
9989 if (tsk->sched_class->moved_group)
9990 tsk->sched_class->moved_group(tsk);
9993 if (unlikely(running))
9994 tsk->sched_class->set_curr_task(rq);
9996 enqueue_task(rq, tsk, 0);
9998 task_rq_unlock(rq, &flags);
10000 #endif /* CONFIG_GROUP_SCHED */
10002 #ifdef CONFIG_FAIR_GROUP_SCHED
10003 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10005 struct cfs_rq *cfs_rq = se->cfs_rq;
10010 dequeue_entity(cfs_rq, se, 0);
10012 se->load.weight = shares;
10013 se->load.inv_weight = 0;
10016 enqueue_entity(cfs_rq, se, 0);
10019 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10021 struct cfs_rq *cfs_rq = se->cfs_rq;
10022 struct rq *rq = cfs_rq->rq;
10023 unsigned long flags;
10025 spin_lock_irqsave(&rq->lock, flags);
10026 __set_se_shares(se, shares);
10027 spin_unlock_irqrestore(&rq->lock, flags);
10030 static DEFINE_MUTEX(shares_mutex);
10032 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10035 unsigned long flags;
10038 * We can't change the weight of the root cgroup.
10043 if (shares < MIN_SHARES)
10044 shares = MIN_SHARES;
10045 else if (shares > MAX_SHARES)
10046 shares = MAX_SHARES;
10048 mutex_lock(&shares_mutex);
10049 if (tg->shares == shares)
10052 spin_lock_irqsave(&task_group_lock, flags);
10053 for_each_possible_cpu(i)
10054 unregister_fair_sched_group(tg, i);
10055 list_del_rcu(&tg->siblings);
10056 spin_unlock_irqrestore(&task_group_lock, flags);
10058 /* wait for any ongoing reference to this group to finish */
10059 synchronize_sched();
10062 * Now we are free to modify the group's share on each cpu
10063 * w/o tripping rebalance_share or load_balance_fair.
10065 tg->shares = shares;
10066 for_each_possible_cpu(i) {
10068 * force a rebalance
10070 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10071 set_se_shares(tg->se[i], shares);
10075 * Enable load balance activity on this group, by inserting it back on
10076 * each cpu's rq->leaf_cfs_rq_list.
10078 spin_lock_irqsave(&task_group_lock, flags);
10079 for_each_possible_cpu(i)
10080 register_fair_sched_group(tg, i);
10081 list_add_rcu(&tg->siblings, &tg->parent->children);
10082 spin_unlock_irqrestore(&task_group_lock, flags);
10084 mutex_unlock(&shares_mutex);
10088 unsigned long sched_group_shares(struct task_group *tg)
10094 #ifdef CONFIG_RT_GROUP_SCHED
10096 * Ensure that the real time constraints are schedulable.
10098 static DEFINE_MUTEX(rt_constraints_mutex);
10100 static unsigned long to_ratio(u64 period, u64 runtime)
10102 if (runtime == RUNTIME_INF)
10105 return div64_u64(runtime << 20, period);
10108 /* Must be called with tasklist_lock held */
10109 static inline int tg_has_rt_tasks(struct task_group *tg)
10111 struct task_struct *g, *p;
10113 do_each_thread(g, p) {
10114 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10116 } while_each_thread(g, p);
10121 struct rt_schedulable_data {
10122 struct task_group *tg;
10127 static int tg_schedulable(struct task_group *tg, void *data)
10129 struct rt_schedulable_data *d = data;
10130 struct task_group *child;
10131 unsigned long total, sum = 0;
10132 u64 period, runtime;
10134 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10135 runtime = tg->rt_bandwidth.rt_runtime;
10138 period = d->rt_period;
10139 runtime = d->rt_runtime;
10142 #ifdef CONFIG_USER_SCHED
10143 if (tg == &root_task_group) {
10144 period = global_rt_period();
10145 runtime = global_rt_runtime();
10150 * Cannot have more runtime than the period.
10152 if (runtime > period && runtime != RUNTIME_INF)
10156 * Ensure we don't starve existing RT tasks.
10158 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10161 total = to_ratio(period, runtime);
10164 * Nobody can have more than the global setting allows.
10166 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10170 * The sum of our children's runtime should not exceed our own.
10172 list_for_each_entry_rcu(child, &tg->children, siblings) {
10173 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10174 runtime = child->rt_bandwidth.rt_runtime;
10176 if (child == d->tg) {
10177 period = d->rt_period;
10178 runtime = d->rt_runtime;
10181 sum += to_ratio(period, runtime);
10190 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10192 struct rt_schedulable_data data = {
10194 .rt_period = period,
10195 .rt_runtime = runtime,
10198 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10201 static int tg_set_bandwidth(struct task_group *tg,
10202 u64 rt_period, u64 rt_runtime)
10206 mutex_lock(&rt_constraints_mutex);
10207 read_lock(&tasklist_lock);
10208 err = __rt_schedulable(tg, rt_period, rt_runtime);
10212 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10213 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10214 tg->rt_bandwidth.rt_runtime = rt_runtime;
10216 for_each_possible_cpu(i) {
10217 struct rt_rq *rt_rq = tg->rt_rq[i];
10219 spin_lock(&rt_rq->rt_runtime_lock);
10220 rt_rq->rt_runtime = rt_runtime;
10221 spin_unlock(&rt_rq->rt_runtime_lock);
10223 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10225 read_unlock(&tasklist_lock);
10226 mutex_unlock(&rt_constraints_mutex);
10231 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10233 u64 rt_runtime, rt_period;
10235 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10236 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10237 if (rt_runtime_us < 0)
10238 rt_runtime = RUNTIME_INF;
10240 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10243 long sched_group_rt_runtime(struct task_group *tg)
10247 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10250 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10251 do_div(rt_runtime_us, NSEC_PER_USEC);
10252 return rt_runtime_us;
10255 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10257 u64 rt_runtime, rt_period;
10259 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10260 rt_runtime = tg->rt_bandwidth.rt_runtime;
10262 if (rt_period == 0)
10265 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10268 long sched_group_rt_period(struct task_group *tg)
10272 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10273 do_div(rt_period_us, NSEC_PER_USEC);
10274 return rt_period_us;
10277 static int sched_rt_global_constraints(void)
10279 u64 runtime, period;
10282 if (sysctl_sched_rt_period <= 0)
10285 runtime = global_rt_runtime();
10286 period = global_rt_period();
10289 * Sanity check on the sysctl variables.
10291 if (runtime > period && runtime != RUNTIME_INF)
10294 mutex_lock(&rt_constraints_mutex);
10295 read_lock(&tasklist_lock);
10296 ret = __rt_schedulable(NULL, 0, 0);
10297 read_unlock(&tasklist_lock);
10298 mutex_unlock(&rt_constraints_mutex);
10303 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10305 /* Don't accept realtime tasks when there is no way for them to run */
10306 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10312 #else /* !CONFIG_RT_GROUP_SCHED */
10313 static int sched_rt_global_constraints(void)
10315 unsigned long flags;
10318 if (sysctl_sched_rt_period <= 0)
10322 * There's always some RT tasks in the root group
10323 * -- migration, kstopmachine etc..
10325 if (sysctl_sched_rt_runtime == 0)
10328 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10329 for_each_possible_cpu(i) {
10330 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10332 spin_lock(&rt_rq->rt_runtime_lock);
10333 rt_rq->rt_runtime = global_rt_runtime();
10334 spin_unlock(&rt_rq->rt_runtime_lock);
10336 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10340 #endif /* CONFIG_RT_GROUP_SCHED */
10342 int sched_rt_handler(struct ctl_table *table, int write,
10343 struct file *filp, void __user *buffer, size_t *lenp,
10347 int old_period, old_runtime;
10348 static DEFINE_MUTEX(mutex);
10350 mutex_lock(&mutex);
10351 old_period = sysctl_sched_rt_period;
10352 old_runtime = sysctl_sched_rt_runtime;
10354 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10356 if (!ret && write) {
10357 ret = sched_rt_global_constraints();
10359 sysctl_sched_rt_period = old_period;
10360 sysctl_sched_rt_runtime = old_runtime;
10362 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10363 def_rt_bandwidth.rt_period =
10364 ns_to_ktime(global_rt_period());
10367 mutex_unlock(&mutex);
10372 #ifdef CONFIG_CGROUP_SCHED
10374 /* return corresponding task_group object of a cgroup */
10375 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10377 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10378 struct task_group, css);
10381 static struct cgroup_subsys_state *
10382 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10384 struct task_group *tg, *parent;
10386 if (!cgrp->parent) {
10387 /* This is early initialization for the top cgroup */
10388 return &init_task_group.css;
10391 parent = cgroup_tg(cgrp->parent);
10392 tg = sched_create_group(parent);
10394 return ERR_PTR(-ENOMEM);
10400 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10402 struct task_group *tg = cgroup_tg(cgrp);
10404 sched_destroy_group(tg);
10408 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10409 struct task_struct *tsk)
10411 #ifdef CONFIG_RT_GROUP_SCHED
10412 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10415 /* We don't support RT-tasks being in separate groups */
10416 if (tsk->sched_class != &fair_sched_class)
10424 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10425 struct cgroup *old_cont, struct task_struct *tsk)
10427 sched_move_task(tsk);
10430 #ifdef CONFIG_FAIR_GROUP_SCHED
10431 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10434 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10437 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10439 struct task_group *tg = cgroup_tg(cgrp);
10441 return (u64) tg->shares;
10443 #endif /* CONFIG_FAIR_GROUP_SCHED */
10445 #ifdef CONFIG_RT_GROUP_SCHED
10446 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10449 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10452 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10454 return sched_group_rt_runtime(cgroup_tg(cgrp));
10457 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10460 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10463 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10465 return sched_group_rt_period(cgroup_tg(cgrp));
10467 #endif /* CONFIG_RT_GROUP_SCHED */
10469 static struct cftype cpu_files[] = {
10470 #ifdef CONFIG_FAIR_GROUP_SCHED
10473 .read_u64 = cpu_shares_read_u64,
10474 .write_u64 = cpu_shares_write_u64,
10477 #ifdef CONFIG_RT_GROUP_SCHED
10479 .name = "rt_runtime_us",
10480 .read_s64 = cpu_rt_runtime_read,
10481 .write_s64 = cpu_rt_runtime_write,
10484 .name = "rt_period_us",
10485 .read_u64 = cpu_rt_period_read_uint,
10486 .write_u64 = cpu_rt_period_write_uint,
10491 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10493 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10496 struct cgroup_subsys cpu_cgroup_subsys = {
10498 .create = cpu_cgroup_create,
10499 .destroy = cpu_cgroup_destroy,
10500 .can_attach = cpu_cgroup_can_attach,
10501 .attach = cpu_cgroup_attach,
10502 .populate = cpu_cgroup_populate,
10503 .subsys_id = cpu_cgroup_subsys_id,
10507 #endif /* CONFIG_CGROUP_SCHED */
10509 #ifdef CONFIG_CGROUP_CPUACCT
10512 * CPU accounting code for task groups.
10514 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10515 * (balbir@in.ibm.com).
10518 /* track cpu usage of a group of tasks and its child groups */
10520 struct cgroup_subsys_state css;
10521 /* cpuusage holds pointer to a u64-type object on every cpu */
10523 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10524 struct cpuacct *parent;
10527 struct cgroup_subsys cpuacct_subsys;
10529 /* return cpu accounting group corresponding to this container */
10530 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10532 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10533 struct cpuacct, css);
10536 /* return cpu accounting group to which this task belongs */
10537 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10539 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10540 struct cpuacct, css);
10543 /* create a new cpu accounting group */
10544 static struct cgroup_subsys_state *cpuacct_create(
10545 struct cgroup_subsys *ss, struct cgroup *cgrp)
10547 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10553 ca->cpuusage = alloc_percpu(u64);
10557 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10558 if (percpu_counter_init(&ca->cpustat[i], 0))
10559 goto out_free_counters;
10562 ca->parent = cgroup_ca(cgrp->parent);
10568 percpu_counter_destroy(&ca->cpustat[i]);
10569 free_percpu(ca->cpuusage);
10573 return ERR_PTR(-ENOMEM);
10576 /* destroy an existing cpu accounting group */
10578 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10580 struct cpuacct *ca = cgroup_ca(cgrp);
10583 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10584 percpu_counter_destroy(&ca->cpustat[i]);
10585 free_percpu(ca->cpuusage);
10589 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10591 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10594 #ifndef CONFIG_64BIT
10596 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10598 spin_lock_irq(&cpu_rq(cpu)->lock);
10600 spin_unlock_irq(&cpu_rq(cpu)->lock);
10608 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10610 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10612 #ifndef CONFIG_64BIT
10614 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10616 spin_lock_irq(&cpu_rq(cpu)->lock);
10618 spin_unlock_irq(&cpu_rq(cpu)->lock);
10624 /* return total cpu usage (in nanoseconds) of a group */
10625 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10627 struct cpuacct *ca = cgroup_ca(cgrp);
10628 u64 totalcpuusage = 0;
10631 for_each_present_cpu(i)
10632 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10634 return totalcpuusage;
10637 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10640 struct cpuacct *ca = cgroup_ca(cgrp);
10649 for_each_present_cpu(i)
10650 cpuacct_cpuusage_write(ca, i, 0);
10656 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10657 struct seq_file *m)
10659 struct cpuacct *ca = cgroup_ca(cgroup);
10663 for_each_present_cpu(i) {
10664 percpu = cpuacct_cpuusage_read(ca, i);
10665 seq_printf(m, "%llu ", (unsigned long long) percpu);
10667 seq_printf(m, "\n");
10671 static const char *cpuacct_stat_desc[] = {
10672 [CPUACCT_STAT_USER] = "user",
10673 [CPUACCT_STAT_SYSTEM] = "system",
10676 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10677 struct cgroup_map_cb *cb)
10679 struct cpuacct *ca = cgroup_ca(cgrp);
10682 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10683 s64 val = percpu_counter_read(&ca->cpustat[i]);
10684 val = cputime64_to_clock_t(val);
10685 cb->fill(cb, cpuacct_stat_desc[i], val);
10690 static struct cftype files[] = {
10693 .read_u64 = cpuusage_read,
10694 .write_u64 = cpuusage_write,
10697 .name = "usage_percpu",
10698 .read_seq_string = cpuacct_percpu_seq_read,
10702 .read_map = cpuacct_stats_show,
10706 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10708 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10712 * charge this task's execution time to its accounting group.
10714 * called with rq->lock held.
10716 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10718 struct cpuacct *ca;
10721 if (unlikely(!cpuacct_subsys.active))
10724 cpu = task_cpu(tsk);
10730 for (; ca; ca = ca->parent) {
10731 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10732 *cpuusage += cputime;
10739 * Charge the system/user time to the task's accounting group.
10741 static void cpuacct_update_stats(struct task_struct *tsk,
10742 enum cpuacct_stat_index idx, cputime_t val)
10744 struct cpuacct *ca;
10746 if (unlikely(!cpuacct_subsys.active))
10753 percpu_counter_add(&ca->cpustat[idx], val);
10759 struct cgroup_subsys cpuacct_subsys = {
10761 .create = cpuacct_create,
10762 .destroy = cpuacct_destroy,
10763 .populate = cpuacct_populate,
10764 .subsys_id = cpuacct_subsys_id,
10766 #endif /* CONFIG_CGROUP_CPUACCT */