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/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
124 static inline int rt_policy(int policy)
126 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
131 static inline int task_has_rt_policy(struct task_struct *p)
133 return rt_policy(p->policy);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array {
140 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
141 struct list_head queue[MAX_RT_PRIO];
144 struct rt_bandwidth {
145 /* nests inside the rq lock: */
146 spinlock_t rt_runtime_lock;
149 struct hrtimer rt_period_timer;
152 static struct rt_bandwidth def_rt_bandwidth;
154 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
156 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
158 struct rt_bandwidth *rt_b =
159 container_of(timer, struct rt_bandwidth, rt_period_timer);
165 now = hrtimer_cb_get_time(timer);
166 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
171 idle = do_sched_rt_period_timer(rt_b, overrun);
174 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
178 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
180 rt_b->rt_period = ns_to_ktime(period);
181 rt_b->rt_runtime = runtime;
183 spin_lock_init(&rt_b->rt_runtime_lock);
185 hrtimer_init(&rt_b->rt_period_timer,
186 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
187 rt_b->rt_period_timer.function = sched_rt_period_timer;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime >= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
199 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
202 if (hrtimer_active(&rt_b->rt_period_timer))
205 spin_lock(&rt_b->rt_runtime_lock);
210 if (hrtimer_active(&rt_b->rt_period_timer))
213 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
214 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
216 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
217 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
218 delta = ktime_to_ns(ktime_sub(hard, soft));
219 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
220 HRTIMER_MODE_ABS_PINNED, 0);
222 spin_unlock(&rt_b->rt_runtime_lock);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
228 hrtimer_cancel(&rt_b->rt_period_timer);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex);
238 #ifdef CONFIG_GROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups);
246 /* task group related information */
248 #ifdef CONFIG_CGROUP_SCHED
249 struct cgroup_subsys_state css;
252 #ifdef CONFIG_USER_SCHED
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
281 /* Helper function to pass uid information to create_sched_user() */
282 void set_tg_uid(struct user_struct *user)
284 user->tg->uid = user->uid;
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
303 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock);
315 static int root_task_group_empty(void)
317 return list_empty(&root_task_group.children);
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 #ifdef CONFIG_USER_SCHED
323 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
324 #else /* !CONFIG_USER_SCHED */
325 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
326 #endif /* CONFIG_USER_SCHED */
329 * A weight of 0 or 1 can cause arithmetics problems.
330 * A weight of a cfs_rq is the sum of weights of which entities
331 * are queued on this cfs_rq, so a weight of a entity should not be
332 * too large, so as the shares value of a task group.
333 * (The default weight is 1024 - so there's no practical
334 * limitation from this.)
337 #define MAX_SHARES (1UL << 18)
339 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
342 /* Default task group.
343 * Every task in system belong to this group at bootup.
345 struct task_group init_task_group;
347 /* return group to which a task belongs */
348 static inline struct task_group *task_group(struct task_struct *p)
350 struct task_group *tg;
352 #ifdef CONFIG_USER_SCHED
354 tg = __task_cred(p)->user->tg;
356 #elif defined(CONFIG_CGROUP_SCHED)
357 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
358 struct task_group, css);
360 tg = &init_task_group;
365 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
366 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
370 p->se.parent = task_group(p)->se[cpu];
373 #ifdef CONFIG_RT_GROUP_SCHED
374 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
375 p->rt.parent = task_group(p)->rt_se[cpu];
382 static int root_task_group_empty(void)
388 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
389 static inline struct task_group *task_group(struct task_struct *p)
394 #endif /* CONFIG_GROUP_SCHED */
396 /* CFS-related fields in a runqueue */
398 struct load_weight load;
399 unsigned long nr_running;
404 struct rb_root tasks_timeline;
405 struct rb_node *rb_leftmost;
407 struct list_head tasks;
408 struct list_head *balance_iterator;
411 * 'curr' points to currently running entity on this cfs_rq.
412 * It is set to NULL otherwise (i.e when none are currently running).
414 struct sched_entity *curr, *next, *last;
416 unsigned int nr_spread_over;
418 #ifdef CONFIG_FAIR_GROUP_SCHED
419 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
422 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
423 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
424 * (like users, containers etc.)
426 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
427 * list is used during load balance.
429 struct list_head leaf_cfs_rq_list;
430 struct task_group *tg; /* group that "owns" this runqueue */
434 * the part of load.weight contributed by tasks
436 unsigned long task_weight;
439 * h_load = weight * f(tg)
441 * Where f(tg) is the recursive weight fraction assigned to
444 unsigned long h_load;
447 * this cpu's part of tg->shares
449 unsigned long shares;
452 * load.weight at the time we set shares
454 unsigned long rq_weight;
459 /* Real-Time classes' related field in a runqueue: */
461 struct rt_prio_array active;
462 unsigned long rt_nr_running;
463 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
465 int curr; /* highest queued rt task prio */
467 int next; /* next highest */
472 unsigned long rt_nr_migratory;
473 unsigned long rt_nr_total;
475 struct plist_head pushable_tasks;
480 /* Nests inside the rq lock: */
481 spinlock_t rt_runtime_lock;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 unsigned long rt_nr_boosted;
487 struct list_head leaf_rt_rq_list;
488 struct task_group *tg;
489 struct sched_rt_entity *rt_se;
496 * We add the notion of a root-domain which will be used to define per-domain
497 * variables. Each exclusive cpuset essentially defines an island domain by
498 * fully partitioning the member cpus from any other cpuset. Whenever a new
499 * exclusive cpuset is created, we also create and attach a new root-domain
506 cpumask_var_t online;
509 * The "RT overload" flag: it gets set if a CPU has more than
510 * one runnable RT task.
512 cpumask_var_t rto_mask;
515 struct cpupri cpupri;
517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
519 * Preferred wake up cpu nominated by sched_mc balance that will be
520 * used when most cpus are idle in the system indicating overall very
521 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
523 unsigned int sched_mc_preferred_wakeup_cpu;
528 * By default the system creates a single root-domain with all cpus as
529 * members (mimicking the global state we have today).
531 static struct root_domain def_root_domain;
536 * This is the main, per-CPU runqueue data structure.
538 * Locking rule: those places that want to lock multiple runqueues
539 * (such as the load balancing or the thread migration code), lock
540 * acquire operations must be ordered by ascending &runqueue.
547 * nr_running and cpu_load should be in the same cacheline because
548 * remote CPUs use both these fields when doing load calculation.
550 unsigned long nr_running;
551 #define CPU_LOAD_IDX_MAX 5
552 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
554 unsigned long last_tick_seen;
555 unsigned char in_nohz_recently;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load;
559 unsigned long nr_load_updates;
561 u64 nr_migrations_in;
566 #ifdef CONFIG_FAIR_GROUP_SCHED
567 /* list of leaf cfs_rq on this cpu: */
568 struct list_head leaf_cfs_rq_list;
570 #ifdef CONFIG_RT_GROUP_SCHED
571 struct list_head leaf_rt_rq_list;
575 * This is part of a global counter where only the total sum
576 * over all CPUs matters. A task can increase this counter on
577 * one CPU and if it got migrated afterwards it may decrease
578 * it on another CPU. Always updated under the runqueue lock:
580 unsigned long nr_uninterruptible;
582 struct task_struct *curr, *idle;
583 unsigned long next_balance;
584 struct mm_struct *prev_mm;
591 struct root_domain *rd;
592 struct sched_domain *sd;
594 unsigned char idle_at_tick;
595 /* For active balancing */
599 /* cpu of this runqueue: */
603 unsigned long avg_load_per_task;
605 struct task_struct *migration_thread;
606 struct list_head migration_queue;
612 /* calc_load related fields */
613 unsigned long calc_load_update;
614 long calc_load_active;
616 #ifdef CONFIG_SCHED_HRTICK
618 int hrtick_csd_pending;
619 struct call_single_data hrtick_csd;
621 struct hrtimer hrtick_timer;
624 #ifdef CONFIG_SCHEDSTATS
626 struct sched_info rq_sched_info;
627 unsigned long long rq_cpu_time;
628 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
630 /* sys_sched_yield() stats */
631 unsigned int yld_count;
633 /* schedule() stats */
634 unsigned int sched_switch;
635 unsigned int sched_count;
636 unsigned int sched_goidle;
638 /* try_to_wake_up() stats */
639 unsigned int ttwu_count;
640 unsigned int ttwu_local;
643 unsigned int bkl_count;
647 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
649 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
651 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
654 static inline int cpu_of(struct rq *rq)
664 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
665 * See detach_destroy_domains: synchronize_sched for details.
667 * The domain tree of any CPU may only be accessed from within
668 * preempt-disabled sections.
670 #define for_each_domain(cpu, __sd) \
671 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
673 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
674 #define this_rq() (&__get_cpu_var(runqueues))
675 #define task_rq(p) cpu_rq(task_cpu(p))
676 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
677 #define raw_rq() (&__raw_get_cpu_var(runqueues))
679 inline void update_rq_clock(struct rq *rq)
681 rq->clock = sched_clock_cpu(cpu_of(rq));
685 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
687 #ifdef CONFIG_SCHED_DEBUG
688 # define const_debug __read_mostly
690 # define const_debug static const
696 * Returns true if the current cpu runqueue is locked.
697 * This interface allows printk to be called with the runqueue lock
698 * held and know whether or not it is OK to wake up the klogd.
700 int runqueue_is_locked(void)
703 struct rq *rq = cpu_rq(cpu);
706 ret = spin_is_locked(&rq->lock);
712 * Debugging: various feature bits
715 #define SCHED_FEAT(name, enabled) \
716 __SCHED_FEAT_##name ,
719 #include "sched_features.h"
724 #define SCHED_FEAT(name, enabled) \
725 (1UL << __SCHED_FEAT_##name) * enabled |
727 const_debug unsigned int sysctl_sched_features =
728 #include "sched_features.h"
733 #ifdef CONFIG_SCHED_DEBUG
734 #define SCHED_FEAT(name, enabled) \
737 static __read_mostly char *sched_feat_names[] = {
738 #include "sched_features.h"
744 static int sched_feat_show(struct seq_file *m, void *v)
748 for (i = 0; sched_feat_names[i]; i++) {
749 if (!(sysctl_sched_features & (1UL << i)))
751 seq_printf(m, "%s ", sched_feat_names[i]);
759 sched_feat_write(struct file *filp, const char __user *ubuf,
760 size_t cnt, loff_t *ppos)
770 if (copy_from_user(&buf, ubuf, cnt))
775 if (strncmp(buf, "NO_", 3) == 0) {
780 for (i = 0; sched_feat_names[i]; i++) {
781 int len = strlen(sched_feat_names[i]);
783 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
785 sysctl_sched_features &= ~(1UL << i);
787 sysctl_sched_features |= (1UL << i);
792 if (!sched_feat_names[i])
800 static int sched_feat_open(struct inode *inode, struct file *filp)
802 return single_open(filp, sched_feat_show, NULL);
805 static struct file_operations sched_feat_fops = {
806 .open = sched_feat_open,
807 .write = sched_feat_write,
810 .release = single_release,
813 static __init int sched_init_debug(void)
815 debugfs_create_file("sched_features", 0644, NULL, NULL,
820 late_initcall(sched_init_debug);
824 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
827 * Number of tasks to iterate in a single balance run.
828 * Limited because this is done with IRQs disabled.
830 const_debug unsigned int sysctl_sched_nr_migrate = 32;
833 * ratelimit for updating the group shares.
836 unsigned int sysctl_sched_shares_ratelimit = 250000;
839 * Inject some fuzzyness into changing the per-cpu group shares
840 * this avoids remote rq-locks at the expense of fairness.
843 unsigned int sysctl_sched_shares_thresh = 4;
846 * period over which we average the RT time consumption, measured
851 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
854 * period over which we measure -rt task cpu usage in us.
857 unsigned int sysctl_sched_rt_period = 1000000;
859 static __read_mostly int scheduler_running;
862 * part of the period that we allow rt tasks to run in us.
865 int sysctl_sched_rt_runtime = 950000;
867 static inline u64 global_rt_period(void)
869 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
872 static inline u64 global_rt_runtime(void)
874 if (sysctl_sched_rt_runtime < 0)
877 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
880 #ifndef prepare_arch_switch
881 # define prepare_arch_switch(next) do { } while (0)
883 #ifndef finish_arch_switch
884 # define finish_arch_switch(prev) do { } while (0)
887 static inline int task_current(struct rq *rq, struct task_struct *p)
889 return rq->curr == p;
892 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
893 static inline int task_running(struct rq *rq, struct task_struct *p)
895 return task_current(rq, p);
898 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
902 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
904 #ifdef CONFIG_DEBUG_SPINLOCK
905 /* this is a valid case when another task releases the spinlock */
906 rq->lock.owner = current;
909 * If we are tracking spinlock dependencies then we have to
910 * fix up the runqueue lock - which gets 'carried over' from
913 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
915 spin_unlock_irq(&rq->lock);
918 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
919 static inline int task_running(struct rq *rq, struct task_struct *p)
924 return task_current(rq, p);
928 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
932 * We can optimise this out completely for !SMP, because the
933 * SMP rebalancing from interrupt is the only thing that cares
938 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
939 spin_unlock_irq(&rq->lock);
941 spin_unlock(&rq->lock);
945 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
949 * After ->oncpu is cleared, the task can be moved to a different CPU.
950 * We must ensure this doesn't happen until the switch is completely
956 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
960 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
963 * __task_rq_lock - lock the runqueue a given task resides on.
964 * Must be called interrupts disabled.
966 static inline struct rq *__task_rq_lock(struct task_struct *p)
970 struct rq *rq = task_rq(p);
971 spin_lock(&rq->lock);
972 if (likely(rq == task_rq(p)))
974 spin_unlock(&rq->lock);
979 * task_rq_lock - lock the runqueue a given task resides on and disable
980 * interrupts. Note the ordering: we can safely lookup the task_rq without
981 * explicitly disabling preemption.
983 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
989 local_irq_save(*flags);
991 spin_lock(&rq->lock);
992 if (likely(rq == task_rq(p)))
994 spin_unlock_irqrestore(&rq->lock, *flags);
998 void task_rq_unlock_wait(struct task_struct *p)
1000 struct rq *rq = task_rq(p);
1002 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1003 spin_unlock_wait(&rq->lock);
1006 static void __task_rq_unlock(struct rq *rq)
1007 __releases(rq->lock)
1009 spin_unlock(&rq->lock);
1012 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1013 __releases(rq->lock)
1015 spin_unlock_irqrestore(&rq->lock, *flags);
1019 * this_rq_lock - lock this runqueue and disable interrupts.
1021 static struct rq *this_rq_lock(void)
1022 __acquires(rq->lock)
1026 local_irq_disable();
1028 spin_lock(&rq->lock);
1033 #ifdef CONFIG_SCHED_HRTICK
1035 * Use HR-timers to deliver accurate preemption points.
1037 * Its all a bit involved since we cannot program an hrt while holding the
1038 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1041 * When we get rescheduled we reprogram the hrtick_timer outside of the
1047 * - enabled by features
1048 * - hrtimer is actually high res
1050 static inline int hrtick_enabled(struct rq *rq)
1052 if (!sched_feat(HRTICK))
1054 if (!cpu_active(cpu_of(rq)))
1056 return hrtimer_is_hres_active(&rq->hrtick_timer);
1059 static void hrtick_clear(struct rq *rq)
1061 if (hrtimer_active(&rq->hrtick_timer))
1062 hrtimer_cancel(&rq->hrtick_timer);
1066 * High-resolution timer tick.
1067 * Runs from hardirq context with interrupts disabled.
1069 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1071 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1073 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1075 spin_lock(&rq->lock);
1076 update_rq_clock(rq);
1077 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1078 spin_unlock(&rq->lock);
1080 return HRTIMER_NORESTART;
1085 * called from hardirq (IPI) context
1087 static void __hrtick_start(void *arg)
1089 struct rq *rq = arg;
1091 spin_lock(&rq->lock);
1092 hrtimer_restart(&rq->hrtick_timer);
1093 rq->hrtick_csd_pending = 0;
1094 spin_unlock(&rq->lock);
1098 * Called to set the hrtick timer state.
1100 * called with rq->lock held and irqs disabled
1102 static void hrtick_start(struct rq *rq, u64 delay)
1104 struct hrtimer *timer = &rq->hrtick_timer;
1105 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1107 hrtimer_set_expires(timer, time);
1109 if (rq == this_rq()) {
1110 hrtimer_restart(timer);
1111 } else if (!rq->hrtick_csd_pending) {
1112 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1113 rq->hrtick_csd_pending = 1;
1118 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1120 int cpu = (int)(long)hcpu;
1123 case CPU_UP_CANCELED:
1124 case CPU_UP_CANCELED_FROZEN:
1125 case CPU_DOWN_PREPARE:
1126 case CPU_DOWN_PREPARE_FROZEN:
1128 case CPU_DEAD_FROZEN:
1129 hrtick_clear(cpu_rq(cpu));
1136 static __init void init_hrtick(void)
1138 hotcpu_notifier(hotplug_hrtick, 0);
1142 * Called to set the hrtick timer state.
1144 * called with rq->lock held and irqs disabled
1146 static void hrtick_start(struct rq *rq, u64 delay)
1148 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1149 HRTIMER_MODE_REL_PINNED, 0);
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SMP */
1157 static void init_rq_hrtick(struct rq *rq)
1160 rq->hrtick_csd_pending = 0;
1162 rq->hrtick_csd.flags = 0;
1163 rq->hrtick_csd.func = __hrtick_start;
1164 rq->hrtick_csd.info = rq;
1167 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1168 rq->hrtick_timer.function = hrtick;
1170 #else /* CONFIG_SCHED_HRTICK */
1171 static inline void hrtick_clear(struct rq *rq)
1175 static inline void init_rq_hrtick(struct rq *rq)
1179 static inline void init_hrtick(void)
1182 #endif /* CONFIG_SCHED_HRTICK */
1185 * resched_task - mark a task 'to be rescheduled now'.
1187 * On UP this means the setting of the need_resched flag, on SMP it
1188 * might also involve a cross-CPU call to trigger the scheduler on
1193 #ifndef tsk_is_polling
1194 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1197 static void resched_task(struct task_struct *p)
1201 assert_spin_locked(&task_rq(p)->lock);
1203 if (test_tsk_need_resched(p))
1206 set_tsk_need_resched(p);
1209 if (cpu == smp_processor_id())
1212 /* NEED_RESCHED must be visible before we test polling */
1214 if (!tsk_is_polling(p))
1215 smp_send_reschedule(cpu);
1218 static void resched_cpu(int cpu)
1220 struct rq *rq = cpu_rq(cpu);
1221 unsigned long flags;
1223 if (!spin_trylock_irqsave(&rq->lock, flags))
1225 resched_task(cpu_curr(cpu));
1226 spin_unlock_irqrestore(&rq->lock, flags);
1231 * When add_timer_on() enqueues a timer into the timer wheel of an
1232 * idle CPU then this timer might expire before the next timer event
1233 * which is scheduled to wake up that CPU. In case of a completely
1234 * idle system the next event might even be infinite time into the
1235 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1236 * leaves the inner idle loop so the newly added timer is taken into
1237 * account when the CPU goes back to idle and evaluates the timer
1238 * wheel for the next timer event.
1240 void wake_up_idle_cpu(int cpu)
1242 struct rq *rq = cpu_rq(cpu);
1244 if (cpu == smp_processor_id())
1248 * This is safe, as this function is called with the timer
1249 * wheel base lock of (cpu) held. When the CPU is on the way
1250 * to idle and has not yet set rq->curr to idle then it will
1251 * be serialized on the timer wheel base lock and take the new
1252 * timer into account automatically.
1254 if (rq->curr != rq->idle)
1258 * We can set TIF_RESCHED on the idle task of the other CPU
1259 * lockless. The worst case is that the other CPU runs the
1260 * idle task through an additional NOOP schedule()
1262 set_tsk_need_resched(rq->idle);
1264 /* NEED_RESCHED must be visible before we test polling */
1266 if (!tsk_is_polling(rq->idle))
1267 smp_send_reschedule(cpu);
1269 #endif /* CONFIG_NO_HZ */
1271 static u64 sched_avg_period(void)
1273 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1276 static void sched_avg_update(struct rq *rq)
1278 s64 period = sched_avg_period();
1280 while ((s64)(rq->clock - rq->age_stamp) > period) {
1281 rq->age_stamp += period;
1286 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1288 rq->rt_avg += rt_delta;
1289 sched_avg_update(rq);
1292 #else /* !CONFIG_SMP */
1293 static void resched_task(struct task_struct *p)
1295 assert_spin_locked(&task_rq(p)->lock);
1296 set_tsk_need_resched(p);
1299 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1302 #endif /* CONFIG_SMP */
1304 #if BITS_PER_LONG == 32
1305 # define WMULT_CONST (~0UL)
1307 # define WMULT_CONST (1UL << 32)
1310 #define WMULT_SHIFT 32
1313 * Shift right and round:
1315 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1318 * delta *= weight / lw
1320 static unsigned long
1321 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1322 struct load_weight *lw)
1326 if (!lw->inv_weight) {
1327 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1330 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1334 tmp = (u64)delta_exec * weight;
1336 * Check whether we'd overflow the 64-bit multiplication:
1338 if (unlikely(tmp > WMULT_CONST))
1339 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1342 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1344 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1347 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1353 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1360 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1361 * of tasks with abnormal "nice" values across CPUs the contribution that
1362 * each task makes to its run queue's load is weighted according to its
1363 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1364 * scaled version of the new time slice allocation that they receive on time
1368 #define WEIGHT_IDLEPRIO 3
1369 #define WMULT_IDLEPRIO 1431655765
1372 * Nice levels are multiplicative, with a gentle 10% change for every
1373 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1374 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1375 * that remained on nice 0.
1377 * The "10% effect" is relative and cumulative: from _any_ nice level,
1378 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1379 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1380 * If a task goes up by ~10% and another task goes down by ~10% then
1381 * the relative distance between them is ~25%.)
1383 static const int prio_to_weight[40] = {
1384 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1385 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1386 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1387 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1388 /* 0 */ 1024, 820, 655, 526, 423,
1389 /* 5 */ 335, 272, 215, 172, 137,
1390 /* 10 */ 110, 87, 70, 56, 45,
1391 /* 15 */ 36, 29, 23, 18, 15,
1395 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1397 * In cases where the weight does not change often, we can use the
1398 * precalculated inverse to speed up arithmetics by turning divisions
1399 * into multiplications:
1401 static const u32 prio_to_wmult[40] = {
1402 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1403 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1404 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1405 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1406 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1407 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1408 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1409 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1412 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1415 * runqueue iterator, to support SMP load-balancing between different
1416 * scheduling classes, without having to expose their internal data
1417 * structures to the load-balancing proper:
1419 struct rq_iterator {
1421 struct task_struct *(*start)(void *);
1422 struct task_struct *(*next)(void *);
1426 static unsigned long
1427 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1428 unsigned long max_load_move, struct sched_domain *sd,
1429 enum cpu_idle_type idle, int *all_pinned,
1430 int *this_best_prio, struct rq_iterator *iterator);
1433 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1434 struct sched_domain *sd, enum cpu_idle_type idle,
1435 struct rq_iterator *iterator);
1438 /* Time spent by the tasks of the cpu accounting group executing in ... */
1439 enum cpuacct_stat_index {
1440 CPUACCT_STAT_USER, /* ... user mode */
1441 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1443 CPUACCT_STAT_NSTATS,
1446 #ifdef CONFIG_CGROUP_CPUACCT
1447 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1448 static void cpuacct_update_stats(struct task_struct *tsk,
1449 enum cpuacct_stat_index idx, cputime_t val);
1451 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1452 static inline void cpuacct_update_stats(struct task_struct *tsk,
1453 enum cpuacct_stat_index idx, cputime_t val) {}
1456 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1458 update_load_add(&rq->load, load);
1461 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1463 update_load_sub(&rq->load, load);
1466 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1467 typedef int (*tg_visitor)(struct task_group *, void *);
1470 * Iterate the full tree, calling @down when first entering a node and @up when
1471 * leaving it for the final time.
1473 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1475 struct task_group *parent, *child;
1479 parent = &root_task_group;
1481 ret = (*down)(parent, data);
1484 list_for_each_entry_rcu(child, &parent->children, siblings) {
1491 ret = (*up)(parent, data);
1496 parent = parent->parent;
1505 static int tg_nop(struct task_group *tg, void *data)
1512 static unsigned long source_load(int cpu, int type);
1513 static unsigned long target_load(int cpu, int type);
1514 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1516 static unsigned long cpu_avg_load_per_task(int cpu)
1518 struct rq *rq = cpu_rq(cpu);
1519 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1522 rq->avg_load_per_task = rq->load.weight / nr_running;
1524 rq->avg_load_per_task = 0;
1526 return rq->avg_load_per_task;
1529 #ifdef CONFIG_FAIR_GROUP_SCHED
1531 struct update_shares_data {
1532 unsigned long rq_weight[NR_CPUS];
1535 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1537 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1540 * Calculate and set the cpu's group shares.
1542 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1543 unsigned long sd_shares,
1544 unsigned long sd_rq_weight,
1545 struct update_shares_data *usd)
1547 unsigned long shares, rq_weight;
1550 rq_weight = usd->rq_weight[cpu];
1553 rq_weight = NICE_0_LOAD;
1557 * \Sum_j shares_j * rq_weight_i
1558 * shares_i = -----------------------------
1559 * \Sum_j rq_weight_j
1561 shares = (sd_shares * rq_weight) / sd_rq_weight;
1562 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1564 if (abs(shares - tg->se[cpu]->load.weight) >
1565 sysctl_sched_shares_thresh) {
1566 struct rq *rq = cpu_rq(cpu);
1567 unsigned long flags;
1569 spin_lock_irqsave(&rq->lock, flags);
1570 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1571 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1572 __set_se_shares(tg->se[cpu], shares);
1573 spin_unlock_irqrestore(&rq->lock, flags);
1578 * Re-compute the task group their per cpu shares over the given domain.
1579 * This needs to be done in a bottom-up fashion because the rq weight of a
1580 * parent group depends on the shares of its child groups.
1582 static int tg_shares_up(struct task_group *tg, void *data)
1584 unsigned long weight, rq_weight = 0, shares = 0;
1585 struct update_shares_data *usd;
1586 struct sched_domain *sd = data;
1587 unsigned long flags;
1593 local_irq_save(flags);
1594 usd = &__get_cpu_var(update_shares_data);
1596 for_each_cpu(i, sched_domain_span(sd)) {
1597 weight = tg->cfs_rq[i]->load.weight;
1598 usd->rq_weight[i] = weight;
1601 * If there are currently no tasks on the cpu pretend there
1602 * is one of average load so that when a new task gets to
1603 * run here it will not get delayed by group starvation.
1606 weight = NICE_0_LOAD;
1608 rq_weight += weight;
1609 shares += tg->cfs_rq[i]->shares;
1612 if ((!shares && rq_weight) || shares > tg->shares)
1613 shares = tg->shares;
1615 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1616 shares = tg->shares;
1618 for_each_cpu(i, sched_domain_span(sd))
1619 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1621 local_irq_restore(flags);
1627 * Compute the cpu's hierarchical load factor for each task group.
1628 * This needs to be done in a top-down fashion because the load of a child
1629 * group is a fraction of its parents load.
1631 static int tg_load_down(struct task_group *tg, void *data)
1634 long cpu = (long)data;
1637 load = cpu_rq(cpu)->load.weight;
1639 load = tg->parent->cfs_rq[cpu]->h_load;
1640 load *= tg->cfs_rq[cpu]->shares;
1641 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1644 tg->cfs_rq[cpu]->h_load = load;
1649 static void update_shares(struct sched_domain *sd)
1654 if (root_task_group_empty())
1657 now = cpu_clock(raw_smp_processor_id());
1658 elapsed = now - sd->last_update;
1660 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1661 sd->last_update = now;
1662 walk_tg_tree(tg_nop, tg_shares_up, sd);
1666 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1668 if (root_task_group_empty())
1671 spin_unlock(&rq->lock);
1673 spin_lock(&rq->lock);
1676 static void update_h_load(long cpu)
1678 if (root_task_group_empty())
1681 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1686 static inline void update_shares(struct sched_domain *sd)
1690 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1696 #ifdef CONFIG_PREEMPT
1699 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1700 * way at the expense of forcing extra atomic operations in all
1701 * invocations. This assures that the double_lock is acquired using the
1702 * same underlying policy as the spinlock_t on this architecture, which
1703 * reduces latency compared to the unfair variant below. However, it
1704 * also adds more overhead and therefore may reduce throughput.
1706 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1707 __releases(this_rq->lock)
1708 __acquires(busiest->lock)
1709 __acquires(this_rq->lock)
1711 spin_unlock(&this_rq->lock);
1712 double_rq_lock(this_rq, busiest);
1719 * Unfair double_lock_balance: Optimizes throughput at the expense of
1720 * latency by eliminating extra atomic operations when the locks are
1721 * already in proper order on entry. This favors lower cpu-ids and will
1722 * grant the double lock to lower cpus over higher ids under contention,
1723 * regardless of entry order into the function.
1725 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1726 __releases(this_rq->lock)
1727 __acquires(busiest->lock)
1728 __acquires(this_rq->lock)
1732 if (unlikely(!spin_trylock(&busiest->lock))) {
1733 if (busiest < this_rq) {
1734 spin_unlock(&this_rq->lock);
1735 spin_lock(&busiest->lock);
1736 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1739 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1744 #endif /* CONFIG_PREEMPT */
1747 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1749 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1751 if (unlikely(!irqs_disabled())) {
1752 /* printk() doesn't work good under rq->lock */
1753 spin_unlock(&this_rq->lock);
1757 return _double_lock_balance(this_rq, busiest);
1760 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1761 __releases(busiest->lock)
1763 spin_unlock(&busiest->lock);
1764 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1768 #ifdef CONFIG_FAIR_GROUP_SCHED
1769 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1772 cfs_rq->shares = shares;
1777 static void calc_load_account_active(struct rq *this_rq);
1779 #include "sched_stats.h"
1780 #include "sched_idletask.c"
1781 #include "sched_fair.c"
1782 #include "sched_rt.c"
1783 #ifdef CONFIG_SCHED_DEBUG
1784 # include "sched_debug.c"
1787 #define sched_class_highest (&rt_sched_class)
1788 #define for_each_class(class) \
1789 for (class = sched_class_highest; class; class = class->next)
1791 static void inc_nr_running(struct rq *rq)
1796 static void dec_nr_running(struct rq *rq)
1801 static void set_load_weight(struct task_struct *p)
1803 if (task_has_rt_policy(p)) {
1804 p->se.load.weight = prio_to_weight[0] * 2;
1805 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1810 * SCHED_IDLE tasks get minimal weight:
1812 if (p->policy == SCHED_IDLE) {
1813 p->se.load.weight = WEIGHT_IDLEPRIO;
1814 p->se.load.inv_weight = WMULT_IDLEPRIO;
1818 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1819 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1822 static void update_avg(u64 *avg, u64 sample)
1824 s64 diff = sample - *avg;
1828 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1831 p->se.start_runtime = p->se.sum_exec_runtime;
1833 sched_info_queued(p);
1834 p->sched_class->enqueue_task(rq, p, wakeup);
1838 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1841 if (p->se.last_wakeup) {
1842 update_avg(&p->se.avg_overlap,
1843 p->se.sum_exec_runtime - p->se.last_wakeup);
1844 p->se.last_wakeup = 0;
1846 update_avg(&p->se.avg_wakeup,
1847 sysctl_sched_wakeup_granularity);
1851 sched_info_dequeued(p);
1852 p->sched_class->dequeue_task(rq, p, sleep);
1857 * __normal_prio - return the priority that is based on the static prio
1859 static inline int __normal_prio(struct task_struct *p)
1861 return p->static_prio;
1865 * Calculate the expected normal priority: i.e. priority
1866 * without taking RT-inheritance into account. Might be
1867 * boosted by interactivity modifiers. Changes upon fork,
1868 * setprio syscalls, and whenever the interactivity
1869 * estimator recalculates.
1871 static inline int normal_prio(struct task_struct *p)
1875 if (task_has_rt_policy(p))
1876 prio = MAX_RT_PRIO-1 - p->rt_priority;
1878 prio = __normal_prio(p);
1883 * Calculate the current priority, i.e. the priority
1884 * taken into account by the scheduler. This value might
1885 * be boosted by RT tasks, or might be boosted by
1886 * interactivity modifiers. Will be RT if the task got
1887 * RT-boosted. If not then it returns p->normal_prio.
1889 static int effective_prio(struct task_struct *p)
1891 p->normal_prio = normal_prio(p);
1893 * If we are RT tasks or we were boosted to RT priority,
1894 * keep the priority unchanged. Otherwise, update priority
1895 * to the normal priority:
1897 if (!rt_prio(p->prio))
1898 return p->normal_prio;
1903 * activate_task - move a task to the runqueue.
1905 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1907 if (task_contributes_to_load(p))
1908 rq->nr_uninterruptible--;
1910 enqueue_task(rq, p, wakeup);
1915 * deactivate_task - remove a task from the runqueue.
1917 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1919 if (task_contributes_to_load(p))
1920 rq->nr_uninterruptible++;
1922 dequeue_task(rq, p, sleep);
1927 * task_curr - is this task currently executing on a CPU?
1928 * @p: the task in question.
1930 inline int task_curr(const struct task_struct *p)
1932 return cpu_curr(task_cpu(p)) == p;
1935 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1937 set_task_rq(p, cpu);
1940 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1941 * successfuly executed on another CPU. We must ensure that updates of
1942 * per-task data have been completed by this moment.
1945 task_thread_info(p)->cpu = cpu;
1949 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1950 const struct sched_class *prev_class,
1951 int oldprio, int running)
1953 if (prev_class != p->sched_class) {
1954 if (prev_class->switched_from)
1955 prev_class->switched_from(rq, p, running);
1956 p->sched_class->switched_to(rq, p, running);
1958 p->sched_class->prio_changed(rq, p, oldprio, running);
1963 /* Used instead of source_load when we know the type == 0 */
1964 static unsigned long weighted_cpuload(const int cpu)
1966 return cpu_rq(cpu)->load.weight;
1970 * Is this task likely cache-hot:
1973 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1978 * Buddy candidates are cache hot:
1980 if (sched_feat(CACHE_HOT_BUDDY) &&
1981 (&p->se == cfs_rq_of(&p->se)->next ||
1982 &p->se == cfs_rq_of(&p->se)->last))
1985 if (p->sched_class != &fair_sched_class)
1988 if (sysctl_sched_migration_cost == -1)
1990 if (sysctl_sched_migration_cost == 0)
1993 delta = now - p->se.exec_start;
1995 return delta < (s64)sysctl_sched_migration_cost;
1999 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2001 int old_cpu = task_cpu(p);
2002 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2003 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2004 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2007 clock_offset = old_rq->clock - new_rq->clock;
2009 trace_sched_migrate_task(p, new_cpu);
2011 #ifdef CONFIG_SCHEDSTATS
2012 if (p->se.wait_start)
2013 p->se.wait_start -= clock_offset;
2014 if (p->se.sleep_start)
2015 p->se.sleep_start -= clock_offset;
2016 if (p->se.block_start)
2017 p->se.block_start -= clock_offset;
2019 if (old_cpu != new_cpu) {
2020 p->se.nr_migrations++;
2021 new_rq->nr_migrations_in++;
2022 #ifdef CONFIG_SCHEDSTATS
2023 if (task_hot(p, old_rq->clock, NULL))
2024 schedstat_inc(p, se.nr_forced2_migrations);
2026 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2029 p->se.vruntime -= old_cfsrq->min_vruntime -
2030 new_cfsrq->min_vruntime;
2032 __set_task_cpu(p, new_cpu);
2035 struct migration_req {
2036 struct list_head list;
2038 struct task_struct *task;
2041 struct completion done;
2045 * The task's runqueue lock must be held.
2046 * Returns true if you have to wait for migration thread.
2049 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2051 struct rq *rq = task_rq(p);
2054 * If the task is not on a runqueue (and not running), then
2055 * it is sufficient to simply update the task's cpu field.
2057 if (!p->se.on_rq && !task_running(rq, p)) {
2058 set_task_cpu(p, dest_cpu);
2062 init_completion(&req->done);
2064 req->dest_cpu = dest_cpu;
2065 list_add(&req->list, &rq->migration_queue);
2071 * wait_task_context_switch - wait for a thread to complete at least one
2074 * @p must not be current.
2076 void wait_task_context_switch(struct task_struct *p)
2078 unsigned long nvcsw, nivcsw, flags;
2086 * The runqueue is assigned before the actual context
2087 * switch. We need to take the runqueue lock.
2089 * We could check initially without the lock but it is
2090 * very likely that we need to take the lock in every
2093 rq = task_rq_lock(p, &flags);
2094 running = task_running(rq, p);
2095 task_rq_unlock(rq, &flags);
2097 if (likely(!running))
2100 * The switch count is incremented before the actual
2101 * context switch. We thus wait for two switches to be
2102 * sure at least one completed.
2104 if ((p->nvcsw - nvcsw) > 1)
2106 if ((p->nivcsw - nivcsw) > 1)
2114 * wait_task_inactive - wait for a thread to unschedule.
2116 * If @match_state is nonzero, it's the @p->state value just checked and
2117 * not expected to change. If it changes, i.e. @p might have woken up,
2118 * then return zero. When we succeed in waiting for @p to be off its CPU,
2119 * we return a positive number (its total switch count). If a second call
2120 * a short while later returns the same number, the caller can be sure that
2121 * @p has remained unscheduled the whole time.
2123 * The caller must ensure that the task *will* unschedule sometime soon,
2124 * else this function might spin for a *long* time. This function can't
2125 * be called with interrupts off, or it may introduce deadlock with
2126 * smp_call_function() if an IPI is sent by the same process we are
2127 * waiting to become inactive.
2129 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2131 unsigned long flags;
2138 * We do the initial early heuristics without holding
2139 * any task-queue locks at all. We'll only try to get
2140 * the runqueue lock when things look like they will
2146 * If the task is actively running on another CPU
2147 * still, just relax and busy-wait without holding
2150 * NOTE! Since we don't hold any locks, it's not
2151 * even sure that "rq" stays as the right runqueue!
2152 * But we don't care, since "task_running()" will
2153 * return false if the runqueue has changed and p
2154 * is actually now running somewhere else!
2156 while (task_running(rq, p)) {
2157 if (match_state && unlikely(p->state != match_state))
2163 * Ok, time to look more closely! We need the rq
2164 * lock now, to be *sure*. If we're wrong, we'll
2165 * just go back and repeat.
2167 rq = task_rq_lock(p, &flags);
2168 trace_sched_wait_task(rq, p);
2169 running = task_running(rq, p);
2170 on_rq = p->se.on_rq;
2172 if (!match_state || p->state == match_state)
2173 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2174 task_rq_unlock(rq, &flags);
2177 * If it changed from the expected state, bail out now.
2179 if (unlikely(!ncsw))
2183 * Was it really running after all now that we
2184 * checked with the proper locks actually held?
2186 * Oops. Go back and try again..
2188 if (unlikely(running)) {
2194 * It's not enough that it's not actively running,
2195 * it must be off the runqueue _entirely_, and not
2198 * So if it was still runnable (but just not actively
2199 * running right now), it's preempted, and we should
2200 * yield - it could be a while.
2202 if (unlikely(on_rq)) {
2203 schedule_timeout_uninterruptible(1);
2208 * Ahh, all good. It wasn't running, and it wasn't
2209 * runnable, which means that it will never become
2210 * running in the future either. We're all done!
2219 * kick_process - kick a running thread to enter/exit the kernel
2220 * @p: the to-be-kicked thread
2222 * Cause a process which is running on another CPU to enter
2223 * kernel-mode, without any delay. (to get signals handled.)
2225 * NOTE: this function doesnt have to take the runqueue lock,
2226 * because all it wants to ensure is that the remote task enters
2227 * the kernel. If the IPI races and the task has been migrated
2228 * to another CPU then no harm is done and the purpose has been
2231 void kick_process(struct task_struct *p)
2237 if ((cpu != smp_processor_id()) && task_curr(p))
2238 smp_send_reschedule(cpu);
2241 EXPORT_SYMBOL_GPL(kick_process);
2244 * Return a low guess at the load of a migration-source cpu weighted
2245 * according to the scheduling class and "nice" value.
2247 * We want to under-estimate the load of migration sources, to
2248 * balance conservatively.
2250 static unsigned long source_load(int cpu, int type)
2252 struct rq *rq = cpu_rq(cpu);
2253 unsigned long total = weighted_cpuload(cpu);
2255 if (type == 0 || !sched_feat(LB_BIAS))
2258 return min(rq->cpu_load[type-1], total);
2262 * Return a high guess at the load of a migration-target cpu weighted
2263 * according to the scheduling class and "nice" value.
2265 static unsigned long target_load(int cpu, int type)
2267 struct rq *rq = cpu_rq(cpu);
2268 unsigned long total = weighted_cpuload(cpu);
2270 if (type == 0 || !sched_feat(LB_BIAS))
2273 return max(rq->cpu_load[type-1], total);
2277 * find_idlest_group finds and returns the least busy CPU group within the
2280 static struct sched_group *
2281 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2283 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2284 unsigned long min_load = ULONG_MAX, this_load = 0;
2285 int load_idx = sd->forkexec_idx;
2286 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2289 unsigned long load, avg_load;
2293 /* Skip over this group if it has no CPUs allowed */
2294 if (!cpumask_intersects(sched_group_cpus(group),
2298 local_group = cpumask_test_cpu(this_cpu,
2299 sched_group_cpus(group));
2301 /* Tally up the load of all CPUs in the group */
2304 for_each_cpu(i, sched_group_cpus(group)) {
2305 /* Bias balancing toward cpus of our domain */
2307 load = source_load(i, load_idx);
2309 load = target_load(i, load_idx);
2314 /* Adjust by relative CPU power of the group */
2315 avg_load = (avg_load * SCHED_LOAD_SCALE) / group->cpu_power;
2318 this_load = avg_load;
2320 } else if (avg_load < min_load) {
2321 min_load = avg_load;
2324 } while (group = group->next, group != sd->groups);
2326 if (!idlest || 100*this_load < imbalance*min_load)
2332 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2335 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2337 unsigned long load, min_load = ULONG_MAX;
2341 /* Traverse only the allowed CPUs */
2342 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2343 load = weighted_cpuload(i);
2345 if (load < min_load || (load == min_load && i == this_cpu)) {
2355 * sched_balance_self: balance the current task (running on cpu) in domains
2356 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2359 * Balance, ie. select the least loaded group.
2361 * Returns the target CPU number, or the same CPU if no balancing is needed.
2363 * preempt must be disabled.
2365 static int sched_balance_self(int cpu, int flag)
2367 struct task_struct *t = current;
2368 struct sched_domain *tmp, *sd = NULL;
2370 for_each_domain(cpu, tmp) {
2372 * If power savings logic is enabled for a domain, stop there.
2374 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2376 if (tmp->flags & flag)
2384 struct sched_group *group;
2385 int new_cpu, weight;
2387 if (!(sd->flags & flag)) {
2392 group = find_idlest_group(sd, t, cpu);
2398 new_cpu = find_idlest_cpu(group, t, cpu);
2399 if (new_cpu == -1 || new_cpu == cpu) {
2400 /* Now try balancing at a lower domain level of cpu */
2405 /* Now try balancing at a lower domain level of new_cpu */
2407 weight = cpumask_weight(sched_domain_span(sd));
2409 for_each_domain(cpu, tmp) {
2410 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2412 if (tmp->flags & flag)
2415 /* while loop will break here if sd == NULL */
2421 #endif /* CONFIG_SMP */
2424 * task_oncpu_function_call - call a function on the cpu on which a task runs
2425 * @p: the task to evaluate
2426 * @func: the function to be called
2427 * @info: the function call argument
2429 * Calls the function @func when the task is currently running. This might
2430 * be on the current CPU, which just calls the function directly
2432 void task_oncpu_function_call(struct task_struct *p,
2433 void (*func) (void *info), void *info)
2440 smp_call_function_single(cpu, func, info, 1);
2445 * try_to_wake_up - wake up a thread
2446 * @p: the to-be-woken-up thread
2447 * @state: the mask of task states that can be woken
2448 * @sync: do a synchronous wakeup?
2450 * Put it on the run-queue if it's not already there. The "current"
2451 * thread is always on the run-queue (except when the actual
2452 * re-schedule is in progress), and as such you're allowed to do
2453 * the simpler "current->state = TASK_RUNNING" to mark yourself
2454 * runnable without the overhead of this.
2456 * returns failure only if the task is already active.
2458 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2460 int cpu, orig_cpu, this_cpu, success = 0;
2461 unsigned long flags;
2465 if (!sched_feat(SYNC_WAKEUPS))
2469 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2470 struct sched_domain *sd;
2472 this_cpu = raw_smp_processor_id();
2475 for_each_domain(this_cpu, sd) {
2476 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2485 rq = task_rq_lock(p, &flags);
2486 update_rq_clock(rq);
2487 old_state = p->state;
2488 if (!(old_state & state))
2496 this_cpu = smp_processor_id();
2499 if (unlikely(task_running(rq, p)))
2502 cpu = p->sched_class->select_task_rq(p, sync);
2503 if (cpu != orig_cpu) {
2504 set_task_cpu(p, cpu);
2505 task_rq_unlock(rq, &flags);
2506 /* might preempt at this point */
2507 rq = task_rq_lock(p, &flags);
2508 old_state = p->state;
2509 if (!(old_state & state))
2514 this_cpu = smp_processor_id();
2518 #ifdef CONFIG_SCHEDSTATS
2519 schedstat_inc(rq, ttwu_count);
2520 if (cpu == this_cpu)
2521 schedstat_inc(rq, ttwu_local);
2523 struct sched_domain *sd;
2524 for_each_domain(this_cpu, sd) {
2525 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2526 schedstat_inc(sd, ttwu_wake_remote);
2531 #endif /* CONFIG_SCHEDSTATS */
2534 #endif /* CONFIG_SMP */
2535 schedstat_inc(p, se.nr_wakeups);
2537 schedstat_inc(p, se.nr_wakeups_sync);
2538 if (orig_cpu != cpu)
2539 schedstat_inc(p, se.nr_wakeups_migrate);
2540 if (cpu == this_cpu)
2541 schedstat_inc(p, se.nr_wakeups_local);
2543 schedstat_inc(p, se.nr_wakeups_remote);
2544 activate_task(rq, p, 1);
2548 * Only attribute actual wakeups done by this task.
2550 if (!in_interrupt()) {
2551 struct sched_entity *se = ¤t->se;
2552 u64 sample = se->sum_exec_runtime;
2554 if (se->last_wakeup)
2555 sample -= se->last_wakeup;
2557 sample -= se->start_runtime;
2558 update_avg(&se->avg_wakeup, sample);
2560 se->last_wakeup = se->sum_exec_runtime;
2564 trace_sched_wakeup(rq, p, success);
2565 check_preempt_curr(rq, p, sync);
2567 p->state = TASK_RUNNING;
2569 if (p->sched_class->task_wake_up)
2570 p->sched_class->task_wake_up(rq, p);
2573 task_rq_unlock(rq, &flags);
2579 * wake_up_process - Wake up a specific process
2580 * @p: The process to be woken up.
2582 * Attempt to wake up the nominated process and move it to the set of runnable
2583 * processes. Returns 1 if the process was woken up, 0 if it was already
2586 * It may be assumed that this function implies a write memory barrier before
2587 * changing the task state if and only if any tasks are woken up.
2589 int wake_up_process(struct task_struct *p)
2591 return try_to_wake_up(p, TASK_ALL, 0);
2593 EXPORT_SYMBOL(wake_up_process);
2595 int wake_up_state(struct task_struct *p, unsigned int state)
2597 return try_to_wake_up(p, state, 0);
2601 * Perform scheduler related setup for a newly forked process p.
2602 * p is forked by current.
2604 * __sched_fork() is basic setup used by init_idle() too:
2606 static void __sched_fork(struct task_struct *p)
2608 p->se.exec_start = 0;
2609 p->se.sum_exec_runtime = 0;
2610 p->se.prev_sum_exec_runtime = 0;
2611 p->se.nr_migrations = 0;
2612 p->se.last_wakeup = 0;
2613 p->se.avg_overlap = 0;
2614 p->se.start_runtime = 0;
2615 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2617 #ifdef CONFIG_SCHEDSTATS
2618 p->se.wait_start = 0;
2620 p->se.wait_count = 0;
2623 p->se.sleep_start = 0;
2624 p->se.sleep_max = 0;
2625 p->se.sum_sleep_runtime = 0;
2627 p->se.block_start = 0;
2628 p->se.block_max = 0;
2630 p->se.slice_max = 0;
2632 p->se.nr_migrations_cold = 0;
2633 p->se.nr_failed_migrations_affine = 0;
2634 p->se.nr_failed_migrations_running = 0;
2635 p->se.nr_failed_migrations_hot = 0;
2636 p->se.nr_forced_migrations = 0;
2637 p->se.nr_forced2_migrations = 0;
2639 p->se.nr_wakeups = 0;
2640 p->se.nr_wakeups_sync = 0;
2641 p->se.nr_wakeups_migrate = 0;
2642 p->se.nr_wakeups_local = 0;
2643 p->se.nr_wakeups_remote = 0;
2644 p->se.nr_wakeups_affine = 0;
2645 p->se.nr_wakeups_affine_attempts = 0;
2646 p->se.nr_wakeups_passive = 0;
2647 p->se.nr_wakeups_idle = 0;
2651 INIT_LIST_HEAD(&p->rt.run_list);
2653 INIT_LIST_HEAD(&p->se.group_node);
2655 #ifdef CONFIG_PREEMPT_NOTIFIERS
2656 INIT_HLIST_HEAD(&p->preempt_notifiers);
2660 * We mark the process as running here, but have not actually
2661 * inserted it onto the runqueue yet. This guarantees that
2662 * nobody will actually run it, and a signal or other external
2663 * event cannot wake it up and insert it on the runqueue either.
2665 p->state = TASK_RUNNING;
2669 * fork()/clone()-time setup:
2671 void sched_fork(struct task_struct *p, int clone_flags)
2673 int cpu = get_cpu();
2678 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2680 set_task_cpu(p, cpu);
2683 * Make sure we do not leak PI boosting priority to the child.
2685 p->prio = current->normal_prio;
2688 * Revert to default priority/policy on fork if requested.
2690 if (unlikely(p->sched_reset_on_fork)) {
2691 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2692 p->policy = SCHED_NORMAL;
2694 if (p->normal_prio < DEFAULT_PRIO)
2695 p->prio = DEFAULT_PRIO;
2697 if (PRIO_TO_NICE(p->static_prio) < 0) {
2698 p->static_prio = NICE_TO_PRIO(0);
2703 * We don't need the reset flag anymore after the fork. It has
2704 * fulfilled its duty:
2706 p->sched_reset_on_fork = 0;
2709 if (!rt_prio(p->prio))
2710 p->sched_class = &fair_sched_class;
2712 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2713 if (likely(sched_info_on()))
2714 memset(&p->sched_info, 0, sizeof(p->sched_info));
2716 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2719 #ifdef CONFIG_PREEMPT
2720 /* Want to start with kernel preemption disabled. */
2721 task_thread_info(p)->preempt_count = 1;
2723 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2729 * wake_up_new_task - wake up a newly created task for the first time.
2731 * This function will do some initial scheduler statistics housekeeping
2732 * that must be done for every newly created context, then puts the task
2733 * on the runqueue and wakes it.
2735 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2737 unsigned long flags;
2740 rq = task_rq_lock(p, &flags);
2741 BUG_ON(p->state != TASK_RUNNING);
2742 update_rq_clock(rq);
2744 p->prio = effective_prio(p);
2746 if (!p->sched_class->task_new || !current->se.on_rq) {
2747 activate_task(rq, p, 0);
2750 * Let the scheduling class do new task startup
2751 * management (if any):
2753 p->sched_class->task_new(rq, p);
2756 trace_sched_wakeup_new(rq, p, 1);
2757 check_preempt_curr(rq, p, 0);
2759 if (p->sched_class->task_wake_up)
2760 p->sched_class->task_wake_up(rq, p);
2762 task_rq_unlock(rq, &flags);
2765 #ifdef CONFIG_PREEMPT_NOTIFIERS
2768 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2769 * @notifier: notifier struct to register
2771 void preempt_notifier_register(struct preempt_notifier *notifier)
2773 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2775 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2778 * preempt_notifier_unregister - no longer interested in preemption notifications
2779 * @notifier: notifier struct to unregister
2781 * This is safe to call from within a preemption notifier.
2783 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2785 hlist_del(¬ifier->link);
2787 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2789 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2791 struct preempt_notifier *notifier;
2792 struct hlist_node *node;
2794 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2795 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2799 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2800 struct task_struct *next)
2802 struct preempt_notifier *notifier;
2803 struct hlist_node *node;
2805 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2806 notifier->ops->sched_out(notifier, next);
2809 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2811 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2816 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2817 struct task_struct *next)
2821 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2824 * prepare_task_switch - prepare to switch tasks
2825 * @rq: the runqueue preparing to switch
2826 * @prev: the current task that is being switched out
2827 * @next: the task we are going to switch to.
2829 * This is called with the rq lock held and interrupts off. It must
2830 * be paired with a subsequent finish_task_switch after the context
2833 * prepare_task_switch sets up locking and calls architecture specific
2837 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2838 struct task_struct *next)
2840 fire_sched_out_preempt_notifiers(prev, next);
2841 prepare_lock_switch(rq, next);
2842 prepare_arch_switch(next);
2846 * finish_task_switch - clean up after a task-switch
2847 * @rq: runqueue associated with task-switch
2848 * @prev: the thread we just switched away from.
2850 * finish_task_switch must be called after the context switch, paired
2851 * with a prepare_task_switch call before the context switch.
2852 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2853 * and do any other architecture-specific cleanup actions.
2855 * Note that we may have delayed dropping an mm in context_switch(). If
2856 * so, we finish that here outside of the runqueue lock. (Doing it
2857 * with the lock held can cause deadlocks; see schedule() for
2860 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2861 __releases(rq->lock)
2863 struct mm_struct *mm = rq->prev_mm;
2869 * A task struct has one reference for the use as "current".
2870 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2871 * schedule one last time. The schedule call will never return, and
2872 * the scheduled task must drop that reference.
2873 * The test for TASK_DEAD must occur while the runqueue locks are
2874 * still held, otherwise prev could be scheduled on another cpu, die
2875 * there before we look at prev->state, and then the reference would
2877 * Manfred Spraul <manfred@colorfullife.com>
2879 prev_state = prev->state;
2880 finish_arch_switch(prev);
2881 perf_counter_task_sched_in(current, cpu_of(rq));
2882 finish_lock_switch(rq, prev);
2884 fire_sched_in_preempt_notifiers(current);
2887 if (unlikely(prev_state == TASK_DEAD)) {
2889 * Remove function-return probe instances associated with this
2890 * task and put them back on the free list.
2892 kprobe_flush_task(prev);
2893 put_task_struct(prev);
2899 /* assumes rq->lock is held */
2900 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2902 if (prev->sched_class->pre_schedule)
2903 prev->sched_class->pre_schedule(rq, prev);
2906 /* rq->lock is NOT held, but preemption is disabled */
2907 static inline void post_schedule(struct rq *rq)
2909 if (rq->post_schedule) {
2910 unsigned long flags;
2912 spin_lock_irqsave(&rq->lock, flags);
2913 if (rq->curr->sched_class->post_schedule)
2914 rq->curr->sched_class->post_schedule(rq);
2915 spin_unlock_irqrestore(&rq->lock, flags);
2917 rq->post_schedule = 0;
2923 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2927 static inline void post_schedule(struct rq *rq)
2934 * schedule_tail - first thing a freshly forked thread must call.
2935 * @prev: the thread we just switched away from.
2937 asmlinkage void schedule_tail(struct task_struct *prev)
2938 __releases(rq->lock)
2940 struct rq *rq = this_rq();
2942 finish_task_switch(rq, prev);
2945 * FIXME: do we need to worry about rq being invalidated by the
2950 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2951 /* In this case, finish_task_switch does not reenable preemption */
2954 if (current->set_child_tid)
2955 put_user(task_pid_vnr(current), current->set_child_tid);
2959 * context_switch - switch to the new MM and the new
2960 * thread's register state.
2963 context_switch(struct rq *rq, struct task_struct *prev,
2964 struct task_struct *next)
2966 struct mm_struct *mm, *oldmm;
2968 prepare_task_switch(rq, prev, next);
2969 trace_sched_switch(rq, prev, next);
2971 oldmm = prev->active_mm;
2973 * For paravirt, this is coupled with an exit in switch_to to
2974 * combine the page table reload and the switch backend into
2977 arch_start_context_switch(prev);
2979 if (unlikely(!mm)) {
2980 next->active_mm = oldmm;
2981 atomic_inc(&oldmm->mm_count);
2982 enter_lazy_tlb(oldmm, next);
2984 switch_mm(oldmm, mm, next);
2986 if (unlikely(!prev->mm)) {
2987 prev->active_mm = NULL;
2988 rq->prev_mm = oldmm;
2991 * Since the runqueue lock will be released by the next
2992 * task (which is an invalid locking op but in the case
2993 * of the scheduler it's an obvious special-case), so we
2994 * do an early lockdep release here:
2996 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2997 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
3000 /* Here we just switch the register state and the stack. */
3001 switch_to(prev, next, prev);
3005 * this_rq must be evaluated again because prev may have moved
3006 * CPUs since it called schedule(), thus the 'rq' on its stack
3007 * frame will be invalid.
3009 finish_task_switch(this_rq(), prev);
3013 * nr_running, nr_uninterruptible and nr_context_switches:
3015 * externally visible scheduler statistics: current number of runnable
3016 * threads, current number of uninterruptible-sleeping threads, total
3017 * number of context switches performed since bootup.
3019 unsigned long nr_running(void)
3021 unsigned long i, sum = 0;
3023 for_each_online_cpu(i)
3024 sum += cpu_rq(i)->nr_running;
3029 unsigned long nr_uninterruptible(void)
3031 unsigned long i, sum = 0;
3033 for_each_possible_cpu(i)
3034 sum += cpu_rq(i)->nr_uninterruptible;
3037 * Since we read the counters lockless, it might be slightly
3038 * inaccurate. Do not allow it to go below zero though:
3040 if (unlikely((long)sum < 0))
3046 unsigned long long nr_context_switches(void)
3049 unsigned long long sum = 0;
3051 for_each_possible_cpu(i)
3052 sum += cpu_rq(i)->nr_switches;
3057 unsigned long nr_iowait(void)
3059 unsigned long i, sum = 0;
3061 for_each_possible_cpu(i)
3062 sum += atomic_read(&cpu_rq(i)->nr_iowait);
3067 /* Variables and functions for calc_load */
3068 static atomic_long_t calc_load_tasks;
3069 static unsigned long calc_load_update;
3070 unsigned long avenrun[3];
3071 EXPORT_SYMBOL(avenrun);
3074 * get_avenrun - get the load average array
3075 * @loads: pointer to dest load array
3076 * @offset: offset to add
3077 * @shift: shift count to shift the result left
3079 * These values are estimates at best, so no need for locking.
3081 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
3083 loads[0] = (avenrun[0] + offset) << shift;
3084 loads[1] = (avenrun[1] + offset) << shift;
3085 loads[2] = (avenrun[2] + offset) << shift;
3088 static unsigned long
3089 calc_load(unsigned long load, unsigned long exp, unsigned long active)
3092 load += active * (FIXED_1 - exp);
3093 return load >> FSHIFT;
3097 * calc_load - update the avenrun load estimates 10 ticks after the
3098 * CPUs have updated calc_load_tasks.
3100 void calc_global_load(void)
3102 unsigned long upd = calc_load_update + 10;
3105 if (time_before(jiffies, upd))
3108 active = atomic_long_read(&calc_load_tasks);
3109 active = active > 0 ? active * FIXED_1 : 0;
3111 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
3112 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
3113 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
3115 calc_load_update += LOAD_FREQ;
3119 * Either called from update_cpu_load() or from a cpu going idle
3121 static void calc_load_account_active(struct rq *this_rq)
3123 long nr_active, delta;
3125 nr_active = this_rq->nr_running;
3126 nr_active += (long) this_rq->nr_uninterruptible;
3128 if (nr_active != this_rq->calc_load_active) {
3129 delta = nr_active - this_rq->calc_load_active;
3130 this_rq->calc_load_active = nr_active;
3131 atomic_long_add(delta, &calc_load_tasks);
3136 * Externally visible per-cpu scheduler statistics:
3137 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3139 u64 cpu_nr_migrations(int cpu)
3141 return cpu_rq(cpu)->nr_migrations_in;
3145 * Update rq->cpu_load[] statistics. This function is usually called every
3146 * scheduler tick (TICK_NSEC).
3148 static void update_cpu_load(struct rq *this_rq)
3150 unsigned long this_load = this_rq->load.weight;
3153 this_rq->nr_load_updates++;
3155 /* Update our load: */
3156 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3157 unsigned long old_load, new_load;
3159 /* scale is effectively 1 << i now, and >> i divides by scale */
3161 old_load = this_rq->cpu_load[i];
3162 new_load = this_load;
3164 * Round up the averaging division if load is increasing. This
3165 * prevents us from getting stuck on 9 if the load is 10, for
3168 if (new_load > old_load)
3169 new_load += scale-1;
3170 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3173 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3174 this_rq->calc_load_update += LOAD_FREQ;
3175 calc_load_account_active(this_rq);
3182 * double_rq_lock - safely lock two runqueues
3184 * Note this does not disable interrupts like task_rq_lock,
3185 * you need to do so manually before calling.
3187 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3188 __acquires(rq1->lock)
3189 __acquires(rq2->lock)
3191 BUG_ON(!irqs_disabled());
3193 spin_lock(&rq1->lock);
3194 __acquire(rq2->lock); /* Fake it out ;) */
3197 spin_lock(&rq1->lock);
3198 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3200 spin_lock(&rq2->lock);
3201 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3204 update_rq_clock(rq1);
3205 update_rq_clock(rq2);
3209 * double_rq_unlock - safely unlock two runqueues
3211 * Note this does not restore interrupts like task_rq_unlock,
3212 * you need to do so manually after calling.
3214 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3215 __releases(rq1->lock)
3216 __releases(rq2->lock)
3218 spin_unlock(&rq1->lock);
3220 spin_unlock(&rq2->lock);
3222 __release(rq2->lock);
3226 * If dest_cpu is allowed for this process, migrate the task to it.
3227 * This is accomplished by forcing the cpu_allowed mask to only
3228 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3229 * the cpu_allowed mask is restored.
3231 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3233 struct migration_req req;
3234 unsigned long flags;
3237 rq = task_rq_lock(p, &flags);
3238 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3239 || unlikely(!cpu_active(dest_cpu)))
3242 /* force the process onto the specified CPU */
3243 if (migrate_task(p, dest_cpu, &req)) {
3244 /* Need to wait for migration thread (might exit: take ref). */
3245 struct task_struct *mt = rq->migration_thread;
3247 get_task_struct(mt);
3248 task_rq_unlock(rq, &flags);
3249 wake_up_process(mt);
3250 put_task_struct(mt);
3251 wait_for_completion(&req.done);
3256 task_rq_unlock(rq, &flags);
3260 * sched_exec - execve() is a valuable balancing opportunity, because at
3261 * this point the task has the smallest effective memory and cache footprint.
3263 void sched_exec(void)
3265 int new_cpu, this_cpu = get_cpu();
3266 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3268 if (new_cpu != this_cpu)
3269 sched_migrate_task(current, new_cpu);
3273 * pull_task - move a task from a remote runqueue to the local runqueue.
3274 * Both runqueues must be locked.
3276 static void pull_task(struct rq *src_rq, struct task_struct *p,
3277 struct rq *this_rq, int this_cpu)
3279 deactivate_task(src_rq, p, 0);
3280 set_task_cpu(p, this_cpu);
3281 activate_task(this_rq, p, 0);
3283 * Note that idle threads have a prio of MAX_PRIO, for this test
3284 * to be always true for them.
3286 check_preempt_curr(this_rq, p, 0);
3290 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3293 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3294 struct sched_domain *sd, enum cpu_idle_type idle,
3297 int tsk_cache_hot = 0;
3299 * We do not migrate tasks that are:
3300 * 1) running (obviously), or
3301 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3302 * 3) are cache-hot on their current CPU.
3304 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3305 schedstat_inc(p, se.nr_failed_migrations_affine);
3310 if (task_running(rq, p)) {
3311 schedstat_inc(p, se.nr_failed_migrations_running);
3316 * Aggressive migration if:
3317 * 1) task is cache cold, or
3318 * 2) too many balance attempts have failed.
3321 tsk_cache_hot = task_hot(p, rq->clock, sd);
3322 if (!tsk_cache_hot ||
3323 sd->nr_balance_failed > sd->cache_nice_tries) {
3324 #ifdef CONFIG_SCHEDSTATS
3325 if (tsk_cache_hot) {
3326 schedstat_inc(sd, lb_hot_gained[idle]);
3327 schedstat_inc(p, se.nr_forced_migrations);
3333 if (tsk_cache_hot) {
3334 schedstat_inc(p, se.nr_failed_migrations_hot);
3340 static unsigned long
3341 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3342 unsigned long max_load_move, struct sched_domain *sd,
3343 enum cpu_idle_type idle, int *all_pinned,
3344 int *this_best_prio, struct rq_iterator *iterator)
3346 int loops = 0, pulled = 0, pinned = 0;
3347 struct task_struct *p;
3348 long rem_load_move = max_load_move;
3350 if (max_load_move == 0)
3356 * Start the load-balancing iterator:
3358 p = iterator->start(iterator->arg);
3360 if (!p || loops++ > sysctl_sched_nr_migrate)
3363 if ((p->se.load.weight >> 1) > rem_load_move ||
3364 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3365 p = iterator->next(iterator->arg);
3369 pull_task(busiest, p, this_rq, this_cpu);
3371 rem_load_move -= p->se.load.weight;
3373 #ifdef CONFIG_PREEMPT
3375 * NEWIDLE balancing is a source of latency, so preemptible kernels
3376 * will stop after the first task is pulled to minimize the critical
3379 if (idle == CPU_NEWLY_IDLE)
3384 * We only want to steal up to the prescribed amount of weighted load.
3386 if (rem_load_move > 0) {
3387 if (p->prio < *this_best_prio)
3388 *this_best_prio = p->prio;
3389 p = iterator->next(iterator->arg);
3394 * Right now, this is one of only two places pull_task() is called,
3395 * so we can safely collect pull_task() stats here rather than
3396 * inside pull_task().
3398 schedstat_add(sd, lb_gained[idle], pulled);
3401 *all_pinned = pinned;
3403 return max_load_move - rem_load_move;
3407 * move_tasks tries to move up to max_load_move weighted load from busiest to
3408 * this_rq, as part of a balancing operation within domain "sd".
3409 * Returns 1 if successful and 0 otherwise.
3411 * Called with both runqueues locked.
3413 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3414 unsigned long max_load_move,
3415 struct sched_domain *sd, enum cpu_idle_type idle,
3418 const struct sched_class *class = sched_class_highest;
3419 unsigned long total_load_moved = 0;
3420 int this_best_prio = this_rq->curr->prio;
3424 class->load_balance(this_rq, this_cpu, busiest,
3425 max_load_move - total_load_moved,
3426 sd, idle, all_pinned, &this_best_prio);
3427 class = class->next;
3429 #ifdef CONFIG_PREEMPT
3431 * NEWIDLE balancing is a source of latency, so preemptible
3432 * kernels will stop after the first task is pulled to minimize
3433 * the critical section.
3435 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3438 } while (class && max_load_move > total_load_moved);
3440 return total_load_moved > 0;
3444 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3445 struct sched_domain *sd, enum cpu_idle_type idle,
3446 struct rq_iterator *iterator)
3448 struct task_struct *p = iterator->start(iterator->arg);
3452 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3453 pull_task(busiest, p, this_rq, this_cpu);
3455 * Right now, this is only the second place pull_task()
3456 * is called, so we can safely collect pull_task()
3457 * stats here rather than inside pull_task().
3459 schedstat_inc(sd, lb_gained[idle]);
3463 p = iterator->next(iterator->arg);
3470 * move_one_task tries to move exactly one task from busiest to this_rq, as
3471 * part of active balancing operations within "domain".
3472 * Returns 1 if successful and 0 otherwise.
3474 * Called with both runqueues locked.
3476 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3477 struct sched_domain *sd, enum cpu_idle_type idle)
3479 const struct sched_class *class;
3481 for_each_class(class) {
3482 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3488 /********** Helpers for find_busiest_group ************************/
3490 * sd_lb_stats - Structure to store the statistics of a sched_domain
3491 * during load balancing.
3493 struct sd_lb_stats {
3494 struct sched_group *busiest; /* Busiest group in this sd */
3495 struct sched_group *this; /* Local group in this sd */
3496 unsigned long total_load; /* Total load of all groups in sd */
3497 unsigned long total_pwr; /* Total power of all groups in sd */
3498 unsigned long avg_load; /* Average load across all groups in sd */
3500 /** Statistics of this group */
3501 unsigned long this_load;
3502 unsigned long this_load_per_task;
3503 unsigned long this_nr_running;
3505 /* Statistics of the busiest group */
3506 unsigned long max_load;
3507 unsigned long busiest_load_per_task;
3508 unsigned long busiest_nr_running;
3510 int group_imb; /* Is there imbalance in this sd */
3511 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3512 int power_savings_balance; /* Is powersave balance needed for this sd */
3513 struct sched_group *group_min; /* Least loaded group in sd */
3514 struct sched_group *group_leader; /* Group which relieves group_min */
3515 unsigned long min_load_per_task; /* load_per_task in group_min */
3516 unsigned long leader_nr_running; /* Nr running of group_leader */
3517 unsigned long min_nr_running; /* Nr running of group_min */
3522 * sg_lb_stats - stats of a sched_group required for load_balancing
3524 struct sg_lb_stats {
3525 unsigned long avg_load; /*Avg load across the CPUs of the group */
3526 unsigned long group_load; /* Total load over the CPUs of the group */
3527 unsigned long sum_nr_running; /* Nr tasks running in the group */
3528 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3529 unsigned long group_capacity;
3530 int group_imb; /* Is there an imbalance in the group ? */
3534 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3535 * @group: The group whose first cpu is to be returned.
3537 static inline unsigned int group_first_cpu(struct sched_group *group)
3539 return cpumask_first(sched_group_cpus(group));
3543 * get_sd_load_idx - Obtain the load index for a given sched domain.
3544 * @sd: The sched_domain whose load_idx is to be obtained.
3545 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3547 static inline int get_sd_load_idx(struct sched_domain *sd,
3548 enum cpu_idle_type idle)
3554 load_idx = sd->busy_idx;
3557 case CPU_NEWLY_IDLE:
3558 load_idx = sd->newidle_idx;
3561 load_idx = sd->idle_idx;
3569 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3571 * init_sd_power_savings_stats - Initialize power savings statistics for
3572 * the given sched_domain, during load balancing.
3574 * @sd: Sched domain whose power-savings statistics are to be initialized.
3575 * @sds: Variable containing the statistics for sd.
3576 * @idle: Idle status of the CPU at which we're performing load-balancing.
3578 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3579 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3582 * Busy processors will not participate in power savings
3585 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3586 sds->power_savings_balance = 0;
3588 sds->power_savings_balance = 1;
3589 sds->min_nr_running = ULONG_MAX;
3590 sds->leader_nr_running = 0;
3595 * update_sd_power_savings_stats - Update the power saving stats for a
3596 * sched_domain while performing load balancing.
3598 * @group: sched_group belonging to the sched_domain under consideration.
3599 * @sds: Variable containing the statistics of the sched_domain
3600 * @local_group: Does group contain the CPU for which we're performing
3602 * @sgs: Variable containing the statistics of the group.
3604 static inline void update_sd_power_savings_stats(struct sched_group *group,
3605 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3608 if (!sds->power_savings_balance)
3612 * If the local group is idle or completely loaded
3613 * no need to do power savings balance at this domain
3615 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3616 !sds->this_nr_running))
3617 sds->power_savings_balance = 0;
3620 * If a group is already running at full capacity or idle,
3621 * don't include that group in power savings calculations
3623 if (!sds->power_savings_balance ||
3624 sgs->sum_nr_running >= sgs->group_capacity ||
3625 !sgs->sum_nr_running)
3629 * Calculate the group which has the least non-idle load.
3630 * This is the group from where we need to pick up the load
3633 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3634 (sgs->sum_nr_running == sds->min_nr_running &&
3635 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3636 sds->group_min = group;
3637 sds->min_nr_running = sgs->sum_nr_running;
3638 sds->min_load_per_task = sgs->sum_weighted_load /
3639 sgs->sum_nr_running;
3643 * Calculate the group which is almost near its
3644 * capacity but still has some space to pick up some load
3645 * from other group and save more power
3647 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3650 if (sgs->sum_nr_running > sds->leader_nr_running ||
3651 (sgs->sum_nr_running == sds->leader_nr_running &&
3652 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3653 sds->group_leader = group;
3654 sds->leader_nr_running = sgs->sum_nr_running;
3659 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3660 * @sds: Variable containing the statistics of the sched_domain
3661 * under consideration.
3662 * @this_cpu: Cpu at which we're currently performing load-balancing.
3663 * @imbalance: Variable to store the imbalance.
3666 * Check if we have potential to perform some power-savings balance.
3667 * If yes, set the busiest group to be the least loaded group in the
3668 * sched_domain, so that it's CPUs can be put to idle.
3670 * Returns 1 if there is potential to perform power-savings balance.
3673 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3674 int this_cpu, unsigned long *imbalance)
3676 if (!sds->power_savings_balance)
3679 if (sds->this != sds->group_leader ||
3680 sds->group_leader == sds->group_min)
3683 *imbalance = sds->min_load_per_task;
3684 sds->busiest = sds->group_min;
3686 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3687 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3688 group_first_cpu(sds->group_leader);
3694 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3695 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3696 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3701 static inline void update_sd_power_savings_stats(struct sched_group *group,
3702 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3707 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3708 int this_cpu, unsigned long *imbalance)
3712 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3714 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3716 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3717 unsigned long smt_gain = sd->smt_gain;
3724 unsigned long scale_rt_power(int cpu)
3726 struct rq *rq = cpu_rq(cpu);
3727 u64 total, available;
3729 sched_avg_update(rq);
3731 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3732 available = total - rq->rt_avg;
3734 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3735 total = SCHED_LOAD_SCALE;
3737 total >>= SCHED_LOAD_SHIFT;
3739 return div_u64(available, total);
3742 static void update_cpu_power(struct sched_domain *sd, int cpu)
3744 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3745 unsigned long power = SCHED_LOAD_SCALE;
3746 struct sched_group *sdg = sd->groups;
3748 /* here we could scale based on cpufreq */
3750 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3751 power *= arch_scale_smt_power(sd, cpu);
3752 power >>= SCHED_LOAD_SHIFT;
3755 power *= scale_rt_power(cpu);
3756 power >>= SCHED_LOAD_SHIFT;
3761 sdg->cpu_power = power;
3764 static void update_group_power(struct sched_domain *sd, int cpu)
3766 struct sched_domain *child = sd->child;
3767 struct sched_group *group, *sdg = sd->groups;
3770 update_cpu_power(sd, cpu);
3776 group = child->groups;
3778 sdg->cpu_power += group->cpu_power;
3779 group = group->next;
3780 } while (group != child->groups);
3784 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3785 * @group: sched_group whose statistics are to be updated.
3786 * @this_cpu: Cpu for which load balance is currently performed.
3787 * @idle: Idle status of this_cpu
3788 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3789 * @sd_idle: Idle status of the sched_domain containing group.
3790 * @local_group: Does group contain this_cpu.
3791 * @cpus: Set of cpus considered for load balancing.
3792 * @balance: Should we balance.
3793 * @sgs: variable to hold the statistics for this group.
3795 static inline void update_sg_lb_stats(struct sched_domain *sd,
3796 struct sched_group *group, int this_cpu,
3797 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3798 int local_group, const struct cpumask *cpus,
3799 int *balance, struct sg_lb_stats *sgs)
3801 unsigned long load, max_cpu_load, min_cpu_load;
3803 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3804 unsigned long sum_avg_load_per_task;
3805 unsigned long avg_load_per_task;
3808 balance_cpu = group_first_cpu(group);
3809 if (balance_cpu == this_cpu)
3810 update_group_power(sd, this_cpu);
3813 /* Tally up the load of all CPUs in the group */
3814 sum_avg_load_per_task = avg_load_per_task = 0;
3816 min_cpu_load = ~0UL;
3818 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3819 struct rq *rq = cpu_rq(i);
3821 if (*sd_idle && rq->nr_running)
3824 /* Bias balancing toward cpus of our domain */
3826 if (idle_cpu(i) && !first_idle_cpu) {
3831 load = target_load(i, load_idx);
3833 load = source_load(i, load_idx);
3834 if (load > max_cpu_load)
3835 max_cpu_load = load;
3836 if (min_cpu_load > load)
3837 min_cpu_load = load;
3840 sgs->group_load += load;
3841 sgs->sum_nr_running += rq->nr_running;
3842 sgs->sum_weighted_load += weighted_cpuload(i);
3844 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3848 * First idle cpu or the first cpu(busiest) in this sched group
3849 * is eligible for doing load balancing at this and above
3850 * domains. In the newly idle case, we will allow all the cpu's
3851 * to do the newly idle load balance.
3853 if (idle != CPU_NEWLY_IDLE && local_group &&
3854 balance_cpu != this_cpu && balance) {
3859 /* Adjust by relative CPU power of the group */
3860 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3864 * Consider the group unbalanced when the imbalance is larger
3865 * than the average weight of two tasks.
3867 * APZ: with cgroup the avg task weight can vary wildly and
3868 * might not be a suitable number - should we keep a
3869 * normalized nr_running number somewhere that negates
3872 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3875 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3878 sgs->group_capacity =
3879 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3883 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3884 * @sd: sched_domain whose statistics are to be updated.
3885 * @this_cpu: Cpu for which load balance is currently performed.
3886 * @idle: Idle status of this_cpu
3887 * @sd_idle: Idle status of the sched_domain containing group.
3888 * @cpus: Set of cpus considered for load balancing.
3889 * @balance: Should we balance.
3890 * @sds: variable to hold the statistics for this sched_domain.
3892 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3893 enum cpu_idle_type idle, int *sd_idle,
3894 const struct cpumask *cpus, int *balance,
3895 struct sd_lb_stats *sds)
3897 struct sched_domain *child = sd->child;
3898 struct sched_group *group = sd->groups;
3899 struct sg_lb_stats sgs;
3900 int load_idx, prefer_sibling = 0;
3902 if (child && child->flags & SD_PREFER_SIBLING)
3905 init_sd_power_savings_stats(sd, sds, idle);
3906 load_idx = get_sd_load_idx(sd, idle);
3911 local_group = cpumask_test_cpu(this_cpu,
3912 sched_group_cpus(group));
3913 memset(&sgs, 0, sizeof(sgs));
3914 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3915 local_group, cpus, balance, &sgs);
3917 if (local_group && balance && !(*balance))
3920 sds->total_load += sgs.group_load;
3921 sds->total_pwr += group->cpu_power;
3924 * In case the child domain prefers tasks go to siblings
3925 * first, lower the group capacity to one so that we'll try
3926 * and move all the excess tasks away.
3929 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3932 sds->this_load = sgs.avg_load;
3934 sds->this_nr_running = sgs.sum_nr_running;
3935 sds->this_load_per_task = sgs.sum_weighted_load;
3936 } else if (sgs.avg_load > sds->max_load &&
3937 (sgs.sum_nr_running > sgs.group_capacity ||
3939 sds->max_load = sgs.avg_load;
3940 sds->busiest = group;
3941 sds->busiest_nr_running = sgs.sum_nr_running;
3942 sds->busiest_load_per_task = sgs.sum_weighted_load;
3943 sds->group_imb = sgs.group_imb;
3946 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3947 group = group->next;
3948 } while (group != sd->groups);
3952 * fix_small_imbalance - Calculate the minor imbalance that exists
3953 * amongst the groups of a sched_domain, during
3955 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3956 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3957 * @imbalance: Variable to store the imbalance.
3959 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3960 int this_cpu, unsigned long *imbalance)
3962 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3963 unsigned int imbn = 2;
3965 if (sds->this_nr_running) {
3966 sds->this_load_per_task /= sds->this_nr_running;
3967 if (sds->busiest_load_per_task >
3968 sds->this_load_per_task)
3971 sds->this_load_per_task =
3972 cpu_avg_load_per_task(this_cpu);
3974 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3975 sds->busiest_load_per_task * imbn) {
3976 *imbalance = sds->busiest_load_per_task;
3981 * OK, we don't have enough imbalance to justify moving tasks,
3982 * however we may be able to increase total CPU power used by
3986 pwr_now += sds->busiest->cpu_power *
3987 min(sds->busiest_load_per_task, sds->max_load);
3988 pwr_now += sds->this->cpu_power *
3989 min(sds->this_load_per_task, sds->this_load);
3990 pwr_now /= SCHED_LOAD_SCALE;
3992 /* Amount of load we'd subtract */
3993 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3994 sds->busiest->cpu_power;
3995 if (sds->max_load > tmp)
3996 pwr_move += sds->busiest->cpu_power *
3997 min(sds->busiest_load_per_task, sds->max_load - tmp);
3999 /* Amount of load we'd add */
4000 if (sds->max_load * sds->busiest->cpu_power <
4001 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
4002 tmp = (sds->max_load * sds->busiest->cpu_power) /
4003 sds->this->cpu_power;
4005 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
4006 sds->this->cpu_power;
4007 pwr_move += sds->this->cpu_power *
4008 min(sds->this_load_per_task, sds->this_load + tmp);
4009 pwr_move /= SCHED_LOAD_SCALE;
4011 /* Move if we gain throughput */
4012 if (pwr_move > pwr_now)
4013 *imbalance = sds->busiest_load_per_task;
4017 * calculate_imbalance - Calculate the amount of imbalance present within the
4018 * groups of a given sched_domain during load balance.
4019 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
4020 * @this_cpu: Cpu for which currently load balance is being performed.
4021 * @imbalance: The variable to store the imbalance.
4023 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
4024 unsigned long *imbalance)
4026 unsigned long max_pull;
4028 * In the presence of smp nice balancing, certain scenarios can have
4029 * max load less than avg load(as we skip the groups at or below
4030 * its cpu_power, while calculating max_load..)
4032 if (sds->max_load < sds->avg_load) {
4034 return fix_small_imbalance(sds, this_cpu, imbalance);
4037 /* Don't want to pull so many tasks that a group would go idle */
4038 max_pull = min(sds->max_load - sds->avg_load,
4039 sds->max_load - sds->busiest_load_per_task);
4041 /* How much load to actually move to equalise the imbalance */
4042 *imbalance = min(max_pull * sds->busiest->cpu_power,
4043 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
4047 * if *imbalance is less than the average load per runnable task
4048 * there is no gaurantee that any tasks will be moved so we'll have
4049 * a think about bumping its value to force at least one task to be
4052 if (*imbalance < sds->busiest_load_per_task)
4053 return fix_small_imbalance(sds, this_cpu, imbalance);
4056 /******* find_busiest_group() helpers end here *********************/
4059 * find_busiest_group - Returns the busiest group within the sched_domain
4060 * if there is an imbalance. If there isn't an imbalance, and
4061 * the user has opted for power-savings, it returns a group whose
4062 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
4063 * such a group exists.
4065 * Also calculates the amount of weighted load which should be moved
4066 * to restore balance.
4068 * @sd: The sched_domain whose busiest group is to be returned.
4069 * @this_cpu: The cpu for which load balancing is currently being performed.
4070 * @imbalance: Variable which stores amount of weighted load which should
4071 * be moved to restore balance/put a group to idle.
4072 * @idle: The idle status of this_cpu.
4073 * @sd_idle: The idleness of sd
4074 * @cpus: The set of CPUs under consideration for load-balancing.
4075 * @balance: Pointer to a variable indicating if this_cpu
4076 * is the appropriate cpu to perform load balancing at this_level.
4078 * Returns: - the busiest group if imbalance exists.
4079 * - If no imbalance and user has opted for power-savings balance,
4080 * return the least loaded group whose CPUs can be
4081 * put to idle by rebalancing its tasks onto our group.
4083 static struct sched_group *
4084 find_busiest_group(struct sched_domain *sd, int this_cpu,
4085 unsigned long *imbalance, enum cpu_idle_type idle,
4086 int *sd_idle, const struct cpumask *cpus, int *balance)
4088 struct sd_lb_stats sds;
4090 memset(&sds, 0, sizeof(sds));
4093 * Compute the various statistics relavent for load balancing at
4096 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
4099 /* Cases where imbalance does not exist from POV of this_cpu */
4100 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4102 * 2) There is no busy sibling group to pull from.
4103 * 3) This group is the busiest group.
4104 * 4) This group is more busy than the avg busieness at this
4106 * 5) The imbalance is within the specified limit.
4107 * 6) Any rebalance would lead to ping-pong
4109 if (balance && !(*balance))
4112 if (!sds.busiest || sds.busiest_nr_running == 0)
4115 if (sds.this_load >= sds.max_load)
4118 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
4120 if (sds.this_load >= sds.avg_load)
4123 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
4126 sds.busiest_load_per_task /= sds.busiest_nr_running;
4128 sds.busiest_load_per_task =
4129 min(sds.busiest_load_per_task, sds.avg_load);
4132 * We're trying to get all the cpus to the average_load, so we don't
4133 * want to push ourselves above the average load, nor do we wish to
4134 * reduce the max loaded cpu below the average load, as either of these
4135 * actions would just result in more rebalancing later, and ping-pong
4136 * tasks around. Thus we look for the minimum possible imbalance.
4137 * Negative imbalances (*we* are more loaded than anyone else) will
4138 * be counted as no imbalance for these purposes -- we can't fix that
4139 * by pulling tasks to us. Be careful of negative numbers as they'll
4140 * appear as very large values with unsigned longs.
4142 if (sds.max_load <= sds.busiest_load_per_task)
4145 /* Looks like there is an imbalance. Compute it */
4146 calculate_imbalance(&sds, this_cpu, imbalance);
4151 * There is no obvious imbalance. But check if we can do some balancing
4154 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4161 static struct sched_group *group_of(int cpu)
4163 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
4171 static unsigned long power_of(int cpu)
4173 struct sched_group *group = group_of(cpu);
4176 return SCHED_LOAD_SCALE;
4178 return group->cpu_power;
4182 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4185 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4186 unsigned long imbalance, const struct cpumask *cpus)
4188 struct rq *busiest = NULL, *rq;
4189 unsigned long max_load = 0;
4192 for_each_cpu(i, sched_group_cpus(group)) {
4193 unsigned long power = power_of(i);
4194 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4197 if (!cpumask_test_cpu(i, cpus))
4201 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4204 if (capacity && rq->nr_running == 1 && wl > imbalance)
4207 if (wl > max_load) {
4217 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4218 * so long as it is large enough.
4220 #define MAX_PINNED_INTERVAL 512
4222 /* Working cpumask for load_balance and load_balance_newidle. */
4223 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4226 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4227 * tasks if there is an imbalance.
4229 static int load_balance(int this_cpu, struct rq *this_rq,
4230 struct sched_domain *sd, enum cpu_idle_type idle,
4233 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4234 struct sched_group *group;
4235 unsigned long imbalance;
4237 unsigned long flags;
4238 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4240 cpumask_setall(cpus);
4243 * When power savings policy is enabled for the parent domain, idle
4244 * sibling can pick up load irrespective of busy siblings. In this case,
4245 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4246 * portraying it as CPU_NOT_IDLE.
4248 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4249 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4252 schedstat_inc(sd, lb_count[idle]);
4256 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4263 schedstat_inc(sd, lb_nobusyg[idle]);
4267 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4269 schedstat_inc(sd, lb_nobusyq[idle]);
4273 BUG_ON(busiest == this_rq);
4275 schedstat_add(sd, lb_imbalance[idle], imbalance);
4278 if (busiest->nr_running > 1) {
4280 * Attempt to move tasks. If find_busiest_group has found
4281 * an imbalance but busiest->nr_running <= 1, the group is
4282 * still unbalanced. ld_moved simply stays zero, so it is
4283 * correctly treated as an imbalance.
4285 local_irq_save(flags);
4286 double_rq_lock(this_rq, busiest);
4287 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4288 imbalance, sd, idle, &all_pinned);
4289 double_rq_unlock(this_rq, busiest);
4290 local_irq_restore(flags);
4293 * some other cpu did the load balance for us.
4295 if (ld_moved && this_cpu != smp_processor_id())
4296 resched_cpu(this_cpu);
4298 /* All tasks on this runqueue were pinned by CPU affinity */
4299 if (unlikely(all_pinned)) {
4300 cpumask_clear_cpu(cpu_of(busiest), cpus);
4301 if (!cpumask_empty(cpus))
4308 schedstat_inc(sd, lb_failed[idle]);
4309 sd->nr_balance_failed++;
4311 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4313 spin_lock_irqsave(&busiest->lock, flags);
4315 /* don't kick the migration_thread, if the curr
4316 * task on busiest cpu can't be moved to this_cpu
4318 if (!cpumask_test_cpu(this_cpu,
4319 &busiest->curr->cpus_allowed)) {
4320 spin_unlock_irqrestore(&busiest->lock, flags);
4322 goto out_one_pinned;
4325 if (!busiest->active_balance) {
4326 busiest->active_balance = 1;
4327 busiest->push_cpu = this_cpu;
4330 spin_unlock_irqrestore(&busiest->lock, flags);
4332 wake_up_process(busiest->migration_thread);
4335 * We've kicked active balancing, reset the failure
4338 sd->nr_balance_failed = sd->cache_nice_tries+1;
4341 sd->nr_balance_failed = 0;
4343 if (likely(!active_balance)) {
4344 /* We were unbalanced, so reset the balancing interval */
4345 sd->balance_interval = sd->min_interval;
4348 * If we've begun active balancing, start to back off. This
4349 * case may not be covered by the all_pinned logic if there
4350 * is only 1 task on the busy runqueue (because we don't call
4353 if (sd->balance_interval < sd->max_interval)
4354 sd->balance_interval *= 2;
4357 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4358 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4364 schedstat_inc(sd, lb_balanced[idle]);
4366 sd->nr_balance_failed = 0;
4369 /* tune up the balancing interval */
4370 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4371 (sd->balance_interval < sd->max_interval))
4372 sd->balance_interval *= 2;
4374 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4375 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4386 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4387 * tasks if there is an imbalance.
4389 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4390 * this_rq is locked.
4393 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4395 struct sched_group *group;
4396 struct rq *busiest = NULL;
4397 unsigned long imbalance;
4401 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4403 cpumask_setall(cpus);
4406 * When power savings policy is enabled for the parent domain, idle
4407 * sibling can pick up load irrespective of busy siblings. In this case,
4408 * let the state of idle sibling percolate up as IDLE, instead of
4409 * portraying it as CPU_NOT_IDLE.
4411 if (sd->flags & SD_SHARE_CPUPOWER &&
4412 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4415 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4417 update_shares_locked(this_rq, sd);
4418 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4419 &sd_idle, cpus, NULL);
4421 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4425 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4427 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4431 BUG_ON(busiest == this_rq);
4433 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4436 if (busiest->nr_running > 1) {
4437 /* Attempt to move tasks */
4438 double_lock_balance(this_rq, busiest);
4439 /* this_rq->clock is already updated */
4440 update_rq_clock(busiest);
4441 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4442 imbalance, sd, CPU_NEWLY_IDLE,
4444 double_unlock_balance(this_rq, busiest);
4446 if (unlikely(all_pinned)) {
4447 cpumask_clear_cpu(cpu_of(busiest), cpus);
4448 if (!cpumask_empty(cpus))
4454 int active_balance = 0;
4456 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4457 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4458 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4461 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4464 if (sd->nr_balance_failed++ < 2)
4468 * The only task running in a non-idle cpu can be moved to this
4469 * cpu in an attempt to completely freeup the other CPU
4470 * package. The same method used to move task in load_balance()
4471 * have been extended for load_balance_newidle() to speedup
4472 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4474 * The package power saving logic comes from
4475 * find_busiest_group(). If there are no imbalance, then
4476 * f_b_g() will return NULL. However when sched_mc={1,2} then
4477 * f_b_g() will select a group from which a running task may be
4478 * pulled to this cpu in order to make the other package idle.
4479 * If there is no opportunity to make a package idle and if
4480 * there are no imbalance, then f_b_g() will return NULL and no
4481 * action will be taken in load_balance_newidle().
4483 * Under normal task pull operation due to imbalance, there
4484 * will be more than one task in the source run queue and
4485 * move_tasks() will succeed. ld_moved will be true and this
4486 * active balance code will not be triggered.
4489 /* Lock busiest in correct order while this_rq is held */
4490 double_lock_balance(this_rq, busiest);
4493 * don't kick the migration_thread, if the curr
4494 * task on busiest cpu can't be moved to this_cpu
4496 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4497 double_unlock_balance(this_rq, busiest);
4502 if (!busiest->active_balance) {
4503 busiest->active_balance = 1;
4504 busiest->push_cpu = this_cpu;
4508 double_unlock_balance(this_rq, busiest);
4510 * Should not call ttwu while holding a rq->lock
4512 spin_unlock(&this_rq->lock);
4514 wake_up_process(busiest->migration_thread);
4515 spin_lock(&this_rq->lock);
4518 sd->nr_balance_failed = 0;
4520 update_shares_locked(this_rq, sd);
4524 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4525 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4526 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4528 sd->nr_balance_failed = 0;
4534 * idle_balance is called by schedule() if this_cpu is about to become
4535 * idle. Attempts to pull tasks from other CPUs.
4537 static void idle_balance(int this_cpu, struct rq *this_rq)
4539 struct sched_domain *sd;
4540 int pulled_task = 0;
4541 unsigned long next_balance = jiffies + HZ;
4543 for_each_domain(this_cpu, sd) {
4544 unsigned long interval;
4546 if (!(sd->flags & SD_LOAD_BALANCE))
4549 if (sd->flags & SD_BALANCE_NEWIDLE)
4550 /* If we've pulled tasks over stop searching: */
4551 pulled_task = load_balance_newidle(this_cpu, this_rq,
4554 interval = msecs_to_jiffies(sd->balance_interval);
4555 if (time_after(next_balance, sd->last_balance + interval))
4556 next_balance = sd->last_balance + interval;
4560 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4562 * We are going idle. next_balance may be set based on
4563 * a busy processor. So reset next_balance.
4565 this_rq->next_balance = next_balance;
4570 * active_load_balance is run by migration threads. It pushes running tasks
4571 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4572 * running on each physical CPU where possible, and avoids physical /
4573 * logical imbalances.
4575 * Called with busiest_rq locked.
4577 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4579 int target_cpu = busiest_rq->push_cpu;
4580 struct sched_domain *sd;
4581 struct rq *target_rq;
4583 /* Is there any task to move? */
4584 if (busiest_rq->nr_running <= 1)
4587 target_rq = cpu_rq(target_cpu);
4590 * This condition is "impossible", if it occurs
4591 * we need to fix it. Originally reported by
4592 * Bjorn Helgaas on a 128-cpu setup.
4594 BUG_ON(busiest_rq == target_rq);
4596 /* move a task from busiest_rq to target_rq */
4597 double_lock_balance(busiest_rq, target_rq);
4598 update_rq_clock(busiest_rq);
4599 update_rq_clock(target_rq);
4601 /* Search for an sd spanning us and the target CPU. */
4602 for_each_domain(target_cpu, sd) {
4603 if ((sd->flags & SD_LOAD_BALANCE) &&
4604 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4609 schedstat_inc(sd, alb_count);
4611 if (move_one_task(target_rq, target_cpu, busiest_rq,
4613 schedstat_inc(sd, alb_pushed);
4615 schedstat_inc(sd, alb_failed);
4617 double_unlock_balance(busiest_rq, target_rq);
4622 atomic_t load_balancer;
4623 cpumask_var_t cpu_mask;
4624 cpumask_var_t ilb_grp_nohz_mask;
4625 } nohz ____cacheline_aligned = {
4626 .load_balancer = ATOMIC_INIT(-1),
4629 int get_nohz_load_balancer(void)
4631 return atomic_read(&nohz.load_balancer);
4634 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4636 * lowest_flag_domain - Return lowest sched_domain containing flag.
4637 * @cpu: The cpu whose lowest level of sched domain is to
4639 * @flag: The flag to check for the lowest sched_domain
4640 * for the given cpu.
4642 * Returns the lowest sched_domain of a cpu which contains the given flag.
4644 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4646 struct sched_domain *sd;
4648 for_each_domain(cpu, sd)
4649 if (sd && (sd->flags & flag))
4656 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4657 * @cpu: The cpu whose domains we're iterating over.
4658 * @sd: variable holding the value of the power_savings_sd
4660 * @flag: The flag to filter the sched_domains to be iterated.
4662 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4663 * set, starting from the lowest sched_domain to the highest.
4665 #define for_each_flag_domain(cpu, sd, flag) \
4666 for (sd = lowest_flag_domain(cpu, flag); \
4667 (sd && (sd->flags & flag)); sd = sd->parent)
4670 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4671 * @ilb_group: group to be checked for semi-idleness
4673 * Returns: 1 if the group is semi-idle. 0 otherwise.
4675 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4676 * and atleast one non-idle CPU. This helper function checks if the given
4677 * sched_group is semi-idle or not.
4679 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4681 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4682 sched_group_cpus(ilb_group));
4685 * A sched_group is semi-idle when it has atleast one busy cpu
4686 * and atleast one idle cpu.
4688 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4691 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4697 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4698 * @cpu: The cpu which is nominating a new idle_load_balancer.
4700 * Returns: Returns the id of the idle load balancer if it exists,
4701 * Else, returns >= nr_cpu_ids.
4703 * This algorithm picks the idle load balancer such that it belongs to a
4704 * semi-idle powersavings sched_domain. The idea is to try and avoid
4705 * completely idle packages/cores just for the purpose of idle load balancing
4706 * when there are other idle cpu's which are better suited for that job.
4708 static int find_new_ilb(int cpu)
4710 struct sched_domain *sd;
4711 struct sched_group *ilb_group;
4714 * Have idle load balancer selection from semi-idle packages only
4715 * when power-aware load balancing is enabled
4717 if (!(sched_smt_power_savings || sched_mc_power_savings))
4721 * Optimize for the case when we have no idle CPUs or only one
4722 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4724 if (cpumask_weight(nohz.cpu_mask) < 2)
4727 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4728 ilb_group = sd->groups;
4731 if (is_semi_idle_group(ilb_group))
4732 return cpumask_first(nohz.ilb_grp_nohz_mask);
4734 ilb_group = ilb_group->next;
4736 } while (ilb_group != sd->groups);
4740 return cpumask_first(nohz.cpu_mask);
4742 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4743 static inline int find_new_ilb(int call_cpu)
4745 return cpumask_first(nohz.cpu_mask);
4750 * This routine will try to nominate the ilb (idle load balancing)
4751 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4752 * load balancing on behalf of all those cpus. If all the cpus in the system
4753 * go into this tickless mode, then there will be no ilb owner (as there is
4754 * no need for one) and all the cpus will sleep till the next wakeup event
4757 * For the ilb owner, tick is not stopped. And this tick will be used
4758 * for idle load balancing. ilb owner will still be part of
4761 * While stopping the tick, this cpu will become the ilb owner if there
4762 * is no other owner. And will be the owner till that cpu becomes busy
4763 * or if all cpus in the system stop their ticks at which point
4764 * there is no need for ilb owner.
4766 * When the ilb owner becomes busy, it nominates another owner, during the
4767 * next busy scheduler_tick()
4769 int select_nohz_load_balancer(int stop_tick)
4771 int cpu = smp_processor_id();
4774 cpu_rq(cpu)->in_nohz_recently = 1;
4776 if (!cpu_active(cpu)) {
4777 if (atomic_read(&nohz.load_balancer) != cpu)
4781 * If we are going offline and still the leader,
4784 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4790 cpumask_set_cpu(cpu, nohz.cpu_mask);
4792 /* time for ilb owner also to sleep */
4793 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4794 if (atomic_read(&nohz.load_balancer) == cpu)
4795 atomic_set(&nohz.load_balancer, -1);
4799 if (atomic_read(&nohz.load_balancer) == -1) {
4800 /* make me the ilb owner */
4801 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4803 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4806 if (!(sched_smt_power_savings ||
4807 sched_mc_power_savings))
4810 * Check to see if there is a more power-efficient
4813 new_ilb = find_new_ilb(cpu);
4814 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4815 atomic_set(&nohz.load_balancer, -1);
4816 resched_cpu(new_ilb);
4822 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4825 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4827 if (atomic_read(&nohz.load_balancer) == cpu)
4828 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4835 static DEFINE_SPINLOCK(balancing);
4838 * It checks each scheduling domain to see if it is due to be balanced,
4839 * and initiates a balancing operation if so.
4841 * Balancing parameters are set up in arch_init_sched_domains.
4843 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4846 struct rq *rq = cpu_rq(cpu);
4847 unsigned long interval;
4848 struct sched_domain *sd;
4849 /* Earliest time when we have to do rebalance again */
4850 unsigned long next_balance = jiffies + 60*HZ;
4851 int update_next_balance = 0;
4854 for_each_domain(cpu, sd) {
4855 if (!(sd->flags & SD_LOAD_BALANCE))
4858 interval = sd->balance_interval;
4859 if (idle != CPU_IDLE)
4860 interval *= sd->busy_factor;
4862 /* scale ms to jiffies */
4863 interval = msecs_to_jiffies(interval);
4864 if (unlikely(!interval))
4866 if (interval > HZ*NR_CPUS/10)
4867 interval = HZ*NR_CPUS/10;
4869 need_serialize = sd->flags & SD_SERIALIZE;
4871 if (need_serialize) {
4872 if (!spin_trylock(&balancing))
4876 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4877 if (load_balance(cpu, rq, sd, idle, &balance)) {
4879 * We've pulled tasks over so either we're no
4880 * longer idle, or one of our SMT siblings is
4883 idle = CPU_NOT_IDLE;
4885 sd->last_balance = jiffies;
4888 spin_unlock(&balancing);
4890 if (time_after(next_balance, sd->last_balance + interval)) {
4891 next_balance = sd->last_balance + interval;
4892 update_next_balance = 1;
4896 * Stop the load balance at this level. There is another
4897 * CPU in our sched group which is doing load balancing more
4905 * next_balance will be updated only when there is a need.
4906 * When the cpu is attached to null domain for ex, it will not be
4909 if (likely(update_next_balance))
4910 rq->next_balance = next_balance;
4914 * run_rebalance_domains is triggered when needed from the scheduler tick.
4915 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4916 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4918 static void run_rebalance_domains(struct softirq_action *h)
4920 int this_cpu = smp_processor_id();
4921 struct rq *this_rq = cpu_rq(this_cpu);
4922 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4923 CPU_IDLE : CPU_NOT_IDLE;
4925 rebalance_domains(this_cpu, idle);
4929 * If this cpu is the owner for idle load balancing, then do the
4930 * balancing on behalf of the other idle cpus whose ticks are
4933 if (this_rq->idle_at_tick &&
4934 atomic_read(&nohz.load_balancer) == this_cpu) {
4938 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4939 if (balance_cpu == this_cpu)
4943 * If this cpu gets work to do, stop the load balancing
4944 * work being done for other cpus. Next load
4945 * balancing owner will pick it up.
4950 rebalance_domains(balance_cpu, CPU_IDLE);
4952 rq = cpu_rq(balance_cpu);
4953 if (time_after(this_rq->next_balance, rq->next_balance))
4954 this_rq->next_balance = rq->next_balance;
4960 static inline int on_null_domain(int cpu)
4962 return !rcu_dereference(cpu_rq(cpu)->sd);
4966 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4968 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4969 * idle load balancing owner or decide to stop the periodic load balancing,
4970 * if the whole system is idle.
4972 static inline void trigger_load_balance(struct rq *rq, int cpu)
4976 * If we were in the nohz mode recently and busy at the current
4977 * scheduler tick, then check if we need to nominate new idle
4980 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4981 rq->in_nohz_recently = 0;
4983 if (atomic_read(&nohz.load_balancer) == cpu) {
4984 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4985 atomic_set(&nohz.load_balancer, -1);
4988 if (atomic_read(&nohz.load_balancer) == -1) {
4989 int ilb = find_new_ilb(cpu);
4991 if (ilb < nr_cpu_ids)
4997 * If this cpu is idle and doing idle load balancing for all the
4998 * cpus with ticks stopped, is it time for that to stop?
5000 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
5001 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
5007 * If this cpu is idle and the idle load balancing is done by
5008 * someone else, then no need raise the SCHED_SOFTIRQ
5010 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
5011 cpumask_test_cpu(cpu, nohz.cpu_mask))
5014 /* Don't need to rebalance while attached to NULL domain */
5015 if (time_after_eq(jiffies, rq->next_balance) &&
5016 likely(!on_null_domain(cpu)))
5017 raise_softirq(SCHED_SOFTIRQ);
5020 #else /* CONFIG_SMP */
5023 * on UP we do not need to balance between CPUs:
5025 static inline void idle_balance(int cpu, struct rq *rq)
5031 DEFINE_PER_CPU(struct kernel_stat, kstat);
5033 EXPORT_PER_CPU_SYMBOL(kstat);
5036 * Return any ns on the sched_clock that have not yet been accounted in
5037 * @p in case that task is currently running.
5039 * Called with task_rq_lock() held on @rq.
5041 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
5045 if (task_current(rq, p)) {
5046 update_rq_clock(rq);
5047 ns = rq->clock - p->se.exec_start;
5055 unsigned long long task_delta_exec(struct task_struct *p)
5057 unsigned long flags;
5061 rq = task_rq_lock(p, &flags);
5062 ns = do_task_delta_exec(p, rq);
5063 task_rq_unlock(rq, &flags);
5069 * Return accounted runtime for the task.
5070 * In case the task is currently running, return the runtime plus current's
5071 * pending runtime that have not been accounted yet.
5073 unsigned long long task_sched_runtime(struct task_struct *p)
5075 unsigned long flags;
5079 rq = task_rq_lock(p, &flags);
5080 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
5081 task_rq_unlock(rq, &flags);
5087 * Return sum_exec_runtime for the thread group.
5088 * In case the task is currently running, return the sum plus current's
5089 * pending runtime that have not been accounted yet.
5091 * Note that the thread group might have other running tasks as well,
5092 * so the return value not includes other pending runtime that other
5093 * running tasks might have.
5095 unsigned long long thread_group_sched_runtime(struct task_struct *p)
5097 struct task_cputime totals;
5098 unsigned long flags;
5102 rq = task_rq_lock(p, &flags);
5103 thread_group_cputime(p, &totals);
5104 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
5105 task_rq_unlock(rq, &flags);
5111 * Account user cpu time to a process.
5112 * @p: the process that the cpu time gets accounted to
5113 * @cputime: the cpu time spent in user space since the last update
5114 * @cputime_scaled: cputime scaled by cpu frequency
5116 void account_user_time(struct task_struct *p, cputime_t cputime,
5117 cputime_t cputime_scaled)
5119 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5122 /* Add user time to process. */
5123 p->utime = cputime_add(p->utime, cputime);
5124 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5125 account_group_user_time(p, cputime);
5127 /* Add user time to cpustat. */
5128 tmp = cputime_to_cputime64(cputime);
5129 if (TASK_NICE(p) > 0)
5130 cpustat->nice = cputime64_add(cpustat->nice, tmp);
5132 cpustat->user = cputime64_add(cpustat->user, tmp);
5134 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
5135 /* Account for user time used */
5136 acct_update_integrals(p);
5140 * Account guest cpu time to a process.
5141 * @p: the process that the cpu time gets accounted to
5142 * @cputime: the cpu time spent in virtual machine since the last update
5143 * @cputime_scaled: cputime scaled by cpu frequency
5145 static void account_guest_time(struct task_struct *p, cputime_t cputime,
5146 cputime_t cputime_scaled)
5149 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5151 tmp = cputime_to_cputime64(cputime);
5153 /* Add guest time to process. */
5154 p->utime = cputime_add(p->utime, cputime);
5155 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5156 account_group_user_time(p, cputime);
5157 p->gtime = cputime_add(p->gtime, cputime);
5159 /* Add guest time to cpustat. */
5160 cpustat->user = cputime64_add(cpustat->user, tmp);
5161 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5165 * Account system cpu time to a process.
5166 * @p: the process that the cpu time gets accounted to
5167 * @hardirq_offset: the offset to subtract from hardirq_count()
5168 * @cputime: the cpu time spent in kernel space since the last update
5169 * @cputime_scaled: cputime scaled by cpu frequency
5171 void account_system_time(struct task_struct *p, int hardirq_offset,
5172 cputime_t cputime, cputime_t cputime_scaled)
5174 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5177 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5178 account_guest_time(p, cputime, cputime_scaled);
5182 /* Add system time to process. */
5183 p->stime = cputime_add(p->stime, cputime);
5184 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5185 account_group_system_time(p, cputime);
5187 /* Add system time to cpustat. */
5188 tmp = cputime_to_cputime64(cputime);
5189 if (hardirq_count() - hardirq_offset)
5190 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5191 else if (softirq_count())
5192 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5194 cpustat->system = cputime64_add(cpustat->system, tmp);
5196 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5198 /* Account for system time used */
5199 acct_update_integrals(p);
5203 * Account for involuntary wait time.
5204 * @steal: the cpu time spent in involuntary wait
5206 void account_steal_time(cputime_t cputime)
5208 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5209 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5211 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5215 * Account for idle time.
5216 * @cputime: the cpu time spent in idle wait
5218 void account_idle_time(cputime_t cputime)
5220 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5221 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5222 struct rq *rq = this_rq();
5224 if (atomic_read(&rq->nr_iowait) > 0)
5225 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5227 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5230 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5233 * Account a single tick of cpu time.
5234 * @p: the process that the cpu time gets accounted to
5235 * @user_tick: indicates if the tick is a user or a system tick
5237 void account_process_tick(struct task_struct *p, int user_tick)
5239 cputime_t one_jiffy = jiffies_to_cputime(1);
5240 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5241 struct rq *rq = this_rq();
5244 account_user_time(p, one_jiffy, one_jiffy_scaled);
5245 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5246 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5249 account_idle_time(one_jiffy);
5253 * Account multiple ticks of steal time.
5254 * @p: the process from which the cpu time has been stolen
5255 * @ticks: number of stolen ticks
5257 void account_steal_ticks(unsigned long ticks)
5259 account_steal_time(jiffies_to_cputime(ticks));
5263 * Account multiple ticks of idle time.
5264 * @ticks: number of stolen ticks
5266 void account_idle_ticks(unsigned long ticks)
5268 account_idle_time(jiffies_to_cputime(ticks));
5274 * Use precise platform statistics if available:
5276 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5277 cputime_t task_utime(struct task_struct *p)
5282 cputime_t task_stime(struct task_struct *p)
5287 cputime_t task_utime(struct task_struct *p)
5289 clock_t utime = cputime_to_clock_t(p->utime),
5290 total = utime + cputime_to_clock_t(p->stime);
5294 * Use CFS's precise accounting:
5296 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5300 do_div(temp, total);
5302 utime = (clock_t)temp;
5304 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5305 return p->prev_utime;
5308 cputime_t task_stime(struct task_struct *p)
5313 * Use CFS's precise accounting. (we subtract utime from
5314 * the total, to make sure the total observed by userspace
5315 * grows monotonically - apps rely on that):
5317 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5318 cputime_to_clock_t(task_utime(p));
5321 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5323 return p->prev_stime;
5327 inline cputime_t task_gtime(struct task_struct *p)
5333 * This function gets called by the timer code, with HZ frequency.
5334 * We call it with interrupts disabled.
5336 * It also gets called by the fork code, when changing the parent's
5339 void scheduler_tick(void)
5341 int cpu = smp_processor_id();
5342 struct rq *rq = cpu_rq(cpu);
5343 struct task_struct *curr = rq->curr;
5347 spin_lock(&rq->lock);
5348 update_rq_clock(rq);
5349 update_cpu_load(rq);
5350 curr->sched_class->task_tick(rq, curr, 0);
5351 spin_unlock(&rq->lock);
5353 perf_counter_task_tick(curr, cpu);
5356 rq->idle_at_tick = idle_cpu(cpu);
5357 trigger_load_balance(rq, cpu);
5361 notrace unsigned long get_parent_ip(unsigned long addr)
5363 if (in_lock_functions(addr)) {
5364 addr = CALLER_ADDR2;
5365 if (in_lock_functions(addr))
5366 addr = CALLER_ADDR3;
5371 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5372 defined(CONFIG_PREEMPT_TRACER))
5374 void __kprobes add_preempt_count(int val)
5376 #ifdef CONFIG_DEBUG_PREEMPT
5380 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5383 preempt_count() += val;
5384 #ifdef CONFIG_DEBUG_PREEMPT
5386 * Spinlock count overflowing soon?
5388 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5391 if (preempt_count() == val)
5392 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5394 EXPORT_SYMBOL(add_preempt_count);
5396 void __kprobes sub_preempt_count(int val)
5398 #ifdef CONFIG_DEBUG_PREEMPT
5402 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5405 * Is the spinlock portion underflowing?
5407 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5408 !(preempt_count() & PREEMPT_MASK)))
5412 if (preempt_count() == val)
5413 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5414 preempt_count() -= val;
5416 EXPORT_SYMBOL(sub_preempt_count);
5421 * Print scheduling while atomic bug:
5423 static noinline void __schedule_bug(struct task_struct *prev)
5425 struct pt_regs *regs = get_irq_regs();
5427 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5428 prev->comm, prev->pid, preempt_count());
5430 debug_show_held_locks(prev);
5432 if (irqs_disabled())
5433 print_irqtrace_events(prev);
5442 * Various schedule()-time debugging checks and statistics:
5444 static inline void schedule_debug(struct task_struct *prev)
5447 * Test if we are atomic. Since do_exit() needs to call into
5448 * schedule() atomically, we ignore that path for now.
5449 * Otherwise, whine if we are scheduling when we should not be.
5451 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5452 __schedule_bug(prev);
5454 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5456 schedstat_inc(this_rq(), sched_count);
5457 #ifdef CONFIG_SCHEDSTATS
5458 if (unlikely(prev->lock_depth >= 0)) {
5459 schedstat_inc(this_rq(), bkl_count);
5460 schedstat_inc(prev, sched_info.bkl_count);
5465 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5467 if (prev->state == TASK_RUNNING) {
5468 u64 runtime = prev->se.sum_exec_runtime;
5470 runtime -= prev->se.prev_sum_exec_runtime;
5471 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5474 * In order to avoid avg_overlap growing stale when we are
5475 * indeed overlapping and hence not getting put to sleep, grow
5476 * the avg_overlap on preemption.
5478 * We use the average preemption runtime because that
5479 * correlates to the amount of cache footprint a task can
5482 update_avg(&prev->se.avg_overlap, runtime);
5484 prev->sched_class->put_prev_task(rq, prev);
5488 * Pick up the highest-prio task:
5490 static inline struct task_struct *
5491 pick_next_task(struct rq *rq)
5493 const struct sched_class *class;
5494 struct task_struct *p;
5497 * Optimization: we know that if all tasks are in
5498 * the fair class we can call that function directly:
5500 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5501 p = fair_sched_class.pick_next_task(rq);
5506 class = sched_class_highest;
5508 p = class->pick_next_task(rq);
5512 * Will never be NULL as the idle class always
5513 * returns a non-NULL p:
5515 class = class->next;
5520 * schedule() is the main scheduler function.
5522 asmlinkage void __sched schedule(void)
5524 struct task_struct *prev, *next;
5525 unsigned long *switch_count;
5531 cpu = smp_processor_id();
5535 switch_count = &prev->nivcsw;
5537 release_kernel_lock(prev);
5538 need_resched_nonpreemptible:
5540 schedule_debug(prev);
5542 if (sched_feat(HRTICK))
5545 spin_lock_irq(&rq->lock);
5546 update_rq_clock(rq);
5547 clear_tsk_need_resched(prev);
5549 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5550 if (unlikely(signal_pending_state(prev->state, prev)))
5551 prev->state = TASK_RUNNING;
5553 deactivate_task(rq, prev, 1);
5554 switch_count = &prev->nvcsw;
5557 pre_schedule(rq, prev);
5559 if (unlikely(!rq->nr_running))
5560 idle_balance(cpu, rq);
5562 put_prev_task(rq, prev);
5563 next = pick_next_task(rq);
5565 if (likely(prev != next)) {
5566 sched_info_switch(prev, next);
5567 perf_counter_task_sched_out(prev, next, cpu);
5573 context_switch(rq, prev, next); /* unlocks the rq */
5575 * the context switch might have flipped the stack from under
5576 * us, hence refresh the local variables.
5578 cpu = smp_processor_id();
5581 spin_unlock_irq(&rq->lock);
5585 if (unlikely(reacquire_kernel_lock(current) < 0))
5586 goto need_resched_nonpreemptible;
5588 preempt_enable_no_resched();
5592 EXPORT_SYMBOL(schedule);
5596 * Look out! "owner" is an entirely speculative pointer
5597 * access and not reliable.
5599 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5604 if (!sched_feat(OWNER_SPIN))
5607 #ifdef CONFIG_DEBUG_PAGEALLOC
5609 * Need to access the cpu field knowing that
5610 * DEBUG_PAGEALLOC could have unmapped it if
5611 * the mutex owner just released it and exited.
5613 if (probe_kernel_address(&owner->cpu, cpu))
5620 * Even if the access succeeded (likely case),
5621 * the cpu field may no longer be valid.
5623 if (cpu >= nr_cpumask_bits)
5627 * We need to validate that we can do a
5628 * get_cpu() and that we have the percpu area.
5630 if (!cpu_online(cpu))
5637 * Owner changed, break to re-assess state.
5639 if (lock->owner != owner)
5643 * Is that owner really running on that cpu?
5645 if (task_thread_info(rq->curr) != owner || need_resched())
5655 #ifdef CONFIG_PREEMPT
5657 * this is the entry point to schedule() from in-kernel preemption
5658 * off of preempt_enable. Kernel preemptions off return from interrupt
5659 * occur there and call schedule directly.
5661 asmlinkage void __sched preempt_schedule(void)
5663 struct thread_info *ti = current_thread_info();
5666 * If there is a non-zero preempt_count or interrupts are disabled,
5667 * we do not want to preempt the current task. Just return..
5669 if (likely(ti->preempt_count || irqs_disabled()))
5673 add_preempt_count(PREEMPT_ACTIVE);
5675 sub_preempt_count(PREEMPT_ACTIVE);
5678 * Check again in case we missed a preemption opportunity
5679 * between schedule and now.
5682 } while (need_resched());
5684 EXPORT_SYMBOL(preempt_schedule);
5687 * this is the entry point to schedule() from kernel preemption
5688 * off of irq context.
5689 * Note, that this is called and return with irqs disabled. This will
5690 * protect us against recursive calling from irq.
5692 asmlinkage void __sched preempt_schedule_irq(void)
5694 struct thread_info *ti = current_thread_info();
5696 /* Catch callers which need to be fixed */
5697 BUG_ON(ti->preempt_count || !irqs_disabled());
5700 add_preempt_count(PREEMPT_ACTIVE);
5703 local_irq_disable();
5704 sub_preempt_count(PREEMPT_ACTIVE);
5707 * Check again in case we missed a preemption opportunity
5708 * between schedule and now.
5711 } while (need_resched());
5714 #endif /* CONFIG_PREEMPT */
5716 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5719 return try_to_wake_up(curr->private, mode, sync);
5721 EXPORT_SYMBOL(default_wake_function);
5724 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5725 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5726 * number) then we wake all the non-exclusive tasks and one exclusive task.
5728 * There are circumstances in which we can try to wake a task which has already
5729 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5730 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5732 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5733 int nr_exclusive, int sync, void *key)
5735 wait_queue_t *curr, *next;
5737 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5738 unsigned flags = curr->flags;
5740 if (curr->func(curr, mode, sync, key) &&
5741 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5747 * __wake_up - wake up threads blocked on a waitqueue.
5749 * @mode: which threads
5750 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5751 * @key: is directly passed to the wakeup function
5753 * It may be assumed that this function implies a write memory barrier before
5754 * changing the task state if and only if any tasks are woken up.
5756 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5757 int nr_exclusive, void *key)
5759 unsigned long flags;
5761 spin_lock_irqsave(&q->lock, flags);
5762 __wake_up_common(q, mode, nr_exclusive, 0, key);
5763 spin_unlock_irqrestore(&q->lock, flags);
5765 EXPORT_SYMBOL(__wake_up);
5768 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5770 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5772 __wake_up_common(q, mode, 1, 0, NULL);
5775 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5777 __wake_up_common(q, mode, 1, 0, key);
5781 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5783 * @mode: which threads
5784 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5785 * @key: opaque value to be passed to wakeup targets
5787 * The sync wakeup differs that the waker knows that it will schedule
5788 * away soon, so while the target thread will be woken up, it will not
5789 * be migrated to another CPU - ie. the two threads are 'synchronized'
5790 * with each other. This can prevent needless bouncing between CPUs.
5792 * On UP it can prevent extra preemption.
5794 * It may be assumed that this function implies a write memory barrier before
5795 * changing the task state if and only if any tasks are woken up.
5797 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5798 int nr_exclusive, void *key)
5800 unsigned long flags;
5806 if (unlikely(!nr_exclusive))
5809 spin_lock_irqsave(&q->lock, flags);
5810 __wake_up_common(q, mode, nr_exclusive, sync, key);
5811 spin_unlock_irqrestore(&q->lock, flags);
5813 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5816 * __wake_up_sync - see __wake_up_sync_key()
5818 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5820 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5822 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5825 * complete: - signals a single thread waiting on this completion
5826 * @x: holds the state of this particular completion
5828 * This will wake up a single thread waiting on this completion. Threads will be
5829 * awakened in the same order in which they were queued.
5831 * See also complete_all(), wait_for_completion() and related routines.
5833 * It may be assumed that this function implies a write memory barrier before
5834 * changing the task state if and only if any tasks are woken up.
5836 void complete(struct completion *x)
5838 unsigned long flags;
5840 spin_lock_irqsave(&x->wait.lock, flags);
5842 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5843 spin_unlock_irqrestore(&x->wait.lock, flags);
5845 EXPORT_SYMBOL(complete);
5848 * complete_all: - signals all threads waiting on this completion
5849 * @x: holds the state of this particular completion
5851 * This will wake up all threads waiting on this particular completion event.
5853 * It may be assumed that this function implies a write memory barrier before
5854 * changing the task state if and only if any tasks are woken up.
5856 void complete_all(struct completion *x)
5858 unsigned long flags;
5860 spin_lock_irqsave(&x->wait.lock, flags);
5861 x->done += UINT_MAX/2;
5862 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5863 spin_unlock_irqrestore(&x->wait.lock, flags);
5865 EXPORT_SYMBOL(complete_all);
5867 static inline long __sched
5868 do_wait_for_common(struct completion *x, long timeout, int state)
5871 DECLARE_WAITQUEUE(wait, current);
5873 wait.flags |= WQ_FLAG_EXCLUSIVE;
5874 __add_wait_queue_tail(&x->wait, &wait);
5876 if (signal_pending_state(state, current)) {
5877 timeout = -ERESTARTSYS;
5880 __set_current_state(state);
5881 spin_unlock_irq(&x->wait.lock);
5882 timeout = schedule_timeout(timeout);
5883 spin_lock_irq(&x->wait.lock);
5884 } while (!x->done && timeout);
5885 __remove_wait_queue(&x->wait, &wait);
5890 return timeout ?: 1;
5894 wait_for_common(struct completion *x, long timeout, int state)
5898 spin_lock_irq(&x->wait.lock);
5899 timeout = do_wait_for_common(x, timeout, state);
5900 spin_unlock_irq(&x->wait.lock);
5905 * wait_for_completion: - waits for completion of a task
5906 * @x: holds the state of this particular completion
5908 * This waits to be signaled for completion of a specific task. It is NOT
5909 * interruptible and there is no timeout.
5911 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5912 * and interrupt capability. Also see complete().
5914 void __sched wait_for_completion(struct completion *x)
5916 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5918 EXPORT_SYMBOL(wait_for_completion);
5921 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5922 * @x: holds the state of this particular completion
5923 * @timeout: timeout value in jiffies
5925 * This waits for either a completion of a specific task to be signaled or for a
5926 * specified timeout to expire. The timeout is in jiffies. It is not
5929 unsigned long __sched
5930 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5932 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5934 EXPORT_SYMBOL(wait_for_completion_timeout);
5937 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5938 * @x: holds the state of this particular completion
5940 * This waits for completion of a specific task to be signaled. It is
5943 int __sched wait_for_completion_interruptible(struct completion *x)
5945 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5946 if (t == -ERESTARTSYS)
5950 EXPORT_SYMBOL(wait_for_completion_interruptible);
5953 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5954 * @x: holds the state of this particular completion
5955 * @timeout: timeout value in jiffies
5957 * This waits for either a completion of a specific task to be signaled or for a
5958 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5960 unsigned long __sched
5961 wait_for_completion_interruptible_timeout(struct completion *x,
5962 unsigned long timeout)
5964 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5966 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5969 * wait_for_completion_killable: - waits for completion of a task (killable)
5970 * @x: holds the state of this particular completion
5972 * This waits to be signaled for completion of a specific task. It can be
5973 * interrupted by a kill signal.
5975 int __sched wait_for_completion_killable(struct completion *x)
5977 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5978 if (t == -ERESTARTSYS)
5982 EXPORT_SYMBOL(wait_for_completion_killable);
5985 * try_wait_for_completion - try to decrement a completion without blocking
5986 * @x: completion structure
5988 * Returns: 0 if a decrement cannot be done without blocking
5989 * 1 if a decrement succeeded.
5991 * If a completion is being used as a counting completion,
5992 * attempt to decrement the counter without blocking. This
5993 * enables us to avoid waiting if the resource the completion
5994 * is protecting is not available.
5996 bool try_wait_for_completion(struct completion *x)
6000 spin_lock_irq(&x->wait.lock);
6005 spin_unlock_irq(&x->wait.lock);
6008 EXPORT_SYMBOL(try_wait_for_completion);
6011 * completion_done - Test to see if a completion has any waiters
6012 * @x: completion structure
6014 * Returns: 0 if there are waiters (wait_for_completion() in progress)
6015 * 1 if there are no waiters.
6018 bool completion_done(struct completion *x)
6022 spin_lock_irq(&x->wait.lock);
6025 spin_unlock_irq(&x->wait.lock);
6028 EXPORT_SYMBOL(completion_done);
6031 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
6033 unsigned long flags;
6036 init_waitqueue_entry(&wait, current);
6038 __set_current_state(state);
6040 spin_lock_irqsave(&q->lock, flags);
6041 __add_wait_queue(q, &wait);
6042 spin_unlock(&q->lock);
6043 timeout = schedule_timeout(timeout);
6044 spin_lock_irq(&q->lock);
6045 __remove_wait_queue(q, &wait);
6046 spin_unlock_irqrestore(&q->lock, flags);
6051 void __sched interruptible_sleep_on(wait_queue_head_t *q)
6053 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6055 EXPORT_SYMBOL(interruptible_sleep_on);
6058 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
6060 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
6062 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
6064 void __sched sleep_on(wait_queue_head_t *q)
6066 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
6068 EXPORT_SYMBOL(sleep_on);
6070 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
6072 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
6074 EXPORT_SYMBOL(sleep_on_timeout);
6076 #ifdef CONFIG_RT_MUTEXES
6079 * rt_mutex_setprio - set the current priority of a task
6081 * @prio: prio value (kernel-internal form)
6083 * This function changes the 'effective' priority of a task. It does
6084 * not touch ->normal_prio like __setscheduler().
6086 * Used by the rt_mutex code to implement priority inheritance logic.
6088 void rt_mutex_setprio(struct task_struct *p, int prio)
6090 unsigned long flags;
6091 int oldprio, on_rq, running;
6093 const struct sched_class *prev_class = p->sched_class;
6095 BUG_ON(prio < 0 || prio > MAX_PRIO);
6097 rq = task_rq_lock(p, &flags);
6098 update_rq_clock(rq);
6101 on_rq = p->se.on_rq;
6102 running = task_current(rq, p);
6104 dequeue_task(rq, p, 0);
6106 p->sched_class->put_prev_task(rq, p);
6109 p->sched_class = &rt_sched_class;
6111 p->sched_class = &fair_sched_class;
6116 p->sched_class->set_curr_task(rq);
6118 enqueue_task(rq, p, 0);
6120 check_class_changed(rq, p, prev_class, oldprio, running);
6122 task_rq_unlock(rq, &flags);
6127 void set_user_nice(struct task_struct *p, long nice)
6129 int old_prio, delta, on_rq;
6130 unsigned long flags;
6133 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
6136 * We have to be careful, if called from sys_setpriority(),
6137 * the task might be in the middle of scheduling on another CPU.
6139 rq = task_rq_lock(p, &flags);
6140 update_rq_clock(rq);
6142 * The RT priorities are set via sched_setscheduler(), but we still
6143 * allow the 'normal' nice value to be set - but as expected
6144 * it wont have any effect on scheduling until the task is
6145 * SCHED_FIFO/SCHED_RR:
6147 if (task_has_rt_policy(p)) {
6148 p->static_prio = NICE_TO_PRIO(nice);
6151 on_rq = p->se.on_rq;
6153 dequeue_task(rq, p, 0);
6155 p->static_prio = NICE_TO_PRIO(nice);
6158 p->prio = effective_prio(p);
6159 delta = p->prio - old_prio;
6162 enqueue_task(rq, p, 0);
6164 * If the task increased its priority or is running and
6165 * lowered its priority, then reschedule its CPU:
6167 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6168 resched_task(rq->curr);
6171 task_rq_unlock(rq, &flags);
6173 EXPORT_SYMBOL(set_user_nice);
6176 * can_nice - check if a task can reduce its nice value
6180 int can_nice(const struct task_struct *p, const int nice)
6182 /* convert nice value [19,-20] to rlimit style value [1,40] */
6183 int nice_rlim = 20 - nice;
6185 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6186 capable(CAP_SYS_NICE));
6189 #ifdef __ARCH_WANT_SYS_NICE
6192 * sys_nice - change the priority of the current process.
6193 * @increment: priority increment
6195 * sys_setpriority is a more generic, but much slower function that
6196 * does similar things.
6198 SYSCALL_DEFINE1(nice, int, increment)
6203 * Setpriority might change our priority at the same moment.
6204 * We don't have to worry. Conceptually one call occurs first
6205 * and we have a single winner.
6207 if (increment < -40)
6212 nice = TASK_NICE(current) + increment;
6218 if (increment < 0 && !can_nice(current, nice))
6221 retval = security_task_setnice(current, nice);
6225 set_user_nice(current, nice);
6232 * task_prio - return the priority value of a given task.
6233 * @p: the task in question.
6235 * This is the priority value as seen by users in /proc.
6236 * RT tasks are offset by -200. Normal tasks are centered
6237 * around 0, value goes from -16 to +15.
6239 int task_prio(const struct task_struct *p)
6241 return p->prio - MAX_RT_PRIO;
6245 * task_nice - return the nice value of a given task.
6246 * @p: the task in question.
6248 int task_nice(const struct task_struct *p)
6250 return TASK_NICE(p);
6252 EXPORT_SYMBOL(task_nice);
6255 * idle_cpu - is a given cpu idle currently?
6256 * @cpu: the processor in question.
6258 int idle_cpu(int cpu)
6260 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6264 * idle_task - return the idle task for a given cpu.
6265 * @cpu: the processor in question.
6267 struct task_struct *idle_task(int cpu)
6269 return cpu_rq(cpu)->idle;
6273 * find_process_by_pid - find a process with a matching PID value.
6274 * @pid: the pid in question.
6276 static struct task_struct *find_process_by_pid(pid_t pid)
6278 return pid ? find_task_by_vpid(pid) : current;
6281 /* Actually do priority change: must hold rq lock. */
6283 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6285 BUG_ON(p->se.on_rq);
6288 switch (p->policy) {
6292 p->sched_class = &fair_sched_class;
6296 p->sched_class = &rt_sched_class;
6300 p->rt_priority = prio;
6301 p->normal_prio = normal_prio(p);
6302 /* we are holding p->pi_lock already */
6303 p->prio = rt_mutex_getprio(p);
6308 * check the target process has a UID that matches the current process's
6310 static bool check_same_owner(struct task_struct *p)
6312 const struct cred *cred = current_cred(), *pcred;
6316 pcred = __task_cred(p);
6317 match = (cred->euid == pcred->euid ||
6318 cred->euid == pcred->uid);
6323 static int __sched_setscheduler(struct task_struct *p, int policy,
6324 struct sched_param *param, bool user)
6326 int retval, oldprio, oldpolicy = -1, on_rq, running;
6327 unsigned long flags;
6328 const struct sched_class *prev_class = p->sched_class;
6332 /* may grab non-irq protected spin_locks */
6333 BUG_ON(in_interrupt());
6335 /* double check policy once rq lock held */
6337 reset_on_fork = p->sched_reset_on_fork;
6338 policy = oldpolicy = p->policy;
6340 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6341 policy &= ~SCHED_RESET_ON_FORK;
6343 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6344 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6345 policy != SCHED_IDLE)
6350 * Valid priorities for SCHED_FIFO and SCHED_RR are
6351 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6352 * SCHED_BATCH and SCHED_IDLE is 0.
6354 if (param->sched_priority < 0 ||
6355 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6356 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6358 if (rt_policy(policy) != (param->sched_priority != 0))
6362 * Allow unprivileged RT tasks to decrease priority:
6364 if (user && !capable(CAP_SYS_NICE)) {
6365 if (rt_policy(policy)) {
6366 unsigned long rlim_rtprio;
6368 if (!lock_task_sighand(p, &flags))
6370 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6371 unlock_task_sighand(p, &flags);
6373 /* can't set/change the rt policy */
6374 if (policy != p->policy && !rlim_rtprio)
6377 /* can't increase priority */
6378 if (param->sched_priority > p->rt_priority &&
6379 param->sched_priority > rlim_rtprio)
6383 * Like positive nice levels, dont allow tasks to
6384 * move out of SCHED_IDLE either:
6386 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6389 /* can't change other user's priorities */
6390 if (!check_same_owner(p))
6393 /* Normal users shall not reset the sched_reset_on_fork flag */
6394 if (p->sched_reset_on_fork && !reset_on_fork)
6399 #ifdef CONFIG_RT_GROUP_SCHED
6401 * Do not allow realtime tasks into groups that have no runtime
6404 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6405 task_group(p)->rt_bandwidth.rt_runtime == 0)
6409 retval = security_task_setscheduler(p, policy, param);
6415 * make sure no PI-waiters arrive (or leave) while we are
6416 * changing the priority of the task:
6418 spin_lock_irqsave(&p->pi_lock, flags);
6420 * To be able to change p->policy safely, the apropriate
6421 * runqueue lock must be held.
6423 rq = __task_rq_lock(p);
6424 /* recheck policy now with rq lock held */
6425 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6426 policy = oldpolicy = -1;
6427 __task_rq_unlock(rq);
6428 spin_unlock_irqrestore(&p->pi_lock, flags);
6431 update_rq_clock(rq);
6432 on_rq = p->se.on_rq;
6433 running = task_current(rq, p);
6435 deactivate_task(rq, p, 0);
6437 p->sched_class->put_prev_task(rq, p);
6439 p->sched_reset_on_fork = reset_on_fork;
6442 __setscheduler(rq, p, policy, param->sched_priority);
6445 p->sched_class->set_curr_task(rq);
6447 activate_task(rq, p, 0);
6449 check_class_changed(rq, p, prev_class, oldprio, running);
6451 __task_rq_unlock(rq);
6452 spin_unlock_irqrestore(&p->pi_lock, flags);
6454 rt_mutex_adjust_pi(p);
6460 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6461 * @p: the task in question.
6462 * @policy: new policy.
6463 * @param: structure containing the new RT priority.
6465 * NOTE that the task may be already dead.
6467 int sched_setscheduler(struct task_struct *p, int policy,
6468 struct sched_param *param)
6470 return __sched_setscheduler(p, policy, param, true);
6472 EXPORT_SYMBOL_GPL(sched_setscheduler);
6475 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6476 * @p: the task in question.
6477 * @policy: new policy.
6478 * @param: structure containing the new RT priority.
6480 * Just like sched_setscheduler, only don't bother checking if the
6481 * current context has permission. For example, this is needed in
6482 * stop_machine(): we create temporary high priority worker threads,
6483 * but our caller might not have that capability.
6485 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6486 struct sched_param *param)
6488 return __sched_setscheduler(p, policy, param, false);
6492 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6494 struct sched_param lparam;
6495 struct task_struct *p;
6498 if (!param || pid < 0)
6500 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6505 p = find_process_by_pid(pid);
6507 retval = sched_setscheduler(p, policy, &lparam);
6514 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6515 * @pid: the pid in question.
6516 * @policy: new policy.
6517 * @param: structure containing the new RT priority.
6519 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6520 struct sched_param __user *, param)
6522 /* negative values for policy are not valid */
6526 return do_sched_setscheduler(pid, policy, param);
6530 * sys_sched_setparam - set/change the RT priority of a thread
6531 * @pid: the pid in question.
6532 * @param: structure containing the new RT priority.
6534 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6536 return do_sched_setscheduler(pid, -1, param);
6540 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6541 * @pid: the pid in question.
6543 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6545 struct task_struct *p;
6552 read_lock(&tasklist_lock);
6553 p = find_process_by_pid(pid);
6555 retval = security_task_getscheduler(p);
6558 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6560 read_unlock(&tasklist_lock);
6565 * sys_sched_getparam - get the RT priority of a thread
6566 * @pid: the pid in question.
6567 * @param: structure containing the RT priority.
6569 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6571 struct sched_param lp;
6572 struct task_struct *p;
6575 if (!param || pid < 0)
6578 read_lock(&tasklist_lock);
6579 p = find_process_by_pid(pid);
6584 retval = security_task_getscheduler(p);
6588 lp.sched_priority = p->rt_priority;
6589 read_unlock(&tasklist_lock);
6592 * This one might sleep, we cannot do it with a spinlock held ...
6594 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6599 read_unlock(&tasklist_lock);
6603 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6605 cpumask_var_t cpus_allowed, new_mask;
6606 struct task_struct *p;
6610 read_lock(&tasklist_lock);
6612 p = find_process_by_pid(pid);
6614 read_unlock(&tasklist_lock);
6620 * It is not safe to call set_cpus_allowed with the
6621 * tasklist_lock held. We will bump the task_struct's
6622 * usage count and then drop tasklist_lock.
6625 read_unlock(&tasklist_lock);
6627 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6631 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6633 goto out_free_cpus_allowed;
6636 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6639 retval = security_task_setscheduler(p, 0, NULL);
6643 cpuset_cpus_allowed(p, cpus_allowed);
6644 cpumask_and(new_mask, in_mask, cpus_allowed);
6646 retval = set_cpus_allowed_ptr(p, new_mask);
6649 cpuset_cpus_allowed(p, cpus_allowed);
6650 if (!cpumask_subset(new_mask, cpus_allowed)) {
6652 * We must have raced with a concurrent cpuset
6653 * update. Just reset the cpus_allowed to the
6654 * cpuset's cpus_allowed
6656 cpumask_copy(new_mask, cpus_allowed);
6661 free_cpumask_var(new_mask);
6662 out_free_cpus_allowed:
6663 free_cpumask_var(cpus_allowed);
6670 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6671 struct cpumask *new_mask)
6673 if (len < cpumask_size())
6674 cpumask_clear(new_mask);
6675 else if (len > cpumask_size())
6676 len = cpumask_size();
6678 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6682 * sys_sched_setaffinity - set the cpu affinity of a process
6683 * @pid: pid of the process
6684 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6685 * @user_mask_ptr: user-space pointer to the new cpu mask
6687 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6688 unsigned long __user *, user_mask_ptr)
6690 cpumask_var_t new_mask;
6693 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6696 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6698 retval = sched_setaffinity(pid, new_mask);
6699 free_cpumask_var(new_mask);
6703 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6705 struct task_struct *p;
6709 read_lock(&tasklist_lock);
6712 p = find_process_by_pid(pid);
6716 retval = security_task_getscheduler(p);
6720 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6723 read_unlock(&tasklist_lock);
6730 * sys_sched_getaffinity - get the cpu affinity of a process
6731 * @pid: pid of the process
6732 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6733 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6735 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6736 unsigned long __user *, user_mask_ptr)
6741 if (len < cpumask_size())
6744 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6747 ret = sched_getaffinity(pid, mask);
6749 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6752 ret = cpumask_size();
6754 free_cpumask_var(mask);
6760 * sys_sched_yield - yield the current processor to other threads.
6762 * This function yields the current CPU to other tasks. If there are no
6763 * other threads running on this CPU then this function will return.
6765 SYSCALL_DEFINE0(sched_yield)
6767 struct rq *rq = this_rq_lock();
6769 schedstat_inc(rq, yld_count);
6770 current->sched_class->yield_task(rq);
6773 * Since we are going to call schedule() anyway, there's
6774 * no need to preempt or enable interrupts:
6776 __release(rq->lock);
6777 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6778 _raw_spin_unlock(&rq->lock);
6779 preempt_enable_no_resched();
6786 static inline int should_resched(void)
6788 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6791 static void __cond_resched(void)
6793 add_preempt_count(PREEMPT_ACTIVE);
6795 sub_preempt_count(PREEMPT_ACTIVE);
6798 int __sched _cond_resched(void)
6800 if (should_resched()) {
6806 EXPORT_SYMBOL(_cond_resched);
6809 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6810 * call schedule, and on return reacquire the lock.
6812 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6813 * operations here to prevent schedule() from being called twice (once via
6814 * spin_unlock(), once by hand).
6816 int __cond_resched_lock(spinlock_t *lock)
6818 int resched = should_resched();
6821 if (spin_needbreak(lock) || resched) {
6832 EXPORT_SYMBOL(__cond_resched_lock);
6834 int __sched __cond_resched_softirq(void)
6836 BUG_ON(!in_softirq());
6838 if (should_resched()) {
6846 EXPORT_SYMBOL(__cond_resched_softirq);
6849 * yield - yield the current processor to other threads.
6851 * This is a shortcut for kernel-space yielding - it marks the
6852 * thread runnable and calls sys_sched_yield().
6854 void __sched yield(void)
6856 set_current_state(TASK_RUNNING);
6859 EXPORT_SYMBOL(yield);
6862 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6863 * that process accounting knows that this is a task in IO wait state.
6865 * But don't do that if it is a deliberate, throttling IO wait (this task
6866 * has set its backing_dev_info: the queue against which it should throttle)
6868 void __sched io_schedule(void)
6870 struct rq *rq = raw_rq();
6872 delayacct_blkio_start();
6873 atomic_inc(&rq->nr_iowait);
6874 current->in_iowait = 1;
6876 current->in_iowait = 0;
6877 atomic_dec(&rq->nr_iowait);
6878 delayacct_blkio_end();
6880 EXPORT_SYMBOL(io_schedule);
6882 long __sched io_schedule_timeout(long timeout)
6884 struct rq *rq = raw_rq();
6887 delayacct_blkio_start();
6888 atomic_inc(&rq->nr_iowait);
6889 current->in_iowait = 1;
6890 ret = schedule_timeout(timeout);
6891 current->in_iowait = 0;
6892 atomic_dec(&rq->nr_iowait);
6893 delayacct_blkio_end();
6898 * sys_sched_get_priority_max - return maximum RT priority.
6899 * @policy: scheduling class.
6901 * this syscall returns the maximum rt_priority that can be used
6902 * by a given scheduling class.
6904 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6911 ret = MAX_USER_RT_PRIO-1;
6923 * sys_sched_get_priority_min - return minimum RT priority.
6924 * @policy: scheduling class.
6926 * this syscall returns the minimum rt_priority that can be used
6927 * by a given scheduling class.
6929 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6947 * sys_sched_rr_get_interval - return the default timeslice of a process.
6948 * @pid: pid of the process.
6949 * @interval: userspace pointer to the timeslice value.
6951 * this syscall writes the default timeslice value of a given process
6952 * into the user-space timespec buffer. A value of '0' means infinity.
6954 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6955 struct timespec __user *, interval)
6957 struct task_struct *p;
6958 unsigned int time_slice;
6966 read_lock(&tasklist_lock);
6967 p = find_process_by_pid(pid);
6971 retval = security_task_getscheduler(p);
6976 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6977 * tasks that are on an otherwise idle runqueue:
6980 if (p->policy == SCHED_RR) {
6981 time_slice = DEF_TIMESLICE;
6982 } else if (p->policy != SCHED_FIFO) {
6983 struct sched_entity *se = &p->se;
6984 unsigned long flags;
6987 rq = task_rq_lock(p, &flags);
6988 if (rq->cfs.load.weight)
6989 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6990 task_rq_unlock(rq, &flags);
6992 read_unlock(&tasklist_lock);
6993 jiffies_to_timespec(time_slice, &t);
6994 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6998 read_unlock(&tasklist_lock);
7002 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
7004 void sched_show_task(struct task_struct *p)
7006 unsigned long free = 0;
7009 state = p->state ? __ffs(p->state) + 1 : 0;
7010 printk(KERN_INFO "%-13.13s %c", p->comm,
7011 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
7012 #if BITS_PER_LONG == 32
7013 if (state == TASK_RUNNING)
7014 printk(KERN_CONT " running ");
7016 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
7018 if (state == TASK_RUNNING)
7019 printk(KERN_CONT " running task ");
7021 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
7023 #ifdef CONFIG_DEBUG_STACK_USAGE
7024 free = stack_not_used(p);
7026 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
7027 task_pid_nr(p), task_pid_nr(p->real_parent),
7028 (unsigned long)task_thread_info(p)->flags);
7030 show_stack(p, NULL);
7033 void show_state_filter(unsigned long state_filter)
7035 struct task_struct *g, *p;
7037 #if BITS_PER_LONG == 32
7039 " task PC stack pid father\n");
7042 " task PC stack pid father\n");
7044 read_lock(&tasklist_lock);
7045 do_each_thread(g, p) {
7047 * reset the NMI-timeout, listing all files on a slow
7048 * console might take alot of time:
7050 touch_nmi_watchdog();
7051 if (!state_filter || (p->state & state_filter))
7053 } while_each_thread(g, p);
7055 touch_all_softlockup_watchdogs();
7057 #ifdef CONFIG_SCHED_DEBUG
7058 sysrq_sched_debug_show();
7060 read_unlock(&tasklist_lock);
7062 * Only show locks if all tasks are dumped:
7064 if (state_filter == -1)
7065 debug_show_all_locks();
7068 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
7070 idle->sched_class = &idle_sched_class;
7074 * init_idle - set up an idle thread for a given CPU
7075 * @idle: task in question
7076 * @cpu: cpu the idle task belongs to
7078 * NOTE: this function does not set the idle thread's NEED_RESCHED
7079 * flag, to make booting more robust.
7081 void __cpuinit init_idle(struct task_struct *idle, int cpu)
7083 struct rq *rq = cpu_rq(cpu);
7084 unsigned long flags;
7086 spin_lock_irqsave(&rq->lock, flags);
7089 idle->se.exec_start = sched_clock();
7091 idle->prio = idle->normal_prio = MAX_PRIO;
7092 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
7093 __set_task_cpu(idle, cpu);
7095 rq->curr = rq->idle = idle;
7096 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7099 spin_unlock_irqrestore(&rq->lock, flags);
7101 /* Set the preempt count _outside_ the spinlocks! */
7102 #if defined(CONFIG_PREEMPT)
7103 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
7105 task_thread_info(idle)->preempt_count = 0;
7108 * The idle tasks have their own, simple scheduling class:
7110 idle->sched_class = &idle_sched_class;
7111 ftrace_graph_init_task(idle);
7115 * In a system that switches off the HZ timer nohz_cpu_mask
7116 * indicates which cpus entered this state. This is used
7117 * in the rcu update to wait only for active cpus. For system
7118 * which do not switch off the HZ timer nohz_cpu_mask should
7119 * always be CPU_BITS_NONE.
7121 cpumask_var_t nohz_cpu_mask;
7124 * Increase the granularity value when there are more CPUs,
7125 * because with more CPUs the 'effective latency' as visible
7126 * to users decreases. But the relationship is not linear,
7127 * so pick a second-best guess by going with the log2 of the
7130 * This idea comes from the SD scheduler of Con Kolivas:
7132 static inline void sched_init_granularity(void)
7134 unsigned int factor = 1 + ilog2(num_online_cpus());
7135 const unsigned long limit = 200000000;
7137 sysctl_sched_min_granularity *= factor;
7138 if (sysctl_sched_min_granularity > limit)
7139 sysctl_sched_min_granularity = limit;
7141 sysctl_sched_latency *= factor;
7142 if (sysctl_sched_latency > limit)
7143 sysctl_sched_latency = limit;
7145 sysctl_sched_wakeup_granularity *= factor;
7147 sysctl_sched_shares_ratelimit *= factor;
7152 * This is how migration works:
7154 * 1) we queue a struct migration_req structure in the source CPU's
7155 * runqueue and wake up that CPU's migration thread.
7156 * 2) we down() the locked semaphore => thread blocks.
7157 * 3) migration thread wakes up (implicitly it forces the migrated
7158 * thread off the CPU)
7159 * 4) it gets the migration request and checks whether the migrated
7160 * task is still in the wrong runqueue.
7161 * 5) if it's in the wrong runqueue then the migration thread removes
7162 * it and puts it into the right queue.
7163 * 6) migration thread up()s the semaphore.
7164 * 7) we wake up and the migration is done.
7168 * Change a given task's CPU affinity. Migrate the thread to a
7169 * proper CPU and schedule it away if the CPU it's executing on
7170 * is removed from the allowed bitmask.
7172 * NOTE: the caller must have a valid reference to the task, the
7173 * task must not exit() & deallocate itself prematurely. The
7174 * call is not atomic; no spinlocks may be held.
7176 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7178 struct migration_req req;
7179 unsigned long flags;
7183 rq = task_rq_lock(p, &flags);
7184 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7189 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7190 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7195 if (p->sched_class->set_cpus_allowed)
7196 p->sched_class->set_cpus_allowed(p, new_mask);
7198 cpumask_copy(&p->cpus_allowed, new_mask);
7199 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7202 /* Can the task run on the task's current CPU? If so, we're done */
7203 if (cpumask_test_cpu(task_cpu(p), new_mask))
7206 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7207 /* Need help from migration thread: drop lock and wait. */
7208 struct task_struct *mt = rq->migration_thread;
7210 get_task_struct(mt);
7211 task_rq_unlock(rq, &flags);
7212 wake_up_process(rq->migration_thread);
7213 put_task_struct(mt);
7214 wait_for_completion(&req.done);
7215 tlb_migrate_finish(p->mm);
7219 task_rq_unlock(rq, &flags);
7223 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7226 * Move (not current) task off this cpu, onto dest cpu. We're doing
7227 * this because either it can't run here any more (set_cpus_allowed()
7228 * away from this CPU, or CPU going down), or because we're
7229 * attempting to rebalance this task on exec (sched_exec).
7231 * So we race with normal scheduler movements, but that's OK, as long
7232 * as the task is no longer on this CPU.
7234 * Returns non-zero if task was successfully migrated.
7236 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7238 struct rq *rq_dest, *rq_src;
7241 if (unlikely(!cpu_active(dest_cpu)))
7244 rq_src = cpu_rq(src_cpu);
7245 rq_dest = cpu_rq(dest_cpu);
7247 double_rq_lock(rq_src, rq_dest);
7248 /* Already moved. */
7249 if (task_cpu(p) != src_cpu)
7251 /* Affinity changed (again). */
7252 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7255 on_rq = p->se.on_rq;
7257 deactivate_task(rq_src, p, 0);
7259 set_task_cpu(p, dest_cpu);
7261 activate_task(rq_dest, p, 0);
7262 check_preempt_curr(rq_dest, p, 0);
7267 double_rq_unlock(rq_src, rq_dest);
7272 * migration_thread - this is a highprio system thread that performs
7273 * thread migration by bumping thread off CPU then 'pushing' onto
7276 static int migration_thread(void *data)
7278 int cpu = (long)data;
7282 BUG_ON(rq->migration_thread != current);
7284 set_current_state(TASK_INTERRUPTIBLE);
7285 while (!kthread_should_stop()) {
7286 struct migration_req *req;
7287 struct list_head *head;
7289 spin_lock_irq(&rq->lock);
7291 if (cpu_is_offline(cpu)) {
7292 spin_unlock_irq(&rq->lock);
7296 if (rq->active_balance) {
7297 active_load_balance(rq, cpu);
7298 rq->active_balance = 0;
7301 head = &rq->migration_queue;
7303 if (list_empty(head)) {
7304 spin_unlock_irq(&rq->lock);
7306 set_current_state(TASK_INTERRUPTIBLE);
7309 req = list_entry(head->next, struct migration_req, list);
7310 list_del_init(head->next);
7312 spin_unlock(&rq->lock);
7313 __migrate_task(req->task, cpu, req->dest_cpu);
7316 complete(&req->done);
7318 __set_current_state(TASK_RUNNING);
7323 #ifdef CONFIG_HOTPLUG_CPU
7325 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7329 local_irq_disable();
7330 ret = __migrate_task(p, src_cpu, dest_cpu);
7336 * Figure out where task on dead CPU should go, use force if necessary.
7338 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7341 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7344 /* Look for allowed, online CPU in same node. */
7345 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7346 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7349 /* Any allowed, online CPU? */
7350 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7351 if (dest_cpu < nr_cpu_ids)
7354 /* No more Mr. Nice Guy. */
7355 if (dest_cpu >= nr_cpu_ids) {
7356 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7357 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7360 * Don't tell them about moving exiting tasks or
7361 * kernel threads (both mm NULL), since they never
7364 if (p->mm && printk_ratelimit()) {
7365 printk(KERN_INFO "process %d (%s) no "
7366 "longer affine to cpu%d\n",
7367 task_pid_nr(p), p->comm, dead_cpu);
7372 /* It can have affinity changed while we were choosing. */
7373 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7378 * While a dead CPU has no uninterruptible tasks queued at this point,
7379 * it might still have a nonzero ->nr_uninterruptible counter, because
7380 * for performance reasons the counter is not stricly tracking tasks to
7381 * their home CPUs. So we just add the counter to another CPU's counter,
7382 * to keep the global sum constant after CPU-down:
7384 static void migrate_nr_uninterruptible(struct rq *rq_src)
7386 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7387 unsigned long flags;
7389 local_irq_save(flags);
7390 double_rq_lock(rq_src, rq_dest);
7391 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7392 rq_src->nr_uninterruptible = 0;
7393 double_rq_unlock(rq_src, rq_dest);
7394 local_irq_restore(flags);
7397 /* Run through task list and migrate tasks from the dead cpu. */
7398 static void migrate_live_tasks(int src_cpu)
7400 struct task_struct *p, *t;
7402 read_lock(&tasklist_lock);
7404 do_each_thread(t, p) {
7408 if (task_cpu(p) == src_cpu)
7409 move_task_off_dead_cpu(src_cpu, p);
7410 } while_each_thread(t, p);
7412 read_unlock(&tasklist_lock);
7416 * Schedules idle task to be the next runnable task on current CPU.
7417 * It does so by boosting its priority to highest possible.
7418 * Used by CPU offline code.
7420 void sched_idle_next(void)
7422 int this_cpu = smp_processor_id();
7423 struct rq *rq = cpu_rq(this_cpu);
7424 struct task_struct *p = rq->idle;
7425 unsigned long flags;
7427 /* cpu has to be offline */
7428 BUG_ON(cpu_online(this_cpu));
7431 * Strictly not necessary since rest of the CPUs are stopped by now
7432 * and interrupts disabled on the current cpu.
7434 spin_lock_irqsave(&rq->lock, flags);
7436 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7438 update_rq_clock(rq);
7439 activate_task(rq, p, 0);
7441 spin_unlock_irqrestore(&rq->lock, flags);
7445 * Ensures that the idle task is using init_mm right before its cpu goes
7448 void idle_task_exit(void)
7450 struct mm_struct *mm = current->active_mm;
7452 BUG_ON(cpu_online(smp_processor_id()));
7455 switch_mm(mm, &init_mm, current);
7459 /* called under rq->lock with disabled interrupts */
7460 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7462 struct rq *rq = cpu_rq(dead_cpu);
7464 /* Must be exiting, otherwise would be on tasklist. */
7465 BUG_ON(!p->exit_state);
7467 /* Cannot have done final schedule yet: would have vanished. */
7468 BUG_ON(p->state == TASK_DEAD);
7473 * Drop lock around migration; if someone else moves it,
7474 * that's OK. No task can be added to this CPU, so iteration is
7477 spin_unlock_irq(&rq->lock);
7478 move_task_off_dead_cpu(dead_cpu, p);
7479 spin_lock_irq(&rq->lock);
7484 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7485 static void migrate_dead_tasks(unsigned int dead_cpu)
7487 struct rq *rq = cpu_rq(dead_cpu);
7488 struct task_struct *next;
7491 if (!rq->nr_running)
7493 update_rq_clock(rq);
7494 next = pick_next_task(rq);
7497 next->sched_class->put_prev_task(rq, next);
7498 migrate_dead(dead_cpu, next);
7504 * remove the tasks which were accounted by rq from calc_load_tasks.
7506 static void calc_global_load_remove(struct rq *rq)
7508 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7509 rq->calc_load_active = 0;
7511 #endif /* CONFIG_HOTPLUG_CPU */
7513 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7515 static struct ctl_table sd_ctl_dir[] = {
7517 .procname = "sched_domain",
7523 static struct ctl_table sd_ctl_root[] = {
7525 .ctl_name = CTL_KERN,
7526 .procname = "kernel",
7528 .child = sd_ctl_dir,
7533 static struct ctl_table *sd_alloc_ctl_entry(int n)
7535 struct ctl_table *entry =
7536 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7541 static void sd_free_ctl_entry(struct ctl_table **tablep)
7543 struct ctl_table *entry;
7546 * In the intermediate directories, both the child directory and
7547 * procname are dynamically allocated and could fail but the mode
7548 * will always be set. In the lowest directory the names are
7549 * static strings and all have proc handlers.
7551 for (entry = *tablep; entry->mode; entry++) {
7553 sd_free_ctl_entry(&entry->child);
7554 if (entry->proc_handler == NULL)
7555 kfree(entry->procname);
7563 set_table_entry(struct ctl_table *entry,
7564 const char *procname, void *data, int maxlen,
7565 mode_t mode, proc_handler *proc_handler)
7567 entry->procname = procname;
7569 entry->maxlen = maxlen;
7571 entry->proc_handler = proc_handler;
7574 static struct ctl_table *
7575 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7577 struct ctl_table *table = sd_alloc_ctl_entry(13);
7582 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7583 sizeof(long), 0644, proc_doulongvec_minmax);
7584 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7585 sizeof(long), 0644, proc_doulongvec_minmax);
7586 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7587 sizeof(int), 0644, proc_dointvec_minmax);
7588 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7589 sizeof(int), 0644, proc_dointvec_minmax);
7590 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7591 sizeof(int), 0644, proc_dointvec_minmax);
7592 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7593 sizeof(int), 0644, proc_dointvec_minmax);
7594 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7595 sizeof(int), 0644, proc_dointvec_minmax);
7596 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7597 sizeof(int), 0644, proc_dointvec_minmax);
7598 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7599 sizeof(int), 0644, proc_dointvec_minmax);
7600 set_table_entry(&table[9], "cache_nice_tries",
7601 &sd->cache_nice_tries,
7602 sizeof(int), 0644, proc_dointvec_minmax);
7603 set_table_entry(&table[10], "flags", &sd->flags,
7604 sizeof(int), 0644, proc_dointvec_minmax);
7605 set_table_entry(&table[11], "name", sd->name,
7606 CORENAME_MAX_SIZE, 0444, proc_dostring);
7607 /* &table[12] is terminator */
7612 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7614 struct ctl_table *entry, *table;
7615 struct sched_domain *sd;
7616 int domain_num = 0, i;
7619 for_each_domain(cpu, sd)
7621 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7626 for_each_domain(cpu, sd) {
7627 snprintf(buf, 32, "domain%d", i);
7628 entry->procname = kstrdup(buf, GFP_KERNEL);
7630 entry->child = sd_alloc_ctl_domain_table(sd);
7637 static struct ctl_table_header *sd_sysctl_header;
7638 static void register_sched_domain_sysctl(void)
7640 int i, cpu_num = num_online_cpus();
7641 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7644 WARN_ON(sd_ctl_dir[0].child);
7645 sd_ctl_dir[0].child = entry;
7650 for_each_online_cpu(i) {
7651 snprintf(buf, 32, "cpu%d", i);
7652 entry->procname = kstrdup(buf, GFP_KERNEL);
7654 entry->child = sd_alloc_ctl_cpu_table(i);
7658 WARN_ON(sd_sysctl_header);
7659 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7662 /* may be called multiple times per register */
7663 static void unregister_sched_domain_sysctl(void)
7665 if (sd_sysctl_header)
7666 unregister_sysctl_table(sd_sysctl_header);
7667 sd_sysctl_header = NULL;
7668 if (sd_ctl_dir[0].child)
7669 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7672 static void register_sched_domain_sysctl(void)
7675 static void unregister_sched_domain_sysctl(void)
7680 static void set_rq_online(struct rq *rq)
7683 const struct sched_class *class;
7685 cpumask_set_cpu(rq->cpu, rq->rd->online);
7688 for_each_class(class) {
7689 if (class->rq_online)
7690 class->rq_online(rq);
7695 static void set_rq_offline(struct rq *rq)
7698 const struct sched_class *class;
7700 for_each_class(class) {
7701 if (class->rq_offline)
7702 class->rq_offline(rq);
7705 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7711 * migration_call - callback that gets triggered when a CPU is added.
7712 * Here we can start up the necessary migration thread for the new CPU.
7714 static int __cpuinit
7715 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7717 struct task_struct *p;
7718 int cpu = (long)hcpu;
7719 unsigned long flags;
7724 case CPU_UP_PREPARE:
7725 case CPU_UP_PREPARE_FROZEN:
7726 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7729 kthread_bind(p, cpu);
7730 /* Must be high prio: stop_machine expects to yield to it. */
7731 rq = task_rq_lock(p, &flags);
7732 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7733 task_rq_unlock(rq, &flags);
7735 cpu_rq(cpu)->migration_thread = p;
7736 rq->calc_load_update = calc_load_update;
7740 case CPU_ONLINE_FROZEN:
7741 /* Strictly unnecessary, as first user will wake it. */
7742 wake_up_process(cpu_rq(cpu)->migration_thread);
7744 /* Update our root-domain */
7746 spin_lock_irqsave(&rq->lock, flags);
7748 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7752 spin_unlock_irqrestore(&rq->lock, flags);
7755 #ifdef CONFIG_HOTPLUG_CPU
7756 case CPU_UP_CANCELED:
7757 case CPU_UP_CANCELED_FROZEN:
7758 if (!cpu_rq(cpu)->migration_thread)
7760 /* Unbind it from offline cpu so it can run. Fall thru. */
7761 kthread_bind(cpu_rq(cpu)->migration_thread,
7762 cpumask_any(cpu_online_mask));
7763 kthread_stop(cpu_rq(cpu)->migration_thread);
7764 put_task_struct(cpu_rq(cpu)->migration_thread);
7765 cpu_rq(cpu)->migration_thread = NULL;
7769 case CPU_DEAD_FROZEN:
7770 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7771 migrate_live_tasks(cpu);
7773 kthread_stop(rq->migration_thread);
7774 put_task_struct(rq->migration_thread);
7775 rq->migration_thread = NULL;
7776 /* Idle task back to normal (off runqueue, low prio) */
7777 spin_lock_irq(&rq->lock);
7778 update_rq_clock(rq);
7779 deactivate_task(rq, rq->idle, 0);
7780 rq->idle->static_prio = MAX_PRIO;
7781 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7782 rq->idle->sched_class = &idle_sched_class;
7783 migrate_dead_tasks(cpu);
7784 spin_unlock_irq(&rq->lock);
7786 migrate_nr_uninterruptible(rq);
7787 BUG_ON(rq->nr_running != 0);
7788 calc_global_load_remove(rq);
7790 * No need to migrate the tasks: it was best-effort if
7791 * they didn't take sched_hotcpu_mutex. Just wake up
7794 spin_lock_irq(&rq->lock);
7795 while (!list_empty(&rq->migration_queue)) {
7796 struct migration_req *req;
7798 req = list_entry(rq->migration_queue.next,
7799 struct migration_req, list);
7800 list_del_init(&req->list);
7801 spin_unlock_irq(&rq->lock);
7802 complete(&req->done);
7803 spin_lock_irq(&rq->lock);
7805 spin_unlock_irq(&rq->lock);
7809 case CPU_DYING_FROZEN:
7810 /* Update our root-domain */
7812 spin_lock_irqsave(&rq->lock, flags);
7814 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7817 spin_unlock_irqrestore(&rq->lock, flags);
7825 * Register at high priority so that task migration (migrate_all_tasks)
7826 * happens before everything else. This has to be lower priority than
7827 * the notifier in the perf_counter subsystem, though.
7829 static struct notifier_block __cpuinitdata migration_notifier = {
7830 .notifier_call = migration_call,
7834 static int __init migration_init(void)
7836 void *cpu = (void *)(long)smp_processor_id();
7839 /* Start one for the boot CPU: */
7840 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7841 BUG_ON(err == NOTIFY_BAD);
7842 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7843 register_cpu_notifier(&migration_notifier);
7847 early_initcall(migration_init);
7852 #ifdef CONFIG_SCHED_DEBUG
7854 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7855 struct cpumask *groupmask)
7857 struct sched_group *group = sd->groups;
7860 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7861 cpumask_clear(groupmask);
7863 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7865 if (!(sd->flags & SD_LOAD_BALANCE)) {
7866 printk("does not load-balance\n");
7868 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7873 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7875 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7876 printk(KERN_ERR "ERROR: domain->span does not contain "
7879 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7880 printk(KERN_ERR "ERROR: domain->groups does not contain"
7884 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7888 printk(KERN_ERR "ERROR: group is NULL\n");
7892 if (!group->cpu_power) {
7893 printk(KERN_CONT "\n");
7894 printk(KERN_ERR "ERROR: domain->cpu_power not "
7899 if (!cpumask_weight(sched_group_cpus(group))) {
7900 printk(KERN_CONT "\n");
7901 printk(KERN_ERR "ERROR: empty group\n");
7905 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7906 printk(KERN_CONT "\n");
7907 printk(KERN_ERR "ERROR: repeated CPUs\n");
7911 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7913 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7915 printk(KERN_CONT " %s", str);
7916 if (group->cpu_power != SCHED_LOAD_SCALE) {
7917 printk(KERN_CONT " (cpu_power = %d)",
7921 group = group->next;
7922 } while (group != sd->groups);
7923 printk(KERN_CONT "\n");
7925 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7926 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7929 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7930 printk(KERN_ERR "ERROR: parent span is not a superset "
7931 "of domain->span\n");
7935 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7937 cpumask_var_t groupmask;
7941 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7945 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7947 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7948 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7953 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7960 free_cpumask_var(groupmask);
7962 #else /* !CONFIG_SCHED_DEBUG */
7963 # define sched_domain_debug(sd, cpu) do { } while (0)
7964 #endif /* CONFIG_SCHED_DEBUG */
7966 static int sd_degenerate(struct sched_domain *sd)
7968 if (cpumask_weight(sched_domain_span(sd)) == 1)
7971 /* Following flags need at least 2 groups */
7972 if (sd->flags & (SD_LOAD_BALANCE |
7973 SD_BALANCE_NEWIDLE |
7977 SD_SHARE_PKG_RESOURCES)) {
7978 if (sd->groups != sd->groups->next)
7982 /* Following flags don't use groups */
7983 if (sd->flags & (SD_WAKE_IDLE |
7992 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7994 unsigned long cflags = sd->flags, pflags = parent->flags;
7996 if (sd_degenerate(parent))
7999 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
8002 /* Does parent contain flags not in child? */
8003 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
8004 if (cflags & SD_WAKE_AFFINE)
8005 pflags &= ~SD_WAKE_BALANCE;
8006 /* Flags needing groups don't count if only 1 group in parent */
8007 if (parent->groups == parent->groups->next) {
8008 pflags &= ~(SD_LOAD_BALANCE |
8009 SD_BALANCE_NEWIDLE |
8013 SD_SHARE_PKG_RESOURCES);
8014 if (nr_node_ids == 1)
8015 pflags &= ~SD_SERIALIZE;
8017 if (~cflags & pflags)
8023 static void free_rootdomain(struct root_domain *rd)
8025 cpupri_cleanup(&rd->cpupri);
8027 free_cpumask_var(rd->rto_mask);
8028 free_cpumask_var(rd->online);
8029 free_cpumask_var(rd->span);
8033 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
8035 struct root_domain *old_rd = NULL;
8036 unsigned long flags;
8038 spin_lock_irqsave(&rq->lock, flags);
8043 if (cpumask_test_cpu(rq->cpu, old_rd->online))
8046 cpumask_clear_cpu(rq->cpu, old_rd->span);
8049 * If we dont want to free the old_rt yet then
8050 * set old_rd to NULL to skip the freeing later
8053 if (!atomic_dec_and_test(&old_rd->refcount))
8057 atomic_inc(&rd->refcount);
8060 cpumask_set_cpu(rq->cpu, rd->span);
8061 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
8064 spin_unlock_irqrestore(&rq->lock, flags);
8067 free_rootdomain(old_rd);
8070 static int init_rootdomain(struct root_domain *rd, bool bootmem)
8072 gfp_t gfp = GFP_KERNEL;
8074 memset(rd, 0, sizeof(*rd));
8079 if (!alloc_cpumask_var(&rd->span, gfp))
8081 if (!alloc_cpumask_var(&rd->online, gfp))
8083 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
8086 if (cpupri_init(&rd->cpupri, bootmem) != 0)
8091 free_cpumask_var(rd->rto_mask);
8093 free_cpumask_var(rd->online);
8095 free_cpumask_var(rd->span);
8100 static void init_defrootdomain(void)
8102 init_rootdomain(&def_root_domain, true);
8104 atomic_set(&def_root_domain.refcount, 1);
8107 static struct root_domain *alloc_rootdomain(void)
8109 struct root_domain *rd;
8111 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
8115 if (init_rootdomain(rd, false) != 0) {
8124 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8125 * hold the hotplug lock.
8128 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8130 struct rq *rq = cpu_rq(cpu);
8131 struct sched_domain *tmp;
8133 /* Remove the sched domains which do not contribute to scheduling. */
8134 for (tmp = sd; tmp; ) {
8135 struct sched_domain *parent = tmp->parent;
8139 if (sd_parent_degenerate(tmp, parent)) {
8140 tmp->parent = parent->parent;
8142 parent->parent->child = tmp;
8147 if (sd && sd_degenerate(sd)) {
8153 sched_domain_debug(sd, cpu);
8155 rq_attach_root(rq, rd);
8156 rcu_assign_pointer(rq->sd, sd);
8159 /* cpus with isolated domains */
8160 static cpumask_var_t cpu_isolated_map;
8162 /* Setup the mask of cpus configured for isolated domains */
8163 static int __init isolated_cpu_setup(char *str)
8165 cpulist_parse(str, cpu_isolated_map);
8169 __setup("isolcpus=", isolated_cpu_setup);
8172 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8173 * to a function which identifies what group(along with sched group) a CPU
8174 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8175 * (due to the fact that we keep track of groups covered with a struct cpumask).
8177 * init_sched_build_groups will build a circular linked list of the groups
8178 * covered by the given span, and will set each group's ->cpumask correctly,
8179 * and ->cpu_power to 0.
8182 init_sched_build_groups(const struct cpumask *span,
8183 const struct cpumask *cpu_map,
8184 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8185 struct sched_group **sg,
8186 struct cpumask *tmpmask),
8187 struct cpumask *covered, struct cpumask *tmpmask)
8189 struct sched_group *first = NULL, *last = NULL;
8192 cpumask_clear(covered);
8194 for_each_cpu(i, span) {
8195 struct sched_group *sg;
8196 int group = group_fn(i, cpu_map, &sg, tmpmask);
8199 if (cpumask_test_cpu(i, covered))
8202 cpumask_clear(sched_group_cpus(sg));
8205 for_each_cpu(j, span) {
8206 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8209 cpumask_set_cpu(j, covered);
8210 cpumask_set_cpu(j, sched_group_cpus(sg));
8221 #define SD_NODES_PER_DOMAIN 16
8226 * find_next_best_node - find the next node to include in a sched_domain
8227 * @node: node whose sched_domain we're building
8228 * @used_nodes: nodes already in the sched_domain
8230 * Find the next node to include in a given scheduling domain. Simply
8231 * finds the closest node not already in the @used_nodes map.
8233 * Should use nodemask_t.
8235 static int find_next_best_node(int node, nodemask_t *used_nodes)
8237 int i, n, val, min_val, best_node = 0;
8241 for (i = 0; i < nr_node_ids; i++) {
8242 /* Start at @node */
8243 n = (node + i) % nr_node_ids;
8245 if (!nr_cpus_node(n))
8248 /* Skip already used nodes */
8249 if (node_isset(n, *used_nodes))
8252 /* Simple min distance search */
8253 val = node_distance(node, n);
8255 if (val < min_val) {
8261 node_set(best_node, *used_nodes);
8266 * sched_domain_node_span - get a cpumask for a node's sched_domain
8267 * @node: node whose cpumask we're constructing
8268 * @span: resulting cpumask
8270 * Given a node, construct a good cpumask for its sched_domain to span. It
8271 * should be one that prevents unnecessary balancing, but also spreads tasks
8274 static void sched_domain_node_span(int node, struct cpumask *span)
8276 nodemask_t used_nodes;
8279 cpumask_clear(span);
8280 nodes_clear(used_nodes);
8282 cpumask_or(span, span, cpumask_of_node(node));
8283 node_set(node, used_nodes);
8285 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8286 int next_node = find_next_best_node(node, &used_nodes);
8288 cpumask_or(span, span, cpumask_of_node(next_node));
8291 #endif /* CONFIG_NUMA */
8293 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8296 * The cpus mask in sched_group and sched_domain hangs off the end.
8298 * ( See the the comments in include/linux/sched.h:struct sched_group
8299 * and struct sched_domain. )
8301 struct static_sched_group {
8302 struct sched_group sg;
8303 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8306 struct static_sched_domain {
8307 struct sched_domain sd;
8308 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8314 cpumask_var_t domainspan;
8315 cpumask_var_t covered;
8316 cpumask_var_t notcovered;
8318 cpumask_var_t nodemask;
8319 cpumask_var_t this_sibling_map;
8320 cpumask_var_t this_core_map;
8321 cpumask_var_t send_covered;
8322 cpumask_var_t tmpmask;
8323 struct sched_group **sched_group_nodes;
8324 struct root_domain *rd;
8328 sa_sched_groups = 0,
8333 sa_this_sibling_map,
8335 sa_sched_group_nodes,
8345 * SMT sched-domains:
8347 #ifdef CONFIG_SCHED_SMT
8348 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8349 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8352 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8353 struct sched_group **sg, struct cpumask *unused)
8356 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8359 #endif /* CONFIG_SCHED_SMT */
8362 * multi-core sched-domains:
8364 #ifdef CONFIG_SCHED_MC
8365 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8366 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8367 #endif /* CONFIG_SCHED_MC */
8369 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8371 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8372 struct sched_group **sg, struct cpumask *mask)
8376 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8377 group = cpumask_first(mask);
8379 *sg = &per_cpu(sched_group_core, group).sg;
8382 #elif defined(CONFIG_SCHED_MC)
8384 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8385 struct sched_group **sg, struct cpumask *unused)
8388 *sg = &per_cpu(sched_group_core, cpu).sg;
8393 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8394 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8397 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8398 struct sched_group **sg, struct cpumask *mask)
8401 #ifdef CONFIG_SCHED_MC
8402 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8403 group = cpumask_first(mask);
8404 #elif defined(CONFIG_SCHED_SMT)
8405 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8406 group = cpumask_first(mask);
8411 *sg = &per_cpu(sched_group_phys, group).sg;
8417 * The init_sched_build_groups can't handle what we want to do with node
8418 * groups, so roll our own. Now each node has its own list of groups which
8419 * gets dynamically allocated.
8421 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8422 static struct sched_group ***sched_group_nodes_bycpu;
8424 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8425 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8427 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8428 struct sched_group **sg,
8429 struct cpumask *nodemask)
8433 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8434 group = cpumask_first(nodemask);
8437 *sg = &per_cpu(sched_group_allnodes, group).sg;
8441 static void init_numa_sched_groups_power(struct sched_group *group_head)
8443 struct sched_group *sg = group_head;
8449 for_each_cpu(j, sched_group_cpus(sg)) {
8450 struct sched_domain *sd;
8452 sd = &per_cpu(phys_domains, j).sd;
8453 if (j != group_first_cpu(sd->groups)) {
8455 * Only add "power" once for each
8461 sg->cpu_power += sd->groups->cpu_power;
8464 } while (sg != group_head);
8467 static int build_numa_sched_groups(struct s_data *d,
8468 const struct cpumask *cpu_map, int num)
8470 struct sched_domain *sd;
8471 struct sched_group *sg, *prev;
8474 cpumask_clear(d->covered);
8475 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8476 if (cpumask_empty(d->nodemask)) {
8477 d->sched_group_nodes[num] = NULL;
8481 sched_domain_node_span(num, d->domainspan);
8482 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8484 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8487 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8491 d->sched_group_nodes[num] = sg;
8493 for_each_cpu(j, d->nodemask) {
8494 sd = &per_cpu(node_domains, j).sd;
8499 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8501 cpumask_or(d->covered, d->covered, d->nodemask);
8504 for (j = 0; j < nr_node_ids; j++) {
8505 n = (num + j) % nr_node_ids;
8506 cpumask_complement(d->notcovered, d->covered);
8507 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8508 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8509 if (cpumask_empty(d->tmpmask))
8511 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8512 if (cpumask_empty(d->tmpmask))
8514 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8518 "Can not alloc domain group for node %d\n", j);
8522 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8523 sg->next = prev->next;
8524 cpumask_or(d->covered, d->covered, d->tmpmask);
8531 #endif /* CONFIG_NUMA */
8534 /* Free memory allocated for various sched_group structures */
8535 static void free_sched_groups(const struct cpumask *cpu_map,
8536 struct cpumask *nodemask)
8540 for_each_cpu(cpu, cpu_map) {
8541 struct sched_group **sched_group_nodes
8542 = sched_group_nodes_bycpu[cpu];
8544 if (!sched_group_nodes)
8547 for (i = 0; i < nr_node_ids; i++) {
8548 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8550 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8551 if (cpumask_empty(nodemask))
8561 if (oldsg != sched_group_nodes[i])
8564 kfree(sched_group_nodes);
8565 sched_group_nodes_bycpu[cpu] = NULL;
8568 #else /* !CONFIG_NUMA */
8569 static void free_sched_groups(const struct cpumask *cpu_map,
8570 struct cpumask *nodemask)
8573 #endif /* CONFIG_NUMA */
8576 * Initialize sched groups cpu_power.
8578 * cpu_power indicates the capacity of sched group, which is used while
8579 * distributing the load between different sched groups in a sched domain.
8580 * Typically cpu_power for all the groups in a sched domain will be same unless
8581 * there are asymmetries in the topology. If there are asymmetries, group
8582 * having more cpu_power will pickup more load compared to the group having
8585 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8587 struct sched_domain *child;
8588 struct sched_group *group;
8592 WARN_ON(!sd || !sd->groups);
8594 if (cpu != group_first_cpu(sd->groups))
8599 sd->groups->cpu_power = 0;
8602 power = SCHED_LOAD_SCALE;
8603 weight = cpumask_weight(sched_domain_span(sd));
8605 * SMT siblings share the power of a single core.
8606 * Usually multiple threads get a better yield out of
8607 * that one core than a single thread would have,
8608 * reflect that in sd->smt_gain.
8610 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8611 power *= sd->smt_gain;
8613 power >>= SCHED_LOAD_SHIFT;
8615 sd->groups->cpu_power += power;
8620 * Add cpu_power of each child group to this groups cpu_power.
8622 group = child->groups;
8624 sd->groups->cpu_power += group->cpu_power;
8625 group = group->next;
8626 } while (group != child->groups);
8630 * Initializers for schedule domains
8631 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8634 #ifdef CONFIG_SCHED_DEBUG
8635 # define SD_INIT_NAME(sd, type) sd->name = #type
8637 # define SD_INIT_NAME(sd, type) do { } while (0)
8640 #define SD_INIT(sd, type) sd_init_##type(sd)
8642 #define SD_INIT_FUNC(type) \
8643 static noinline void sd_init_##type(struct sched_domain *sd) \
8645 memset(sd, 0, sizeof(*sd)); \
8646 *sd = SD_##type##_INIT; \
8647 sd->level = SD_LV_##type; \
8648 SD_INIT_NAME(sd, type); \
8653 SD_INIT_FUNC(ALLNODES)
8656 #ifdef CONFIG_SCHED_SMT
8657 SD_INIT_FUNC(SIBLING)
8659 #ifdef CONFIG_SCHED_MC
8663 static int default_relax_domain_level = -1;
8665 static int __init setup_relax_domain_level(char *str)
8669 val = simple_strtoul(str, NULL, 0);
8670 if (val < SD_LV_MAX)
8671 default_relax_domain_level = val;
8675 __setup("relax_domain_level=", setup_relax_domain_level);
8677 static void set_domain_attribute(struct sched_domain *sd,
8678 struct sched_domain_attr *attr)
8682 if (!attr || attr->relax_domain_level < 0) {
8683 if (default_relax_domain_level < 0)
8686 request = default_relax_domain_level;
8688 request = attr->relax_domain_level;
8689 if (request < sd->level) {
8690 /* turn off idle balance on this domain */
8691 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8693 /* turn on idle balance on this domain */
8694 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8698 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8699 const struct cpumask *cpu_map)
8702 case sa_sched_groups:
8703 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8704 d->sched_group_nodes = NULL;
8706 free_rootdomain(d->rd); /* fall through */
8708 free_cpumask_var(d->tmpmask); /* fall through */
8709 case sa_send_covered:
8710 free_cpumask_var(d->send_covered); /* fall through */
8711 case sa_this_core_map:
8712 free_cpumask_var(d->this_core_map); /* fall through */
8713 case sa_this_sibling_map:
8714 free_cpumask_var(d->this_sibling_map); /* fall through */
8716 free_cpumask_var(d->nodemask); /* fall through */
8717 case sa_sched_group_nodes:
8719 kfree(d->sched_group_nodes); /* fall through */
8721 free_cpumask_var(d->notcovered); /* fall through */
8723 free_cpumask_var(d->covered); /* fall through */
8725 free_cpumask_var(d->domainspan); /* fall through */
8732 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8733 const struct cpumask *cpu_map)
8736 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8738 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8739 return sa_domainspan;
8740 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8742 /* Allocate the per-node list of sched groups */
8743 d->sched_group_nodes = kcalloc(nr_node_ids,
8744 sizeof(struct sched_group *), GFP_KERNEL);
8745 if (!d->sched_group_nodes) {
8746 printk(KERN_WARNING "Can not alloc sched group node list\n");
8747 return sa_notcovered;
8749 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8751 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8752 return sa_sched_group_nodes;
8753 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8755 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8756 return sa_this_sibling_map;
8757 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8758 return sa_this_core_map;
8759 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8760 return sa_send_covered;
8761 d->rd = alloc_rootdomain();
8763 printk(KERN_WARNING "Cannot alloc root domain\n");
8766 return sa_rootdomain;
8769 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8770 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8772 struct sched_domain *sd = NULL;
8774 struct sched_domain *parent;
8777 if (cpumask_weight(cpu_map) >
8778 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8779 sd = &per_cpu(allnodes_domains, i).sd;
8780 SD_INIT(sd, ALLNODES);
8781 set_domain_attribute(sd, attr);
8782 cpumask_copy(sched_domain_span(sd), cpu_map);
8783 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8788 sd = &per_cpu(node_domains, i).sd;
8790 set_domain_attribute(sd, attr);
8791 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8792 sd->parent = parent;
8795 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8800 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8801 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8802 struct sched_domain *parent, int i)
8804 struct sched_domain *sd;
8805 sd = &per_cpu(phys_domains, i).sd;
8807 set_domain_attribute(sd, attr);
8808 cpumask_copy(sched_domain_span(sd), d->nodemask);
8809 sd->parent = parent;
8812 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8816 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8817 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8818 struct sched_domain *parent, int i)
8820 struct sched_domain *sd = parent;
8821 #ifdef CONFIG_SCHED_MC
8822 sd = &per_cpu(core_domains, i).sd;
8824 set_domain_attribute(sd, attr);
8825 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8826 sd->parent = parent;
8828 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8833 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8834 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8835 struct sched_domain *parent, int i)
8837 struct sched_domain *sd = parent;
8838 #ifdef CONFIG_SCHED_SMT
8839 sd = &per_cpu(cpu_domains, i).sd;
8840 SD_INIT(sd, SIBLING);
8841 set_domain_attribute(sd, attr);
8842 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8843 sd->parent = parent;
8845 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8850 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8851 const struct cpumask *cpu_map, int cpu)
8854 #ifdef CONFIG_SCHED_SMT
8855 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8856 cpumask_and(d->this_sibling_map, cpu_map,
8857 topology_thread_cpumask(cpu));
8858 if (cpu == cpumask_first(d->this_sibling_map))
8859 init_sched_build_groups(d->this_sibling_map, cpu_map,
8861 d->send_covered, d->tmpmask);
8864 #ifdef CONFIG_SCHED_MC
8865 case SD_LV_MC: /* set up multi-core groups */
8866 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8867 if (cpu == cpumask_first(d->this_core_map))
8868 init_sched_build_groups(d->this_core_map, cpu_map,
8870 d->send_covered, d->tmpmask);
8873 case SD_LV_CPU: /* set up physical groups */
8874 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8875 if (!cpumask_empty(d->nodemask))
8876 init_sched_build_groups(d->nodemask, cpu_map,
8878 d->send_covered, d->tmpmask);
8881 case SD_LV_ALLNODES:
8882 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8883 d->send_covered, d->tmpmask);
8892 * Build sched domains for a given set of cpus and attach the sched domains
8893 * to the individual cpus
8895 static int __build_sched_domains(const struct cpumask *cpu_map,
8896 struct sched_domain_attr *attr)
8898 enum s_alloc alloc_state = sa_none;
8900 struct sched_domain *sd;
8906 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8907 if (alloc_state != sa_rootdomain)
8909 alloc_state = sa_sched_groups;
8912 * Set up domains for cpus specified by the cpu_map.
8914 for_each_cpu(i, cpu_map) {
8915 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8918 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8919 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8920 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8921 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8924 for_each_cpu(i, cpu_map) {
8925 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8926 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8929 /* Set up physical groups */
8930 for (i = 0; i < nr_node_ids; i++)
8931 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8934 /* Set up node groups */
8936 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8938 for (i = 0; i < nr_node_ids; i++)
8939 if (build_numa_sched_groups(&d, cpu_map, i))
8943 /* Calculate CPU power for physical packages and nodes */
8944 #ifdef CONFIG_SCHED_SMT
8945 for_each_cpu(i, cpu_map) {
8946 sd = &per_cpu(cpu_domains, i).sd;
8947 init_sched_groups_power(i, sd);
8950 #ifdef CONFIG_SCHED_MC
8951 for_each_cpu(i, cpu_map) {
8952 sd = &per_cpu(core_domains, i).sd;
8953 init_sched_groups_power(i, sd);
8957 for_each_cpu(i, cpu_map) {
8958 sd = &per_cpu(phys_domains, i).sd;
8959 init_sched_groups_power(i, sd);
8963 for (i = 0; i < nr_node_ids; i++)
8964 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8966 if (d.sd_allnodes) {
8967 struct sched_group *sg;
8969 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8971 init_numa_sched_groups_power(sg);
8975 /* Attach the domains */
8976 for_each_cpu(i, cpu_map) {
8977 #ifdef CONFIG_SCHED_SMT
8978 sd = &per_cpu(cpu_domains, i).sd;
8979 #elif defined(CONFIG_SCHED_MC)
8980 sd = &per_cpu(core_domains, i).sd;
8982 sd = &per_cpu(phys_domains, i).sd;
8984 cpu_attach_domain(sd, d.rd, i);
8987 d.sched_group_nodes = NULL; /* don't free this we still need it */
8988 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8992 __free_domain_allocs(&d, alloc_state, cpu_map);
8996 static int build_sched_domains(const struct cpumask *cpu_map)
8998 return __build_sched_domains(cpu_map, NULL);
9001 static struct cpumask *doms_cur; /* current sched domains */
9002 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
9003 static struct sched_domain_attr *dattr_cur;
9004 /* attribues of custom domains in 'doms_cur' */
9007 * Special case: If a kmalloc of a doms_cur partition (array of
9008 * cpumask) fails, then fallback to a single sched domain,
9009 * as determined by the single cpumask fallback_doms.
9011 static cpumask_var_t fallback_doms;
9014 * arch_update_cpu_topology lets virtualized architectures update the
9015 * cpu core maps. It is supposed to return 1 if the topology changed
9016 * or 0 if it stayed the same.
9018 int __attribute__((weak)) arch_update_cpu_topology(void)
9024 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9025 * For now this just excludes isolated cpus, but could be used to
9026 * exclude other special cases in the future.
9028 static int arch_init_sched_domains(const struct cpumask *cpu_map)
9032 arch_update_cpu_topology();
9034 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
9036 doms_cur = fallback_doms;
9037 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
9039 err = build_sched_domains(doms_cur);
9040 register_sched_domain_sysctl();
9045 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
9046 struct cpumask *tmpmask)
9048 free_sched_groups(cpu_map, tmpmask);
9052 * Detach sched domains from a group of cpus specified in cpu_map
9053 * These cpus will now be attached to the NULL domain
9055 static void detach_destroy_domains(const struct cpumask *cpu_map)
9057 /* Save because hotplug lock held. */
9058 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
9061 for_each_cpu(i, cpu_map)
9062 cpu_attach_domain(NULL, &def_root_domain, i);
9063 synchronize_sched();
9064 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
9067 /* handle null as "default" */
9068 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
9069 struct sched_domain_attr *new, int idx_new)
9071 struct sched_domain_attr tmp;
9078 return !memcmp(cur ? (cur + idx_cur) : &tmp,
9079 new ? (new + idx_new) : &tmp,
9080 sizeof(struct sched_domain_attr));
9084 * Partition sched domains as specified by the 'ndoms_new'
9085 * cpumasks in the array doms_new[] of cpumasks. This compares
9086 * doms_new[] to the current sched domain partitioning, doms_cur[].
9087 * It destroys each deleted domain and builds each new domain.
9089 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9090 * The masks don't intersect (don't overlap.) We should setup one
9091 * sched domain for each mask. CPUs not in any of the cpumasks will
9092 * not be load balanced. If the same cpumask appears both in the
9093 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9096 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9097 * ownership of it and will kfree it when done with it. If the caller
9098 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9099 * ndoms_new == 1, and partition_sched_domains() will fallback to
9100 * the single partition 'fallback_doms', it also forces the domains
9103 * If doms_new == NULL it will be replaced with cpu_online_mask.
9104 * ndoms_new == 0 is a special case for destroying existing domains,
9105 * and it will not create the default domain.
9107 * Call with hotplug lock held
9109 /* FIXME: Change to struct cpumask *doms_new[] */
9110 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
9111 struct sched_domain_attr *dattr_new)
9116 mutex_lock(&sched_domains_mutex);
9118 /* always unregister in case we don't destroy any domains */
9119 unregister_sched_domain_sysctl();
9121 /* Let architecture update cpu core mappings. */
9122 new_topology = arch_update_cpu_topology();
9124 n = doms_new ? ndoms_new : 0;
9126 /* Destroy deleted domains */
9127 for (i = 0; i < ndoms_cur; i++) {
9128 for (j = 0; j < n && !new_topology; j++) {
9129 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9130 && dattrs_equal(dattr_cur, i, dattr_new, j))
9133 /* no match - a current sched domain not in new doms_new[] */
9134 detach_destroy_domains(doms_cur + i);
9139 if (doms_new == NULL) {
9141 doms_new = fallback_doms;
9142 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9143 WARN_ON_ONCE(dattr_new);
9146 /* Build new domains */
9147 for (i = 0; i < ndoms_new; i++) {
9148 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9149 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9150 && dattrs_equal(dattr_new, i, dattr_cur, j))
9153 /* no match - add a new doms_new */
9154 __build_sched_domains(doms_new + i,
9155 dattr_new ? dattr_new + i : NULL);
9160 /* Remember the new sched domains */
9161 if (doms_cur != fallback_doms)
9163 kfree(dattr_cur); /* kfree(NULL) is safe */
9164 doms_cur = doms_new;
9165 dattr_cur = dattr_new;
9166 ndoms_cur = ndoms_new;
9168 register_sched_domain_sysctl();
9170 mutex_unlock(&sched_domains_mutex);
9173 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9174 static void arch_reinit_sched_domains(void)
9178 /* Destroy domains first to force the rebuild */
9179 partition_sched_domains(0, NULL, NULL);
9181 rebuild_sched_domains();
9185 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9187 unsigned int level = 0;
9189 if (sscanf(buf, "%u", &level) != 1)
9193 * level is always be positive so don't check for
9194 * level < POWERSAVINGS_BALANCE_NONE which is 0
9195 * What happens on 0 or 1 byte write,
9196 * need to check for count as well?
9199 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9203 sched_smt_power_savings = level;
9205 sched_mc_power_savings = level;
9207 arch_reinit_sched_domains();
9212 #ifdef CONFIG_SCHED_MC
9213 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9216 return sprintf(page, "%u\n", sched_mc_power_savings);
9218 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9219 const char *buf, size_t count)
9221 return sched_power_savings_store(buf, count, 0);
9223 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9224 sched_mc_power_savings_show,
9225 sched_mc_power_savings_store);
9228 #ifdef CONFIG_SCHED_SMT
9229 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9232 return sprintf(page, "%u\n", sched_smt_power_savings);
9234 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9235 const char *buf, size_t count)
9237 return sched_power_savings_store(buf, count, 1);
9239 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9240 sched_smt_power_savings_show,
9241 sched_smt_power_savings_store);
9244 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9248 #ifdef CONFIG_SCHED_SMT
9250 err = sysfs_create_file(&cls->kset.kobj,
9251 &attr_sched_smt_power_savings.attr);
9253 #ifdef CONFIG_SCHED_MC
9254 if (!err && mc_capable())
9255 err = sysfs_create_file(&cls->kset.kobj,
9256 &attr_sched_mc_power_savings.attr);
9260 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9262 #ifndef CONFIG_CPUSETS
9264 * Add online and remove offline CPUs from the scheduler domains.
9265 * When cpusets are enabled they take over this function.
9267 static int update_sched_domains(struct notifier_block *nfb,
9268 unsigned long action, void *hcpu)
9272 case CPU_ONLINE_FROZEN:
9274 case CPU_DEAD_FROZEN:
9275 partition_sched_domains(1, NULL, NULL);
9284 static int update_runtime(struct notifier_block *nfb,
9285 unsigned long action, void *hcpu)
9287 int cpu = (int)(long)hcpu;
9290 case CPU_DOWN_PREPARE:
9291 case CPU_DOWN_PREPARE_FROZEN:
9292 disable_runtime(cpu_rq(cpu));
9295 case CPU_DOWN_FAILED:
9296 case CPU_DOWN_FAILED_FROZEN:
9298 case CPU_ONLINE_FROZEN:
9299 enable_runtime(cpu_rq(cpu));
9307 void __init sched_init_smp(void)
9309 cpumask_var_t non_isolated_cpus;
9311 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9313 #if defined(CONFIG_NUMA)
9314 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9316 BUG_ON(sched_group_nodes_bycpu == NULL);
9319 mutex_lock(&sched_domains_mutex);
9320 arch_init_sched_domains(cpu_online_mask);
9321 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9322 if (cpumask_empty(non_isolated_cpus))
9323 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9324 mutex_unlock(&sched_domains_mutex);
9327 #ifndef CONFIG_CPUSETS
9328 /* XXX: Theoretical race here - CPU may be hotplugged now */
9329 hotcpu_notifier(update_sched_domains, 0);
9332 /* RT runtime code needs to handle some hotplug events */
9333 hotcpu_notifier(update_runtime, 0);
9337 /* Move init over to a non-isolated CPU */
9338 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9340 sched_init_granularity();
9341 free_cpumask_var(non_isolated_cpus);
9343 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9344 init_sched_rt_class();
9347 void __init sched_init_smp(void)
9349 sched_init_granularity();
9351 #endif /* CONFIG_SMP */
9353 const_debug unsigned int sysctl_timer_migration = 1;
9355 int in_sched_functions(unsigned long addr)
9357 return in_lock_functions(addr) ||
9358 (addr >= (unsigned long)__sched_text_start
9359 && addr < (unsigned long)__sched_text_end);
9362 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9364 cfs_rq->tasks_timeline = RB_ROOT;
9365 INIT_LIST_HEAD(&cfs_rq->tasks);
9366 #ifdef CONFIG_FAIR_GROUP_SCHED
9369 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9372 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9374 struct rt_prio_array *array;
9377 array = &rt_rq->active;
9378 for (i = 0; i < MAX_RT_PRIO; i++) {
9379 INIT_LIST_HEAD(array->queue + i);
9380 __clear_bit(i, array->bitmap);
9382 /* delimiter for bitsearch: */
9383 __set_bit(MAX_RT_PRIO, array->bitmap);
9385 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9386 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9388 rt_rq->highest_prio.next = MAX_RT_PRIO;
9392 rt_rq->rt_nr_migratory = 0;
9393 rt_rq->overloaded = 0;
9394 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9398 rt_rq->rt_throttled = 0;
9399 rt_rq->rt_runtime = 0;
9400 spin_lock_init(&rt_rq->rt_runtime_lock);
9402 #ifdef CONFIG_RT_GROUP_SCHED
9403 rt_rq->rt_nr_boosted = 0;
9408 #ifdef CONFIG_FAIR_GROUP_SCHED
9409 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9410 struct sched_entity *se, int cpu, int add,
9411 struct sched_entity *parent)
9413 struct rq *rq = cpu_rq(cpu);
9414 tg->cfs_rq[cpu] = cfs_rq;
9415 init_cfs_rq(cfs_rq, rq);
9418 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9421 /* se could be NULL for init_task_group */
9426 se->cfs_rq = &rq->cfs;
9428 se->cfs_rq = parent->my_q;
9431 se->load.weight = tg->shares;
9432 se->load.inv_weight = 0;
9433 se->parent = parent;
9437 #ifdef CONFIG_RT_GROUP_SCHED
9438 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9439 struct sched_rt_entity *rt_se, int cpu, int add,
9440 struct sched_rt_entity *parent)
9442 struct rq *rq = cpu_rq(cpu);
9444 tg->rt_rq[cpu] = rt_rq;
9445 init_rt_rq(rt_rq, rq);
9447 rt_rq->rt_se = rt_se;
9448 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9450 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9452 tg->rt_se[cpu] = rt_se;
9457 rt_se->rt_rq = &rq->rt;
9459 rt_se->rt_rq = parent->my_q;
9461 rt_se->my_q = rt_rq;
9462 rt_se->parent = parent;
9463 INIT_LIST_HEAD(&rt_se->run_list);
9467 void __init sched_init(void)
9470 unsigned long alloc_size = 0, ptr;
9472 #ifdef CONFIG_FAIR_GROUP_SCHED
9473 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9475 #ifdef CONFIG_RT_GROUP_SCHED
9476 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9478 #ifdef CONFIG_USER_SCHED
9481 #ifdef CONFIG_CPUMASK_OFFSTACK
9482 alloc_size += num_possible_cpus() * cpumask_size();
9485 * As sched_init() is called before page_alloc is setup,
9486 * we use alloc_bootmem().
9489 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9491 #ifdef CONFIG_FAIR_GROUP_SCHED
9492 init_task_group.se = (struct sched_entity **)ptr;
9493 ptr += nr_cpu_ids * sizeof(void **);
9495 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9496 ptr += nr_cpu_ids * sizeof(void **);
9498 #ifdef CONFIG_USER_SCHED
9499 root_task_group.se = (struct sched_entity **)ptr;
9500 ptr += nr_cpu_ids * sizeof(void **);
9502 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9503 ptr += nr_cpu_ids * sizeof(void **);
9504 #endif /* CONFIG_USER_SCHED */
9505 #endif /* CONFIG_FAIR_GROUP_SCHED */
9506 #ifdef CONFIG_RT_GROUP_SCHED
9507 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9508 ptr += nr_cpu_ids * sizeof(void **);
9510 init_task_group.rt_rq = (struct rt_rq **)ptr;
9511 ptr += nr_cpu_ids * sizeof(void **);
9513 #ifdef CONFIG_USER_SCHED
9514 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9515 ptr += nr_cpu_ids * sizeof(void **);
9517 root_task_group.rt_rq = (struct rt_rq **)ptr;
9518 ptr += nr_cpu_ids * sizeof(void **);
9519 #endif /* CONFIG_USER_SCHED */
9520 #endif /* CONFIG_RT_GROUP_SCHED */
9521 #ifdef CONFIG_CPUMASK_OFFSTACK
9522 for_each_possible_cpu(i) {
9523 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9524 ptr += cpumask_size();
9526 #endif /* CONFIG_CPUMASK_OFFSTACK */
9530 init_defrootdomain();
9533 init_rt_bandwidth(&def_rt_bandwidth,
9534 global_rt_period(), global_rt_runtime());
9536 #ifdef CONFIG_RT_GROUP_SCHED
9537 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9538 global_rt_period(), global_rt_runtime());
9539 #ifdef CONFIG_USER_SCHED
9540 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9541 global_rt_period(), RUNTIME_INF);
9542 #endif /* CONFIG_USER_SCHED */
9543 #endif /* CONFIG_RT_GROUP_SCHED */
9545 #ifdef CONFIG_GROUP_SCHED
9546 list_add(&init_task_group.list, &task_groups);
9547 INIT_LIST_HEAD(&init_task_group.children);
9549 #ifdef CONFIG_USER_SCHED
9550 INIT_LIST_HEAD(&root_task_group.children);
9551 init_task_group.parent = &root_task_group;
9552 list_add(&init_task_group.siblings, &root_task_group.children);
9553 #endif /* CONFIG_USER_SCHED */
9554 #endif /* CONFIG_GROUP_SCHED */
9556 for_each_possible_cpu(i) {
9560 spin_lock_init(&rq->lock);
9562 rq->calc_load_active = 0;
9563 rq->calc_load_update = jiffies + LOAD_FREQ;
9564 init_cfs_rq(&rq->cfs, rq);
9565 init_rt_rq(&rq->rt, rq);
9566 #ifdef CONFIG_FAIR_GROUP_SCHED
9567 init_task_group.shares = init_task_group_load;
9568 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9569 #ifdef CONFIG_CGROUP_SCHED
9571 * How much cpu bandwidth does init_task_group get?
9573 * In case of task-groups formed thr' the cgroup filesystem, it
9574 * gets 100% of the cpu resources in the system. This overall
9575 * system cpu resource is divided among the tasks of
9576 * init_task_group and its child task-groups in a fair manner,
9577 * based on each entity's (task or task-group's) weight
9578 * (se->load.weight).
9580 * In other words, if init_task_group has 10 tasks of weight
9581 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9582 * then A0's share of the cpu resource is:
9584 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9586 * We achieve this by letting init_task_group's tasks sit
9587 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9589 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9590 #elif defined CONFIG_USER_SCHED
9591 root_task_group.shares = NICE_0_LOAD;
9592 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9594 * In case of task-groups formed thr' the user id of tasks,
9595 * init_task_group represents tasks belonging to root user.
9596 * Hence it forms a sibling of all subsequent groups formed.
9597 * In this case, init_task_group gets only a fraction of overall
9598 * system cpu resource, based on the weight assigned to root
9599 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9600 * by letting tasks of init_task_group sit in a separate cfs_rq
9601 * (init_tg_cfs_rq) and having one entity represent this group of
9602 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9604 init_tg_cfs_entry(&init_task_group,
9605 &per_cpu(init_tg_cfs_rq, i),
9606 &per_cpu(init_sched_entity, i), i, 1,
9607 root_task_group.se[i]);
9610 #endif /* CONFIG_FAIR_GROUP_SCHED */
9612 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9613 #ifdef CONFIG_RT_GROUP_SCHED
9614 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9615 #ifdef CONFIG_CGROUP_SCHED
9616 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9617 #elif defined CONFIG_USER_SCHED
9618 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9619 init_tg_rt_entry(&init_task_group,
9620 &per_cpu(init_rt_rq, i),
9621 &per_cpu(init_sched_rt_entity, i), i, 1,
9622 root_task_group.rt_se[i]);
9626 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9627 rq->cpu_load[j] = 0;
9631 rq->post_schedule = 0;
9632 rq->active_balance = 0;
9633 rq->next_balance = jiffies;
9637 rq->migration_thread = NULL;
9638 INIT_LIST_HEAD(&rq->migration_queue);
9639 rq_attach_root(rq, &def_root_domain);
9642 atomic_set(&rq->nr_iowait, 0);
9645 set_load_weight(&init_task);
9647 #ifdef CONFIG_PREEMPT_NOTIFIERS
9648 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9652 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9655 #ifdef CONFIG_RT_MUTEXES
9656 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9660 * The boot idle thread does lazy MMU switching as well:
9662 atomic_inc(&init_mm.mm_count);
9663 enter_lazy_tlb(&init_mm, current);
9666 * Make us the idle thread. Technically, schedule() should not be
9667 * called from this thread, however somewhere below it might be,
9668 * but because we are the idle thread, we just pick up running again
9669 * when this runqueue becomes "idle".
9671 init_idle(current, smp_processor_id());
9673 calc_load_update = jiffies + LOAD_FREQ;
9676 * During early bootup we pretend to be a normal task:
9678 current->sched_class = &fair_sched_class;
9680 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9681 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9684 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9685 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9687 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9690 perf_counter_init();
9692 scheduler_running = 1;
9695 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9696 static inline int preempt_count_equals(int preempt_offset)
9698 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9700 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9703 void __might_sleep(char *file, int line, int preempt_offset)
9706 static unsigned long prev_jiffy; /* ratelimiting */
9708 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9709 system_state != SYSTEM_RUNNING || oops_in_progress)
9711 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9713 prev_jiffy = jiffies;
9716 "BUG: sleeping function called from invalid context at %s:%d\n",
9719 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9720 in_atomic(), irqs_disabled(),
9721 current->pid, current->comm);
9723 debug_show_held_locks(current);
9724 if (irqs_disabled())
9725 print_irqtrace_events(current);
9729 EXPORT_SYMBOL(__might_sleep);
9732 #ifdef CONFIG_MAGIC_SYSRQ
9733 static void normalize_task(struct rq *rq, struct task_struct *p)
9737 update_rq_clock(rq);
9738 on_rq = p->se.on_rq;
9740 deactivate_task(rq, p, 0);
9741 __setscheduler(rq, p, SCHED_NORMAL, 0);
9743 activate_task(rq, p, 0);
9744 resched_task(rq->curr);
9748 void normalize_rt_tasks(void)
9750 struct task_struct *g, *p;
9751 unsigned long flags;
9754 read_lock_irqsave(&tasklist_lock, flags);
9755 do_each_thread(g, p) {
9757 * Only normalize user tasks:
9762 p->se.exec_start = 0;
9763 #ifdef CONFIG_SCHEDSTATS
9764 p->se.wait_start = 0;
9765 p->se.sleep_start = 0;
9766 p->se.block_start = 0;
9771 * Renice negative nice level userspace
9774 if (TASK_NICE(p) < 0 && p->mm)
9775 set_user_nice(p, 0);
9779 spin_lock(&p->pi_lock);
9780 rq = __task_rq_lock(p);
9782 normalize_task(rq, p);
9784 __task_rq_unlock(rq);
9785 spin_unlock(&p->pi_lock);
9786 } while_each_thread(g, p);
9788 read_unlock_irqrestore(&tasklist_lock, flags);
9791 #endif /* CONFIG_MAGIC_SYSRQ */
9795 * These functions are only useful for the IA64 MCA handling.
9797 * They can only be called when the whole system has been
9798 * stopped - every CPU needs to be quiescent, and no scheduling
9799 * activity can take place. Using them for anything else would
9800 * be a serious bug, and as a result, they aren't even visible
9801 * under any other configuration.
9805 * curr_task - return the current task for a given cpu.
9806 * @cpu: the processor in question.
9808 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9810 struct task_struct *curr_task(int cpu)
9812 return cpu_curr(cpu);
9816 * set_curr_task - set the current task for a given cpu.
9817 * @cpu: the processor in question.
9818 * @p: the task pointer to set.
9820 * Description: This function must only be used when non-maskable interrupts
9821 * are serviced on a separate stack. It allows the architecture to switch the
9822 * notion of the current task on a cpu in a non-blocking manner. This function
9823 * must be called with all CPU's synchronized, and interrupts disabled, the
9824 * and caller must save the original value of the current task (see
9825 * curr_task() above) and restore that value before reenabling interrupts and
9826 * re-starting the system.
9828 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9830 void set_curr_task(int cpu, struct task_struct *p)
9837 #ifdef CONFIG_FAIR_GROUP_SCHED
9838 static void free_fair_sched_group(struct task_group *tg)
9842 for_each_possible_cpu(i) {
9844 kfree(tg->cfs_rq[i]);
9854 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9856 struct cfs_rq *cfs_rq;
9857 struct sched_entity *se;
9861 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9864 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9868 tg->shares = NICE_0_LOAD;
9870 for_each_possible_cpu(i) {
9873 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9874 GFP_KERNEL, cpu_to_node(i));
9878 se = kzalloc_node(sizeof(struct sched_entity),
9879 GFP_KERNEL, cpu_to_node(i));
9883 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9892 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9894 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9895 &cpu_rq(cpu)->leaf_cfs_rq_list);
9898 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9900 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9902 #else /* !CONFG_FAIR_GROUP_SCHED */
9903 static inline void free_fair_sched_group(struct task_group *tg)
9908 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9913 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9917 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9920 #endif /* CONFIG_FAIR_GROUP_SCHED */
9922 #ifdef CONFIG_RT_GROUP_SCHED
9923 static void free_rt_sched_group(struct task_group *tg)
9927 destroy_rt_bandwidth(&tg->rt_bandwidth);
9929 for_each_possible_cpu(i) {
9931 kfree(tg->rt_rq[i]);
9933 kfree(tg->rt_se[i]);
9941 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9943 struct rt_rq *rt_rq;
9944 struct sched_rt_entity *rt_se;
9948 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9951 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9955 init_rt_bandwidth(&tg->rt_bandwidth,
9956 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9958 for_each_possible_cpu(i) {
9961 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9962 GFP_KERNEL, cpu_to_node(i));
9966 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9967 GFP_KERNEL, cpu_to_node(i));
9971 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9980 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9982 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9983 &cpu_rq(cpu)->leaf_rt_rq_list);
9986 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9988 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9990 #else /* !CONFIG_RT_GROUP_SCHED */
9991 static inline void free_rt_sched_group(struct task_group *tg)
9996 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
10001 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
10005 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
10008 #endif /* CONFIG_RT_GROUP_SCHED */
10010 #ifdef CONFIG_GROUP_SCHED
10011 static void free_sched_group(struct task_group *tg)
10013 free_fair_sched_group(tg);
10014 free_rt_sched_group(tg);
10018 /* allocate runqueue etc for a new task group */
10019 struct task_group *sched_create_group(struct task_group *parent)
10021 struct task_group *tg;
10022 unsigned long flags;
10025 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
10027 return ERR_PTR(-ENOMEM);
10029 if (!alloc_fair_sched_group(tg, parent))
10032 if (!alloc_rt_sched_group(tg, parent))
10035 spin_lock_irqsave(&task_group_lock, flags);
10036 for_each_possible_cpu(i) {
10037 register_fair_sched_group(tg, i);
10038 register_rt_sched_group(tg, i);
10040 list_add_rcu(&tg->list, &task_groups);
10042 WARN_ON(!parent); /* root should already exist */
10044 tg->parent = parent;
10045 INIT_LIST_HEAD(&tg->children);
10046 list_add_rcu(&tg->siblings, &parent->children);
10047 spin_unlock_irqrestore(&task_group_lock, flags);
10052 free_sched_group(tg);
10053 return ERR_PTR(-ENOMEM);
10056 /* rcu callback to free various structures associated with a task group */
10057 static void free_sched_group_rcu(struct rcu_head *rhp)
10059 /* now it should be safe to free those cfs_rqs */
10060 free_sched_group(container_of(rhp, struct task_group, rcu));
10063 /* Destroy runqueue etc associated with a task group */
10064 void sched_destroy_group(struct task_group *tg)
10066 unsigned long flags;
10069 spin_lock_irqsave(&task_group_lock, flags);
10070 for_each_possible_cpu(i) {
10071 unregister_fair_sched_group(tg, i);
10072 unregister_rt_sched_group(tg, i);
10074 list_del_rcu(&tg->list);
10075 list_del_rcu(&tg->siblings);
10076 spin_unlock_irqrestore(&task_group_lock, flags);
10078 /* wait for possible concurrent references to cfs_rqs complete */
10079 call_rcu(&tg->rcu, free_sched_group_rcu);
10082 /* change task's runqueue when it moves between groups.
10083 * The caller of this function should have put the task in its new group
10084 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10085 * reflect its new group.
10087 void sched_move_task(struct task_struct *tsk)
10089 int on_rq, running;
10090 unsigned long flags;
10093 rq = task_rq_lock(tsk, &flags);
10095 update_rq_clock(rq);
10097 running = task_current(rq, tsk);
10098 on_rq = tsk->se.on_rq;
10101 dequeue_task(rq, tsk, 0);
10102 if (unlikely(running))
10103 tsk->sched_class->put_prev_task(rq, tsk);
10105 set_task_rq(tsk, task_cpu(tsk));
10107 #ifdef CONFIG_FAIR_GROUP_SCHED
10108 if (tsk->sched_class->moved_group)
10109 tsk->sched_class->moved_group(tsk);
10112 if (unlikely(running))
10113 tsk->sched_class->set_curr_task(rq);
10115 enqueue_task(rq, tsk, 0);
10117 task_rq_unlock(rq, &flags);
10119 #endif /* CONFIG_GROUP_SCHED */
10121 #ifdef CONFIG_FAIR_GROUP_SCHED
10122 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
10124 struct cfs_rq *cfs_rq = se->cfs_rq;
10129 dequeue_entity(cfs_rq, se, 0);
10131 se->load.weight = shares;
10132 se->load.inv_weight = 0;
10135 enqueue_entity(cfs_rq, se, 0);
10138 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10140 struct cfs_rq *cfs_rq = se->cfs_rq;
10141 struct rq *rq = cfs_rq->rq;
10142 unsigned long flags;
10144 spin_lock_irqsave(&rq->lock, flags);
10145 __set_se_shares(se, shares);
10146 spin_unlock_irqrestore(&rq->lock, flags);
10149 static DEFINE_MUTEX(shares_mutex);
10151 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10154 unsigned long flags;
10157 * We can't change the weight of the root cgroup.
10162 if (shares < MIN_SHARES)
10163 shares = MIN_SHARES;
10164 else if (shares > MAX_SHARES)
10165 shares = MAX_SHARES;
10167 mutex_lock(&shares_mutex);
10168 if (tg->shares == shares)
10171 spin_lock_irqsave(&task_group_lock, flags);
10172 for_each_possible_cpu(i)
10173 unregister_fair_sched_group(tg, i);
10174 list_del_rcu(&tg->siblings);
10175 spin_unlock_irqrestore(&task_group_lock, flags);
10177 /* wait for any ongoing reference to this group to finish */
10178 synchronize_sched();
10181 * Now we are free to modify the group's share on each cpu
10182 * w/o tripping rebalance_share or load_balance_fair.
10184 tg->shares = shares;
10185 for_each_possible_cpu(i) {
10187 * force a rebalance
10189 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10190 set_se_shares(tg->se[i], shares);
10194 * Enable load balance activity on this group, by inserting it back on
10195 * each cpu's rq->leaf_cfs_rq_list.
10197 spin_lock_irqsave(&task_group_lock, flags);
10198 for_each_possible_cpu(i)
10199 register_fair_sched_group(tg, i);
10200 list_add_rcu(&tg->siblings, &tg->parent->children);
10201 spin_unlock_irqrestore(&task_group_lock, flags);
10203 mutex_unlock(&shares_mutex);
10207 unsigned long sched_group_shares(struct task_group *tg)
10213 #ifdef CONFIG_RT_GROUP_SCHED
10215 * Ensure that the real time constraints are schedulable.
10217 static DEFINE_MUTEX(rt_constraints_mutex);
10219 static unsigned long to_ratio(u64 period, u64 runtime)
10221 if (runtime == RUNTIME_INF)
10224 return div64_u64(runtime << 20, period);
10227 /* Must be called with tasklist_lock held */
10228 static inline int tg_has_rt_tasks(struct task_group *tg)
10230 struct task_struct *g, *p;
10232 do_each_thread(g, p) {
10233 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10235 } while_each_thread(g, p);
10240 struct rt_schedulable_data {
10241 struct task_group *tg;
10246 static int tg_schedulable(struct task_group *tg, void *data)
10248 struct rt_schedulable_data *d = data;
10249 struct task_group *child;
10250 unsigned long total, sum = 0;
10251 u64 period, runtime;
10253 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10254 runtime = tg->rt_bandwidth.rt_runtime;
10257 period = d->rt_period;
10258 runtime = d->rt_runtime;
10261 #ifdef CONFIG_USER_SCHED
10262 if (tg == &root_task_group) {
10263 period = global_rt_period();
10264 runtime = global_rt_runtime();
10269 * Cannot have more runtime than the period.
10271 if (runtime > period && runtime != RUNTIME_INF)
10275 * Ensure we don't starve existing RT tasks.
10277 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10280 total = to_ratio(period, runtime);
10283 * Nobody can have more than the global setting allows.
10285 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10289 * The sum of our children's runtime should not exceed our own.
10291 list_for_each_entry_rcu(child, &tg->children, siblings) {
10292 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10293 runtime = child->rt_bandwidth.rt_runtime;
10295 if (child == d->tg) {
10296 period = d->rt_period;
10297 runtime = d->rt_runtime;
10300 sum += to_ratio(period, runtime);
10309 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10311 struct rt_schedulable_data data = {
10313 .rt_period = period,
10314 .rt_runtime = runtime,
10317 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10320 static int tg_set_bandwidth(struct task_group *tg,
10321 u64 rt_period, u64 rt_runtime)
10325 mutex_lock(&rt_constraints_mutex);
10326 read_lock(&tasklist_lock);
10327 err = __rt_schedulable(tg, rt_period, rt_runtime);
10331 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10332 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10333 tg->rt_bandwidth.rt_runtime = rt_runtime;
10335 for_each_possible_cpu(i) {
10336 struct rt_rq *rt_rq = tg->rt_rq[i];
10338 spin_lock(&rt_rq->rt_runtime_lock);
10339 rt_rq->rt_runtime = rt_runtime;
10340 spin_unlock(&rt_rq->rt_runtime_lock);
10342 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10344 read_unlock(&tasklist_lock);
10345 mutex_unlock(&rt_constraints_mutex);
10350 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10352 u64 rt_runtime, rt_period;
10354 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10355 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10356 if (rt_runtime_us < 0)
10357 rt_runtime = RUNTIME_INF;
10359 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10362 long sched_group_rt_runtime(struct task_group *tg)
10366 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10369 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10370 do_div(rt_runtime_us, NSEC_PER_USEC);
10371 return rt_runtime_us;
10374 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10376 u64 rt_runtime, rt_period;
10378 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10379 rt_runtime = tg->rt_bandwidth.rt_runtime;
10381 if (rt_period == 0)
10384 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10387 long sched_group_rt_period(struct task_group *tg)
10391 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10392 do_div(rt_period_us, NSEC_PER_USEC);
10393 return rt_period_us;
10396 static int sched_rt_global_constraints(void)
10398 u64 runtime, period;
10401 if (sysctl_sched_rt_period <= 0)
10404 runtime = global_rt_runtime();
10405 period = global_rt_period();
10408 * Sanity check on the sysctl variables.
10410 if (runtime > period && runtime != RUNTIME_INF)
10413 mutex_lock(&rt_constraints_mutex);
10414 read_lock(&tasklist_lock);
10415 ret = __rt_schedulable(NULL, 0, 0);
10416 read_unlock(&tasklist_lock);
10417 mutex_unlock(&rt_constraints_mutex);
10422 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10424 /* Don't accept realtime tasks when there is no way for them to run */
10425 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10431 #else /* !CONFIG_RT_GROUP_SCHED */
10432 static int sched_rt_global_constraints(void)
10434 unsigned long flags;
10437 if (sysctl_sched_rt_period <= 0)
10441 * There's always some RT tasks in the root group
10442 * -- migration, kstopmachine etc..
10444 if (sysctl_sched_rt_runtime == 0)
10447 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10448 for_each_possible_cpu(i) {
10449 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10451 spin_lock(&rt_rq->rt_runtime_lock);
10452 rt_rq->rt_runtime = global_rt_runtime();
10453 spin_unlock(&rt_rq->rt_runtime_lock);
10455 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10459 #endif /* CONFIG_RT_GROUP_SCHED */
10461 int sched_rt_handler(struct ctl_table *table, int write,
10462 struct file *filp, void __user *buffer, size_t *lenp,
10466 int old_period, old_runtime;
10467 static DEFINE_MUTEX(mutex);
10469 mutex_lock(&mutex);
10470 old_period = sysctl_sched_rt_period;
10471 old_runtime = sysctl_sched_rt_runtime;
10473 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10475 if (!ret && write) {
10476 ret = sched_rt_global_constraints();
10478 sysctl_sched_rt_period = old_period;
10479 sysctl_sched_rt_runtime = old_runtime;
10481 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10482 def_rt_bandwidth.rt_period =
10483 ns_to_ktime(global_rt_period());
10486 mutex_unlock(&mutex);
10491 #ifdef CONFIG_CGROUP_SCHED
10493 /* return corresponding task_group object of a cgroup */
10494 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10496 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10497 struct task_group, css);
10500 static struct cgroup_subsys_state *
10501 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10503 struct task_group *tg, *parent;
10505 if (!cgrp->parent) {
10506 /* This is early initialization for the top cgroup */
10507 return &init_task_group.css;
10510 parent = cgroup_tg(cgrp->parent);
10511 tg = sched_create_group(parent);
10513 return ERR_PTR(-ENOMEM);
10519 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10521 struct task_group *tg = cgroup_tg(cgrp);
10523 sched_destroy_group(tg);
10527 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10528 struct task_struct *tsk)
10530 #ifdef CONFIG_RT_GROUP_SCHED
10531 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10534 /* We don't support RT-tasks being in separate groups */
10535 if (tsk->sched_class != &fair_sched_class)
10543 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10544 struct cgroup *old_cont, struct task_struct *tsk)
10546 sched_move_task(tsk);
10549 #ifdef CONFIG_FAIR_GROUP_SCHED
10550 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10553 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10556 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10558 struct task_group *tg = cgroup_tg(cgrp);
10560 return (u64) tg->shares;
10562 #endif /* CONFIG_FAIR_GROUP_SCHED */
10564 #ifdef CONFIG_RT_GROUP_SCHED
10565 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10568 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10571 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10573 return sched_group_rt_runtime(cgroup_tg(cgrp));
10576 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10579 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10582 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10584 return sched_group_rt_period(cgroup_tg(cgrp));
10586 #endif /* CONFIG_RT_GROUP_SCHED */
10588 static struct cftype cpu_files[] = {
10589 #ifdef CONFIG_FAIR_GROUP_SCHED
10592 .read_u64 = cpu_shares_read_u64,
10593 .write_u64 = cpu_shares_write_u64,
10596 #ifdef CONFIG_RT_GROUP_SCHED
10598 .name = "rt_runtime_us",
10599 .read_s64 = cpu_rt_runtime_read,
10600 .write_s64 = cpu_rt_runtime_write,
10603 .name = "rt_period_us",
10604 .read_u64 = cpu_rt_period_read_uint,
10605 .write_u64 = cpu_rt_period_write_uint,
10610 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10612 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10615 struct cgroup_subsys cpu_cgroup_subsys = {
10617 .create = cpu_cgroup_create,
10618 .destroy = cpu_cgroup_destroy,
10619 .can_attach = cpu_cgroup_can_attach,
10620 .attach = cpu_cgroup_attach,
10621 .populate = cpu_cgroup_populate,
10622 .subsys_id = cpu_cgroup_subsys_id,
10626 #endif /* CONFIG_CGROUP_SCHED */
10628 #ifdef CONFIG_CGROUP_CPUACCT
10631 * CPU accounting code for task groups.
10633 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10634 * (balbir@in.ibm.com).
10637 /* track cpu usage of a group of tasks and its child groups */
10639 struct cgroup_subsys_state css;
10640 /* cpuusage holds pointer to a u64-type object on every cpu */
10642 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10643 struct cpuacct *parent;
10646 struct cgroup_subsys cpuacct_subsys;
10648 /* return cpu accounting group corresponding to this container */
10649 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10651 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10652 struct cpuacct, css);
10655 /* return cpu accounting group to which this task belongs */
10656 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10658 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10659 struct cpuacct, css);
10662 /* create a new cpu accounting group */
10663 static struct cgroup_subsys_state *cpuacct_create(
10664 struct cgroup_subsys *ss, struct cgroup *cgrp)
10666 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10672 ca->cpuusage = alloc_percpu(u64);
10676 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10677 if (percpu_counter_init(&ca->cpustat[i], 0))
10678 goto out_free_counters;
10681 ca->parent = cgroup_ca(cgrp->parent);
10687 percpu_counter_destroy(&ca->cpustat[i]);
10688 free_percpu(ca->cpuusage);
10692 return ERR_PTR(-ENOMEM);
10695 /* destroy an existing cpu accounting group */
10697 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10699 struct cpuacct *ca = cgroup_ca(cgrp);
10702 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10703 percpu_counter_destroy(&ca->cpustat[i]);
10704 free_percpu(ca->cpuusage);
10708 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10710 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10713 #ifndef CONFIG_64BIT
10715 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10717 spin_lock_irq(&cpu_rq(cpu)->lock);
10719 spin_unlock_irq(&cpu_rq(cpu)->lock);
10727 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10729 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10731 #ifndef CONFIG_64BIT
10733 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10735 spin_lock_irq(&cpu_rq(cpu)->lock);
10737 spin_unlock_irq(&cpu_rq(cpu)->lock);
10743 /* return total cpu usage (in nanoseconds) of a group */
10744 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10746 struct cpuacct *ca = cgroup_ca(cgrp);
10747 u64 totalcpuusage = 0;
10750 for_each_present_cpu(i)
10751 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10753 return totalcpuusage;
10756 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10759 struct cpuacct *ca = cgroup_ca(cgrp);
10768 for_each_present_cpu(i)
10769 cpuacct_cpuusage_write(ca, i, 0);
10775 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10776 struct seq_file *m)
10778 struct cpuacct *ca = cgroup_ca(cgroup);
10782 for_each_present_cpu(i) {
10783 percpu = cpuacct_cpuusage_read(ca, i);
10784 seq_printf(m, "%llu ", (unsigned long long) percpu);
10786 seq_printf(m, "\n");
10790 static const char *cpuacct_stat_desc[] = {
10791 [CPUACCT_STAT_USER] = "user",
10792 [CPUACCT_STAT_SYSTEM] = "system",
10795 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10796 struct cgroup_map_cb *cb)
10798 struct cpuacct *ca = cgroup_ca(cgrp);
10801 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10802 s64 val = percpu_counter_read(&ca->cpustat[i]);
10803 val = cputime64_to_clock_t(val);
10804 cb->fill(cb, cpuacct_stat_desc[i], val);
10809 static struct cftype files[] = {
10812 .read_u64 = cpuusage_read,
10813 .write_u64 = cpuusage_write,
10816 .name = "usage_percpu",
10817 .read_seq_string = cpuacct_percpu_seq_read,
10821 .read_map = cpuacct_stats_show,
10825 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10827 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10831 * charge this task's execution time to its accounting group.
10833 * called with rq->lock held.
10835 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10837 struct cpuacct *ca;
10840 if (unlikely(!cpuacct_subsys.active))
10843 cpu = task_cpu(tsk);
10849 for (; ca; ca = ca->parent) {
10850 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10851 *cpuusage += cputime;
10858 * Charge the system/user time to the task's accounting group.
10860 static void cpuacct_update_stats(struct task_struct *tsk,
10861 enum cpuacct_stat_index idx, cputime_t val)
10863 struct cpuacct *ca;
10865 if (unlikely(!cpuacct_subsys.active))
10872 percpu_counter_add(&ca->cpustat[idx], val);
10878 struct cgroup_subsys cpuacct_subsys = {
10880 .create = cpuacct_create,
10881 .destroy = cpuacct_destroy,
10882 .populate = cpuacct_populate,
10883 .subsys_id = cpuacct_subsys_id,
10885 #endif /* CONFIG_CGROUP_CPUACCT */