4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
234 if (hrtimer_active(&rt_b->rt_period_timer))
237 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
238 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 hrtimer_start_expires(&rt_b->rt_period_timer,
242 spin_unlock(&rt_b->rt_runtime_lock);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
248 hrtimer_cancel(&rt_b->rt_period_timer);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity **rt_se;
286 struct rt_rq **rt_rq;
288 struct rt_bandwidth rt_bandwidth;
292 struct list_head list;
294 struct task_group *parent;
295 struct list_head siblings;
296 struct list_head children;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct *user)
304 user->tg->uid = user->uid;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
323 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock);
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 #ifdef CONFIG_USER_SCHED
336 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
337 #else /* !CONFIG_USER_SCHED */
338 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
339 #endif /* CONFIG_USER_SCHED */
342 * A weight of 0 or 1 can cause arithmetics problems.
343 * A weight of a cfs_rq is the sum of weights of which entities
344 * are queued on this cfs_rq, so a weight of a entity should not be
345 * too large, so as the shares value of a task group.
346 * (The default weight is 1024 - so there's no practical
347 * limitation from this.)
350 #define MAX_SHARES (1UL << 18)
352 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
355 /* Default task group.
356 * Every task in system belong to this group at bootup.
358 struct task_group init_task_group;
360 /* return group to which a task belongs */
361 static inline struct task_group *task_group(struct task_struct *p)
363 struct task_group *tg;
365 #ifdef CONFIG_USER_SCHED
367 tg = __task_cred(p)->user->tg;
369 #elif defined(CONFIG_CGROUP_SCHED)
370 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
371 struct task_group, css);
373 tg = &init_task_group;
378 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
383 p->se.parent = task_group(p)->se[cpu];
386 #ifdef CONFIG_RT_GROUP_SCHED
387 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
388 p->rt.parent = task_group(p)->rt_se[cpu];
394 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
395 static inline struct task_group *task_group(struct task_struct *p)
400 #endif /* CONFIG_GROUP_SCHED */
402 /* CFS-related fields in a runqueue */
404 struct load_weight load;
405 unsigned long nr_running;
410 struct rb_root tasks_timeline;
411 struct rb_node *rb_leftmost;
413 struct list_head tasks;
414 struct list_head *balance_iterator;
417 * 'curr' points to currently running entity on this cfs_rq.
418 * It is set to NULL otherwise (i.e when none are currently running).
420 struct sched_entity *curr, *next, *last;
422 unsigned int nr_spread_over;
424 #ifdef CONFIG_FAIR_GROUP_SCHED
425 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
428 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
429 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
430 * (like users, containers etc.)
432 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
433 * list is used during load balance.
435 struct list_head leaf_cfs_rq_list;
436 struct task_group *tg; /* group that "owns" this runqueue */
440 * the part of load.weight contributed by tasks
442 unsigned long task_weight;
445 * h_load = weight * f(tg)
447 * Where f(tg) is the recursive weight fraction assigned to
450 unsigned long h_load;
453 * this cpu's part of tg->shares
455 unsigned long shares;
458 * load.weight at the time we set shares
460 unsigned long rq_weight;
465 /* Real-Time classes' related field in a runqueue: */
467 struct rt_prio_array active;
468 unsigned long rt_nr_running;
469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
471 int curr; /* highest queued rt task prio */
473 int next; /* next highest */
478 unsigned long rt_nr_migratory;
480 struct plist_head pushable_tasks;
485 /* Nests inside the rq lock: */
486 spinlock_t rt_runtime_lock;
488 #ifdef CONFIG_RT_GROUP_SCHED
489 unsigned long rt_nr_boosted;
492 struct list_head leaf_rt_rq_list;
493 struct task_group *tg;
494 struct sched_rt_entity *rt_se;
501 * We add the notion of a root-domain which will be used to define per-domain
502 * variables. Each exclusive cpuset essentially defines an island domain by
503 * fully partitioning the member cpus from any other cpuset. Whenever a new
504 * exclusive cpuset is created, we also create and attach a new root-domain
511 cpumask_var_t online;
514 * The "RT overload" flag: it gets set if a CPU has more than
515 * one runnable RT task.
517 cpumask_var_t rto_mask;
520 struct cpupri cpupri;
522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
524 * Preferred wake up cpu nominated by sched_mc balance that will be
525 * used when most cpus are idle in the system indicating overall very
526 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
528 unsigned int sched_mc_preferred_wakeup_cpu;
533 * By default the system creates a single root-domain with all cpus as
534 * members (mimicking the global state we have today).
536 static struct root_domain def_root_domain;
541 * This is the main, per-CPU runqueue data structure.
543 * Locking rule: those places that want to lock multiple runqueues
544 * (such as the load balancing or the thread migration code), lock
545 * acquire operations must be ordered by ascending &runqueue.
552 * nr_running and cpu_load should be in the same cacheline because
553 * remote CPUs use both these fields when doing load calculation.
555 unsigned long nr_running;
556 #define CPU_LOAD_IDX_MAX 5
557 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
559 unsigned long last_tick_seen;
560 unsigned char in_nohz_recently;
562 /* capture load from *all* tasks on this cpu: */
563 struct load_weight load;
564 unsigned long nr_load_updates;
570 #ifdef CONFIG_FAIR_GROUP_SCHED
571 /* list of leaf cfs_rq on this cpu: */
572 struct list_head leaf_cfs_rq_list;
574 #ifdef CONFIG_RT_GROUP_SCHED
575 struct list_head leaf_rt_rq_list;
579 * This is part of a global counter where only the total sum
580 * over all CPUs matters. A task can increase this counter on
581 * one CPU and if it got migrated afterwards it may decrease
582 * it on another CPU. Always updated under the runqueue lock:
584 unsigned long nr_uninterruptible;
586 struct task_struct *curr, *idle;
587 unsigned long next_balance;
588 struct mm_struct *prev_mm;
595 struct root_domain *rd;
596 struct sched_domain *sd;
598 unsigned char idle_at_tick;
599 /* For active balancing */
602 /* cpu of this runqueue: */
606 unsigned long avg_load_per_task;
608 struct task_struct *migration_thread;
609 struct list_head migration_queue;
612 #ifdef CONFIG_SCHED_HRTICK
614 int hrtick_csd_pending;
615 struct call_single_data hrtick_csd;
617 struct hrtimer hrtick_timer;
620 #ifdef CONFIG_SCHEDSTATS
622 struct sched_info rq_sched_info;
623 unsigned long long rq_cpu_time;
624 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
626 /* sys_sched_yield() stats */
627 unsigned int yld_exp_empty;
628 unsigned int yld_act_empty;
629 unsigned int yld_both_empty;
630 unsigned int yld_count;
632 /* schedule() stats */
633 unsigned int sched_switch;
634 unsigned int sched_count;
635 unsigned int sched_goidle;
637 /* try_to_wake_up() stats */
638 unsigned int ttwu_count;
639 unsigned int ttwu_local;
642 unsigned int bkl_count;
646 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
648 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
650 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
653 static inline int cpu_of(struct rq *rq)
663 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
664 * See detach_destroy_domains: synchronize_sched for details.
666 * The domain tree of any CPU may only be accessed from within
667 * preempt-disabled sections.
669 #define for_each_domain(cpu, __sd) \
670 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
672 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
673 #define this_rq() (&__get_cpu_var(runqueues))
674 #define task_rq(p) cpu_rq(task_cpu(p))
675 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
677 static inline void update_rq_clock(struct rq *rq)
679 rq->clock = sched_clock_cpu(cpu_of(rq));
683 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
685 #ifdef CONFIG_SCHED_DEBUG
686 # define const_debug __read_mostly
688 # define const_debug static const
694 * Returns true if the current cpu runqueue is locked.
695 * This interface allows printk to be called with the runqueue lock
696 * held and know whether or not it is OK to wake up the klogd.
698 int runqueue_is_locked(void)
701 struct rq *rq = cpu_rq(cpu);
704 ret = spin_is_locked(&rq->lock);
710 * Debugging: various feature bits
713 #define SCHED_FEAT(name, enabled) \
714 __SCHED_FEAT_##name ,
717 #include "sched_features.h"
722 #define SCHED_FEAT(name, enabled) \
723 (1UL << __SCHED_FEAT_##name) * enabled |
725 const_debug unsigned int sysctl_sched_features =
726 #include "sched_features.h"
731 #ifdef CONFIG_SCHED_DEBUG
732 #define SCHED_FEAT(name, enabled) \
735 static __read_mostly char *sched_feat_names[] = {
736 #include "sched_features.h"
742 static int sched_feat_show(struct seq_file *m, void *v)
746 for (i = 0; sched_feat_names[i]; i++) {
747 if (!(sysctl_sched_features & (1UL << i)))
749 seq_printf(m, "%s ", sched_feat_names[i]);
757 sched_feat_write(struct file *filp, const char __user *ubuf,
758 size_t cnt, loff_t *ppos)
768 if (copy_from_user(&buf, ubuf, cnt))
773 if (strncmp(buf, "NO_", 3) == 0) {
778 for (i = 0; sched_feat_names[i]; i++) {
779 int len = strlen(sched_feat_names[i]);
781 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
783 sysctl_sched_features &= ~(1UL << i);
785 sysctl_sched_features |= (1UL << i);
790 if (!sched_feat_names[i])
798 static int sched_feat_open(struct inode *inode, struct file *filp)
800 return single_open(filp, sched_feat_show, NULL);
803 static struct file_operations sched_feat_fops = {
804 .open = sched_feat_open,
805 .write = sched_feat_write,
808 .release = single_release,
811 static __init int sched_init_debug(void)
813 debugfs_create_file("sched_features", 0644, NULL, NULL,
818 late_initcall(sched_init_debug);
822 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
825 * Number of tasks to iterate in a single balance run.
826 * Limited because this is done with IRQs disabled.
828 const_debug unsigned int sysctl_sched_nr_migrate = 32;
831 * ratelimit for updating the group shares.
834 unsigned int sysctl_sched_shares_ratelimit = 250000;
837 * Inject some fuzzyness into changing the per-cpu group shares
838 * this avoids remote rq-locks at the expense of fairness.
841 unsigned int sysctl_sched_shares_thresh = 4;
844 * period over which we measure -rt task cpu usage in us.
847 unsigned int sysctl_sched_rt_period = 1000000;
849 static __read_mostly int scheduler_running;
852 * part of the period that we allow rt tasks to run in us.
855 int sysctl_sched_rt_runtime = 950000;
857 static inline u64 global_rt_period(void)
859 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
862 static inline u64 global_rt_runtime(void)
864 if (sysctl_sched_rt_runtime < 0)
867 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
870 #ifndef prepare_arch_switch
871 # define prepare_arch_switch(next) do { } while (0)
873 #ifndef finish_arch_switch
874 # define finish_arch_switch(prev) do { } while (0)
877 static inline int task_current(struct rq *rq, struct task_struct *p)
879 return rq->curr == p;
882 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
883 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
888 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
892 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
894 #ifdef CONFIG_DEBUG_SPINLOCK
895 /* this is a valid case when another task releases the spinlock */
896 rq->lock.owner = current;
899 * If we are tracking spinlock dependencies then we have to
900 * fix up the runqueue lock - which gets 'carried over' from
903 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
905 spin_unlock_irq(&rq->lock);
908 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
909 static inline int task_running(struct rq *rq, struct task_struct *p)
914 return task_current(rq, p);
918 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
922 * We can optimise this out completely for !SMP, because the
923 * SMP rebalancing from interrupt is the only thing that cares
928 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 spin_unlock_irq(&rq->lock);
931 spin_unlock(&rq->lock);
935 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
939 * After ->oncpu is cleared, the task can be moved to a different CPU.
940 * We must ensure this doesn't happen until the switch is completely
946 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
950 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
953 * __task_rq_lock - lock the runqueue a given task resides on.
954 * Must be called interrupts disabled.
956 static inline struct rq *__task_rq_lock(struct task_struct *p)
960 struct rq *rq = task_rq(p);
961 spin_lock(&rq->lock);
962 if (likely(rq == task_rq(p)))
964 spin_unlock(&rq->lock);
969 * task_rq_lock - lock the runqueue a given task resides on and disable
970 * interrupts. Note the ordering: we can safely lookup the task_rq without
971 * explicitly disabling preemption.
973 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
979 local_irq_save(*flags);
981 spin_lock(&rq->lock);
982 if (likely(rq == task_rq(p)))
984 spin_unlock_irqrestore(&rq->lock, *flags);
988 void task_rq_unlock_wait(struct task_struct *p)
990 struct rq *rq = task_rq(p);
992 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
993 spin_unlock_wait(&rq->lock);
996 static void __task_rq_unlock(struct rq *rq)
999 spin_unlock(&rq->lock);
1002 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1003 __releases(rq->lock)
1005 spin_unlock_irqrestore(&rq->lock, *flags);
1009 * this_rq_lock - lock this runqueue and disable interrupts.
1011 static struct rq *this_rq_lock(void)
1012 __acquires(rq->lock)
1016 local_irq_disable();
1018 spin_lock(&rq->lock);
1023 #ifdef CONFIG_SCHED_HRTICK
1025 * Use HR-timers to deliver accurate preemption points.
1027 * Its all a bit involved since we cannot program an hrt while holding the
1028 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1031 * When we get rescheduled we reprogram the hrtick_timer outside of the
1037 * - enabled by features
1038 * - hrtimer is actually high res
1040 static inline int hrtick_enabled(struct rq *rq)
1042 if (!sched_feat(HRTICK))
1044 if (!cpu_active(cpu_of(rq)))
1046 return hrtimer_is_hres_active(&rq->hrtick_timer);
1049 static void hrtick_clear(struct rq *rq)
1051 if (hrtimer_active(&rq->hrtick_timer))
1052 hrtimer_cancel(&rq->hrtick_timer);
1056 * High-resolution timer tick.
1057 * Runs from hardirq context with interrupts disabled.
1059 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1061 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1065 spin_lock(&rq->lock);
1066 update_rq_clock(rq);
1067 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1068 spin_unlock(&rq->lock);
1070 return HRTIMER_NORESTART;
1075 * called from hardirq (IPI) context
1077 static void __hrtick_start(void *arg)
1079 struct rq *rq = arg;
1081 spin_lock(&rq->lock);
1082 hrtimer_restart(&rq->hrtick_timer);
1083 rq->hrtick_csd_pending = 0;
1084 spin_unlock(&rq->lock);
1088 * Called to set the hrtick timer state.
1090 * called with rq->lock held and irqs disabled
1092 static void hrtick_start(struct rq *rq, u64 delay)
1094 struct hrtimer *timer = &rq->hrtick_timer;
1095 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1097 hrtimer_set_expires(timer, time);
1099 if (rq == this_rq()) {
1100 hrtimer_restart(timer);
1101 } else if (!rq->hrtick_csd_pending) {
1102 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1103 rq->hrtick_csd_pending = 1;
1108 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1110 int cpu = (int)(long)hcpu;
1113 case CPU_UP_CANCELED:
1114 case CPU_UP_CANCELED_FROZEN:
1115 case CPU_DOWN_PREPARE:
1116 case CPU_DOWN_PREPARE_FROZEN:
1118 case CPU_DEAD_FROZEN:
1119 hrtick_clear(cpu_rq(cpu));
1126 static __init void init_hrtick(void)
1128 hotcpu_notifier(hotplug_hrtick, 0);
1132 * Called to set the hrtick timer state.
1134 * called with rq->lock held and irqs disabled
1136 static void hrtick_start(struct rq *rq, u64 delay)
1138 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SMP */
1146 static void init_rq_hrtick(struct rq *rq)
1149 rq->hrtick_csd_pending = 0;
1151 rq->hrtick_csd.flags = 0;
1152 rq->hrtick_csd.func = __hrtick_start;
1153 rq->hrtick_csd.info = rq;
1156 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1157 rq->hrtick_timer.function = hrtick;
1159 #else /* CONFIG_SCHED_HRTICK */
1160 static inline void hrtick_clear(struct rq *rq)
1164 static inline void init_rq_hrtick(struct rq *rq)
1168 static inline void init_hrtick(void)
1171 #endif /* CONFIG_SCHED_HRTICK */
1174 * resched_task - mark a task 'to be rescheduled now'.
1176 * On UP this means the setting of the need_resched flag, on SMP it
1177 * might also involve a cross-CPU call to trigger the scheduler on
1182 #ifndef tsk_is_polling
1183 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1186 static void resched_task(struct task_struct *p)
1190 assert_spin_locked(&task_rq(p)->lock);
1192 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1195 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1198 if (cpu == smp_processor_id())
1201 /* NEED_RESCHED must be visible before we test polling */
1203 if (!tsk_is_polling(p))
1204 smp_send_reschedule(cpu);
1207 static void resched_cpu(int cpu)
1209 struct rq *rq = cpu_rq(cpu);
1210 unsigned long flags;
1212 if (!spin_trylock_irqsave(&rq->lock, flags))
1214 resched_task(cpu_curr(cpu));
1215 spin_unlock_irqrestore(&rq->lock, flags);
1220 * When add_timer_on() enqueues a timer into the timer wheel of an
1221 * idle CPU then this timer might expire before the next timer event
1222 * which is scheduled to wake up that CPU. In case of a completely
1223 * idle system the next event might even be infinite time into the
1224 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1225 * leaves the inner idle loop so the newly added timer is taken into
1226 * account when the CPU goes back to idle and evaluates the timer
1227 * wheel for the next timer event.
1229 void wake_up_idle_cpu(int cpu)
1231 struct rq *rq = cpu_rq(cpu);
1233 if (cpu == smp_processor_id())
1237 * This is safe, as this function is called with the timer
1238 * wheel base lock of (cpu) held. When the CPU is on the way
1239 * to idle and has not yet set rq->curr to idle then it will
1240 * be serialized on the timer wheel base lock and take the new
1241 * timer into account automatically.
1243 if (rq->curr != rq->idle)
1247 * We can set TIF_RESCHED on the idle task of the other CPU
1248 * lockless. The worst case is that the other CPU runs the
1249 * idle task through an additional NOOP schedule()
1251 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1253 /* NEED_RESCHED must be visible before we test polling */
1255 if (!tsk_is_polling(rq->idle))
1256 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 #else /* !CONFIG_SMP */
1261 static void resched_task(struct task_struct *p)
1263 assert_spin_locked(&task_rq(p)->lock);
1264 set_tsk_need_resched(p);
1266 #endif /* CONFIG_SMP */
1268 #if BITS_PER_LONG == 32
1269 # define WMULT_CONST (~0UL)
1271 # define WMULT_CONST (1UL << 32)
1274 #define WMULT_SHIFT 32
1277 * Shift right and round:
1279 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1282 * delta *= weight / lw
1284 static unsigned long
1285 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1286 struct load_weight *lw)
1290 if (!lw->inv_weight) {
1291 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1294 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1298 tmp = (u64)delta_exec * weight;
1300 * Check whether we'd overflow the 64-bit multiplication:
1302 if (unlikely(tmp > WMULT_CONST))
1303 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1306 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1308 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1311 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1317 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1324 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1325 * of tasks with abnormal "nice" values across CPUs the contribution that
1326 * each task makes to its run queue's load is weighted according to its
1327 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1328 * scaled version of the new time slice allocation that they receive on time
1332 #define WEIGHT_IDLEPRIO 3
1333 #define WMULT_IDLEPRIO 1431655765
1336 * Nice levels are multiplicative, with a gentle 10% change for every
1337 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1338 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1339 * that remained on nice 0.
1341 * The "10% effect" is relative and cumulative: from _any_ nice level,
1342 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1343 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1344 * If a task goes up by ~10% and another task goes down by ~10% then
1345 * the relative distance between them is ~25%.)
1347 static const int prio_to_weight[40] = {
1348 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1349 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1350 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1351 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1352 /* 0 */ 1024, 820, 655, 526, 423,
1353 /* 5 */ 335, 272, 215, 172, 137,
1354 /* 10 */ 110, 87, 70, 56, 45,
1355 /* 15 */ 36, 29, 23, 18, 15,
1359 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1361 * In cases where the weight does not change often, we can use the
1362 * precalculated inverse to speed up arithmetics by turning divisions
1363 * into multiplications:
1365 static const u32 prio_to_wmult[40] = {
1366 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1367 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1368 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1369 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1370 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1371 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1372 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1373 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1376 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1379 * runqueue iterator, to support SMP load-balancing between different
1380 * scheduling classes, without having to expose their internal data
1381 * structures to the load-balancing proper:
1383 struct rq_iterator {
1385 struct task_struct *(*start)(void *);
1386 struct task_struct *(*next)(void *);
1390 static unsigned long
1391 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1392 unsigned long max_load_move, struct sched_domain *sd,
1393 enum cpu_idle_type idle, int *all_pinned,
1394 int *this_best_prio, struct rq_iterator *iterator);
1397 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1398 struct sched_domain *sd, enum cpu_idle_type idle,
1399 struct rq_iterator *iterator);
1402 #ifdef CONFIG_CGROUP_CPUACCT
1403 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1405 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1408 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1410 update_load_add(&rq->load, load);
1413 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1415 update_load_sub(&rq->load, load);
1418 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1419 typedef int (*tg_visitor)(struct task_group *, void *);
1422 * Iterate the full tree, calling @down when first entering a node and @up when
1423 * leaving it for the final time.
1425 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1427 struct task_group *parent, *child;
1431 parent = &root_task_group;
1433 ret = (*down)(parent, data);
1436 list_for_each_entry_rcu(child, &parent->children, siblings) {
1443 ret = (*up)(parent, data);
1448 parent = parent->parent;
1457 static int tg_nop(struct task_group *tg, void *data)
1464 static unsigned long source_load(int cpu, int type);
1465 static unsigned long target_load(int cpu, int type);
1466 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1468 static unsigned long cpu_avg_load_per_task(int cpu)
1470 struct rq *rq = cpu_rq(cpu);
1471 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1474 rq->avg_load_per_task = rq->load.weight / nr_running;
1476 rq->avg_load_per_task = 0;
1478 return rq->avg_load_per_task;
1481 #ifdef CONFIG_FAIR_GROUP_SCHED
1483 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1486 * Calculate and set the cpu's group shares.
1489 update_group_shares_cpu(struct task_group *tg, int cpu,
1490 unsigned long sd_shares, unsigned long sd_rq_weight)
1492 unsigned long shares;
1493 unsigned long rq_weight;
1498 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1501 * \Sum shares * rq_weight
1502 * shares = -----------------------
1506 shares = (sd_shares * rq_weight) / sd_rq_weight;
1507 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1509 if (abs(shares - tg->se[cpu]->load.weight) >
1510 sysctl_sched_shares_thresh) {
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long flags;
1514 spin_lock_irqsave(&rq->lock, flags);
1515 tg->cfs_rq[cpu]->shares = shares;
1517 __set_se_shares(tg->se[cpu], shares);
1518 spin_unlock_irqrestore(&rq->lock, flags);
1523 * Re-compute the task group their per cpu shares over the given domain.
1524 * This needs to be done in a bottom-up fashion because the rq weight of a
1525 * parent group depends on the shares of its child groups.
1527 static int tg_shares_up(struct task_group *tg, void *data)
1529 unsigned long weight, rq_weight = 0;
1530 unsigned long shares = 0;
1531 struct sched_domain *sd = data;
1534 for_each_cpu(i, sched_domain_span(sd)) {
1536 * If there are currently no tasks on the cpu pretend there
1537 * is one of average load so that when a new task gets to
1538 * run here it will not get delayed by group starvation.
1540 weight = tg->cfs_rq[i]->load.weight;
1542 weight = NICE_0_LOAD;
1544 tg->cfs_rq[i]->rq_weight = weight;
1545 rq_weight += weight;
1546 shares += tg->cfs_rq[i]->shares;
1549 if ((!shares && rq_weight) || shares > tg->shares)
1550 shares = tg->shares;
1552 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1553 shares = tg->shares;
1555 for_each_cpu(i, sched_domain_span(sd))
1556 update_group_shares_cpu(tg, i, shares, rq_weight);
1562 * Compute the cpu's hierarchical load factor for each task group.
1563 * This needs to be done in a top-down fashion because the load of a child
1564 * group is a fraction of its parents load.
1566 static int tg_load_down(struct task_group *tg, void *data)
1569 long cpu = (long)data;
1572 load = cpu_rq(cpu)->load.weight;
1574 load = tg->parent->cfs_rq[cpu]->h_load;
1575 load *= tg->cfs_rq[cpu]->shares;
1576 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1579 tg->cfs_rq[cpu]->h_load = load;
1584 static void update_shares(struct sched_domain *sd)
1586 u64 now = cpu_clock(raw_smp_processor_id());
1587 s64 elapsed = now - sd->last_update;
1589 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1590 sd->last_update = now;
1591 walk_tg_tree(tg_nop, tg_shares_up, sd);
1595 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1597 spin_unlock(&rq->lock);
1599 spin_lock(&rq->lock);
1602 static void update_h_load(long cpu)
1604 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1609 static inline void update_shares(struct sched_domain *sd)
1613 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1619 #ifdef CONFIG_PREEMPT
1622 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1623 * way at the expense of forcing extra atomic operations in all
1624 * invocations. This assures that the double_lock is acquired using the
1625 * same underlying policy as the spinlock_t on this architecture, which
1626 * reduces latency compared to the unfair variant below. However, it
1627 * also adds more overhead and therefore may reduce throughput.
1629 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1630 __releases(this_rq->lock)
1631 __acquires(busiest->lock)
1632 __acquires(this_rq->lock)
1634 spin_unlock(&this_rq->lock);
1635 double_rq_lock(this_rq, busiest);
1642 * Unfair double_lock_balance: Optimizes throughput at the expense of
1643 * latency by eliminating extra atomic operations when the locks are
1644 * already in proper order on entry. This favors lower cpu-ids and will
1645 * grant the double lock to lower cpus over higher ids under contention,
1646 * regardless of entry order into the function.
1648 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1649 __releases(this_rq->lock)
1650 __acquires(busiest->lock)
1651 __acquires(this_rq->lock)
1655 if (unlikely(!spin_trylock(&busiest->lock))) {
1656 if (busiest < this_rq) {
1657 spin_unlock(&this_rq->lock);
1658 spin_lock(&busiest->lock);
1659 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1662 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1667 #endif /* CONFIG_PREEMPT */
1670 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1672 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1674 if (unlikely(!irqs_disabled())) {
1675 /* printk() doesn't work good under rq->lock */
1676 spin_unlock(&this_rq->lock);
1680 return _double_lock_balance(this_rq, busiest);
1683 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1684 __releases(busiest->lock)
1686 spin_unlock(&busiest->lock);
1687 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1691 #ifdef CONFIG_FAIR_GROUP_SCHED
1692 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1695 cfs_rq->shares = shares;
1700 #include "sched_stats.h"
1701 #include "sched_idletask.c"
1702 #include "sched_fair.c"
1703 #include "sched_rt.c"
1704 #ifdef CONFIG_SCHED_DEBUG
1705 # include "sched_debug.c"
1708 #define sched_class_highest (&rt_sched_class)
1709 #define for_each_class(class) \
1710 for (class = sched_class_highest; class; class = class->next)
1712 static void inc_nr_running(struct rq *rq)
1717 static void dec_nr_running(struct rq *rq)
1722 static void set_load_weight(struct task_struct *p)
1724 if (task_has_rt_policy(p)) {
1725 p->se.load.weight = prio_to_weight[0] * 2;
1726 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1731 * SCHED_IDLE tasks get minimal weight:
1733 if (p->policy == SCHED_IDLE) {
1734 p->se.load.weight = WEIGHT_IDLEPRIO;
1735 p->se.load.inv_weight = WMULT_IDLEPRIO;
1739 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1740 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1743 static void update_avg(u64 *avg, u64 sample)
1745 s64 diff = sample - *avg;
1749 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1752 p->se.start_runtime = p->se.sum_exec_runtime;
1754 sched_info_queued(p);
1755 p->sched_class->enqueue_task(rq, p, wakeup);
1759 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1762 if (p->se.last_wakeup) {
1763 update_avg(&p->se.avg_overlap,
1764 p->se.sum_exec_runtime - p->se.last_wakeup);
1765 p->se.last_wakeup = 0;
1767 update_avg(&p->se.avg_wakeup,
1768 sysctl_sched_wakeup_granularity);
1772 sched_info_dequeued(p);
1773 p->sched_class->dequeue_task(rq, p, sleep);
1778 * __normal_prio - return the priority that is based on the static prio
1780 static inline int __normal_prio(struct task_struct *p)
1782 return p->static_prio;
1786 * Calculate the expected normal priority: i.e. priority
1787 * without taking RT-inheritance into account. Might be
1788 * boosted by interactivity modifiers. Changes upon fork,
1789 * setprio syscalls, and whenever the interactivity
1790 * estimator recalculates.
1792 static inline int normal_prio(struct task_struct *p)
1796 if (task_has_rt_policy(p))
1797 prio = MAX_RT_PRIO-1 - p->rt_priority;
1799 prio = __normal_prio(p);
1804 * Calculate the current priority, i.e. the priority
1805 * taken into account by the scheduler. This value might
1806 * be boosted by RT tasks, or might be boosted by
1807 * interactivity modifiers. Will be RT if the task got
1808 * RT-boosted. If not then it returns p->normal_prio.
1810 static int effective_prio(struct task_struct *p)
1812 p->normal_prio = normal_prio(p);
1814 * If we are RT tasks or we were boosted to RT priority,
1815 * keep the priority unchanged. Otherwise, update priority
1816 * to the normal priority:
1818 if (!rt_prio(p->prio))
1819 return p->normal_prio;
1824 * activate_task - move a task to the runqueue.
1826 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1828 if (task_contributes_to_load(p))
1829 rq->nr_uninterruptible--;
1831 enqueue_task(rq, p, wakeup);
1836 * deactivate_task - remove a task from the runqueue.
1838 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1840 if (task_contributes_to_load(p))
1841 rq->nr_uninterruptible++;
1843 dequeue_task(rq, p, sleep);
1848 * task_curr - is this task currently executing on a CPU?
1849 * @p: the task in question.
1851 inline int task_curr(const struct task_struct *p)
1853 return cpu_curr(task_cpu(p)) == p;
1856 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1858 set_task_rq(p, cpu);
1861 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1862 * successfuly executed on another CPU. We must ensure that updates of
1863 * per-task data have been completed by this moment.
1866 task_thread_info(p)->cpu = cpu;
1870 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1871 const struct sched_class *prev_class,
1872 int oldprio, int running)
1874 if (prev_class != p->sched_class) {
1875 if (prev_class->switched_from)
1876 prev_class->switched_from(rq, p, running);
1877 p->sched_class->switched_to(rq, p, running);
1879 p->sched_class->prio_changed(rq, p, oldprio, running);
1884 /* Used instead of source_load when we know the type == 0 */
1885 static unsigned long weighted_cpuload(const int cpu)
1887 return cpu_rq(cpu)->load.weight;
1891 * Is this task likely cache-hot:
1894 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1899 * Buddy candidates are cache hot:
1901 if (sched_feat(CACHE_HOT_BUDDY) &&
1902 (&p->se == cfs_rq_of(&p->se)->next ||
1903 &p->se == cfs_rq_of(&p->se)->last))
1906 if (p->sched_class != &fair_sched_class)
1909 if (sysctl_sched_migration_cost == -1)
1911 if (sysctl_sched_migration_cost == 0)
1914 delta = now - p->se.exec_start;
1916 return delta < (s64)sysctl_sched_migration_cost;
1920 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1922 int old_cpu = task_cpu(p);
1923 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1924 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1925 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1928 clock_offset = old_rq->clock - new_rq->clock;
1930 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1932 #ifdef CONFIG_SCHEDSTATS
1933 if (p->se.wait_start)
1934 p->se.wait_start -= clock_offset;
1935 if (p->se.sleep_start)
1936 p->se.sleep_start -= clock_offset;
1937 if (p->se.block_start)
1938 p->se.block_start -= clock_offset;
1939 if (old_cpu != new_cpu) {
1940 schedstat_inc(p, se.nr_migrations);
1941 if (task_hot(p, old_rq->clock, NULL))
1942 schedstat_inc(p, se.nr_forced2_migrations);
1945 p->se.vruntime -= old_cfsrq->min_vruntime -
1946 new_cfsrq->min_vruntime;
1948 __set_task_cpu(p, new_cpu);
1951 struct migration_req {
1952 struct list_head list;
1954 struct task_struct *task;
1957 struct completion done;
1961 * The task's runqueue lock must be held.
1962 * Returns true if you have to wait for migration thread.
1965 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1967 struct rq *rq = task_rq(p);
1970 * If the task is not on a runqueue (and not running), then
1971 * it is sufficient to simply update the task's cpu field.
1973 if (!p->se.on_rq && !task_running(rq, p)) {
1974 set_task_cpu(p, dest_cpu);
1978 init_completion(&req->done);
1980 req->dest_cpu = dest_cpu;
1981 list_add(&req->list, &rq->migration_queue);
1987 * wait_task_inactive - wait for a thread to unschedule.
1989 * If @match_state is nonzero, it's the @p->state value just checked and
1990 * not expected to change. If it changes, i.e. @p might have woken up,
1991 * then return zero. When we succeed in waiting for @p to be off its CPU,
1992 * we return a positive number (its total switch count). If a second call
1993 * a short while later returns the same number, the caller can be sure that
1994 * @p has remained unscheduled the whole time.
1996 * The caller must ensure that the task *will* unschedule sometime soon,
1997 * else this function might spin for a *long* time. This function can't
1998 * be called with interrupts off, or it may introduce deadlock with
1999 * smp_call_function() if an IPI is sent by the same process we are
2000 * waiting to become inactive.
2002 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2004 unsigned long flags;
2011 * We do the initial early heuristics without holding
2012 * any task-queue locks at all. We'll only try to get
2013 * the runqueue lock when things look like they will
2019 * If the task is actively running on another CPU
2020 * still, just relax and busy-wait without holding
2023 * NOTE! Since we don't hold any locks, it's not
2024 * even sure that "rq" stays as the right runqueue!
2025 * But we don't care, since "task_running()" will
2026 * return false if the runqueue has changed and p
2027 * is actually now running somewhere else!
2029 while (task_running(rq, p)) {
2030 if (match_state && unlikely(p->state != match_state))
2036 * Ok, time to look more closely! We need the rq
2037 * lock now, to be *sure*. If we're wrong, we'll
2038 * just go back and repeat.
2040 rq = task_rq_lock(p, &flags);
2041 trace_sched_wait_task(rq, p);
2042 running = task_running(rq, p);
2043 on_rq = p->se.on_rq;
2045 if (!match_state || p->state == match_state)
2046 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2047 task_rq_unlock(rq, &flags);
2050 * If it changed from the expected state, bail out now.
2052 if (unlikely(!ncsw))
2056 * Was it really running after all now that we
2057 * checked with the proper locks actually held?
2059 * Oops. Go back and try again..
2061 if (unlikely(running)) {
2067 * It's not enough that it's not actively running,
2068 * it must be off the runqueue _entirely_, and not
2071 * So if it wa still runnable (but just not actively
2072 * running right now), it's preempted, and we should
2073 * yield - it could be a while.
2075 if (unlikely(on_rq)) {
2076 schedule_timeout_uninterruptible(1);
2081 * Ahh, all good. It wasn't running, and it wasn't
2082 * runnable, which means that it will never become
2083 * running in the future either. We're all done!
2092 * kick_process - kick a running thread to enter/exit the kernel
2093 * @p: the to-be-kicked thread
2095 * Cause a process which is running on another CPU to enter
2096 * kernel-mode, without any delay. (to get signals handled.)
2098 * NOTE: this function doesnt have to take the runqueue lock,
2099 * because all it wants to ensure is that the remote task enters
2100 * the kernel. If the IPI races and the task has been migrated
2101 * to another CPU then no harm is done and the purpose has been
2104 void kick_process(struct task_struct *p)
2110 if ((cpu != smp_processor_id()) && task_curr(p))
2111 smp_send_reschedule(cpu);
2116 * Return a low guess at the load of a migration-source cpu weighted
2117 * according to the scheduling class and "nice" value.
2119 * We want to under-estimate the load of migration sources, to
2120 * balance conservatively.
2122 static unsigned long source_load(int cpu, int type)
2124 struct rq *rq = cpu_rq(cpu);
2125 unsigned long total = weighted_cpuload(cpu);
2127 if (type == 0 || !sched_feat(LB_BIAS))
2130 return min(rq->cpu_load[type-1], total);
2134 * Return a high guess at the load of a migration-target cpu weighted
2135 * according to the scheduling class and "nice" value.
2137 static unsigned long target_load(int cpu, int type)
2139 struct rq *rq = cpu_rq(cpu);
2140 unsigned long total = weighted_cpuload(cpu);
2142 if (type == 0 || !sched_feat(LB_BIAS))
2145 return max(rq->cpu_load[type-1], total);
2149 * find_idlest_group finds and returns the least busy CPU group within the
2152 static struct sched_group *
2153 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2155 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2156 unsigned long min_load = ULONG_MAX, this_load = 0;
2157 int load_idx = sd->forkexec_idx;
2158 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2161 unsigned long load, avg_load;
2165 /* Skip over this group if it has no CPUs allowed */
2166 if (!cpumask_intersects(sched_group_cpus(group),
2170 local_group = cpumask_test_cpu(this_cpu,
2171 sched_group_cpus(group));
2173 /* Tally up the load of all CPUs in the group */
2176 for_each_cpu(i, sched_group_cpus(group)) {
2177 /* Bias balancing toward cpus of our domain */
2179 load = source_load(i, load_idx);
2181 load = target_load(i, load_idx);
2186 /* Adjust by relative CPU power of the group */
2187 avg_load = sg_div_cpu_power(group,
2188 avg_load * SCHED_LOAD_SCALE);
2191 this_load = avg_load;
2193 } else if (avg_load < min_load) {
2194 min_load = avg_load;
2197 } while (group = group->next, group != sd->groups);
2199 if (!idlest || 100*this_load < imbalance*min_load)
2205 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2208 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2210 unsigned long load, min_load = ULONG_MAX;
2214 /* Traverse only the allowed CPUs */
2215 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2216 load = weighted_cpuload(i);
2218 if (load < min_load || (load == min_load && i == this_cpu)) {
2228 * sched_balance_self: balance the current task (running on cpu) in domains
2229 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2232 * Balance, ie. select the least loaded group.
2234 * Returns the target CPU number, or the same CPU if no balancing is needed.
2236 * preempt must be disabled.
2238 static int sched_balance_self(int cpu, int flag)
2240 struct task_struct *t = current;
2241 struct sched_domain *tmp, *sd = NULL;
2243 for_each_domain(cpu, tmp) {
2245 * If power savings logic is enabled for a domain, stop there.
2247 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2249 if (tmp->flags & flag)
2257 struct sched_group *group;
2258 int new_cpu, weight;
2260 if (!(sd->flags & flag)) {
2265 group = find_idlest_group(sd, t, cpu);
2271 new_cpu = find_idlest_cpu(group, t, cpu);
2272 if (new_cpu == -1 || new_cpu == cpu) {
2273 /* Now try balancing at a lower domain level of cpu */
2278 /* Now try balancing at a lower domain level of new_cpu */
2280 weight = cpumask_weight(sched_domain_span(sd));
2282 for_each_domain(cpu, tmp) {
2283 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2285 if (tmp->flags & flag)
2288 /* while loop will break here if sd == NULL */
2294 #endif /* CONFIG_SMP */
2297 * try_to_wake_up - wake up a thread
2298 * @p: the to-be-woken-up thread
2299 * @state: the mask of task states that can be woken
2300 * @sync: do a synchronous wakeup?
2302 * Put it on the run-queue if it's not already there. The "current"
2303 * thread is always on the run-queue (except when the actual
2304 * re-schedule is in progress), and as such you're allowed to do
2305 * the simpler "current->state = TASK_RUNNING" to mark yourself
2306 * runnable without the overhead of this.
2308 * returns failure only if the task is already active.
2310 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2312 int cpu, orig_cpu, this_cpu, success = 0;
2313 unsigned long flags;
2317 if (!sched_feat(SYNC_WAKEUPS))
2321 if (sched_feat(LB_WAKEUP_UPDATE)) {
2322 struct sched_domain *sd;
2324 this_cpu = raw_smp_processor_id();
2327 for_each_domain(this_cpu, sd) {
2328 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2337 rq = task_rq_lock(p, &flags);
2338 update_rq_clock(rq);
2339 old_state = p->state;
2340 if (!(old_state & state))
2348 this_cpu = smp_processor_id();
2351 if (unlikely(task_running(rq, p)))
2354 cpu = p->sched_class->select_task_rq(p, sync);
2355 if (cpu != orig_cpu) {
2356 set_task_cpu(p, cpu);
2357 task_rq_unlock(rq, &flags);
2358 /* might preempt at this point */
2359 rq = task_rq_lock(p, &flags);
2360 old_state = p->state;
2361 if (!(old_state & state))
2366 this_cpu = smp_processor_id();
2370 #ifdef CONFIG_SCHEDSTATS
2371 schedstat_inc(rq, ttwu_count);
2372 if (cpu == this_cpu)
2373 schedstat_inc(rq, ttwu_local);
2375 struct sched_domain *sd;
2376 for_each_domain(this_cpu, sd) {
2377 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2378 schedstat_inc(sd, ttwu_wake_remote);
2383 #endif /* CONFIG_SCHEDSTATS */
2386 #endif /* CONFIG_SMP */
2387 schedstat_inc(p, se.nr_wakeups);
2389 schedstat_inc(p, se.nr_wakeups_sync);
2390 if (orig_cpu != cpu)
2391 schedstat_inc(p, se.nr_wakeups_migrate);
2392 if (cpu == this_cpu)
2393 schedstat_inc(p, se.nr_wakeups_local);
2395 schedstat_inc(p, se.nr_wakeups_remote);
2396 activate_task(rq, p, 1);
2400 * Only attribute actual wakeups done by this task.
2402 if (!in_interrupt()) {
2403 struct sched_entity *se = ¤t->se;
2404 u64 sample = se->sum_exec_runtime;
2406 if (se->last_wakeup)
2407 sample -= se->last_wakeup;
2409 sample -= se->start_runtime;
2410 update_avg(&se->avg_wakeup, sample);
2412 se->last_wakeup = se->sum_exec_runtime;
2416 trace_sched_wakeup(rq, p, success);
2417 check_preempt_curr(rq, p, sync);
2419 p->state = TASK_RUNNING;
2421 if (p->sched_class->task_wake_up)
2422 p->sched_class->task_wake_up(rq, p);
2425 task_rq_unlock(rq, &flags);
2430 int wake_up_process(struct task_struct *p)
2432 return try_to_wake_up(p, TASK_ALL, 0);
2434 EXPORT_SYMBOL(wake_up_process);
2436 int wake_up_state(struct task_struct *p, unsigned int state)
2438 return try_to_wake_up(p, state, 0);
2442 * Perform scheduler related setup for a newly forked process p.
2443 * p is forked by current.
2445 * __sched_fork() is basic setup used by init_idle() too:
2447 static void __sched_fork(struct task_struct *p)
2449 p->se.exec_start = 0;
2450 p->se.sum_exec_runtime = 0;
2451 p->se.prev_sum_exec_runtime = 0;
2452 p->se.last_wakeup = 0;
2453 p->se.avg_overlap = 0;
2454 p->se.start_runtime = 0;
2455 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2457 #ifdef CONFIG_SCHEDSTATS
2458 p->se.wait_start = 0;
2459 p->se.sum_sleep_runtime = 0;
2460 p->se.sleep_start = 0;
2461 p->se.block_start = 0;
2462 p->se.sleep_max = 0;
2463 p->se.block_max = 0;
2465 p->se.slice_max = 0;
2469 INIT_LIST_HEAD(&p->rt.run_list);
2471 INIT_LIST_HEAD(&p->se.group_node);
2473 #ifdef CONFIG_PREEMPT_NOTIFIERS
2474 INIT_HLIST_HEAD(&p->preempt_notifiers);
2478 * We mark the process as running here, but have not actually
2479 * inserted it onto the runqueue yet. This guarantees that
2480 * nobody will actually run it, and a signal or other external
2481 * event cannot wake it up and insert it on the runqueue either.
2483 p->state = TASK_RUNNING;
2487 * fork()/clone()-time setup:
2489 void sched_fork(struct task_struct *p, int clone_flags)
2491 int cpu = get_cpu();
2496 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2498 set_task_cpu(p, cpu);
2501 * Make sure we do not leak PI boosting priority to the child:
2503 p->prio = current->normal_prio;
2504 if (!rt_prio(p->prio))
2505 p->sched_class = &fair_sched_class;
2507 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2508 if (likely(sched_info_on()))
2509 memset(&p->sched_info, 0, sizeof(p->sched_info));
2511 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2514 #ifdef CONFIG_PREEMPT
2515 /* Want to start with kernel preemption disabled. */
2516 task_thread_info(p)->preempt_count = 1;
2518 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2524 * wake_up_new_task - wake up a newly created task for the first time.
2526 * This function will do some initial scheduler statistics housekeeping
2527 * that must be done for every newly created context, then puts the task
2528 * on the runqueue and wakes it.
2530 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2532 unsigned long flags;
2535 rq = task_rq_lock(p, &flags);
2536 BUG_ON(p->state != TASK_RUNNING);
2537 update_rq_clock(rq);
2539 p->prio = effective_prio(p);
2541 if (!p->sched_class->task_new || !current->se.on_rq) {
2542 activate_task(rq, p, 0);
2545 * Let the scheduling class do new task startup
2546 * management (if any):
2548 p->sched_class->task_new(rq, p);
2551 trace_sched_wakeup_new(rq, p, 1);
2552 check_preempt_curr(rq, p, 0);
2554 if (p->sched_class->task_wake_up)
2555 p->sched_class->task_wake_up(rq, p);
2557 task_rq_unlock(rq, &flags);
2560 #ifdef CONFIG_PREEMPT_NOTIFIERS
2563 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2564 * @notifier: notifier struct to register
2566 void preempt_notifier_register(struct preempt_notifier *notifier)
2568 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2570 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2573 * preempt_notifier_unregister - no longer interested in preemption notifications
2574 * @notifier: notifier struct to unregister
2576 * This is safe to call from within a preemption notifier.
2578 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2580 hlist_del(¬ifier->link);
2582 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2584 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2586 struct preempt_notifier *notifier;
2587 struct hlist_node *node;
2589 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2590 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2594 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2595 struct task_struct *next)
2597 struct preempt_notifier *notifier;
2598 struct hlist_node *node;
2600 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2601 notifier->ops->sched_out(notifier, next);
2604 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2606 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2611 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2612 struct task_struct *next)
2616 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2619 * prepare_task_switch - prepare to switch tasks
2620 * @rq: the runqueue preparing to switch
2621 * @prev: the current task that is being switched out
2622 * @next: the task we are going to switch to.
2624 * This is called with the rq lock held and interrupts off. It must
2625 * be paired with a subsequent finish_task_switch after the context
2628 * prepare_task_switch sets up locking and calls architecture specific
2632 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2633 struct task_struct *next)
2635 fire_sched_out_preempt_notifiers(prev, next);
2636 prepare_lock_switch(rq, next);
2637 prepare_arch_switch(next);
2641 * finish_task_switch - clean up after a task-switch
2642 * @rq: runqueue associated with task-switch
2643 * @prev: the thread we just switched away from.
2645 * finish_task_switch must be called after the context switch, paired
2646 * with a prepare_task_switch call before the context switch.
2647 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2648 * and do any other architecture-specific cleanup actions.
2650 * Note that we may have delayed dropping an mm in context_switch(). If
2651 * so, we finish that here outside of the runqueue lock. (Doing it
2652 * with the lock held can cause deadlocks; see schedule() for
2655 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2656 __releases(rq->lock)
2658 struct mm_struct *mm = rq->prev_mm;
2661 int post_schedule = 0;
2663 if (current->sched_class->needs_post_schedule)
2664 post_schedule = current->sched_class->needs_post_schedule(rq);
2670 * A task struct has one reference for the use as "current".
2671 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2672 * schedule one last time. The schedule call will never return, and
2673 * the scheduled task must drop that reference.
2674 * The test for TASK_DEAD must occur while the runqueue locks are
2675 * still held, otherwise prev could be scheduled on another cpu, die
2676 * there before we look at prev->state, and then the reference would
2678 * Manfred Spraul <manfred@colorfullife.com>
2680 prev_state = prev->state;
2681 finish_arch_switch(prev);
2682 finish_lock_switch(rq, prev);
2685 current->sched_class->post_schedule(rq);
2688 fire_sched_in_preempt_notifiers(current);
2691 if (unlikely(prev_state == TASK_DEAD)) {
2693 * Remove function-return probe instances associated with this
2694 * task and put them back on the free list.
2696 kprobe_flush_task(prev);
2697 put_task_struct(prev);
2702 * schedule_tail - first thing a freshly forked thread must call.
2703 * @prev: the thread we just switched away from.
2705 asmlinkage void schedule_tail(struct task_struct *prev)
2706 __releases(rq->lock)
2708 struct rq *rq = this_rq();
2710 finish_task_switch(rq, prev);
2711 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2712 /* In this case, finish_task_switch does not reenable preemption */
2715 if (current->set_child_tid)
2716 put_user(task_pid_vnr(current), current->set_child_tid);
2720 * context_switch - switch to the new MM and the new
2721 * thread's register state.
2724 context_switch(struct rq *rq, struct task_struct *prev,
2725 struct task_struct *next)
2727 struct mm_struct *mm, *oldmm;
2729 prepare_task_switch(rq, prev, next);
2730 trace_sched_switch(rq, prev, next);
2732 oldmm = prev->active_mm;
2734 * For paravirt, this is coupled with an exit in switch_to to
2735 * combine the page table reload and the switch backend into
2738 arch_enter_lazy_cpu_mode();
2740 if (unlikely(!mm)) {
2741 next->active_mm = oldmm;
2742 atomic_inc(&oldmm->mm_count);
2743 enter_lazy_tlb(oldmm, next);
2745 switch_mm(oldmm, mm, next);
2747 if (unlikely(!prev->mm)) {
2748 prev->active_mm = NULL;
2749 rq->prev_mm = oldmm;
2752 * Since the runqueue lock will be released by the next
2753 * task (which is an invalid locking op but in the case
2754 * of the scheduler it's an obvious special-case), so we
2755 * do an early lockdep release here:
2757 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2758 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2761 /* Here we just switch the register state and the stack. */
2762 switch_to(prev, next, prev);
2766 * this_rq must be evaluated again because prev may have moved
2767 * CPUs since it called schedule(), thus the 'rq' on its stack
2768 * frame will be invalid.
2770 finish_task_switch(this_rq(), prev);
2774 * nr_running, nr_uninterruptible and nr_context_switches:
2776 * externally visible scheduler statistics: current number of runnable
2777 * threads, current number of uninterruptible-sleeping threads, total
2778 * number of context switches performed since bootup.
2780 unsigned long nr_running(void)
2782 unsigned long i, sum = 0;
2784 for_each_online_cpu(i)
2785 sum += cpu_rq(i)->nr_running;
2790 unsigned long nr_uninterruptible(void)
2792 unsigned long i, sum = 0;
2794 for_each_possible_cpu(i)
2795 sum += cpu_rq(i)->nr_uninterruptible;
2798 * Since we read the counters lockless, it might be slightly
2799 * inaccurate. Do not allow it to go below zero though:
2801 if (unlikely((long)sum < 0))
2807 unsigned long long nr_context_switches(void)
2810 unsigned long long sum = 0;
2812 for_each_possible_cpu(i)
2813 sum += cpu_rq(i)->nr_switches;
2818 unsigned long nr_iowait(void)
2820 unsigned long i, sum = 0;
2822 for_each_possible_cpu(i)
2823 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2828 unsigned long nr_active(void)
2830 unsigned long i, running = 0, uninterruptible = 0;
2832 for_each_online_cpu(i) {
2833 running += cpu_rq(i)->nr_running;
2834 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2837 if (unlikely((long)uninterruptible < 0))
2838 uninterruptible = 0;
2840 return running + uninterruptible;
2844 * Update rq->cpu_load[] statistics. This function is usually called every
2845 * scheduler tick (TICK_NSEC).
2847 static void update_cpu_load(struct rq *this_rq)
2849 unsigned long this_load = this_rq->load.weight;
2852 this_rq->nr_load_updates++;
2854 /* Update our load: */
2855 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2856 unsigned long old_load, new_load;
2858 /* scale is effectively 1 << i now, and >> i divides by scale */
2860 old_load = this_rq->cpu_load[i];
2861 new_load = this_load;
2863 * Round up the averaging division if load is increasing. This
2864 * prevents us from getting stuck on 9 if the load is 10, for
2867 if (new_load > old_load)
2868 new_load += scale-1;
2869 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2876 * double_rq_lock - safely lock two runqueues
2878 * Note this does not disable interrupts like task_rq_lock,
2879 * you need to do so manually before calling.
2881 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2882 __acquires(rq1->lock)
2883 __acquires(rq2->lock)
2885 BUG_ON(!irqs_disabled());
2887 spin_lock(&rq1->lock);
2888 __acquire(rq2->lock); /* Fake it out ;) */
2891 spin_lock(&rq1->lock);
2892 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2894 spin_lock(&rq2->lock);
2895 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2898 update_rq_clock(rq1);
2899 update_rq_clock(rq2);
2903 * double_rq_unlock - safely unlock two runqueues
2905 * Note this does not restore interrupts like task_rq_unlock,
2906 * you need to do so manually after calling.
2908 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2909 __releases(rq1->lock)
2910 __releases(rq2->lock)
2912 spin_unlock(&rq1->lock);
2914 spin_unlock(&rq2->lock);
2916 __release(rq2->lock);
2920 * If dest_cpu is allowed for this process, migrate the task to it.
2921 * This is accomplished by forcing the cpu_allowed mask to only
2922 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2923 * the cpu_allowed mask is restored.
2925 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2927 struct migration_req req;
2928 unsigned long flags;
2931 rq = task_rq_lock(p, &flags);
2932 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2933 || unlikely(!cpu_active(dest_cpu)))
2936 /* force the process onto the specified CPU */
2937 if (migrate_task(p, dest_cpu, &req)) {
2938 /* Need to wait for migration thread (might exit: take ref). */
2939 struct task_struct *mt = rq->migration_thread;
2941 get_task_struct(mt);
2942 task_rq_unlock(rq, &flags);
2943 wake_up_process(mt);
2944 put_task_struct(mt);
2945 wait_for_completion(&req.done);
2950 task_rq_unlock(rq, &flags);
2954 * sched_exec - execve() is a valuable balancing opportunity, because at
2955 * this point the task has the smallest effective memory and cache footprint.
2957 void sched_exec(void)
2959 int new_cpu, this_cpu = get_cpu();
2960 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2962 if (new_cpu != this_cpu)
2963 sched_migrate_task(current, new_cpu);
2967 * pull_task - move a task from a remote runqueue to the local runqueue.
2968 * Both runqueues must be locked.
2970 static void pull_task(struct rq *src_rq, struct task_struct *p,
2971 struct rq *this_rq, int this_cpu)
2973 deactivate_task(src_rq, p, 0);
2974 set_task_cpu(p, this_cpu);
2975 activate_task(this_rq, p, 0);
2977 * Note that idle threads have a prio of MAX_PRIO, for this test
2978 * to be always true for them.
2980 check_preempt_curr(this_rq, p, 0);
2984 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2987 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2988 struct sched_domain *sd, enum cpu_idle_type idle,
2992 * We do not migrate tasks that are:
2993 * 1) running (obviously), or
2994 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2995 * 3) are cache-hot on their current CPU.
2997 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2998 schedstat_inc(p, se.nr_failed_migrations_affine);
3003 if (task_running(rq, p)) {
3004 schedstat_inc(p, se.nr_failed_migrations_running);
3009 * Aggressive migration if:
3010 * 1) task is cache cold, or
3011 * 2) too many balance attempts have failed.
3014 if (!task_hot(p, rq->clock, sd) ||
3015 sd->nr_balance_failed > sd->cache_nice_tries) {
3016 #ifdef CONFIG_SCHEDSTATS
3017 if (task_hot(p, rq->clock, sd)) {
3018 schedstat_inc(sd, lb_hot_gained[idle]);
3019 schedstat_inc(p, se.nr_forced_migrations);
3025 if (task_hot(p, rq->clock, sd)) {
3026 schedstat_inc(p, se.nr_failed_migrations_hot);
3032 static unsigned long
3033 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3034 unsigned long max_load_move, struct sched_domain *sd,
3035 enum cpu_idle_type idle, int *all_pinned,
3036 int *this_best_prio, struct rq_iterator *iterator)
3038 int loops = 0, pulled = 0, pinned = 0;
3039 struct task_struct *p;
3040 long rem_load_move = max_load_move;
3042 if (max_load_move == 0)
3048 * Start the load-balancing iterator:
3050 p = iterator->start(iterator->arg);
3052 if (!p || loops++ > sysctl_sched_nr_migrate)
3055 if ((p->se.load.weight >> 1) > rem_load_move ||
3056 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3057 p = iterator->next(iterator->arg);
3061 pull_task(busiest, p, this_rq, this_cpu);
3063 rem_load_move -= p->se.load.weight;
3065 #ifdef CONFIG_PREEMPT
3067 * NEWIDLE balancing is a source of latency, so preemptible kernels
3068 * will stop after the first task is pulled to minimize the critical
3071 if (idle == CPU_NEWLY_IDLE)
3076 * We only want to steal up to the prescribed amount of weighted load.
3078 if (rem_load_move > 0) {
3079 if (p->prio < *this_best_prio)
3080 *this_best_prio = p->prio;
3081 p = iterator->next(iterator->arg);
3086 * Right now, this is one of only two places pull_task() is called,
3087 * so we can safely collect pull_task() stats here rather than
3088 * inside pull_task().
3090 schedstat_add(sd, lb_gained[idle], pulled);
3093 *all_pinned = pinned;
3095 return max_load_move - rem_load_move;
3099 * move_tasks tries to move up to max_load_move weighted load from busiest to
3100 * this_rq, as part of a balancing operation within domain "sd".
3101 * Returns 1 if successful and 0 otherwise.
3103 * Called with both runqueues locked.
3105 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3106 unsigned long max_load_move,
3107 struct sched_domain *sd, enum cpu_idle_type idle,
3110 const struct sched_class *class = sched_class_highest;
3111 unsigned long total_load_moved = 0;
3112 int this_best_prio = this_rq->curr->prio;
3116 class->load_balance(this_rq, this_cpu, busiest,
3117 max_load_move - total_load_moved,
3118 sd, idle, all_pinned, &this_best_prio);
3119 class = class->next;
3121 #ifdef CONFIG_PREEMPT
3123 * NEWIDLE balancing is a source of latency, so preemptible
3124 * kernels will stop after the first task is pulled to minimize
3125 * the critical section.
3127 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3130 } while (class && max_load_move > total_load_moved);
3132 return total_load_moved > 0;
3136 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3137 struct sched_domain *sd, enum cpu_idle_type idle,
3138 struct rq_iterator *iterator)
3140 struct task_struct *p = iterator->start(iterator->arg);
3144 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3145 pull_task(busiest, p, this_rq, this_cpu);
3147 * Right now, this is only the second place pull_task()
3148 * is called, so we can safely collect pull_task()
3149 * stats here rather than inside pull_task().
3151 schedstat_inc(sd, lb_gained[idle]);
3155 p = iterator->next(iterator->arg);
3162 * move_one_task tries to move exactly one task from busiest to this_rq, as
3163 * part of active balancing operations within "domain".
3164 * Returns 1 if successful and 0 otherwise.
3166 * Called with both runqueues locked.
3168 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3169 struct sched_domain *sd, enum cpu_idle_type idle)
3171 const struct sched_class *class;
3173 for (class = sched_class_highest; class; class = class->next)
3174 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3181 * find_busiest_group finds and returns the busiest CPU group within the
3182 * domain. It calculates and returns the amount of weighted load which
3183 * should be moved to restore balance via the imbalance parameter.
3185 static struct sched_group *
3186 find_busiest_group(struct sched_domain *sd, int this_cpu,
3187 unsigned long *imbalance, enum cpu_idle_type idle,
3188 int *sd_idle, const struct cpumask *cpus, int *balance)
3190 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3191 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3192 unsigned long max_pull;
3193 unsigned long busiest_load_per_task, busiest_nr_running;
3194 unsigned long this_load_per_task, this_nr_running;
3195 int load_idx, group_imb = 0;
3196 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3197 int power_savings_balance = 1;
3198 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3199 unsigned long min_nr_running = ULONG_MAX;
3200 struct sched_group *group_min = NULL, *group_leader = NULL;
3203 max_load = this_load = total_load = total_pwr = 0;
3204 busiest_load_per_task = busiest_nr_running = 0;
3205 this_load_per_task = this_nr_running = 0;
3207 if (idle == CPU_NOT_IDLE)
3208 load_idx = sd->busy_idx;
3209 else if (idle == CPU_NEWLY_IDLE)
3210 load_idx = sd->newidle_idx;
3212 load_idx = sd->idle_idx;
3215 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3218 int __group_imb = 0;
3219 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3220 unsigned long sum_nr_running, sum_weighted_load;
3221 unsigned long sum_avg_load_per_task;
3222 unsigned long avg_load_per_task;
3224 local_group = cpumask_test_cpu(this_cpu,
3225 sched_group_cpus(group));
3228 balance_cpu = cpumask_first(sched_group_cpus(group));
3230 /* Tally up the load of all CPUs in the group */
3231 sum_weighted_load = sum_nr_running = avg_load = 0;
3232 sum_avg_load_per_task = avg_load_per_task = 0;
3235 min_cpu_load = ~0UL;
3237 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3238 struct rq *rq = cpu_rq(i);
3240 if (*sd_idle && rq->nr_running)
3243 /* Bias balancing toward cpus of our domain */
3245 if (idle_cpu(i) && !first_idle_cpu) {
3250 load = target_load(i, load_idx);
3252 load = source_load(i, load_idx);
3253 if (load > max_cpu_load)
3254 max_cpu_load = load;
3255 if (min_cpu_load > load)
3256 min_cpu_load = load;
3260 sum_nr_running += rq->nr_running;
3261 sum_weighted_load += weighted_cpuload(i);
3263 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3267 * First idle cpu or the first cpu(busiest) in this sched group
3268 * is eligible for doing load balancing at this and above
3269 * domains. In the newly idle case, we will allow all the cpu's
3270 * to do the newly idle load balance.
3272 if (idle != CPU_NEWLY_IDLE && local_group &&
3273 balance_cpu != this_cpu && balance) {
3278 total_load += avg_load;
3279 total_pwr += group->__cpu_power;
3281 /* Adjust by relative CPU power of the group */
3282 avg_load = sg_div_cpu_power(group,
3283 avg_load * SCHED_LOAD_SCALE);
3287 * Consider the group unbalanced when the imbalance is larger
3288 * than the average weight of two tasks.
3290 * APZ: with cgroup the avg task weight can vary wildly and
3291 * might not be a suitable number - should we keep a
3292 * normalized nr_running number somewhere that negates
3295 avg_load_per_task = sg_div_cpu_power(group,
3296 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3298 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3301 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3304 this_load = avg_load;
3306 this_nr_running = sum_nr_running;
3307 this_load_per_task = sum_weighted_load;
3308 } else if (avg_load > max_load &&
3309 (sum_nr_running > group_capacity || __group_imb)) {
3310 max_load = avg_load;
3312 busiest_nr_running = sum_nr_running;
3313 busiest_load_per_task = sum_weighted_load;
3314 group_imb = __group_imb;
3317 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3319 * Busy processors will not participate in power savings
3322 if (idle == CPU_NOT_IDLE ||
3323 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3327 * If the local group is idle or completely loaded
3328 * no need to do power savings balance at this domain
3330 if (local_group && (this_nr_running >= group_capacity ||
3332 power_savings_balance = 0;
3335 * If a group is already running at full capacity or idle,
3336 * don't include that group in power savings calculations
3338 if (!power_savings_balance || sum_nr_running >= group_capacity
3343 * Calculate the group which has the least non-idle load.
3344 * This is the group from where we need to pick up the load
3347 if ((sum_nr_running < min_nr_running) ||
3348 (sum_nr_running == min_nr_running &&
3349 cpumask_first(sched_group_cpus(group)) >
3350 cpumask_first(sched_group_cpus(group_min)))) {
3352 min_nr_running = sum_nr_running;
3353 min_load_per_task = sum_weighted_load /
3358 * Calculate the group which is almost near its
3359 * capacity but still has some space to pick up some load
3360 * from other group and save more power
3362 if (sum_nr_running <= group_capacity - 1) {
3363 if (sum_nr_running > leader_nr_running ||
3364 (sum_nr_running == leader_nr_running &&
3365 cpumask_first(sched_group_cpus(group)) <
3366 cpumask_first(sched_group_cpus(group_leader)))) {
3367 group_leader = group;
3368 leader_nr_running = sum_nr_running;
3373 group = group->next;
3374 } while (group != sd->groups);
3376 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3379 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3381 if (this_load >= avg_load ||
3382 100*max_load <= sd->imbalance_pct*this_load)
3385 busiest_load_per_task /= busiest_nr_running;
3387 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3390 * We're trying to get all the cpus to the average_load, so we don't
3391 * want to push ourselves above the average load, nor do we wish to
3392 * reduce the max loaded cpu below the average load, as either of these
3393 * actions would just result in more rebalancing later, and ping-pong
3394 * tasks around. Thus we look for the minimum possible imbalance.
3395 * Negative imbalances (*we* are more loaded than anyone else) will
3396 * be counted as no imbalance for these purposes -- we can't fix that
3397 * by pulling tasks to us. Be careful of negative numbers as they'll
3398 * appear as very large values with unsigned longs.
3400 if (max_load <= busiest_load_per_task)
3404 * In the presence of smp nice balancing, certain scenarios can have
3405 * max load less than avg load(as we skip the groups at or below
3406 * its cpu_power, while calculating max_load..)
3408 if (max_load < avg_load) {
3410 goto small_imbalance;
3413 /* Don't want to pull so many tasks that a group would go idle */
3414 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3416 /* How much load to actually move to equalise the imbalance */
3417 *imbalance = min(max_pull * busiest->__cpu_power,
3418 (avg_load - this_load) * this->__cpu_power)
3422 * if *imbalance is less than the average load per runnable task
3423 * there is no gaurantee that any tasks will be moved so we'll have
3424 * a think about bumping its value to force at least one task to be
3427 if (*imbalance < busiest_load_per_task) {
3428 unsigned long tmp, pwr_now, pwr_move;
3432 pwr_move = pwr_now = 0;
3434 if (this_nr_running) {
3435 this_load_per_task /= this_nr_running;
3436 if (busiest_load_per_task > this_load_per_task)
3439 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3441 if (max_load - this_load + busiest_load_per_task >=
3442 busiest_load_per_task * imbn) {
3443 *imbalance = busiest_load_per_task;
3448 * OK, we don't have enough imbalance to justify moving tasks,
3449 * however we may be able to increase total CPU power used by
3453 pwr_now += busiest->__cpu_power *
3454 min(busiest_load_per_task, max_load);
3455 pwr_now += this->__cpu_power *
3456 min(this_load_per_task, this_load);
3457 pwr_now /= SCHED_LOAD_SCALE;
3459 /* Amount of load we'd subtract */
3460 tmp = sg_div_cpu_power(busiest,
3461 busiest_load_per_task * SCHED_LOAD_SCALE);
3463 pwr_move += busiest->__cpu_power *
3464 min(busiest_load_per_task, max_load - tmp);
3466 /* Amount of load we'd add */
3467 if (max_load * busiest->__cpu_power <
3468 busiest_load_per_task * SCHED_LOAD_SCALE)
3469 tmp = sg_div_cpu_power(this,
3470 max_load * busiest->__cpu_power);
3472 tmp = sg_div_cpu_power(this,
3473 busiest_load_per_task * SCHED_LOAD_SCALE);
3474 pwr_move += this->__cpu_power *
3475 min(this_load_per_task, this_load + tmp);
3476 pwr_move /= SCHED_LOAD_SCALE;
3478 /* Move if we gain throughput */
3479 if (pwr_move > pwr_now)
3480 *imbalance = busiest_load_per_task;
3486 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3487 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3490 if (this == group_leader && group_leader != group_min) {
3491 *imbalance = min_load_per_task;
3492 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3493 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3494 cpumask_first(sched_group_cpus(group_leader));
3505 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3508 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3509 unsigned long imbalance, const struct cpumask *cpus)
3511 struct rq *busiest = NULL, *rq;
3512 unsigned long max_load = 0;
3515 for_each_cpu(i, sched_group_cpus(group)) {
3518 if (!cpumask_test_cpu(i, cpus))
3522 wl = weighted_cpuload(i);
3524 if (rq->nr_running == 1 && wl > imbalance)
3527 if (wl > max_load) {
3537 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3538 * so long as it is large enough.
3540 #define MAX_PINNED_INTERVAL 512
3543 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3544 * tasks if there is an imbalance.
3546 static int load_balance(int this_cpu, struct rq *this_rq,
3547 struct sched_domain *sd, enum cpu_idle_type idle,
3548 int *balance, struct cpumask *cpus)
3550 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3551 struct sched_group *group;
3552 unsigned long imbalance;
3554 unsigned long flags;
3556 cpumask_setall(cpus);
3559 * When power savings policy is enabled for the parent domain, idle
3560 * sibling can pick up load irrespective of busy siblings. In this case,
3561 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3562 * portraying it as CPU_NOT_IDLE.
3564 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3565 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3568 schedstat_inc(sd, lb_count[idle]);
3572 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3579 schedstat_inc(sd, lb_nobusyg[idle]);
3583 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3585 schedstat_inc(sd, lb_nobusyq[idle]);
3589 BUG_ON(busiest == this_rq);
3591 schedstat_add(sd, lb_imbalance[idle], imbalance);
3594 if (busiest->nr_running > 1) {
3596 * Attempt to move tasks. If find_busiest_group has found
3597 * an imbalance but busiest->nr_running <= 1, the group is
3598 * still unbalanced. ld_moved simply stays zero, so it is
3599 * correctly treated as an imbalance.
3601 local_irq_save(flags);
3602 double_rq_lock(this_rq, busiest);
3603 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3604 imbalance, sd, idle, &all_pinned);
3605 double_rq_unlock(this_rq, busiest);
3606 local_irq_restore(flags);
3609 * some other cpu did the load balance for us.
3611 if (ld_moved && this_cpu != smp_processor_id())
3612 resched_cpu(this_cpu);
3614 /* All tasks on this runqueue were pinned by CPU affinity */
3615 if (unlikely(all_pinned)) {
3616 cpumask_clear_cpu(cpu_of(busiest), cpus);
3617 if (!cpumask_empty(cpus))
3624 schedstat_inc(sd, lb_failed[idle]);
3625 sd->nr_balance_failed++;
3627 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3629 spin_lock_irqsave(&busiest->lock, flags);
3631 /* don't kick the migration_thread, if the curr
3632 * task on busiest cpu can't be moved to this_cpu
3634 if (!cpumask_test_cpu(this_cpu,
3635 &busiest->curr->cpus_allowed)) {
3636 spin_unlock_irqrestore(&busiest->lock, flags);
3638 goto out_one_pinned;
3641 if (!busiest->active_balance) {
3642 busiest->active_balance = 1;
3643 busiest->push_cpu = this_cpu;
3646 spin_unlock_irqrestore(&busiest->lock, flags);
3648 wake_up_process(busiest->migration_thread);
3651 * We've kicked active balancing, reset the failure
3654 sd->nr_balance_failed = sd->cache_nice_tries+1;
3657 sd->nr_balance_failed = 0;
3659 if (likely(!active_balance)) {
3660 /* We were unbalanced, so reset the balancing interval */
3661 sd->balance_interval = sd->min_interval;
3664 * If we've begun active balancing, start to back off. This
3665 * case may not be covered by the all_pinned logic if there
3666 * is only 1 task on the busy runqueue (because we don't call
3669 if (sd->balance_interval < sd->max_interval)
3670 sd->balance_interval *= 2;
3673 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3674 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3680 schedstat_inc(sd, lb_balanced[idle]);
3682 sd->nr_balance_failed = 0;
3685 /* tune up the balancing interval */
3686 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3687 (sd->balance_interval < sd->max_interval))
3688 sd->balance_interval *= 2;
3690 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3691 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3702 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3703 * tasks if there is an imbalance.
3705 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3706 * this_rq is locked.
3709 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3710 struct cpumask *cpus)
3712 struct sched_group *group;
3713 struct rq *busiest = NULL;
3714 unsigned long imbalance;
3719 cpumask_setall(cpus);
3722 * When power savings policy is enabled for the parent domain, idle
3723 * sibling can pick up load irrespective of busy siblings. In this case,
3724 * let the state of idle sibling percolate up as IDLE, instead of
3725 * portraying it as CPU_NOT_IDLE.
3727 if (sd->flags & SD_SHARE_CPUPOWER &&
3728 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3731 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3733 update_shares_locked(this_rq, sd);
3734 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3735 &sd_idle, cpus, NULL);
3737 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3741 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3743 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3747 BUG_ON(busiest == this_rq);
3749 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3752 if (busiest->nr_running > 1) {
3753 /* Attempt to move tasks */
3754 double_lock_balance(this_rq, busiest);
3755 /* this_rq->clock is already updated */
3756 update_rq_clock(busiest);
3757 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3758 imbalance, sd, CPU_NEWLY_IDLE,
3760 double_unlock_balance(this_rq, busiest);
3762 if (unlikely(all_pinned)) {
3763 cpumask_clear_cpu(cpu_of(busiest), cpus);
3764 if (!cpumask_empty(cpus))
3770 int active_balance = 0;
3772 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3773 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3774 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3777 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3780 if (sd->nr_balance_failed++ < 2)
3784 * The only task running in a non-idle cpu can be moved to this
3785 * cpu in an attempt to completely freeup the other CPU
3786 * package. The same method used to move task in load_balance()
3787 * have been extended for load_balance_newidle() to speedup
3788 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3790 * The package power saving logic comes from
3791 * find_busiest_group(). If there are no imbalance, then
3792 * f_b_g() will return NULL. However when sched_mc={1,2} then
3793 * f_b_g() will select a group from which a running task may be
3794 * pulled to this cpu in order to make the other package idle.
3795 * If there is no opportunity to make a package idle and if
3796 * there are no imbalance, then f_b_g() will return NULL and no
3797 * action will be taken in load_balance_newidle().
3799 * Under normal task pull operation due to imbalance, there
3800 * will be more than one task in the source run queue and
3801 * move_tasks() will succeed. ld_moved will be true and this
3802 * active balance code will not be triggered.
3805 /* Lock busiest in correct order while this_rq is held */
3806 double_lock_balance(this_rq, busiest);
3809 * don't kick the migration_thread, if the curr
3810 * task on busiest cpu can't be moved to this_cpu
3812 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3813 double_unlock_balance(this_rq, busiest);
3818 if (!busiest->active_balance) {
3819 busiest->active_balance = 1;
3820 busiest->push_cpu = this_cpu;
3824 double_unlock_balance(this_rq, busiest);
3826 * Should not call ttwu while holding a rq->lock
3828 spin_unlock(&this_rq->lock);
3830 wake_up_process(busiest->migration_thread);
3831 spin_lock(&this_rq->lock);
3834 sd->nr_balance_failed = 0;
3836 update_shares_locked(this_rq, sd);
3840 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3841 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3842 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3844 sd->nr_balance_failed = 0;
3850 * idle_balance is called by schedule() if this_cpu is about to become
3851 * idle. Attempts to pull tasks from other CPUs.
3853 static void idle_balance(int this_cpu, struct rq *this_rq)
3855 struct sched_domain *sd;
3856 int pulled_task = 0;
3857 unsigned long next_balance = jiffies + HZ;
3858 cpumask_var_t tmpmask;
3860 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3863 for_each_domain(this_cpu, sd) {
3864 unsigned long interval;
3866 if (!(sd->flags & SD_LOAD_BALANCE))
3869 if (sd->flags & SD_BALANCE_NEWIDLE)
3870 /* If we've pulled tasks over stop searching: */
3871 pulled_task = load_balance_newidle(this_cpu, this_rq,
3874 interval = msecs_to_jiffies(sd->balance_interval);
3875 if (time_after(next_balance, sd->last_balance + interval))
3876 next_balance = sd->last_balance + interval;
3880 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3882 * We are going idle. next_balance may be set based on
3883 * a busy processor. So reset next_balance.
3885 this_rq->next_balance = next_balance;
3887 free_cpumask_var(tmpmask);
3891 * active_load_balance is run by migration threads. It pushes running tasks
3892 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3893 * running on each physical CPU where possible, and avoids physical /
3894 * logical imbalances.
3896 * Called with busiest_rq locked.
3898 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3900 int target_cpu = busiest_rq->push_cpu;
3901 struct sched_domain *sd;
3902 struct rq *target_rq;
3904 /* Is there any task to move? */
3905 if (busiest_rq->nr_running <= 1)
3908 target_rq = cpu_rq(target_cpu);
3911 * This condition is "impossible", if it occurs
3912 * we need to fix it. Originally reported by
3913 * Bjorn Helgaas on a 128-cpu setup.
3915 BUG_ON(busiest_rq == target_rq);
3917 /* move a task from busiest_rq to target_rq */
3918 double_lock_balance(busiest_rq, target_rq);
3919 update_rq_clock(busiest_rq);
3920 update_rq_clock(target_rq);
3922 /* Search for an sd spanning us and the target CPU. */
3923 for_each_domain(target_cpu, sd) {
3924 if ((sd->flags & SD_LOAD_BALANCE) &&
3925 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3930 schedstat_inc(sd, alb_count);
3932 if (move_one_task(target_rq, target_cpu, busiest_rq,
3934 schedstat_inc(sd, alb_pushed);
3936 schedstat_inc(sd, alb_failed);
3938 double_unlock_balance(busiest_rq, target_rq);
3943 atomic_t load_balancer;
3944 cpumask_var_t cpu_mask;
3945 } nohz ____cacheline_aligned = {
3946 .load_balancer = ATOMIC_INIT(-1),
3950 * This routine will try to nominate the ilb (idle load balancing)
3951 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3952 * load balancing on behalf of all those cpus. If all the cpus in the system
3953 * go into this tickless mode, then there will be no ilb owner (as there is
3954 * no need for one) and all the cpus will sleep till the next wakeup event
3957 * For the ilb owner, tick is not stopped. And this tick will be used
3958 * for idle load balancing. ilb owner will still be part of
3961 * While stopping the tick, this cpu will become the ilb owner if there
3962 * is no other owner. And will be the owner till that cpu becomes busy
3963 * or if all cpus in the system stop their ticks at which point
3964 * there is no need for ilb owner.
3966 * When the ilb owner becomes busy, it nominates another owner, during the
3967 * next busy scheduler_tick()
3969 int select_nohz_load_balancer(int stop_tick)
3971 int cpu = smp_processor_id();
3974 cpu_rq(cpu)->in_nohz_recently = 1;
3976 if (!cpu_active(cpu)) {
3977 if (atomic_read(&nohz.load_balancer) != cpu)
3981 * If we are going offline and still the leader,
3984 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3990 cpumask_set_cpu(cpu, nohz.cpu_mask);
3992 /* time for ilb owner also to sleep */
3993 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3994 if (atomic_read(&nohz.load_balancer) == cpu)
3995 atomic_set(&nohz.load_balancer, -1);
3999 if (atomic_read(&nohz.load_balancer) == -1) {
4000 /* make me the ilb owner */
4001 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4003 } else if (atomic_read(&nohz.load_balancer) == cpu)
4006 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4009 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4011 if (atomic_read(&nohz.load_balancer) == cpu)
4012 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4019 static DEFINE_SPINLOCK(balancing);
4022 * It checks each scheduling domain to see if it is due to be balanced,
4023 * and initiates a balancing operation if so.
4025 * Balancing parameters are set up in arch_init_sched_domains.
4027 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4030 struct rq *rq = cpu_rq(cpu);
4031 unsigned long interval;
4032 struct sched_domain *sd;
4033 /* Earliest time when we have to do rebalance again */
4034 unsigned long next_balance = jiffies + 60*HZ;
4035 int update_next_balance = 0;
4039 /* Fails alloc? Rebalancing probably not a priority right now. */
4040 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4043 for_each_domain(cpu, sd) {
4044 if (!(sd->flags & SD_LOAD_BALANCE))
4047 interval = sd->balance_interval;
4048 if (idle != CPU_IDLE)
4049 interval *= sd->busy_factor;
4051 /* scale ms to jiffies */
4052 interval = msecs_to_jiffies(interval);
4053 if (unlikely(!interval))
4055 if (interval > HZ*NR_CPUS/10)
4056 interval = HZ*NR_CPUS/10;
4058 need_serialize = sd->flags & SD_SERIALIZE;
4060 if (need_serialize) {
4061 if (!spin_trylock(&balancing))
4065 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4066 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4068 * We've pulled tasks over so either we're no
4069 * longer idle, or one of our SMT siblings is
4072 idle = CPU_NOT_IDLE;
4074 sd->last_balance = jiffies;
4077 spin_unlock(&balancing);
4079 if (time_after(next_balance, sd->last_balance + interval)) {
4080 next_balance = sd->last_balance + interval;
4081 update_next_balance = 1;
4085 * Stop the load balance at this level. There is another
4086 * CPU in our sched group which is doing load balancing more
4094 * next_balance will be updated only when there is a need.
4095 * When the cpu is attached to null domain for ex, it will not be
4098 if (likely(update_next_balance))
4099 rq->next_balance = next_balance;
4101 free_cpumask_var(tmp);
4105 * run_rebalance_domains is triggered when needed from the scheduler tick.
4106 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4107 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4109 static void run_rebalance_domains(struct softirq_action *h)
4111 int this_cpu = smp_processor_id();
4112 struct rq *this_rq = cpu_rq(this_cpu);
4113 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4114 CPU_IDLE : CPU_NOT_IDLE;
4116 rebalance_domains(this_cpu, idle);
4120 * If this cpu is the owner for idle load balancing, then do the
4121 * balancing on behalf of the other idle cpus whose ticks are
4124 if (this_rq->idle_at_tick &&
4125 atomic_read(&nohz.load_balancer) == this_cpu) {
4129 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4130 if (balance_cpu == this_cpu)
4134 * If this cpu gets work to do, stop the load balancing
4135 * work being done for other cpus. Next load
4136 * balancing owner will pick it up.
4141 rebalance_domains(balance_cpu, CPU_IDLE);
4143 rq = cpu_rq(balance_cpu);
4144 if (time_after(this_rq->next_balance, rq->next_balance))
4145 this_rq->next_balance = rq->next_balance;
4152 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4154 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4155 * idle load balancing owner or decide to stop the periodic load balancing,
4156 * if the whole system is idle.
4158 static inline void trigger_load_balance(struct rq *rq, int cpu)
4162 * If we were in the nohz mode recently and busy at the current
4163 * scheduler tick, then check if we need to nominate new idle
4166 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4167 rq->in_nohz_recently = 0;
4169 if (atomic_read(&nohz.load_balancer) == cpu) {
4170 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4171 atomic_set(&nohz.load_balancer, -1);
4174 if (atomic_read(&nohz.load_balancer) == -1) {
4176 * simple selection for now: Nominate the
4177 * first cpu in the nohz list to be the next
4180 * TBD: Traverse the sched domains and nominate
4181 * the nearest cpu in the nohz.cpu_mask.
4183 int ilb = cpumask_first(nohz.cpu_mask);
4185 if (ilb < nr_cpu_ids)
4191 * If this cpu is idle and doing idle load balancing for all the
4192 * cpus with ticks stopped, is it time for that to stop?
4194 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4195 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4201 * If this cpu is idle and the idle load balancing is done by
4202 * someone else, then no need raise the SCHED_SOFTIRQ
4204 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4205 cpumask_test_cpu(cpu, nohz.cpu_mask))
4208 if (time_after_eq(jiffies, rq->next_balance))
4209 raise_softirq(SCHED_SOFTIRQ);
4212 #else /* CONFIG_SMP */
4215 * on UP we do not need to balance between CPUs:
4217 static inline void idle_balance(int cpu, struct rq *rq)
4223 DEFINE_PER_CPU(struct kernel_stat, kstat);
4225 EXPORT_PER_CPU_SYMBOL(kstat);
4228 * Return any ns on the sched_clock that have not yet been banked in
4229 * @p in case that task is currently running.
4231 unsigned long long task_delta_exec(struct task_struct *p)
4233 unsigned long flags;
4237 rq = task_rq_lock(p, &flags);
4239 if (task_current(rq, p)) {
4242 update_rq_clock(rq);
4243 delta_exec = rq->clock - p->se.exec_start;
4244 if ((s64)delta_exec > 0)
4248 task_rq_unlock(rq, &flags);
4254 * Account user cpu time to a process.
4255 * @p: the process that the cpu time gets accounted to
4256 * @cputime: the cpu time spent in user space since the last update
4257 * @cputime_scaled: cputime scaled by cpu frequency
4259 void account_user_time(struct task_struct *p, cputime_t cputime,
4260 cputime_t cputime_scaled)
4262 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4265 /* Add user time to process. */
4266 p->utime = cputime_add(p->utime, cputime);
4267 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4268 account_group_user_time(p, cputime);
4270 /* Add user time to cpustat. */
4271 tmp = cputime_to_cputime64(cputime);
4272 if (TASK_NICE(p) > 0)
4273 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4275 cpustat->user = cputime64_add(cpustat->user, tmp);
4276 /* Account for user time used */
4277 acct_update_integrals(p);
4281 * Account guest cpu time to a process.
4282 * @p: the process that the cpu time gets accounted to
4283 * @cputime: the cpu time spent in virtual machine since the last update
4284 * @cputime_scaled: cputime scaled by cpu frequency
4286 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4287 cputime_t cputime_scaled)
4290 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4292 tmp = cputime_to_cputime64(cputime);
4294 /* Add guest time to process. */
4295 p->utime = cputime_add(p->utime, cputime);
4296 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4297 account_group_user_time(p, cputime);
4298 p->gtime = cputime_add(p->gtime, cputime);
4300 /* Add guest time to cpustat. */
4301 cpustat->user = cputime64_add(cpustat->user, tmp);
4302 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4306 * Account system cpu time to a process.
4307 * @p: the process that the cpu time gets accounted to
4308 * @hardirq_offset: the offset to subtract from hardirq_count()
4309 * @cputime: the cpu time spent in kernel space since the last update
4310 * @cputime_scaled: cputime scaled by cpu frequency
4312 void account_system_time(struct task_struct *p, int hardirq_offset,
4313 cputime_t cputime, cputime_t cputime_scaled)
4315 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4318 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4319 account_guest_time(p, cputime, cputime_scaled);
4323 /* Add system time to process. */
4324 p->stime = cputime_add(p->stime, cputime);
4325 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4326 account_group_system_time(p, cputime);
4328 /* Add system time to cpustat. */
4329 tmp = cputime_to_cputime64(cputime);
4330 if (hardirq_count() - hardirq_offset)
4331 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4332 else if (softirq_count())
4333 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4335 cpustat->system = cputime64_add(cpustat->system, tmp);
4337 /* Account for system time used */
4338 acct_update_integrals(p);
4342 * Account for involuntary wait time.
4343 * @steal: the cpu time spent in involuntary wait
4345 void account_steal_time(cputime_t cputime)
4347 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4348 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4350 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4354 * Account for idle time.
4355 * @cputime: the cpu time spent in idle wait
4357 void account_idle_time(cputime_t cputime)
4359 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4360 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4361 struct rq *rq = this_rq();
4363 if (atomic_read(&rq->nr_iowait) > 0)
4364 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4366 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4369 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4372 * Account a single tick of cpu time.
4373 * @p: the process that the cpu time gets accounted to
4374 * @user_tick: indicates if the tick is a user or a system tick
4376 void account_process_tick(struct task_struct *p, int user_tick)
4378 cputime_t one_jiffy = jiffies_to_cputime(1);
4379 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4380 struct rq *rq = this_rq();
4383 account_user_time(p, one_jiffy, one_jiffy_scaled);
4384 else if (p != rq->idle)
4385 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4388 account_idle_time(one_jiffy);
4392 * Account multiple ticks of steal time.
4393 * @p: the process from which the cpu time has been stolen
4394 * @ticks: number of stolen ticks
4396 void account_steal_ticks(unsigned long ticks)
4398 account_steal_time(jiffies_to_cputime(ticks));
4402 * Account multiple ticks of idle time.
4403 * @ticks: number of stolen ticks
4405 void account_idle_ticks(unsigned long ticks)
4407 account_idle_time(jiffies_to_cputime(ticks));
4413 * Use precise platform statistics if available:
4415 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4416 cputime_t task_utime(struct task_struct *p)
4421 cputime_t task_stime(struct task_struct *p)
4426 cputime_t task_utime(struct task_struct *p)
4428 clock_t utime = cputime_to_clock_t(p->utime),
4429 total = utime + cputime_to_clock_t(p->stime);
4433 * Use CFS's precise accounting:
4435 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4439 do_div(temp, total);
4441 utime = (clock_t)temp;
4443 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4444 return p->prev_utime;
4447 cputime_t task_stime(struct task_struct *p)
4452 * Use CFS's precise accounting. (we subtract utime from
4453 * the total, to make sure the total observed by userspace
4454 * grows monotonically - apps rely on that):
4456 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4457 cputime_to_clock_t(task_utime(p));
4460 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4462 return p->prev_stime;
4466 inline cputime_t task_gtime(struct task_struct *p)
4472 * This function gets called by the timer code, with HZ frequency.
4473 * We call it with interrupts disabled.
4475 * It also gets called by the fork code, when changing the parent's
4478 void scheduler_tick(void)
4480 int cpu = smp_processor_id();
4481 struct rq *rq = cpu_rq(cpu);
4482 struct task_struct *curr = rq->curr;
4486 spin_lock(&rq->lock);
4487 update_rq_clock(rq);
4488 update_cpu_load(rq);
4489 curr->sched_class->task_tick(rq, curr, 0);
4490 spin_unlock(&rq->lock);
4493 rq->idle_at_tick = idle_cpu(cpu);
4494 trigger_load_balance(rq, cpu);
4498 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4499 defined(CONFIG_PREEMPT_TRACER))
4501 static inline unsigned long get_parent_ip(unsigned long addr)
4503 if (in_lock_functions(addr)) {
4504 addr = CALLER_ADDR2;
4505 if (in_lock_functions(addr))
4506 addr = CALLER_ADDR3;
4511 void __kprobes add_preempt_count(int val)
4513 #ifdef CONFIG_DEBUG_PREEMPT
4517 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4520 preempt_count() += val;
4521 #ifdef CONFIG_DEBUG_PREEMPT
4523 * Spinlock count overflowing soon?
4525 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4528 if (preempt_count() == val)
4529 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4531 EXPORT_SYMBOL(add_preempt_count);
4533 void __kprobes sub_preempt_count(int val)
4535 #ifdef CONFIG_DEBUG_PREEMPT
4539 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4542 * Is the spinlock portion underflowing?
4544 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4545 !(preempt_count() & PREEMPT_MASK)))
4549 if (preempt_count() == val)
4550 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4551 preempt_count() -= val;
4553 EXPORT_SYMBOL(sub_preempt_count);
4558 * Print scheduling while atomic bug:
4560 static noinline void __schedule_bug(struct task_struct *prev)
4562 struct pt_regs *regs = get_irq_regs();
4564 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4565 prev->comm, prev->pid, preempt_count());
4567 debug_show_held_locks(prev);
4569 if (irqs_disabled())
4570 print_irqtrace_events(prev);
4579 * Various schedule()-time debugging checks and statistics:
4581 static inline void schedule_debug(struct task_struct *prev)
4584 * Test if we are atomic. Since do_exit() needs to call into
4585 * schedule() atomically, we ignore that path for now.
4586 * Otherwise, whine if we are scheduling when we should not be.
4588 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4589 __schedule_bug(prev);
4591 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4593 schedstat_inc(this_rq(), sched_count);
4594 #ifdef CONFIG_SCHEDSTATS
4595 if (unlikely(prev->lock_depth >= 0)) {
4596 schedstat_inc(this_rq(), bkl_count);
4597 schedstat_inc(prev, sched_info.bkl_count);
4603 * Pick up the highest-prio task:
4605 static inline struct task_struct *
4606 pick_next_task(struct rq *rq, struct task_struct *prev)
4608 const struct sched_class *class;
4609 struct task_struct *p;
4612 * Optimization: we know that if all tasks are in
4613 * the fair class we can call that function directly:
4615 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4616 p = fair_sched_class.pick_next_task(rq);
4621 class = sched_class_highest;
4623 p = class->pick_next_task(rq);
4627 * Will never be NULL as the idle class always
4628 * returns a non-NULL p:
4630 class = class->next;
4635 * schedule() is the main scheduler function.
4637 asmlinkage void __sched schedule(void)
4639 struct task_struct *prev, *next;
4640 unsigned long *switch_count;
4646 cpu = smp_processor_id();
4650 switch_count = &prev->nivcsw;
4652 release_kernel_lock(prev);
4653 need_resched_nonpreemptible:
4655 schedule_debug(prev);
4657 if (sched_feat(HRTICK))
4660 spin_lock_irq(&rq->lock);
4661 update_rq_clock(rq);
4662 clear_tsk_need_resched(prev);
4664 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4665 if (unlikely(signal_pending_state(prev->state, prev)))
4666 prev->state = TASK_RUNNING;
4668 deactivate_task(rq, prev, 1);
4669 switch_count = &prev->nvcsw;
4673 if (prev->sched_class->pre_schedule)
4674 prev->sched_class->pre_schedule(rq, prev);
4677 if (unlikely(!rq->nr_running))
4678 idle_balance(cpu, rq);
4680 prev->sched_class->put_prev_task(rq, prev);
4681 next = pick_next_task(rq, prev);
4683 if (likely(prev != next)) {
4684 sched_info_switch(prev, next);
4690 context_switch(rq, prev, next); /* unlocks the rq */
4692 * the context switch might have flipped the stack from under
4693 * us, hence refresh the local variables.
4695 cpu = smp_processor_id();
4698 spin_unlock_irq(&rq->lock);
4700 if (unlikely(reacquire_kernel_lock(current) < 0))
4701 goto need_resched_nonpreemptible;
4703 preempt_enable_no_resched();
4704 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4707 EXPORT_SYMBOL(schedule);
4709 #ifdef CONFIG_PREEMPT
4711 * this is the entry point to schedule() from in-kernel preemption
4712 * off of preempt_enable. Kernel preemptions off return from interrupt
4713 * occur there and call schedule directly.
4715 asmlinkage void __sched preempt_schedule(void)
4717 struct thread_info *ti = current_thread_info();
4720 * If there is a non-zero preempt_count or interrupts are disabled,
4721 * we do not want to preempt the current task. Just return..
4723 if (likely(ti->preempt_count || irqs_disabled()))
4727 add_preempt_count(PREEMPT_ACTIVE);
4729 sub_preempt_count(PREEMPT_ACTIVE);
4732 * Check again in case we missed a preemption opportunity
4733 * between schedule and now.
4736 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4738 EXPORT_SYMBOL(preempt_schedule);
4741 * this is the entry point to schedule() from kernel preemption
4742 * off of irq context.
4743 * Note, that this is called and return with irqs disabled. This will
4744 * protect us against recursive calling from irq.
4746 asmlinkage void __sched preempt_schedule_irq(void)
4748 struct thread_info *ti = current_thread_info();
4750 /* Catch callers which need to be fixed */
4751 BUG_ON(ti->preempt_count || !irqs_disabled());
4754 add_preempt_count(PREEMPT_ACTIVE);
4757 local_irq_disable();
4758 sub_preempt_count(PREEMPT_ACTIVE);
4761 * Check again in case we missed a preemption opportunity
4762 * between schedule and now.
4765 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4768 #endif /* CONFIG_PREEMPT */
4770 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4773 return try_to_wake_up(curr->private, mode, sync);
4775 EXPORT_SYMBOL(default_wake_function);
4778 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4779 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4780 * number) then we wake all the non-exclusive tasks and one exclusive task.
4782 * There are circumstances in which we can try to wake a task which has already
4783 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4784 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4786 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4787 int nr_exclusive, int sync, void *key)
4789 wait_queue_t *curr, *next;
4791 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4792 unsigned flags = curr->flags;
4794 if (curr->func(curr, mode, sync, key) &&
4795 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4801 * __wake_up - wake up threads blocked on a waitqueue.
4803 * @mode: which threads
4804 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4805 * @key: is directly passed to the wakeup function
4807 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4808 int nr_exclusive, void *key)
4810 unsigned long flags;
4812 spin_lock_irqsave(&q->lock, flags);
4813 __wake_up_common(q, mode, nr_exclusive, 0, key);
4814 spin_unlock_irqrestore(&q->lock, flags);
4816 EXPORT_SYMBOL(__wake_up);
4819 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4821 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4823 __wake_up_common(q, mode, 1, 0, NULL);
4827 * __wake_up_sync - wake up threads blocked on a waitqueue.
4829 * @mode: which threads
4830 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4832 * The sync wakeup differs that the waker knows that it will schedule
4833 * away soon, so while the target thread will be woken up, it will not
4834 * be migrated to another CPU - ie. the two threads are 'synchronized'
4835 * with each other. This can prevent needless bouncing between CPUs.
4837 * On UP it can prevent extra preemption.
4840 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4842 unsigned long flags;
4848 if (unlikely(!nr_exclusive))
4851 spin_lock_irqsave(&q->lock, flags);
4852 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4853 spin_unlock_irqrestore(&q->lock, flags);
4855 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4858 * complete: - signals a single thread waiting on this completion
4859 * @x: holds the state of this particular completion
4861 * This will wake up a single thread waiting on this completion. Threads will be
4862 * awakened in the same order in which they were queued.
4864 * See also complete_all(), wait_for_completion() and related routines.
4866 void complete(struct completion *x)
4868 unsigned long flags;
4870 spin_lock_irqsave(&x->wait.lock, flags);
4872 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4873 spin_unlock_irqrestore(&x->wait.lock, flags);
4875 EXPORT_SYMBOL(complete);
4878 * complete_all: - signals all threads waiting on this completion
4879 * @x: holds the state of this particular completion
4881 * This will wake up all threads waiting on this particular completion event.
4883 void complete_all(struct completion *x)
4885 unsigned long flags;
4887 spin_lock_irqsave(&x->wait.lock, flags);
4888 x->done += UINT_MAX/2;
4889 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4890 spin_unlock_irqrestore(&x->wait.lock, flags);
4892 EXPORT_SYMBOL(complete_all);
4894 static inline long __sched
4895 do_wait_for_common(struct completion *x, long timeout, int state)
4898 DECLARE_WAITQUEUE(wait, current);
4900 wait.flags |= WQ_FLAG_EXCLUSIVE;
4901 __add_wait_queue_tail(&x->wait, &wait);
4903 if (signal_pending_state(state, current)) {
4904 timeout = -ERESTARTSYS;
4907 __set_current_state(state);
4908 spin_unlock_irq(&x->wait.lock);
4909 timeout = schedule_timeout(timeout);
4910 spin_lock_irq(&x->wait.lock);
4911 } while (!x->done && timeout);
4912 __remove_wait_queue(&x->wait, &wait);
4917 return timeout ?: 1;
4921 wait_for_common(struct completion *x, long timeout, int state)
4925 spin_lock_irq(&x->wait.lock);
4926 timeout = do_wait_for_common(x, timeout, state);
4927 spin_unlock_irq(&x->wait.lock);
4932 * wait_for_completion: - waits for completion of a task
4933 * @x: holds the state of this particular completion
4935 * This waits to be signaled for completion of a specific task. It is NOT
4936 * interruptible and there is no timeout.
4938 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4939 * and interrupt capability. Also see complete().
4941 void __sched wait_for_completion(struct completion *x)
4943 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4945 EXPORT_SYMBOL(wait_for_completion);
4948 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4949 * @x: holds the state of this particular completion
4950 * @timeout: timeout value in jiffies
4952 * This waits for either a completion of a specific task to be signaled or for a
4953 * specified timeout to expire. The timeout is in jiffies. It is not
4956 unsigned long __sched
4957 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4959 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4961 EXPORT_SYMBOL(wait_for_completion_timeout);
4964 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4965 * @x: holds the state of this particular completion
4967 * This waits for completion of a specific task to be signaled. It is
4970 int __sched wait_for_completion_interruptible(struct completion *x)
4972 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4973 if (t == -ERESTARTSYS)
4977 EXPORT_SYMBOL(wait_for_completion_interruptible);
4980 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4981 * @x: holds the state of this particular completion
4982 * @timeout: timeout value in jiffies
4984 * This waits for either a completion of a specific task to be signaled or for a
4985 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4987 unsigned long __sched
4988 wait_for_completion_interruptible_timeout(struct completion *x,
4989 unsigned long timeout)
4991 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4993 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4996 * wait_for_completion_killable: - waits for completion of a task (killable)
4997 * @x: holds the state of this particular completion
4999 * This waits to be signaled for completion of a specific task. It can be
5000 * interrupted by a kill signal.
5002 int __sched wait_for_completion_killable(struct completion *x)
5004 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5005 if (t == -ERESTARTSYS)
5009 EXPORT_SYMBOL(wait_for_completion_killable);
5012 * try_wait_for_completion - try to decrement a completion without blocking
5013 * @x: completion structure
5015 * Returns: 0 if a decrement cannot be done without blocking
5016 * 1 if a decrement succeeded.
5018 * If a completion is being used as a counting completion,
5019 * attempt to decrement the counter without blocking. This
5020 * enables us to avoid waiting if the resource the completion
5021 * is protecting is not available.
5023 bool try_wait_for_completion(struct completion *x)
5027 spin_lock_irq(&x->wait.lock);
5032 spin_unlock_irq(&x->wait.lock);
5035 EXPORT_SYMBOL(try_wait_for_completion);
5038 * completion_done - Test to see if a completion has any waiters
5039 * @x: completion structure
5041 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5042 * 1 if there are no waiters.
5045 bool completion_done(struct completion *x)
5049 spin_lock_irq(&x->wait.lock);
5052 spin_unlock_irq(&x->wait.lock);
5055 EXPORT_SYMBOL(completion_done);
5058 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5060 unsigned long flags;
5063 init_waitqueue_entry(&wait, current);
5065 __set_current_state(state);
5067 spin_lock_irqsave(&q->lock, flags);
5068 __add_wait_queue(q, &wait);
5069 spin_unlock(&q->lock);
5070 timeout = schedule_timeout(timeout);
5071 spin_lock_irq(&q->lock);
5072 __remove_wait_queue(q, &wait);
5073 spin_unlock_irqrestore(&q->lock, flags);
5078 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5080 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5082 EXPORT_SYMBOL(interruptible_sleep_on);
5085 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5087 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5089 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5091 void __sched sleep_on(wait_queue_head_t *q)
5093 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5095 EXPORT_SYMBOL(sleep_on);
5097 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5099 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5101 EXPORT_SYMBOL(sleep_on_timeout);
5103 #ifdef CONFIG_RT_MUTEXES
5106 * rt_mutex_setprio - set the current priority of a task
5108 * @prio: prio value (kernel-internal form)
5110 * This function changes the 'effective' priority of a task. It does
5111 * not touch ->normal_prio like __setscheduler().
5113 * Used by the rt_mutex code to implement priority inheritance logic.
5115 void rt_mutex_setprio(struct task_struct *p, int prio)
5117 unsigned long flags;
5118 int oldprio, on_rq, running;
5120 const struct sched_class *prev_class = p->sched_class;
5122 BUG_ON(prio < 0 || prio > MAX_PRIO);
5124 rq = task_rq_lock(p, &flags);
5125 update_rq_clock(rq);
5128 on_rq = p->se.on_rq;
5129 running = task_current(rq, p);
5131 dequeue_task(rq, p, 0);
5133 p->sched_class->put_prev_task(rq, p);
5136 p->sched_class = &rt_sched_class;
5138 p->sched_class = &fair_sched_class;
5143 p->sched_class->set_curr_task(rq);
5145 enqueue_task(rq, p, 0);
5147 check_class_changed(rq, p, prev_class, oldprio, running);
5149 task_rq_unlock(rq, &flags);
5154 void set_user_nice(struct task_struct *p, long nice)
5156 int old_prio, delta, on_rq;
5157 unsigned long flags;
5160 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5163 * We have to be careful, if called from sys_setpriority(),
5164 * the task might be in the middle of scheduling on another CPU.
5166 rq = task_rq_lock(p, &flags);
5167 update_rq_clock(rq);
5169 * The RT priorities are set via sched_setscheduler(), but we still
5170 * allow the 'normal' nice value to be set - but as expected
5171 * it wont have any effect on scheduling until the task is
5172 * SCHED_FIFO/SCHED_RR:
5174 if (task_has_rt_policy(p)) {
5175 p->static_prio = NICE_TO_PRIO(nice);
5178 on_rq = p->se.on_rq;
5180 dequeue_task(rq, p, 0);
5182 p->static_prio = NICE_TO_PRIO(nice);
5185 p->prio = effective_prio(p);
5186 delta = p->prio - old_prio;
5189 enqueue_task(rq, p, 0);
5191 * If the task increased its priority or is running and
5192 * lowered its priority, then reschedule its CPU:
5194 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5195 resched_task(rq->curr);
5198 task_rq_unlock(rq, &flags);
5200 EXPORT_SYMBOL(set_user_nice);
5203 * can_nice - check if a task can reduce its nice value
5207 int can_nice(const struct task_struct *p, const int nice)
5209 /* convert nice value [19,-20] to rlimit style value [1,40] */
5210 int nice_rlim = 20 - nice;
5212 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5213 capable(CAP_SYS_NICE));
5216 #ifdef __ARCH_WANT_SYS_NICE
5219 * sys_nice - change the priority of the current process.
5220 * @increment: priority increment
5222 * sys_setpriority is a more generic, but much slower function that
5223 * does similar things.
5225 SYSCALL_DEFINE1(nice, int, increment)
5230 * Setpriority might change our priority at the same moment.
5231 * We don't have to worry. Conceptually one call occurs first
5232 * and we have a single winner.
5234 if (increment < -40)
5239 nice = PRIO_TO_NICE(current->static_prio) + increment;
5245 if (increment < 0 && !can_nice(current, nice))
5248 retval = security_task_setnice(current, nice);
5252 set_user_nice(current, nice);
5259 * task_prio - return the priority value of a given task.
5260 * @p: the task in question.
5262 * This is the priority value as seen by users in /proc.
5263 * RT tasks are offset by -200. Normal tasks are centered
5264 * around 0, value goes from -16 to +15.
5266 int task_prio(const struct task_struct *p)
5268 return p->prio - MAX_RT_PRIO;
5272 * task_nice - return the nice value of a given task.
5273 * @p: the task in question.
5275 int task_nice(const struct task_struct *p)
5277 return TASK_NICE(p);
5279 EXPORT_SYMBOL(task_nice);
5282 * idle_cpu - is a given cpu idle currently?
5283 * @cpu: the processor in question.
5285 int idle_cpu(int cpu)
5287 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5291 * idle_task - return the idle task for a given cpu.
5292 * @cpu: the processor in question.
5294 struct task_struct *idle_task(int cpu)
5296 return cpu_rq(cpu)->idle;
5300 * find_process_by_pid - find a process with a matching PID value.
5301 * @pid: the pid in question.
5303 static struct task_struct *find_process_by_pid(pid_t pid)
5305 return pid ? find_task_by_vpid(pid) : current;
5308 /* Actually do priority change: must hold rq lock. */
5310 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5312 BUG_ON(p->se.on_rq);
5315 switch (p->policy) {
5319 p->sched_class = &fair_sched_class;
5323 p->sched_class = &rt_sched_class;
5327 p->rt_priority = prio;
5328 p->normal_prio = normal_prio(p);
5329 /* we are holding p->pi_lock already */
5330 p->prio = rt_mutex_getprio(p);
5335 * check the target process has a UID that matches the current process's
5337 static bool check_same_owner(struct task_struct *p)
5339 const struct cred *cred = current_cred(), *pcred;
5343 pcred = __task_cred(p);
5344 match = (cred->euid == pcred->euid ||
5345 cred->euid == pcred->uid);
5350 static int __sched_setscheduler(struct task_struct *p, int policy,
5351 struct sched_param *param, bool user)
5353 int retval, oldprio, oldpolicy = -1, on_rq, running;
5354 unsigned long flags;
5355 const struct sched_class *prev_class = p->sched_class;
5358 /* may grab non-irq protected spin_locks */
5359 BUG_ON(in_interrupt());
5361 /* double check policy once rq lock held */
5363 policy = oldpolicy = p->policy;
5364 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5365 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5366 policy != SCHED_IDLE)
5369 * Valid priorities for SCHED_FIFO and SCHED_RR are
5370 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5371 * SCHED_BATCH and SCHED_IDLE is 0.
5373 if (param->sched_priority < 0 ||
5374 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5375 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5377 if (rt_policy(policy) != (param->sched_priority != 0))
5381 * Allow unprivileged RT tasks to decrease priority:
5383 if (user && !capable(CAP_SYS_NICE)) {
5384 if (rt_policy(policy)) {
5385 unsigned long rlim_rtprio;
5387 if (!lock_task_sighand(p, &flags))
5389 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5390 unlock_task_sighand(p, &flags);
5392 /* can't set/change the rt policy */
5393 if (policy != p->policy && !rlim_rtprio)
5396 /* can't increase priority */
5397 if (param->sched_priority > p->rt_priority &&
5398 param->sched_priority > rlim_rtprio)
5402 * Like positive nice levels, dont allow tasks to
5403 * move out of SCHED_IDLE either:
5405 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5408 /* can't change other user's priorities */
5409 if (!check_same_owner(p))
5414 #ifdef CONFIG_RT_GROUP_SCHED
5416 * Do not allow realtime tasks into groups that have no runtime
5419 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5420 task_group(p)->rt_bandwidth.rt_runtime == 0)
5424 retval = security_task_setscheduler(p, policy, param);
5430 * make sure no PI-waiters arrive (or leave) while we are
5431 * changing the priority of the task:
5433 spin_lock_irqsave(&p->pi_lock, flags);
5435 * To be able to change p->policy safely, the apropriate
5436 * runqueue lock must be held.
5438 rq = __task_rq_lock(p);
5439 /* recheck policy now with rq lock held */
5440 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5441 policy = oldpolicy = -1;
5442 __task_rq_unlock(rq);
5443 spin_unlock_irqrestore(&p->pi_lock, flags);
5446 update_rq_clock(rq);
5447 on_rq = p->se.on_rq;
5448 running = task_current(rq, p);
5450 deactivate_task(rq, p, 0);
5452 p->sched_class->put_prev_task(rq, p);
5455 __setscheduler(rq, p, policy, param->sched_priority);
5458 p->sched_class->set_curr_task(rq);
5460 activate_task(rq, p, 0);
5462 check_class_changed(rq, p, prev_class, oldprio, running);
5464 __task_rq_unlock(rq);
5465 spin_unlock_irqrestore(&p->pi_lock, flags);
5467 rt_mutex_adjust_pi(p);
5473 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5474 * @p: the task in question.
5475 * @policy: new policy.
5476 * @param: structure containing the new RT priority.
5478 * NOTE that the task may be already dead.
5480 int sched_setscheduler(struct task_struct *p, int policy,
5481 struct sched_param *param)
5483 return __sched_setscheduler(p, policy, param, true);
5485 EXPORT_SYMBOL_GPL(sched_setscheduler);
5488 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5489 * @p: the task in question.
5490 * @policy: new policy.
5491 * @param: structure containing the new RT priority.
5493 * Just like sched_setscheduler, only don't bother checking if the
5494 * current context has permission. For example, this is needed in
5495 * stop_machine(): we create temporary high priority worker threads,
5496 * but our caller might not have that capability.
5498 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5499 struct sched_param *param)
5501 return __sched_setscheduler(p, policy, param, false);
5505 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5507 struct sched_param lparam;
5508 struct task_struct *p;
5511 if (!param || pid < 0)
5513 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5518 p = find_process_by_pid(pid);
5520 retval = sched_setscheduler(p, policy, &lparam);
5527 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5528 * @pid: the pid in question.
5529 * @policy: new policy.
5530 * @param: structure containing the new RT priority.
5532 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5533 struct sched_param __user *, param)
5535 /* negative values for policy are not valid */
5539 return do_sched_setscheduler(pid, policy, param);
5543 * sys_sched_setparam - set/change the RT priority of a thread
5544 * @pid: the pid in question.
5545 * @param: structure containing the new RT priority.
5547 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5549 return do_sched_setscheduler(pid, -1, param);
5553 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5554 * @pid: the pid in question.
5556 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5558 struct task_struct *p;
5565 read_lock(&tasklist_lock);
5566 p = find_process_by_pid(pid);
5568 retval = security_task_getscheduler(p);
5572 read_unlock(&tasklist_lock);
5577 * sys_sched_getscheduler - get the RT priority of a thread
5578 * @pid: the pid in question.
5579 * @param: structure containing the RT priority.
5581 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5583 struct sched_param lp;
5584 struct task_struct *p;
5587 if (!param || pid < 0)
5590 read_lock(&tasklist_lock);
5591 p = find_process_by_pid(pid);
5596 retval = security_task_getscheduler(p);
5600 lp.sched_priority = p->rt_priority;
5601 read_unlock(&tasklist_lock);
5604 * This one might sleep, we cannot do it with a spinlock held ...
5606 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5611 read_unlock(&tasklist_lock);
5615 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5617 cpumask_var_t cpus_allowed, new_mask;
5618 struct task_struct *p;
5622 read_lock(&tasklist_lock);
5624 p = find_process_by_pid(pid);
5626 read_unlock(&tasklist_lock);
5632 * It is not safe to call set_cpus_allowed with the
5633 * tasklist_lock held. We will bump the task_struct's
5634 * usage count and then drop tasklist_lock.
5637 read_unlock(&tasklist_lock);
5639 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5643 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5645 goto out_free_cpus_allowed;
5648 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5651 retval = security_task_setscheduler(p, 0, NULL);
5655 cpuset_cpus_allowed(p, cpus_allowed);
5656 cpumask_and(new_mask, in_mask, cpus_allowed);
5658 retval = set_cpus_allowed_ptr(p, new_mask);
5661 cpuset_cpus_allowed(p, cpus_allowed);
5662 if (!cpumask_subset(new_mask, cpus_allowed)) {
5664 * We must have raced with a concurrent cpuset
5665 * update. Just reset the cpus_allowed to the
5666 * cpuset's cpus_allowed
5668 cpumask_copy(new_mask, cpus_allowed);
5673 free_cpumask_var(new_mask);
5674 out_free_cpus_allowed:
5675 free_cpumask_var(cpus_allowed);
5682 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5683 struct cpumask *new_mask)
5685 if (len < cpumask_size())
5686 cpumask_clear(new_mask);
5687 else if (len > cpumask_size())
5688 len = cpumask_size();
5690 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5694 * sys_sched_setaffinity - set the cpu affinity of a process
5695 * @pid: pid of the process
5696 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5697 * @user_mask_ptr: user-space pointer to the new cpu mask
5699 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5700 unsigned long __user *, user_mask_ptr)
5702 cpumask_var_t new_mask;
5705 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5708 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5710 retval = sched_setaffinity(pid, new_mask);
5711 free_cpumask_var(new_mask);
5715 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5717 struct task_struct *p;
5721 read_lock(&tasklist_lock);
5724 p = find_process_by_pid(pid);
5728 retval = security_task_getscheduler(p);
5732 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5735 read_unlock(&tasklist_lock);
5742 * sys_sched_getaffinity - get the cpu affinity of a process
5743 * @pid: pid of the process
5744 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5745 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5747 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5748 unsigned long __user *, user_mask_ptr)
5753 if (len < cpumask_size())
5756 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5759 ret = sched_getaffinity(pid, mask);
5761 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5764 ret = cpumask_size();
5766 free_cpumask_var(mask);
5772 * sys_sched_yield - yield the current processor to other threads.
5774 * This function yields the current CPU to other tasks. If there are no
5775 * other threads running on this CPU then this function will return.
5777 SYSCALL_DEFINE0(sched_yield)
5779 struct rq *rq = this_rq_lock();
5781 schedstat_inc(rq, yld_count);
5782 current->sched_class->yield_task(rq);
5785 * Since we are going to call schedule() anyway, there's
5786 * no need to preempt or enable interrupts:
5788 __release(rq->lock);
5789 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5790 _raw_spin_unlock(&rq->lock);
5791 preempt_enable_no_resched();
5798 static void __cond_resched(void)
5800 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5801 __might_sleep(__FILE__, __LINE__);
5804 * The BKS might be reacquired before we have dropped
5805 * PREEMPT_ACTIVE, which could trigger a second
5806 * cond_resched() call.
5809 add_preempt_count(PREEMPT_ACTIVE);
5811 sub_preempt_count(PREEMPT_ACTIVE);
5812 } while (need_resched());
5815 int __sched _cond_resched(void)
5817 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5818 system_state == SYSTEM_RUNNING) {
5824 EXPORT_SYMBOL(_cond_resched);
5827 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5828 * call schedule, and on return reacquire the lock.
5830 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5831 * operations here to prevent schedule() from being called twice (once via
5832 * spin_unlock(), once by hand).
5834 int cond_resched_lock(spinlock_t *lock)
5836 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5839 if (spin_needbreak(lock) || resched) {
5841 if (resched && need_resched())
5850 EXPORT_SYMBOL(cond_resched_lock);
5852 int __sched cond_resched_softirq(void)
5854 BUG_ON(!in_softirq());
5856 if (need_resched() && system_state == SYSTEM_RUNNING) {
5864 EXPORT_SYMBOL(cond_resched_softirq);
5867 * yield - yield the current processor to other threads.
5869 * This is a shortcut for kernel-space yielding - it marks the
5870 * thread runnable and calls sys_sched_yield().
5872 void __sched yield(void)
5874 set_current_state(TASK_RUNNING);
5877 EXPORT_SYMBOL(yield);
5880 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5881 * that process accounting knows that this is a task in IO wait state.
5883 * But don't do that if it is a deliberate, throttling IO wait (this task
5884 * has set its backing_dev_info: the queue against which it should throttle)
5886 void __sched io_schedule(void)
5888 struct rq *rq = &__raw_get_cpu_var(runqueues);
5890 delayacct_blkio_start();
5891 atomic_inc(&rq->nr_iowait);
5893 atomic_dec(&rq->nr_iowait);
5894 delayacct_blkio_end();
5896 EXPORT_SYMBOL(io_schedule);
5898 long __sched io_schedule_timeout(long timeout)
5900 struct rq *rq = &__raw_get_cpu_var(runqueues);
5903 delayacct_blkio_start();
5904 atomic_inc(&rq->nr_iowait);
5905 ret = schedule_timeout(timeout);
5906 atomic_dec(&rq->nr_iowait);
5907 delayacct_blkio_end();
5912 * sys_sched_get_priority_max - return maximum RT priority.
5913 * @policy: scheduling class.
5915 * this syscall returns the maximum rt_priority that can be used
5916 * by a given scheduling class.
5918 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5925 ret = MAX_USER_RT_PRIO-1;
5937 * sys_sched_get_priority_min - return minimum RT priority.
5938 * @policy: scheduling class.
5940 * this syscall returns the minimum rt_priority that can be used
5941 * by a given scheduling class.
5943 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5961 * sys_sched_rr_get_interval - return the default timeslice of a process.
5962 * @pid: pid of the process.
5963 * @interval: userspace pointer to the timeslice value.
5965 * this syscall writes the default timeslice value of a given process
5966 * into the user-space timespec buffer. A value of '0' means infinity.
5968 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5969 struct timespec __user *, interval)
5971 struct task_struct *p;
5972 unsigned int time_slice;
5980 read_lock(&tasklist_lock);
5981 p = find_process_by_pid(pid);
5985 retval = security_task_getscheduler(p);
5990 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5991 * tasks that are on an otherwise idle runqueue:
5994 if (p->policy == SCHED_RR) {
5995 time_slice = DEF_TIMESLICE;
5996 } else if (p->policy != SCHED_FIFO) {
5997 struct sched_entity *se = &p->se;
5998 unsigned long flags;
6001 rq = task_rq_lock(p, &flags);
6002 if (rq->cfs.load.weight)
6003 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6004 task_rq_unlock(rq, &flags);
6006 read_unlock(&tasklist_lock);
6007 jiffies_to_timespec(time_slice, &t);
6008 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6012 read_unlock(&tasklist_lock);
6016 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6018 void sched_show_task(struct task_struct *p)
6020 unsigned long free = 0;
6023 state = p->state ? __ffs(p->state) + 1 : 0;
6024 printk(KERN_INFO "%-13.13s %c", p->comm,
6025 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6026 #if BITS_PER_LONG == 32
6027 if (state == TASK_RUNNING)
6028 printk(KERN_CONT " running ");
6030 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6032 if (state == TASK_RUNNING)
6033 printk(KERN_CONT " running task ");
6035 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6037 #ifdef CONFIG_DEBUG_STACK_USAGE
6039 unsigned long *n = end_of_stack(p);
6042 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6045 printk(KERN_CONT "%5lu %5d %6d\n", free,
6046 task_pid_nr(p), task_pid_nr(p->real_parent));
6048 show_stack(p, NULL);
6051 void show_state_filter(unsigned long state_filter)
6053 struct task_struct *g, *p;
6055 #if BITS_PER_LONG == 32
6057 " task PC stack pid father\n");
6060 " task PC stack pid father\n");
6062 read_lock(&tasklist_lock);
6063 do_each_thread(g, p) {
6065 * reset the NMI-timeout, listing all files on a slow
6066 * console might take alot of time:
6068 touch_nmi_watchdog();
6069 if (!state_filter || (p->state & state_filter))
6071 } while_each_thread(g, p);
6073 touch_all_softlockup_watchdogs();
6075 #ifdef CONFIG_SCHED_DEBUG
6076 sysrq_sched_debug_show();
6078 read_unlock(&tasklist_lock);
6080 * Only show locks if all tasks are dumped:
6082 if (state_filter == -1)
6083 debug_show_all_locks();
6086 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6088 idle->sched_class = &idle_sched_class;
6092 * init_idle - set up an idle thread for a given CPU
6093 * @idle: task in question
6094 * @cpu: cpu the idle task belongs to
6096 * NOTE: this function does not set the idle thread's NEED_RESCHED
6097 * flag, to make booting more robust.
6099 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6101 struct rq *rq = cpu_rq(cpu);
6102 unsigned long flags;
6104 spin_lock_irqsave(&rq->lock, flags);
6107 idle->se.exec_start = sched_clock();
6109 idle->prio = idle->normal_prio = MAX_PRIO;
6110 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6111 __set_task_cpu(idle, cpu);
6113 rq->curr = rq->idle = idle;
6114 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6117 spin_unlock_irqrestore(&rq->lock, flags);
6119 /* Set the preempt count _outside_ the spinlocks! */
6120 #if defined(CONFIG_PREEMPT)
6121 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6123 task_thread_info(idle)->preempt_count = 0;
6126 * The idle tasks have their own, simple scheduling class:
6128 idle->sched_class = &idle_sched_class;
6129 ftrace_graph_init_task(idle);
6133 * In a system that switches off the HZ timer nohz_cpu_mask
6134 * indicates which cpus entered this state. This is used
6135 * in the rcu update to wait only for active cpus. For system
6136 * which do not switch off the HZ timer nohz_cpu_mask should
6137 * always be CPU_BITS_NONE.
6139 cpumask_var_t nohz_cpu_mask;
6142 * Increase the granularity value when there are more CPUs,
6143 * because with more CPUs the 'effective latency' as visible
6144 * to users decreases. But the relationship is not linear,
6145 * so pick a second-best guess by going with the log2 of the
6148 * This idea comes from the SD scheduler of Con Kolivas:
6150 static inline void sched_init_granularity(void)
6152 unsigned int factor = 1 + ilog2(num_online_cpus());
6153 const unsigned long limit = 200000000;
6155 sysctl_sched_min_granularity *= factor;
6156 if (sysctl_sched_min_granularity > limit)
6157 sysctl_sched_min_granularity = limit;
6159 sysctl_sched_latency *= factor;
6160 if (sysctl_sched_latency > limit)
6161 sysctl_sched_latency = limit;
6163 sysctl_sched_wakeup_granularity *= factor;
6165 sysctl_sched_shares_ratelimit *= factor;
6170 * This is how migration works:
6172 * 1) we queue a struct migration_req structure in the source CPU's
6173 * runqueue and wake up that CPU's migration thread.
6174 * 2) we down() the locked semaphore => thread blocks.
6175 * 3) migration thread wakes up (implicitly it forces the migrated
6176 * thread off the CPU)
6177 * 4) it gets the migration request and checks whether the migrated
6178 * task is still in the wrong runqueue.
6179 * 5) if it's in the wrong runqueue then the migration thread removes
6180 * it and puts it into the right queue.
6181 * 6) migration thread up()s the semaphore.
6182 * 7) we wake up and the migration is done.
6186 * Change a given task's CPU affinity. Migrate the thread to a
6187 * proper CPU and schedule it away if the CPU it's executing on
6188 * is removed from the allowed bitmask.
6190 * NOTE: the caller must have a valid reference to the task, the
6191 * task must not exit() & deallocate itself prematurely. The
6192 * call is not atomic; no spinlocks may be held.
6194 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6196 struct migration_req req;
6197 unsigned long flags;
6201 rq = task_rq_lock(p, &flags);
6202 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6207 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6208 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6213 if (p->sched_class->set_cpus_allowed)
6214 p->sched_class->set_cpus_allowed(p, new_mask);
6216 cpumask_copy(&p->cpus_allowed, new_mask);
6217 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6220 /* Can the task run on the task's current CPU? If so, we're done */
6221 if (cpumask_test_cpu(task_cpu(p), new_mask))
6224 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6225 /* Need help from migration thread: drop lock and wait. */
6226 task_rq_unlock(rq, &flags);
6227 wake_up_process(rq->migration_thread);
6228 wait_for_completion(&req.done);
6229 tlb_migrate_finish(p->mm);
6233 task_rq_unlock(rq, &flags);
6237 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6240 * Move (not current) task off this cpu, onto dest cpu. We're doing
6241 * this because either it can't run here any more (set_cpus_allowed()
6242 * away from this CPU, or CPU going down), or because we're
6243 * attempting to rebalance this task on exec (sched_exec).
6245 * So we race with normal scheduler movements, but that's OK, as long
6246 * as the task is no longer on this CPU.
6248 * Returns non-zero if task was successfully migrated.
6250 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6252 struct rq *rq_dest, *rq_src;
6255 if (unlikely(!cpu_active(dest_cpu)))
6258 rq_src = cpu_rq(src_cpu);
6259 rq_dest = cpu_rq(dest_cpu);
6261 double_rq_lock(rq_src, rq_dest);
6262 /* Already moved. */
6263 if (task_cpu(p) != src_cpu)
6265 /* Affinity changed (again). */
6266 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6269 on_rq = p->se.on_rq;
6271 deactivate_task(rq_src, p, 0);
6273 set_task_cpu(p, dest_cpu);
6275 activate_task(rq_dest, p, 0);
6276 check_preempt_curr(rq_dest, p, 0);
6281 double_rq_unlock(rq_src, rq_dest);
6286 * migration_thread - this is a highprio system thread that performs
6287 * thread migration by bumping thread off CPU then 'pushing' onto
6290 static int migration_thread(void *data)
6292 int cpu = (long)data;
6296 BUG_ON(rq->migration_thread != current);
6298 set_current_state(TASK_INTERRUPTIBLE);
6299 while (!kthread_should_stop()) {
6300 struct migration_req *req;
6301 struct list_head *head;
6303 spin_lock_irq(&rq->lock);
6305 if (cpu_is_offline(cpu)) {
6306 spin_unlock_irq(&rq->lock);
6310 if (rq->active_balance) {
6311 active_load_balance(rq, cpu);
6312 rq->active_balance = 0;
6315 head = &rq->migration_queue;
6317 if (list_empty(head)) {
6318 spin_unlock_irq(&rq->lock);
6320 set_current_state(TASK_INTERRUPTIBLE);
6323 req = list_entry(head->next, struct migration_req, list);
6324 list_del_init(head->next);
6326 spin_unlock(&rq->lock);
6327 __migrate_task(req->task, cpu, req->dest_cpu);
6330 complete(&req->done);
6332 __set_current_state(TASK_RUNNING);
6336 /* Wait for kthread_stop */
6337 set_current_state(TASK_INTERRUPTIBLE);
6338 while (!kthread_should_stop()) {
6340 set_current_state(TASK_INTERRUPTIBLE);
6342 __set_current_state(TASK_RUNNING);
6346 #ifdef CONFIG_HOTPLUG_CPU
6348 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6352 local_irq_disable();
6353 ret = __migrate_task(p, src_cpu, dest_cpu);
6359 * Figure out where task on dead CPU should go, use force if necessary.
6361 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6364 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6367 /* Look for allowed, online CPU in same node. */
6368 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6369 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6372 /* Any allowed, online CPU? */
6373 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6374 if (dest_cpu < nr_cpu_ids)
6377 /* No more Mr. Nice Guy. */
6378 if (dest_cpu >= nr_cpu_ids) {
6379 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6380 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6383 * Don't tell them about moving exiting tasks or
6384 * kernel threads (both mm NULL), since they never
6387 if (p->mm && printk_ratelimit()) {
6388 printk(KERN_INFO "process %d (%s) no "
6389 "longer affine to cpu%d\n",
6390 task_pid_nr(p), p->comm, dead_cpu);
6395 /* It can have affinity changed while we were choosing. */
6396 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6401 * While a dead CPU has no uninterruptible tasks queued at this point,
6402 * it might still have a nonzero ->nr_uninterruptible counter, because
6403 * for performance reasons the counter is not stricly tracking tasks to
6404 * their home CPUs. So we just add the counter to another CPU's counter,
6405 * to keep the global sum constant after CPU-down:
6407 static void migrate_nr_uninterruptible(struct rq *rq_src)
6409 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6410 unsigned long flags;
6412 local_irq_save(flags);
6413 double_rq_lock(rq_src, rq_dest);
6414 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6415 rq_src->nr_uninterruptible = 0;
6416 double_rq_unlock(rq_src, rq_dest);
6417 local_irq_restore(flags);
6420 /* Run through task list and migrate tasks from the dead cpu. */
6421 static void migrate_live_tasks(int src_cpu)
6423 struct task_struct *p, *t;
6425 read_lock(&tasklist_lock);
6427 do_each_thread(t, p) {
6431 if (task_cpu(p) == src_cpu)
6432 move_task_off_dead_cpu(src_cpu, p);
6433 } while_each_thread(t, p);
6435 read_unlock(&tasklist_lock);
6439 * Schedules idle task to be the next runnable task on current CPU.
6440 * It does so by boosting its priority to highest possible.
6441 * Used by CPU offline code.
6443 void sched_idle_next(void)
6445 int this_cpu = smp_processor_id();
6446 struct rq *rq = cpu_rq(this_cpu);
6447 struct task_struct *p = rq->idle;
6448 unsigned long flags;
6450 /* cpu has to be offline */
6451 BUG_ON(cpu_online(this_cpu));
6454 * Strictly not necessary since rest of the CPUs are stopped by now
6455 * and interrupts disabled on the current cpu.
6457 spin_lock_irqsave(&rq->lock, flags);
6459 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6461 update_rq_clock(rq);
6462 activate_task(rq, p, 0);
6464 spin_unlock_irqrestore(&rq->lock, flags);
6468 * Ensures that the idle task is using init_mm right before its cpu goes
6471 void idle_task_exit(void)
6473 struct mm_struct *mm = current->active_mm;
6475 BUG_ON(cpu_online(smp_processor_id()));
6478 switch_mm(mm, &init_mm, current);
6482 /* called under rq->lock with disabled interrupts */
6483 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6485 struct rq *rq = cpu_rq(dead_cpu);
6487 /* Must be exiting, otherwise would be on tasklist. */
6488 BUG_ON(!p->exit_state);
6490 /* Cannot have done final schedule yet: would have vanished. */
6491 BUG_ON(p->state == TASK_DEAD);
6496 * Drop lock around migration; if someone else moves it,
6497 * that's OK. No task can be added to this CPU, so iteration is
6500 spin_unlock_irq(&rq->lock);
6501 move_task_off_dead_cpu(dead_cpu, p);
6502 spin_lock_irq(&rq->lock);
6507 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6508 static void migrate_dead_tasks(unsigned int dead_cpu)
6510 struct rq *rq = cpu_rq(dead_cpu);
6511 struct task_struct *next;
6514 if (!rq->nr_running)
6516 update_rq_clock(rq);
6517 next = pick_next_task(rq, rq->curr);
6520 next->sched_class->put_prev_task(rq, next);
6521 migrate_dead(dead_cpu, next);
6525 #endif /* CONFIG_HOTPLUG_CPU */
6527 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6529 static struct ctl_table sd_ctl_dir[] = {
6531 .procname = "sched_domain",
6537 static struct ctl_table sd_ctl_root[] = {
6539 .ctl_name = CTL_KERN,
6540 .procname = "kernel",
6542 .child = sd_ctl_dir,
6547 static struct ctl_table *sd_alloc_ctl_entry(int n)
6549 struct ctl_table *entry =
6550 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6555 static void sd_free_ctl_entry(struct ctl_table **tablep)
6557 struct ctl_table *entry;
6560 * In the intermediate directories, both the child directory and
6561 * procname are dynamically allocated and could fail but the mode
6562 * will always be set. In the lowest directory the names are
6563 * static strings and all have proc handlers.
6565 for (entry = *tablep; entry->mode; entry++) {
6567 sd_free_ctl_entry(&entry->child);
6568 if (entry->proc_handler == NULL)
6569 kfree(entry->procname);
6577 set_table_entry(struct ctl_table *entry,
6578 const char *procname, void *data, int maxlen,
6579 mode_t mode, proc_handler *proc_handler)
6581 entry->procname = procname;
6583 entry->maxlen = maxlen;
6585 entry->proc_handler = proc_handler;
6588 static struct ctl_table *
6589 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6591 struct ctl_table *table = sd_alloc_ctl_entry(13);
6596 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6597 sizeof(long), 0644, proc_doulongvec_minmax);
6598 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6599 sizeof(long), 0644, proc_doulongvec_minmax);
6600 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6601 sizeof(int), 0644, proc_dointvec_minmax);
6602 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6603 sizeof(int), 0644, proc_dointvec_minmax);
6604 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6605 sizeof(int), 0644, proc_dointvec_minmax);
6606 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6607 sizeof(int), 0644, proc_dointvec_minmax);
6608 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6609 sizeof(int), 0644, proc_dointvec_minmax);
6610 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6611 sizeof(int), 0644, proc_dointvec_minmax);
6612 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6613 sizeof(int), 0644, proc_dointvec_minmax);
6614 set_table_entry(&table[9], "cache_nice_tries",
6615 &sd->cache_nice_tries,
6616 sizeof(int), 0644, proc_dointvec_minmax);
6617 set_table_entry(&table[10], "flags", &sd->flags,
6618 sizeof(int), 0644, proc_dointvec_minmax);
6619 set_table_entry(&table[11], "name", sd->name,
6620 CORENAME_MAX_SIZE, 0444, proc_dostring);
6621 /* &table[12] is terminator */
6626 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6628 struct ctl_table *entry, *table;
6629 struct sched_domain *sd;
6630 int domain_num = 0, i;
6633 for_each_domain(cpu, sd)
6635 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6640 for_each_domain(cpu, sd) {
6641 snprintf(buf, 32, "domain%d", i);
6642 entry->procname = kstrdup(buf, GFP_KERNEL);
6644 entry->child = sd_alloc_ctl_domain_table(sd);
6651 static struct ctl_table_header *sd_sysctl_header;
6652 static void register_sched_domain_sysctl(void)
6654 int i, cpu_num = num_online_cpus();
6655 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6658 WARN_ON(sd_ctl_dir[0].child);
6659 sd_ctl_dir[0].child = entry;
6664 for_each_online_cpu(i) {
6665 snprintf(buf, 32, "cpu%d", i);
6666 entry->procname = kstrdup(buf, GFP_KERNEL);
6668 entry->child = sd_alloc_ctl_cpu_table(i);
6672 WARN_ON(sd_sysctl_header);
6673 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6676 /* may be called multiple times per register */
6677 static void unregister_sched_domain_sysctl(void)
6679 if (sd_sysctl_header)
6680 unregister_sysctl_table(sd_sysctl_header);
6681 sd_sysctl_header = NULL;
6682 if (sd_ctl_dir[0].child)
6683 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6686 static void register_sched_domain_sysctl(void)
6689 static void unregister_sched_domain_sysctl(void)
6694 static void set_rq_online(struct rq *rq)
6697 const struct sched_class *class;
6699 cpumask_set_cpu(rq->cpu, rq->rd->online);
6702 for_each_class(class) {
6703 if (class->rq_online)
6704 class->rq_online(rq);
6709 static void set_rq_offline(struct rq *rq)
6712 const struct sched_class *class;
6714 for_each_class(class) {
6715 if (class->rq_offline)
6716 class->rq_offline(rq);
6719 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6725 * migration_call - callback that gets triggered when a CPU is added.
6726 * Here we can start up the necessary migration thread for the new CPU.
6728 static int __cpuinit
6729 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6731 struct task_struct *p;
6732 int cpu = (long)hcpu;
6733 unsigned long flags;
6738 case CPU_UP_PREPARE:
6739 case CPU_UP_PREPARE_FROZEN:
6740 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6743 kthread_bind(p, cpu);
6744 /* Must be high prio: stop_machine expects to yield to it. */
6745 rq = task_rq_lock(p, &flags);
6746 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6747 task_rq_unlock(rq, &flags);
6748 cpu_rq(cpu)->migration_thread = p;
6752 case CPU_ONLINE_FROZEN:
6753 /* Strictly unnecessary, as first user will wake it. */
6754 wake_up_process(cpu_rq(cpu)->migration_thread);
6756 /* Update our root-domain */
6758 spin_lock_irqsave(&rq->lock, flags);
6760 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6764 spin_unlock_irqrestore(&rq->lock, flags);
6767 #ifdef CONFIG_HOTPLUG_CPU
6768 case CPU_UP_CANCELED:
6769 case CPU_UP_CANCELED_FROZEN:
6770 if (!cpu_rq(cpu)->migration_thread)
6772 /* Unbind it from offline cpu so it can run. Fall thru. */
6773 kthread_bind(cpu_rq(cpu)->migration_thread,
6774 cpumask_any(cpu_online_mask));
6775 kthread_stop(cpu_rq(cpu)->migration_thread);
6776 cpu_rq(cpu)->migration_thread = NULL;
6780 case CPU_DEAD_FROZEN:
6781 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6782 migrate_live_tasks(cpu);
6784 kthread_stop(rq->migration_thread);
6785 rq->migration_thread = NULL;
6786 /* Idle task back to normal (off runqueue, low prio) */
6787 spin_lock_irq(&rq->lock);
6788 update_rq_clock(rq);
6789 deactivate_task(rq, rq->idle, 0);
6790 rq->idle->static_prio = MAX_PRIO;
6791 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6792 rq->idle->sched_class = &idle_sched_class;
6793 migrate_dead_tasks(cpu);
6794 spin_unlock_irq(&rq->lock);
6796 migrate_nr_uninterruptible(rq);
6797 BUG_ON(rq->nr_running != 0);
6800 * No need to migrate the tasks: it was best-effort if
6801 * they didn't take sched_hotcpu_mutex. Just wake up
6804 spin_lock_irq(&rq->lock);
6805 while (!list_empty(&rq->migration_queue)) {
6806 struct migration_req *req;
6808 req = list_entry(rq->migration_queue.next,
6809 struct migration_req, list);
6810 list_del_init(&req->list);
6811 spin_unlock_irq(&rq->lock);
6812 complete(&req->done);
6813 spin_lock_irq(&rq->lock);
6815 spin_unlock_irq(&rq->lock);
6819 case CPU_DYING_FROZEN:
6820 /* Update our root-domain */
6822 spin_lock_irqsave(&rq->lock, flags);
6824 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6827 spin_unlock_irqrestore(&rq->lock, flags);
6834 /* Register at highest priority so that task migration (migrate_all_tasks)
6835 * happens before everything else.
6837 static struct notifier_block __cpuinitdata migration_notifier = {
6838 .notifier_call = migration_call,
6842 static int __init migration_init(void)
6844 void *cpu = (void *)(long)smp_processor_id();
6847 /* Start one for the boot CPU: */
6848 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6849 BUG_ON(err == NOTIFY_BAD);
6850 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6851 register_cpu_notifier(&migration_notifier);
6855 early_initcall(migration_init);
6860 #ifdef CONFIG_SCHED_DEBUG
6862 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6863 struct cpumask *groupmask)
6865 struct sched_group *group = sd->groups;
6868 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6869 cpumask_clear(groupmask);
6871 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6873 if (!(sd->flags & SD_LOAD_BALANCE)) {
6874 printk("does not load-balance\n");
6876 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6881 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6883 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6884 printk(KERN_ERR "ERROR: domain->span does not contain "
6887 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6888 printk(KERN_ERR "ERROR: domain->groups does not contain"
6892 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6896 printk(KERN_ERR "ERROR: group is NULL\n");
6900 if (!group->__cpu_power) {
6901 printk(KERN_CONT "\n");
6902 printk(KERN_ERR "ERROR: domain->cpu_power not "
6907 if (!cpumask_weight(sched_group_cpus(group))) {
6908 printk(KERN_CONT "\n");
6909 printk(KERN_ERR "ERROR: empty group\n");
6913 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6914 printk(KERN_CONT "\n");
6915 printk(KERN_ERR "ERROR: repeated CPUs\n");
6919 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6921 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6922 printk(KERN_CONT " %s", str);
6924 group = group->next;
6925 } while (group != sd->groups);
6926 printk(KERN_CONT "\n");
6928 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6929 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6932 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6933 printk(KERN_ERR "ERROR: parent span is not a superset "
6934 "of domain->span\n");
6938 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6940 cpumask_var_t groupmask;
6944 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6948 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6950 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6951 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6956 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6963 free_cpumask_var(groupmask);
6965 #else /* !CONFIG_SCHED_DEBUG */
6966 # define sched_domain_debug(sd, cpu) do { } while (0)
6967 #endif /* CONFIG_SCHED_DEBUG */
6969 static int sd_degenerate(struct sched_domain *sd)
6971 if (cpumask_weight(sched_domain_span(sd)) == 1)
6974 /* Following flags need at least 2 groups */
6975 if (sd->flags & (SD_LOAD_BALANCE |
6976 SD_BALANCE_NEWIDLE |
6980 SD_SHARE_PKG_RESOURCES)) {
6981 if (sd->groups != sd->groups->next)
6985 /* Following flags don't use groups */
6986 if (sd->flags & (SD_WAKE_IDLE |
6995 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6997 unsigned long cflags = sd->flags, pflags = parent->flags;
6999 if (sd_degenerate(parent))
7002 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7005 /* Does parent contain flags not in child? */
7006 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7007 if (cflags & SD_WAKE_AFFINE)
7008 pflags &= ~SD_WAKE_BALANCE;
7009 /* Flags needing groups don't count if only 1 group in parent */
7010 if (parent->groups == parent->groups->next) {
7011 pflags &= ~(SD_LOAD_BALANCE |
7012 SD_BALANCE_NEWIDLE |
7016 SD_SHARE_PKG_RESOURCES);
7017 if (nr_node_ids == 1)
7018 pflags &= ~SD_SERIALIZE;
7020 if (~cflags & pflags)
7026 static void free_rootdomain(struct root_domain *rd)
7028 cpupri_cleanup(&rd->cpupri);
7030 free_cpumask_var(rd->rto_mask);
7031 free_cpumask_var(rd->online);
7032 free_cpumask_var(rd->span);
7036 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7038 struct root_domain *old_rd = NULL;
7039 unsigned long flags;
7041 spin_lock_irqsave(&rq->lock, flags);
7046 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7049 cpumask_clear_cpu(rq->cpu, old_rd->span);
7052 * If we dont want to free the old_rt yet then
7053 * set old_rd to NULL to skip the freeing later
7056 if (!atomic_dec_and_test(&old_rd->refcount))
7060 atomic_inc(&rd->refcount);
7063 cpumask_set_cpu(rq->cpu, rd->span);
7064 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7067 spin_unlock_irqrestore(&rq->lock, flags);
7070 free_rootdomain(old_rd);
7073 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7075 memset(rd, 0, sizeof(*rd));
7078 alloc_bootmem_cpumask_var(&def_root_domain.span);
7079 alloc_bootmem_cpumask_var(&def_root_domain.online);
7080 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7081 cpupri_init(&rd->cpupri, true);
7085 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7087 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7089 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7092 if (cpupri_init(&rd->cpupri, false) != 0)
7097 free_cpumask_var(rd->rto_mask);
7099 free_cpumask_var(rd->online);
7101 free_cpumask_var(rd->span);
7106 static void init_defrootdomain(void)
7108 init_rootdomain(&def_root_domain, true);
7110 atomic_set(&def_root_domain.refcount, 1);
7113 static struct root_domain *alloc_rootdomain(void)
7115 struct root_domain *rd;
7117 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7121 if (init_rootdomain(rd, false) != 0) {
7130 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7131 * hold the hotplug lock.
7134 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7136 struct rq *rq = cpu_rq(cpu);
7137 struct sched_domain *tmp;
7139 /* Remove the sched domains which do not contribute to scheduling. */
7140 for (tmp = sd; tmp; ) {
7141 struct sched_domain *parent = tmp->parent;
7145 if (sd_parent_degenerate(tmp, parent)) {
7146 tmp->parent = parent->parent;
7148 parent->parent->child = tmp;
7153 if (sd && sd_degenerate(sd)) {
7159 sched_domain_debug(sd, cpu);
7161 rq_attach_root(rq, rd);
7162 rcu_assign_pointer(rq->sd, sd);
7165 /* cpus with isolated domains */
7166 static cpumask_var_t cpu_isolated_map;
7168 /* Setup the mask of cpus configured for isolated domains */
7169 static int __init isolated_cpu_setup(char *str)
7171 cpulist_parse(str, cpu_isolated_map);
7175 __setup("isolcpus=", isolated_cpu_setup);
7178 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7179 * to a function which identifies what group(along with sched group) a CPU
7180 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7181 * (due to the fact that we keep track of groups covered with a struct cpumask).
7183 * init_sched_build_groups will build a circular linked list of the groups
7184 * covered by the given span, and will set each group's ->cpumask correctly,
7185 * and ->cpu_power to 0.
7188 init_sched_build_groups(const struct cpumask *span,
7189 const struct cpumask *cpu_map,
7190 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7191 struct sched_group **sg,
7192 struct cpumask *tmpmask),
7193 struct cpumask *covered, struct cpumask *tmpmask)
7195 struct sched_group *first = NULL, *last = NULL;
7198 cpumask_clear(covered);
7200 for_each_cpu(i, span) {
7201 struct sched_group *sg;
7202 int group = group_fn(i, cpu_map, &sg, tmpmask);
7205 if (cpumask_test_cpu(i, covered))
7208 cpumask_clear(sched_group_cpus(sg));
7209 sg->__cpu_power = 0;
7211 for_each_cpu(j, span) {
7212 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7215 cpumask_set_cpu(j, covered);
7216 cpumask_set_cpu(j, sched_group_cpus(sg));
7227 #define SD_NODES_PER_DOMAIN 16
7232 * find_next_best_node - find the next node to include in a sched_domain
7233 * @node: node whose sched_domain we're building
7234 * @used_nodes: nodes already in the sched_domain
7236 * Find the next node to include in a given scheduling domain. Simply
7237 * finds the closest node not already in the @used_nodes map.
7239 * Should use nodemask_t.
7241 static int find_next_best_node(int node, nodemask_t *used_nodes)
7243 int i, n, val, min_val, best_node = 0;
7247 for (i = 0; i < nr_node_ids; i++) {
7248 /* Start at @node */
7249 n = (node + i) % nr_node_ids;
7251 if (!nr_cpus_node(n))
7254 /* Skip already used nodes */
7255 if (node_isset(n, *used_nodes))
7258 /* Simple min distance search */
7259 val = node_distance(node, n);
7261 if (val < min_val) {
7267 node_set(best_node, *used_nodes);
7272 * sched_domain_node_span - get a cpumask for a node's sched_domain
7273 * @node: node whose cpumask we're constructing
7274 * @span: resulting cpumask
7276 * Given a node, construct a good cpumask for its sched_domain to span. It
7277 * should be one that prevents unnecessary balancing, but also spreads tasks
7280 static void sched_domain_node_span(int node, struct cpumask *span)
7282 nodemask_t used_nodes;
7285 cpumask_clear(span);
7286 nodes_clear(used_nodes);
7288 cpumask_or(span, span, cpumask_of_node(node));
7289 node_set(node, used_nodes);
7291 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7292 int next_node = find_next_best_node(node, &used_nodes);
7294 cpumask_or(span, span, cpumask_of_node(next_node));
7297 #endif /* CONFIG_NUMA */
7299 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7302 * The cpus mask in sched_group and sched_domain hangs off the end.
7303 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7304 * for nr_cpu_ids < CONFIG_NR_CPUS.
7306 struct static_sched_group {
7307 struct sched_group sg;
7308 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7311 struct static_sched_domain {
7312 struct sched_domain sd;
7313 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7317 * SMT sched-domains:
7319 #ifdef CONFIG_SCHED_SMT
7320 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7321 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7324 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7325 struct sched_group **sg, struct cpumask *unused)
7328 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7331 #endif /* CONFIG_SCHED_SMT */
7334 * multi-core sched-domains:
7336 #ifdef CONFIG_SCHED_MC
7337 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7338 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7339 #endif /* CONFIG_SCHED_MC */
7341 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7343 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7344 struct sched_group **sg, struct cpumask *mask)
7348 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7349 group = cpumask_first(mask);
7351 *sg = &per_cpu(sched_group_core, group).sg;
7354 #elif defined(CONFIG_SCHED_MC)
7356 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7357 struct sched_group **sg, struct cpumask *unused)
7360 *sg = &per_cpu(sched_group_core, cpu).sg;
7365 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7366 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7369 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7370 struct sched_group **sg, struct cpumask *mask)
7373 #ifdef CONFIG_SCHED_MC
7374 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7375 group = cpumask_first(mask);
7376 #elif defined(CONFIG_SCHED_SMT)
7377 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7378 group = cpumask_first(mask);
7383 *sg = &per_cpu(sched_group_phys, group).sg;
7389 * The init_sched_build_groups can't handle what we want to do with node
7390 * groups, so roll our own. Now each node has its own list of groups which
7391 * gets dynamically allocated.
7393 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7394 static struct sched_group ***sched_group_nodes_bycpu;
7396 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7397 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7399 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7400 struct sched_group **sg,
7401 struct cpumask *nodemask)
7405 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7406 group = cpumask_first(nodemask);
7409 *sg = &per_cpu(sched_group_allnodes, group).sg;
7413 static void init_numa_sched_groups_power(struct sched_group *group_head)
7415 struct sched_group *sg = group_head;
7421 for_each_cpu(j, sched_group_cpus(sg)) {
7422 struct sched_domain *sd;
7424 sd = &per_cpu(phys_domains, j).sd;
7425 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7427 * Only add "power" once for each
7433 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7436 } while (sg != group_head);
7438 #endif /* CONFIG_NUMA */
7441 /* Free memory allocated for various sched_group structures */
7442 static void free_sched_groups(const struct cpumask *cpu_map,
7443 struct cpumask *nodemask)
7447 for_each_cpu(cpu, cpu_map) {
7448 struct sched_group **sched_group_nodes
7449 = sched_group_nodes_bycpu[cpu];
7451 if (!sched_group_nodes)
7454 for (i = 0; i < nr_node_ids; i++) {
7455 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7457 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7458 if (cpumask_empty(nodemask))
7468 if (oldsg != sched_group_nodes[i])
7471 kfree(sched_group_nodes);
7472 sched_group_nodes_bycpu[cpu] = NULL;
7475 #else /* !CONFIG_NUMA */
7476 static void free_sched_groups(const struct cpumask *cpu_map,
7477 struct cpumask *nodemask)
7480 #endif /* CONFIG_NUMA */
7483 * Initialize sched groups cpu_power.
7485 * cpu_power indicates the capacity of sched group, which is used while
7486 * distributing the load between different sched groups in a sched domain.
7487 * Typically cpu_power for all the groups in a sched domain will be same unless
7488 * there are asymmetries in the topology. If there are asymmetries, group
7489 * having more cpu_power will pickup more load compared to the group having
7492 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7493 * the maximum number of tasks a group can handle in the presence of other idle
7494 * or lightly loaded groups in the same sched domain.
7496 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7498 struct sched_domain *child;
7499 struct sched_group *group;
7501 WARN_ON(!sd || !sd->groups);
7503 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7508 sd->groups->__cpu_power = 0;
7511 * For perf policy, if the groups in child domain share resources
7512 * (for example cores sharing some portions of the cache hierarchy
7513 * or SMT), then set this domain groups cpu_power such that each group
7514 * can handle only one task, when there are other idle groups in the
7515 * same sched domain.
7517 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7519 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7520 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7525 * add cpu_power of each child group to this groups cpu_power
7527 group = child->groups;
7529 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7530 group = group->next;
7531 } while (group != child->groups);
7535 * Initializers for schedule domains
7536 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7539 #ifdef CONFIG_SCHED_DEBUG
7540 # define SD_INIT_NAME(sd, type) sd->name = #type
7542 # define SD_INIT_NAME(sd, type) do { } while (0)
7545 #define SD_INIT(sd, type) sd_init_##type(sd)
7547 #define SD_INIT_FUNC(type) \
7548 static noinline void sd_init_##type(struct sched_domain *sd) \
7550 memset(sd, 0, sizeof(*sd)); \
7551 *sd = SD_##type##_INIT; \
7552 sd->level = SD_LV_##type; \
7553 SD_INIT_NAME(sd, type); \
7558 SD_INIT_FUNC(ALLNODES)
7561 #ifdef CONFIG_SCHED_SMT
7562 SD_INIT_FUNC(SIBLING)
7564 #ifdef CONFIG_SCHED_MC
7568 static int default_relax_domain_level = -1;
7570 static int __init setup_relax_domain_level(char *str)
7574 val = simple_strtoul(str, NULL, 0);
7575 if (val < SD_LV_MAX)
7576 default_relax_domain_level = val;
7580 __setup("relax_domain_level=", setup_relax_domain_level);
7582 static void set_domain_attribute(struct sched_domain *sd,
7583 struct sched_domain_attr *attr)
7587 if (!attr || attr->relax_domain_level < 0) {
7588 if (default_relax_domain_level < 0)
7591 request = default_relax_domain_level;
7593 request = attr->relax_domain_level;
7594 if (request < sd->level) {
7595 /* turn off idle balance on this domain */
7596 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7598 /* turn on idle balance on this domain */
7599 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7604 * Build sched domains for a given set of cpus and attach the sched domains
7605 * to the individual cpus
7607 static int __build_sched_domains(const struct cpumask *cpu_map,
7608 struct sched_domain_attr *attr)
7610 int i, err = -ENOMEM;
7611 struct root_domain *rd;
7612 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7615 cpumask_var_t domainspan, covered, notcovered;
7616 struct sched_group **sched_group_nodes = NULL;
7617 int sd_allnodes = 0;
7619 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7621 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7622 goto free_domainspan;
7623 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7627 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7628 goto free_notcovered;
7629 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7631 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7632 goto free_this_sibling_map;
7633 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7634 goto free_this_core_map;
7635 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7636 goto free_send_covered;
7640 * Allocate the per-node list of sched groups
7642 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7644 if (!sched_group_nodes) {
7645 printk(KERN_WARNING "Can not alloc sched group node list\n");
7650 rd = alloc_rootdomain();
7652 printk(KERN_WARNING "Cannot alloc root domain\n");
7653 goto free_sched_groups;
7657 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7661 * Set up domains for cpus specified by the cpu_map.
7663 for_each_cpu(i, cpu_map) {
7664 struct sched_domain *sd = NULL, *p;
7666 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7669 if (cpumask_weight(cpu_map) >
7670 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7671 sd = &per_cpu(allnodes_domains, i).sd;
7672 SD_INIT(sd, ALLNODES);
7673 set_domain_attribute(sd, attr);
7674 cpumask_copy(sched_domain_span(sd), cpu_map);
7675 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7681 sd = &per_cpu(node_domains, i).sd;
7683 set_domain_attribute(sd, attr);
7684 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7688 cpumask_and(sched_domain_span(sd),
7689 sched_domain_span(sd), cpu_map);
7693 sd = &per_cpu(phys_domains, i).sd;
7695 set_domain_attribute(sd, attr);
7696 cpumask_copy(sched_domain_span(sd), nodemask);
7700 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7702 #ifdef CONFIG_SCHED_MC
7704 sd = &per_cpu(core_domains, i).sd;
7706 set_domain_attribute(sd, attr);
7707 cpumask_and(sched_domain_span(sd), cpu_map,
7708 cpu_coregroup_mask(i));
7711 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7714 #ifdef CONFIG_SCHED_SMT
7716 sd = &per_cpu(cpu_domains, i).sd;
7717 SD_INIT(sd, SIBLING);
7718 set_domain_attribute(sd, attr);
7719 cpumask_and(sched_domain_span(sd),
7720 &per_cpu(cpu_sibling_map, i), cpu_map);
7723 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7727 #ifdef CONFIG_SCHED_SMT
7728 /* Set up CPU (sibling) groups */
7729 for_each_cpu(i, cpu_map) {
7730 cpumask_and(this_sibling_map,
7731 &per_cpu(cpu_sibling_map, i), cpu_map);
7732 if (i != cpumask_first(this_sibling_map))
7735 init_sched_build_groups(this_sibling_map, cpu_map,
7737 send_covered, tmpmask);
7741 #ifdef CONFIG_SCHED_MC
7742 /* Set up multi-core groups */
7743 for_each_cpu(i, cpu_map) {
7744 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7745 if (i != cpumask_first(this_core_map))
7748 init_sched_build_groups(this_core_map, cpu_map,
7750 send_covered, tmpmask);
7754 /* Set up physical groups */
7755 for (i = 0; i < nr_node_ids; i++) {
7756 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7757 if (cpumask_empty(nodemask))
7760 init_sched_build_groups(nodemask, cpu_map,
7762 send_covered, tmpmask);
7766 /* Set up node groups */
7768 init_sched_build_groups(cpu_map, cpu_map,
7769 &cpu_to_allnodes_group,
7770 send_covered, tmpmask);
7773 for (i = 0; i < nr_node_ids; i++) {
7774 /* Set up node groups */
7775 struct sched_group *sg, *prev;
7778 cpumask_clear(covered);
7779 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7780 if (cpumask_empty(nodemask)) {
7781 sched_group_nodes[i] = NULL;
7785 sched_domain_node_span(i, domainspan);
7786 cpumask_and(domainspan, domainspan, cpu_map);
7788 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7791 printk(KERN_WARNING "Can not alloc domain group for "
7795 sched_group_nodes[i] = sg;
7796 for_each_cpu(j, nodemask) {
7797 struct sched_domain *sd;
7799 sd = &per_cpu(node_domains, j).sd;
7802 sg->__cpu_power = 0;
7803 cpumask_copy(sched_group_cpus(sg), nodemask);
7805 cpumask_or(covered, covered, nodemask);
7808 for (j = 0; j < nr_node_ids; j++) {
7809 int n = (i + j) % nr_node_ids;
7811 cpumask_complement(notcovered, covered);
7812 cpumask_and(tmpmask, notcovered, cpu_map);
7813 cpumask_and(tmpmask, tmpmask, domainspan);
7814 if (cpumask_empty(tmpmask))
7817 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7818 if (cpumask_empty(tmpmask))
7821 sg = kmalloc_node(sizeof(struct sched_group) +
7826 "Can not alloc domain group for node %d\n", j);
7829 sg->__cpu_power = 0;
7830 cpumask_copy(sched_group_cpus(sg), tmpmask);
7831 sg->next = prev->next;
7832 cpumask_or(covered, covered, tmpmask);
7839 /* Calculate CPU power for physical packages and nodes */
7840 #ifdef CONFIG_SCHED_SMT
7841 for_each_cpu(i, cpu_map) {
7842 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7844 init_sched_groups_power(i, sd);
7847 #ifdef CONFIG_SCHED_MC
7848 for_each_cpu(i, cpu_map) {
7849 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7851 init_sched_groups_power(i, sd);
7855 for_each_cpu(i, cpu_map) {
7856 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7858 init_sched_groups_power(i, sd);
7862 for (i = 0; i < nr_node_ids; i++)
7863 init_numa_sched_groups_power(sched_group_nodes[i]);
7866 struct sched_group *sg;
7868 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7870 init_numa_sched_groups_power(sg);
7874 /* Attach the domains */
7875 for_each_cpu(i, cpu_map) {
7876 struct sched_domain *sd;
7877 #ifdef CONFIG_SCHED_SMT
7878 sd = &per_cpu(cpu_domains, i).sd;
7879 #elif defined(CONFIG_SCHED_MC)
7880 sd = &per_cpu(core_domains, i).sd;
7882 sd = &per_cpu(phys_domains, i).sd;
7884 cpu_attach_domain(sd, rd, i);
7890 free_cpumask_var(tmpmask);
7892 free_cpumask_var(send_covered);
7894 free_cpumask_var(this_core_map);
7895 free_this_sibling_map:
7896 free_cpumask_var(this_sibling_map);
7898 free_cpumask_var(nodemask);
7901 free_cpumask_var(notcovered);
7903 free_cpumask_var(covered);
7905 free_cpumask_var(domainspan);
7912 kfree(sched_group_nodes);
7918 free_sched_groups(cpu_map, tmpmask);
7919 free_rootdomain(rd);
7924 static int build_sched_domains(const struct cpumask *cpu_map)
7926 return __build_sched_domains(cpu_map, NULL);
7929 static struct cpumask *doms_cur; /* current sched domains */
7930 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7931 static struct sched_domain_attr *dattr_cur;
7932 /* attribues of custom domains in 'doms_cur' */
7935 * Special case: If a kmalloc of a doms_cur partition (array of
7936 * cpumask) fails, then fallback to a single sched domain,
7937 * as determined by the single cpumask fallback_doms.
7939 static cpumask_var_t fallback_doms;
7942 * arch_update_cpu_topology lets virtualized architectures update the
7943 * cpu core maps. It is supposed to return 1 if the topology changed
7944 * or 0 if it stayed the same.
7946 int __attribute__((weak)) arch_update_cpu_topology(void)
7952 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7953 * For now this just excludes isolated cpus, but could be used to
7954 * exclude other special cases in the future.
7956 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7960 arch_update_cpu_topology();
7962 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7964 doms_cur = fallback_doms;
7965 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7967 err = build_sched_domains(doms_cur);
7968 register_sched_domain_sysctl();
7973 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7974 struct cpumask *tmpmask)
7976 free_sched_groups(cpu_map, tmpmask);
7980 * Detach sched domains from a group of cpus specified in cpu_map
7981 * These cpus will now be attached to the NULL domain
7983 static void detach_destroy_domains(const struct cpumask *cpu_map)
7985 /* Save because hotplug lock held. */
7986 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7989 for_each_cpu(i, cpu_map)
7990 cpu_attach_domain(NULL, &def_root_domain, i);
7991 synchronize_sched();
7992 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7995 /* handle null as "default" */
7996 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7997 struct sched_domain_attr *new, int idx_new)
7999 struct sched_domain_attr tmp;
8006 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8007 new ? (new + idx_new) : &tmp,
8008 sizeof(struct sched_domain_attr));
8012 * Partition sched domains as specified by the 'ndoms_new'
8013 * cpumasks in the array doms_new[] of cpumasks. This compares
8014 * doms_new[] to the current sched domain partitioning, doms_cur[].
8015 * It destroys each deleted domain and builds each new domain.
8017 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8018 * The masks don't intersect (don't overlap.) We should setup one
8019 * sched domain for each mask. CPUs not in any of the cpumasks will
8020 * not be load balanced. If the same cpumask appears both in the
8021 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8024 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8025 * ownership of it and will kfree it when done with it. If the caller
8026 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8027 * ndoms_new == 1, and partition_sched_domains() will fallback to
8028 * the single partition 'fallback_doms', it also forces the domains
8031 * If doms_new == NULL it will be replaced with cpu_online_mask.
8032 * ndoms_new == 0 is a special case for destroying existing domains,
8033 * and it will not create the default domain.
8035 * Call with hotplug lock held
8037 /* FIXME: Change to struct cpumask *doms_new[] */
8038 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8039 struct sched_domain_attr *dattr_new)
8044 mutex_lock(&sched_domains_mutex);
8046 /* always unregister in case we don't destroy any domains */
8047 unregister_sched_domain_sysctl();
8049 /* Let architecture update cpu core mappings. */
8050 new_topology = arch_update_cpu_topology();
8052 n = doms_new ? ndoms_new : 0;
8054 /* Destroy deleted domains */
8055 for (i = 0; i < ndoms_cur; i++) {
8056 for (j = 0; j < n && !new_topology; j++) {
8057 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8058 && dattrs_equal(dattr_cur, i, dattr_new, j))
8061 /* no match - a current sched domain not in new doms_new[] */
8062 detach_destroy_domains(doms_cur + i);
8067 if (doms_new == NULL) {
8069 doms_new = fallback_doms;
8070 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8071 WARN_ON_ONCE(dattr_new);
8074 /* Build new domains */
8075 for (i = 0; i < ndoms_new; i++) {
8076 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8077 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8078 && dattrs_equal(dattr_new, i, dattr_cur, j))
8081 /* no match - add a new doms_new */
8082 __build_sched_domains(doms_new + i,
8083 dattr_new ? dattr_new + i : NULL);
8088 /* Remember the new sched domains */
8089 if (doms_cur != fallback_doms)
8091 kfree(dattr_cur); /* kfree(NULL) is safe */
8092 doms_cur = doms_new;
8093 dattr_cur = dattr_new;
8094 ndoms_cur = ndoms_new;
8096 register_sched_domain_sysctl();
8098 mutex_unlock(&sched_domains_mutex);
8101 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8102 static void arch_reinit_sched_domains(void)
8106 /* Destroy domains first to force the rebuild */
8107 partition_sched_domains(0, NULL, NULL);
8109 rebuild_sched_domains();
8113 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8115 unsigned int level = 0;
8117 if (sscanf(buf, "%u", &level) != 1)
8121 * level is always be positive so don't check for
8122 * level < POWERSAVINGS_BALANCE_NONE which is 0
8123 * What happens on 0 or 1 byte write,
8124 * need to check for count as well?
8127 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8131 sched_smt_power_savings = level;
8133 sched_mc_power_savings = level;
8135 arch_reinit_sched_domains();
8140 #ifdef CONFIG_SCHED_MC
8141 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8144 return sprintf(page, "%u\n", sched_mc_power_savings);
8146 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8147 const char *buf, size_t count)
8149 return sched_power_savings_store(buf, count, 0);
8151 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8152 sched_mc_power_savings_show,
8153 sched_mc_power_savings_store);
8156 #ifdef CONFIG_SCHED_SMT
8157 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8160 return sprintf(page, "%u\n", sched_smt_power_savings);
8162 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8163 const char *buf, size_t count)
8165 return sched_power_savings_store(buf, count, 1);
8167 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8168 sched_smt_power_savings_show,
8169 sched_smt_power_savings_store);
8172 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8176 #ifdef CONFIG_SCHED_SMT
8178 err = sysfs_create_file(&cls->kset.kobj,
8179 &attr_sched_smt_power_savings.attr);
8181 #ifdef CONFIG_SCHED_MC
8182 if (!err && mc_capable())
8183 err = sysfs_create_file(&cls->kset.kobj,
8184 &attr_sched_mc_power_savings.attr);
8188 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8190 #ifndef CONFIG_CPUSETS
8192 * Add online and remove offline CPUs from the scheduler domains.
8193 * When cpusets are enabled they take over this function.
8195 static int update_sched_domains(struct notifier_block *nfb,
8196 unsigned long action, void *hcpu)
8200 case CPU_ONLINE_FROZEN:
8202 case CPU_DEAD_FROZEN:
8203 partition_sched_domains(1, NULL, NULL);
8212 static int update_runtime(struct notifier_block *nfb,
8213 unsigned long action, void *hcpu)
8215 int cpu = (int)(long)hcpu;
8218 case CPU_DOWN_PREPARE:
8219 case CPU_DOWN_PREPARE_FROZEN:
8220 disable_runtime(cpu_rq(cpu));
8223 case CPU_DOWN_FAILED:
8224 case CPU_DOWN_FAILED_FROZEN:
8226 case CPU_ONLINE_FROZEN:
8227 enable_runtime(cpu_rq(cpu));
8235 void __init sched_init_smp(void)
8237 cpumask_var_t non_isolated_cpus;
8239 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8241 #if defined(CONFIG_NUMA)
8242 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8244 BUG_ON(sched_group_nodes_bycpu == NULL);
8247 mutex_lock(&sched_domains_mutex);
8248 arch_init_sched_domains(cpu_online_mask);
8249 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8250 if (cpumask_empty(non_isolated_cpus))
8251 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8252 mutex_unlock(&sched_domains_mutex);
8255 #ifndef CONFIG_CPUSETS
8256 /* XXX: Theoretical race here - CPU may be hotplugged now */
8257 hotcpu_notifier(update_sched_domains, 0);
8260 /* RT runtime code needs to handle some hotplug events */
8261 hotcpu_notifier(update_runtime, 0);
8265 /* Move init over to a non-isolated CPU */
8266 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8268 sched_init_granularity();
8269 free_cpumask_var(non_isolated_cpus);
8271 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8272 init_sched_rt_class();
8275 void __init sched_init_smp(void)
8277 sched_init_granularity();
8279 #endif /* CONFIG_SMP */
8281 int in_sched_functions(unsigned long addr)
8283 return in_lock_functions(addr) ||
8284 (addr >= (unsigned long)__sched_text_start
8285 && addr < (unsigned long)__sched_text_end);
8288 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8290 cfs_rq->tasks_timeline = RB_ROOT;
8291 INIT_LIST_HEAD(&cfs_rq->tasks);
8292 #ifdef CONFIG_FAIR_GROUP_SCHED
8295 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8298 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8300 struct rt_prio_array *array;
8303 array = &rt_rq->active;
8304 for (i = 0; i < MAX_RT_PRIO; i++) {
8305 INIT_LIST_HEAD(array->queue + i);
8306 __clear_bit(i, array->bitmap);
8308 /* delimiter for bitsearch: */
8309 __set_bit(MAX_RT_PRIO, array->bitmap);
8311 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8312 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8314 rt_rq->highest_prio.next = MAX_RT_PRIO;
8318 rt_rq->rt_nr_migratory = 0;
8319 rt_rq->overloaded = 0;
8320 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8324 rt_rq->rt_throttled = 0;
8325 rt_rq->rt_runtime = 0;
8326 spin_lock_init(&rt_rq->rt_runtime_lock);
8328 #ifdef CONFIG_RT_GROUP_SCHED
8329 rt_rq->rt_nr_boosted = 0;
8334 #ifdef CONFIG_FAIR_GROUP_SCHED
8335 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8336 struct sched_entity *se, int cpu, int add,
8337 struct sched_entity *parent)
8339 struct rq *rq = cpu_rq(cpu);
8340 tg->cfs_rq[cpu] = cfs_rq;
8341 init_cfs_rq(cfs_rq, rq);
8344 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8347 /* se could be NULL for init_task_group */
8352 se->cfs_rq = &rq->cfs;
8354 se->cfs_rq = parent->my_q;
8357 se->load.weight = tg->shares;
8358 se->load.inv_weight = 0;
8359 se->parent = parent;
8363 #ifdef CONFIG_RT_GROUP_SCHED
8364 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8365 struct sched_rt_entity *rt_se, int cpu, int add,
8366 struct sched_rt_entity *parent)
8368 struct rq *rq = cpu_rq(cpu);
8370 tg->rt_rq[cpu] = rt_rq;
8371 init_rt_rq(rt_rq, rq);
8373 rt_rq->rt_se = rt_se;
8374 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8376 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8378 tg->rt_se[cpu] = rt_se;
8383 rt_se->rt_rq = &rq->rt;
8385 rt_se->rt_rq = parent->my_q;
8387 rt_se->my_q = rt_rq;
8388 rt_se->parent = parent;
8389 INIT_LIST_HEAD(&rt_se->run_list);
8393 void __init sched_init(void)
8396 unsigned long alloc_size = 0, ptr;
8398 #ifdef CONFIG_FAIR_GROUP_SCHED
8399 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8401 #ifdef CONFIG_RT_GROUP_SCHED
8402 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8404 #ifdef CONFIG_USER_SCHED
8408 * As sched_init() is called before page_alloc is setup,
8409 * we use alloc_bootmem().
8412 ptr = (unsigned long)alloc_bootmem(alloc_size);
8414 #ifdef CONFIG_FAIR_GROUP_SCHED
8415 init_task_group.se = (struct sched_entity **)ptr;
8416 ptr += nr_cpu_ids * sizeof(void **);
8418 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8419 ptr += nr_cpu_ids * sizeof(void **);
8421 #ifdef CONFIG_USER_SCHED
8422 root_task_group.se = (struct sched_entity **)ptr;
8423 ptr += nr_cpu_ids * sizeof(void **);
8425 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8426 ptr += nr_cpu_ids * sizeof(void **);
8427 #endif /* CONFIG_USER_SCHED */
8428 #endif /* CONFIG_FAIR_GROUP_SCHED */
8429 #ifdef CONFIG_RT_GROUP_SCHED
8430 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8431 ptr += nr_cpu_ids * sizeof(void **);
8433 init_task_group.rt_rq = (struct rt_rq **)ptr;
8434 ptr += nr_cpu_ids * sizeof(void **);
8436 #ifdef CONFIG_USER_SCHED
8437 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8438 ptr += nr_cpu_ids * sizeof(void **);
8440 root_task_group.rt_rq = (struct rt_rq **)ptr;
8441 ptr += nr_cpu_ids * sizeof(void **);
8442 #endif /* CONFIG_USER_SCHED */
8443 #endif /* CONFIG_RT_GROUP_SCHED */
8447 init_defrootdomain();
8450 init_rt_bandwidth(&def_rt_bandwidth,
8451 global_rt_period(), global_rt_runtime());
8453 #ifdef CONFIG_RT_GROUP_SCHED
8454 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8455 global_rt_period(), global_rt_runtime());
8456 #ifdef CONFIG_USER_SCHED
8457 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8458 global_rt_period(), RUNTIME_INF);
8459 #endif /* CONFIG_USER_SCHED */
8460 #endif /* CONFIG_RT_GROUP_SCHED */
8462 #ifdef CONFIG_GROUP_SCHED
8463 list_add(&init_task_group.list, &task_groups);
8464 INIT_LIST_HEAD(&init_task_group.children);
8466 #ifdef CONFIG_USER_SCHED
8467 INIT_LIST_HEAD(&root_task_group.children);
8468 init_task_group.parent = &root_task_group;
8469 list_add(&init_task_group.siblings, &root_task_group.children);
8470 #endif /* CONFIG_USER_SCHED */
8471 #endif /* CONFIG_GROUP_SCHED */
8473 for_each_possible_cpu(i) {
8477 spin_lock_init(&rq->lock);
8479 init_cfs_rq(&rq->cfs, rq);
8480 init_rt_rq(&rq->rt, rq);
8481 #ifdef CONFIG_FAIR_GROUP_SCHED
8482 init_task_group.shares = init_task_group_load;
8483 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8484 #ifdef CONFIG_CGROUP_SCHED
8486 * How much cpu bandwidth does init_task_group get?
8488 * In case of task-groups formed thr' the cgroup filesystem, it
8489 * gets 100% of the cpu resources in the system. This overall
8490 * system cpu resource is divided among the tasks of
8491 * init_task_group and its child task-groups in a fair manner,
8492 * based on each entity's (task or task-group's) weight
8493 * (se->load.weight).
8495 * In other words, if init_task_group has 10 tasks of weight
8496 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8497 * then A0's share of the cpu resource is:
8499 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8501 * We achieve this by letting init_task_group's tasks sit
8502 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8504 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8505 #elif defined CONFIG_USER_SCHED
8506 root_task_group.shares = NICE_0_LOAD;
8507 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8509 * In case of task-groups formed thr' the user id of tasks,
8510 * init_task_group represents tasks belonging to root user.
8511 * Hence it forms a sibling of all subsequent groups formed.
8512 * In this case, init_task_group gets only a fraction of overall
8513 * system cpu resource, based on the weight assigned to root
8514 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8515 * by letting tasks of init_task_group sit in a separate cfs_rq
8516 * (init_cfs_rq) and having one entity represent this group of
8517 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8519 init_tg_cfs_entry(&init_task_group,
8520 &per_cpu(init_cfs_rq, i),
8521 &per_cpu(init_sched_entity, i), i, 1,
8522 root_task_group.se[i]);
8525 #endif /* CONFIG_FAIR_GROUP_SCHED */
8527 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8528 #ifdef CONFIG_RT_GROUP_SCHED
8529 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8530 #ifdef CONFIG_CGROUP_SCHED
8531 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8532 #elif defined CONFIG_USER_SCHED
8533 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8534 init_tg_rt_entry(&init_task_group,
8535 &per_cpu(init_rt_rq, i),
8536 &per_cpu(init_sched_rt_entity, i), i, 1,
8537 root_task_group.rt_se[i]);
8541 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8542 rq->cpu_load[j] = 0;
8546 rq->active_balance = 0;
8547 rq->next_balance = jiffies;
8551 rq->migration_thread = NULL;
8552 INIT_LIST_HEAD(&rq->migration_queue);
8553 rq_attach_root(rq, &def_root_domain);
8556 atomic_set(&rq->nr_iowait, 0);
8559 set_load_weight(&init_task);
8561 #ifdef CONFIG_PREEMPT_NOTIFIERS
8562 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8566 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8569 #ifdef CONFIG_RT_MUTEXES
8570 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8574 * The boot idle thread does lazy MMU switching as well:
8576 atomic_inc(&init_mm.mm_count);
8577 enter_lazy_tlb(&init_mm, current);
8580 * Make us the idle thread. Technically, schedule() should not be
8581 * called from this thread, however somewhere below it might be,
8582 * but because we are the idle thread, we just pick up running again
8583 * when this runqueue becomes "idle".
8585 init_idle(current, smp_processor_id());
8587 * During early bootup we pretend to be a normal task:
8589 current->sched_class = &fair_sched_class;
8591 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8592 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8595 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8597 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8600 scheduler_running = 1;
8603 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8604 void __might_sleep(char *file, int line)
8607 static unsigned long prev_jiffy; /* ratelimiting */
8609 if ((!in_atomic() && !irqs_disabled()) ||
8610 system_state != SYSTEM_RUNNING || oops_in_progress)
8612 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8614 prev_jiffy = jiffies;
8617 "BUG: sleeping function called from invalid context at %s:%d\n",
8620 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8621 in_atomic(), irqs_disabled(),
8622 current->pid, current->comm);
8624 debug_show_held_locks(current);
8625 if (irqs_disabled())
8626 print_irqtrace_events(current);
8630 EXPORT_SYMBOL(__might_sleep);
8633 #ifdef CONFIG_MAGIC_SYSRQ
8634 static void normalize_task(struct rq *rq, struct task_struct *p)
8638 update_rq_clock(rq);
8639 on_rq = p->se.on_rq;
8641 deactivate_task(rq, p, 0);
8642 __setscheduler(rq, p, SCHED_NORMAL, 0);
8644 activate_task(rq, p, 0);
8645 resched_task(rq->curr);
8649 void normalize_rt_tasks(void)
8651 struct task_struct *g, *p;
8652 unsigned long flags;
8655 read_lock_irqsave(&tasklist_lock, flags);
8656 do_each_thread(g, p) {
8658 * Only normalize user tasks:
8663 p->se.exec_start = 0;
8664 #ifdef CONFIG_SCHEDSTATS
8665 p->se.wait_start = 0;
8666 p->se.sleep_start = 0;
8667 p->se.block_start = 0;
8672 * Renice negative nice level userspace
8675 if (TASK_NICE(p) < 0 && p->mm)
8676 set_user_nice(p, 0);
8680 spin_lock(&p->pi_lock);
8681 rq = __task_rq_lock(p);
8683 normalize_task(rq, p);
8685 __task_rq_unlock(rq);
8686 spin_unlock(&p->pi_lock);
8687 } while_each_thread(g, p);
8689 read_unlock_irqrestore(&tasklist_lock, flags);
8692 #endif /* CONFIG_MAGIC_SYSRQ */
8696 * These functions are only useful for the IA64 MCA handling.
8698 * They can only be called when the whole system has been
8699 * stopped - every CPU needs to be quiescent, and no scheduling
8700 * activity can take place. Using them for anything else would
8701 * be a serious bug, and as a result, they aren't even visible
8702 * under any other configuration.
8706 * curr_task - return the current task for a given cpu.
8707 * @cpu: the processor in question.
8709 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8711 struct task_struct *curr_task(int cpu)
8713 return cpu_curr(cpu);
8717 * set_curr_task - set the current task for a given cpu.
8718 * @cpu: the processor in question.
8719 * @p: the task pointer to set.
8721 * Description: This function must only be used when non-maskable interrupts
8722 * are serviced on a separate stack. It allows the architecture to switch the
8723 * notion of the current task on a cpu in a non-blocking manner. This function
8724 * must be called with all CPU's synchronized, and interrupts disabled, the
8725 * and caller must save the original value of the current task (see
8726 * curr_task() above) and restore that value before reenabling interrupts and
8727 * re-starting the system.
8729 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8731 void set_curr_task(int cpu, struct task_struct *p)
8738 #ifdef CONFIG_FAIR_GROUP_SCHED
8739 static void free_fair_sched_group(struct task_group *tg)
8743 for_each_possible_cpu(i) {
8745 kfree(tg->cfs_rq[i]);
8755 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8757 struct cfs_rq *cfs_rq;
8758 struct sched_entity *se;
8762 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8765 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8769 tg->shares = NICE_0_LOAD;
8771 for_each_possible_cpu(i) {
8774 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8775 GFP_KERNEL, cpu_to_node(i));
8779 se = kzalloc_node(sizeof(struct sched_entity),
8780 GFP_KERNEL, cpu_to_node(i));
8784 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8793 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8795 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8796 &cpu_rq(cpu)->leaf_cfs_rq_list);
8799 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8801 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8803 #else /* !CONFG_FAIR_GROUP_SCHED */
8804 static inline void free_fair_sched_group(struct task_group *tg)
8809 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8814 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8818 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8821 #endif /* CONFIG_FAIR_GROUP_SCHED */
8823 #ifdef CONFIG_RT_GROUP_SCHED
8824 static void free_rt_sched_group(struct task_group *tg)
8828 destroy_rt_bandwidth(&tg->rt_bandwidth);
8830 for_each_possible_cpu(i) {
8832 kfree(tg->rt_rq[i]);
8834 kfree(tg->rt_se[i]);
8842 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8844 struct rt_rq *rt_rq;
8845 struct sched_rt_entity *rt_se;
8849 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8852 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8856 init_rt_bandwidth(&tg->rt_bandwidth,
8857 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8859 for_each_possible_cpu(i) {
8862 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8863 GFP_KERNEL, cpu_to_node(i));
8867 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8868 GFP_KERNEL, cpu_to_node(i));
8872 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8881 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8883 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8884 &cpu_rq(cpu)->leaf_rt_rq_list);
8887 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8889 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8891 #else /* !CONFIG_RT_GROUP_SCHED */
8892 static inline void free_rt_sched_group(struct task_group *tg)
8897 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8902 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8906 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8909 #endif /* CONFIG_RT_GROUP_SCHED */
8911 #ifdef CONFIG_GROUP_SCHED
8912 static void free_sched_group(struct task_group *tg)
8914 free_fair_sched_group(tg);
8915 free_rt_sched_group(tg);
8919 /* allocate runqueue etc for a new task group */
8920 struct task_group *sched_create_group(struct task_group *parent)
8922 struct task_group *tg;
8923 unsigned long flags;
8926 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8928 return ERR_PTR(-ENOMEM);
8930 if (!alloc_fair_sched_group(tg, parent))
8933 if (!alloc_rt_sched_group(tg, parent))
8936 spin_lock_irqsave(&task_group_lock, flags);
8937 for_each_possible_cpu(i) {
8938 register_fair_sched_group(tg, i);
8939 register_rt_sched_group(tg, i);
8941 list_add_rcu(&tg->list, &task_groups);
8943 WARN_ON(!parent); /* root should already exist */
8945 tg->parent = parent;
8946 INIT_LIST_HEAD(&tg->children);
8947 list_add_rcu(&tg->siblings, &parent->children);
8948 spin_unlock_irqrestore(&task_group_lock, flags);
8953 free_sched_group(tg);
8954 return ERR_PTR(-ENOMEM);
8957 /* rcu callback to free various structures associated with a task group */
8958 static void free_sched_group_rcu(struct rcu_head *rhp)
8960 /* now it should be safe to free those cfs_rqs */
8961 free_sched_group(container_of(rhp, struct task_group, rcu));
8964 /* Destroy runqueue etc associated with a task group */
8965 void sched_destroy_group(struct task_group *tg)
8967 unsigned long flags;
8970 spin_lock_irqsave(&task_group_lock, flags);
8971 for_each_possible_cpu(i) {
8972 unregister_fair_sched_group(tg, i);
8973 unregister_rt_sched_group(tg, i);
8975 list_del_rcu(&tg->list);
8976 list_del_rcu(&tg->siblings);
8977 spin_unlock_irqrestore(&task_group_lock, flags);
8979 /* wait for possible concurrent references to cfs_rqs complete */
8980 call_rcu(&tg->rcu, free_sched_group_rcu);
8983 /* change task's runqueue when it moves between groups.
8984 * The caller of this function should have put the task in its new group
8985 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8986 * reflect its new group.
8988 void sched_move_task(struct task_struct *tsk)
8991 unsigned long flags;
8994 rq = task_rq_lock(tsk, &flags);
8996 update_rq_clock(rq);
8998 running = task_current(rq, tsk);
8999 on_rq = tsk->se.on_rq;
9002 dequeue_task(rq, tsk, 0);
9003 if (unlikely(running))
9004 tsk->sched_class->put_prev_task(rq, tsk);
9006 set_task_rq(tsk, task_cpu(tsk));
9008 #ifdef CONFIG_FAIR_GROUP_SCHED
9009 if (tsk->sched_class->moved_group)
9010 tsk->sched_class->moved_group(tsk);
9013 if (unlikely(running))
9014 tsk->sched_class->set_curr_task(rq);
9016 enqueue_task(rq, tsk, 0);
9018 task_rq_unlock(rq, &flags);
9020 #endif /* CONFIG_GROUP_SCHED */
9022 #ifdef CONFIG_FAIR_GROUP_SCHED
9023 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9025 struct cfs_rq *cfs_rq = se->cfs_rq;
9030 dequeue_entity(cfs_rq, se, 0);
9032 se->load.weight = shares;
9033 se->load.inv_weight = 0;
9036 enqueue_entity(cfs_rq, se, 0);
9039 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9041 struct cfs_rq *cfs_rq = se->cfs_rq;
9042 struct rq *rq = cfs_rq->rq;
9043 unsigned long flags;
9045 spin_lock_irqsave(&rq->lock, flags);
9046 __set_se_shares(se, shares);
9047 spin_unlock_irqrestore(&rq->lock, flags);
9050 static DEFINE_MUTEX(shares_mutex);
9052 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9055 unsigned long flags;
9058 * We can't change the weight of the root cgroup.
9063 if (shares < MIN_SHARES)
9064 shares = MIN_SHARES;
9065 else if (shares > MAX_SHARES)
9066 shares = MAX_SHARES;
9068 mutex_lock(&shares_mutex);
9069 if (tg->shares == shares)
9072 spin_lock_irqsave(&task_group_lock, flags);
9073 for_each_possible_cpu(i)
9074 unregister_fair_sched_group(tg, i);
9075 list_del_rcu(&tg->siblings);
9076 spin_unlock_irqrestore(&task_group_lock, flags);
9078 /* wait for any ongoing reference to this group to finish */
9079 synchronize_sched();
9082 * Now we are free to modify the group's share on each cpu
9083 * w/o tripping rebalance_share or load_balance_fair.
9085 tg->shares = shares;
9086 for_each_possible_cpu(i) {
9090 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9091 set_se_shares(tg->se[i], shares);
9095 * Enable load balance activity on this group, by inserting it back on
9096 * each cpu's rq->leaf_cfs_rq_list.
9098 spin_lock_irqsave(&task_group_lock, flags);
9099 for_each_possible_cpu(i)
9100 register_fair_sched_group(tg, i);
9101 list_add_rcu(&tg->siblings, &tg->parent->children);
9102 spin_unlock_irqrestore(&task_group_lock, flags);
9104 mutex_unlock(&shares_mutex);
9108 unsigned long sched_group_shares(struct task_group *tg)
9114 #ifdef CONFIG_RT_GROUP_SCHED
9116 * Ensure that the real time constraints are schedulable.
9118 static DEFINE_MUTEX(rt_constraints_mutex);
9120 static unsigned long to_ratio(u64 period, u64 runtime)
9122 if (runtime == RUNTIME_INF)
9125 return div64_u64(runtime << 20, period);
9128 /* Must be called with tasklist_lock held */
9129 static inline int tg_has_rt_tasks(struct task_group *tg)
9131 struct task_struct *g, *p;
9133 do_each_thread(g, p) {
9134 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9136 } while_each_thread(g, p);
9141 struct rt_schedulable_data {
9142 struct task_group *tg;
9147 static int tg_schedulable(struct task_group *tg, void *data)
9149 struct rt_schedulable_data *d = data;
9150 struct task_group *child;
9151 unsigned long total, sum = 0;
9152 u64 period, runtime;
9154 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9155 runtime = tg->rt_bandwidth.rt_runtime;
9158 period = d->rt_period;
9159 runtime = d->rt_runtime;
9162 #ifdef CONFIG_USER_SCHED
9163 if (tg == &root_task_group) {
9164 period = global_rt_period();
9165 runtime = global_rt_runtime();
9170 * Cannot have more runtime than the period.
9172 if (runtime > period && runtime != RUNTIME_INF)
9176 * Ensure we don't starve existing RT tasks.
9178 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9181 total = to_ratio(period, runtime);
9184 * Nobody can have more than the global setting allows.
9186 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9190 * The sum of our children's runtime should not exceed our own.
9192 list_for_each_entry_rcu(child, &tg->children, siblings) {
9193 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9194 runtime = child->rt_bandwidth.rt_runtime;
9196 if (child == d->tg) {
9197 period = d->rt_period;
9198 runtime = d->rt_runtime;
9201 sum += to_ratio(period, runtime);
9210 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9212 struct rt_schedulable_data data = {
9214 .rt_period = period,
9215 .rt_runtime = runtime,
9218 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9221 static int tg_set_bandwidth(struct task_group *tg,
9222 u64 rt_period, u64 rt_runtime)
9226 mutex_lock(&rt_constraints_mutex);
9227 read_lock(&tasklist_lock);
9228 err = __rt_schedulable(tg, rt_period, rt_runtime);
9232 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9233 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9234 tg->rt_bandwidth.rt_runtime = rt_runtime;
9236 for_each_possible_cpu(i) {
9237 struct rt_rq *rt_rq = tg->rt_rq[i];
9239 spin_lock(&rt_rq->rt_runtime_lock);
9240 rt_rq->rt_runtime = rt_runtime;
9241 spin_unlock(&rt_rq->rt_runtime_lock);
9243 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9245 read_unlock(&tasklist_lock);
9246 mutex_unlock(&rt_constraints_mutex);
9251 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9253 u64 rt_runtime, rt_period;
9255 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9256 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9257 if (rt_runtime_us < 0)
9258 rt_runtime = RUNTIME_INF;
9260 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9263 long sched_group_rt_runtime(struct task_group *tg)
9267 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9270 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9271 do_div(rt_runtime_us, NSEC_PER_USEC);
9272 return rt_runtime_us;
9275 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9277 u64 rt_runtime, rt_period;
9279 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9280 rt_runtime = tg->rt_bandwidth.rt_runtime;
9285 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9288 long sched_group_rt_period(struct task_group *tg)
9292 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9293 do_div(rt_period_us, NSEC_PER_USEC);
9294 return rt_period_us;
9297 static int sched_rt_global_constraints(void)
9299 u64 runtime, period;
9302 if (sysctl_sched_rt_period <= 0)
9305 runtime = global_rt_runtime();
9306 period = global_rt_period();
9309 * Sanity check on the sysctl variables.
9311 if (runtime > period && runtime != RUNTIME_INF)
9314 mutex_lock(&rt_constraints_mutex);
9315 read_lock(&tasklist_lock);
9316 ret = __rt_schedulable(NULL, 0, 0);
9317 read_unlock(&tasklist_lock);
9318 mutex_unlock(&rt_constraints_mutex);
9322 #else /* !CONFIG_RT_GROUP_SCHED */
9323 static int sched_rt_global_constraints(void)
9325 unsigned long flags;
9328 if (sysctl_sched_rt_period <= 0)
9331 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9332 for_each_possible_cpu(i) {
9333 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9335 spin_lock(&rt_rq->rt_runtime_lock);
9336 rt_rq->rt_runtime = global_rt_runtime();
9337 spin_unlock(&rt_rq->rt_runtime_lock);
9339 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9343 #endif /* CONFIG_RT_GROUP_SCHED */
9345 int sched_rt_handler(struct ctl_table *table, int write,
9346 struct file *filp, void __user *buffer, size_t *lenp,
9350 int old_period, old_runtime;
9351 static DEFINE_MUTEX(mutex);
9354 old_period = sysctl_sched_rt_period;
9355 old_runtime = sysctl_sched_rt_runtime;
9357 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9359 if (!ret && write) {
9360 ret = sched_rt_global_constraints();
9362 sysctl_sched_rt_period = old_period;
9363 sysctl_sched_rt_runtime = old_runtime;
9365 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9366 def_rt_bandwidth.rt_period =
9367 ns_to_ktime(global_rt_period());
9370 mutex_unlock(&mutex);
9375 #ifdef CONFIG_CGROUP_SCHED
9377 /* return corresponding task_group object of a cgroup */
9378 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9380 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9381 struct task_group, css);
9384 static struct cgroup_subsys_state *
9385 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9387 struct task_group *tg, *parent;
9389 if (!cgrp->parent) {
9390 /* This is early initialization for the top cgroup */
9391 return &init_task_group.css;
9394 parent = cgroup_tg(cgrp->parent);
9395 tg = sched_create_group(parent);
9397 return ERR_PTR(-ENOMEM);
9403 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9405 struct task_group *tg = cgroup_tg(cgrp);
9407 sched_destroy_group(tg);
9411 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9412 struct task_struct *tsk)
9414 #ifdef CONFIG_RT_GROUP_SCHED
9415 /* Don't accept realtime tasks when there is no way for them to run */
9416 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9419 /* We don't support RT-tasks being in separate groups */
9420 if (tsk->sched_class != &fair_sched_class)
9428 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9429 struct cgroup *old_cont, struct task_struct *tsk)
9431 sched_move_task(tsk);
9434 #ifdef CONFIG_FAIR_GROUP_SCHED
9435 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9438 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9441 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9443 struct task_group *tg = cgroup_tg(cgrp);
9445 return (u64) tg->shares;
9447 #endif /* CONFIG_FAIR_GROUP_SCHED */
9449 #ifdef CONFIG_RT_GROUP_SCHED
9450 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9453 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9456 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9458 return sched_group_rt_runtime(cgroup_tg(cgrp));
9461 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9464 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9467 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9469 return sched_group_rt_period(cgroup_tg(cgrp));
9471 #endif /* CONFIG_RT_GROUP_SCHED */
9473 static struct cftype cpu_files[] = {
9474 #ifdef CONFIG_FAIR_GROUP_SCHED
9477 .read_u64 = cpu_shares_read_u64,
9478 .write_u64 = cpu_shares_write_u64,
9481 #ifdef CONFIG_RT_GROUP_SCHED
9483 .name = "rt_runtime_us",
9484 .read_s64 = cpu_rt_runtime_read,
9485 .write_s64 = cpu_rt_runtime_write,
9488 .name = "rt_period_us",
9489 .read_u64 = cpu_rt_period_read_uint,
9490 .write_u64 = cpu_rt_period_write_uint,
9495 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9497 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9500 struct cgroup_subsys cpu_cgroup_subsys = {
9502 .create = cpu_cgroup_create,
9503 .destroy = cpu_cgroup_destroy,
9504 .can_attach = cpu_cgroup_can_attach,
9505 .attach = cpu_cgroup_attach,
9506 .populate = cpu_cgroup_populate,
9507 .subsys_id = cpu_cgroup_subsys_id,
9511 #endif /* CONFIG_CGROUP_SCHED */
9513 #ifdef CONFIG_CGROUP_CPUACCT
9516 * CPU accounting code for task groups.
9518 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9519 * (balbir@in.ibm.com).
9522 /* track cpu usage of a group of tasks and its child groups */
9524 struct cgroup_subsys_state css;
9525 /* cpuusage holds pointer to a u64-type object on every cpu */
9527 struct cpuacct *parent;
9530 struct cgroup_subsys cpuacct_subsys;
9532 /* return cpu accounting group corresponding to this container */
9533 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9535 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9536 struct cpuacct, css);
9539 /* return cpu accounting group to which this task belongs */
9540 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9542 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9543 struct cpuacct, css);
9546 /* create a new cpu accounting group */
9547 static struct cgroup_subsys_state *cpuacct_create(
9548 struct cgroup_subsys *ss, struct cgroup *cgrp)
9550 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9553 return ERR_PTR(-ENOMEM);
9555 ca->cpuusage = alloc_percpu(u64);
9556 if (!ca->cpuusage) {
9558 return ERR_PTR(-ENOMEM);
9562 ca->parent = cgroup_ca(cgrp->parent);
9567 /* destroy an existing cpu accounting group */
9569 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9571 struct cpuacct *ca = cgroup_ca(cgrp);
9573 free_percpu(ca->cpuusage);
9577 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9579 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9582 #ifndef CONFIG_64BIT
9584 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9586 spin_lock_irq(&cpu_rq(cpu)->lock);
9588 spin_unlock_irq(&cpu_rq(cpu)->lock);
9596 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9598 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9600 #ifndef CONFIG_64BIT
9602 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9604 spin_lock_irq(&cpu_rq(cpu)->lock);
9606 spin_unlock_irq(&cpu_rq(cpu)->lock);
9612 /* return total cpu usage (in nanoseconds) of a group */
9613 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9615 struct cpuacct *ca = cgroup_ca(cgrp);
9616 u64 totalcpuusage = 0;
9619 for_each_present_cpu(i)
9620 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9622 return totalcpuusage;
9625 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9628 struct cpuacct *ca = cgroup_ca(cgrp);
9637 for_each_present_cpu(i)
9638 cpuacct_cpuusage_write(ca, i, 0);
9644 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9647 struct cpuacct *ca = cgroup_ca(cgroup);
9651 for_each_present_cpu(i) {
9652 percpu = cpuacct_cpuusage_read(ca, i);
9653 seq_printf(m, "%llu ", (unsigned long long) percpu);
9655 seq_printf(m, "\n");
9659 static struct cftype files[] = {
9662 .read_u64 = cpuusage_read,
9663 .write_u64 = cpuusage_write,
9666 .name = "usage_percpu",
9667 .read_seq_string = cpuacct_percpu_seq_read,
9672 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9674 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9678 * charge this task's execution time to its accounting group.
9680 * called with rq->lock held.
9682 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9687 if (!cpuacct_subsys.active)
9690 cpu = task_cpu(tsk);
9693 for (; ca; ca = ca->parent) {
9694 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9695 *cpuusage += cputime;
9699 struct cgroup_subsys cpuacct_subsys = {
9701 .create = cpuacct_create,
9702 .destroy = cpuacct_destroy,
9703 .populate = cpuacct_populate,
9704 .subsys_id = cpuacct_subsys_id,
9706 #endif /* CONFIG_CGROUP_CPUACCT */