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 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
212 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_UNLOCKED;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime >= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
224 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
227 if (hrtimer_active(&rt_b->rt_period_timer))
230 spin_lock(&rt_b->rt_runtime_lock);
232 if (hrtimer_active(&rt_b->rt_period_timer))
235 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
236 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
237 hrtimer_start_expires(&rt_b->rt_period_timer,
240 spin_unlock(&rt_b->rt_runtime_lock);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
246 hrtimer_cancel(&rt_b->rt_period_timer);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css;
270 #ifdef CONFIG_USER_SCHED
274 #ifdef CONFIG_FAIR_GROUP_SCHED
275 /* schedulable entities of this group on each cpu */
276 struct sched_entity **se;
277 /* runqueue "owned" by this group on each cpu */
278 struct cfs_rq **cfs_rq;
279 unsigned long shares;
282 #ifdef CONFIG_RT_GROUP_SCHED
283 struct sched_rt_entity **rt_se;
284 struct rt_rq **rt_rq;
286 struct rt_bandwidth rt_bandwidth;
290 struct list_head list;
292 struct task_group *parent;
293 struct list_head siblings;
294 struct list_head children;
297 #ifdef CONFIG_USER_SCHED
299 /* Helper function to pass uid information to create_sched_user() */
300 void set_tg_uid(struct user_struct *user)
302 user->tg->uid = user->uid;
307 * Every UID task group (including init_task_group aka UID-0) will
308 * be a child to this group.
310 struct task_group root_task_group;
312 #ifdef CONFIG_FAIR_GROUP_SCHED
313 /* Default task group's sched entity on each cpu */
314 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
315 /* Default task group's cfs_rq on each cpu */
316 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
317 #endif /* CONFIG_FAIR_GROUP_SCHED */
319 #ifdef CONFIG_RT_GROUP_SCHED
320 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
321 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
322 #endif /* CONFIG_RT_GROUP_SCHED */
323 #else /* !CONFIG_USER_SCHED */
324 #define root_task_group init_task_group
325 #endif /* CONFIG_USER_SCHED */
327 /* task_group_lock serializes add/remove of task groups and also changes to
328 * a task group's cpu shares.
330 static DEFINE_SPINLOCK(task_group_lock);
332 #ifdef CONFIG_FAIR_GROUP_SCHED
333 #ifdef CONFIG_USER_SCHED
334 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
335 #else /* !CONFIG_USER_SCHED */
336 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
337 #endif /* CONFIG_USER_SCHED */
340 * A weight of 0 or 1 can cause arithmetics problems.
341 * A weight of a cfs_rq is the sum of weights of which entities
342 * are queued on this cfs_rq, so a weight of a entity should not be
343 * too large, so as the shares value of a task group.
344 * (The default weight is 1024 - so there's no practical
345 * limitation from this.)
348 #define MAX_SHARES (1UL << 18)
350 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
353 /* Default task group.
354 * Every task in system belong to this group at bootup.
356 struct task_group init_task_group;
358 /* return group to which a task belongs */
359 static inline struct task_group *task_group(struct task_struct *p)
361 struct task_group *tg;
363 #ifdef CONFIG_USER_SCHED
365 #elif defined(CONFIG_CGROUP_SCHED)
366 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
367 struct task_group, css);
369 tg = &init_task_group;
374 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
375 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
377 #ifdef CONFIG_FAIR_GROUP_SCHED
378 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
379 p->se.parent = task_group(p)->se[cpu];
382 #ifdef CONFIG_RT_GROUP_SCHED
383 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
384 p->rt.parent = task_group(p)->rt_se[cpu];
390 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
391 static inline struct task_group *task_group(struct task_struct *p)
396 #endif /* CONFIG_GROUP_SCHED */
398 /* CFS-related fields in a runqueue */
400 struct load_weight load;
401 unsigned long nr_running;
406 struct rb_root tasks_timeline;
407 struct rb_node *rb_leftmost;
409 struct list_head tasks;
410 struct list_head *balance_iterator;
413 * 'curr' points to currently running entity on this cfs_rq.
414 * It is set to NULL otherwise (i.e when none are currently running).
416 struct sched_entity *curr, *next, *last;
418 unsigned int nr_spread_over;
420 #ifdef CONFIG_FAIR_GROUP_SCHED
421 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
424 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
425 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
426 * (like users, containers etc.)
428 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
429 * list is used during load balance.
431 struct list_head leaf_cfs_rq_list;
432 struct task_group *tg; /* group that "owns" this runqueue */
436 * the part of load.weight contributed by tasks
438 unsigned long task_weight;
441 * h_load = weight * f(tg)
443 * Where f(tg) is the recursive weight fraction assigned to
446 unsigned long h_load;
449 * this cpu's part of tg->shares
451 unsigned long shares;
454 * load.weight at the time we set shares
456 unsigned long rq_weight;
461 /* Real-Time classes' related field in a runqueue: */
463 struct rt_prio_array active;
464 unsigned long rt_nr_running;
465 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
467 int curr; /* highest queued rt task prio */
468 int next; /* next highest */
472 unsigned long rt_nr_migratory;
478 /* Nests inside the rq lock: */
479 spinlock_t rt_runtime_lock;
481 #ifdef CONFIG_RT_GROUP_SCHED
482 unsigned long rt_nr_boosted;
485 struct list_head leaf_rt_rq_list;
486 struct task_group *tg;
487 struct sched_rt_entity *rt_se;
494 * We add the notion of a root-domain which will be used to define per-domain
495 * variables. Each exclusive cpuset essentially defines an island domain by
496 * fully partitioning the member cpus from any other cpuset. Whenever a new
497 * exclusive cpuset is created, we also create and attach a new root-domain
504 cpumask_var_t online;
507 * The "RT overload" flag: it gets set if a CPU has more than
508 * one runnable RT task.
510 cpumask_var_t rto_mask;
513 struct cpupri cpupri;
515 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
517 * Preferred wake up cpu nominated by sched_mc balance that will be
518 * used when most cpus are idle in the system indicating overall very
519 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
521 unsigned int sched_mc_preferred_wakeup_cpu;
526 * By default the system creates a single root-domain with all cpus as
527 * members (mimicking the global state we have today).
529 static struct root_domain def_root_domain;
534 * This is the main, per-CPU runqueue data structure.
536 * Locking rule: those places that want to lock multiple runqueues
537 * (such as the load balancing or the thread migration code), lock
538 * acquire operations must be ordered by ascending &runqueue.
545 * nr_running and cpu_load should be in the same cacheline because
546 * remote CPUs use both these fields when doing load calculation.
548 unsigned long nr_running;
549 #define CPU_LOAD_IDX_MAX 5
550 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
551 unsigned char idle_at_tick;
553 unsigned long last_tick_seen;
554 unsigned char in_nohz_recently;
556 /* capture load from *all* tasks on this cpu: */
557 struct load_weight load;
558 unsigned long nr_load_updates;
564 #ifdef CONFIG_FAIR_GROUP_SCHED
565 /* list of leaf cfs_rq on this cpu: */
566 struct list_head leaf_cfs_rq_list;
568 #ifdef CONFIG_RT_GROUP_SCHED
569 struct list_head leaf_rt_rq_list;
573 * This is part of a global counter where only the total sum
574 * over all CPUs matters. A task can increase this counter on
575 * one CPU and if it got migrated afterwards it may decrease
576 * it on another CPU. Always updated under the runqueue lock:
578 unsigned long nr_uninterruptible;
580 struct task_struct *curr, *idle;
581 unsigned long next_balance;
582 struct mm_struct *prev_mm;
589 struct root_domain *rd;
590 struct sched_domain *sd;
592 /* For active balancing */
595 /* cpu of this runqueue: */
599 unsigned long avg_load_per_task;
601 struct task_struct *migration_thread;
602 struct list_head migration_queue;
605 #ifdef CONFIG_SCHED_HRTICK
607 int hrtick_csd_pending;
608 struct call_single_data hrtick_csd;
610 struct hrtimer hrtick_timer;
613 #ifdef CONFIG_SCHEDSTATS
615 struct sched_info rq_sched_info;
617 /* sys_sched_yield() stats */
618 unsigned int yld_exp_empty;
619 unsigned int yld_act_empty;
620 unsigned int yld_both_empty;
621 unsigned int yld_count;
623 /* schedule() stats */
624 unsigned int sched_switch;
625 unsigned int sched_count;
626 unsigned int sched_goidle;
628 /* try_to_wake_up() stats */
629 unsigned int ttwu_count;
630 unsigned int ttwu_local;
633 unsigned int bkl_count;
637 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
639 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
641 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
644 static inline int cpu_of(struct rq *rq)
654 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
655 * See detach_destroy_domains: synchronize_sched for details.
657 * The domain tree of any CPU may only be accessed from within
658 * preempt-disabled sections.
660 #define for_each_domain(cpu, __sd) \
661 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
663 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
664 #define this_rq() (&__get_cpu_var(runqueues))
665 #define task_rq(p) cpu_rq(task_cpu(p))
666 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
668 static inline void update_rq_clock(struct rq *rq)
670 rq->clock = sched_clock_cpu(cpu_of(rq));
674 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
676 #ifdef CONFIG_SCHED_DEBUG
677 # define const_debug __read_mostly
679 # define const_debug static const
685 * Returns true if the current cpu runqueue is locked.
686 * This interface allows printk to be called with the runqueue lock
687 * held and know whether or not it is OK to wake up the klogd.
689 int runqueue_is_locked(void)
692 struct rq *rq = cpu_rq(cpu);
695 ret = spin_is_locked(&rq->lock);
701 * Debugging: various feature bits
704 #define SCHED_FEAT(name, enabled) \
705 __SCHED_FEAT_##name ,
708 #include "sched_features.h"
713 #define SCHED_FEAT(name, enabled) \
714 (1UL << __SCHED_FEAT_##name) * enabled |
716 const_debug unsigned int sysctl_sched_features =
717 #include "sched_features.h"
722 #ifdef CONFIG_SCHED_DEBUG
723 #define SCHED_FEAT(name, enabled) \
726 static __read_mostly char *sched_feat_names[] = {
727 #include "sched_features.h"
733 static int sched_feat_show(struct seq_file *m, void *v)
737 for (i = 0; sched_feat_names[i]; i++) {
738 if (!(sysctl_sched_features & (1UL << i)))
740 seq_printf(m, "%s ", sched_feat_names[i]);
748 sched_feat_write(struct file *filp, const char __user *ubuf,
749 size_t cnt, loff_t *ppos)
759 if (copy_from_user(&buf, ubuf, cnt))
764 if (strncmp(buf, "NO_", 3) == 0) {
769 for (i = 0; sched_feat_names[i]; i++) {
770 int len = strlen(sched_feat_names[i]);
772 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
774 sysctl_sched_features &= ~(1UL << i);
776 sysctl_sched_features |= (1UL << i);
781 if (!sched_feat_names[i])
789 static int sched_feat_open(struct inode *inode, struct file *filp)
791 return single_open(filp, sched_feat_show, NULL);
794 static struct file_operations sched_feat_fops = {
795 .open = sched_feat_open,
796 .write = sched_feat_write,
799 .release = single_release,
802 static __init int sched_init_debug(void)
804 debugfs_create_file("sched_features", 0644, NULL, NULL,
809 late_initcall(sched_init_debug);
813 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
816 * Number of tasks to iterate in a single balance run.
817 * Limited because this is done with IRQs disabled.
819 const_debug unsigned int sysctl_sched_nr_migrate = 32;
822 * ratelimit for updating the group shares.
825 unsigned int sysctl_sched_shares_ratelimit = 250000;
828 * Inject some fuzzyness into changing the per-cpu group shares
829 * this avoids remote rq-locks at the expense of fairness.
832 unsigned int sysctl_sched_shares_thresh = 4;
835 * period over which we measure -rt task cpu usage in us.
838 unsigned int sysctl_sched_rt_period = 1000000;
840 static __read_mostly int scheduler_running;
843 * part of the period that we allow rt tasks to run in us.
846 int sysctl_sched_rt_runtime = 950000;
848 static inline u64 global_rt_period(void)
850 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
853 static inline u64 global_rt_runtime(void)
855 if (sysctl_sched_rt_runtime < 0)
858 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
861 #ifndef prepare_arch_switch
862 # define prepare_arch_switch(next) do { } while (0)
864 #ifndef finish_arch_switch
865 # define finish_arch_switch(prev) do { } while (0)
868 static inline int task_current(struct rq *rq, struct task_struct *p)
870 return rq->curr == p;
873 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
874 static inline int task_running(struct rq *rq, struct task_struct *p)
876 return task_current(rq, p);
879 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
883 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
885 #ifdef CONFIG_DEBUG_SPINLOCK
886 /* this is a valid case when another task releases the spinlock */
887 rq->lock.owner = current;
890 * If we are tracking spinlock dependencies then we have to
891 * fix up the runqueue lock - which gets 'carried over' from
894 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
896 spin_unlock_irq(&rq->lock);
899 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
900 static inline int task_running(struct rq *rq, struct task_struct *p)
905 return task_current(rq, p);
909 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
913 * We can optimise this out completely for !SMP, because the
914 * SMP rebalancing from interrupt is the only thing that cares
919 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
920 spin_unlock_irq(&rq->lock);
922 spin_unlock(&rq->lock);
926 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
930 * After ->oncpu is cleared, the task can be moved to a different CPU.
931 * We must ensure this doesn't happen until the switch is completely
937 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
941 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
944 * __task_rq_lock - lock the runqueue a given task resides on.
945 * Must be called interrupts disabled.
947 static inline struct rq *__task_rq_lock(struct task_struct *p)
951 struct rq *rq = task_rq(p);
952 spin_lock(&rq->lock);
953 if (likely(rq == task_rq(p)))
955 spin_unlock(&rq->lock);
960 * task_rq_lock - lock the runqueue a given task resides on and disable
961 * interrupts. Note the ordering: we can safely lookup the task_rq without
962 * explicitly disabling preemption.
964 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
970 local_irq_save(*flags);
972 spin_lock(&rq->lock);
973 if (likely(rq == task_rq(p)))
975 spin_unlock_irqrestore(&rq->lock, *flags);
979 void task_rq_unlock_wait(struct task_struct *p)
981 struct rq *rq = task_rq(p);
983 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
984 spin_unlock_wait(&rq->lock);
987 static void __task_rq_unlock(struct rq *rq)
990 spin_unlock(&rq->lock);
993 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
996 spin_unlock_irqrestore(&rq->lock, *flags);
1000 * this_rq_lock - lock this runqueue and disable interrupts.
1002 static struct rq *this_rq_lock(void)
1003 __acquires(rq->lock)
1007 local_irq_disable();
1009 spin_lock(&rq->lock);
1014 #ifdef CONFIG_SCHED_HRTICK
1016 * Use HR-timers to deliver accurate preemption points.
1018 * Its all a bit involved since we cannot program an hrt while holding the
1019 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1022 * When we get rescheduled we reprogram the hrtick_timer outside of the
1028 * - enabled by features
1029 * - hrtimer is actually high res
1031 static inline int hrtick_enabled(struct rq *rq)
1033 if (!sched_feat(HRTICK))
1035 if (!cpu_active(cpu_of(rq)))
1037 return hrtimer_is_hres_active(&rq->hrtick_timer);
1040 static void hrtick_clear(struct rq *rq)
1042 if (hrtimer_active(&rq->hrtick_timer))
1043 hrtimer_cancel(&rq->hrtick_timer);
1047 * High-resolution timer tick.
1048 * Runs from hardirq context with interrupts disabled.
1050 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1052 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1054 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1056 spin_lock(&rq->lock);
1057 update_rq_clock(rq);
1058 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1059 spin_unlock(&rq->lock);
1061 return HRTIMER_NORESTART;
1066 * called from hardirq (IPI) context
1068 static void __hrtick_start(void *arg)
1070 struct rq *rq = arg;
1072 spin_lock(&rq->lock);
1073 hrtimer_restart(&rq->hrtick_timer);
1074 rq->hrtick_csd_pending = 0;
1075 spin_unlock(&rq->lock);
1079 * Called to set the hrtick timer state.
1081 * called with rq->lock held and irqs disabled
1083 static void hrtick_start(struct rq *rq, u64 delay)
1085 struct hrtimer *timer = &rq->hrtick_timer;
1086 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1088 hrtimer_set_expires(timer, time);
1090 if (rq == this_rq()) {
1091 hrtimer_restart(timer);
1092 } else if (!rq->hrtick_csd_pending) {
1093 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1094 rq->hrtick_csd_pending = 1;
1099 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1101 int cpu = (int)(long)hcpu;
1104 case CPU_UP_CANCELED:
1105 case CPU_UP_CANCELED_FROZEN:
1106 case CPU_DOWN_PREPARE:
1107 case CPU_DOWN_PREPARE_FROZEN:
1109 case CPU_DEAD_FROZEN:
1110 hrtick_clear(cpu_rq(cpu));
1117 static __init void init_hrtick(void)
1119 hotcpu_notifier(hotplug_hrtick, 0);
1123 * Called to set the hrtick timer state.
1125 * called with rq->lock held and irqs disabled
1127 static void hrtick_start(struct rq *rq, u64 delay)
1129 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1132 static inline void init_hrtick(void)
1135 #endif /* CONFIG_SMP */
1137 static void init_rq_hrtick(struct rq *rq)
1140 rq->hrtick_csd_pending = 0;
1142 rq->hrtick_csd.flags = 0;
1143 rq->hrtick_csd.func = __hrtick_start;
1144 rq->hrtick_csd.info = rq;
1147 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1148 rq->hrtick_timer.function = hrtick;
1149 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_PERCPU;
1151 #else /* CONFIG_SCHED_HRTICK */
1152 static inline void hrtick_clear(struct rq *rq)
1156 static inline void init_rq_hrtick(struct rq *rq)
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SCHED_HRTICK */
1166 * resched_task - mark a task 'to be rescheduled now'.
1168 * On UP this means the setting of the need_resched flag, on SMP it
1169 * might also involve a cross-CPU call to trigger the scheduler on
1174 #ifndef tsk_is_polling
1175 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1178 static void resched_task(struct task_struct *p)
1182 assert_spin_locked(&task_rq(p)->lock);
1184 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1187 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1190 if (cpu == smp_processor_id())
1193 /* NEED_RESCHED must be visible before we test polling */
1195 if (!tsk_is_polling(p))
1196 smp_send_reschedule(cpu);
1199 static void resched_cpu(int cpu)
1201 struct rq *rq = cpu_rq(cpu);
1202 unsigned long flags;
1204 if (!spin_trylock_irqsave(&rq->lock, flags))
1206 resched_task(cpu_curr(cpu));
1207 spin_unlock_irqrestore(&rq->lock, flags);
1212 * When add_timer_on() enqueues a timer into the timer wheel of an
1213 * idle CPU then this timer might expire before the next timer event
1214 * which is scheduled to wake up that CPU. In case of a completely
1215 * idle system the next event might even be infinite time into the
1216 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1217 * leaves the inner idle loop so the newly added timer is taken into
1218 * account when the CPU goes back to idle and evaluates the timer
1219 * wheel for the next timer event.
1221 void wake_up_idle_cpu(int cpu)
1223 struct rq *rq = cpu_rq(cpu);
1225 if (cpu == smp_processor_id())
1229 * This is safe, as this function is called with the timer
1230 * wheel base lock of (cpu) held. When the CPU is on the way
1231 * to idle and has not yet set rq->curr to idle then it will
1232 * be serialized on the timer wheel base lock and take the new
1233 * timer into account automatically.
1235 if (rq->curr != rq->idle)
1239 * We can set TIF_RESCHED on the idle task of the other CPU
1240 * lockless. The worst case is that the other CPU runs the
1241 * idle task through an additional NOOP schedule()
1243 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1245 /* NEED_RESCHED must be visible before we test polling */
1247 if (!tsk_is_polling(rq->idle))
1248 smp_send_reschedule(cpu);
1250 #endif /* CONFIG_NO_HZ */
1252 #else /* !CONFIG_SMP */
1253 static void resched_task(struct task_struct *p)
1255 assert_spin_locked(&task_rq(p)->lock);
1256 set_tsk_need_resched(p);
1258 #endif /* CONFIG_SMP */
1260 #if BITS_PER_LONG == 32
1261 # define WMULT_CONST (~0UL)
1263 # define WMULT_CONST (1UL << 32)
1266 #define WMULT_SHIFT 32
1269 * Shift right and round:
1271 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1274 * delta *= weight / lw
1276 static unsigned long
1277 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1278 struct load_weight *lw)
1282 if (!lw->inv_weight) {
1283 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1286 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1290 tmp = (u64)delta_exec * weight;
1292 * Check whether we'd overflow the 64-bit multiplication:
1294 if (unlikely(tmp > WMULT_CONST))
1295 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1298 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1300 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1303 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1309 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1316 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1317 * of tasks with abnormal "nice" values across CPUs the contribution that
1318 * each task makes to its run queue's load is weighted according to its
1319 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1320 * scaled version of the new time slice allocation that they receive on time
1324 #define WEIGHT_IDLEPRIO 2
1325 #define WMULT_IDLEPRIO (1 << 31)
1328 * Nice levels are multiplicative, with a gentle 10% change for every
1329 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1330 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1331 * that remained on nice 0.
1333 * The "10% effect" is relative and cumulative: from _any_ nice level,
1334 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1335 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1336 * If a task goes up by ~10% and another task goes down by ~10% then
1337 * the relative distance between them is ~25%.)
1339 static const int prio_to_weight[40] = {
1340 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1341 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1342 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1343 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1344 /* 0 */ 1024, 820, 655, 526, 423,
1345 /* 5 */ 335, 272, 215, 172, 137,
1346 /* 10 */ 110, 87, 70, 56, 45,
1347 /* 15 */ 36, 29, 23, 18, 15,
1351 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1353 * In cases where the weight does not change often, we can use the
1354 * precalculated inverse to speed up arithmetics by turning divisions
1355 * into multiplications:
1357 static const u32 prio_to_wmult[40] = {
1358 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1359 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1360 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1361 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1362 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1363 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1364 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1365 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1368 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1371 * runqueue iterator, to support SMP load-balancing between different
1372 * scheduling classes, without having to expose their internal data
1373 * structures to the load-balancing proper:
1375 struct rq_iterator {
1377 struct task_struct *(*start)(void *);
1378 struct task_struct *(*next)(void *);
1382 static unsigned long
1383 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1384 unsigned long max_load_move, struct sched_domain *sd,
1385 enum cpu_idle_type idle, int *all_pinned,
1386 int *this_best_prio, struct rq_iterator *iterator);
1389 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1390 struct sched_domain *sd, enum cpu_idle_type idle,
1391 struct rq_iterator *iterator);
1394 #ifdef CONFIG_CGROUP_CPUACCT
1395 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1397 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1400 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1402 update_load_add(&rq->load, load);
1405 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1407 update_load_sub(&rq->load, load);
1410 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1411 typedef int (*tg_visitor)(struct task_group *, void *);
1414 * Iterate the full tree, calling @down when first entering a node and @up when
1415 * leaving it for the final time.
1417 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1419 struct task_group *parent, *child;
1423 parent = &root_task_group;
1425 ret = (*down)(parent, data);
1428 list_for_each_entry_rcu(child, &parent->children, siblings) {
1435 ret = (*up)(parent, data);
1440 parent = parent->parent;
1449 static int tg_nop(struct task_group *tg, void *data)
1456 static unsigned long source_load(int cpu, int type);
1457 static unsigned long target_load(int cpu, int type);
1458 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1460 static unsigned long cpu_avg_load_per_task(int cpu)
1462 struct rq *rq = cpu_rq(cpu);
1463 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1466 rq->avg_load_per_task = rq->load.weight / nr_running;
1468 rq->avg_load_per_task = 0;
1470 return rq->avg_load_per_task;
1473 #ifdef CONFIG_FAIR_GROUP_SCHED
1475 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1478 * Calculate and set the cpu's group shares.
1481 update_group_shares_cpu(struct task_group *tg, int cpu,
1482 unsigned long sd_shares, unsigned long sd_rq_weight)
1484 unsigned long shares;
1485 unsigned long rq_weight;
1490 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1493 * \Sum shares * rq_weight
1494 * shares = -----------------------
1498 shares = (sd_shares * rq_weight) / sd_rq_weight;
1499 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1501 if (abs(shares - tg->se[cpu]->load.weight) >
1502 sysctl_sched_shares_thresh) {
1503 struct rq *rq = cpu_rq(cpu);
1504 unsigned long flags;
1506 spin_lock_irqsave(&rq->lock, flags);
1507 tg->cfs_rq[cpu]->shares = shares;
1509 __set_se_shares(tg->se[cpu], shares);
1510 spin_unlock_irqrestore(&rq->lock, flags);
1515 * Re-compute the task group their per cpu shares over the given domain.
1516 * This needs to be done in a bottom-up fashion because the rq weight of a
1517 * parent group depends on the shares of its child groups.
1519 static int tg_shares_up(struct task_group *tg, void *data)
1521 unsigned long weight, rq_weight = 0;
1522 unsigned long shares = 0;
1523 struct sched_domain *sd = data;
1526 for_each_cpu(i, sched_domain_span(sd)) {
1528 * If there are currently no tasks on the cpu pretend there
1529 * is one of average load so that when a new task gets to
1530 * run here it will not get delayed by group starvation.
1532 weight = tg->cfs_rq[i]->load.weight;
1534 weight = NICE_0_LOAD;
1536 tg->cfs_rq[i]->rq_weight = weight;
1537 rq_weight += weight;
1538 shares += tg->cfs_rq[i]->shares;
1541 if ((!shares && rq_weight) || shares > tg->shares)
1542 shares = tg->shares;
1544 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1545 shares = tg->shares;
1547 for_each_cpu(i, sched_domain_span(sd))
1548 update_group_shares_cpu(tg, i, shares, rq_weight);
1554 * Compute the cpu's hierarchical load factor for each task group.
1555 * This needs to be done in a top-down fashion because the load of a child
1556 * group is a fraction of its parents load.
1558 static int tg_load_down(struct task_group *tg, void *data)
1561 long cpu = (long)data;
1564 load = cpu_rq(cpu)->load.weight;
1566 load = tg->parent->cfs_rq[cpu]->h_load;
1567 load *= tg->cfs_rq[cpu]->shares;
1568 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1571 tg->cfs_rq[cpu]->h_load = load;
1576 static void update_shares(struct sched_domain *sd)
1578 u64 now = cpu_clock(raw_smp_processor_id());
1579 s64 elapsed = now - sd->last_update;
1581 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1582 sd->last_update = now;
1583 walk_tg_tree(tg_nop, tg_shares_up, sd);
1587 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1589 spin_unlock(&rq->lock);
1591 spin_lock(&rq->lock);
1594 static void update_h_load(long cpu)
1596 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1601 static inline void update_shares(struct sched_domain *sd)
1605 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1612 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1614 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1615 __releases(this_rq->lock)
1616 __acquires(busiest->lock)
1617 __acquires(this_rq->lock)
1621 if (unlikely(!irqs_disabled())) {
1622 /* printk() doesn't work good under rq->lock */
1623 spin_unlock(&this_rq->lock);
1626 if (unlikely(!spin_trylock(&busiest->lock))) {
1627 if (busiest < this_rq) {
1628 spin_unlock(&this_rq->lock);
1629 spin_lock(&busiest->lock);
1630 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1633 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1638 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1639 __releases(busiest->lock)
1641 spin_unlock(&busiest->lock);
1642 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1646 #ifdef CONFIG_FAIR_GROUP_SCHED
1647 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1650 cfs_rq->shares = shares;
1655 #include "sched_stats.h"
1656 #include "sched_idletask.c"
1657 #include "sched_fair.c"
1658 #include "sched_rt.c"
1659 #ifdef CONFIG_SCHED_DEBUG
1660 # include "sched_debug.c"
1663 #define sched_class_highest (&rt_sched_class)
1664 #define for_each_class(class) \
1665 for (class = sched_class_highest; class; class = class->next)
1667 static void inc_nr_running(struct rq *rq)
1672 static void dec_nr_running(struct rq *rq)
1677 static void set_load_weight(struct task_struct *p)
1679 if (task_has_rt_policy(p)) {
1680 p->se.load.weight = prio_to_weight[0] * 2;
1681 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1686 * SCHED_IDLE tasks get minimal weight:
1688 if (p->policy == SCHED_IDLE) {
1689 p->se.load.weight = WEIGHT_IDLEPRIO;
1690 p->se.load.inv_weight = WMULT_IDLEPRIO;
1694 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1695 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1698 static void update_avg(u64 *avg, u64 sample)
1700 s64 diff = sample - *avg;
1704 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1706 sched_info_queued(p);
1707 p->sched_class->enqueue_task(rq, p, wakeup);
1711 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1713 if (sleep && p->se.last_wakeup) {
1714 update_avg(&p->se.avg_overlap,
1715 p->se.sum_exec_runtime - p->se.last_wakeup);
1716 p->se.last_wakeup = 0;
1719 sched_info_dequeued(p);
1720 p->sched_class->dequeue_task(rq, p, sleep);
1725 * __normal_prio - return the priority that is based on the static prio
1727 static inline int __normal_prio(struct task_struct *p)
1729 return p->static_prio;
1733 * Calculate the expected normal priority: i.e. priority
1734 * without taking RT-inheritance into account. Might be
1735 * boosted by interactivity modifiers. Changes upon fork,
1736 * setprio syscalls, and whenever the interactivity
1737 * estimator recalculates.
1739 static inline int normal_prio(struct task_struct *p)
1743 if (task_has_rt_policy(p))
1744 prio = MAX_RT_PRIO-1 - p->rt_priority;
1746 prio = __normal_prio(p);
1751 * Calculate the current priority, i.e. the priority
1752 * taken into account by the scheduler. This value might
1753 * be boosted by RT tasks, or might be boosted by
1754 * interactivity modifiers. Will be RT if the task got
1755 * RT-boosted. If not then it returns p->normal_prio.
1757 static int effective_prio(struct task_struct *p)
1759 p->normal_prio = normal_prio(p);
1761 * If we are RT tasks or we were boosted to RT priority,
1762 * keep the priority unchanged. Otherwise, update priority
1763 * to the normal priority:
1765 if (!rt_prio(p->prio))
1766 return p->normal_prio;
1771 * activate_task - move a task to the runqueue.
1773 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1775 if (task_contributes_to_load(p))
1776 rq->nr_uninterruptible--;
1778 enqueue_task(rq, p, wakeup);
1783 * deactivate_task - remove a task from the runqueue.
1785 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1787 if (task_contributes_to_load(p))
1788 rq->nr_uninterruptible++;
1790 dequeue_task(rq, p, sleep);
1795 * task_curr - is this task currently executing on a CPU?
1796 * @p: the task in question.
1798 inline int task_curr(const struct task_struct *p)
1800 return cpu_curr(task_cpu(p)) == p;
1803 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1805 set_task_rq(p, cpu);
1808 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1809 * successfuly executed on another CPU. We must ensure that updates of
1810 * per-task data have been completed by this moment.
1813 task_thread_info(p)->cpu = cpu;
1817 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1818 const struct sched_class *prev_class,
1819 int oldprio, int running)
1821 if (prev_class != p->sched_class) {
1822 if (prev_class->switched_from)
1823 prev_class->switched_from(rq, p, running);
1824 p->sched_class->switched_to(rq, p, running);
1826 p->sched_class->prio_changed(rq, p, oldprio, running);
1831 /* Used instead of source_load when we know the type == 0 */
1832 static unsigned long weighted_cpuload(const int cpu)
1834 return cpu_rq(cpu)->load.weight;
1838 * Is this task likely cache-hot:
1841 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1846 * Buddy candidates are cache hot:
1848 if (sched_feat(CACHE_HOT_BUDDY) &&
1849 (&p->se == cfs_rq_of(&p->se)->next ||
1850 &p->se == cfs_rq_of(&p->se)->last))
1853 if (p->sched_class != &fair_sched_class)
1856 if (sysctl_sched_migration_cost == -1)
1858 if (sysctl_sched_migration_cost == 0)
1861 delta = now - p->se.exec_start;
1863 return delta < (s64)sysctl_sched_migration_cost;
1867 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1869 int old_cpu = task_cpu(p);
1870 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1871 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1872 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1875 clock_offset = old_rq->clock - new_rq->clock;
1877 #ifdef CONFIG_SCHEDSTATS
1878 if (p->se.wait_start)
1879 p->se.wait_start -= clock_offset;
1880 if (p->se.sleep_start)
1881 p->se.sleep_start -= clock_offset;
1882 if (p->se.block_start)
1883 p->se.block_start -= clock_offset;
1884 if (old_cpu != new_cpu) {
1885 schedstat_inc(p, se.nr_migrations);
1886 if (task_hot(p, old_rq->clock, NULL))
1887 schedstat_inc(p, se.nr_forced2_migrations);
1890 p->se.vruntime -= old_cfsrq->min_vruntime -
1891 new_cfsrq->min_vruntime;
1893 __set_task_cpu(p, new_cpu);
1896 struct migration_req {
1897 struct list_head list;
1899 struct task_struct *task;
1902 struct completion done;
1906 * The task's runqueue lock must be held.
1907 * Returns true if you have to wait for migration thread.
1910 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1912 struct rq *rq = task_rq(p);
1915 * If the task is not on a runqueue (and not running), then
1916 * it is sufficient to simply update the task's cpu field.
1918 if (!p->se.on_rq && !task_running(rq, p)) {
1919 set_task_cpu(p, dest_cpu);
1923 init_completion(&req->done);
1925 req->dest_cpu = dest_cpu;
1926 list_add(&req->list, &rq->migration_queue);
1932 * wait_task_inactive - wait for a thread to unschedule.
1934 * If @match_state is nonzero, it's the @p->state value just checked and
1935 * not expected to change. If it changes, i.e. @p might have woken up,
1936 * then return zero. When we succeed in waiting for @p to be off its CPU,
1937 * we return a positive number (its total switch count). If a second call
1938 * a short while later returns the same number, the caller can be sure that
1939 * @p has remained unscheduled the whole time.
1941 * The caller must ensure that the task *will* unschedule sometime soon,
1942 * else this function might spin for a *long* time. This function can't
1943 * be called with interrupts off, or it may introduce deadlock with
1944 * smp_call_function() if an IPI is sent by the same process we are
1945 * waiting to become inactive.
1947 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1949 unsigned long flags;
1956 * We do the initial early heuristics without holding
1957 * any task-queue locks at all. We'll only try to get
1958 * the runqueue lock when things look like they will
1964 * If the task is actively running on another CPU
1965 * still, just relax and busy-wait without holding
1968 * NOTE! Since we don't hold any locks, it's not
1969 * even sure that "rq" stays as the right runqueue!
1970 * But we don't care, since "task_running()" will
1971 * return false if the runqueue has changed and p
1972 * is actually now running somewhere else!
1974 while (task_running(rq, p)) {
1975 if (match_state && unlikely(p->state != match_state))
1981 * Ok, time to look more closely! We need the rq
1982 * lock now, to be *sure*. If we're wrong, we'll
1983 * just go back and repeat.
1985 rq = task_rq_lock(p, &flags);
1986 trace_sched_wait_task(rq, p);
1987 running = task_running(rq, p);
1988 on_rq = p->se.on_rq;
1990 if (!match_state || p->state == match_state)
1991 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1992 task_rq_unlock(rq, &flags);
1995 * If it changed from the expected state, bail out now.
1997 if (unlikely(!ncsw))
2001 * Was it really running after all now that we
2002 * checked with the proper locks actually held?
2004 * Oops. Go back and try again..
2006 if (unlikely(running)) {
2012 * It's not enough that it's not actively running,
2013 * it must be off the runqueue _entirely_, and not
2016 * So if it wa still runnable (but just not actively
2017 * running right now), it's preempted, and we should
2018 * yield - it could be a while.
2020 if (unlikely(on_rq)) {
2021 schedule_timeout_uninterruptible(1);
2026 * Ahh, all good. It wasn't running, and it wasn't
2027 * runnable, which means that it will never become
2028 * running in the future either. We're all done!
2037 * kick_process - kick a running thread to enter/exit the kernel
2038 * @p: the to-be-kicked thread
2040 * Cause a process which is running on another CPU to enter
2041 * kernel-mode, without any delay. (to get signals handled.)
2043 * NOTE: this function doesnt have to take the runqueue lock,
2044 * because all it wants to ensure is that the remote task enters
2045 * the kernel. If the IPI races and the task has been migrated
2046 * to another CPU then no harm is done and the purpose has been
2049 void kick_process(struct task_struct *p)
2055 if ((cpu != smp_processor_id()) && task_curr(p))
2056 smp_send_reschedule(cpu);
2061 * Return a low guess at the load of a migration-source cpu weighted
2062 * according to the scheduling class and "nice" value.
2064 * We want to under-estimate the load of migration sources, to
2065 * balance conservatively.
2067 static unsigned long source_load(int cpu, int type)
2069 struct rq *rq = cpu_rq(cpu);
2070 unsigned long total = weighted_cpuload(cpu);
2072 if (type == 0 || !sched_feat(LB_BIAS))
2075 return min(rq->cpu_load[type-1], total);
2079 * Return a high guess at the load of a migration-target cpu weighted
2080 * according to the scheduling class and "nice" value.
2082 static unsigned long target_load(int cpu, int type)
2084 struct rq *rq = cpu_rq(cpu);
2085 unsigned long total = weighted_cpuload(cpu);
2087 if (type == 0 || !sched_feat(LB_BIAS))
2090 return max(rq->cpu_load[type-1], total);
2094 * find_idlest_group finds and returns the least busy CPU group within the
2097 static struct sched_group *
2098 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2100 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2101 unsigned long min_load = ULONG_MAX, this_load = 0;
2102 int load_idx = sd->forkexec_idx;
2103 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2106 unsigned long load, avg_load;
2110 /* Skip over this group if it has no CPUs allowed */
2111 if (!cpumask_intersects(sched_group_cpus(group),
2115 local_group = cpumask_test_cpu(this_cpu,
2116 sched_group_cpus(group));
2118 /* Tally up the load of all CPUs in the group */
2121 for_each_cpu(i, sched_group_cpus(group)) {
2122 /* Bias balancing toward cpus of our domain */
2124 load = source_load(i, load_idx);
2126 load = target_load(i, load_idx);
2131 /* Adjust by relative CPU power of the group */
2132 avg_load = sg_div_cpu_power(group,
2133 avg_load * SCHED_LOAD_SCALE);
2136 this_load = avg_load;
2138 } else if (avg_load < min_load) {
2139 min_load = avg_load;
2142 } while (group = group->next, group != sd->groups);
2144 if (!idlest || 100*this_load < imbalance*min_load)
2150 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2153 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2155 unsigned long load, min_load = ULONG_MAX;
2159 /* Traverse only the allowed CPUs */
2160 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2161 load = weighted_cpuload(i);
2163 if (load < min_load || (load == min_load && i == this_cpu)) {
2173 * sched_balance_self: balance the current task (running on cpu) in domains
2174 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2177 * Balance, ie. select the least loaded group.
2179 * Returns the target CPU number, or the same CPU if no balancing is needed.
2181 * preempt must be disabled.
2183 static int sched_balance_self(int cpu, int flag)
2185 struct task_struct *t = current;
2186 struct sched_domain *tmp, *sd = NULL;
2188 for_each_domain(cpu, tmp) {
2190 * If power savings logic is enabled for a domain, stop there.
2192 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2194 if (tmp->flags & flag)
2202 struct sched_group *group;
2203 int new_cpu, weight;
2205 if (!(sd->flags & flag)) {
2210 group = find_idlest_group(sd, t, cpu);
2216 new_cpu = find_idlest_cpu(group, t, cpu);
2217 if (new_cpu == -1 || new_cpu == cpu) {
2218 /* Now try balancing at a lower domain level of cpu */
2223 /* Now try balancing at a lower domain level of new_cpu */
2225 weight = cpumask_weight(sched_domain_span(sd));
2227 for_each_domain(cpu, tmp) {
2228 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2230 if (tmp->flags & flag)
2233 /* while loop will break here if sd == NULL */
2239 #endif /* CONFIG_SMP */
2242 * try_to_wake_up - wake up a thread
2243 * @p: the to-be-woken-up thread
2244 * @state: the mask of task states that can be woken
2245 * @sync: do a synchronous wakeup?
2247 * Put it on the run-queue if it's not already there. The "current"
2248 * thread is always on the run-queue (except when the actual
2249 * re-schedule is in progress), and as such you're allowed to do
2250 * the simpler "current->state = TASK_RUNNING" to mark yourself
2251 * runnable without the overhead of this.
2253 * returns failure only if the task is already active.
2255 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2257 int cpu, orig_cpu, this_cpu, success = 0;
2258 unsigned long flags;
2262 if (!sched_feat(SYNC_WAKEUPS))
2266 if (sched_feat(LB_WAKEUP_UPDATE)) {
2267 struct sched_domain *sd;
2269 this_cpu = raw_smp_processor_id();
2272 for_each_domain(this_cpu, sd) {
2273 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2282 rq = task_rq_lock(p, &flags);
2283 old_state = p->state;
2284 if (!(old_state & state))
2292 this_cpu = smp_processor_id();
2295 if (unlikely(task_running(rq, p)))
2298 cpu = p->sched_class->select_task_rq(p, sync);
2299 if (cpu != orig_cpu) {
2300 set_task_cpu(p, cpu);
2301 task_rq_unlock(rq, &flags);
2302 /* might preempt at this point */
2303 rq = task_rq_lock(p, &flags);
2304 old_state = p->state;
2305 if (!(old_state & state))
2310 this_cpu = smp_processor_id();
2314 #ifdef CONFIG_SCHEDSTATS
2315 schedstat_inc(rq, ttwu_count);
2316 if (cpu == this_cpu)
2317 schedstat_inc(rq, ttwu_local);
2319 struct sched_domain *sd;
2320 for_each_domain(this_cpu, sd) {
2321 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2322 schedstat_inc(sd, ttwu_wake_remote);
2327 #endif /* CONFIG_SCHEDSTATS */
2330 #endif /* CONFIG_SMP */
2331 schedstat_inc(p, se.nr_wakeups);
2333 schedstat_inc(p, se.nr_wakeups_sync);
2334 if (orig_cpu != cpu)
2335 schedstat_inc(p, se.nr_wakeups_migrate);
2336 if (cpu == this_cpu)
2337 schedstat_inc(p, se.nr_wakeups_local);
2339 schedstat_inc(p, se.nr_wakeups_remote);
2340 update_rq_clock(rq);
2341 activate_task(rq, p, 1);
2345 trace_sched_wakeup(rq, p);
2346 check_preempt_curr(rq, p, sync);
2348 p->state = TASK_RUNNING;
2350 if (p->sched_class->task_wake_up)
2351 p->sched_class->task_wake_up(rq, p);
2354 current->se.last_wakeup = current->se.sum_exec_runtime;
2356 task_rq_unlock(rq, &flags);
2361 int wake_up_process(struct task_struct *p)
2363 return try_to_wake_up(p, TASK_ALL, 0);
2365 EXPORT_SYMBOL(wake_up_process);
2367 int wake_up_state(struct task_struct *p, unsigned int state)
2369 return try_to_wake_up(p, state, 0);
2373 * Perform scheduler related setup for a newly forked process p.
2374 * p is forked by current.
2376 * __sched_fork() is basic setup used by init_idle() too:
2378 static void __sched_fork(struct task_struct *p)
2380 p->se.exec_start = 0;
2381 p->se.sum_exec_runtime = 0;
2382 p->se.prev_sum_exec_runtime = 0;
2383 p->se.last_wakeup = 0;
2384 p->se.avg_overlap = 0;
2386 #ifdef CONFIG_SCHEDSTATS
2387 p->se.wait_start = 0;
2388 p->se.sum_sleep_runtime = 0;
2389 p->se.sleep_start = 0;
2390 p->se.block_start = 0;
2391 p->se.sleep_max = 0;
2392 p->se.block_max = 0;
2394 p->se.slice_max = 0;
2398 INIT_LIST_HEAD(&p->rt.run_list);
2400 INIT_LIST_HEAD(&p->se.group_node);
2402 #ifdef CONFIG_PREEMPT_NOTIFIERS
2403 INIT_HLIST_HEAD(&p->preempt_notifiers);
2407 * We mark the process as running here, but have not actually
2408 * inserted it onto the runqueue yet. This guarantees that
2409 * nobody will actually run it, and a signal or other external
2410 * event cannot wake it up and insert it on the runqueue either.
2412 p->state = TASK_RUNNING;
2416 * fork()/clone()-time setup:
2418 void sched_fork(struct task_struct *p, int clone_flags)
2420 int cpu = get_cpu();
2425 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2427 set_task_cpu(p, cpu);
2430 * Make sure we do not leak PI boosting priority to the child:
2432 p->prio = current->normal_prio;
2433 if (!rt_prio(p->prio))
2434 p->sched_class = &fair_sched_class;
2436 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2437 if (likely(sched_info_on()))
2438 memset(&p->sched_info, 0, sizeof(p->sched_info));
2440 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2443 #ifdef CONFIG_PREEMPT
2444 /* Want to start with kernel preemption disabled. */
2445 task_thread_info(p)->preempt_count = 1;
2451 * wake_up_new_task - wake up a newly created task for the first time.
2453 * This function will do some initial scheduler statistics housekeeping
2454 * that must be done for every newly created context, then puts the task
2455 * on the runqueue and wakes it.
2457 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2459 unsigned long flags;
2462 rq = task_rq_lock(p, &flags);
2463 BUG_ON(p->state != TASK_RUNNING);
2464 update_rq_clock(rq);
2466 p->prio = effective_prio(p);
2468 if (!p->sched_class->task_new || !current->se.on_rq) {
2469 activate_task(rq, p, 0);
2472 * Let the scheduling class do new task startup
2473 * management (if any):
2475 p->sched_class->task_new(rq, p);
2478 trace_sched_wakeup_new(rq, p);
2479 check_preempt_curr(rq, p, 0);
2481 if (p->sched_class->task_wake_up)
2482 p->sched_class->task_wake_up(rq, p);
2484 task_rq_unlock(rq, &flags);
2487 #ifdef CONFIG_PREEMPT_NOTIFIERS
2490 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2491 * @notifier: notifier struct to register
2493 void preempt_notifier_register(struct preempt_notifier *notifier)
2495 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2497 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2500 * preempt_notifier_unregister - no longer interested in preemption notifications
2501 * @notifier: notifier struct to unregister
2503 * This is safe to call from within a preemption notifier.
2505 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2507 hlist_del(¬ifier->link);
2509 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2511 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2513 struct preempt_notifier *notifier;
2514 struct hlist_node *node;
2516 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2517 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2521 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2522 struct task_struct *next)
2524 struct preempt_notifier *notifier;
2525 struct hlist_node *node;
2527 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2528 notifier->ops->sched_out(notifier, next);
2531 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2533 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2538 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2539 struct task_struct *next)
2543 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2546 * prepare_task_switch - prepare to switch tasks
2547 * @rq: the runqueue preparing to switch
2548 * @prev: the current task that is being switched out
2549 * @next: the task we are going to switch to.
2551 * This is called with the rq lock held and interrupts off. It must
2552 * be paired with a subsequent finish_task_switch after the context
2555 * prepare_task_switch sets up locking and calls architecture specific
2559 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2560 struct task_struct *next)
2562 fire_sched_out_preempt_notifiers(prev, next);
2563 prepare_lock_switch(rq, next);
2564 prepare_arch_switch(next);
2568 * finish_task_switch - clean up after a task-switch
2569 * @rq: runqueue associated with task-switch
2570 * @prev: the thread we just switched away from.
2572 * finish_task_switch must be called after the context switch, paired
2573 * with a prepare_task_switch call before the context switch.
2574 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2575 * and do any other architecture-specific cleanup actions.
2577 * Note that we may have delayed dropping an mm in context_switch(). If
2578 * so, we finish that here outside of the runqueue lock. (Doing it
2579 * with the lock held can cause deadlocks; see schedule() for
2582 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2583 __releases(rq->lock)
2585 struct mm_struct *mm = rq->prev_mm;
2591 * A task struct has one reference for the use as "current".
2592 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2593 * schedule one last time. The schedule call will never return, and
2594 * the scheduled task must drop that reference.
2595 * The test for TASK_DEAD must occur while the runqueue locks are
2596 * still held, otherwise prev could be scheduled on another cpu, die
2597 * there before we look at prev->state, and then the reference would
2599 * Manfred Spraul <manfred@colorfullife.com>
2601 prev_state = prev->state;
2602 finish_arch_switch(prev);
2603 finish_lock_switch(rq, prev);
2605 if (current->sched_class->post_schedule)
2606 current->sched_class->post_schedule(rq);
2609 fire_sched_in_preempt_notifiers(current);
2612 if (unlikely(prev_state == TASK_DEAD)) {
2614 * Remove function-return probe instances associated with this
2615 * task and put them back on the free list.
2617 kprobe_flush_task(prev);
2618 put_task_struct(prev);
2623 * schedule_tail - first thing a freshly forked thread must call.
2624 * @prev: the thread we just switched away from.
2626 asmlinkage void schedule_tail(struct task_struct *prev)
2627 __releases(rq->lock)
2629 struct rq *rq = this_rq();
2631 finish_task_switch(rq, prev);
2632 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2633 /* In this case, finish_task_switch does not reenable preemption */
2636 if (current->set_child_tid)
2637 put_user(task_pid_vnr(current), current->set_child_tid);
2641 * context_switch - switch to the new MM and the new
2642 * thread's register state.
2645 context_switch(struct rq *rq, struct task_struct *prev,
2646 struct task_struct *next)
2648 struct mm_struct *mm, *oldmm;
2650 prepare_task_switch(rq, prev, next);
2651 trace_sched_switch(rq, prev, next);
2653 oldmm = prev->active_mm;
2655 * For paravirt, this is coupled with an exit in switch_to to
2656 * combine the page table reload and the switch backend into
2659 arch_enter_lazy_cpu_mode();
2661 if (unlikely(!mm)) {
2662 next->active_mm = oldmm;
2663 atomic_inc(&oldmm->mm_count);
2664 enter_lazy_tlb(oldmm, next);
2666 switch_mm(oldmm, mm, next);
2668 if (unlikely(!prev->mm)) {
2669 prev->active_mm = NULL;
2670 rq->prev_mm = oldmm;
2673 * Since the runqueue lock will be released by the next
2674 * task (which is an invalid locking op but in the case
2675 * of the scheduler it's an obvious special-case), so we
2676 * do an early lockdep release here:
2678 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2679 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2682 /* Here we just switch the register state and the stack. */
2683 switch_to(prev, next, prev);
2687 * this_rq must be evaluated again because prev may have moved
2688 * CPUs since it called schedule(), thus the 'rq' on its stack
2689 * frame will be invalid.
2691 finish_task_switch(this_rq(), prev);
2695 * nr_running, nr_uninterruptible and nr_context_switches:
2697 * externally visible scheduler statistics: current number of runnable
2698 * threads, current number of uninterruptible-sleeping threads, total
2699 * number of context switches performed since bootup.
2701 unsigned long nr_running(void)
2703 unsigned long i, sum = 0;
2705 for_each_online_cpu(i)
2706 sum += cpu_rq(i)->nr_running;
2711 unsigned long nr_uninterruptible(void)
2713 unsigned long i, sum = 0;
2715 for_each_possible_cpu(i)
2716 sum += cpu_rq(i)->nr_uninterruptible;
2719 * Since we read the counters lockless, it might be slightly
2720 * inaccurate. Do not allow it to go below zero though:
2722 if (unlikely((long)sum < 0))
2728 unsigned long long nr_context_switches(void)
2731 unsigned long long sum = 0;
2733 for_each_possible_cpu(i)
2734 sum += cpu_rq(i)->nr_switches;
2739 unsigned long nr_iowait(void)
2741 unsigned long i, sum = 0;
2743 for_each_possible_cpu(i)
2744 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2749 unsigned long nr_active(void)
2751 unsigned long i, running = 0, uninterruptible = 0;
2753 for_each_online_cpu(i) {
2754 running += cpu_rq(i)->nr_running;
2755 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2758 if (unlikely((long)uninterruptible < 0))
2759 uninterruptible = 0;
2761 return running + uninterruptible;
2765 * Update rq->cpu_load[] statistics. This function is usually called every
2766 * scheduler tick (TICK_NSEC).
2768 static void update_cpu_load(struct rq *this_rq)
2770 unsigned long this_load = this_rq->load.weight;
2773 this_rq->nr_load_updates++;
2775 /* Update our load: */
2776 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2777 unsigned long old_load, new_load;
2779 /* scale is effectively 1 << i now, and >> i divides by scale */
2781 old_load = this_rq->cpu_load[i];
2782 new_load = this_load;
2784 * Round up the averaging division if load is increasing. This
2785 * prevents us from getting stuck on 9 if the load is 10, for
2788 if (new_load > old_load)
2789 new_load += scale-1;
2790 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2797 * double_rq_lock - safely lock two runqueues
2799 * Note this does not disable interrupts like task_rq_lock,
2800 * you need to do so manually before calling.
2802 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2803 __acquires(rq1->lock)
2804 __acquires(rq2->lock)
2806 BUG_ON(!irqs_disabled());
2808 spin_lock(&rq1->lock);
2809 __acquire(rq2->lock); /* Fake it out ;) */
2812 spin_lock(&rq1->lock);
2813 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2815 spin_lock(&rq2->lock);
2816 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2819 update_rq_clock(rq1);
2820 update_rq_clock(rq2);
2824 * double_rq_unlock - safely unlock two runqueues
2826 * Note this does not restore interrupts like task_rq_unlock,
2827 * you need to do so manually after calling.
2829 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2830 __releases(rq1->lock)
2831 __releases(rq2->lock)
2833 spin_unlock(&rq1->lock);
2835 spin_unlock(&rq2->lock);
2837 __release(rq2->lock);
2841 * If dest_cpu is allowed for this process, migrate the task to it.
2842 * This is accomplished by forcing the cpu_allowed mask to only
2843 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2844 * the cpu_allowed mask is restored.
2846 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2848 struct migration_req req;
2849 unsigned long flags;
2852 rq = task_rq_lock(p, &flags);
2853 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2854 || unlikely(!cpu_active(dest_cpu)))
2857 trace_sched_migrate_task(rq, p, dest_cpu);
2858 /* force the process onto the specified CPU */
2859 if (migrate_task(p, dest_cpu, &req)) {
2860 /* Need to wait for migration thread (might exit: take ref). */
2861 struct task_struct *mt = rq->migration_thread;
2863 get_task_struct(mt);
2864 task_rq_unlock(rq, &flags);
2865 wake_up_process(mt);
2866 put_task_struct(mt);
2867 wait_for_completion(&req.done);
2872 task_rq_unlock(rq, &flags);
2876 * sched_exec - execve() is a valuable balancing opportunity, because at
2877 * this point the task has the smallest effective memory and cache footprint.
2879 void sched_exec(void)
2881 int new_cpu, this_cpu = get_cpu();
2882 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2884 if (new_cpu != this_cpu)
2885 sched_migrate_task(current, new_cpu);
2889 * pull_task - move a task from a remote runqueue to the local runqueue.
2890 * Both runqueues must be locked.
2892 static void pull_task(struct rq *src_rq, struct task_struct *p,
2893 struct rq *this_rq, int this_cpu)
2895 deactivate_task(src_rq, p, 0);
2896 set_task_cpu(p, this_cpu);
2897 activate_task(this_rq, p, 0);
2899 * Note that idle threads have a prio of MAX_PRIO, for this test
2900 * to be always true for them.
2902 check_preempt_curr(this_rq, p, 0);
2906 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2909 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2910 struct sched_domain *sd, enum cpu_idle_type idle,
2914 * We do not migrate tasks that are:
2915 * 1) running (obviously), or
2916 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2917 * 3) are cache-hot on their current CPU.
2919 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2920 schedstat_inc(p, se.nr_failed_migrations_affine);
2925 if (task_running(rq, p)) {
2926 schedstat_inc(p, se.nr_failed_migrations_running);
2931 * Aggressive migration if:
2932 * 1) task is cache cold, or
2933 * 2) too many balance attempts have failed.
2936 if (!task_hot(p, rq->clock, sd) ||
2937 sd->nr_balance_failed > sd->cache_nice_tries) {
2938 #ifdef CONFIG_SCHEDSTATS
2939 if (task_hot(p, rq->clock, sd)) {
2940 schedstat_inc(sd, lb_hot_gained[idle]);
2941 schedstat_inc(p, se.nr_forced_migrations);
2947 if (task_hot(p, rq->clock, sd)) {
2948 schedstat_inc(p, se.nr_failed_migrations_hot);
2954 static unsigned long
2955 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2956 unsigned long max_load_move, struct sched_domain *sd,
2957 enum cpu_idle_type idle, int *all_pinned,
2958 int *this_best_prio, struct rq_iterator *iterator)
2960 int loops = 0, pulled = 0, pinned = 0;
2961 struct task_struct *p;
2962 long rem_load_move = max_load_move;
2964 if (max_load_move == 0)
2970 * Start the load-balancing iterator:
2972 p = iterator->start(iterator->arg);
2974 if (!p || loops++ > sysctl_sched_nr_migrate)
2977 if ((p->se.load.weight >> 1) > rem_load_move ||
2978 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2979 p = iterator->next(iterator->arg);
2983 pull_task(busiest, p, this_rq, this_cpu);
2985 rem_load_move -= p->se.load.weight;
2987 #ifdef CONFIG_PREEMPT
2989 * NEWIDLE balancing is a source of latency, so preemptible kernels
2990 * will stop after the first task is pulled to minimize the critical
2993 if (idle == CPU_NEWLY_IDLE)
2998 * We only want to steal up to the prescribed amount of weighted load.
3000 if (rem_load_move > 0) {
3001 if (p->prio < *this_best_prio)
3002 *this_best_prio = p->prio;
3003 p = iterator->next(iterator->arg);
3008 * Right now, this is one of only two places pull_task() is called,
3009 * so we can safely collect pull_task() stats here rather than
3010 * inside pull_task().
3012 schedstat_add(sd, lb_gained[idle], pulled);
3015 *all_pinned = pinned;
3017 return max_load_move - rem_load_move;
3021 * move_tasks tries to move up to max_load_move weighted load from busiest to
3022 * this_rq, as part of a balancing operation within domain "sd".
3023 * Returns 1 if successful and 0 otherwise.
3025 * Called with both runqueues locked.
3027 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3028 unsigned long max_load_move,
3029 struct sched_domain *sd, enum cpu_idle_type idle,
3032 const struct sched_class *class = sched_class_highest;
3033 unsigned long total_load_moved = 0;
3034 int this_best_prio = this_rq->curr->prio;
3038 class->load_balance(this_rq, this_cpu, busiest,
3039 max_load_move - total_load_moved,
3040 sd, idle, all_pinned, &this_best_prio);
3041 class = class->next;
3043 #ifdef CONFIG_PREEMPT
3045 * NEWIDLE balancing is a source of latency, so preemptible
3046 * kernels will stop after the first task is pulled to minimize
3047 * the critical section.
3049 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3052 } while (class && max_load_move > total_load_moved);
3054 return total_load_moved > 0;
3058 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3059 struct sched_domain *sd, enum cpu_idle_type idle,
3060 struct rq_iterator *iterator)
3062 struct task_struct *p = iterator->start(iterator->arg);
3066 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3067 pull_task(busiest, p, this_rq, this_cpu);
3069 * Right now, this is only the second place pull_task()
3070 * is called, so we can safely collect pull_task()
3071 * stats here rather than inside pull_task().
3073 schedstat_inc(sd, lb_gained[idle]);
3077 p = iterator->next(iterator->arg);
3084 * move_one_task tries to move exactly one task from busiest to this_rq, as
3085 * part of active balancing operations within "domain".
3086 * Returns 1 if successful and 0 otherwise.
3088 * Called with both runqueues locked.
3090 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3091 struct sched_domain *sd, enum cpu_idle_type idle)
3093 const struct sched_class *class;
3095 for (class = sched_class_highest; class; class = class->next)
3096 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3103 * find_busiest_group finds and returns the busiest CPU group within the
3104 * domain. It calculates and returns the amount of weighted load which
3105 * should be moved to restore balance via the imbalance parameter.
3107 static struct sched_group *
3108 find_busiest_group(struct sched_domain *sd, int this_cpu,
3109 unsigned long *imbalance, enum cpu_idle_type idle,
3110 int *sd_idle, const struct cpumask *cpus, int *balance)
3112 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3113 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3114 unsigned long max_pull;
3115 unsigned long busiest_load_per_task, busiest_nr_running;
3116 unsigned long this_load_per_task, this_nr_running;
3117 int load_idx, group_imb = 0;
3118 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3119 int power_savings_balance = 1;
3120 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3121 unsigned long min_nr_running = ULONG_MAX;
3122 struct sched_group *group_min = NULL, *group_leader = NULL;
3125 max_load = this_load = total_load = total_pwr = 0;
3126 busiest_load_per_task = busiest_nr_running = 0;
3127 this_load_per_task = this_nr_running = 0;
3129 if (idle == CPU_NOT_IDLE)
3130 load_idx = sd->busy_idx;
3131 else if (idle == CPU_NEWLY_IDLE)
3132 load_idx = sd->newidle_idx;
3134 load_idx = sd->idle_idx;
3137 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3140 int __group_imb = 0;
3141 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3142 unsigned long sum_nr_running, sum_weighted_load;
3143 unsigned long sum_avg_load_per_task;
3144 unsigned long avg_load_per_task;
3146 local_group = cpumask_test_cpu(this_cpu,
3147 sched_group_cpus(group));
3150 balance_cpu = cpumask_first(sched_group_cpus(group));
3152 /* Tally up the load of all CPUs in the group */
3153 sum_weighted_load = sum_nr_running = avg_load = 0;
3154 sum_avg_load_per_task = avg_load_per_task = 0;
3157 min_cpu_load = ~0UL;
3159 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3160 struct rq *rq = cpu_rq(i);
3162 if (*sd_idle && rq->nr_running)
3165 /* Bias balancing toward cpus of our domain */
3167 if (idle_cpu(i) && !first_idle_cpu) {
3172 load = target_load(i, load_idx);
3174 load = source_load(i, load_idx);
3175 if (load > max_cpu_load)
3176 max_cpu_load = load;
3177 if (min_cpu_load > load)
3178 min_cpu_load = load;
3182 sum_nr_running += rq->nr_running;
3183 sum_weighted_load += weighted_cpuload(i);
3185 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3189 * First idle cpu or the first cpu(busiest) in this sched group
3190 * is eligible for doing load balancing at this and above
3191 * domains. In the newly idle case, we will allow all the cpu's
3192 * to do the newly idle load balance.
3194 if (idle != CPU_NEWLY_IDLE && local_group &&
3195 balance_cpu != this_cpu && balance) {
3200 total_load += avg_load;
3201 total_pwr += group->__cpu_power;
3203 /* Adjust by relative CPU power of the group */
3204 avg_load = sg_div_cpu_power(group,
3205 avg_load * SCHED_LOAD_SCALE);
3209 * Consider the group unbalanced when the imbalance is larger
3210 * than the average weight of two tasks.
3212 * APZ: with cgroup the avg task weight can vary wildly and
3213 * might not be a suitable number - should we keep a
3214 * normalized nr_running number somewhere that negates
3217 avg_load_per_task = sg_div_cpu_power(group,
3218 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3220 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3223 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3226 this_load = avg_load;
3228 this_nr_running = sum_nr_running;
3229 this_load_per_task = sum_weighted_load;
3230 } else if (avg_load > max_load &&
3231 (sum_nr_running > group_capacity || __group_imb)) {
3232 max_load = avg_load;
3234 busiest_nr_running = sum_nr_running;
3235 busiest_load_per_task = sum_weighted_load;
3236 group_imb = __group_imb;
3239 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3241 * Busy processors will not participate in power savings
3244 if (idle == CPU_NOT_IDLE ||
3245 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3249 * If the local group is idle or completely loaded
3250 * no need to do power savings balance at this domain
3252 if (local_group && (this_nr_running >= group_capacity ||
3254 power_savings_balance = 0;
3257 * If a group is already running at full capacity or idle,
3258 * don't include that group in power savings calculations
3260 if (!power_savings_balance || sum_nr_running >= group_capacity
3265 * Calculate the group which has the least non-idle load.
3266 * This is the group from where we need to pick up the load
3269 if ((sum_nr_running < min_nr_running) ||
3270 (sum_nr_running == min_nr_running &&
3271 cpumask_first(sched_group_cpus(group)) >
3272 cpumask_first(sched_group_cpus(group_min)))) {
3274 min_nr_running = sum_nr_running;
3275 min_load_per_task = sum_weighted_load /
3280 * Calculate the group which is almost near its
3281 * capacity but still has some space to pick up some load
3282 * from other group and save more power
3284 if (sum_nr_running <= group_capacity - 1) {
3285 if (sum_nr_running > leader_nr_running ||
3286 (sum_nr_running == leader_nr_running &&
3287 cpumask_first(sched_group_cpus(group)) <
3288 cpumask_first(sched_group_cpus(group_leader)))) {
3289 group_leader = group;
3290 leader_nr_running = sum_nr_running;
3295 group = group->next;
3296 } while (group != sd->groups);
3298 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3301 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3303 if (this_load >= avg_load ||
3304 100*max_load <= sd->imbalance_pct*this_load)
3307 busiest_load_per_task /= busiest_nr_running;
3309 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3312 * We're trying to get all the cpus to the average_load, so we don't
3313 * want to push ourselves above the average load, nor do we wish to
3314 * reduce the max loaded cpu below the average load, as either of these
3315 * actions would just result in more rebalancing later, and ping-pong
3316 * tasks around. Thus we look for the minimum possible imbalance.
3317 * Negative imbalances (*we* are more loaded than anyone else) will
3318 * be counted as no imbalance for these purposes -- we can't fix that
3319 * by pulling tasks to us. Be careful of negative numbers as they'll
3320 * appear as very large values with unsigned longs.
3322 if (max_load <= busiest_load_per_task)
3326 * In the presence of smp nice balancing, certain scenarios can have
3327 * max load less than avg load(as we skip the groups at or below
3328 * its cpu_power, while calculating max_load..)
3330 if (max_load < avg_load) {
3332 goto small_imbalance;
3335 /* Don't want to pull so many tasks that a group would go idle */
3336 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3338 /* How much load to actually move to equalise the imbalance */
3339 *imbalance = min(max_pull * busiest->__cpu_power,
3340 (avg_load - this_load) * this->__cpu_power)
3344 * if *imbalance is less than the average load per runnable task
3345 * there is no gaurantee that any tasks will be moved so we'll have
3346 * a think about bumping its value to force at least one task to be
3349 if (*imbalance < busiest_load_per_task) {
3350 unsigned long tmp, pwr_now, pwr_move;
3354 pwr_move = pwr_now = 0;
3356 if (this_nr_running) {
3357 this_load_per_task /= this_nr_running;
3358 if (busiest_load_per_task > this_load_per_task)
3361 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3363 if (max_load - this_load + busiest_load_per_task >=
3364 busiest_load_per_task * imbn) {
3365 *imbalance = busiest_load_per_task;
3370 * OK, we don't have enough imbalance to justify moving tasks,
3371 * however we may be able to increase total CPU power used by
3375 pwr_now += busiest->__cpu_power *
3376 min(busiest_load_per_task, max_load);
3377 pwr_now += this->__cpu_power *
3378 min(this_load_per_task, this_load);
3379 pwr_now /= SCHED_LOAD_SCALE;
3381 /* Amount of load we'd subtract */
3382 tmp = sg_div_cpu_power(busiest,
3383 busiest_load_per_task * SCHED_LOAD_SCALE);
3385 pwr_move += busiest->__cpu_power *
3386 min(busiest_load_per_task, max_load - tmp);
3388 /* Amount of load we'd add */
3389 if (max_load * busiest->__cpu_power <
3390 busiest_load_per_task * SCHED_LOAD_SCALE)
3391 tmp = sg_div_cpu_power(this,
3392 max_load * busiest->__cpu_power);
3394 tmp = sg_div_cpu_power(this,
3395 busiest_load_per_task * SCHED_LOAD_SCALE);
3396 pwr_move += this->__cpu_power *
3397 min(this_load_per_task, this_load + tmp);
3398 pwr_move /= SCHED_LOAD_SCALE;
3400 /* Move if we gain throughput */
3401 if (pwr_move > pwr_now)
3402 *imbalance = busiest_load_per_task;
3408 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3409 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3412 if (this == group_leader && group_leader != group_min) {
3413 *imbalance = min_load_per_task;
3414 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3415 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3416 cpumask_first(sched_group_cpus(group_leader));
3427 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3430 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3431 unsigned long imbalance, const struct cpumask *cpus)
3433 struct rq *busiest = NULL, *rq;
3434 unsigned long max_load = 0;
3437 for_each_cpu(i, sched_group_cpus(group)) {
3440 if (!cpumask_test_cpu(i, cpus))
3444 wl = weighted_cpuload(i);
3446 if (rq->nr_running == 1 && wl > imbalance)
3449 if (wl > max_load) {
3459 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3460 * so long as it is large enough.
3462 #define MAX_PINNED_INTERVAL 512
3465 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3466 * tasks if there is an imbalance.
3468 static int load_balance(int this_cpu, struct rq *this_rq,
3469 struct sched_domain *sd, enum cpu_idle_type idle,
3470 int *balance, struct cpumask *cpus)
3472 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3473 struct sched_group *group;
3474 unsigned long imbalance;
3476 unsigned long flags;
3478 cpumask_setall(cpus);
3481 * When power savings policy is enabled for the parent domain, idle
3482 * sibling can pick up load irrespective of busy siblings. In this case,
3483 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3484 * portraying it as CPU_NOT_IDLE.
3486 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3487 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3490 schedstat_inc(sd, lb_count[idle]);
3494 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3501 schedstat_inc(sd, lb_nobusyg[idle]);
3505 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3507 schedstat_inc(sd, lb_nobusyq[idle]);
3511 BUG_ON(busiest == this_rq);
3513 schedstat_add(sd, lb_imbalance[idle], imbalance);
3516 if (busiest->nr_running > 1) {
3518 * Attempt to move tasks. If find_busiest_group has found
3519 * an imbalance but busiest->nr_running <= 1, the group is
3520 * still unbalanced. ld_moved simply stays zero, so it is
3521 * correctly treated as an imbalance.
3523 local_irq_save(flags);
3524 double_rq_lock(this_rq, busiest);
3525 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3526 imbalance, sd, idle, &all_pinned);
3527 double_rq_unlock(this_rq, busiest);
3528 local_irq_restore(flags);
3531 * some other cpu did the load balance for us.
3533 if (ld_moved && this_cpu != smp_processor_id())
3534 resched_cpu(this_cpu);
3536 /* All tasks on this runqueue were pinned by CPU affinity */
3537 if (unlikely(all_pinned)) {
3538 cpumask_clear_cpu(cpu_of(busiest), cpus);
3539 if (!cpumask_empty(cpus))
3546 schedstat_inc(sd, lb_failed[idle]);
3547 sd->nr_balance_failed++;
3549 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3551 spin_lock_irqsave(&busiest->lock, flags);
3553 /* don't kick the migration_thread, if the curr
3554 * task on busiest cpu can't be moved to this_cpu
3556 if (!cpumask_test_cpu(this_cpu,
3557 &busiest->curr->cpus_allowed)) {
3558 spin_unlock_irqrestore(&busiest->lock, flags);
3560 goto out_one_pinned;
3563 if (!busiest->active_balance) {
3564 busiest->active_balance = 1;
3565 busiest->push_cpu = this_cpu;
3568 spin_unlock_irqrestore(&busiest->lock, flags);
3570 wake_up_process(busiest->migration_thread);
3573 * We've kicked active balancing, reset the failure
3576 sd->nr_balance_failed = sd->cache_nice_tries+1;
3579 sd->nr_balance_failed = 0;
3581 if (likely(!active_balance)) {
3582 /* We were unbalanced, so reset the balancing interval */
3583 sd->balance_interval = sd->min_interval;
3586 * If we've begun active balancing, start to back off. This
3587 * case may not be covered by the all_pinned logic if there
3588 * is only 1 task on the busy runqueue (because we don't call
3591 if (sd->balance_interval < sd->max_interval)
3592 sd->balance_interval *= 2;
3595 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3596 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3602 schedstat_inc(sd, lb_balanced[idle]);
3604 sd->nr_balance_failed = 0;
3607 /* tune up the balancing interval */
3608 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3609 (sd->balance_interval < sd->max_interval))
3610 sd->balance_interval *= 2;
3612 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3613 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3624 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3625 * tasks if there is an imbalance.
3627 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3628 * this_rq is locked.
3631 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3632 struct cpumask *cpus)
3634 struct sched_group *group;
3635 struct rq *busiest = NULL;
3636 unsigned long imbalance;
3641 cpumask_setall(cpus);
3644 * When power savings policy is enabled for the parent domain, idle
3645 * sibling can pick up load irrespective of busy siblings. In this case,
3646 * let the state of idle sibling percolate up as IDLE, instead of
3647 * portraying it as CPU_NOT_IDLE.
3649 if (sd->flags & SD_SHARE_CPUPOWER &&
3650 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3653 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3655 update_shares_locked(this_rq, sd);
3656 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3657 &sd_idle, cpus, NULL);
3659 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3663 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3665 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3669 BUG_ON(busiest == this_rq);
3671 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3674 if (busiest->nr_running > 1) {
3675 /* Attempt to move tasks */
3676 double_lock_balance(this_rq, busiest);
3677 /* this_rq->clock is already updated */
3678 update_rq_clock(busiest);
3679 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3680 imbalance, sd, CPU_NEWLY_IDLE,
3682 double_unlock_balance(this_rq, busiest);
3684 if (unlikely(all_pinned)) {
3685 cpumask_clear_cpu(cpu_of(busiest), cpus);
3686 if (!cpumask_empty(cpus))
3692 int active_balance = 0;
3694 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3695 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3696 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3699 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3702 if (sd->nr_balance_failed++ < 2)
3706 * The only task running in a non-idle cpu can be moved to this
3707 * cpu in an attempt to completely freeup the other CPU
3708 * package. The same method used to move task in load_balance()
3709 * have been extended for load_balance_newidle() to speedup
3710 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3712 * The package power saving logic comes from
3713 * find_busiest_group(). If there are no imbalance, then
3714 * f_b_g() will return NULL. However when sched_mc={1,2} then
3715 * f_b_g() will select a group from which a running task may be
3716 * pulled to this cpu in order to make the other package idle.
3717 * If there is no opportunity to make a package idle and if
3718 * there are no imbalance, then f_b_g() will return NULL and no
3719 * action will be taken in load_balance_newidle().
3721 * Under normal task pull operation due to imbalance, there
3722 * will be more than one task in the source run queue and
3723 * move_tasks() will succeed. ld_moved will be true and this
3724 * active balance code will not be triggered.
3727 /* Lock busiest in correct order while this_rq is held */
3728 double_lock_balance(this_rq, busiest);
3731 * don't kick the migration_thread, if the curr
3732 * task on busiest cpu can't be moved to this_cpu
3734 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3735 double_unlock_balance(this_rq, busiest);
3740 if (!busiest->active_balance) {
3741 busiest->active_balance = 1;
3742 busiest->push_cpu = this_cpu;
3746 double_unlock_balance(this_rq, busiest);
3748 wake_up_process(busiest->migration_thread);
3751 sd->nr_balance_failed = 0;
3753 update_shares_locked(this_rq, sd);
3757 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3758 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3759 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3761 sd->nr_balance_failed = 0;
3767 * idle_balance is called by schedule() if this_cpu is about to become
3768 * idle. Attempts to pull tasks from other CPUs.
3770 static void idle_balance(int this_cpu, struct rq *this_rq)
3772 struct sched_domain *sd;
3773 int pulled_task = 0;
3774 unsigned long next_balance = jiffies + HZ;
3775 cpumask_var_t tmpmask;
3777 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3780 for_each_domain(this_cpu, sd) {
3781 unsigned long interval;
3783 if (!(sd->flags & SD_LOAD_BALANCE))
3786 if (sd->flags & SD_BALANCE_NEWIDLE)
3787 /* If we've pulled tasks over stop searching: */
3788 pulled_task = load_balance_newidle(this_cpu, this_rq,
3791 interval = msecs_to_jiffies(sd->balance_interval);
3792 if (time_after(next_balance, sd->last_balance + interval))
3793 next_balance = sd->last_balance + interval;
3797 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3799 * We are going idle. next_balance may be set based on
3800 * a busy processor. So reset next_balance.
3802 this_rq->next_balance = next_balance;
3804 free_cpumask_var(tmpmask);
3808 * active_load_balance is run by migration threads. It pushes running tasks
3809 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3810 * running on each physical CPU where possible, and avoids physical /
3811 * logical imbalances.
3813 * Called with busiest_rq locked.
3815 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3817 int target_cpu = busiest_rq->push_cpu;
3818 struct sched_domain *sd;
3819 struct rq *target_rq;
3821 /* Is there any task to move? */
3822 if (busiest_rq->nr_running <= 1)
3825 target_rq = cpu_rq(target_cpu);
3828 * This condition is "impossible", if it occurs
3829 * we need to fix it. Originally reported by
3830 * Bjorn Helgaas on a 128-cpu setup.
3832 BUG_ON(busiest_rq == target_rq);
3834 /* move a task from busiest_rq to target_rq */
3835 double_lock_balance(busiest_rq, target_rq);
3836 update_rq_clock(busiest_rq);
3837 update_rq_clock(target_rq);
3839 /* Search for an sd spanning us and the target CPU. */
3840 for_each_domain(target_cpu, sd) {
3841 if ((sd->flags & SD_LOAD_BALANCE) &&
3842 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3847 schedstat_inc(sd, alb_count);
3849 if (move_one_task(target_rq, target_cpu, busiest_rq,
3851 schedstat_inc(sd, alb_pushed);
3853 schedstat_inc(sd, alb_failed);
3855 double_unlock_balance(busiest_rq, target_rq);
3860 atomic_t load_balancer;
3861 cpumask_var_t cpu_mask;
3862 } nohz ____cacheline_aligned = {
3863 .load_balancer = ATOMIC_INIT(-1),
3867 * This routine will try to nominate the ilb (idle load balancing)
3868 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3869 * load balancing on behalf of all those cpus. If all the cpus in the system
3870 * go into this tickless mode, then there will be no ilb owner (as there is
3871 * no need for one) and all the cpus will sleep till the next wakeup event
3874 * For the ilb owner, tick is not stopped. And this tick will be used
3875 * for idle load balancing. ilb owner will still be part of
3878 * While stopping the tick, this cpu will become the ilb owner if there
3879 * is no other owner. And will be the owner till that cpu becomes busy
3880 * or if all cpus in the system stop their ticks at which point
3881 * there is no need for ilb owner.
3883 * When the ilb owner becomes busy, it nominates another owner, during the
3884 * next busy scheduler_tick()
3886 int select_nohz_load_balancer(int stop_tick)
3888 int cpu = smp_processor_id();
3891 cpumask_set_cpu(cpu, nohz.cpu_mask);
3892 cpu_rq(cpu)->in_nohz_recently = 1;
3895 * If we are going offline and still the leader, give up!
3897 if (!cpu_active(cpu) &&
3898 atomic_read(&nohz.load_balancer) == cpu) {
3899 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3904 /* time for ilb owner also to sleep */
3905 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3906 if (atomic_read(&nohz.load_balancer) == cpu)
3907 atomic_set(&nohz.load_balancer, -1);
3911 if (atomic_read(&nohz.load_balancer) == -1) {
3912 /* make me the ilb owner */
3913 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3915 } else if (atomic_read(&nohz.load_balancer) == cpu)
3918 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3921 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3923 if (atomic_read(&nohz.load_balancer) == cpu)
3924 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3931 static DEFINE_SPINLOCK(balancing);
3934 * It checks each scheduling domain to see if it is due to be balanced,
3935 * and initiates a balancing operation if so.
3937 * Balancing parameters are set up in arch_init_sched_domains.
3939 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3942 struct rq *rq = cpu_rq(cpu);
3943 unsigned long interval;
3944 struct sched_domain *sd;
3945 /* Earliest time when we have to do rebalance again */
3946 unsigned long next_balance = jiffies + 60*HZ;
3947 int update_next_balance = 0;
3951 /* Fails alloc? Rebalancing probably not a priority right now. */
3952 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3955 for_each_domain(cpu, sd) {
3956 if (!(sd->flags & SD_LOAD_BALANCE))
3959 interval = sd->balance_interval;
3960 if (idle != CPU_IDLE)
3961 interval *= sd->busy_factor;
3963 /* scale ms to jiffies */
3964 interval = msecs_to_jiffies(interval);
3965 if (unlikely(!interval))
3967 if (interval > HZ*NR_CPUS/10)
3968 interval = HZ*NR_CPUS/10;
3970 need_serialize = sd->flags & SD_SERIALIZE;
3972 if (need_serialize) {
3973 if (!spin_trylock(&balancing))
3977 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3978 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
3980 * We've pulled tasks over so either we're no
3981 * longer idle, or one of our SMT siblings is
3984 idle = CPU_NOT_IDLE;
3986 sd->last_balance = jiffies;
3989 spin_unlock(&balancing);
3991 if (time_after(next_balance, sd->last_balance + interval)) {
3992 next_balance = sd->last_balance + interval;
3993 update_next_balance = 1;
3997 * Stop the load balance at this level. There is another
3998 * CPU in our sched group which is doing load balancing more
4006 * next_balance will be updated only when there is a need.
4007 * When the cpu is attached to null domain for ex, it will not be
4010 if (likely(update_next_balance))
4011 rq->next_balance = next_balance;
4013 free_cpumask_var(tmp);
4017 * run_rebalance_domains is triggered when needed from the scheduler tick.
4018 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4019 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4021 static void run_rebalance_domains(struct softirq_action *h)
4023 int this_cpu = smp_processor_id();
4024 struct rq *this_rq = cpu_rq(this_cpu);
4025 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4026 CPU_IDLE : CPU_NOT_IDLE;
4028 rebalance_domains(this_cpu, idle);
4032 * If this cpu is the owner for idle load balancing, then do the
4033 * balancing on behalf of the other idle cpus whose ticks are
4036 if (this_rq->idle_at_tick &&
4037 atomic_read(&nohz.load_balancer) == this_cpu) {
4041 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4042 if (balance_cpu == this_cpu)
4046 * If this cpu gets work to do, stop the load balancing
4047 * work being done for other cpus. Next load
4048 * balancing owner will pick it up.
4053 rebalance_domains(balance_cpu, CPU_IDLE);
4055 rq = cpu_rq(balance_cpu);
4056 if (time_after(this_rq->next_balance, rq->next_balance))
4057 this_rq->next_balance = rq->next_balance;
4064 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4066 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4067 * idle load balancing owner or decide to stop the periodic load balancing,
4068 * if the whole system is idle.
4070 static inline void trigger_load_balance(struct rq *rq, int cpu)
4074 * If we were in the nohz mode recently and busy at the current
4075 * scheduler tick, then check if we need to nominate new idle
4078 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4079 rq->in_nohz_recently = 0;
4081 if (atomic_read(&nohz.load_balancer) == cpu) {
4082 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4083 atomic_set(&nohz.load_balancer, -1);
4086 if (atomic_read(&nohz.load_balancer) == -1) {
4088 * simple selection for now: Nominate the
4089 * first cpu in the nohz list to be the next
4092 * TBD: Traverse the sched domains and nominate
4093 * the nearest cpu in the nohz.cpu_mask.
4095 int ilb = cpumask_first(nohz.cpu_mask);
4097 if (ilb < nr_cpu_ids)
4103 * If this cpu is idle and doing idle load balancing for all the
4104 * cpus with ticks stopped, is it time for that to stop?
4106 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4107 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4113 * If this cpu is idle and the idle load balancing is done by
4114 * someone else, then no need raise the SCHED_SOFTIRQ
4116 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4117 cpumask_test_cpu(cpu, nohz.cpu_mask))
4120 if (time_after_eq(jiffies, rq->next_balance))
4121 raise_softirq(SCHED_SOFTIRQ);
4124 #else /* CONFIG_SMP */
4127 * on UP we do not need to balance between CPUs:
4129 static inline void idle_balance(int cpu, struct rq *rq)
4135 DEFINE_PER_CPU(struct kernel_stat, kstat);
4137 EXPORT_PER_CPU_SYMBOL(kstat);
4140 * Return any ns on the sched_clock that have not yet been banked in
4141 * @p in case that task is currently running.
4143 unsigned long long task_delta_exec(struct task_struct *p)
4145 unsigned long flags;
4149 rq = task_rq_lock(p, &flags);
4151 if (task_current(rq, p)) {
4154 update_rq_clock(rq);
4155 delta_exec = rq->clock - p->se.exec_start;
4156 if ((s64)delta_exec > 0)
4160 task_rq_unlock(rq, &flags);
4166 * Account user cpu time to a process.
4167 * @p: the process that the cpu time gets accounted to
4168 * @cputime: the cpu time spent in user space since the last update
4170 void account_user_time(struct task_struct *p, cputime_t cputime)
4172 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4175 p->utime = cputime_add(p->utime, cputime);
4176 account_group_user_time(p, cputime);
4178 /* Add user time to cpustat. */
4179 tmp = cputime_to_cputime64(cputime);
4180 if (TASK_NICE(p) > 0)
4181 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4183 cpustat->user = cputime64_add(cpustat->user, tmp);
4184 /* Account for user time used */
4185 acct_update_integrals(p);
4189 * Account guest cpu time to a process.
4190 * @p: the process that the cpu time gets accounted to
4191 * @cputime: the cpu time spent in virtual machine since the last update
4193 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4196 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4198 tmp = cputime_to_cputime64(cputime);
4200 p->utime = cputime_add(p->utime, cputime);
4201 account_group_user_time(p, cputime);
4202 p->gtime = cputime_add(p->gtime, cputime);
4204 cpustat->user = cputime64_add(cpustat->user, tmp);
4205 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4209 * Account scaled user cpu time to a process.
4210 * @p: the process that the cpu time gets accounted to
4211 * @cputime: the cpu time spent in user space since the last update
4213 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4215 p->utimescaled = cputime_add(p->utimescaled, cputime);
4219 * Account system cpu time to a process.
4220 * @p: the process that the cpu time gets accounted to
4221 * @hardirq_offset: the offset to subtract from hardirq_count()
4222 * @cputime: the cpu time spent in kernel space since the last update
4224 void account_system_time(struct task_struct *p, int hardirq_offset,
4227 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4228 struct rq *rq = this_rq();
4231 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4232 account_guest_time(p, cputime);
4236 p->stime = cputime_add(p->stime, cputime);
4237 account_group_system_time(p, cputime);
4239 /* Add system time to cpustat. */
4240 tmp = cputime_to_cputime64(cputime);
4241 if (hardirq_count() - hardirq_offset)
4242 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4243 else if (softirq_count())
4244 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4245 else if (p != rq->idle)
4246 cpustat->system = cputime64_add(cpustat->system, tmp);
4247 else if (atomic_read(&rq->nr_iowait) > 0)
4248 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4250 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4251 /* Account for system time used */
4252 acct_update_integrals(p);
4256 * Account scaled system cpu time to a process.
4257 * @p: the process that the cpu time gets accounted to
4258 * @hardirq_offset: the offset to subtract from hardirq_count()
4259 * @cputime: the cpu time spent in kernel space since the last update
4261 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4263 p->stimescaled = cputime_add(p->stimescaled, cputime);
4267 * Account for involuntary wait time.
4268 * @p: the process from which the cpu time has been stolen
4269 * @steal: the cpu time spent in involuntary wait
4271 void account_steal_time(struct task_struct *p, cputime_t steal)
4273 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4274 cputime64_t tmp = cputime_to_cputime64(steal);
4275 struct rq *rq = this_rq();
4277 if (p == rq->idle) {
4278 p->stime = cputime_add(p->stime, steal);
4279 if (atomic_read(&rq->nr_iowait) > 0)
4280 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4282 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4284 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4288 * Use precise platform statistics if available:
4290 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4291 cputime_t task_utime(struct task_struct *p)
4296 cputime_t task_stime(struct task_struct *p)
4301 cputime_t task_utime(struct task_struct *p)
4303 clock_t utime = cputime_to_clock_t(p->utime),
4304 total = utime + cputime_to_clock_t(p->stime);
4308 * Use CFS's precise accounting:
4310 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4314 do_div(temp, total);
4316 utime = (clock_t)temp;
4318 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4319 return p->prev_utime;
4322 cputime_t task_stime(struct task_struct *p)
4327 * Use CFS's precise accounting. (we subtract utime from
4328 * the total, to make sure the total observed by userspace
4329 * grows monotonically - apps rely on that):
4331 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4332 cputime_to_clock_t(task_utime(p));
4335 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4337 return p->prev_stime;
4341 inline cputime_t task_gtime(struct task_struct *p)
4347 * This function gets called by the timer code, with HZ frequency.
4348 * We call it with interrupts disabled.
4350 * It also gets called by the fork code, when changing the parent's
4353 void scheduler_tick(void)
4355 int cpu = smp_processor_id();
4356 struct rq *rq = cpu_rq(cpu);
4357 struct task_struct *curr = rq->curr;
4361 spin_lock(&rq->lock);
4362 update_rq_clock(rq);
4363 update_cpu_load(rq);
4364 curr->sched_class->task_tick(rq, curr, 0);
4365 spin_unlock(&rq->lock);
4368 rq->idle_at_tick = idle_cpu(cpu);
4369 trigger_load_balance(rq, cpu);
4373 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4374 defined(CONFIG_PREEMPT_TRACER))
4376 static inline unsigned long get_parent_ip(unsigned long addr)
4378 if (in_lock_functions(addr)) {
4379 addr = CALLER_ADDR2;
4380 if (in_lock_functions(addr))
4381 addr = CALLER_ADDR3;
4386 void __kprobes add_preempt_count(int val)
4388 #ifdef CONFIG_DEBUG_PREEMPT
4392 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4395 preempt_count() += val;
4396 #ifdef CONFIG_DEBUG_PREEMPT
4398 * Spinlock count overflowing soon?
4400 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4403 if (preempt_count() == val)
4404 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4406 EXPORT_SYMBOL(add_preempt_count);
4408 void __kprobes sub_preempt_count(int val)
4410 #ifdef CONFIG_DEBUG_PREEMPT
4414 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4417 * Is the spinlock portion underflowing?
4419 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4420 !(preempt_count() & PREEMPT_MASK)))
4424 if (preempt_count() == val)
4425 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4426 preempt_count() -= val;
4428 EXPORT_SYMBOL(sub_preempt_count);
4433 * Print scheduling while atomic bug:
4435 static noinline void __schedule_bug(struct task_struct *prev)
4437 struct pt_regs *regs = get_irq_regs();
4439 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4440 prev->comm, prev->pid, preempt_count());
4442 debug_show_held_locks(prev);
4444 if (irqs_disabled())
4445 print_irqtrace_events(prev);
4454 * Various schedule()-time debugging checks and statistics:
4456 static inline void schedule_debug(struct task_struct *prev)
4459 * Test if we are atomic. Since do_exit() needs to call into
4460 * schedule() atomically, we ignore that path for now.
4461 * Otherwise, whine if we are scheduling when we should not be.
4463 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4464 __schedule_bug(prev);
4466 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4468 schedstat_inc(this_rq(), sched_count);
4469 #ifdef CONFIG_SCHEDSTATS
4470 if (unlikely(prev->lock_depth >= 0)) {
4471 schedstat_inc(this_rq(), bkl_count);
4472 schedstat_inc(prev, sched_info.bkl_count);
4478 * Pick up the highest-prio task:
4480 static inline struct task_struct *
4481 pick_next_task(struct rq *rq, struct task_struct *prev)
4483 const struct sched_class *class;
4484 struct task_struct *p;
4487 * Optimization: we know that if all tasks are in
4488 * the fair class we can call that function directly:
4490 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4491 p = fair_sched_class.pick_next_task(rq);
4496 class = sched_class_highest;
4498 p = class->pick_next_task(rq);
4502 * Will never be NULL as the idle class always
4503 * returns a non-NULL p:
4505 class = class->next;
4510 * schedule() is the main scheduler function.
4512 asmlinkage void __sched schedule(void)
4514 struct task_struct *prev, *next;
4515 unsigned long *switch_count;
4521 cpu = smp_processor_id();
4525 switch_count = &prev->nivcsw;
4527 release_kernel_lock(prev);
4528 need_resched_nonpreemptible:
4530 schedule_debug(prev);
4532 if (sched_feat(HRTICK))
4535 spin_lock_irq(&rq->lock);
4536 update_rq_clock(rq);
4537 clear_tsk_need_resched(prev);
4539 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4540 if (unlikely(signal_pending_state(prev->state, prev)))
4541 prev->state = TASK_RUNNING;
4543 deactivate_task(rq, prev, 1);
4544 switch_count = &prev->nvcsw;
4548 if (prev->sched_class->pre_schedule)
4549 prev->sched_class->pre_schedule(rq, prev);
4552 if (unlikely(!rq->nr_running))
4553 idle_balance(cpu, rq);
4555 prev->sched_class->put_prev_task(rq, prev);
4556 next = pick_next_task(rq, prev);
4558 if (likely(prev != next)) {
4559 sched_info_switch(prev, next);
4565 context_switch(rq, prev, next); /* unlocks the rq */
4567 * the context switch might have flipped the stack from under
4568 * us, hence refresh the local variables.
4570 cpu = smp_processor_id();
4573 spin_unlock_irq(&rq->lock);
4575 if (unlikely(reacquire_kernel_lock(current) < 0))
4576 goto need_resched_nonpreemptible;
4578 preempt_enable_no_resched();
4579 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4582 EXPORT_SYMBOL(schedule);
4584 #ifdef CONFIG_PREEMPT
4586 * this is the entry point to schedule() from in-kernel preemption
4587 * off of preempt_enable. Kernel preemptions off return from interrupt
4588 * occur there and call schedule directly.
4590 asmlinkage void __sched preempt_schedule(void)
4592 struct thread_info *ti = current_thread_info();
4595 * If there is a non-zero preempt_count or interrupts are disabled,
4596 * we do not want to preempt the current task. Just return..
4598 if (likely(ti->preempt_count || irqs_disabled()))
4602 add_preempt_count(PREEMPT_ACTIVE);
4604 sub_preempt_count(PREEMPT_ACTIVE);
4607 * Check again in case we missed a preemption opportunity
4608 * between schedule and now.
4611 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4613 EXPORT_SYMBOL(preempt_schedule);
4616 * this is the entry point to schedule() from kernel preemption
4617 * off of irq context.
4618 * Note, that this is called and return with irqs disabled. This will
4619 * protect us against recursive calling from irq.
4621 asmlinkage void __sched preempt_schedule_irq(void)
4623 struct thread_info *ti = current_thread_info();
4625 /* Catch callers which need to be fixed */
4626 BUG_ON(ti->preempt_count || !irqs_disabled());
4629 add_preempt_count(PREEMPT_ACTIVE);
4632 local_irq_disable();
4633 sub_preempt_count(PREEMPT_ACTIVE);
4636 * Check again in case we missed a preemption opportunity
4637 * between schedule and now.
4640 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4643 #endif /* CONFIG_PREEMPT */
4645 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4648 return try_to_wake_up(curr->private, mode, sync);
4650 EXPORT_SYMBOL(default_wake_function);
4653 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4654 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4655 * number) then we wake all the non-exclusive tasks and one exclusive task.
4657 * There are circumstances in which we can try to wake a task which has already
4658 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4659 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4661 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4662 int nr_exclusive, int sync, void *key)
4664 wait_queue_t *curr, *next;
4666 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4667 unsigned flags = curr->flags;
4669 if (curr->func(curr, mode, sync, key) &&
4670 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4676 * __wake_up - wake up threads blocked on a waitqueue.
4678 * @mode: which threads
4679 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4680 * @key: is directly passed to the wakeup function
4682 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4683 int nr_exclusive, void *key)
4685 unsigned long flags;
4687 spin_lock_irqsave(&q->lock, flags);
4688 __wake_up_common(q, mode, nr_exclusive, 0, key);
4689 spin_unlock_irqrestore(&q->lock, flags);
4691 EXPORT_SYMBOL(__wake_up);
4694 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4696 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4698 __wake_up_common(q, mode, 1, 0, NULL);
4702 * __wake_up_sync - wake up threads blocked on a waitqueue.
4704 * @mode: which threads
4705 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4707 * The sync wakeup differs that the waker knows that it will schedule
4708 * away soon, so while the target thread will be woken up, it will not
4709 * be migrated to another CPU - ie. the two threads are 'synchronized'
4710 * with each other. This can prevent needless bouncing between CPUs.
4712 * On UP it can prevent extra preemption.
4715 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4717 unsigned long flags;
4723 if (unlikely(!nr_exclusive))
4726 spin_lock_irqsave(&q->lock, flags);
4727 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4728 spin_unlock_irqrestore(&q->lock, flags);
4730 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4733 * complete: - signals a single thread waiting on this completion
4734 * @x: holds the state of this particular completion
4736 * This will wake up a single thread waiting on this completion. Threads will be
4737 * awakened in the same order in which they were queued.
4739 * See also complete_all(), wait_for_completion() and related routines.
4741 void complete(struct completion *x)
4743 unsigned long flags;
4745 spin_lock_irqsave(&x->wait.lock, flags);
4747 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4748 spin_unlock_irqrestore(&x->wait.lock, flags);
4750 EXPORT_SYMBOL(complete);
4753 * complete_all: - signals all threads waiting on this completion
4754 * @x: holds the state of this particular completion
4756 * This will wake up all threads waiting on this particular completion event.
4758 void complete_all(struct completion *x)
4760 unsigned long flags;
4762 spin_lock_irqsave(&x->wait.lock, flags);
4763 x->done += UINT_MAX/2;
4764 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4765 spin_unlock_irqrestore(&x->wait.lock, flags);
4767 EXPORT_SYMBOL(complete_all);
4769 static inline long __sched
4770 do_wait_for_common(struct completion *x, long timeout, int state)
4773 DECLARE_WAITQUEUE(wait, current);
4775 wait.flags |= WQ_FLAG_EXCLUSIVE;
4776 __add_wait_queue_tail(&x->wait, &wait);
4778 if (signal_pending_state(state, current)) {
4779 timeout = -ERESTARTSYS;
4782 __set_current_state(state);
4783 spin_unlock_irq(&x->wait.lock);
4784 timeout = schedule_timeout(timeout);
4785 spin_lock_irq(&x->wait.lock);
4786 } while (!x->done && timeout);
4787 __remove_wait_queue(&x->wait, &wait);
4792 return timeout ?: 1;
4796 wait_for_common(struct completion *x, long timeout, int state)
4800 spin_lock_irq(&x->wait.lock);
4801 timeout = do_wait_for_common(x, timeout, state);
4802 spin_unlock_irq(&x->wait.lock);
4807 * wait_for_completion: - waits for completion of a task
4808 * @x: holds the state of this particular completion
4810 * This waits to be signaled for completion of a specific task. It is NOT
4811 * interruptible and there is no timeout.
4813 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4814 * and interrupt capability. Also see complete().
4816 void __sched wait_for_completion(struct completion *x)
4818 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4820 EXPORT_SYMBOL(wait_for_completion);
4823 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4824 * @x: holds the state of this particular completion
4825 * @timeout: timeout value in jiffies
4827 * This waits for either a completion of a specific task to be signaled or for a
4828 * specified timeout to expire. The timeout is in jiffies. It is not
4831 unsigned long __sched
4832 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4834 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4836 EXPORT_SYMBOL(wait_for_completion_timeout);
4839 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4840 * @x: holds the state of this particular completion
4842 * This waits for completion of a specific task to be signaled. It is
4845 int __sched wait_for_completion_interruptible(struct completion *x)
4847 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4848 if (t == -ERESTARTSYS)
4852 EXPORT_SYMBOL(wait_for_completion_interruptible);
4855 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4856 * @x: holds the state of this particular completion
4857 * @timeout: timeout value in jiffies
4859 * This waits for either a completion of a specific task to be signaled or for a
4860 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4862 unsigned long __sched
4863 wait_for_completion_interruptible_timeout(struct completion *x,
4864 unsigned long timeout)
4866 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4868 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4871 * wait_for_completion_killable: - waits for completion of a task (killable)
4872 * @x: holds the state of this particular completion
4874 * This waits to be signaled for completion of a specific task. It can be
4875 * interrupted by a kill signal.
4877 int __sched wait_for_completion_killable(struct completion *x)
4879 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4880 if (t == -ERESTARTSYS)
4884 EXPORT_SYMBOL(wait_for_completion_killable);
4887 * try_wait_for_completion - try to decrement a completion without blocking
4888 * @x: completion structure
4890 * Returns: 0 if a decrement cannot be done without blocking
4891 * 1 if a decrement succeeded.
4893 * If a completion is being used as a counting completion,
4894 * attempt to decrement the counter without blocking. This
4895 * enables us to avoid waiting if the resource the completion
4896 * is protecting is not available.
4898 bool try_wait_for_completion(struct completion *x)
4902 spin_lock_irq(&x->wait.lock);
4907 spin_unlock_irq(&x->wait.lock);
4910 EXPORT_SYMBOL(try_wait_for_completion);
4913 * completion_done - Test to see if a completion has any waiters
4914 * @x: completion structure
4916 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4917 * 1 if there are no waiters.
4920 bool completion_done(struct completion *x)
4924 spin_lock_irq(&x->wait.lock);
4927 spin_unlock_irq(&x->wait.lock);
4930 EXPORT_SYMBOL(completion_done);
4933 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4935 unsigned long flags;
4938 init_waitqueue_entry(&wait, current);
4940 __set_current_state(state);
4942 spin_lock_irqsave(&q->lock, flags);
4943 __add_wait_queue(q, &wait);
4944 spin_unlock(&q->lock);
4945 timeout = schedule_timeout(timeout);
4946 spin_lock_irq(&q->lock);
4947 __remove_wait_queue(q, &wait);
4948 spin_unlock_irqrestore(&q->lock, flags);
4953 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4955 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4957 EXPORT_SYMBOL(interruptible_sleep_on);
4960 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4962 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4964 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4966 void __sched sleep_on(wait_queue_head_t *q)
4968 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4970 EXPORT_SYMBOL(sleep_on);
4972 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4974 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4976 EXPORT_SYMBOL(sleep_on_timeout);
4978 #ifdef CONFIG_RT_MUTEXES
4981 * rt_mutex_setprio - set the current priority of a task
4983 * @prio: prio value (kernel-internal form)
4985 * This function changes the 'effective' priority of a task. It does
4986 * not touch ->normal_prio like __setscheduler().
4988 * Used by the rt_mutex code to implement priority inheritance logic.
4990 void rt_mutex_setprio(struct task_struct *p, int prio)
4992 unsigned long flags;
4993 int oldprio, on_rq, running;
4995 const struct sched_class *prev_class = p->sched_class;
4997 BUG_ON(prio < 0 || prio > MAX_PRIO);
4999 rq = task_rq_lock(p, &flags);
5000 update_rq_clock(rq);
5003 on_rq = p->se.on_rq;
5004 running = task_current(rq, p);
5006 dequeue_task(rq, p, 0);
5008 p->sched_class->put_prev_task(rq, p);
5011 p->sched_class = &rt_sched_class;
5013 p->sched_class = &fair_sched_class;
5018 p->sched_class->set_curr_task(rq);
5020 enqueue_task(rq, p, 0);
5022 check_class_changed(rq, p, prev_class, oldprio, running);
5024 task_rq_unlock(rq, &flags);
5029 void set_user_nice(struct task_struct *p, long nice)
5031 int old_prio, delta, on_rq;
5032 unsigned long flags;
5035 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5038 * We have to be careful, if called from sys_setpriority(),
5039 * the task might be in the middle of scheduling on another CPU.
5041 rq = task_rq_lock(p, &flags);
5042 update_rq_clock(rq);
5044 * The RT priorities are set via sched_setscheduler(), but we still
5045 * allow the 'normal' nice value to be set - but as expected
5046 * it wont have any effect on scheduling until the task is
5047 * SCHED_FIFO/SCHED_RR:
5049 if (task_has_rt_policy(p)) {
5050 p->static_prio = NICE_TO_PRIO(nice);
5053 on_rq = p->se.on_rq;
5055 dequeue_task(rq, p, 0);
5057 p->static_prio = NICE_TO_PRIO(nice);
5060 p->prio = effective_prio(p);
5061 delta = p->prio - old_prio;
5064 enqueue_task(rq, p, 0);
5066 * If the task increased its priority or is running and
5067 * lowered its priority, then reschedule its CPU:
5069 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5070 resched_task(rq->curr);
5073 task_rq_unlock(rq, &flags);
5075 EXPORT_SYMBOL(set_user_nice);
5078 * can_nice - check if a task can reduce its nice value
5082 int can_nice(const struct task_struct *p, const int nice)
5084 /* convert nice value [19,-20] to rlimit style value [1,40] */
5085 int nice_rlim = 20 - nice;
5087 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5088 capable(CAP_SYS_NICE));
5091 #ifdef __ARCH_WANT_SYS_NICE
5094 * sys_nice - change the priority of the current process.
5095 * @increment: priority increment
5097 * sys_setpriority is a more generic, but much slower function that
5098 * does similar things.
5100 asmlinkage long sys_nice(int increment)
5105 * Setpriority might change our priority at the same moment.
5106 * We don't have to worry. Conceptually one call occurs first
5107 * and we have a single winner.
5109 if (increment < -40)
5114 nice = PRIO_TO_NICE(current->static_prio) + increment;
5120 if (increment < 0 && !can_nice(current, nice))
5123 retval = security_task_setnice(current, nice);
5127 set_user_nice(current, nice);
5134 * task_prio - return the priority value of a given task.
5135 * @p: the task in question.
5137 * This is the priority value as seen by users in /proc.
5138 * RT tasks are offset by -200. Normal tasks are centered
5139 * around 0, value goes from -16 to +15.
5141 int task_prio(const struct task_struct *p)
5143 return p->prio - MAX_RT_PRIO;
5147 * task_nice - return the nice value of a given task.
5148 * @p: the task in question.
5150 int task_nice(const struct task_struct *p)
5152 return TASK_NICE(p);
5154 EXPORT_SYMBOL(task_nice);
5157 * idle_cpu - is a given cpu idle currently?
5158 * @cpu: the processor in question.
5160 int idle_cpu(int cpu)
5162 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5166 * idle_task - return the idle task for a given cpu.
5167 * @cpu: the processor in question.
5169 struct task_struct *idle_task(int cpu)
5171 return cpu_rq(cpu)->idle;
5175 * find_process_by_pid - find a process with a matching PID value.
5176 * @pid: the pid in question.
5178 static struct task_struct *find_process_by_pid(pid_t pid)
5180 return pid ? find_task_by_vpid(pid) : current;
5183 /* Actually do priority change: must hold rq lock. */
5185 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5187 BUG_ON(p->se.on_rq);
5190 switch (p->policy) {
5194 p->sched_class = &fair_sched_class;
5198 p->sched_class = &rt_sched_class;
5202 p->rt_priority = prio;
5203 p->normal_prio = normal_prio(p);
5204 /* we are holding p->pi_lock already */
5205 p->prio = rt_mutex_getprio(p);
5209 static int __sched_setscheduler(struct task_struct *p, int policy,
5210 struct sched_param *param, bool user)
5212 int retval, oldprio, oldpolicy = -1, on_rq, running;
5213 unsigned long flags;
5214 const struct sched_class *prev_class = p->sched_class;
5217 /* may grab non-irq protected spin_locks */
5218 BUG_ON(in_interrupt());
5220 /* double check policy once rq lock held */
5222 policy = oldpolicy = p->policy;
5223 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5224 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5225 policy != SCHED_IDLE)
5228 * Valid priorities for SCHED_FIFO and SCHED_RR are
5229 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5230 * SCHED_BATCH and SCHED_IDLE is 0.
5232 if (param->sched_priority < 0 ||
5233 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5234 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5236 if (rt_policy(policy) != (param->sched_priority != 0))
5240 * Allow unprivileged RT tasks to decrease priority:
5242 if (user && !capable(CAP_SYS_NICE)) {
5243 if (rt_policy(policy)) {
5244 unsigned long rlim_rtprio;
5246 if (!lock_task_sighand(p, &flags))
5248 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5249 unlock_task_sighand(p, &flags);
5251 /* can't set/change the rt policy */
5252 if (policy != p->policy && !rlim_rtprio)
5255 /* can't increase priority */
5256 if (param->sched_priority > p->rt_priority &&
5257 param->sched_priority > rlim_rtprio)
5261 * Like positive nice levels, dont allow tasks to
5262 * move out of SCHED_IDLE either:
5264 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5267 /* can't change other user's priorities */
5268 if ((current->euid != p->euid) &&
5269 (current->euid != p->uid))
5274 #ifdef CONFIG_RT_GROUP_SCHED
5276 * Do not allow realtime tasks into groups that have no runtime
5279 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5280 task_group(p)->rt_bandwidth.rt_runtime == 0)
5284 retval = security_task_setscheduler(p, policy, param);
5290 * make sure no PI-waiters arrive (or leave) while we are
5291 * changing the priority of the task:
5293 spin_lock_irqsave(&p->pi_lock, flags);
5295 * To be able to change p->policy safely, the apropriate
5296 * runqueue lock must be held.
5298 rq = __task_rq_lock(p);
5299 /* recheck policy now with rq lock held */
5300 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5301 policy = oldpolicy = -1;
5302 __task_rq_unlock(rq);
5303 spin_unlock_irqrestore(&p->pi_lock, flags);
5306 update_rq_clock(rq);
5307 on_rq = p->se.on_rq;
5308 running = task_current(rq, p);
5310 deactivate_task(rq, p, 0);
5312 p->sched_class->put_prev_task(rq, p);
5315 __setscheduler(rq, p, policy, param->sched_priority);
5318 p->sched_class->set_curr_task(rq);
5320 activate_task(rq, p, 0);
5322 check_class_changed(rq, p, prev_class, oldprio, running);
5324 __task_rq_unlock(rq);
5325 spin_unlock_irqrestore(&p->pi_lock, flags);
5327 rt_mutex_adjust_pi(p);
5333 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5334 * @p: the task in question.
5335 * @policy: new policy.
5336 * @param: structure containing the new RT priority.
5338 * NOTE that the task may be already dead.
5340 int sched_setscheduler(struct task_struct *p, int policy,
5341 struct sched_param *param)
5343 return __sched_setscheduler(p, policy, param, true);
5345 EXPORT_SYMBOL_GPL(sched_setscheduler);
5348 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5349 * @p: the task in question.
5350 * @policy: new policy.
5351 * @param: structure containing the new RT priority.
5353 * Just like sched_setscheduler, only don't bother checking if the
5354 * current context has permission. For example, this is needed in
5355 * stop_machine(): we create temporary high priority worker threads,
5356 * but our caller might not have that capability.
5358 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5359 struct sched_param *param)
5361 return __sched_setscheduler(p, policy, param, false);
5365 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5367 struct sched_param lparam;
5368 struct task_struct *p;
5371 if (!param || pid < 0)
5373 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5378 p = find_process_by_pid(pid);
5380 retval = sched_setscheduler(p, policy, &lparam);
5387 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5388 * @pid: the pid in question.
5389 * @policy: new policy.
5390 * @param: structure containing the new RT priority.
5393 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5395 /* negative values for policy are not valid */
5399 return do_sched_setscheduler(pid, policy, param);
5403 * sys_sched_setparam - set/change the RT priority of a thread
5404 * @pid: the pid in question.
5405 * @param: structure containing the new RT priority.
5407 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5409 return do_sched_setscheduler(pid, -1, param);
5413 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5414 * @pid: the pid in question.
5416 asmlinkage long sys_sched_getscheduler(pid_t pid)
5418 struct task_struct *p;
5425 read_lock(&tasklist_lock);
5426 p = find_process_by_pid(pid);
5428 retval = security_task_getscheduler(p);
5432 read_unlock(&tasklist_lock);
5437 * sys_sched_getscheduler - get the RT priority of a thread
5438 * @pid: the pid in question.
5439 * @param: structure containing the RT priority.
5441 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5443 struct sched_param lp;
5444 struct task_struct *p;
5447 if (!param || pid < 0)
5450 read_lock(&tasklist_lock);
5451 p = find_process_by_pid(pid);
5456 retval = security_task_getscheduler(p);
5460 lp.sched_priority = p->rt_priority;
5461 read_unlock(&tasklist_lock);
5464 * This one might sleep, we cannot do it with a spinlock held ...
5466 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5471 read_unlock(&tasklist_lock);
5475 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5477 cpumask_var_t cpus_allowed, new_mask;
5478 struct task_struct *p;
5482 read_lock(&tasklist_lock);
5484 p = find_process_by_pid(pid);
5486 read_unlock(&tasklist_lock);
5492 * It is not safe to call set_cpus_allowed with the
5493 * tasklist_lock held. We will bump the task_struct's
5494 * usage count and then drop tasklist_lock.
5497 read_unlock(&tasklist_lock);
5499 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5503 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5505 goto out_free_cpus_allowed;
5508 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5509 !capable(CAP_SYS_NICE))
5512 retval = security_task_setscheduler(p, 0, NULL);
5516 cpuset_cpus_allowed(p, cpus_allowed);
5517 cpumask_and(new_mask, in_mask, cpus_allowed);
5519 retval = set_cpus_allowed_ptr(p, new_mask);
5522 cpuset_cpus_allowed(p, cpus_allowed);
5523 if (!cpumask_subset(new_mask, cpus_allowed)) {
5525 * We must have raced with a concurrent cpuset
5526 * update. Just reset the cpus_allowed to the
5527 * cpuset's cpus_allowed
5529 cpumask_copy(new_mask, cpus_allowed);
5534 free_cpumask_var(new_mask);
5535 out_free_cpus_allowed:
5536 free_cpumask_var(cpus_allowed);
5543 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5544 struct cpumask *new_mask)
5546 if (len < cpumask_size())
5547 cpumask_clear(new_mask);
5548 else if (len > cpumask_size())
5549 len = cpumask_size();
5551 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5555 * sys_sched_setaffinity - set the cpu affinity of a process
5556 * @pid: pid of the process
5557 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5558 * @user_mask_ptr: user-space pointer to the new cpu mask
5560 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5561 unsigned long __user *user_mask_ptr)
5563 cpumask_var_t new_mask;
5566 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5569 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5571 retval = sched_setaffinity(pid, new_mask);
5572 free_cpumask_var(new_mask);
5576 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5578 struct task_struct *p;
5582 read_lock(&tasklist_lock);
5585 p = find_process_by_pid(pid);
5589 retval = security_task_getscheduler(p);
5593 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5596 read_unlock(&tasklist_lock);
5603 * sys_sched_getaffinity - get the cpu affinity of a process
5604 * @pid: pid of the process
5605 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5606 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5608 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5609 unsigned long __user *user_mask_ptr)
5614 if (len < cpumask_size())
5617 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5620 ret = sched_getaffinity(pid, mask);
5622 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5625 ret = cpumask_size();
5627 free_cpumask_var(mask);
5633 * sys_sched_yield - yield the current processor to other threads.
5635 * This function yields the current CPU to other tasks. If there are no
5636 * other threads running on this CPU then this function will return.
5638 asmlinkage long sys_sched_yield(void)
5640 struct rq *rq = this_rq_lock();
5642 schedstat_inc(rq, yld_count);
5643 current->sched_class->yield_task(rq);
5646 * Since we are going to call schedule() anyway, there's
5647 * no need to preempt or enable interrupts:
5649 __release(rq->lock);
5650 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5651 _raw_spin_unlock(&rq->lock);
5652 preempt_enable_no_resched();
5659 static void __cond_resched(void)
5661 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5662 __might_sleep(__FILE__, __LINE__);
5665 * The BKS might be reacquired before we have dropped
5666 * PREEMPT_ACTIVE, which could trigger a second
5667 * cond_resched() call.
5670 add_preempt_count(PREEMPT_ACTIVE);
5672 sub_preempt_count(PREEMPT_ACTIVE);
5673 } while (need_resched());
5676 int __sched _cond_resched(void)
5678 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5679 system_state == SYSTEM_RUNNING) {
5685 EXPORT_SYMBOL(_cond_resched);
5688 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5689 * call schedule, and on return reacquire the lock.
5691 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5692 * operations here to prevent schedule() from being called twice (once via
5693 * spin_unlock(), once by hand).
5695 int cond_resched_lock(spinlock_t *lock)
5697 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5700 if (spin_needbreak(lock) || resched) {
5702 if (resched && need_resched())
5711 EXPORT_SYMBOL(cond_resched_lock);
5713 int __sched cond_resched_softirq(void)
5715 BUG_ON(!in_softirq());
5717 if (need_resched() && system_state == SYSTEM_RUNNING) {
5725 EXPORT_SYMBOL(cond_resched_softirq);
5728 * yield - yield the current processor to other threads.
5730 * This is a shortcut for kernel-space yielding - it marks the
5731 * thread runnable and calls sys_sched_yield().
5733 void __sched yield(void)
5735 set_current_state(TASK_RUNNING);
5738 EXPORT_SYMBOL(yield);
5741 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5742 * that process accounting knows that this is a task in IO wait state.
5744 * But don't do that if it is a deliberate, throttling IO wait (this task
5745 * has set its backing_dev_info: the queue against which it should throttle)
5747 void __sched io_schedule(void)
5749 struct rq *rq = &__raw_get_cpu_var(runqueues);
5751 delayacct_blkio_start();
5752 atomic_inc(&rq->nr_iowait);
5754 atomic_dec(&rq->nr_iowait);
5755 delayacct_blkio_end();
5757 EXPORT_SYMBOL(io_schedule);
5759 long __sched io_schedule_timeout(long timeout)
5761 struct rq *rq = &__raw_get_cpu_var(runqueues);
5764 delayacct_blkio_start();
5765 atomic_inc(&rq->nr_iowait);
5766 ret = schedule_timeout(timeout);
5767 atomic_dec(&rq->nr_iowait);
5768 delayacct_blkio_end();
5773 * sys_sched_get_priority_max - return maximum RT priority.
5774 * @policy: scheduling class.
5776 * this syscall returns the maximum rt_priority that can be used
5777 * by a given scheduling class.
5779 asmlinkage long sys_sched_get_priority_max(int policy)
5786 ret = MAX_USER_RT_PRIO-1;
5798 * sys_sched_get_priority_min - return minimum RT priority.
5799 * @policy: scheduling class.
5801 * this syscall returns the minimum rt_priority that can be used
5802 * by a given scheduling class.
5804 asmlinkage long sys_sched_get_priority_min(int policy)
5822 * sys_sched_rr_get_interval - return the default timeslice of a process.
5823 * @pid: pid of the process.
5824 * @interval: userspace pointer to the timeslice value.
5826 * this syscall writes the default timeslice value of a given process
5827 * into the user-space timespec buffer. A value of '0' means infinity.
5830 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5832 struct task_struct *p;
5833 unsigned int time_slice;
5841 read_lock(&tasklist_lock);
5842 p = find_process_by_pid(pid);
5846 retval = security_task_getscheduler(p);
5851 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5852 * tasks that are on an otherwise idle runqueue:
5855 if (p->policy == SCHED_RR) {
5856 time_slice = DEF_TIMESLICE;
5857 } else if (p->policy != SCHED_FIFO) {
5858 struct sched_entity *se = &p->se;
5859 unsigned long flags;
5862 rq = task_rq_lock(p, &flags);
5863 if (rq->cfs.load.weight)
5864 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5865 task_rq_unlock(rq, &flags);
5867 read_unlock(&tasklist_lock);
5868 jiffies_to_timespec(time_slice, &t);
5869 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5873 read_unlock(&tasklist_lock);
5877 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5879 void sched_show_task(struct task_struct *p)
5881 unsigned long free = 0;
5884 state = p->state ? __ffs(p->state) + 1 : 0;
5885 printk(KERN_INFO "%-13.13s %c", p->comm,
5886 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5887 #if BITS_PER_LONG == 32
5888 if (state == TASK_RUNNING)
5889 printk(KERN_CONT " running ");
5891 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5893 if (state == TASK_RUNNING)
5894 printk(KERN_CONT " running task ");
5896 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5898 #ifdef CONFIG_DEBUG_STACK_USAGE
5900 unsigned long *n = end_of_stack(p);
5903 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5906 printk(KERN_CONT "%5lu %5d %6d\n", free,
5907 task_pid_nr(p), task_pid_nr(p->real_parent));
5909 show_stack(p, NULL);
5912 void show_state_filter(unsigned long state_filter)
5914 struct task_struct *g, *p;
5916 #if BITS_PER_LONG == 32
5918 " task PC stack pid father\n");
5921 " task PC stack pid father\n");
5923 read_lock(&tasklist_lock);
5924 do_each_thread(g, p) {
5926 * reset the NMI-timeout, listing all files on a slow
5927 * console might take alot of time:
5929 touch_nmi_watchdog();
5930 if (!state_filter || (p->state & state_filter))
5932 } while_each_thread(g, p);
5934 touch_all_softlockup_watchdogs();
5936 #ifdef CONFIG_SCHED_DEBUG
5937 sysrq_sched_debug_show();
5939 read_unlock(&tasklist_lock);
5941 * Only show locks if all tasks are dumped:
5943 if (state_filter == -1)
5944 debug_show_all_locks();
5947 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5949 idle->sched_class = &idle_sched_class;
5953 * init_idle - set up an idle thread for a given CPU
5954 * @idle: task in question
5955 * @cpu: cpu the idle task belongs to
5957 * NOTE: this function does not set the idle thread's NEED_RESCHED
5958 * flag, to make booting more robust.
5960 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5962 struct rq *rq = cpu_rq(cpu);
5963 unsigned long flags;
5965 spin_lock_irqsave(&rq->lock, flags);
5968 idle->se.exec_start = sched_clock();
5970 idle->prio = idle->normal_prio = MAX_PRIO;
5971 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
5972 __set_task_cpu(idle, cpu);
5974 rq->curr = rq->idle = idle;
5975 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5978 spin_unlock_irqrestore(&rq->lock, flags);
5980 /* Set the preempt count _outside_ the spinlocks! */
5981 #if defined(CONFIG_PREEMPT)
5982 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5984 task_thread_info(idle)->preempt_count = 0;
5987 * The idle tasks have their own, simple scheduling class:
5989 idle->sched_class = &idle_sched_class;
5990 ftrace_graph_init_task(idle);
5994 * In a system that switches off the HZ timer nohz_cpu_mask
5995 * indicates which cpus entered this state. This is used
5996 * in the rcu update to wait only for active cpus. For system
5997 * which do not switch off the HZ timer nohz_cpu_mask should
5998 * always be CPU_BITS_NONE.
6000 cpumask_var_t nohz_cpu_mask;
6003 * Increase the granularity value when there are more CPUs,
6004 * because with more CPUs the 'effective latency' as visible
6005 * to users decreases. But the relationship is not linear,
6006 * so pick a second-best guess by going with the log2 of the
6009 * This idea comes from the SD scheduler of Con Kolivas:
6011 static inline void sched_init_granularity(void)
6013 unsigned int factor = 1 + ilog2(num_online_cpus());
6014 const unsigned long limit = 200000000;
6016 sysctl_sched_min_granularity *= factor;
6017 if (sysctl_sched_min_granularity > limit)
6018 sysctl_sched_min_granularity = limit;
6020 sysctl_sched_latency *= factor;
6021 if (sysctl_sched_latency > limit)
6022 sysctl_sched_latency = limit;
6024 sysctl_sched_wakeup_granularity *= factor;
6026 sysctl_sched_shares_ratelimit *= factor;
6031 * This is how migration works:
6033 * 1) we queue a struct migration_req structure in the source CPU's
6034 * runqueue and wake up that CPU's migration thread.
6035 * 2) we down() the locked semaphore => thread blocks.
6036 * 3) migration thread wakes up (implicitly it forces the migrated
6037 * thread off the CPU)
6038 * 4) it gets the migration request and checks whether the migrated
6039 * task is still in the wrong runqueue.
6040 * 5) if it's in the wrong runqueue then the migration thread removes
6041 * it and puts it into the right queue.
6042 * 6) migration thread up()s the semaphore.
6043 * 7) we wake up and the migration is done.
6047 * Change a given task's CPU affinity. Migrate the thread to a
6048 * proper CPU and schedule it away if the CPU it's executing on
6049 * is removed from the allowed bitmask.
6051 * NOTE: the caller must have a valid reference to the task, the
6052 * task must not exit() & deallocate itself prematurely. The
6053 * call is not atomic; no spinlocks may be held.
6055 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6057 struct migration_req req;
6058 unsigned long flags;
6062 rq = task_rq_lock(p, &flags);
6063 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6068 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6069 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6074 if (p->sched_class->set_cpus_allowed)
6075 p->sched_class->set_cpus_allowed(p, new_mask);
6077 cpumask_copy(&p->cpus_allowed, new_mask);
6078 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6081 /* Can the task run on the task's current CPU? If so, we're done */
6082 if (cpumask_test_cpu(task_cpu(p), new_mask))
6085 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6086 /* Need help from migration thread: drop lock and wait. */
6087 task_rq_unlock(rq, &flags);
6088 wake_up_process(rq->migration_thread);
6089 wait_for_completion(&req.done);
6090 tlb_migrate_finish(p->mm);
6094 task_rq_unlock(rq, &flags);
6098 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6101 * Move (not current) task off this cpu, onto dest cpu. We're doing
6102 * this because either it can't run here any more (set_cpus_allowed()
6103 * away from this CPU, or CPU going down), or because we're
6104 * attempting to rebalance this task on exec (sched_exec).
6106 * So we race with normal scheduler movements, but that's OK, as long
6107 * as the task is no longer on this CPU.
6109 * Returns non-zero if task was successfully migrated.
6111 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6113 struct rq *rq_dest, *rq_src;
6116 if (unlikely(!cpu_active(dest_cpu)))
6119 rq_src = cpu_rq(src_cpu);
6120 rq_dest = cpu_rq(dest_cpu);
6122 double_rq_lock(rq_src, rq_dest);
6123 /* Already moved. */
6124 if (task_cpu(p) != src_cpu)
6126 /* Affinity changed (again). */
6127 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6130 on_rq = p->se.on_rq;
6132 deactivate_task(rq_src, p, 0);
6134 set_task_cpu(p, dest_cpu);
6136 activate_task(rq_dest, p, 0);
6137 check_preempt_curr(rq_dest, p, 0);
6142 double_rq_unlock(rq_src, rq_dest);
6147 * migration_thread - this is a highprio system thread that performs
6148 * thread migration by bumping thread off CPU then 'pushing' onto
6151 static int migration_thread(void *data)
6153 int cpu = (long)data;
6157 BUG_ON(rq->migration_thread != current);
6159 set_current_state(TASK_INTERRUPTIBLE);
6160 while (!kthread_should_stop()) {
6161 struct migration_req *req;
6162 struct list_head *head;
6164 spin_lock_irq(&rq->lock);
6166 if (cpu_is_offline(cpu)) {
6167 spin_unlock_irq(&rq->lock);
6171 if (rq->active_balance) {
6172 active_load_balance(rq, cpu);
6173 rq->active_balance = 0;
6176 head = &rq->migration_queue;
6178 if (list_empty(head)) {
6179 spin_unlock_irq(&rq->lock);
6181 set_current_state(TASK_INTERRUPTIBLE);
6184 req = list_entry(head->next, struct migration_req, list);
6185 list_del_init(head->next);
6187 spin_unlock(&rq->lock);
6188 __migrate_task(req->task, cpu, req->dest_cpu);
6191 complete(&req->done);
6193 __set_current_state(TASK_RUNNING);
6197 /* Wait for kthread_stop */
6198 set_current_state(TASK_INTERRUPTIBLE);
6199 while (!kthread_should_stop()) {
6201 set_current_state(TASK_INTERRUPTIBLE);
6203 __set_current_state(TASK_RUNNING);
6207 #ifdef CONFIG_HOTPLUG_CPU
6209 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6213 local_irq_disable();
6214 ret = __migrate_task(p, src_cpu, dest_cpu);
6220 * Figure out where task on dead CPU should go, use force if necessary.
6222 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6225 /* FIXME: Use cpumask_of_node here. */
6226 cpumask_t _nodemask = node_to_cpumask(cpu_to_node(dead_cpu));
6227 const struct cpumask *nodemask = &_nodemask;
6230 /* Look for allowed, online CPU in same node. */
6231 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6232 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6235 /* Any allowed, online CPU? */
6236 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6237 if (dest_cpu < nr_cpu_ids)
6240 /* No more Mr. Nice Guy. */
6241 if (dest_cpu >= nr_cpu_ids) {
6242 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6243 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6246 * Don't tell them about moving exiting tasks or
6247 * kernel threads (both mm NULL), since they never
6250 if (p->mm && printk_ratelimit()) {
6251 printk(KERN_INFO "process %d (%s) no "
6252 "longer affine to cpu%d\n",
6253 task_pid_nr(p), p->comm, dead_cpu);
6258 /* It can have affinity changed while we were choosing. */
6259 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6264 * While a dead CPU has no uninterruptible tasks queued at this point,
6265 * it might still have a nonzero ->nr_uninterruptible counter, because
6266 * for performance reasons the counter is not stricly tracking tasks to
6267 * their home CPUs. So we just add the counter to another CPU's counter,
6268 * to keep the global sum constant after CPU-down:
6270 static void migrate_nr_uninterruptible(struct rq *rq_src)
6272 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6273 unsigned long flags;
6275 local_irq_save(flags);
6276 double_rq_lock(rq_src, rq_dest);
6277 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6278 rq_src->nr_uninterruptible = 0;
6279 double_rq_unlock(rq_src, rq_dest);
6280 local_irq_restore(flags);
6283 /* Run through task list and migrate tasks from the dead cpu. */
6284 static void migrate_live_tasks(int src_cpu)
6286 struct task_struct *p, *t;
6288 read_lock(&tasklist_lock);
6290 do_each_thread(t, p) {
6294 if (task_cpu(p) == src_cpu)
6295 move_task_off_dead_cpu(src_cpu, p);
6296 } while_each_thread(t, p);
6298 read_unlock(&tasklist_lock);
6302 * Schedules idle task to be the next runnable task on current CPU.
6303 * It does so by boosting its priority to highest possible.
6304 * Used by CPU offline code.
6306 void sched_idle_next(void)
6308 int this_cpu = smp_processor_id();
6309 struct rq *rq = cpu_rq(this_cpu);
6310 struct task_struct *p = rq->idle;
6311 unsigned long flags;
6313 /* cpu has to be offline */
6314 BUG_ON(cpu_online(this_cpu));
6317 * Strictly not necessary since rest of the CPUs are stopped by now
6318 * and interrupts disabled on the current cpu.
6320 spin_lock_irqsave(&rq->lock, flags);
6322 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6324 update_rq_clock(rq);
6325 activate_task(rq, p, 0);
6327 spin_unlock_irqrestore(&rq->lock, flags);
6331 * Ensures that the idle task is using init_mm right before its cpu goes
6334 void idle_task_exit(void)
6336 struct mm_struct *mm = current->active_mm;
6338 BUG_ON(cpu_online(smp_processor_id()));
6341 switch_mm(mm, &init_mm, current);
6345 /* called under rq->lock with disabled interrupts */
6346 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6348 struct rq *rq = cpu_rq(dead_cpu);
6350 /* Must be exiting, otherwise would be on tasklist. */
6351 BUG_ON(!p->exit_state);
6353 /* Cannot have done final schedule yet: would have vanished. */
6354 BUG_ON(p->state == TASK_DEAD);
6359 * Drop lock around migration; if someone else moves it,
6360 * that's OK. No task can be added to this CPU, so iteration is
6363 spin_unlock_irq(&rq->lock);
6364 move_task_off_dead_cpu(dead_cpu, p);
6365 spin_lock_irq(&rq->lock);
6370 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6371 static void migrate_dead_tasks(unsigned int dead_cpu)
6373 struct rq *rq = cpu_rq(dead_cpu);
6374 struct task_struct *next;
6377 if (!rq->nr_running)
6379 update_rq_clock(rq);
6380 next = pick_next_task(rq, rq->curr);
6383 next->sched_class->put_prev_task(rq, next);
6384 migrate_dead(dead_cpu, next);
6388 #endif /* CONFIG_HOTPLUG_CPU */
6390 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6392 static struct ctl_table sd_ctl_dir[] = {
6394 .procname = "sched_domain",
6400 static struct ctl_table sd_ctl_root[] = {
6402 .ctl_name = CTL_KERN,
6403 .procname = "kernel",
6405 .child = sd_ctl_dir,
6410 static struct ctl_table *sd_alloc_ctl_entry(int n)
6412 struct ctl_table *entry =
6413 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6418 static void sd_free_ctl_entry(struct ctl_table **tablep)
6420 struct ctl_table *entry;
6423 * In the intermediate directories, both the child directory and
6424 * procname are dynamically allocated and could fail but the mode
6425 * will always be set. In the lowest directory the names are
6426 * static strings and all have proc handlers.
6428 for (entry = *tablep; entry->mode; entry++) {
6430 sd_free_ctl_entry(&entry->child);
6431 if (entry->proc_handler == NULL)
6432 kfree(entry->procname);
6440 set_table_entry(struct ctl_table *entry,
6441 const char *procname, void *data, int maxlen,
6442 mode_t mode, proc_handler *proc_handler)
6444 entry->procname = procname;
6446 entry->maxlen = maxlen;
6448 entry->proc_handler = proc_handler;
6451 static struct ctl_table *
6452 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6454 struct ctl_table *table = sd_alloc_ctl_entry(13);
6459 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6460 sizeof(long), 0644, proc_doulongvec_minmax);
6461 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6462 sizeof(long), 0644, proc_doulongvec_minmax);
6463 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6464 sizeof(int), 0644, proc_dointvec_minmax);
6465 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6466 sizeof(int), 0644, proc_dointvec_minmax);
6467 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6468 sizeof(int), 0644, proc_dointvec_minmax);
6469 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6470 sizeof(int), 0644, proc_dointvec_minmax);
6471 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6472 sizeof(int), 0644, proc_dointvec_minmax);
6473 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6474 sizeof(int), 0644, proc_dointvec_minmax);
6475 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6476 sizeof(int), 0644, proc_dointvec_minmax);
6477 set_table_entry(&table[9], "cache_nice_tries",
6478 &sd->cache_nice_tries,
6479 sizeof(int), 0644, proc_dointvec_minmax);
6480 set_table_entry(&table[10], "flags", &sd->flags,
6481 sizeof(int), 0644, proc_dointvec_minmax);
6482 set_table_entry(&table[11], "name", sd->name,
6483 CORENAME_MAX_SIZE, 0444, proc_dostring);
6484 /* &table[12] is terminator */
6489 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6491 struct ctl_table *entry, *table;
6492 struct sched_domain *sd;
6493 int domain_num = 0, i;
6496 for_each_domain(cpu, sd)
6498 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6503 for_each_domain(cpu, sd) {
6504 snprintf(buf, 32, "domain%d", i);
6505 entry->procname = kstrdup(buf, GFP_KERNEL);
6507 entry->child = sd_alloc_ctl_domain_table(sd);
6514 static struct ctl_table_header *sd_sysctl_header;
6515 static void register_sched_domain_sysctl(void)
6517 int i, cpu_num = num_online_cpus();
6518 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6521 WARN_ON(sd_ctl_dir[0].child);
6522 sd_ctl_dir[0].child = entry;
6527 for_each_online_cpu(i) {
6528 snprintf(buf, 32, "cpu%d", i);
6529 entry->procname = kstrdup(buf, GFP_KERNEL);
6531 entry->child = sd_alloc_ctl_cpu_table(i);
6535 WARN_ON(sd_sysctl_header);
6536 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6539 /* may be called multiple times per register */
6540 static void unregister_sched_domain_sysctl(void)
6542 if (sd_sysctl_header)
6543 unregister_sysctl_table(sd_sysctl_header);
6544 sd_sysctl_header = NULL;
6545 if (sd_ctl_dir[0].child)
6546 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6549 static void register_sched_domain_sysctl(void)
6552 static void unregister_sched_domain_sysctl(void)
6557 static void set_rq_online(struct rq *rq)
6560 const struct sched_class *class;
6562 cpumask_set_cpu(rq->cpu, rq->rd->online);
6565 for_each_class(class) {
6566 if (class->rq_online)
6567 class->rq_online(rq);
6572 static void set_rq_offline(struct rq *rq)
6575 const struct sched_class *class;
6577 for_each_class(class) {
6578 if (class->rq_offline)
6579 class->rq_offline(rq);
6582 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6588 * migration_call - callback that gets triggered when a CPU is added.
6589 * Here we can start up the necessary migration thread for the new CPU.
6591 static int __cpuinit
6592 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6594 struct task_struct *p;
6595 int cpu = (long)hcpu;
6596 unsigned long flags;
6601 case CPU_UP_PREPARE:
6602 case CPU_UP_PREPARE_FROZEN:
6603 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6606 kthread_bind(p, cpu);
6607 /* Must be high prio: stop_machine expects to yield to it. */
6608 rq = task_rq_lock(p, &flags);
6609 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6610 task_rq_unlock(rq, &flags);
6611 cpu_rq(cpu)->migration_thread = p;
6615 case CPU_ONLINE_FROZEN:
6616 /* Strictly unnecessary, as first user will wake it. */
6617 wake_up_process(cpu_rq(cpu)->migration_thread);
6619 /* Update our root-domain */
6621 spin_lock_irqsave(&rq->lock, flags);
6623 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6627 spin_unlock_irqrestore(&rq->lock, flags);
6630 #ifdef CONFIG_HOTPLUG_CPU
6631 case CPU_UP_CANCELED:
6632 case CPU_UP_CANCELED_FROZEN:
6633 if (!cpu_rq(cpu)->migration_thread)
6635 /* Unbind it from offline cpu so it can run. Fall thru. */
6636 kthread_bind(cpu_rq(cpu)->migration_thread,
6637 cpumask_any(cpu_online_mask));
6638 kthread_stop(cpu_rq(cpu)->migration_thread);
6639 cpu_rq(cpu)->migration_thread = NULL;
6643 case CPU_DEAD_FROZEN:
6644 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6645 migrate_live_tasks(cpu);
6647 kthread_stop(rq->migration_thread);
6648 rq->migration_thread = NULL;
6649 /* Idle task back to normal (off runqueue, low prio) */
6650 spin_lock_irq(&rq->lock);
6651 update_rq_clock(rq);
6652 deactivate_task(rq, rq->idle, 0);
6653 rq->idle->static_prio = MAX_PRIO;
6654 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6655 rq->idle->sched_class = &idle_sched_class;
6656 migrate_dead_tasks(cpu);
6657 spin_unlock_irq(&rq->lock);
6659 migrate_nr_uninterruptible(rq);
6660 BUG_ON(rq->nr_running != 0);
6663 * No need to migrate the tasks: it was best-effort if
6664 * they didn't take sched_hotcpu_mutex. Just wake up
6667 spin_lock_irq(&rq->lock);
6668 while (!list_empty(&rq->migration_queue)) {
6669 struct migration_req *req;
6671 req = list_entry(rq->migration_queue.next,
6672 struct migration_req, list);
6673 list_del_init(&req->list);
6674 spin_unlock_irq(&rq->lock);
6675 complete(&req->done);
6676 spin_lock_irq(&rq->lock);
6678 spin_unlock_irq(&rq->lock);
6682 case CPU_DYING_FROZEN:
6683 /* Update our root-domain */
6685 spin_lock_irqsave(&rq->lock, flags);
6687 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6690 spin_unlock_irqrestore(&rq->lock, flags);
6697 /* Register at highest priority so that task migration (migrate_all_tasks)
6698 * happens before everything else.
6700 static struct notifier_block __cpuinitdata migration_notifier = {
6701 .notifier_call = migration_call,
6705 static int __init migration_init(void)
6707 void *cpu = (void *)(long)smp_processor_id();
6710 /* Start one for the boot CPU: */
6711 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6712 BUG_ON(err == NOTIFY_BAD);
6713 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6714 register_cpu_notifier(&migration_notifier);
6718 early_initcall(migration_init);
6723 #ifdef CONFIG_SCHED_DEBUG
6725 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6726 struct cpumask *groupmask)
6728 struct sched_group *group = sd->groups;
6731 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6732 cpumask_clear(groupmask);
6734 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6736 if (!(sd->flags & SD_LOAD_BALANCE)) {
6737 printk("does not load-balance\n");
6739 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6744 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6746 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6747 printk(KERN_ERR "ERROR: domain->span does not contain "
6750 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6751 printk(KERN_ERR "ERROR: domain->groups does not contain"
6755 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6759 printk(KERN_ERR "ERROR: group is NULL\n");
6763 if (!group->__cpu_power) {
6764 printk(KERN_CONT "\n");
6765 printk(KERN_ERR "ERROR: domain->cpu_power not "
6770 if (!cpumask_weight(sched_group_cpus(group))) {
6771 printk(KERN_CONT "\n");
6772 printk(KERN_ERR "ERROR: empty group\n");
6776 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6777 printk(KERN_CONT "\n");
6778 printk(KERN_ERR "ERROR: repeated CPUs\n");
6782 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6784 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6785 printk(KERN_CONT " %s", str);
6787 group = group->next;
6788 } while (group != sd->groups);
6789 printk(KERN_CONT "\n");
6791 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6792 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6795 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6796 printk(KERN_ERR "ERROR: parent span is not a superset "
6797 "of domain->span\n");
6801 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6803 cpumask_var_t groupmask;
6807 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6811 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6813 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6814 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6819 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6826 free_cpumask_var(groupmask);
6828 #else /* !CONFIG_SCHED_DEBUG */
6829 # define sched_domain_debug(sd, cpu) do { } while (0)
6830 #endif /* CONFIG_SCHED_DEBUG */
6832 static int sd_degenerate(struct sched_domain *sd)
6834 if (cpumask_weight(sched_domain_span(sd)) == 1)
6837 /* Following flags need at least 2 groups */
6838 if (sd->flags & (SD_LOAD_BALANCE |
6839 SD_BALANCE_NEWIDLE |
6843 SD_SHARE_PKG_RESOURCES)) {
6844 if (sd->groups != sd->groups->next)
6848 /* Following flags don't use groups */
6849 if (sd->flags & (SD_WAKE_IDLE |
6858 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6860 unsigned long cflags = sd->flags, pflags = parent->flags;
6862 if (sd_degenerate(parent))
6865 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6868 /* Does parent contain flags not in child? */
6869 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6870 if (cflags & SD_WAKE_AFFINE)
6871 pflags &= ~SD_WAKE_BALANCE;
6872 /* Flags needing groups don't count if only 1 group in parent */
6873 if (parent->groups == parent->groups->next) {
6874 pflags &= ~(SD_LOAD_BALANCE |
6875 SD_BALANCE_NEWIDLE |
6879 SD_SHARE_PKG_RESOURCES);
6880 if (nr_node_ids == 1)
6881 pflags &= ~SD_SERIALIZE;
6883 if (~cflags & pflags)
6889 static void free_rootdomain(struct root_domain *rd)
6891 cpupri_cleanup(&rd->cpupri);
6893 free_cpumask_var(rd->rto_mask);
6894 free_cpumask_var(rd->online);
6895 free_cpumask_var(rd->span);
6899 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6901 unsigned long flags;
6903 spin_lock_irqsave(&rq->lock, flags);
6906 struct root_domain *old_rd = rq->rd;
6908 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6911 cpumask_clear_cpu(rq->cpu, old_rd->span);
6913 if (atomic_dec_and_test(&old_rd->refcount))
6914 free_rootdomain(old_rd);
6917 atomic_inc(&rd->refcount);
6920 cpumask_set_cpu(rq->cpu, rd->span);
6921 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
6924 spin_unlock_irqrestore(&rq->lock, flags);
6927 static int init_rootdomain(struct root_domain *rd, bool bootmem)
6929 memset(rd, 0, sizeof(*rd));
6932 alloc_bootmem_cpumask_var(&def_root_domain.span);
6933 alloc_bootmem_cpumask_var(&def_root_domain.online);
6934 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
6935 cpupri_init(&rd->cpupri, true);
6939 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6941 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6943 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6946 if (cpupri_init(&rd->cpupri, false) != 0)
6951 free_cpumask_var(rd->rto_mask);
6953 free_cpumask_var(rd->online);
6955 free_cpumask_var(rd->span);
6961 static void init_defrootdomain(void)
6963 init_rootdomain(&def_root_domain, true);
6965 atomic_set(&def_root_domain.refcount, 1);
6968 static struct root_domain *alloc_rootdomain(void)
6970 struct root_domain *rd;
6972 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6976 if (init_rootdomain(rd, false) != 0) {
6985 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6986 * hold the hotplug lock.
6989 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6991 struct rq *rq = cpu_rq(cpu);
6992 struct sched_domain *tmp;
6994 /* Remove the sched domains which do not contribute to scheduling. */
6995 for (tmp = sd; tmp; ) {
6996 struct sched_domain *parent = tmp->parent;
7000 if (sd_parent_degenerate(tmp, parent)) {
7001 tmp->parent = parent->parent;
7003 parent->parent->child = tmp;
7008 if (sd && sd_degenerate(sd)) {
7014 sched_domain_debug(sd, cpu);
7016 rq_attach_root(rq, rd);
7017 rcu_assign_pointer(rq->sd, sd);
7020 /* cpus with isolated domains */
7021 static cpumask_var_t cpu_isolated_map;
7023 /* Setup the mask of cpus configured for isolated domains */
7024 static int __init isolated_cpu_setup(char *str)
7026 cpulist_parse(str, cpu_isolated_map);
7030 __setup("isolcpus=", isolated_cpu_setup);
7033 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7034 * to a function which identifies what group(along with sched group) a CPU
7035 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7036 * (due to the fact that we keep track of groups covered with a struct cpumask).
7038 * init_sched_build_groups will build a circular linked list of the groups
7039 * covered by the given span, and will set each group's ->cpumask correctly,
7040 * and ->cpu_power to 0.
7043 init_sched_build_groups(const struct cpumask *span,
7044 const struct cpumask *cpu_map,
7045 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7046 struct sched_group **sg,
7047 struct cpumask *tmpmask),
7048 struct cpumask *covered, struct cpumask *tmpmask)
7050 struct sched_group *first = NULL, *last = NULL;
7053 cpumask_clear(covered);
7055 for_each_cpu(i, span) {
7056 struct sched_group *sg;
7057 int group = group_fn(i, cpu_map, &sg, tmpmask);
7060 if (cpumask_test_cpu(i, covered))
7063 cpumask_clear(sched_group_cpus(sg));
7064 sg->__cpu_power = 0;
7066 for_each_cpu(j, span) {
7067 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7070 cpumask_set_cpu(j, covered);
7071 cpumask_set_cpu(j, sched_group_cpus(sg));
7082 #define SD_NODES_PER_DOMAIN 16
7087 * find_next_best_node - find the next node to include in a sched_domain
7088 * @node: node whose sched_domain we're building
7089 * @used_nodes: nodes already in the sched_domain
7091 * Find the next node to include in a given scheduling domain. Simply
7092 * finds the closest node not already in the @used_nodes map.
7094 * Should use nodemask_t.
7096 static int find_next_best_node(int node, nodemask_t *used_nodes)
7098 int i, n, val, min_val, best_node = 0;
7102 for (i = 0; i < nr_node_ids; i++) {
7103 /* Start at @node */
7104 n = (node + i) % nr_node_ids;
7106 if (!nr_cpus_node(n))
7109 /* Skip already used nodes */
7110 if (node_isset(n, *used_nodes))
7113 /* Simple min distance search */
7114 val = node_distance(node, n);
7116 if (val < min_val) {
7122 node_set(best_node, *used_nodes);
7127 * sched_domain_node_span - get a cpumask for a node's sched_domain
7128 * @node: node whose cpumask we're constructing
7129 * @span: resulting cpumask
7131 * Given a node, construct a good cpumask for its sched_domain to span. It
7132 * should be one that prevents unnecessary balancing, but also spreads tasks
7135 static void sched_domain_node_span(int node, struct cpumask *span)
7137 nodemask_t used_nodes;
7138 /* FIXME: use cpumask_of_node() */
7139 node_to_cpumask_ptr(nodemask, node);
7143 nodes_clear(used_nodes);
7145 cpus_or(*span, *span, *nodemask);
7146 node_set(node, used_nodes);
7148 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7149 int next_node = find_next_best_node(node, &used_nodes);
7151 node_to_cpumask_ptr_next(nodemask, next_node);
7152 cpus_or(*span, *span, *nodemask);
7155 #endif /* CONFIG_NUMA */
7157 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7160 * The cpus mask in sched_group and sched_domain hangs off the end.
7161 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7162 * for nr_cpu_ids < CONFIG_NR_CPUS.
7164 struct static_sched_group {
7165 struct sched_group sg;
7166 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7169 struct static_sched_domain {
7170 struct sched_domain sd;
7171 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7175 * SMT sched-domains:
7177 #ifdef CONFIG_SCHED_SMT
7178 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7179 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7182 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7183 struct sched_group **sg, struct cpumask *unused)
7186 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7189 #endif /* CONFIG_SCHED_SMT */
7192 * multi-core sched-domains:
7194 #ifdef CONFIG_SCHED_MC
7195 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7196 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7197 #endif /* CONFIG_SCHED_MC */
7199 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7201 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7202 struct sched_group **sg, struct cpumask *mask)
7206 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7207 group = cpumask_first(mask);
7209 *sg = &per_cpu(sched_group_core, group).sg;
7212 #elif defined(CONFIG_SCHED_MC)
7214 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7215 struct sched_group **sg, struct cpumask *unused)
7218 *sg = &per_cpu(sched_group_core, cpu).sg;
7223 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7224 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7227 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7228 struct sched_group **sg, struct cpumask *mask)
7231 #ifdef CONFIG_SCHED_MC
7232 /* FIXME: Use cpu_coregroup_mask. */
7233 *mask = cpu_coregroup_map(cpu);
7234 cpus_and(*mask, *mask, *cpu_map);
7235 group = cpumask_first(mask);
7236 #elif defined(CONFIG_SCHED_SMT)
7237 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7238 group = cpumask_first(mask);
7243 *sg = &per_cpu(sched_group_phys, group).sg;
7249 * The init_sched_build_groups can't handle what we want to do with node
7250 * groups, so roll our own. Now each node has its own list of groups which
7251 * gets dynamically allocated.
7253 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7254 static struct sched_group ***sched_group_nodes_bycpu;
7256 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7257 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7259 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7260 struct sched_group **sg,
7261 struct cpumask *nodemask)
7264 /* FIXME: use cpumask_of_node */
7265 node_to_cpumask_ptr(pnodemask, cpu_to_node(cpu));
7267 cpumask_and(nodemask, pnodemask, cpu_map);
7268 group = cpumask_first(nodemask);
7271 *sg = &per_cpu(sched_group_allnodes, group).sg;
7275 static void init_numa_sched_groups_power(struct sched_group *group_head)
7277 struct sched_group *sg = group_head;
7283 for_each_cpu(j, sched_group_cpus(sg)) {
7284 struct sched_domain *sd;
7286 sd = &per_cpu(phys_domains, j).sd;
7287 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7289 * Only add "power" once for each
7295 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7298 } while (sg != group_head);
7300 #endif /* CONFIG_NUMA */
7303 /* Free memory allocated for various sched_group structures */
7304 static void free_sched_groups(const struct cpumask *cpu_map,
7305 struct cpumask *nodemask)
7309 for_each_cpu(cpu, cpu_map) {
7310 struct sched_group **sched_group_nodes
7311 = sched_group_nodes_bycpu[cpu];
7313 if (!sched_group_nodes)
7316 for (i = 0; i < nr_node_ids; i++) {
7317 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7318 /* FIXME: Use cpumask_of_node */
7319 node_to_cpumask_ptr(pnodemask, i);
7321 cpus_and(*nodemask, *pnodemask, *cpu_map);
7322 if (cpumask_empty(nodemask))
7332 if (oldsg != sched_group_nodes[i])
7335 kfree(sched_group_nodes);
7336 sched_group_nodes_bycpu[cpu] = NULL;
7339 #else /* !CONFIG_NUMA */
7340 static void free_sched_groups(const struct cpumask *cpu_map,
7341 struct cpumask *nodemask)
7344 #endif /* CONFIG_NUMA */
7347 * Initialize sched groups cpu_power.
7349 * cpu_power indicates the capacity of sched group, which is used while
7350 * distributing the load between different sched groups in a sched domain.
7351 * Typically cpu_power for all the groups in a sched domain will be same unless
7352 * there are asymmetries in the topology. If there are asymmetries, group
7353 * having more cpu_power will pickup more load compared to the group having
7356 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7357 * the maximum number of tasks a group can handle in the presence of other idle
7358 * or lightly loaded groups in the same sched domain.
7360 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7362 struct sched_domain *child;
7363 struct sched_group *group;
7365 WARN_ON(!sd || !sd->groups);
7367 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7372 sd->groups->__cpu_power = 0;
7375 * For perf policy, if the groups in child domain share resources
7376 * (for example cores sharing some portions of the cache hierarchy
7377 * or SMT), then set this domain groups cpu_power such that each group
7378 * can handle only one task, when there are other idle groups in the
7379 * same sched domain.
7381 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7383 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7384 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7389 * add cpu_power of each child group to this groups cpu_power
7391 group = child->groups;
7393 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7394 group = group->next;
7395 } while (group != child->groups);
7399 * Initializers for schedule domains
7400 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7403 #ifdef CONFIG_SCHED_DEBUG
7404 # define SD_INIT_NAME(sd, type) sd->name = #type
7406 # define SD_INIT_NAME(sd, type) do { } while (0)
7409 #define SD_INIT(sd, type) sd_init_##type(sd)
7411 #define SD_INIT_FUNC(type) \
7412 static noinline void sd_init_##type(struct sched_domain *sd) \
7414 memset(sd, 0, sizeof(*sd)); \
7415 *sd = SD_##type##_INIT; \
7416 sd->level = SD_LV_##type; \
7417 SD_INIT_NAME(sd, type); \
7422 SD_INIT_FUNC(ALLNODES)
7425 #ifdef CONFIG_SCHED_SMT
7426 SD_INIT_FUNC(SIBLING)
7428 #ifdef CONFIG_SCHED_MC
7432 static int default_relax_domain_level = -1;
7434 static int __init setup_relax_domain_level(char *str)
7438 val = simple_strtoul(str, NULL, 0);
7439 if (val < SD_LV_MAX)
7440 default_relax_domain_level = val;
7444 __setup("relax_domain_level=", setup_relax_domain_level);
7446 static void set_domain_attribute(struct sched_domain *sd,
7447 struct sched_domain_attr *attr)
7451 if (!attr || attr->relax_domain_level < 0) {
7452 if (default_relax_domain_level < 0)
7455 request = default_relax_domain_level;
7457 request = attr->relax_domain_level;
7458 if (request < sd->level) {
7459 /* turn off idle balance on this domain */
7460 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7462 /* turn on idle balance on this domain */
7463 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7468 * Build sched domains for a given set of cpus and attach the sched domains
7469 * to the individual cpus
7471 static int __build_sched_domains(const struct cpumask *cpu_map,
7472 struct sched_domain_attr *attr)
7474 int i, err = -ENOMEM;
7475 struct root_domain *rd;
7476 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7479 cpumask_var_t domainspan, covered, notcovered;
7480 struct sched_group **sched_group_nodes = NULL;
7481 int sd_allnodes = 0;
7483 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7485 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7486 goto free_domainspan;
7487 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7491 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7492 goto free_notcovered;
7493 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7495 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7496 goto free_this_sibling_map;
7497 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7498 goto free_this_core_map;
7499 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7500 goto free_send_covered;
7504 * Allocate the per-node list of sched groups
7506 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7508 if (!sched_group_nodes) {
7509 printk(KERN_WARNING "Can not alloc sched group node list\n");
7514 rd = alloc_rootdomain();
7516 printk(KERN_WARNING "Cannot alloc root domain\n");
7517 goto free_sched_groups;
7521 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7525 * Set up domains for cpus specified by the cpu_map.
7527 for_each_cpu(i, cpu_map) {
7528 struct sched_domain *sd = NULL, *p;
7530 /* FIXME: use cpumask_of_node */
7531 *nodemask = node_to_cpumask(cpu_to_node(i));
7532 cpus_and(*nodemask, *nodemask, *cpu_map);
7535 if (cpumask_weight(cpu_map) >
7536 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7537 sd = &per_cpu(allnodes_domains, i);
7538 SD_INIT(sd, ALLNODES);
7539 set_domain_attribute(sd, attr);
7540 cpumask_copy(sched_domain_span(sd), cpu_map);
7541 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7547 sd = &per_cpu(node_domains, i);
7549 set_domain_attribute(sd, attr);
7550 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7554 cpumask_and(sched_domain_span(sd),
7555 sched_domain_span(sd), cpu_map);
7559 sd = &per_cpu(phys_domains, i).sd;
7561 set_domain_attribute(sd, attr);
7562 cpumask_copy(sched_domain_span(sd), nodemask);
7566 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7568 #ifdef CONFIG_SCHED_MC
7570 sd = &per_cpu(core_domains, i).sd;
7572 set_domain_attribute(sd, attr);
7573 *sched_domain_span(sd) = cpu_coregroup_map(i);
7574 cpumask_and(sched_domain_span(sd),
7575 sched_domain_span(sd), cpu_map);
7578 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7581 #ifdef CONFIG_SCHED_SMT
7583 sd = &per_cpu(cpu_domains, i).sd;
7584 SD_INIT(sd, SIBLING);
7585 set_domain_attribute(sd, attr);
7586 cpumask_and(sched_domain_span(sd),
7587 &per_cpu(cpu_sibling_map, i), cpu_map);
7590 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7594 #ifdef CONFIG_SCHED_SMT
7595 /* Set up CPU (sibling) groups */
7596 for_each_cpu(i, cpu_map) {
7597 cpumask_and(this_sibling_map,
7598 &per_cpu(cpu_sibling_map, i), cpu_map);
7599 if (i != cpumask_first(this_sibling_map))
7602 init_sched_build_groups(this_sibling_map, cpu_map,
7604 send_covered, tmpmask);
7608 #ifdef CONFIG_SCHED_MC
7609 /* Set up multi-core groups */
7610 for_each_cpu(i, cpu_map) {
7611 /* FIXME: Use cpu_coregroup_mask */
7612 *this_core_map = cpu_coregroup_map(i);
7613 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7614 if (i != cpumask_first(this_core_map))
7617 init_sched_build_groups(this_core_map, cpu_map,
7619 send_covered, tmpmask);
7623 /* Set up physical groups */
7624 for (i = 0; i < nr_node_ids; i++) {
7625 /* FIXME: Use cpumask_of_node */
7626 *nodemask = node_to_cpumask(i);
7627 cpus_and(*nodemask, *nodemask, *cpu_map);
7628 if (cpumask_empty(nodemask))
7631 init_sched_build_groups(nodemask, cpu_map,
7633 send_covered, tmpmask);
7637 /* Set up node groups */
7639 init_sched_build_groups(cpu_map, cpu_map,
7640 &cpu_to_allnodes_group,
7641 send_covered, tmpmask);
7644 for (i = 0; i < nr_node_ids; i++) {
7645 /* Set up node groups */
7646 struct sched_group *sg, *prev;
7649 /* FIXME: Use cpumask_of_node */
7650 *nodemask = node_to_cpumask(i);
7651 cpumask_clear(covered);
7653 cpus_and(*nodemask, *nodemask, *cpu_map);
7654 if (cpumask_empty(nodemask)) {
7655 sched_group_nodes[i] = NULL;
7659 sched_domain_node_span(i, domainspan);
7660 cpumask_and(domainspan, domainspan, cpu_map);
7662 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7665 printk(KERN_WARNING "Can not alloc domain group for "
7669 sched_group_nodes[i] = sg;
7670 for_each_cpu(j, nodemask) {
7671 struct sched_domain *sd;
7673 sd = &per_cpu(node_domains, j);
7676 sg->__cpu_power = 0;
7677 cpumask_copy(sched_group_cpus(sg), nodemask);
7679 cpumask_or(covered, covered, nodemask);
7682 for (j = 0; j < nr_node_ids; j++) {
7683 int n = (i + j) % nr_node_ids;
7684 /* FIXME: Use cpumask_of_node */
7685 node_to_cpumask_ptr(pnodemask, n);
7687 cpumask_complement(notcovered, covered);
7688 cpumask_and(tmpmask, notcovered, cpu_map);
7689 cpumask_and(tmpmask, tmpmask, domainspan);
7690 if (cpumask_empty(tmpmask))
7693 cpumask_and(tmpmask, tmpmask, pnodemask);
7694 if (cpumask_empty(tmpmask))
7697 sg = kmalloc_node(sizeof(struct sched_group) +
7702 "Can not alloc domain group for node %d\n", j);
7705 sg->__cpu_power = 0;
7706 cpumask_copy(sched_group_cpus(sg), tmpmask);
7707 sg->next = prev->next;
7708 cpumask_or(covered, covered, tmpmask);
7715 /* Calculate CPU power for physical packages and nodes */
7716 #ifdef CONFIG_SCHED_SMT
7717 for_each_cpu(i, cpu_map) {
7718 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7720 init_sched_groups_power(i, sd);
7723 #ifdef CONFIG_SCHED_MC
7724 for_each_cpu(i, cpu_map) {
7725 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7727 init_sched_groups_power(i, sd);
7731 for_each_cpu(i, cpu_map) {
7732 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7734 init_sched_groups_power(i, sd);
7738 for (i = 0; i < nr_node_ids; i++)
7739 init_numa_sched_groups_power(sched_group_nodes[i]);
7742 struct sched_group *sg;
7744 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7746 init_numa_sched_groups_power(sg);
7750 /* Attach the domains */
7751 for_each_cpu(i, cpu_map) {
7752 struct sched_domain *sd;
7753 #ifdef CONFIG_SCHED_SMT
7754 sd = &per_cpu(cpu_domains, i).sd;
7755 #elif defined(CONFIG_SCHED_MC)
7756 sd = &per_cpu(core_domains, i).sd;
7758 sd = &per_cpu(phys_domains, i).sd;
7760 cpu_attach_domain(sd, rd, i);
7766 free_cpumask_var(tmpmask);
7768 free_cpumask_var(send_covered);
7770 free_cpumask_var(this_core_map);
7771 free_this_sibling_map:
7772 free_cpumask_var(this_sibling_map);
7774 free_cpumask_var(nodemask);
7777 free_cpumask_var(notcovered);
7779 free_cpumask_var(covered);
7781 free_cpumask_var(domainspan);
7788 kfree(sched_group_nodes);
7794 free_sched_groups(cpu_map, tmpmask);
7795 free_rootdomain(rd);
7800 static int build_sched_domains(const struct cpumask *cpu_map)
7802 return __build_sched_domains(cpu_map, NULL);
7805 static struct cpumask *doms_cur; /* current sched domains */
7806 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7807 static struct sched_domain_attr *dattr_cur;
7808 /* attribues of custom domains in 'doms_cur' */
7811 * Special case: If a kmalloc of a doms_cur partition (array of
7812 * cpumask) fails, then fallback to a single sched domain,
7813 * as determined by the single cpumask fallback_doms.
7815 static cpumask_var_t fallback_doms;
7818 * arch_update_cpu_topology lets virtualized architectures update the
7819 * cpu core maps. It is supposed to return 1 if the topology changed
7820 * or 0 if it stayed the same.
7822 int __attribute__((weak)) arch_update_cpu_topology(void)
7828 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7829 * For now this just excludes isolated cpus, but could be used to
7830 * exclude other special cases in the future.
7832 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7836 arch_update_cpu_topology();
7838 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7840 doms_cur = fallback_doms;
7841 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7843 err = build_sched_domains(doms_cur);
7844 register_sched_domain_sysctl();
7849 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7850 struct cpumask *tmpmask)
7852 free_sched_groups(cpu_map, tmpmask);
7856 * Detach sched domains from a group of cpus specified in cpu_map
7857 * These cpus will now be attached to the NULL domain
7859 static void detach_destroy_domains(const struct cpumask *cpu_map)
7861 /* Save because hotplug lock held. */
7862 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7865 for_each_cpu(i, cpu_map)
7866 cpu_attach_domain(NULL, &def_root_domain, i);
7867 synchronize_sched();
7868 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7871 /* handle null as "default" */
7872 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7873 struct sched_domain_attr *new, int idx_new)
7875 struct sched_domain_attr tmp;
7882 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7883 new ? (new + idx_new) : &tmp,
7884 sizeof(struct sched_domain_attr));
7888 * Partition sched domains as specified by the 'ndoms_new'
7889 * cpumasks in the array doms_new[] of cpumasks. This compares
7890 * doms_new[] to the current sched domain partitioning, doms_cur[].
7891 * It destroys each deleted domain and builds each new domain.
7893 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7894 * The masks don't intersect (don't overlap.) We should setup one
7895 * sched domain for each mask. CPUs not in any of the cpumasks will
7896 * not be load balanced. If the same cpumask appears both in the
7897 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7900 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7901 * ownership of it and will kfree it when done with it. If the caller
7902 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7903 * ndoms_new == 1, and partition_sched_domains() will fallback to
7904 * the single partition 'fallback_doms', it also forces the domains
7907 * If doms_new == NULL it will be replaced with cpu_online_mask.
7908 * ndoms_new == 0 is a special case for destroying existing domains,
7909 * and it will not create the default domain.
7911 * Call with hotplug lock held
7913 /* FIXME: Change to struct cpumask *doms_new[] */
7914 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
7915 struct sched_domain_attr *dattr_new)
7920 mutex_lock(&sched_domains_mutex);
7922 /* always unregister in case we don't destroy any domains */
7923 unregister_sched_domain_sysctl();
7925 /* Let architecture update cpu core mappings. */
7926 new_topology = arch_update_cpu_topology();
7928 n = doms_new ? ndoms_new : 0;
7930 /* Destroy deleted domains */
7931 for (i = 0; i < ndoms_cur; i++) {
7932 for (j = 0; j < n && !new_topology; j++) {
7933 if (cpumask_equal(&doms_cur[i], &doms_new[j])
7934 && dattrs_equal(dattr_cur, i, dattr_new, j))
7937 /* no match - a current sched domain not in new doms_new[] */
7938 detach_destroy_domains(doms_cur + i);
7943 if (doms_new == NULL) {
7945 doms_new = fallback_doms;
7946 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
7947 WARN_ON_ONCE(dattr_new);
7950 /* Build new domains */
7951 for (i = 0; i < ndoms_new; i++) {
7952 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7953 if (cpumask_equal(&doms_new[i], &doms_cur[j])
7954 && dattrs_equal(dattr_new, i, dattr_cur, j))
7957 /* no match - add a new doms_new */
7958 __build_sched_domains(doms_new + i,
7959 dattr_new ? dattr_new + i : NULL);
7964 /* Remember the new sched domains */
7965 if (doms_cur != fallback_doms)
7967 kfree(dattr_cur); /* kfree(NULL) is safe */
7968 doms_cur = doms_new;
7969 dattr_cur = dattr_new;
7970 ndoms_cur = ndoms_new;
7972 register_sched_domain_sysctl();
7974 mutex_unlock(&sched_domains_mutex);
7977 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7978 int arch_reinit_sched_domains(void)
7982 /* Destroy domains first to force the rebuild */
7983 partition_sched_domains(0, NULL, NULL);
7985 rebuild_sched_domains();
7991 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7994 unsigned int level = 0;
7996 if (sscanf(buf, "%u", &level) != 1)
8000 * level is always be positive so don't check for
8001 * level < POWERSAVINGS_BALANCE_NONE which is 0
8002 * What happens on 0 or 1 byte write,
8003 * need to check for count as well?
8006 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8010 sched_smt_power_savings = level;
8012 sched_mc_power_savings = level;
8014 ret = arch_reinit_sched_domains();
8016 return ret ? ret : count;
8019 #ifdef CONFIG_SCHED_MC
8020 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8023 return sprintf(page, "%u\n", sched_mc_power_savings);
8025 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8026 const char *buf, size_t count)
8028 return sched_power_savings_store(buf, count, 0);
8030 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8031 sched_mc_power_savings_show,
8032 sched_mc_power_savings_store);
8035 #ifdef CONFIG_SCHED_SMT
8036 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8039 return sprintf(page, "%u\n", sched_smt_power_savings);
8041 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8042 const char *buf, size_t count)
8044 return sched_power_savings_store(buf, count, 1);
8046 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8047 sched_smt_power_savings_show,
8048 sched_smt_power_savings_store);
8051 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8055 #ifdef CONFIG_SCHED_SMT
8057 err = sysfs_create_file(&cls->kset.kobj,
8058 &attr_sched_smt_power_savings.attr);
8060 #ifdef CONFIG_SCHED_MC
8061 if (!err && mc_capable())
8062 err = sysfs_create_file(&cls->kset.kobj,
8063 &attr_sched_mc_power_savings.attr);
8067 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8069 #ifndef CONFIG_CPUSETS
8071 * Add online and remove offline CPUs from the scheduler domains.
8072 * When cpusets are enabled they take over this function.
8074 static int update_sched_domains(struct notifier_block *nfb,
8075 unsigned long action, void *hcpu)
8079 case CPU_ONLINE_FROZEN:
8081 case CPU_DEAD_FROZEN:
8082 partition_sched_domains(1, NULL, NULL);
8091 static int update_runtime(struct notifier_block *nfb,
8092 unsigned long action, void *hcpu)
8094 int cpu = (int)(long)hcpu;
8097 case CPU_DOWN_PREPARE:
8098 case CPU_DOWN_PREPARE_FROZEN:
8099 disable_runtime(cpu_rq(cpu));
8102 case CPU_DOWN_FAILED:
8103 case CPU_DOWN_FAILED_FROZEN:
8105 case CPU_ONLINE_FROZEN:
8106 enable_runtime(cpu_rq(cpu));
8114 void __init sched_init_smp(void)
8116 cpumask_var_t non_isolated_cpus;
8118 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8120 #if defined(CONFIG_NUMA)
8121 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8123 BUG_ON(sched_group_nodes_bycpu == NULL);
8126 mutex_lock(&sched_domains_mutex);
8127 arch_init_sched_domains(cpu_online_mask);
8128 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8129 if (cpumask_empty(non_isolated_cpus))
8130 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8131 mutex_unlock(&sched_domains_mutex);
8134 #ifndef CONFIG_CPUSETS
8135 /* XXX: Theoretical race here - CPU may be hotplugged now */
8136 hotcpu_notifier(update_sched_domains, 0);
8139 /* RT runtime code needs to handle some hotplug events */
8140 hotcpu_notifier(update_runtime, 0);
8144 /* Move init over to a non-isolated CPU */
8145 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8147 sched_init_granularity();
8148 free_cpumask_var(non_isolated_cpus);
8150 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8151 init_sched_rt_class();
8154 void __init sched_init_smp(void)
8156 sched_init_granularity();
8158 #endif /* CONFIG_SMP */
8160 int in_sched_functions(unsigned long addr)
8162 return in_lock_functions(addr) ||
8163 (addr >= (unsigned long)__sched_text_start
8164 && addr < (unsigned long)__sched_text_end);
8167 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8169 cfs_rq->tasks_timeline = RB_ROOT;
8170 INIT_LIST_HEAD(&cfs_rq->tasks);
8171 #ifdef CONFIG_FAIR_GROUP_SCHED
8174 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8177 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8179 struct rt_prio_array *array;
8182 array = &rt_rq->active;
8183 for (i = 0; i < MAX_RT_PRIO; i++) {
8184 INIT_LIST_HEAD(array->queue + i);
8185 __clear_bit(i, array->bitmap);
8187 /* delimiter for bitsearch: */
8188 __set_bit(MAX_RT_PRIO, array->bitmap);
8190 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8191 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8192 rt_rq->highest_prio.next = MAX_RT_PRIO;
8195 rt_rq->rt_nr_migratory = 0;
8196 rt_rq->overloaded = 0;
8200 rt_rq->rt_throttled = 0;
8201 rt_rq->rt_runtime = 0;
8202 spin_lock_init(&rt_rq->rt_runtime_lock);
8204 #ifdef CONFIG_RT_GROUP_SCHED
8205 rt_rq->rt_nr_boosted = 0;
8210 #ifdef CONFIG_FAIR_GROUP_SCHED
8211 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8212 struct sched_entity *se, int cpu, int add,
8213 struct sched_entity *parent)
8215 struct rq *rq = cpu_rq(cpu);
8216 tg->cfs_rq[cpu] = cfs_rq;
8217 init_cfs_rq(cfs_rq, rq);
8220 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8223 /* se could be NULL for init_task_group */
8228 se->cfs_rq = &rq->cfs;
8230 se->cfs_rq = parent->my_q;
8233 se->load.weight = tg->shares;
8234 se->load.inv_weight = 0;
8235 se->parent = parent;
8239 #ifdef CONFIG_RT_GROUP_SCHED
8240 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8241 struct sched_rt_entity *rt_se, int cpu, int add,
8242 struct sched_rt_entity *parent)
8244 struct rq *rq = cpu_rq(cpu);
8246 tg->rt_rq[cpu] = rt_rq;
8247 init_rt_rq(rt_rq, rq);
8249 rt_rq->rt_se = rt_se;
8250 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8252 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8254 tg->rt_se[cpu] = rt_se;
8259 rt_se->rt_rq = &rq->rt;
8261 rt_se->rt_rq = parent->my_q;
8263 rt_se->my_q = rt_rq;
8264 rt_se->parent = parent;
8265 INIT_LIST_HEAD(&rt_se->run_list);
8269 void __init sched_init(void)
8272 unsigned long alloc_size = 0, ptr;
8274 #ifdef CONFIG_FAIR_GROUP_SCHED
8275 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8277 #ifdef CONFIG_RT_GROUP_SCHED
8278 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8280 #ifdef CONFIG_USER_SCHED
8284 * As sched_init() is called before page_alloc is setup,
8285 * we use alloc_bootmem().
8288 ptr = (unsigned long)alloc_bootmem(alloc_size);
8290 #ifdef CONFIG_FAIR_GROUP_SCHED
8291 init_task_group.se = (struct sched_entity **)ptr;
8292 ptr += nr_cpu_ids * sizeof(void **);
8294 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8295 ptr += nr_cpu_ids * sizeof(void **);
8297 #ifdef CONFIG_USER_SCHED
8298 root_task_group.se = (struct sched_entity **)ptr;
8299 ptr += nr_cpu_ids * sizeof(void **);
8301 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8302 ptr += nr_cpu_ids * sizeof(void **);
8303 #endif /* CONFIG_USER_SCHED */
8304 #endif /* CONFIG_FAIR_GROUP_SCHED */
8305 #ifdef CONFIG_RT_GROUP_SCHED
8306 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8307 ptr += nr_cpu_ids * sizeof(void **);
8309 init_task_group.rt_rq = (struct rt_rq **)ptr;
8310 ptr += nr_cpu_ids * sizeof(void **);
8312 #ifdef CONFIG_USER_SCHED
8313 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8314 ptr += nr_cpu_ids * sizeof(void **);
8316 root_task_group.rt_rq = (struct rt_rq **)ptr;
8317 ptr += nr_cpu_ids * sizeof(void **);
8318 #endif /* CONFIG_USER_SCHED */
8319 #endif /* CONFIG_RT_GROUP_SCHED */
8323 init_defrootdomain();
8326 init_rt_bandwidth(&def_rt_bandwidth,
8327 global_rt_period(), global_rt_runtime());
8329 #ifdef CONFIG_RT_GROUP_SCHED
8330 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8331 global_rt_period(), global_rt_runtime());
8332 #ifdef CONFIG_USER_SCHED
8333 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8334 global_rt_period(), RUNTIME_INF);
8335 #endif /* CONFIG_USER_SCHED */
8336 #endif /* CONFIG_RT_GROUP_SCHED */
8338 #ifdef CONFIG_GROUP_SCHED
8339 list_add(&init_task_group.list, &task_groups);
8340 INIT_LIST_HEAD(&init_task_group.children);
8342 #ifdef CONFIG_USER_SCHED
8343 INIT_LIST_HEAD(&root_task_group.children);
8344 init_task_group.parent = &root_task_group;
8345 list_add(&init_task_group.siblings, &root_task_group.children);
8346 #endif /* CONFIG_USER_SCHED */
8347 #endif /* CONFIG_GROUP_SCHED */
8349 for_each_possible_cpu(i) {
8353 spin_lock_init(&rq->lock);
8355 init_cfs_rq(&rq->cfs, rq);
8356 init_rt_rq(&rq->rt, rq);
8357 #ifdef CONFIG_FAIR_GROUP_SCHED
8358 init_task_group.shares = init_task_group_load;
8359 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8360 #ifdef CONFIG_CGROUP_SCHED
8362 * How much cpu bandwidth does init_task_group get?
8364 * In case of task-groups formed thr' the cgroup filesystem, it
8365 * gets 100% of the cpu resources in the system. This overall
8366 * system cpu resource is divided among the tasks of
8367 * init_task_group and its child task-groups in a fair manner,
8368 * based on each entity's (task or task-group's) weight
8369 * (se->load.weight).
8371 * In other words, if init_task_group has 10 tasks of weight
8372 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8373 * then A0's share of the cpu resource is:
8375 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8377 * We achieve this by letting init_task_group's tasks sit
8378 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8380 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8381 #elif defined CONFIG_USER_SCHED
8382 root_task_group.shares = NICE_0_LOAD;
8383 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8385 * In case of task-groups formed thr' the user id of tasks,
8386 * init_task_group represents tasks belonging to root user.
8387 * Hence it forms a sibling of all subsequent groups formed.
8388 * In this case, init_task_group gets only a fraction of overall
8389 * system cpu resource, based on the weight assigned to root
8390 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8391 * by letting tasks of init_task_group sit in a separate cfs_rq
8392 * (init_cfs_rq) and having one entity represent this group of
8393 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8395 init_tg_cfs_entry(&init_task_group,
8396 &per_cpu(init_cfs_rq, i),
8397 &per_cpu(init_sched_entity, i), i, 1,
8398 root_task_group.se[i]);
8401 #endif /* CONFIG_FAIR_GROUP_SCHED */
8403 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8404 #ifdef CONFIG_RT_GROUP_SCHED
8405 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8406 #ifdef CONFIG_CGROUP_SCHED
8407 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8408 #elif defined CONFIG_USER_SCHED
8409 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8410 init_tg_rt_entry(&init_task_group,
8411 &per_cpu(init_rt_rq, i),
8412 &per_cpu(init_sched_rt_entity, i), i, 1,
8413 root_task_group.rt_se[i]);
8417 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8418 rq->cpu_load[j] = 0;
8422 rq->active_balance = 0;
8423 rq->next_balance = jiffies;
8427 rq->migration_thread = NULL;
8428 INIT_LIST_HEAD(&rq->migration_queue);
8429 rq_attach_root(rq, &def_root_domain);
8432 atomic_set(&rq->nr_iowait, 0);
8435 set_load_weight(&init_task);
8437 #ifdef CONFIG_PREEMPT_NOTIFIERS
8438 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8442 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8445 #ifdef CONFIG_RT_MUTEXES
8446 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8450 * The boot idle thread does lazy MMU switching as well:
8452 atomic_inc(&init_mm.mm_count);
8453 enter_lazy_tlb(&init_mm, current);
8456 * Make us the idle thread. Technically, schedule() should not be
8457 * called from this thread, however somewhere below it might be,
8458 * but because we are the idle thread, we just pick up running again
8459 * when this runqueue becomes "idle".
8461 init_idle(current, smp_processor_id());
8463 * During early bootup we pretend to be a normal task:
8465 current->sched_class = &fair_sched_class;
8467 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8468 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8471 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8473 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8476 scheduler_running = 1;
8479 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8480 void __might_sleep(char *file, int line)
8483 static unsigned long prev_jiffy; /* ratelimiting */
8485 if ((!in_atomic() && !irqs_disabled()) ||
8486 system_state != SYSTEM_RUNNING || oops_in_progress)
8488 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8490 prev_jiffy = jiffies;
8493 "BUG: sleeping function called from invalid context at %s:%d\n",
8496 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8497 in_atomic(), irqs_disabled(),
8498 current->pid, current->comm);
8500 debug_show_held_locks(current);
8501 if (irqs_disabled())
8502 print_irqtrace_events(current);
8506 EXPORT_SYMBOL(__might_sleep);
8509 #ifdef CONFIG_MAGIC_SYSRQ
8510 static void normalize_task(struct rq *rq, struct task_struct *p)
8514 update_rq_clock(rq);
8515 on_rq = p->se.on_rq;
8517 deactivate_task(rq, p, 0);
8518 __setscheduler(rq, p, SCHED_NORMAL, 0);
8520 activate_task(rq, p, 0);
8521 resched_task(rq->curr);
8525 void normalize_rt_tasks(void)
8527 struct task_struct *g, *p;
8528 unsigned long flags;
8531 read_lock_irqsave(&tasklist_lock, flags);
8532 do_each_thread(g, p) {
8534 * Only normalize user tasks:
8539 p->se.exec_start = 0;
8540 #ifdef CONFIG_SCHEDSTATS
8541 p->se.wait_start = 0;
8542 p->se.sleep_start = 0;
8543 p->se.block_start = 0;
8548 * Renice negative nice level userspace
8551 if (TASK_NICE(p) < 0 && p->mm)
8552 set_user_nice(p, 0);
8556 spin_lock(&p->pi_lock);
8557 rq = __task_rq_lock(p);
8559 normalize_task(rq, p);
8561 __task_rq_unlock(rq);
8562 spin_unlock(&p->pi_lock);
8563 } while_each_thread(g, p);
8565 read_unlock_irqrestore(&tasklist_lock, flags);
8568 #endif /* CONFIG_MAGIC_SYSRQ */
8572 * These functions are only useful for the IA64 MCA handling.
8574 * They can only be called when the whole system has been
8575 * stopped - every CPU needs to be quiescent, and no scheduling
8576 * activity can take place. Using them for anything else would
8577 * be a serious bug, and as a result, they aren't even visible
8578 * under any other configuration.
8582 * curr_task - return the current task for a given cpu.
8583 * @cpu: the processor in question.
8585 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8587 struct task_struct *curr_task(int cpu)
8589 return cpu_curr(cpu);
8593 * set_curr_task - set the current task for a given cpu.
8594 * @cpu: the processor in question.
8595 * @p: the task pointer to set.
8597 * Description: This function must only be used when non-maskable interrupts
8598 * are serviced on a separate stack. It allows the architecture to switch the
8599 * notion of the current task on a cpu in a non-blocking manner. This function
8600 * must be called with all CPU's synchronized, and interrupts disabled, the
8601 * and caller must save the original value of the current task (see
8602 * curr_task() above) and restore that value before reenabling interrupts and
8603 * re-starting the system.
8605 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8607 void set_curr_task(int cpu, struct task_struct *p)
8614 #ifdef CONFIG_FAIR_GROUP_SCHED
8615 static void free_fair_sched_group(struct task_group *tg)
8619 for_each_possible_cpu(i) {
8621 kfree(tg->cfs_rq[i]);
8631 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8633 struct cfs_rq *cfs_rq;
8634 struct sched_entity *se;
8638 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8641 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8645 tg->shares = NICE_0_LOAD;
8647 for_each_possible_cpu(i) {
8650 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8651 GFP_KERNEL, cpu_to_node(i));
8655 se = kzalloc_node(sizeof(struct sched_entity),
8656 GFP_KERNEL, cpu_to_node(i));
8660 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8669 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8671 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8672 &cpu_rq(cpu)->leaf_cfs_rq_list);
8675 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8677 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8679 #else /* !CONFG_FAIR_GROUP_SCHED */
8680 static inline void free_fair_sched_group(struct task_group *tg)
8685 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8690 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8694 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8697 #endif /* CONFIG_FAIR_GROUP_SCHED */
8699 #ifdef CONFIG_RT_GROUP_SCHED
8700 static void free_rt_sched_group(struct task_group *tg)
8704 destroy_rt_bandwidth(&tg->rt_bandwidth);
8706 for_each_possible_cpu(i) {
8708 kfree(tg->rt_rq[i]);
8710 kfree(tg->rt_se[i]);
8718 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8720 struct rt_rq *rt_rq;
8721 struct sched_rt_entity *rt_se;
8725 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8728 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8732 init_rt_bandwidth(&tg->rt_bandwidth,
8733 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8735 for_each_possible_cpu(i) {
8738 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8739 GFP_KERNEL, cpu_to_node(i));
8743 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8744 GFP_KERNEL, cpu_to_node(i));
8748 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8757 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8759 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8760 &cpu_rq(cpu)->leaf_rt_rq_list);
8763 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8765 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8767 #else /* !CONFIG_RT_GROUP_SCHED */
8768 static inline void free_rt_sched_group(struct task_group *tg)
8773 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8778 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8782 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8785 #endif /* CONFIG_RT_GROUP_SCHED */
8787 #ifdef CONFIG_GROUP_SCHED
8788 static void free_sched_group(struct task_group *tg)
8790 free_fair_sched_group(tg);
8791 free_rt_sched_group(tg);
8795 /* allocate runqueue etc for a new task group */
8796 struct task_group *sched_create_group(struct task_group *parent)
8798 struct task_group *tg;
8799 unsigned long flags;
8802 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8804 return ERR_PTR(-ENOMEM);
8806 if (!alloc_fair_sched_group(tg, parent))
8809 if (!alloc_rt_sched_group(tg, parent))
8812 spin_lock_irqsave(&task_group_lock, flags);
8813 for_each_possible_cpu(i) {
8814 register_fair_sched_group(tg, i);
8815 register_rt_sched_group(tg, i);
8817 list_add_rcu(&tg->list, &task_groups);
8819 WARN_ON(!parent); /* root should already exist */
8821 tg->parent = parent;
8822 INIT_LIST_HEAD(&tg->children);
8823 list_add_rcu(&tg->siblings, &parent->children);
8824 spin_unlock_irqrestore(&task_group_lock, flags);
8829 free_sched_group(tg);
8830 return ERR_PTR(-ENOMEM);
8833 /* rcu callback to free various structures associated with a task group */
8834 static void free_sched_group_rcu(struct rcu_head *rhp)
8836 /* now it should be safe to free those cfs_rqs */
8837 free_sched_group(container_of(rhp, struct task_group, rcu));
8840 /* Destroy runqueue etc associated with a task group */
8841 void sched_destroy_group(struct task_group *tg)
8843 unsigned long flags;
8846 spin_lock_irqsave(&task_group_lock, flags);
8847 for_each_possible_cpu(i) {
8848 unregister_fair_sched_group(tg, i);
8849 unregister_rt_sched_group(tg, i);
8851 list_del_rcu(&tg->list);
8852 list_del_rcu(&tg->siblings);
8853 spin_unlock_irqrestore(&task_group_lock, flags);
8855 /* wait for possible concurrent references to cfs_rqs complete */
8856 call_rcu(&tg->rcu, free_sched_group_rcu);
8859 /* change task's runqueue when it moves between groups.
8860 * The caller of this function should have put the task in its new group
8861 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8862 * reflect its new group.
8864 void sched_move_task(struct task_struct *tsk)
8867 unsigned long flags;
8870 rq = task_rq_lock(tsk, &flags);
8872 update_rq_clock(rq);
8874 running = task_current(rq, tsk);
8875 on_rq = tsk->se.on_rq;
8878 dequeue_task(rq, tsk, 0);
8879 if (unlikely(running))
8880 tsk->sched_class->put_prev_task(rq, tsk);
8882 set_task_rq(tsk, task_cpu(tsk));
8884 #ifdef CONFIG_FAIR_GROUP_SCHED
8885 if (tsk->sched_class->moved_group)
8886 tsk->sched_class->moved_group(tsk);
8889 if (unlikely(running))
8890 tsk->sched_class->set_curr_task(rq);
8892 enqueue_task(rq, tsk, 0);
8894 task_rq_unlock(rq, &flags);
8896 #endif /* CONFIG_GROUP_SCHED */
8898 #ifdef CONFIG_FAIR_GROUP_SCHED
8899 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8901 struct cfs_rq *cfs_rq = se->cfs_rq;
8906 dequeue_entity(cfs_rq, se, 0);
8908 se->load.weight = shares;
8909 se->load.inv_weight = 0;
8912 enqueue_entity(cfs_rq, se, 0);
8915 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8917 struct cfs_rq *cfs_rq = se->cfs_rq;
8918 struct rq *rq = cfs_rq->rq;
8919 unsigned long flags;
8921 spin_lock_irqsave(&rq->lock, flags);
8922 __set_se_shares(se, shares);
8923 spin_unlock_irqrestore(&rq->lock, flags);
8926 static DEFINE_MUTEX(shares_mutex);
8928 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8931 unsigned long flags;
8934 * We can't change the weight of the root cgroup.
8939 if (shares < MIN_SHARES)
8940 shares = MIN_SHARES;
8941 else if (shares > MAX_SHARES)
8942 shares = MAX_SHARES;
8944 mutex_lock(&shares_mutex);
8945 if (tg->shares == shares)
8948 spin_lock_irqsave(&task_group_lock, flags);
8949 for_each_possible_cpu(i)
8950 unregister_fair_sched_group(tg, i);
8951 list_del_rcu(&tg->siblings);
8952 spin_unlock_irqrestore(&task_group_lock, flags);
8954 /* wait for any ongoing reference to this group to finish */
8955 synchronize_sched();
8958 * Now we are free to modify the group's share on each cpu
8959 * w/o tripping rebalance_share or load_balance_fair.
8961 tg->shares = shares;
8962 for_each_possible_cpu(i) {
8966 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8967 set_se_shares(tg->se[i], shares);
8971 * Enable load balance activity on this group, by inserting it back on
8972 * each cpu's rq->leaf_cfs_rq_list.
8974 spin_lock_irqsave(&task_group_lock, flags);
8975 for_each_possible_cpu(i)
8976 register_fair_sched_group(tg, i);
8977 list_add_rcu(&tg->siblings, &tg->parent->children);
8978 spin_unlock_irqrestore(&task_group_lock, flags);
8980 mutex_unlock(&shares_mutex);
8984 unsigned long sched_group_shares(struct task_group *tg)
8990 #ifdef CONFIG_RT_GROUP_SCHED
8992 * Ensure that the real time constraints are schedulable.
8994 static DEFINE_MUTEX(rt_constraints_mutex);
8996 static unsigned long to_ratio(u64 period, u64 runtime)
8998 if (runtime == RUNTIME_INF)
9001 return div64_u64(runtime << 20, period);
9004 /* Must be called with tasklist_lock held */
9005 static inline int tg_has_rt_tasks(struct task_group *tg)
9007 struct task_struct *g, *p;
9009 do_each_thread(g, p) {
9010 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9012 } while_each_thread(g, p);
9017 struct rt_schedulable_data {
9018 struct task_group *tg;
9023 static int tg_schedulable(struct task_group *tg, void *data)
9025 struct rt_schedulable_data *d = data;
9026 struct task_group *child;
9027 unsigned long total, sum = 0;
9028 u64 period, runtime;
9030 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9031 runtime = tg->rt_bandwidth.rt_runtime;
9034 period = d->rt_period;
9035 runtime = d->rt_runtime;
9039 * Cannot have more runtime than the period.
9041 if (runtime > period && runtime != RUNTIME_INF)
9045 * Ensure we don't starve existing RT tasks.
9047 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9050 total = to_ratio(period, runtime);
9053 * Nobody can have more than the global setting allows.
9055 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9059 * The sum of our children's runtime should not exceed our own.
9061 list_for_each_entry_rcu(child, &tg->children, siblings) {
9062 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9063 runtime = child->rt_bandwidth.rt_runtime;
9065 if (child == d->tg) {
9066 period = d->rt_period;
9067 runtime = d->rt_runtime;
9070 sum += to_ratio(period, runtime);
9079 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9081 struct rt_schedulable_data data = {
9083 .rt_period = period,
9084 .rt_runtime = runtime,
9087 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9090 static int tg_set_bandwidth(struct task_group *tg,
9091 u64 rt_period, u64 rt_runtime)
9095 mutex_lock(&rt_constraints_mutex);
9096 read_lock(&tasklist_lock);
9097 err = __rt_schedulable(tg, rt_period, rt_runtime);
9101 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9102 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9103 tg->rt_bandwidth.rt_runtime = rt_runtime;
9105 for_each_possible_cpu(i) {
9106 struct rt_rq *rt_rq = tg->rt_rq[i];
9108 spin_lock(&rt_rq->rt_runtime_lock);
9109 rt_rq->rt_runtime = rt_runtime;
9110 spin_unlock(&rt_rq->rt_runtime_lock);
9112 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9114 read_unlock(&tasklist_lock);
9115 mutex_unlock(&rt_constraints_mutex);
9120 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9122 u64 rt_runtime, rt_period;
9124 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9125 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9126 if (rt_runtime_us < 0)
9127 rt_runtime = RUNTIME_INF;
9129 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9132 long sched_group_rt_runtime(struct task_group *tg)
9136 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9139 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9140 do_div(rt_runtime_us, NSEC_PER_USEC);
9141 return rt_runtime_us;
9144 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9146 u64 rt_runtime, rt_period;
9148 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9149 rt_runtime = tg->rt_bandwidth.rt_runtime;
9154 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9157 long sched_group_rt_period(struct task_group *tg)
9161 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9162 do_div(rt_period_us, NSEC_PER_USEC);
9163 return rt_period_us;
9166 static int sched_rt_global_constraints(void)
9168 u64 runtime, period;
9171 if (sysctl_sched_rt_period <= 0)
9174 runtime = global_rt_runtime();
9175 period = global_rt_period();
9178 * Sanity check on the sysctl variables.
9180 if (runtime > period && runtime != RUNTIME_INF)
9183 mutex_lock(&rt_constraints_mutex);
9184 read_lock(&tasklist_lock);
9185 ret = __rt_schedulable(NULL, 0, 0);
9186 read_unlock(&tasklist_lock);
9187 mutex_unlock(&rt_constraints_mutex);
9191 #else /* !CONFIG_RT_GROUP_SCHED */
9192 static int sched_rt_global_constraints(void)
9194 unsigned long flags;
9197 if (sysctl_sched_rt_period <= 0)
9200 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9201 for_each_possible_cpu(i) {
9202 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9204 spin_lock(&rt_rq->rt_runtime_lock);
9205 rt_rq->rt_runtime = global_rt_runtime();
9206 spin_unlock(&rt_rq->rt_runtime_lock);
9208 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9212 #endif /* CONFIG_RT_GROUP_SCHED */
9214 int sched_rt_handler(struct ctl_table *table, int write,
9215 struct file *filp, void __user *buffer, size_t *lenp,
9219 int old_period, old_runtime;
9220 static DEFINE_MUTEX(mutex);
9223 old_period = sysctl_sched_rt_period;
9224 old_runtime = sysctl_sched_rt_runtime;
9226 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9228 if (!ret && write) {
9229 ret = sched_rt_global_constraints();
9231 sysctl_sched_rt_period = old_period;
9232 sysctl_sched_rt_runtime = old_runtime;
9234 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9235 def_rt_bandwidth.rt_period =
9236 ns_to_ktime(global_rt_period());
9239 mutex_unlock(&mutex);
9244 #ifdef CONFIG_CGROUP_SCHED
9246 /* return corresponding task_group object of a cgroup */
9247 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9249 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9250 struct task_group, css);
9253 static struct cgroup_subsys_state *
9254 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9256 struct task_group *tg, *parent;
9258 if (!cgrp->parent) {
9259 /* This is early initialization for the top cgroup */
9260 return &init_task_group.css;
9263 parent = cgroup_tg(cgrp->parent);
9264 tg = sched_create_group(parent);
9266 return ERR_PTR(-ENOMEM);
9272 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9274 struct task_group *tg = cgroup_tg(cgrp);
9276 sched_destroy_group(tg);
9280 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9281 struct task_struct *tsk)
9283 #ifdef CONFIG_RT_GROUP_SCHED
9284 /* Don't accept realtime tasks when there is no way for them to run */
9285 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9288 /* We don't support RT-tasks being in separate groups */
9289 if (tsk->sched_class != &fair_sched_class)
9297 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9298 struct cgroup *old_cont, struct task_struct *tsk)
9300 sched_move_task(tsk);
9303 #ifdef CONFIG_FAIR_GROUP_SCHED
9304 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9307 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9310 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9312 struct task_group *tg = cgroup_tg(cgrp);
9314 return (u64) tg->shares;
9316 #endif /* CONFIG_FAIR_GROUP_SCHED */
9318 #ifdef CONFIG_RT_GROUP_SCHED
9319 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9322 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9325 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9327 return sched_group_rt_runtime(cgroup_tg(cgrp));
9330 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9333 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9336 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9338 return sched_group_rt_period(cgroup_tg(cgrp));
9340 #endif /* CONFIG_RT_GROUP_SCHED */
9342 static struct cftype cpu_files[] = {
9343 #ifdef CONFIG_FAIR_GROUP_SCHED
9346 .read_u64 = cpu_shares_read_u64,
9347 .write_u64 = cpu_shares_write_u64,
9350 #ifdef CONFIG_RT_GROUP_SCHED
9352 .name = "rt_runtime_us",
9353 .read_s64 = cpu_rt_runtime_read,
9354 .write_s64 = cpu_rt_runtime_write,
9357 .name = "rt_period_us",
9358 .read_u64 = cpu_rt_period_read_uint,
9359 .write_u64 = cpu_rt_period_write_uint,
9364 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9366 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9369 struct cgroup_subsys cpu_cgroup_subsys = {
9371 .create = cpu_cgroup_create,
9372 .destroy = cpu_cgroup_destroy,
9373 .can_attach = cpu_cgroup_can_attach,
9374 .attach = cpu_cgroup_attach,
9375 .populate = cpu_cgroup_populate,
9376 .subsys_id = cpu_cgroup_subsys_id,
9380 #endif /* CONFIG_CGROUP_SCHED */
9382 #ifdef CONFIG_CGROUP_CPUACCT
9385 * CPU accounting code for task groups.
9387 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9388 * (balbir@in.ibm.com).
9391 /* track cpu usage of a group of tasks and its child groups */
9393 struct cgroup_subsys_state css;
9394 /* cpuusage holds pointer to a u64-type object on every cpu */
9396 struct cpuacct *parent;
9399 struct cgroup_subsys cpuacct_subsys;
9401 /* return cpu accounting group corresponding to this container */
9402 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9404 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9405 struct cpuacct, css);
9408 /* return cpu accounting group to which this task belongs */
9409 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9411 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9412 struct cpuacct, css);
9415 /* create a new cpu accounting group */
9416 static struct cgroup_subsys_state *cpuacct_create(
9417 struct cgroup_subsys *ss, struct cgroup *cgrp)
9419 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9422 return ERR_PTR(-ENOMEM);
9424 ca->cpuusage = alloc_percpu(u64);
9425 if (!ca->cpuusage) {
9427 return ERR_PTR(-ENOMEM);
9431 ca->parent = cgroup_ca(cgrp->parent);
9436 /* destroy an existing cpu accounting group */
9438 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9440 struct cpuacct *ca = cgroup_ca(cgrp);
9442 free_percpu(ca->cpuusage);
9446 /* return total cpu usage (in nanoseconds) of a group */
9447 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9449 struct cpuacct *ca = cgroup_ca(cgrp);
9450 u64 totalcpuusage = 0;
9453 for_each_possible_cpu(i) {
9454 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9457 * Take rq->lock to make 64-bit addition safe on 32-bit
9460 spin_lock_irq(&cpu_rq(i)->lock);
9461 totalcpuusage += *cpuusage;
9462 spin_unlock_irq(&cpu_rq(i)->lock);
9465 return totalcpuusage;
9468 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9471 struct cpuacct *ca = cgroup_ca(cgrp);
9480 for_each_possible_cpu(i) {
9481 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9483 spin_lock_irq(&cpu_rq(i)->lock);
9485 spin_unlock_irq(&cpu_rq(i)->lock);
9491 static struct cftype files[] = {
9494 .read_u64 = cpuusage_read,
9495 .write_u64 = cpuusage_write,
9499 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9501 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9505 * charge this task's execution time to its accounting group.
9507 * called with rq->lock held.
9509 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9514 if (!cpuacct_subsys.active)
9517 cpu = task_cpu(tsk);
9520 for (; ca; ca = ca->parent) {
9521 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9522 *cpuusage += cputime;
9526 struct cgroup_subsys cpuacct_subsys = {
9528 .create = cpuacct_create,
9529 .destroy = cpuacct_destroy,
9530 .populate = cpuacct_populate,
9531 .subsys_id = cpuacct_subsys_id,
9533 #endif /* CONFIG_CGROUP_CPUACCT */