4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_counter.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy)
124 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
129 static inline int task_has_rt_policy(struct task_struct *p)
131 return rt_policy(p->policy);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array {
138 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
139 struct list_head queue[MAX_RT_PRIO];
142 struct rt_bandwidth {
143 /* nests inside the rq lock: */
144 spinlock_t rt_runtime_lock;
147 struct hrtimer rt_period_timer;
150 static struct rt_bandwidth def_rt_bandwidth;
152 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
154 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
156 struct rt_bandwidth *rt_b =
157 container_of(timer, struct rt_bandwidth, rt_period_timer);
163 now = hrtimer_cb_get_time(timer);
164 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
169 idle = do_sched_rt_period_timer(rt_b, overrun);
172 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
176 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
178 rt_b->rt_period = ns_to_ktime(period);
179 rt_b->rt_runtime = runtime;
181 spin_lock_init(&rt_b->rt_runtime_lock);
183 hrtimer_init(&rt_b->rt_period_timer,
184 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
185 rt_b->rt_period_timer.function = sched_rt_period_timer;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime >= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
197 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
200 if (hrtimer_active(&rt_b->rt_period_timer))
203 spin_lock(&rt_b->rt_runtime_lock);
208 if (hrtimer_active(&rt_b->rt_period_timer))
211 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
212 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
214 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
215 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
216 delta = ktime_to_ns(ktime_sub(hard, soft));
217 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
218 HRTIMER_MODE_ABS_PINNED, 0);
220 spin_unlock(&rt_b->rt_runtime_lock);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
226 hrtimer_cancel(&rt_b->rt_period_timer);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex);
236 #ifdef CONFIG_GROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups);
244 /* task group related information */
246 #ifdef CONFIG_CGROUP_SCHED
247 struct cgroup_subsys_state css;
250 #ifdef CONFIG_USER_SCHED
254 #ifdef CONFIG_FAIR_GROUP_SCHED
255 /* schedulable entities of this group on each cpu */
256 struct sched_entity **se;
257 /* runqueue "owned" by this group on each cpu */
258 struct cfs_rq **cfs_rq;
259 unsigned long shares;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity **rt_se;
264 struct rt_rq **rt_rq;
266 struct rt_bandwidth rt_bandwidth;
270 struct list_head list;
272 struct task_group *parent;
273 struct list_head siblings;
274 struct list_head children;
277 #ifdef CONFIG_USER_SCHED
279 /* Helper function to pass uid information to create_sched_user() */
280 void set_tg_uid(struct user_struct *user)
282 user->tg->uid = user->uid;
287 * Every UID task group (including init_task_group aka UID-0) will
288 * be a child to this group.
290 struct task_group root_task_group;
292 #ifdef CONFIG_FAIR_GROUP_SCHED
293 /* Default task group's sched entity on each cpu */
294 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
295 /* Default task group's cfs_rq on each cpu */
296 static DEFINE_PER_CPU(struct cfs_rq, init_tg_cfs_rq) ____cacheline_aligned_in_smp;
297 #endif /* CONFIG_FAIR_GROUP_SCHED */
299 #ifdef CONFIG_RT_GROUP_SCHED
300 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
301 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
302 #endif /* CONFIG_RT_GROUP_SCHED */
303 #else /* !CONFIG_USER_SCHED */
304 #define root_task_group init_task_group
305 #endif /* CONFIG_USER_SCHED */
307 /* task_group_lock serializes add/remove of task groups and also changes to
308 * a task group's cpu shares.
310 static DEFINE_SPINLOCK(task_group_lock);
313 static int root_task_group_empty(void)
315 return list_empty(&root_task_group.children);
319 #ifdef CONFIG_FAIR_GROUP_SCHED
320 #ifdef CONFIG_USER_SCHED
321 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
322 #else /* !CONFIG_USER_SCHED */
323 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
324 #endif /* CONFIG_USER_SCHED */
327 * A weight of 0 or 1 can cause arithmetics problems.
328 * A weight of a cfs_rq is the sum of weights of which entities
329 * are queued on this cfs_rq, so a weight of a entity should not be
330 * too large, so as the shares value of a task group.
331 * (The default weight is 1024 - so there's no practical
332 * limitation from this.)
335 #define MAX_SHARES (1UL << 18)
337 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
340 /* Default task group.
341 * Every task in system belong to this group at bootup.
343 struct task_group init_task_group;
345 /* return group to which a task belongs */
346 static inline struct task_group *task_group(struct task_struct *p)
348 struct task_group *tg;
350 #ifdef CONFIG_USER_SCHED
352 tg = __task_cred(p)->user->tg;
354 #elif defined(CONFIG_CGROUP_SCHED)
355 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
356 struct task_group, css);
358 tg = &init_task_group;
363 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
364 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
368 p->se.parent = task_group(p)->se[cpu];
371 #ifdef CONFIG_RT_GROUP_SCHED
372 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
373 p->rt.parent = task_group(p)->rt_se[cpu];
380 static int root_task_group_empty(void)
386 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
387 static inline struct task_group *task_group(struct task_struct *p)
392 #endif /* CONFIG_GROUP_SCHED */
394 /* CFS-related fields in a runqueue */
396 struct load_weight load;
397 unsigned long nr_running;
402 struct rb_root tasks_timeline;
403 struct rb_node *rb_leftmost;
405 struct list_head tasks;
406 struct list_head *balance_iterator;
409 * 'curr' points to currently running entity on this cfs_rq.
410 * It is set to NULL otherwise (i.e when none are currently running).
412 struct sched_entity *curr, *next, *last;
414 unsigned int nr_spread_over;
416 #ifdef CONFIG_FAIR_GROUP_SCHED
417 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
420 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
421 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
422 * (like users, containers etc.)
424 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
425 * list is used during load balance.
427 struct list_head leaf_cfs_rq_list;
428 struct task_group *tg; /* group that "owns" this runqueue */
432 * the part of load.weight contributed by tasks
434 unsigned long task_weight;
437 * h_load = weight * f(tg)
439 * Where f(tg) is the recursive weight fraction assigned to
442 unsigned long h_load;
445 * this cpu's part of tg->shares
447 unsigned long shares;
450 * load.weight at the time we set shares
452 unsigned long rq_weight;
457 /* Real-Time classes' related field in a runqueue: */
459 struct rt_prio_array active;
460 unsigned long rt_nr_running;
461 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
463 int curr; /* highest queued rt task prio */
465 int next; /* next highest */
470 unsigned long rt_nr_migratory;
471 unsigned long rt_nr_total;
473 struct plist_head pushable_tasks;
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];
552 unsigned long last_tick_seen;
553 unsigned char in_nohz_recently;
555 /* capture load from *all* tasks on this cpu: */
556 struct load_weight load;
557 unsigned long nr_load_updates;
559 u64 nr_migrations_in;
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 unsigned char idle_at_tick;
593 /* For active balancing */
597 /* cpu of this runqueue: */
601 unsigned long avg_load_per_task;
603 struct task_struct *migration_thread;
604 struct list_head migration_queue;
610 /* calc_load related fields */
611 unsigned long calc_load_update;
612 long calc_load_active;
614 #ifdef CONFIG_SCHED_HRTICK
616 int hrtick_csd_pending;
617 struct call_single_data hrtick_csd;
619 struct hrtimer hrtick_timer;
622 #ifdef CONFIG_SCHEDSTATS
624 struct sched_info rq_sched_info;
625 unsigned long long rq_cpu_time;
626 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
628 /* sys_sched_yield() stats */
629 unsigned int yld_count;
631 /* schedule() stats */
632 unsigned int sched_switch;
633 unsigned int sched_count;
634 unsigned int sched_goidle;
636 /* try_to_wake_up() stats */
637 unsigned int ttwu_count;
638 unsigned int ttwu_local;
641 unsigned int bkl_count;
645 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
647 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
649 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
652 static inline int cpu_of(struct rq *rq)
662 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
663 * See detach_destroy_domains: synchronize_sched for details.
665 * The domain tree of any CPU may only be accessed from within
666 * preempt-disabled sections.
668 #define for_each_domain(cpu, __sd) \
669 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
671 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
672 #define this_rq() (&__get_cpu_var(runqueues))
673 #define task_rq(p) cpu_rq(task_cpu(p))
674 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
675 #define raw_rq() (&__raw_get_cpu_var(runqueues))
677 inline void update_rq_clock(struct rq *rq)
679 rq->clock = sched_clock_cpu(cpu_of(rq));
683 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
685 #ifdef CONFIG_SCHED_DEBUG
686 # define const_debug __read_mostly
688 # define const_debug static const
694 * Returns true if the current cpu runqueue is locked.
695 * This interface allows printk to be called with the runqueue lock
696 * held and know whether or not it is OK to wake up the klogd.
698 int runqueue_is_locked(void)
701 struct rq *rq = cpu_rq(cpu);
704 ret = spin_is_locked(&rq->lock);
710 * Debugging: various feature bits
713 #define SCHED_FEAT(name, enabled) \
714 __SCHED_FEAT_##name ,
717 #include "sched_features.h"
722 #define SCHED_FEAT(name, enabled) \
723 (1UL << __SCHED_FEAT_##name) * enabled |
725 const_debug unsigned int sysctl_sched_features =
726 #include "sched_features.h"
731 #ifdef CONFIG_SCHED_DEBUG
732 #define SCHED_FEAT(name, enabled) \
735 static __read_mostly char *sched_feat_names[] = {
736 #include "sched_features.h"
742 static int sched_feat_show(struct seq_file *m, void *v)
746 for (i = 0; sched_feat_names[i]; i++) {
747 if (!(sysctl_sched_features & (1UL << i)))
749 seq_printf(m, "%s ", sched_feat_names[i]);
757 sched_feat_write(struct file *filp, const char __user *ubuf,
758 size_t cnt, loff_t *ppos)
768 if (copy_from_user(&buf, ubuf, cnt))
773 if (strncmp(buf, "NO_", 3) == 0) {
778 for (i = 0; sched_feat_names[i]; i++) {
779 int len = strlen(sched_feat_names[i]);
781 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
783 sysctl_sched_features &= ~(1UL << i);
785 sysctl_sched_features |= (1UL << i);
790 if (!sched_feat_names[i])
798 static int sched_feat_open(struct inode *inode, struct file *filp)
800 return single_open(filp, sched_feat_show, NULL);
803 static struct file_operations sched_feat_fops = {
804 .open = sched_feat_open,
805 .write = sched_feat_write,
808 .release = single_release,
811 static __init int sched_init_debug(void)
813 debugfs_create_file("sched_features", 0644, NULL, NULL,
818 late_initcall(sched_init_debug);
822 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
825 * Number of tasks to iterate in a single balance run.
826 * Limited because this is done with IRQs disabled.
828 const_debug unsigned int sysctl_sched_nr_migrate = 32;
831 * ratelimit for updating the group shares.
834 unsigned int sysctl_sched_shares_ratelimit = 250000;
837 * Inject some fuzzyness into changing the per-cpu group shares
838 * this avoids remote rq-locks at the expense of fairness.
841 unsigned int sysctl_sched_shares_thresh = 4;
844 * period over which we average the RT time consumption, measured
849 const_debug unsigned int sysctl_sched_time_avg = MSEC_PER_SEC;
852 * period over which we measure -rt task cpu usage in us.
855 unsigned int sysctl_sched_rt_period = 1000000;
857 static __read_mostly int scheduler_running;
860 * part of the period that we allow rt tasks to run in us.
863 int sysctl_sched_rt_runtime = 950000;
865 static inline u64 global_rt_period(void)
867 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
870 static inline u64 global_rt_runtime(void)
872 if (sysctl_sched_rt_runtime < 0)
875 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
878 #ifndef prepare_arch_switch
879 # define prepare_arch_switch(next) do { } while (0)
881 #ifndef finish_arch_switch
882 # define finish_arch_switch(prev) do { } while (0)
885 static inline int task_current(struct rq *rq, struct task_struct *p)
887 return rq->curr == p;
890 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
891 static inline int task_running(struct rq *rq, struct task_struct *p)
893 return task_current(rq, p);
896 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
900 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
902 #ifdef CONFIG_DEBUG_SPINLOCK
903 /* this is a valid case when another task releases the spinlock */
904 rq->lock.owner = current;
907 * If we are tracking spinlock dependencies then we have to
908 * fix up the runqueue lock - which gets 'carried over' from
911 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
913 spin_unlock_irq(&rq->lock);
916 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
917 static inline int task_running(struct rq *rq, struct task_struct *p)
922 return task_current(rq, p);
926 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
930 * We can optimise this out completely for !SMP, because the
931 * SMP rebalancing from interrupt is the only thing that cares
936 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
937 spin_unlock_irq(&rq->lock);
939 spin_unlock(&rq->lock);
943 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
947 * After ->oncpu is cleared, the task can be moved to a different CPU.
948 * We must ensure this doesn't happen until the switch is completely
954 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
958 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
961 * __task_rq_lock - lock the runqueue a given task resides on.
962 * Must be called interrupts disabled.
964 static inline struct rq *__task_rq_lock(struct task_struct *p)
968 struct rq *rq = task_rq(p);
969 spin_lock(&rq->lock);
970 if (likely(rq == task_rq(p)))
972 spin_unlock(&rq->lock);
977 * task_rq_lock - lock the runqueue a given task resides on and disable
978 * interrupts. Note the ordering: we can safely lookup the task_rq without
979 * explicitly disabling preemption.
981 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
987 local_irq_save(*flags);
989 spin_lock(&rq->lock);
990 if (likely(rq == task_rq(p)))
992 spin_unlock_irqrestore(&rq->lock, *flags);
996 void task_rq_unlock_wait(struct task_struct *p)
998 struct rq *rq = task_rq(p);
1000 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1001 spin_unlock_wait(&rq->lock);
1004 static void __task_rq_unlock(struct rq *rq)
1005 __releases(rq->lock)
1007 spin_unlock(&rq->lock);
1010 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1011 __releases(rq->lock)
1013 spin_unlock_irqrestore(&rq->lock, *flags);
1017 * this_rq_lock - lock this runqueue and disable interrupts.
1019 static struct rq *this_rq_lock(void)
1020 __acquires(rq->lock)
1024 local_irq_disable();
1026 spin_lock(&rq->lock);
1031 #ifdef CONFIG_SCHED_HRTICK
1033 * Use HR-timers to deliver accurate preemption points.
1035 * Its all a bit involved since we cannot program an hrt while holding the
1036 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1039 * When we get rescheduled we reprogram the hrtick_timer outside of the
1045 * - enabled by features
1046 * - hrtimer is actually high res
1048 static inline int hrtick_enabled(struct rq *rq)
1050 if (!sched_feat(HRTICK))
1052 if (!cpu_active(cpu_of(rq)))
1054 return hrtimer_is_hres_active(&rq->hrtick_timer);
1057 static void hrtick_clear(struct rq *rq)
1059 if (hrtimer_active(&rq->hrtick_timer))
1060 hrtimer_cancel(&rq->hrtick_timer);
1064 * High-resolution timer tick.
1065 * Runs from hardirq context with interrupts disabled.
1067 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1069 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1071 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1073 spin_lock(&rq->lock);
1074 update_rq_clock(rq);
1075 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1076 spin_unlock(&rq->lock);
1078 return HRTIMER_NORESTART;
1083 * called from hardirq (IPI) context
1085 static void __hrtick_start(void *arg)
1087 struct rq *rq = arg;
1089 spin_lock(&rq->lock);
1090 hrtimer_restart(&rq->hrtick_timer);
1091 rq->hrtick_csd_pending = 0;
1092 spin_unlock(&rq->lock);
1096 * Called to set the hrtick timer state.
1098 * called with rq->lock held and irqs disabled
1100 static void hrtick_start(struct rq *rq, u64 delay)
1102 struct hrtimer *timer = &rq->hrtick_timer;
1103 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1105 hrtimer_set_expires(timer, time);
1107 if (rq == this_rq()) {
1108 hrtimer_restart(timer);
1109 } else if (!rq->hrtick_csd_pending) {
1110 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1111 rq->hrtick_csd_pending = 1;
1116 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1118 int cpu = (int)(long)hcpu;
1121 case CPU_UP_CANCELED:
1122 case CPU_UP_CANCELED_FROZEN:
1123 case CPU_DOWN_PREPARE:
1124 case CPU_DOWN_PREPARE_FROZEN:
1126 case CPU_DEAD_FROZEN:
1127 hrtick_clear(cpu_rq(cpu));
1134 static __init void init_hrtick(void)
1136 hotcpu_notifier(hotplug_hrtick, 0);
1140 * Called to set the hrtick timer state.
1142 * called with rq->lock held and irqs disabled
1144 static void hrtick_start(struct rq *rq, u64 delay)
1146 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1147 HRTIMER_MODE_REL_PINNED, 0);
1150 static inline void init_hrtick(void)
1153 #endif /* CONFIG_SMP */
1155 static void init_rq_hrtick(struct rq *rq)
1158 rq->hrtick_csd_pending = 0;
1160 rq->hrtick_csd.flags = 0;
1161 rq->hrtick_csd.func = __hrtick_start;
1162 rq->hrtick_csd.info = rq;
1165 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1166 rq->hrtick_timer.function = hrtick;
1168 #else /* CONFIG_SCHED_HRTICK */
1169 static inline void hrtick_clear(struct rq *rq)
1173 static inline void init_rq_hrtick(struct rq *rq)
1177 static inline void init_hrtick(void)
1180 #endif /* CONFIG_SCHED_HRTICK */
1183 * resched_task - mark a task 'to be rescheduled now'.
1185 * On UP this means the setting of the need_resched flag, on SMP it
1186 * might also involve a cross-CPU call to trigger the scheduler on
1191 #ifndef tsk_is_polling
1192 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1195 static void resched_task(struct task_struct *p)
1199 assert_spin_locked(&task_rq(p)->lock);
1201 if (test_tsk_need_resched(p))
1204 set_tsk_need_resched(p);
1207 if (cpu == smp_processor_id())
1210 /* NEED_RESCHED must be visible before we test polling */
1212 if (!tsk_is_polling(p))
1213 smp_send_reschedule(cpu);
1216 static void resched_cpu(int cpu)
1218 struct rq *rq = cpu_rq(cpu);
1219 unsigned long flags;
1221 if (!spin_trylock_irqsave(&rq->lock, flags))
1223 resched_task(cpu_curr(cpu));
1224 spin_unlock_irqrestore(&rq->lock, flags);
1229 * When add_timer_on() enqueues a timer into the timer wheel of an
1230 * idle CPU then this timer might expire before the next timer event
1231 * which is scheduled to wake up that CPU. In case of a completely
1232 * idle system the next event might even be infinite time into the
1233 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1234 * leaves the inner idle loop so the newly added timer is taken into
1235 * account when the CPU goes back to idle and evaluates the timer
1236 * wheel for the next timer event.
1238 void wake_up_idle_cpu(int cpu)
1240 struct rq *rq = cpu_rq(cpu);
1242 if (cpu == smp_processor_id())
1246 * This is safe, as this function is called with the timer
1247 * wheel base lock of (cpu) held. When the CPU is on the way
1248 * to idle and has not yet set rq->curr to idle then it will
1249 * be serialized on the timer wheel base lock and take the new
1250 * timer into account automatically.
1252 if (rq->curr != rq->idle)
1256 * We can set TIF_RESCHED on the idle task of the other CPU
1257 * lockless. The worst case is that the other CPU runs the
1258 * idle task through an additional NOOP schedule()
1260 set_tsk_need_resched(rq->idle);
1262 /* NEED_RESCHED must be visible before we test polling */
1264 if (!tsk_is_polling(rq->idle))
1265 smp_send_reschedule(cpu);
1267 #endif /* CONFIG_NO_HZ */
1269 static u64 sched_avg_period(void)
1271 return (u64)sysctl_sched_time_avg * NSEC_PER_MSEC / 2;
1274 static void sched_avg_update(struct rq *rq)
1276 s64 period = sched_avg_period();
1278 while ((s64)(rq->clock - rq->age_stamp) > period) {
1279 rq->age_stamp += period;
1284 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1286 rq->rt_avg += rt_delta;
1287 sched_avg_update(rq);
1290 #else /* !CONFIG_SMP */
1291 static void resched_task(struct task_struct *p)
1293 assert_spin_locked(&task_rq(p)->lock);
1294 set_tsk_need_resched(p);
1297 static void sched_rt_avg_update(struct rq *rq, u64 rt_delta)
1300 #endif /* CONFIG_SMP */
1302 #if BITS_PER_LONG == 32
1303 # define WMULT_CONST (~0UL)
1305 # define WMULT_CONST (1UL << 32)
1308 #define WMULT_SHIFT 32
1311 * Shift right and round:
1313 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1316 * delta *= weight / lw
1318 static unsigned long
1319 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1320 struct load_weight *lw)
1324 if (!lw->inv_weight) {
1325 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1328 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1332 tmp = (u64)delta_exec * weight;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp > WMULT_CONST))
1337 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1340 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1342 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1345 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1351 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1358 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1359 * of tasks with abnormal "nice" values across CPUs the contribution that
1360 * each task makes to its run queue's load is weighted according to its
1361 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1362 * scaled version of the new time slice allocation that they receive on time
1366 #define WEIGHT_IDLEPRIO 3
1367 #define WMULT_IDLEPRIO 1431655765
1370 * Nice levels are multiplicative, with a gentle 10% change for every
1371 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1372 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1373 * that remained on nice 0.
1375 * The "10% effect" is relative and cumulative: from _any_ nice level,
1376 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1377 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1378 * If a task goes up by ~10% and another task goes down by ~10% then
1379 * the relative distance between them is ~25%.)
1381 static const int prio_to_weight[40] = {
1382 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1383 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1384 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1385 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1386 /* 0 */ 1024, 820, 655, 526, 423,
1387 /* 5 */ 335, 272, 215, 172, 137,
1388 /* 10 */ 110, 87, 70, 56, 45,
1389 /* 15 */ 36, 29, 23, 18, 15,
1393 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1395 * In cases where the weight does not change often, we can use the
1396 * precalculated inverse to speed up arithmetics by turning divisions
1397 * into multiplications:
1399 static const u32 prio_to_wmult[40] = {
1400 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1401 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1402 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1403 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1404 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1405 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1406 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1407 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1410 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1413 * runqueue iterator, to support SMP load-balancing between different
1414 * scheduling classes, without having to expose their internal data
1415 * structures to the load-balancing proper:
1417 struct rq_iterator {
1419 struct task_struct *(*start)(void *);
1420 struct task_struct *(*next)(void *);
1424 static unsigned long
1425 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1426 unsigned long max_load_move, struct sched_domain *sd,
1427 enum cpu_idle_type idle, int *all_pinned,
1428 int *this_best_prio, struct rq_iterator *iterator);
1431 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1432 struct sched_domain *sd, enum cpu_idle_type idle,
1433 struct rq_iterator *iterator);
1436 /* Time spent by the tasks of the cpu accounting group executing in ... */
1437 enum cpuacct_stat_index {
1438 CPUACCT_STAT_USER, /* ... user mode */
1439 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1441 CPUACCT_STAT_NSTATS,
1444 #ifdef CONFIG_CGROUP_CPUACCT
1445 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1446 static void cpuacct_update_stats(struct task_struct *tsk,
1447 enum cpuacct_stat_index idx, cputime_t val);
1449 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1450 static inline void cpuacct_update_stats(struct task_struct *tsk,
1451 enum cpuacct_stat_index idx, cputime_t val) {}
1454 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1456 update_load_add(&rq->load, load);
1459 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1461 update_load_sub(&rq->load, load);
1464 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1465 typedef int (*tg_visitor)(struct task_group *, void *);
1468 * Iterate the full tree, calling @down when first entering a node and @up when
1469 * leaving it for the final time.
1471 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1473 struct task_group *parent, *child;
1477 parent = &root_task_group;
1479 ret = (*down)(parent, data);
1482 list_for_each_entry_rcu(child, &parent->children, siblings) {
1489 ret = (*up)(parent, data);
1494 parent = parent->parent;
1503 static int tg_nop(struct task_group *tg, void *data)
1510 /* Used instead of source_load when we know the type == 0 */
1511 static unsigned long weighted_cpuload(const int cpu)
1513 return cpu_rq(cpu)->load.weight;
1517 * Return a low guess at the load of a migration-source cpu weighted
1518 * according to the scheduling class and "nice" value.
1520 * We want to under-estimate the load of migration sources, to
1521 * balance conservatively.
1523 static unsigned long source_load(int cpu, int type)
1525 struct rq *rq = cpu_rq(cpu);
1526 unsigned long total = weighted_cpuload(cpu);
1528 if (type == 0 || !sched_feat(LB_BIAS))
1531 return min(rq->cpu_load[type-1], total);
1535 * Return a high guess at the load of a migration-target cpu weighted
1536 * according to the scheduling class and "nice" value.
1538 static unsigned long target_load(int cpu, int type)
1540 struct rq *rq = cpu_rq(cpu);
1541 unsigned long total = weighted_cpuload(cpu);
1543 if (type == 0 || !sched_feat(LB_BIAS))
1546 return max(rq->cpu_load[type-1], total);
1549 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1551 static unsigned long cpu_avg_load_per_task(int cpu)
1553 struct rq *rq = cpu_rq(cpu);
1554 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1557 rq->avg_load_per_task = rq->load.weight / nr_running;
1559 rq->avg_load_per_task = 0;
1561 return rq->avg_load_per_task;
1564 #ifdef CONFIG_FAIR_GROUP_SCHED
1566 struct update_shares_data {
1567 unsigned long rq_weight[NR_CPUS];
1570 static DEFINE_PER_CPU(struct update_shares_data, update_shares_data);
1572 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1575 * Calculate and set the cpu's group shares.
1577 static void update_group_shares_cpu(struct task_group *tg, int cpu,
1578 unsigned long sd_shares,
1579 unsigned long sd_rq_weight,
1580 struct update_shares_data *usd)
1582 unsigned long shares, rq_weight;
1585 rq_weight = usd->rq_weight[cpu];
1588 rq_weight = NICE_0_LOAD;
1592 * \Sum_j shares_j * rq_weight_i
1593 * shares_i = -----------------------------
1594 * \Sum_j rq_weight_j
1596 shares = (sd_shares * rq_weight) / sd_rq_weight;
1597 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1599 if (abs(shares - tg->se[cpu]->load.weight) >
1600 sysctl_sched_shares_thresh) {
1601 struct rq *rq = cpu_rq(cpu);
1602 unsigned long flags;
1604 spin_lock_irqsave(&rq->lock, flags);
1605 tg->cfs_rq[cpu]->rq_weight = boost ? 0 : rq_weight;
1606 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1607 __set_se_shares(tg->se[cpu], shares);
1608 spin_unlock_irqrestore(&rq->lock, flags);
1613 * Re-compute the task group their per cpu shares over the given domain.
1614 * This needs to be done in a bottom-up fashion because the rq weight of a
1615 * parent group depends on the shares of its child groups.
1617 static int tg_shares_up(struct task_group *tg, void *data)
1619 unsigned long weight, rq_weight = 0, shares = 0;
1620 struct update_shares_data *usd;
1621 struct sched_domain *sd = data;
1622 unsigned long flags;
1628 local_irq_save(flags);
1629 usd = &__get_cpu_var(update_shares_data);
1631 for_each_cpu(i, sched_domain_span(sd)) {
1632 weight = tg->cfs_rq[i]->load.weight;
1633 usd->rq_weight[i] = weight;
1636 * If there are currently no tasks on the cpu pretend there
1637 * is one of average load so that when a new task gets to
1638 * run here it will not get delayed by group starvation.
1641 weight = NICE_0_LOAD;
1643 rq_weight += weight;
1644 shares += tg->cfs_rq[i]->shares;
1647 if ((!shares && rq_weight) || shares > tg->shares)
1648 shares = tg->shares;
1650 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1651 shares = tg->shares;
1653 for_each_cpu(i, sched_domain_span(sd))
1654 update_group_shares_cpu(tg, i, shares, rq_weight, usd);
1656 local_irq_restore(flags);
1662 * Compute the cpu's hierarchical load factor for each task group.
1663 * This needs to be done in a top-down fashion because the load of a child
1664 * group is a fraction of its parents load.
1666 static int tg_load_down(struct task_group *tg, void *data)
1669 long cpu = (long)data;
1672 load = cpu_rq(cpu)->load.weight;
1674 load = tg->parent->cfs_rq[cpu]->h_load;
1675 load *= tg->cfs_rq[cpu]->shares;
1676 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1679 tg->cfs_rq[cpu]->h_load = load;
1684 static void update_shares(struct sched_domain *sd)
1689 if (root_task_group_empty())
1692 now = cpu_clock(raw_smp_processor_id());
1693 elapsed = now - sd->last_update;
1695 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1696 sd->last_update = now;
1697 walk_tg_tree(tg_nop, tg_shares_up, sd);
1701 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1703 if (root_task_group_empty())
1706 spin_unlock(&rq->lock);
1708 spin_lock(&rq->lock);
1711 static void update_h_load(long cpu)
1713 if (root_task_group_empty())
1716 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1721 static inline void update_shares(struct sched_domain *sd)
1725 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1731 #ifdef CONFIG_PREEMPT
1733 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
1736 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1737 * way at the expense of forcing extra atomic operations in all
1738 * invocations. This assures that the double_lock is acquired using the
1739 * same underlying policy as the spinlock_t on this architecture, which
1740 * reduces latency compared to the unfair variant below. However, it
1741 * also adds more overhead and therefore may reduce throughput.
1743 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1744 __releases(this_rq->lock)
1745 __acquires(busiest->lock)
1746 __acquires(this_rq->lock)
1748 spin_unlock(&this_rq->lock);
1749 double_rq_lock(this_rq, busiest);
1756 * Unfair double_lock_balance: Optimizes throughput at the expense of
1757 * latency by eliminating extra atomic operations when the locks are
1758 * already in proper order on entry. This favors lower cpu-ids and will
1759 * grant the double lock to lower cpus over higher ids under contention,
1760 * regardless of entry order into the function.
1762 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1763 __releases(this_rq->lock)
1764 __acquires(busiest->lock)
1765 __acquires(this_rq->lock)
1769 if (unlikely(!spin_trylock(&busiest->lock))) {
1770 if (busiest < this_rq) {
1771 spin_unlock(&this_rq->lock);
1772 spin_lock(&busiest->lock);
1773 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1776 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1781 #endif /* CONFIG_PREEMPT */
1784 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1786 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1788 if (unlikely(!irqs_disabled())) {
1789 /* printk() doesn't work good under rq->lock */
1790 spin_unlock(&this_rq->lock);
1794 return _double_lock_balance(this_rq, busiest);
1797 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1798 __releases(busiest->lock)
1800 spin_unlock(&busiest->lock);
1801 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1805 #ifdef CONFIG_FAIR_GROUP_SCHED
1806 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1809 cfs_rq->shares = shares;
1814 static void calc_load_account_active(struct rq *this_rq);
1816 #include "sched_stats.h"
1817 #include "sched_idletask.c"
1818 #include "sched_fair.c"
1819 #include "sched_rt.c"
1820 #ifdef CONFIG_SCHED_DEBUG
1821 # include "sched_debug.c"
1824 #define sched_class_highest (&rt_sched_class)
1825 #define for_each_class(class) \
1826 for (class = sched_class_highest; class; class = class->next)
1828 static void inc_nr_running(struct rq *rq)
1833 static void dec_nr_running(struct rq *rq)
1838 static void set_load_weight(struct task_struct *p)
1840 if (task_has_rt_policy(p)) {
1841 p->se.load.weight = prio_to_weight[0] * 2;
1842 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1847 * SCHED_IDLE tasks get minimal weight:
1849 if (p->policy == SCHED_IDLE) {
1850 p->se.load.weight = WEIGHT_IDLEPRIO;
1851 p->se.load.inv_weight = WMULT_IDLEPRIO;
1855 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1856 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1859 static void update_avg(u64 *avg, u64 sample)
1861 s64 diff = sample - *avg;
1865 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1868 p->se.start_runtime = p->se.sum_exec_runtime;
1870 sched_info_queued(p);
1871 p->sched_class->enqueue_task(rq, p, wakeup);
1875 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1878 if (p->se.last_wakeup) {
1879 update_avg(&p->se.avg_overlap,
1880 p->se.sum_exec_runtime - p->se.last_wakeup);
1881 p->se.last_wakeup = 0;
1883 update_avg(&p->se.avg_wakeup,
1884 sysctl_sched_wakeup_granularity);
1888 sched_info_dequeued(p);
1889 p->sched_class->dequeue_task(rq, p, sleep);
1894 * __normal_prio - return the priority that is based on the static prio
1896 static inline int __normal_prio(struct task_struct *p)
1898 return p->static_prio;
1902 * Calculate the expected normal priority: i.e. priority
1903 * without taking RT-inheritance into account. Might be
1904 * boosted by interactivity modifiers. Changes upon fork,
1905 * setprio syscalls, and whenever the interactivity
1906 * estimator recalculates.
1908 static inline int normal_prio(struct task_struct *p)
1912 if (task_has_rt_policy(p))
1913 prio = MAX_RT_PRIO-1 - p->rt_priority;
1915 prio = __normal_prio(p);
1920 * Calculate the current priority, i.e. the priority
1921 * taken into account by the scheduler. This value might
1922 * be boosted by RT tasks, or might be boosted by
1923 * interactivity modifiers. Will be RT if the task got
1924 * RT-boosted. If not then it returns p->normal_prio.
1926 static int effective_prio(struct task_struct *p)
1928 p->normal_prio = normal_prio(p);
1930 * If we are RT tasks or we were boosted to RT priority,
1931 * keep the priority unchanged. Otherwise, update priority
1932 * to the normal priority:
1934 if (!rt_prio(p->prio))
1935 return p->normal_prio;
1940 * activate_task - move a task to the runqueue.
1942 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1944 if (task_contributes_to_load(p))
1945 rq->nr_uninterruptible--;
1947 enqueue_task(rq, p, wakeup);
1952 * deactivate_task - remove a task from the runqueue.
1954 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1956 if (task_contributes_to_load(p))
1957 rq->nr_uninterruptible++;
1959 dequeue_task(rq, p, sleep);
1964 * task_curr - is this task currently executing on a CPU?
1965 * @p: the task in question.
1967 inline int task_curr(const struct task_struct *p)
1969 return cpu_curr(task_cpu(p)) == p;
1972 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1974 set_task_rq(p, cpu);
1977 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1978 * successfuly executed on another CPU. We must ensure that updates of
1979 * per-task data have been completed by this moment.
1982 task_thread_info(p)->cpu = cpu;
1986 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1987 const struct sched_class *prev_class,
1988 int oldprio, int running)
1990 if (prev_class != p->sched_class) {
1991 if (prev_class->switched_from)
1992 prev_class->switched_from(rq, p, running);
1993 p->sched_class->switched_to(rq, p, running);
1995 p->sched_class->prio_changed(rq, p, oldprio, running);
2000 * Is this task likely cache-hot:
2003 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY) &&
2011 (&p->se == cfs_rq_of(&p->se)->next ||
2012 &p->se == cfs_rq_of(&p->se)->last))
2015 if (p->sched_class != &fair_sched_class)
2018 if (sysctl_sched_migration_cost == -1)
2020 if (sysctl_sched_migration_cost == 0)
2023 delta = now - p->se.exec_start;
2025 return delta < (s64)sysctl_sched_migration_cost;
2029 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2031 int old_cpu = task_cpu(p);
2032 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2033 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2034 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2037 clock_offset = old_rq->clock - new_rq->clock;
2039 trace_sched_migrate_task(p, new_cpu);
2041 #ifdef CONFIG_SCHEDSTATS
2042 if (p->se.wait_start)
2043 p->se.wait_start -= clock_offset;
2044 if (p->se.sleep_start)
2045 p->se.sleep_start -= clock_offset;
2046 if (p->se.block_start)
2047 p->se.block_start -= clock_offset;
2049 if (old_cpu != new_cpu) {
2050 p->se.nr_migrations++;
2051 new_rq->nr_migrations_in++;
2052 #ifdef CONFIG_SCHEDSTATS
2053 if (task_hot(p, old_rq->clock, NULL))
2054 schedstat_inc(p, se.nr_forced2_migrations);
2056 perf_swcounter_event(PERF_COUNT_SW_CPU_MIGRATIONS,
2059 p->se.vruntime -= old_cfsrq->min_vruntime -
2060 new_cfsrq->min_vruntime;
2062 __set_task_cpu(p, new_cpu);
2065 struct migration_req {
2066 struct list_head list;
2068 struct task_struct *task;
2071 struct completion done;
2075 * The task's runqueue lock must be held.
2076 * Returns true if you have to wait for migration thread.
2079 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2081 struct rq *rq = task_rq(p);
2084 * If the task is not on a runqueue (and not running), then
2085 * it is sufficient to simply update the task's cpu field.
2087 if (!p->se.on_rq && !task_running(rq, p)) {
2088 set_task_cpu(p, dest_cpu);
2092 init_completion(&req->done);
2094 req->dest_cpu = dest_cpu;
2095 list_add(&req->list, &rq->migration_queue);
2101 * wait_task_context_switch - wait for a thread to complete at least one
2104 * @p must not be current.
2106 void wait_task_context_switch(struct task_struct *p)
2108 unsigned long nvcsw, nivcsw, flags;
2116 * The runqueue is assigned before the actual context
2117 * switch. We need to take the runqueue lock.
2119 * We could check initially without the lock but it is
2120 * very likely that we need to take the lock in every
2123 rq = task_rq_lock(p, &flags);
2124 running = task_running(rq, p);
2125 task_rq_unlock(rq, &flags);
2127 if (likely(!running))
2130 * The switch count is incremented before the actual
2131 * context switch. We thus wait for two switches to be
2132 * sure at least one completed.
2134 if ((p->nvcsw - nvcsw) > 1)
2136 if ((p->nivcsw - nivcsw) > 1)
2144 * wait_task_inactive - wait for a thread to unschedule.
2146 * If @match_state is nonzero, it's the @p->state value just checked and
2147 * not expected to change. If it changes, i.e. @p might have woken up,
2148 * then return zero. When we succeed in waiting for @p to be off its CPU,
2149 * we return a positive number (its total switch count). If a second call
2150 * a short while later returns the same number, the caller can be sure that
2151 * @p has remained unscheduled the whole time.
2153 * The caller must ensure that the task *will* unschedule sometime soon,
2154 * else this function might spin for a *long* time. This function can't
2155 * be called with interrupts off, or it may introduce deadlock with
2156 * smp_call_function() if an IPI is sent by the same process we are
2157 * waiting to become inactive.
2159 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2161 unsigned long flags;
2168 * We do the initial early heuristics without holding
2169 * any task-queue locks at all. We'll only try to get
2170 * the runqueue lock when things look like they will
2176 * If the task is actively running on another CPU
2177 * still, just relax and busy-wait without holding
2180 * NOTE! Since we don't hold any locks, it's not
2181 * even sure that "rq" stays as the right runqueue!
2182 * But we don't care, since "task_running()" will
2183 * return false if the runqueue has changed and p
2184 * is actually now running somewhere else!
2186 while (task_running(rq, p)) {
2187 if (match_state && unlikely(p->state != match_state))
2193 * Ok, time to look more closely! We need the rq
2194 * lock now, to be *sure*. If we're wrong, we'll
2195 * just go back and repeat.
2197 rq = task_rq_lock(p, &flags);
2198 trace_sched_wait_task(rq, p);
2199 running = task_running(rq, p);
2200 on_rq = p->se.on_rq;
2202 if (!match_state || p->state == match_state)
2203 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2204 task_rq_unlock(rq, &flags);
2207 * If it changed from the expected state, bail out now.
2209 if (unlikely(!ncsw))
2213 * Was it really running after all now that we
2214 * checked with the proper locks actually held?
2216 * Oops. Go back and try again..
2218 if (unlikely(running)) {
2224 * It's not enough that it's not actively running,
2225 * it must be off the runqueue _entirely_, and not
2228 * So if it was still runnable (but just not actively
2229 * running right now), it's preempted, and we should
2230 * yield - it could be a while.
2232 if (unlikely(on_rq)) {
2233 schedule_timeout_uninterruptible(1);
2238 * Ahh, all good. It wasn't running, and it wasn't
2239 * runnable, which means that it will never become
2240 * running in the future either. We're all done!
2249 * kick_process - kick a running thread to enter/exit the kernel
2250 * @p: the to-be-kicked thread
2252 * Cause a process which is running on another CPU to enter
2253 * kernel-mode, without any delay. (to get signals handled.)
2255 * NOTE: this function doesnt have to take the runqueue lock,
2256 * because all it wants to ensure is that the remote task enters
2257 * the kernel. If the IPI races and the task has been migrated
2258 * to another CPU then no harm is done and the purpose has been
2261 void kick_process(struct task_struct *p)
2267 if ((cpu != smp_processor_id()) && task_curr(p))
2268 smp_send_reschedule(cpu);
2271 EXPORT_SYMBOL_GPL(kick_process);
2272 #endif /* CONFIG_SMP */
2275 * task_oncpu_function_call - call a function on the cpu on which a task runs
2276 * @p: the task to evaluate
2277 * @func: the function to be called
2278 * @info: the function call argument
2280 * Calls the function @func when the task is currently running. This might
2281 * be on the current CPU, which just calls the function directly
2283 void task_oncpu_function_call(struct task_struct *p,
2284 void (*func) (void *info), void *info)
2291 smp_call_function_single(cpu, func, info, 1);
2296 * try_to_wake_up - wake up a thread
2297 * @p: the to-be-woken-up thread
2298 * @state: the mask of task states that can be woken
2299 * @sync: do a synchronous wakeup?
2301 * Put it on the run-queue if it's not already there. The "current"
2302 * thread is always on the run-queue (except when the actual
2303 * re-schedule is in progress), and as such you're allowed to do
2304 * the simpler "current->state = TASK_RUNNING" to mark yourself
2305 * runnable without the overhead of this.
2307 * returns failure only if the task is already active.
2309 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2311 int cpu, orig_cpu, this_cpu, success = 0;
2312 unsigned long flags;
2315 if (!sched_feat(SYNC_WAKEUPS))
2319 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2320 struct sched_domain *sd;
2322 this_cpu = raw_smp_processor_id();
2325 for_each_domain(this_cpu, sd) {
2326 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2334 this_cpu = get_cpu();
2337 rq = task_rq_lock(p, &flags);
2338 update_rq_clock(rq);
2339 if (!(p->state & state))
2349 if (unlikely(task_running(rq, p)))
2353 * In order to handle concurrent wakeups and release the rq->lock
2354 * we put the task in TASK_WAKING state.
2356 p->state = TASK_WAKING;
2357 task_rq_unlock(rq, &flags);
2359 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_WAKE, sync);
2360 if (cpu != orig_cpu)
2361 set_task_cpu(p, cpu);
2363 rq = task_rq_lock(p, &flags);
2364 WARN_ON(p->state != TASK_WAKING);
2367 #ifdef CONFIG_SCHEDSTATS
2368 schedstat_inc(rq, ttwu_count);
2369 if (cpu == this_cpu)
2370 schedstat_inc(rq, ttwu_local);
2372 struct sched_domain *sd;
2373 for_each_domain(this_cpu, sd) {
2374 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2375 schedstat_inc(sd, ttwu_wake_remote);
2380 #endif /* CONFIG_SCHEDSTATS */
2383 #endif /* CONFIG_SMP */
2384 schedstat_inc(p, se.nr_wakeups);
2386 schedstat_inc(p, se.nr_wakeups_sync);
2387 if (orig_cpu != cpu)
2388 schedstat_inc(p, se.nr_wakeups_migrate);
2389 if (cpu == this_cpu)
2390 schedstat_inc(p, se.nr_wakeups_local);
2392 schedstat_inc(p, se.nr_wakeups_remote);
2393 activate_task(rq, p, 1);
2397 * Only attribute actual wakeups done by this task.
2399 if (!in_interrupt()) {
2400 struct sched_entity *se = ¤t->se;
2401 u64 sample = se->sum_exec_runtime;
2403 if (se->last_wakeup)
2404 sample -= se->last_wakeup;
2406 sample -= se->start_runtime;
2407 update_avg(&se->avg_wakeup, sample);
2409 se->last_wakeup = se->sum_exec_runtime;
2413 trace_sched_wakeup(rq, p, success);
2414 check_preempt_curr(rq, p, sync);
2416 p->state = TASK_RUNNING;
2418 if (p->sched_class->task_wake_up)
2419 p->sched_class->task_wake_up(rq, p);
2422 task_rq_unlock(rq, &flags);
2429 * wake_up_process - Wake up a specific process
2430 * @p: The process to be woken up.
2432 * Attempt to wake up the nominated process and move it to the set of runnable
2433 * processes. Returns 1 if the process was woken up, 0 if it was already
2436 * It may be assumed that this function implies a write memory barrier before
2437 * changing the task state if and only if any tasks are woken up.
2439 int wake_up_process(struct task_struct *p)
2441 return try_to_wake_up(p, TASK_ALL, 0);
2443 EXPORT_SYMBOL(wake_up_process);
2445 int wake_up_state(struct task_struct *p, unsigned int state)
2447 return try_to_wake_up(p, state, 0);
2451 * Perform scheduler related setup for a newly forked process p.
2452 * p is forked by current.
2454 * __sched_fork() is basic setup used by init_idle() too:
2456 static void __sched_fork(struct task_struct *p)
2458 p->se.exec_start = 0;
2459 p->se.sum_exec_runtime = 0;
2460 p->se.prev_sum_exec_runtime = 0;
2461 p->se.nr_migrations = 0;
2462 p->se.last_wakeup = 0;
2463 p->se.avg_overlap = 0;
2464 p->se.start_runtime = 0;
2465 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2467 #ifdef CONFIG_SCHEDSTATS
2468 p->se.wait_start = 0;
2470 p->se.wait_count = 0;
2473 p->se.sleep_start = 0;
2474 p->se.sleep_max = 0;
2475 p->se.sum_sleep_runtime = 0;
2477 p->se.block_start = 0;
2478 p->se.block_max = 0;
2480 p->se.slice_max = 0;
2482 p->se.nr_migrations_cold = 0;
2483 p->se.nr_failed_migrations_affine = 0;
2484 p->se.nr_failed_migrations_running = 0;
2485 p->se.nr_failed_migrations_hot = 0;
2486 p->se.nr_forced_migrations = 0;
2487 p->se.nr_forced2_migrations = 0;
2489 p->se.nr_wakeups = 0;
2490 p->se.nr_wakeups_sync = 0;
2491 p->se.nr_wakeups_migrate = 0;
2492 p->se.nr_wakeups_local = 0;
2493 p->se.nr_wakeups_remote = 0;
2494 p->se.nr_wakeups_affine = 0;
2495 p->se.nr_wakeups_affine_attempts = 0;
2496 p->se.nr_wakeups_passive = 0;
2497 p->se.nr_wakeups_idle = 0;
2501 INIT_LIST_HEAD(&p->rt.run_list);
2503 INIT_LIST_HEAD(&p->se.group_node);
2505 #ifdef CONFIG_PREEMPT_NOTIFIERS
2506 INIT_HLIST_HEAD(&p->preempt_notifiers);
2510 * We mark the process as running here, but have not actually
2511 * inserted it onto the runqueue yet. This guarantees that
2512 * nobody will actually run it, and a signal or other external
2513 * event cannot wake it up and insert it on the runqueue either.
2515 p->state = TASK_RUNNING;
2519 * fork()/clone()-time setup:
2521 void sched_fork(struct task_struct *p, int clone_flags)
2523 int cpu = get_cpu();
2528 * Make sure we do not leak PI boosting priority to the child.
2530 p->prio = current->normal_prio;
2533 * Revert to default priority/policy on fork if requested.
2535 if (unlikely(p->sched_reset_on_fork)) {
2536 if (p->policy == SCHED_FIFO || p->policy == SCHED_RR)
2537 p->policy = SCHED_NORMAL;
2539 if (p->normal_prio < DEFAULT_PRIO)
2540 p->prio = DEFAULT_PRIO;
2542 if (PRIO_TO_NICE(p->static_prio) < 0) {
2543 p->static_prio = NICE_TO_PRIO(0);
2548 * We don't need the reset flag anymore after the fork. It has
2549 * fulfilled its duty:
2551 p->sched_reset_on_fork = 0;
2554 if (!rt_prio(p->prio))
2555 p->sched_class = &fair_sched_class;
2558 cpu = p->sched_class->select_task_rq(p, SD_BALANCE_FORK, 0);
2560 set_task_cpu(p, cpu);
2562 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2563 if (likely(sched_info_on()))
2564 memset(&p->sched_info, 0, sizeof(p->sched_info));
2566 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2569 #ifdef CONFIG_PREEMPT
2570 /* Want to start with kernel preemption disabled. */
2571 task_thread_info(p)->preempt_count = 1;
2573 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2579 * wake_up_new_task - wake up a newly created task for the first time.
2581 * This function will do some initial scheduler statistics housekeeping
2582 * that must be done for every newly created context, then puts the task
2583 * on the runqueue and wakes it.
2585 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2587 unsigned long flags;
2590 rq = task_rq_lock(p, &flags);
2591 BUG_ON(p->state != TASK_RUNNING);
2592 update_rq_clock(rq);
2594 p->prio = effective_prio(p);
2596 if (!p->sched_class->task_new || !current->se.on_rq) {
2597 activate_task(rq, p, 0);
2600 * Let the scheduling class do new task startup
2601 * management (if any):
2603 p->sched_class->task_new(rq, p);
2606 trace_sched_wakeup_new(rq, p, 1);
2607 check_preempt_curr(rq, p, 0);
2609 if (p->sched_class->task_wake_up)
2610 p->sched_class->task_wake_up(rq, p);
2612 task_rq_unlock(rq, &flags);
2615 #ifdef CONFIG_PREEMPT_NOTIFIERS
2618 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2619 * @notifier: notifier struct to register
2621 void preempt_notifier_register(struct preempt_notifier *notifier)
2623 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2625 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2628 * preempt_notifier_unregister - no longer interested in preemption notifications
2629 * @notifier: notifier struct to unregister
2631 * This is safe to call from within a preemption notifier.
2633 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2635 hlist_del(¬ifier->link);
2637 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2639 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2641 struct preempt_notifier *notifier;
2642 struct hlist_node *node;
2644 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2645 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2649 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2650 struct task_struct *next)
2652 struct preempt_notifier *notifier;
2653 struct hlist_node *node;
2655 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2656 notifier->ops->sched_out(notifier, next);
2659 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2661 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2666 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2667 struct task_struct *next)
2671 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2674 * prepare_task_switch - prepare to switch tasks
2675 * @rq: the runqueue preparing to switch
2676 * @prev: the current task that is being switched out
2677 * @next: the task we are going to switch to.
2679 * This is called with the rq lock held and interrupts off. It must
2680 * be paired with a subsequent finish_task_switch after the context
2683 * prepare_task_switch sets up locking and calls architecture specific
2687 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2688 struct task_struct *next)
2690 fire_sched_out_preempt_notifiers(prev, next);
2691 prepare_lock_switch(rq, next);
2692 prepare_arch_switch(next);
2696 * finish_task_switch - clean up after a task-switch
2697 * @rq: runqueue associated with task-switch
2698 * @prev: the thread we just switched away from.
2700 * finish_task_switch must be called after the context switch, paired
2701 * with a prepare_task_switch call before the context switch.
2702 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2703 * and do any other architecture-specific cleanup actions.
2705 * Note that we may have delayed dropping an mm in context_switch(). If
2706 * so, we finish that here outside of the runqueue lock. (Doing it
2707 * with the lock held can cause deadlocks; see schedule() for
2710 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2711 __releases(rq->lock)
2713 struct mm_struct *mm = rq->prev_mm;
2719 * A task struct has one reference for the use as "current".
2720 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2721 * schedule one last time. The schedule call will never return, and
2722 * the scheduled task must drop that reference.
2723 * The test for TASK_DEAD must occur while the runqueue locks are
2724 * still held, otherwise prev could be scheduled on another cpu, die
2725 * there before we look at prev->state, and then the reference would
2727 * Manfred Spraul <manfred@colorfullife.com>
2729 prev_state = prev->state;
2730 finish_arch_switch(prev);
2731 perf_counter_task_sched_in(current, cpu_of(rq));
2732 finish_lock_switch(rq, prev);
2734 fire_sched_in_preempt_notifiers(current);
2737 if (unlikely(prev_state == TASK_DEAD)) {
2739 * Remove function-return probe instances associated with this
2740 * task and put them back on the free list.
2742 kprobe_flush_task(prev);
2743 put_task_struct(prev);
2749 /* assumes rq->lock is held */
2750 static inline void pre_schedule(struct rq *rq, struct task_struct *prev)
2752 if (prev->sched_class->pre_schedule)
2753 prev->sched_class->pre_schedule(rq, prev);
2756 /* rq->lock is NOT held, but preemption is disabled */
2757 static inline void post_schedule(struct rq *rq)
2759 if (rq->post_schedule) {
2760 unsigned long flags;
2762 spin_lock_irqsave(&rq->lock, flags);
2763 if (rq->curr->sched_class->post_schedule)
2764 rq->curr->sched_class->post_schedule(rq);
2765 spin_unlock_irqrestore(&rq->lock, flags);
2767 rq->post_schedule = 0;
2773 static inline void pre_schedule(struct rq *rq, struct task_struct *p)
2777 static inline void post_schedule(struct rq *rq)
2784 * schedule_tail - first thing a freshly forked thread must call.
2785 * @prev: the thread we just switched away from.
2787 asmlinkage void schedule_tail(struct task_struct *prev)
2788 __releases(rq->lock)
2790 struct rq *rq = this_rq();
2792 finish_task_switch(rq, prev);
2795 * FIXME: do we need to worry about rq being invalidated by the
2800 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2801 /* In this case, finish_task_switch does not reenable preemption */
2804 if (current->set_child_tid)
2805 put_user(task_pid_vnr(current), current->set_child_tid);
2809 * context_switch - switch to the new MM and the new
2810 * thread's register state.
2813 context_switch(struct rq *rq, struct task_struct *prev,
2814 struct task_struct *next)
2816 struct mm_struct *mm, *oldmm;
2818 prepare_task_switch(rq, prev, next);
2819 trace_sched_switch(rq, prev, next);
2821 oldmm = prev->active_mm;
2823 * For paravirt, this is coupled with an exit in switch_to to
2824 * combine the page table reload and the switch backend into
2827 arch_start_context_switch(prev);
2829 if (unlikely(!mm)) {
2830 next->active_mm = oldmm;
2831 atomic_inc(&oldmm->mm_count);
2832 enter_lazy_tlb(oldmm, next);
2834 switch_mm(oldmm, mm, next);
2836 if (unlikely(!prev->mm)) {
2837 prev->active_mm = NULL;
2838 rq->prev_mm = oldmm;
2841 * Since the runqueue lock will be released by the next
2842 * task (which is an invalid locking op but in the case
2843 * of the scheduler it's an obvious special-case), so we
2844 * do an early lockdep release here:
2846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2847 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2850 /* Here we just switch the register state and the stack. */
2851 switch_to(prev, next, prev);
2855 * this_rq must be evaluated again because prev may have moved
2856 * CPUs since it called schedule(), thus the 'rq' on its stack
2857 * frame will be invalid.
2859 finish_task_switch(this_rq(), prev);
2863 * nr_running, nr_uninterruptible and nr_context_switches:
2865 * externally visible scheduler statistics: current number of runnable
2866 * threads, current number of uninterruptible-sleeping threads, total
2867 * number of context switches performed since bootup.
2869 unsigned long nr_running(void)
2871 unsigned long i, sum = 0;
2873 for_each_online_cpu(i)
2874 sum += cpu_rq(i)->nr_running;
2879 unsigned long nr_uninterruptible(void)
2881 unsigned long i, sum = 0;
2883 for_each_possible_cpu(i)
2884 sum += cpu_rq(i)->nr_uninterruptible;
2887 * Since we read the counters lockless, it might be slightly
2888 * inaccurate. Do not allow it to go below zero though:
2890 if (unlikely((long)sum < 0))
2896 unsigned long long nr_context_switches(void)
2899 unsigned long long sum = 0;
2901 for_each_possible_cpu(i)
2902 sum += cpu_rq(i)->nr_switches;
2907 unsigned long nr_iowait(void)
2909 unsigned long i, sum = 0;
2911 for_each_possible_cpu(i)
2912 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2917 /* Variables and functions for calc_load */
2918 static atomic_long_t calc_load_tasks;
2919 static unsigned long calc_load_update;
2920 unsigned long avenrun[3];
2921 EXPORT_SYMBOL(avenrun);
2924 * get_avenrun - get the load average array
2925 * @loads: pointer to dest load array
2926 * @offset: offset to add
2927 * @shift: shift count to shift the result left
2929 * These values are estimates at best, so no need for locking.
2931 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2933 loads[0] = (avenrun[0] + offset) << shift;
2934 loads[1] = (avenrun[1] + offset) << shift;
2935 loads[2] = (avenrun[2] + offset) << shift;
2938 static unsigned long
2939 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2942 load += active * (FIXED_1 - exp);
2943 return load >> FSHIFT;
2947 * calc_load - update the avenrun load estimates 10 ticks after the
2948 * CPUs have updated calc_load_tasks.
2950 void calc_global_load(void)
2952 unsigned long upd = calc_load_update + 10;
2955 if (time_before(jiffies, upd))
2958 active = atomic_long_read(&calc_load_tasks);
2959 active = active > 0 ? active * FIXED_1 : 0;
2961 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2962 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2963 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2965 calc_load_update += LOAD_FREQ;
2969 * Either called from update_cpu_load() or from a cpu going idle
2971 static void calc_load_account_active(struct rq *this_rq)
2973 long nr_active, delta;
2975 nr_active = this_rq->nr_running;
2976 nr_active += (long) this_rq->nr_uninterruptible;
2978 if (nr_active != this_rq->calc_load_active) {
2979 delta = nr_active - this_rq->calc_load_active;
2980 this_rq->calc_load_active = nr_active;
2981 atomic_long_add(delta, &calc_load_tasks);
2986 * Externally visible per-cpu scheduler statistics:
2987 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2989 u64 cpu_nr_migrations(int cpu)
2991 return cpu_rq(cpu)->nr_migrations_in;
2995 * Update rq->cpu_load[] statistics. This function is usually called every
2996 * scheduler tick (TICK_NSEC).
2998 static void update_cpu_load(struct rq *this_rq)
3000 unsigned long this_load = this_rq->load.weight;
3003 this_rq->nr_load_updates++;
3005 /* Update our load: */
3006 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
3007 unsigned long old_load, new_load;
3009 /* scale is effectively 1 << i now, and >> i divides by scale */
3011 old_load = this_rq->cpu_load[i];
3012 new_load = this_load;
3014 * Round up the averaging division if load is increasing. This
3015 * prevents us from getting stuck on 9 if the load is 10, for
3018 if (new_load > old_load)
3019 new_load += scale-1;
3020 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3023 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
3024 this_rq->calc_load_update += LOAD_FREQ;
3025 calc_load_account_active(this_rq);
3032 * double_rq_lock - safely lock two runqueues
3034 * Note this does not disable interrupts like task_rq_lock,
3035 * you need to do so manually before calling.
3037 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3038 __acquires(rq1->lock)
3039 __acquires(rq2->lock)
3041 BUG_ON(!irqs_disabled());
3043 spin_lock(&rq1->lock);
3044 __acquire(rq2->lock); /* Fake it out ;) */
3047 spin_lock(&rq1->lock);
3048 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
3050 spin_lock(&rq2->lock);
3051 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
3054 update_rq_clock(rq1);
3055 update_rq_clock(rq2);
3059 * double_rq_unlock - safely unlock two runqueues
3061 * Note this does not restore interrupts like task_rq_unlock,
3062 * you need to do so manually after calling.
3064 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3065 __releases(rq1->lock)
3066 __releases(rq2->lock)
3068 spin_unlock(&rq1->lock);
3070 spin_unlock(&rq2->lock);
3072 __release(rq2->lock);
3076 * If dest_cpu is allowed for this process, migrate the task to it.
3077 * This is accomplished by forcing the cpu_allowed mask to only
3078 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3079 * the cpu_allowed mask is restored.
3081 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3083 struct migration_req req;
3084 unsigned long flags;
3087 rq = task_rq_lock(p, &flags);
3088 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3089 || unlikely(!cpu_active(dest_cpu)))
3092 /* force the process onto the specified CPU */
3093 if (migrate_task(p, dest_cpu, &req)) {
3094 /* Need to wait for migration thread (might exit: take ref). */
3095 struct task_struct *mt = rq->migration_thread;
3097 get_task_struct(mt);
3098 task_rq_unlock(rq, &flags);
3099 wake_up_process(mt);
3100 put_task_struct(mt);
3101 wait_for_completion(&req.done);
3106 task_rq_unlock(rq, &flags);
3110 * sched_exec - execve() is a valuable balancing opportunity, because at
3111 * this point the task has the smallest effective memory and cache footprint.
3113 void sched_exec(void)
3115 int new_cpu, this_cpu = get_cpu();
3116 new_cpu = current->sched_class->select_task_rq(current, SD_BALANCE_EXEC, 0);
3118 if (new_cpu != this_cpu)
3119 sched_migrate_task(current, new_cpu);
3123 * pull_task - move a task from a remote runqueue to the local runqueue.
3124 * Both runqueues must be locked.
3126 static void pull_task(struct rq *src_rq, struct task_struct *p,
3127 struct rq *this_rq, int this_cpu)
3129 deactivate_task(src_rq, p, 0);
3130 set_task_cpu(p, this_cpu);
3131 activate_task(this_rq, p, 0);
3133 * Note that idle threads have a prio of MAX_PRIO, for this test
3134 * to be always true for them.
3136 check_preempt_curr(this_rq, p, 0);
3140 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3143 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3144 struct sched_domain *sd, enum cpu_idle_type idle,
3147 int tsk_cache_hot = 0;
3149 * We do not migrate tasks that are:
3150 * 1) running (obviously), or
3151 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3152 * 3) are cache-hot on their current CPU.
3154 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3155 schedstat_inc(p, se.nr_failed_migrations_affine);
3160 if (task_running(rq, p)) {
3161 schedstat_inc(p, se.nr_failed_migrations_running);
3166 * Aggressive migration if:
3167 * 1) task is cache cold, or
3168 * 2) too many balance attempts have failed.
3171 tsk_cache_hot = task_hot(p, rq->clock, sd);
3172 if (!tsk_cache_hot ||
3173 sd->nr_balance_failed > sd->cache_nice_tries) {
3174 #ifdef CONFIG_SCHEDSTATS
3175 if (tsk_cache_hot) {
3176 schedstat_inc(sd, lb_hot_gained[idle]);
3177 schedstat_inc(p, se.nr_forced_migrations);
3183 if (tsk_cache_hot) {
3184 schedstat_inc(p, se.nr_failed_migrations_hot);
3190 static unsigned long
3191 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3192 unsigned long max_load_move, struct sched_domain *sd,
3193 enum cpu_idle_type idle, int *all_pinned,
3194 int *this_best_prio, struct rq_iterator *iterator)
3196 int loops = 0, pulled = 0, pinned = 0;
3197 struct task_struct *p;
3198 long rem_load_move = max_load_move;
3200 if (max_load_move == 0)
3206 * Start the load-balancing iterator:
3208 p = iterator->start(iterator->arg);
3210 if (!p || loops++ > sysctl_sched_nr_migrate)
3213 if ((p->se.load.weight >> 1) > rem_load_move ||
3214 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3215 p = iterator->next(iterator->arg);
3219 pull_task(busiest, p, this_rq, this_cpu);
3221 rem_load_move -= p->se.load.weight;
3223 #ifdef CONFIG_PREEMPT
3225 * NEWIDLE balancing is a source of latency, so preemptible kernels
3226 * will stop after the first task is pulled to minimize the critical
3229 if (idle == CPU_NEWLY_IDLE)
3234 * We only want to steal up to the prescribed amount of weighted load.
3236 if (rem_load_move > 0) {
3237 if (p->prio < *this_best_prio)
3238 *this_best_prio = p->prio;
3239 p = iterator->next(iterator->arg);
3244 * Right now, this is one of only two places pull_task() is called,
3245 * so we can safely collect pull_task() stats here rather than
3246 * inside pull_task().
3248 schedstat_add(sd, lb_gained[idle], pulled);
3251 *all_pinned = pinned;
3253 return max_load_move - rem_load_move;
3257 * move_tasks tries to move up to max_load_move weighted load from busiest to
3258 * this_rq, as part of a balancing operation within domain "sd".
3259 * Returns 1 if successful and 0 otherwise.
3261 * Called with both runqueues locked.
3263 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3264 unsigned long max_load_move,
3265 struct sched_domain *sd, enum cpu_idle_type idle,
3268 const struct sched_class *class = sched_class_highest;
3269 unsigned long total_load_moved = 0;
3270 int this_best_prio = this_rq->curr->prio;
3274 class->load_balance(this_rq, this_cpu, busiest,
3275 max_load_move - total_load_moved,
3276 sd, idle, all_pinned, &this_best_prio);
3277 class = class->next;
3279 #ifdef CONFIG_PREEMPT
3281 * NEWIDLE balancing is a source of latency, so preemptible
3282 * kernels will stop after the first task is pulled to minimize
3283 * the critical section.
3285 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3288 } while (class && max_load_move > total_load_moved);
3290 return total_load_moved > 0;
3294 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3295 struct sched_domain *sd, enum cpu_idle_type idle,
3296 struct rq_iterator *iterator)
3298 struct task_struct *p = iterator->start(iterator->arg);
3302 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3303 pull_task(busiest, p, this_rq, this_cpu);
3305 * Right now, this is only the second place pull_task()
3306 * is called, so we can safely collect pull_task()
3307 * stats here rather than inside pull_task().
3309 schedstat_inc(sd, lb_gained[idle]);
3313 p = iterator->next(iterator->arg);
3320 * move_one_task tries to move exactly one task from busiest to this_rq, as
3321 * part of active balancing operations within "domain".
3322 * Returns 1 if successful and 0 otherwise.
3324 * Called with both runqueues locked.
3326 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3327 struct sched_domain *sd, enum cpu_idle_type idle)
3329 const struct sched_class *class;
3331 for_each_class(class) {
3332 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3338 /********** Helpers for find_busiest_group ************************/
3340 * sd_lb_stats - Structure to store the statistics of a sched_domain
3341 * during load balancing.
3343 struct sd_lb_stats {
3344 struct sched_group *busiest; /* Busiest group in this sd */
3345 struct sched_group *this; /* Local group in this sd */
3346 unsigned long total_load; /* Total load of all groups in sd */
3347 unsigned long total_pwr; /* Total power of all groups in sd */
3348 unsigned long avg_load; /* Average load across all groups in sd */
3350 /** Statistics of this group */
3351 unsigned long this_load;
3352 unsigned long this_load_per_task;
3353 unsigned long this_nr_running;
3355 /* Statistics of the busiest group */
3356 unsigned long max_load;
3357 unsigned long busiest_load_per_task;
3358 unsigned long busiest_nr_running;
3360 int group_imb; /* Is there imbalance in this sd */
3361 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3362 int power_savings_balance; /* Is powersave balance needed for this sd */
3363 struct sched_group *group_min; /* Least loaded group in sd */
3364 struct sched_group *group_leader; /* Group which relieves group_min */
3365 unsigned long min_load_per_task; /* load_per_task in group_min */
3366 unsigned long leader_nr_running; /* Nr running of group_leader */
3367 unsigned long min_nr_running; /* Nr running of group_min */
3372 * sg_lb_stats - stats of a sched_group required for load_balancing
3374 struct sg_lb_stats {
3375 unsigned long avg_load; /*Avg load across the CPUs of the group */
3376 unsigned long group_load; /* Total load over the CPUs of the group */
3377 unsigned long sum_nr_running; /* Nr tasks running in the group */
3378 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3379 unsigned long group_capacity;
3380 int group_imb; /* Is there an imbalance in the group ? */
3384 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3385 * @group: The group whose first cpu is to be returned.
3387 static inline unsigned int group_first_cpu(struct sched_group *group)
3389 return cpumask_first(sched_group_cpus(group));
3393 * get_sd_load_idx - Obtain the load index for a given sched domain.
3394 * @sd: The sched_domain whose load_idx is to be obtained.
3395 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3397 static inline int get_sd_load_idx(struct sched_domain *sd,
3398 enum cpu_idle_type idle)
3404 load_idx = sd->busy_idx;
3407 case CPU_NEWLY_IDLE:
3408 load_idx = sd->newidle_idx;
3411 load_idx = sd->idle_idx;
3419 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3421 * init_sd_power_savings_stats - Initialize power savings statistics for
3422 * the given sched_domain, during load balancing.
3424 * @sd: Sched domain whose power-savings statistics are to be initialized.
3425 * @sds: Variable containing the statistics for sd.
3426 * @idle: Idle status of the CPU at which we're performing load-balancing.
3428 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3429 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3432 * Busy processors will not participate in power savings
3435 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3436 sds->power_savings_balance = 0;
3438 sds->power_savings_balance = 1;
3439 sds->min_nr_running = ULONG_MAX;
3440 sds->leader_nr_running = 0;
3445 * update_sd_power_savings_stats - Update the power saving stats for a
3446 * sched_domain while performing load balancing.
3448 * @group: sched_group belonging to the sched_domain under consideration.
3449 * @sds: Variable containing the statistics of the sched_domain
3450 * @local_group: Does group contain the CPU for which we're performing
3452 * @sgs: Variable containing the statistics of the group.
3454 static inline void update_sd_power_savings_stats(struct sched_group *group,
3455 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3458 if (!sds->power_savings_balance)
3462 * If the local group is idle or completely loaded
3463 * no need to do power savings balance at this domain
3465 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3466 !sds->this_nr_running))
3467 sds->power_savings_balance = 0;
3470 * If a group is already running at full capacity or idle,
3471 * don't include that group in power savings calculations
3473 if (!sds->power_savings_balance ||
3474 sgs->sum_nr_running >= sgs->group_capacity ||
3475 !sgs->sum_nr_running)
3479 * Calculate the group which has the least non-idle load.
3480 * This is the group from where we need to pick up the load
3483 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3484 (sgs->sum_nr_running == sds->min_nr_running &&
3485 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3486 sds->group_min = group;
3487 sds->min_nr_running = sgs->sum_nr_running;
3488 sds->min_load_per_task = sgs->sum_weighted_load /
3489 sgs->sum_nr_running;
3493 * Calculate the group which is almost near its
3494 * capacity but still has some space to pick up some load
3495 * from other group and save more power
3497 if (sgs->sum_nr_running + 1 > sgs->group_capacity)
3500 if (sgs->sum_nr_running > sds->leader_nr_running ||
3501 (sgs->sum_nr_running == sds->leader_nr_running &&
3502 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3503 sds->group_leader = group;
3504 sds->leader_nr_running = sgs->sum_nr_running;
3509 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3510 * @sds: Variable containing the statistics of the sched_domain
3511 * under consideration.
3512 * @this_cpu: Cpu at which we're currently performing load-balancing.
3513 * @imbalance: Variable to store the imbalance.
3516 * Check if we have potential to perform some power-savings balance.
3517 * If yes, set the busiest group to be the least loaded group in the
3518 * sched_domain, so that it's CPUs can be put to idle.
3520 * Returns 1 if there is potential to perform power-savings balance.
3523 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3524 int this_cpu, unsigned long *imbalance)
3526 if (!sds->power_savings_balance)
3529 if (sds->this != sds->group_leader ||
3530 sds->group_leader == sds->group_min)
3533 *imbalance = sds->min_load_per_task;
3534 sds->busiest = sds->group_min;
3536 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3537 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3538 group_first_cpu(sds->group_leader);
3544 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3545 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3546 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3551 static inline void update_sd_power_savings_stats(struct sched_group *group,
3552 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3557 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3558 int this_cpu, unsigned long *imbalance)
3562 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3564 unsigned long __weak arch_scale_smt_power(struct sched_domain *sd, int cpu)
3566 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3567 unsigned long smt_gain = sd->smt_gain;
3574 unsigned long scale_rt_power(int cpu)
3576 struct rq *rq = cpu_rq(cpu);
3577 u64 total, available;
3579 sched_avg_update(rq);
3581 total = sched_avg_period() + (rq->clock - rq->age_stamp);
3582 available = total - rq->rt_avg;
3584 if (unlikely((s64)total < SCHED_LOAD_SCALE))
3585 total = SCHED_LOAD_SCALE;
3587 total >>= SCHED_LOAD_SHIFT;
3589 return div_u64(available, total);
3592 static void update_cpu_power(struct sched_domain *sd, int cpu)
3594 unsigned long weight = cpumask_weight(sched_domain_span(sd));
3595 unsigned long power = SCHED_LOAD_SCALE;
3596 struct sched_group *sdg = sd->groups;
3598 /* here we could scale based on cpufreq */
3600 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
3601 power *= arch_scale_smt_power(sd, cpu);
3602 power >>= SCHED_LOAD_SHIFT;
3605 power *= scale_rt_power(cpu);
3606 power >>= SCHED_LOAD_SHIFT;
3611 sdg->cpu_power = power;
3614 static void update_group_power(struct sched_domain *sd, int cpu)
3616 struct sched_domain *child = sd->child;
3617 struct sched_group *group, *sdg = sd->groups;
3618 unsigned long power;
3621 update_cpu_power(sd, cpu);
3627 group = child->groups;
3629 power += group->cpu_power;
3630 group = group->next;
3631 } while (group != child->groups);
3633 sdg->cpu_power = power;
3637 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3638 * @group: sched_group whose statistics are to be updated.
3639 * @this_cpu: Cpu for which load balance is currently performed.
3640 * @idle: Idle status of this_cpu
3641 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3642 * @sd_idle: Idle status of the sched_domain containing group.
3643 * @local_group: Does group contain this_cpu.
3644 * @cpus: Set of cpus considered for load balancing.
3645 * @balance: Should we balance.
3646 * @sgs: variable to hold the statistics for this group.
3648 static inline void update_sg_lb_stats(struct sched_domain *sd,
3649 struct sched_group *group, int this_cpu,
3650 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3651 int local_group, const struct cpumask *cpus,
3652 int *balance, struct sg_lb_stats *sgs)
3654 unsigned long load, max_cpu_load, min_cpu_load;
3656 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3657 unsigned long sum_avg_load_per_task;
3658 unsigned long avg_load_per_task;
3661 balance_cpu = group_first_cpu(group);
3662 if (balance_cpu == this_cpu)
3663 update_group_power(sd, this_cpu);
3666 /* Tally up the load of all CPUs in the group */
3667 sum_avg_load_per_task = avg_load_per_task = 0;
3669 min_cpu_load = ~0UL;
3671 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3672 struct rq *rq = cpu_rq(i);
3674 if (*sd_idle && rq->nr_running)
3677 /* Bias balancing toward cpus of our domain */
3679 if (idle_cpu(i) && !first_idle_cpu) {
3684 load = target_load(i, load_idx);
3686 load = source_load(i, load_idx);
3687 if (load > max_cpu_load)
3688 max_cpu_load = load;
3689 if (min_cpu_load > load)
3690 min_cpu_load = load;
3693 sgs->group_load += load;
3694 sgs->sum_nr_running += rq->nr_running;
3695 sgs->sum_weighted_load += weighted_cpuload(i);
3697 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3701 * First idle cpu or the first cpu(busiest) in this sched group
3702 * is eligible for doing load balancing at this and above
3703 * domains. In the newly idle case, we will allow all the cpu's
3704 * to do the newly idle load balance.
3706 if (idle != CPU_NEWLY_IDLE && local_group &&
3707 balance_cpu != this_cpu && balance) {
3712 /* Adjust by relative CPU power of the group */
3713 sgs->avg_load = (sgs->group_load * SCHED_LOAD_SCALE) / group->cpu_power;
3717 * Consider the group unbalanced when the imbalance is larger
3718 * than the average weight of two tasks.
3720 * APZ: with cgroup the avg task weight can vary wildly and
3721 * might not be a suitable number - should we keep a
3722 * normalized nr_running number somewhere that negates
3725 avg_load_per_task = (sum_avg_load_per_task * SCHED_LOAD_SCALE) /
3728 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3731 sgs->group_capacity =
3732 DIV_ROUND_CLOSEST(group->cpu_power, SCHED_LOAD_SCALE);
3736 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3737 * @sd: sched_domain whose statistics are to be updated.
3738 * @this_cpu: Cpu for which load balance is currently performed.
3739 * @idle: Idle status of this_cpu
3740 * @sd_idle: Idle status of the sched_domain containing group.
3741 * @cpus: Set of cpus considered for load balancing.
3742 * @balance: Should we balance.
3743 * @sds: variable to hold the statistics for this sched_domain.
3745 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3746 enum cpu_idle_type idle, int *sd_idle,
3747 const struct cpumask *cpus, int *balance,
3748 struct sd_lb_stats *sds)
3750 struct sched_domain *child = sd->child;
3751 struct sched_group *group = sd->groups;
3752 struct sg_lb_stats sgs;
3753 int load_idx, prefer_sibling = 0;
3755 if (child && child->flags & SD_PREFER_SIBLING)
3758 init_sd_power_savings_stats(sd, sds, idle);
3759 load_idx = get_sd_load_idx(sd, idle);
3764 local_group = cpumask_test_cpu(this_cpu,
3765 sched_group_cpus(group));
3766 memset(&sgs, 0, sizeof(sgs));
3767 update_sg_lb_stats(sd, group, this_cpu, idle, load_idx, sd_idle,
3768 local_group, cpus, balance, &sgs);
3770 if (local_group && balance && !(*balance))
3773 sds->total_load += sgs.group_load;
3774 sds->total_pwr += group->cpu_power;
3777 * In case the child domain prefers tasks go to siblings
3778 * first, lower the group capacity to one so that we'll try
3779 * and move all the excess tasks away.
3782 sgs.group_capacity = min(sgs.group_capacity, 1UL);
3785 sds->this_load = sgs.avg_load;
3787 sds->this_nr_running = sgs.sum_nr_running;
3788 sds->this_load_per_task = sgs.sum_weighted_load;
3789 } else if (sgs.avg_load > sds->max_load &&
3790 (sgs.sum_nr_running > sgs.group_capacity ||
3792 sds->max_load = sgs.avg_load;
3793 sds->busiest = group;
3794 sds->busiest_nr_running = sgs.sum_nr_running;
3795 sds->busiest_load_per_task = sgs.sum_weighted_load;
3796 sds->group_imb = sgs.group_imb;
3799 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3800 group = group->next;
3801 } while (group != sd->groups);
3805 * fix_small_imbalance - Calculate the minor imbalance that exists
3806 * amongst the groups of a sched_domain, during
3808 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3809 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3810 * @imbalance: Variable to store the imbalance.
3812 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3813 int this_cpu, unsigned long *imbalance)
3815 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3816 unsigned int imbn = 2;
3818 if (sds->this_nr_running) {
3819 sds->this_load_per_task /= sds->this_nr_running;
3820 if (sds->busiest_load_per_task >
3821 sds->this_load_per_task)
3824 sds->this_load_per_task =
3825 cpu_avg_load_per_task(this_cpu);
3827 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3828 sds->busiest_load_per_task * imbn) {
3829 *imbalance = sds->busiest_load_per_task;
3834 * OK, we don't have enough imbalance to justify moving tasks,
3835 * however we may be able to increase total CPU power used by
3839 pwr_now += sds->busiest->cpu_power *
3840 min(sds->busiest_load_per_task, sds->max_load);
3841 pwr_now += sds->this->cpu_power *
3842 min(sds->this_load_per_task, sds->this_load);
3843 pwr_now /= SCHED_LOAD_SCALE;
3845 /* Amount of load we'd subtract */
3846 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3847 sds->busiest->cpu_power;
3848 if (sds->max_load > tmp)
3849 pwr_move += sds->busiest->cpu_power *
3850 min(sds->busiest_load_per_task, sds->max_load - tmp);
3852 /* Amount of load we'd add */
3853 if (sds->max_load * sds->busiest->cpu_power <
3854 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3855 tmp = (sds->max_load * sds->busiest->cpu_power) /
3856 sds->this->cpu_power;
3858 tmp = (sds->busiest_load_per_task * SCHED_LOAD_SCALE) /
3859 sds->this->cpu_power;
3860 pwr_move += sds->this->cpu_power *
3861 min(sds->this_load_per_task, sds->this_load + tmp);
3862 pwr_move /= SCHED_LOAD_SCALE;
3864 /* Move if we gain throughput */
3865 if (pwr_move > pwr_now)
3866 *imbalance = sds->busiest_load_per_task;
3870 * calculate_imbalance - Calculate the amount of imbalance present within the
3871 * groups of a given sched_domain during load balance.
3872 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3873 * @this_cpu: Cpu for which currently load balance is being performed.
3874 * @imbalance: The variable to store the imbalance.
3876 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3877 unsigned long *imbalance)
3879 unsigned long max_pull;
3881 * In the presence of smp nice balancing, certain scenarios can have
3882 * max load less than avg load(as we skip the groups at or below
3883 * its cpu_power, while calculating max_load..)
3885 if (sds->max_load < sds->avg_load) {
3887 return fix_small_imbalance(sds, this_cpu, imbalance);
3890 /* Don't want to pull so many tasks that a group would go idle */
3891 max_pull = min(sds->max_load - sds->avg_load,
3892 sds->max_load - sds->busiest_load_per_task);
3894 /* How much load to actually move to equalise the imbalance */
3895 *imbalance = min(max_pull * sds->busiest->cpu_power,
3896 (sds->avg_load - sds->this_load) * sds->this->cpu_power)
3900 * if *imbalance is less than the average load per runnable task
3901 * there is no gaurantee that any tasks will be moved so we'll have
3902 * a think about bumping its value to force at least one task to be
3905 if (*imbalance < sds->busiest_load_per_task)
3906 return fix_small_imbalance(sds, this_cpu, imbalance);
3909 /******* find_busiest_group() helpers end here *********************/
3912 * find_busiest_group - Returns the busiest group within the sched_domain
3913 * if there is an imbalance. If there isn't an imbalance, and
3914 * the user has opted for power-savings, it returns a group whose
3915 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3916 * such a group exists.
3918 * Also calculates the amount of weighted load which should be moved
3919 * to restore balance.
3921 * @sd: The sched_domain whose busiest group is to be returned.
3922 * @this_cpu: The cpu for which load balancing is currently being performed.
3923 * @imbalance: Variable which stores amount of weighted load which should
3924 * be moved to restore balance/put a group to idle.
3925 * @idle: The idle status of this_cpu.
3926 * @sd_idle: The idleness of sd
3927 * @cpus: The set of CPUs under consideration for load-balancing.
3928 * @balance: Pointer to a variable indicating if this_cpu
3929 * is the appropriate cpu to perform load balancing at this_level.
3931 * Returns: - the busiest group if imbalance exists.
3932 * - If no imbalance and user has opted for power-savings balance,
3933 * return the least loaded group whose CPUs can be
3934 * put to idle by rebalancing its tasks onto our group.
3936 static struct sched_group *
3937 find_busiest_group(struct sched_domain *sd, int this_cpu,
3938 unsigned long *imbalance, enum cpu_idle_type idle,
3939 int *sd_idle, const struct cpumask *cpus, int *balance)
3941 struct sd_lb_stats sds;
3943 memset(&sds, 0, sizeof(sds));
3946 * Compute the various statistics relavent for load balancing at
3949 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3952 /* Cases where imbalance does not exist from POV of this_cpu */
3953 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3955 * 2) There is no busy sibling group to pull from.
3956 * 3) This group is the busiest group.
3957 * 4) This group is more busy than the avg busieness at this
3959 * 5) The imbalance is within the specified limit.
3960 * 6) Any rebalance would lead to ping-pong
3962 if (balance && !(*balance))
3965 if (!sds.busiest || sds.busiest_nr_running == 0)
3968 if (sds.this_load >= sds.max_load)
3971 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3973 if (sds.this_load >= sds.avg_load)
3976 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3979 sds.busiest_load_per_task /= sds.busiest_nr_running;
3981 sds.busiest_load_per_task =
3982 min(sds.busiest_load_per_task, sds.avg_load);
3985 * We're trying to get all the cpus to the average_load, so we don't
3986 * want to push ourselves above the average load, nor do we wish to
3987 * reduce the max loaded cpu below the average load, as either of these
3988 * actions would just result in more rebalancing later, and ping-pong
3989 * tasks around. Thus we look for the minimum possible imbalance.
3990 * Negative imbalances (*we* are more loaded than anyone else) will
3991 * be counted as no imbalance for these purposes -- we can't fix that
3992 * by pulling tasks to us. Be careful of negative numbers as they'll
3993 * appear as very large values with unsigned longs.
3995 if (sds.max_load <= sds.busiest_load_per_task)
3998 /* Looks like there is an imbalance. Compute it */
3999 calculate_imbalance(&sds, this_cpu, imbalance);
4004 * There is no obvious imbalance. But check if we can do some balancing
4007 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
4014 static struct sched_group *group_of(int cpu)
4016 struct sched_domain *sd = rcu_dereference(cpu_rq(cpu)->sd);
4024 static unsigned long power_of(int cpu)
4026 struct sched_group *group = group_of(cpu);
4029 return SCHED_LOAD_SCALE;
4031 return group->cpu_power;
4035 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4038 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
4039 unsigned long imbalance, const struct cpumask *cpus)
4041 struct rq *busiest = NULL, *rq;
4042 unsigned long max_load = 0;
4045 for_each_cpu(i, sched_group_cpus(group)) {
4046 unsigned long power = power_of(i);
4047 unsigned long capacity = DIV_ROUND_CLOSEST(power, SCHED_LOAD_SCALE);
4050 if (!cpumask_test_cpu(i, cpus))
4054 wl = weighted_cpuload(i) * SCHED_LOAD_SCALE;
4057 if (capacity && rq->nr_running == 1 && wl > imbalance)
4060 if (wl > max_load) {
4070 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4071 * so long as it is large enough.
4073 #define MAX_PINNED_INTERVAL 512
4075 /* Working cpumask for load_balance and load_balance_newidle. */
4076 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
4079 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4080 * tasks if there is an imbalance.
4082 static int load_balance(int this_cpu, struct rq *this_rq,
4083 struct sched_domain *sd, enum cpu_idle_type idle,
4086 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
4087 struct sched_group *group;
4088 unsigned long imbalance;
4090 unsigned long flags;
4091 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4093 cpumask_setall(cpus);
4096 * When power savings policy is enabled for the parent domain, idle
4097 * sibling can pick up load irrespective of busy siblings. In this case,
4098 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4099 * portraying it as CPU_NOT_IDLE.
4101 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
4102 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4105 schedstat_inc(sd, lb_count[idle]);
4109 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
4116 schedstat_inc(sd, lb_nobusyg[idle]);
4120 busiest = find_busiest_queue(group, idle, imbalance, cpus);
4122 schedstat_inc(sd, lb_nobusyq[idle]);
4126 BUG_ON(busiest == this_rq);
4128 schedstat_add(sd, lb_imbalance[idle], imbalance);
4131 if (busiest->nr_running > 1) {
4133 * Attempt to move tasks. If find_busiest_group has found
4134 * an imbalance but busiest->nr_running <= 1, the group is
4135 * still unbalanced. ld_moved simply stays zero, so it is
4136 * correctly treated as an imbalance.
4138 local_irq_save(flags);
4139 double_rq_lock(this_rq, busiest);
4140 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4141 imbalance, sd, idle, &all_pinned);
4142 double_rq_unlock(this_rq, busiest);
4143 local_irq_restore(flags);
4146 * some other cpu did the load balance for us.
4148 if (ld_moved && this_cpu != smp_processor_id())
4149 resched_cpu(this_cpu);
4151 /* All tasks on this runqueue were pinned by CPU affinity */
4152 if (unlikely(all_pinned)) {
4153 cpumask_clear_cpu(cpu_of(busiest), cpus);
4154 if (!cpumask_empty(cpus))
4161 schedstat_inc(sd, lb_failed[idle]);
4162 sd->nr_balance_failed++;
4164 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
4166 spin_lock_irqsave(&busiest->lock, flags);
4168 /* don't kick the migration_thread, if the curr
4169 * task on busiest cpu can't be moved to this_cpu
4171 if (!cpumask_test_cpu(this_cpu,
4172 &busiest->curr->cpus_allowed)) {
4173 spin_unlock_irqrestore(&busiest->lock, flags);
4175 goto out_one_pinned;
4178 if (!busiest->active_balance) {
4179 busiest->active_balance = 1;
4180 busiest->push_cpu = this_cpu;
4183 spin_unlock_irqrestore(&busiest->lock, flags);
4185 wake_up_process(busiest->migration_thread);
4188 * We've kicked active balancing, reset the failure
4191 sd->nr_balance_failed = sd->cache_nice_tries+1;
4194 sd->nr_balance_failed = 0;
4196 if (likely(!active_balance)) {
4197 /* We were unbalanced, so reset the balancing interval */
4198 sd->balance_interval = sd->min_interval;
4201 * If we've begun active balancing, start to back off. This
4202 * case may not be covered by the all_pinned logic if there
4203 * is only 1 task on the busy runqueue (because we don't call
4206 if (sd->balance_interval < sd->max_interval)
4207 sd->balance_interval *= 2;
4210 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4211 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4217 schedstat_inc(sd, lb_balanced[idle]);
4219 sd->nr_balance_failed = 0;
4222 /* tune up the balancing interval */
4223 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4224 (sd->balance_interval < sd->max_interval))
4225 sd->balance_interval *= 2;
4227 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4228 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4239 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4240 * tasks if there is an imbalance.
4242 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4243 * this_rq is locked.
4246 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4248 struct sched_group *group;
4249 struct rq *busiest = NULL;
4250 unsigned long imbalance;
4254 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4256 cpumask_setall(cpus);
4259 * When power savings policy is enabled for the parent domain, idle
4260 * sibling can pick up load irrespective of busy siblings. In this case,
4261 * let the state of idle sibling percolate up as IDLE, instead of
4262 * portraying it as CPU_NOT_IDLE.
4264 if (sd->flags & SD_SHARE_CPUPOWER &&
4265 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4268 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4270 update_shares_locked(this_rq, sd);
4271 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4272 &sd_idle, cpus, NULL);
4274 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4278 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4280 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4284 BUG_ON(busiest == this_rq);
4286 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4289 if (busiest->nr_running > 1) {
4290 /* Attempt to move tasks */
4291 double_lock_balance(this_rq, busiest);
4292 /* this_rq->clock is already updated */
4293 update_rq_clock(busiest);
4294 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4295 imbalance, sd, CPU_NEWLY_IDLE,
4297 double_unlock_balance(this_rq, busiest);
4299 if (unlikely(all_pinned)) {
4300 cpumask_clear_cpu(cpu_of(busiest), cpus);
4301 if (!cpumask_empty(cpus))
4307 int active_balance = 0;
4309 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4310 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4311 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4314 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4317 if (sd->nr_balance_failed++ < 2)
4321 * The only task running in a non-idle cpu can be moved to this
4322 * cpu in an attempt to completely freeup the other CPU
4323 * package. The same method used to move task in load_balance()
4324 * have been extended for load_balance_newidle() to speedup
4325 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4327 * The package power saving logic comes from
4328 * find_busiest_group(). If there are no imbalance, then
4329 * f_b_g() will return NULL. However when sched_mc={1,2} then
4330 * f_b_g() will select a group from which a running task may be
4331 * pulled to this cpu in order to make the other package idle.
4332 * If there is no opportunity to make a package idle and if
4333 * there are no imbalance, then f_b_g() will return NULL and no
4334 * action will be taken in load_balance_newidle().
4336 * Under normal task pull operation due to imbalance, there
4337 * will be more than one task in the source run queue and
4338 * move_tasks() will succeed. ld_moved will be true and this
4339 * active balance code will not be triggered.
4342 /* Lock busiest in correct order while this_rq is held */
4343 double_lock_balance(this_rq, busiest);
4346 * don't kick the migration_thread, if the curr
4347 * task on busiest cpu can't be moved to this_cpu
4349 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4350 double_unlock_balance(this_rq, busiest);
4355 if (!busiest->active_balance) {
4356 busiest->active_balance = 1;
4357 busiest->push_cpu = this_cpu;
4361 double_unlock_balance(this_rq, busiest);
4363 * Should not call ttwu while holding a rq->lock
4365 spin_unlock(&this_rq->lock);
4367 wake_up_process(busiest->migration_thread);
4368 spin_lock(&this_rq->lock);
4371 sd->nr_balance_failed = 0;
4373 update_shares_locked(this_rq, sd);
4377 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4378 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4379 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4381 sd->nr_balance_failed = 0;
4387 * idle_balance is called by schedule() if this_cpu is about to become
4388 * idle. Attempts to pull tasks from other CPUs.
4390 static void idle_balance(int this_cpu, struct rq *this_rq)
4392 struct sched_domain *sd;
4393 int pulled_task = 0;
4394 unsigned long next_balance = jiffies + HZ;
4396 for_each_domain(this_cpu, sd) {
4397 unsigned long interval;
4399 if (!(sd->flags & SD_LOAD_BALANCE))
4402 if (sd->flags & SD_BALANCE_NEWIDLE)
4403 /* If we've pulled tasks over stop searching: */
4404 pulled_task = load_balance_newidle(this_cpu, this_rq,
4407 interval = msecs_to_jiffies(sd->balance_interval);
4408 if (time_after(next_balance, sd->last_balance + interval))
4409 next_balance = sd->last_balance + interval;
4413 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4415 * We are going idle. next_balance may be set based on
4416 * a busy processor. So reset next_balance.
4418 this_rq->next_balance = next_balance;
4423 * active_load_balance is run by migration threads. It pushes running tasks
4424 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4425 * running on each physical CPU where possible, and avoids physical /
4426 * logical imbalances.
4428 * Called with busiest_rq locked.
4430 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4432 int target_cpu = busiest_rq->push_cpu;
4433 struct sched_domain *sd;
4434 struct rq *target_rq;
4436 /* Is there any task to move? */
4437 if (busiest_rq->nr_running <= 1)
4440 target_rq = cpu_rq(target_cpu);
4443 * This condition is "impossible", if it occurs
4444 * we need to fix it. Originally reported by
4445 * Bjorn Helgaas on a 128-cpu setup.
4447 BUG_ON(busiest_rq == target_rq);
4449 /* move a task from busiest_rq to target_rq */
4450 double_lock_balance(busiest_rq, target_rq);
4451 update_rq_clock(busiest_rq);
4452 update_rq_clock(target_rq);
4454 /* Search for an sd spanning us and the target CPU. */
4455 for_each_domain(target_cpu, sd) {
4456 if ((sd->flags & SD_LOAD_BALANCE) &&
4457 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4462 schedstat_inc(sd, alb_count);
4464 if (move_one_task(target_rq, target_cpu, busiest_rq,
4466 schedstat_inc(sd, alb_pushed);
4468 schedstat_inc(sd, alb_failed);
4470 double_unlock_balance(busiest_rq, target_rq);
4475 atomic_t load_balancer;
4476 cpumask_var_t cpu_mask;
4477 cpumask_var_t ilb_grp_nohz_mask;
4478 } nohz ____cacheline_aligned = {
4479 .load_balancer = ATOMIC_INIT(-1),
4482 int get_nohz_load_balancer(void)
4484 return atomic_read(&nohz.load_balancer);
4487 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4489 * lowest_flag_domain - Return lowest sched_domain containing flag.
4490 * @cpu: The cpu whose lowest level of sched domain is to
4492 * @flag: The flag to check for the lowest sched_domain
4493 * for the given cpu.
4495 * Returns the lowest sched_domain of a cpu which contains the given flag.
4497 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4499 struct sched_domain *sd;
4501 for_each_domain(cpu, sd)
4502 if (sd && (sd->flags & flag))
4509 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4510 * @cpu: The cpu whose domains we're iterating over.
4511 * @sd: variable holding the value of the power_savings_sd
4513 * @flag: The flag to filter the sched_domains to be iterated.
4515 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4516 * set, starting from the lowest sched_domain to the highest.
4518 #define for_each_flag_domain(cpu, sd, flag) \
4519 for (sd = lowest_flag_domain(cpu, flag); \
4520 (sd && (sd->flags & flag)); sd = sd->parent)
4523 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4524 * @ilb_group: group to be checked for semi-idleness
4526 * Returns: 1 if the group is semi-idle. 0 otherwise.
4528 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4529 * and atleast one non-idle CPU. This helper function checks if the given
4530 * sched_group is semi-idle or not.
4532 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4534 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4535 sched_group_cpus(ilb_group));
4538 * A sched_group is semi-idle when it has atleast one busy cpu
4539 * and atleast one idle cpu.
4541 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4544 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4550 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4551 * @cpu: The cpu which is nominating a new idle_load_balancer.
4553 * Returns: Returns the id of the idle load balancer if it exists,
4554 * Else, returns >= nr_cpu_ids.
4556 * This algorithm picks the idle load balancer such that it belongs to a
4557 * semi-idle powersavings sched_domain. The idea is to try and avoid
4558 * completely idle packages/cores just for the purpose of idle load balancing
4559 * when there are other idle cpu's which are better suited for that job.
4561 static int find_new_ilb(int cpu)
4563 struct sched_domain *sd;
4564 struct sched_group *ilb_group;
4567 * Have idle load balancer selection from semi-idle packages only
4568 * when power-aware load balancing is enabled
4570 if (!(sched_smt_power_savings || sched_mc_power_savings))
4574 * Optimize for the case when we have no idle CPUs or only one
4575 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4577 if (cpumask_weight(nohz.cpu_mask) < 2)
4580 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4581 ilb_group = sd->groups;
4584 if (is_semi_idle_group(ilb_group))
4585 return cpumask_first(nohz.ilb_grp_nohz_mask);
4587 ilb_group = ilb_group->next;
4589 } while (ilb_group != sd->groups);
4593 return cpumask_first(nohz.cpu_mask);
4595 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4596 static inline int find_new_ilb(int call_cpu)
4598 return cpumask_first(nohz.cpu_mask);
4603 * This routine will try to nominate the ilb (idle load balancing)
4604 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4605 * load balancing on behalf of all those cpus. If all the cpus in the system
4606 * go into this tickless mode, then there will be no ilb owner (as there is
4607 * no need for one) and all the cpus will sleep till the next wakeup event
4610 * For the ilb owner, tick is not stopped. And this tick will be used
4611 * for idle load balancing. ilb owner will still be part of
4614 * While stopping the tick, this cpu will become the ilb owner if there
4615 * is no other owner. And will be the owner till that cpu becomes busy
4616 * or if all cpus in the system stop their ticks at which point
4617 * there is no need for ilb owner.
4619 * When the ilb owner becomes busy, it nominates another owner, during the
4620 * next busy scheduler_tick()
4622 int select_nohz_load_balancer(int stop_tick)
4624 int cpu = smp_processor_id();
4627 cpu_rq(cpu)->in_nohz_recently = 1;
4629 if (!cpu_active(cpu)) {
4630 if (atomic_read(&nohz.load_balancer) != cpu)
4634 * If we are going offline and still the leader,
4637 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4643 cpumask_set_cpu(cpu, nohz.cpu_mask);
4645 /* time for ilb owner also to sleep */
4646 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4647 if (atomic_read(&nohz.load_balancer) == cpu)
4648 atomic_set(&nohz.load_balancer, -1);
4652 if (atomic_read(&nohz.load_balancer) == -1) {
4653 /* make me the ilb owner */
4654 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4656 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4659 if (!(sched_smt_power_savings ||
4660 sched_mc_power_savings))
4663 * Check to see if there is a more power-efficient
4666 new_ilb = find_new_ilb(cpu);
4667 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4668 atomic_set(&nohz.load_balancer, -1);
4669 resched_cpu(new_ilb);
4675 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4678 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4680 if (atomic_read(&nohz.load_balancer) == cpu)
4681 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4688 static DEFINE_SPINLOCK(balancing);
4691 * It checks each scheduling domain to see if it is due to be balanced,
4692 * and initiates a balancing operation if so.
4694 * Balancing parameters are set up in arch_init_sched_domains.
4696 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4699 struct rq *rq = cpu_rq(cpu);
4700 unsigned long interval;
4701 struct sched_domain *sd;
4702 /* Earliest time when we have to do rebalance again */
4703 unsigned long next_balance = jiffies + 60*HZ;
4704 int update_next_balance = 0;
4707 for_each_domain(cpu, sd) {
4708 if (!(sd->flags & SD_LOAD_BALANCE))
4711 interval = sd->balance_interval;
4712 if (idle != CPU_IDLE)
4713 interval *= sd->busy_factor;
4715 /* scale ms to jiffies */
4716 interval = msecs_to_jiffies(interval);
4717 if (unlikely(!interval))
4719 if (interval > HZ*NR_CPUS/10)
4720 interval = HZ*NR_CPUS/10;
4722 need_serialize = sd->flags & SD_SERIALIZE;
4724 if (need_serialize) {
4725 if (!spin_trylock(&balancing))
4729 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4730 if (load_balance(cpu, rq, sd, idle, &balance)) {
4732 * We've pulled tasks over so either we're no
4733 * longer idle, or one of our SMT siblings is
4736 idle = CPU_NOT_IDLE;
4738 sd->last_balance = jiffies;
4741 spin_unlock(&balancing);
4743 if (time_after(next_balance, sd->last_balance + interval)) {
4744 next_balance = sd->last_balance + interval;
4745 update_next_balance = 1;
4749 * Stop the load balance at this level. There is another
4750 * CPU in our sched group which is doing load balancing more
4758 * next_balance will be updated only when there is a need.
4759 * When the cpu is attached to null domain for ex, it will not be
4762 if (likely(update_next_balance))
4763 rq->next_balance = next_balance;
4767 * run_rebalance_domains is triggered when needed from the scheduler tick.
4768 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4769 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4771 static void run_rebalance_domains(struct softirq_action *h)
4773 int this_cpu = smp_processor_id();
4774 struct rq *this_rq = cpu_rq(this_cpu);
4775 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4776 CPU_IDLE : CPU_NOT_IDLE;
4778 rebalance_domains(this_cpu, idle);
4782 * If this cpu is the owner for idle load balancing, then do the
4783 * balancing on behalf of the other idle cpus whose ticks are
4786 if (this_rq->idle_at_tick &&
4787 atomic_read(&nohz.load_balancer) == this_cpu) {
4791 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4792 if (balance_cpu == this_cpu)
4796 * If this cpu gets work to do, stop the load balancing
4797 * work being done for other cpus. Next load
4798 * balancing owner will pick it up.
4803 rebalance_domains(balance_cpu, CPU_IDLE);
4805 rq = cpu_rq(balance_cpu);
4806 if (time_after(this_rq->next_balance, rq->next_balance))
4807 this_rq->next_balance = rq->next_balance;
4813 static inline int on_null_domain(int cpu)
4815 return !rcu_dereference(cpu_rq(cpu)->sd);
4819 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4821 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4822 * idle load balancing owner or decide to stop the periodic load balancing,
4823 * if the whole system is idle.
4825 static inline void trigger_load_balance(struct rq *rq, int cpu)
4829 * If we were in the nohz mode recently and busy at the current
4830 * scheduler tick, then check if we need to nominate new idle
4833 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4834 rq->in_nohz_recently = 0;
4836 if (atomic_read(&nohz.load_balancer) == cpu) {
4837 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4838 atomic_set(&nohz.load_balancer, -1);
4841 if (atomic_read(&nohz.load_balancer) == -1) {
4842 int ilb = find_new_ilb(cpu);
4844 if (ilb < nr_cpu_ids)
4850 * If this cpu is idle and doing idle load balancing for all the
4851 * cpus with ticks stopped, is it time for that to stop?
4853 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4854 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4860 * If this cpu is idle and the idle load balancing is done by
4861 * someone else, then no need raise the SCHED_SOFTIRQ
4863 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4864 cpumask_test_cpu(cpu, nohz.cpu_mask))
4867 /* Don't need to rebalance while attached to NULL domain */
4868 if (time_after_eq(jiffies, rq->next_balance) &&
4869 likely(!on_null_domain(cpu)))
4870 raise_softirq(SCHED_SOFTIRQ);
4873 #else /* CONFIG_SMP */
4876 * on UP we do not need to balance between CPUs:
4878 static inline void idle_balance(int cpu, struct rq *rq)
4884 DEFINE_PER_CPU(struct kernel_stat, kstat);
4886 EXPORT_PER_CPU_SYMBOL(kstat);
4889 * Return any ns on the sched_clock that have not yet been accounted in
4890 * @p in case that task is currently running.
4892 * Called with task_rq_lock() held on @rq.
4894 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4898 if (task_current(rq, p)) {
4899 update_rq_clock(rq);
4900 ns = rq->clock - p->se.exec_start;
4908 unsigned long long task_delta_exec(struct task_struct *p)
4910 unsigned long flags;
4914 rq = task_rq_lock(p, &flags);
4915 ns = do_task_delta_exec(p, rq);
4916 task_rq_unlock(rq, &flags);
4922 * Return accounted runtime for the task.
4923 * In case the task is currently running, return the runtime plus current's
4924 * pending runtime that have not been accounted yet.
4926 unsigned long long task_sched_runtime(struct task_struct *p)
4928 unsigned long flags;
4932 rq = task_rq_lock(p, &flags);
4933 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4934 task_rq_unlock(rq, &flags);
4940 * Return sum_exec_runtime for the thread group.
4941 * In case the task is currently running, return the sum plus current's
4942 * pending runtime that have not been accounted yet.
4944 * Note that the thread group might have other running tasks as well,
4945 * so the return value not includes other pending runtime that other
4946 * running tasks might have.
4948 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4950 struct task_cputime totals;
4951 unsigned long flags;
4955 rq = task_rq_lock(p, &flags);
4956 thread_group_cputime(p, &totals);
4957 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4958 task_rq_unlock(rq, &flags);
4964 * Account user cpu time to a process.
4965 * @p: the process that the cpu time gets accounted to
4966 * @cputime: the cpu time spent in user space since the last update
4967 * @cputime_scaled: cputime scaled by cpu frequency
4969 void account_user_time(struct task_struct *p, cputime_t cputime,
4970 cputime_t cputime_scaled)
4972 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4975 /* Add user time to process. */
4976 p->utime = cputime_add(p->utime, cputime);
4977 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4978 account_group_user_time(p, cputime);
4980 /* Add user time to cpustat. */
4981 tmp = cputime_to_cputime64(cputime);
4982 if (TASK_NICE(p) > 0)
4983 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4985 cpustat->user = cputime64_add(cpustat->user, tmp);
4987 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4988 /* Account for user time used */
4989 acct_update_integrals(p);
4993 * Account guest cpu time to a process.
4994 * @p: the process that the cpu time gets accounted to
4995 * @cputime: the cpu time spent in virtual machine since the last update
4996 * @cputime_scaled: cputime scaled by cpu frequency
4998 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4999 cputime_t cputime_scaled)
5002 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5004 tmp = cputime_to_cputime64(cputime);
5006 /* Add guest time to process. */
5007 p->utime = cputime_add(p->utime, cputime);
5008 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
5009 account_group_user_time(p, cputime);
5010 p->gtime = cputime_add(p->gtime, cputime);
5012 /* Add guest time to cpustat. */
5013 cpustat->user = cputime64_add(cpustat->user, tmp);
5014 cpustat->guest = cputime64_add(cpustat->guest, tmp);
5018 * Account system cpu time to a process.
5019 * @p: the process that the cpu time gets accounted to
5020 * @hardirq_offset: the offset to subtract from hardirq_count()
5021 * @cputime: the cpu time spent in kernel space since the last update
5022 * @cputime_scaled: cputime scaled by cpu frequency
5024 void account_system_time(struct task_struct *p, int hardirq_offset,
5025 cputime_t cputime, cputime_t cputime_scaled)
5027 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5030 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
5031 account_guest_time(p, cputime, cputime_scaled);
5035 /* Add system time to process. */
5036 p->stime = cputime_add(p->stime, cputime);
5037 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
5038 account_group_system_time(p, cputime);
5040 /* Add system time to cpustat. */
5041 tmp = cputime_to_cputime64(cputime);
5042 if (hardirq_count() - hardirq_offset)
5043 cpustat->irq = cputime64_add(cpustat->irq, tmp);
5044 else if (softirq_count())
5045 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
5047 cpustat->system = cputime64_add(cpustat->system, tmp);
5049 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
5051 /* Account for system time used */
5052 acct_update_integrals(p);
5056 * Account for involuntary wait time.
5057 * @steal: the cpu time spent in involuntary wait
5059 void account_steal_time(cputime_t cputime)
5061 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5062 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5064 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
5068 * Account for idle time.
5069 * @cputime: the cpu time spent in idle wait
5071 void account_idle_time(cputime_t cputime)
5073 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
5074 cputime64_t cputime64 = cputime_to_cputime64(cputime);
5075 struct rq *rq = this_rq();
5077 if (atomic_read(&rq->nr_iowait) > 0)
5078 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
5080 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
5083 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5086 * Account a single tick of cpu time.
5087 * @p: the process that the cpu time gets accounted to
5088 * @user_tick: indicates if the tick is a user or a system tick
5090 void account_process_tick(struct task_struct *p, int user_tick)
5092 cputime_t one_jiffy = jiffies_to_cputime(1);
5093 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
5094 struct rq *rq = this_rq();
5097 account_user_time(p, one_jiffy, one_jiffy_scaled);
5098 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
5099 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
5102 account_idle_time(one_jiffy);
5106 * Account multiple ticks of steal time.
5107 * @p: the process from which the cpu time has been stolen
5108 * @ticks: number of stolen ticks
5110 void account_steal_ticks(unsigned long ticks)
5112 account_steal_time(jiffies_to_cputime(ticks));
5116 * Account multiple ticks of idle time.
5117 * @ticks: number of stolen ticks
5119 void account_idle_ticks(unsigned long ticks)
5121 account_idle_time(jiffies_to_cputime(ticks));
5127 * Use precise platform statistics if available:
5129 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5130 cputime_t task_utime(struct task_struct *p)
5135 cputime_t task_stime(struct task_struct *p)
5140 cputime_t task_utime(struct task_struct *p)
5142 clock_t utime = cputime_to_clock_t(p->utime),
5143 total = utime + cputime_to_clock_t(p->stime);
5147 * Use CFS's precise accounting:
5149 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
5153 do_div(temp, total);
5155 utime = (clock_t)temp;
5157 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
5158 return p->prev_utime;
5161 cputime_t task_stime(struct task_struct *p)
5166 * Use CFS's precise accounting. (we subtract utime from
5167 * the total, to make sure the total observed by userspace
5168 * grows monotonically - apps rely on that):
5170 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
5171 cputime_to_clock_t(task_utime(p));
5174 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5176 return p->prev_stime;
5180 inline cputime_t task_gtime(struct task_struct *p)
5186 * This function gets called by the timer code, with HZ frequency.
5187 * We call it with interrupts disabled.
5189 * It also gets called by the fork code, when changing the parent's
5192 void scheduler_tick(void)
5194 int cpu = smp_processor_id();
5195 struct rq *rq = cpu_rq(cpu);
5196 struct task_struct *curr = rq->curr;
5200 spin_lock(&rq->lock);
5201 update_rq_clock(rq);
5202 update_cpu_load(rq);
5203 curr->sched_class->task_tick(rq, curr, 0);
5204 spin_unlock(&rq->lock);
5206 perf_counter_task_tick(curr, cpu);
5209 rq->idle_at_tick = idle_cpu(cpu);
5210 trigger_load_balance(rq, cpu);
5214 notrace unsigned long get_parent_ip(unsigned long addr)
5216 if (in_lock_functions(addr)) {
5217 addr = CALLER_ADDR2;
5218 if (in_lock_functions(addr))
5219 addr = CALLER_ADDR3;
5224 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5225 defined(CONFIG_PREEMPT_TRACER))
5227 void __kprobes add_preempt_count(int val)
5229 #ifdef CONFIG_DEBUG_PREEMPT
5233 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5236 preempt_count() += val;
5237 #ifdef CONFIG_DEBUG_PREEMPT
5239 * Spinlock count overflowing soon?
5241 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5244 if (preempt_count() == val)
5245 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5247 EXPORT_SYMBOL(add_preempt_count);
5249 void __kprobes sub_preempt_count(int val)
5251 #ifdef CONFIG_DEBUG_PREEMPT
5255 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5258 * Is the spinlock portion underflowing?
5260 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5261 !(preempt_count() & PREEMPT_MASK)))
5265 if (preempt_count() == val)
5266 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5267 preempt_count() -= val;
5269 EXPORT_SYMBOL(sub_preempt_count);
5274 * Print scheduling while atomic bug:
5276 static noinline void __schedule_bug(struct task_struct *prev)
5278 struct pt_regs *regs = get_irq_regs();
5280 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5281 prev->comm, prev->pid, preempt_count());
5283 debug_show_held_locks(prev);
5285 if (irqs_disabled())
5286 print_irqtrace_events(prev);
5295 * Various schedule()-time debugging checks and statistics:
5297 static inline void schedule_debug(struct task_struct *prev)
5300 * Test if we are atomic. Since do_exit() needs to call into
5301 * schedule() atomically, we ignore that path for now.
5302 * Otherwise, whine if we are scheduling when we should not be.
5304 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5305 __schedule_bug(prev);
5307 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5309 schedstat_inc(this_rq(), sched_count);
5310 #ifdef CONFIG_SCHEDSTATS
5311 if (unlikely(prev->lock_depth >= 0)) {
5312 schedstat_inc(this_rq(), bkl_count);
5313 schedstat_inc(prev, sched_info.bkl_count);
5318 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5320 if (prev->state == TASK_RUNNING) {
5321 u64 runtime = prev->se.sum_exec_runtime;
5323 runtime -= prev->se.prev_sum_exec_runtime;
5324 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5327 * In order to avoid avg_overlap growing stale when we are
5328 * indeed overlapping and hence not getting put to sleep, grow
5329 * the avg_overlap on preemption.
5331 * We use the average preemption runtime because that
5332 * correlates to the amount of cache footprint a task can
5335 update_avg(&prev->se.avg_overlap, runtime);
5337 prev->sched_class->put_prev_task(rq, prev);
5341 * Pick up the highest-prio task:
5343 static inline struct task_struct *
5344 pick_next_task(struct rq *rq)
5346 const struct sched_class *class;
5347 struct task_struct *p;
5350 * Optimization: we know that if all tasks are in
5351 * the fair class we can call that function directly:
5353 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5354 p = fair_sched_class.pick_next_task(rq);
5359 class = sched_class_highest;
5361 p = class->pick_next_task(rq);
5365 * Will never be NULL as the idle class always
5366 * returns a non-NULL p:
5368 class = class->next;
5373 * schedule() is the main scheduler function.
5375 asmlinkage void __sched schedule(void)
5377 struct task_struct *prev, *next;
5378 unsigned long *switch_count;
5384 cpu = smp_processor_id();
5388 switch_count = &prev->nivcsw;
5390 release_kernel_lock(prev);
5391 need_resched_nonpreemptible:
5393 schedule_debug(prev);
5395 if (sched_feat(HRTICK))
5398 spin_lock_irq(&rq->lock);
5399 update_rq_clock(rq);
5400 clear_tsk_need_resched(prev);
5402 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5403 if (unlikely(signal_pending_state(prev->state, prev)))
5404 prev->state = TASK_RUNNING;
5406 deactivate_task(rq, prev, 1);
5407 switch_count = &prev->nvcsw;
5410 pre_schedule(rq, prev);
5412 if (unlikely(!rq->nr_running))
5413 idle_balance(cpu, rq);
5415 put_prev_task(rq, prev);
5416 next = pick_next_task(rq);
5418 if (likely(prev != next)) {
5419 sched_info_switch(prev, next);
5420 perf_counter_task_sched_out(prev, next, cpu);
5426 context_switch(rq, prev, next); /* unlocks the rq */
5428 * the context switch might have flipped the stack from under
5429 * us, hence refresh the local variables.
5431 cpu = smp_processor_id();
5434 spin_unlock_irq(&rq->lock);
5438 if (unlikely(reacquire_kernel_lock(current) < 0))
5439 goto need_resched_nonpreemptible;
5441 preempt_enable_no_resched();
5445 EXPORT_SYMBOL(schedule);
5449 * Look out! "owner" is an entirely speculative pointer
5450 * access and not reliable.
5452 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5457 if (!sched_feat(OWNER_SPIN))
5460 #ifdef CONFIG_DEBUG_PAGEALLOC
5462 * Need to access the cpu field knowing that
5463 * DEBUG_PAGEALLOC could have unmapped it if
5464 * the mutex owner just released it and exited.
5466 if (probe_kernel_address(&owner->cpu, cpu))
5473 * Even if the access succeeded (likely case),
5474 * the cpu field may no longer be valid.
5476 if (cpu >= nr_cpumask_bits)
5480 * We need to validate that we can do a
5481 * get_cpu() and that we have the percpu area.
5483 if (!cpu_online(cpu))
5490 * Owner changed, break to re-assess state.
5492 if (lock->owner != owner)
5496 * Is that owner really running on that cpu?
5498 if (task_thread_info(rq->curr) != owner || need_resched())
5508 #ifdef CONFIG_PREEMPT
5510 * this is the entry point to schedule() from in-kernel preemption
5511 * off of preempt_enable. Kernel preemptions off return from interrupt
5512 * occur there and call schedule directly.
5514 asmlinkage void __sched preempt_schedule(void)
5516 struct thread_info *ti = current_thread_info();
5519 * If there is a non-zero preempt_count or interrupts are disabled,
5520 * we do not want to preempt the current task. Just return..
5522 if (likely(ti->preempt_count || irqs_disabled()))
5526 add_preempt_count(PREEMPT_ACTIVE);
5528 sub_preempt_count(PREEMPT_ACTIVE);
5531 * Check again in case we missed a preemption opportunity
5532 * between schedule and now.
5535 } while (need_resched());
5537 EXPORT_SYMBOL(preempt_schedule);
5540 * this is the entry point to schedule() from kernel preemption
5541 * off of irq context.
5542 * Note, that this is called and return with irqs disabled. This will
5543 * protect us against recursive calling from irq.
5545 asmlinkage void __sched preempt_schedule_irq(void)
5547 struct thread_info *ti = current_thread_info();
5549 /* Catch callers which need to be fixed */
5550 BUG_ON(ti->preempt_count || !irqs_disabled());
5553 add_preempt_count(PREEMPT_ACTIVE);
5556 local_irq_disable();
5557 sub_preempt_count(PREEMPT_ACTIVE);
5560 * Check again in case we missed a preemption opportunity
5561 * between schedule and now.
5564 } while (need_resched());
5567 #endif /* CONFIG_PREEMPT */
5569 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5572 return try_to_wake_up(curr->private, mode, sync);
5574 EXPORT_SYMBOL(default_wake_function);
5577 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5578 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5579 * number) then we wake all the non-exclusive tasks and one exclusive task.
5581 * There are circumstances in which we can try to wake a task which has already
5582 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5583 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5585 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5586 int nr_exclusive, int sync, void *key)
5588 wait_queue_t *curr, *next;
5590 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5591 unsigned flags = curr->flags;
5593 if (curr->func(curr, mode, sync, key) &&
5594 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5600 * __wake_up - wake up threads blocked on a waitqueue.
5602 * @mode: which threads
5603 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5604 * @key: is directly passed to the wakeup function
5606 * It may be assumed that this function implies a write memory barrier before
5607 * changing the task state if and only if any tasks are woken up.
5609 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5610 int nr_exclusive, void *key)
5612 unsigned long flags;
5614 spin_lock_irqsave(&q->lock, flags);
5615 __wake_up_common(q, mode, nr_exclusive, 0, key);
5616 spin_unlock_irqrestore(&q->lock, flags);
5618 EXPORT_SYMBOL(__wake_up);
5621 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5623 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5625 __wake_up_common(q, mode, 1, 0, NULL);
5628 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5630 __wake_up_common(q, mode, 1, 0, key);
5634 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5636 * @mode: which threads
5637 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5638 * @key: opaque value to be passed to wakeup targets
5640 * The sync wakeup differs that the waker knows that it will schedule
5641 * away soon, so while the target thread will be woken up, it will not
5642 * be migrated to another CPU - ie. the two threads are 'synchronized'
5643 * with each other. This can prevent needless bouncing between CPUs.
5645 * On UP it can prevent extra preemption.
5647 * It may be assumed that this function implies a write memory barrier before
5648 * changing the task state if and only if any tasks are woken up.
5650 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5651 int nr_exclusive, void *key)
5653 unsigned long flags;
5659 if (unlikely(!nr_exclusive))
5662 spin_lock_irqsave(&q->lock, flags);
5663 __wake_up_common(q, mode, nr_exclusive, sync, key);
5664 spin_unlock_irqrestore(&q->lock, flags);
5666 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5669 * __wake_up_sync - see __wake_up_sync_key()
5671 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5673 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5675 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5678 * complete: - signals a single thread waiting on this completion
5679 * @x: holds the state of this particular completion
5681 * This will wake up a single thread waiting on this completion. Threads will be
5682 * awakened in the same order in which they were queued.
5684 * See also complete_all(), wait_for_completion() and related routines.
5686 * It may be assumed that this function implies a write memory barrier before
5687 * changing the task state if and only if any tasks are woken up.
5689 void complete(struct completion *x)
5691 unsigned long flags;
5693 spin_lock_irqsave(&x->wait.lock, flags);
5695 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5696 spin_unlock_irqrestore(&x->wait.lock, flags);
5698 EXPORT_SYMBOL(complete);
5701 * complete_all: - signals all threads waiting on this completion
5702 * @x: holds the state of this particular completion
5704 * This will wake up all threads waiting on this particular completion event.
5706 * It may be assumed that this function implies a write memory barrier before
5707 * changing the task state if and only if any tasks are woken up.
5709 void complete_all(struct completion *x)
5711 unsigned long flags;
5713 spin_lock_irqsave(&x->wait.lock, flags);
5714 x->done += UINT_MAX/2;
5715 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5716 spin_unlock_irqrestore(&x->wait.lock, flags);
5718 EXPORT_SYMBOL(complete_all);
5720 static inline long __sched
5721 do_wait_for_common(struct completion *x, long timeout, int state)
5724 DECLARE_WAITQUEUE(wait, current);
5726 wait.flags |= WQ_FLAG_EXCLUSIVE;
5727 __add_wait_queue_tail(&x->wait, &wait);
5729 if (signal_pending_state(state, current)) {
5730 timeout = -ERESTARTSYS;
5733 __set_current_state(state);
5734 spin_unlock_irq(&x->wait.lock);
5735 timeout = schedule_timeout(timeout);
5736 spin_lock_irq(&x->wait.lock);
5737 } while (!x->done && timeout);
5738 __remove_wait_queue(&x->wait, &wait);
5743 return timeout ?: 1;
5747 wait_for_common(struct completion *x, long timeout, int state)
5751 spin_lock_irq(&x->wait.lock);
5752 timeout = do_wait_for_common(x, timeout, state);
5753 spin_unlock_irq(&x->wait.lock);
5758 * wait_for_completion: - waits for completion of a task
5759 * @x: holds the state of this particular completion
5761 * This waits to be signaled for completion of a specific task. It is NOT
5762 * interruptible and there is no timeout.
5764 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5765 * and interrupt capability. Also see complete().
5767 void __sched wait_for_completion(struct completion *x)
5769 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5771 EXPORT_SYMBOL(wait_for_completion);
5774 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5775 * @x: holds the state of this particular completion
5776 * @timeout: timeout value in jiffies
5778 * This waits for either a completion of a specific task to be signaled or for a
5779 * specified timeout to expire. The timeout is in jiffies. It is not
5782 unsigned long __sched
5783 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5785 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5787 EXPORT_SYMBOL(wait_for_completion_timeout);
5790 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5791 * @x: holds the state of this particular completion
5793 * This waits for completion of a specific task to be signaled. It is
5796 int __sched wait_for_completion_interruptible(struct completion *x)
5798 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5799 if (t == -ERESTARTSYS)
5803 EXPORT_SYMBOL(wait_for_completion_interruptible);
5806 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5807 * @x: holds the state of this particular completion
5808 * @timeout: timeout value in jiffies
5810 * This waits for either a completion of a specific task to be signaled or for a
5811 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5813 unsigned long __sched
5814 wait_for_completion_interruptible_timeout(struct completion *x,
5815 unsigned long timeout)
5817 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5819 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5822 * wait_for_completion_killable: - waits for completion of a task (killable)
5823 * @x: holds the state of this particular completion
5825 * This waits to be signaled for completion of a specific task. It can be
5826 * interrupted by a kill signal.
5828 int __sched wait_for_completion_killable(struct completion *x)
5830 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5831 if (t == -ERESTARTSYS)
5835 EXPORT_SYMBOL(wait_for_completion_killable);
5838 * try_wait_for_completion - try to decrement a completion without blocking
5839 * @x: completion structure
5841 * Returns: 0 if a decrement cannot be done without blocking
5842 * 1 if a decrement succeeded.
5844 * If a completion is being used as a counting completion,
5845 * attempt to decrement the counter without blocking. This
5846 * enables us to avoid waiting if the resource the completion
5847 * is protecting is not available.
5849 bool try_wait_for_completion(struct completion *x)
5853 spin_lock_irq(&x->wait.lock);
5858 spin_unlock_irq(&x->wait.lock);
5861 EXPORT_SYMBOL(try_wait_for_completion);
5864 * completion_done - Test to see if a completion has any waiters
5865 * @x: completion structure
5867 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5868 * 1 if there are no waiters.
5871 bool completion_done(struct completion *x)
5875 spin_lock_irq(&x->wait.lock);
5878 spin_unlock_irq(&x->wait.lock);
5881 EXPORT_SYMBOL(completion_done);
5884 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5886 unsigned long flags;
5889 init_waitqueue_entry(&wait, current);
5891 __set_current_state(state);
5893 spin_lock_irqsave(&q->lock, flags);
5894 __add_wait_queue(q, &wait);
5895 spin_unlock(&q->lock);
5896 timeout = schedule_timeout(timeout);
5897 spin_lock_irq(&q->lock);
5898 __remove_wait_queue(q, &wait);
5899 spin_unlock_irqrestore(&q->lock, flags);
5904 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5906 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5908 EXPORT_SYMBOL(interruptible_sleep_on);
5911 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5913 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5915 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5917 void __sched sleep_on(wait_queue_head_t *q)
5919 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5921 EXPORT_SYMBOL(sleep_on);
5923 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5925 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5927 EXPORT_SYMBOL(sleep_on_timeout);
5929 #ifdef CONFIG_RT_MUTEXES
5932 * rt_mutex_setprio - set the current priority of a task
5934 * @prio: prio value (kernel-internal form)
5936 * This function changes the 'effective' priority of a task. It does
5937 * not touch ->normal_prio like __setscheduler().
5939 * Used by the rt_mutex code to implement priority inheritance logic.
5941 void rt_mutex_setprio(struct task_struct *p, int prio)
5943 unsigned long flags;
5944 int oldprio, on_rq, running;
5946 const struct sched_class *prev_class = p->sched_class;
5948 BUG_ON(prio < 0 || prio > MAX_PRIO);
5950 rq = task_rq_lock(p, &flags);
5951 update_rq_clock(rq);
5954 on_rq = p->se.on_rq;
5955 running = task_current(rq, p);
5957 dequeue_task(rq, p, 0);
5959 p->sched_class->put_prev_task(rq, p);
5962 p->sched_class = &rt_sched_class;
5964 p->sched_class = &fair_sched_class;
5969 p->sched_class->set_curr_task(rq);
5971 enqueue_task(rq, p, 0);
5973 check_class_changed(rq, p, prev_class, oldprio, running);
5975 task_rq_unlock(rq, &flags);
5980 void set_user_nice(struct task_struct *p, long nice)
5982 int old_prio, delta, on_rq;
5983 unsigned long flags;
5986 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5989 * We have to be careful, if called from sys_setpriority(),
5990 * the task might be in the middle of scheduling on another CPU.
5992 rq = task_rq_lock(p, &flags);
5993 update_rq_clock(rq);
5995 * The RT priorities are set via sched_setscheduler(), but we still
5996 * allow the 'normal' nice value to be set - but as expected
5997 * it wont have any effect on scheduling until the task is
5998 * SCHED_FIFO/SCHED_RR:
6000 if (task_has_rt_policy(p)) {
6001 p->static_prio = NICE_TO_PRIO(nice);
6004 on_rq = p->se.on_rq;
6006 dequeue_task(rq, p, 0);
6008 p->static_prio = NICE_TO_PRIO(nice);
6011 p->prio = effective_prio(p);
6012 delta = p->prio - old_prio;
6015 enqueue_task(rq, p, 0);
6017 * If the task increased its priority or is running and
6018 * lowered its priority, then reschedule its CPU:
6020 if (delta < 0 || (delta > 0 && task_running(rq, p)))
6021 resched_task(rq->curr);
6024 task_rq_unlock(rq, &flags);
6026 EXPORT_SYMBOL(set_user_nice);
6029 * can_nice - check if a task can reduce its nice value
6033 int can_nice(const struct task_struct *p, const int nice)
6035 /* convert nice value [19,-20] to rlimit style value [1,40] */
6036 int nice_rlim = 20 - nice;
6038 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
6039 capable(CAP_SYS_NICE));
6042 #ifdef __ARCH_WANT_SYS_NICE
6045 * sys_nice - change the priority of the current process.
6046 * @increment: priority increment
6048 * sys_setpriority is a more generic, but much slower function that
6049 * does similar things.
6051 SYSCALL_DEFINE1(nice, int, increment)
6056 * Setpriority might change our priority at the same moment.
6057 * We don't have to worry. Conceptually one call occurs first
6058 * and we have a single winner.
6060 if (increment < -40)
6065 nice = TASK_NICE(current) + increment;
6071 if (increment < 0 && !can_nice(current, nice))
6074 retval = security_task_setnice(current, nice);
6078 set_user_nice(current, nice);
6085 * task_prio - return the priority value of a given task.
6086 * @p: the task in question.
6088 * This is the priority value as seen by users in /proc.
6089 * RT tasks are offset by -200. Normal tasks are centered
6090 * around 0, value goes from -16 to +15.
6092 int task_prio(const struct task_struct *p)
6094 return p->prio - MAX_RT_PRIO;
6098 * task_nice - return the nice value of a given task.
6099 * @p: the task in question.
6101 int task_nice(const struct task_struct *p)
6103 return TASK_NICE(p);
6105 EXPORT_SYMBOL(task_nice);
6108 * idle_cpu - is a given cpu idle currently?
6109 * @cpu: the processor in question.
6111 int idle_cpu(int cpu)
6113 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
6117 * idle_task - return the idle task for a given cpu.
6118 * @cpu: the processor in question.
6120 struct task_struct *idle_task(int cpu)
6122 return cpu_rq(cpu)->idle;
6126 * find_process_by_pid - find a process with a matching PID value.
6127 * @pid: the pid in question.
6129 static struct task_struct *find_process_by_pid(pid_t pid)
6131 return pid ? find_task_by_vpid(pid) : current;
6134 /* Actually do priority change: must hold rq lock. */
6136 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
6138 BUG_ON(p->se.on_rq);
6141 switch (p->policy) {
6145 p->sched_class = &fair_sched_class;
6149 p->sched_class = &rt_sched_class;
6153 p->rt_priority = prio;
6154 p->normal_prio = normal_prio(p);
6155 /* we are holding p->pi_lock already */
6156 p->prio = rt_mutex_getprio(p);
6161 * check the target process has a UID that matches the current process's
6163 static bool check_same_owner(struct task_struct *p)
6165 const struct cred *cred = current_cred(), *pcred;
6169 pcred = __task_cred(p);
6170 match = (cred->euid == pcred->euid ||
6171 cred->euid == pcred->uid);
6176 static int __sched_setscheduler(struct task_struct *p, int policy,
6177 struct sched_param *param, bool user)
6179 int retval, oldprio, oldpolicy = -1, on_rq, running;
6180 unsigned long flags;
6181 const struct sched_class *prev_class = p->sched_class;
6185 /* may grab non-irq protected spin_locks */
6186 BUG_ON(in_interrupt());
6188 /* double check policy once rq lock held */
6190 reset_on_fork = p->sched_reset_on_fork;
6191 policy = oldpolicy = p->policy;
6193 reset_on_fork = !!(policy & SCHED_RESET_ON_FORK);
6194 policy &= ~SCHED_RESET_ON_FORK;
6196 if (policy != SCHED_FIFO && policy != SCHED_RR &&
6197 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6198 policy != SCHED_IDLE)
6203 * Valid priorities for SCHED_FIFO and SCHED_RR are
6204 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6205 * SCHED_BATCH and SCHED_IDLE is 0.
6207 if (param->sched_priority < 0 ||
6208 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6209 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6211 if (rt_policy(policy) != (param->sched_priority != 0))
6215 * Allow unprivileged RT tasks to decrease priority:
6217 if (user && !capable(CAP_SYS_NICE)) {
6218 if (rt_policy(policy)) {
6219 unsigned long rlim_rtprio;
6221 if (!lock_task_sighand(p, &flags))
6223 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6224 unlock_task_sighand(p, &flags);
6226 /* can't set/change the rt policy */
6227 if (policy != p->policy && !rlim_rtprio)
6230 /* can't increase priority */
6231 if (param->sched_priority > p->rt_priority &&
6232 param->sched_priority > rlim_rtprio)
6236 * Like positive nice levels, dont allow tasks to
6237 * move out of SCHED_IDLE either:
6239 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6242 /* can't change other user's priorities */
6243 if (!check_same_owner(p))
6246 /* Normal users shall not reset the sched_reset_on_fork flag */
6247 if (p->sched_reset_on_fork && !reset_on_fork)
6252 #ifdef CONFIG_RT_GROUP_SCHED
6254 * Do not allow realtime tasks into groups that have no runtime
6257 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6258 task_group(p)->rt_bandwidth.rt_runtime == 0)
6262 retval = security_task_setscheduler(p, policy, param);
6268 * make sure no PI-waiters arrive (or leave) while we are
6269 * changing the priority of the task:
6271 spin_lock_irqsave(&p->pi_lock, flags);
6273 * To be able to change p->policy safely, the apropriate
6274 * runqueue lock must be held.
6276 rq = __task_rq_lock(p);
6277 /* recheck policy now with rq lock held */
6278 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6279 policy = oldpolicy = -1;
6280 __task_rq_unlock(rq);
6281 spin_unlock_irqrestore(&p->pi_lock, flags);
6284 update_rq_clock(rq);
6285 on_rq = p->se.on_rq;
6286 running = task_current(rq, p);
6288 deactivate_task(rq, p, 0);
6290 p->sched_class->put_prev_task(rq, p);
6292 p->sched_reset_on_fork = reset_on_fork;
6295 __setscheduler(rq, p, policy, param->sched_priority);
6298 p->sched_class->set_curr_task(rq);
6300 activate_task(rq, p, 0);
6302 check_class_changed(rq, p, prev_class, oldprio, running);
6304 __task_rq_unlock(rq);
6305 spin_unlock_irqrestore(&p->pi_lock, flags);
6307 rt_mutex_adjust_pi(p);
6313 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6314 * @p: the task in question.
6315 * @policy: new policy.
6316 * @param: structure containing the new RT priority.
6318 * NOTE that the task may be already dead.
6320 int sched_setscheduler(struct task_struct *p, int policy,
6321 struct sched_param *param)
6323 return __sched_setscheduler(p, policy, param, true);
6325 EXPORT_SYMBOL_GPL(sched_setscheduler);
6328 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6329 * @p: the task in question.
6330 * @policy: new policy.
6331 * @param: structure containing the new RT priority.
6333 * Just like sched_setscheduler, only don't bother checking if the
6334 * current context has permission. For example, this is needed in
6335 * stop_machine(): we create temporary high priority worker threads,
6336 * but our caller might not have that capability.
6338 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6339 struct sched_param *param)
6341 return __sched_setscheduler(p, policy, param, false);
6345 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6347 struct sched_param lparam;
6348 struct task_struct *p;
6351 if (!param || pid < 0)
6353 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6358 p = find_process_by_pid(pid);
6360 retval = sched_setscheduler(p, policy, &lparam);
6367 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6368 * @pid: the pid in question.
6369 * @policy: new policy.
6370 * @param: structure containing the new RT priority.
6372 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6373 struct sched_param __user *, param)
6375 /* negative values for policy are not valid */
6379 return do_sched_setscheduler(pid, policy, param);
6383 * sys_sched_setparam - set/change the RT priority of a thread
6384 * @pid: the pid in question.
6385 * @param: structure containing the new RT priority.
6387 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6389 return do_sched_setscheduler(pid, -1, param);
6393 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6394 * @pid: the pid in question.
6396 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6398 struct task_struct *p;
6405 read_lock(&tasklist_lock);
6406 p = find_process_by_pid(pid);
6408 retval = security_task_getscheduler(p);
6411 | (p->sched_reset_on_fork ? SCHED_RESET_ON_FORK : 0);
6413 read_unlock(&tasklist_lock);
6418 * sys_sched_getparam - get the RT priority of a thread
6419 * @pid: the pid in question.
6420 * @param: structure containing the RT priority.
6422 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6424 struct sched_param lp;
6425 struct task_struct *p;
6428 if (!param || pid < 0)
6431 read_lock(&tasklist_lock);
6432 p = find_process_by_pid(pid);
6437 retval = security_task_getscheduler(p);
6441 lp.sched_priority = p->rt_priority;
6442 read_unlock(&tasklist_lock);
6445 * This one might sleep, we cannot do it with a spinlock held ...
6447 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6452 read_unlock(&tasklist_lock);
6456 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6458 cpumask_var_t cpus_allowed, new_mask;
6459 struct task_struct *p;
6463 read_lock(&tasklist_lock);
6465 p = find_process_by_pid(pid);
6467 read_unlock(&tasklist_lock);
6473 * It is not safe to call set_cpus_allowed with the
6474 * tasklist_lock held. We will bump the task_struct's
6475 * usage count and then drop tasklist_lock.
6478 read_unlock(&tasklist_lock);
6480 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6484 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6486 goto out_free_cpus_allowed;
6489 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6492 retval = security_task_setscheduler(p, 0, NULL);
6496 cpuset_cpus_allowed(p, cpus_allowed);
6497 cpumask_and(new_mask, in_mask, cpus_allowed);
6499 retval = set_cpus_allowed_ptr(p, new_mask);
6502 cpuset_cpus_allowed(p, cpus_allowed);
6503 if (!cpumask_subset(new_mask, cpus_allowed)) {
6505 * We must have raced with a concurrent cpuset
6506 * update. Just reset the cpus_allowed to the
6507 * cpuset's cpus_allowed
6509 cpumask_copy(new_mask, cpus_allowed);
6514 free_cpumask_var(new_mask);
6515 out_free_cpus_allowed:
6516 free_cpumask_var(cpus_allowed);
6523 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6524 struct cpumask *new_mask)
6526 if (len < cpumask_size())
6527 cpumask_clear(new_mask);
6528 else if (len > cpumask_size())
6529 len = cpumask_size();
6531 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6535 * sys_sched_setaffinity - set the cpu affinity of a process
6536 * @pid: pid of the process
6537 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6538 * @user_mask_ptr: user-space pointer to the new cpu mask
6540 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6541 unsigned long __user *, user_mask_ptr)
6543 cpumask_var_t new_mask;
6546 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6549 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6551 retval = sched_setaffinity(pid, new_mask);
6552 free_cpumask_var(new_mask);
6556 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6558 struct task_struct *p;
6562 read_lock(&tasklist_lock);
6565 p = find_process_by_pid(pid);
6569 retval = security_task_getscheduler(p);
6573 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6576 read_unlock(&tasklist_lock);
6583 * sys_sched_getaffinity - get the cpu affinity of a process
6584 * @pid: pid of the process
6585 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6586 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6588 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6589 unsigned long __user *, user_mask_ptr)
6594 if (len < cpumask_size())
6597 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6600 ret = sched_getaffinity(pid, mask);
6602 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6605 ret = cpumask_size();
6607 free_cpumask_var(mask);
6613 * sys_sched_yield - yield the current processor to other threads.
6615 * This function yields the current CPU to other tasks. If there are no
6616 * other threads running on this CPU then this function will return.
6618 SYSCALL_DEFINE0(sched_yield)
6620 struct rq *rq = this_rq_lock();
6622 schedstat_inc(rq, yld_count);
6623 current->sched_class->yield_task(rq);
6626 * Since we are going to call schedule() anyway, there's
6627 * no need to preempt or enable interrupts:
6629 __release(rq->lock);
6630 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6631 _raw_spin_unlock(&rq->lock);
6632 preempt_enable_no_resched();
6639 static inline int should_resched(void)
6641 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE);
6644 static void __cond_resched(void)
6646 add_preempt_count(PREEMPT_ACTIVE);
6648 sub_preempt_count(PREEMPT_ACTIVE);
6651 int __sched _cond_resched(void)
6653 if (should_resched()) {
6659 EXPORT_SYMBOL(_cond_resched);
6662 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6663 * call schedule, and on return reacquire the lock.
6665 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6666 * operations here to prevent schedule() from being called twice (once via
6667 * spin_unlock(), once by hand).
6669 int __cond_resched_lock(spinlock_t *lock)
6671 int resched = should_resched();
6674 lockdep_assert_held(lock);
6676 if (spin_needbreak(lock) || resched) {
6687 EXPORT_SYMBOL(__cond_resched_lock);
6689 int __sched __cond_resched_softirq(void)
6691 BUG_ON(!in_softirq());
6693 if (should_resched()) {
6701 EXPORT_SYMBOL(__cond_resched_softirq);
6704 * yield - yield the current processor to other threads.
6706 * This is a shortcut for kernel-space yielding - it marks the
6707 * thread runnable and calls sys_sched_yield().
6709 void __sched yield(void)
6711 set_current_state(TASK_RUNNING);
6714 EXPORT_SYMBOL(yield);
6717 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6718 * that process accounting knows that this is a task in IO wait state.
6720 * But don't do that if it is a deliberate, throttling IO wait (this task
6721 * has set its backing_dev_info: the queue against which it should throttle)
6723 void __sched io_schedule(void)
6725 struct rq *rq = raw_rq();
6727 delayacct_blkio_start();
6728 atomic_inc(&rq->nr_iowait);
6729 current->in_iowait = 1;
6731 current->in_iowait = 0;
6732 atomic_dec(&rq->nr_iowait);
6733 delayacct_blkio_end();
6735 EXPORT_SYMBOL(io_schedule);
6737 long __sched io_schedule_timeout(long timeout)
6739 struct rq *rq = raw_rq();
6742 delayacct_blkio_start();
6743 atomic_inc(&rq->nr_iowait);
6744 current->in_iowait = 1;
6745 ret = schedule_timeout(timeout);
6746 current->in_iowait = 0;
6747 atomic_dec(&rq->nr_iowait);
6748 delayacct_blkio_end();
6753 * sys_sched_get_priority_max - return maximum RT priority.
6754 * @policy: scheduling class.
6756 * this syscall returns the maximum rt_priority that can be used
6757 * by a given scheduling class.
6759 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6766 ret = MAX_USER_RT_PRIO-1;
6778 * sys_sched_get_priority_min - return minimum RT priority.
6779 * @policy: scheduling class.
6781 * this syscall returns the minimum rt_priority that can be used
6782 * by a given scheduling class.
6784 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6802 * sys_sched_rr_get_interval - return the default timeslice of a process.
6803 * @pid: pid of the process.
6804 * @interval: userspace pointer to the timeslice value.
6806 * this syscall writes the default timeslice value of a given process
6807 * into the user-space timespec buffer. A value of '0' means infinity.
6809 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6810 struct timespec __user *, interval)
6812 struct task_struct *p;
6813 unsigned int time_slice;
6821 read_lock(&tasklist_lock);
6822 p = find_process_by_pid(pid);
6826 retval = security_task_getscheduler(p);
6831 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6832 * tasks that are on an otherwise idle runqueue:
6835 if (p->policy == SCHED_RR) {
6836 time_slice = DEF_TIMESLICE;
6837 } else if (p->policy != SCHED_FIFO) {
6838 struct sched_entity *se = &p->se;
6839 unsigned long flags;
6842 rq = task_rq_lock(p, &flags);
6843 if (rq->cfs.load.weight)
6844 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6845 task_rq_unlock(rq, &flags);
6847 read_unlock(&tasklist_lock);
6848 jiffies_to_timespec(time_slice, &t);
6849 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6853 read_unlock(&tasklist_lock);
6857 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6859 void sched_show_task(struct task_struct *p)
6861 unsigned long free = 0;
6864 state = p->state ? __ffs(p->state) + 1 : 0;
6865 printk(KERN_INFO "%-13.13s %c", p->comm,
6866 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6867 #if BITS_PER_LONG == 32
6868 if (state == TASK_RUNNING)
6869 printk(KERN_CONT " running ");
6871 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6873 if (state == TASK_RUNNING)
6874 printk(KERN_CONT " running task ");
6876 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6878 #ifdef CONFIG_DEBUG_STACK_USAGE
6879 free = stack_not_used(p);
6881 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6882 task_pid_nr(p), task_pid_nr(p->real_parent),
6883 (unsigned long)task_thread_info(p)->flags);
6885 show_stack(p, NULL);
6888 void show_state_filter(unsigned long state_filter)
6890 struct task_struct *g, *p;
6892 #if BITS_PER_LONG == 32
6894 " task PC stack pid father\n");
6897 " task PC stack pid father\n");
6899 read_lock(&tasklist_lock);
6900 do_each_thread(g, p) {
6902 * reset the NMI-timeout, listing all files on a slow
6903 * console might take alot of time:
6905 touch_nmi_watchdog();
6906 if (!state_filter || (p->state & state_filter))
6908 } while_each_thread(g, p);
6910 touch_all_softlockup_watchdogs();
6912 #ifdef CONFIG_SCHED_DEBUG
6913 sysrq_sched_debug_show();
6915 read_unlock(&tasklist_lock);
6917 * Only show locks if all tasks are dumped:
6919 if (state_filter == -1)
6920 debug_show_all_locks();
6923 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6925 idle->sched_class = &idle_sched_class;
6929 * init_idle - set up an idle thread for a given CPU
6930 * @idle: task in question
6931 * @cpu: cpu the idle task belongs to
6933 * NOTE: this function does not set the idle thread's NEED_RESCHED
6934 * flag, to make booting more robust.
6936 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6938 struct rq *rq = cpu_rq(cpu);
6939 unsigned long flags;
6941 spin_lock_irqsave(&rq->lock, flags);
6944 idle->se.exec_start = sched_clock();
6946 idle->prio = idle->normal_prio = MAX_PRIO;
6947 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6948 __set_task_cpu(idle, cpu);
6950 rq->curr = rq->idle = idle;
6951 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6954 spin_unlock_irqrestore(&rq->lock, flags);
6956 /* Set the preempt count _outside_ the spinlocks! */
6957 #if defined(CONFIG_PREEMPT)
6958 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6960 task_thread_info(idle)->preempt_count = 0;
6963 * The idle tasks have their own, simple scheduling class:
6965 idle->sched_class = &idle_sched_class;
6966 ftrace_graph_init_task(idle);
6970 * In a system that switches off the HZ timer nohz_cpu_mask
6971 * indicates which cpus entered this state. This is used
6972 * in the rcu update to wait only for active cpus. For system
6973 * which do not switch off the HZ timer nohz_cpu_mask should
6974 * always be CPU_BITS_NONE.
6976 cpumask_var_t nohz_cpu_mask;
6979 * Increase the granularity value when there are more CPUs,
6980 * because with more CPUs the 'effective latency' as visible
6981 * to users decreases. But the relationship is not linear,
6982 * so pick a second-best guess by going with the log2 of the
6985 * This idea comes from the SD scheduler of Con Kolivas:
6987 static inline void sched_init_granularity(void)
6989 unsigned int factor = 1 + ilog2(num_online_cpus());
6990 const unsigned long limit = 200000000;
6992 sysctl_sched_min_granularity *= factor;
6993 if (sysctl_sched_min_granularity > limit)
6994 sysctl_sched_min_granularity = limit;
6996 sysctl_sched_latency *= factor;
6997 if (sysctl_sched_latency > limit)
6998 sysctl_sched_latency = limit;
7000 sysctl_sched_wakeup_granularity *= factor;
7002 sysctl_sched_shares_ratelimit *= factor;
7007 * This is how migration works:
7009 * 1) we queue a struct migration_req structure in the source CPU's
7010 * runqueue and wake up that CPU's migration thread.
7011 * 2) we down() the locked semaphore => thread blocks.
7012 * 3) migration thread wakes up (implicitly it forces the migrated
7013 * thread off the CPU)
7014 * 4) it gets the migration request and checks whether the migrated
7015 * task is still in the wrong runqueue.
7016 * 5) if it's in the wrong runqueue then the migration thread removes
7017 * it and puts it into the right queue.
7018 * 6) migration thread up()s the semaphore.
7019 * 7) we wake up and the migration is done.
7023 * Change a given task's CPU affinity. Migrate the thread to a
7024 * proper CPU and schedule it away if the CPU it's executing on
7025 * is removed from the allowed bitmask.
7027 * NOTE: the caller must have a valid reference to the task, the
7028 * task must not exit() & deallocate itself prematurely. The
7029 * call is not atomic; no spinlocks may be held.
7031 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
7033 struct migration_req req;
7034 unsigned long flags;
7038 rq = task_rq_lock(p, &flags);
7039 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
7044 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
7045 !cpumask_equal(&p->cpus_allowed, new_mask))) {
7050 if (p->sched_class->set_cpus_allowed)
7051 p->sched_class->set_cpus_allowed(p, new_mask);
7053 cpumask_copy(&p->cpus_allowed, new_mask);
7054 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
7057 /* Can the task run on the task's current CPU? If so, we're done */
7058 if (cpumask_test_cpu(task_cpu(p), new_mask))
7061 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
7062 /* Need help from migration thread: drop lock and wait. */
7063 struct task_struct *mt = rq->migration_thread;
7065 get_task_struct(mt);
7066 task_rq_unlock(rq, &flags);
7067 wake_up_process(rq->migration_thread);
7068 put_task_struct(mt);
7069 wait_for_completion(&req.done);
7070 tlb_migrate_finish(p->mm);
7074 task_rq_unlock(rq, &flags);
7078 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
7081 * Move (not current) task off this cpu, onto dest cpu. We're doing
7082 * this because either it can't run here any more (set_cpus_allowed()
7083 * away from this CPU, or CPU going down), or because we're
7084 * attempting to rebalance this task on exec (sched_exec).
7086 * So we race with normal scheduler movements, but that's OK, as long
7087 * as the task is no longer on this CPU.
7089 * Returns non-zero if task was successfully migrated.
7091 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
7093 struct rq *rq_dest, *rq_src;
7096 if (unlikely(!cpu_active(dest_cpu)))
7099 rq_src = cpu_rq(src_cpu);
7100 rq_dest = cpu_rq(dest_cpu);
7102 double_rq_lock(rq_src, rq_dest);
7103 /* Already moved. */
7104 if (task_cpu(p) != src_cpu)
7106 /* Affinity changed (again). */
7107 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7110 on_rq = p->se.on_rq;
7112 deactivate_task(rq_src, p, 0);
7114 set_task_cpu(p, dest_cpu);
7116 activate_task(rq_dest, p, 0);
7117 check_preempt_curr(rq_dest, p, 0);
7122 double_rq_unlock(rq_src, rq_dest);
7126 #define RCU_MIGRATION_IDLE 0
7127 #define RCU_MIGRATION_NEED_QS 1
7128 #define RCU_MIGRATION_GOT_QS 2
7129 #define RCU_MIGRATION_MUST_SYNC 3
7132 * migration_thread - this is a highprio system thread that performs
7133 * thread migration by bumping thread off CPU then 'pushing' onto
7136 static int migration_thread(void *data)
7139 int cpu = (long)data;
7143 BUG_ON(rq->migration_thread != current);
7145 set_current_state(TASK_INTERRUPTIBLE);
7146 while (!kthread_should_stop()) {
7147 struct migration_req *req;
7148 struct list_head *head;
7150 spin_lock_irq(&rq->lock);
7152 if (cpu_is_offline(cpu)) {
7153 spin_unlock_irq(&rq->lock);
7157 if (rq->active_balance) {
7158 active_load_balance(rq, cpu);
7159 rq->active_balance = 0;
7162 head = &rq->migration_queue;
7164 if (list_empty(head)) {
7165 spin_unlock_irq(&rq->lock);
7167 set_current_state(TASK_INTERRUPTIBLE);
7170 req = list_entry(head->next, struct migration_req, list);
7171 list_del_init(head->next);
7173 if (req->task != NULL) {
7174 spin_unlock(&rq->lock);
7175 __migrate_task(req->task, cpu, req->dest_cpu);
7176 } else if (likely(cpu == (badcpu = smp_processor_id()))) {
7177 req->dest_cpu = RCU_MIGRATION_GOT_QS;
7178 spin_unlock(&rq->lock);
7180 req->dest_cpu = RCU_MIGRATION_MUST_SYNC;
7181 spin_unlock(&rq->lock);
7182 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu, cpu);
7186 complete(&req->done);
7188 __set_current_state(TASK_RUNNING);
7193 #ifdef CONFIG_HOTPLUG_CPU
7195 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
7199 local_irq_disable();
7200 ret = __migrate_task(p, src_cpu, dest_cpu);
7206 * Figure out where task on dead CPU should go, use force if necessary.
7208 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
7211 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7214 /* Look for allowed, online CPU in same node. */
7215 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7216 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7219 /* Any allowed, online CPU? */
7220 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7221 if (dest_cpu < nr_cpu_ids)
7224 /* No more Mr. Nice Guy. */
7225 if (dest_cpu >= nr_cpu_ids) {
7226 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7227 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7230 * Don't tell them about moving exiting tasks or
7231 * kernel threads (both mm NULL), since they never
7234 if (p->mm && printk_ratelimit()) {
7235 printk(KERN_INFO "process %d (%s) no "
7236 "longer affine to cpu%d\n",
7237 task_pid_nr(p), p->comm, dead_cpu);
7242 /* It can have affinity changed while we were choosing. */
7243 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7248 * While a dead CPU has no uninterruptible tasks queued at this point,
7249 * it might still have a nonzero ->nr_uninterruptible counter, because
7250 * for performance reasons the counter is not stricly tracking tasks to
7251 * their home CPUs. So we just add the counter to another CPU's counter,
7252 * to keep the global sum constant after CPU-down:
7254 static void migrate_nr_uninterruptible(struct rq *rq_src)
7256 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7257 unsigned long flags;
7259 local_irq_save(flags);
7260 double_rq_lock(rq_src, rq_dest);
7261 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7262 rq_src->nr_uninterruptible = 0;
7263 double_rq_unlock(rq_src, rq_dest);
7264 local_irq_restore(flags);
7267 /* Run through task list and migrate tasks from the dead cpu. */
7268 static void migrate_live_tasks(int src_cpu)
7270 struct task_struct *p, *t;
7272 read_lock(&tasklist_lock);
7274 do_each_thread(t, p) {
7278 if (task_cpu(p) == src_cpu)
7279 move_task_off_dead_cpu(src_cpu, p);
7280 } while_each_thread(t, p);
7282 read_unlock(&tasklist_lock);
7286 * Schedules idle task to be the next runnable task on current CPU.
7287 * It does so by boosting its priority to highest possible.
7288 * Used by CPU offline code.
7290 void sched_idle_next(void)
7292 int this_cpu = smp_processor_id();
7293 struct rq *rq = cpu_rq(this_cpu);
7294 struct task_struct *p = rq->idle;
7295 unsigned long flags;
7297 /* cpu has to be offline */
7298 BUG_ON(cpu_online(this_cpu));
7301 * Strictly not necessary since rest of the CPUs are stopped by now
7302 * and interrupts disabled on the current cpu.
7304 spin_lock_irqsave(&rq->lock, flags);
7306 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7308 update_rq_clock(rq);
7309 activate_task(rq, p, 0);
7311 spin_unlock_irqrestore(&rq->lock, flags);
7315 * Ensures that the idle task is using init_mm right before its cpu goes
7318 void idle_task_exit(void)
7320 struct mm_struct *mm = current->active_mm;
7322 BUG_ON(cpu_online(smp_processor_id()));
7325 switch_mm(mm, &init_mm, current);
7329 /* called under rq->lock with disabled interrupts */
7330 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7332 struct rq *rq = cpu_rq(dead_cpu);
7334 /* Must be exiting, otherwise would be on tasklist. */
7335 BUG_ON(!p->exit_state);
7337 /* Cannot have done final schedule yet: would have vanished. */
7338 BUG_ON(p->state == TASK_DEAD);
7343 * Drop lock around migration; if someone else moves it,
7344 * that's OK. No task can be added to this CPU, so iteration is
7347 spin_unlock_irq(&rq->lock);
7348 move_task_off_dead_cpu(dead_cpu, p);
7349 spin_lock_irq(&rq->lock);
7354 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7355 static void migrate_dead_tasks(unsigned int dead_cpu)
7357 struct rq *rq = cpu_rq(dead_cpu);
7358 struct task_struct *next;
7361 if (!rq->nr_running)
7363 update_rq_clock(rq);
7364 next = pick_next_task(rq);
7367 next->sched_class->put_prev_task(rq, next);
7368 migrate_dead(dead_cpu, next);
7374 * remove the tasks which were accounted by rq from calc_load_tasks.
7376 static void calc_global_load_remove(struct rq *rq)
7378 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7379 rq->calc_load_active = 0;
7381 #endif /* CONFIG_HOTPLUG_CPU */
7383 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7385 static struct ctl_table sd_ctl_dir[] = {
7387 .procname = "sched_domain",
7393 static struct ctl_table sd_ctl_root[] = {
7395 .ctl_name = CTL_KERN,
7396 .procname = "kernel",
7398 .child = sd_ctl_dir,
7403 static struct ctl_table *sd_alloc_ctl_entry(int n)
7405 struct ctl_table *entry =
7406 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7411 static void sd_free_ctl_entry(struct ctl_table **tablep)
7413 struct ctl_table *entry;
7416 * In the intermediate directories, both the child directory and
7417 * procname are dynamically allocated and could fail but the mode
7418 * will always be set. In the lowest directory the names are
7419 * static strings and all have proc handlers.
7421 for (entry = *tablep; entry->mode; entry++) {
7423 sd_free_ctl_entry(&entry->child);
7424 if (entry->proc_handler == NULL)
7425 kfree(entry->procname);
7433 set_table_entry(struct ctl_table *entry,
7434 const char *procname, void *data, int maxlen,
7435 mode_t mode, proc_handler *proc_handler)
7437 entry->procname = procname;
7439 entry->maxlen = maxlen;
7441 entry->proc_handler = proc_handler;
7444 static struct ctl_table *
7445 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7447 struct ctl_table *table = sd_alloc_ctl_entry(13);
7452 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7453 sizeof(long), 0644, proc_doulongvec_minmax);
7454 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7455 sizeof(long), 0644, proc_doulongvec_minmax);
7456 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7457 sizeof(int), 0644, proc_dointvec_minmax);
7458 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7459 sizeof(int), 0644, proc_dointvec_minmax);
7460 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7461 sizeof(int), 0644, proc_dointvec_minmax);
7462 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7463 sizeof(int), 0644, proc_dointvec_minmax);
7464 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7465 sizeof(int), 0644, proc_dointvec_minmax);
7466 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7467 sizeof(int), 0644, proc_dointvec_minmax);
7468 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7469 sizeof(int), 0644, proc_dointvec_minmax);
7470 set_table_entry(&table[9], "cache_nice_tries",
7471 &sd->cache_nice_tries,
7472 sizeof(int), 0644, proc_dointvec_minmax);
7473 set_table_entry(&table[10], "flags", &sd->flags,
7474 sizeof(int), 0644, proc_dointvec_minmax);
7475 set_table_entry(&table[11], "name", sd->name,
7476 CORENAME_MAX_SIZE, 0444, proc_dostring);
7477 /* &table[12] is terminator */
7482 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7484 struct ctl_table *entry, *table;
7485 struct sched_domain *sd;
7486 int domain_num = 0, i;
7489 for_each_domain(cpu, sd)
7491 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7496 for_each_domain(cpu, sd) {
7497 snprintf(buf, 32, "domain%d", i);
7498 entry->procname = kstrdup(buf, GFP_KERNEL);
7500 entry->child = sd_alloc_ctl_domain_table(sd);
7507 static struct ctl_table_header *sd_sysctl_header;
7508 static void register_sched_domain_sysctl(void)
7510 int i, cpu_num = num_online_cpus();
7511 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7514 WARN_ON(sd_ctl_dir[0].child);
7515 sd_ctl_dir[0].child = entry;
7520 for_each_online_cpu(i) {
7521 snprintf(buf, 32, "cpu%d", i);
7522 entry->procname = kstrdup(buf, GFP_KERNEL);
7524 entry->child = sd_alloc_ctl_cpu_table(i);
7528 WARN_ON(sd_sysctl_header);
7529 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7532 /* may be called multiple times per register */
7533 static void unregister_sched_domain_sysctl(void)
7535 if (sd_sysctl_header)
7536 unregister_sysctl_table(sd_sysctl_header);
7537 sd_sysctl_header = NULL;
7538 if (sd_ctl_dir[0].child)
7539 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7542 static void register_sched_domain_sysctl(void)
7545 static void unregister_sched_domain_sysctl(void)
7550 static void set_rq_online(struct rq *rq)
7553 const struct sched_class *class;
7555 cpumask_set_cpu(rq->cpu, rq->rd->online);
7558 for_each_class(class) {
7559 if (class->rq_online)
7560 class->rq_online(rq);
7565 static void set_rq_offline(struct rq *rq)
7568 const struct sched_class *class;
7570 for_each_class(class) {
7571 if (class->rq_offline)
7572 class->rq_offline(rq);
7575 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7581 * migration_call - callback that gets triggered when a CPU is added.
7582 * Here we can start up the necessary migration thread for the new CPU.
7584 static int __cpuinit
7585 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7587 struct task_struct *p;
7588 int cpu = (long)hcpu;
7589 unsigned long flags;
7594 case CPU_UP_PREPARE:
7595 case CPU_UP_PREPARE_FROZEN:
7596 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7599 kthread_bind(p, cpu);
7600 /* Must be high prio: stop_machine expects to yield to it. */
7601 rq = task_rq_lock(p, &flags);
7602 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7603 task_rq_unlock(rq, &flags);
7605 cpu_rq(cpu)->migration_thread = p;
7606 rq->calc_load_update = calc_load_update;
7610 case CPU_ONLINE_FROZEN:
7611 /* Strictly unnecessary, as first user will wake it. */
7612 wake_up_process(cpu_rq(cpu)->migration_thread);
7614 /* Update our root-domain */
7616 spin_lock_irqsave(&rq->lock, flags);
7618 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7622 spin_unlock_irqrestore(&rq->lock, flags);
7625 #ifdef CONFIG_HOTPLUG_CPU
7626 case CPU_UP_CANCELED:
7627 case CPU_UP_CANCELED_FROZEN:
7628 if (!cpu_rq(cpu)->migration_thread)
7630 /* Unbind it from offline cpu so it can run. Fall thru. */
7631 kthread_bind(cpu_rq(cpu)->migration_thread,
7632 cpumask_any(cpu_online_mask));
7633 kthread_stop(cpu_rq(cpu)->migration_thread);
7634 put_task_struct(cpu_rq(cpu)->migration_thread);
7635 cpu_rq(cpu)->migration_thread = NULL;
7639 case CPU_DEAD_FROZEN:
7640 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7641 migrate_live_tasks(cpu);
7643 kthread_stop(rq->migration_thread);
7644 put_task_struct(rq->migration_thread);
7645 rq->migration_thread = NULL;
7646 /* Idle task back to normal (off runqueue, low prio) */
7647 spin_lock_irq(&rq->lock);
7648 update_rq_clock(rq);
7649 deactivate_task(rq, rq->idle, 0);
7650 rq->idle->static_prio = MAX_PRIO;
7651 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7652 rq->idle->sched_class = &idle_sched_class;
7653 migrate_dead_tasks(cpu);
7654 spin_unlock_irq(&rq->lock);
7656 migrate_nr_uninterruptible(rq);
7657 BUG_ON(rq->nr_running != 0);
7658 calc_global_load_remove(rq);
7660 * No need to migrate the tasks: it was best-effort if
7661 * they didn't take sched_hotcpu_mutex. Just wake up
7664 spin_lock_irq(&rq->lock);
7665 while (!list_empty(&rq->migration_queue)) {
7666 struct migration_req *req;
7668 req = list_entry(rq->migration_queue.next,
7669 struct migration_req, list);
7670 list_del_init(&req->list);
7671 spin_unlock_irq(&rq->lock);
7672 complete(&req->done);
7673 spin_lock_irq(&rq->lock);
7675 spin_unlock_irq(&rq->lock);
7679 case CPU_DYING_FROZEN:
7680 /* Update our root-domain */
7682 spin_lock_irqsave(&rq->lock, flags);
7684 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7687 spin_unlock_irqrestore(&rq->lock, flags);
7695 * Register at high priority so that task migration (migrate_all_tasks)
7696 * happens before everything else. This has to be lower priority than
7697 * the notifier in the perf_counter subsystem, though.
7699 static struct notifier_block __cpuinitdata migration_notifier = {
7700 .notifier_call = migration_call,
7704 static int __init migration_init(void)
7706 void *cpu = (void *)(long)smp_processor_id();
7709 /* Start one for the boot CPU: */
7710 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7711 BUG_ON(err == NOTIFY_BAD);
7712 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7713 register_cpu_notifier(&migration_notifier);
7717 early_initcall(migration_init);
7722 #ifdef CONFIG_SCHED_DEBUG
7724 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7725 struct cpumask *groupmask)
7727 struct sched_group *group = sd->groups;
7730 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7731 cpumask_clear(groupmask);
7733 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7735 if (!(sd->flags & SD_LOAD_BALANCE)) {
7736 printk("does not load-balance\n");
7738 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7743 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7745 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7746 printk(KERN_ERR "ERROR: domain->span does not contain "
7749 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7750 printk(KERN_ERR "ERROR: domain->groups does not contain"
7754 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7758 printk(KERN_ERR "ERROR: group is NULL\n");
7762 if (!group->cpu_power) {
7763 printk(KERN_CONT "\n");
7764 printk(KERN_ERR "ERROR: domain->cpu_power not "
7769 if (!cpumask_weight(sched_group_cpus(group))) {
7770 printk(KERN_CONT "\n");
7771 printk(KERN_ERR "ERROR: empty group\n");
7775 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7776 printk(KERN_CONT "\n");
7777 printk(KERN_ERR "ERROR: repeated CPUs\n");
7781 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7783 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7785 printk(KERN_CONT " %s", str);
7786 if (group->cpu_power != SCHED_LOAD_SCALE) {
7787 printk(KERN_CONT " (cpu_power = %d)",
7791 group = group->next;
7792 } while (group != sd->groups);
7793 printk(KERN_CONT "\n");
7795 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7796 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7799 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7800 printk(KERN_ERR "ERROR: parent span is not a superset "
7801 "of domain->span\n");
7805 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7807 cpumask_var_t groupmask;
7811 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7815 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7817 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7818 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7823 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7830 free_cpumask_var(groupmask);
7832 #else /* !CONFIG_SCHED_DEBUG */
7833 # define sched_domain_debug(sd, cpu) do { } while (0)
7834 #endif /* CONFIG_SCHED_DEBUG */
7836 static int sd_degenerate(struct sched_domain *sd)
7838 if (cpumask_weight(sched_domain_span(sd)) == 1)
7841 /* Following flags need at least 2 groups */
7842 if (sd->flags & (SD_LOAD_BALANCE |
7843 SD_BALANCE_NEWIDLE |
7847 SD_SHARE_PKG_RESOURCES)) {
7848 if (sd->groups != sd->groups->next)
7852 /* Following flags don't use groups */
7853 if (sd->flags & (SD_WAKE_IDLE |
7862 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7864 unsigned long cflags = sd->flags, pflags = parent->flags;
7866 if (sd_degenerate(parent))
7869 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7872 /* Does parent contain flags not in child? */
7873 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7874 if (cflags & SD_WAKE_AFFINE)
7875 pflags &= ~SD_WAKE_BALANCE;
7876 /* Flags needing groups don't count if only 1 group in parent */
7877 if (parent->groups == parent->groups->next) {
7878 pflags &= ~(SD_LOAD_BALANCE |
7879 SD_BALANCE_NEWIDLE |
7883 SD_SHARE_PKG_RESOURCES);
7884 if (nr_node_ids == 1)
7885 pflags &= ~SD_SERIALIZE;
7887 if (~cflags & pflags)
7893 static void free_rootdomain(struct root_domain *rd)
7895 cpupri_cleanup(&rd->cpupri);
7897 free_cpumask_var(rd->rto_mask);
7898 free_cpumask_var(rd->online);
7899 free_cpumask_var(rd->span);
7903 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7905 struct root_domain *old_rd = NULL;
7906 unsigned long flags;
7908 spin_lock_irqsave(&rq->lock, flags);
7913 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7916 cpumask_clear_cpu(rq->cpu, old_rd->span);
7919 * If we dont want to free the old_rt yet then
7920 * set old_rd to NULL to skip the freeing later
7923 if (!atomic_dec_and_test(&old_rd->refcount))
7927 atomic_inc(&rd->refcount);
7930 cpumask_set_cpu(rq->cpu, rd->span);
7931 if (cpumask_test_cpu(rq->cpu, cpu_active_mask))
7934 spin_unlock_irqrestore(&rq->lock, flags);
7937 free_rootdomain(old_rd);
7940 static int init_rootdomain(struct root_domain *rd, bool bootmem)
7942 gfp_t gfp = GFP_KERNEL;
7944 memset(rd, 0, sizeof(*rd));
7949 if (!alloc_cpumask_var(&rd->span, gfp))
7951 if (!alloc_cpumask_var(&rd->online, gfp))
7953 if (!alloc_cpumask_var(&rd->rto_mask, gfp))
7956 if (cpupri_init(&rd->cpupri, bootmem) != 0)
7961 free_cpumask_var(rd->rto_mask);
7963 free_cpumask_var(rd->online);
7965 free_cpumask_var(rd->span);
7970 static void init_defrootdomain(void)
7972 init_rootdomain(&def_root_domain, true);
7974 atomic_set(&def_root_domain.refcount, 1);
7977 static struct root_domain *alloc_rootdomain(void)
7979 struct root_domain *rd;
7981 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7985 if (init_rootdomain(rd, false) != 0) {
7994 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7995 * hold the hotplug lock.
7998 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
8000 struct rq *rq = cpu_rq(cpu);
8001 struct sched_domain *tmp;
8003 /* Remove the sched domains which do not contribute to scheduling. */
8004 for (tmp = sd; tmp; ) {
8005 struct sched_domain *parent = tmp->parent;
8009 if (sd_parent_degenerate(tmp, parent)) {
8010 tmp->parent = parent->parent;
8012 parent->parent->child = tmp;
8017 if (sd && sd_degenerate(sd)) {
8023 sched_domain_debug(sd, cpu);
8025 rq_attach_root(rq, rd);
8026 rcu_assign_pointer(rq->sd, sd);
8029 /* cpus with isolated domains */
8030 static cpumask_var_t cpu_isolated_map;
8032 /* Setup the mask of cpus configured for isolated domains */
8033 static int __init isolated_cpu_setup(char *str)
8035 cpulist_parse(str, cpu_isolated_map);
8039 __setup("isolcpus=", isolated_cpu_setup);
8042 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8043 * to a function which identifies what group(along with sched group) a CPU
8044 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8045 * (due to the fact that we keep track of groups covered with a struct cpumask).
8047 * init_sched_build_groups will build a circular linked list of the groups
8048 * covered by the given span, and will set each group's ->cpumask correctly,
8049 * and ->cpu_power to 0.
8052 init_sched_build_groups(const struct cpumask *span,
8053 const struct cpumask *cpu_map,
8054 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
8055 struct sched_group **sg,
8056 struct cpumask *tmpmask),
8057 struct cpumask *covered, struct cpumask *tmpmask)
8059 struct sched_group *first = NULL, *last = NULL;
8062 cpumask_clear(covered);
8064 for_each_cpu(i, span) {
8065 struct sched_group *sg;
8066 int group = group_fn(i, cpu_map, &sg, tmpmask);
8069 if (cpumask_test_cpu(i, covered))
8072 cpumask_clear(sched_group_cpus(sg));
8075 for_each_cpu(j, span) {
8076 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
8079 cpumask_set_cpu(j, covered);
8080 cpumask_set_cpu(j, sched_group_cpus(sg));
8091 #define SD_NODES_PER_DOMAIN 16
8096 * find_next_best_node - find the next node to include in a sched_domain
8097 * @node: node whose sched_domain we're building
8098 * @used_nodes: nodes already in the sched_domain
8100 * Find the next node to include in a given scheduling domain. Simply
8101 * finds the closest node not already in the @used_nodes map.
8103 * Should use nodemask_t.
8105 static int find_next_best_node(int node, nodemask_t *used_nodes)
8107 int i, n, val, min_val, best_node = 0;
8111 for (i = 0; i < nr_node_ids; i++) {
8112 /* Start at @node */
8113 n = (node + i) % nr_node_ids;
8115 if (!nr_cpus_node(n))
8118 /* Skip already used nodes */
8119 if (node_isset(n, *used_nodes))
8122 /* Simple min distance search */
8123 val = node_distance(node, n);
8125 if (val < min_val) {
8131 node_set(best_node, *used_nodes);
8136 * sched_domain_node_span - get a cpumask for a node's sched_domain
8137 * @node: node whose cpumask we're constructing
8138 * @span: resulting cpumask
8140 * Given a node, construct a good cpumask for its sched_domain to span. It
8141 * should be one that prevents unnecessary balancing, but also spreads tasks
8144 static void sched_domain_node_span(int node, struct cpumask *span)
8146 nodemask_t used_nodes;
8149 cpumask_clear(span);
8150 nodes_clear(used_nodes);
8152 cpumask_or(span, span, cpumask_of_node(node));
8153 node_set(node, used_nodes);
8155 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
8156 int next_node = find_next_best_node(node, &used_nodes);
8158 cpumask_or(span, span, cpumask_of_node(next_node));
8161 #endif /* CONFIG_NUMA */
8163 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
8166 * The cpus mask in sched_group and sched_domain hangs off the end.
8168 * ( See the the comments in include/linux/sched.h:struct sched_group
8169 * and struct sched_domain. )
8171 struct static_sched_group {
8172 struct sched_group sg;
8173 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
8176 struct static_sched_domain {
8177 struct sched_domain sd;
8178 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
8184 cpumask_var_t domainspan;
8185 cpumask_var_t covered;
8186 cpumask_var_t notcovered;
8188 cpumask_var_t nodemask;
8189 cpumask_var_t this_sibling_map;
8190 cpumask_var_t this_core_map;
8191 cpumask_var_t send_covered;
8192 cpumask_var_t tmpmask;
8193 struct sched_group **sched_group_nodes;
8194 struct root_domain *rd;
8198 sa_sched_groups = 0,
8203 sa_this_sibling_map,
8205 sa_sched_group_nodes,
8215 * SMT sched-domains:
8217 #ifdef CONFIG_SCHED_SMT
8218 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
8219 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
8222 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
8223 struct sched_group **sg, struct cpumask *unused)
8226 *sg = &per_cpu(sched_group_cpus, cpu).sg;
8229 #endif /* CONFIG_SCHED_SMT */
8232 * multi-core sched-domains:
8234 #ifdef CONFIG_SCHED_MC
8235 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
8236 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
8237 #endif /* CONFIG_SCHED_MC */
8239 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8241 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8242 struct sched_group **sg, struct cpumask *mask)
8246 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8247 group = cpumask_first(mask);
8249 *sg = &per_cpu(sched_group_core, group).sg;
8252 #elif defined(CONFIG_SCHED_MC)
8254 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8255 struct sched_group **sg, struct cpumask *unused)
8258 *sg = &per_cpu(sched_group_core, cpu).sg;
8263 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8264 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8267 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8268 struct sched_group **sg, struct cpumask *mask)
8271 #ifdef CONFIG_SCHED_MC
8272 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8273 group = cpumask_first(mask);
8274 #elif defined(CONFIG_SCHED_SMT)
8275 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8276 group = cpumask_first(mask);
8281 *sg = &per_cpu(sched_group_phys, group).sg;
8287 * The init_sched_build_groups can't handle what we want to do with node
8288 * groups, so roll our own. Now each node has its own list of groups which
8289 * gets dynamically allocated.
8291 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8292 static struct sched_group ***sched_group_nodes_bycpu;
8294 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8295 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8297 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8298 struct sched_group **sg,
8299 struct cpumask *nodemask)
8303 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8304 group = cpumask_first(nodemask);
8307 *sg = &per_cpu(sched_group_allnodes, group).sg;
8311 static void init_numa_sched_groups_power(struct sched_group *group_head)
8313 struct sched_group *sg = group_head;
8319 for_each_cpu(j, sched_group_cpus(sg)) {
8320 struct sched_domain *sd;
8322 sd = &per_cpu(phys_domains, j).sd;
8323 if (j != group_first_cpu(sd->groups)) {
8325 * Only add "power" once for each
8331 sg->cpu_power += sd->groups->cpu_power;
8334 } while (sg != group_head);
8337 static int build_numa_sched_groups(struct s_data *d,
8338 const struct cpumask *cpu_map, int num)
8340 struct sched_domain *sd;
8341 struct sched_group *sg, *prev;
8344 cpumask_clear(d->covered);
8345 cpumask_and(d->nodemask, cpumask_of_node(num), cpu_map);
8346 if (cpumask_empty(d->nodemask)) {
8347 d->sched_group_nodes[num] = NULL;
8351 sched_domain_node_span(num, d->domainspan);
8352 cpumask_and(d->domainspan, d->domainspan, cpu_map);
8354 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8357 printk(KERN_WARNING "Can not alloc domain group for node %d\n",
8361 d->sched_group_nodes[num] = sg;
8363 for_each_cpu(j, d->nodemask) {
8364 sd = &per_cpu(node_domains, j).sd;
8369 cpumask_copy(sched_group_cpus(sg), d->nodemask);
8371 cpumask_or(d->covered, d->covered, d->nodemask);
8374 for (j = 0; j < nr_node_ids; j++) {
8375 n = (num + j) % nr_node_ids;
8376 cpumask_complement(d->notcovered, d->covered);
8377 cpumask_and(d->tmpmask, d->notcovered, cpu_map);
8378 cpumask_and(d->tmpmask, d->tmpmask, d->domainspan);
8379 if (cpumask_empty(d->tmpmask))
8381 cpumask_and(d->tmpmask, d->tmpmask, cpumask_of_node(n));
8382 if (cpumask_empty(d->tmpmask))
8384 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8388 "Can not alloc domain group for node %d\n", j);
8392 cpumask_copy(sched_group_cpus(sg), d->tmpmask);
8393 sg->next = prev->next;
8394 cpumask_or(d->covered, d->covered, d->tmpmask);
8401 #endif /* CONFIG_NUMA */
8404 /* Free memory allocated for various sched_group structures */
8405 static void free_sched_groups(const struct cpumask *cpu_map,
8406 struct cpumask *nodemask)
8410 for_each_cpu(cpu, cpu_map) {
8411 struct sched_group **sched_group_nodes
8412 = sched_group_nodes_bycpu[cpu];
8414 if (!sched_group_nodes)
8417 for (i = 0; i < nr_node_ids; i++) {
8418 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8420 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8421 if (cpumask_empty(nodemask))
8431 if (oldsg != sched_group_nodes[i])
8434 kfree(sched_group_nodes);
8435 sched_group_nodes_bycpu[cpu] = NULL;
8438 #else /* !CONFIG_NUMA */
8439 static void free_sched_groups(const struct cpumask *cpu_map,
8440 struct cpumask *nodemask)
8443 #endif /* CONFIG_NUMA */
8446 * Initialize sched groups cpu_power.
8448 * cpu_power indicates the capacity of sched group, which is used while
8449 * distributing the load between different sched groups in a sched domain.
8450 * Typically cpu_power for all the groups in a sched domain will be same unless
8451 * there are asymmetries in the topology. If there are asymmetries, group
8452 * having more cpu_power will pickup more load compared to the group having
8455 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8457 struct sched_domain *child;
8458 struct sched_group *group;
8462 WARN_ON(!sd || !sd->groups);
8464 if (cpu != group_first_cpu(sd->groups))
8469 sd->groups->cpu_power = 0;
8472 power = SCHED_LOAD_SCALE;
8473 weight = cpumask_weight(sched_domain_span(sd));
8475 * SMT siblings share the power of a single core.
8476 * Usually multiple threads get a better yield out of
8477 * that one core than a single thread would have,
8478 * reflect that in sd->smt_gain.
8480 if ((sd->flags & SD_SHARE_CPUPOWER) && weight > 1) {
8481 power *= sd->smt_gain;
8483 power >>= SCHED_LOAD_SHIFT;
8485 sd->groups->cpu_power += power;
8490 * Add cpu_power of each child group to this groups cpu_power.
8492 group = child->groups;
8494 sd->groups->cpu_power += group->cpu_power;
8495 group = group->next;
8496 } while (group != child->groups);
8500 * Initializers for schedule domains
8501 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8504 #ifdef CONFIG_SCHED_DEBUG
8505 # define SD_INIT_NAME(sd, type) sd->name = #type
8507 # define SD_INIT_NAME(sd, type) do { } while (0)
8510 #define SD_INIT(sd, type) sd_init_##type(sd)
8512 #define SD_INIT_FUNC(type) \
8513 static noinline void sd_init_##type(struct sched_domain *sd) \
8515 memset(sd, 0, sizeof(*sd)); \
8516 *sd = SD_##type##_INIT; \
8517 sd->level = SD_LV_##type; \
8518 SD_INIT_NAME(sd, type); \
8523 SD_INIT_FUNC(ALLNODES)
8526 #ifdef CONFIG_SCHED_SMT
8527 SD_INIT_FUNC(SIBLING)
8529 #ifdef CONFIG_SCHED_MC
8533 static int default_relax_domain_level = -1;
8535 static int __init setup_relax_domain_level(char *str)
8539 val = simple_strtoul(str, NULL, 0);
8540 if (val < SD_LV_MAX)
8541 default_relax_domain_level = val;
8545 __setup("relax_domain_level=", setup_relax_domain_level);
8547 static void set_domain_attribute(struct sched_domain *sd,
8548 struct sched_domain_attr *attr)
8552 if (!attr || attr->relax_domain_level < 0) {
8553 if (default_relax_domain_level < 0)
8556 request = default_relax_domain_level;
8558 request = attr->relax_domain_level;
8559 if (request < sd->level) {
8560 /* turn off idle balance on this domain */
8561 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8563 /* turn on idle balance on this domain */
8564 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8568 static void __free_domain_allocs(struct s_data *d, enum s_alloc what,
8569 const struct cpumask *cpu_map)
8572 case sa_sched_groups:
8573 free_sched_groups(cpu_map, d->tmpmask); /* fall through */
8574 d->sched_group_nodes = NULL;
8576 free_rootdomain(d->rd); /* fall through */
8578 free_cpumask_var(d->tmpmask); /* fall through */
8579 case sa_send_covered:
8580 free_cpumask_var(d->send_covered); /* fall through */
8581 case sa_this_core_map:
8582 free_cpumask_var(d->this_core_map); /* fall through */
8583 case sa_this_sibling_map:
8584 free_cpumask_var(d->this_sibling_map); /* fall through */
8586 free_cpumask_var(d->nodemask); /* fall through */
8587 case sa_sched_group_nodes:
8589 kfree(d->sched_group_nodes); /* fall through */
8591 free_cpumask_var(d->notcovered); /* fall through */
8593 free_cpumask_var(d->covered); /* fall through */
8595 free_cpumask_var(d->domainspan); /* fall through */
8602 static enum s_alloc __visit_domain_allocation_hell(struct s_data *d,
8603 const struct cpumask *cpu_map)
8606 if (!alloc_cpumask_var(&d->domainspan, GFP_KERNEL))
8608 if (!alloc_cpumask_var(&d->covered, GFP_KERNEL))
8609 return sa_domainspan;
8610 if (!alloc_cpumask_var(&d->notcovered, GFP_KERNEL))
8612 /* Allocate the per-node list of sched groups */
8613 d->sched_group_nodes = kcalloc(nr_node_ids,
8614 sizeof(struct sched_group *), GFP_KERNEL);
8615 if (!d->sched_group_nodes) {
8616 printk(KERN_WARNING "Can not alloc sched group node list\n");
8617 return sa_notcovered;
8619 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = d->sched_group_nodes;
8621 if (!alloc_cpumask_var(&d->nodemask, GFP_KERNEL))
8622 return sa_sched_group_nodes;
8623 if (!alloc_cpumask_var(&d->this_sibling_map, GFP_KERNEL))
8625 if (!alloc_cpumask_var(&d->this_core_map, GFP_KERNEL))
8626 return sa_this_sibling_map;
8627 if (!alloc_cpumask_var(&d->send_covered, GFP_KERNEL))
8628 return sa_this_core_map;
8629 if (!alloc_cpumask_var(&d->tmpmask, GFP_KERNEL))
8630 return sa_send_covered;
8631 d->rd = alloc_rootdomain();
8633 printk(KERN_WARNING "Cannot alloc root domain\n");
8636 return sa_rootdomain;
8639 static struct sched_domain *__build_numa_sched_domains(struct s_data *d,
8640 const struct cpumask *cpu_map, struct sched_domain_attr *attr, int i)
8642 struct sched_domain *sd = NULL;
8644 struct sched_domain *parent;
8647 if (cpumask_weight(cpu_map) >
8648 SD_NODES_PER_DOMAIN * cpumask_weight(d->nodemask)) {
8649 sd = &per_cpu(allnodes_domains, i).sd;
8650 SD_INIT(sd, ALLNODES);
8651 set_domain_attribute(sd, attr);
8652 cpumask_copy(sched_domain_span(sd), cpu_map);
8653 cpu_to_allnodes_group(i, cpu_map, &sd->groups, d->tmpmask);
8658 sd = &per_cpu(node_domains, i).sd;
8660 set_domain_attribute(sd, attr);
8661 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8662 sd->parent = parent;
8665 cpumask_and(sched_domain_span(sd), sched_domain_span(sd), cpu_map);
8670 static struct sched_domain *__build_cpu_sched_domain(struct s_data *d,
8671 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8672 struct sched_domain *parent, int i)
8674 struct sched_domain *sd;
8675 sd = &per_cpu(phys_domains, i).sd;
8677 set_domain_attribute(sd, attr);
8678 cpumask_copy(sched_domain_span(sd), d->nodemask);
8679 sd->parent = parent;
8682 cpu_to_phys_group(i, cpu_map, &sd->groups, d->tmpmask);
8686 static struct sched_domain *__build_mc_sched_domain(struct s_data *d,
8687 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8688 struct sched_domain *parent, int i)
8690 struct sched_domain *sd = parent;
8691 #ifdef CONFIG_SCHED_MC
8692 sd = &per_cpu(core_domains, i).sd;
8694 set_domain_attribute(sd, attr);
8695 cpumask_and(sched_domain_span(sd), cpu_map, cpu_coregroup_mask(i));
8696 sd->parent = parent;
8698 cpu_to_core_group(i, cpu_map, &sd->groups, d->tmpmask);
8703 static struct sched_domain *__build_smt_sched_domain(struct s_data *d,
8704 const struct cpumask *cpu_map, struct sched_domain_attr *attr,
8705 struct sched_domain *parent, int i)
8707 struct sched_domain *sd = parent;
8708 #ifdef CONFIG_SCHED_SMT
8709 sd = &per_cpu(cpu_domains, i).sd;
8710 SD_INIT(sd, SIBLING);
8711 set_domain_attribute(sd, attr);
8712 cpumask_and(sched_domain_span(sd), cpu_map, topology_thread_cpumask(i));
8713 sd->parent = parent;
8715 cpu_to_cpu_group(i, cpu_map, &sd->groups, d->tmpmask);
8720 static void build_sched_groups(struct s_data *d, enum sched_domain_level l,
8721 const struct cpumask *cpu_map, int cpu)
8724 #ifdef CONFIG_SCHED_SMT
8725 case SD_LV_SIBLING: /* set up CPU (sibling) groups */
8726 cpumask_and(d->this_sibling_map, cpu_map,
8727 topology_thread_cpumask(cpu));
8728 if (cpu == cpumask_first(d->this_sibling_map))
8729 init_sched_build_groups(d->this_sibling_map, cpu_map,
8731 d->send_covered, d->tmpmask);
8734 #ifdef CONFIG_SCHED_MC
8735 case SD_LV_MC: /* set up multi-core groups */
8736 cpumask_and(d->this_core_map, cpu_map, cpu_coregroup_mask(cpu));
8737 if (cpu == cpumask_first(d->this_core_map))
8738 init_sched_build_groups(d->this_core_map, cpu_map,
8740 d->send_covered, d->tmpmask);
8743 case SD_LV_CPU: /* set up physical groups */
8744 cpumask_and(d->nodemask, cpumask_of_node(cpu), cpu_map);
8745 if (!cpumask_empty(d->nodemask))
8746 init_sched_build_groups(d->nodemask, cpu_map,
8748 d->send_covered, d->tmpmask);
8751 case SD_LV_ALLNODES:
8752 init_sched_build_groups(cpu_map, cpu_map, &cpu_to_allnodes_group,
8753 d->send_covered, d->tmpmask);
8762 * Build sched domains for a given set of cpus and attach the sched domains
8763 * to the individual cpus
8765 static int __build_sched_domains(const struct cpumask *cpu_map,
8766 struct sched_domain_attr *attr)
8768 enum s_alloc alloc_state = sa_none;
8770 struct sched_domain *sd;
8776 alloc_state = __visit_domain_allocation_hell(&d, cpu_map);
8777 if (alloc_state != sa_rootdomain)
8779 alloc_state = sa_sched_groups;
8782 * Set up domains for cpus specified by the cpu_map.
8784 for_each_cpu(i, cpu_map) {
8785 cpumask_and(d.nodemask, cpumask_of_node(cpu_to_node(i)),
8788 sd = __build_numa_sched_domains(&d, cpu_map, attr, i);
8789 sd = __build_cpu_sched_domain(&d, cpu_map, attr, sd, i);
8790 sd = __build_mc_sched_domain(&d, cpu_map, attr, sd, i);
8791 sd = __build_smt_sched_domain(&d, cpu_map, attr, sd, i);
8794 for_each_cpu(i, cpu_map) {
8795 build_sched_groups(&d, SD_LV_SIBLING, cpu_map, i);
8796 build_sched_groups(&d, SD_LV_MC, cpu_map, i);
8799 /* Set up physical groups */
8800 for (i = 0; i < nr_node_ids; i++)
8801 build_sched_groups(&d, SD_LV_CPU, cpu_map, i);
8804 /* Set up node groups */
8806 build_sched_groups(&d, SD_LV_ALLNODES, cpu_map, 0);
8808 for (i = 0; i < nr_node_ids; i++)
8809 if (build_numa_sched_groups(&d, cpu_map, i))
8813 /* Calculate CPU power for physical packages and nodes */
8814 #ifdef CONFIG_SCHED_SMT
8815 for_each_cpu(i, cpu_map) {
8816 sd = &per_cpu(cpu_domains, i).sd;
8817 init_sched_groups_power(i, sd);
8820 #ifdef CONFIG_SCHED_MC
8821 for_each_cpu(i, cpu_map) {
8822 sd = &per_cpu(core_domains, i).sd;
8823 init_sched_groups_power(i, sd);
8827 for_each_cpu(i, cpu_map) {
8828 sd = &per_cpu(phys_domains, i).sd;
8829 init_sched_groups_power(i, sd);
8833 for (i = 0; i < nr_node_ids; i++)
8834 init_numa_sched_groups_power(d.sched_group_nodes[i]);
8836 if (d.sd_allnodes) {
8837 struct sched_group *sg;
8839 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8841 init_numa_sched_groups_power(sg);
8845 /* Attach the domains */
8846 for_each_cpu(i, cpu_map) {
8847 #ifdef CONFIG_SCHED_SMT
8848 sd = &per_cpu(cpu_domains, i).sd;
8849 #elif defined(CONFIG_SCHED_MC)
8850 sd = &per_cpu(core_domains, i).sd;
8852 sd = &per_cpu(phys_domains, i).sd;
8854 cpu_attach_domain(sd, d.rd, i);
8857 d.sched_group_nodes = NULL; /* don't free this we still need it */
8858 __free_domain_allocs(&d, sa_tmpmask, cpu_map);
8862 __free_domain_allocs(&d, alloc_state, cpu_map);
8866 static int build_sched_domains(const struct cpumask *cpu_map)
8868 return __build_sched_domains(cpu_map, NULL);
8871 static struct cpumask *doms_cur; /* current sched domains */
8872 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8873 static struct sched_domain_attr *dattr_cur;
8874 /* attribues of custom domains in 'doms_cur' */
8877 * Special case: If a kmalloc of a doms_cur partition (array of
8878 * cpumask) fails, then fallback to a single sched domain,
8879 * as determined by the single cpumask fallback_doms.
8881 static cpumask_var_t fallback_doms;
8884 * arch_update_cpu_topology lets virtualized architectures update the
8885 * cpu core maps. It is supposed to return 1 if the topology changed
8886 * or 0 if it stayed the same.
8888 int __attribute__((weak)) arch_update_cpu_topology(void)
8894 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8895 * For now this just excludes isolated cpus, but could be used to
8896 * exclude other special cases in the future.
8898 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8902 arch_update_cpu_topology();
8904 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8906 doms_cur = fallback_doms;
8907 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8909 err = build_sched_domains(doms_cur);
8910 register_sched_domain_sysctl();
8915 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8916 struct cpumask *tmpmask)
8918 free_sched_groups(cpu_map, tmpmask);
8922 * Detach sched domains from a group of cpus specified in cpu_map
8923 * These cpus will now be attached to the NULL domain
8925 static void detach_destroy_domains(const struct cpumask *cpu_map)
8927 /* Save because hotplug lock held. */
8928 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8931 for_each_cpu(i, cpu_map)
8932 cpu_attach_domain(NULL, &def_root_domain, i);
8933 synchronize_sched();
8934 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8937 /* handle null as "default" */
8938 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8939 struct sched_domain_attr *new, int idx_new)
8941 struct sched_domain_attr tmp;
8948 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8949 new ? (new + idx_new) : &tmp,
8950 sizeof(struct sched_domain_attr));
8954 * Partition sched domains as specified by the 'ndoms_new'
8955 * cpumasks in the array doms_new[] of cpumasks. This compares
8956 * doms_new[] to the current sched domain partitioning, doms_cur[].
8957 * It destroys each deleted domain and builds each new domain.
8959 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8960 * The masks don't intersect (don't overlap.) We should setup one
8961 * sched domain for each mask. CPUs not in any of the cpumasks will
8962 * not be load balanced. If the same cpumask appears both in the
8963 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8966 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8967 * ownership of it and will kfree it when done with it. If the caller
8968 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8969 * ndoms_new == 1, and partition_sched_domains() will fallback to
8970 * the single partition 'fallback_doms', it also forces the domains
8973 * If doms_new == NULL it will be replaced with cpu_online_mask.
8974 * ndoms_new == 0 is a special case for destroying existing domains,
8975 * and it will not create the default domain.
8977 * Call with hotplug lock held
8979 /* FIXME: Change to struct cpumask *doms_new[] */
8980 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8981 struct sched_domain_attr *dattr_new)
8986 mutex_lock(&sched_domains_mutex);
8988 /* always unregister in case we don't destroy any domains */
8989 unregister_sched_domain_sysctl();
8991 /* Let architecture update cpu core mappings. */
8992 new_topology = arch_update_cpu_topology();
8994 n = doms_new ? ndoms_new : 0;
8996 /* Destroy deleted domains */
8997 for (i = 0; i < ndoms_cur; i++) {
8998 for (j = 0; j < n && !new_topology; j++) {
8999 if (cpumask_equal(&doms_cur[i], &doms_new[j])
9000 && dattrs_equal(dattr_cur, i, dattr_new, j))
9003 /* no match - a current sched domain not in new doms_new[] */
9004 detach_destroy_domains(doms_cur + i);
9009 if (doms_new == NULL) {
9011 doms_new = fallback_doms;
9012 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
9013 WARN_ON_ONCE(dattr_new);
9016 /* Build new domains */
9017 for (i = 0; i < ndoms_new; i++) {
9018 for (j = 0; j < ndoms_cur && !new_topology; j++) {
9019 if (cpumask_equal(&doms_new[i], &doms_cur[j])
9020 && dattrs_equal(dattr_new, i, dattr_cur, j))
9023 /* no match - add a new doms_new */
9024 __build_sched_domains(doms_new + i,
9025 dattr_new ? dattr_new + i : NULL);
9030 /* Remember the new sched domains */
9031 if (doms_cur != fallback_doms)
9033 kfree(dattr_cur); /* kfree(NULL) is safe */
9034 doms_cur = doms_new;
9035 dattr_cur = dattr_new;
9036 ndoms_cur = ndoms_new;
9038 register_sched_domain_sysctl();
9040 mutex_unlock(&sched_domains_mutex);
9043 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9044 static void arch_reinit_sched_domains(void)
9048 /* Destroy domains first to force the rebuild */
9049 partition_sched_domains(0, NULL, NULL);
9051 rebuild_sched_domains();
9055 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
9057 unsigned int level = 0;
9059 if (sscanf(buf, "%u", &level) != 1)
9063 * level is always be positive so don't check for
9064 * level < POWERSAVINGS_BALANCE_NONE which is 0
9065 * What happens on 0 or 1 byte write,
9066 * need to check for count as well?
9069 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
9073 sched_smt_power_savings = level;
9075 sched_mc_power_savings = level;
9077 arch_reinit_sched_domains();
9082 #ifdef CONFIG_SCHED_MC
9083 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
9086 return sprintf(page, "%u\n", sched_mc_power_savings);
9088 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
9089 const char *buf, size_t count)
9091 return sched_power_savings_store(buf, count, 0);
9093 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
9094 sched_mc_power_savings_show,
9095 sched_mc_power_savings_store);
9098 #ifdef CONFIG_SCHED_SMT
9099 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
9102 return sprintf(page, "%u\n", sched_smt_power_savings);
9104 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
9105 const char *buf, size_t count)
9107 return sched_power_savings_store(buf, count, 1);
9109 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
9110 sched_smt_power_savings_show,
9111 sched_smt_power_savings_store);
9114 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
9118 #ifdef CONFIG_SCHED_SMT
9120 err = sysfs_create_file(&cls->kset.kobj,
9121 &attr_sched_smt_power_savings.attr);
9123 #ifdef CONFIG_SCHED_MC
9124 if (!err && mc_capable())
9125 err = sysfs_create_file(&cls->kset.kobj,
9126 &attr_sched_mc_power_savings.attr);
9130 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9132 #ifndef CONFIG_CPUSETS
9134 * Add online and remove offline CPUs from the scheduler domains.
9135 * When cpusets are enabled they take over this function.
9137 static int update_sched_domains(struct notifier_block *nfb,
9138 unsigned long action, void *hcpu)
9142 case CPU_ONLINE_FROZEN:
9144 case CPU_DEAD_FROZEN:
9145 partition_sched_domains(1, NULL, NULL);
9154 static int update_runtime(struct notifier_block *nfb,
9155 unsigned long action, void *hcpu)
9157 int cpu = (int)(long)hcpu;
9160 case CPU_DOWN_PREPARE:
9161 case CPU_DOWN_PREPARE_FROZEN:
9162 disable_runtime(cpu_rq(cpu));
9165 case CPU_DOWN_FAILED:
9166 case CPU_DOWN_FAILED_FROZEN:
9168 case CPU_ONLINE_FROZEN:
9169 enable_runtime(cpu_rq(cpu));
9177 void __init sched_init_smp(void)
9179 cpumask_var_t non_isolated_cpus;
9181 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
9183 #if defined(CONFIG_NUMA)
9184 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
9186 BUG_ON(sched_group_nodes_bycpu == NULL);
9189 mutex_lock(&sched_domains_mutex);
9190 arch_init_sched_domains(cpu_online_mask);
9191 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
9192 if (cpumask_empty(non_isolated_cpus))
9193 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
9194 mutex_unlock(&sched_domains_mutex);
9197 #ifndef CONFIG_CPUSETS
9198 /* XXX: Theoretical race here - CPU may be hotplugged now */
9199 hotcpu_notifier(update_sched_domains, 0);
9202 /* RT runtime code needs to handle some hotplug events */
9203 hotcpu_notifier(update_runtime, 0);
9207 /* Move init over to a non-isolated CPU */
9208 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
9210 sched_init_granularity();
9211 free_cpumask_var(non_isolated_cpus);
9213 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
9214 init_sched_rt_class();
9217 void __init sched_init_smp(void)
9219 sched_init_granularity();
9221 #endif /* CONFIG_SMP */
9223 const_debug unsigned int sysctl_timer_migration = 1;
9225 int in_sched_functions(unsigned long addr)
9227 return in_lock_functions(addr) ||
9228 (addr >= (unsigned long)__sched_text_start
9229 && addr < (unsigned long)__sched_text_end);
9232 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
9234 cfs_rq->tasks_timeline = RB_ROOT;
9235 INIT_LIST_HEAD(&cfs_rq->tasks);
9236 #ifdef CONFIG_FAIR_GROUP_SCHED
9239 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
9242 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
9244 struct rt_prio_array *array;
9247 array = &rt_rq->active;
9248 for (i = 0; i < MAX_RT_PRIO; i++) {
9249 INIT_LIST_HEAD(array->queue + i);
9250 __clear_bit(i, array->bitmap);
9252 /* delimiter for bitsearch: */
9253 __set_bit(MAX_RT_PRIO, array->bitmap);
9255 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9256 rt_rq->highest_prio.curr = MAX_RT_PRIO;
9258 rt_rq->highest_prio.next = MAX_RT_PRIO;
9262 rt_rq->rt_nr_migratory = 0;
9263 rt_rq->overloaded = 0;
9264 plist_head_init(&rt_rq->pushable_tasks, &rq->lock);
9268 rt_rq->rt_throttled = 0;
9269 rt_rq->rt_runtime = 0;
9270 spin_lock_init(&rt_rq->rt_runtime_lock);
9272 #ifdef CONFIG_RT_GROUP_SCHED
9273 rt_rq->rt_nr_boosted = 0;
9278 #ifdef CONFIG_FAIR_GROUP_SCHED
9279 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
9280 struct sched_entity *se, int cpu, int add,
9281 struct sched_entity *parent)
9283 struct rq *rq = cpu_rq(cpu);
9284 tg->cfs_rq[cpu] = cfs_rq;
9285 init_cfs_rq(cfs_rq, rq);
9288 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
9291 /* se could be NULL for init_task_group */
9296 se->cfs_rq = &rq->cfs;
9298 se->cfs_rq = parent->my_q;
9301 se->load.weight = tg->shares;
9302 se->load.inv_weight = 0;
9303 se->parent = parent;
9307 #ifdef CONFIG_RT_GROUP_SCHED
9308 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9309 struct sched_rt_entity *rt_se, int cpu, int add,
9310 struct sched_rt_entity *parent)
9312 struct rq *rq = cpu_rq(cpu);
9314 tg->rt_rq[cpu] = rt_rq;
9315 init_rt_rq(rt_rq, rq);
9317 rt_rq->rt_se = rt_se;
9318 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9320 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9322 tg->rt_se[cpu] = rt_se;
9327 rt_se->rt_rq = &rq->rt;
9329 rt_se->rt_rq = parent->my_q;
9331 rt_se->my_q = rt_rq;
9332 rt_se->parent = parent;
9333 INIT_LIST_HEAD(&rt_se->run_list);
9337 void __init sched_init(void)
9340 unsigned long alloc_size = 0, ptr;
9342 #ifdef CONFIG_FAIR_GROUP_SCHED
9343 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9345 #ifdef CONFIG_RT_GROUP_SCHED
9346 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9348 #ifdef CONFIG_USER_SCHED
9351 #ifdef CONFIG_CPUMASK_OFFSTACK
9352 alloc_size += num_possible_cpus() * cpumask_size();
9355 * As sched_init() is called before page_alloc is setup,
9356 * we use alloc_bootmem().
9359 ptr = (unsigned long)kzalloc(alloc_size, GFP_NOWAIT);
9361 #ifdef CONFIG_FAIR_GROUP_SCHED
9362 init_task_group.se = (struct sched_entity **)ptr;
9363 ptr += nr_cpu_ids * sizeof(void **);
9365 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9366 ptr += nr_cpu_ids * sizeof(void **);
9368 #ifdef CONFIG_USER_SCHED
9369 root_task_group.se = (struct sched_entity **)ptr;
9370 ptr += nr_cpu_ids * sizeof(void **);
9372 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9373 ptr += nr_cpu_ids * sizeof(void **);
9374 #endif /* CONFIG_USER_SCHED */
9375 #endif /* CONFIG_FAIR_GROUP_SCHED */
9376 #ifdef CONFIG_RT_GROUP_SCHED
9377 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9378 ptr += nr_cpu_ids * sizeof(void **);
9380 init_task_group.rt_rq = (struct rt_rq **)ptr;
9381 ptr += nr_cpu_ids * sizeof(void **);
9383 #ifdef CONFIG_USER_SCHED
9384 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9385 ptr += nr_cpu_ids * sizeof(void **);
9387 root_task_group.rt_rq = (struct rt_rq **)ptr;
9388 ptr += nr_cpu_ids * sizeof(void **);
9389 #endif /* CONFIG_USER_SCHED */
9390 #endif /* CONFIG_RT_GROUP_SCHED */
9391 #ifdef CONFIG_CPUMASK_OFFSTACK
9392 for_each_possible_cpu(i) {
9393 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9394 ptr += cpumask_size();
9396 #endif /* CONFIG_CPUMASK_OFFSTACK */
9400 init_defrootdomain();
9403 init_rt_bandwidth(&def_rt_bandwidth,
9404 global_rt_period(), global_rt_runtime());
9406 #ifdef CONFIG_RT_GROUP_SCHED
9407 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9408 global_rt_period(), global_rt_runtime());
9409 #ifdef CONFIG_USER_SCHED
9410 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9411 global_rt_period(), RUNTIME_INF);
9412 #endif /* CONFIG_USER_SCHED */
9413 #endif /* CONFIG_RT_GROUP_SCHED */
9415 #ifdef CONFIG_GROUP_SCHED
9416 list_add(&init_task_group.list, &task_groups);
9417 INIT_LIST_HEAD(&init_task_group.children);
9419 #ifdef CONFIG_USER_SCHED
9420 INIT_LIST_HEAD(&root_task_group.children);
9421 init_task_group.parent = &root_task_group;
9422 list_add(&init_task_group.siblings, &root_task_group.children);
9423 #endif /* CONFIG_USER_SCHED */
9424 #endif /* CONFIG_GROUP_SCHED */
9426 for_each_possible_cpu(i) {
9430 spin_lock_init(&rq->lock);
9432 rq->calc_load_active = 0;
9433 rq->calc_load_update = jiffies + LOAD_FREQ;
9434 init_cfs_rq(&rq->cfs, rq);
9435 init_rt_rq(&rq->rt, rq);
9436 #ifdef CONFIG_FAIR_GROUP_SCHED
9437 init_task_group.shares = init_task_group_load;
9438 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9439 #ifdef CONFIG_CGROUP_SCHED
9441 * How much cpu bandwidth does init_task_group get?
9443 * In case of task-groups formed thr' the cgroup filesystem, it
9444 * gets 100% of the cpu resources in the system. This overall
9445 * system cpu resource is divided among the tasks of
9446 * init_task_group and its child task-groups in a fair manner,
9447 * based on each entity's (task or task-group's) weight
9448 * (se->load.weight).
9450 * In other words, if init_task_group has 10 tasks of weight
9451 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9452 * then A0's share of the cpu resource is:
9454 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9456 * We achieve this by letting init_task_group's tasks sit
9457 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9459 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9460 #elif defined CONFIG_USER_SCHED
9461 root_task_group.shares = NICE_0_LOAD;
9462 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9464 * In case of task-groups formed thr' the user id of tasks,
9465 * init_task_group represents tasks belonging to root user.
9466 * Hence it forms a sibling of all subsequent groups formed.
9467 * In this case, init_task_group gets only a fraction of overall
9468 * system cpu resource, based on the weight assigned to root
9469 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9470 * by letting tasks of init_task_group sit in a separate cfs_rq
9471 * (init_tg_cfs_rq) and having one entity represent this group of
9472 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9474 init_tg_cfs_entry(&init_task_group,
9475 &per_cpu(init_tg_cfs_rq, i),
9476 &per_cpu(init_sched_entity, i), i, 1,
9477 root_task_group.se[i]);
9480 #endif /* CONFIG_FAIR_GROUP_SCHED */
9482 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9483 #ifdef CONFIG_RT_GROUP_SCHED
9484 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9485 #ifdef CONFIG_CGROUP_SCHED
9486 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9487 #elif defined CONFIG_USER_SCHED
9488 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9489 init_tg_rt_entry(&init_task_group,
9490 &per_cpu(init_rt_rq, i),
9491 &per_cpu(init_sched_rt_entity, i), i, 1,
9492 root_task_group.rt_se[i]);
9496 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9497 rq->cpu_load[j] = 0;
9501 rq->post_schedule = 0;
9502 rq->active_balance = 0;
9503 rq->next_balance = jiffies;
9507 rq->migration_thread = NULL;
9508 INIT_LIST_HEAD(&rq->migration_queue);
9509 rq_attach_root(rq, &def_root_domain);
9512 atomic_set(&rq->nr_iowait, 0);
9515 set_load_weight(&init_task);
9517 #ifdef CONFIG_PREEMPT_NOTIFIERS
9518 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9522 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9525 #ifdef CONFIG_RT_MUTEXES
9526 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9530 * The boot idle thread does lazy MMU switching as well:
9532 atomic_inc(&init_mm.mm_count);
9533 enter_lazy_tlb(&init_mm, current);
9536 * Make us the idle thread. Technically, schedule() should not be
9537 * called from this thread, however somewhere below it might be,
9538 * but because we are the idle thread, we just pick up running again
9539 * when this runqueue becomes "idle".
9541 init_idle(current, smp_processor_id());
9543 calc_load_update = jiffies + LOAD_FREQ;
9546 * During early bootup we pretend to be a normal task:
9548 current->sched_class = &fair_sched_class;
9550 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9551 alloc_cpumask_var(&nohz_cpu_mask, GFP_NOWAIT);
9554 alloc_cpumask_var(&nohz.cpu_mask, GFP_NOWAIT);
9555 alloc_cpumask_var(&nohz.ilb_grp_nohz_mask, GFP_NOWAIT);
9557 alloc_cpumask_var(&cpu_isolated_map, GFP_NOWAIT);
9560 perf_counter_init();
9562 scheduler_running = 1;
9565 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9566 static inline int preempt_count_equals(int preempt_offset)
9568 int nested = preempt_count() & ~PREEMPT_ACTIVE;
9570 return (nested == PREEMPT_INATOMIC_BASE + preempt_offset);
9573 void __might_sleep(char *file, int line, int preempt_offset)
9576 static unsigned long prev_jiffy; /* ratelimiting */
9578 if ((preempt_count_equals(preempt_offset) && !irqs_disabled()) ||
9579 system_state != SYSTEM_RUNNING || oops_in_progress)
9581 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9583 prev_jiffy = jiffies;
9586 "BUG: sleeping function called from invalid context at %s:%d\n",
9589 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9590 in_atomic(), irqs_disabled(),
9591 current->pid, current->comm);
9593 debug_show_held_locks(current);
9594 if (irqs_disabled())
9595 print_irqtrace_events(current);
9599 EXPORT_SYMBOL(__might_sleep);
9602 #ifdef CONFIG_MAGIC_SYSRQ
9603 static void normalize_task(struct rq *rq, struct task_struct *p)
9607 update_rq_clock(rq);
9608 on_rq = p->se.on_rq;
9610 deactivate_task(rq, p, 0);
9611 __setscheduler(rq, p, SCHED_NORMAL, 0);
9613 activate_task(rq, p, 0);
9614 resched_task(rq->curr);
9618 void normalize_rt_tasks(void)
9620 struct task_struct *g, *p;
9621 unsigned long flags;
9624 read_lock_irqsave(&tasklist_lock, flags);
9625 do_each_thread(g, p) {
9627 * Only normalize user tasks:
9632 p->se.exec_start = 0;
9633 #ifdef CONFIG_SCHEDSTATS
9634 p->se.wait_start = 0;
9635 p->se.sleep_start = 0;
9636 p->se.block_start = 0;
9641 * Renice negative nice level userspace
9644 if (TASK_NICE(p) < 0 && p->mm)
9645 set_user_nice(p, 0);
9649 spin_lock(&p->pi_lock);
9650 rq = __task_rq_lock(p);
9652 normalize_task(rq, p);
9654 __task_rq_unlock(rq);
9655 spin_unlock(&p->pi_lock);
9656 } while_each_thread(g, p);
9658 read_unlock_irqrestore(&tasklist_lock, flags);
9661 #endif /* CONFIG_MAGIC_SYSRQ */
9665 * These functions are only useful for the IA64 MCA handling.
9667 * They can only be called when the whole system has been
9668 * stopped - every CPU needs to be quiescent, and no scheduling
9669 * activity can take place. Using them for anything else would
9670 * be a serious bug, and as a result, they aren't even visible
9671 * under any other configuration.
9675 * curr_task - return the current task for a given cpu.
9676 * @cpu: the processor in question.
9678 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9680 struct task_struct *curr_task(int cpu)
9682 return cpu_curr(cpu);
9686 * set_curr_task - set the current task for a given cpu.
9687 * @cpu: the processor in question.
9688 * @p: the task pointer to set.
9690 * Description: This function must only be used when non-maskable interrupts
9691 * are serviced on a separate stack. It allows the architecture to switch the
9692 * notion of the current task on a cpu in a non-blocking manner. This function
9693 * must be called with all CPU's synchronized, and interrupts disabled, the
9694 * and caller must save the original value of the current task (see
9695 * curr_task() above) and restore that value before reenabling interrupts and
9696 * re-starting the system.
9698 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9700 void set_curr_task(int cpu, struct task_struct *p)
9707 #ifdef CONFIG_FAIR_GROUP_SCHED
9708 static void free_fair_sched_group(struct task_group *tg)
9712 for_each_possible_cpu(i) {
9714 kfree(tg->cfs_rq[i]);
9724 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9726 struct cfs_rq *cfs_rq;
9727 struct sched_entity *se;
9731 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9734 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9738 tg->shares = NICE_0_LOAD;
9740 for_each_possible_cpu(i) {
9743 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9744 GFP_KERNEL, cpu_to_node(i));
9748 se = kzalloc_node(sizeof(struct sched_entity),
9749 GFP_KERNEL, cpu_to_node(i));
9753 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9762 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9764 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9765 &cpu_rq(cpu)->leaf_cfs_rq_list);
9768 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9770 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9772 #else /* !CONFG_FAIR_GROUP_SCHED */
9773 static inline void free_fair_sched_group(struct task_group *tg)
9778 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9783 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9787 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9790 #endif /* CONFIG_FAIR_GROUP_SCHED */
9792 #ifdef CONFIG_RT_GROUP_SCHED
9793 static void free_rt_sched_group(struct task_group *tg)
9797 destroy_rt_bandwidth(&tg->rt_bandwidth);
9799 for_each_possible_cpu(i) {
9801 kfree(tg->rt_rq[i]);
9803 kfree(tg->rt_se[i]);
9811 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9813 struct rt_rq *rt_rq;
9814 struct sched_rt_entity *rt_se;
9818 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9821 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9825 init_rt_bandwidth(&tg->rt_bandwidth,
9826 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9828 for_each_possible_cpu(i) {
9831 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9832 GFP_KERNEL, cpu_to_node(i));
9836 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9837 GFP_KERNEL, cpu_to_node(i));
9841 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9850 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9852 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9853 &cpu_rq(cpu)->leaf_rt_rq_list);
9856 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9858 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9860 #else /* !CONFIG_RT_GROUP_SCHED */
9861 static inline void free_rt_sched_group(struct task_group *tg)
9866 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9871 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9875 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9878 #endif /* CONFIG_RT_GROUP_SCHED */
9880 #ifdef CONFIG_GROUP_SCHED
9881 static void free_sched_group(struct task_group *tg)
9883 free_fair_sched_group(tg);
9884 free_rt_sched_group(tg);
9888 /* allocate runqueue etc for a new task group */
9889 struct task_group *sched_create_group(struct task_group *parent)
9891 struct task_group *tg;
9892 unsigned long flags;
9895 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9897 return ERR_PTR(-ENOMEM);
9899 if (!alloc_fair_sched_group(tg, parent))
9902 if (!alloc_rt_sched_group(tg, parent))
9905 spin_lock_irqsave(&task_group_lock, flags);
9906 for_each_possible_cpu(i) {
9907 register_fair_sched_group(tg, i);
9908 register_rt_sched_group(tg, i);
9910 list_add_rcu(&tg->list, &task_groups);
9912 WARN_ON(!parent); /* root should already exist */
9914 tg->parent = parent;
9915 INIT_LIST_HEAD(&tg->children);
9916 list_add_rcu(&tg->siblings, &parent->children);
9917 spin_unlock_irqrestore(&task_group_lock, flags);
9922 free_sched_group(tg);
9923 return ERR_PTR(-ENOMEM);
9926 /* rcu callback to free various structures associated with a task group */
9927 static void free_sched_group_rcu(struct rcu_head *rhp)
9929 /* now it should be safe to free those cfs_rqs */
9930 free_sched_group(container_of(rhp, struct task_group, rcu));
9933 /* Destroy runqueue etc associated with a task group */
9934 void sched_destroy_group(struct task_group *tg)
9936 unsigned long flags;
9939 spin_lock_irqsave(&task_group_lock, flags);
9940 for_each_possible_cpu(i) {
9941 unregister_fair_sched_group(tg, i);
9942 unregister_rt_sched_group(tg, i);
9944 list_del_rcu(&tg->list);
9945 list_del_rcu(&tg->siblings);
9946 spin_unlock_irqrestore(&task_group_lock, flags);
9948 /* wait for possible concurrent references to cfs_rqs complete */
9949 call_rcu(&tg->rcu, free_sched_group_rcu);
9952 /* change task's runqueue when it moves between groups.
9953 * The caller of this function should have put the task in its new group
9954 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9955 * reflect its new group.
9957 void sched_move_task(struct task_struct *tsk)
9960 unsigned long flags;
9963 rq = task_rq_lock(tsk, &flags);
9965 update_rq_clock(rq);
9967 running = task_current(rq, tsk);
9968 on_rq = tsk->se.on_rq;
9971 dequeue_task(rq, tsk, 0);
9972 if (unlikely(running))
9973 tsk->sched_class->put_prev_task(rq, tsk);
9975 set_task_rq(tsk, task_cpu(tsk));
9977 #ifdef CONFIG_FAIR_GROUP_SCHED
9978 if (tsk->sched_class->moved_group)
9979 tsk->sched_class->moved_group(tsk);
9982 if (unlikely(running))
9983 tsk->sched_class->set_curr_task(rq);
9985 enqueue_task(rq, tsk, 0);
9987 task_rq_unlock(rq, &flags);
9989 #endif /* CONFIG_GROUP_SCHED */
9991 #ifdef CONFIG_FAIR_GROUP_SCHED
9992 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9994 struct cfs_rq *cfs_rq = se->cfs_rq;
9999 dequeue_entity(cfs_rq, se, 0);
10001 se->load.weight = shares;
10002 se->load.inv_weight = 0;
10005 enqueue_entity(cfs_rq, se, 0);
10008 static void set_se_shares(struct sched_entity *se, unsigned long shares)
10010 struct cfs_rq *cfs_rq = se->cfs_rq;
10011 struct rq *rq = cfs_rq->rq;
10012 unsigned long flags;
10014 spin_lock_irqsave(&rq->lock, flags);
10015 __set_se_shares(se, shares);
10016 spin_unlock_irqrestore(&rq->lock, flags);
10019 static DEFINE_MUTEX(shares_mutex);
10021 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
10024 unsigned long flags;
10027 * We can't change the weight of the root cgroup.
10032 if (shares < MIN_SHARES)
10033 shares = MIN_SHARES;
10034 else if (shares > MAX_SHARES)
10035 shares = MAX_SHARES;
10037 mutex_lock(&shares_mutex);
10038 if (tg->shares == shares)
10041 spin_lock_irqsave(&task_group_lock, flags);
10042 for_each_possible_cpu(i)
10043 unregister_fair_sched_group(tg, i);
10044 list_del_rcu(&tg->siblings);
10045 spin_unlock_irqrestore(&task_group_lock, flags);
10047 /* wait for any ongoing reference to this group to finish */
10048 synchronize_sched();
10051 * Now we are free to modify the group's share on each cpu
10052 * w/o tripping rebalance_share or load_balance_fair.
10054 tg->shares = shares;
10055 for_each_possible_cpu(i) {
10057 * force a rebalance
10059 cfs_rq_set_shares(tg->cfs_rq[i], 0);
10060 set_se_shares(tg->se[i], shares);
10064 * Enable load balance activity on this group, by inserting it back on
10065 * each cpu's rq->leaf_cfs_rq_list.
10067 spin_lock_irqsave(&task_group_lock, flags);
10068 for_each_possible_cpu(i)
10069 register_fair_sched_group(tg, i);
10070 list_add_rcu(&tg->siblings, &tg->parent->children);
10071 spin_unlock_irqrestore(&task_group_lock, flags);
10073 mutex_unlock(&shares_mutex);
10077 unsigned long sched_group_shares(struct task_group *tg)
10083 #ifdef CONFIG_RT_GROUP_SCHED
10085 * Ensure that the real time constraints are schedulable.
10087 static DEFINE_MUTEX(rt_constraints_mutex);
10089 static unsigned long to_ratio(u64 period, u64 runtime)
10091 if (runtime == RUNTIME_INF)
10094 return div64_u64(runtime << 20, period);
10097 /* Must be called with tasklist_lock held */
10098 static inline int tg_has_rt_tasks(struct task_group *tg)
10100 struct task_struct *g, *p;
10102 do_each_thread(g, p) {
10103 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
10105 } while_each_thread(g, p);
10110 struct rt_schedulable_data {
10111 struct task_group *tg;
10116 static int tg_schedulable(struct task_group *tg, void *data)
10118 struct rt_schedulable_data *d = data;
10119 struct task_group *child;
10120 unsigned long total, sum = 0;
10121 u64 period, runtime;
10123 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10124 runtime = tg->rt_bandwidth.rt_runtime;
10127 period = d->rt_period;
10128 runtime = d->rt_runtime;
10131 #ifdef CONFIG_USER_SCHED
10132 if (tg == &root_task_group) {
10133 period = global_rt_period();
10134 runtime = global_rt_runtime();
10139 * Cannot have more runtime than the period.
10141 if (runtime > period && runtime != RUNTIME_INF)
10145 * Ensure we don't starve existing RT tasks.
10147 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
10150 total = to_ratio(period, runtime);
10153 * Nobody can have more than the global setting allows.
10155 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
10159 * The sum of our children's runtime should not exceed our own.
10161 list_for_each_entry_rcu(child, &tg->children, siblings) {
10162 period = ktime_to_ns(child->rt_bandwidth.rt_period);
10163 runtime = child->rt_bandwidth.rt_runtime;
10165 if (child == d->tg) {
10166 period = d->rt_period;
10167 runtime = d->rt_runtime;
10170 sum += to_ratio(period, runtime);
10179 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
10181 struct rt_schedulable_data data = {
10183 .rt_period = period,
10184 .rt_runtime = runtime,
10187 return walk_tg_tree(tg_schedulable, tg_nop, &data);
10190 static int tg_set_bandwidth(struct task_group *tg,
10191 u64 rt_period, u64 rt_runtime)
10195 mutex_lock(&rt_constraints_mutex);
10196 read_lock(&tasklist_lock);
10197 err = __rt_schedulable(tg, rt_period, rt_runtime);
10201 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10202 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
10203 tg->rt_bandwidth.rt_runtime = rt_runtime;
10205 for_each_possible_cpu(i) {
10206 struct rt_rq *rt_rq = tg->rt_rq[i];
10208 spin_lock(&rt_rq->rt_runtime_lock);
10209 rt_rq->rt_runtime = rt_runtime;
10210 spin_unlock(&rt_rq->rt_runtime_lock);
10212 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
10214 read_unlock(&tasklist_lock);
10215 mutex_unlock(&rt_constraints_mutex);
10220 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
10222 u64 rt_runtime, rt_period;
10224 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
10225 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
10226 if (rt_runtime_us < 0)
10227 rt_runtime = RUNTIME_INF;
10229 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10232 long sched_group_rt_runtime(struct task_group *tg)
10236 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
10239 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
10240 do_div(rt_runtime_us, NSEC_PER_USEC);
10241 return rt_runtime_us;
10244 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
10246 u64 rt_runtime, rt_period;
10248 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
10249 rt_runtime = tg->rt_bandwidth.rt_runtime;
10251 if (rt_period == 0)
10254 return tg_set_bandwidth(tg, rt_period, rt_runtime);
10257 long sched_group_rt_period(struct task_group *tg)
10261 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
10262 do_div(rt_period_us, NSEC_PER_USEC);
10263 return rt_period_us;
10266 static int sched_rt_global_constraints(void)
10268 u64 runtime, period;
10271 if (sysctl_sched_rt_period <= 0)
10274 runtime = global_rt_runtime();
10275 period = global_rt_period();
10278 * Sanity check on the sysctl variables.
10280 if (runtime > period && runtime != RUNTIME_INF)
10283 mutex_lock(&rt_constraints_mutex);
10284 read_lock(&tasklist_lock);
10285 ret = __rt_schedulable(NULL, 0, 0);
10286 read_unlock(&tasklist_lock);
10287 mutex_unlock(&rt_constraints_mutex);
10292 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
10294 /* Don't accept realtime tasks when there is no way for them to run */
10295 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
10301 #else /* !CONFIG_RT_GROUP_SCHED */
10302 static int sched_rt_global_constraints(void)
10304 unsigned long flags;
10307 if (sysctl_sched_rt_period <= 0)
10311 * There's always some RT tasks in the root group
10312 * -- migration, kstopmachine etc..
10314 if (sysctl_sched_rt_runtime == 0)
10317 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10318 for_each_possible_cpu(i) {
10319 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10321 spin_lock(&rt_rq->rt_runtime_lock);
10322 rt_rq->rt_runtime = global_rt_runtime();
10323 spin_unlock(&rt_rq->rt_runtime_lock);
10325 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10329 #endif /* CONFIG_RT_GROUP_SCHED */
10331 int sched_rt_handler(struct ctl_table *table, int write,
10332 struct file *filp, void __user *buffer, size_t *lenp,
10336 int old_period, old_runtime;
10337 static DEFINE_MUTEX(mutex);
10339 mutex_lock(&mutex);
10340 old_period = sysctl_sched_rt_period;
10341 old_runtime = sysctl_sched_rt_runtime;
10343 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10345 if (!ret && write) {
10346 ret = sched_rt_global_constraints();
10348 sysctl_sched_rt_period = old_period;
10349 sysctl_sched_rt_runtime = old_runtime;
10351 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10352 def_rt_bandwidth.rt_period =
10353 ns_to_ktime(global_rt_period());
10356 mutex_unlock(&mutex);
10361 #ifdef CONFIG_CGROUP_SCHED
10363 /* return corresponding task_group object of a cgroup */
10364 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10366 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10367 struct task_group, css);
10370 static struct cgroup_subsys_state *
10371 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10373 struct task_group *tg, *parent;
10375 if (!cgrp->parent) {
10376 /* This is early initialization for the top cgroup */
10377 return &init_task_group.css;
10380 parent = cgroup_tg(cgrp->parent);
10381 tg = sched_create_group(parent);
10383 return ERR_PTR(-ENOMEM);
10389 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10391 struct task_group *tg = cgroup_tg(cgrp);
10393 sched_destroy_group(tg);
10397 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10398 struct task_struct *tsk)
10400 #ifdef CONFIG_RT_GROUP_SCHED
10401 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10404 /* We don't support RT-tasks being in separate groups */
10405 if (tsk->sched_class != &fair_sched_class)
10413 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10414 struct cgroup *old_cont, struct task_struct *tsk)
10416 sched_move_task(tsk);
10419 #ifdef CONFIG_FAIR_GROUP_SCHED
10420 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10423 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10426 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10428 struct task_group *tg = cgroup_tg(cgrp);
10430 return (u64) tg->shares;
10432 #endif /* CONFIG_FAIR_GROUP_SCHED */
10434 #ifdef CONFIG_RT_GROUP_SCHED
10435 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10438 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10441 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10443 return sched_group_rt_runtime(cgroup_tg(cgrp));
10446 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10449 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10452 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10454 return sched_group_rt_period(cgroup_tg(cgrp));
10456 #endif /* CONFIG_RT_GROUP_SCHED */
10458 static struct cftype cpu_files[] = {
10459 #ifdef CONFIG_FAIR_GROUP_SCHED
10462 .read_u64 = cpu_shares_read_u64,
10463 .write_u64 = cpu_shares_write_u64,
10466 #ifdef CONFIG_RT_GROUP_SCHED
10468 .name = "rt_runtime_us",
10469 .read_s64 = cpu_rt_runtime_read,
10470 .write_s64 = cpu_rt_runtime_write,
10473 .name = "rt_period_us",
10474 .read_u64 = cpu_rt_period_read_uint,
10475 .write_u64 = cpu_rt_period_write_uint,
10480 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10482 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10485 struct cgroup_subsys cpu_cgroup_subsys = {
10487 .create = cpu_cgroup_create,
10488 .destroy = cpu_cgroup_destroy,
10489 .can_attach = cpu_cgroup_can_attach,
10490 .attach = cpu_cgroup_attach,
10491 .populate = cpu_cgroup_populate,
10492 .subsys_id = cpu_cgroup_subsys_id,
10496 #endif /* CONFIG_CGROUP_SCHED */
10498 #ifdef CONFIG_CGROUP_CPUACCT
10501 * CPU accounting code for task groups.
10503 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10504 * (balbir@in.ibm.com).
10507 /* track cpu usage of a group of tasks and its child groups */
10509 struct cgroup_subsys_state css;
10510 /* cpuusage holds pointer to a u64-type object on every cpu */
10512 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10513 struct cpuacct *parent;
10516 struct cgroup_subsys cpuacct_subsys;
10518 /* return cpu accounting group corresponding to this container */
10519 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10521 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10522 struct cpuacct, css);
10525 /* return cpu accounting group to which this task belongs */
10526 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10528 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10529 struct cpuacct, css);
10532 /* create a new cpu accounting group */
10533 static struct cgroup_subsys_state *cpuacct_create(
10534 struct cgroup_subsys *ss, struct cgroup *cgrp)
10536 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10542 ca->cpuusage = alloc_percpu(u64);
10546 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10547 if (percpu_counter_init(&ca->cpustat[i], 0))
10548 goto out_free_counters;
10551 ca->parent = cgroup_ca(cgrp->parent);
10557 percpu_counter_destroy(&ca->cpustat[i]);
10558 free_percpu(ca->cpuusage);
10562 return ERR_PTR(-ENOMEM);
10565 /* destroy an existing cpu accounting group */
10567 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10569 struct cpuacct *ca = cgroup_ca(cgrp);
10572 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10573 percpu_counter_destroy(&ca->cpustat[i]);
10574 free_percpu(ca->cpuusage);
10578 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10580 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10583 #ifndef CONFIG_64BIT
10585 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10587 spin_lock_irq(&cpu_rq(cpu)->lock);
10589 spin_unlock_irq(&cpu_rq(cpu)->lock);
10597 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10599 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10601 #ifndef CONFIG_64BIT
10603 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10605 spin_lock_irq(&cpu_rq(cpu)->lock);
10607 spin_unlock_irq(&cpu_rq(cpu)->lock);
10613 /* return total cpu usage (in nanoseconds) of a group */
10614 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10616 struct cpuacct *ca = cgroup_ca(cgrp);
10617 u64 totalcpuusage = 0;
10620 for_each_present_cpu(i)
10621 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10623 return totalcpuusage;
10626 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10629 struct cpuacct *ca = cgroup_ca(cgrp);
10638 for_each_present_cpu(i)
10639 cpuacct_cpuusage_write(ca, i, 0);
10645 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10646 struct seq_file *m)
10648 struct cpuacct *ca = cgroup_ca(cgroup);
10652 for_each_present_cpu(i) {
10653 percpu = cpuacct_cpuusage_read(ca, i);
10654 seq_printf(m, "%llu ", (unsigned long long) percpu);
10656 seq_printf(m, "\n");
10660 static const char *cpuacct_stat_desc[] = {
10661 [CPUACCT_STAT_USER] = "user",
10662 [CPUACCT_STAT_SYSTEM] = "system",
10665 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10666 struct cgroup_map_cb *cb)
10668 struct cpuacct *ca = cgroup_ca(cgrp);
10671 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10672 s64 val = percpu_counter_read(&ca->cpustat[i]);
10673 val = cputime64_to_clock_t(val);
10674 cb->fill(cb, cpuacct_stat_desc[i], val);
10679 static struct cftype files[] = {
10682 .read_u64 = cpuusage_read,
10683 .write_u64 = cpuusage_write,
10686 .name = "usage_percpu",
10687 .read_seq_string = cpuacct_percpu_seq_read,
10691 .read_map = cpuacct_stats_show,
10695 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10697 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10701 * charge this task's execution time to its accounting group.
10703 * called with rq->lock held.
10705 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10707 struct cpuacct *ca;
10710 if (unlikely(!cpuacct_subsys.active))
10713 cpu = task_cpu(tsk);
10719 for (; ca; ca = ca->parent) {
10720 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10721 *cpuusage += cputime;
10728 * Charge the system/user time to the task's accounting group.
10730 static void cpuacct_update_stats(struct task_struct *tsk,
10731 enum cpuacct_stat_index idx, cputime_t val)
10733 struct cpuacct *ca;
10735 if (unlikely(!cpuacct_subsys.active))
10742 percpu_counter_add(&ca->cpustat[idx], val);
10748 struct cgroup_subsys cpuacct_subsys = {
10750 .create = cpuacct_create,
10751 .destroy = cpuacct_destroy,
10752 .populate = cpuacct_populate,
10753 .subsys_id = cpuacct_subsys_id,
10755 #endif /* CONFIG_CGROUP_CPUACCT */
10759 int rcu_expedited_torture_stats(char *page)
10763 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10765 void synchronize_sched_expedited(void)
10768 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10770 #else /* #ifndef CONFIG_SMP */
10772 static DEFINE_PER_CPU(struct migration_req, rcu_migration_req);
10773 static DEFINE_MUTEX(rcu_sched_expedited_mutex);
10775 #define RCU_EXPEDITED_STATE_POST -2
10776 #define RCU_EXPEDITED_STATE_IDLE -1
10778 static int rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10780 int rcu_expedited_torture_stats(char *page)
10785 cnt += sprintf(&page[cnt], "state: %d /", rcu_expedited_state);
10786 for_each_online_cpu(cpu) {
10787 cnt += sprintf(&page[cnt], " %d:%d",
10788 cpu, per_cpu(rcu_migration_req, cpu).dest_cpu);
10790 cnt += sprintf(&page[cnt], "\n");
10793 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats);
10795 static long synchronize_sched_expedited_count;
10798 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10799 * approach to force grace period to end quickly. This consumes
10800 * significant time on all CPUs, and is thus not recommended for
10801 * any sort of common-case code.
10803 * Note that it is illegal to call this function while holding any
10804 * lock that is acquired by a CPU-hotplug notifier. Failing to
10805 * observe this restriction will result in deadlock.
10807 void synchronize_sched_expedited(void)
10810 unsigned long flags;
10811 bool need_full_sync = 0;
10813 struct migration_req *req;
10817 smp_mb(); /* ensure prior mod happens before capturing snap. */
10818 snap = ACCESS_ONCE(synchronize_sched_expedited_count) + 1;
10820 while (!mutex_trylock(&rcu_sched_expedited_mutex)) {
10822 if (trycount++ < 10)
10823 udelay(trycount * num_online_cpus());
10825 synchronize_sched();
10828 if (ACCESS_ONCE(synchronize_sched_expedited_count) - snap > 0) {
10829 smp_mb(); /* ensure test happens before caller kfree */
10834 rcu_expedited_state = RCU_EXPEDITED_STATE_POST;
10835 for_each_online_cpu(cpu) {
10837 req = &per_cpu(rcu_migration_req, cpu);
10838 init_completion(&req->done);
10840 req->dest_cpu = RCU_MIGRATION_NEED_QS;
10841 spin_lock_irqsave(&rq->lock, flags);
10842 list_add(&req->list, &rq->migration_queue);
10843 spin_unlock_irqrestore(&rq->lock, flags);
10844 wake_up_process(rq->migration_thread);
10846 for_each_online_cpu(cpu) {
10847 rcu_expedited_state = cpu;
10848 req = &per_cpu(rcu_migration_req, cpu);
10850 wait_for_completion(&req->done);
10851 spin_lock_irqsave(&rq->lock, flags);
10852 if (unlikely(req->dest_cpu == RCU_MIGRATION_MUST_SYNC))
10853 need_full_sync = 1;
10854 req->dest_cpu = RCU_MIGRATION_IDLE;
10855 spin_unlock_irqrestore(&rq->lock, flags);
10857 rcu_expedited_state = RCU_EXPEDITED_STATE_IDLE;
10858 mutex_unlock(&rcu_sched_expedited_mutex);
10860 if (need_full_sync)
10861 synchronize_sched();
10863 EXPORT_SYMBOL_GPL(synchronize_sched_expedited);
10865 #endif /* #else #ifndef CONFIG_SMP */