4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
234 if (hrtimer_active(&rt_b->rt_period_timer))
237 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
238 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 hrtimer_start_expires(&rt_b->rt_period_timer,
242 spin_unlock(&rt_b->rt_runtime_lock);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
248 hrtimer_cancel(&rt_b->rt_period_timer);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity **rt_se;
286 struct rt_rq **rt_rq;
288 struct rt_bandwidth rt_bandwidth;
292 struct list_head list;
294 struct task_group *parent;
295 struct list_head siblings;
296 struct list_head children;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct *user)
304 user->tg->uid = user->uid;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
323 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock);
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 #ifdef CONFIG_USER_SCHED
336 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
337 #else /* !CONFIG_USER_SCHED */
338 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
339 #endif /* CONFIG_USER_SCHED */
342 * A weight of 0 or 1 can cause arithmetics problems.
343 * A weight of a cfs_rq is the sum of weights of which entities
344 * are queued on this cfs_rq, so a weight of a entity should not be
345 * too large, so as the shares value of a task group.
346 * (The default weight is 1024 - so there's no practical
347 * limitation from this.)
350 #define MAX_SHARES (1UL << 18)
352 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
355 /* Default task group.
356 * Every task in system belong to this group at bootup.
358 struct task_group init_task_group;
360 /* return group to which a task belongs */
361 static inline struct task_group *task_group(struct task_struct *p)
363 struct task_group *tg;
365 #ifdef CONFIG_USER_SCHED
367 tg = __task_cred(p)->user->tg;
369 #elif defined(CONFIG_CGROUP_SCHED)
370 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
371 struct task_group, css);
373 tg = &init_task_group;
378 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
383 p->se.parent = task_group(p)->se[cpu];
386 #ifdef CONFIG_RT_GROUP_SCHED
387 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
388 p->rt.parent = task_group(p)->rt_se[cpu];
394 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
395 static inline struct task_group *task_group(struct task_struct *p)
400 #endif /* CONFIG_GROUP_SCHED */
402 /* CFS-related fields in a runqueue */
404 struct load_weight load;
405 unsigned long nr_running;
410 struct rb_root tasks_timeline;
411 struct rb_node *rb_leftmost;
413 struct list_head tasks;
414 struct list_head *balance_iterator;
417 * 'curr' points to currently running entity on this cfs_rq.
418 * It is set to NULL otherwise (i.e when none are currently running).
420 struct sched_entity *curr, *next, *last;
422 unsigned int nr_spread_over;
424 #ifdef CONFIG_FAIR_GROUP_SCHED
425 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
428 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
429 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
430 * (like users, containers etc.)
432 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
433 * list is used during load balance.
435 struct list_head leaf_cfs_rq_list;
436 struct task_group *tg; /* group that "owns" this runqueue */
440 * the part of load.weight contributed by tasks
442 unsigned long task_weight;
445 * h_load = weight * f(tg)
447 * Where f(tg) is the recursive weight fraction assigned to
450 unsigned long h_load;
453 * this cpu's part of tg->shares
455 unsigned long shares;
458 * load.weight at the time we set shares
460 unsigned long rq_weight;
465 /* Real-Time classes' related field in a runqueue: */
467 struct rt_prio_array active;
468 unsigned long rt_nr_running;
469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
471 int curr; /* highest queued rt task prio */
473 int next; /* next highest */
478 unsigned long rt_nr_migratory;
480 struct plist_head pushable_tasks;
485 /* Nests inside the rq lock: */
486 spinlock_t rt_runtime_lock;
488 #ifdef CONFIG_RT_GROUP_SCHED
489 unsigned long rt_nr_boosted;
492 struct list_head leaf_rt_rq_list;
493 struct task_group *tg;
494 struct sched_rt_entity *rt_se;
501 * We add the notion of a root-domain which will be used to define per-domain
502 * variables. Each exclusive cpuset essentially defines an island domain by
503 * fully partitioning the member cpus from any other cpuset. Whenever a new
504 * exclusive cpuset is created, we also create and attach a new root-domain
511 cpumask_var_t online;
514 * The "RT overload" flag: it gets set if a CPU has more than
515 * one runnable RT task.
517 cpumask_var_t rto_mask;
520 struct cpupri cpupri;
522 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
524 * Preferred wake up cpu nominated by sched_mc balance that will be
525 * used when most cpus are idle in the system indicating overall very
526 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
528 unsigned int sched_mc_preferred_wakeup_cpu;
533 * By default the system creates a single root-domain with all cpus as
534 * members (mimicking the global state we have today).
536 static struct root_domain def_root_domain;
541 * This is the main, per-CPU runqueue data structure.
543 * Locking rule: those places that want to lock multiple runqueues
544 * (such as the load balancing or the thread migration code), lock
545 * acquire operations must be ordered by ascending &runqueue.
552 * nr_running and cpu_load should be in the same cacheline because
553 * remote CPUs use both these fields when doing load calculation.
555 unsigned long nr_running;
556 #define CPU_LOAD_IDX_MAX 5
557 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
558 unsigned char idle_at_tick;
560 unsigned long last_tick_seen;
561 unsigned char in_nohz_recently;
563 /* capture load from *all* tasks on this cpu: */
564 struct load_weight load;
565 unsigned long nr_load_updates;
571 #ifdef CONFIG_FAIR_GROUP_SCHED
572 /* list of leaf cfs_rq on this cpu: */
573 struct list_head leaf_cfs_rq_list;
575 #ifdef CONFIG_RT_GROUP_SCHED
576 struct list_head leaf_rt_rq_list;
580 * This is part of a global counter where only the total sum
581 * over all CPUs matters. A task can increase this counter on
582 * one CPU and if it got migrated afterwards it may decrease
583 * it on another CPU. Always updated under the runqueue lock:
585 unsigned long nr_uninterruptible;
587 struct task_struct *curr, *idle;
588 unsigned long next_balance;
589 struct mm_struct *prev_mm;
596 struct root_domain *rd;
597 struct sched_domain *sd;
599 /* For active balancing */
602 /* cpu of this runqueue: */
606 unsigned long avg_load_per_task;
608 struct task_struct *migration_thread;
609 struct list_head migration_queue;
612 #ifdef CONFIG_SCHED_HRTICK
614 int hrtick_csd_pending;
615 struct call_single_data hrtick_csd;
617 struct hrtimer hrtick_timer;
620 #ifdef CONFIG_SCHEDSTATS
622 struct sched_info rq_sched_info;
623 unsigned long long rq_cpu_time;
624 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
626 /* sys_sched_yield() stats */
627 unsigned int yld_exp_empty;
628 unsigned int yld_act_empty;
629 unsigned int yld_both_empty;
630 unsigned int yld_count;
632 /* schedule() stats */
633 unsigned int sched_switch;
634 unsigned int sched_count;
635 unsigned int sched_goidle;
637 /* try_to_wake_up() stats */
638 unsigned int ttwu_count;
639 unsigned int ttwu_local;
642 unsigned int bkl_count;
646 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
648 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
650 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
653 static inline int cpu_of(struct rq *rq)
663 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
664 * See detach_destroy_domains: synchronize_sched for details.
666 * The domain tree of any CPU may only be accessed from within
667 * preempt-disabled sections.
669 #define for_each_domain(cpu, __sd) \
670 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
672 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
673 #define this_rq() (&__get_cpu_var(runqueues))
674 #define task_rq(p) cpu_rq(task_cpu(p))
675 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
677 static inline void update_rq_clock(struct rq *rq)
679 rq->clock = sched_clock_cpu(cpu_of(rq));
683 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
685 #ifdef CONFIG_SCHED_DEBUG
686 # define const_debug __read_mostly
688 # define const_debug static const
694 * Returns true if the current cpu runqueue is locked.
695 * This interface allows printk to be called with the runqueue lock
696 * held and know whether or not it is OK to wake up the klogd.
698 int runqueue_is_locked(void)
701 struct rq *rq = cpu_rq(cpu);
704 ret = spin_is_locked(&rq->lock);
710 * Debugging: various feature bits
713 #define SCHED_FEAT(name, enabled) \
714 __SCHED_FEAT_##name ,
717 #include "sched_features.h"
722 #define SCHED_FEAT(name, enabled) \
723 (1UL << __SCHED_FEAT_##name) * enabled |
725 const_debug unsigned int sysctl_sched_features =
726 #include "sched_features.h"
731 #ifdef CONFIG_SCHED_DEBUG
732 #define SCHED_FEAT(name, enabled) \
735 static __read_mostly char *sched_feat_names[] = {
736 #include "sched_features.h"
742 static int sched_feat_show(struct seq_file *m, void *v)
746 for (i = 0; sched_feat_names[i]; i++) {
747 if (!(sysctl_sched_features & (1UL << i)))
749 seq_printf(m, "%s ", sched_feat_names[i]);
757 sched_feat_write(struct file *filp, const char __user *ubuf,
758 size_t cnt, loff_t *ppos)
768 if (copy_from_user(&buf, ubuf, cnt))
773 if (strncmp(buf, "NO_", 3) == 0) {
778 for (i = 0; sched_feat_names[i]; i++) {
779 int len = strlen(sched_feat_names[i]);
781 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
783 sysctl_sched_features &= ~(1UL << i);
785 sysctl_sched_features |= (1UL << i);
790 if (!sched_feat_names[i])
798 static int sched_feat_open(struct inode *inode, struct file *filp)
800 return single_open(filp, sched_feat_show, NULL);
803 static struct file_operations sched_feat_fops = {
804 .open = sched_feat_open,
805 .write = sched_feat_write,
808 .release = single_release,
811 static __init int sched_init_debug(void)
813 debugfs_create_file("sched_features", 0644, NULL, NULL,
818 late_initcall(sched_init_debug);
822 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
825 * Number of tasks to iterate in a single balance run.
826 * Limited because this is done with IRQs disabled.
828 const_debug unsigned int sysctl_sched_nr_migrate = 32;
831 * ratelimit for updating the group shares.
834 unsigned int sysctl_sched_shares_ratelimit = 250000;
837 * Inject some fuzzyness into changing the per-cpu group shares
838 * this avoids remote rq-locks at the expense of fairness.
841 unsigned int sysctl_sched_shares_thresh = 4;
844 * period over which we measure -rt task cpu usage in us.
847 unsigned int sysctl_sched_rt_period = 1000000;
849 static __read_mostly int scheduler_running;
852 * part of the period that we allow rt tasks to run in us.
855 int sysctl_sched_rt_runtime = 950000;
857 static inline u64 global_rt_period(void)
859 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
862 static inline u64 global_rt_runtime(void)
864 if (sysctl_sched_rt_runtime < 0)
867 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
870 #ifndef prepare_arch_switch
871 # define prepare_arch_switch(next) do { } while (0)
873 #ifndef finish_arch_switch
874 # define finish_arch_switch(prev) do { } while (0)
877 static inline int task_current(struct rq *rq, struct task_struct *p)
879 return rq->curr == p;
882 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
883 static inline int task_running(struct rq *rq, struct task_struct *p)
885 return task_current(rq, p);
888 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
892 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
894 #ifdef CONFIG_DEBUG_SPINLOCK
895 /* this is a valid case when another task releases the spinlock */
896 rq->lock.owner = current;
899 * If we are tracking spinlock dependencies then we have to
900 * fix up the runqueue lock - which gets 'carried over' from
903 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
905 spin_unlock_irq(&rq->lock);
908 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
909 static inline int task_running(struct rq *rq, struct task_struct *p)
914 return task_current(rq, p);
918 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
922 * We can optimise this out completely for !SMP, because the
923 * SMP rebalancing from interrupt is the only thing that cares
928 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
929 spin_unlock_irq(&rq->lock);
931 spin_unlock(&rq->lock);
935 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
939 * After ->oncpu is cleared, the task can be moved to a different CPU.
940 * We must ensure this doesn't happen until the switch is completely
946 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
950 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
953 * __task_rq_lock - lock the runqueue a given task resides on.
954 * Must be called interrupts disabled.
956 static inline struct rq *__task_rq_lock(struct task_struct *p)
960 struct rq *rq = task_rq(p);
961 spin_lock(&rq->lock);
962 if (likely(rq == task_rq(p)))
964 spin_unlock(&rq->lock);
969 * task_rq_lock - lock the runqueue a given task resides on and disable
970 * interrupts. Note the ordering: we can safely lookup the task_rq without
971 * explicitly disabling preemption.
973 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
979 local_irq_save(*flags);
981 spin_lock(&rq->lock);
982 if (likely(rq == task_rq(p)))
984 spin_unlock_irqrestore(&rq->lock, *flags);
988 void task_rq_unlock_wait(struct task_struct *p)
990 struct rq *rq = task_rq(p);
992 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
993 spin_unlock_wait(&rq->lock);
996 static void __task_rq_unlock(struct rq *rq)
999 spin_unlock(&rq->lock);
1002 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1003 __releases(rq->lock)
1005 spin_unlock_irqrestore(&rq->lock, *flags);
1009 * this_rq_lock - lock this runqueue and disable interrupts.
1011 static struct rq *this_rq_lock(void)
1012 __acquires(rq->lock)
1016 local_irq_disable();
1018 spin_lock(&rq->lock);
1023 #ifdef CONFIG_SCHED_HRTICK
1025 * Use HR-timers to deliver accurate preemption points.
1027 * Its all a bit involved since we cannot program an hrt while holding the
1028 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1031 * When we get rescheduled we reprogram the hrtick_timer outside of the
1037 * - enabled by features
1038 * - hrtimer is actually high res
1040 static inline int hrtick_enabled(struct rq *rq)
1042 if (!sched_feat(HRTICK))
1044 if (!cpu_active(cpu_of(rq)))
1046 return hrtimer_is_hres_active(&rq->hrtick_timer);
1049 static void hrtick_clear(struct rq *rq)
1051 if (hrtimer_active(&rq->hrtick_timer))
1052 hrtimer_cancel(&rq->hrtick_timer);
1056 * High-resolution timer tick.
1057 * Runs from hardirq context with interrupts disabled.
1059 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1061 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1063 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1065 spin_lock(&rq->lock);
1066 update_rq_clock(rq);
1067 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1068 spin_unlock(&rq->lock);
1070 return HRTIMER_NORESTART;
1075 * called from hardirq (IPI) context
1077 static void __hrtick_start(void *arg)
1079 struct rq *rq = arg;
1081 spin_lock(&rq->lock);
1082 hrtimer_restart(&rq->hrtick_timer);
1083 rq->hrtick_csd_pending = 0;
1084 spin_unlock(&rq->lock);
1088 * Called to set the hrtick timer state.
1090 * called with rq->lock held and irqs disabled
1092 static void hrtick_start(struct rq *rq, u64 delay)
1094 struct hrtimer *timer = &rq->hrtick_timer;
1095 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1097 hrtimer_set_expires(timer, time);
1099 if (rq == this_rq()) {
1100 hrtimer_restart(timer);
1101 } else if (!rq->hrtick_csd_pending) {
1102 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1103 rq->hrtick_csd_pending = 1;
1108 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1110 int cpu = (int)(long)hcpu;
1113 case CPU_UP_CANCELED:
1114 case CPU_UP_CANCELED_FROZEN:
1115 case CPU_DOWN_PREPARE:
1116 case CPU_DOWN_PREPARE_FROZEN:
1118 case CPU_DEAD_FROZEN:
1119 hrtick_clear(cpu_rq(cpu));
1126 static __init void init_hrtick(void)
1128 hotcpu_notifier(hotplug_hrtick, 0);
1132 * Called to set the hrtick timer state.
1134 * called with rq->lock held and irqs disabled
1136 static void hrtick_start(struct rq *rq, u64 delay)
1138 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SMP */
1146 static void init_rq_hrtick(struct rq *rq)
1149 rq->hrtick_csd_pending = 0;
1151 rq->hrtick_csd.flags = 0;
1152 rq->hrtick_csd.func = __hrtick_start;
1153 rq->hrtick_csd.info = rq;
1156 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1157 rq->hrtick_timer.function = hrtick;
1159 #else /* CONFIG_SCHED_HRTICK */
1160 static inline void hrtick_clear(struct rq *rq)
1164 static inline void init_rq_hrtick(struct rq *rq)
1168 static inline void init_hrtick(void)
1171 #endif /* CONFIG_SCHED_HRTICK */
1174 * resched_task - mark a task 'to be rescheduled now'.
1176 * On UP this means the setting of the need_resched flag, on SMP it
1177 * might also involve a cross-CPU call to trigger the scheduler on
1182 #ifndef tsk_is_polling
1183 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1186 static void resched_task(struct task_struct *p)
1190 assert_spin_locked(&task_rq(p)->lock);
1192 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1195 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1198 if (cpu == smp_processor_id())
1201 /* NEED_RESCHED must be visible before we test polling */
1203 if (!tsk_is_polling(p))
1204 smp_send_reschedule(cpu);
1207 static void resched_cpu(int cpu)
1209 struct rq *rq = cpu_rq(cpu);
1210 unsigned long flags;
1212 if (!spin_trylock_irqsave(&rq->lock, flags))
1214 resched_task(cpu_curr(cpu));
1215 spin_unlock_irqrestore(&rq->lock, flags);
1220 * When add_timer_on() enqueues a timer into the timer wheel of an
1221 * idle CPU then this timer might expire before the next timer event
1222 * which is scheduled to wake up that CPU. In case of a completely
1223 * idle system the next event might even be infinite time into the
1224 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1225 * leaves the inner idle loop so the newly added timer is taken into
1226 * account when the CPU goes back to idle and evaluates the timer
1227 * wheel for the next timer event.
1229 void wake_up_idle_cpu(int cpu)
1231 struct rq *rq = cpu_rq(cpu);
1233 if (cpu == smp_processor_id())
1237 * This is safe, as this function is called with the timer
1238 * wheel base lock of (cpu) held. When the CPU is on the way
1239 * to idle and has not yet set rq->curr to idle then it will
1240 * be serialized on the timer wheel base lock and take the new
1241 * timer into account automatically.
1243 if (rq->curr != rq->idle)
1247 * We can set TIF_RESCHED on the idle task of the other CPU
1248 * lockless. The worst case is that the other CPU runs the
1249 * idle task through an additional NOOP schedule()
1251 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1253 /* NEED_RESCHED must be visible before we test polling */
1255 if (!tsk_is_polling(rq->idle))
1256 smp_send_reschedule(cpu);
1258 #endif /* CONFIG_NO_HZ */
1260 #else /* !CONFIG_SMP */
1261 static void resched_task(struct task_struct *p)
1263 assert_spin_locked(&task_rq(p)->lock);
1264 set_tsk_need_resched(p);
1266 #endif /* CONFIG_SMP */
1268 #if BITS_PER_LONG == 32
1269 # define WMULT_CONST (~0UL)
1271 # define WMULT_CONST (1UL << 32)
1274 #define WMULT_SHIFT 32
1277 * Shift right and round:
1279 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1282 * delta *= weight / lw
1284 static unsigned long
1285 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1286 struct load_weight *lw)
1290 if (!lw->inv_weight) {
1291 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1294 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1298 tmp = (u64)delta_exec * weight;
1300 * Check whether we'd overflow the 64-bit multiplication:
1302 if (unlikely(tmp > WMULT_CONST))
1303 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1306 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1308 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1311 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1317 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1324 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1325 * of tasks with abnormal "nice" values across CPUs the contribution that
1326 * each task makes to its run queue's load is weighted according to its
1327 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1328 * scaled version of the new time slice allocation that they receive on time
1332 #define WEIGHT_IDLEPRIO 3
1333 #define WMULT_IDLEPRIO 1431655765
1336 * Nice levels are multiplicative, with a gentle 10% change for every
1337 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1338 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1339 * that remained on nice 0.
1341 * The "10% effect" is relative and cumulative: from _any_ nice level,
1342 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1343 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1344 * If a task goes up by ~10% and another task goes down by ~10% then
1345 * the relative distance between them is ~25%.)
1347 static const int prio_to_weight[40] = {
1348 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1349 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1350 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1351 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1352 /* 0 */ 1024, 820, 655, 526, 423,
1353 /* 5 */ 335, 272, 215, 172, 137,
1354 /* 10 */ 110, 87, 70, 56, 45,
1355 /* 15 */ 36, 29, 23, 18, 15,
1359 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1361 * In cases where the weight does not change often, we can use the
1362 * precalculated inverse to speed up arithmetics by turning divisions
1363 * into multiplications:
1365 static const u32 prio_to_wmult[40] = {
1366 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1367 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1368 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1369 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1370 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1371 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1372 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1373 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1376 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1379 * runqueue iterator, to support SMP load-balancing between different
1380 * scheduling classes, without having to expose their internal data
1381 * structures to the load-balancing proper:
1383 struct rq_iterator {
1385 struct task_struct *(*start)(void *);
1386 struct task_struct *(*next)(void *);
1390 static unsigned long
1391 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1392 unsigned long max_load_move, struct sched_domain *sd,
1393 enum cpu_idle_type idle, int *all_pinned,
1394 int *this_best_prio, struct rq_iterator *iterator);
1397 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1398 struct sched_domain *sd, enum cpu_idle_type idle,
1399 struct rq_iterator *iterator);
1402 #ifdef CONFIG_CGROUP_CPUACCT
1403 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1405 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1408 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1410 update_load_add(&rq->load, load);
1413 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1415 update_load_sub(&rq->load, load);
1418 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1419 typedef int (*tg_visitor)(struct task_group *, void *);
1422 * Iterate the full tree, calling @down when first entering a node and @up when
1423 * leaving it for the final time.
1425 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1427 struct task_group *parent, *child;
1431 parent = &root_task_group;
1433 ret = (*down)(parent, data);
1436 list_for_each_entry_rcu(child, &parent->children, siblings) {
1443 ret = (*up)(parent, data);
1448 parent = parent->parent;
1457 static int tg_nop(struct task_group *tg, void *data)
1464 static unsigned long source_load(int cpu, int type);
1465 static unsigned long target_load(int cpu, int type);
1466 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1468 static unsigned long cpu_avg_load_per_task(int cpu)
1470 struct rq *rq = cpu_rq(cpu);
1471 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1474 rq->avg_load_per_task = rq->load.weight / nr_running;
1476 rq->avg_load_per_task = 0;
1478 return rq->avg_load_per_task;
1481 #ifdef CONFIG_FAIR_GROUP_SCHED
1483 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1486 * Calculate and set the cpu's group shares.
1489 update_group_shares_cpu(struct task_group *tg, int cpu,
1490 unsigned long sd_shares, unsigned long sd_rq_weight)
1492 unsigned long shares;
1493 unsigned long rq_weight;
1498 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1501 * \Sum shares * rq_weight
1502 * shares = -----------------------
1506 shares = (sd_shares * rq_weight) / sd_rq_weight;
1507 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1509 if (abs(shares - tg->se[cpu]->load.weight) >
1510 sysctl_sched_shares_thresh) {
1511 struct rq *rq = cpu_rq(cpu);
1512 unsigned long flags;
1514 spin_lock_irqsave(&rq->lock, flags);
1515 tg->cfs_rq[cpu]->shares = shares;
1517 __set_se_shares(tg->se[cpu], shares);
1518 spin_unlock_irqrestore(&rq->lock, flags);
1523 * Re-compute the task group their per cpu shares over the given domain.
1524 * This needs to be done in a bottom-up fashion because the rq weight of a
1525 * parent group depends on the shares of its child groups.
1527 static int tg_shares_up(struct task_group *tg, void *data)
1529 unsigned long weight, rq_weight = 0;
1530 unsigned long shares = 0;
1531 struct sched_domain *sd = data;
1534 for_each_cpu(i, sched_domain_span(sd)) {
1536 * If there are currently no tasks on the cpu pretend there
1537 * is one of average load so that when a new task gets to
1538 * run here it will not get delayed by group starvation.
1540 weight = tg->cfs_rq[i]->load.weight;
1542 weight = NICE_0_LOAD;
1544 tg->cfs_rq[i]->rq_weight = weight;
1545 rq_weight += weight;
1546 shares += tg->cfs_rq[i]->shares;
1549 if ((!shares && rq_weight) || shares > tg->shares)
1550 shares = tg->shares;
1552 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1553 shares = tg->shares;
1555 for_each_cpu(i, sched_domain_span(sd))
1556 update_group_shares_cpu(tg, i, shares, rq_weight);
1562 * Compute the cpu's hierarchical load factor for each task group.
1563 * This needs to be done in a top-down fashion because the load of a child
1564 * group is a fraction of its parents load.
1566 static int tg_load_down(struct task_group *tg, void *data)
1569 long cpu = (long)data;
1572 load = cpu_rq(cpu)->load.weight;
1574 load = tg->parent->cfs_rq[cpu]->h_load;
1575 load *= tg->cfs_rq[cpu]->shares;
1576 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1579 tg->cfs_rq[cpu]->h_load = load;
1584 static void update_shares(struct sched_domain *sd)
1586 u64 now = cpu_clock(raw_smp_processor_id());
1587 s64 elapsed = now - sd->last_update;
1589 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1590 sd->last_update = now;
1591 walk_tg_tree(tg_nop, tg_shares_up, sd);
1595 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1597 spin_unlock(&rq->lock);
1599 spin_lock(&rq->lock);
1602 static void update_h_load(long cpu)
1604 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1609 static inline void update_shares(struct sched_domain *sd)
1613 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1619 #ifdef CONFIG_PREEMPT
1622 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1623 * way at the expense of forcing extra atomic operations in all
1624 * invocations. This assures that the double_lock is acquired using the
1625 * same underlying policy as the spinlock_t on this architecture, which
1626 * reduces latency compared to the unfair variant below. However, it
1627 * also adds more overhead and therefore may reduce throughput.
1629 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1630 __releases(this_rq->lock)
1631 __acquires(busiest->lock)
1632 __acquires(this_rq->lock)
1634 spin_unlock(&this_rq->lock);
1635 double_rq_lock(this_rq, busiest);
1642 * Unfair double_lock_balance: Optimizes throughput at the expense of
1643 * latency by eliminating extra atomic operations when the locks are
1644 * already in proper order on entry. This favors lower cpu-ids and will
1645 * grant the double lock to lower cpus over higher ids under contention,
1646 * regardless of entry order into the function.
1648 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1649 __releases(this_rq->lock)
1650 __acquires(busiest->lock)
1651 __acquires(this_rq->lock)
1655 if (unlikely(!spin_trylock(&busiest->lock))) {
1656 if (busiest < this_rq) {
1657 spin_unlock(&this_rq->lock);
1658 spin_lock(&busiest->lock);
1659 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1662 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1667 #endif /* CONFIG_PREEMPT */
1670 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1672 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1674 if (unlikely(!irqs_disabled())) {
1675 /* printk() doesn't work good under rq->lock */
1676 spin_unlock(&this_rq->lock);
1680 return _double_lock_balance(this_rq, busiest);
1683 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1684 __releases(busiest->lock)
1686 spin_unlock(&busiest->lock);
1687 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1691 #ifdef CONFIG_FAIR_GROUP_SCHED
1692 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1695 cfs_rq->shares = shares;
1700 #include "sched_stats.h"
1701 #include "sched_idletask.c"
1702 #include "sched_fair.c"
1703 #include "sched_rt.c"
1704 #ifdef CONFIG_SCHED_DEBUG
1705 # include "sched_debug.c"
1708 #define sched_class_highest (&rt_sched_class)
1709 #define for_each_class(class) \
1710 for (class = sched_class_highest; class; class = class->next)
1712 static void inc_nr_running(struct rq *rq)
1717 static void dec_nr_running(struct rq *rq)
1722 static void set_load_weight(struct task_struct *p)
1724 if (task_has_rt_policy(p)) {
1725 p->se.load.weight = prio_to_weight[0] * 2;
1726 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1731 * SCHED_IDLE tasks get minimal weight:
1733 if (p->policy == SCHED_IDLE) {
1734 p->se.load.weight = WEIGHT_IDLEPRIO;
1735 p->se.load.inv_weight = WMULT_IDLEPRIO;
1739 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1740 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1743 static void update_avg(u64 *avg, u64 sample)
1745 s64 diff = sample - *avg;
1749 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1752 p->se.start_runtime = p->se.sum_exec_runtime;
1754 sched_info_queued(p);
1755 p->sched_class->enqueue_task(rq, p, wakeup);
1759 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1762 if (p->se.last_wakeup) {
1763 update_avg(&p->se.avg_overlap,
1764 p->se.sum_exec_runtime - p->se.last_wakeup);
1765 p->se.last_wakeup = 0;
1767 update_avg(&p->se.avg_wakeup,
1768 sysctl_sched_wakeup_granularity);
1772 sched_info_dequeued(p);
1773 p->sched_class->dequeue_task(rq, p, sleep);
1778 * __normal_prio - return the priority that is based on the static prio
1780 static inline int __normal_prio(struct task_struct *p)
1782 return p->static_prio;
1786 * Calculate the expected normal priority: i.e. priority
1787 * without taking RT-inheritance into account. Might be
1788 * boosted by interactivity modifiers. Changes upon fork,
1789 * setprio syscalls, and whenever the interactivity
1790 * estimator recalculates.
1792 static inline int normal_prio(struct task_struct *p)
1796 if (task_has_rt_policy(p))
1797 prio = MAX_RT_PRIO-1 - p->rt_priority;
1799 prio = __normal_prio(p);
1804 * Calculate the current priority, i.e. the priority
1805 * taken into account by the scheduler. This value might
1806 * be boosted by RT tasks, or might be boosted by
1807 * interactivity modifiers. Will be RT if the task got
1808 * RT-boosted. If not then it returns p->normal_prio.
1810 static int effective_prio(struct task_struct *p)
1812 p->normal_prio = normal_prio(p);
1814 * If we are RT tasks or we were boosted to RT priority,
1815 * keep the priority unchanged. Otherwise, update priority
1816 * to the normal priority:
1818 if (!rt_prio(p->prio))
1819 return p->normal_prio;
1824 * activate_task - move a task to the runqueue.
1826 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1828 if (task_contributes_to_load(p))
1829 rq->nr_uninterruptible--;
1831 enqueue_task(rq, p, wakeup);
1836 * deactivate_task - remove a task from the runqueue.
1838 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1840 if (task_contributes_to_load(p))
1841 rq->nr_uninterruptible++;
1843 dequeue_task(rq, p, sleep);
1848 * task_curr - is this task currently executing on a CPU?
1849 * @p: the task in question.
1851 inline int task_curr(const struct task_struct *p)
1853 return cpu_curr(task_cpu(p)) == p;
1856 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1858 set_task_rq(p, cpu);
1861 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1862 * successfuly executed on another CPU. We must ensure that updates of
1863 * per-task data have been completed by this moment.
1866 task_thread_info(p)->cpu = cpu;
1870 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1871 const struct sched_class *prev_class,
1872 int oldprio, int running)
1874 if (prev_class != p->sched_class) {
1875 if (prev_class->switched_from)
1876 prev_class->switched_from(rq, p, running);
1877 p->sched_class->switched_to(rq, p, running);
1879 p->sched_class->prio_changed(rq, p, oldprio, running);
1884 /* Used instead of source_load when we know the type == 0 */
1885 static unsigned long weighted_cpuload(const int cpu)
1887 return cpu_rq(cpu)->load.weight;
1891 * Is this task likely cache-hot:
1894 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1899 * Buddy candidates are cache hot:
1901 if (sched_feat(CACHE_HOT_BUDDY) &&
1902 (&p->se == cfs_rq_of(&p->se)->next ||
1903 &p->se == cfs_rq_of(&p->se)->last))
1906 if (p->sched_class != &fair_sched_class)
1909 if (sysctl_sched_migration_cost == -1)
1911 if (sysctl_sched_migration_cost == 0)
1914 delta = now - p->se.exec_start;
1916 return delta < (s64)sysctl_sched_migration_cost;
1920 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1922 int old_cpu = task_cpu(p);
1923 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1924 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1925 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1928 clock_offset = old_rq->clock - new_rq->clock;
1930 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1932 #ifdef CONFIG_SCHEDSTATS
1933 if (p->se.wait_start)
1934 p->se.wait_start -= clock_offset;
1935 if (p->se.sleep_start)
1936 p->se.sleep_start -= clock_offset;
1937 if (p->se.block_start)
1938 p->se.block_start -= clock_offset;
1939 if (old_cpu != new_cpu) {
1940 schedstat_inc(p, se.nr_migrations);
1941 if (task_hot(p, old_rq->clock, NULL))
1942 schedstat_inc(p, se.nr_forced2_migrations);
1945 p->se.vruntime -= old_cfsrq->min_vruntime -
1946 new_cfsrq->min_vruntime;
1948 __set_task_cpu(p, new_cpu);
1951 struct migration_req {
1952 struct list_head list;
1954 struct task_struct *task;
1957 struct completion done;
1961 * The task's runqueue lock must be held.
1962 * Returns true if you have to wait for migration thread.
1965 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1967 struct rq *rq = task_rq(p);
1970 * If the task is not on a runqueue (and not running), then
1971 * it is sufficient to simply update the task's cpu field.
1973 if (!p->se.on_rq && !task_running(rq, p)) {
1974 set_task_cpu(p, dest_cpu);
1978 init_completion(&req->done);
1980 req->dest_cpu = dest_cpu;
1981 list_add(&req->list, &rq->migration_queue);
1987 * wait_task_inactive - wait for a thread to unschedule.
1989 * If @match_state is nonzero, it's the @p->state value just checked and
1990 * not expected to change. If it changes, i.e. @p might have woken up,
1991 * then return zero. When we succeed in waiting for @p to be off its CPU,
1992 * we return a positive number (its total switch count). If a second call
1993 * a short while later returns the same number, the caller can be sure that
1994 * @p has remained unscheduled the whole time.
1996 * The caller must ensure that the task *will* unschedule sometime soon,
1997 * else this function might spin for a *long* time. This function can't
1998 * be called with interrupts off, or it may introduce deadlock with
1999 * smp_call_function() if an IPI is sent by the same process we are
2000 * waiting to become inactive.
2002 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2004 unsigned long flags;
2011 * We do the initial early heuristics without holding
2012 * any task-queue locks at all. We'll only try to get
2013 * the runqueue lock when things look like they will
2019 * If the task is actively running on another CPU
2020 * still, just relax and busy-wait without holding
2023 * NOTE! Since we don't hold any locks, it's not
2024 * even sure that "rq" stays as the right runqueue!
2025 * But we don't care, since "task_running()" will
2026 * return false if the runqueue has changed and p
2027 * is actually now running somewhere else!
2029 while (task_running(rq, p)) {
2030 if (match_state && unlikely(p->state != match_state))
2036 * Ok, time to look more closely! We need the rq
2037 * lock now, to be *sure*. If we're wrong, we'll
2038 * just go back and repeat.
2040 rq = task_rq_lock(p, &flags);
2041 trace_sched_wait_task(rq, p);
2042 running = task_running(rq, p);
2043 on_rq = p->se.on_rq;
2045 if (!match_state || p->state == match_state)
2046 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2047 task_rq_unlock(rq, &flags);
2050 * If it changed from the expected state, bail out now.
2052 if (unlikely(!ncsw))
2056 * Was it really running after all now that we
2057 * checked with the proper locks actually held?
2059 * Oops. Go back and try again..
2061 if (unlikely(running)) {
2067 * It's not enough that it's not actively running,
2068 * it must be off the runqueue _entirely_, and not
2071 * So if it wa still runnable (but just not actively
2072 * running right now), it's preempted, and we should
2073 * yield - it could be a while.
2075 if (unlikely(on_rq)) {
2076 schedule_timeout_uninterruptible(1);
2081 * Ahh, all good. It wasn't running, and it wasn't
2082 * runnable, which means that it will never become
2083 * running in the future either. We're all done!
2092 * kick_process - kick a running thread to enter/exit the kernel
2093 * @p: the to-be-kicked thread
2095 * Cause a process which is running on another CPU to enter
2096 * kernel-mode, without any delay. (to get signals handled.)
2098 * NOTE: this function doesnt have to take the runqueue lock,
2099 * because all it wants to ensure is that the remote task enters
2100 * the kernel. If the IPI races and the task has been migrated
2101 * to another CPU then no harm is done and the purpose has been
2104 void kick_process(struct task_struct *p)
2110 if ((cpu != smp_processor_id()) && task_curr(p))
2111 smp_send_reschedule(cpu);
2116 * Return a low guess at the load of a migration-source cpu weighted
2117 * according to the scheduling class and "nice" value.
2119 * We want to under-estimate the load of migration sources, to
2120 * balance conservatively.
2122 static unsigned long source_load(int cpu, int type)
2124 struct rq *rq = cpu_rq(cpu);
2125 unsigned long total = weighted_cpuload(cpu);
2127 if (type == 0 || !sched_feat(LB_BIAS))
2130 return min(rq->cpu_load[type-1], total);
2134 * Return a high guess at the load of a migration-target cpu weighted
2135 * according to the scheduling class and "nice" value.
2137 static unsigned long target_load(int cpu, int type)
2139 struct rq *rq = cpu_rq(cpu);
2140 unsigned long total = weighted_cpuload(cpu);
2142 if (type == 0 || !sched_feat(LB_BIAS))
2145 return max(rq->cpu_load[type-1], total);
2149 * find_idlest_group finds and returns the least busy CPU group within the
2152 static struct sched_group *
2153 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2155 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2156 unsigned long min_load = ULONG_MAX, this_load = 0;
2157 int load_idx = sd->forkexec_idx;
2158 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2161 unsigned long load, avg_load;
2165 /* Skip over this group if it has no CPUs allowed */
2166 if (!cpumask_intersects(sched_group_cpus(group),
2170 local_group = cpumask_test_cpu(this_cpu,
2171 sched_group_cpus(group));
2173 /* Tally up the load of all CPUs in the group */
2176 for_each_cpu(i, sched_group_cpus(group)) {
2177 /* Bias balancing toward cpus of our domain */
2179 load = source_load(i, load_idx);
2181 load = target_load(i, load_idx);
2186 /* Adjust by relative CPU power of the group */
2187 avg_load = sg_div_cpu_power(group,
2188 avg_load * SCHED_LOAD_SCALE);
2191 this_load = avg_load;
2193 } else if (avg_load < min_load) {
2194 min_load = avg_load;
2197 } while (group = group->next, group != sd->groups);
2199 if (!idlest || 100*this_load < imbalance*min_load)
2205 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2208 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2210 unsigned long load, min_load = ULONG_MAX;
2214 /* Traverse only the allowed CPUs */
2215 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2216 load = weighted_cpuload(i);
2218 if (load < min_load || (load == min_load && i == this_cpu)) {
2228 * sched_balance_self: balance the current task (running on cpu) in domains
2229 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2232 * Balance, ie. select the least loaded group.
2234 * Returns the target CPU number, or the same CPU if no balancing is needed.
2236 * preempt must be disabled.
2238 static int sched_balance_self(int cpu, int flag)
2240 struct task_struct *t = current;
2241 struct sched_domain *tmp, *sd = NULL;
2243 for_each_domain(cpu, tmp) {
2245 * If power savings logic is enabled for a domain, stop there.
2247 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2249 if (tmp->flags & flag)
2257 struct sched_group *group;
2258 int new_cpu, weight;
2260 if (!(sd->flags & flag)) {
2265 group = find_idlest_group(sd, t, cpu);
2271 new_cpu = find_idlest_cpu(group, t, cpu);
2272 if (new_cpu == -1 || new_cpu == cpu) {
2273 /* Now try balancing at a lower domain level of cpu */
2278 /* Now try balancing at a lower domain level of new_cpu */
2280 weight = cpumask_weight(sched_domain_span(sd));
2282 for_each_domain(cpu, tmp) {
2283 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2285 if (tmp->flags & flag)
2288 /* while loop will break here if sd == NULL */
2294 #endif /* CONFIG_SMP */
2297 * try_to_wake_up - wake up a thread
2298 * @p: the to-be-woken-up thread
2299 * @state: the mask of task states that can be woken
2300 * @sync: do a synchronous wakeup?
2302 * Put it on the run-queue if it's not already there. The "current"
2303 * thread is always on the run-queue (except when the actual
2304 * re-schedule is in progress), and as such you're allowed to do
2305 * the simpler "current->state = TASK_RUNNING" to mark yourself
2306 * runnable without the overhead of this.
2308 * returns failure only if the task is already active.
2310 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2312 int cpu, orig_cpu, this_cpu, success = 0;
2313 unsigned long flags;
2317 if (!sched_feat(SYNC_WAKEUPS))
2321 if (current->se.avg_overlap < sysctl_sched_migration_cost &&
2322 p->se.avg_overlap < sysctl_sched_migration_cost)
2325 if (current->se.avg_overlap >= sysctl_sched_migration_cost ||
2326 p->se.avg_overlap >= sysctl_sched_migration_cost)
2331 if (sched_feat(LB_WAKEUP_UPDATE)) {
2332 struct sched_domain *sd;
2334 this_cpu = raw_smp_processor_id();
2337 for_each_domain(this_cpu, sd) {
2338 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2347 rq = task_rq_lock(p, &flags);
2348 update_rq_clock(rq);
2349 old_state = p->state;
2350 if (!(old_state & state))
2358 this_cpu = smp_processor_id();
2361 if (unlikely(task_running(rq, p)))
2364 cpu = p->sched_class->select_task_rq(p, sync);
2365 if (cpu != orig_cpu) {
2366 set_task_cpu(p, cpu);
2367 task_rq_unlock(rq, &flags);
2368 /* might preempt at this point */
2369 rq = task_rq_lock(p, &flags);
2370 old_state = p->state;
2371 if (!(old_state & state))
2376 this_cpu = smp_processor_id();
2380 #ifdef CONFIG_SCHEDSTATS
2381 schedstat_inc(rq, ttwu_count);
2382 if (cpu == this_cpu)
2383 schedstat_inc(rq, ttwu_local);
2385 struct sched_domain *sd;
2386 for_each_domain(this_cpu, sd) {
2387 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2388 schedstat_inc(sd, ttwu_wake_remote);
2393 #endif /* CONFIG_SCHEDSTATS */
2396 #endif /* CONFIG_SMP */
2397 schedstat_inc(p, se.nr_wakeups);
2399 schedstat_inc(p, se.nr_wakeups_sync);
2400 if (orig_cpu != cpu)
2401 schedstat_inc(p, se.nr_wakeups_migrate);
2402 if (cpu == this_cpu)
2403 schedstat_inc(p, se.nr_wakeups_local);
2405 schedstat_inc(p, se.nr_wakeups_remote);
2406 activate_task(rq, p, 1);
2410 * Only attribute actual wakeups done by this task.
2412 if (!in_interrupt()) {
2413 struct sched_entity *se = ¤t->se;
2414 u64 sample = se->sum_exec_runtime;
2416 if (se->last_wakeup)
2417 sample -= se->last_wakeup;
2419 sample -= se->start_runtime;
2420 update_avg(&se->avg_wakeup, sample);
2422 se->last_wakeup = se->sum_exec_runtime;
2426 trace_sched_wakeup(rq, p, success);
2427 check_preempt_curr(rq, p, sync);
2429 p->state = TASK_RUNNING;
2431 if (p->sched_class->task_wake_up)
2432 p->sched_class->task_wake_up(rq, p);
2435 task_rq_unlock(rq, &flags);
2440 int wake_up_process(struct task_struct *p)
2442 return try_to_wake_up(p, TASK_ALL, 0);
2444 EXPORT_SYMBOL(wake_up_process);
2446 int wake_up_state(struct task_struct *p, unsigned int state)
2448 return try_to_wake_up(p, state, 0);
2452 * Perform scheduler related setup for a newly forked process p.
2453 * p is forked by current.
2455 * __sched_fork() is basic setup used by init_idle() too:
2457 static void __sched_fork(struct task_struct *p)
2459 p->se.exec_start = 0;
2460 p->se.sum_exec_runtime = 0;
2461 p->se.prev_sum_exec_runtime = 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;
2469 p->se.sum_sleep_runtime = 0;
2470 p->se.sleep_start = 0;
2471 p->se.block_start = 0;
2472 p->se.sleep_max = 0;
2473 p->se.block_max = 0;
2475 p->se.slice_max = 0;
2479 INIT_LIST_HEAD(&p->rt.run_list);
2481 INIT_LIST_HEAD(&p->se.group_node);
2483 #ifdef CONFIG_PREEMPT_NOTIFIERS
2484 INIT_HLIST_HEAD(&p->preempt_notifiers);
2488 * We mark the process as running here, but have not actually
2489 * inserted it onto the runqueue yet. This guarantees that
2490 * nobody will actually run it, and a signal or other external
2491 * event cannot wake it up and insert it on the runqueue either.
2493 p->state = TASK_RUNNING;
2497 * fork()/clone()-time setup:
2499 void sched_fork(struct task_struct *p, int clone_flags)
2501 int cpu = get_cpu();
2506 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2508 set_task_cpu(p, cpu);
2511 * Make sure we do not leak PI boosting priority to the child:
2513 p->prio = current->normal_prio;
2514 if (!rt_prio(p->prio))
2515 p->sched_class = &fair_sched_class;
2517 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2518 if (likely(sched_info_on()))
2519 memset(&p->sched_info, 0, sizeof(p->sched_info));
2521 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2524 #ifdef CONFIG_PREEMPT
2525 /* Want to start with kernel preemption disabled. */
2526 task_thread_info(p)->preempt_count = 1;
2528 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2534 * wake_up_new_task - wake up a newly created task for the first time.
2536 * This function will do some initial scheduler statistics housekeeping
2537 * that must be done for every newly created context, then puts the task
2538 * on the runqueue and wakes it.
2540 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2542 unsigned long flags;
2545 rq = task_rq_lock(p, &flags);
2546 BUG_ON(p->state != TASK_RUNNING);
2547 update_rq_clock(rq);
2549 p->prio = effective_prio(p);
2551 if (!p->sched_class->task_new || !current->se.on_rq) {
2552 activate_task(rq, p, 0);
2555 * Let the scheduling class do new task startup
2556 * management (if any):
2558 p->sched_class->task_new(rq, p);
2561 trace_sched_wakeup_new(rq, p, 1);
2562 check_preempt_curr(rq, p, 0);
2564 if (p->sched_class->task_wake_up)
2565 p->sched_class->task_wake_up(rq, p);
2567 task_rq_unlock(rq, &flags);
2570 #ifdef CONFIG_PREEMPT_NOTIFIERS
2573 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2574 * @notifier: notifier struct to register
2576 void preempt_notifier_register(struct preempt_notifier *notifier)
2578 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2580 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2583 * preempt_notifier_unregister - no longer interested in preemption notifications
2584 * @notifier: notifier struct to unregister
2586 * This is safe to call from within a preemption notifier.
2588 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2590 hlist_del(¬ifier->link);
2592 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2594 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2596 struct preempt_notifier *notifier;
2597 struct hlist_node *node;
2599 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2600 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2604 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2605 struct task_struct *next)
2607 struct preempt_notifier *notifier;
2608 struct hlist_node *node;
2610 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2611 notifier->ops->sched_out(notifier, next);
2614 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2616 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2621 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2622 struct task_struct *next)
2626 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2629 * prepare_task_switch - prepare to switch tasks
2630 * @rq: the runqueue preparing to switch
2631 * @prev: the current task that is being switched out
2632 * @next: the task we are going to switch to.
2634 * This is called with the rq lock held and interrupts off. It must
2635 * be paired with a subsequent finish_task_switch after the context
2638 * prepare_task_switch sets up locking and calls architecture specific
2642 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2643 struct task_struct *next)
2645 fire_sched_out_preempt_notifiers(prev, next);
2646 prepare_lock_switch(rq, next);
2647 prepare_arch_switch(next);
2651 * finish_task_switch - clean up after a task-switch
2652 * @rq: runqueue associated with task-switch
2653 * @prev: the thread we just switched away from.
2655 * finish_task_switch must be called after the context switch, paired
2656 * with a prepare_task_switch call before the context switch.
2657 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2658 * and do any other architecture-specific cleanup actions.
2660 * Note that we may have delayed dropping an mm in context_switch(). If
2661 * so, we finish that here outside of the runqueue lock. (Doing it
2662 * with the lock held can cause deadlocks; see schedule() for
2665 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2666 __releases(rq->lock)
2668 struct mm_struct *mm = rq->prev_mm;
2671 int post_schedule = 0;
2673 if (current->sched_class->needs_post_schedule)
2674 post_schedule = current->sched_class->needs_post_schedule(rq);
2680 * A task struct has one reference for the use as "current".
2681 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2682 * schedule one last time. The schedule call will never return, and
2683 * the scheduled task must drop that reference.
2684 * The test for TASK_DEAD must occur while the runqueue locks are
2685 * still held, otherwise prev could be scheduled on another cpu, die
2686 * there before we look at prev->state, and then the reference would
2688 * Manfred Spraul <manfred@colorfullife.com>
2690 prev_state = prev->state;
2691 finish_arch_switch(prev);
2692 finish_lock_switch(rq, prev);
2695 current->sched_class->post_schedule(rq);
2698 fire_sched_in_preempt_notifiers(current);
2701 if (unlikely(prev_state == TASK_DEAD)) {
2703 * Remove function-return probe instances associated with this
2704 * task and put them back on the free list.
2706 kprobe_flush_task(prev);
2707 put_task_struct(prev);
2712 * schedule_tail - first thing a freshly forked thread must call.
2713 * @prev: the thread we just switched away from.
2715 asmlinkage void schedule_tail(struct task_struct *prev)
2716 __releases(rq->lock)
2718 struct rq *rq = this_rq();
2720 finish_task_switch(rq, prev);
2721 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2722 /* In this case, finish_task_switch does not reenable preemption */
2725 if (current->set_child_tid)
2726 put_user(task_pid_vnr(current), current->set_child_tid);
2730 * context_switch - switch to the new MM and the new
2731 * thread's register state.
2734 context_switch(struct rq *rq, struct task_struct *prev,
2735 struct task_struct *next)
2737 struct mm_struct *mm, *oldmm;
2739 prepare_task_switch(rq, prev, next);
2740 trace_sched_switch(rq, prev, next);
2742 oldmm = prev->active_mm;
2744 * For paravirt, this is coupled with an exit in switch_to to
2745 * combine the page table reload and the switch backend into
2748 arch_enter_lazy_cpu_mode();
2750 if (unlikely(!mm)) {
2751 next->active_mm = oldmm;
2752 atomic_inc(&oldmm->mm_count);
2753 enter_lazy_tlb(oldmm, next);
2755 switch_mm(oldmm, mm, next);
2757 if (unlikely(!prev->mm)) {
2758 prev->active_mm = NULL;
2759 rq->prev_mm = oldmm;
2762 * Since the runqueue lock will be released by the next
2763 * task (which is an invalid locking op but in the case
2764 * of the scheduler it's an obvious special-case), so we
2765 * do an early lockdep release here:
2767 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2768 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2771 /* Here we just switch the register state and the stack. */
2772 switch_to(prev, next, prev);
2776 * this_rq must be evaluated again because prev may have moved
2777 * CPUs since it called schedule(), thus the 'rq' on its stack
2778 * frame will be invalid.
2780 finish_task_switch(this_rq(), prev);
2784 * nr_running, nr_uninterruptible and nr_context_switches:
2786 * externally visible scheduler statistics: current number of runnable
2787 * threads, current number of uninterruptible-sleeping threads, total
2788 * number of context switches performed since bootup.
2790 unsigned long nr_running(void)
2792 unsigned long i, sum = 0;
2794 for_each_online_cpu(i)
2795 sum += cpu_rq(i)->nr_running;
2800 unsigned long nr_uninterruptible(void)
2802 unsigned long i, sum = 0;
2804 for_each_possible_cpu(i)
2805 sum += cpu_rq(i)->nr_uninterruptible;
2808 * Since we read the counters lockless, it might be slightly
2809 * inaccurate. Do not allow it to go below zero though:
2811 if (unlikely((long)sum < 0))
2817 unsigned long long nr_context_switches(void)
2820 unsigned long long sum = 0;
2822 for_each_possible_cpu(i)
2823 sum += cpu_rq(i)->nr_switches;
2828 unsigned long nr_iowait(void)
2830 unsigned long i, sum = 0;
2832 for_each_possible_cpu(i)
2833 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2838 unsigned long nr_active(void)
2840 unsigned long i, running = 0, uninterruptible = 0;
2842 for_each_online_cpu(i) {
2843 running += cpu_rq(i)->nr_running;
2844 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2847 if (unlikely((long)uninterruptible < 0))
2848 uninterruptible = 0;
2850 return running + uninterruptible;
2854 * Update rq->cpu_load[] statistics. This function is usually called every
2855 * scheduler tick (TICK_NSEC).
2857 static void update_cpu_load(struct rq *this_rq)
2859 unsigned long this_load = this_rq->load.weight;
2862 this_rq->nr_load_updates++;
2864 /* Update our load: */
2865 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2866 unsigned long old_load, new_load;
2868 /* scale is effectively 1 << i now, and >> i divides by scale */
2870 old_load = this_rq->cpu_load[i];
2871 new_load = this_load;
2873 * Round up the averaging division if load is increasing. This
2874 * prevents us from getting stuck on 9 if the load is 10, for
2877 if (new_load > old_load)
2878 new_load += scale-1;
2879 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2886 * double_rq_lock - safely lock two runqueues
2888 * Note this does not disable interrupts like task_rq_lock,
2889 * you need to do so manually before calling.
2891 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2892 __acquires(rq1->lock)
2893 __acquires(rq2->lock)
2895 BUG_ON(!irqs_disabled());
2897 spin_lock(&rq1->lock);
2898 __acquire(rq2->lock); /* Fake it out ;) */
2901 spin_lock(&rq1->lock);
2902 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2904 spin_lock(&rq2->lock);
2905 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2908 update_rq_clock(rq1);
2909 update_rq_clock(rq2);
2913 * double_rq_unlock - safely unlock two runqueues
2915 * Note this does not restore interrupts like task_rq_unlock,
2916 * you need to do so manually after calling.
2918 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2919 __releases(rq1->lock)
2920 __releases(rq2->lock)
2922 spin_unlock(&rq1->lock);
2924 spin_unlock(&rq2->lock);
2926 __release(rq2->lock);
2930 * If dest_cpu is allowed for this process, migrate the task to it.
2931 * This is accomplished by forcing the cpu_allowed mask to only
2932 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2933 * the cpu_allowed mask is restored.
2935 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2937 struct migration_req req;
2938 unsigned long flags;
2941 rq = task_rq_lock(p, &flags);
2942 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2943 || unlikely(!cpu_active(dest_cpu)))
2946 /* force the process onto the specified CPU */
2947 if (migrate_task(p, dest_cpu, &req)) {
2948 /* Need to wait for migration thread (might exit: take ref). */
2949 struct task_struct *mt = rq->migration_thread;
2951 get_task_struct(mt);
2952 task_rq_unlock(rq, &flags);
2953 wake_up_process(mt);
2954 put_task_struct(mt);
2955 wait_for_completion(&req.done);
2960 task_rq_unlock(rq, &flags);
2964 * sched_exec - execve() is a valuable balancing opportunity, because at
2965 * this point the task has the smallest effective memory and cache footprint.
2967 void sched_exec(void)
2969 int new_cpu, this_cpu = get_cpu();
2970 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2972 if (new_cpu != this_cpu)
2973 sched_migrate_task(current, new_cpu);
2977 * pull_task - move a task from a remote runqueue to the local runqueue.
2978 * Both runqueues must be locked.
2980 static void pull_task(struct rq *src_rq, struct task_struct *p,
2981 struct rq *this_rq, int this_cpu)
2983 deactivate_task(src_rq, p, 0);
2984 set_task_cpu(p, this_cpu);
2985 activate_task(this_rq, p, 0);
2987 * Note that idle threads have a prio of MAX_PRIO, for this test
2988 * to be always true for them.
2990 check_preempt_curr(this_rq, p, 0);
2994 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2997 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2998 struct sched_domain *sd, enum cpu_idle_type idle,
3002 * We do not migrate tasks that are:
3003 * 1) running (obviously), or
3004 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3005 * 3) are cache-hot on their current CPU.
3007 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3008 schedstat_inc(p, se.nr_failed_migrations_affine);
3013 if (task_running(rq, p)) {
3014 schedstat_inc(p, se.nr_failed_migrations_running);
3019 * Aggressive migration if:
3020 * 1) task is cache cold, or
3021 * 2) too many balance attempts have failed.
3024 if (!task_hot(p, rq->clock, sd) ||
3025 sd->nr_balance_failed > sd->cache_nice_tries) {
3026 #ifdef CONFIG_SCHEDSTATS
3027 if (task_hot(p, rq->clock, sd)) {
3028 schedstat_inc(sd, lb_hot_gained[idle]);
3029 schedstat_inc(p, se.nr_forced_migrations);
3035 if (task_hot(p, rq->clock, sd)) {
3036 schedstat_inc(p, se.nr_failed_migrations_hot);
3042 static unsigned long
3043 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3044 unsigned long max_load_move, struct sched_domain *sd,
3045 enum cpu_idle_type idle, int *all_pinned,
3046 int *this_best_prio, struct rq_iterator *iterator)
3048 int loops = 0, pulled = 0, pinned = 0;
3049 struct task_struct *p;
3050 long rem_load_move = max_load_move;
3052 if (max_load_move == 0)
3058 * Start the load-balancing iterator:
3060 p = iterator->start(iterator->arg);
3062 if (!p || loops++ > sysctl_sched_nr_migrate)
3065 if ((p->se.load.weight >> 1) > rem_load_move ||
3066 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3067 p = iterator->next(iterator->arg);
3071 pull_task(busiest, p, this_rq, this_cpu);
3073 rem_load_move -= p->se.load.weight;
3075 #ifdef CONFIG_PREEMPT
3077 * NEWIDLE balancing is a source of latency, so preemptible kernels
3078 * will stop after the first task is pulled to minimize the critical
3081 if (idle == CPU_NEWLY_IDLE)
3086 * We only want to steal up to the prescribed amount of weighted load.
3088 if (rem_load_move > 0) {
3089 if (p->prio < *this_best_prio)
3090 *this_best_prio = p->prio;
3091 p = iterator->next(iterator->arg);
3096 * Right now, this is one of only two places pull_task() is called,
3097 * so we can safely collect pull_task() stats here rather than
3098 * inside pull_task().
3100 schedstat_add(sd, lb_gained[idle], pulled);
3103 *all_pinned = pinned;
3105 return max_load_move - rem_load_move;
3109 * move_tasks tries to move up to max_load_move weighted load from busiest to
3110 * this_rq, as part of a balancing operation within domain "sd".
3111 * Returns 1 if successful and 0 otherwise.
3113 * Called with both runqueues locked.
3115 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3116 unsigned long max_load_move,
3117 struct sched_domain *sd, enum cpu_idle_type idle,
3120 const struct sched_class *class = sched_class_highest;
3121 unsigned long total_load_moved = 0;
3122 int this_best_prio = this_rq->curr->prio;
3126 class->load_balance(this_rq, this_cpu, busiest,
3127 max_load_move - total_load_moved,
3128 sd, idle, all_pinned, &this_best_prio);
3129 class = class->next;
3131 #ifdef CONFIG_PREEMPT
3133 * NEWIDLE balancing is a source of latency, so preemptible
3134 * kernels will stop after the first task is pulled to minimize
3135 * the critical section.
3137 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3140 } while (class && max_load_move > total_load_moved);
3142 return total_load_moved > 0;
3146 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3147 struct sched_domain *sd, enum cpu_idle_type idle,
3148 struct rq_iterator *iterator)
3150 struct task_struct *p = iterator->start(iterator->arg);
3154 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3155 pull_task(busiest, p, this_rq, this_cpu);
3157 * Right now, this is only the second place pull_task()
3158 * is called, so we can safely collect pull_task()
3159 * stats here rather than inside pull_task().
3161 schedstat_inc(sd, lb_gained[idle]);
3165 p = iterator->next(iterator->arg);
3172 * move_one_task tries to move exactly one task from busiest to this_rq, as
3173 * part of active balancing operations within "domain".
3174 * Returns 1 if successful and 0 otherwise.
3176 * Called with both runqueues locked.
3178 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3179 struct sched_domain *sd, enum cpu_idle_type idle)
3181 const struct sched_class *class;
3183 for (class = sched_class_highest; class; class = class->next)
3184 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3191 * find_busiest_group finds and returns the busiest CPU group within the
3192 * domain. It calculates and returns the amount of weighted load which
3193 * should be moved to restore balance via the imbalance parameter.
3195 static struct sched_group *
3196 find_busiest_group(struct sched_domain *sd, int this_cpu,
3197 unsigned long *imbalance, enum cpu_idle_type idle,
3198 int *sd_idle, const struct cpumask *cpus, int *balance)
3200 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3201 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3202 unsigned long max_pull;
3203 unsigned long busiest_load_per_task, busiest_nr_running;
3204 unsigned long this_load_per_task, this_nr_running;
3205 int load_idx, group_imb = 0;
3206 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3207 int power_savings_balance = 1;
3208 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3209 unsigned long min_nr_running = ULONG_MAX;
3210 struct sched_group *group_min = NULL, *group_leader = NULL;
3213 max_load = this_load = total_load = total_pwr = 0;
3214 busiest_load_per_task = busiest_nr_running = 0;
3215 this_load_per_task = this_nr_running = 0;
3217 if (idle == CPU_NOT_IDLE)
3218 load_idx = sd->busy_idx;
3219 else if (idle == CPU_NEWLY_IDLE)
3220 load_idx = sd->newidle_idx;
3222 load_idx = sd->idle_idx;
3225 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3228 int __group_imb = 0;
3229 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3230 unsigned long sum_nr_running, sum_weighted_load;
3231 unsigned long sum_avg_load_per_task;
3232 unsigned long avg_load_per_task;
3234 local_group = cpumask_test_cpu(this_cpu,
3235 sched_group_cpus(group));
3238 balance_cpu = cpumask_first(sched_group_cpus(group));
3240 /* Tally up the load of all CPUs in the group */
3241 sum_weighted_load = sum_nr_running = avg_load = 0;
3242 sum_avg_load_per_task = avg_load_per_task = 0;
3245 min_cpu_load = ~0UL;
3247 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3248 struct rq *rq = cpu_rq(i);
3250 if (*sd_idle && rq->nr_running)
3253 /* Bias balancing toward cpus of our domain */
3255 if (idle_cpu(i) && !first_idle_cpu) {
3260 load = target_load(i, load_idx);
3262 load = source_load(i, load_idx);
3263 if (load > max_cpu_load)
3264 max_cpu_load = load;
3265 if (min_cpu_load > load)
3266 min_cpu_load = load;
3270 sum_nr_running += rq->nr_running;
3271 sum_weighted_load += weighted_cpuload(i);
3273 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3277 * First idle cpu or the first cpu(busiest) in this sched group
3278 * is eligible for doing load balancing at this and above
3279 * domains. In the newly idle case, we will allow all the cpu's
3280 * to do the newly idle load balance.
3282 if (idle != CPU_NEWLY_IDLE && local_group &&
3283 balance_cpu != this_cpu && balance) {
3288 total_load += avg_load;
3289 total_pwr += group->__cpu_power;
3291 /* Adjust by relative CPU power of the group */
3292 avg_load = sg_div_cpu_power(group,
3293 avg_load * SCHED_LOAD_SCALE);
3297 * Consider the group unbalanced when the imbalance is larger
3298 * than the average weight of two tasks.
3300 * APZ: with cgroup the avg task weight can vary wildly and
3301 * might not be a suitable number - should we keep a
3302 * normalized nr_running number somewhere that negates
3305 avg_load_per_task = sg_div_cpu_power(group,
3306 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3308 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3311 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3314 this_load = avg_load;
3316 this_nr_running = sum_nr_running;
3317 this_load_per_task = sum_weighted_load;
3318 } else if (avg_load > max_load &&
3319 (sum_nr_running > group_capacity || __group_imb)) {
3320 max_load = avg_load;
3322 busiest_nr_running = sum_nr_running;
3323 busiest_load_per_task = sum_weighted_load;
3324 group_imb = __group_imb;
3327 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3329 * Busy processors will not participate in power savings
3332 if (idle == CPU_NOT_IDLE ||
3333 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3337 * If the local group is idle or completely loaded
3338 * no need to do power savings balance at this domain
3340 if (local_group && (this_nr_running >= group_capacity ||
3342 power_savings_balance = 0;
3345 * If a group is already running at full capacity or idle,
3346 * don't include that group in power savings calculations
3348 if (!power_savings_balance || sum_nr_running >= group_capacity
3353 * Calculate the group which has the least non-idle load.
3354 * This is the group from where we need to pick up the load
3357 if ((sum_nr_running < min_nr_running) ||
3358 (sum_nr_running == min_nr_running &&
3359 cpumask_first(sched_group_cpus(group)) >
3360 cpumask_first(sched_group_cpus(group_min)))) {
3362 min_nr_running = sum_nr_running;
3363 min_load_per_task = sum_weighted_load /
3368 * Calculate the group which is almost near its
3369 * capacity but still has some space to pick up some load
3370 * from other group and save more power
3372 if (sum_nr_running <= group_capacity - 1) {
3373 if (sum_nr_running > leader_nr_running ||
3374 (sum_nr_running == leader_nr_running &&
3375 cpumask_first(sched_group_cpus(group)) <
3376 cpumask_first(sched_group_cpus(group_leader)))) {
3377 group_leader = group;
3378 leader_nr_running = sum_nr_running;
3383 group = group->next;
3384 } while (group != sd->groups);
3386 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3389 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3391 if (this_load >= avg_load ||
3392 100*max_load <= sd->imbalance_pct*this_load)
3395 busiest_load_per_task /= busiest_nr_running;
3397 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3400 * We're trying to get all the cpus to the average_load, so we don't
3401 * want to push ourselves above the average load, nor do we wish to
3402 * reduce the max loaded cpu below the average load, as either of these
3403 * actions would just result in more rebalancing later, and ping-pong
3404 * tasks around. Thus we look for the minimum possible imbalance.
3405 * Negative imbalances (*we* are more loaded than anyone else) will
3406 * be counted as no imbalance for these purposes -- we can't fix that
3407 * by pulling tasks to us. Be careful of negative numbers as they'll
3408 * appear as very large values with unsigned longs.
3410 if (max_load <= busiest_load_per_task)
3414 * In the presence of smp nice balancing, certain scenarios can have
3415 * max load less than avg load(as we skip the groups at or below
3416 * its cpu_power, while calculating max_load..)
3418 if (max_load < avg_load) {
3420 goto small_imbalance;
3423 /* Don't want to pull so many tasks that a group would go idle */
3424 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3426 /* How much load to actually move to equalise the imbalance */
3427 *imbalance = min(max_pull * busiest->__cpu_power,
3428 (avg_load - this_load) * this->__cpu_power)
3432 * if *imbalance is less than the average load per runnable task
3433 * there is no gaurantee that any tasks will be moved so we'll have
3434 * a think about bumping its value to force at least one task to be
3437 if (*imbalance < busiest_load_per_task) {
3438 unsigned long tmp, pwr_now, pwr_move;
3442 pwr_move = pwr_now = 0;
3444 if (this_nr_running) {
3445 this_load_per_task /= this_nr_running;
3446 if (busiest_load_per_task > this_load_per_task)
3449 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3451 if (max_load - this_load + busiest_load_per_task >=
3452 busiest_load_per_task * imbn) {
3453 *imbalance = busiest_load_per_task;
3458 * OK, we don't have enough imbalance to justify moving tasks,
3459 * however we may be able to increase total CPU power used by
3463 pwr_now += busiest->__cpu_power *
3464 min(busiest_load_per_task, max_load);
3465 pwr_now += this->__cpu_power *
3466 min(this_load_per_task, this_load);
3467 pwr_now /= SCHED_LOAD_SCALE;
3469 /* Amount of load we'd subtract */
3470 tmp = sg_div_cpu_power(busiest,
3471 busiest_load_per_task * SCHED_LOAD_SCALE);
3473 pwr_move += busiest->__cpu_power *
3474 min(busiest_load_per_task, max_load - tmp);
3476 /* Amount of load we'd add */
3477 if (max_load * busiest->__cpu_power <
3478 busiest_load_per_task * SCHED_LOAD_SCALE)
3479 tmp = sg_div_cpu_power(this,
3480 max_load * busiest->__cpu_power);
3482 tmp = sg_div_cpu_power(this,
3483 busiest_load_per_task * SCHED_LOAD_SCALE);
3484 pwr_move += this->__cpu_power *
3485 min(this_load_per_task, this_load + tmp);
3486 pwr_move /= SCHED_LOAD_SCALE;
3488 /* Move if we gain throughput */
3489 if (pwr_move > pwr_now)
3490 *imbalance = busiest_load_per_task;
3496 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3497 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3500 if (this == group_leader && group_leader != group_min) {
3501 *imbalance = min_load_per_task;
3502 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3503 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3504 cpumask_first(sched_group_cpus(group_leader));
3515 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3518 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3519 unsigned long imbalance, const struct cpumask *cpus)
3521 struct rq *busiest = NULL, *rq;
3522 unsigned long max_load = 0;
3525 for_each_cpu(i, sched_group_cpus(group)) {
3528 if (!cpumask_test_cpu(i, cpus))
3532 wl = weighted_cpuload(i);
3534 if (rq->nr_running == 1 && wl > imbalance)
3537 if (wl > max_load) {
3547 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3548 * so long as it is large enough.
3550 #define MAX_PINNED_INTERVAL 512
3553 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3554 * tasks if there is an imbalance.
3556 static int load_balance(int this_cpu, struct rq *this_rq,
3557 struct sched_domain *sd, enum cpu_idle_type idle,
3558 int *balance, struct cpumask *cpus)
3560 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3561 struct sched_group *group;
3562 unsigned long imbalance;
3564 unsigned long flags;
3566 cpumask_setall(cpus);
3569 * When power savings policy is enabled for the parent domain, idle
3570 * sibling can pick up load irrespective of busy siblings. In this case,
3571 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3572 * portraying it as CPU_NOT_IDLE.
3574 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3575 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3578 schedstat_inc(sd, lb_count[idle]);
3582 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3589 schedstat_inc(sd, lb_nobusyg[idle]);
3593 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3595 schedstat_inc(sd, lb_nobusyq[idle]);
3599 BUG_ON(busiest == this_rq);
3601 schedstat_add(sd, lb_imbalance[idle], imbalance);
3604 if (busiest->nr_running > 1) {
3606 * Attempt to move tasks. If find_busiest_group has found
3607 * an imbalance but busiest->nr_running <= 1, the group is
3608 * still unbalanced. ld_moved simply stays zero, so it is
3609 * correctly treated as an imbalance.
3611 local_irq_save(flags);
3612 double_rq_lock(this_rq, busiest);
3613 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3614 imbalance, sd, idle, &all_pinned);
3615 double_rq_unlock(this_rq, busiest);
3616 local_irq_restore(flags);
3619 * some other cpu did the load balance for us.
3621 if (ld_moved && this_cpu != smp_processor_id())
3622 resched_cpu(this_cpu);
3624 /* All tasks on this runqueue were pinned by CPU affinity */
3625 if (unlikely(all_pinned)) {
3626 cpumask_clear_cpu(cpu_of(busiest), cpus);
3627 if (!cpumask_empty(cpus))
3634 schedstat_inc(sd, lb_failed[idle]);
3635 sd->nr_balance_failed++;
3637 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3639 spin_lock_irqsave(&busiest->lock, flags);
3641 /* don't kick the migration_thread, if the curr
3642 * task on busiest cpu can't be moved to this_cpu
3644 if (!cpumask_test_cpu(this_cpu,
3645 &busiest->curr->cpus_allowed)) {
3646 spin_unlock_irqrestore(&busiest->lock, flags);
3648 goto out_one_pinned;
3651 if (!busiest->active_balance) {
3652 busiest->active_balance = 1;
3653 busiest->push_cpu = this_cpu;
3656 spin_unlock_irqrestore(&busiest->lock, flags);
3658 wake_up_process(busiest->migration_thread);
3661 * We've kicked active balancing, reset the failure
3664 sd->nr_balance_failed = sd->cache_nice_tries+1;
3667 sd->nr_balance_failed = 0;
3669 if (likely(!active_balance)) {
3670 /* We were unbalanced, so reset the balancing interval */
3671 sd->balance_interval = sd->min_interval;
3674 * If we've begun active balancing, start to back off. This
3675 * case may not be covered by the all_pinned logic if there
3676 * is only 1 task on the busy runqueue (because we don't call
3679 if (sd->balance_interval < sd->max_interval)
3680 sd->balance_interval *= 2;
3683 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3684 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3690 schedstat_inc(sd, lb_balanced[idle]);
3692 sd->nr_balance_failed = 0;
3695 /* tune up the balancing interval */
3696 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3697 (sd->balance_interval < sd->max_interval))
3698 sd->balance_interval *= 2;
3700 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3701 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3712 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3713 * tasks if there is an imbalance.
3715 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3716 * this_rq is locked.
3719 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3720 struct cpumask *cpus)
3722 struct sched_group *group;
3723 struct rq *busiest = NULL;
3724 unsigned long imbalance;
3729 cpumask_setall(cpus);
3732 * When power savings policy is enabled for the parent domain, idle
3733 * sibling can pick up load irrespective of busy siblings. In this case,
3734 * let the state of idle sibling percolate up as IDLE, instead of
3735 * portraying it as CPU_NOT_IDLE.
3737 if (sd->flags & SD_SHARE_CPUPOWER &&
3738 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3741 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3743 update_shares_locked(this_rq, sd);
3744 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3745 &sd_idle, cpus, NULL);
3747 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3751 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3753 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3757 BUG_ON(busiest == this_rq);
3759 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3762 if (busiest->nr_running > 1) {
3763 /* Attempt to move tasks */
3764 double_lock_balance(this_rq, busiest);
3765 /* this_rq->clock is already updated */
3766 update_rq_clock(busiest);
3767 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3768 imbalance, sd, CPU_NEWLY_IDLE,
3770 double_unlock_balance(this_rq, busiest);
3772 if (unlikely(all_pinned)) {
3773 cpumask_clear_cpu(cpu_of(busiest), cpus);
3774 if (!cpumask_empty(cpus))
3780 int active_balance = 0;
3782 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3783 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3784 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3787 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3790 if (sd->nr_balance_failed++ < 2)
3794 * The only task running in a non-idle cpu can be moved to this
3795 * cpu in an attempt to completely freeup the other CPU
3796 * package. The same method used to move task in load_balance()
3797 * have been extended for load_balance_newidle() to speedup
3798 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3800 * The package power saving logic comes from
3801 * find_busiest_group(). If there are no imbalance, then
3802 * f_b_g() will return NULL. However when sched_mc={1,2} then
3803 * f_b_g() will select a group from which a running task may be
3804 * pulled to this cpu in order to make the other package idle.
3805 * If there is no opportunity to make a package idle and if
3806 * there are no imbalance, then f_b_g() will return NULL and no
3807 * action will be taken in load_balance_newidle().
3809 * Under normal task pull operation due to imbalance, there
3810 * will be more than one task in the source run queue and
3811 * move_tasks() will succeed. ld_moved will be true and this
3812 * active balance code will not be triggered.
3815 /* Lock busiest in correct order while this_rq is held */
3816 double_lock_balance(this_rq, busiest);
3819 * don't kick the migration_thread, if the curr
3820 * task on busiest cpu can't be moved to this_cpu
3822 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3823 double_unlock_balance(this_rq, busiest);
3828 if (!busiest->active_balance) {
3829 busiest->active_balance = 1;
3830 busiest->push_cpu = this_cpu;
3834 double_unlock_balance(this_rq, busiest);
3836 * Should not call ttwu while holding a rq->lock
3838 spin_unlock(&this_rq->lock);
3840 wake_up_process(busiest->migration_thread);
3841 spin_lock(&this_rq->lock);
3844 sd->nr_balance_failed = 0;
3846 update_shares_locked(this_rq, sd);
3850 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3851 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3852 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3854 sd->nr_balance_failed = 0;
3860 * idle_balance is called by schedule() if this_cpu is about to become
3861 * idle. Attempts to pull tasks from other CPUs.
3863 static void idle_balance(int this_cpu, struct rq *this_rq)
3865 struct sched_domain *sd;
3866 int pulled_task = 0;
3867 unsigned long next_balance = jiffies + HZ;
3868 cpumask_var_t tmpmask;
3870 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3873 for_each_domain(this_cpu, sd) {
3874 unsigned long interval;
3876 if (!(sd->flags & SD_LOAD_BALANCE))
3879 if (sd->flags & SD_BALANCE_NEWIDLE)
3880 /* If we've pulled tasks over stop searching: */
3881 pulled_task = load_balance_newidle(this_cpu, this_rq,
3884 interval = msecs_to_jiffies(sd->balance_interval);
3885 if (time_after(next_balance, sd->last_balance + interval))
3886 next_balance = sd->last_balance + interval;
3890 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3892 * We are going idle. next_balance may be set based on
3893 * a busy processor. So reset next_balance.
3895 this_rq->next_balance = next_balance;
3897 free_cpumask_var(tmpmask);
3901 * active_load_balance is run by migration threads. It pushes running tasks
3902 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3903 * running on each physical CPU where possible, and avoids physical /
3904 * logical imbalances.
3906 * Called with busiest_rq locked.
3908 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3910 int target_cpu = busiest_rq->push_cpu;
3911 struct sched_domain *sd;
3912 struct rq *target_rq;
3914 /* Is there any task to move? */
3915 if (busiest_rq->nr_running <= 1)
3918 target_rq = cpu_rq(target_cpu);
3921 * This condition is "impossible", if it occurs
3922 * we need to fix it. Originally reported by
3923 * Bjorn Helgaas on a 128-cpu setup.
3925 BUG_ON(busiest_rq == target_rq);
3927 /* move a task from busiest_rq to target_rq */
3928 double_lock_balance(busiest_rq, target_rq);
3929 update_rq_clock(busiest_rq);
3930 update_rq_clock(target_rq);
3932 /* Search for an sd spanning us and the target CPU. */
3933 for_each_domain(target_cpu, sd) {
3934 if ((sd->flags & SD_LOAD_BALANCE) &&
3935 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3940 schedstat_inc(sd, alb_count);
3942 if (move_one_task(target_rq, target_cpu, busiest_rq,
3944 schedstat_inc(sd, alb_pushed);
3946 schedstat_inc(sd, alb_failed);
3948 double_unlock_balance(busiest_rq, target_rq);
3953 atomic_t load_balancer;
3954 cpumask_var_t cpu_mask;
3955 } nohz ____cacheline_aligned = {
3956 .load_balancer = ATOMIC_INIT(-1),
3960 * This routine will try to nominate the ilb (idle load balancing)
3961 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3962 * load balancing on behalf of all those cpus. If all the cpus in the system
3963 * go into this tickless mode, then there will be no ilb owner (as there is
3964 * no need for one) and all the cpus will sleep till the next wakeup event
3967 * For the ilb owner, tick is not stopped. And this tick will be used
3968 * for idle load balancing. ilb owner will still be part of
3971 * While stopping the tick, this cpu will become the ilb owner if there
3972 * is no other owner. And will be the owner till that cpu becomes busy
3973 * or if all cpus in the system stop their ticks at which point
3974 * there is no need for ilb owner.
3976 * When the ilb owner becomes busy, it nominates another owner, during the
3977 * next busy scheduler_tick()
3979 int select_nohz_load_balancer(int stop_tick)
3981 int cpu = smp_processor_id();
3984 cpu_rq(cpu)->in_nohz_recently = 1;
3986 if (!cpu_active(cpu)) {
3987 if (atomic_read(&nohz.load_balancer) != cpu)
3991 * If we are going offline and still the leader,
3994 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4000 cpumask_set_cpu(cpu, nohz.cpu_mask);
4002 /* time for ilb owner also to sleep */
4003 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4004 if (atomic_read(&nohz.load_balancer) == cpu)
4005 atomic_set(&nohz.load_balancer, -1);
4009 if (atomic_read(&nohz.load_balancer) == -1) {
4010 /* make me the ilb owner */
4011 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4013 } else if (atomic_read(&nohz.load_balancer) == cpu)
4016 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4019 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4021 if (atomic_read(&nohz.load_balancer) == cpu)
4022 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4029 static DEFINE_SPINLOCK(balancing);
4032 * It checks each scheduling domain to see if it is due to be balanced,
4033 * and initiates a balancing operation if so.
4035 * Balancing parameters are set up in arch_init_sched_domains.
4037 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4040 struct rq *rq = cpu_rq(cpu);
4041 unsigned long interval;
4042 struct sched_domain *sd;
4043 /* Earliest time when we have to do rebalance again */
4044 unsigned long next_balance = jiffies + 60*HZ;
4045 int update_next_balance = 0;
4049 /* Fails alloc? Rebalancing probably not a priority right now. */
4050 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4053 for_each_domain(cpu, sd) {
4054 if (!(sd->flags & SD_LOAD_BALANCE))
4057 interval = sd->balance_interval;
4058 if (idle != CPU_IDLE)
4059 interval *= sd->busy_factor;
4061 /* scale ms to jiffies */
4062 interval = msecs_to_jiffies(interval);
4063 if (unlikely(!interval))
4065 if (interval > HZ*NR_CPUS/10)
4066 interval = HZ*NR_CPUS/10;
4068 need_serialize = sd->flags & SD_SERIALIZE;
4070 if (need_serialize) {
4071 if (!spin_trylock(&balancing))
4075 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4076 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4078 * We've pulled tasks over so either we're no
4079 * longer idle, or one of our SMT siblings is
4082 idle = CPU_NOT_IDLE;
4084 sd->last_balance = jiffies;
4087 spin_unlock(&balancing);
4089 if (time_after(next_balance, sd->last_balance + interval)) {
4090 next_balance = sd->last_balance + interval;
4091 update_next_balance = 1;
4095 * Stop the load balance at this level. There is another
4096 * CPU in our sched group which is doing load balancing more
4104 * next_balance will be updated only when there is a need.
4105 * When the cpu is attached to null domain for ex, it will not be
4108 if (likely(update_next_balance))
4109 rq->next_balance = next_balance;
4111 free_cpumask_var(tmp);
4115 * run_rebalance_domains is triggered when needed from the scheduler tick.
4116 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4117 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4119 static void run_rebalance_domains(struct softirq_action *h)
4121 int this_cpu = smp_processor_id();
4122 struct rq *this_rq = cpu_rq(this_cpu);
4123 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4124 CPU_IDLE : CPU_NOT_IDLE;
4126 rebalance_domains(this_cpu, idle);
4130 * If this cpu is the owner for idle load balancing, then do the
4131 * balancing on behalf of the other idle cpus whose ticks are
4134 if (this_rq->idle_at_tick &&
4135 atomic_read(&nohz.load_balancer) == this_cpu) {
4139 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4140 if (balance_cpu == this_cpu)
4144 * If this cpu gets work to do, stop the load balancing
4145 * work being done for other cpus. Next load
4146 * balancing owner will pick it up.
4151 rebalance_domains(balance_cpu, CPU_IDLE);
4153 rq = cpu_rq(balance_cpu);
4154 if (time_after(this_rq->next_balance, rq->next_balance))
4155 this_rq->next_balance = rq->next_balance;
4162 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4164 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4165 * idle load balancing owner or decide to stop the periodic load balancing,
4166 * if the whole system is idle.
4168 static inline void trigger_load_balance(struct rq *rq, int cpu)
4172 * If we were in the nohz mode recently and busy at the current
4173 * scheduler tick, then check if we need to nominate new idle
4176 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4177 rq->in_nohz_recently = 0;
4179 if (atomic_read(&nohz.load_balancer) == cpu) {
4180 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4181 atomic_set(&nohz.load_balancer, -1);
4184 if (atomic_read(&nohz.load_balancer) == -1) {
4186 * simple selection for now: Nominate the
4187 * first cpu in the nohz list to be the next
4190 * TBD: Traverse the sched domains and nominate
4191 * the nearest cpu in the nohz.cpu_mask.
4193 int ilb = cpumask_first(nohz.cpu_mask);
4195 if (ilb < nr_cpu_ids)
4201 * If this cpu is idle and doing idle load balancing for all the
4202 * cpus with ticks stopped, is it time for that to stop?
4204 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4205 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4211 * If this cpu is idle and the idle load balancing is done by
4212 * someone else, then no need raise the SCHED_SOFTIRQ
4214 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4215 cpumask_test_cpu(cpu, nohz.cpu_mask))
4218 if (time_after_eq(jiffies, rq->next_balance))
4219 raise_softirq(SCHED_SOFTIRQ);
4222 #else /* CONFIG_SMP */
4225 * on UP we do not need to balance between CPUs:
4227 static inline void idle_balance(int cpu, struct rq *rq)
4233 DEFINE_PER_CPU(struct kernel_stat, kstat);
4235 EXPORT_PER_CPU_SYMBOL(kstat);
4238 * Return any ns on the sched_clock that have not yet been banked in
4239 * @p in case that task is currently running.
4241 unsigned long long task_delta_exec(struct task_struct *p)
4243 unsigned long flags;
4247 rq = task_rq_lock(p, &flags);
4249 if (task_current(rq, p)) {
4252 update_rq_clock(rq);
4253 delta_exec = rq->clock - p->se.exec_start;
4254 if ((s64)delta_exec > 0)
4258 task_rq_unlock(rq, &flags);
4264 * Account user cpu time to a process.
4265 * @p: the process that the cpu time gets accounted to
4266 * @cputime: the cpu time spent in user space since the last update
4267 * @cputime_scaled: cputime scaled by cpu frequency
4269 void account_user_time(struct task_struct *p, cputime_t cputime,
4270 cputime_t cputime_scaled)
4272 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4275 /* Add user time to process. */
4276 p->utime = cputime_add(p->utime, cputime);
4277 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4278 account_group_user_time(p, cputime);
4280 /* Add user time to cpustat. */
4281 tmp = cputime_to_cputime64(cputime);
4282 if (TASK_NICE(p) > 0)
4283 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4285 cpustat->user = cputime64_add(cpustat->user, tmp);
4286 /* Account for user time used */
4287 acct_update_integrals(p);
4291 * Account guest cpu time to a process.
4292 * @p: the process that the cpu time gets accounted to
4293 * @cputime: the cpu time spent in virtual machine since the last update
4294 * @cputime_scaled: cputime scaled by cpu frequency
4296 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4297 cputime_t cputime_scaled)
4300 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4302 tmp = cputime_to_cputime64(cputime);
4304 /* Add guest time to process. */
4305 p->utime = cputime_add(p->utime, cputime);
4306 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4307 account_group_user_time(p, cputime);
4308 p->gtime = cputime_add(p->gtime, cputime);
4310 /* Add guest time to cpustat. */
4311 cpustat->user = cputime64_add(cpustat->user, tmp);
4312 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4316 * Account system cpu time to a process.
4317 * @p: the process that the cpu time gets accounted to
4318 * @hardirq_offset: the offset to subtract from hardirq_count()
4319 * @cputime: the cpu time spent in kernel space since the last update
4320 * @cputime_scaled: cputime scaled by cpu frequency
4322 void account_system_time(struct task_struct *p, int hardirq_offset,
4323 cputime_t cputime, cputime_t cputime_scaled)
4325 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4328 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4329 account_guest_time(p, cputime, cputime_scaled);
4333 /* Add system time to process. */
4334 p->stime = cputime_add(p->stime, cputime);
4335 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4336 account_group_system_time(p, cputime);
4338 /* Add system time to cpustat. */
4339 tmp = cputime_to_cputime64(cputime);
4340 if (hardirq_count() - hardirq_offset)
4341 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4342 else if (softirq_count())
4343 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4345 cpustat->system = cputime64_add(cpustat->system, tmp);
4347 /* Account for system time used */
4348 acct_update_integrals(p);
4352 * Account for involuntary wait time.
4353 * @steal: the cpu time spent in involuntary wait
4355 void account_steal_time(cputime_t cputime)
4357 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4358 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4360 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4364 * Account for idle time.
4365 * @cputime: the cpu time spent in idle wait
4367 void account_idle_time(cputime_t cputime)
4369 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4370 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4371 struct rq *rq = this_rq();
4373 if (atomic_read(&rq->nr_iowait) > 0)
4374 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4376 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4379 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4382 * Account a single tick of cpu time.
4383 * @p: the process that the cpu time gets accounted to
4384 * @user_tick: indicates if the tick is a user or a system tick
4386 void account_process_tick(struct task_struct *p, int user_tick)
4388 cputime_t one_jiffy = jiffies_to_cputime(1);
4389 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4390 struct rq *rq = this_rq();
4393 account_user_time(p, one_jiffy, one_jiffy_scaled);
4394 else if (p != rq->idle)
4395 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4398 account_idle_time(one_jiffy);
4402 * Account multiple ticks of steal time.
4403 * @p: the process from which the cpu time has been stolen
4404 * @ticks: number of stolen ticks
4406 void account_steal_ticks(unsigned long ticks)
4408 account_steal_time(jiffies_to_cputime(ticks));
4412 * Account multiple ticks of idle time.
4413 * @ticks: number of stolen ticks
4415 void account_idle_ticks(unsigned long ticks)
4417 account_idle_time(jiffies_to_cputime(ticks));
4423 * Use precise platform statistics if available:
4425 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4426 cputime_t task_utime(struct task_struct *p)
4431 cputime_t task_stime(struct task_struct *p)
4436 cputime_t task_utime(struct task_struct *p)
4438 clock_t utime = cputime_to_clock_t(p->utime),
4439 total = utime + cputime_to_clock_t(p->stime);
4443 * Use CFS's precise accounting:
4445 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4449 do_div(temp, total);
4451 utime = (clock_t)temp;
4453 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4454 return p->prev_utime;
4457 cputime_t task_stime(struct task_struct *p)
4462 * Use CFS's precise accounting. (we subtract utime from
4463 * the total, to make sure the total observed by userspace
4464 * grows monotonically - apps rely on that):
4466 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4467 cputime_to_clock_t(task_utime(p));
4470 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4472 return p->prev_stime;
4476 inline cputime_t task_gtime(struct task_struct *p)
4482 * This function gets called by the timer code, with HZ frequency.
4483 * We call it with interrupts disabled.
4485 * It also gets called by the fork code, when changing the parent's
4488 void scheduler_tick(void)
4490 int cpu = smp_processor_id();
4491 struct rq *rq = cpu_rq(cpu);
4492 struct task_struct *curr = rq->curr;
4496 spin_lock(&rq->lock);
4497 update_rq_clock(rq);
4498 update_cpu_load(rq);
4499 curr->sched_class->task_tick(rq, curr, 0);
4500 spin_unlock(&rq->lock);
4503 rq->idle_at_tick = idle_cpu(cpu);
4504 trigger_load_balance(rq, cpu);
4508 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4509 defined(CONFIG_PREEMPT_TRACER))
4511 static inline unsigned long get_parent_ip(unsigned long addr)
4513 if (in_lock_functions(addr)) {
4514 addr = CALLER_ADDR2;
4515 if (in_lock_functions(addr))
4516 addr = CALLER_ADDR3;
4521 void __kprobes add_preempt_count(int val)
4523 #ifdef CONFIG_DEBUG_PREEMPT
4527 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4530 preempt_count() += val;
4531 #ifdef CONFIG_DEBUG_PREEMPT
4533 * Spinlock count overflowing soon?
4535 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4538 if (preempt_count() == val)
4539 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4541 EXPORT_SYMBOL(add_preempt_count);
4543 void __kprobes sub_preempt_count(int val)
4545 #ifdef CONFIG_DEBUG_PREEMPT
4549 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4552 * Is the spinlock portion underflowing?
4554 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4555 !(preempt_count() & PREEMPT_MASK)))
4559 if (preempt_count() == val)
4560 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4561 preempt_count() -= val;
4563 EXPORT_SYMBOL(sub_preempt_count);
4568 * Print scheduling while atomic bug:
4570 static noinline void __schedule_bug(struct task_struct *prev)
4572 struct pt_regs *regs = get_irq_regs();
4574 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4575 prev->comm, prev->pid, preempt_count());
4577 debug_show_held_locks(prev);
4579 if (irqs_disabled())
4580 print_irqtrace_events(prev);
4589 * Various schedule()-time debugging checks and statistics:
4591 static inline void schedule_debug(struct task_struct *prev)
4594 * Test if we are atomic. Since do_exit() needs to call into
4595 * schedule() atomically, we ignore that path for now.
4596 * Otherwise, whine if we are scheduling when we should not be.
4598 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4599 __schedule_bug(prev);
4601 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4603 schedstat_inc(this_rq(), sched_count);
4604 #ifdef CONFIG_SCHEDSTATS
4605 if (unlikely(prev->lock_depth >= 0)) {
4606 schedstat_inc(this_rq(), bkl_count);
4607 schedstat_inc(prev, sched_info.bkl_count);
4613 * Pick up the highest-prio task:
4615 static inline struct task_struct *
4616 pick_next_task(struct rq *rq, struct task_struct *prev)
4618 const struct sched_class *class;
4619 struct task_struct *p;
4622 * Optimization: we know that if all tasks are in
4623 * the fair class we can call that function directly:
4625 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4626 p = fair_sched_class.pick_next_task(rq);
4631 class = sched_class_highest;
4633 p = class->pick_next_task(rq);
4637 * Will never be NULL as the idle class always
4638 * returns a non-NULL p:
4640 class = class->next;
4645 * schedule() is the main scheduler function.
4647 asmlinkage void __sched schedule(void)
4649 struct task_struct *prev, *next;
4650 unsigned long *switch_count;
4656 cpu = smp_processor_id();
4660 switch_count = &prev->nivcsw;
4662 release_kernel_lock(prev);
4663 need_resched_nonpreemptible:
4665 schedule_debug(prev);
4667 if (sched_feat(HRTICK))
4670 spin_lock_irq(&rq->lock);
4671 update_rq_clock(rq);
4672 clear_tsk_need_resched(prev);
4674 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4675 if (unlikely(signal_pending_state(prev->state, prev)))
4676 prev->state = TASK_RUNNING;
4678 deactivate_task(rq, prev, 1);
4679 switch_count = &prev->nvcsw;
4683 if (prev->sched_class->pre_schedule)
4684 prev->sched_class->pre_schedule(rq, prev);
4687 if (unlikely(!rq->nr_running))
4688 idle_balance(cpu, rq);
4690 prev->sched_class->put_prev_task(rq, prev);
4691 next = pick_next_task(rq, prev);
4693 if (likely(prev != next)) {
4694 sched_info_switch(prev, next);
4700 context_switch(rq, prev, next); /* unlocks the rq */
4702 * the context switch might have flipped the stack from under
4703 * us, hence refresh the local variables.
4705 cpu = smp_processor_id();
4708 spin_unlock_irq(&rq->lock);
4710 if (unlikely(reacquire_kernel_lock(current) < 0))
4711 goto need_resched_nonpreemptible;
4713 preempt_enable_no_resched();
4714 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4717 EXPORT_SYMBOL(schedule);
4719 #ifdef CONFIG_PREEMPT
4721 * this is the entry point to schedule() from in-kernel preemption
4722 * off of preempt_enable. Kernel preemptions off return from interrupt
4723 * occur there and call schedule directly.
4725 asmlinkage void __sched preempt_schedule(void)
4727 struct thread_info *ti = current_thread_info();
4730 * If there is a non-zero preempt_count or interrupts are disabled,
4731 * we do not want to preempt the current task. Just return..
4733 if (likely(ti->preempt_count || irqs_disabled()))
4737 add_preempt_count(PREEMPT_ACTIVE);
4739 sub_preempt_count(PREEMPT_ACTIVE);
4742 * Check again in case we missed a preemption opportunity
4743 * between schedule and now.
4746 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4748 EXPORT_SYMBOL(preempt_schedule);
4751 * this is the entry point to schedule() from kernel preemption
4752 * off of irq context.
4753 * Note, that this is called and return with irqs disabled. This will
4754 * protect us against recursive calling from irq.
4756 asmlinkage void __sched preempt_schedule_irq(void)
4758 struct thread_info *ti = current_thread_info();
4760 /* Catch callers which need to be fixed */
4761 BUG_ON(ti->preempt_count || !irqs_disabled());
4764 add_preempt_count(PREEMPT_ACTIVE);
4767 local_irq_disable();
4768 sub_preempt_count(PREEMPT_ACTIVE);
4771 * Check again in case we missed a preemption opportunity
4772 * between schedule and now.
4775 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4778 #endif /* CONFIG_PREEMPT */
4780 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4783 return try_to_wake_up(curr->private, mode, sync);
4785 EXPORT_SYMBOL(default_wake_function);
4788 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4789 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4790 * number) then we wake all the non-exclusive tasks and one exclusive task.
4792 * There are circumstances in which we can try to wake a task which has already
4793 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4794 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4796 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4797 int nr_exclusive, int sync, void *key)
4799 wait_queue_t *curr, *next;
4801 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4802 unsigned flags = curr->flags;
4804 if (curr->func(curr, mode, sync, key) &&
4805 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4811 * __wake_up - wake up threads blocked on a waitqueue.
4813 * @mode: which threads
4814 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4815 * @key: is directly passed to the wakeup function
4817 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4818 int nr_exclusive, void *key)
4820 unsigned long flags;
4822 spin_lock_irqsave(&q->lock, flags);
4823 __wake_up_common(q, mode, nr_exclusive, 0, key);
4824 spin_unlock_irqrestore(&q->lock, flags);
4826 EXPORT_SYMBOL(__wake_up);
4829 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4831 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4833 __wake_up_common(q, mode, 1, 0, NULL);
4837 * __wake_up_sync - wake up threads blocked on a waitqueue.
4839 * @mode: which threads
4840 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4842 * The sync wakeup differs that the waker knows that it will schedule
4843 * away soon, so while the target thread will be woken up, it will not
4844 * be migrated to another CPU - ie. the two threads are 'synchronized'
4845 * with each other. This can prevent needless bouncing between CPUs.
4847 * On UP it can prevent extra preemption.
4850 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4852 unsigned long flags;
4858 if (unlikely(!nr_exclusive))
4861 spin_lock_irqsave(&q->lock, flags);
4862 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4863 spin_unlock_irqrestore(&q->lock, flags);
4865 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4868 * complete: - signals a single thread waiting on this completion
4869 * @x: holds the state of this particular completion
4871 * This will wake up a single thread waiting on this completion. Threads will be
4872 * awakened in the same order in which they were queued.
4874 * See also complete_all(), wait_for_completion() and related routines.
4876 void complete(struct completion *x)
4878 unsigned long flags;
4880 spin_lock_irqsave(&x->wait.lock, flags);
4882 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4883 spin_unlock_irqrestore(&x->wait.lock, flags);
4885 EXPORT_SYMBOL(complete);
4888 * complete_all: - signals all threads waiting on this completion
4889 * @x: holds the state of this particular completion
4891 * This will wake up all threads waiting on this particular completion event.
4893 void complete_all(struct completion *x)
4895 unsigned long flags;
4897 spin_lock_irqsave(&x->wait.lock, flags);
4898 x->done += UINT_MAX/2;
4899 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4900 spin_unlock_irqrestore(&x->wait.lock, flags);
4902 EXPORT_SYMBOL(complete_all);
4904 static inline long __sched
4905 do_wait_for_common(struct completion *x, long timeout, int state)
4908 DECLARE_WAITQUEUE(wait, current);
4910 wait.flags |= WQ_FLAG_EXCLUSIVE;
4911 __add_wait_queue_tail(&x->wait, &wait);
4913 if (signal_pending_state(state, current)) {
4914 timeout = -ERESTARTSYS;
4917 __set_current_state(state);
4918 spin_unlock_irq(&x->wait.lock);
4919 timeout = schedule_timeout(timeout);
4920 spin_lock_irq(&x->wait.lock);
4921 } while (!x->done && timeout);
4922 __remove_wait_queue(&x->wait, &wait);
4927 return timeout ?: 1;
4931 wait_for_common(struct completion *x, long timeout, int state)
4935 spin_lock_irq(&x->wait.lock);
4936 timeout = do_wait_for_common(x, timeout, state);
4937 spin_unlock_irq(&x->wait.lock);
4942 * wait_for_completion: - waits for completion of a task
4943 * @x: holds the state of this particular completion
4945 * This waits to be signaled for completion of a specific task. It is NOT
4946 * interruptible and there is no timeout.
4948 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4949 * and interrupt capability. Also see complete().
4951 void __sched wait_for_completion(struct completion *x)
4953 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4955 EXPORT_SYMBOL(wait_for_completion);
4958 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4959 * @x: holds the state of this particular completion
4960 * @timeout: timeout value in jiffies
4962 * This waits for either a completion of a specific task to be signaled or for a
4963 * specified timeout to expire. The timeout is in jiffies. It is not
4966 unsigned long __sched
4967 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4969 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4971 EXPORT_SYMBOL(wait_for_completion_timeout);
4974 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4975 * @x: holds the state of this particular completion
4977 * This waits for completion of a specific task to be signaled. It is
4980 int __sched wait_for_completion_interruptible(struct completion *x)
4982 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4983 if (t == -ERESTARTSYS)
4987 EXPORT_SYMBOL(wait_for_completion_interruptible);
4990 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4991 * @x: holds the state of this particular completion
4992 * @timeout: timeout value in jiffies
4994 * This waits for either a completion of a specific task to be signaled or for a
4995 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4997 unsigned long __sched
4998 wait_for_completion_interruptible_timeout(struct completion *x,
4999 unsigned long timeout)
5001 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5003 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5006 * wait_for_completion_killable: - waits for completion of a task (killable)
5007 * @x: holds the state of this particular completion
5009 * This waits to be signaled for completion of a specific task. It can be
5010 * interrupted by a kill signal.
5012 int __sched wait_for_completion_killable(struct completion *x)
5014 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5015 if (t == -ERESTARTSYS)
5019 EXPORT_SYMBOL(wait_for_completion_killable);
5022 * try_wait_for_completion - try to decrement a completion without blocking
5023 * @x: completion structure
5025 * Returns: 0 if a decrement cannot be done without blocking
5026 * 1 if a decrement succeeded.
5028 * If a completion is being used as a counting completion,
5029 * attempt to decrement the counter without blocking. This
5030 * enables us to avoid waiting if the resource the completion
5031 * is protecting is not available.
5033 bool try_wait_for_completion(struct completion *x)
5037 spin_lock_irq(&x->wait.lock);
5042 spin_unlock_irq(&x->wait.lock);
5045 EXPORT_SYMBOL(try_wait_for_completion);
5048 * completion_done - Test to see if a completion has any waiters
5049 * @x: completion structure
5051 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5052 * 1 if there are no waiters.
5055 bool completion_done(struct completion *x)
5059 spin_lock_irq(&x->wait.lock);
5062 spin_unlock_irq(&x->wait.lock);
5065 EXPORT_SYMBOL(completion_done);
5068 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5070 unsigned long flags;
5073 init_waitqueue_entry(&wait, current);
5075 __set_current_state(state);
5077 spin_lock_irqsave(&q->lock, flags);
5078 __add_wait_queue(q, &wait);
5079 spin_unlock(&q->lock);
5080 timeout = schedule_timeout(timeout);
5081 spin_lock_irq(&q->lock);
5082 __remove_wait_queue(q, &wait);
5083 spin_unlock_irqrestore(&q->lock, flags);
5088 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5090 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5092 EXPORT_SYMBOL(interruptible_sleep_on);
5095 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5097 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5099 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5101 void __sched sleep_on(wait_queue_head_t *q)
5103 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5105 EXPORT_SYMBOL(sleep_on);
5107 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5109 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5111 EXPORT_SYMBOL(sleep_on_timeout);
5113 #ifdef CONFIG_RT_MUTEXES
5116 * rt_mutex_setprio - set the current priority of a task
5118 * @prio: prio value (kernel-internal form)
5120 * This function changes the 'effective' priority of a task. It does
5121 * not touch ->normal_prio like __setscheduler().
5123 * Used by the rt_mutex code to implement priority inheritance logic.
5125 void rt_mutex_setprio(struct task_struct *p, int prio)
5127 unsigned long flags;
5128 int oldprio, on_rq, running;
5130 const struct sched_class *prev_class = p->sched_class;
5132 BUG_ON(prio < 0 || prio > MAX_PRIO);
5134 rq = task_rq_lock(p, &flags);
5135 update_rq_clock(rq);
5138 on_rq = p->se.on_rq;
5139 running = task_current(rq, p);
5141 dequeue_task(rq, p, 0);
5143 p->sched_class->put_prev_task(rq, p);
5146 p->sched_class = &rt_sched_class;
5148 p->sched_class = &fair_sched_class;
5153 p->sched_class->set_curr_task(rq);
5155 enqueue_task(rq, p, 0);
5157 check_class_changed(rq, p, prev_class, oldprio, running);
5159 task_rq_unlock(rq, &flags);
5164 void set_user_nice(struct task_struct *p, long nice)
5166 int old_prio, delta, on_rq;
5167 unsigned long flags;
5170 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5173 * We have to be careful, if called from sys_setpriority(),
5174 * the task might be in the middle of scheduling on another CPU.
5176 rq = task_rq_lock(p, &flags);
5177 update_rq_clock(rq);
5179 * The RT priorities are set via sched_setscheduler(), but we still
5180 * allow the 'normal' nice value to be set - but as expected
5181 * it wont have any effect on scheduling until the task is
5182 * SCHED_FIFO/SCHED_RR:
5184 if (task_has_rt_policy(p)) {
5185 p->static_prio = NICE_TO_PRIO(nice);
5188 on_rq = p->se.on_rq;
5190 dequeue_task(rq, p, 0);
5192 p->static_prio = NICE_TO_PRIO(nice);
5195 p->prio = effective_prio(p);
5196 delta = p->prio - old_prio;
5199 enqueue_task(rq, p, 0);
5201 * If the task increased its priority or is running and
5202 * lowered its priority, then reschedule its CPU:
5204 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5205 resched_task(rq->curr);
5208 task_rq_unlock(rq, &flags);
5210 EXPORT_SYMBOL(set_user_nice);
5213 * can_nice - check if a task can reduce its nice value
5217 int can_nice(const struct task_struct *p, const int nice)
5219 /* convert nice value [19,-20] to rlimit style value [1,40] */
5220 int nice_rlim = 20 - nice;
5222 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5223 capable(CAP_SYS_NICE));
5226 #ifdef __ARCH_WANT_SYS_NICE
5229 * sys_nice - change the priority of the current process.
5230 * @increment: priority increment
5232 * sys_setpriority is a more generic, but much slower function that
5233 * does similar things.
5235 SYSCALL_DEFINE1(nice, int, increment)
5240 * Setpriority might change our priority at the same moment.
5241 * We don't have to worry. Conceptually one call occurs first
5242 * and we have a single winner.
5244 if (increment < -40)
5249 nice = PRIO_TO_NICE(current->static_prio) + increment;
5255 if (increment < 0 && !can_nice(current, nice))
5258 retval = security_task_setnice(current, nice);
5262 set_user_nice(current, nice);
5269 * task_prio - return the priority value of a given task.
5270 * @p: the task in question.
5272 * This is the priority value as seen by users in /proc.
5273 * RT tasks are offset by -200. Normal tasks are centered
5274 * around 0, value goes from -16 to +15.
5276 int task_prio(const struct task_struct *p)
5278 return p->prio - MAX_RT_PRIO;
5282 * task_nice - return the nice value of a given task.
5283 * @p: the task in question.
5285 int task_nice(const struct task_struct *p)
5287 return TASK_NICE(p);
5289 EXPORT_SYMBOL(task_nice);
5292 * idle_cpu - is a given cpu idle currently?
5293 * @cpu: the processor in question.
5295 int idle_cpu(int cpu)
5297 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5301 * idle_task - return the idle task for a given cpu.
5302 * @cpu: the processor in question.
5304 struct task_struct *idle_task(int cpu)
5306 return cpu_rq(cpu)->idle;
5310 * find_process_by_pid - find a process with a matching PID value.
5311 * @pid: the pid in question.
5313 static struct task_struct *find_process_by_pid(pid_t pid)
5315 return pid ? find_task_by_vpid(pid) : current;
5318 /* Actually do priority change: must hold rq lock. */
5320 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5322 BUG_ON(p->se.on_rq);
5325 switch (p->policy) {
5329 p->sched_class = &fair_sched_class;
5333 p->sched_class = &rt_sched_class;
5337 p->rt_priority = prio;
5338 p->normal_prio = normal_prio(p);
5339 /* we are holding p->pi_lock already */
5340 p->prio = rt_mutex_getprio(p);
5345 * check the target process has a UID that matches the current process's
5347 static bool check_same_owner(struct task_struct *p)
5349 const struct cred *cred = current_cred(), *pcred;
5353 pcred = __task_cred(p);
5354 match = (cred->euid == pcred->euid ||
5355 cred->euid == pcred->uid);
5360 static int __sched_setscheduler(struct task_struct *p, int policy,
5361 struct sched_param *param, bool user)
5363 int retval, oldprio, oldpolicy = -1, on_rq, running;
5364 unsigned long flags;
5365 const struct sched_class *prev_class = p->sched_class;
5368 /* may grab non-irq protected spin_locks */
5369 BUG_ON(in_interrupt());
5371 /* double check policy once rq lock held */
5373 policy = oldpolicy = p->policy;
5374 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5375 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5376 policy != SCHED_IDLE)
5379 * Valid priorities for SCHED_FIFO and SCHED_RR are
5380 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5381 * SCHED_BATCH and SCHED_IDLE is 0.
5383 if (param->sched_priority < 0 ||
5384 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5385 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5387 if (rt_policy(policy) != (param->sched_priority != 0))
5391 * Allow unprivileged RT tasks to decrease priority:
5393 if (user && !capable(CAP_SYS_NICE)) {
5394 if (rt_policy(policy)) {
5395 unsigned long rlim_rtprio;
5397 if (!lock_task_sighand(p, &flags))
5399 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5400 unlock_task_sighand(p, &flags);
5402 /* can't set/change the rt policy */
5403 if (policy != p->policy && !rlim_rtprio)
5406 /* can't increase priority */
5407 if (param->sched_priority > p->rt_priority &&
5408 param->sched_priority > rlim_rtprio)
5412 * Like positive nice levels, dont allow tasks to
5413 * move out of SCHED_IDLE either:
5415 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5418 /* can't change other user's priorities */
5419 if (!check_same_owner(p))
5424 #ifdef CONFIG_RT_GROUP_SCHED
5426 * Do not allow realtime tasks into groups that have no runtime
5429 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5430 task_group(p)->rt_bandwidth.rt_runtime == 0)
5434 retval = security_task_setscheduler(p, policy, param);
5440 * make sure no PI-waiters arrive (or leave) while we are
5441 * changing the priority of the task:
5443 spin_lock_irqsave(&p->pi_lock, flags);
5445 * To be able to change p->policy safely, the apropriate
5446 * runqueue lock must be held.
5448 rq = __task_rq_lock(p);
5449 /* recheck policy now with rq lock held */
5450 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5451 policy = oldpolicy = -1;
5452 __task_rq_unlock(rq);
5453 spin_unlock_irqrestore(&p->pi_lock, flags);
5456 update_rq_clock(rq);
5457 on_rq = p->se.on_rq;
5458 running = task_current(rq, p);
5460 deactivate_task(rq, p, 0);
5462 p->sched_class->put_prev_task(rq, p);
5465 __setscheduler(rq, p, policy, param->sched_priority);
5468 p->sched_class->set_curr_task(rq);
5470 activate_task(rq, p, 0);
5472 check_class_changed(rq, p, prev_class, oldprio, running);
5474 __task_rq_unlock(rq);
5475 spin_unlock_irqrestore(&p->pi_lock, flags);
5477 rt_mutex_adjust_pi(p);
5483 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5484 * @p: the task in question.
5485 * @policy: new policy.
5486 * @param: structure containing the new RT priority.
5488 * NOTE that the task may be already dead.
5490 int sched_setscheduler(struct task_struct *p, int policy,
5491 struct sched_param *param)
5493 return __sched_setscheduler(p, policy, param, true);
5495 EXPORT_SYMBOL_GPL(sched_setscheduler);
5498 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5499 * @p: the task in question.
5500 * @policy: new policy.
5501 * @param: structure containing the new RT priority.
5503 * Just like sched_setscheduler, only don't bother checking if the
5504 * current context has permission. For example, this is needed in
5505 * stop_machine(): we create temporary high priority worker threads,
5506 * but our caller might not have that capability.
5508 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5509 struct sched_param *param)
5511 return __sched_setscheduler(p, policy, param, false);
5515 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5517 struct sched_param lparam;
5518 struct task_struct *p;
5521 if (!param || pid < 0)
5523 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5528 p = find_process_by_pid(pid);
5530 retval = sched_setscheduler(p, policy, &lparam);
5537 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5538 * @pid: the pid in question.
5539 * @policy: new policy.
5540 * @param: structure containing the new RT priority.
5542 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5543 struct sched_param __user *, param)
5545 /* negative values for policy are not valid */
5549 return do_sched_setscheduler(pid, policy, param);
5553 * sys_sched_setparam - set/change the RT priority of a thread
5554 * @pid: the pid in question.
5555 * @param: structure containing the new RT priority.
5557 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5559 return do_sched_setscheduler(pid, -1, param);
5563 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5564 * @pid: the pid in question.
5566 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5568 struct task_struct *p;
5575 read_lock(&tasklist_lock);
5576 p = find_process_by_pid(pid);
5578 retval = security_task_getscheduler(p);
5582 read_unlock(&tasklist_lock);
5587 * sys_sched_getscheduler - get the RT priority of a thread
5588 * @pid: the pid in question.
5589 * @param: structure containing the RT priority.
5591 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5593 struct sched_param lp;
5594 struct task_struct *p;
5597 if (!param || pid < 0)
5600 read_lock(&tasklist_lock);
5601 p = find_process_by_pid(pid);
5606 retval = security_task_getscheduler(p);
5610 lp.sched_priority = p->rt_priority;
5611 read_unlock(&tasklist_lock);
5614 * This one might sleep, we cannot do it with a spinlock held ...
5616 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5621 read_unlock(&tasklist_lock);
5625 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5627 cpumask_var_t cpus_allowed, new_mask;
5628 struct task_struct *p;
5632 read_lock(&tasklist_lock);
5634 p = find_process_by_pid(pid);
5636 read_unlock(&tasklist_lock);
5642 * It is not safe to call set_cpus_allowed with the
5643 * tasklist_lock held. We will bump the task_struct's
5644 * usage count and then drop tasklist_lock.
5647 read_unlock(&tasklist_lock);
5649 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5653 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5655 goto out_free_cpus_allowed;
5658 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5661 retval = security_task_setscheduler(p, 0, NULL);
5665 cpuset_cpus_allowed(p, cpus_allowed);
5666 cpumask_and(new_mask, in_mask, cpus_allowed);
5668 retval = set_cpus_allowed_ptr(p, new_mask);
5671 cpuset_cpus_allowed(p, cpus_allowed);
5672 if (!cpumask_subset(new_mask, cpus_allowed)) {
5674 * We must have raced with a concurrent cpuset
5675 * update. Just reset the cpus_allowed to the
5676 * cpuset's cpus_allowed
5678 cpumask_copy(new_mask, cpus_allowed);
5683 free_cpumask_var(new_mask);
5684 out_free_cpus_allowed:
5685 free_cpumask_var(cpus_allowed);
5692 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5693 struct cpumask *new_mask)
5695 if (len < cpumask_size())
5696 cpumask_clear(new_mask);
5697 else if (len > cpumask_size())
5698 len = cpumask_size();
5700 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5704 * sys_sched_setaffinity - set the cpu affinity of a process
5705 * @pid: pid of the process
5706 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5707 * @user_mask_ptr: user-space pointer to the new cpu mask
5709 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5710 unsigned long __user *, user_mask_ptr)
5712 cpumask_var_t new_mask;
5715 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5718 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5720 retval = sched_setaffinity(pid, new_mask);
5721 free_cpumask_var(new_mask);
5725 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5727 struct task_struct *p;
5731 read_lock(&tasklist_lock);
5734 p = find_process_by_pid(pid);
5738 retval = security_task_getscheduler(p);
5742 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5745 read_unlock(&tasklist_lock);
5752 * sys_sched_getaffinity - get the cpu affinity of a process
5753 * @pid: pid of the process
5754 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5755 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5757 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5758 unsigned long __user *, user_mask_ptr)
5763 if (len < cpumask_size())
5766 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5769 ret = sched_getaffinity(pid, mask);
5771 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5774 ret = cpumask_size();
5776 free_cpumask_var(mask);
5782 * sys_sched_yield - yield the current processor to other threads.
5784 * This function yields the current CPU to other tasks. If there are no
5785 * other threads running on this CPU then this function will return.
5787 SYSCALL_DEFINE0(sched_yield)
5789 struct rq *rq = this_rq_lock();
5791 schedstat_inc(rq, yld_count);
5792 current->sched_class->yield_task(rq);
5795 * Since we are going to call schedule() anyway, there's
5796 * no need to preempt or enable interrupts:
5798 __release(rq->lock);
5799 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5800 _raw_spin_unlock(&rq->lock);
5801 preempt_enable_no_resched();
5808 static void __cond_resched(void)
5810 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5811 __might_sleep(__FILE__, __LINE__);
5814 * The BKS might be reacquired before we have dropped
5815 * PREEMPT_ACTIVE, which could trigger a second
5816 * cond_resched() call.
5819 add_preempt_count(PREEMPT_ACTIVE);
5821 sub_preempt_count(PREEMPT_ACTIVE);
5822 } while (need_resched());
5825 int __sched _cond_resched(void)
5827 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5828 system_state == SYSTEM_RUNNING) {
5834 EXPORT_SYMBOL(_cond_resched);
5837 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5838 * call schedule, and on return reacquire the lock.
5840 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5841 * operations here to prevent schedule() from being called twice (once via
5842 * spin_unlock(), once by hand).
5844 int cond_resched_lock(spinlock_t *lock)
5846 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5849 if (spin_needbreak(lock) || resched) {
5851 if (resched && need_resched())
5860 EXPORT_SYMBOL(cond_resched_lock);
5862 int __sched cond_resched_softirq(void)
5864 BUG_ON(!in_softirq());
5866 if (need_resched() && system_state == SYSTEM_RUNNING) {
5874 EXPORT_SYMBOL(cond_resched_softirq);
5877 * yield - yield the current processor to other threads.
5879 * This is a shortcut for kernel-space yielding - it marks the
5880 * thread runnable and calls sys_sched_yield().
5882 void __sched yield(void)
5884 set_current_state(TASK_RUNNING);
5887 EXPORT_SYMBOL(yield);
5890 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5891 * that process accounting knows that this is a task in IO wait state.
5893 * But don't do that if it is a deliberate, throttling IO wait (this task
5894 * has set its backing_dev_info: the queue against which it should throttle)
5896 void __sched io_schedule(void)
5898 struct rq *rq = &__raw_get_cpu_var(runqueues);
5900 delayacct_blkio_start();
5901 atomic_inc(&rq->nr_iowait);
5903 atomic_dec(&rq->nr_iowait);
5904 delayacct_blkio_end();
5906 EXPORT_SYMBOL(io_schedule);
5908 long __sched io_schedule_timeout(long timeout)
5910 struct rq *rq = &__raw_get_cpu_var(runqueues);
5913 delayacct_blkio_start();
5914 atomic_inc(&rq->nr_iowait);
5915 ret = schedule_timeout(timeout);
5916 atomic_dec(&rq->nr_iowait);
5917 delayacct_blkio_end();
5922 * sys_sched_get_priority_max - return maximum RT priority.
5923 * @policy: scheduling class.
5925 * this syscall returns the maximum rt_priority that can be used
5926 * by a given scheduling class.
5928 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5935 ret = MAX_USER_RT_PRIO-1;
5947 * sys_sched_get_priority_min - return minimum RT priority.
5948 * @policy: scheduling class.
5950 * this syscall returns the minimum rt_priority that can be used
5951 * by a given scheduling class.
5953 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5971 * sys_sched_rr_get_interval - return the default timeslice of a process.
5972 * @pid: pid of the process.
5973 * @interval: userspace pointer to the timeslice value.
5975 * this syscall writes the default timeslice value of a given process
5976 * into the user-space timespec buffer. A value of '0' means infinity.
5978 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5979 struct timespec __user *, interval)
5981 struct task_struct *p;
5982 unsigned int time_slice;
5990 read_lock(&tasklist_lock);
5991 p = find_process_by_pid(pid);
5995 retval = security_task_getscheduler(p);
6000 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6001 * tasks that are on an otherwise idle runqueue:
6004 if (p->policy == SCHED_RR) {
6005 time_slice = DEF_TIMESLICE;
6006 } else if (p->policy != SCHED_FIFO) {
6007 struct sched_entity *se = &p->se;
6008 unsigned long flags;
6011 rq = task_rq_lock(p, &flags);
6012 if (rq->cfs.load.weight)
6013 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6014 task_rq_unlock(rq, &flags);
6016 read_unlock(&tasklist_lock);
6017 jiffies_to_timespec(time_slice, &t);
6018 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6022 read_unlock(&tasklist_lock);
6026 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6028 void sched_show_task(struct task_struct *p)
6030 unsigned long free = 0;
6033 state = p->state ? __ffs(p->state) + 1 : 0;
6034 printk(KERN_INFO "%-13.13s %c", p->comm,
6035 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6036 #if BITS_PER_LONG == 32
6037 if (state == TASK_RUNNING)
6038 printk(KERN_CONT " running ");
6040 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6042 if (state == TASK_RUNNING)
6043 printk(KERN_CONT " running task ");
6045 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6047 #ifdef CONFIG_DEBUG_STACK_USAGE
6049 unsigned long *n = end_of_stack(p);
6052 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6055 printk(KERN_CONT "%5lu %5d %6d\n", free,
6056 task_pid_nr(p), task_pid_nr(p->real_parent));
6058 show_stack(p, NULL);
6061 void show_state_filter(unsigned long state_filter)
6063 struct task_struct *g, *p;
6065 #if BITS_PER_LONG == 32
6067 " task PC stack pid father\n");
6070 " task PC stack pid father\n");
6072 read_lock(&tasklist_lock);
6073 do_each_thread(g, p) {
6075 * reset the NMI-timeout, listing all files on a slow
6076 * console might take alot of time:
6078 touch_nmi_watchdog();
6079 if (!state_filter || (p->state & state_filter))
6081 } while_each_thread(g, p);
6083 touch_all_softlockup_watchdogs();
6085 #ifdef CONFIG_SCHED_DEBUG
6086 sysrq_sched_debug_show();
6088 read_unlock(&tasklist_lock);
6090 * Only show locks if all tasks are dumped:
6092 if (state_filter == -1)
6093 debug_show_all_locks();
6096 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6098 idle->sched_class = &idle_sched_class;
6102 * init_idle - set up an idle thread for a given CPU
6103 * @idle: task in question
6104 * @cpu: cpu the idle task belongs to
6106 * NOTE: this function does not set the idle thread's NEED_RESCHED
6107 * flag, to make booting more robust.
6109 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6111 struct rq *rq = cpu_rq(cpu);
6112 unsigned long flags;
6114 spin_lock_irqsave(&rq->lock, flags);
6117 idle->se.exec_start = sched_clock();
6119 idle->prio = idle->normal_prio = MAX_PRIO;
6120 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6121 __set_task_cpu(idle, cpu);
6123 rq->curr = rq->idle = idle;
6124 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6127 spin_unlock_irqrestore(&rq->lock, flags);
6129 /* Set the preempt count _outside_ the spinlocks! */
6130 #if defined(CONFIG_PREEMPT)
6131 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6133 task_thread_info(idle)->preempt_count = 0;
6136 * The idle tasks have their own, simple scheduling class:
6138 idle->sched_class = &idle_sched_class;
6139 ftrace_graph_init_task(idle);
6143 * In a system that switches off the HZ timer nohz_cpu_mask
6144 * indicates which cpus entered this state. This is used
6145 * in the rcu update to wait only for active cpus. For system
6146 * which do not switch off the HZ timer nohz_cpu_mask should
6147 * always be CPU_BITS_NONE.
6149 cpumask_var_t nohz_cpu_mask;
6152 * Increase the granularity value when there are more CPUs,
6153 * because with more CPUs the 'effective latency' as visible
6154 * to users decreases. But the relationship is not linear,
6155 * so pick a second-best guess by going with the log2 of the
6158 * This idea comes from the SD scheduler of Con Kolivas:
6160 static inline void sched_init_granularity(void)
6162 unsigned int factor = 1 + ilog2(num_online_cpus());
6163 const unsigned long limit = 200000000;
6165 sysctl_sched_min_granularity *= factor;
6166 if (sysctl_sched_min_granularity > limit)
6167 sysctl_sched_min_granularity = limit;
6169 sysctl_sched_latency *= factor;
6170 if (sysctl_sched_latency > limit)
6171 sysctl_sched_latency = limit;
6173 sysctl_sched_wakeup_granularity *= factor;
6175 sysctl_sched_shares_ratelimit *= factor;
6180 * This is how migration works:
6182 * 1) we queue a struct migration_req structure in the source CPU's
6183 * runqueue and wake up that CPU's migration thread.
6184 * 2) we down() the locked semaphore => thread blocks.
6185 * 3) migration thread wakes up (implicitly it forces the migrated
6186 * thread off the CPU)
6187 * 4) it gets the migration request and checks whether the migrated
6188 * task is still in the wrong runqueue.
6189 * 5) if it's in the wrong runqueue then the migration thread removes
6190 * it and puts it into the right queue.
6191 * 6) migration thread up()s the semaphore.
6192 * 7) we wake up and the migration is done.
6196 * Change a given task's CPU affinity. Migrate the thread to a
6197 * proper CPU and schedule it away if the CPU it's executing on
6198 * is removed from the allowed bitmask.
6200 * NOTE: the caller must have a valid reference to the task, the
6201 * task must not exit() & deallocate itself prematurely. The
6202 * call is not atomic; no spinlocks may be held.
6204 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6206 struct migration_req req;
6207 unsigned long flags;
6211 rq = task_rq_lock(p, &flags);
6212 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6217 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6218 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6223 if (p->sched_class->set_cpus_allowed)
6224 p->sched_class->set_cpus_allowed(p, new_mask);
6226 cpumask_copy(&p->cpus_allowed, new_mask);
6227 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6230 /* Can the task run on the task's current CPU? If so, we're done */
6231 if (cpumask_test_cpu(task_cpu(p), new_mask))
6234 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6235 /* Need help from migration thread: drop lock and wait. */
6236 task_rq_unlock(rq, &flags);
6237 wake_up_process(rq->migration_thread);
6238 wait_for_completion(&req.done);
6239 tlb_migrate_finish(p->mm);
6243 task_rq_unlock(rq, &flags);
6247 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6250 * Move (not current) task off this cpu, onto dest cpu. We're doing
6251 * this because either it can't run here any more (set_cpus_allowed()
6252 * away from this CPU, or CPU going down), or because we're
6253 * attempting to rebalance this task on exec (sched_exec).
6255 * So we race with normal scheduler movements, but that's OK, as long
6256 * as the task is no longer on this CPU.
6258 * Returns non-zero if task was successfully migrated.
6260 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6262 struct rq *rq_dest, *rq_src;
6265 if (unlikely(!cpu_active(dest_cpu)))
6268 rq_src = cpu_rq(src_cpu);
6269 rq_dest = cpu_rq(dest_cpu);
6271 double_rq_lock(rq_src, rq_dest);
6272 /* Already moved. */
6273 if (task_cpu(p) != src_cpu)
6275 /* Affinity changed (again). */
6276 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6279 on_rq = p->se.on_rq;
6281 deactivate_task(rq_src, p, 0);
6283 set_task_cpu(p, dest_cpu);
6285 activate_task(rq_dest, p, 0);
6286 check_preempt_curr(rq_dest, p, 0);
6291 double_rq_unlock(rq_src, rq_dest);
6296 * migration_thread - this is a highprio system thread that performs
6297 * thread migration by bumping thread off CPU then 'pushing' onto
6300 static int migration_thread(void *data)
6302 int cpu = (long)data;
6306 BUG_ON(rq->migration_thread != current);
6308 set_current_state(TASK_INTERRUPTIBLE);
6309 while (!kthread_should_stop()) {
6310 struct migration_req *req;
6311 struct list_head *head;
6313 spin_lock_irq(&rq->lock);
6315 if (cpu_is_offline(cpu)) {
6316 spin_unlock_irq(&rq->lock);
6320 if (rq->active_balance) {
6321 active_load_balance(rq, cpu);
6322 rq->active_balance = 0;
6325 head = &rq->migration_queue;
6327 if (list_empty(head)) {
6328 spin_unlock_irq(&rq->lock);
6330 set_current_state(TASK_INTERRUPTIBLE);
6333 req = list_entry(head->next, struct migration_req, list);
6334 list_del_init(head->next);
6336 spin_unlock(&rq->lock);
6337 __migrate_task(req->task, cpu, req->dest_cpu);
6340 complete(&req->done);
6342 __set_current_state(TASK_RUNNING);
6346 /* Wait for kthread_stop */
6347 set_current_state(TASK_INTERRUPTIBLE);
6348 while (!kthread_should_stop()) {
6350 set_current_state(TASK_INTERRUPTIBLE);
6352 __set_current_state(TASK_RUNNING);
6356 #ifdef CONFIG_HOTPLUG_CPU
6358 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6362 local_irq_disable();
6363 ret = __migrate_task(p, src_cpu, dest_cpu);
6369 * Figure out where task on dead CPU should go, use force if necessary.
6371 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6374 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6377 /* Look for allowed, online CPU in same node. */
6378 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6379 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6382 /* Any allowed, online CPU? */
6383 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6384 if (dest_cpu < nr_cpu_ids)
6387 /* No more Mr. Nice Guy. */
6388 if (dest_cpu >= nr_cpu_ids) {
6389 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6390 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6393 * Don't tell them about moving exiting tasks or
6394 * kernel threads (both mm NULL), since they never
6397 if (p->mm && printk_ratelimit()) {
6398 printk(KERN_INFO "process %d (%s) no "
6399 "longer affine to cpu%d\n",
6400 task_pid_nr(p), p->comm, dead_cpu);
6405 /* It can have affinity changed while we were choosing. */
6406 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6411 * While a dead CPU has no uninterruptible tasks queued at this point,
6412 * it might still have a nonzero ->nr_uninterruptible counter, because
6413 * for performance reasons the counter is not stricly tracking tasks to
6414 * their home CPUs. So we just add the counter to another CPU's counter,
6415 * to keep the global sum constant after CPU-down:
6417 static void migrate_nr_uninterruptible(struct rq *rq_src)
6419 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6420 unsigned long flags;
6422 local_irq_save(flags);
6423 double_rq_lock(rq_src, rq_dest);
6424 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6425 rq_src->nr_uninterruptible = 0;
6426 double_rq_unlock(rq_src, rq_dest);
6427 local_irq_restore(flags);
6430 /* Run through task list and migrate tasks from the dead cpu. */
6431 static void migrate_live_tasks(int src_cpu)
6433 struct task_struct *p, *t;
6435 read_lock(&tasklist_lock);
6437 do_each_thread(t, p) {
6441 if (task_cpu(p) == src_cpu)
6442 move_task_off_dead_cpu(src_cpu, p);
6443 } while_each_thread(t, p);
6445 read_unlock(&tasklist_lock);
6449 * Schedules idle task to be the next runnable task on current CPU.
6450 * It does so by boosting its priority to highest possible.
6451 * Used by CPU offline code.
6453 void sched_idle_next(void)
6455 int this_cpu = smp_processor_id();
6456 struct rq *rq = cpu_rq(this_cpu);
6457 struct task_struct *p = rq->idle;
6458 unsigned long flags;
6460 /* cpu has to be offline */
6461 BUG_ON(cpu_online(this_cpu));
6464 * Strictly not necessary since rest of the CPUs are stopped by now
6465 * and interrupts disabled on the current cpu.
6467 spin_lock_irqsave(&rq->lock, flags);
6469 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6471 update_rq_clock(rq);
6472 activate_task(rq, p, 0);
6474 spin_unlock_irqrestore(&rq->lock, flags);
6478 * Ensures that the idle task is using init_mm right before its cpu goes
6481 void idle_task_exit(void)
6483 struct mm_struct *mm = current->active_mm;
6485 BUG_ON(cpu_online(smp_processor_id()));
6488 switch_mm(mm, &init_mm, current);
6492 /* called under rq->lock with disabled interrupts */
6493 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6495 struct rq *rq = cpu_rq(dead_cpu);
6497 /* Must be exiting, otherwise would be on tasklist. */
6498 BUG_ON(!p->exit_state);
6500 /* Cannot have done final schedule yet: would have vanished. */
6501 BUG_ON(p->state == TASK_DEAD);
6506 * Drop lock around migration; if someone else moves it,
6507 * that's OK. No task can be added to this CPU, so iteration is
6510 spin_unlock_irq(&rq->lock);
6511 move_task_off_dead_cpu(dead_cpu, p);
6512 spin_lock_irq(&rq->lock);
6517 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6518 static void migrate_dead_tasks(unsigned int dead_cpu)
6520 struct rq *rq = cpu_rq(dead_cpu);
6521 struct task_struct *next;
6524 if (!rq->nr_running)
6526 update_rq_clock(rq);
6527 next = pick_next_task(rq, rq->curr);
6530 next->sched_class->put_prev_task(rq, next);
6531 migrate_dead(dead_cpu, next);
6535 #endif /* CONFIG_HOTPLUG_CPU */
6537 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6539 static struct ctl_table sd_ctl_dir[] = {
6541 .procname = "sched_domain",
6547 static struct ctl_table sd_ctl_root[] = {
6549 .ctl_name = CTL_KERN,
6550 .procname = "kernel",
6552 .child = sd_ctl_dir,
6557 static struct ctl_table *sd_alloc_ctl_entry(int n)
6559 struct ctl_table *entry =
6560 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6565 static void sd_free_ctl_entry(struct ctl_table **tablep)
6567 struct ctl_table *entry;
6570 * In the intermediate directories, both the child directory and
6571 * procname are dynamically allocated and could fail but the mode
6572 * will always be set. In the lowest directory the names are
6573 * static strings and all have proc handlers.
6575 for (entry = *tablep; entry->mode; entry++) {
6577 sd_free_ctl_entry(&entry->child);
6578 if (entry->proc_handler == NULL)
6579 kfree(entry->procname);
6587 set_table_entry(struct ctl_table *entry,
6588 const char *procname, void *data, int maxlen,
6589 mode_t mode, proc_handler *proc_handler)
6591 entry->procname = procname;
6593 entry->maxlen = maxlen;
6595 entry->proc_handler = proc_handler;
6598 static struct ctl_table *
6599 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6601 struct ctl_table *table = sd_alloc_ctl_entry(13);
6606 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6607 sizeof(long), 0644, proc_doulongvec_minmax);
6608 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6609 sizeof(long), 0644, proc_doulongvec_minmax);
6610 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6611 sizeof(int), 0644, proc_dointvec_minmax);
6612 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6613 sizeof(int), 0644, proc_dointvec_minmax);
6614 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6615 sizeof(int), 0644, proc_dointvec_minmax);
6616 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6617 sizeof(int), 0644, proc_dointvec_minmax);
6618 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6619 sizeof(int), 0644, proc_dointvec_minmax);
6620 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6621 sizeof(int), 0644, proc_dointvec_minmax);
6622 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6623 sizeof(int), 0644, proc_dointvec_minmax);
6624 set_table_entry(&table[9], "cache_nice_tries",
6625 &sd->cache_nice_tries,
6626 sizeof(int), 0644, proc_dointvec_minmax);
6627 set_table_entry(&table[10], "flags", &sd->flags,
6628 sizeof(int), 0644, proc_dointvec_minmax);
6629 set_table_entry(&table[11], "name", sd->name,
6630 CORENAME_MAX_SIZE, 0444, proc_dostring);
6631 /* &table[12] is terminator */
6636 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6638 struct ctl_table *entry, *table;
6639 struct sched_domain *sd;
6640 int domain_num = 0, i;
6643 for_each_domain(cpu, sd)
6645 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6650 for_each_domain(cpu, sd) {
6651 snprintf(buf, 32, "domain%d", i);
6652 entry->procname = kstrdup(buf, GFP_KERNEL);
6654 entry->child = sd_alloc_ctl_domain_table(sd);
6661 static struct ctl_table_header *sd_sysctl_header;
6662 static void register_sched_domain_sysctl(void)
6664 int i, cpu_num = num_online_cpus();
6665 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6668 WARN_ON(sd_ctl_dir[0].child);
6669 sd_ctl_dir[0].child = entry;
6674 for_each_online_cpu(i) {
6675 snprintf(buf, 32, "cpu%d", i);
6676 entry->procname = kstrdup(buf, GFP_KERNEL);
6678 entry->child = sd_alloc_ctl_cpu_table(i);
6682 WARN_ON(sd_sysctl_header);
6683 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6686 /* may be called multiple times per register */
6687 static void unregister_sched_domain_sysctl(void)
6689 if (sd_sysctl_header)
6690 unregister_sysctl_table(sd_sysctl_header);
6691 sd_sysctl_header = NULL;
6692 if (sd_ctl_dir[0].child)
6693 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6696 static void register_sched_domain_sysctl(void)
6699 static void unregister_sched_domain_sysctl(void)
6704 static void set_rq_online(struct rq *rq)
6707 const struct sched_class *class;
6709 cpumask_set_cpu(rq->cpu, rq->rd->online);
6712 for_each_class(class) {
6713 if (class->rq_online)
6714 class->rq_online(rq);
6719 static void set_rq_offline(struct rq *rq)
6722 const struct sched_class *class;
6724 for_each_class(class) {
6725 if (class->rq_offline)
6726 class->rq_offline(rq);
6729 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6735 * migration_call - callback that gets triggered when a CPU is added.
6736 * Here we can start up the necessary migration thread for the new CPU.
6738 static int __cpuinit
6739 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6741 struct task_struct *p;
6742 int cpu = (long)hcpu;
6743 unsigned long flags;
6748 case CPU_UP_PREPARE:
6749 case CPU_UP_PREPARE_FROZEN:
6750 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6753 kthread_bind(p, cpu);
6754 /* Must be high prio: stop_machine expects to yield to it. */
6755 rq = task_rq_lock(p, &flags);
6756 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6757 task_rq_unlock(rq, &flags);
6758 cpu_rq(cpu)->migration_thread = p;
6762 case CPU_ONLINE_FROZEN:
6763 /* Strictly unnecessary, as first user will wake it. */
6764 wake_up_process(cpu_rq(cpu)->migration_thread);
6766 /* Update our root-domain */
6768 spin_lock_irqsave(&rq->lock, flags);
6770 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6774 spin_unlock_irqrestore(&rq->lock, flags);
6777 #ifdef CONFIG_HOTPLUG_CPU
6778 case CPU_UP_CANCELED:
6779 case CPU_UP_CANCELED_FROZEN:
6780 if (!cpu_rq(cpu)->migration_thread)
6782 /* Unbind it from offline cpu so it can run. Fall thru. */
6783 kthread_bind(cpu_rq(cpu)->migration_thread,
6784 cpumask_any(cpu_online_mask));
6785 kthread_stop(cpu_rq(cpu)->migration_thread);
6786 cpu_rq(cpu)->migration_thread = NULL;
6790 case CPU_DEAD_FROZEN:
6791 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6792 migrate_live_tasks(cpu);
6794 kthread_stop(rq->migration_thread);
6795 rq->migration_thread = NULL;
6796 /* Idle task back to normal (off runqueue, low prio) */
6797 spin_lock_irq(&rq->lock);
6798 update_rq_clock(rq);
6799 deactivate_task(rq, rq->idle, 0);
6800 rq->idle->static_prio = MAX_PRIO;
6801 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6802 rq->idle->sched_class = &idle_sched_class;
6803 migrate_dead_tasks(cpu);
6804 spin_unlock_irq(&rq->lock);
6806 migrate_nr_uninterruptible(rq);
6807 BUG_ON(rq->nr_running != 0);
6810 * No need to migrate the tasks: it was best-effort if
6811 * they didn't take sched_hotcpu_mutex. Just wake up
6814 spin_lock_irq(&rq->lock);
6815 while (!list_empty(&rq->migration_queue)) {
6816 struct migration_req *req;
6818 req = list_entry(rq->migration_queue.next,
6819 struct migration_req, list);
6820 list_del_init(&req->list);
6821 spin_unlock_irq(&rq->lock);
6822 complete(&req->done);
6823 spin_lock_irq(&rq->lock);
6825 spin_unlock_irq(&rq->lock);
6829 case CPU_DYING_FROZEN:
6830 /* Update our root-domain */
6832 spin_lock_irqsave(&rq->lock, flags);
6834 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6837 spin_unlock_irqrestore(&rq->lock, flags);
6844 /* Register at highest priority so that task migration (migrate_all_tasks)
6845 * happens before everything else.
6847 static struct notifier_block __cpuinitdata migration_notifier = {
6848 .notifier_call = migration_call,
6852 static int __init migration_init(void)
6854 void *cpu = (void *)(long)smp_processor_id();
6857 /* Start one for the boot CPU: */
6858 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6859 BUG_ON(err == NOTIFY_BAD);
6860 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6861 register_cpu_notifier(&migration_notifier);
6865 early_initcall(migration_init);
6870 #ifdef CONFIG_SCHED_DEBUG
6872 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6873 struct cpumask *groupmask)
6875 struct sched_group *group = sd->groups;
6878 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6879 cpumask_clear(groupmask);
6881 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6883 if (!(sd->flags & SD_LOAD_BALANCE)) {
6884 printk("does not load-balance\n");
6886 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6891 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6893 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6894 printk(KERN_ERR "ERROR: domain->span does not contain "
6897 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6898 printk(KERN_ERR "ERROR: domain->groups does not contain"
6902 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6906 printk(KERN_ERR "ERROR: group is NULL\n");
6910 if (!group->__cpu_power) {
6911 printk(KERN_CONT "\n");
6912 printk(KERN_ERR "ERROR: domain->cpu_power not "
6917 if (!cpumask_weight(sched_group_cpus(group))) {
6918 printk(KERN_CONT "\n");
6919 printk(KERN_ERR "ERROR: empty group\n");
6923 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6924 printk(KERN_CONT "\n");
6925 printk(KERN_ERR "ERROR: repeated CPUs\n");
6929 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6931 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6932 printk(KERN_CONT " %s", str);
6934 group = group->next;
6935 } while (group != sd->groups);
6936 printk(KERN_CONT "\n");
6938 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6939 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6942 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6943 printk(KERN_ERR "ERROR: parent span is not a superset "
6944 "of domain->span\n");
6948 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6950 cpumask_var_t groupmask;
6954 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6958 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6960 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6961 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6966 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6973 free_cpumask_var(groupmask);
6975 #else /* !CONFIG_SCHED_DEBUG */
6976 # define sched_domain_debug(sd, cpu) do { } while (0)
6977 #endif /* CONFIG_SCHED_DEBUG */
6979 static int sd_degenerate(struct sched_domain *sd)
6981 if (cpumask_weight(sched_domain_span(sd)) == 1)
6984 /* Following flags need at least 2 groups */
6985 if (sd->flags & (SD_LOAD_BALANCE |
6986 SD_BALANCE_NEWIDLE |
6990 SD_SHARE_PKG_RESOURCES)) {
6991 if (sd->groups != sd->groups->next)
6995 /* Following flags don't use groups */
6996 if (sd->flags & (SD_WAKE_IDLE |
7005 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7007 unsigned long cflags = sd->flags, pflags = parent->flags;
7009 if (sd_degenerate(parent))
7012 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7015 /* Does parent contain flags not in child? */
7016 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7017 if (cflags & SD_WAKE_AFFINE)
7018 pflags &= ~SD_WAKE_BALANCE;
7019 /* Flags needing groups don't count if only 1 group in parent */
7020 if (parent->groups == parent->groups->next) {
7021 pflags &= ~(SD_LOAD_BALANCE |
7022 SD_BALANCE_NEWIDLE |
7026 SD_SHARE_PKG_RESOURCES);
7027 if (nr_node_ids == 1)
7028 pflags &= ~SD_SERIALIZE;
7030 if (~cflags & pflags)
7036 static void free_rootdomain(struct root_domain *rd)
7038 cpupri_cleanup(&rd->cpupri);
7040 free_cpumask_var(rd->rto_mask);
7041 free_cpumask_var(rd->online);
7042 free_cpumask_var(rd->span);
7046 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7048 unsigned long flags;
7050 spin_lock_irqsave(&rq->lock, flags);
7053 struct root_domain *old_rd = rq->rd;
7055 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7058 cpumask_clear_cpu(rq->cpu, old_rd->span);
7060 if (atomic_dec_and_test(&old_rd->refcount))
7061 free_rootdomain(old_rd);
7064 atomic_inc(&rd->refcount);
7067 cpumask_set_cpu(rq->cpu, rd->span);
7068 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7071 spin_unlock_irqrestore(&rq->lock, flags);
7074 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7076 memset(rd, 0, sizeof(*rd));
7079 alloc_bootmem_cpumask_var(&def_root_domain.span);
7080 alloc_bootmem_cpumask_var(&def_root_domain.online);
7081 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7082 cpupri_init(&rd->cpupri, true);
7086 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7088 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7090 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7093 if (cpupri_init(&rd->cpupri, false) != 0)
7098 free_cpumask_var(rd->rto_mask);
7100 free_cpumask_var(rd->online);
7102 free_cpumask_var(rd->span);
7107 static void init_defrootdomain(void)
7109 init_rootdomain(&def_root_domain, true);
7111 atomic_set(&def_root_domain.refcount, 1);
7114 static struct root_domain *alloc_rootdomain(void)
7116 struct root_domain *rd;
7118 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7122 if (init_rootdomain(rd, false) != 0) {
7131 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7132 * hold the hotplug lock.
7135 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7137 struct rq *rq = cpu_rq(cpu);
7138 struct sched_domain *tmp;
7140 /* Remove the sched domains which do not contribute to scheduling. */
7141 for (tmp = sd; tmp; ) {
7142 struct sched_domain *parent = tmp->parent;
7146 if (sd_parent_degenerate(tmp, parent)) {
7147 tmp->parent = parent->parent;
7149 parent->parent->child = tmp;
7154 if (sd && sd_degenerate(sd)) {
7160 sched_domain_debug(sd, cpu);
7162 rq_attach_root(rq, rd);
7163 rcu_assign_pointer(rq->sd, sd);
7166 /* cpus with isolated domains */
7167 static cpumask_var_t cpu_isolated_map;
7169 /* Setup the mask of cpus configured for isolated domains */
7170 static int __init isolated_cpu_setup(char *str)
7172 cpulist_parse(str, cpu_isolated_map);
7176 __setup("isolcpus=", isolated_cpu_setup);
7179 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7180 * to a function which identifies what group(along with sched group) a CPU
7181 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7182 * (due to the fact that we keep track of groups covered with a struct cpumask).
7184 * init_sched_build_groups will build a circular linked list of the groups
7185 * covered by the given span, and will set each group's ->cpumask correctly,
7186 * and ->cpu_power to 0.
7189 init_sched_build_groups(const struct cpumask *span,
7190 const struct cpumask *cpu_map,
7191 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7192 struct sched_group **sg,
7193 struct cpumask *tmpmask),
7194 struct cpumask *covered, struct cpumask *tmpmask)
7196 struct sched_group *first = NULL, *last = NULL;
7199 cpumask_clear(covered);
7201 for_each_cpu(i, span) {
7202 struct sched_group *sg;
7203 int group = group_fn(i, cpu_map, &sg, tmpmask);
7206 if (cpumask_test_cpu(i, covered))
7209 cpumask_clear(sched_group_cpus(sg));
7210 sg->__cpu_power = 0;
7212 for_each_cpu(j, span) {
7213 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7216 cpumask_set_cpu(j, covered);
7217 cpumask_set_cpu(j, sched_group_cpus(sg));
7228 #define SD_NODES_PER_DOMAIN 16
7233 * find_next_best_node - find the next node to include in a sched_domain
7234 * @node: node whose sched_domain we're building
7235 * @used_nodes: nodes already in the sched_domain
7237 * Find the next node to include in a given scheduling domain. Simply
7238 * finds the closest node not already in the @used_nodes map.
7240 * Should use nodemask_t.
7242 static int find_next_best_node(int node, nodemask_t *used_nodes)
7244 int i, n, val, min_val, best_node = 0;
7248 for (i = 0; i < nr_node_ids; i++) {
7249 /* Start at @node */
7250 n = (node + i) % nr_node_ids;
7252 if (!nr_cpus_node(n))
7255 /* Skip already used nodes */
7256 if (node_isset(n, *used_nodes))
7259 /* Simple min distance search */
7260 val = node_distance(node, n);
7262 if (val < min_val) {
7268 node_set(best_node, *used_nodes);
7273 * sched_domain_node_span - get a cpumask for a node's sched_domain
7274 * @node: node whose cpumask we're constructing
7275 * @span: resulting cpumask
7277 * Given a node, construct a good cpumask for its sched_domain to span. It
7278 * should be one that prevents unnecessary balancing, but also spreads tasks
7281 static void sched_domain_node_span(int node, struct cpumask *span)
7283 nodemask_t used_nodes;
7286 cpumask_clear(span);
7287 nodes_clear(used_nodes);
7289 cpumask_or(span, span, cpumask_of_node(node));
7290 node_set(node, used_nodes);
7292 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7293 int next_node = find_next_best_node(node, &used_nodes);
7295 cpumask_or(span, span, cpumask_of_node(next_node));
7298 #endif /* CONFIG_NUMA */
7300 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7303 * The cpus mask in sched_group and sched_domain hangs off the end.
7304 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7305 * for nr_cpu_ids < CONFIG_NR_CPUS.
7307 struct static_sched_group {
7308 struct sched_group sg;
7309 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7312 struct static_sched_domain {
7313 struct sched_domain sd;
7314 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7318 * SMT sched-domains:
7320 #ifdef CONFIG_SCHED_SMT
7321 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7322 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7325 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7326 struct sched_group **sg, struct cpumask *unused)
7329 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7332 #endif /* CONFIG_SCHED_SMT */
7335 * multi-core sched-domains:
7337 #ifdef CONFIG_SCHED_MC
7338 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7339 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7340 #endif /* CONFIG_SCHED_MC */
7342 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7344 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7345 struct sched_group **sg, struct cpumask *mask)
7349 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7350 group = cpumask_first(mask);
7352 *sg = &per_cpu(sched_group_core, group).sg;
7355 #elif defined(CONFIG_SCHED_MC)
7357 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7358 struct sched_group **sg, struct cpumask *unused)
7361 *sg = &per_cpu(sched_group_core, cpu).sg;
7366 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7367 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7370 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7371 struct sched_group **sg, struct cpumask *mask)
7374 #ifdef CONFIG_SCHED_MC
7375 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7376 group = cpumask_first(mask);
7377 #elif defined(CONFIG_SCHED_SMT)
7378 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7379 group = cpumask_first(mask);
7384 *sg = &per_cpu(sched_group_phys, group).sg;
7390 * The init_sched_build_groups can't handle what we want to do with node
7391 * groups, so roll our own. Now each node has its own list of groups which
7392 * gets dynamically allocated.
7394 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7395 static struct sched_group ***sched_group_nodes_bycpu;
7397 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7398 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7400 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7401 struct sched_group **sg,
7402 struct cpumask *nodemask)
7406 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7407 group = cpumask_first(nodemask);
7410 *sg = &per_cpu(sched_group_allnodes, group).sg;
7414 static void init_numa_sched_groups_power(struct sched_group *group_head)
7416 struct sched_group *sg = group_head;
7422 for_each_cpu(j, sched_group_cpus(sg)) {
7423 struct sched_domain *sd;
7425 sd = &per_cpu(phys_domains, j).sd;
7426 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7428 * Only add "power" once for each
7434 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7437 } while (sg != group_head);
7439 #endif /* CONFIG_NUMA */
7442 /* Free memory allocated for various sched_group structures */
7443 static void free_sched_groups(const struct cpumask *cpu_map,
7444 struct cpumask *nodemask)
7448 for_each_cpu(cpu, cpu_map) {
7449 struct sched_group **sched_group_nodes
7450 = sched_group_nodes_bycpu[cpu];
7452 if (!sched_group_nodes)
7455 for (i = 0; i < nr_node_ids; i++) {
7456 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7458 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7459 if (cpumask_empty(nodemask))
7469 if (oldsg != sched_group_nodes[i])
7472 kfree(sched_group_nodes);
7473 sched_group_nodes_bycpu[cpu] = NULL;
7476 #else /* !CONFIG_NUMA */
7477 static void free_sched_groups(const struct cpumask *cpu_map,
7478 struct cpumask *nodemask)
7481 #endif /* CONFIG_NUMA */
7484 * Initialize sched groups cpu_power.
7486 * cpu_power indicates the capacity of sched group, which is used while
7487 * distributing the load between different sched groups in a sched domain.
7488 * Typically cpu_power for all the groups in a sched domain will be same unless
7489 * there are asymmetries in the topology. If there are asymmetries, group
7490 * having more cpu_power will pickup more load compared to the group having
7493 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7494 * the maximum number of tasks a group can handle in the presence of other idle
7495 * or lightly loaded groups in the same sched domain.
7497 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7499 struct sched_domain *child;
7500 struct sched_group *group;
7502 WARN_ON(!sd || !sd->groups);
7504 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7509 sd->groups->__cpu_power = 0;
7512 * For perf policy, if the groups in child domain share resources
7513 * (for example cores sharing some portions of the cache hierarchy
7514 * or SMT), then set this domain groups cpu_power such that each group
7515 * can handle only one task, when there are other idle groups in the
7516 * same sched domain.
7518 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7520 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7521 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7526 * add cpu_power of each child group to this groups cpu_power
7528 group = child->groups;
7530 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7531 group = group->next;
7532 } while (group != child->groups);
7536 * Initializers for schedule domains
7537 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7540 #ifdef CONFIG_SCHED_DEBUG
7541 # define SD_INIT_NAME(sd, type) sd->name = #type
7543 # define SD_INIT_NAME(sd, type) do { } while (0)
7546 #define SD_INIT(sd, type) sd_init_##type(sd)
7548 #define SD_INIT_FUNC(type) \
7549 static noinline void sd_init_##type(struct sched_domain *sd) \
7551 memset(sd, 0, sizeof(*sd)); \
7552 *sd = SD_##type##_INIT; \
7553 sd->level = SD_LV_##type; \
7554 SD_INIT_NAME(sd, type); \
7559 SD_INIT_FUNC(ALLNODES)
7562 #ifdef CONFIG_SCHED_SMT
7563 SD_INIT_FUNC(SIBLING)
7565 #ifdef CONFIG_SCHED_MC
7569 static int default_relax_domain_level = -1;
7571 static int __init setup_relax_domain_level(char *str)
7575 val = simple_strtoul(str, NULL, 0);
7576 if (val < SD_LV_MAX)
7577 default_relax_domain_level = val;
7581 __setup("relax_domain_level=", setup_relax_domain_level);
7583 static void set_domain_attribute(struct sched_domain *sd,
7584 struct sched_domain_attr *attr)
7588 if (!attr || attr->relax_domain_level < 0) {
7589 if (default_relax_domain_level < 0)
7592 request = default_relax_domain_level;
7594 request = attr->relax_domain_level;
7595 if (request < sd->level) {
7596 /* turn off idle balance on this domain */
7597 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7599 /* turn on idle balance on this domain */
7600 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7605 * Build sched domains for a given set of cpus and attach the sched domains
7606 * to the individual cpus
7608 static int __build_sched_domains(const struct cpumask *cpu_map,
7609 struct sched_domain_attr *attr)
7611 int i, err = -ENOMEM;
7612 struct root_domain *rd;
7613 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7616 cpumask_var_t domainspan, covered, notcovered;
7617 struct sched_group **sched_group_nodes = NULL;
7618 int sd_allnodes = 0;
7620 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7622 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7623 goto free_domainspan;
7624 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7628 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7629 goto free_notcovered;
7630 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7632 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7633 goto free_this_sibling_map;
7634 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7635 goto free_this_core_map;
7636 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7637 goto free_send_covered;
7641 * Allocate the per-node list of sched groups
7643 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7645 if (!sched_group_nodes) {
7646 printk(KERN_WARNING "Can not alloc sched group node list\n");
7651 rd = alloc_rootdomain();
7653 printk(KERN_WARNING "Cannot alloc root domain\n");
7654 goto free_sched_groups;
7658 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7662 * Set up domains for cpus specified by the cpu_map.
7664 for_each_cpu(i, cpu_map) {
7665 struct sched_domain *sd = NULL, *p;
7667 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7670 if (cpumask_weight(cpu_map) >
7671 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7672 sd = &per_cpu(allnodes_domains, i).sd;
7673 SD_INIT(sd, ALLNODES);
7674 set_domain_attribute(sd, attr);
7675 cpumask_copy(sched_domain_span(sd), cpu_map);
7676 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7682 sd = &per_cpu(node_domains, i).sd;
7684 set_domain_attribute(sd, attr);
7685 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7689 cpumask_and(sched_domain_span(sd),
7690 sched_domain_span(sd), cpu_map);
7694 sd = &per_cpu(phys_domains, i).sd;
7696 set_domain_attribute(sd, attr);
7697 cpumask_copy(sched_domain_span(sd), nodemask);
7701 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7703 #ifdef CONFIG_SCHED_MC
7705 sd = &per_cpu(core_domains, i).sd;
7707 set_domain_attribute(sd, attr);
7708 cpumask_and(sched_domain_span(sd), cpu_map,
7709 cpu_coregroup_mask(i));
7712 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7715 #ifdef CONFIG_SCHED_SMT
7717 sd = &per_cpu(cpu_domains, i).sd;
7718 SD_INIT(sd, SIBLING);
7719 set_domain_attribute(sd, attr);
7720 cpumask_and(sched_domain_span(sd),
7721 &per_cpu(cpu_sibling_map, i), cpu_map);
7724 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7728 #ifdef CONFIG_SCHED_SMT
7729 /* Set up CPU (sibling) groups */
7730 for_each_cpu(i, cpu_map) {
7731 cpumask_and(this_sibling_map,
7732 &per_cpu(cpu_sibling_map, i), cpu_map);
7733 if (i != cpumask_first(this_sibling_map))
7736 init_sched_build_groups(this_sibling_map, cpu_map,
7738 send_covered, tmpmask);
7742 #ifdef CONFIG_SCHED_MC
7743 /* Set up multi-core groups */
7744 for_each_cpu(i, cpu_map) {
7745 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7746 if (i != cpumask_first(this_core_map))
7749 init_sched_build_groups(this_core_map, cpu_map,
7751 send_covered, tmpmask);
7755 /* Set up physical groups */
7756 for (i = 0; i < nr_node_ids; i++) {
7757 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7758 if (cpumask_empty(nodemask))
7761 init_sched_build_groups(nodemask, cpu_map,
7763 send_covered, tmpmask);
7767 /* Set up node groups */
7769 init_sched_build_groups(cpu_map, cpu_map,
7770 &cpu_to_allnodes_group,
7771 send_covered, tmpmask);
7774 for (i = 0; i < nr_node_ids; i++) {
7775 /* Set up node groups */
7776 struct sched_group *sg, *prev;
7779 cpumask_clear(covered);
7780 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7781 if (cpumask_empty(nodemask)) {
7782 sched_group_nodes[i] = NULL;
7786 sched_domain_node_span(i, domainspan);
7787 cpumask_and(domainspan, domainspan, cpu_map);
7789 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7792 printk(KERN_WARNING "Can not alloc domain group for "
7796 sched_group_nodes[i] = sg;
7797 for_each_cpu(j, nodemask) {
7798 struct sched_domain *sd;
7800 sd = &per_cpu(node_domains, j).sd;
7803 sg->__cpu_power = 0;
7804 cpumask_copy(sched_group_cpus(sg), nodemask);
7806 cpumask_or(covered, covered, nodemask);
7809 for (j = 0; j < nr_node_ids; j++) {
7810 int n = (i + j) % nr_node_ids;
7812 cpumask_complement(notcovered, covered);
7813 cpumask_and(tmpmask, notcovered, cpu_map);
7814 cpumask_and(tmpmask, tmpmask, domainspan);
7815 if (cpumask_empty(tmpmask))
7818 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7819 if (cpumask_empty(tmpmask))
7822 sg = kmalloc_node(sizeof(struct sched_group) +
7827 "Can not alloc domain group for node %d\n", j);
7830 sg->__cpu_power = 0;
7831 cpumask_copy(sched_group_cpus(sg), tmpmask);
7832 sg->next = prev->next;
7833 cpumask_or(covered, covered, tmpmask);
7840 /* Calculate CPU power for physical packages and nodes */
7841 #ifdef CONFIG_SCHED_SMT
7842 for_each_cpu(i, cpu_map) {
7843 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7845 init_sched_groups_power(i, sd);
7848 #ifdef CONFIG_SCHED_MC
7849 for_each_cpu(i, cpu_map) {
7850 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7852 init_sched_groups_power(i, sd);
7856 for_each_cpu(i, cpu_map) {
7857 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7859 init_sched_groups_power(i, sd);
7863 for (i = 0; i < nr_node_ids; i++)
7864 init_numa_sched_groups_power(sched_group_nodes[i]);
7867 struct sched_group *sg;
7869 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7871 init_numa_sched_groups_power(sg);
7875 /* Attach the domains */
7876 for_each_cpu(i, cpu_map) {
7877 struct sched_domain *sd;
7878 #ifdef CONFIG_SCHED_SMT
7879 sd = &per_cpu(cpu_domains, i).sd;
7880 #elif defined(CONFIG_SCHED_MC)
7881 sd = &per_cpu(core_domains, i).sd;
7883 sd = &per_cpu(phys_domains, i).sd;
7885 cpu_attach_domain(sd, rd, i);
7891 free_cpumask_var(tmpmask);
7893 free_cpumask_var(send_covered);
7895 free_cpumask_var(this_core_map);
7896 free_this_sibling_map:
7897 free_cpumask_var(this_sibling_map);
7899 free_cpumask_var(nodemask);
7902 free_cpumask_var(notcovered);
7904 free_cpumask_var(covered);
7906 free_cpumask_var(domainspan);
7913 kfree(sched_group_nodes);
7919 free_sched_groups(cpu_map, tmpmask);
7920 free_rootdomain(rd);
7925 static int build_sched_domains(const struct cpumask *cpu_map)
7927 return __build_sched_domains(cpu_map, NULL);
7930 static struct cpumask *doms_cur; /* current sched domains */
7931 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7932 static struct sched_domain_attr *dattr_cur;
7933 /* attribues of custom domains in 'doms_cur' */
7936 * Special case: If a kmalloc of a doms_cur partition (array of
7937 * cpumask) fails, then fallback to a single sched domain,
7938 * as determined by the single cpumask fallback_doms.
7940 static cpumask_var_t fallback_doms;
7943 * arch_update_cpu_topology lets virtualized architectures update the
7944 * cpu core maps. It is supposed to return 1 if the topology changed
7945 * or 0 if it stayed the same.
7947 int __attribute__((weak)) arch_update_cpu_topology(void)
7953 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7954 * For now this just excludes isolated cpus, but could be used to
7955 * exclude other special cases in the future.
7957 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7961 arch_update_cpu_topology();
7963 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7965 doms_cur = fallback_doms;
7966 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7968 err = build_sched_domains(doms_cur);
7969 register_sched_domain_sysctl();
7974 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7975 struct cpumask *tmpmask)
7977 free_sched_groups(cpu_map, tmpmask);
7981 * Detach sched domains from a group of cpus specified in cpu_map
7982 * These cpus will now be attached to the NULL domain
7984 static void detach_destroy_domains(const struct cpumask *cpu_map)
7986 /* Save because hotplug lock held. */
7987 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7990 for_each_cpu(i, cpu_map)
7991 cpu_attach_domain(NULL, &def_root_domain, i);
7992 synchronize_sched();
7993 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7996 /* handle null as "default" */
7997 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7998 struct sched_domain_attr *new, int idx_new)
8000 struct sched_domain_attr tmp;
8007 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8008 new ? (new + idx_new) : &tmp,
8009 sizeof(struct sched_domain_attr));
8013 * Partition sched domains as specified by the 'ndoms_new'
8014 * cpumasks in the array doms_new[] of cpumasks. This compares
8015 * doms_new[] to the current sched domain partitioning, doms_cur[].
8016 * It destroys each deleted domain and builds each new domain.
8018 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8019 * The masks don't intersect (don't overlap.) We should setup one
8020 * sched domain for each mask. CPUs not in any of the cpumasks will
8021 * not be load balanced. If the same cpumask appears both in the
8022 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8025 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8026 * ownership of it and will kfree it when done with it. If the caller
8027 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8028 * ndoms_new == 1, and partition_sched_domains() will fallback to
8029 * the single partition 'fallback_doms', it also forces the domains
8032 * If doms_new == NULL it will be replaced with cpu_online_mask.
8033 * ndoms_new == 0 is a special case for destroying existing domains,
8034 * and it will not create the default domain.
8036 * Call with hotplug lock held
8038 /* FIXME: Change to struct cpumask *doms_new[] */
8039 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8040 struct sched_domain_attr *dattr_new)
8045 mutex_lock(&sched_domains_mutex);
8047 /* always unregister in case we don't destroy any domains */
8048 unregister_sched_domain_sysctl();
8050 /* Let architecture update cpu core mappings. */
8051 new_topology = arch_update_cpu_topology();
8053 n = doms_new ? ndoms_new : 0;
8055 /* Destroy deleted domains */
8056 for (i = 0; i < ndoms_cur; i++) {
8057 for (j = 0; j < n && !new_topology; j++) {
8058 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8059 && dattrs_equal(dattr_cur, i, dattr_new, j))
8062 /* no match - a current sched domain not in new doms_new[] */
8063 detach_destroy_domains(doms_cur + i);
8068 if (doms_new == NULL) {
8070 doms_new = fallback_doms;
8071 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8072 WARN_ON_ONCE(dattr_new);
8075 /* Build new domains */
8076 for (i = 0; i < ndoms_new; i++) {
8077 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8078 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8079 && dattrs_equal(dattr_new, i, dattr_cur, j))
8082 /* no match - add a new doms_new */
8083 __build_sched_domains(doms_new + i,
8084 dattr_new ? dattr_new + i : NULL);
8089 /* Remember the new sched domains */
8090 if (doms_cur != fallback_doms)
8092 kfree(dattr_cur); /* kfree(NULL) is safe */
8093 doms_cur = doms_new;
8094 dattr_cur = dattr_new;
8095 ndoms_cur = ndoms_new;
8097 register_sched_domain_sysctl();
8099 mutex_unlock(&sched_domains_mutex);
8102 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8103 static void arch_reinit_sched_domains(void)
8107 /* Destroy domains first to force the rebuild */
8108 partition_sched_domains(0, NULL, NULL);
8110 rebuild_sched_domains();
8114 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8116 unsigned int level = 0;
8118 if (sscanf(buf, "%u", &level) != 1)
8122 * level is always be positive so don't check for
8123 * level < POWERSAVINGS_BALANCE_NONE which is 0
8124 * What happens on 0 or 1 byte write,
8125 * need to check for count as well?
8128 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8132 sched_smt_power_savings = level;
8134 sched_mc_power_savings = level;
8136 arch_reinit_sched_domains();
8141 #ifdef CONFIG_SCHED_MC
8142 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8145 return sprintf(page, "%u\n", sched_mc_power_savings);
8147 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8148 const char *buf, size_t count)
8150 return sched_power_savings_store(buf, count, 0);
8152 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8153 sched_mc_power_savings_show,
8154 sched_mc_power_savings_store);
8157 #ifdef CONFIG_SCHED_SMT
8158 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8161 return sprintf(page, "%u\n", sched_smt_power_savings);
8163 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8164 const char *buf, size_t count)
8166 return sched_power_savings_store(buf, count, 1);
8168 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8169 sched_smt_power_savings_show,
8170 sched_smt_power_savings_store);
8173 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8177 #ifdef CONFIG_SCHED_SMT
8179 err = sysfs_create_file(&cls->kset.kobj,
8180 &attr_sched_smt_power_savings.attr);
8182 #ifdef CONFIG_SCHED_MC
8183 if (!err && mc_capable())
8184 err = sysfs_create_file(&cls->kset.kobj,
8185 &attr_sched_mc_power_savings.attr);
8189 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8191 #ifndef CONFIG_CPUSETS
8193 * Add online and remove offline CPUs from the scheduler domains.
8194 * When cpusets are enabled they take over this function.
8196 static int update_sched_domains(struct notifier_block *nfb,
8197 unsigned long action, void *hcpu)
8201 case CPU_ONLINE_FROZEN:
8203 case CPU_DEAD_FROZEN:
8204 partition_sched_domains(1, NULL, NULL);
8213 static int update_runtime(struct notifier_block *nfb,
8214 unsigned long action, void *hcpu)
8216 int cpu = (int)(long)hcpu;
8219 case CPU_DOWN_PREPARE:
8220 case CPU_DOWN_PREPARE_FROZEN:
8221 disable_runtime(cpu_rq(cpu));
8224 case CPU_DOWN_FAILED:
8225 case CPU_DOWN_FAILED_FROZEN:
8227 case CPU_ONLINE_FROZEN:
8228 enable_runtime(cpu_rq(cpu));
8236 void __init sched_init_smp(void)
8238 cpumask_var_t non_isolated_cpus;
8240 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8242 #if defined(CONFIG_NUMA)
8243 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8245 BUG_ON(sched_group_nodes_bycpu == NULL);
8248 mutex_lock(&sched_domains_mutex);
8249 arch_init_sched_domains(cpu_online_mask);
8250 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8251 if (cpumask_empty(non_isolated_cpus))
8252 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8253 mutex_unlock(&sched_domains_mutex);
8256 #ifndef CONFIG_CPUSETS
8257 /* XXX: Theoretical race here - CPU may be hotplugged now */
8258 hotcpu_notifier(update_sched_domains, 0);
8261 /* RT runtime code needs to handle some hotplug events */
8262 hotcpu_notifier(update_runtime, 0);
8266 /* Move init over to a non-isolated CPU */
8267 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8269 sched_init_granularity();
8270 free_cpumask_var(non_isolated_cpus);
8272 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8273 init_sched_rt_class();
8276 void __init sched_init_smp(void)
8278 sched_init_granularity();
8280 #endif /* CONFIG_SMP */
8282 int in_sched_functions(unsigned long addr)
8284 return in_lock_functions(addr) ||
8285 (addr >= (unsigned long)__sched_text_start
8286 && addr < (unsigned long)__sched_text_end);
8289 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8291 cfs_rq->tasks_timeline = RB_ROOT;
8292 INIT_LIST_HEAD(&cfs_rq->tasks);
8293 #ifdef CONFIG_FAIR_GROUP_SCHED
8296 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8299 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8301 struct rt_prio_array *array;
8304 array = &rt_rq->active;
8305 for (i = 0; i < MAX_RT_PRIO; i++) {
8306 INIT_LIST_HEAD(array->queue + i);
8307 __clear_bit(i, array->bitmap);
8309 /* delimiter for bitsearch: */
8310 __set_bit(MAX_RT_PRIO, array->bitmap);
8312 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8313 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8315 rt_rq->highest_prio.next = MAX_RT_PRIO;
8319 rt_rq->rt_nr_migratory = 0;
8320 rt_rq->overloaded = 0;
8321 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8325 rt_rq->rt_throttled = 0;
8326 rt_rq->rt_runtime = 0;
8327 spin_lock_init(&rt_rq->rt_runtime_lock);
8329 #ifdef CONFIG_RT_GROUP_SCHED
8330 rt_rq->rt_nr_boosted = 0;
8335 #ifdef CONFIG_FAIR_GROUP_SCHED
8336 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8337 struct sched_entity *se, int cpu, int add,
8338 struct sched_entity *parent)
8340 struct rq *rq = cpu_rq(cpu);
8341 tg->cfs_rq[cpu] = cfs_rq;
8342 init_cfs_rq(cfs_rq, rq);
8345 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8348 /* se could be NULL for init_task_group */
8353 se->cfs_rq = &rq->cfs;
8355 se->cfs_rq = parent->my_q;
8358 se->load.weight = tg->shares;
8359 se->load.inv_weight = 0;
8360 se->parent = parent;
8364 #ifdef CONFIG_RT_GROUP_SCHED
8365 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8366 struct sched_rt_entity *rt_se, int cpu, int add,
8367 struct sched_rt_entity *parent)
8369 struct rq *rq = cpu_rq(cpu);
8371 tg->rt_rq[cpu] = rt_rq;
8372 init_rt_rq(rt_rq, rq);
8374 rt_rq->rt_se = rt_se;
8375 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8377 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8379 tg->rt_se[cpu] = rt_se;
8384 rt_se->rt_rq = &rq->rt;
8386 rt_se->rt_rq = parent->my_q;
8388 rt_se->my_q = rt_rq;
8389 rt_se->parent = parent;
8390 INIT_LIST_HEAD(&rt_se->run_list);
8394 void __init sched_init(void)
8397 unsigned long alloc_size = 0, ptr;
8399 #ifdef CONFIG_FAIR_GROUP_SCHED
8400 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8402 #ifdef CONFIG_RT_GROUP_SCHED
8403 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8405 #ifdef CONFIG_USER_SCHED
8409 * As sched_init() is called before page_alloc is setup,
8410 * we use alloc_bootmem().
8413 ptr = (unsigned long)alloc_bootmem(alloc_size);
8415 #ifdef CONFIG_FAIR_GROUP_SCHED
8416 init_task_group.se = (struct sched_entity **)ptr;
8417 ptr += nr_cpu_ids * sizeof(void **);
8419 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8420 ptr += nr_cpu_ids * sizeof(void **);
8422 #ifdef CONFIG_USER_SCHED
8423 root_task_group.se = (struct sched_entity **)ptr;
8424 ptr += nr_cpu_ids * sizeof(void **);
8426 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8427 ptr += nr_cpu_ids * sizeof(void **);
8428 #endif /* CONFIG_USER_SCHED */
8429 #endif /* CONFIG_FAIR_GROUP_SCHED */
8430 #ifdef CONFIG_RT_GROUP_SCHED
8431 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8432 ptr += nr_cpu_ids * sizeof(void **);
8434 init_task_group.rt_rq = (struct rt_rq **)ptr;
8435 ptr += nr_cpu_ids * sizeof(void **);
8437 #ifdef CONFIG_USER_SCHED
8438 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8439 ptr += nr_cpu_ids * sizeof(void **);
8441 root_task_group.rt_rq = (struct rt_rq **)ptr;
8442 ptr += nr_cpu_ids * sizeof(void **);
8443 #endif /* CONFIG_USER_SCHED */
8444 #endif /* CONFIG_RT_GROUP_SCHED */
8448 init_defrootdomain();
8451 init_rt_bandwidth(&def_rt_bandwidth,
8452 global_rt_period(), global_rt_runtime());
8454 #ifdef CONFIG_RT_GROUP_SCHED
8455 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8456 global_rt_period(), global_rt_runtime());
8457 #ifdef CONFIG_USER_SCHED
8458 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8459 global_rt_period(), RUNTIME_INF);
8460 #endif /* CONFIG_USER_SCHED */
8461 #endif /* CONFIG_RT_GROUP_SCHED */
8463 #ifdef CONFIG_GROUP_SCHED
8464 list_add(&init_task_group.list, &task_groups);
8465 INIT_LIST_HEAD(&init_task_group.children);
8467 #ifdef CONFIG_USER_SCHED
8468 INIT_LIST_HEAD(&root_task_group.children);
8469 init_task_group.parent = &root_task_group;
8470 list_add(&init_task_group.siblings, &root_task_group.children);
8471 #endif /* CONFIG_USER_SCHED */
8472 #endif /* CONFIG_GROUP_SCHED */
8474 for_each_possible_cpu(i) {
8478 spin_lock_init(&rq->lock);
8480 init_cfs_rq(&rq->cfs, rq);
8481 init_rt_rq(&rq->rt, rq);
8482 #ifdef CONFIG_FAIR_GROUP_SCHED
8483 init_task_group.shares = init_task_group_load;
8484 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8485 #ifdef CONFIG_CGROUP_SCHED
8487 * How much cpu bandwidth does init_task_group get?
8489 * In case of task-groups formed thr' the cgroup filesystem, it
8490 * gets 100% of the cpu resources in the system. This overall
8491 * system cpu resource is divided among the tasks of
8492 * init_task_group and its child task-groups in a fair manner,
8493 * based on each entity's (task or task-group's) weight
8494 * (se->load.weight).
8496 * In other words, if init_task_group has 10 tasks of weight
8497 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8498 * then A0's share of the cpu resource is:
8500 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8502 * We achieve this by letting init_task_group's tasks sit
8503 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8505 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8506 #elif defined CONFIG_USER_SCHED
8507 root_task_group.shares = NICE_0_LOAD;
8508 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8510 * In case of task-groups formed thr' the user id of tasks,
8511 * init_task_group represents tasks belonging to root user.
8512 * Hence it forms a sibling of all subsequent groups formed.
8513 * In this case, init_task_group gets only a fraction of overall
8514 * system cpu resource, based on the weight assigned to root
8515 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8516 * by letting tasks of init_task_group sit in a separate cfs_rq
8517 * (init_cfs_rq) and having one entity represent this group of
8518 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8520 init_tg_cfs_entry(&init_task_group,
8521 &per_cpu(init_cfs_rq, i),
8522 &per_cpu(init_sched_entity, i), i, 1,
8523 root_task_group.se[i]);
8526 #endif /* CONFIG_FAIR_GROUP_SCHED */
8528 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8529 #ifdef CONFIG_RT_GROUP_SCHED
8530 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8531 #ifdef CONFIG_CGROUP_SCHED
8532 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8533 #elif defined CONFIG_USER_SCHED
8534 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8535 init_tg_rt_entry(&init_task_group,
8536 &per_cpu(init_rt_rq, i),
8537 &per_cpu(init_sched_rt_entity, i), i, 1,
8538 root_task_group.rt_se[i]);
8542 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8543 rq->cpu_load[j] = 0;
8547 rq->active_balance = 0;
8548 rq->next_balance = jiffies;
8552 rq->migration_thread = NULL;
8553 INIT_LIST_HEAD(&rq->migration_queue);
8554 rq_attach_root(rq, &def_root_domain);
8557 atomic_set(&rq->nr_iowait, 0);
8560 set_load_weight(&init_task);
8562 #ifdef CONFIG_PREEMPT_NOTIFIERS
8563 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8567 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8570 #ifdef CONFIG_RT_MUTEXES
8571 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8575 * The boot idle thread does lazy MMU switching as well:
8577 atomic_inc(&init_mm.mm_count);
8578 enter_lazy_tlb(&init_mm, current);
8581 * Make us the idle thread. Technically, schedule() should not be
8582 * called from this thread, however somewhere below it might be,
8583 * but because we are the idle thread, we just pick up running again
8584 * when this runqueue becomes "idle".
8586 init_idle(current, smp_processor_id());
8588 * During early bootup we pretend to be a normal task:
8590 current->sched_class = &fair_sched_class;
8592 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8593 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8596 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8598 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8601 scheduler_running = 1;
8604 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8605 void __might_sleep(char *file, int line)
8608 static unsigned long prev_jiffy; /* ratelimiting */
8610 if ((!in_atomic() && !irqs_disabled()) ||
8611 system_state != SYSTEM_RUNNING || oops_in_progress)
8613 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8615 prev_jiffy = jiffies;
8618 "BUG: sleeping function called from invalid context at %s:%d\n",
8621 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8622 in_atomic(), irqs_disabled(),
8623 current->pid, current->comm);
8625 debug_show_held_locks(current);
8626 if (irqs_disabled())
8627 print_irqtrace_events(current);
8631 EXPORT_SYMBOL(__might_sleep);
8634 #ifdef CONFIG_MAGIC_SYSRQ
8635 static void normalize_task(struct rq *rq, struct task_struct *p)
8639 update_rq_clock(rq);
8640 on_rq = p->se.on_rq;
8642 deactivate_task(rq, p, 0);
8643 __setscheduler(rq, p, SCHED_NORMAL, 0);
8645 activate_task(rq, p, 0);
8646 resched_task(rq->curr);
8650 void normalize_rt_tasks(void)
8652 struct task_struct *g, *p;
8653 unsigned long flags;
8656 read_lock_irqsave(&tasklist_lock, flags);
8657 do_each_thread(g, p) {
8659 * Only normalize user tasks:
8664 p->se.exec_start = 0;
8665 #ifdef CONFIG_SCHEDSTATS
8666 p->se.wait_start = 0;
8667 p->se.sleep_start = 0;
8668 p->se.block_start = 0;
8673 * Renice negative nice level userspace
8676 if (TASK_NICE(p) < 0 && p->mm)
8677 set_user_nice(p, 0);
8681 spin_lock(&p->pi_lock);
8682 rq = __task_rq_lock(p);
8684 normalize_task(rq, p);
8686 __task_rq_unlock(rq);
8687 spin_unlock(&p->pi_lock);
8688 } while_each_thread(g, p);
8690 read_unlock_irqrestore(&tasklist_lock, flags);
8693 #endif /* CONFIG_MAGIC_SYSRQ */
8697 * These functions are only useful for the IA64 MCA handling.
8699 * They can only be called when the whole system has been
8700 * stopped - every CPU needs to be quiescent, and no scheduling
8701 * activity can take place. Using them for anything else would
8702 * be a serious bug, and as a result, they aren't even visible
8703 * under any other configuration.
8707 * curr_task - return the current task for a given cpu.
8708 * @cpu: the processor in question.
8710 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8712 struct task_struct *curr_task(int cpu)
8714 return cpu_curr(cpu);
8718 * set_curr_task - set the current task for a given cpu.
8719 * @cpu: the processor in question.
8720 * @p: the task pointer to set.
8722 * Description: This function must only be used when non-maskable interrupts
8723 * are serviced on a separate stack. It allows the architecture to switch the
8724 * notion of the current task on a cpu in a non-blocking manner. This function
8725 * must be called with all CPU's synchronized, and interrupts disabled, the
8726 * and caller must save the original value of the current task (see
8727 * curr_task() above) and restore that value before reenabling interrupts and
8728 * re-starting the system.
8730 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8732 void set_curr_task(int cpu, struct task_struct *p)
8739 #ifdef CONFIG_FAIR_GROUP_SCHED
8740 static void free_fair_sched_group(struct task_group *tg)
8744 for_each_possible_cpu(i) {
8746 kfree(tg->cfs_rq[i]);
8756 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8758 struct cfs_rq *cfs_rq;
8759 struct sched_entity *se;
8763 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8766 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8770 tg->shares = NICE_0_LOAD;
8772 for_each_possible_cpu(i) {
8775 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8776 GFP_KERNEL, cpu_to_node(i));
8780 se = kzalloc_node(sizeof(struct sched_entity),
8781 GFP_KERNEL, cpu_to_node(i));
8785 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8794 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8796 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8797 &cpu_rq(cpu)->leaf_cfs_rq_list);
8800 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8802 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8804 #else /* !CONFG_FAIR_GROUP_SCHED */
8805 static inline void free_fair_sched_group(struct task_group *tg)
8810 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8815 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8819 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8822 #endif /* CONFIG_FAIR_GROUP_SCHED */
8824 #ifdef CONFIG_RT_GROUP_SCHED
8825 static void free_rt_sched_group(struct task_group *tg)
8829 destroy_rt_bandwidth(&tg->rt_bandwidth);
8831 for_each_possible_cpu(i) {
8833 kfree(tg->rt_rq[i]);
8835 kfree(tg->rt_se[i]);
8843 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8845 struct rt_rq *rt_rq;
8846 struct sched_rt_entity *rt_se;
8850 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8853 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8857 init_rt_bandwidth(&tg->rt_bandwidth,
8858 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8860 for_each_possible_cpu(i) {
8863 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8864 GFP_KERNEL, cpu_to_node(i));
8868 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8869 GFP_KERNEL, cpu_to_node(i));
8873 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8882 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8884 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8885 &cpu_rq(cpu)->leaf_rt_rq_list);
8888 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8890 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8892 #else /* !CONFIG_RT_GROUP_SCHED */
8893 static inline void free_rt_sched_group(struct task_group *tg)
8898 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8903 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8907 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8910 #endif /* CONFIG_RT_GROUP_SCHED */
8912 #ifdef CONFIG_GROUP_SCHED
8913 static void free_sched_group(struct task_group *tg)
8915 free_fair_sched_group(tg);
8916 free_rt_sched_group(tg);
8920 /* allocate runqueue etc for a new task group */
8921 struct task_group *sched_create_group(struct task_group *parent)
8923 struct task_group *tg;
8924 unsigned long flags;
8927 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8929 return ERR_PTR(-ENOMEM);
8931 if (!alloc_fair_sched_group(tg, parent))
8934 if (!alloc_rt_sched_group(tg, parent))
8937 spin_lock_irqsave(&task_group_lock, flags);
8938 for_each_possible_cpu(i) {
8939 register_fair_sched_group(tg, i);
8940 register_rt_sched_group(tg, i);
8942 list_add_rcu(&tg->list, &task_groups);
8944 WARN_ON(!parent); /* root should already exist */
8946 tg->parent = parent;
8947 INIT_LIST_HEAD(&tg->children);
8948 list_add_rcu(&tg->siblings, &parent->children);
8949 spin_unlock_irqrestore(&task_group_lock, flags);
8954 free_sched_group(tg);
8955 return ERR_PTR(-ENOMEM);
8958 /* rcu callback to free various structures associated with a task group */
8959 static void free_sched_group_rcu(struct rcu_head *rhp)
8961 /* now it should be safe to free those cfs_rqs */
8962 free_sched_group(container_of(rhp, struct task_group, rcu));
8965 /* Destroy runqueue etc associated with a task group */
8966 void sched_destroy_group(struct task_group *tg)
8968 unsigned long flags;
8971 spin_lock_irqsave(&task_group_lock, flags);
8972 for_each_possible_cpu(i) {
8973 unregister_fair_sched_group(tg, i);
8974 unregister_rt_sched_group(tg, i);
8976 list_del_rcu(&tg->list);
8977 list_del_rcu(&tg->siblings);
8978 spin_unlock_irqrestore(&task_group_lock, flags);
8980 /* wait for possible concurrent references to cfs_rqs complete */
8981 call_rcu(&tg->rcu, free_sched_group_rcu);
8984 /* change task's runqueue when it moves between groups.
8985 * The caller of this function should have put the task in its new group
8986 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8987 * reflect its new group.
8989 void sched_move_task(struct task_struct *tsk)
8992 unsigned long flags;
8995 rq = task_rq_lock(tsk, &flags);
8997 update_rq_clock(rq);
8999 running = task_current(rq, tsk);
9000 on_rq = tsk->se.on_rq;
9003 dequeue_task(rq, tsk, 0);
9004 if (unlikely(running))
9005 tsk->sched_class->put_prev_task(rq, tsk);
9007 set_task_rq(tsk, task_cpu(tsk));
9009 #ifdef CONFIG_FAIR_GROUP_SCHED
9010 if (tsk->sched_class->moved_group)
9011 tsk->sched_class->moved_group(tsk);
9014 if (unlikely(running))
9015 tsk->sched_class->set_curr_task(rq);
9017 enqueue_task(rq, tsk, 0);
9019 task_rq_unlock(rq, &flags);
9021 #endif /* CONFIG_GROUP_SCHED */
9023 #ifdef CONFIG_FAIR_GROUP_SCHED
9024 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9026 struct cfs_rq *cfs_rq = se->cfs_rq;
9031 dequeue_entity(cfs_rq, se, 0);
9033 se->load.weight = shares;
9034 se->load.inv_weight = 0;
9037 enqueue_entity(cfs_rq, se, 0);
9040 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9042 struct cfs_rq *cfs_rq = se->cfs_rq;
9043 struct rq *rq = cfs_rq->rq;
9044 unsigned long flags;
9046 spin_lock_irqsave(&rq->lock, flags);
9047 __set_se_shares(se, shares);
9048 spin_unlock_irqrestore(&rq->lock, flags);
9051 static DEFINE_MUTEX(shares_mutex);
9053 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9056 unsigned long flags;
9059 * We can't change the weight of the root cgroup.
9064 if (shares < MIN_SHARES)
9065 shares = MIN_SHARES;
9066 else if (shares > MAX_SHARES)
9067 shares = MAX_SHARES;
9069 mutex_lock(&shares_mutex);
9070 if (tg->shares == shares)
9073 spin_lock_irqsave(&task_group_lock, flags);
9074 for_each_possible_cpu(i)
9075 unregister_fair_sched_group(tg, i);
9076 list_del_rcu(&tg->siblings);
9077 spin_unlock_irqrestore(&task_group_lock, flags);
9079 /* wait for any ongoing reference to this group to finish */
9080 synchronize_sched();
9083 * Now we are free to modify the group's share on each cpu
9084 * w/o tripping rebalance_share or load_balance_fair.
9086 tg->shares = shares;
9087 for_each_possible_cpu(i) {
9091 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9092 set_se_shares(tg->se[i], shares);
9096 * Enable load balance activity on this group, by inserting it back on
9097 * each cpu's rq->leaf_cfs_rq_list.
9099 spin_lock_irqsave(&task_group_lock, flags);
9100 for_each_possible_cpu(i)
9101 register_fair_sched_group(tg, i);
9102 list_add_rcu(&tg->siblings, &tg->parent->children);
9103 spin_unlock_irqrestore(&task_group_lock, flags);
9105 mutex_unlock(&shares_mutex);
9109 unsigned long sched_group_shares(struct task_group *tg)
9115 #ifdef CONFIG_RT_GROUP_SCHED
9117 * Ensure that the real time constraints are schedulable.
9119 static DEFINE_MUTEX(rt_constraints_mutex);
9121 static unsigned long to_ratio(u64 period, u64 runtime)
9123 if (runtime == RUNTIME_INF)
9126 return div64_u64(runtime << 20, period);
9129 /* Must be called with tasklist_lock held */
9130 static inline int tg_has_rt_tasks(struct task_group *tg)
9132 struct task_struct *g, *p;
9134 do_each_thread(g, p) {
9135 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9137 } while_each_thread(g, p);
9142 struct rt_schedulable_data {
9143 struct task_group *tg;
9148 static int tg_schedulable(struct task_group *tg, void *data)
9150 struct rt_schedulable_data *d = data;
9151 struct task_group *child;
9152 unsigned long total, sum = 0;
9153 u64 period, runtime;
9155 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9156 runtime = tg->rt_bandwidth.rt_runtime;
9159 period = d->rt_period;
9160 runtime = d->rt_runtime;
9163 #ifdef CONFIG_USER_SCHED
9164 if (tg == &root_task_group) {
9165 period = global_rt_period();
9166 runtime = global_rt_runtime();
9171 * Cannot have more runtime than the period.
9173 if (runtime > period && runtime != RUNTIME_INF)
9177 * Ensure we don't starve existing RT tasks.
9179 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9182 total = to_ratio(period, runtime);
9185 * Nobody can have more than the global setting allows.
9187 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9191 * The sum of our children's runtime should not exceed our own.
9193 list_for_each_entry_rcu(child, &tg->children, siblings) {
9194 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9195 runtime = child->rt_bandwidth.rt_runtime;
9197 if (child == d->tg) {
9198 period = d->rt_period;
9199 runtime = d->rt_runtime;
9202 sum += to_ratio(period, runtime);
9211 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9213 struct rt_schedulable_data data = {
9215 .rt_period = period,
9216 .rt_runtime = runtime,
9219 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9222 static int tg_set_bandwidth(struct task_group *tg,
9223 u64 rt_period, u64 rt_runtime)
9227 mutex_lock(&rt_constraints_mutex);
9228 read_lock(&tasklist_lock);
9229 err = __rt_schedulable(tg, rt_period, rt_runtime);
9233 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9234 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9235 tg->rt_bandwidth.rt_runtime = rt_runtime;
9237 for_each_possible_cpu(i) {
9238 struct rt_rq *rt_rq = tg->rt_rq[i];
9240 spin_lock(&rt_rq->rt_runtime_lock);
9241 rt_rq->rt_runtime = rt_runtime;
9242 spin_unlock(&rt_rq->rt_runtime_lock);
9244 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9246 read_unlock(&tasklist_lock);
9247 mutex_unlock(&rt_constraints_mutex);
9252 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9254 u64 rt_runtime, rt_period;
9256 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9257 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9258 if (rt_runtime_us < 0)
9259 rt_runtime = RUNTIME_INF;
9261 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9264 long sched_group_rt_runtime(struct task_group *tg)
9268 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9271 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9272 do_div(rt_runtime_us, NSEC_PER_USEC);
9273 return rt_runtime_us;
9276 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9278 u64 rt_runtime, rt_period;
9280 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9281 rt_runtime = tg->rt_bandwidth.rt_runtime;
9286 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9289 long sched_group_rt_period(struct task_group *tg)
9293 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9294 do_div(rt_period_us, NSEC_PER_USEC);
9295 return rt_period_us;
9298 static int sched_rt_global_constraints(void)
9300 u64 runtime, period;
9303 if (sysctl_sched_rt_period <= 0)
9306 runtime = global_rt_runtime();
9307 period = global_rt_period();
9310 * Sanity check on the sysctl variables.
9312 if (runtime > period && runtime != RUNTIME_INF)
9315 mutex_lock(&rt_constraints_mutex);
9316 read_lock(&tasklist_lock);
9317 ret = __rt_schedulable(NULL, 0, 0);
9318 read_unlock(&tasklist_lock);
9319 mutex_unlock(&rt_constraints_mutex);
9323 #else /* !CONFIG_RT_GROUP_SCHED */
9324 static int sched_rt_global_constraints(void)
9326 unsigned long flags;
9329 if (sysctl_sched_rt_period <= 0)
9332 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9333 for_each_possible_cpu(i) {
9334 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9336 spin_lock(&rt_rq->rt_runtime_lock);
9337 rt_rq->rt_runtime = global_rt_runtime();
9338 spin_unlock(&rt_rq->rt_runtime_lock);
9340 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9344 #endif /* CONFIG_RT_GROUP_SCHED */
9346 int sched_rt_handler(struct ctl_table *table, int write,
9347 struct file *filp, void __user *buffer, size_t *lenp,
9351 int old_period, old_runtime;
9352 static DEFINE_MUTEX(mutex);
9355 old_period = sysctl_sched_rt_period;
9356 old_runtime = sysctl_sched_rt_runtime;
9358 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9360 if (!ret && write) {
9361 ret = sched_rt_global_constraints();
9363 sysctl_sched_rt_period = old_period;
9364 sysctl_sched_rt_runtime = old_runtime;
9366 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9367 def_rt_bandwidth.rt_period =
9368 ns_to_ktime(global_rt_period());
9371 mutex_unlock(&mutex);
9376 #ifdef CONFIG_CGROUP_SCHED
9378 /* return corresponding task_group object of a cgroup */
9379 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9381 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9382 struct task_group, css);
9385 static struct cgroup_subsys_state *
9386 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9388 struct task_group *tg, *parent;
9390 if (!cgrp->parent) {
9391 /* This is early initialization for the top cgroup */
9392 return &init_task_group.css;
9395 parent = cgroup_tg(cgrp->parent);
9396 tg = sched_create_group(parent);
9398 return ERR_PTR(-ENOMEM);
9404 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9406 struct task_group *tg = cgroup_tg(cgrp);
9408 sched_destroy_group(tg);
9412 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9413 struct task_struct *tsk)
9415 #ifdef CONFIG_RT_GROUP_SCHED
9416 /* Don't accept realtime tasks when there is no way for them to run */
9417 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9420 /* We don't support RT-tasks being in separate groups */
9421 if (tsk->sched_class != &fair_sched_class)
9429 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9430 struct cgroup *old_cont, struct task_struct *tsk)
9432 sched_move_task(tsk);
9435 #ifdef CONFIG_FAIR_GROUP_SCHED
9436 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9439 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9442 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9444 struct task_group *tg = cgroup_tg(cgrp);
9446 return (u64) tg->shares;
9448 #endif /* CONFIG_FAIR_GROUP_SCHED */
9450 #ifdef CONFIG_RT_GROUP_SCHED
9451 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9454 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9457 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9459 return sched_group_rt_runtime(cgroup_tg(cgrp));
9462 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9465 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9468 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9470 return sched_group_rt_period(cgroup_tg(cgrp));
9472 #endif /* CONFIG_RT_GROUP_SCHED */
9474 static struct cftype cpu_files[] = {
9475 #ifdef CONFIG_FAIR_GROUP_SCHED
9478 .read_u64 = cpu_shares_read_u64,
9479 .write_u64 = cpu_shares_write_u64,
9482 #ifdef CONFIG_RT_GROUP_SCHED
9484 .name = "rt_runtime_us",
9485 .read_s64 = cpu_rt_runtime_read,
9486 .write_s64 = cpu_rt_runtime_write,
9489 .name = "rt_period_us",
9490 .read_u64 = cpu_rt_period_read_uint,
9491 .write_u64 = cpu_rt_period_write_uint,
9496 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9498 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9501 struct cgroup_subsys cpu_cgroup_subsys = {
9503 .create = cpu_cgroup_create,
9504 .destroy = cpu_cgroup_destroy,
9505 .can_attach = cpu_cgroup_can_attach,
9506 .attach = cpu_cgroup_attach,
9507 .populate = cpu_cgroup_populate,
9508 .subsys_id = cpu_cgroup_subsys_id,
9512 #endif /* CONFIG_CGROUP_SCHED */
9514 #ifdef CONFIG_CGROUP_CPUACCT
9517 * CPU accounting code for task groups.
9519 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9520 * (balbir@in.ibm.com).
9523 /* track cpu usage of a group of tasks and its child groups */
9525 struct cgroup_subsys_state css;
9526 /* cpuusage holds pointer to a u64-type object on every cpu */
9528 struct cpuacct *parent;
9531 struct cgroup_subsys cpuacct_subsys;
9533 /* return cpu accounting group corresponding to this container */
9534 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9536 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9537 struct cpuacct, css);
9540 /* return cpu accounting group to which this task belongs */
9541 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9543 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9544 struct cpuacct, css);
9547 /* create a new cpu accounting group */
9548 static struct cgroup_subsys_state *cpuacct_create(
9549 struct cgroup_subsys *ss, struct cgroup *cgrp)
9551 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9554 return ERR_PTR(-ENOMEM);
9556 ca->cpuusage = alloc_percpu(u64);
9557 if (!ca->cpuusage) {
9559 return ERR_PTR(-ENOMEM);
9563 ca->parent = cgroup_ca(cgrp->parent);
9568 /* destroy an existing cpu accounting group */
9570 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9572 struct cpuacct *ca = cgroup_ca(cgrp);
9574 free_percpu(ca->cpuusage);
9578 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9580 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9583 #ifndef CONFIG_64BIT
9585 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9587 spin_lock_irq(&cpu_rq(cpu)->lock);
9589 spin_unlock_irq(&cpu_rq(cpu)->lock);
9597 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9599 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9601 #ifndef CONFIG_64BIT
9603 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9605 spin_lock_irq(&cpu_rq(cpu)->lock);
9607 spin_unlock_irq(&cpu_rq(cpu)->lock);
9613 /* return total cpu usage (in nanoseconds) of a group */
9614 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9616 struct cpuacct *ca = cgroup_ca(cgrp);
9617 u64 totalcpuusage = 0;
9620 for_each_present_cpu(i)
9621 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9623 return totalcpuusage;
9626 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9629 struct cpuacct *ca = cgroup_ca(cgrp);
9638 for_each_present_cpu(i)
9639 cpuacct_cpuusage_write(ca, i, 0);
9645 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9648 struct cpuacct *ca = cgroup_ca(cgroup);
9652 for_each_present_cpu(i) {
9653 percpu = cpuacct_cpuusage_read(ca, i);
9654 seq_printf(m, "%llu ", (unsigned long long) percpu);
9656 seq_printf(m, "\n");
9660 static struct cftype files[] = {
9663 .read_u64 = cpuusage_read,
9664 .write_u64 = cpuusage_write,
9667 .name = "usage_percpu",
9668 .read_seq_string = cpuacct_percpu_seq_read,
9673 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9675 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9679 * charge this task's execution time to its accounting group.
9681 * called with rq->lock held.
9683 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9688 if (!cpuacct_subsys.active)
9691 cpu = task_cpu(tsk);
9694 for (; ca; ca = ca->parent) {
9695 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9696 *cpuusage += cputime;
9700 struct cgroup_subsys cpuacct_subsys = {
9702 .create = cpuacct_create,
9703 .destroy = cpuacct_destroy,
9704 .populate = cpuacct_populate,
9705 .subsys_id = cpuacct_subsys_id,
9707 #endif /* CONFIG_CGROUP_CPUACCT */