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
470 int highest_prio; /* highest queued rt task prio */
473 unsigned long rt_nr_migratory;
479 /* Nests inside the rq lock: */
480 spinlock_t rt_runtime_lock;
482 #ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted;
486 struct list_head leaf_rt_rq_list;
487 struct task_group *tg;
488 struct sched_rt_entity *rt_se;
495 * We add the notion of a root-domain which will be used to define per-domain
496 * variables. Each exclusive cpuset essentially defines an island domain by
497 * fully partitioning the member cpus from any other cpuset. Whenever a new
498 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask;
514 struct cpupri cpupri;
516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
518 * Preferred wake up cpu nominated by sched_mc balance that will be
519 * used when most cpus are idle in the system indicating overall very
520 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
522 unsigned int sched_mc_preferred_wakeup_cpu;
527 * By default the system creates a single root-domain with all cpus as
528 * members (mimicking the global state we have today).
530 static struct root_domain def_root_domain;
535 * This is the main, per-CPU runqueue data structure.
537 * Locking rule: those places that want to lock multiple runqueues
538 * (such as the load balancing or the thread migration code), lock
539 * acquire operations must be ordered by ascending &runqueue.
546 * nr_running and cpu_load should be in the same cacheline because
547 * remote CPUs use both these fields when doing load calculation.
549 unsigned long nr_running;
550 #define CPU_LOAD_IDX_MAX 5
551 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
552 unsigned char idle_at_tick;
554 unsigned long last_tick_seen;
555 unsigned char in_nohz_recently;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load;
559 unsigned long nr_load_updates;
565 #ifdef CONFIG_FAIR_GROUP_SCHED
566 /* list of leaf cfs_rq on this cpu: */
567 struct list_head leaf_cfs_rq_list;
569 #ifdef CONFIG_RT_GROUP_SCHED
570 struct list_head leaf_rt_rq_list;
574 * This is part of a global counter where only the total sum
575 * over all CPUs matters. A task can increase this counter on
576 * one CPU and if it got migrated afterwards it may decrease
577 * it on another CPU. Always updated under the runqueue lock:
579 unsigned long nr_uninterruptible;
581 struct task_struct *curr, *idle;
582 unsigned long next_balance;
583 struct mm_struct *prev_mm;
590 struct root_domain *rd;
591 struct sched_domain *sd;
593 /* For active balancing */
596 /* cpu of this runqueue: */
600 unsigned long avg_load_per_task;
602 struct task_struct *migration_thread;
603 struct list_head migration_queue;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending;
609 struct call_single_data hrtick_csd;
611 struct hrtimer hrtick_timer;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info;
617 unsigned long long rq_cpu_time;
618 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
620 /* sys_sched_yield() stats */
621 unsigned int yld_exp_empty;
622 unsigned int yld_act_empty;
623 unsigned int yld_both_empty;
624 unsigned int yld_count;
626 /* schedule() stats */
627 unsigned int sched_switch;
628 unsigned int sched_count;
629 unsigned int sched_goidle;
631 /* try_to_wake_up() stats */
632 unsigned int ttwu_count;
633 unsigned int ttwu_local;
636 unsigned int bkl_count;
640 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
642 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
644 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
647 static inline int cpu_of(struct rq *rq)
657 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
658 * See detach_destroy_domains: synchronize_sched for details.
660 * The domain tree of any CPU may only be accessed from within
661 * preempt-disabled sections.
663 #define for_each_domain(cpu, __sd) \
664 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
666 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
667 #define this_rq() (&__get_cpu_var(runqueues))
668 #define task_rq(p) cpu_rq(task_cpu(p))
669 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
671 static inline void update_rq_clock(struct rq *rq)
673 rq->clock = sched_clock_cpu(cpu_of(rq));
677 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
679 #ifdef CONFIG_SCHED_DEBUG
680 # define const_debug __read_mostly
682 # define const_debug static const
688 * Returns true if the current cpu runqueue is locked.
689 * This interface allows printk to be called with the runqueue lock
690 * held and know whether or not it is OK to wake up the klogd.
692 int runqueue_is_locked(void)
695 struct rq *rq = cpu_rq(cpu);
698 ret = spin_is_locked(&rq->lock);
704 * Debugging: various feature bits
707 #define SCHED_FEAT(name, enabled) \
708 __SCHED_FEAT_##name ,
711 #include "sched_features.h"
716 #define SCHED_FEAT(name, enabled) \
717 (1UL << __SCHED_FEAT_##name) * enabled |
719 const_debug unsigned int sysctl_sched_features =
720 #include "sched_features.h"
725 #ifdef CONFIG_SCHED_DEBUG
726 #define SCHED_FEAT(name, enabled) \
729 static __read_mostly char *sched_feat_names[] = {
730 #include "sched_features.h"
736 static int sched_feat_show(struct seq_file *m, void *v)
740 for (i = 0; sched_feat_names[i]; i++) {
741 if (!(sysctl_sched_features & (1UL << i)))
743 seq_printf(m, "%s ", sched_feat_names[i]);
751 sched_feat_write(struct file *filp, const char __user *ubuf,
752 size_t cnt, loff_t *ppos)
762 if (copy_from_user(&buf, ubuf, cnt))
767 if (strncmp(buf, "NO_", 3) == 0) {
772 for (i = 0; sched_feat_names[i]; i++) {
773 int len = strlen(sched_feat_names[i]);
775 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
777 sysctl_sched_features &= ~(1UL << i);
779 sysctl_sched_features |= (1UL << i);
784 if (!sched_feat_names[i])
792 static int sched_feat_open(struct inode *inode, struct file *filp)
794 return single_open(filp, sched_feat_show, NULL);
797 static struct file_operations sched_feat_fops = {
798 .open = sched_feat_open,
799 .write = sched_feat_write,
802 .release = single_release,
805 static __init int sched_init_debug(void)
807 debugfs_create_file("sched_features", 0644, NULL, NULL,
812 late_initcall(sched_init_debug);
816 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
819 * Number of tasks to iterate in a single balance run.
820 * Limited because this is done with IRQs disabled.
822 const_debug unsigned int sysctl_sched_nr_migrate = 32;
825 * ratelimit for updating the group shares.
828 unsigned int sysctl_sched_shares_ratelimit = 250000;
831 * Inject some fuzzyness into changing the per-cpu group shares
832 * this avoids remote rq-locks at the expense of fairness.
835 unsigned int sysctl_sched_shares_thresh = 4;
838 * period over which we measure -rt task cpu usage in us.
841 unsigned int sysctl_sched_rt_period = 1000000;
843 static __read_mostly int scheduler_running;
846 * part of the period that we allow rt tasks to run in us.
849 int sysctl_sched_rt_runtime = 950000;
851 static inline u64 global_rt_period(void)
853 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
856 static inline u64 global_rt_runtime(void)
858 if (sysctl_sched_rt_runtime < 0)
861 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
864 #ifndef prepare_arch_switch
865 # define prepare_arch_switch(next) do { } while (0)
867 #ifndef finish_arch_switch
868 # define finish_arch_switch(prev) do { } while (0)
871 static inline int task_current(struct rq *rq, struct task_struct *p)
873 return rq->curr == p;
876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
877 static inline int task_running(struct rq *rq, struct task_struct *p)
879 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
888 #ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq->lock.owner = current;
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
897 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
899 spin_unlock_irq(&rq->lock);
902 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
903 static inline int task_running(struct rq *rq, struct task_struct *p)
908 return task_current(rq, p);
912 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
922 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq->lock);
925 spin_unlock(&rq->lock);
929 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
940 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
944 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
950 static inline struct rq *__task_rq_lock(struct task_struct *p)
954 struct rq *rq = task_rq(p);
955 spin_lock(&rq->lock);
956 if (likely(rq == task_rq(p)))
958 spin_unlock(&rq->lock);
963 * task_rq_lock - lock the runqueue a given task resides on and disable
964 * interrupts. Note the ordering: we can safely lookup the task_rq without
965 * explicitly disabling preemption.
967 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
973 local_irq_save(*flags);
975 spin_lock(&rq->lock);
976 if (likely(rq == task_rq(p)))
978 spin_unlock_irqrestore(&rq->lock, *flags);
982 void task_rq_unlock_wait(struct task_struct *p)
984 struct rq *rq = task_rq(p);
986 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
987 spin_unlock_wait(&rq->lock);
990 static void __task_rq_unlock(struct rq *rq)
993 spin_unlock(&rq->lock);
996 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
999 spin_unlock_irqrestore(&rq->lock, *flags);
1003 * this_rq_lock - lock this runqueue and disable interrupts.
1005 static struct rq *this_rq_lock(void)
1006 __acquires(rq->lock)
1010 local_irq_disable();
1012 spin_lock(&rq->lock);
1017 #ifdef CONFIG_SCHED_HRTICK
1019 * Use HR-timers to deliver accurate preemption points.
1021 * Its all a bit involved since we cannot program an hrt while holding the
1022 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1025 * When we get rescheduled we reprogram the hrtick_timer outside of the
1031 * - enabled by features
1032 * - hrtimer is actually high res
1034 static inline int hrtick_enabled(struct rq *rq)
1036 if (!sched_feat(HRTICK))
1038 if (!cpu_active(cpu_of(rq)))
1040 return hrtimer_is_hres_active(&rq->hrtick_timer);
1043 static void hrtick_clear(struct rq *rq)
1045 if (hrtimer_active(&rq->hrtick_timer))
1046 hrtimer_cancel(&rq->hrtick_timer);
1050 * High-resolution timer tick.
1051 * Runs from hardirq context with interrupts disabled.
1053 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1055 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1057 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1059 spin_lock(&rq->lock);
1060 update_rq_clock(rq);
1061 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1062 spin_unlock(&rq->lock);
1064 return HRTIMER_NORESTART;
1069 * called from hardirq (IPI) context
1071 static void __hrtick_start(void *arg)
1073 struct rq *rq = arg;
1075 spin_lock(&rq->lock);
1076 hrtimer_restart(&rq->hrtick_timer);
1077 rq->hrtick_csd_pending = 0;
1078 spin_unlock(&rq->lock);
1082 * Called to set the hrtick timer state.
1084 * called with rq->lock held and irqs disabled
1086 static void hrtick_start(struct rq *rq, u64 delay)
1088 struct hrtimer *timer = &rq->hrtick_timer;
1089 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1091 hrtimer_set_expires(timer, time);
1093 if (rq == this_rq()) {
1094 hrtimer_restart(timer);
1095 } else if (!rq->hrtick_csd_pending) {
1096 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1097 rq->hrtick_csd_pending = 1;
1102 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1104 int cpu = (int)(long)hcpu;
1107 case CPU_UP_CANCELED:
1108 case CPU_UP_CANCELED_FROZEN:
1109 case CPU_DOWN_PREPARE:
1110 case CPU_DOWN_PREPARE_FROZEN:
1112 case CPU_DEAD_FROZEN:
1113 hrtick_clear(cpu_rq(cpu));
1120 static __init void init_hrtick(void)
1122 hotcpu_notifier(hotplug_hrtick, 0);
1126 * Called to set the hrtick timer state.
1128 * called with rq->lock held and irqs disabled
1130 static void hrtick_start(struct rq *rq, u64 delay)
1132 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq *rq)
1143 rq->hrtick_csd_pending = 0;
1145 rq->hrtick_csd.flags = 0;
1146 rq->hrtick_csd.func = __hrtick_start;
1147 rq->hrtick_csd.info = rq;
1150 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1151 rq->hrtick_timer.function = hrtick;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq *rq)
1158 static inline void init_rq_hrtick(struct rq *rq)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct *p)
1184 assert_spin_locked(&task_rq(p)->lock);
1186 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1189 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1192 if (cpu == smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p))
1198 smp_send_reschedule(cpu);
1201 static void resched_cpu(int cpu)
1203 struct rq *rq = cpu_rq(cpu);
1204 unsigned long flags;
1206 if (!spin_trylock_irqsave(&rq->lock, flags))
1208 resched_task(cpu_curr(cpu));
1209 spin_unlock_irqrestore(&rq->lock, flags);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1227 if (cpu == smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq->curr != rq->idle)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq->idle))
1250 smp_send_reschedule(cpu);
1252 #endif /* CONFIG_NO_HZ */
1254 #else /* !CONFIG_SMP */
1255 static void resched_task(struct task_struct *p)
1257 assert_spin_locked(&task_rq(p)->lock);
1258 set_tsk_need_resched(p);
1260 #endif /* CONFIG_SMP */
1262 #if BITS_PER_LONG == 32
1263 # define WMULT_CONST (~0UL)
1265 # define WMULT_CONST (1UL << 32)
1268 #define WMULT_SHIFT 32
1271 * Shift right and round:
1273 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1276 * delta *= weight / lw
1278 static unsigned long
1279 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1280 struct load_weight *lw)
1284 if (!lw->inv_weight) {
1285 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1288 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1292 tmp = (u64)delta_exec * weight;
1294 * Check whether we'd overflow the 64-bit multiplication:
1296 if (unlikely(tmp > WMULT_CONST))
1297 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1300 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1302 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1305 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1311 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1318 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1319 * of tasks with abnormal "nice" values across CPUs the contribution that
1320 * each task makes to its run queue's load is weighted according to its
1321 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1322 * scaled version of the new time slice allocation that they receive on time
1326 #define WEIGHT_IDLEPRIO 3
1327 #define WMULT_IDLEPRIO 1431655765
1330 * Nice levels are multiplicative, with a gentle 10% change for every
1331 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1332 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1333 * that remained on nice 0.
1335 * The "10% effect" is relative and cumulative: from _any_ nice level,
1336 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1337 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1338 * If a task goes up by ~10% and another task goes down by ~10% then
1339 * the relative distance between them is ~25%.)
1341 static const int prio_to_weight[40] = {
1342 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1343 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1344 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1345 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1346 /* 0 */ 1024, 820, 655, 526, 423,
1347 /* 5 */ 335, 272, 215, 172, 137,
1348 /* 10 */ 110, 87, 70, 56, 45,
1349 /* 15 */ 36, 29, 23, 18, 15,
1353 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1355 * In cases where the weight does not change often, we can use the
1356 * precalculated inverse to speed up arithmetics by turning divisions
1357 * into multiplications:
1359 static const u32 prio_to_wmult[40] = {
1360 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1361 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1362 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1363 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1364 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1365 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1366 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1367 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1370 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1373 * runqueue iterator, to support SMP load-balancing between different
1374 * scheduling classes, without having to expose their internal data
1375 * structures to the load-balancing proper:
1377 struct rq_iterator {
1379 struct task_struct *(*start)(void *);
1380 struct task_struct *(*next)(void *);
1384 static unsigned long
1385 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1386 unsigned long max_load_move, struct sched_domain *sd,
1387 enum cpu_idle_type idle, int *all_pinned,
1388 int *this_best_prio, struct rq_iterator *iterator);
1391 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1392 struct sched_domain *sd, enum cpu_idle_type idle,
1393 struct rq_iterator *iterator);
1396 #ifdef CONFIG_CGROUP_CPUACCT
1397 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1399 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1402 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1404 update_load_add(&rq->load, load);
1407 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1409 update_load_sub(&rq->load, load);
1412 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1413 typedef int (*tg_visitor)(struct task_group *, void *);
1416 * Iterate the full tree, calling @down when first entering a node and @up when
1417 * leaving it for the final time.
1419 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1421 struct task_group *parent, *child;
1425 parent = &root_task_group;
1427 ret = (*down)(parent, data);
1430 list_for_each_entry_rcu(child, &parent->children, siblings) {
1437 ret = (*up)(parent, data);
1442 parent = parent->parent;
1451 static int tg_nop(struct task_group *tg, void *data)
1458 static unsigned long source_load(int cpu, int type);
1459 static unsigned long target_load(int cpu, int type);
1460 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1462 static unsigned long cpu_avg_load_per_task(int cpu)
1464 struct rq *rq = cpu_rq(cpu);
1465 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1468 rq->avg_load_per_task = rq->load.weight / nr_running;
1470 rq->avg_load_per_task = 0;
1472 return rq->avg_load_per_task;
1475 #ifdef CONFIG_FAIR_GROUP_SCHED
1477 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1480 * Calculate and set the cpu's group shares.
1483 update_group_shares_cpu(struct task_group *tg, int cpu,
1484 unsigned long sd_shares, unsigned long sd_rq_weight)
1486 unsigned long shares;
1487 unsigned long rq_weight;
1492 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1495 * \Sum shares * rq_weight
1496 * shares = -----------------------
1500 shares = (sd_shares * rq_weight) / sd_rq_weight;
1501 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1503 if (abs(shares - tg->se[cpu]->load.weight) >
1504 sysctl_sched_shares_thresh) {
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long flags;
1508 spin_lock_irqsave(&rq->lock, flags);
1509 tg->cfs_rq[cpu]->shares = shares;
1511 __set_se_shares(tg->se[cpu], shares);
1512 spin_unlock_irqrestore(&rq->lock, flags);
1517 * Re-compute the task group their per cpu shares over the given domain.
1518 * This needs to be done in a bottom-up fashion because the rq weight of a
1519 * parent group depends on the shares of its child groups.
1521 static int tg_shares_up(struct task_group *tg, void *data)
1523 unsigned long weight, rq_weight = 0;
1524 unsigned long shares = 0;
1525 struct sched_domain *sd = data;
1528 for_each_cpu(i, sched_domain_span(sd)) {
1530 * If there are currently no tasks on the cpu pretend there
1531 * is one of average load so that when a new task gets to
1532 * run here it will not get delayed by group starvation.
1534 weight = tg->cfs_rq[i]->load.weight;
1536 weight = NICE_0_LOAD;
1538 tg->cfs_rq[i]->rq_weight = weight;
1539 rq_weight += weight;
1540 shares += tg->cfs_rq[i]->shares;
1543 if ((!shares && rq_weight) || shares > tg->shares)
1544 shares = tg->shares;
1546 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1547 shares = tg->shares;
1549 for_each_cpu(i, sched_domain_span(sd))
1550 update_group_shares_cpu(tg, i, shares, rq_weight);
1556 * Compute the cpu's hierarchical load factor for each task group.
1557 * This needs to be done in a top-down fashion because the load of a child
1558 * group is a fraction of its parents load.
1560 static int tg_load_down(struct task_group *tg, void *data)
1563 long cpu = (long)data;
1566 load = cpu_rq(cpu)->load.weight;
1568 load = tg->parent->cfs_rq[cpu]->h_load;
1569 load *= tg->cfs_rq[cpu]->shares;
1570 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1573 tg->cfs_rq[cpu]->h_load = load;
1578 static void update_shares(struct sched_domain *sd)
1580 u64 now = cpu_clock(raw_smp_processor_id());
1581 s64 elapsed = now - sd->last_update;
1583 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1584 sd->last_update = now;
1585 walk_tg_tree(tg_nop, tg_shares_up, sd);
1589 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1591 spin_unlock(&rq->lock);
1593 spin_lock(&rq->lock);
1596 static void update_h_load(long cpu)
1598 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1603 static inline void update_shares(struct sched_domain *sd)
1607 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1614 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1616 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1617 __releases(this_rq->lock)
1618 __acquires(busiest->lock)
1619 __acquires(this_rq->lock)
1623 if (unlikely(!irqs_disabled())) {
1624 /* printk() doesn't work good under rq->lock */
1625 spin_unlock(&this_rq->lock);
1628 if (unlikely(!spin_trylock(&busiest->lock))) {
1629 if (busiest < this_rq) {
1630 spin_unlock(&this_rq->lock);
1631 spin_lock(&busiest->lock);
1632 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1635 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1640 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1641 __releases(busiest->lock)
1643 spin_unlock(&busiest->lock);
1644 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1648 #ifdef CONFIG_FAIR_GROUP_SCHED
1649 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1652 cfs_rq->shares = shares;
1657 #include "sched_stats.h"
1658 #include "sched_idletask.c"
1659 #include "sched_fair.c"
1660 #include "sched_rt.c"
1661 #ifdef CONFIG_SCHED_DEBUG
1662 # include "sched_debug.c"
1665 #define sched_class_highest (&rt_sched_class)
1666 #define for_each_class(class) \
1667 for (class = sched_class_highest; class; class = class->next)
1669 static void inc_nr_running(struct rq *rq)
1674 static void dec_nr_running(struct rq *rq)
1679 static void set_load_weight(struct task_struct *p)
1681 if (task_has_rt_policy(p)) {
1682 p->se.load.weight = prio_to_weight[0] * 2;
1683 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1688 * SCHED_IDLE tasks get minimal weight:
1690 if (p->policy == SCHED_IDLE) {
1691 p->se.load.weight = WEIGHT_IDLEPRIO;
1692 p->se.load.inv_weight = WMULT_IDLEPRIO;
1696 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1697 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1700 static void update_avg(u64 *avg, u64 sample)
1702 s64 diff = sample - *avg;
1706 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1708 sched_info_queued(p);
1709 p->sched_class->enqueue_task(rq, p, wakeup);
1713 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1715 if (sleep && p->se.last_wakeup) {
1716 update_avg(&p->se.avg_overlap,
1717 p->se.sum_exec_runtime - p->se.last_wakeup);
1718 p->se.last_wakeup = 0;
1721 sched_info_dequeued(p);
1722 p->sched_class->dequeue_task(rq, p, sleep);
1727 * __normal_prio - return the priority that is based on the static prio
1729 static inline int __normal_prio(struct task_struct *p)
1731 return p->static_prio;
1735 * Calculate the expected normal priority: i.e. priority
1736 * without taking RT-inheritance into account. Might be
1737 * boosted by interactivity modifiers. Changes upon fork,
1738 * setprio syscalls, and whenever the interactivity
1739 * estimator recalculates.
1741 static inline int normal_prio(struct task_struct *p)
1745 if (task_has_rt_policy(p))
1746 prio = MAX_RT_PRIO-1 - p->rt_priority;
1748 prio = __normal_prio(p);
1753 * Calculate the current priority, i.e. the priority
1754 * taken into account by the scheduler. This value might
1755 * be boosted by RT tasks, or might be boosted by
1756 * interactivity modifiers. Will be RT if the task got
1757 * RT-boosted. If not then it returns p->normal_prio.
1759 static int effective_prio(struct task_struct *p)
1761 p->normal_prio = normal_prio(p);
1763 * If we are RT tasks or we were boosted to RT priority,
1764 * keep the priority unchanged. Otherwise, update priority
1765 * to the normal priority:
1767 if (!rt_prio(p->prio))
1768 return p->normal_prio;
1773 * activate_task - move a task to the runqueue.
1775 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1777 if (task_contributes_to_load(p))
1778 rq->nr_uninterruptible--;
1780 enqueue_task(rq, p, wakeup);
1785 * deactivate_task - remove a task from the runqueue.
1787 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1789 if (task_contributes_to_load(p))
1790 rq->nr_uninterruptible++;
1792 dequeue_task(rq, p, sleep);
1797 * task_curr - is this task currently executing on a CPU?
1798 * @p: the task in question.
1800 inline int task_curr(const struct task_struct *p)
1802 return cpu_curr(task_cpu(p)) == p;
1805 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1807 set_task_rq(p, cpu);
1810 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1811 * successfuly executed on another CPU. We must ensure that updates of
1812 * per-task data have been completed by this moment.
1815 task_thread_info(p)->cpu = cpu;
1819 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1820 const struct sched_class *prev_class,
1821 int oldprio, int running)
1823 if (prev_class != p->sched_class) {
1824 if (prev_class->switched_from)
1825 prev_class->switched_from(rq, p, running);
1826 p->sched_class->switched_to(rq, p, running);
1828 p->sched_class->prio_changed(rq, p, oldprio, running);
1833 /* Used instead of source_load when we know the type == 0 */
1834 static unsigned long weighted_cpuload(const int cpu)
1836 return cpu_rq(cpu)->load.weight;
1840 * Is this task likely cache-hot:
1843 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1848 * Buddy candidates are cache hot:
1850 if (sched_feat(CACHE_HOT_BUDDY) &&
1851 (&p->se == cfs_rq_of(&p->se)->next ||
1852 &p->se == cfs_rq_of(&p->se)->last))
1855 if (p->sched_class != &fair_sched_class)
1858 if (sysctl_sched_migration_cost == -1)
1860 if (sysctl_sched_migration_cost == 0)
1863 delta = now - p->se.exec_start;
1865 return delta < (s64)sysctl_sched_migration_cost;
1869 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1871 int old_cpu = task_cpu(p);
1872 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1873 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1874 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1877 clock_offset = old_rq->clock - new_rq->clock;
1879 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1881 #ifdef CONFIG_SCHEDSTATS
1882 if (p->se.wait_start)
1883 p->se.wait_start -= clock_offset;
1884 if (p->se.sleep_start)
1885 p->se.sleep_start -= clock_offset;
1886 if (p->se.block_start)
1887 p->se.block_start -= clock_offset;
1888 if (old_cpu != new_cpu) {
1889 schedstat_inc(p, se.nr_migrations);
1890 if (task_hot(p, old_rq->clock, NULL))
1891 schedstat_inc(p, se.nr_forced2_migrations);
1894 p->se.vruntime -= old_cfsrq->min_vruntime -
1895 new_cfsrq->min_vruntime;
1897 __set_task_cpu(p, new_cpu);
1900 struct migration_req {
1901 struct list_head list;
1903 struct task_struct *task;
1906 struct completion done;
1910 * The task's runqueue lock must be held.
1911 * Returns true if you have to wait for migration thread.
1914 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1916 struct rq *rq = task_rq(p);
1919 * If the task is not on a runqueue (and not running), then
1920 * it is sufficient to simply update the task's cpu field.
1922 if (!p->se.on_rq && !task_running(rq, p)) {
1923 set_task_cpu(p, dest_cpu);
1927 init_completion(&req->done);
1929 req->dest_cpu = dest_cpu;
1930 list_add(&req->list, &rq->migration_queue);
1936 * wait_task_inactive - wait for a thread to unschedule.
1938 * If @match_state is nonzero, it's the @p->state value just checked and
1939 * not expected to change. If it changes, i.e. @p might have woken up,
1940 * then return zero. When we succeed in waiting for @p to be off its CPU,
1941 * we return a positive number (its total switch count). If a second call
1942 * a short while later returns the same number, the caller can be sure that
1943 * @p has remained unscheduled the whole time.
1945 * The caller must ensure that the task *will* unschedule sometime soon,
1946 * else this function might spin for a *long* time. This function can't
1947 * be called with interrupts off, or it may introduce deadlock with
1948 * smp_call_function() if an IPI is sent by the same process we are
1949 * waiting to become inactive.
1951 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1953 unsigned long flags;
1960 * We do the initial early heuristics without holding
1961 * any task-queue locks at all. We'll only try to get
1962 * the runqueue lock when things look like they will
1968 * If the task is actively running on another CPU
1969 * still, just relax and busy-wait without holding
1972 * NOTE! Since we don't hold any locks, it's not
1973 * even sure that "rq" stays as the right runqueue!
1974 * But we don't care, since "task_running()" will
1975 * return false if the runqueue has changed and p
1976 * is actually now running somewhere else!
1978 while (task_running(rq, p)) {
1979 if (match_state && unlikely(p->state != match_state))
1985 * Ok, time to look more closely! We need the rq
1986 * lock now, to be *sure*. If we're wrong, we'll
1987 * just go back and repeat.
1989 rq = task_rq_lock(p, &flags);
1990 trace_sched_wait_task(rq, p);
1991 running = task_running(rq, p);
1992 on_rq = p->se.on_rq;
1994 if (!match_state || p->state == match_state)
1995 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1996 task_rq_unlock(rq, &flags);
1999 * If it changed from the expected state, bail out now.
2001 if (unlikely(!ncsw))
2005 * Was it really running after all now that we
2006 * checked with the proper locks actually held?
2008 * Oops. Go back and try again..
2010 if (unlikely(running)) {
2016 * It's not enough that it's not actively running,
2017 * it must be off the runqueue _entirely_, and not
2020 * So if it wa still runnable (but just not actively
2021 * running right now), it's preempted, and we should
2022 * yield - it could be a while.
2024 if (unlikely(on_rq)) {
2025 schedule_timeout_uninterruptible(1);
2030 * Ahh, all good. It wasn't running, and it wasn't
2031 * runnable, which means that it will never become
2032 * running in the future either. We're all done!
2041 * kick_process - kick a running thread to enter/exit the kernel
2042 * @p: the to-be-kicked thread
2044 * Cause a process which is running on another CPU to enter
2045 * kernel-mode, without any delay. (to get signals handled.)
2047 * NOTE: this function doesnt have to take the runqueue lock,
2048 * because all it wants to ensure is that the remote task enters
2049 * the kernel. If the IPI races and the task has been migrated
2050 * to another CPU then no harm is done and the purpose has been
2053 void kick_process(struct task_struct *p)
2059 if ((cpu != smp_processor_id()) && task_curr(p))
2060 smp_send_reschedule(cpu);
2065 * Return a low guess at the load of a migration-source cpu weighted
2066 * according to the scheduling class and "nice" value.
2068 * We want to under-estimate the load of migration sources, to
2069 * balance conservatively.
2071 static unsigned long source_load(int cpu, int type)
2073 struct rq *rq = cpu_rq(cpu);
2074 unsigned long total = weighted_cpuload(cpu);
2076 if (type == 0 || !sched_feat(LB_BIAS))
2079 return min(rq->cpu_load[type-1], total);
2083 * Return a high guess at the load of a migration-target cpu weighted
2084 * according to the scheduling class and "nice" value.
2086 static unsigned long target_load(int cpu, int type)
2088 struct rq *rq = cpu_rq(cpu);
2089 unsigned long total = weighted_cpuload(cpu);
2091 if (type == 0 || !sched_feat(LB_BIAS))
2094 return max(rq->cpu_load[type-1], total);
2098 * find_idlest_group finds and returns the least busy CPU group within the
2101 static struct sched_group *
2102 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2104 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2105 unsigned long min_load = ULONG_MAX, this_load = 0;
2106 int load_idx = sd->forkexec_idx;
2107 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2110 unsigned long load, avg_load;
2114 /* Skip over this group if it has no CPUs allowed */
2115 if (!cpumask_intersects(sched_group_cpus(group),
2119 local_group = cpumask_test_cpu(this_cpu,
2120 sched_group_cpus(group));
2122 /* Tally up the load of all CPUs in the group */
2125 for_each_cpu(i, sched_group_cpus(group)) {
2126 /* Bias balancing toward cpus of our domain */
2128 load = source_load(i, load_idx);
2130 load = target_load(i, load_idx);
2135 /* Adjust by relative CPU power of the group */
2136 avg_load = sg_div_cpu_power(group,
2137 avg_load * SCHED_LOAD_SCALE);
2140 this_load = avg_load;
2142 } else if (avg_load < min_load) {
2143 min_load = avg_load;
2146 } while (group = group->next, group != sd->groups);
2148 if (!idlest || 100*this_load < imbalance*min_load)
2154 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2157 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2159 unsigned long load, min_load = ULONG_MAX;
2163 /* Traverse only the allowed CPUs */
2164 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2165 load = weighted_cpuload(i);
2167 if (load < min_load || (load == min_load && i == this_cpu)) {
2177 * sched_balance_self: balance the current task (running on cpu) in domains
2178 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2181 * Balance, ie. select the least loaded group.
2183 * Returns the target CPU number, or the same CPU if no balancing is needed.
2185 * preempt must be disabled.
2187 static int sched_balance_self(int cpu, int flag)
2189 struct task_struct *t = current;
2190 struct sched_domain *tmp, *sd = NULL;
2192 for_each_domain(cpu, tmp) {
2194 * If power savings logic is enabled for a domain, stop there.
2196 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2198 if (tmp->flags & flag)
2206 struct sched_group *group;
2207 int new_cpu, weight;
2209 if (!(sd->flags & flag)) {
2214 group = find_idlest_group(sd, t, cpu);
2220 new_cpu = find_idlest_cpu(group, t, cpu);
2221 if (new_cpu == -1 || new_cpu == cpu) {
2222 /* Now try balancing at a lower domain level of cpu */
2227 /* Now try balancing at a lower domain level of new_cpu */
2229 weight = cpumask_weight(sched_domain_span(sd));
2231 for_each_domain(cpu, tmp) {
2232 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2234 if (tmp->flags & flag)
2237 /* while loop will break here if sd == NULL */
2243 #endif /* CONFIG_SMP */
2246 * try_to_wake_up - wake up a thread
2247 * @p: the to-be-woken-up thread
2248 * @state: the mask of task states that can be woken
2249 * @sync: do a synchronous wakeup?
2251 * Put it on the run-queue if it's not already there. The "current"
2252 * thread is always on the run-queue (except when the actual
2253 * re-schedule is in progress), and as such you're allowed to do
2254 * the simpler "current->state = TASK_RUNNING" to mark yourself
2255 * runnable without the overhead of this.
2257 * returns failure only if the task is already active.
2259 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2261 int cpu, orig_cpu, this_cpu, success = 0;
2262 unsigned long flags;
2266 if (!sched_feat(SYNC_WAKEUPS))
2269 if (!sync && (current->se.avg_overlap < sysctl_sched_migration_cost &&
2270 p->se.avg_overlap < sysctl_sched_migration_cost))
2274 if (sched_feat(LB_WAKEUP_UPDATE)) {
2275 struct sched_domain *sd;
2277 this_cpu = raw_smp_processor_id();
2280 for_each_domain(this_cpu, sd) {
2281 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2290 rq = task_rq_lock(p, &flags);
2291 update_rq_clock(rq);
2292 old_state = p->state;
2293 if (!(old_state & state))
2301 this_cpu = smp_processor_id();
2304 if (unlikely(task_running(rq, p)))
2307 cpu = p->sched_class->select_task_rq(p, sync);
2308 if (cpu != orig_cpu) {
2309 set_task_cpu(p, cpu);
2310 task_rq_unlock(rq, &flags);
2311 /* might preempt at this point */
2312 rq = task_rq_lock(p, &flags);
2313 old_state = p->state;
2314 if (!(old_state & state))
2319 this_cpu = smp_processor_id();
2323 #ifdef CONFIG_SCHEDSTATS
2324 schedstat_inc(rq, ttwu_count);
2325 if (cpu == this_cpu)
2326 schedstat_inc(rq, ttwu_local);
2328 struct sched_domain *sd;
2329 for_each_domain(this_cpu, sd) {
2330 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2331 schedstat_inc(sd, ttwu_wake_remote);
2336 #endif /* CONFIG_SCHEDSTATS */
2339 #endif /* CONFIG_SMP */
2340 schedstat_inc(p, se.nr_wakeups);
2342 schedstat_inc(p, se.nr_wakeups_sync);
2343 if (orig_cpu != cpu)
2344 schedstat_inc(p, se.nr_wakeups_migrate);
2345 if (cpu == this_cpu)
2346 schedstat_inc(p, se.nr_wakeups_local);
2348 schedstat_inc(p, se.nr_wakeups_remote);
2349 activate_task(rq, p, 1);
2353 trace_sched_wakeup(rq, p, success);
2354 check_preempt_curr(rq, p, sync);
2356 p->state = TASK_RUNNING;
2358 if (p->sched_class->task_wake_up)
2359 p->sched_class->task_wake_up(rq, p);
2362 current->se.last_wakeup = current->se.sum_exec_runtime;
2364 task_rq_unlock(rq, &flags);
2369 int wake_up_process(struct task_struct *p)
2371 return try_to_wake_up(p, TASK_ALL, 0);
2373 EXPORT_SYMBOL(wake_up_process);
2375 int wake_up_state(struct task_struct *p, unsigned int state)
2377 return try_to_wake_up(p, state, 0);
2381 * Perform scheduler related setup for a newly forked process p.
2382 * p is forked by current.
2384 * __sched_fork() is basic setup used by init_idle() too:
2386 static void __sched_fork(struct task_struct *p)
2388 p->se.exec_start = 0;
2389 p->se.sum_exec_runtime = 0;
2390 p->se.prev_sum_exec_runtime = 0;
2391 p->se.last_wakeup = 0;
2392 p->se.avg_overlap = 0;
2394 #ifdef CONFIG_SCHEDSTATS
2395 p->se.wait_start = 0;
2396 p->se.sum_sleep_runtime = 0;
2397 p->se.sleep_start = 0;
2398 p->se.block_start = 0;
2399 p->se.sleep_max = 0;
2400 p->se.block_max = 0;
2402 p->se.slice_max = 0;
2406 INIT_LIST_HEAD(&p->rt.run_list);
2408 INIT_LIST_HEAD(&p->se.group_node);
2410 #ifdef CONFIG_PREEMPT_NOTIFIERS
2411 INIT_HLIST_HEAD(&p->preempt_notifiers);
2415 * We mark the process as running here, but have not actually
2416 * inserted it onto the runqueue yet. This guarantees that
2417 * nobody will actually run it, and a signal or other external
2418 * event cannot wake it up and insert it on the runqueue either.
2420 p->state = TASK_RUNNING;
2424 * fork()/clone()-time setup:
2426 void sched_fork(struct task_struct *p, int clone_flags)
2428 int cpu = get_cpu();
2433 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2435 set_task_cpu(p, cpu);
2438 * Make sure we do not leak PI boosting priority to the child:
2440 p->prio = current->normal_prio;
2441 if (!rt_prio(p->prio))
2442 p->sched_class = &fair_sched_class;
2444 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2445 if (likely(sched_info_on()))
2446 memset(&p->sched_info, 0, sizeof(p->sched_info));
2448 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2451 #ifdef CONFIG_PREEMPT
2452 /* Want to start with kernel preemption disabled. */
2453 task_thread_info(p)->preempt_count = 1;
2459 * wake_up_new_task - wake up a newly created task for the first time.
2461 * This function will do some initial scheduler statistics housekeeping
2462 * that must be done for every newly created context, then puts the task
2463 * on the runqueue and wakes it.
2465 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2467 unsigned long flags;
2470 rq = task_rq_lock(p, &flags);
2471 BUG_ON(p->state != TASK_RUNNING);
2472 update_rq_clock(rq);
2474 p->prio = effective_prio(p);
2476 if (!p->sched_class->task_new || !current->se.on_rq) {
2477 activate_task(rq, p, 0);
2480 * Let the scheduling class do new task startup
2481 * management (if any):
2483 p->sched_class->task_new(rq, p);
2486 trace_sched_wakeup_new(rq, p, 1);
2487 check_preempt_curr(rq, p, 0);
2489 if (p->sched_class->task_wake_up)
2490 p->sched_class->task_wake_up(rq, p);
2492 task_rq_unlock(rq, &flags);
2495 #ifdef CONFIG_PREEMPT_NOTIFIERS
2498 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2499 * @notifier: notifier struct to register
2501 void preempt_notifier_register(struct preempt_notifier *notifier)
2503 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2505 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2508 * preempt_notifier_unregister - no longer interested in preemption notifications
2509 * @notifier: notifier struct to unregister
2511 * This is safe to call from within a preemption notifier.
2513 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2515 hlist_del(¬ifier->link);
2517 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2519 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2521 struct preempt_notifier *notifier;
2522 struct hlist_node *node;
2524 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2525 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2529 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2530 struct task_struct *next)
2532 struct preempt_notifier *notifier;
2533 struct hlist_node *node;
2535 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2536 notifier->ops->sched_out(notifier, next);
2539 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2541 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2546 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2547 struct task_struct *next)
2551 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2554 * prepare_task_switch - prepare to switch tasks
2555 * @rq: the runqueue preparing to switch
2556 * @prev: the current task that is being switched out
2557 * @next: the task we are going to switch to.
2559 * This is called with the rq lock held and interrupts off. It must
2560 * be paired with a subsequent finish_task_switch after the context
2563 * prepare_task_switch sets up locking and calls architecture specific
2567 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2568 struct task_struct *next)
2570 fire_sched_out_preempt_notifiers(prev, next);
2571 prepare_lock_switch(rq, next);
2572 prepare_arch_switch(next);
2576 * finish_task_switch - clean up after a task-switch
2577 * @rq: runqueue associated with task-switch
2578 * @prev: the thread we just switched away from.
2580 * finish_task_switch must be called after the context switch, paired
2581 * with a prepare_task_switch call before the context switch.
2582 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2583 * and do any other architecture-specific cleanup actions.
2585 * Note that we may have delayed dropping an mm in context_switch(). If
2586 * so, we finish that here outside of the runqueue lock. (Doing it
2587 * with the lock held can cause deadlocks; see schedule() for
2590 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2591 __releases(rq->lock)
2593 struct mm_struct *mm = rq->prev_mm;
2599 * A task struct has one reference for the use as "current".
2600 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2601 * schedule one last time. The schedule call will never return, and
2602 * the scheduled task must drop that reference.
2603 * The test for TASK_DEAD must occur while the runqueue locks are
2604 * still held, otherwise prev could be scheduled on another cpu, die
2605 * there before we look at prev->state, and then the reference would
2607 * Manfred Spraul <manfred@colorfullife.com>
2609 prev_state = prev->state;
2610 finish_arch_switch(prev);
2611 finish_lock_switch(rq, prev);
2613 if (current->sched_class->post_schedule)
2614 current->sched_class->post_schedule(rq);
2617 fire_sched_in_preempt_notifiers(current);
2620 if (unlikely(prev_state == TASK_DEAD)) {
2622 * Remove function-return probe instances associated with this
2623 * task and put them back on the free list.
2625 kprobe_flush_task(prev);
2626 put_task_struct(prev);
2631 * schedule_tail - first thing a freshly forked thread must call.
2632 * @prev: the thread we just switched away from.
2634 asmlinkage void schedule_tail(struct task_struct *prev)
2635 __releases(rq->lock)
2637 struct rq *rq = this_rq();
2639 finish_task_switch(rq, prev);
2640 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2641 /* In this case, finish_task_switch does not reenable preemption */
2644 if (current->set_child_tid)
2645 put_user(task_pid_vnr(current), current->set_child_tid);
2649 * context_switch - switch to the new MM and the new
2650 * thread's register state.
2653 context_switch(struct rq *rq, struct task_struct *prev,
2654 struct task_struct *next)
2656 struct mm_struct *mm, *oldmm;
2658 prepare_task_switch(rq, prev, next);
2659 trace_sched_switch(rq, prev, next);
2661 oldmm = prev->active_mm;
2663 * For paravirt, this is coupled with an exit in switch_to to
2664 * combine the page table reload and the switch backend into
2667 arch_enter_lazy_cpu_mode();
2669 if (unlikely(!mm)) {
2670 next->active_mm = oldmm;
2671 atomic_inc(&oldmm->mm_count);
2672 enter_lazy_tlb(oldmm, next);
2674 switch_mm(oldmm, mm, next);
2676 if (unlikely(!prev->mm)) {
2677 prev->active_mm = NULL;
2678 rq->prev_mm = oldmm;
2681 * Since the runqueue lock will be released by the next
2682 * task (which is an invalid locking op but in the case
2683 * of the scheduler it's an obvious special-case), so we
2684 * do an early lockdep release here:
2686 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2687 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2690 /* Here we just switch the register state and the stack. */
2691 switch_to(prev, next, prev);
2695 * this_rq must be evaluated again because prev may have moved
2696 * CPUs since it called schedule(), thus the 'rq' on its stack
2697 * frame will be invalid.
2699 finish_task_switch(this_rq(), prev);
2703 * nr_running, nr_uninterruptible and nr_context_switches:
2705 * externally visible scheduler statistics: current number of runnable
2706 * threads, current number of uninterruptible-sleeping threads, total
2707 * number of context switches performed since bootup.
2709 unsigned long nr_running(void)
2711 unsigned long i, sum = 0;
2713 for_each_online_cpu(i)
2714 sum += cpu_rq(i)->nr_running;
2719 unsigned long nr_uninterruptible(void)
2721 unsigned long i, sum = 0;
2723 for_each_possible_cpu(i)
2724 sum += cpu_rq(i)->nr_uninterruptible;
2727 * Since we read the counters lockless, it might be slightly
2728 * inaccurate. Do not allow it to go below zero though:
2730 if (unlikely((long)sum < 0))
2736 unsigned long long nr_context_switches(void)
2739 unsigned long long sum = 0;
2741 for_each_possible_cpu(i)
2742 sum += cpu_rq(i)->nr_switches;
2747 unsigned long nr_iowait(void)
2749 unsigned long i, sum = 0;
2751 for_each_possible_cpu(i)
2752 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2757 unsigned long nr_active(void)
2759 unsigned long i, running = 0, uninterruptible = 0;
2761 for_each_online_cpu(i) {
2762 running += cpu_rq(i)->nr_running;
2763 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2766 if (unlikely((long)uninterruptible < 0))
2767 uninterruptible = 0;
2769 return running + uninterruptible;
2773 * Update rq->cpu_load[] statistics. This function is usually called every
2774 * scheduler tick (TICK_NSEC).
2776 static void update_cpu_load(struct rq *this_rq)
2778 unsigned long this_load = this_rq->load.weight;
2781 this_rq->nr_load_updates++;
2783 /* Update our load: */
2784 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2785 unsigned long old_load, new_load;
2787 /* scale is effectively 1 << i now, and >> i divides by scale */
2789 old_load = this_rq->cpu_load[i];
2790 new_load = this_load;
2792 * Round up the averaging division if load is increasing. This
2793 * prevents us from getting stuck on 9 if the load is 10, for
2796 if (new_load > old_load)
2797 new_load += scale-1;
2798 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2805 * double_rq_lock - safely lock two runqueues
2807 * Note this does not disable interrupts like task_rq_lock,
2808 * you need to do so manually before calling.
2810 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2811 __acquires(rq1->lock)
2812 __acquires(rq2->lock)
2814 BUG_ON(!irqs_disabled());
2816 spin_lock(&rq1->lock);
2817 __acquire(rq2->lock); /* Fake it out ;) */
2820 spin_lock(&rq1->lock);
2821 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2823 spin_lock(&rq2->lock);
2824 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2827 update_rq_clock(rq1);
2828 update_rq_clock(rq2);
2832 * double_rq_unlock - safely unlock two runqueues
2834 * Note this does not restore interrupts like task_rq_unlock,
2835 * you need to do so manually after calling.
2837 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2838 __releases(rq1->lock)
2839 __releases(rq2->lock)
2841 spin_unlock(&rq1->lock);
2843 spin_unlock(&rq2->lock);
2845 __release(rq2->lock);
2849 * If dest_cpu is allowed for this process, migrate the task to it.
2850 * This is accomplished by forcing the cpu_allowed mask to only
2851 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2852 * the cpu_allowed mask is restored.
2854 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2856 struct migration_req req;
2857 unsigned long flags;
2860 rq = task_rq_lock(p, &flags);
2861 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2862 || unlikely(!cpu_active(dest_cpu)))
2865 /* force the process onto the specified CPU */
2866 if (migrate_task(p, dest_cpu, &req)) {
2867 /* Need to wait for migration thread (might exit: take ref). */
2868 struct task_struct *mt = rq->migration_thread;
2870 get_task_struct(mt);
2871 task_rq_unlock(rq, &flags);
2872 wake_up_process(mt);
2873 put_task_struct(mt);
2874 wait_for_completion(&req.done);
2879 task_rq_unlock(rq, &flags);
2883 * sched_exec - execve() is a valuable balancing opportunity, because at
2884 * this point the task has the smallest effective memory and cache footprint.
2886 void sched_exec(void)
2888 int new_cpu, this_cpu = get_cpu();
2889 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2891 if (new_cpu != this_cpu)
2892 sched_migrate_task(current, new_cpu);
2896 * pull_task - move a task from a remote runqueue to the local runqueue.
2897 * Both runqueues must be locked.
2899 static void pull_task(struct rq *src_rq, struct task_struct *p,
2900 struct rq *this_rq, int this_cpu)
2902 deactivate_task(src_rq, p, 0);
2903 set_task_cpu(p, this_cpu);
2904 activate_task(this_rq, p, 0);
2906 * Note that idle threads have a prio of MAX_PRIO, for this test
2907 * to be always true for them.
2909 check_preempt_curr(this_rq, p, 0);
2913 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2916 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2917 struct sched_domain *sd, enum cpu_idle_type idle,
2921 * We do not migrate tasks that are:
2922 * 1) running (obviously), or
2923 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2924 * 3) are cache-hot on their current CPU.
2926 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2927 schedstat_inc(p, se.nr_failed_migrations_affine);
2932 if (task_running(rq, p)) {
2933 schedstat_inc(p, se.nr_failed_migrations_running);
2938 * Aggressive migration if:
2939 * 1) task is cache cold, or
2940 * 2) too many balance attempts have failed.
2943 if (!task_hot(p, rq->clock, sd) ||
2944 sd->nr_balance_failed > sd->cache_nice_tries) {
2945 #ifdef CONFIG_SCHEDSTATS
2946 if (task_hot(p, rq->clock, sd)) {
2947 schedstat_inc(sd, lb_hot_gained[idle]);
2948 schedstat_inc(p, se.nr_forced_migrations);
2954 if (task_hot(p, rq->clock, sd)) {
2955 schedstat_inc(p, se.nr_failed_migrations_hot);
2961 static unsigned long
2962 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2963 unsigned long max_load_move, struct sched_domain *sd,
2964 enum cpu_idle_type idle, int *all_pinned,
2965 int *this_best_prio, struct rq_iterator *iterator)
2967 int loops = 0, pulled = 0, pinned = 0;
2968 struct task_struct *p;
2969 long rem_load_move = max_load_move;
2971 if (max_load_move == 0)
2977 * Start the load-balancing iterator:
2979 p = iterator->start(iterator->arg);
2981 if (!p || loops++ > sysctl_sched_nr_migrate)
2984 if ((p->se.load.weight >> 1) > rem_load_move ||
2985 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2986 p = iterator->next(iterator->arg);
2990 pull_task(busiest, p, this_rq, this_cpu);
2992 rem_load_move -= p->se.load.weight;
2995 * We only want to steal up to the prescribed amount of weighted load.
2997 if (rem_load_move > 0) {
2998 if (p->prio < *this_best_prio)
2999 *this_best_prio = p->prio;
3000 p = iterator->next(iterator->arg);
3005 * Right now, this is one of only two places pull_task() is called,
3006 * so we can safely collect pull_task() stats here rather than
3007 * inside pull_task().
3009 schedstat_add(sd, lb_gained[idle], pulled);
3012 *all_pinned = pinned;
3014 return max_load_move - rem_load_move;
3018 * move_tasks tries to move up to max_load_move weighted load from busiest to
3019 * this_rq, as part of a balancing operation within domain "sd".
3020 * Returns 1 if successful and 0 otherwise.
3022 * Called with both runqueues locked.
3024 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3025 unsigned long max_load_move,
3026 struct sched_domain *sd, enum cpu_idle_type idle,
3029 const struct sched_class *class = sched_class_highest;
3030 unsigned long total_load_moved = 0;
3031 int this_best_prio = this_rq->curr->prio;
3035 class->load_balance(this_rq, this_cpu, busiest,
3036 max_load_move - total_load_moved,
3037 sd, idle, all_pinned, &this_best_prio);
3038 class = class->next;
3040 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3043 } while (class && max_load_move > total_load_moved);
3045 return total_load_moved > 0;
3049 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3050 struct sched_domain *sd, enum cpu_idle_type idle,
3051 struct rq_iterator *iterator)
3053 struct task_struct *p = iterator->start(iterator->arg);
3057 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3058 pull_task(busiest, p, this_rq, this_cpu);
3060 * Right now, this is only the second place pull_task()
3061 * is called, so we can safely collect pull_task()
3062 * stats here rather than inside pull_task().
3064 schedstat_inc(sd, lb_gained[idle]);
3068 p = iterator->next(iterator->arg);
3075 * move_one_task tries to move exactly one task from busiest to this_rq, as
3076 * part of active balancing operations within "domain".
3077 * Returns 1 if successful and 0 otherwise.
3079 * Called with both runqueues locked.
3081 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3082 struct sched_domain *sd, enum cpu_idle_type idle)
3084 const struct sched_class *class;
3086 for (class = sched_class_highest; class; class = class->next)
3087 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3094 * find_busiest_group finds and returns the busiest CPU group within the
3095 * domain. It calculates and returns the amount of weighted load which
3096 * should be moved to restore balance via the imbalance parameter.
3098 static struct sched_group *
3099 find_busiest_group(struct sched_domain *sd, int this_cpu,
3100 unsigned long *imbalance, enum cpu_idle_type idle,
3101 int *sd_idle, const struct cpumask *cpus, int *balance)
3103 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3104 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3105 unsigned long max_pull;
3106 unsigned long busiest_load_per_task, busiest_nr_running;
3107 unsigned long this_load_per_task, this_nr_running;
3108 int load_idx, group_imb = 0;
3109 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3110 int power_savings_balance = 1;
3111 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3112 unsigned long min_nr_running = ULONG_MAX;
3113 struct sched_group *group_min = NULL, *group_leader = NULL;
3116 max_load = this_load = total_load = total_pwr = 0;
3117 busiest_load_per_task = busiest_nr_running = 0;
3118 this_load_per_task = this_nr_running = 0;
3120 if (idle == CPU_NOT_IDLE)
3121 load_idx = sd->busy_idx;
3122 else if (idle == CPU_NEWLY_IDLE)
3123 load_idx = sd->newidle_idx;
3125 load_idx = sd->idle_idx;
3128 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3131 int __group_imb = 0;
3132 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3133 unsigned long sum_nr_running, sum_weighted_load;
3134 unsigned long sum_avg_load_per_task;
3135 unsigned long avg_load_per_task;
3137 local_group = cpumask_test_cpu(this_cpu,
3138 sched_group_cpus(group));
3141 balance_cpu = cpumask_first(sched_group_cpus(group));
3143 /* Tally up the load of all CPUs in the group */
3144 sum_weighted_load = sum_nr_running = avg_load = 0;
3145 sum_avg_load_per_task = avg_load_per_task = 0;
3148 min_cpu_load = ~0UL;
3150 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3151 struct rq *rq = cpu_rq(i);
3153 if (*sd_idle && rq->nr_running)
3156 /* Bias balancing toward cpus of our domain */
3158 if (idle_cpu(i) && !first_idle_cpu) {
3163 load = target_load(i, load_idx);
3165 load = source_load(i, load_idx);
3166 if (load > max_cpu_load)
3167 max_cpu_load = load;
3168 if (min_cpu_load > load)
3169 min_cpu_load = load;
3173 sum_nr_running += rq->nr_running;
3174 sum_weighted_load += weighted_cpuload(i);
3176 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3180 * First idle cpu or the first cpu(busiest) in this sched group
3181 * is eligible for doing load balancing at this and above
3182 * domains. In the newly idle case, we will allow all the cpu's
3183 * to do the newly idle load balance.
3185 if (idle != CPU_NEWLY_IDLE && local_group &&
3186 balance_cpu != this_cpu && balance) {
3191 total_load += avg_load;
3192 total_pwr += group->__cpu_power;
3194 /* Adjust by relative CPU power of the group */
3195 avg_load = sg_div_cpu_power(group,
3196 avg_load * SCHED_LOAD_SCALE);
3200 * Consider the group unbalanced when the imbalance is larger
3201 * than the average weight of two tasks.
3203 * APZ: with cgroup the avg task weight can vary wildly and
3204 * might not be a suitable number - should we keep a
3205 * normalized nr_running number somewhere that negates
3208 avg_load_per_task = sg_div_cpu_power(group,
3209 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3211 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3214 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3217 this_load = avg_load;
3219 this_nr_running = sum_nr_running;
3220 this_load_per_task = sum_weighted_load;
3221 } else if (avg_load > max_load &&
3222 (sum_nr_running > group_capacity || __group_imb)) {
3223 max_load = avg_load;
3225 busiest_nr_running = sum_nr_running;
3226 busiest_load_per_task = sum_weighted_load;
3227 group_imb = __group_imb;
3230 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3232 * Busy processors will not participate in power savings
3235 if (idle == CPU_NOT_IDLE ||
3236 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3240 * If the local group is idle or completely loaded
3241 * no need to do power savings balance at this domain
3243 if (local_group && (this_nr_running >= group_capacity ||
3245 power_savings_balance = 0;
3248 * If a group is already running at full capacity or idle,
3249 * don't include that group in power savings calculations
3251 if (!power_savings_balance || sum_nr_running >= group_capacity
3256 * Calculate the group which has the least non-idle load.
3257 * This is the group from where we need to pick up the load
3260 if ((sum_nr_running < min_nr_running) ||
3261 (sum_nr_running == min_nr_running &&
3262 cpumask_first(sched_group_cpus(group)) >
3263 cpumask_first(sched_group_cpus(group_min)))) {
3265 min_nr_running = sum_nr_running;
3266 min_load_per_task = sum_weighted_load /
3271 * Calculate the group which is almost near its
3272 * capacity but still has some space to pick up some load
3273 * from other group and save more power
3275 if (sum_nr_running <= group_capacity - 1) {
3276 if (sum_nr_running > leader_nr_running ||
3277 (sum_nr_running == leader_nr_running &&
3278 cpumask_first(sched_group_cpus(group)) <
3279 cpumask_first(sched_group_cpus(group_leader)))) {
3280 group_leader = group;
3281 leader_nr_running = sum_nr_running;
3286 group = group->next;
3287 } while (group != sd->groups);
3289 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3292 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3294 if (this_load >= avg_load ||
3295 100*max_load <= sd->imbalance_pct*this_load)
3298 busiest_load_per_task /= busiest_nr_running;
3300 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3303 * We're trying to get all the cpus to the average_load, so we don't
3304 * want to push ourselves above the average load, nor do we wish to
3305 * reduce the max loaded cpu below the average load, as either of these
3306 * actions would just result in more rebalancing later, and ping-pong
3307 * tasks around. Thus we look for the minimum possible imbalance.
3308 * Negative imbalances (*we* are more loaded than anyone else) will
3309 * be counted as no imbalance for these purposes -- we can't fix that
3310 * by pulling tasks to us. Be careful of negative numbers as they'll
3311 * appear as very large values with unsigned longs.
3313 if (max_load <= busiest_load_per_task)
3317 * In the presence of smp nice balancing, certain scenarios can have
3318 * max load less than avg load(as we skip the groups at or below
3319 * its cpu_power, while calculating max_load..)
3321 if (max_load < avg_load) {
3323 goto small_imbalance;
3326 /* Don't want to pull so many tasks that a group would go idle */
3327 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3329 /* How much load to actually move to equalise the imbalance */
3330 *imbalance = min(max_pull * busiest->__cpu_power,
3331 (avg_load - this_load) * this->__cpu_power)
3335 * if *imbalance is less than the average load per runnable task
3336 * there is no gaurantee that any tasks will be moved so we'll have
3337 * a think about bumping its value to force at least one task to be
3340 if (*imbalance < busiest_load_per_task) {
3341 unsigned long tmp, pwr_now, pwr_move;
3345 pwr_move = pwr_now = 0;
3347 if (this_nr_running) {
3348 this_load_per_task /= this_nr_running;
3349 if (busiest_load_per_task > this_load_per_task)
3352 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3354 if (max_load - this_load + busiest_load_per_task >=
3355 busiest_load_per_task * imbn) {
3356 *imbalance = busiest_load_per_task;
3361 * OK, we don't have enough imbalance to justify moving tasks,
3362 * however we may be able to increase total CPU power used by
3366 pwr_now += busiest->__cpu_power *
3367 min(busiest_load_per_task, max_load);
3368 pwr_now += this->__cpu_power *
3369 min(this_load_per_task, this_load);
3370 pwr_now /= SCHED_LOAD_SCALE;
3372 /* Amount of load we'd subtract */
3373 tmp = sg_div_cpu_power(busiest,
3374 busiest_load_per_task * SCHED_LOAD_SCALE);
3376 pwr_move += busiest->__cpu_power *
3377 min(busiest_load_per_task, max_load - tmp);
3379 /* Amount of load we'd add */
3380 if (max_load * busiest->__cpu_power <
3381 busiest_load_per_task * SCHED_LOAD_SCALE)
3382 tmp = sg_div_cpu_power(this,
3383 max_load * busiest->__cpu_power);
3385 tmp = sg_div_cpu_power(this,
3386 busiest_load_per_task * SCHED_LOAD_SCALE);
3387 pwr_move += this->__cpu_power *
3388 min(this_load_per_task, this_load + tmp);
3389 pwr_move /= SCHED_LOAD_SCALE;
3391 /* Move if we gain throughput */
3392 if (pwr_move > pwr_now)
3393 *imbalance = busiest_load_per_task;
3399 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3400 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3403 if (this == group_leader && group_leader != group_min) {
3404 *imbalance = min_load_per_task;
3405 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3406 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3407 cpumask_first(sched_group_cpus(group_leader));
3418 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3421 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3422 unsigned long imbalance, const struct cpumask *cpus)
3424 struct rq *busiest = NULL, *rq;
3425 unsigned long max_load = 0;
3428 for_each_cpu(i, sched_group_cpus(group)) {
3431 if (!cpumask_test_cpu(i, cpus))
3435 wl = weighted_cpuload(i);
3437 if (rq->nr_running == 1 && wl > imbalance)
3440 if (wl > max_load) {
3450 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3451 * so long as it is large enough.
3453 #define MAX_PINNED_INTERVAL 512
3456 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3457 * tasks if there is an imbalance.
3459 static int load_balance(int this_cpu, struct rq *this_rq,
3460 struct sched_domain *sd, enum cpu_idle_type idle,
3461 int *balance, struct cpumask *cpus)
3463 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3464 struct sched_group *group;
3465 unsigned long imbalance;
3467 unsigned long flags;
3469 cpumask_setall(cpus);
3472 * When power savings policy is enabled for the parent domain, idle
3473 * sibling can pick up load irrespective of busy siblings. In this case,
3474 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3475 * portraying it as CPU_NOT_IDLE.
3477 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3478 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3481 schedstat_inc(sd, lb_count[idle]);
3485 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3492 schedstat_inc(sd, lb_nobusyg[idle]);
3496 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3498 schedstat_inc(sd, lb_nobusyq[idle]);
3502 BUG_ON(busiest == this_rq);
3504 schedstat_add(sd, lb_imbalance[idle], imbalance);
3507 if (busiest->nr_running > 1) {
3509 * Attempt to move tasks. If find_busiest_group has found
3510 * an imbalance but busiest->nr_running <= 1, the group is
3511 * still unbalanced. ld_moved simply stays zero, so it is
3512 * correctly treated as an imbalance.
3514 local_irq_save(flags);
3515 double_rq_lock(this_rq, busiest);
3516 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3517 imbalance, sd, idle, &all_pinned);
3518 double_rq_unlock(this_rq, busiest);
3519 local_irq_restore(flags);
3522 * some other cpu did the load balance for us.
3524 if (ld_moved && this_cpu != smp_processor_id())
3525 resched_cpu(this_cpu);
3527 /* All tasks on this runqueue were pinned by CPU affinity */
3528 if (unlikely(all_pinned)) {
3529 cpumask_clear_cpu(cpu_of(busiest), cpus);
3530 if (!cpumask_empty(cpus))
3537 schedstat_inc(sd, lb_failed[idle]);
3538 sd->nr_balance_failed++;
3540 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3542 spin_lock_irqsave(&busiest->lock, flags);
3544 /* don't kick the migration_thread, if the curr
3545 * task on busiest cpu can't be moved to this_cpu
3547 if (!cpumask_test_cpu(this_cpu,
3548 &busiest->curr->cpus_allowed)) {
3549 spin_unlock_irqrestore(&busiest->lock, flags);
3551 goto out_one_pinned;
3554 if (!busiest->active_balance) {
3555 busiest->active_balance = 1;
3556 busiest->push_cpu = this_cpu;
3559 spin_unlock_irqrestore(&busiest->lock, flags);
3561 wake_up_process(busiest->migration_thread);
3564 * We've kicked active balancing, reset the failure
3567 sd->nr_balance_failed = sd->cache_nice_tries+1;
3570 sd->nr_balance_failed = 0;
3572 if (likely(!active_balance)) {
3573 /* We were unbalanced, so reset the balancing interval */
3574 sd->balance_interval = sd->min_interval;
3577 * If we've begun active balancing, start to back off. This
3578 * case may not be covered by the all_pinned logic if there
3579 * is only 1 task on the busy runqueue (because we don't call
3582 if (sd->balance_interval < sd->max_interval)
3583 sd->balance_interval *= 2;
3586 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3587 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3593 schedstat_inc(sd, lb_balanced[idle]);
3595 sd->nr_balance_failed = 0;
3598 /* tune up the balancing interval */
3599 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3600 (sd->balance_interval < sd->max_interval))
3601 sd->balance_interval *= 2;
3603 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3604 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3615 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3616 * tasks if there is an imbalance.
3618 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3619 * this_rq is locked.
3622 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3623 struct cpumask *cpus)
3625 struct sched_group *group;
3626 struct rq *busiest = NULL;
3627 unsigned long imbalance;
3632 cpumask_setall(cpus);
3635 * When power savings policy is enabled for the parent domain, idle
3636 * sibling can pick up load irrespective of busy siblings. In this case,
3637 * let the state of idle sibling percolate up as IDLE, instead of
3638 * portraying it as CPU_NOT_IDLE.
3640 if (sd->flags & SD_SHARE_CPUPOWER &&
3641 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3644 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3646 update_shares_locked(this_rq, sd);
3647 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3648 &sd_idle, cpus, NULL);
3650 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3654 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3656 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3660 BUG_ON(busiest == this_rq);
3662 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3665 if (busiest->nr_running > 1) {
3666 /* Attempt to move tasks */
3667 double_lock_balance(this_rq, busiest);
3668 /* this_rq->clock is already updated */
3669 update_rq_clock(busiest);
3670 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3671 imbalance, sd, CPU_NEWLY_IDLE,
3673 double_unlock_balance(this_rq, busiest);
3675 if (unlikely(all_pinned)) {
3676 cpumask_clear_cpu(cpu_of(busiest), cpus);
3677 if (!cpumask_empty(cpus))
3683 int active_balance = 0;
3685 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3686 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3687 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3690 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3693 if (sd->nr_balance_failed++ < 2)
3697 * The only task running in a non-idle cpu can be moved to this
3698 * cpu in an attempt to completely freeup the other CPU
3699 * package. The same method used to move task in load_balance()
3700 * have been extended for load_balance_newidle() to speedup
3701 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3703 * The package power saving logic comes from
3704 * find_busiest_group(). If there are no imbalance, then
3705 * f_b_g() will return NULL. However when sched_mc={1,2} then
3706 * f_b_g() will select a group from which a running task may be
3707 * pulled to this cpu in order to make the other package idle.
3708 * If there is no opportunity to make a package idle and if
3709 * there are no imbalance, then f_b_g() will return NULL and no
3710 * action will be taken in load_balance_newidle().
3712 * Under normal task pull operation due to imbalance, there
3713 * will be more than one task in the source run queue and
3714 * move_tasks() will succeed. ld_moved will be true and this
3715 * active balance code will not be triggered.
3718 /* Lock busiest in correct order while this_rq is held */
3719 double_lock_balance(this_rq, busiest);
3722 * don't kick the migration_thread, if the curr
3723 * task on busiest cpu can't be moved to this_cpu
3725 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3726 double_unlock_balance(this_rq, busiest);
3731 if (!busiest->active_balance) {
3732 busiest->active_balance = 1;
3733 busiest->push_cpu = this_cpu;
3737 double_unlock_balance(this_rq, busiest);
3739 * Should not call ttwu while holding a rq->lock
3741 spin_unlock(&this_rq->lock);
3743 wake_up_process(busiest->migration_thread);
3744 spin_lock(&this_rq->lock);
3747 sd->nr_balance_failed = 0;
3749 update_shares_locked(this_rq, sd);
3753 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3754 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3755 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3757 sd->nr_balance_failed = 0;
3763 * idle_balance is called by schedule() if this_cpu is about to become
3764 * idle. Attempts to pull tasks from other CPUs.
3766 static void idle_balance(int this_cpu, struct rq *this_rq)
3768 struct sched_domain *sd;
3769 int pulled_task = 0;
3770 unsigned long next_balance = jiffies + HZ;
3771 cpumask_var_t tmpmask;
3773 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3776 for_each_domain(this_cpu, sd) {
3777 unsigned long interval;
3779 if (!(sd->flags & SD_LOAD_BALANCE))
3782 if (sd->flags & SD_BALANCE_NEWIDLE)
3783 /* If we've pulled tasks over stop searching: */
3784 pulled_task = load_balance_newidle(this_cpu, this_rq,
3787 interval = msecs_to_jiffies(sd->balance_interval);
3788 if (time_after(next_balance, sd->last_balance + interval))
3789 next_balance = sd->last_balance + interval;
3793 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3795 * We are going idle. next_balance may be set based on
3796 * a busy processor. So reset next_balance.
3798 this_rq->next_balance = next_balance;
3800 free_cpumask_var(tmpmask);
3804 * active_load_balance is run by migration threads. It pushes running tasks
3805 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3806 * running on each physical CPU where possible, and avoids physical /
3807 * logical imbalances.
3809 * Called with busiest_rq locked.
3811 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3813 int target_cpu = busiest_rq->push_cpu;
3814 struct sched_domain *sd;
3815 struct rq *target_rq;
3817 /* Is there any task to move? */
3818 if (busiest_rq->nr_running <= 1)
3821 target_rq = cpu_rq(target_cpu);
3824 * This condition is "impossible", if it occurs
3825 * we need to fix it. Originally reported by
3826 * Bjorn Helgaas on a 128-cpu setup.
3828 BUG_ON(busiest_rq == target_rq);
3830 /* move a task from busiest_rq to target_rq */
3831 double_lock_balance(busiest_rq, target_rq);
3832 update_rq_clock(busiest_rq);
3833 update_rq_clock(target_rq);
3835 /* Search for an sd spanning us and the target CPU. */
3836 for_each_domain(target_cpu, sd) {
3837 if ((sd->flags & SD_LOAD_BALANCE) &&
3838 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3843 schedstat_inc(sd, alb_count);
3845 if (move_one_task(target_rq, target_cpu, busiest_rq,
3847 schedstat_inc(sd, alb_pushed);
3849 schedstat_inc(sd, alb_failed);
3851 double_unlock_balance(busiest_rq, target_rq);
3856 atomic_t load_balancer;
3857 cpumask_var_t cpu_mask;
3858 } nohz ____cacheline_aligned = {
3859 .load_balancer = ATOMIC_INIT(-1),
3863 * This routine will try to nominate the ilb (idle load balancing)
3864 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3865 * load balancing on behalf of all those cpus. If all the cpus in the system
3866 * go into this tickless mode, then there will be no ilb owner (as there is
3867 * no need for one) and all the cpus will sleep till the next wakeup event
3870 * For the ilb owner, tick is not stopped. And this tick will be used
3871 * for idle load balancing. ilb owner will still be part of
3874 * While stopping the tick, this cpu will become the ilb owner if there
3875 * is no other owner. And will be the owner till that cpu becomes busy
3876 * or if all cpus in the system stop their ticks at which point
3877 * there is no need for ilb owner.
3879 * When the ilb owner becomes busy, it nominates another owner, during the
3880 * next busy scheduler_tick()
3882 int select_nohz_load_balancer(int stop_tick)
3884 int cpu = smp_processor_id();
3887 cpumask_set_cpu(cpu, nohz.cpu_mask);
3888 cpu_rq(cpu)->in_nohz_recently = 1;
3891 * If we are going offline and still the leader, give up!
3893 if (!cpu_active(cpu) &&
3894 atomic_read(&nohz.load_balancer) == cpu) {
3895 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3900 /* time for ilb owner also to sleep */
3901 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3902 if (atomic_read(&nohz.load_balancer) == cpu)
3903 atomic_set(&nohz.load_balancer, -1);
3907 if (atomic_read(&nohz.load_balancer) == -1) {
3908 /* make me the ilb owner */
3909 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3911 } else if (atomic_read(&nohz.load_balancer) == cpu)
3914 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3917 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3919 if (atomic_read(&nohz.load_balancer) == cpu)
3920 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3927 static DEFINE_SPINLOCK(balancing);
3930 * It checks each scheduling domain to see if it is due to be balanced,
3931 * and initiates a balancing operation if so.
3933 * Balancing parameters are set up in arch_init_sched_domains.
3935 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3938 struct rq *rq = cpu_rq(cpu);
3939 unsigned long interval;
3940 struct sched_domain *sd;
3941 /* Earliest time when we have to do rebalance again */
3942 unsigned long next_balance = jiffies + 60*HZ;
3943 int update_next_balance = 0;
3947 /* Fails alloc? Rebalancing probably not a priority right now. */
3948 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3951 for_each_domain(cpu, sd) {
3952 if (!(sd->flags & SD_LOAD_BALANCE))
3955 interval = sd->balance_interval;
3956 if (idle != CPU_IDLE)
3957 interval *= sd->busy_factor;
3959 /* scale ms to jiffies */
3960 interval = msecs_to_jiffies(interval);
3961 if (unlikely(!interval))
3963 if (interval > HZ*NR_CPUS/10)
3964 interval = HZ*NR_CPUS/10;
3966 need_serialize = sd->flags & SD_SERIALIZE;
3968 if (need_serialize) {
3969 if (!spin_trylock(&balancing))
3973 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3974 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
3976 * We've pulled tasks over so either we're no
3977 * longer idle, or one of our SMT siblings is
3980 idle = CPU_NOT_IDLE;
3982 sd->last_balance = jiffies;
3985 spin_unlock(&balancing);
3987 if (time_after(next_balance, sd->last_balance + interval)) {
3988 next_balance = sd->last_balance + interval;
3989 update_next_balance = 1;
3993 * Stop the load balance at this level. There is another
3994 * CPU in our sched group which is doing load balancing more
4002 * next_balance will be updated only when there is a need.
4003 * When the cpu is attached to null domain for ex, it will not be
4006 if (likely(update_next_balance))
4007 rq->next_balance = next_balance;
4009 free_cpumask_var(tmp);
4013 * run_rebalance_domains is triggered when needed from the scheduler tick.
4014 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4015 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4017 static void run_rebalance_domains(struct softirq_action *h)
4019 int this_cpu = smp_processor_id();
4020 struct rq *this_rq = cpu_rq(this_cpu);
4021 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4022 CPU_IDLE : CPU_NOT_IDLE;
4024 rebalance_domains(this_cpu, idle);
4028 * If this cpu is the owner for idle load balancing, then do the
4029 * balancing on behalf of the other idle cpus whose ticks are
4032 if (this_rq->idle_at_tick &&
4033 atomic_read(&nohz.load_balancer) == this_cpu) {
4037 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4038 if (balance_cpu == this_cpu)
4042 * If this cpu gets work to do, stop the load balancing
4043 * work being done for other cpus. Next load
4044 * balancing owner will pick it up.
4049 rebalance_domains(balance_cpu, CPU_IDLE);
4051 rq = cpu_rq(balance_cpu);
4052 if (time_after(this_rq->next_balance, rq->next_balance))
4053 this_rq->next_balance = rq->next_balance;
4060 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4062 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4063 * idle load balancing owner or decide to stop the periodic load balancing,
4064 * if the whole system is idle.
4066 static inline void trigger_load_balance(struct rq *rq, int cpu)
4070 * If we were in the nohz mode recently and busy at the current
4071 * scheduler tick, then check if we need to nominate new idle
4074 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4075 rq->in_nohz_recently = 0;
4077 if (atomic_read(&nohz.load_balancer) == cpu) {
4078 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4079 atomic_set(&nohz.load_balancer, -1);
4082 if (atomic_read(&nohz.load_balancer) == -1) {
4084 * simple selection for now: Nominate the
4085 * first cpu in the nohz list to be the next
4088 * TBD: Traverse the sched domains and nominate
4089 * the nearest cpu in the nohz.cpu_mask.
4091 int ilb = cpumask_first(nohz.cpu_mask);
4093 if (ilb < nr_cpu_ids)
4099 * If this cpu is idle and doing idle load balancing for all the
4100 * cpus with ticks stopped, is it time for that to stop?
4102 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4103 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4109 * If this cpu is idle and the idle load balancing is done by
4110 * someone else, then no need raise the SCHED_SOFTIRQ
4112 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4113 cpumask_test_cpu(cpu, nohz.cpu_mask))
4116 if (time_after_eq(jiffies, rq->next_balance))
4117 raise_softirq(SCHED_SOFTIRQ);
4120 #else /* CONFIG_SMP */
4123 * on UP we do not need to balance between CPUs:
4125 static inline void idle_balance(int cpu, struct rq *rq)
4131 DEFINE_PER_CPU(struct kernel_stat, kstat);
4133 EXPORT_PER_CPU_SYMBOL(kstat);
4136 * Return any ns on the sched_clock that have not yet been banked in
4137 * @p in case that task is currently running.
4139 unsigned long long task_delta_exec(struct task_struct *p)
4141 unsigned long flags;
4145 rq = task_rq_lock(p, &flags);
4147 if (task_current(rq, p)) {
4150 update_rq_clock(rq);
4151 delta_exec = rq->clock - p->se.exec_start;
4152 if ((s64)delta_exec > 0)
4156 task_rq_unlock(rq, &flags);
4162 * Account user cpu time to a process.
4163 * @p: the process that the cpu time gets accounted to
4164 * @cputime: the cpu time spent in user space since the last update
4165 * @cputime_scaled: cputime scaled by cpu frequency
4167 void account_user_time(struct task_struct *p, cputime_t cputime,
4168 cputime_t cputime_scaled)
4170 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4173 /* Add user time to process. */
4174 p->utime = cputime_add(p->utime, cputime);
4175 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4176 account_group_user_time(p, cputime);
4178 /* Add user time to cpustat. */
4179 tmp = cputime_to_cputime64(cputime);
4180 if (TASK_NICE(p) > 0)
4181 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4183 cpustat->user = cputime64_add(cpustat->user, tmp);
4184 /* Account for user time used */
4185 acct_update_integrals(p);
4189 * Account guest cpu time to a process.
4190 * @p: the process that the cpu time gets accounted to
4191 * @cputime: the cpu time spent in virtual machine since the last update
4192 * @cputime_scaled: cputime scaled by cpu frequency
4194 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4195 cputime_t cputime_scaled)
4198 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4200 tmp = cputime_to_cputime64(cputime);
4202 /* Add guest time to process. */
4203 p->utime = cputime_add(p->utime, cputime);
4204 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4205 account_group_user_time(p, cputime);
4206 p->gtime = cputime_add(p->gtime, cputime);
4208 /* Add guest time to cpustat. */
4209 cpustat->user = cputime64_add(cpustat->user, tmp);
4210 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4214 * Account system cpu time to a process.
4215 * @p: the process that the cpu time gets accounted to
4216 * @hardirq_offset: the offset to subtract from hardirq_count()
4217 * @cputime: the cpu time spent in kernel space since the last update
4218 * @cputime_scaled: cputime scaled by cpu frequency
4220 void account_system_time(struct task_struct *p, int hardirq_offset,
4221 cputime_t cputime, cputime_t cputime_scaled)
4223 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4226 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4227 account_guest_time(p, cputime, cputime_scaled);
4231 /* Add system time to process. */
4232 p->stime = cputime_add(p->stime, cputime);
4233 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4234 account_group_system_time(p, cputime);
4236 /* Add system time to cpustat. */
4237 tmp = cputime_to_cputime64(cputime);
4238 if (hardirq_count() - hardirq_offset)
4239 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4240 else if (softirq_count())
4241 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4243 cpustat->system = cputime64_add(cpustat->system, tmp);
4245 /* Account for system time used */
4246 acct_update_integrals(p);
4250 * Account for involuntary wait time.
4251 * @steal: the cpu time spent in involuntary wait
4253 void account_steal_time(cputime_t cputime)
4255 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4256 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4258 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4262 * Account for idle time.
4263 * @cputime: the cpu time spent in idle wait
4265 void account_idle_time(cputime_t cputime)
4267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4268 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4269 struct rq *rq = this_rq();
4271 if (atomic_read(&rq->nr_iowait) > 0)
4272 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4274 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4277 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4280 * Account a single tick of cpu time.
4281 * @p: the process that the cpu time gets accounted to
4282 * @user_tick: indicates if the tick is a user or a system tick
4284 void account_process_tick(struct task_struct *p, int user_tick)
4286 cputime_t one_jiffy = jiffies_to_cputime(1);
4287 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4288 struct rq *rq = this_rq();
4291 account_user_time(p, one_jiffy, one_jiffy_scaled);
4292 else if (p != rq->idle)
4293 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4296 account_idle_time(one_jiffy);
4300 * Account multiple ticks of steal time.
4301 * @p: the process from which the cpu time has been stolen
4302 * @ticks: number of stolen ticks
4304 void account_steal_ticks(unsigned long ticks)
4306 account_steal_time(jiffies_to_cputime(ticks));
4310 * Account multiple ticks of idle time.
4311 * @ticks: number of stolen ticks
4313 void account_idle_ticks(unsigned long ticks)
4315 account_idle_time(jiffies_to_cputime(ticks));
4321 * Use precise platform statistics if available:
4323 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4324 cputime_t task_utime(struct task_struct *p)
4329 cputime_t task_stime(struct task_struct *p)
4334 cputime_t task_utime(struct task_struct *p)
4336 clock_t utime = cputime_to_clock_t(p->utime),
4337 total = utime + cputime_to_clock_t(p->stime);
4341 * Use CFS's precise accounting:
4343 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4347 do_div(temp, total);
4349 utime = (clock_t)temp;
4351 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4352 return p->prev_utime;
4355 cputime_t task_stime(struct task_struct *p)
4360 * Use CFS's precise accounting. (we subtract utime from
4361 * the total, to make sure the total observed by userspace
4362 * grows monotonically - apps rely on that):
4364 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4365 cputime_to_clock_t(task_utime(p));
4368 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4370 return p->prev_stime;
4374 inline cputime_t task_gtime(struct task_struct *p)
4380 * This function gets called by the timer code, with HZ frequency.
4381 * We call it with interrupts disabled.
4383 * It also gets called by the fork code, when changing the parent's
4386 void scheduler_tick(void)
4388 int cpu = smp_processor_id();
4389 struct rq *rq = cpu_rq(cpu);
4390 struct task_struct *curr = rq->curr;
4394 spin_lock(&rq->lock);
4395 update_rq_clock(rq);
4396 update_cpu_load(rq);
4397 curr->sched_class->task_tick(rq, curr, 0);
4398 spin_unlock(&rq->lock);
4401 rq->idle_at_tick = idle_cpu(cpu);
4402 trigger_load_balance(rq, cpu);
4406 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4407 defined(CONFIG_PREEMPT_TRACER))
4409 static inline unsigned long get_parent_ip(unsigned long addr)
4411 if (in_lock_functions(addr)) {
4412 addr = CALLER_ADDR2;
4413 if (in_lock_functions(addr))
4414 addr = CALLER_ADDR3;
4419 void __kprobes add_preempt_count(int val)
4421 #ifdef CONFIG_DEBUG_PREEMPT
4425 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4428 preempt_count() += val;
4429 #ifdef CONFIG_DEBUG_PREEMPT
4431 * Spinlock count overflowing soon?
4433 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4436 if (preempt_count() == val)
4437 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4439 EXPORT_SYMBOL(add_preempt_count);
4441 void __kprobes sub_preempt_count(int val)
4443 #ifdef CONFIG_DEBUG_PREEMPT
4447 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4450 * Is the spinlock portion underflowing?
4452 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4453 !(preempt_count() & PREEMPT_MASK)))
4457 if (preempt_count() == val)
4458 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4459 preempt_count() -= val;
4461 EXPORT_SYMBOL(sub_preempt_count);
4466 * Print scheduling while atomic bug:
4468 static noinline void __schedule_bug(struct task_struct *prev)
4470 struct pt_regs *regs = get_irq_regs();
4472 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4473 prev->comm, prev->pid, preempt_count());
4475 debug_show_held_locks(prev);
4477 if (irqs_disabled())
4478 print_irqtrace_events(prev);
4487 * Various schedule()-time debugging checks and statistics:
4489 static inline void schedule_debug(struct task_struct *prev)
4492 * Test if we are atomic. Since do_exit() needs to call into
4493 * schedule() atomically, we ignore that path for now.
4494 * Otherwise, whine if we are scheduling when we should not be.
4496 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4497 __schedule_bug(prev);
4499 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4501 schedstat_inc(this_rq(), sched_count);
4502 #ifdef CONFIG_SCHEDSTATS
4503 if (unlikely(prev->lock_depth >= 0)) {
4504 schedstat_inc(this_rq(), bkl_count);
4505 schedstat_inc(prev, sched_info.bkl_count);
4511 * Pick up the highest-prio task:
4513 static inline struct task_struct *
4514 pick_next_task(struct rq *rq, struct task_struct *prev)
4516 const struct sched_class *class;
4517 struct task_struct *p;
4520 * Optimization: we know that if all tasks are in
4521 * the fair class we can call that function directly:
4523 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4524 p = fair_sched_class.pick_next_task(rq);
4529 class = sched_class_highest;
4531 p = class->pick_next_task(rq);
4535 * Will never be NULL as the idle class always
4536 * returns a non-NULL p:
4538 class = class->next;
4543 * schedule() is the main scheduler function.
4545 asmlinkage void __sched schedule(void)
4547 struct task_struct *prev, *next;
4548 unsigned long *switch_count;
4554 cpu = smp_processor_id();
4558 switch_count = &prev->nivcsw;
4560 release_kernel_lock(prev);
4561 need_resched_nonpreemptible:
4563 schedule_debug(prev);
4565 if (sched_feat(HRTICK))
4568 spin_lock_irq(&rq->lock);
4569 update_rq_clock(rq);
4570 clear_tsk_need_resched(prev);
4572 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4573 if (unlikely(signal_pending_state(prev->state, prev)))
4574 prev->state = TASK_RUNNING;
4576 deactivate_task(rq, prev, 1);
4577 switch_count = &prev->nvcsw;
4581 if (prev->sched_class->pre_schedule)
4582 prev->sched_class->pre_schedule(rq, prev);
4585 if (unlikely(!rq->nr_running))
4586 idle_balance(cpu, rq);
4588 prev->sched_class->put_prev_task(rq, prev);
4589 next = pick_next_task(rq, prev);
4591 if (likely(prev != next)) {
4592 sched_info_switch(prev, next);
4598 context_switch(rq, prev, next); /* unlocks the rq */
4600 * the context switch might have flipped the stack from under
4601 * us, hence refresh the local variables.
4603 cpu = smp_processor_id();
4606 spin_unlock_irq(&rq->lock);
4608 if (unlikely(reacquire_kernel_lock(current) < 0))
4609 goto need_resched_nonpreemptible;
4611 preempt_enable_no_resched();
4612 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4615 EXPORT_SYMBOL(schedule);
4617 #ifdef CONFIG_PREEMPT
4619 * this is the entry point to schedule() from in-kernel preemption
4620 * off of preempt_enable. Kernel preemptions off return from interrupt
4621 * occur there and call schedule directly.
4623 asmlinkage void __sched preempt_schedule(void)
4625 struct thread_info *ti = current_thread_info();
4628 * If there is a non-zero preempt_count or interrupts are disabled,
4629 * we do not want to preempt the current task. Just return..
4631 if (likely(ti->preempt_count || irqs_disabled()))
4635 add_preempt_count(PREEMPT_ACTIVE);
4637 sub_preempt_count(PREEMPT_ACTIVE);
4640 * Check again in case we missed a preemption opportunity
4641 * between schedule and now.
4644 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4646 EXPORT_SYMBOL(preempt_schedule);
4649 * this is the entry point to schedule() from kernel preemption
4650 * off of irq context.
4651 * Note, that this is called and return with irqs disabled. This will
4652 * protect us against recursive calling from irq.
4654 asmlinkage void __sched preempt_schedule_irq(void)
4656 struct thread_info *ti = current_thread_info();
4658 /* Catch callers which need to be fixed */
4659 BUG_ON(ti->preempt_count || !irqs_disabled());
4662 add_preempt_count(PREEMPT_ACTIVE);
4665 local_irq_disable();
4666 sub_preempt_count(PREEMPT_ACTIVE);
4669 * Check again in case we missed a preemption opportunity
4670 * between schedule and now.
4673 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4676 #endif /* CONFIG_PREEMPT */
4678 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4681 return try_to_wake_up(curr->private, mode, sync);
4683 EXPORT_SYMBOL(default_wake_function);
4686 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4687 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4688 * number) then we wake all the non-exclusive tasks and one exclusive task.
4690 * There are circumstances in which we can try to wake a task which has already
4691 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4692 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4694 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4695 int nr_exclusive, int sync, void *key)
4697 wait_queue_t *curr, *next;
4699 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4700 unsigned flags = curr->flags;
4702 if (curr->func(curr, mode, sync, key) &&
4703 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4709 * __wake_up - wake up threads blocked on a waitqueue.
4711 * @mode: which threads
4712 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4713 * @key: is directly passed to the wakeup function
4715 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4716 int nr_exclusive, void *key)
4718 unsigned long flags;
4720 spin_lock_irqsave(&q->lock, flags);
4721 __wake_up_common(q, mode, nr_exclusive, 0, key);
4722 spin_unlock_irqrestore(&q->lock, flags);
4724 EXPORT_SYMBOL(__wake_up);
4727 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4729 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4731 __wake_up_common(q, mode, 1, 0, NULL);
4735 * __wake_up_sync - wake up threads blocked on a waitqueue.
4737 * @mode: which threads
4738 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4740 * The sync wakeup differs that the waker knows that it will schedule
4741 * away soon, so while the target thread will be woken up, it will not
4742 * be migrated to another CPU - ie. the two threads are 'synchronized'
4743 * with each other. This can prevent needless bouncing between CPUs.
4745 * On UP it can prevent extra preemption.
4748 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4750 unsigned long flags;
4756 if (unlikely(!nr_exclusive))
4759 spin_lock_irqsave(&q->lock, flags);
4760 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4761 spin_unlock_irqrestore(&q->lock, flags);
4763 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4766 * complete: - signals a single thread waiting on this completion
4767 * @x: holds the state of this particular completion
4769 * This will wake up a single thread waiting on this completion. Threads will be
4770 * awakened in the same order in which they were queued.
4772 * See also complete_all(), wait_for_completion() and related routines.
4774 void complete(struct completion *x)
4776 unsigned long flags;
4778 spin_lock_irqsave(&x->wait.lock, flags);
4780 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4781 spin_unlock_irqrestore(&x->wait.lock, flags);
4783 EXPORT_SYMBOL(complete);
4786 * complete_all: - signals all threads waiting on this completion
4787 * @x: holds the state of this particular completion
4789 * This will wake up all threads waiting on this particular completion event.
4791 void complete_all(struct completion *x)
4793 unsigned long flags;
4795 spin_lock_irqsave(&x->wait.lock, flags);
4796 x->done += UINT_MAX/2;
4797 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4798 spin_unlock_irqrestore(&x->wait.lock, flags);
4800 EXPORT_SYMBOL(complete_all);
4802 static inline long __sched
4803 do_wait_for_common(struct completion *x, long timeout, int state)
4806 DECLARE_WAITQUEUE(wait, current);
4808 wait.flags |= WQ_FLAG_EXCLUSIVE;
4809 __add_wait_queue_tail(&x->wait, &wait);
4811 if (signal_pending_state(state, current)) {
4812 timeout = -ERESTARTSYS;
4815 __set_current_state(state);
4816 spin_unlock_irq(&x->wait.lock);
4817 timeout = schedule_timeout(timeout);
4818 spin_lock_irq(&x->wait.lock);
4819 } while (!x->done && timeout);
4820 __remove_wait_queue(&x->wait, &wait);
4825 return timeout ?: 1;
4829 wait_for_common(struct completion *x, long timeout, int state)
4833 spin_lock_irq(&x->wait.lock);
4834 timeout = do_wait_for_common(x, timeout, state);
4835 spin_unlock_irq(&x->wait.lock);
4840 * wait_for_completion: - waits for completion of a task
4841 * @x: holds the state of this particular completion
4843 * This waits to be signaled for completion of a specific task. It is NOT
4844 * interruptible and there is no timeout.
4846 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4847 * and interrupt capability. Also see complete().
4849 void __sched wait_for_completion(struct completion *x)
4851 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4853 EXPORT_SYMBOL(wait_for_completion);
4856 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4857 * @x: holds the state of this particular completion
4858 * @timeout: timeout value in jiffies
4860 * This waits for either a completion of a specific task to be signaled or for a
4861 * specified timeout to expire. The timeout is in jiffies. It is not
4864 unsigned long __sched
4865 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4867 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4869 EXPORT_SYMBOL(wait_for_completion_timeout);
4872 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4873 * @x: holds the state of this particular completion
4875 * This waits for completion of a specific task to be signaled. It is
4878 int __sched wait_for_completion_interruptible(struct completion *x)
4880 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4881 if (t == -ERESTARTSYS)
4885 EXPORT_SYMBOL(wait_for_completion_interruptible);
4888 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4889 * @x: holds the state of this particular completion
4890 * @timeout: timeout value in jiffies
4892 * This waits for either a completion of a specific task to be signaled or for a
4893 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4895 unsigned long __sched
4896 wait_for_completion_interruptible_timeout(struct completion *x,
4897 unsigned long timeout)
4899 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4901 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4904 * wait_for_completion_killable: - waits for completion of a task (killable)
4905 * @x: holds the state of this particular completion
4907 * This waits to be signaled for completion of a specific task. It can be
4908 * interrupted by a kill signal.
4910 int __sched wait_for_completion_killable(struct completion *x)
4912 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4913 if (t == -ERESTARTSYS)
4917 EXPORT_SYMBOL(wait_for_completion_killable);
4920 * try_wait_for_completion - try to decrement a completion without blocking
4921 * @x: completion structure
4923 * Returns: 0 if a decrement cannot be done without blocking
4924 * 1 if a decrement succeeded.
4926 * If a completion is being used as a counting completion,
4927 * attempt to decrement the counter without blocking. This
4928 * enables us to avoid waiting if the resource the completion
4929 * is protecting is not available.
4931 bool try_wait_for_completion(struct completion *x)
4935 spin_lock_irq(&x->wait.lock);
4940 spin_unlock_irq(&x->wait.lock);
4943 EXPORT_SYMBOL(try_wait_for_completion);
4946 * completion_done - Test to see if a completion has any waiters
4947 * @x: completion structure
4949 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4950 * 1 if there are no waiters.
4953 bool completion_done(struct completion *x)
4957 spin_lock_irq(&x->wait.lock);
4960 spin_unlock_irq(&x->wait.lock);
4963 EXPORT_SYMBOL(completion_done);
4966 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4968 unsigned long flags;
4971 init_waitqueue_entry(&wait, current);
4973 __set_current_state(state);
4975 spin_lock_irqsave(&q->lock, flags);
4976 __add_wait_queue(q, &wait);
4977 spin_unlock(&q->lock);
4978 timeout = schedule_timeout(timeout);
4979 spin_lock_irq(&q->lock);
4980 __remove_wait_queue(q, &wait);
4981 spin_unlock_irqrestore(&q->lock, flags);
4986 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4988 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4990 EXPORT_SYMBOL(interruptible_sleep_on);
4993 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4995 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4997 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4999 void __sched sleep_on(wait_queue_head_t *q)
5001 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5003 EXPORT_SYMBOL(sleep_on);
5005 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5007 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5009 EXPORT_SYMBOL(sleep_on_timeout);
5011 #ifdef CONFIG_RT_MUTEXES
5014 * rt_mutex_setprio - set the current priority of a task
5016 * @prio: prio value (kernel-internal form)
5018 * This function changes the 'effective' priority of a task. It does
5019 * not touch ->normal_prio like __setscheduler().
5021 * Used by the rt_mutex code to implement priority inheritance logic.
5023 void rt_mutex_setprio(struct task_struct *p, int prio)
5025 unsigned long flags;
5026 int oldprio, on_rq, running;
5028 const struct sched_class *prev_class = p->sched_class;
5030 BUG_ON(prio < 0 || prio > MAX_PRIO);
5032 rq = task_rq_lock(p, &flags);
5033 update_rq_clock(rq);
5036 on_rq = p->se.on_rq;
5037 running = task_current(rq, p);
5039 dequeue_task(rq, p, 0);
5041 p->sched_class->put_prev_task(rq, p);
5044 p->sched_class = &rt_sched_class;
5046 p->sched_class = &fair_sched_class;
5051 p->sched_class->set_curr_task(rq);
5053 enqueue_task(rq, p, 0);
5055 check_class_changed(rq, p, prev_class, oldprio, running);
5057 task_rq_unlock(rq, &flags);
5062 void set_user_nice(struct task_struct *p, long nice)
5064 int old_prio, delta, on_rq;
5065 unsigned long flags;
5068 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5071 * We have to be careful, if called from sys_setpriority(),
5072 * the task might be in the middle of scheduling on another CPU.
5074 rq = task_rq_lock(p, &flags);
5075 update_rq_clock(rq);
5077 * The RT priorities are set via sched_setscheduler(), but we still
5078 * allow the 'normal' nice value to be set - but as expected
5079 * it wont have any effect on scheduling until the task is
5080 * SCHED_FIFO/SCHED_RR:
5082 if (task_has_rt_policy(p)) {
5083 p->static_prio = NICE_TO_PRIO(nice);
5086 on_rq = p->se.on_rq;
5088 dequeue_task(rq, p, 0);
5090 p->static_prio = NICE_TO_PRIO(nice);
5093 p->prio = effective_prio(p);
5094 delta = p->prio - old_prio;
5097 enqueue_task(rq, p, 0);
5099 * If the task increased its priority or is running and
5100 * lowered its priority, then reschedule its CPU:
5102 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5103 resched_task(rq->curr);
5106 task_rq_unlock(rq, &flags);
5108 EXPORT_SYMBOL(set_user_nice);
5111 * can_nice - check if a task can reduce its nice value
5115 int can_nice(const struct task_struct *p, const int nice)
5117 /* convert nice value [19,-20] to rlimit style value [1,40] */
5118 int nice_rlim = 20 - nice;
5120 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5121 capable(CAP_SYS_NICE));
5124 #ifdef __ARCH_WANT_SYS_NICE
5127 * sys_nice - change the priority of the current process.
5128 * @increment: priority increment
5130 * sys_setpriority is a more generic, but much slower function that
5131 * does similar things.
5133 SYSCALL_DEFINE1(nice, int, increment)
5138 * Setpriority might change our priority at the same moment.
5139 * We don't have to worry. Conceptually one call occurs first
5140 * and we have a single winner.
5142 if (increment < -40)
5147 nice = PRIO_TO_NICE(current->static_prio) + increment;
5153 if (increment < 0 && !can_nice(current, nice))
5156 retval = security_task_setnice(current, nice);
5160 set_user_nice(current, nice);
5167 * task_prio - return the priority value of a given task.
5168 * @p: the task in question.
5170 * This is the priority value as seen by users in /proc.
5171 * RT tasks are offset by -200. Normal tasks are centered
5172 * around 0, value goes from -16 to +15.
5174 int task_prio(const struct task_struct *p)
5176 return p->prio - MAX_RT_PRIO;
5180 * task_nice - return the nice value of a given task.
5181 * @p: the task in question.
5183 int task_nice(const struct task_struct *p)
5185 return TASK_NICE(p);
5187 EXPORT_SYMBOL(task_nice);
5190 * idle_cpu - is a given cpu idle currently?
5191 * @cpu: the processor in question.
5193 int idle_cpu(int cpu)
5195 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5199 * idle_task - return the idle task for a given cpu.
5200 * @cpu: the processor in question.
5202 struct task_struct *idle_task(int cpu)
5204 return cpu_rq(cpu)->idle;
5208 * find_process_by_pid - find a process with a matching PID value.
5209 * @pid: the pid in question.
5211 static struct task_struct *find_process_by_pid(pid_t pid)
5213 return pid ? find_task_by_vpid(pid) : current;
5216 /* Actually do priority change: must hold rq lock. */
5218 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5220 BUG_ON(p->se.on_rq);
5223 switch (p->policy) {
5227 p->sched_class = &fair_sched_class;
5231 p->sched_class = &rt_sched_class;
5235 p->rt_priority = prio;
5236 p->normal_prio = normal_prio(p);
5237 /* we are holding p->pi_lock already */
5238 p->prio = rt_mutex_getprio(p);
5243 * check the target process has a UID that matches the current process's
5245 static bool check_same_owner(struct task_struct *p)
5247 const struct cred *cred = current_cred(), *pcred;
5251 pcred = __task_cred(p);
5252 match = (cred->euid == pcred->euid ||
5253 cred->euid == pcred->uid);
5258 static int __sched_setscheduler(struct task_struct *p, int policy,
5259 struct sched_param *param, bool user)
5261 int retval, oldprio, oldpolicy = -1, on_rq, running;
5262 unsigned long flags;
5263 const struct sched_class *prev_class = p->sched_class;
5266 /* may grab non-irq protected spin_locks */
5267 BUG_ON(in_interrupt());
5269 /* double check policy once rq lock held */
5271 policy = oldpolicy = p->policy;
5272 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5273 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5274 policy != SCHED_IDLE)
5277 * Valid priorities for SCHED_FIFO and SCHED_RR are
5278 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5279 * SCHED_BATCH and SCHED_IDLE is 0.
5281 if (param->sched_priority < 0 ||
5282 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5283 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5285 if (rt_policy(policy) != (param->sched_priority != 0))
5289 * Allow unprivileged RT tasks to decrease priority:
5291 if (user && !capable(CAP_SYS_NICE)) {
5292 if (rt_policy(policy)) {
5293 unsigned long rlim_rtprio;
5295 if (!lock_task_sighand(p, &flags))
5297 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5298 unlock_task_sighand(p, &flags);
5300 /* can't set/change the rt policy */
5301 if (policy != p->policy && !rlim_rtprio)
5304 /* can't increase priority */
5305 if (param->sched_priority > p->rt_priority &&
5306 param->sched_priority > rlim_rtprio)
5310 * Like positive nice levels, dont allow tasks to
5311 * move out of SCHED_IDLE either:
5313 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5316 /* can't change other user's priorities */
5317 if (!check_same_owner(p))
5322 #ifdef CONFIG_RT_GROUP_SCHED
5324 * Do not allow realtime tasks into groups that have no runtime
5327 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5328 task_group(p)->rt_bandwidth.rt_runtime == 0)
5332 retval = security_task_setscheduler(p, policy, param);
5338 * make sure no PI-waiters arrive (or leave) while we are
5339 * changing the priority of the task:
5341 spin_lock_irqsave(&p->pi_lock, flags);
5343 * To be able to change p->policy safely, the apropriate
5344 * runqueue lock must be held.
5346 rq = __task_rq_lock(p);
5347 /* recheck policy now with rq lock held */
5348 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5349 policy = oldpolicy = -1;
5350 __task_rq_unlock(rq);
5351 spin_unlock_irqrestore(&p->pi_lock, flags);
5354 update_rq_clock(rq);
5355 on_rq = p->se.on_rq;
5356 running = task_current(rq, p);
5358 deactivate_task(rq, p, 0);
5360 p->sched_class->put_prev_task(rq, p);
5363 __setscheduler(rq, p, policy, param->sched_priority);
5366 p->sched_class->set_curr_task(rq);
5368 activate_task(rq, p, 0);
5370 check_class_changed(rq, p, prev_class, oldprio, running);
5372 __task_rq_unlock(rq);
5373 spin_unlock_irqrestore(&p->pi_lock, flags);
5375 rt_mutex_adjust_pi(p);
5381 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5382 * @p: the task in question.
5383 * @policy: new policy.
5384 * @param: structure containing the new RT priority.
5386 * NOTE that the task may be already dead.
5388 int sched_setscheduler(struct task_struct *p, int policy,
5389 struct sched_param *param)
5391 return __sched_setscheduler(p, policy, param, true);
5393 EXPORT_SYMBOL_GPL(sched_setscheduler);
5396 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5397 * @p: the task in question.
5398 * @policy: new policy.
5399 * @param: structure containing the new RT priority.
5401 * Just like sched_setscheduler, only don't bother checking if the
5402 * current context has permission. For example, this is needed in
5403 * stop_machine(): we create temporary high priority worker threads,
5404 * but our caller might not have that capability.
5406 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5407 struct sched_param *param)
5409 return __sched_setscheduler(p, policy, param, false);
5413 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5415 struct sched_param lparam;
5416 struct task_struct *p;
5419 if (!param || pid < 0)
5421 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5426 p = find_process_by_pid(pid);
5428 retval = sched_setscheduler(p, policy, &lparam);
5435 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5436 * @pid: the pid in question.
5437 * @policy: new policy.
5438 * @param: structure containing the new RT priority.
5440 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5441 struct sched_param __user *, param)
5443 /* negative values for policy are not valid */
5447 return do_sched_setscheduler(pid, policy, param);
5451 * sys_sched_setparam - set/change the RT priority of a thread
5452 * @pid: the pid in question.
5453 * @param: structure containing the new RT priority.
5455 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5457 return do_sched_setscheduler(pid, -1, param);
5461 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5462 * @pid: the pid in question.
5464 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5466 struct task_struct *p;
5473 read_lock(&tasklist_lock);
5474 p = find_process_by_pid(pid);
5476 retval = security_task_getscheduler(p);
5480 read_unlock(&tasklist_lock);
5485 * sys_sched_getscheduler - get the RT priority of a thread
5486 * @pid: the pid in question.
5487 * @param: structure containing the RT priority.
5489 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5491 struct sched_param lp;
5492 struct task_struct *p;
5495 if (!param || pid < 0)
5498 read_lock(&tasklist_lock);
5499 p = find_process_by_pid(pid);
5504 retval = security_task_getscheduler(p);
5508 lp.sched_priority = p->rt_priority;
5509 read_unlock(&tasklist_lock);
5512 * This one might sleep, we cannot do it with a spinlock held ...
5514 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5519 read_unlock(&tasklist_lock);
5523 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5525 cpumask_var_t cpus_allowed, new_mask;
5526 struct task_struct *p;
5530 read_lock(&tasklist_lock);
5532 p = find_process_by_pid(pid);
5534 read_unlock(&tasklist_lock);
5540 * It is not safe to call set_cpus_allowed with the
5541 * tasklist_lock held. We will bump the task_struct's
5542 * usage count and then drop tasklist_lock.
5545 read_unlock(&tasklist_lock);
5547 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5551 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5553 goto out_free_cpus_allowed;
5556 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5559 retval = security_task_setscheduler(p, 0, NULL);
5563 cpuset_cpus_allowed(p, cpus_allowed);
5564 cpumask_and(new_mask, in_mask, cpus_allowed);
5566 retval = set_cpus_allowed_ptr(p, new_mask);
5569 cpuset_cpus_allowed(p, cpus_allowed);
5570 if (!cpumask_subset(new_mask, cpus_allowed)) {
5572 * We must have raced with a concurrent cpuset
5573 * update. Just reset the cpus_allowed to the
5574 * cpuset's cpus_allowed
5576 cpumask_copy(new_mask, cpus_allowed);
5581 free_cpumask_var(new_mask);
5582 out_free_cpus_allowed:
5583 free_cpumask_var(cpus_allowed);
5590 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5591 struct cpumask *new_mask)
5593 if (len < cpumask_size())
5594 cpumask_clear(new_mask);
5595 else if (len > cpumask_size())
5596 len = cpumask_size();
5598 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5602 * sys_sched_setaffinity - set the cpu affinity of a process
5603 * @pid: pid of the process
5604 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5605 * @user_mask_ptr: user-space pointer to the new cpu mask
5607 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5608 unsigned long __user *, user_mask_ptr)
5610 cpumask_var_t new_mask;
5613 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5616 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5618 retval = sched_setaffinity(pid, new_mask);
5619 free_cpumask_var(new_mask);
5623 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5625 struct task_struct *p;
5629 read_lock(&tasklist_lock);
5632 p = find_process_by_pid(pid);
5636 retval = security_task_getscheduler(p);
5640 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5643 read_unlock(&tasklist_lock);
5650 * sys_sched_getaffinity - get the cpu affinity of a process
5651 * @pid: pid of the process
5652 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5653 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5655 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5656 unsigned long __user *, user_mask_ptr)
5661 if (len < cpumask_size())
5664 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5667 ret = sched_getaffinity(pid, mask);
5669 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5672 ret = cpumask_size();
5674 free_cpumask_var(mask);
5680 * sys_sched_yield - yield the current processor to other threads.
5682 * This function yields the current CPU to other tasks. If there are no
5683 * other threads running on this CPU then this function will return.
5685 SYSCALL_DEFINE0(sched_yield)
5687 struct rq *rq = this_rq_lock();
5689 schedstat_inc(rq, yld_count);
5690 current->sched_class->yield_task(rq);
5693 * Since we are going to call schedule() anyway, there's
5694 * no need to preempt or enable interrupts:
5696 __release(rq->lock);
5697 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5698 _raw_spin_unlock(&rq->lock);
5699 preempt_enable_no_resched();
5706 static void __cond_resched(void)
5708 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5709 __might_sleep(__FILE__, __LINE__);
5712 * The BKS might be reacquired before we have dropped
5713 * PREEMPT_ACTIVE, which could trigger a second
5714 * cond_resched() call.
5717 add_preempt_count(PREEMPT_ACTIVE);
5719 sub_preempt_count(PREEMPT_ACTIVE);
5720 } while (need_resched());
5723 int __sched _cond_resched(void)
5725 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5726 system_state == SYSTEM_RUNNING) {
5732 EXPORT_SYMBOL(_cond_resched);
5735 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5736 * call schedule, and on return reacquire the lock.
5738 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5739 * operations here to prevent schedule() from being called twice (once via
5740 * spin_unlock(), once by hand).
5742 int cond_resched_lock(spinlock_t *lock)
5744 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5747 if (spin_needbreak(lock) || resched) {
5749 if (resched && need_resched())
5758 EXPORT_SYMBOL(cond_resched_lock);
5760 int __sched cond_resched_softirq(void)
5762 BUG_ON(!in_softirq());
5764 if (need_resched() && system_state == SYSTEM_RUNNING) {
5772 EXPORT_SYMBOL(cond_resched_softirq);
5775 * yield - yield the current processor to other threads.
5777 * This is a shortcut for kernel-space yielding - it marks the
5778 * thread runnable and calls sys_sched_yield().
5780 void __sched yield(void)
5782 set_current_state(TASK_RUNNING);
5785 EXPORT_SYMBOL(yield);
5788 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5789 * that process accounting knows that this is a task in IO wait state.
5791 * But don't do that if it is a deliberate, throttling IO wait (this task
5792 * has set its backing_dev_info: the queue against which it should throttle)
5794 void __sched io_schedule(void)
5796 struct rq *rq = &__raw_get_cpu_var(runqueues);
5798 delayacct_blkio_start();
5799 atomic_inc(&rq->nr_iowait);
5801 atomic_dec(&rq->nr_iowait);
5802 delayacct_blkio_end();
5804 EXPORT_SYMBOL(io_schedule);
5806 long __sched io_schedule_timeout(long timeout)
5808 struct rq *rq = &__raw_get_cpu_var(runqueues);
5811 delayacct_blkio_start();
5812 atomic_inc(&rq->nr_iowait);
5813 ret = schedule_timeout(timeout);
5814 atomic_dec(&rq->nr_iowait);
5815 delayacct_blkio_end();
5820 * sys_sched_get_priority_max - return maximum RT priority.
5821 * @policy: scheduling class.
5823 * this syscall returns the maximum rt_priority that can be used
5824 * by a given scheduling class.
5826 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5833 ret = MAX_USER_RT_PRIO-1;
5845 * sys_sched_get_priority_min - return minimum RT priority.
5846 * @policy: scheduling class.
5848 * this syscall returns the minimum rt_priority that can be used
5849 * by a given scheduling class.
5851 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
5869 * sys_sched_rr_get_interval - return the default timeslice of a process.
5870 * @pid: pid of the process.
5871 * @interval: userspace pointer to the timeslice value.
5873 * this syscall writes the default timeslice value of a given process
5874 * into the user-space timespec buffer. A value of '0' means infinity.
5876 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
5877 struct timespec __user *, interval)
5879 struct task_struct *p;
5880 unsigned int time_slice;
5888 read_lock(&tasklist_lock);
5889 p = find_process_by_pid(pid);
5893 retval = security_task_getscheduler(p);
5898 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5899 * tasks that are on an otherwise idle runqueue:
5902 if (p->policy == SCHED_RR) {
5903 time_slice = DEF_TIMESLICE;
5904 } else if (p->policy != SCHED_FIFO) {
5905 struct sched_entity *se = &p->se;
5906 unsigned long flags;
5909 rq = task_rq_lock(p, &flags);
5910 if (rq->cfs.load.weight)
5911 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5912 task_rq_unlock(rq, &flags);
5914 read_unlock(&tasklist_lock);
5915 jiffies_to_timespec(time_slice, &t);
5916 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5920 read_unlock(&tasklist_lock);
5924 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5926 void sched_show_task(struct task_struct *p)
5928 unsigned long free = 0;
5931 state = p->state ? __ffs(p->state) + 1 : 0;
5932 printk(KERN_INFO "%-13.13s %c", p->comm,
5933 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5934 #if BITS_PER_LONG == 32
5935 if (state == TASK_RUNNING)
5936 printk(KERN_CONT " running ");
5938 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5940 if (state == TASK_RUNNING)
5941 printk(KERN_CONT " running task ");
5943 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5945 #ifdef CONFIG_DEBUG_STACK_USAGE
5947 unsigned long *n = end_of_stack(p);
5950 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5953 printk(KERN_CONT "%5lu %5d %6d\n", free,
5954 task_pid_nr(p), task_pid_nr(p->real_parent));
5956 show_stack(p, NULL);
5959 void show_state_filter(unsigned long state_filter)
5961 struct task_struct *g, *p;
5963 #if BITS_PER_LONG == 32
5965 " task PC stack pid father\n");
5968 " task PC stack pid father\n");
5970 read_lock(&tasklist_lock);
5971 do_each_thread(g, p) {
5973 * reset the NMI-timeout, listing all files on a slow
5974 * console might take alot of time:
5976 touch_nmi_watchdog();
5977 if (!state_filter || (p->state & state_filter))
5979 } while_each_thread(g, p);
5981 touch_all_softlockup_watchdogs();
5983 #ifdef CONFIG_SCHED_DEBUG
5984 sysrq_sched_debug_show();
5986 read_unlock(&tasklist_lock);
5988 * Only show locks if all tasks are dumped:
5990 if (state_filter == -1)
5991 debug_show_all_locks();
5994 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5996 idle->sched_class = &idle_sched_class;
6000 * init_idle - set up an idle thread for a given CPU
6001 * @idle: task in question
6002 * @cpu: cpu the idle task belongs to
6004 * NOTE: this function does not set the idle thread's NEED_RESCHED
6005 * flag, to make booting more robust.
6007 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6009 struct rq *rq = cpu_rq(cpu);
6010 unsigned long flags;
6012 spin_lock_irqsave(&rq->lock, flags);
6015 idle->se.exec_start = sched_clock();
6017 idle->prio = idle->normal_prio = MAX_PRIO;
6018 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6019 __set_task_cpu(idle, cpu);
6021 rq->curr = rq->idle = idle;
6022 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6025 spin_unlock_irqrestore(&rq->lock, flags);
6027 /* Set the preempt count _outside_ the spinlocks! */
6028 #if defined(CONFIG_PREEMPT)
6029 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6031 task_thread_info(idle)->preempt_count = 0;
6034 * The idle tasks have their own, simple scheduling class:
6036 idle->sched_class = &idle_sched_class;
6037 ftrace_graph_init_task(idle);
6041 * In a system that switches off the HZ timer nohz_cpu_mask
6042 * indicates which cpus entered this state. This is used
6043 * in the rcu update to wait only for active cpus. For system
6044 * which do not switch off the HZ timer nohz_cpu_mask should
6045 * always be CPU_BITS_NONE.
6047 cpumask_var_t nohz_cpu_mask;
6050 * Increase the granularity value when there are more CPUs,
6051 * because with more CPUs the 'effective latency' as visible
6052 * to users decreases. But the relationship is not linear,
6053 * so pick a second-best guess by going with the log2 of the
6056 * This idea comes from the SD scheduler of Con Kolivas:
6058 static inline void sched_init_granularity(void)
6060 unsigned int factor = 1 + ilog2(num_online_cpus());
6061 const unsigned long limit = 200000000;
6063 sysctl_sched_min_granularity *= factor;
6064 if (sysctl_sched_min_granularity > limit)
6065 sysctl_sched_min_granularity = limit;
6067 sysctl_sched_latency *= factor;
6068 if (sysctl_sched_latency > limit)
6069 sysctl_sched_latency = limit;
6071 sysctl_sched_wakeup_granularity *= factor;
6073 sysctl_sched_shares_ratelimit *= factor;
6078 * This is how migration works:
6080 * 1) we queue a struct migration_req structure in the source CPU's
6081 * runqueue and wake up that CPU's migration thread.
6082 * 2) we down() the locked semaphore => thread blocks.
6083 * 3) migration thread wakes up (implicitly it forces the migrated
6084 * thread off the CPU)
6085 * 4) it gets the migration request and checks whether the migrated
6086 * task is still in the wrong runqueue.
6087 * 5) if it's in the wrong runqueue then the migration thread removes
6088 * it and puts it into the right queue.
6089 * 6) migration thread up()s the semaphore.
6090 * 7) we wake up and the migration is done.
6094 * Change a given task's CPU affinity. Migrate the thread to a
6095 * proper CPU and schedule it away if the CPU it's executing on
6096 * is removed from the allowed bitmask.
6098 * NOTE: the caller must have a valid reference to the task, the
6099 * task must not exit() & deallocate itself prematurely. The
6100 * call is not atomic; no spinlocks may be held.
6102 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6104 struct migration_req req;
6105 unsigned long flags;
6109 rq = task_rq_lock(p, &flags);
6110 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6115 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6116 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6121 if (p->sched_class->set_cpus_allowed)
6122 p->sched_class->set_cpus_allowed(p, new_mask);
6124 cpumask_copy(&p->cpus_allowed, new_mask);
6125 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6128 /* Can the task run on the task's current CPU? If so, we're done */
6129 if (cpumask_test_cpu(task_cpu(p), new_mask))
6132 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6133 /* Need help from migration thread: drop lock and wait. */
6134 task_rq_unlock(rq, &flags);
6135 wake_up_process(rq->migration_thread);
6136 wait_for_completion(&req.done);
6137 tlb_migrate_finish(p->mm);
6141 task_rq_unlock(rq, &flags);
6145 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6148 * Move (not current) task off this cpu, onto dest cpu. We're doing
6149 * this because either it can't run here any more (set_cpus_allowed()
6150 * away from this CPU, or CPU going down), or because we're
6151 * attempting to rebalance this task on exec (sched_exec).
6153 * So we race with normal scheduler movements, but that's OK, as long
6154 * as the task is no longer on this CPU.
6156 * Returns non-zero if task was successfully migrated.
6158 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6160 struct rq *rq_dest, *rq_src;
6163 if (unlikely(!cpu_active(dest_cpu)))
6166 rq_src = cpu_rq(src_cpu);
6167 rq_dest = cpu_rq(dest_cpu);
6169 double_rq_lock(rq_src, rq_dest);
6170 /* Already moved. */
6171 if (task_cpu(p) != src_cpu)
6173 /* Affinity changed (again). */
6174 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6177 on_rq = p->se.on_rq;
6179 deactivate_task(rq_src, p, 0);
6181 set_task_cpu(p, dest_cpu);
6183 activate_task(rq_dest, p, 0);
6184 check_preempt_curr(rq_dest, p, 0);
6189 double_rq_unlock(rq_src, rq_dest);
6194 * migration_thread - this is a highprio system thread that performs
6195 * thread migration by bumping thread off CPU then 'pushing' onto
6198 static int migration_thread(void *data)
6200 int cpu = (long)data;
6204 BUG_ON(rq->migration_thread != current);
6206 set_current_state(TASK_INTERRUPTIBLE);
6207 while (!kthread_should_stop()) {
6208 struct migration_req *req;
6209 struct list_head *head;
6211 spin_lock_irq(&rq->lock);
6213 if (cpu_is_offline(cpu)) {
6214 spin_unlock_irq(&rq->lock);
6218 if (rq->active_balance) {
6219 active_load_balance(rq, cpu);
6220 rq->active_balance = 0;
6223 head = &rq->migration_queue;
6225 if (list_empty(head)) {
6226 spin_unlock_irq(&rq->lock);
6228 set_current_state(TASK_INTERRUPTIBLE);
6231 req = list_entry(head->next, struct migration_req, list);
6232 list_del_init(head->next);
6234 spin_unlock(&rq->lock);
6235 __migrate_task(req->task, cpu, req->dest_cpu);
6238 complete(&req->done);
6240 __set_current_state(TASK_RUNNING);
6244 /* Wait for kthread_stop */
6245 set_current_state(TASK_INTERRUPTIBLE);
6246 while (!kthread_should_stop()) {
6248 set_current_state(TASK_INTERRUPTIBLE);
6250 __set_current_state(TASK_RUNNING);
6254 #ifdef CONFIG_HOTPLUG_CPU
6256 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6260 local_irq_disable();
6261 ret = __migrate_task(p, src_cpu, dest_cpu);
6267 * Figure out where task on dead CPU should go, use force if necessary.
6269 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6272 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6275 /* Look for allowed, online CPU in same node. */
6276 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6277 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6280 /* Any allowed, online CPU? */
6281 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6282 if (dest_cpu < nr_cpu_ids)
6285 /* No more Mr. Nice Guy. */
6286 if (dest_cpu >= nr_cpu_ids) {
6287 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6288 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6291 * Don't tell them about moving exiting tasks or
6292 * kernel threads (both mm NULL), since they never
6295 if (p->mm && printk_ratelimit()) {
6296 printk(KERN_INFO "process %d (%s) no "
6297 "longer affine to cpu%d\n",
6298 task_pid_nr(p), p->comm, dead_cpu);
6303 /* It can have affinity changed while we were choosing. */
6304 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6309 * While a dead CPU has no uninterruptible tasks queued at this point,
6310 * it might still have a nonzero ->nr_uninterruptible counter, because
6311 * for performance reasons the counter is not stricly tracking tasks to
6312 * their home CPUs. So we just add the counter to another CPU's counter,
6313 * to keep the global sum constant after CPU-down:
6315 static void migrate_nr_uninterruptible(struct rq *rq_src)
6317 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6318 unsigned long flags;
6320 local_irq_save(flags);
6321 double_rq_lock(rq_src, rq_dest);
6322 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6323 rq_src->nr_uninterruptible = 0;
6324 double_rq_unlock(rq_src, rq_dest);
6325 local_irq_restore(flags);
6328 /* Run through task list and migrate tasks from the dead cpu. */
6329 static void migrate_live_tasks(int src_cpu)
6331 struct task_struct *p, *t;
6333 read_lock(&tasklist_lock);
6335 do_each_thread(t, p) {
6339 if (task_cpu(p) == src_cpu)
6340 move_task_off_dead_cpu(src_cpu, p);
6341 } while_each_thread(t, p);
6343 read_unlock(&tasklist_lock);
6347 * Schedules idle task to be the next runnable task on current CPU.
6348 * It does so by boosting its priority to highest possible.
6349 * Used by CPU offline code.
6351 void sched_idle_next(void)
6353 int this_cpu = smp_processor_id();
6354 struct rq *rq = cpu_rq(this_cpu);
6355 struct task_struct *p = rq->idle;
6356 unsigned long flags;
6358 /* cpu has to be offline */
6359 BUG_ON(cpu_online(this_cpu));
6362 * Strictly not necessary since rest of the CPUs are stopped by now
6363 * and interrupts disabled on the current cpu.
6365 spin_lock_irqsave(&rq->lock, flags);
6367 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6369 update_rq_clock(rq);
6370 activate_task(rq, p, 0);
6372 spin_unlock_irqrestore(&rq->lock, flags);
6376 * Ensures that the idle task is using init_mm right before its cpu goes
6379 void idle_task_exit(void)
6381 struct mm_struct *mm = current->active_mm;
6383 BUG_ON(cpu_online(smp_processor_id()));
6386 switch_mm(mm, &init_mm, current);
6390 /* called under rq->lock with disabled interrupts */
6391 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6393 struct rq *rq = cpu_rq(dead_cpu);
6395 /* Must be exiting, otherwise would be on tasklist. */
6396 BUG_ON(!p->exit_state);
6398 /* Cannot have done final schedule yet: would have vanished. */
6399 BUG_ON(p->state == TASK_DEAD);
6404 * Drop lock around migration; if someone else moves it,
6405 * that's OK. No task can be added to this CPU, so iteration is
6408 spin_unlock_irq(&rq->lock);
6409 move_task_off_dead_cpu(dead_cpu, p);
6410 spin_lock_irq(&rq->lock);
6415 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6416 static void migrate_dead_tasks(unsigned int dead_cpu)
6418 struct rq *rq = cpu_rq(dead_cpu);
6419 struct task_struct *next;
6422 if (!rq->nr_running)
6424 update_rq_clock(rq);
6425 next = pick_next_task(rq, rq->curr);
6428 next->sched_class->put_prev_task(rq, next);
6429 migrate_dead(dead_cpu, next);
6433 #endif /* CONFIG_HOTPLUG_CPU */
6435 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6437 static struct ctl_table sd_ctl_dir[] = {
6439 .procname = "sched_domain",
6445 static struct ctl_table sd_ctl_root[] = {
6447 .ctl_name = CTL_KERN,
6448 .procname = "kernel",
6450 .child = sd_ctl_dir,
6455 static struct ctl_table *sd_alloc_ctl_entry(int n)
6457 struct ctl_table *entry =
6458 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6463 static void sd_free_ctl_entry(struct ctl_table **tablep)
6465 struct ctl_table *entry;
6468 * In the intermediate directories, both the child directory and
6469 * procname are dynamically allocated and could fail but the mode
6470 * will always be set. In the lowest directory the names are
6471 * static strings and all have proc handlers.
6473 for (entry = *tablep; entry->mode; entry++) {
6475 sd_free_ctl_entry(&entry->child);
6476 if (entry->proc_handler == NULL)
6477 kfree(entry->procname);
6485 set_table_entry(struct ctl_table *entry,
6486 const char *procname, void *data, int maxlen,
6487 mode_t mode, proc_handler *proc_handler)
6489 entry->procname = procname;
6491 entry->maxlen = maxlen;
6493 entry->proc_handler = proc_handler;
6496 static struct ctl_table *
6497 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6499 struct ctl_table *table = sd_alloc_ctl_entry(13);
6504 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6505 sizeof(long), 0644, proc_doulongvec_minmax);
6506 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6507 sizeof(long), 0644, proc_doulongvec_minmax);
6508 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6509 sizeof(int), 0644, proc_dointvec_minmax);
6510 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6511 sizeof(int), 0644, proc_dointvec_minmax);
6512 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6513 sizeof(int), 0644, proc_dointvec_minmax);
6514 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6515 sizeof(int), 0644, proc_dointvec_minmax);
6516 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6517 sizeof(int), 0644, proc_dointvec_minmax);
6518 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6519 sizeof(int), 0644, proc_dointvec_minmax);
6520 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6521 sizeof(int), 0644, proc_dointvec_minmax);
6522 set_table_entry(&table[9], "cache_nice_tries",
6523 &sd->cache_nice_tries,
6524 sizeof(int), 0644, proc_dointvec_minmax);
6525 set_table_entry(&table[10], "flags", &sd->flags,
6526 sizeof(int), 0644, proc_dointvec_minmax);
6527 set_table_entry(&table[11], "name", sd->name,
6528 CORENAME_MAX_SIZE, 0444, proc_dostring);
6529 /* &table[12] is terminator */
6534 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6536 struct ctl_table *entry, *table;
6537 struct sched_domain *sd;
6538 int domain_num = 0, i;
6541 for_each_domain(cpu, sd)
6543 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6548 for_each_domain(cpu, sd) {
6549 snprintf(buf, 32, "domain%d", i);
6550 entry->procname = kstrdup(buf, GFP_KERNEL);
6552 entry->child = sd_alloc_ctl_domain_table(sd);
6559 static struct ctl_table_header *sd_sysctl_header;
6560 static void register_sched_domain_sysctl(void)
6562 int i, cpu_num = num_online_cpus();
6563 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6566 WARN_ON(sd_ctl_dir[0].child);
6567 sd_ctl_dir[0].child = entry;
6572 for_each_online_cpu(i) {
6573 snprintf(buf, 32, "cpu%d", i);
6574 entry->procname = kstrdup(buf, GFP_KERNEL);
6576 entry->child = sd_alloc_ctl_cpu_table(i);
6580 WARN_ON(sd_sysctl_header);
6581 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6584 /* may be called multiple times per register */
6585 static void unregister_sched_domain_sysctl(void)
6587 if (sd_sysctl_header)
6588 unregister_sysctl_table(sd_sysctl_header);
6589 sd_sysctl_header = NULL;
6590 if (sd_ctl_dir[0].child)
6591 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6594 static void register_sched_domain_sysctl(void)
6597 static void unregister_sched_domain_sysctl(void)
6602 static void set_rq_online(struct rq *rq)
6605 const struct sched_class *class;
6607 cpumask_set_cpu(rq->cpu, rq->rd->online);
6610 for_each_class(class) {
6611 if (class->rq_online)
6612 class->rq_online(rq);
6617 static void set_rq_offline(struct rq *rq)
6620 const struct sched_class *class;
6622 for_each_class(class) {
6623 if (class->rq_offline)
6624 class->rq_offline(rq);
6627 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6633 * migration_call - callback that gets triggered when a CPU is added.
6634 * Here we can start up the necessary migration thread for the new CPU.
6636 static int __cpuinit
6637 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6639 struct task_struct *p;
6640 int cpu = (long)hcpu;
6641 unsigned long flags;
6646 case CPU_UP_PREPARE:
6647 case CPU_UP_PREPARE_FROZEN:
6648 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6651 kthread_bind(p, cpu);
6652 /* Must be high prio: stop_machine expects to yield to it. */
6653 rq = task_rq_lock(p, &flags);
6654 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6655 task_rq_unlock(rq, &flags);
6656 cpu_rq(cpu)->migration_thread = p;
6660 case CPU_ONLINE_FROZEN:
6661 /* Strictly unnecessary, as first user will wake it. */
6662 wake_up_process(cpu_rq(cpu)->migration_thread);
6664 /* Update our root-domain */
6666 spin_lock_irqsave(&rq->lock, flags);
6668 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6672 spin_unlock_irqrestore(&rq->lock, flags);
6675 #ifdef CONFIG_HOTPLUG_CPU
6676 case CPU_UP_CANCELED:
6677 case CPU_UP_CANCELED_FROZEN:
6678 if (!cpu_rq(cpu)->migration_thread)
6680 /* Unbind it from offline cpu so it can run. Fall thru. */
6681 kthread_bind(cpu_rq(cpu)->migration_thread,
6682 cpumask_any(cpu_online_mask));
6683 kthread_stop(cpu_rq(cpu)->migration_thread);
6684 cpu_rq(cpu)->migration_thread = NULL;
6688 case CPU_DEAD_FROZEN:
6689 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6690 migrate_live_tasks(cpu);
6692 kthread_stop(rq->migration_thread);
6693 rq->migration_thread = NULL;
6694 /* Idle task back to normal (off runqueue, low prio) */
6695 spin_lock_irq(&rq->lock);
6696 update_rq_clock(rq);
6697 deactivate_task(rq, rq->idle, 0);
6698 rq->idle->static_prio = MAX_PRIO;
6699 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6700 rq->idle->sched_class = &idle_sched_class;
6701 migrate_dead_tasks(cpu);
6702 spin_unlock_irq(&rq->lock);
6704 migrate_nr_uninterruptible(rq);
6705 BUG_ON(rq->nr_running != 0);
6708 * No need to migrate the tasks: it was best-effort if
6709 * they didn't take sched_hotcpu_mutex. Just wake up
6712 spin_lock_irq(&rq->lock);
6713 while (!list_empty(&rq->migration_queue)) {
6714 struct migration_req *req;
6716 req = list_entry(rq->migration_queue.next,
6717 struct migration_req, list);
6718 list_del_init(&req->list);
6719 spin_unlock_irq(&rq->lock);
6720 complete(&req->done);
6721 spin_lock_irq(&rq->lock);
6723 spin_unlock_irq(&rq->lock);
6727 case CPU_DYING_FROZEN:
6728 /* Update our root-domain */
6730 spin_lock_irqsave(&rq->lock, flags);
6732 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6735 spin_unlock_irqrestore(&rq->lock, flags);
6742 /* Register at highest priority so that task migration (migrate_all_tasks)
6743 * happens before everything else.
6745 static struct notifier_block __cpuinitdata migration_notifier = {
6746 .notifier_call = migration_call,
6750 static int __init migration_init(void)
6752 void *cpu = (void *)(long)smp_processor_id();
6755 /* Start one for the boot CPU: */
6756 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6757 BUG_ON(err == NOTIFY_BAD);
6758 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6759 register_cpu_notifier(&migration_notifier);
6763 early_initcall(migration_init);
6768 #ifdef CONFIG_SCHED_DEBUG
6770 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6771 struct cpumask *groupmask)
6773 struct sched_group *group = sd->groups;
6776 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6777 cpumask_clear(groupmask);
6779 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6781 if (!(sd->flags & SD_LOAD_BALANCE)) {
6782 printk("does not load-balance\n");
6784 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6789 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6791 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6792 printk(KERN_ERR "ERROR: domain->span does not contain "
6795 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6796 printk(KERN_ERR "ERROR: domain->groups does not contain"
6800 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6804 printk(KERN_ERR "ERROR: group is NULL\n");
6808 if (!group->__cpu_power) {
6809 printk(KERN_CONT "\n");
6810 printk(KERN_ERR "ERROR: domain->cpu_power not "
6815 if (!cpumask_weight(sched_group_cpus(group))) {
6816 printk(KERN_CONT "\n");
6817 printk(KERN_ERR "ERROR: empty group\n");
6821 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6822 printk(KERN_CONT "\n");
6823 printk(KERN_ERR "ERROR: repeated CPUs\n");
6827 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6829 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6830 printk(KERN_CONT " %s", str);
6832 group = group->next;
6833 } while (group != sd->groups);
6834 printk(KERN_CONT "\n");
6836 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6837 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6840 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6841 printk(KERN_ERR "ERROR: parent span is not a superset "
6842 "of domain->span\n");
6846 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6848 cpumask_var_t groupmask;
6852 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6856 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6858 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6859 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6864 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6871 free_cpumask_var(groupmask);
6873 #else /* !CONFIG_SCHED_DEBUG */
6874 # define sched_domain_debug(sd, cpu) do { } while (0)
6875 #endif /* CONFIG_SCHED_DEBUG */
6877 static int sd_degenerate(struct sched_domain *sd)
6879 if (cpumask_weight(sched_domain_span(sd)) == 1)
6882 /* Following flags need at least 2 groups */
6883 if (sd->flags & (SD_LOAD_BALANCE |
6884 SD_BALANCE_NEWIDLE |
6888 SD_SHARE_PKG_RESOURCES)) {
6889 if (sd->groups != sd->groups->next)
6893 /* Following flags don't use groups */
6894 if (sd->flags & (SD_WAKE_IDLE |
6903 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6905 unsigned long cflags = sd->flags, pflags = parent->flags;
6907 if (sd_degenerate(parent))
6910 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6913 /* Does parent contain flags not in child? */
6914 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6915 if (cflags & SD_WAKE_AFFINE)
6916 pflags &= ~SD_WAKE_BALANCE;
6917 /* Flags needing groups don't count if only 1 group in parent */
6918 if (parent->groups == parent->groups->next) {
6919 pflags &= ~(SD_LOAD_BALANCE |
6920 SD_BALANCE_NEWIDLE |
6924 SD_SHARE_PKG_RESOURCES);
6925 if (nr_node_ids == 1)
6926 pflags &= ~SD_SERIALIZE;
6928 if (~cflags & pflags)
6934 static void free_rootdomain(struct root_domain *rd)
6936 cpupri_cleanup(&rd->cpupri);
6938 free_cpumask_var(rd->rto_mask);
6939 free_cpumask_var(rd->online);
6940 free_cpumask_var(rd->span);
6944 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6946 unsigned long flags;
6948 spin_lock_irqsave(&rq->lock, flags);
6951 struct root_domain *old_rd = rq->rd;
6953 if (cpumask_test_cpu(rq->cpu, old_rd->online))
6956 cpumask_clear_cpu(rq->cpu, old_rd->span);
6958 if (atomic_dec_and_test(&old_rd->refcount))
6959 free_rootdomain(old_rd);
6962 atomic_inc(&rd->refcount);
6965 cpumask_set_cpu(rq->cpu, rd->span);
6966 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
6969 spin_unlock_irqrestore(&rq->lock, flags);
6972 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
6974 memset(rd, 0, sizeof(*rd));
6977 alloc_bootmem_cpumask_var(&def_root_domain.span);
6978 alloc_bootmem_cpumask_var(&def_root_domain.online);
6979 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
6980 cpupri_init(&rd->cpupri, true);
6984 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
6986 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
6988 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
6991 if (cpupri_init(&rd->cpupri, false) != 0)
6996 free_cpumask_var(rd->rto_mask);
6998 free_cpumask_var(rd->online);
7000 free_cpumask_var(rd->span);
7005 static void init_defrootdomain(void)
7007 init_rootdomain(&def_root_domain, true);
7009 atomic_set(&def_root_domain.refcount, 1);
7012 static struct root_domain *alloc_rootdomain(void)
7014 struct root_domain *rd;
7016 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7020 if (init_rootdomain(rd, false) != 0) {
7029 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7030 * hold the hotplug lock.
7033 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7035 struct rq *rq = cpu_rq(cpu);
7036 struct sched_domain *tmp;
7038 /* Remove the sched domains which do not contribute to scheduling. */
7039 for (tmp = sd; tmp; ) {
7040 struct sched_domain *parent = tmp->parent;
7044 if (sd_parent_degenerate(tmp, parent)) {
7045 tmp->parent = parent->parent;
7047 parent->parent->child = tmp;
7052 if (sd && sd_degenerate(sd)) {
7058 sched_domain_debug(sd, cpu);
7060 rq_attach_root(rq, rd);
7061 rcu_assign_pointer(rq->sd, sd);
7064 /* cpus with isolated domains */
7065 static cpumask_var_t cpu_isolated_map;
7067 /* Setup the mask of cpus configured for isolated domains */
7068 static int __init isolated_cpu_setup(char *str)
7070 cpulist_parse(str, cpu_isolated_map);
7074 __setup("isolcpus=", isolated_cpu_setup);
7077 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7078 * to a function which identifies what group(along with sched group) a CPU
7079 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7080 * (due to the fact that we keep track of groups covered with a struct cpumask).
7082 * init_sched_build_groups will build a circular linked list of the groups
7083 * covered by the given span, and will set each group's ->cpumask correctly,
7084 * and ->cpu_power to 0.
7087 init_sched_build_groups(const struct cpumask *span,
7088 const struct cpumask *cpu_map,
7089 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7090 struct sched_group **sg,
7091 struct cpumask *tmpmask),
7092 struct cpumask *covered, struct cpumask *tmpmask)
7094 struct sched_group *first = NULL, *last = NULL;
7097 cpumask_clear(covered);
7099 for_each_cpu(i, span) {
7100 struct sched_group *sg;
7101 int group = group_fn(i, cpu_map, &sg, tmpmask);
7104 if (cpumask_test_cpu(i, covered))
7107 cpumask_clear(sched_group_cpus(sg));
7108 sg->__cpu_power = 0;
7110 for_each_cpu(j, span) {
7111 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7114 cpumask_set_cpu(j, covered);
7115 cpumask_set_cpu(j, sched_group_cpus(sg));
7126 #define SD_NODES_PER_DOMAIN 16
7131 * find_next_best_node - find the next node to include in a sched_domain
7132 * @node: node whose sched_domain we're building
7133 * @used_nodes: nodes already in the sched_domain
7135 * Find the next node to include in a given scheduling domain. Simply
7136 * finds the closest node not already in the @used_nodes map.
7138 * Should use nodemask_t.
7140 static int find_next_best_node(int node, nodemask_t *used_nodes)
7142 int i, n, val, min_val, best_node = 0;
7146 for (i = 0; i < nr_node_ids; i++) {
7147 /* Start at @node */
7148 n = (node + i) % nr_node_ids;
7150 if (!nr_cpus_node(n))
7153 /* Skip already used nodes */
7154 if (node_isset(n, *used_nodes))
7157 /* Simple min distance search */
7158 val = node_distance(node, n);
7160 if (val < min_val) {
7166 node_set(best_node, *used_nodes);
7171 * sched_domain_node_span - get a cpumask for a node's sched_domain
7172 * @node: node whose cpumask we're constructing
7173 * @span: resulting cpumask
7175 * Given a node, construct a good cpumask for its sched_domain to span. It
7176 * should be one that prevents unnecessary balancing, but also spreads tasks
7179 static void sched_domain_node_span(int node, struct cpumask *span)
7181 nodemask_t used_nodes;
7184 cpumask_clear(span);
7185 nodes_clear(used_nodes);
7187 cpumask_or(span, span, cpumask_of_node(node));
7188 node_set(node, used_nodes);
7190 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7191 int next_node = find_next_best_node(node, &used_nodes);
7193 cpumask_or(span, span, cpumask_of_node(next_node));
7196 #endif /* CONFIG_NUMA */
7198 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7201 * The cpus mask in sched_group and sched_domain hangs off the end.
7202 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7203 * for nr_cpu_ids < CONFIG_NR_CPUS.
7205 struct static_sched_group {
7206 struct sched_group sg;
7207 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7210 struct static_sched_domain {
7211 struct sched_domain sd;
7212 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7216 * SMT sched-domains:
7218 #ifdef CONFIG_SCHED_SMT
7219 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7220 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7223 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7224 struct sched_group **sg, struct cpumask *unused)
7227 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7230 #endif /* CONFIG_SCHED_SMT */
7233 * multi-core sched-domains:
7235 #ifdef CONFIG_SCHED_MC
7236 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7237 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7238 #endif /* CONFIG_SCHED_MC */
7240 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7242 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7243 struct sched_group **sg, struct cpumask *mask)
7247 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7248 group = cpumask_first(mask);
7250 *sg = &per_cpu(sched_group_core, group).sg;
7253 #elif defined(CONFIG_SCHED_MC)
7255 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7256 struct sched_group **sg, struct cpumask *unused)
7259 *sg = &per_cpu(sched_group_core, cpu).sg;
7264 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7265 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7268 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7269 struct sched_group **sg, struct cpumask *mask)
7272 #ifdef CONFIG_SCHED_MC
7273 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7274 group = cpumask_first(mask);
7275 #elif defined(CONFIG_SCHED_SMT)
7276 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7277 group = cpumask_first(mask);
7282 *sg = &per_cpu(sched_group_phys, group).sg;
7288 * The init_sched_build_groups can't handle what we want to do with node
7289 * groups, so roll our own. Now each node has its own list of groups which
7290 * gets dynamically allocated.
7292 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7293 static struct sched_group ***sched_group_nodes_bycpu;
7295 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7296 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7298 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7299 struct sched_group **sg,
7300 struct cpumask *nodemask)
7304 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7305 group = cpumask_first(nodemask);
7308 *sg = &per_cpu(sched_group_allnodes, group).sg;
7312 static void init_numa_sched_groups_power(struct sched_group *group_head)
7314 struct sched_group *sg = group_head;
7320 for_each_cpu(j, sched_group_cpus(sg)) {
7321 struct sched_domain *sd;
7323 sd = &per_cpu(phys_domains, j).sd;
7324 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7326 * Only add "power" once for each
7332 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7335 } while (sg != group_head);
7337 #endif /* CONFIG_NUMA */
7340 /* Free memory allocated for various sched_group structures */
7341 static void free_sched_groups(const struct cpumask *cpu_map,
7342 struct cpumask *nodemask)
7346 for_each_cpu(cpu, cpu_map) {
7347 struct sched_group **sched_group_nodes
7348 = sched_group_nodes_bycpu[cpu];
7350 if (!sched_group_nodes)
7353 for (i = 0; i < nr_node_ids; i++) {
7354 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7356 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7357 if (cpumask_empty(nodemask))
7367 if (oldsg != sched_group_nodes[i])
7370 kfree(sched_group_nodes);
7371 sched_group_nodes_bycpu[cpu] = NULL;
7374 #else /* !CONFIG_NUMA */
7375 static void free_sched_groups(const struct cpumask *cpu_map,
7376 struct cpumask *nodemask)
7379 #endif /* CONFIG_NUMA */
7382 * Initialize sched groups cpu_power.
7384 * cpu_power indicates the capacity of sched group, which is used while
7385 * distributing the load between different sched groups in a sched domain.
7386 * Typically cpu_power for all the groups in a sched domain will be same unless
7387 * there are asymmetries in the topology. If there are asymmetries, group
7388 * having more cpu_power will pickup more load compared to the group having
7391 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7392 * the maximum number of tasks a group can handle in the presence of other idle
7393 * or lightly loaded groups in the same sched domain.
7395 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7397 struct sched_domain *child;
7398 struct sched_group *group;
7400 WARN_ON(!sd || !sd->groups);
7402 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7407 sd->groups->__cpu_power = 0;
7410 * For perf policy, if the groups in child domain share resources
7411 * (for example cores sharing some portions of the cache hierarchy
7412 * or SMT), then set this domain groups cpu_power such that each group
7413 * can handle only one task, when there are other idle groups in the
7414 * same sched domain.
7416 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7418 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7419 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7424 * add cpu_power of each child group to this groups cpu_power
7426 group = child->groups;
7428 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7429 group = group->next;
7430 } while (group != child->groups);
7434 * Initializers for schedule domains
7435 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7438 #ifdef CONFIG_SCHED_DEBUG
7439 # define SD_INIT_NAME(sd, type) sd->name = #type
7441 # define SD_INIT_NAME(sd, type) do { } while (0)
7444 #define SD_INIT(sd, type) sd_init_##type(sd)
7446 #define SD_INIT_FUNC(type) \
7447 static noinline void sd_init_##type(struct sched_domain *sd) \
7449 memset(sd, 0, sizeof(*sd)); \
7450 *sd = SD_##type##_INIT; \
7451 sd->level = SD_LV_##type; \
7452 SD_INIT_NAME(sd, type); \
7457 SD_INIT_FUNC(ALLNODES)
7460 #ifdef CONFIG_SCHED_SMT
7461 SD_INIT_FUNC(SIBLING)
7463 #ifdef CONFIG_SCHED_MC
7467 static int default_relax_domain_level = -1;
7469 static int __init setup_relax_domain_level(char *str)
7473 val = simple_strtoul(str, NULL, 0);
7474 if (val < SD_LV_MAX)
7475 default_relax_domain_level = val;
7479 __setup("relax_domain_level=", setup_relax_domain_level);
7481 static void set_domain_attribute(struct sched_domain *sd,
7482 struct sched_domain_attr *attr)
7486 if (!attr || attr->relax_domain_level < 0) {
7487 if (default_relax_domain_level < 0)
7490 request = default_relax_domain_level;
7492 request = attr->relax_domain_level;
7493 if (request < sd->level) {
7494 /* turn off idle balance on this domain */
7495 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7497 /* turn on idle balance on this domain */
7498 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7503 * Build sched domains for a given set of cpus and attach the sched domains
7504 * to the individual cpus
7506 static int __build_sched_domains(const struct cpumask *cpu_map,
7507 struct sched_domain_attr *attr)
7509 int i, err = -ENOMEM;
7510 struct root_domain *rd;
7511 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7514 cpumask_var_t domainspan, covered, notcovered;
7515 struct sched_group **sched_group_nodes = NULL;
7516 int sd_allnodes = 0;
7518 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7520 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7521 goto free_domainspan;
7522 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7526 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7527 goto free_notcovered;
7528 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7530 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7531 goto free_this_sibling_map;
7532 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7533 goto free_this_core_map;
7534 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7535 goto free_send_covered;
7539 * Allocate the per-node list of sched groups
7541 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7543 if (!sched_group_nodes) {
7544 printk(KERN_WARNING "Can not alloc sched group node list\n");
7549 rd = alloc_rootdomain();
7551 printk(KERN_WARNING "Cannot alloc root domain\n");
7552 goto free_sched_groups;
7556 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7560 * Set up domains for cpus specified by the cpu_map.
7562 for_each_cpu(i, cpu_map) {
7563 struct sched_domain *sd = NULL, *p;
7565 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7568 if (cpumask_weight(cpu_map) >
7569 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7570 sd = &per_cpu(allnodes_domains, i).sd;
7571 SD_INIT(sd, ALLNODES);
7572 set_domain_attribute(sd, attr);
7573 cpumask_copy(sched_domain_span(sd), cpu_map);
7574 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7580 sd = &per_cpu(node_domains, i).sd;
7582 set_domain_attribute(sd, attr);
7583 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7587 cpumask_and(sched_domain_span(sd),
7588 sched_domain_span(sd), cpu_map);
7592 sd = &per_cpu(phys_domains, i).sd;
7594 set_domain_attribute(sd, attr);
7595 cpumask_copy(sched_domain_span(sd), nodemask);
7599 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7601 #ifdef CONFIG_SCHED_MC
7603 sd = &per_cpu(core_domains, i).sd;
7605 set_domain_attribute(sd, attr);
7606 cpumask_and(sched_domain_span(sd), cpu_map,
7607 cpu_coregroup_mask(i));
7610 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7613 #ifdef CONFIG_SCHED_SMT
7615 sd = &per_cpu(cpu_domains, i).sd;
7616 SD_INIT(sd, SIBLING);
7617 set_domain_attribute(sd, attr);
7618 cpumask_and(sched_domain_span(sd),
7619 &per_cpu(cpu_sibling_map, i), cpu_map);
7622 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7626 #ifdef CONFIG_SCHED_SMT
7627 /* Set up CPU (sibling) groups */
7628 for_each_cpu(i, cpu_map) {
7629 cpumask_and(this_sibling_map,
7630 &per_cpu(cpu_sibling_map, i), cpu_map);
7631 if (i != cpumask_first(this_sibling_map))
7634 init_sched_build_groups(this_sibling_map, cpu_map,
7636 send_covered, tmpmask);
7640 #ifdef CONFIG_SCHED_MC
7641 /* Set up multi-core groups */
7642 for_each_cpu(i, cpu_map) {
7643 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7644 if (i != cpumask_first(this_core_map))
7647 init_sched_build_groups(this_core_map, cpu_map,
7649 send_covered, tmpmask);
7653 /* Set up physical groups */
7654 for (i = 0; i < nr_node_ids; i++) {
7655 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7656 if (cpumask_empty(nodemask))
7659 init_sched_build_groups(nodemask, cpu_map,
7661 send_covered, tmpmask);
7665 /* Set up node groups */
7667 init_sched_build_groups(cpu_map, cpu_map,
7668 &cpu_to_allnodes_group,
7669 send_covered, tmpmask);
7672 for (i = 0; i < nr_node_ids; i++) {
7673 /* Set up node groups */
7674 struct sched_group *sg, *prev;
7677 cpumask_clear(covered);
7678 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7679 if (cpumask_empty(nodemask)) {
7680 sched_group_nodes[i] = NULL;
7684 sched_domain_node_span(i, domainspan);
7685 cpumask_and(domainspan, domainspan, cpu_map);
7687 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7690 printk(KERN_WARNING "Can not alloc domain group for "
7694 sched_group_nodes[i] = sg;
7695 for_each_cpu(j, nodemask) {
7696 struct sched_domain *sd;
7698 sd = &per_cpu(node_domains, j).sd;
7701 sg->__cpu_power = 0;
7702 cpumask_copy(sched_group_cpus(sg), nodemask);
7704 cpumask_or(covered, covered, nodemask);
7707 for (j = 0; j < nr_node_ids; j++) {
7708 int n = (i + j) % nr_node_ids;
7710 cpumask_complement(notcovered, covered);
7711 cpumask_and(tmpmask, notcovered, cpu_map);
7712 cpumask_and(tmpmask, tmpmask, domainspan);
7713 if (cpumask_empty(tmpmask))
7716 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7717 if (cpumask_empty(tmpmask))
7720 sg = kmalloc_node(sizeof(struct sched_group) +
7725 "Can not alloc domain group for node %d\n", j);
7728 sg->__cpu_power = 0;
7729 cpumask_copy(sched_group_cpus(sg), tmpmask);
7730 sg->next = prev->next;
7731 cpumask_or(covered, covered, tmpmask);
7738 /* Calculate CPU power for physical packages and nodes */
7739 #ifdef CONFIG_SCHED_SMT
7740 for_each_cpu(i, cpu_map) {
7741 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7743 init_sched_groups_power(i, sd);
7746 #ifdef CONFIG_SCHED_MC
7747 for_each_cpu(i, cpu_map) {
7748 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7750 init_sched_groups_power(i, sd);
7754 for_each_cpu(i, cpu_map) {
7755 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7757 init_sched_groups_power(i, sd);
7761 for (i = 0; i < nr_node_ids; i++)
7762 init_numa_sched_groups_power(sched_group_nodes[i]);
7765 struct sched_group *sg;
7767 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7769 init_numa_sched_groups_power(sg);
7773 /* Attach the domains */
7774 for_each_cpu(i, cpu_map) {
7775 struct sched_domain *sd;
7776 #ifdef CONFIG_SCHED_SMT
7777 sd = &per_cpu(cpu_domains, i).sd;
7778 #elif defined(CONFIG_SCHED_MC)
7779 sd = &per_cpu(core_domains, i).sd;
7781 sd = &per_cpu(phys_domains, i).sd;
7783 cpu_attach_domain(sd, rd, i);
7789 free_cpumask_var(tmpmask);
7791 free_cpumask_var(send_covered);
7793 free_cpumask_var(this_core_map);
7794 free_this_sibling_map:
7795 free_cpumask_var(this_sibling_map);
7797 free_cpumask_var(nodemask);
7800 free_cpumask_var(notcovered);
7802 free_cpumask_var(covered);
7804 free_cpumask_var(domainspan);
7811 kfree(sched_group_nodes);
7817 free_sched_groups(cpu_map, tmpmask);
7818 free_rootdomain(rd);
7823 static int build_sched_domains(const struct cpumask *cpu_map)
7825 return __build_sched_domains(cpu_map, NULL);
7828 static struct cpumask *doms_cur; /* current sched domains */
7829 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7830 static struct sched_domain_attr *dattr_cur;
7831 /* attribues of custom domains in 'doms_cur' */
7834 * Special case: If a kmalloc of a doms_cur partition (array of
7835 * cpumask) fails, then fallback to a single sched domain,
7836 * as determined by the single cpumask fallback_doms.
7838 static cpumask_var_t fallback_doms;
7841 * arch_update_cpu_topology lets virtualized architectures update the
7842 * cpu core maps. It is supposed to return 1 if the topology changed
7843 * or 0 if it stayed the same.
7845 int __attribute__((weak)) arch_update_cpu_topology(void)
7851 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7852 * For now this just excludes isolated cpus, but could be used to
7853 * exclude other special cases in the future.
7855 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7859 arch_update_cpu_topology();
7861 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7863 doms_cur = fallback_doms;
7864 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7866 err = build_sched_domains(doms_cur);
7867 register_sched_domain_sysctl();
7872 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7873 struct cpumask *tmpmask)
7875 free_sched_groups(cpu_map, tmpmask);
7879 * Detach sched domains from a group of cpus specified in cpu_map
7880 * These cpus will now be attached to the NULL domain
7882 static void detach_destroy_domains(const struct cpumask *cpu_map)
7884 /* Save because hotplug lock held. */
7885 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7888 for_each_cpu(i, cpu_map)
7889 cpu_attach_domain(NULL, &def_root_domain, i);
7890 synchronize_sched();
7891 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7894 /* handle null as "default" */
7895 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7896 struct sched_domain_attr *new, int idx_new)
7898 struct sched_domain_attr tmp;
7905 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7906 new ? (new + idx_new) : &tmp,
7907 sizeof(struct sched_domain_attr));
7911 * Partition sched domains as specified by the 'ndoms_new'
7912 * cpumasks in the array doms_new[] of cpumasks. This compares
7913 * doms_new[] to the current sched domain partitioning, doms_cur[].
7914 * It destroys each deleted domain and builds each new domain.
7916 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7917 * The masks don't intersect (don't overlap.) We should setup one
7918 * sched domain for each mask. CPUs not in any of the cpumasks will
7919 * not be load balanced. If the same cpumask appears both in the
7920 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7923 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7924 * ownership of it and will kfree it when done with it. If the caller
7925 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7926 * ndoms_new == 1, and partition_sched_domains() will fallback to
7927 * the single partition 'fallback_doms', it also forces the domains
7930 * If doms_new == NULL it will be replaced with cpu_online_mask.
7931 * ndoms_new == 0 is a special case for destroying existing domains,
7932 * and it will not create the default domain.
7934 * Call with hotplug lock held
7936 /* FIXME: Change to struct cpumask *doms_new[] */
7937 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
7938 struct sched_domain_attr *dattr_new)
7943 mutex_lock(&sched_domains_mutex);
7945 /* always unregister in case we don't destroy any domains */
7946 unregister_sched_domain_sysctl();
7948 /* Let architecture update cpu core mappings. */
7949 new_topology = arch_update_cpu_topology();
7951 n = doms_new ? ndoms_new : 0;
7953 /* Destroy deleted domains */
7954 for (i = 0; i < ndoms_cur; i++) {
7955 for (j = 0; j < n && !new_topology; j++) {
7956 if (cpumask_equal(&doms_cur[i], &doms_new[j])
7957 && dattrs_equal(dattr_cur, i, dattr_new, j))
7960 /* no match - a current sched domain not in new doms_new[] */
7961 detach_destroy_domains(doms_cur + i);
7966 if (doms_new == NULL) {
7968 doms_new = fallback_doms;
7969 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
7970 WARN_ON_ONCE(dattr_new);
7973 /* Build new domains */
7974 for (i = 0; i < ndoms_new; i++) {
7975 for (j = 0; j < ndoms_cur && !new_topology; j++) {
7976 if (cpumask_equal(&doms_new[i], &doms_cur[j])
7977 && dattrs_equal(dattr_new, i, dattr_cur, j))
7980 /* no match - add a new doms_new */
7981 __build_sched_domains(doms_new + i,
7982 dattr_new ? dattr_new + i : NULL);
7987 /* Remember the new sched domains */
7988 if (doms_cur != fallback_doms)
7990 kfree(dattr_cur); /* kfree(NULL) is safe */
7991 doms_cur = doms_new;
7992 dattr_cur = dattr_new;
7993 ndoms_cur = ndoms_new;
7995 register_sched_domain_sysctl();
7997 mutex_unlock(&sched_domains_mutex);
8000 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8001 static void arch_reinit_sched_domains(void)
8005 /* Destroy domains first to force the rebuild */
8006 partition_sched_domains(0, NULL, NULL);
8008 rebuild_sched_domains();
8012 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8014 unsigned int level = 0;
8016 if (sscanf(buf, "%u", &level) != 1)
8020 * level is always be positive so don't check for
8021 * level < POWERSAVINGS_BALANCE_NONE which is 0
8022 * What happens on 0 or 1 byte write,
8023 * need to check for count as well?
8026 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8030 sched_smt_power_savings = level;
8032 sched_mc_power_savings = level;
8034 arch_reinit_sched_domains();
8039 #ifdef CONFIG_SCHED_MC
8040 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8043 return sprintf(page, "%u\n", sched_mc_power_savings);
8045 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8046 const char *buf, size_t count)
8048 return sched_power_savings_store(buf, count, 0);
8050 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8051 sched_mc_power_savings_show,
8052 sched_mc_power_savings_store);
8055 #ifdef CONFIG_SCHED_SMT
8056 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8059 return sprintf(page, "%u\n", sched_smt_power_savings);
8061 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8062 const char *buf, size_t count)
8064 return sched_power_savings_store(buf, count, 1);
8066 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8067 sched_smt_power_savings_show,
8068 sched_smt_power_savings_store);
8071 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8075 #ifdef CONFIG_SCHED_SMT
8077 err = sysfs_create_file(&cls->kset.kobj,
8078 &attr_sched_smt_power_savings.attr);
8080 #ifdef CONFIG_SCHED_MC
8081 if (!err && mc_capable())
8082 err = sysfs_create_file(&cls->kset.kobj,
8083 &attr_sched_mc_power_savings.attr);
8087 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8089 #ifndef CONFIG_CPUSETS
8091 * Add online and remove offline CPUs from the scheduler domains.
8092 * When cpusets are enabled they take over this function.
8094 static int update_sched_domains(struct notifier_block *nfb,
8095 unsigned long action, void *hcpu)
8099 case CPU_ONLINE_FROZEN:
8101 case CPU_DEAD_FROZEN:
8102 partition_sched_domains(1, NULL, NULL);
8111 static int update_runtime(struct notifier_block *nfb,
8112 unsigned long action, void *hcpu)
8114 int cpu = (int)(long)hcpu;
8117 case CPU_DOWN_PREPARE:
8118 case CPU_DOWN_PREPARE_FROZEN:
8119 disable_runtime(cpu_rq(cpu));
8122 case CPU_DOWN_FAILED:
8123 case CPU_DOWN_FAILED_FROZEN:
8125 case CPU_ONLINE_FROZEN:
8126 enable_runtime(cpu_rq(cpu));
8134 void __init sched_init_smp(void)
8136 cpumask_var_t non_isolated_cpus;
8138 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8140 #if defined(CONFIG_NUMA)
8141 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8143 BUG_ON(sched_group_nodes_bycpu == NULL);
8146 mutex_lock(&sched_domains_mutex);
8147 arch_init_sched_domains(cpu_online_mask);
8148 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8149 if (cpumask_empty(non_isolated_cpus))
8150 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8151 mutex_unlock(&sched_domains_mutex);
8154 #ifndef CONFIG_CPUSETS
8155 /* XXX: Theoretical race here - CPU may be hotplugged now */
8156 hotcpu_notifier(update_sched_domains, 0);
8159 /* RT runtime code needs to handle some hotplug events */
8160 hotcpu_notifier(update_runtime, 0);
8164 /* Move init over to a non-isolated CPU */
8165 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8167 sched_init_granularity();
8168 free_cpumask_var(non_isolated_cpus);
8170 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8171 init_sched_rt_class();
8174 void __init sched_init_smp(void)
8176 sched_init_granularity();
8178 #endif /* CONFIG_SMP */
8180 int in_sched_functions(unsigned long addr)
8182 return in_lock_functions(addr) ||
8183 (addr >= (unsigned long)__sched_text_start
8184 && addr < (unsigned long)__sched_text_end);
8187 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8189 cfs_rq->tasks_timeline = RB_ROOT;
8190 INIT_LIST_HEAD(&cfs_rq->tasks);
8191 #ifdef CONFIG_FAIR_GROUP_SCHED
8194 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8197 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8199 struct rt_prio_array *array;
8202 array = &rt_rq->active;
8203 for (i = 0; i < MAX_RT_PRIO; i++) {
8204 INIT_LIST_HEAD(array->queue + i);
8205 __clear_bit(i, array->bitmap);
8207 /* delimiter for bitsearch: */
8208 __set_bit(MAX_RT_PRIO, array->bitmap);
8210 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8211 rt_rq->highest_prio = MAX_RT_PRIO;
8214 rt_rq->rt_nr_migratory = 0;
8215 rt_rq->overloaded = 0;
8219 rt_rq->rt_throttled = 0;
8220 rt_rq->rt_runtime = 0;
8221 spin_lock_init(&rt_rq->rt_runtime_lock);
8223 #ifdef CONFIG_RT_GROUP_SCHED
8224 rt_rq->rt_nr_boosted = 0;
8229 #ifdef CONFIG_FAIR_GROUP_SCHED
8230 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8231 struct sched_entity *se, int cpu, int add,
8232 struct sched_entity *parent)
8234 struct rq *rq = cpu_rq(cpu);
8235 tg->cfs_rq[cpu] = cfs_rq;
8236 init_cfs_rq(cfs_rq, rq);
8239 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8242 /* se could be NULL for init_task_group */
8247 se->cfs_rq = &rq->cfs;
8249 se->cfs_rq = parent->my_q;
8252 se->load.weight = tg->shares;
8253 se->load.inv_weight = 0;
8254 se->parent = parent;
8258 #ifdef CONFIG_RT_GROUP_SCHED
8259 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8260 struct sched_rt_entity *rt_se, int cpu, int add,
8261 struct sched_rt_entity *parent)
8263 struct rq *rq = cpu_rq(cpu);
8265 tg->rt_rq[cpu] = rt_rq;
8266 init_rt_rq(rt_rq, rq);
8268 rt_rq->rt_se = rt_se;
8269 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8271 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8273 tg->rt_se[cpu] = rt_se;
8278 rt_se->rt_rq = &rq->rt;
8280 rt_se->rt_rq = parent->my_q;
8282 rt_se->my_q = rt_rq;
8283 rt_se->parent = parent;
8284 INIT_LIST_HEAD(&rt_se->run_list);
8288 void __init sched_init(void)
8291 unsigned long alloc_size = 0, ptr;
8293 #ifdef CONFIG_FAIR_GROUP_SCHED
8294 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8296 #ifdef CONFIG_RT_GROUP_SCHED
8297 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8299 #ifdef CONFIG_USER_SCHED
8303 * As sched_init() is called before page_alloc is setup,
8304 * we use alloc_bootmem().
8307 ptr = (unsigned long)alloc_bootmem(alloc_size);
8309 #ifdef CONFIG_FAIR_GROUP_SCHED
8310 init_task_group.se = (struct sched_entity **)ptr;
8311 ptr += nr_cpu_ids * sizeof(void **);
8313 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8314 ptr += nr_cpu_ids * sizeof(void **);
8316 #ifdef CONFIG_USER_SCHED
8317 root_task_group.se = (struct sched_entity **)ptr;
8318 ptr += nr_cpu_ids * sizeof(void **);
8320 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8321 ptr += nr_cpu_ids * sizeof(void **);
8322 #endif /* CONFIG_USER_SCHED */
8323 #endif /* CONFIG_FAIR_GROUP_SCHED */
8324 #ifdef CONFIG_RT_GROUP_SCHED
8325 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8326 ptr += nr_cpu_ids * sizeof(void **);
8328 init_task_group.rt_rq = (struct rt_rq **)ptr;
8329 ptr += nr_cpu_ids * sizeof(void **);
8331 #ifdef CONFIG_USER_SCHED
8332 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8333 ptr += nr_cpu_ids * sizeof(void **);
8335 root_task_group.rt_rq = (struct rt_rq **)ptr;
8336 ptr += nr_cpu_ids * sizeof(void **);
8337 #endif /* CONFIG_USER_SCHED */
8338 #endif /* CONFIG_RT_GROUP_SCHED */
8342 init_defrootdomain();
8345 init_rt_bandwidth(&def_rt_bandwidth,
8346 global_rt_period(), global_rt_runtime());
8348 #ifdef CONFIG_RT_GROUP_SCHED
8349 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8350 global_rt_period(), global_rt_runtime());
8351 #ifdef CONFIG_USER_SCHED
8352 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8353 global_rt_period(), RUNTIME_INF);
8354 #endif /* CONFIG_USER_SCHED */
8355 #endif /* CONFIG_RT_GROUP_SCHED */
8357 #ifdef CONFIG_GROUP_SCHED
8358 list_add(&init_task_group.list, &task_groups);
8359 INIT_LIST_HEAD(&init_task_group.children);
8361 #ifdef CONFIG_USER_SCHED
8362 INIT_LIST_HEAD(&root_task_group.children);
8363 init_task_group.parent = &root_task_group;
8364 list_add(&init_task_group.siblings, &root_task_group.children);
8365 #endif /* CONFIG_USER_SCHED */
8366 #endif /* CONFIG_GROUP_SCHED */
8368 for_each_possible_cpu(i) {
8372 spin_lock_init(&rq->lock);
8374 init_cfs_rq(&rq->cfs, rq);
8375 init_rt_rq(&rq->rt, rq);
8376 #ifdef CONFIG_FAIR_GROUP_SCHED
8377 init_task_group.shares = init_task_group_load;
8378 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8379 #ifdef CONFIG_CGROUP_SCHED
8381 * How much cpu bandwidth does init_task_group get?
8383 * In case of task-groups formed thr' the cgroup filesystem, it
8384 * gets 100% of the cpu resources in the system. This overall
8385 * system cpu resource is divided among the tasks of
8386 * init_task_group and its child task-groups in a fair manner,
8387 * based on each entity's (task or task-group's) weight
8388 * (se->load.weight).
8390 * In other words, if init_task_group has 10 tasks of weight
8391 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8392 * then A0's share of the cpu resource is:
8394 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8396 * We achieve this by letting init_task_group's tasks sit
8397 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8399 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8400 #elif defined CONFIG_USER_SCHED
8401 root_task_group.shares = NICE_0_LOAD;
8402 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8404 * In case of task-groups formed thr' the user id of tasks,
8405 * init_task_group represents tasks belonging to root user.
8406 * Hence it forms a sibling of all subsequent groups formed.
8407 * In this case, init_task_group gets only a fraction of overall
8408 * system cpu resource, based on the weight assigned to root
8409 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8410 * by letting tasks of init_task_group sit in a separate cfs_rq
8411 * (init_cfs_rq) and having one entity represent this group of
8412 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8414 init_tg_cfs_entry(&init_task_group,
8415 &per_cpu(init_cfs_rq, i),
8416 &per_cpu(init_sched_entity, i), i, 1,
8417 root_task_group.se[i]);
8420 #endif /* CONFIG_FAIR_GROUP_SCHED */
8422 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8423 #ifdef CONFIG_RT_GROUP_SCHED
8424 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8425 #ifdef CONFIG_CGROUP_SCHED
8426 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8427 #elif defined CONFIG_USER_SCHED
8428 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8429 init_tg_rt_entry(&init_task_group,
8430 &per_cpu(init_rt_rq, i),
8431 &per_cpu(init_sched_rt_entity, i), i, 1,
8432 root_task_group.rt_se[i]);
8436 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8437 rq->cpu_load[j] = 0;
8441 rq->active_balance = 0;
8442 rq->next_balance = jiffies;
8446 rq->migration_thread = NULL;
8447 INIT_LIST_HEAD(&rq->migration_queue);
8448 rq_attach_root(rq, &def_root_domain);
8451 atomic_set(&rq->nr_iowait, 0);
8454 set_load_weight(&init_task);
8456 #ifdef CONFIG_PREEMPT_NOTIFIERS
8457 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8461 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8464 #ifdef CONFIG_RT_MUTEXES
8465 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8469 * The boot idle thread does lazy MMU switching as well:
8471 atomic_inc(&init_mm.mm_count);
8472 enter_lazy_tlb(&init_mm, current);
8475 * Make us the idle thread. Technically, schedule() should not be
8476 * called from this thread, however somewhere below it might be,
8477 * but because we are the idle thread, we just pick up running again
8478 * when this runqueue becomes "idle".
8480 init_idle(current, smp_processor_id());
8482 * During early bootup we pretend to be a normal task:
8484 current->sched_class = &fair_sched_class;
8486 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8487 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8490 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8492 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8495 scheduler_running = 1;
8498 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8499 void __might_sleep(char *file, int line)
8502 static unsigned long prev_jiffy; /* ratelimiting */
8504 if ((!in_atomic() && !irqs_disabled()) ||
8505 system_state != SYSTEM_RUNNING || oops_in_progress)
8507 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8509 prev_jiffy = jiffies;
8512 "BUG: sleeping function called from invalid context at %s:%d\n",
8515 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8516 in_atomic(), irqs_disabled(),
8517 current->pid, current->comm);
8519 debug_show_held_locks(current);
8520 if (irqs_disabled())
8521 print_irqtrace_events(current);
8525 EXPORT_SYMBOL(__might_sleep);
8528 #ifdef CONFIG_MAGIC_SYSRQ
8529 static void normalize_task(struct rq *rq, struct task_struct *p)
8533 update_rq_clock(rq);
8534 on_rq = p->se.on_rq;
8536 deactivate_task(rq, p, 0);
8537 __setscheduler(rq, p, SCHED_NORMAL, 0);
8539 activate_task(rq, p, 0);
8540 resched_task(rq->curr);
8544 void normalize_rt_tasks(void)
8546 struct task_struct *g, *p;
8547 unsigned long flags;
8550 read_lock_irqsave(&tasklist_lock, flags);
8551 do_each_thread(g, p) {
8553 * Only normalize user tasks:
8558 p->se.exec_start = 0;
8559 #ifdef CONFIG_SCHEDSTATS
8560 p->se.wait_start = 0;
8561 p->se.sleep_start = 0;
8562 p->se.block_start = 0;
8567 * Renice negative nice level userspace
8570 if (TASK_NICE(p) < 0 && p->mm)
8571 set_user_nice(p, 0);
8575 spin_lock(&p->pi_lock);
8576 rq = __task_rq_lock(p);
8578 normalize_task(rq, p);
8580 __task_rq_unlock(rq);
8581 spin_unlock(&p->pi_lock);
8582 } while_each_thread(g, p);
8584 read_unlock_irqrestore(&tasklist_lock, flags);
8587 #endif /* CONFIG_MAGIC_SYSRQ */
8591 * These functions are only useful for the IA64 MCA handling.
8593 * They can only be called when the whole system has been
8594 * stopped - every CPU needs to be quiescent, and no scheduling
8595 * activity can take place. Using them for anything else would
8596 * be a serious bug, and as a result, they aren't even visible
8597 * under any other configuration.
8601 * curr_task - return the current task for a given cpu.
8602 * @cpu: the processor in question.
8604 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8606 struct task_struct *curr_task(int cpu)
8608 return cpu_curr(cpu);
8612 * set_curr_task - set the current task for a given cpu.
8613 * @cpu: the processor in question.
8614 * @p: the task pointer to set.
8616 * Description: This function must only be used when non-maskable interrupts
8617 * are serviced on a separate stack. It allows the architecture to switch the
8618 * notion of the current task on a cpu in a non-blocking manner. This function
8619 * must be called with all CPU's synchronized, and interrupts disabled, the
8620 * and caller must save the original value of the current task (see
8621 * curr_task() above) and restore that value before reenabling interrupts and
8622 * re-starting the system.
8624 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8626 void set_curr_task(int cpu, struct task_struct *p)
8633 #ifdef CONFIG_FAIR_GROUP_SCHED
8634 static void free_fair_sched_group(struct task_group *tg)
8638 for_each_possible_cpu(i) {
8640 kfree(tg->cfs_rq[i]);
8650 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8652 struct cfs_rq *cfs_rq;
8653 struct sched_entity *se;
8657 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8660 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8664 tg->shares = NICE_0_LOAD;
8666 for_each_possible_cpu(i) {
8669 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8670 GFP_KERNEL, cpu_to_node(i));
8674 se = kzalloc_node(sizeof(struct sched_entity),
8675 GFP_KERNEL, cpu_to_node(i));
8679 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8688 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8690 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8691 &cpu_rq(cpu)->leaf_cfs_rq_list);
8694 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8696 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8698 #else /* !CONFG_FAIR_GROUP_SCHED */
8699 static inline void free_fair_sched_group(struct task_group *tg)
8704 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8709 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8713 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8716 #endif /* CONFIG_FAIR_GROUP_SCHED */
8718 #ifdef CONFIG_RT_GROUP_SCHED
8719 static void free_rt_sched_group(struct task_group *tg)
8723 destroy_rt_bandwidth(&tg->rt_bandwidth);
8725 for_each_possible_cpu(i) {
8727 kfree(tg->rt_rq[i]);
8729 kfree(tg->rt_se[i]);
8737 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8739 struct rt_rq *rt_rq;
8740 struct sched_rt_entity *rt_se;
8744 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8747 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8751 init_rt_bandwidth(&tg->rt_bandwidth,
8752 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8754 for_each_possible_cpu(i) {
8757 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8758 GFP_KERNEL, cpu_to_node(i));
8762 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8763 GFP_KERNEL, cpu_to_node(i));
8767 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8776 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8778 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8779 &cpu_rq(cpu)->leaf_rt_rq_list);
8782 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8784 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8786 #else /* !CONFIG_RT_GROUP_SCHED */
8787 static inline void free_rt_sched_group(struct task_group *tg)
8792 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8797 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8801 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8804 #endif /* CONFIG_RT_GROUP_SCHED */
8806 #ifdef CONFIG_GROUP_SCHED
8807 static void free_sched_group(struct task_group *tg)
8809 free_fair_sched_group(tg);
8810 free_rt_sched_group(tg);
8814 /* allocate runqueue etc for a new task group */
8815 struct task_group *sched_create_group(struct task_group *parent)
8817 struct task_group *tg;
8818 unsigned long flags;
8821 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8823 return ERR_PTR(-ENOMEM);
8825 if (!alloc_fair_sched_group(tg, parent))
8828 if (!alloc_rt_sched_group(tg, parent))
8831 spin_lock_irqsave(&task_group_lock, flags);
8832 for_each_possible_cpu(i) {
8833 register_fair_sched_group(tg, i);
8834 register_rt_sched_group(tg, i);
8836 list_add_rcu(&tg->list, &task_groups);
8838 WARN_ON(!parent); /* root should already exist */
8840 tg->parent = parent;
8841 INIT_LIST_HEAD(&tg->children);
8842 list_add_rcu(&tg->siblings, &parent->children);
8843 spin_unlock_irqrestore(&task_group_lock, flags);
8848 free_sched_group(tg);
8849 return ERR_PTR(-ENOMEM);
8852 /* rcu callback to free various structures associated with a task group */
8853 static void free_sched_group_rcu(struct rcu_head *rhp)
8855 /* now it should be safe to free those cfs_rqs */
8856 free_sched_group(container_of(rhp, struct task_group, rcu));
8859 /* Destroy runqueue etc associated with a task group */
8860 void sched_destroy_group(struct task_group *tg)
8862 unsigned long flags;
8865 spin_lock_irqsave(&task_group_lock, flags);
8866 for_each_possible_cpu(i) {
8867 unregister_fair_sched_group(tg, i);
8868 unregister_rt_sched_group(tg, i);
8870 list_del_rcu(&tg->list);
8871 list_del_rcu(&tg->siblings);
8872 spin_unlock_irqrestore(&task_group_lock, flags);
8874 /* wait for possible concurrent references to cfs_rqs complete */
8875 call_rcu(&tg->rcu, free_sched_group_rcu);
8878 /* change task's runqueue when it moves between groups.
8879 * The caller of this function should have put the task in its new group
8880 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8881 * reflect its new group.
8883 void sched_move_task(struct task_struct *tsk)
8886 unsigned long flags;
8889 rq = task_rq_lock(tsk, &flags);
8891 update_rq_clock(rq);
8893 running = task_current(rq, tsk);
8894 on_rq = tsk->se.on_rq;
8897 dequeue_task(rq, tsk, 0);
8898 if (unlikely(running))
8899 tsk->sched_class->put_prev_task(rq, tsk);
8901 set_task_rq(tsk, task_cpu(tsk));
8903 #ifdef CONFIG_FAIR_GROUP_SCHED
8904 if (tsk->sched_class->moved_group)
8905 tsk->sched_class->moved_group(tsk);
8908 if (unlikely(running))
8909 tsk->sched_class->set_curr_task(rq);
8911 enqueue_task(rq, tsk, 0);
8913 task_rq_unlock(rq, &flags);
8915 #endif /* CONFIG_GROUP_SCHED */
8917 #ifdef CONFIG_FAIR_GROUP_SCHED
8918 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8920 struct cfs_rq *cfs_rq = se->cfs_rq;
8925 dequeue_entity(cfs_rq, se, 0);
8927 se->load.weight = shares;
8928 se->load.inv_weight = 0;
8931 enqueue_entity(cfs_rq, se, 0);
8934 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8936 struct cfs_rq *cfs_rq = se->cfs_rq;
8937 struct rq *rq = cfs_rq->rq;
8938 unsigned long flags;
8940 spin_lock_irqsave(&rq->lock, flags);
8941 __set_se_shares(se, shares);
8942 spin_unlock_irqrestore(&rq->lock, flags);
8945 static DEFINE_MUTEX(shares_mutex);
8947 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8950 unsigned long flags;
8953 * We can't change the weight of the root cgroup.
8958 if (shares < MIN_SHARES)
8959 shares = MIN_SHARES;
8960 else if (shares > MAX_SHARES)
8961 shares = MAX_SHARES;
8963 mutex_lock(&shares_mutex);
8964 if (tg->shares == shares)
8967 spin_lock_irqsave(&task_group_lock, flags);
8968 for_each_possible_cpu(i)
8969 unregister_fair_sched_group(tg, i);
8970 list_del_rcu(&tg->siblings);
8971 spin_unlock_irqrestore(&task_group_lock, flags);
8973 /* wait for any ongoing reference to this group to finish */
8974 synchronize_sched();
8977 * Now we are free to modify the group's share on each cpu
8978 * w/o tripping rebalance_share or load_balance_fair.
8980 tg->shares = shares;
8981 for_each_possible_cpu(i) {
8985 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8986 set_se_shares(tg->se[i], shares);
8990 * Enable load balance activity on this group, by inserting it back on
8991 * each cpu's rq->leaf_cfs_rq_list.
8993 spin_lock_irqsave(&task_group_lock, flags);
8994 for_each_possible_cpu(i)
8995 register_fair_sched_group(tg, i);
8996 list_add_rcu(&tg->siblings, &tg->parent->children);
8997 spin_unlock_irqrestore(&task_group_lock, flags);
8999 mutex_unlock(&shares_mutex);
9003 unsigned long sched_group_shares(struct task_group *tg)
9009 #ifdef CONFIG_RT_GROUP_SCHED
9011 * Ensure that the real time constraints are schedulable.
9013 static DEFINE_MUTEX(rt_constraints_mutex);
9015 static unsigned long to_ratio(u64 period, u64 runtime)
9017 if (runtime == RUNTIME_INF)
9020 return div64_u64(runtime << 20, period);
9023 /* Must be called with tasklist_lock held */
9024 static inline int tg_has_rt_tasks(struct task_group *tg)
9026 struct task_struct *g, *p;
9028 do_each_thread(g, p) {
9029 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9031 } while_each_thread(g, p);
9036 struct rt_schedulable_data {
9037 struct task_group *tg;
9042 static int tg_schedulable(struct task_group *tg, void *data)
9044 struct rt_schedulable_data *d = data;
9045 struct task_group *child;
9046 unsigned long total, sum = 0;
9047 u64 period, runtime;
9049 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9050 runtime = tg->rt_bandwidth.rt_runtime;
9053 period = d->rt_period;
9054 runtime = d->rt_runtime;
9057 #ifdef CONFIG_USER_SCHED
9058 if (tg == &root_task_group) {
9059 period = global_rt_period();
9060 runtime = global_rt_runtime();
9065 * Cannot have more runtime than the period.
9067 if (runtime > period && runtime != RUNTIME_INF)
9071 * Ensure we don't starve existing RT tasks.
9073 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9076 total = to_ratio(period, runtime);
9079 * Nobody can have more than the global setting allows.
9081 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9085 * The sum of our children's runtime should not exceed our own.
9087 list_for_each_entry_rcu(child, &tg->children, siblings) {
9088 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9089 runtime = child->rt_bandwidth.rt_runtime;
9091 if (child == d->tg) {
9092 period = d->rt_period;
9093 runtime = d->rt_runtime;
9096 sum += to_ratio(period, runtime);
9105 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9107 struct rt_schedulable_data data = {
9109 .rt_period = period,
9110 .rt_runtime = runtime,
9113 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9116 static int tg_set_bandwidth(struct task_group *tg,
9117 u64 rt_period, u64 rt_runtime)
9121 mutex_lock(&rt_constraints_mutex);
9122 read_lock(&tasklist_lock);
9123 err = __rt_schedulable(tg, rt_period, rt_runtime);
9127 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9128 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9129 tg->rt_bandwidth.rt_runtime = rt_runtime;
9131 for_each_possible_cpu(i) {
9132 struct rt_rq *rt_rq = tg->rt_rq[i];
9134 spin_lock(&rt_rq->rt_runtime_lock);
9135 rt_rq->rt_runtime = rt_runtime;
9136 spin_unlock(&rt_rq->rt_runtime_lock);
9138 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9140 read_unlock(&tasklist_lock);
9141 mutex_unlock(&rt_constraints_mutex);
9146 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9148 u64 rt_runtime, rt_period;
9150 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9151 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9152 if (rt_runtime_us < 0)
9153 rt_runtime = RUNTIME_INF;
9155 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9158 long sched_group_rt_runtime(struct task_group *tg)
9162 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9165 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9166 do_div(rt_runtime_us, NSEC_PER_USEC);
9167 return rt_runtime_us;
9170 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9172 u64 rt_runtime, rt_period;
9174 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9175 rt_runtime = tg->rt_bandwidth.rt_runtime;
9180 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9183 long sched_group_rt_period(struct task_group *tg)
9187 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9188 do_div(rt_period_us, NSEC_PER_USEC);
9189 return rt_period_us;
9192 static int sched_rt_global_constraints(void)
9194 u64 runtime, period;
9197 if (sysctl_sched_rt_period <= 0)
9200 runtime = global_rt_runtime();
9201 period = global_rt_period();
9204 * Sanity check on the sysctl variables.
9206 if (runtime > period && runtime != RUNTIME_INF)
9209 mutex_lock(&rt_constraints_mutex);
9210 read_lock(&tasklist_lock);
9211 ret = __rt_schedulable(NULL, 0, 0);
9212 read_unlock(&tasklist_lock);
9213 mutex_unlock(&rt_constraints_mutex);
9217 #else /* !CONFIG_RT_GROUP_SCHED */
9218 static int sched_rt_global_constraints(void)
9220 unsigned long flags;
9223 if (sysctl_sched_rt_period <= 0)
9226 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9227 for_each_possible_cpu(i) {
9228 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9230 spin_lock(&rt_rq->rt_runtime_lock);
9231 rt_rq->rt_runtime = global_rt_runtime();
9232 spin_unlock(&rt_rq->rt_runtime_lock);
9234 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9238 #endif /* CONFIG_RT_GROUP_SCHED */
9240 int sched_rt_handler(struct ctl_table *table, int write,
9241 struct file *filp, void __user *buffer, size_t *lenp,
9245 int old_period, old_runtime;
9246 static DEFINE_MUTEX(mutex);
9249 old_period = sysctl_sched_rt_period;
9250 old_runtime = sysctl_sched_rt_runtime;
9252 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9254 if (!ret && write) {
9255 ret = sched_rt_global_constraints();
9257 sysctl_sched_rt_period = old_period;
9258 sysctl_sched_rt_runtime = old_runtime;
9260 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9261 def_rt_bandwidth.rt_period =
9262 ns_to_ktime(global_rt_period());
9265 mutex_unlock(&mutex);
9270 #ifdef CONFIG_CGROUP_SCHED
9272 /* return corresponding task_group object of a cgroup */
9273 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9275 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9276 struct task_group, css);
9279 static struct cgroup_subsys_state *
9280 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9282 struct task_group *tg, *parent;
9284 if (!cgrp->parent) {
9285 /* This is early initialization for the top cgroup */
9286 return &init_task_group.css;
9289 parent = cgroup_tg(cgrp->parent);
9290 tg = sched_create_group(parent);
9292 return ERR_PTR(-ENOMEM);
9298 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9300 struct task_group *tg = cgroup_tg(cgrp);
9302 sched_destroy_group(tg);
9306 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9307 struct task_struct *tsk)
9309 #ifdef CONFIG_RT_GROUP_SCHED
9310 /* Don't accept realtime tasks when there is no way for them to run */
9311 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9314 /* We don't support RT-tasks being in separate groups */
9315 if (tsk->sched_class != &fair_sched_class)
9323 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9324 struct cgroup *old_cont, struct task_struct *tsk)
9326 sched_move_task(tsk);
9329 #ifdef CONFIG_FAIR_GROUP_SCHED
9330 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9333 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9336 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9338 struct task_group *tg = cgroup_tg(cgrp);
9340 return (u64) tg->shares;
9342 #endif /* CONFIG_FAIR_GROUP_SCHED */
9344 #ifdef CONFIG_RT_GROUP_SCHED
9345 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9348 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9351 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9353 return sched_group_rt_runtime(cgroup_tg(cgrp));
9356 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9359 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9362 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9364 return sched_group_rt_period(cgroup_tg(cgrp));
9366 #endif /* CONFIG_RT_GROUP_SCHED */
9368 static struct cftype cpu_files[] = {
9369 #ifdef CONFIG_FAIR_GROUP_SCHED
9372 .read_u64 = cpu_shares_read_u64,
9373 .write_u64 = cpu_shares_write_u64,
9376 #ifdef CONFIG_RT_GROUP_SCHED
9378 .name = "rt_runtime_us",
9379 .read_s64 = cpu_rt_runtime_read,
9380 .write_s64 = cpu_rt_runtime_write,
9383 .name = "rt_period_us",
9384 .read_u64 = cpu_rt_period_read_uint,
9385 .write_u64 = cpu_rt_period_write_uint,
9390 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9392 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9395 struct cgroup_subsys cpu_cgroup_subsys = {
9397 .create = cpu_cgroup_create,
9398 .destroy = cpu_cgroup_destroy,
9399 .can_attach = cpu_cgroup_can_attach,
9400 .attach = cpu_cgroup_attach,
9401 .populate = cpu_cgroup_populate,
9402 .subsys_id = cpu_cgroup_subsys_id,
9406 #endif /* CONFIG_CGROUP_SCHED */
9408 #ifdef CONFIG_CGROUP_CPUACCT
9411 * CPU accounting code for task groups.
9413 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9414 * (balbir@in.ibm.com).
9417 /* track cpu usage of a group of tasks and its child groups */
9419 struct cgroup_subsys_state css;
9420 /* cpuusage holds pointer to a u64-type object on every cpu */
9422 struct cpuacct *parent;
9425 struct cgroup_subsys cpuacct_subsys;
9427 /* return cpu accounting group corresponding to this container */
9428 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9430 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9431 struct cpuacct, css);
9434 /* return cpu accounting group to which this task belongs */
9435 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9437 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9438 struct cpuacct, css);
9441 /* create a new cpu accounting group */
9442 static struct cgroup_subsys_state *cpuacct_create(
9443 struct cgroup_subsys *ss, struct cgroup *cgrp)
9445 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9448 return ERR_PTR(-ENOMEM);
9450 ca->cpuusage = alloc_percpu(u64);
9451 if (!ca->cpuusage) {
9453 return ERR_PTR(-ENOMEM);
9457 ca->parent = cgroup_ca(cgrp->parent);
9462 /* destroy an existing cpu accounting group */
9464 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9466 struct cpuacct *ca = cgroup_ca(cgrp);
9468 free_percpu(ca->cpuusage);
9472 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9474 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9477 #ifndef CONFIG_64BIT
9479 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9481 spin_lock_irq(&cpu_rq(cpu)->lock);
9483 spin_unlock_irq(&cpu_rq(cpu)->lock);
9491 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9493 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9495 #ifndef CONFIG_64BIT
9497 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9499 spin_lock_irq(&cpu_rq(cpu)->lock);
9501 spin_unlock_irq(&cpu_rq(cpu)->lock);
9507 /* return total cpu usage (in nanoseconds) of a group */
9508 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9510 struct cpuacct *ca = cgroup_ca(cgrp);
9511 u64 totalcpuusage = 0;
9514 for_each_present_cpu(i)
9515 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9517 return totalcpuusage;
9520 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9523 struct cpuacct *ca = cgroup_ca(cgrp);
9532 for_each_present_cpu(i)
9533 cpuacct_cpuusage_write(ca, i, 0);
9539 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9542 struct cpuacct *ca = cgroup_ca(cgroup);
9546 for_each_present_cpu(i) {
9547 percpu = cpuacct_cpuusage_read(ca, i);
9548 seq_printf(m, "%llu ", (unsigned long long) percpu);
9550 seq_printf(m, "\n");
9554 static struct cftype files[] = {
9557 .read_u64 = cpuusage_read,
9558 .write_u64 = cpuusage_write,
9561 .name = "usage_percpu",
9562 .read_seq_string = cpuacct_percpu_seq_read,
9567 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9569 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9573 * charge this task's execution time to its accounting group.
9575 * called with rq->lock held.
9577 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9582 if (!cpuacct_subsys.active)
9585 cpu = task_cpu(tsk);
9588 for (; ca; ca = ca->parent) {
9589 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9590 *cpuusage += cputime;
9594 struct cgroup_subsys cpuacct_subsys = {
9596 .create = cpuacct_create,
9597 .destroy = cpuacct_destroy,
9598 .populate = cpuacct_populate,
9599 .subsys_id = cpuacct_subsys_id,
9601 #endif /* CONFIG_CGROUP_CPUACCT */