4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
134 return reciprocal_divide(load, sg->reciprocal_cpu_power);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
143 sg->__cpu_power += val;
144 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
148 static inline int rt_policy(int policy)
150 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
155 static inline int task_has_rt_policy(struct task_struct *p)
157 return rt_policy(p->policy);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array {
164 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
165 struct list_head queue[MAX_RT_PRIO];
168 struct rt_bandwidth {
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock;
173 struct hrtimer rt_period_timer;
176 static struct rt_bandwidth def_rt_bandwidth;
178 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
180 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
182 struct rt_bandwidth *rt_b =
183 container_of(timer, struct rt_bandwidth, rt_period_timer);
189 now = hrtimer_cb_get_time(timer);
190 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
195 idle = do_sched_rt_period_timer(rt_b, overrun);
198 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
202 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
204 rt_b->rt_period = ns_to_ktime(period);
205 rt_b->rt_runtime = runtime;
207 spin_lock_init(&rt_b->rt_runtime_lock);
209 hrtimer_init(&rt_b->rt_period_timer,
210 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
211 rt_b->rt_period_timer.function = sched_rt_period_timer;
214 static inline int rt_bandwidth_enabled(void)
216 return sysctl_sched_rt_runtime >= 0;
219 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
223 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
226 if (hrtimer_active(&rt_b->rt_period_timer))
229 spin_lock(&rt_b->rt_runtime_lock);
231 if (hrtimer_active(&rt_b->rt_period_timer))
234 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
235 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
236 hrtimer_start_expires(&rt_b->rt_period_timer,
239 spin_unlock(&rt_b->rt_runtime_lock);
242 #ifdef CONFIG_RT_GROUP_SCHED
243 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
245 hrtimer_cancel(&rt_b->rt_period_timer);
250 * sched_domains_mutex serializes calls to arch_init_sched_domains,
251 * detach_destroy_domains and partition_sched_domains.
253 static DEFINE_MUTEX(sched_domains_mutex);
255 #ifdef CONFIG_GROUP_SCHED
257 #include <linux/cgroup.h>
261 static LIST_HEAD(task_groups);
263 /* task group related information */
265 #ifdef CONFIG_CGROUP_SCHED
266 struct cgroup_subsys_state css;
269 #ifdef CONFIG_USER_SCHED
273 #ifdef CONFIG_FAIR_GROUP_SCHED
274 /* schedulable entities of this group on each cpu */
275 struct sched_entity **se;
276 /* runqueue "owned" by this group on each cpu */
277 struct cfs_rq **cfs_rq;
278 unsigned long shares;
281 #ifdef CONFIG_RT_GROUP_SCHED
282 struct sched_rt_entity **rt_se;
283 struct rt_rq **rt_rq;
285 struct rt_bandwidth rt_bandwidth;
289 struct list_head list;
291 struct task_group *parent;
292 struct list_head siblings;
293 struct list_head children;
296 #ifdef CONFIG_USER_SCHED
298 /* Helper function to pass uid information to create_sched_user() */
299 void set_tg_uid(struct user_struct *user)
301 user->tg->uid = user->uid;
306 * Every UID task group (including init_task_group aka UID-0) will
307 * be a child to this group.
309 struct task_group root_task_group;
311 #ifdef CONFIG_FAIR_GROUP_SCHED
312 /* Default task group's sched entity on each cpu */
313 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
314 /* Default task group's cfs_rq on each cpu */
315 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
316 #endif /* CONFIG_FAIR_GROUP_SCHED */
318 #ifdef CONFIG_RT_GROUP_SCHED
319 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
320 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
321 #endif /* CONFIG_RT_GROUP_SCHED */
322 #else /* !CONFIG_USER_SCHED */
323 #define root_task_group init_task_group
324 #endif /* CONFIG_USER_SCHED */
326 /* task_group_lock serializes add/remove of task groups and also changes to
327 * a task group's cpu shares.
329 static DEFINE_SPINLOCK(task_group_lock);
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 #ifdef CONFIG_USER_SCHED
333 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
334 #else /* !CONFIG_USER_SCHED */
335 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
336 #endif /* CONFIG_USER_SCHED */
339 * A weight of 0 or 1 can cause arithmetics problems.
340 * A weight of a cfs_rq is the sum of weights of which entities
341 * are queued on this cfs_rq, so a weight of a entity should not be
342 * too large, so as the shares value of a task group.
343 * (The default weight is 1024 - so there's no practical
344 * limitation from this.)
347 #define MAX_SHARES (1UL << 18)
349 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
352 /* Default task group.
353 * Every task in system belong to this group at bootup.
355 struct task_group init_task_group;
357 /* return group to which a task belongs */
358 static inline struct task_group *task_group(struct task_struct *p)
360 struct task_group *tg;
362 #ifdef CONFIG_USER_SCHED
364 tg = __task_cred(p)->user->tg;
366 #elif defined(CONFIG_CGROUP_SCHED)
367 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
368 struct task_group, css);
370 tg = &init_task_group;
375 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
376 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
378 #ifdef CONFIG_FAIR_GROUP_SCHED
379 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
380 p->se.parent = task_group(p)->se[cpu];
383 #ifdef CONFIG_RT_GROUP_SCHED
384 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
385 p->rt.parent = task_group(p)->rt_se[cpu];
391 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
392 static inline struct task_group *task_group(struct task_struct *p)
397 #endif /* CONFIG_GROUP_SCHED */
399 /* CFS-related fields in a runqueue */
401 struct load_weight load;
402 unsigned long nr_running;
407 struct rb_root tasks_timeline;
408 struct rb_node *rb_leftmost;
410 struct list_head tasks;
411 struct list_head *balance_iterator;
414 * 'curr' points to currently running entity on this cfs_rq.
415 * It is set to NULL otherwise (i.e when none are currently running).
417 struct sched_entity *curr, *next, *last;
419 unsigned int nr_spread_over;
421 #ifdef CONFIG_FAIR_GROUP_SCHED
422 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
425 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
426 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
427 * (like users, containers etc.)
429 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
430 * list is used during load balance.
432 struct list_head leaf_cfs_rq_list;
433 struct task_group *tg; /* group that "owns" this runqueue */
437 * the part of load.weight contributed by tasks
439 unsigned long task_weight;
442 * h_load = weight * f(tg)
444 * Where f(tg) is the recursive weight fraction assigned to
447 unsigned long h_load;
450 * this cpu's part of tg->shares
452 unsigned long shares;
455 * load.weight at the time we set shares
457 unsigned long rq_weight;
462 /* Real-Time classes' related field in a runqueue: */
464 struct rt_prio_array active;
465 unsigned long rt_nr_running;
466 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
468 int curr; /* highest queued rt task prio */
469 int next; /* next highest */
473 unsigned long rt_nr_migratory;
475 struct plist_head pushable_tasks;
480 /* Nests inside the rq lock: */
481 spinlock_t rt_runtime_lock;
483 #ifdef CONFIG_RT_GROUP_SCHED
484 unsigned long rt_nr_boosted;
487 struct list_head leaf_rt_rq_list;
488 struct task_group *tg;
489 struct sched_rt_entity *rt_se;
496 * We add the notion of a root-domain which will be used to define per-domain
497 * variables. Each exclusive cpuset essentially defines an island domain by
498 * fully partitioning the member cpus from any other cpuset. Whenever a new
499 * exclusive cpuset is created, we also create and attach a new root-domain
506 cpumask_var_t online;
509 * The "RT overload" flag: it gets set if a CPU has more than
510 * one runnable RT task.
512 cpumask_var_t rto_mask;
515 struct cpupri cpupri;
517 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
519 * Preferred wake up cpu nominated by sched_mc balance that will be
520 * used when most cpus are idle in the system indicating overall very
521 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
523 unsigned int sched_mc_preferred_wakeup_cpu;
528 * By default the system creates a single root-domain with all cpus as
529 * members (mimicking the global state we have today).
531 static struct root_domain def_root_domain;
536 * This is the main, per-CPU runqueue data structure.
538 * Locking rule: those places that want to lock multiple runqueues
539 * (such as the load balancing or the thread migration code), lock
540 * acquire operations must be ordered by ascending &runqueue.
547 * nr_running and cpu_load should be in the same cacheline because
548 * remote CPUs use both these fields when doing load calculation.
550 unsigned long nr_running;
551 #define CPU_LOAD_IDX_MAX 5
552 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
553 unsigned char idle_at_tick;
555 unsigned long last_tick_seen;
556 unsigned char in_nohz_recently;
558 /* capture load from *all* tasks on this cpu: */
559 struct load_weight load;
560 unsigned long nr_load_updates;
566 #ifdef CONFIG_FAIR_GROUP_SCHED
567 /* list of leaf cfs_rq on this cpu: */
568 struct list_head leaf_cfs_rq_list;
570 #ifdef CONFIG_RT_GROUP_SCHED
571 struct list_head leaf_rt_rq_list;
575 * This is part of a global counter where only the total sum
576 * over all CPUs matters. A task can increase this counter on
577 * one CPU and if it got migrated afterwards it may decrease
578 * it on another CPU. Always updated under the runqueue lock:
580 unsigned long nr_uninterruptible;
582 struct task_struct *curr, *idle;
583 unsigned long next_balance;
584 struct mm_struct *prev_mm;
591 struct root_domain *rd;
592 struct sched_domain *sd;
594 /* For active balancing */
597 /* cpu of this runqueue: */
601 unsigned long avg_load_per_task;
603 struct task_struct *migration_thread;
604 struct list_head migration_queue;
607 #ifdef CONFIG_SCHED_HRTICK
609 int hrtick_csd_pending;
610 struct call_single_data hrtick_csd;
612 struct hrtimer hrtick_timer;
615 #ifdef CONFIG_SCHEDSTATS
617 struct sched_info rq_sched_info;
618 unsigned long long rq_cpu_time;
619 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
621 /* sys_sched_yield() stats */
622 unsigned int yld_exp_empty;
623 unsigned int yld_act_empty;
624 unsigned int yld_both_empty;
625 unsigned int yld_count;
627 /* schedule() stats */
628 unsigned int sched_switch;
629 unsigned int sched_count;
630 unsigned int sched_goidle;
632 /* try_to_wake_up() stats */
633 unsigned int ttwu_count;
634 unsigned int ttwu_local;
637 unsigned int bkl_count;
641 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
643 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
645 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
648 static inline int cpu_of(struct rq *rq)
658 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
659 * See detach_destroy_domains: synchronize_sched for details.
661 * The domain tree of any CPU may only be accessed from within
662 * preempt-disabled sections.
664 #define for_each_domain(cpu, __sd) \
665 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
667 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
668 #define this_rq() (&__get_cpu_var(runqueues))
669 #define task_rq(p) cpu_rq(task_cpu(p))
670 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
672 static inline void update_rq_clock(struct rq *rq)
674 rq->clock = sched_clock_cpu(cpu_of(rq));
678 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
680 #ifdef CONFIG_SCHED_DEBUG
681 # define const_debug __read_mostly
683 # define const_debug static const
689 * Returns true if the current cpu runqueue is locked.
690 * This interface allows printk to be called with the runqueue lock
691 * held and know whether or not it is OK to wake up the klogd.
693 int runqueue_is_locked(void)
696 struct rq *rq = cpu_rq(cpu);
699 ret = spin_is_locked(&rq->lock);
705 * Debugging: various feature bits
708 #define SCHED_FEAT(name, enabled) \
709 __SCHED_FEAT_##name ,
712 #include "sched_features.h"
717 #define SCHED_FEAT(name, enabled) \
718 (1UL << __SCHED_FEAT_##name) * enabled |
720 const_debug unsigned int sysctl_sched_features =
721 #include "sched_features.h"
726 #ifdef CONFIG_SCHED_DEBUG
727 #define SCHED_FEAT(name, enabled) \
730 static __read_mostly char *sched_feat_names[] = {
731 #include "sched_features.h"
737 static int sched_feat_show(struct seq_file *m, void *v)
741 for (i = 0; sched_feat_names[i]; i++) {
742 if (!(sysctl_sched_features & (1UL << i)))
744 seq_printf(m, "%s ", sched_feat_names[i]);
752 sched_feat_write(struct file *filp, const char __user *ubuf,
753 size_t cnt, loff_t *ppos)
763 if (copy_from_user(&buf, ubuf, cnt))
768 if (strncmp(buf, "NO_", 3) == 0) {
773 for (i = 0; sched_feat_names[i]; i++) {
774 int len = strlen(sched_feat_names[i]);
776 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
778 sysctl_sched_features &= ~(1UL << i);
780 sysctl_sched_features |= (1UL << i);
785 if (!sched_feat_names[i])
793 static int sched_feat_open(struct inode *inode, struct file *filp)
795 return single_open(filp, sched_feat_show, NULL);
798 static struct file_operations sched_feat_fops = {
799 .open = sched_feat_open,
800 .write = sched_feat_write,
803 .release = single_release,
806 static __init int sched_init_debug(void)
808 debugfs_create_file("sched_features", 0644, NULL, NULL,
813 late_initcall(sched_init_debug);
817 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
820 * Number of tasks to iterate in a single balance run.
821 * Limited because this is done with IRQs disabled.
823 const_debug unsigned int sysctl_sched_nr_migrate = 32;
826 * ratelimit for updating the group shares.
829 unsigned int sysctl_sched_shares_ratelimit = 250000;
832 * Inject some fuzzyness into changing the per-cpu group shares
833 * this avoids remote rq-locks at the expense of fairness.
836 unsigned int sysctl_sched_shares_thresh = 4;
839 * period over which we measure -rt task cpu usage in us.
842 unsigned int sysctl_sched_rt_period = 1000000;
844 static __read_mostly int scheduler_running;
847 * part of the period that we allow rt tasks to run in us.
850 int sysctl_sched_rt_runtime = 950000;
852 static inline u64 global_rt_period(void)
854 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
857 static inline u64 global_rt_runtime(void)
859 if (sysctl_sched_rt_runtime < 0)
862 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
865 #ifndef prepare_arch_switch
866 # define prepare_arch_switch(next) do { } while (0)
868 #ifndef finish_arch_switch
869 # define finish_arch_switch(prev) do { } while (0)
872 static inline int task_current(struct rq *rq, struct task_struct *p)
874 return rq->curr == p;
877 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
878 static inline int task_running(struct rq *rq, struct task_struct *p)
880 return task_current(rq, p);
883 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
887 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
889 #ifdef CONFIG_DEBUG_SPINLOCK
890 /* this is a valid case when another task releases the spinlock */
891 rq->lock.owner = current;
894 * If we are tracking spinlock dependencies then we have to
895 * fix up the runqueue lock - which gets 'carried over' from
898 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
900 spin_unlock_irq(&rq->lock);
903 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
904 static inline int task_running(struct rq *rq, struct task_struct *p)
909 return task_current(rq, p);
913 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
917 * We can optimise this out completely for !SMP, because the
918 * SMP rebalancing from interrupt is the only thing that cares
923 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
924 spin_unlock_irq(&rq->lock);
926 spin_unlock(&rq->lock);
930 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
934 * After ->oncpu is cleared, the task can be moved to a different CPU.
935 * We must ensure this doesn't happen until the switch is completely
941 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
945 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
948 * __task_rq_lock - lock the runqueue a given task resides on.
949 * Must be called interrupts disabled.
951 static inline struct rq *__task_rq_lock(struct task_struct *p)
955 struct rq *rq = task_rq(p);
956 spin_lock(&rq->lock);
957 if (likely(rq == task_rq(p)))
959 spin_unlock(&rq->lock);
964 * task_rq_lock - lock the runqueue a given task resides on and disable
965 * interrupts. Note the ordering: we can safely lookup the task_rq without
966 * explicitly disabling preemption.
968 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
974 local_irq_save(*flags);
976 spin_lock(&rq->lock);
977 if (likely(rq == task_rq(p)))
979 spin_unlock_irqrestore(&rq->lock, *flags);
983 void task_rq_unlock_wait(struct task_struct *p)
985 struct rq *rq = task_rq(p);
987 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
988 spin_unlock_wait(&rq->lock);
991 static void __task_rq_unlock(struct rq *rq)
994 spin_unlock(&rq->lock);
997 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1000 spin_unlock_irqrestore(&rq->lock, *flags);
1004 * this_rq_lock - lock this runqueue and disable interrupts.
1006 static struct rq *this_rq_lock(void)
1007 __acquires(rq->lock)
1011 local_irq_disable();
1013 spin_lock(&rq->lock);
1018 #ifdef CONFIG_SCHED_HRTICK
1020 * Use HR-timers to deliver accurate preemption points.
1022 * Its all a bit involved since we cannot program an hrt while holding the
1023 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1026 * When we get rescheduled we reprogram the hrtick_timer outside of the
1032 * - enabled by features
1033 * - hrtimer is actually high res
1035 static inline int hrtick_enabled(struct rq *rq)
1037 if (!sched_feat(HRTICK))
1039 if (!cpu_active(cpu_of(rq)))
1041 return hrtimer_is_hres_active(&rq->hrtick_timer);
1044 static void hrtick_clear(struct rq *rq)
1046 if (hrtimer_active(&rq->hrtick_timer))
1047 hrtimer_cancel(&rq->hrtick_timer);
1051 * High-resolution timer tick.
1052 * Runs from hardirq context with interrupts disabled.
1054 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1056 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1058 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1060 spin_lock(&rq->lock);
1061 update_rq_clock(rq);
1062 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1063 spin_unlock(&rq->lock);
1065 return HRTIMER_NORESTART;
1070 * called from hardirq (IPI) context
1072 static void __hrtick_start(void *arg)
1074 struct rq *rq = arg;
1076 spin_lock(&rq->lock);
1077 hrtimer_restart(&rq->hrtick_timer);
1078 rq->hrtick_csd_pending = 0;
1079 spin_unlock(&rq->lock);
1083 * Called to set the hrtick timer state.
1085 * called with rq->lock held and irqs disabled
1087 static void hrtick_start(struct rq *rq, u64 delay)
1089 struct hrtimer *timer = &rq->hrtick_timer;
1090 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1092 hrtimer_set_expires(timer, time);
1094 if (rq == this_rq()) {
1095 hrtimer_restart(timer);
1096 } else if (!rq->hrtick_csd_pending) {
1097 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1098 rq->hrtick_csd_pending = 1;
1103 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1105 int cpu = (int)(long)hcpu;
1108 case CPU_UP_CANCELED:
1109 case CPU_UP_CANCELED_FROZEN:
1110 case CPU_DOWN_PREPARE:
1111 case CPU_DOWN_PREPARE_FROZEN:
1113 case CPU_DEAD_FROZEN:
1114 hrtick_clear(cpu_rq(cpu));
1121 static __init void init_hrtick(void)
1123 hotcpu_notifier(hotplug_hrtick, 0);
1127 * Called to set the hrtick timer state.
1129 * called with rq->lock held and irqs disabled
1131 static void hrtick_start(struct rq *rq, u64 delay)
1133 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1136 static inline void init_hrtick(void)
1139 #endif /* CONFIG_SMP */
1141 static void init_rq_hrtick(struct rq *rq)
1144 rq->hrtick_csd_pending = 0;
1146 rq->hrtick_csd.flags = 0;
1147 rq->hrtick_csd.func = __hrtick_start;
1148 rq->hrtick_csd.info = rq;
1151 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1152 rq->hrtick_timer.function = hrtick;
1154 #else /* CONFIG_SCHED_HRTICK */
1155 static inline void hrtick_clear(struct rq *rq)
1159 static inline void init_rq_hrtick(struct rq *rq)
1163 static inline void init_hrtick(void)
1166 #endif /* CONFIG_SCHED_HRTICK */
1169 * resched_task - mark a task 'to be rescheduled now'.
1171 * On UP this means the setting of the need_resched flag, on SMP it
1172 * might also involve a cross-CPU call to trigger the scheduler on
1177 #ifndef tsk_is_polling
1178 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1181 static void resched_task(struct task_struct *p)
1185 assert_spin_locked(&task_rq(p)->lock);
1187 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1190 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1193 if (cpu == smp_processor_id())
1196 /* NEED_RESCHED must be visible before we test polling */
1198 if (!tsk_is_polling(p))
1199 smp_send_reschedule(cpu);
1202 static void resched_cpu(int cpu)
1204 struct rq *rq = cpu_rq(cpu);
1205 unsigned long flags;
1207 if (!spin_trylock_irqsave(&rq->lock, flags))
1209 resched_task(cpu_curr(cpu));
1210 spin_unlock_irqrestore(&rq->lock, flags);
1215 * When add_timer_on() enqueues a timer into the timer wheel of an
1216 * idle CPU then this timer might expire before the next timer event
1217 * which is scheduled to wake up that CPU. In case of a completely
1218 * idle system the next event might even be infinite time into the
1219 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1220 * leaves the inner idle loop so the newly added timer is taken into
1221 * account when the CPU goes back to idle and evaluates the timer
1222 * wheel for the next timer event.
1224 void wake_up_idle_cpu(int cpu)
1226 struct rq *rq = cpu_rq(cpu);
1228 if (cpu == smp_processor_id())
1232 * This is safe, as this function is called with the timer
1233 * wheel base lock of (cpu) held. When the CPU is on the way
1234 * to idle and has not yet set rq->curr to idle then it will
1235 * be serialized on the timer wheel base lock and take the new
1236 * timer into account automatically.
1238 if (rq->curr != rq->idle)
1242 * We can set TIF_RESCHED on the idle task of the other CPU
1243 * lockless. The worst case is that the other CPU runs the
1244 * idle task through an additional NOOP schedule()
1246 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1248 /* NEED_RESCHED must be visible before we test polling */
1250 if (!tsk_is_polling(rq->idle))
1251 smp_send_reschedule(cpu);
1253 #endif /* CONFIG_NO_HZ */
1255 #else /* !CONFIG_SMP */
1256 static void resched_task(struct task_struct *p)
1258 assert_spin_locked(&task_rq(p)->lock);
1259 set_tsk_need_resched(p);
1261 #endif /* CONFIG_SMP */
1263 #if BITS_PER_LONG == 32
1264 # define WMULT_CONST (~0UL)
1266 # define WMULT_CONST (1UL << 32)
1269 #define WMULT_SHIFT 32
1272 * Shift right and round:
1274 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1277 * delta *= weight / lw
1279 static unsigned long
1280 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1281 struct load_weight *lw)
1285 if (!lw->inv_weight) {
1286 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1289 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1293 tmp = (u64)delta_exec * weight;
1295 * Check whether we'd overflow the 64-bit multiplication:
1297 if (unlikely(tmp > WMULT_CONST))
1298 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1301 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1303 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1306 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1312 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1319 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1320 * of tasks with abnormal "nice" values across CPUs the contribution that
1321 * each task makes to its run queue's load is weighted according to its
1322 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1323 * scaled version of the new time slice allocation that they receive on time
1327 #define WEIGHT_IDLEPRIO 2
1328 #define WMULT_IDLEPRIO (1 << 31)
1331 * Nice levels are multiplicative, with a gentle 10% change for every
1332 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1333 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1334 * that remained on nice 0.
1336 * The "10% effect" is relative and cumulative: from _any_ nice level,
1337 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1338 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1339 * If a task goes up by ~10% and another task goes down by ~10% then
1340 * the relative distance between them is ~25%.)
1342 static const int prio_to_weight[40] = {
1343 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1344 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1345 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1346 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1347 /* 0 */ 1024, 820, 655, 526, 423,
1348 /* 5 */ 335, 272, 215, 172, 137,
1349 /* 10 */ 110, 87, 70, 56, 45,
1350 /* 15 */ 36, 29, 23, 18, 15,
1354 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1356 * In cases where the weight does not change often, we can use the
1357 * precalculated inverse to speed up arithmetics by turning divisions
1358 * into multiplications:
1360 static const u32 prio_to_wmult[40] = {
1361 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1362 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1363 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1364 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1365 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1366 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1367 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1368 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1371 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1374 * runqueue iterator, to support SMP load-balancing between different
1375 * scheduling classes, without having to expose their internal data
1376 * structures to the load-balancing proper:
1378 struct rq_iterator {
1380 struct task_struct *(*start)(void *);
1381 struct task_struct *(*next)(void *);
1385 static unsigned long
1386 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1387 unsigned long max_load_move, struct sched_domain *sd,
1388 enum cpu_idle_type idle, int *all_pinned,
1389 int *this_best_prio, struct rq_iterator *iterator);
1392 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1393 struct sched_domain *sd, enum cpu_idle_type idle,
1394 struct rq_iterator *iterator);
1397 #ifdef CONFIG_CGROUP_CPUACCT
1398 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1400 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1403 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1405 update_load_add(&rq->load, load);
1408 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1410 update_load_sub(&rq->load, load);
1413 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1414 typedef int (*tg_visitor)(struct task_group *, void *);
1417 * Iterate the full tree, calling @down when first entering a node and @up when
1418 * leaving it for the final time.
1420 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1422 struct task_group *parent, *child;
1426 parent = &root_task_group;
1428 ret = (*down)(parent, data);
1431 list_for_each_entry_rcu(child, &parent->children, siblings) {
1438 ret = (*up)(parent, data);
1443 parent = parent->parent;
1452 static int tg_nop(struct task_group *tg, void *data)
1459 static unsigned long source_load(int cpu, int type);
1460 static unsigned long target_load(int cpu, int type);
1461 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1463 static unsigned long cpu_avg_load_per_task(int cpu)
1465 struct rq *rq = cpu_rq(cpu);
1466 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1469 rq->avg_load_per_task = rq->load.weight / nr_running;
1471 rq->avg_load_per_task = 0;
1473 return rq->avg_load_per_task;
1476 #ifdef CONFIG_FAIR_GROUP_SCHED
1478 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1481 * Calculate and set the cpu's group shares.
1484 update_group_shares_cpu(struct task_group *tg, int cpu,
1485 unsigned long sd_shares, unsigned long sd_rq_weight)
1487 unsigned long shares;
1488 unsigned long rq_weight;
1493 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1496 * \Sum shares * rq_weight
1497 * shares = -----------------------
1501 shares = (sd_shares * rq_weight) / sd_rq_weight;
1502 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1504 if (abs(shares - tg->se[cpu]->load.weight) >
1505 sysctl_sched_shares_thresh) {
1506 struct rq *rq = cpu_rq(cpu);
1507 unsigned long flags;
1509 spin_lock_irqsave(&rq->lock, flags);
1510 tg->cfs_rq[cpu]->shares = shares;
1512 __set_se_shares(tg->se[cpu], shares);
1513 spin_unlock_irqrestore(&rq->lock, flags);
1518 * Re-compute the task group their per cpu shares over the given domain.
1519 * This needs to be done in a bottom-up fashion because the rq weight of a
1520 * parent group depends on the shares of its child groups.
1522 static int tg_shares_up(struct task_group *tg, void *data)
1524 unsigned long weight, rq_weight = 0;
1525 unsigned long shares = 0;
1526 struct sched_domain *sd = data;
1529 for_each_cpu(i, sched_domain_span(sd)) {
1531 * If there are currently no tasks on the cpu pretend there
1532 * is one of average load so that when a new task gets to
1533 * run here it will not get delayed by group starvation.
1535 weight = tg->cfs_rq[i]->load.weight;
1537 weight = NICE_0_LOAD;
1539 tg->cfs_rq[i]->rq_weight = weight;
1540 rq_weight += weight;
1541 shares += tg->cfs_rq[i]->shares;
1544 if ((!shares && rq_weight) || shares > tg->shares)
1545 shares = tg->shares;
1547 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1548 shares = tg->shares;
1550 for_each_cpu(i, sched_domain_span(sd))
1551 update_group_shares_cpu(tg, i, shares, rq_weight);
1557 * Compute the cpu's hierarchical load factor for each task group.
1558 * This needs to be done in a top-down fashion because the load of a child
1559 * group is a fraction of its parents load.
1561 static int tg_load_down(struct task_group *tg, void *data)
1564 long cpu = (long)data;
1567 load = cpu_rq(cpu)->load.weight;
1569 load = tg->parent->cfs_rq[cpu]->h_load;
1570 load *= tg->cfs_rq[cpu]->shares;
1571 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1574 tg->cfs_rq[cpu]->h_load = load;
1579 static void update_shares(struct sched_domain *sd)
1581 u64 now = cpu_clock(raw_smp_processor_id());
1582 s64 elapsed = now - sd->last_update;
1584 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1585 sd->last_update = now;
1586 walk_tg_tree(tg_nop, tg_shares_up, sd);
1590 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1592 spin_unlock(&rq->lock);
1594 spin_lock(&rq->lock);
1597 static void update_h_load(long cpu)
1599 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1604 static inline void update_shares(struct sched_domain *sd)
1608 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1614 #ifdef CONFIG_PREEMPT
1617 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1618 * way at the expense of forcing extra atomic operations in all
1619 * invocations. This assures that the double_lock is acquired using the
1620 * same underlying policy as the spinlock_t on this architecture, which
1621 * reduces latency compared to the unfair variant below. However, it
1622 * also adds more overhead and therefore may reduce throughput.
1624 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1625 __releases(this_rq->lock)
1626 __acquires(busiest->lock)
1627 __acquires(this_rq->lock)
1629 spin_unlock(&this_rq->lock);
1630 double_rq_lock(this_rq, busiest);
1637 * Unfair double_lock_balance: Optimizes throughput at the expense of
1638 * latency by eliminating extra atomic operations when the locks are
1639 * already in proper order on entry. This favors lower cpu-ids and will
1640 * grant the double lock to lower cpus over higher ids under contention,
1641 * regardless of entry order into the function.
1643 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1644 __releases(this_rq->lock)
1645 __acquires(busiest->lock)
1646 __acquires(this_rq->lock)
1650 if (unlikely(!spin_trylock(&busiest->lock))) {
1651 if (busiest < this_rq) {
1652 spin_unlock(&this_rq->lock);
1653 spin_lock(&busiest->lock);
1654 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1657 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1662 #endif /* CONFIG_PREEMPT */
1665 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1667 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1669 if (unlikely(!irqs_disabled())) {
1670 /* printk() doesn't work good under rq->lock */
1671 spin_unlock(&this_rq->lock);
1675 return _double_lock_balance(this_rq, busiest);
1678 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1679 __releases(busiest->lock)
1681 spin_unlock(&busiest->lock);
1682 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1686 #ifdef CONFIG_FAIR_GROUP_SCHED
1687 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1690 cfs_rq->shares = shares;
1695 #include "sched_stats.h"
1696 #include "sched_idletask.c"
1697 #include "sched_fair.c"
1698 #include "sched_rt.c"
1699 #ifdef CONFIG_SCHED_DEBUG
1700 # include "sched_debug.c"
1703 #define sched_class_highest (&rt_sched_class)
1704 #define for_each_class(class) \
1705 for (class = sched_class_highest; class; class = class->next)
1707 static void inc_nr_running(struct rq *rq)
1712 static void dec_nr_running(struct rq *rq)
1717 static void set_load_weight(struct task_struct *p)
1719 if (task_has_rt_policy(p)) {
1720 p->se.load.weight = prio_to_weight[0] * 2;
1721 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1726 * SCHED_IDLE tasks get minimal weight:
1728 if (p->policy == SCHED_IDLE) {
1729 p->se.load.weight = WEIGHT_IDLEPRIO;
1730 p->se.load.inv_weight = WMULT_IDLEPRIO;
1734 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1735 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1738 static void update_avg(u64 *avg, u64 sample)
1740 s64 diff = sample - *avg;
1744 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1746 sched_info_queued(p);
1747 p->sched_class->enqueue_task(rq, p, wakeup);
1751 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1753 if (sleep && p->se.last_wakeup) {
1754 update_avg(&p->se.avg_overlap,
1755 p->se.sum_exec_runtime - p->se.last_wakeup);
1756 p->se.last_wakeup = 0;
1759 sched_info_dequeued(p);
1760 p->sched_class->dequeue_task(rq, p, sleep);
1765 * __normal_prio - return the priority that is based on the static prio
1767 static inline int __normal_prio(struct task_struct *p)
1769 return p->static_prio;
1773 * Calculate the expected normal priority: i.e. priority
1774 * without taking RT-inheritance into account. Might be
1775 * boosted by interactivity modifiers. Changes upon fork,
1776 * setprio syscalls, and whenever the interactivity
1777 * estimator recalculates.
1779 static inline int normal_prio(struct task_struct *p)
1783 if (task_has_rt_policy(p))
1784 prio = MAX_RT_PRIO-1 - p->rt_priority;
1786 prio = __normal_prio(p);
1791 * Calculate the current priority, i.e. the priority
1792 * taken into account by the scheduler. This value might
1793 * be boosted by RT tasks, or might be boosted by
1794 * interactivity modifiers. Will be RT if the task got
1795 * RT-boosted. If not then it returns p->normal_prio.
1797 static int effective_prio(struct task_struct *p)
1799 p->normal_prio = normal_prio(p);
1801 * If we are RT tasks or we were boosted to RT priority,
1802 * keep the priority unchanged. Otherwise, update priority
1803 * to the normal priority:
1805 if (!rt_prio(p->prio))
1806 return p->normal_prio;
1811 * activate_task - move a task to the runqueue.
1813 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1815 if (task_contributes_to_load(p))
1816 rq->nr_uninterruptible--;
1818 enqueue_task(rq, p, wakeup);
1823 * deactivate_task - remove a task from the runqueue.
1825 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1827 if (task_contributes_to_load(p))
1828 rq->nr_uninterruptible++;
1830 dequeue_task(rq, p, sleep);
1835 * task_curr - is this task currently executing on a CPU?
1836 * @p: the task in question.
1838 inline int task_curr(const struct task_struct *p)
1840 return cpu_curr(task_cpu(p)) == p;
1843 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1845 set_task_rq(p, cpu);
1848 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1849 * successfuly executed on another CPU. We must ensure that updates of
1850 * per-task data have been completed by this moment.
1853 task_thread_info(p)->cpu = cpu;
1857 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1858 const struct sched_class *prev_class,
1859 int oldprio, int running)
1861 if (prev_class != p->sched_class) {
1862 if (prev_class->switched_from)
1863 prev_class->switched_from(rq, p, running);
1864 p->sched_class->switched_to(rq, p, running);
1866 p->sched_class->prio_changed(rq, p, oldprio, running);
1871 /* Used instead of source_load when we know the type == 0 */
1872 static unsigned long weighted_cpuload(const int cpu)
1874 return cpu_rq(cpu)->load.weight;
1878 * Is this task likely cache-hot:
1881 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1886 * Buddy candidates are cache hot:
1888 if (sched_feat(CACHE_HOT_BUDDY) &&
1889 (&p->se == cfs_rq_of(&p->se)->next ||
1890 &p->se == cfs_rq_of(&p->se)->last))
1893 if (p->sched_class != &fair_sched_class)
1896 if (sysctl_sched_migration_cost == -1)
1898 if (sysctl_sched_migration_cost == 0)
1901 delta = now - p->se.exec_start;
1903 return delta < (s64)sysctl_sched_migration_cost;
1907 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1909 int old_cpu = task_cpu(p);
1910 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1911 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1912 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1915 clock_offset = old_rq->clock - new_rq->clock;
1917 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1919 #ifdef CONFIG_SCHEDSTATS
1920 if (p->se.wait_start)
1921 p->se.wait_start -= clock_offset;
1922 if (p->se.sleep_start)
1923 p->se.sleep_start -= clock_offset;
1924 if (p->se.block_start)
1925 p->se.block_start -= clock_offset;
1926 if (old_cpu != new_cpu) {
1927 schedstat_inc(p, se.nr_migrations);
1928 if (task_hot(p, old_rq->clock, NULL))
1929 schedstat_inc(p, se.nr_forced2_migrations);
1932 p->se.vruntime -= old_cfsrq->min_vruntime -
1933 new_cfsrq->min_vruntime;
1935 __set_task_cpu(p, new_cpu);
1938 struct migration_req {
1939 struct list_head list;
1941 struct task_struct *task;
1944 struct completion done;
1948 * The task's runqueue lock must be held.
1949 * Returns true if you have to wait for migration thread.
1952 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1954 struct rq *rq = task_rq(p);
1957 * If the task is not on a runqueue (and not running), then
1958 * it is sufficient to simply update the task's cpu field.
1960 if (!p->se.on_rq && !task_running(rq, p)) {
1961 set_task_cpu(p, dest_cpu);
1965 init_completion(&req->done);
1967 req->dest_cpu = dest_cpu;
1968 list_add(&req->list, &rq->migration_queue);
1974 * wait_task_inactive - wait for a thread to unschedule.
1976 * If @match_state is nonzero, it's the @p->state value just checked and
1977 * not expected to change. If it changes, i.e. @p might have woken up,
1978 * then return zero. When we succeed in waiting for @p to be off its CPU,
1979 * we return a positive number (its total switch count). If a second call
1980 * a short while later returns the same number, the caller can be sure that
1981 * @p has remained unscheduled the whole time.
1983 * The caller must ensure that the task *will* unschedule sometime soon,
1984 * else this function might spin for a *long* time. This function can't
1985 * be called with interrupts off, or it may introduce deadlock with
1986 * smp_call_function() if an IPI is sent by the same process we are
1987 * waiting to become inactive.
1989 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1991 unsigned long flags;
1998 * We do the initial early heuristics without holding
1999 * any task-queue locks at all. We'll only try to get
2000 * the runqueue lock when things look like they will
2006 * If the task is actively running on another CPU
2007 * still, just relax and busy-wait without holding
2010 * NOTE! Since we don't hold any locks, it's not
2011 * even sure that "rq" stays as the right runqueue!
2012 * But we don't care, since "task_running()" will
2013 * return false if the runqueue has changed and p
2014 * is actually now running somewhere else!
2016 while (task_running(rq, p)) {
2017 if (match_state && unlikely(p->state != match_state))
2023 * Ok, time to look more closely! We need the rq
2024 * lock now, to be *sure*. If we're wrong, we'll
2025 * just go back and repeat.
2027 rq = task_rq_lock(p, &flags);
2028 trace_sched_wait_task(rq, p);
2029 running = task_running(rq, p);
2030 on_rq = p->se.on_rq;
2032 if (!match_state || p->state == match_state)
2033 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2034 task_rq_unlock(rq, &flags);
2037 * If it changed from the expected state, bail out now.
2039 if (unlikely(!ncsw))
2043 * Was it really running after all now that we
2044 * checked with the proper locks actually held?
2046 * Oops. Go back and try again..
2048 if (unlikely(running)) {
2054 * It's not enough that it's not actively running,
2055 * it must be off the runqueue _entirely_, and not
2058 * So if it wa still runnable (but just not actively
2059 * running right now), it's preempted, and we should
2060 * yield - it could be a while.
2062 if (unlikely(on_rq)) {
2063 schedule_timeout_uninterruptible(1);
2068 * Ahh, all good. It wasn't running, and it wasn't
2069 * runnable, which means that it will never become
2070 * running in the future either. We're all done!
2079 * kick_process - kick a running thread to enter/exit the kernel
2080 * @p: the to-be-kicked thread
2082 * Cause a process which is running on another CPU to enter
2083 * kernel-mode, without any delay. (to get signals handled.)
2085 * NOTE: this function doesnt have to take the runqueue lock,
2086 * because all it wants to ensure is that the remote task enters
2087 * the kernel. If the IPI races and the task has been migrated
2088 * to another CPU then no harm is done and the purpose has been
2091 void kick_process(struct task_struct *p)
2097 if ((cpu != smp_processor_id()) && task_curr(p))
2098 smp_send_reschedule(cpu);
2103 * Return a low guess at the load of a migration-source cpu weighted
2104 * according to the scheduling class and "nice" value.
2106 * We want to under-estimate the load of migration sources, to
2107 * balance conservatively.
2109 static unsigned long source_load(int cpu, int type)
2111 struct rq *rq = cpu_rq(cpu);
2112 unsigned long total = weighted_cpuload(cpu);
2114 if (type == 0 || !sched_feat(LB_BIAS))
2117 return min(rq->cpu_load[type-1], total);
2121 * Return a high guess at the load of a migration-target cpu weighted
2122 * according to the scheduling class and "nice" value.
2124 static unsigned long target_load(int cpu, int type)
2126 struct rq *rq = cpu_rq(cpu);
2127 unsigned long total = weighted_cpuload(cpu);
2129 if (type == 0 || !sched_feat(LB_BIAS))
2132 return max(rq->cpu_load[type-1], total);
2136 * find_idlest_group finds and returns the least busy CPU group within the
2139 static struct sched_group *
2140 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2142 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2143 unsigned long min_load = ULONG_MAX, this_load = 0;
2144 int load_idx = sd->forkexec_idx;
2145 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2148 unsigned long load, avg_load;
2152 /* Skip over this group if it has no CPUs allowed */
2153 if (!cpumask_intersects(sched_group_cpus(group),
2157 local_group = cpumask_test_cpu(this_cpu,
2158 sched_group_cpus(group));
2160 /* Tally up the load of all CPUs in the group */
2163 for_each_cpu(i, sched_group_cpus(group)) {
2164 /* Bias balancing toward cpus of our domain */
2166 load = source_load(i, load_idx);
2168 load = target_load(i, load_idx);
2173 /* Adjust by relative CPU power of the group */
2174 avg_load = sg_div_cpu_power(group,
2175 avg_load * SCHED_LOAD_SCALE);
2178 this_load = avg_load;
2180 } else if (avg_load < min_load) {
2181 min_load = avg_load;
2184 } while (group = group->next, group != sd->groups);
2186 if (!idlest || 100*this_load < imbalance*min_load)
2192 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2195 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2197 unsigned long load, min_load = ULONG_MAX;
2201 /* Traverse only the allowed CPUs */
2202 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2203 load = weighted_cpuload(i);
2205 if (load < min_load || (load == min_load && i == this_cpu)) {
2215 * sched_balance_self: balance the current task (running on cpu) in domains
2216 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2219 * Balance, ie. select the least loaded group.
2221 * Returns the target CPU number, or the same CPU if no balancing is needed.
2223 * preempt must be disabled.
2225 static int sched_balance_self(int cpu, int flag)
2227 struct task_struct *t = current;
2228 struct sched_domain *tmp, *sd = NULL;
2230 for_each_domain(cpu, tmp) {
2232 * If power savings logic is enabled for a domain, stop there.
2234 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2236 if (tmp->flags & flag)
2244 struct sched_group *group;
2245 int new_cpu, weight;
2247 if (!(sd->flags & flag)) {
2252 group = find_idlest_group(sd, t, cpu);
2258 new_cpu = find_idlest_cpu(group, t, cpu);
2259 if (new_cpu == -1 || new_cpu == cpu) {
2260 /* Now try balancing at a lower domain level of cpu */
2265 /* Now try balancing at a lower domain level of new_cpu */
2267 weight = cpumask_weight(sched_domain_span(sd));
2269 for_each_domain(cpu, tmp) {
2270 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2272 if (tmp->flags & flag)
2275 /* while loop will break here if sd == NULL */
2281 #endif /* CONFIG_SMP */
2284 * try_to_wake_up - wake up a thread
2285 * @p: the to-be-woken-up thread
2286 * @state: the mask of task states that can be woken
2287 * @sync: do a synchronous wakeup?
2289 * Put it on the run-queue if it's not already there. The "current"
2290 * thread is always on the run-queue (except when the actual
2291 * re-schedule is in progress), and as such you're allowed to do
2292 * the simpler "current->state = TASK_RUNNING" to mark yourself
2293 * runnable without the overhead of this.
2295 * returns failure only if the task is already active.
2297 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2299 int cpu, orig_cpu, this_cpu, success = 0;
2300 unsigned long flags;
2304 if (!sched_feat(SYNC_WAKEUPS))
2308 if (sched_feat(LB_WAKEUP_UPDATE)) {
2309 struct sched_domain *sd;
2311 this_cpu = raw_smp_processor_id();
2314 for_each_domain(this_cpu, sd) {
2315 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2324 rq = task_rq_lock(p, &flags);
2325 update_rq_clock(rq);
2326 old_state = p->state;
2327 if (!(old_state & state))
2335 this_cpu = smp_processor_id();
2338 if (unlikely(task_running(rq, p)))
2341 cpu = p->sched_class->select_task_rq(p, sync);
2342 if (cpu != orig_cpu) {
2343 set_task_cpu(p, cpu);
2344 task_rq_unlock(rq, &flags);
2345 /* might preempt at this point */
2346 rq = task_rq_lock(p, &flags);
2347 old_state = p->state;
2348 if (!(old_state & state))
2353 this_cpu = smp_processor_id();
2357 #ifdef CONFIG_SCHEDSTATS
2358 schedstat_inc(rq, ttwu_count);
2359 if (cpu == this_cpu)
2360 schedstat_inc(rq, ttwu_local);
2362 struct sched_domain *sd;
2363 for_each_domain(this_cpu, sd) {
2364 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2365 schedstat_inc(sd, ttwu_wake_remote);
2370 #endif /* CONFIG_SCHEDSTATS */
2373 #endif /* CONFIG_SMP */
2374 schedstat_inc(p, se.nr_wakeups);
2376 schedstat_inc(p, se.nr_wakeups_sync);
2377 if (orig_cpu != cpu)
2378 schedstat_inc(p, se.nr_wakeups_migrate);
2379 if (cpu == this_cpu)
2380 schedstat_inc(p, se.nr_wakeups_local);
2382 schedstat_inc(p, se.nr_wakeups_remote);
2383 activate_task(rq, p, 1);
2387 trace_sched_wakeup(rq, p, success);
2388 check_preempt_curr(rq, p, sync);
2390 p->state = TASK_RUNNING;
2392 if (p->sched_class->task_wake_up)
2393 p->sched_class->task_wake_up(rq, p);
2396 current->se.last_wakeup = current->se.sum_exec_runtime;
2398 task_rq_unlock(rq, &flags);
2403 int wake_up_process(struct task_struct *p)
2405 return try_to_wake_up(p, TASK_ALL, 0);
2407 EXPORT_SYMBOL(wake_up_process);
2409 int wake_up_state(struct task_struct *p, unsigned int state)
2411 return try_to_wake_up(p, state, 0);
2415 * Perform scheduler related setup for a newly forked process p.
2416 * p is forked by current.
2418 * __sched_fork() is basic setup used by init_idle() too:
2420 static void __sched_fork(struct task_struct *p)
2422 p->se.exec_start = 0;
2423 p->se.sum_exec_runtime = 0;
2424 p->se.prev_sum_exec_runtime = 0;
2425 p->se.last_wakeup = 0;
2426 p->se.avg_overlap = 0;
2428 #ifdef CONFIG_SCHEDSTATS
2429 p->se.wait_start = 0;
2430 p->se.sum_sleep_runtime = 0;
2431 p->se.sleep_start = 0;
2432 p->se.block_start = 0;
2433 p->se.sleep_max = 0;
2434 p->se.block_max = 0;
2436 p->se.slice_max = 0;
2440 INIT_LIST_HEAD(&p->rt.run_list);
2442 INIT_LIST_HEAD(&p->se.group_node);
2444 #ifdef CONFIG_PREEMPT_NOTIFIERS
2445 INIT_HLIST_HEAD(&p->preempt_notifiers);
2449 * We mark the process as running here, but have not actually
2450 * inserted it onto the runqueue yet. This guarantees that
2451 * nobody will actually run it, and a signal or other external
2452 * event cannot wake it up and insert it on the runqueue either.
2454 p->state = TASK_RUNNING;
2458 * fork()/clone()-time setup:
2460 void sched_fork(struct task_struct *p, int clone_flags)
2462 int cpu = get_cpu();
2467 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2469 set_task_cpu(p, cpu);
2472 * Make sure we do not leak PI boosting priority to the child:
2474 p->prio = current->normal_prio;
2475 if (!rt_prio(p->prio))
2476 p->sched_class = &fair_sched_class;
2478 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2479 if (likely(sched_info_on()))
2480 memset(&p->sched_info, 0, sizeof(p->sched_info));
2482 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2485 #ifdef CONFIG_PREEMPT
2486 /* Want to start with kernel preemption disabled. */
2487 task_thread_info(p)->preempt_count = 1;
2489 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2495 * wake_up_new_task - wake up a newly created task for the first time.
2497 * This function will do some initial scheduler statistics housekeeping
2498 * that must be done for every newly created context, then puts the task
2499 * on the runqueue and wakes it.
2501 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2503 unsigned long flags;
2506 rq = task_rq_lock(p, &flags);
2507 BUG_ON(p->state != TASK_RUNNING);
2508 update_rq_clock(rq);
2510 p->prio = effective_prio(p);
2512 if (!p->sched_class->task_new || !current->se.on_rq) {
2513 activate_task(rq, p, 0);
2516 * Let the scheduling class do new task startup
2517 * management (if any):
2519 p->sched_class->task_new(rq, p);
2522 trace_sched_wakeup_new(rq, p, 1);
2523 check_preempt_curr(rq, p, 0);
2525 if (p->sched_class->task_wake_up)
2526 p->sched_class->task_wake_up(rq, p);
2528 task_rq_unlock(rq, &flags);
2531 #ifdef CONFIG_PREEMPT_NOTIFIERS
2534 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2535 * @notifier: notifier struct to register
2537 void preempt_notifier_register(struct preempt_notifier *notifier)
2539 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2541 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2544 * preempt_notifier_unregister - no longer interested in preemption notifications
2545 * @notifier: notifier struct to unregister
2547 * This is safe to call from within a preemption notifier.
2549 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2551 hlist_del(¬ifier->link);
2553 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2555 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2557 struct preempt_notifier *notifier;
2558 struct hlist_node *node;
2560 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2561 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2565 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2566 struct task_struct *next)
2568 struct preempt_notifier *notifier;
2569 struct hlist_node *node;
2571 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2572 notifier->ops->sched_out(notifier, next);
2575 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2577 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2582 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2583 struct task_struct *next)
2587 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2590 * prepare_task_switch - prepare to switch tasks
2591 * @rq: the runqueue preparing to switch
2592 * @prev: the current task that is being switched out
2593 * @next: the task we are going to switch to.
2595 * This is called with the rq lock held and interrupts off. It must
2596 * be paired with a subsequent finish_task_switch after the context
2599 * prepare_task_switch sets up locking and calls architecture specific
2603 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2604 struct task_struct *next)
2606 fire_sched_out_preempt_notifiers(prev, next);
2607 prepare_lock_switch(rq, next);
2608 prepare_arch_switch(next);
2612 * finish_task_switch - clean up after a task-switch
2613 * @rq: runqueue associated with task-switch
2614 * @prev: the thread we just switched away from.
2616 * finish_task_switch must be called after the context switch, paired
2617 * with a prepare_task_switch call before the context switch.
2618 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2619 * and do any other architecture-specific cleanup actions.
2621 * Note that we may have delayed dropping an mm in context_switch(). If
2622 * so, we finish that here outside of the runqueue lock. (Doing it
2623 * with the lock held can cause deadlocks; see schedule() for
2626 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2627 __releases(rq->lock)
2629 struct mm_struct *mm = rq->prev_mm;
2632 int post_schedule = 0;
2634 if (current->sched_class->needs_post_schedule)
2635 post_schedule = current->sched_class->needs_post_schedule(rq);
2641 * A task struct has one reference for the use as "current".
2642 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2643 * schedule one last time. The schedule call will never return, and
2644 * the scheduled task must drop that reference.
2645 * The test for TASK_DEAD must occur while the runqueue locks are
2646 * still held, otherwise prev could be scheduled on another cpu, die
2647 * there before we look at prev->state, and then the reference would
2649 * Manfred Spraul <manfred@colorfullife.com>
2651 prev_state = prev->state;
2652 finish_arch_switch(prev);
2653 finish_lock_switch(rq, prev);
2656 current->sched_class->post_schedule(rq);
2659 fire_sched_in_preempt_notifiers(current);
2662 if (unlikely(prev_state == TASK_DEAD)) {
2664 * Remove function-return probe instances associated with this
2665 * task and put them back on the free list.
2667 kprobe_flush_task(prev);
2668 put_task_struct(prev);
2673 * schedule_tail - first thing a freshly forked thread must call.
2674 * @prev: the thread we just switched away from.
2676 asmlinkage void schedule_tail(struct task_struct *prev)
2677 __releases(rq->lock)
2679 struct rq *rq = this_rq();
2681 finish_task_switch(rq, prev);
2682 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2683 /* In this case, finish_task_switch does not reenable preemption */
2686 if (current->set_child_tid)
2687 put_user(task_pid_vnr(current), current->set_child_tid);
2691 * context_switch - switch to the new MM and the new
2692 * thread's register state.
2695 context_switch(struct rq *rq, struct task_struct *prev,
2696 struct task_struct *next)
2698 struct mm_struct *mm, *oldmm;
2700 prepare_task_switch(rq, prev, next);
2701 trace_sched_switch(rq, prev, next);
2703 oldmm = prev->active_mm;
2705 * For paravirt, this is coupled with an exit in switch_to to
2706 * combine the page table reload and the switch backend into
2709 arch_enter_lazy_cpu_mode();
2711 if (unlikely(!mm)) {
2712 next->active_mm = oldmm;
2713 atomic_inc(&oldmm->mm_count);
2714 enter_lazy_tlb(oldmm, next);
2716 switch_mm(oldmm, mm, next);
2718 if (unlikely(!prev->mm)) {
2719 prev->active_mm = NULL;
2720 rq->prev_mm = oldmm;
2723 * Since the runqueue lock will be released by the next
2724 * task (which is an invalid locking op but in the case
2725 * of the scheduler it's an obvious special-case), so we
2726 * do an early lockdep release here:
2728 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2729 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2732 /* Here we just switch the register state and the stack. */
2733 switch_to(prev, next, prev);
2737 * this_rq must be evaluated again because prev may have moved
2738 * CPUs since it called schedule(), thus the 'rq' on its stack
2739 * frame will be invalid.
2741 finish_task_switch(this_rq(), prev);
2745 * nr_running, nr_uninterruptible and nr_context_switches:
2747 * externally visible scheduler statistics: current number of runnable
2748 * threads, current number of uninterruptible-sleeping threads, total
2749 * number of context switches performed since bootup.
2751 unsigned long nr_running(void)
2753 unsigned long i, sum = 0;
2755 for_each_online_cpu(i)
2756 sum += cpu_rq(i)->nr_running;
2761 unsigned long nr_uninterruptible(void)
2763 unsigned long i, sum = 0;
2765 for_each_possible_cpu(i)
2766 sum += cpu_rq(i)->nr_uninterruptible;
2769 * Since we read the counters lockless, it might be slightly
2770 * inaccurate. Do not allow it to go below zero though:
2772 if (unlikely((long)sum < 0))
2778 unsigned long long nr_context_switches(void)
2781 unsigned long long sum = 0;
2783 for_each_possible_cpu(i)
2784 sum += cpu_rq(i)->nr_switches;
2789 unsigned long nr_iowait(void)
2791 unsigned long i, sum = 0;
2793 for_each_possible_cpu(i)
2794 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2799 unsigned long nr_active(void)
2801 unsigned long i, running = 0, uninterruptible = 0;
2803 for_each_online_cpu(i) {
2804 running += cpu_rq(i)->nr_running;
2805 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2808 if (unlikely((long)uninterruptible < 0))
2809 uninterruptible = 0;
2811 return running + uninterruptible;
2815 * Update rq->cpu_load[] statistics. This function is usually called every
2816 * scheduler tick (TICK_NSEC).
2818 static void update_cpu_load(struct rq *this_rq)
2820 unsigned long this_load = this_rq->load.weight;
2823 this_rq->nr_load_updates++;
2825 /* Update our load: */
2826 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2827 unsigned long old_load, new_load;
2829 /* scale is effectively 1 << i now, and >> i divides by scale */
2831 old_load = this_rq->cpu_load[i];
2832 new_load = this_load;
2834 * Round up the averaging division if load is increasing. This
2835 * prevents us from getting stuck on 9 if the load is 10, for
2838 if (new_load > old_load)
2839 new_load += scale-1;
2840 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2847 * double_rq_lock - safely lock two runqueues
2849 * Note this does not disable interrupts like task_rq_lock,
2850 * you need to do so manually before calling.
2852 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2853 __acquires(rq1->lock)
2854 __acquires(rq2->lock)
2856 BUG_ON(!irqs_disabled());
2858 spin_lock(&rq1->lock);
2859 __acquire(rq2->lock); /* Fake it out ;) */
2862 spin_lock(&rq1->lock);
2863 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2865 spin_lock(&rq2->lock);
2866 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2869 update_rq_clock(rq1);
2870 update_rq_clock(rq2);
2874 * double_rq_unlock - safely unlock two runqueues
2876 * Note this does not restore interrupts like task_rq_unlock,
2877 * you need to do so manually after calling.
2879 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2880 __releases(rq1->lock)
2881 __releases(rq2->lock)
2883 spin_unlock(&rq1->lock);
2885 spin_unlock(&rq2->lock);
2887 __release(rq2->lock);
2891 * If dest_cpu is allowed for this process, migrate the task to it.
2892 * This is accomplished by forcing the cpu_allowed mask to only
2893 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2894 * the cpu_allowed mask is restored.
2896 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2898 struct migration_req req;
2899 unsigned long flags;
2902 rq = task_rq_lock(p, &flags);
2903 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2904 || unlikely(!cpu_active(dest_cpu)))
2907 /* force the process onto the specified CPU */
2908 if (migrate_task(p, dest_cpu, &req)) {
2909 /* Need to wait for migration thread (might exit: take ref). */
2910 struct task_struct *mt = rq->migration_thread;
2912 get_task_struct(mt);
2913 task_rq_unlock(rq, &flags);
2914 wake_up_process(mt);
2915 put_task_struct(mt);
2916 wait_for_completion(&req.done);
2921 task_rq_unlock(rq, &flags);
2925 * sched_exec - execve() is a valuable balancing opportunity, because at
2926 * this point the task has the smallest effective memory and cache footprint.
2928 void sched_exec(void)
2930 int new_cpu, this_cpu = get_cpu();
2931 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2933 if (new_cpu != this_cpu)
2934 sched_migrate_task(current, new_cpu);
2938 * pull_task - move a task from a remote runqueue to the local runqueue.
2939 * Both runqueues must be locked.
2941 static void pull_task(struct rq *src_rq, struct task_struct *p,
2942 struct rq *this_rq, int this_cpu)
2944 deactivate_task(src_rq, p, 0);
2945 set_task_cpu(p, this_cpu);
2946 activate_task(this_rq, p, 0);
2948 * Note that idle threads have a prio of MAX_PRIO, for this test
2949 * to be always true for them.
2951 check_preempt_curr(this_rq, p, 0);
2955 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2958 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2959 struct sched_domain *sd, enum cpu_idle_type idle,
2963 * We do not migrate tasks that are:
2964 * 1) running (obviously), or
2965 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2966 * 3) are cache-hot on their current CPU.
2968 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2969 schedstat_inc(p, se.nr_failed_migrations_affine);
2974 if (task_running(rq, p)) {
2975 schedstat_inc(p, se.nr_failed_migrations_running);
2980 * Aggressive migration if:
2981 * 1) task is cache cold, or
2982 * 2) too many balance attempts have failed.
2985 if (!task_hot(p, rq->clock, sd) ||
2986 sd->nr_balance_failed > sd->cache_nice_tries) {
2987 #ifdef CONFIG_SCHEDSTATS
2988 if (task_hot(p, rq->clock, sd)) {
2989 schedstat_inc(sd, lb_hot_gained[idle]);
2990 schedstat_inc(p, se.nr_forced_migrations);
2996 if (task_hot(p, rq->clock, sd)) {
2997 schedstat_inc(p, se.nr_failed_migrations_hot);
3003 static unsigned long
3004 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3005 unsigned long max_load_move, struct sched_domain *sd,
3006 enum cpu_idle_type idle, int *all_pinned,
3007 int *this_best_prio, struct rq_iterator *iterator)
3009 int loops = 0, pulled = 0, pinned = 0;
3010 struct task_struct *p;
3011 long rem_load_move = max_load_move;
3013 if (max_load_move == 0)
3019 * Start the load-balancing iterator:
3021 p = iterator->start(iterator->arg);
3023 if (!p || loops++ > sysctl_sched_nr_migrate)
3026 if ((p->se.load.weight >> 1) > rem_load_move ||
3027 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3028 p = iterator->next(iterator->arg);
3032 pull_task(busiest, p, this_rq, this_cpu);
3034 rem_load_move -= p->se.load.weight;
3036 #ifdef CONFIG_PREEMPT
3038 * NEWIDLE balancing is a source of latency, so preemptible kernels
3039 * will stop after the first task is pulled to minimize the critical
3042 if (idle == CPU_NEWLY_IDLE)
3047 * We only want to steal up to the prescribed amount of weighted load.
3049 if (rem_load_move > 0) {
3050 if (p->prio < *this_best_prio)
3051 *this_best_prio = p->prio;
3052 p = iterator->next(iterator->arg);
3057 * Right now, this is one of only two places pull_task() is called,
3058 * so we can safely collect pull_task() stats here rather than
3059 * inside pull_task().
3061 schedstat_add(sd, lb_gained[idle], pulled);
3064 *all_pinned = pinned;
3066 return max_load_move - rem_load_move;
3070 * move_tasks tries to move up to max_load_move weighted load from busiest to
3071 * this_rq, as part of a balancing operation within domain "sd".
3072 * Returns 1 if successful and 0 otherwise.
3074 * Called with both runqueues locked.
3076 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3077 unsigned long max_load_move,
3078 struct sched_domain *sd, enum cpu_idle_type idle,
3081 const struct sched_class *class = sched_class_highest;
3082 unsigned long total_load_moved = 0;
3083 int this_best_prio = this_rq->curr->prio;
3087 class->load_balance(this_rq, this_cpu, busiest,
3088 max_load_move - total_load_moved,
3089 sd, idle, all_pinned, &this_best_prio);
3090 class = class->next;
3092 #ifdef CONFIG_PREEMPT
3094 * NEWIDLE balancing is a source of latency, so preemptible
3095 * kernels will stop after the first task is pulled to minimize
3096 * the critical section.
3098 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3101 } while (class && max_load_move > total_load_moved);
3103 return total_load_moved > 0;
3107 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3108 struct sched_domain *sd, enum cpu_idle_type idle,
3109 struct rq_iterator *iterator)
3111 struct task_struct *p = iterator->start(iterator->arg);
3115 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3116 pull_task(busiest, p, this_rq, this_cpu);
3118 * Right now, this is only the second place pull_task()
3119 * is called, so we can safely collect pull_task()
3120 * stats here rather than inside pull_task().
3122 schedstat_inc(sd, lb_gained[idle]);
3126 p = iterator->next(iterator->arg);
3133 * move_one_task tries to move exactly one task from busiest to this_rq, as
3134 * part of active balancing operations within "domain".
3135 * Returns 1 if successful and 0 otherwise.
3137 * Called with both runqueues locked.
3139 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3140 struct sched_domain *sd, enum cpu_idle_type idle)
3142 const struct sched_class *class;
3144 for (class = sched_class_highest; class; class = class->next)
3145 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3152 * find_busiest_group finds and returns the busiest CPU group within the
3153 * domain. It calculates and returns the amount of weighted load which
3154 * should be moved to restore balance via the imbalance parameter.
3156 static struct sched_group *
3157 find_busiest_group(struct sched_domain *sd, int this_cpu,
3158 unsigned long *imbalance, enum cpu_idle_type idle,
3159 int *sd_idle, const struct cpumask *cpus, int *balance)
3161 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3162 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3163 unsigned long max_pull;
3164 unsigned long busiest_load_per_task, busiest_nr_running;
3165 unsigned long this_load_per_task, this_nr_running;
3166 int load_idx, group_imb = 0;
3167 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3168 int power_savings_balance = 1;
3169 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3170 unsigned long min_nr_running = ULONG_MAX;
3171 struct sched_group *group_min = NULL, *group_leader = NULL;
3174 max_load = this_load = total_load = total_pwr = 0;
3175 busiest_load_per_task = busiest_nr_running = 0;
3176 this_load_per_task = this_nr_running = 0;
3178 if (idle == CPU_NOT_IDLE)
3179 load_idx = sd->busy_idx;
3180 else if (idle == CPU_NEWLY_IDLE)
3181 load_idx = sd->newidle_idx;
3183 load_idx = sd->idle_idx;
3186 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3189 int __group_imb = 0;
3190 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3191 unsigned long sum_nr_running, sum_weighted_load;
3192 unsigned long sum_avg_load_per_task;
3193 unsigned long avg_load_per_task;
3195 local_group = cpumask_test_cpu(this_cpu,
3196 sched_group_cpus(group));
3199 balance_cpu = cpumask_first(sched_group_cpus(group));
3201 /* Tally up the load of all CPUs in the group */
3202 sum_weighted_load = sum_nr_running = avg_load = 0;
3203 sum_avg_load_per_task = avg_load_per_task = 0;
3206 min_cpu_load = ~0UL;
3208 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3209 struct rq *rq = cpu_rq(i);
3211 if (*sd_idle && rq->nr_running)
3214 /* Bias balancing toward cpus of our domain */
3216 if (idle_cpu(i) && !first_idle_cpu) {
3221 load = target_load(i, load_idx);
3223 load = source_load(i, load_idx);
3224 if (load > max_cpu_load)
3225 max_cpu_load = load;
3226 if (min_cpu_load > load)
3227 min_cpu_load = load;
3231 sum_nr_running += rq->nr_running;
3232 sum_weighted_load += weighted_cpuload(i);
3234 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3238 * First idle cpu or the first cpu(busiest) in this sched group
3239 * is eligible for doing load balancing at this and above
3240 * domains. In the newly idle case, we will allow all the cpu's
3241 * to do the newly idle load balance.
3243 if (idle != CPU_NEWLY_IDLE && local_group &&
3244 balance_cpu != this_cpu && balance) {
3249 total_load += avg_load;
3250 total_pwr += group->__cpu_power;
3252 /* Adjust by relative CPU power of the group */
3253 avg_load = sg_div_cpu_power(group,
3254 avg_load * SCHED_LOAD_SCALE);
3258 * Consider the group unbalanced when the imbalance is larger
3259 * than the average weight of two tasks.
3261 * APZ: with cgroup the avg task weight can vary wildly and
3262 * might not be a suitable number - should we keep a
3263 * normalized nr_running number somewhere that negates
3266 avg_load_per_task = sg_div_cpu_power(group,
3267 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3269 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3272 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3275 this_load = avg_load;
3277 this_nr_running = sum_nr_running;
3278 this_load_per_task = sum_weighted_load;
3279 } else if (avg_load > max_load &&
3280 (sum_nr_running > group_capacity || __group_imb)) {
3281 max_load = avg_load;
3283 busiest_nr_running = sum_nr_running;
3284 busiest_load_per_task = sum_weighted_load;
3285 group_imb = __group_imb;
3288 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3290 * Busy processors will not participate in power savings
3293 if (idle == CPU_NOT_IDLE ||
3294 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3298 * If the local group is idle or completely loaded
3299 * no need to do power savings balance at this domain
3301 if (local_group && (this_nr_running >= group_capacity ||
3303 power_savings_balance = 0;
3306 * If a group is already running at full capacity or idle,
3307 * don't include that group in power savings calculations
3309 if (!power_savings_balance || sum_nr_running >= group_capacity
3314 * Calculate the group which has the least non-idle load.
3315 * This is the group from where we need to pick up the load
3318 if ((sum_nr_running < min_nr_running) ||
3319 (sum_nr_running == min_nr_running &&
3320 cpumask_first(sched_group_cpus(group)) >
3321 cpumask_first(sched_group_cpus(group_min)))) {
3323 min_nr_running = sum_nr_running;
3324 min_load_per_task = sum_weighted_load /
3329 * Calculate the group which is almost near its
3330 * capacity but still has some space to pick up some load
3331 * from other group and save more power
3333 if (sum_nr_running <= group_capacity - 1) {
3334 if (sum_nr_running > leader_nr_running ||
3335 (sum_nr_running == leader_nr_running &&
3336 cpumask_first(sched_group_cpus(group)) <
3337 cpumask_first(sched_group_cpus(group_leader)))) {
3338 group_leader = group;
3339 leader_nr_running = sum_nr_running;
3344 group = group->next;
3345 } while (group != sd->groups);
3347 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3350 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3352 if (this_load >= avg_load ||
3353 100*max_load <= sd->imbalance_pct*this_load)
3356 busiest_load_per_task /= busiest_nr_running;
3358 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3361 * We're trying to get all the cpus to the average_load, so we don't
3362 * want to push ourselves above the average load, nor do we wish to
3363 * reduce the max loaded cpu below the average load, as either of these
3364 * actions would just result in more rebalancing later, and ping-pong
3365 * tasks around. Thus we look for the minimum possible imbalance.
3366 * Negative imbalances (*we* are more loaded than anyone else) will
3367 * be counted as no imbalance for these purposes -- we can't fix that
3368 * by pulling tasks to us. Be careful of negative numbers as they'll
3369 * appear as very large values with unsigned longs.
3371 if (max_load <= busiest_load_per_task)
3375 * In the presence of smp nice balancing, certain scenarios can have
3376 * max load less than avg load(as we skip the groups at or below
3377 * its cpu_power, while calculating max_load..)
3379 if (max_load < avg_load) {
3381 goto small_imbalance;
3384 /* Don't want to pull so many tasks that a group would go idle */
3385 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3387 /* How much load to actually move to equalise the imbalance */
3388 *imbalance = min(max_pull * busiest->__cpu_power,
3389 (avg_load - this_load) * this->__cpu_power)
3393 * if *imbalance is less than the average load per runnable task
3394 * there is no gaurantee that any tasks will be moved so we'll have
3395 * a think about bumping its value to force at least one task to be
3398 if (*imbalance < busiest_load_per_task) {
3399 unsigned long tmp, pwr_now, pwr_move;
3403 pwr_move = pwr_now = 0;
3405 if (this_nr_running) {
3406 this_load_per_task /= this_nr_running;
3407 if (busiest_load_per_task > this_load_per_task)
3410 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3412 if (max_load - this_load + busiest_load_per_task >=
3413 busiest_load_per_task * imbn) {
3414 *imbalance = busiest_load_per_task;
3419 * OK, we don't have enough imbalance to justify moving tasks,
3420 * however we may be able to increase total CPU power used by
3424 pwr_now += busiest->__cpu_power *
3425 min(busiest_load_per_task, max_load);
3426 pwr_now += this->__cpu_power *
3427 min(this_load_per_task, this_load);
3428 pwr_now /= SCHED_LOAD_SCALE;
3430 /* Amount of load we'd subtract */
3431 tmp = sg_div_cpu_power(busiest,
3432 busiest_load_per_task * SCHED_LOAD_SCALE);
3434 pwr_move += busiest->__cpu_power *
3435 min(busiest_load_per_task, max_load - tmp);
3437 /* Amount of load we'd add */
3438 if (max_load * busiest->__cpu_power <
3439 busiest_load_per_task * SCHED_LOAD_SCALE)
3440 tmp = sg_div_cpu_power(this,
3441 max_load * busiest->__cpu_power);
3443 tmp = sg_div_cpu_power(this,
3444 busiest_load_per_task * SCHED_LOAD_SCALE);
3445 pwr_move += this->__cpu_power *
3446 min(this_load_per_task, this_load + tmp);
3447 pwr_move /= SCHED_LOAD_SCALE;
3449 /* Move if we gain throughput */
3450 if (pwr_move > pwr_now)
3451 *imbalance = busiest_load_per_task;
3457 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3458 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3461 if (this == group_leader && group_leader != group_min) {
3462 *imbalance = min_load_per_task;
3463 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3464 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3465 cpumask_first(sched_group_cpus(group_leader));
3476 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3479 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3480 unsigned long imbalance, const struct cpumask *cpus)
3482 struct rq *busiest = NULL, *rq;
3483 unsigned long max_load = 0;
3486 for_each_cpu(i, sched_group_cpus(group)) {
3489 if (!cpumask_test_cpu(i, cpus))
3493 wl = weighted_cpuload(i);
3495 if (rq->nr_running == 1 && wl > imbalance)
3498 if (wl > max_load) {
3508 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3509 * so long as it is large enough.
3511 #define MAX_PINNED_INTERVAL 512
3514 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3515 * tasks if there is an imbalance.
3517 static int load_balance(int this_cpu, struct rq *this_rq,
3518 struct sched_domain *sd, enum cpu_idle_type idle,
3519 int *balance, struct cpumask *cpus)
3521 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3522 struct sched_group *group;
3523 unsigned long imbalance;
3525 unsigned long flags;
3527 cpumask_setall(cpus);
3530 * When power savings policy is enabled for the parent domain, idle
3531 * sibling can pick up load irrespective of busy siblings. In this case,
3532 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3533 * portraying it as CPU_NOT_IDLE.
3535 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3536 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3539 schedstat_inc(sd, lb_count[idle]);
3543 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3550 schedstat_inc(sd, lb_nobusyg[idle]);
3554 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3556 schedstat_inc(sd, lb_nobusyq[idle]);
3560 BUG_ON(busiest == this_rq);
3562 schedstat_add(sd, lb_imbalance[idle], imbalance);
3565 if (busiest->nr_running > 1) {
3567 * Attempt to move tasks. If find_busiest_group has found
3568 * an imbalance but busiest->nr_running <= 1, the group is
3569 * still unbalanced. ld_moved simply stays zero, so it is
3570 * correctly treated as an imbalance.
3572 local_irq_save(flags);
3573 double_rq_lock(this_rq, busiest);
3574 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3575 imbalance, sd, idle, &all_pinned);
3576 double_rq_unlock(this_rq, busiest);
3577 local_irq_restore(flags);
3580 * some other cpu did the load balance for us.
3582 if (ld_moved && this_cpu != smp_processor_id())
3583 resched_cpu(this_cpu);
3585 /* All tasks on this runqueue were pinned by CPU affinity */
3586 if (unlikely(all_pinned)) {
3587 cpumask_clear_cpu(cpu_of(busiest), cpus);
3588 if (!cpumask_empty(cpus))
3595 schedstat_inc(sd, lb_failed[idle]);
3596 sd->nr_balance_failed++;
3598 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3600 spin_lock_irqsave(&busiest->lock, flags);
3602 /* don't kick the migration_thread, if the curr
3603 * task on busiest cpu can't be moved to this_cpu
3605 if (!cpumask_test_cpu(this_cpu,
3606 &busiest->curr->cpus_allowed)) {
3607 spin_unlock_irqrestore(&busiest->lock, flags);
3609 goto out_one_pinned;
3612 if (!busiest->active_balance) {
3613 busiest->active_balance = 1;
3614 busiest->push_cpu = this_cpu;
3617 spin_unlock_irqrestore(&busiest->lock, flags);
3619 wake_up_process(busiest->migration_thread);
3622 * We've kicked active balancing, reset the failure
3625 sd->nr_balance_failed = sd->cache_nice_tries+1;
3628 sd->nr_balance_failed = 0;
3630 if (likely(!active_balance)) {
3631 /* We were unbalanced, so reset the balancing interval */
3632 sd->balance_interval = sd->min_interval;
3635 * If we've begun active balancing, start to back off. This
3636 * case may not be covered by the all_pinned logic if there
3637 * is only 1 task on the busy runqueue (because we don't call
3640 if (sd->balance_interval < sd->max_interval)
3641 sd->balance_interval *= 2;
3644 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3645 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3651 schedstat_inc(sd, lb_balanced[idle]);
3653 sd->nr_balance_failed = 0;
3656 /* tune up the balancing interval */
3657 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3658 (sd->balance_interval < sd->max_interval))
3659 sd->balance_interval *= 2;
3661 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3662 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3673 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3674 * tasks if there is an imbalance.
3676 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3677 * this_rq is locked.
3680 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3681 struct cpumask *cpus)
3683 struct sched_group *group;
3684 struct rq *busiest = NULL;
3685 unsigned long imbalance;
3690 cpumask_setall(cpus);
3693 * When power savings policy is enabled for the parent domain, idle
3694 * sibling can pick up load irrespective of busy siblings. In this case,
3695 * let the state of idle sibling percolate up as IDLE, instead of
3696 * portraying it as CPU_NOT_IDLE.
3698 if (sd->flags & SD_SHARE_CPUPOWER &&
3699 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3702 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3704 update_shares_locked(this_rq, sd);
3705 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3706 &sd_idle, cpus, NULL);
3708 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3712 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3714 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3718 BUG_ON(busiest == this_rq);
3720 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3723 if (busiest->nr_running > 1) {
3724 /* Attempt to move tasks */
3725 double_lock_balance(this_rq, busiest);
3726 /* this_rq->clock is already updated */
3727 update_rq_clock(busiest);
3728 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3729 imbalance, sd, CPU_NEWLY_IDLE,
3731 double_unlock_balance(this_rq, busiest);
3733 if (unlikely(all_pinned)) {
3734 cpumask_clear_cpu(cpu_of(busiest), cpus);
3735 if (!cpumask_empty(cpus))
3741 int active_balance = 0;
3743 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3744 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3745 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3748 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3751 if (sd->nr_balance_failed++ < 2)
3755 * The only task running in a non-idle cpu can be moved to this
3756 * cpu in an attempt to completely freeup the other CPU
3757 * package. The same method used to move task in load_balance()
3758 * have been extended for load_balance_newidle() to speedup
3759 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3761 * The package power saving logic comes from
3762 * find_busiest_group(). If there are no imbalance, then
3763 * f_b_g() will return NULL. However when sched_mc={1,2} then
3764 * f_b_g() will select a group from which a running task may be
3765 * pulled to this cpu in order to make the other package idle.
3766 * If there is no opportunity to make a package idle and if
3767 * there are no imbalance, then f_b_g() will return NULL and no
3768 * action will be taken in load_balance_newidle().
3770 * Under normal task pull operation due to imbalance, there
3771 * will be more than one task in the source run queue and
3772 * move_tasks() will succeed. ld_moved will be true and this
3773 * active balance code will not be triggered.
3776 /* Lock busiest in correct order while this_rq is held */
3777 double_lock_balance(this_rq, busiest);
3780 * don't kick the migration_thread, if the curr
3781 * task on busiest cpu can't be moved to this_cpu
3783 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3784 double_unlock_balance(this_rq, busiest);
3789 if (!busiest->active_balance) {
3790 busiest->active_balance = 1;
3791 busiest->push_cpu = this_cpu;
3795 double_unlock_balance(this_rq, busiest);
3797 * Should not call ttwu while holding a rq->lock
3799 spin_unlock(&this_rq->lock);
3801 wake_up_process(busiest->migration_thread);
3802 spin_lock(&this_rq->lock);
3805 sd->nr_balance_failed = 0;
3807 update_shares_locked(this_rq, sd);
3811 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3812 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3813 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3815 sd->nr_balance_failed = 0;
3821 * idle_balance is called by schedule() if this_cpu is about to become
3822 * idle. Attempts to pull tasks from other CPUs.
3824 static void idle_balance(int this_cpu, struct rq *this_rq)
3826 struct sched_domain *sd;
3827 int pulled_task = 0;
3828 unsigned long next_balance = jiffies + HZ;
3829 cpumask_var_t tmpmask;
3831 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3834 for_each_domain(this_cpu, sd) {
3835 unsigned long interval;
3837 if (!(sd->flags & SD_LOAD_BALANCE))
3840 if (sd->flags & SD_BALANCE_NEWIDLE)
3841 /* If we've pulled tasks over stop searching: */
3842 pulled_task = load_balance_newidle(this_cpu, this_rq,
3845 interval = msecs_to_jiffies(sd->balance_interval);
3846 if (time_after(next_balance, sd->last_balance + interval))
3847 next_balance = sd->last_balance + interval;
3851 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3853 * We are going idle. next_balance may be set based on
3854 * a busy processor. So reset next_balance.
3856 this_rq->next_balance = next_balance;
3858 free_cpumask_var(tmpmask);
3862 * active_load_balance is run by migration threads. It pushes running tasks
3863 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3864 * running on each physical CPU where possible, and avoids physical /
3865 * logical imbalances.
3867 * Called with busiest_rq locked.
3869 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3871 int target_cpu = busiest_rq->push_cpu;
3872 struct sched_domain *sd;
3873 struct rq *target_rq;
3875 /* Is there any task to move? */
3876 if (busiest_rq->nr_running <= 1)
3879 target_rq = cpu_rq(target_cpu);
3882 * This condition is "impossible", if it occurs
3883 * we need to fix it. Originally reported by
3884 * Bjorn Helgaas on a 128-cpu setup.
3886 BUG_ON(busiest_rq == target_rq);
3888 /* move a task from busiest_rq to target_rq */
3889 double_lock_balance(busiest_rq, target_rq);
3890 update_rq_clock(busiest_rq);
3891 update_rq_clock(target_rq);
3893 /* Search for an sd spanning us and the target CPU. */
3894 for_each_domain(target_cpu, sd) {
3895 if ((sd->flags & SD_LOAD_BALANCE) &&
3896 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3901 schedstat_inc(sd, alb_count);
3903 if (move_one_task(target_rq, target_cpu, busiest_rq,
3905 schedstat_inc(sd, alb_pushed);
3907 schedstat_inc(sd, alb_failed);
3909 double_unlock_balance(busiest_rq, target_rq);
3914 atomic_t load_balancer;
3915 cpumask_var_t cpu_mask;
3916 } nohz ____cacheline_aligned = {
3917 .load_balancer = ATOMIC_INIT(-1),
3921 * This routine will try to nominate the ilb (idle load balancing)
3922 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3923 * load balancing on behalf of all those cpus. If all the cpus in the system
3924 * go into this tickless mode, then there will be no ilb owner (as there is
3925 * no need for one) and all the cpus will sleep till the next wakeup event
3928 * For the ilb owner, tick is not stopped. And this tick will be used
3929 * for idle load balancing. ilb owner will still be part of
3932 * While stopping the tick, this cpu will become the ilb owner if there
3933 * is no other owner. And will be the owner till that cpu becomes busy
3934 * or if all cpus in the system stop their ticks at which point
3935 * there is no need for ilb owner.
3937 * When the ilb owner becomes busy, it nominates another owner, during the
3938 * next busy scheduler_tick()
3940 int select_nohz_load_balancer(int stop_tick)
3942 int cpu = smp_processor_id();
3945 cpumask_set_cpu(cpu, nohz.cpu_mask);
3946 cpu_rq(cpu)->in_nohz_recently = 1;
3949 * If we are going offline and still the leader, give up!
3951 if (!cpu_active(cpu) &&
3952 atomic_read(&nohz.load_balancer) == cpu) {
3953 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3958 /* time for ilb owner also to sleep */
3959 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3960 if (atomic_read(&nohz.load_balancer) == cpu)
3961 atomic_set(&nohz.load_balancer, -1);
3965 if (atomic_read(&nohz.load_balancer) == -1) {
3966 /* make me the ilb owner */
3967 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3969 } else if (atomic_read(&nohz.load_balancer) == cpu)
3972 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3975 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3977 if (atomic_read(&nohz.load_balancer) == cpu)
3978 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3985 static DEFINE_SPINLOCK(balancing);
3988 * It checks each scheduling domain to see if it is due to be balanced,
3989 * and initiates a balancing operation if so.
3991 * Balancing parameters are set up in arch_init_sched_domains.
3993 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3996 struct rq *rq = cpu_rq(cpu);
3997 unsigned long interval;
3998 struct sched_domain *sd;
3999 /* Earliest time when we have to do rebalance again */
4000 unsigned long next_balance = jiffies + 60*HZ;
4001 int update_next_balance = 0;
4005 /* Fails alloc? Rebalancing probably not a priority right now. */
4006 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
4009 for_each_domain(cpu, sd) {
4010 if (!(sd->flags & SD_LOAD_BALANCE))
4013 interval = sd->balance_interval;
4014 if (idle != CPU_IDLE)
4015 interval *= sd->busy_factor;
4017 /* scale ms to jiffies */
4018 interval = msecs_to_jiffies(interval);
4019 if (unlikely(!interval))
4021 if (interval > HZ*NR_CPUS/10)
4022 interval = HZ*NR_CPUS/10;
4024 need_serialize = sd->flags & SD_SERIALIZE;
4026 if (need_serialize) {
4027 if (!spin_trylock(&balancing))
4031 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4032 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
4034 * We've pulled tasks over so either we're no
4035 * longer idle, or one of our SMT siblings is
4038 idle = CPU_NOT_IDLE;
4040 sd->last_balance = jiffies;
4043 spin_unlock(&balancing);
4045 if (time_after(next_balance, sd->last_balance + interval)) {
4046 next_balance = sd->last_balance + interval;
4047 update_next_balance = 1;
4051 * Stop the load balance at this level. There is another
4052 * CPU in our sched group which is doing load balancing more
4060 * next_balance will be updated only when there is a need.
4061 * When the cpu is attached to null domain for ex, it will not be
4064 if (likely(update_next_balance))
4065 rq->next_balance = next_balance;
4067 free_cpumask_var(tmp);
4071 * run_rebalance_domains is triggered when needed from the scheduler tick.
4072 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4073 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4075 static void run_rebalance_domains(struct softirq_action *h)
4077 int this_cpu = smp_processor_id();
4078 struct rq *this_rq = cpu_rq(this_cpu);
4079 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4080 CPU_IDLE : CPU_NOT_IDLE;
4082 rebalance_domains(this_cpu, idle);
4086 * If this cpu is the owner for idle load balancing, then do the
4087 * balancing on behalf of the other idle cpus whose ticks are
4090 if (this_rq->idle_at_tick &&
4091 atomic_read(&nohz.load_balancer) == this_cpu) {
4095 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4096 if (balance_cpu == this_cpu)
4100 * If this cpu gets work to do, stop the load balancing
4101 * work being done for other cpus. Next load
4102 * balancing owner will pick it up.
4107 rebalance_domains(balance_cpu, CPU_IDLE);
4109 rq = cpu_rq(balance_cpu);
4110 if (time_after(this_rq->next_balance, rq->next_balance))
4111 this_rq->next_balance = rq->next_balance;
4118 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4120 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4121 * idle load balancing owner or decide to stop the periodic load balancing,
4122 * if the whole system is idle.
4124 static inline void trigger_load_balance(struct rq *rq, int cpu)
4128 * If we were in the nohz mode recently and busy at the current
4129 * scheduler tick, then check if we need to nominate new idle
4132 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4133 rq->in_nohz_recently = 0;
4135 if (atomic_read(&nohz.load_balancer) == cpu) {
4136 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4137 atomic_set(&nohz.load_balancer, -1);
4140 if (atomic_read(&nohz.load_balancer) == -1) {
4142 * simple selection for now: Nominate the
4143 * first cpu in the nohz list to be the next
4146 * TBD: Traverse the sched domains and nominate
4147 * the nearest cpu in the nohz.cpu_mask.
4149 int ilb = cpumask_first(nohz.cpu_mask);
4151 if (ilb < nr_cpu_ids)
4157 * If this cpu is idle and doing idle load balancing for all the
4158 * cpus with ticks stopped, is it time for that to stop?
4160 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4161 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4167 * If this cpu is idle and the idle load balancing is done by
4168 * someone else, then no need raise the SCHED_SOFTIRQ
4170 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4171 cpumask_test_cpu(cpu, nohz.cpu_mask))
4174 if (time_after_eq(jiffies, rq->next_balance))
4175 raise_softirq(SCHED_SOFTIRQ);
4178 #else /* CONFIG_SMP */
4181 * on UP we do not need to balance between CPUs:
4183 static inline void idle_balance(int cpu, struct rq *rq)
4189 DEFINE_PER_CPU(struct kernel_stat, kstat);
4191 EXPORT_PER_CPU_SYMBOL(kstat);
4194 * Return any ns on the sched_clock that have not yet been banked in
4195 * @p in case that task is currently running.
4197 unsigned long long task_delta_exec(struct task_struct *p)
4199 unsigned long flags;
4203 rq = task_rq_lock(p, &flags);
4205 if (task_current(rq, p)) {
4208 update_rq_clock(rq);
4209 delta_exec = rq->clock - p->se.exec_start;
4210 if ((s64)delta_exec > 0)
4214 task_rq_unlock(rq, &flags);
4220 * Account user cpu time to a process.
4221 * @p: the process that the cpu time gets accounted to
4222 * @cputime: the cpu time spent in user space since the last update
4223 * @cputime_scaled: cputime scaled by cpu frequency
4225 void account_user_time(struct task_struct *p, cputime_t cputime,
4226 cputime_t cputime_scaled)
4228 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4231 /* Add user time to process. */
4232 p->utime = cputime_add(p->utime, cputime);
4233 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4234 account_group_user_time(p, cputime);
4236 /* Add user time to cpustat. */
4237 tmp = cputime_to_cputime64(cputime);
4238 if (TASK_NICE(p) > 0)
4239 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4241 cpustat->user = cputime64_add(cpustat->user, tmp);
4242 /* Account for user time used */
4243 acct_update_integrals(p);
4247 * Account guest cpu time to a process.
4248 * @p: the process that the cpu time gets accounted to
4249 * @cputime: the cpu time spent in virtual machine since the last update
4250 * @cputime_scaled: cputime scaled by cpu frequency
4252 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4253 cputime_t cputime_scaled)
4256 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4258 tmp = cputime_to_cputime64(cputime);
4260 /* Add guest time to process. */
4261 p->utime = cputime_add(p->utime, cputime);
4262 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4263 account_group_user_time(p, cputime);
4264 p->gtime = cputime_add(p->gtime, cputime);
4266 /* Add guest time to cpustat. */
4267 cpustat->user = cputime64_add(cpustat->user, tmp);
4268 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4272 * Account system cpu time to a process.
4273 * @p: the process that the cpu time gets accounted to
4274 * @hardirq_offset: the offset to subtract from hardirq_count()
4275 * @cputime: the cpu time spent in kernel space since the last update
4276 * @cputime_scaled: cputime scaled by cpu frequency
4278 void account_system_time(struct task_struct *p, int hardirq_offset,
4279 cputime_t cputime, cputime_t cputime_scaled)
4281 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4284 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4285 account_guest_time(p, cputime, cputime_scaled);
4289 /* Add system time to process. */
4290 p->stime = cputime_add(p->stime, cputime);
4291 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4292 account_group_system_time(p, cputime);
4294 /* Add system time to cpustat. */
4295 tmp = cputime_to_cputime64(cputime);
4296 if (hardirq_count() - hardirq_offset)
4297 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4298 else if (softirq_count())
4299 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4301 cpustat->system = cputime64_add(cpustat->system, tmp);
4303 /* Account for system time used */
4304 acct_update_integrals(p);
4308 * Account for involuntary wait time.
4309 * @steal: the cpu time spent in involuntary wait
4311 void account_steal_time(cputime_t cputime)
4313 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4314 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4316 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4320 * Account for idle time.
4321 * @cputime: the cpu time spent in idle wait
4323 void account_idle_time(cputime_t cputime)
4325 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4326 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4327 struct rq *rq = this_rq();
4329 if (atomic_read(&rq->nr_iowait) > 0)
4330 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4332 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4335 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4338 * Account a single tick of cpu time.
4339 * @p: the process that the cpu time gets accounted to
4340 * @user_tick: indicates if the tick is a user or a system tick
4342 void account_process_tick(struct task_struct *p, int user_tick)
4344 cputime_t one_jiffy = jiffies_to_cputime(1);
4345 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4346 struct rq *rq = this_rq();
4349 account_user_time(p, one_jiffy, one_jiffy_scaled);
4350 else if (p != rq->idle)
4351 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4354 account_idle_time(one_jiffy);
4358 * Account multiple ticks of steal time.
4359 * @p: the process from which the cpu time has been stolen
4360 * @ticks: number of stolen ticks
4362 void account_steal_ticks(unsigned long ticks)
4364 account_steal_time(jiffies_to_cputime(ticks));
4368 * Account multiple ticks of idle time.
4369 * @ticks: number of stolen ticks
4371 void account_idle_ticks(unsigned long ticks)
4373 account_idle_time(jiffies_to_cputime(ticks));
4379 * Use precise platform statistics if available:
4381 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4382 cputime_t task_utime(struct task_struct *p)
4387 cputime_t task_stime(struct task_struct *p)
4392 cputime_t task_utime(struct task_struct *p)
4394 clock_t utime = cputime_to_clock_t(p->utime),
4395 total = utime + cputime_to_clock_t(p->stime);
4399 * Use CFS's precise accounting:
4401 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4405 do_div(temp, total);
4407 utime = (clock_t)temp;
4409 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4410 return p->prev_utime;
4413 cputime_t task_stime(struct task_struct *p)
4418 * Use CFS's precise accounting. (we subtract utime from
4419 * the total, to make sure the total observed by userspace
4420 * grows monotonically - apps rely on that):
4422 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4423 cputime_to_clock_t(task_utime(p));
4426 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4428 return p->prev_stime;
4432 inline cputime_t task_gtime(struct task_struct *p)
4438 * This function gets called by the timer code, with HZ frequency.
4439 * We call it with interrupts disabled.
4441 * It also gets called by the fork code, when changing the parent's
4444 void scheduler_tick(void)
4446 int cpu = smp_processor_id();
4447 struct rq *rq = cpu_rq(cpu);
4448 struct task_struct *curr = rq->curr;
4452 spin_lock(&rq->lock);
4453 update_rq_clock(rq);
4454 update_cpu_load(rq);
4455 curr->sched_class->task_tick(rq, curr, 0);
4456 spin_unlock(&rq->lock);
4459 rq->idle_at_tick = idle_cpu(cpu);
4460 trigger_load_balance(rq, cpu);
4464 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4465 defined(CONFIG_PREEMPT_TRACER))
4467 static inline unsigned long get_parent_ip(unsigned long addr)
4469 if (in_lock_functions(addr)) {
4470 addr = CALLER_ADDR2;
4471 if (in_lock_functions(addr))
4472 addr = CALLER_ADDR3;
4477 void __kprobes add_preempt_count(int val)
4479 #ifdef CONFIG_DEBUG_PREEMPT
4483 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4486 preempt_count() += val;
4487 #ifdef CONFIG_DEBUG_PREEMPT
4489 * Spinlock count overflowing soon?
4491 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4494 if (preempt_count() == val)
4495 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4497 EXPORT_SYMBOL(add_preempt_count);
4499 void __kprobes sub_preempt_count(int val)
4501 #ifdef CONFIG_DEBUG_PREEMPT
4505 if (DEBUG_LOCKS_WARN_ON(val > preempt_count() - (!!kernel_locked())))
4508 * Is the spinlock portion underflowing?
4510 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4511 !(preempt_count() & PREEMPT_MASK)))
4515 if (preempt_count() == val)
4516 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4517 preempt_count() -= val;
4519 EXPORT_SYMBOL(sub_preempt_count);
4524 * Print scheduling while atomic bug:
4526 static noinline void __schedule_bug(struct task_struct *prev)
4528 struct pt_regs *regs = get_irq_regs();
4530 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4531 prev->comm, prev->pid, preempt_count());
4533 debug_show_held_locks(prev);
4535 if (irqs_disabled())
4536 print_irqtrace_events(prev);
4545 * Various schedule()-time debugging checks and statistics:
4547 static inline void schedule_debug(struct task_struct *prev)
4550 * Test if we are atomic. Since do_exit() needs to call into
4551 * schedule() atomically, we ignore that path for now.
4552 * Otherwise, whine if we are scheduling when we should not be.
4554 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4555 __schedule_bug(prev);
4557 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4559 schedstat_inc(this_rq(), sched_count);
4560 #ifdef CONFIG_SCHEDSTATS
4561 if (unlikely(prev->lock_depth >= 0)) {
4562 schedstat_inc(this_rq(), bkl_count);
4563 schedstat_inc(prev, sched_info.bkl_count);
4569 * Pick up the highest-prio task:
4571 static inline struct task_struct *
4572 pick_next_task(struct rq *rq, struct task_struct *prev)
4574 const struct sched_class *class;
4575 struct task_struct *p;
4578 * Optimization: we know that if all tasks are in
4579 * the fair class we can call that function directly:
4581 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4582 p = fair_sched_class.pick_next_task(rq);
4587 class = sched_class_highest;
4589 p = class->pick_next_task(rq);
4593 * Will never be NULL as the idle class always
4594 * returns a non-NULL p:
4596 class = class->next;
4601 * schedule() is the main scheduler function.
4603 asmlinkage void __sched schedule(void)
4605 struct task_struct *prev, *next;
4606 unsigned long *switch_count;
4612 cpu = smp_processor_id();
4616 switch_count = &prev->nivcsw;
4618 release_kernel_lock(prev);
4619 need_resched_nonpreemptible:
4621 schedule_debug(prev);
4623 if (sched_feat(HRTICK))
4626 spin_lock_irq(&rq->lock);
4627 update_rq_clock(rq);
4628 clear_tsk_need_resched(prev);
4630 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4631 if (unlikely(signal_pending_state(prev->state, prev)))
4632 prev->state = TASK_RUNNING;
4634 deactivate_task(rq, prev, 1);
4635 switch_count = &prev->nvcsw;
4639 if (prev->sched_class->pre_schedule)
4640 prev->sched_class->pre_schedule(rq, prev);
4643 if (unlikely(!rq->nr_running))
4644 idle_balance(cpu, rq);
4646 prev->sched_class->put_prev_task(rq, prev);
4647 next = pick_next_task(rq, prev);
4649 if (likely(prev != next)) {
4650 sched_info_switch(prev, next);
4656 context_switch(rq, prev, next); /* unlocks the rq */
4658 * the context switch might have flipped the stack from under
4659 * us, hence refresh the local variables.
4661 cpu = smp_processor_id();
4664 spin_unlock_irq(&rq->lock);
4666 if (unlikely(reacquire_kernel_lock(current) < 0))
4667 goto need_resched_nonpreemptible;
4669 preempt_enable_no_resched();
4670 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4673 EXPORT_SYMBOL(schedule);
4675 #ifdef CONFIG_PREEMPT
4677 * this is the entry point to schedule() from in-kernel preemption
4678 * off of preempt_enable. Kernel preemptions off return from interrupt
4679 * occur there and call schedule directly.
4681 asmlinkage void __sched preempt_schedule(void)
4683 struct thread_info *ti = current_thread_info();
4686 * If there is a non-zero preempt_count or interrupts are disabled,
4687 * we do not want to preempt the current task. Just return..
4689 if (likely(ti->preempt_count || irqs_disabled()))
4693 add_preempt_count(PREEMPT_ACTIVE);
4695 sub_preempt_count(PREEMPT_ACTIVE);
4698 * Check again in case we missed a preemption opportunity
4699 * between schedule and now.
4702 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4704 EXPORT_SYMBOL(preempt_schedule);
4707 * this is the entry point to schedule() from kernel preemption
4708 * off of irq context.
4709 * Note, that this is called and return with irqs disabled. This will
4710 * protect us against recursive calling from irq.
4712 asmlinkage void __sched preempt_schedule_irq(void)
4714 struct thread_info *ti = current_thread_info();
4716 /* Catch callers which need to be fixed */
4717 BUG_ON(ti->preempt_count || !irqs_disabled());
4720 add_preempt_count(PREEMPT_ACTIVE);
4723 local_irq_disable();
4724 sub_preempt_count(PREEMPT_ACTIVE);
4727 * Check again in case we missed a preemption opportunity
4728 * between schedule and now.
4731 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4734 #endif /* CONFIG_PREEMPT */
4736 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4739 return try_to_wake_up(curr->private, mode, sync);
4741 EXPORT_SYMBOL(default_wake_function);
4744 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4745 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4746 * number) then we wake all the non-exclusive tasks and one exclusive task.
4748 * There are circumstances in which we can try to wake a task which has already
4749 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4750 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4752 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4753 int nr_exclusive, int sync, void *key)
4755 wait_queue_t *curr, *next;
4757 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4758 unsigned flags = curr->flags;
4760 if (curr->func(curr, mode, sync, key) &&
4761 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4767 * __wake_up - wake up threads blocked on a waitqueue.
4769 * @mode: which threads
4770 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4771 * @key: is directly passed to the wakeup function
4773 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4774 int nr_exclusive, void *key)
4776 unsigned long flags;
4778 spin_lock_irqsave(&q->lock, flags);
4779 __wake_up_common(q, mode, nr_exclusive, 0, key);
4780 spin_unlock_irqrestore(&q->lock, flags);
4782 EXPORT_SYMBOL(__wake_up);
4785 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4787 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4789 __wake_up_common(q, mode, 1, 0, NULL);
4793 * __wake_up_sync - wake up threads blocked on a waitqueue.
4795 * @mode: which threads
4796 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4798 * The sync wakeup differs that the waker knows that it will schedule
4799 * away soon, so while the target thread will be woken up, it will not
4800 * be migrated to another CPU - ie. the two threads are 'synchronized'
4801 * with each other. This can prevent needless bouncing between CPUs.
4803 * On UP it can prevent extra preemption.
4806 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4808 unsigned long flags;
4814 if (unlikely(!nr_exclusive))
4817 spin_lock_irqsave(&q->lock, flags);
4818 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4819 spin_unlock_irqrestore(&q->lock, flags);
4821 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4824 * complete: - signals a single thread waiting on this completion
4825 * @x: holds the state of this particular completion
4827 * This will wake up a single thread waiting on this completion. Threads will be
4828 * awakened in the same order in which they were queued.
4830 * See also complete_all(), wait_for_completion() and related routines.
4832 void complete(struct completion *x)
4834 unsigned long flags;
4836 spin_lock_irqsave(&x->wait.lock, flags);
4838 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4839 spin_unlock_irqrestore(&x->wait.lock, flags);
4841 EXPORT_SYMBOL(complete);
4844 * complete_all: - signals all threads waiting on this completion
4845 * @x: holds the state of this particular completion
4847 * This will wake up all threads waiting on this particular completion event.
4849 void complete_all(struct completion *x)
4851 unsigned long flags;
4853 spin_lock_irqsave(&x->wait.lock, flags);
4854 x->done += UINT_MAX/2;
4855 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4856 spin_unlock_irqrestore(&x->wait.lock, flags);
4858 EXPORT_SYMBOL(complete_all);
4860 static inline long __sched
4861 do_wait_for_common(struct completion *x, long timeout, int state)
4864 DECLARE_WAITQUEUE(wait, current);
4866 wait.flags |= WQ_FLAG_EXCLUSIVE;
4867 __add_wait_queue_tail(&x->wait, &wait);
4869 if (signal_pending_state(state, current)) {
4870 timeout = -ERESTARTSYS;
4873 __set_current_state(state);
4874 spin_unlock_irq(&x->wait.lock);
4875 timeout = schedule_timeout(timeout);
4876 spin_lock_irq(&x->wait.lock);
4877 } while (!x->done && timeout);
4878 __remove_wait_queue(&x->wait, &wait);
4883 return timeout ?: 1;
4887 wait_for_common(struct completion *x, long timeout, int state)
4891 spin_lock_irq(&x->wait.lock);
4892 timeout = do_wait_for_common(x, timeout, state);
4893 spin_unlock_irq(&x->wait.lock);
4898 * wait_for_completion: - waits for completion of a task
4899 * @x: holds the state of this particular completion
4901 * This waits to be signaled for completion of a specific task. It is NOT
4902 * interruptible and there is no timeout.
4904 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4905 * and interrupt capability. Also see complete().
4907 void __sched wait_for_completion(struct completion *x)
4909 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4911 EXPORT_SYMBOL(wait_for_completion);
4914 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4915 * @x: holds the state of this particular completion
4916 * @timeout: timeout value in jiffies
4918 * This waits for either a completion of a specific task to be signaled or for a
4919 * specified timeout to expire. The timeout is in jiffies. It is not
4922 unsigned long __sched
4923 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4925 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4927 EXPORT_SYMBOL(wait_for_completion_timeout);
4930 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4931 * @x: holds the state of this particular completion
4933 * This waits for completion of a specific task to be signaled. It is
4936 int __sched wait_for_completion_interruptible(struct completion *x)
4938 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4939 if (t == -ERESTARTSYS)
4943 EXPORT_SYMBOL(wait_for_completion_interruptible);
4946 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4947 * @x: holds the state of this particular completion
4948 * @timeout: timeout value in jiffies
4950 * This waits for either a completion of a specific task to be signaled or for a
4951 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4953 unsigned long __sched
4954 wait_for_completion_interruptible_timeout(struct completion *x,
4955 unsigned long timeout)
4957 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4959 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4962 * wait_for_completion_killable: - waits for completion of a task (killable)
4963 * @x: holds the state of this particular completion
4965 * This waits to be signaled for completion of a specific task. It can be
4966 * interrupted by a kill signal.
4968 int __sched wait_for_completion_killable(struct completion *x)
4970 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4971 if (t == -ERESTARTSYS)
4975 EXPORT_SYMBOL(wait_for_completion_killable);
4978 * try_wait_for_completion - try to decrement a completion without blocking
4979 * @x: completion structure
4981 * Returns: 0 if a decrement cannot be done without blocking
4982 * 1 if a decrement succeeded.
4984 * If a completion is being used as a counting completion,
4985 * attempt to decrement the counter without blocking. This
4986 * enables us to avoid waiting if the resource the completion
4987 * is protecting is not available.
4989 bool try_wait_for_completion(struct completion *x)
4993 spin_lock_irq(&x->wait.lock);
4998 spin_unlock_irq(&x->wait.lock);
5001 EXPORT_SYMBOL(try_wait_for_completion);
5004 * completion_done - Test to see if a completion has any waiters
5005 * @x: completion structure
5007 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5008 * 1 if there are no waiters.
5011 bool completion_done(struct completion *x)
5015 spin_lock_irq(&x->wait.lock);
5018 spin_unlock_irq(&x->wait.lock);
5021 EXPORT_SYMBOL(completion_done);
5024 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5026 unsigned long flags;
5029 init_waitqueue_entry(&wait, current);
5031 __set_current_state(state);
5033 spin_lock_irqsave(&q->lock, flags);
5034 __add_wait_queue(q, &wait);
5035 spin_unlock(&q->lock);
5036 timeout = schedule_timeout(timeout);
5037 spin_lock_irq(&q->lock);
5038 __remove_wait_queue(q, &wait);
5039 spin_unlock_irqrestore(&q->lock, flags);
5044 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5046 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5048 EXPORT_SYMBOL(interruptible_sleep_on);
5051 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5053 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5055 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5057 void __sched sleep_on(wait_queue_head_t *q)
5059 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5061 EXPORT_SYMBOL(sleep_on);
5063 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5065 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5067 EXPORT_SYMBOL(sleep_on_timeout);
5069 #ifdef CONFIG_RT_MUTEXES
5072 * rt_mutex_setprio - set the current priority of a task
5074 * @prio: prio value (kernel-internal form)
5076 * This function changes the 'effective' priority of a task. It does
5077 * not touch ->normal_prio like __setscheduler().
5079 * Used by the rt_mutex code to implement priority inheritance logic.
5081 void rt_mutex_setprio(struct task_struct *p, int prio)
5083 unsigned long flags;
5084 int oldprio, on_rq, running;
5086 const struct sched_class *prev_class = p->sched_class;
5088 BUG_ON(prio < 0 || prio > MAX_PRIO);
5090 rq = task_rq_lock(p, &flags);
5091 update_rq_clock(rq);
5094 on_rq = p->se.on_rq;
5095 running = task_current(rq, p);
5097 dequeue_task(rq, p, 0);
5099 p->sched_class->put_prev_task(rq, p);
5102 p->sched_class = &rt_sched_class;
5104 p->sched_class = &fair_sched_class;
5109 p->sched_class->set_curr_task(rq);
5111 enqueue_task(rq, p, 0);
5113 check_class_changed(rq, p, prev_class, oldprio, running);
5115 task_rq_unlock(rq, &flags);
5120 void set_user_nice(struct task_struct *p, long nice)
5122 int old_prio, delta, on_rq;
5123 unsigned long flags;
5126 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5129 * We have to be careful, if called from sys_setpriority(),
5130 * the task might be in the middle of scheduling on another CPU.
5132 rq = task_rq_lock(p, &flags);
5133 update_rq_clock(rq);
5135 * The RT priorities are set via sched_setscheduler(), but we still
5136 * allow the 'normal' nice value to be set - but as expected
5137 * it wont have any effect on scheduling until the task is
5138 * SCHED_FIFO/SCHED_RR:
5140 if (task_has_rt_policy(p)) {
5141 p->static_prio = NICE_TO_PRIO(nice);
5144 on_rq = p->se.on_rq;
5146 dequeue_task(rq, p, 0);
5148 p->static_prio = NICE_TO_PRIO(nice);
5151 p->prio = effective_prio(p);
5152 delta = p->prio - old_prio;
5155 enqueue_task(rq, p, 0);
5157 * If the task increased its priority or is running and
5158 * lowered its priority, then reschedule its CPU:
5160 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5161 resched_task(rq->curr);
5164 task_rq_unlock(rq, &flags);
5166 EXPORT_SYMBOL(set_user_nice);
5169 * can_nice - check if a task can reduce its nice value
5173 int can_nice(const struct task_struct *p, const int nice)
5175 /* convert nice value [19,-20] to rlimit style value [1,40] */
5176 int nice_rlim = 20 - nice;
5178 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5179 capable(CAP_SYS_NICE));
5182 #ifdef __ARCH_WANT_SYS_NICE
5185 * sys_nice - change the priority of the current process.
5186 * @increment: priority increment
5188 * sys_setpriority is a more generic, but much slower function that
5189 * does similar things.
5191 asmlinkage long sys_nice(int increment)
5196 * Setpriority might change our priority at the same moment.
5197 * We don't have to worry. Conceptually one call occurs first
5198 * and we have a single winner.
5200 if (increment < -40)
5205 nice = PRIO_TO_NICE(current->static_prio) + increment;
5211 if (increment < 0 && !can_nice(current, nice))
5214 retval = security_task_setnice(current, nice);
5218 set_user_nice(current, nice);
5225 * task_prio - return the priority value of a given task.
5226 * @p: the task in question.
5228 * This is the priority value as seen by users in /proc.
5229 * RT tasks are offset by -200. Normal tasks are centered
5230 * around 0, value goes from -16 to +15.
5232 int task_prio(const struct task_struct *p)
5234 return p->prio - MAX_RT_PRIO;
5238 * task_nice - return the nice value of a given task.
5239 * @p: the task in question.
5241 int task_nice(const struct task_struct *p)
5243 return TASK_NICE(p);
5245 EXPORT_SYMBOL(task_nice);
5248 * idle_cpu - is a given cpu idle currently?
5249 * @cpu: the processor in question.
5251 int idle_cpu(int cpu)
5253 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5257 * idle_task - return the idle task for a given cpu.
5258 * @cpu: the processor in question.
5260 struct task_struct *idle_task(int cpu)
5262 return cpu_rq(cpu)->idle;
5266 * find_process_by_pid - find a process with a matching PID value.
5267 * @pid: the pid in question.
5269 static struct task_struct *find_process_by_pid(pid_t pid)
5271 return pid ? find_task_by_vpid(pid) : current;
5274 /* Actually do priority change: must hold rq lock. */
5276 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5278 BUG_ON(p->se.on_rq);
5281 switch (p->policy) {
5285 p->sched_class = &fair_sched_class;
5289 p->sched_class = &rt_sched_class;
5293 p->rt_priority = prio;
5294 p->normal_prio = normal_prio(p);
5295 /* we are holding p->pi_lock already */
5296 p->prio = rt_mutex_getprio(p);
5301 * check the target process has a UID that matches the current process's
5303 static bool check_same_owner(struct task_struct *p)
5305 const struct cred *cred = current_cred(), *pcred;
5309 pcred = __task_cred(p);
5310 match = (cred->euid == pcred->euid ||
5311 cred->euid == pcred->uid);
5316 static int __sched_setscheduler(struct task_struct *p, int policy,
5317 struct sched_param *param, bool user)
5319 int retval, oldprio, oldpolicy = -1, on_rq, running;
5320 unsigned long flags;
5321 const struct sched_class *prev_class = p->sched_class;
5324 /* may grab non-irq protected spin_locks */
5325 BUG_ON(in_interrupt());
5327 /* double check policy once rq lock held */
5329 policy = oldpolicy = p->policy;
5330 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5331 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5332 policy != SCHED_IDLE)
5335 * Valid priorities for SCHED_FIFO and SCHED_RR are
5336 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5337 * SCHED_BATCH and SCHED_IDLE is 0.
5339 if (param->sched_priority < 0 ||
5340 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5341 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5343 if (rt_policy(policy) != (param->sched_priority != 0))
5347 * Allow unprivileged RT tasks to decrease priority:
5349 if (user && !capable(CAP_SYS_NICE)) {
5350 if (rt_policy(policy)) {
5351 unsigned long rlim_rtprio;
5353 if (!lock_task_sighand(p, &flags))
5355 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5356 unlock_task_sighand(p, &flags);
5358 /* can't set/change the rt policy */
5359 if (policy != p->policy && !rlim_rtprio)
5362 /* can't increase priority */
5363 if (param->sched_priority > p->rt_priority &&
5364 param->sched_priority > rlim_rtprio)
5368 * Like positive nice levels, dont allow tasks to
5369 * move out of SCHED_IDLE either:
5371 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5374 /* can't change other user's priorities */
5375 if (!check_same_owner(p))
5380 #ifdef CONFIG_RT_GROUP_SCHED
5382 * Do not allow realtime tasks into groups that have no runtime
5385 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5386 task_group(p)->rt_bandwidth.rt_runtime == 0)
5390 retval = security_task_setscheduler(p, policy, param);
5396 * make sure no PI-waiters arrive (or leave) while we are
5397 * changing the priority of the task:
5399 spin_lock_irqsave(&p->pi_lock, flags);
5401 * To be able to change p->policy safely, the apropriate
5402 * runqueue lock must be held.
5404 rq = __task_rq_lock(p);
5405 /* recheck policy now with rq lock held */
5406 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5407 policy = oldpolicy = -1;
5408 __task_rq_unlock(rq);
5409 spin_unlock_irqrestore(&p->pi_lock, flags);
5412 update_rq_clock(rq);
5413 on_rq = p->se.on_rq;
5414 running = task_current(rq, p);
5416 deactivate_task(rq, p, 0);
5418 p->sched_class->put_prev_task(rq, p);
5421 __setscheduler(rq, p, policy, param->sched_priority);
5424 p->sched_class->set_curr_task(rq);
5426 activate_task(rq, p, 0);
5428 check_class_changed(rq, p, prev_class, oldprio, running);
5430 __task_rq_unlock(rq);
5431 spin_unlock_irqrestore(&p->pi_lock, flags);
5433 rt_mutex_adjust_pi(p);
5439 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5440 * @p: the task in question.
5441 * @policy: new policy.
5442 * @param: structure containing the new RT priority.
5444 * NOTE that the task may be already dead.
5446 int sched_setscheduler(struct task_struct *p, int policy,
5447 struct sched_param *param)
5449 return __sched_setscheduler(p, policy, param, true);
5451 EXPORT_SYMBOL_GPL(sched_setscheduler);
5454 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5455 * @p: the task in question.
5456 * @policy: new policy.
5457 * @param: structure containing the new RT priority.
5459 * Just like sched_setscheduler, only don't bother checking if the
5460 * current context has permission. For example, this is needed in
5461 * stop_machine(): we create temporary high priority worker threads,
5462 * but our caller might not have that capability.
5464 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5465 struct sched_param *param)
5467 return __sched_setscheduler(p, policy, param, false);
5471 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5473 struct sched_param lparam;
5474 struct task_struct *p;
5477 if (!param || pid < 0)
5479 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5484 p = find_process_by_pid(pid);
5486 retval = sched_setscheduler(p, policy, &lparam);
5493 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5494 * @pid: the pid in question.
5495 * @policy: new policy.
5496 * @param: structure containing the new RT priority.
5499 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5501 /* negative values for policy are not valid */
5505 return do_sched_setscheduler(pid, policy, param);
5509 * sys_sched_setparam - set/change the RT priority of a thread
5510 * @pid: the pid in question.
5511 * @param: structure containing the new RT priority.
5513 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5515 return do_sched_setscheduler(pid, -1, param);
5519 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5520 * @pid: the pid in question.
5522 asmlinkage long sys_sched_getscheduler(pid_t pid)
5524 struct task_struct *p;
5531 read_lock(&tasklist_lock);
5532 p = find_process_by_pid(pid);
5534 retval = security_task_getscheduler(p);
5538 read_unlock(&tasklist_lock);
5543 * sys_sched_getscheduler - get the RT priority of a thread
5544 * @pid: the pid in question.
5545 * @param: structure containing the RT priority.
5547 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5549 struct sched_param lp;
5550 struct task_struct *p;
5553 if (!param || pid < 0)
5556 read_lock(&tasklist_lock);
5557 p = find_process_by_pid(pid);
5562 retval = security_task_getscheduler(p);
5566 lp.sched_priority = p->rt_priority;
5567 read_unlock(&tasklist_lock);
5570 * This one might sleep, we cannot do it with a spinlock held ...
5572 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5577 read_unlock(&tasklist_lock);
5581 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5583 cpumask_var_t cpus_allowed, new_mask;
5584 struct task_struct *p;
5588 read_lock(&tasklist_lock);
5590 p = find_process_by_pid(pid);
5592 read_unlock(&tasklist_lock);
5598 * It is not safe to call set_cpus_allowed with the
5599 * tasklist_lock held. We will bump the task_struct's
5600 * usage count and then drop tasklist_lock.
5603 read_unlock(&tasklist_lock);
5605 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5609 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5611 goto out_free_cpus_allowed;
5614 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5617 retval = security_task_setscheduler(p, 0, NULL);
5621 cpuset_cpus_allowed(p, cpus_allowed);
5622 cpumask_and(new_mask, in_mask, cpus_allowed);
5624 retval = set_cpus_allowed_ptr(p, new_mask);
5627 cpuset_cpus_allowed(p, cpus_allowed);
5628 if (!cpumask_subset(new_mask, cpus_allowed)) {
5630 * We must have raced with a concurrent cpuset
5631 * update. Just reset the cpus_allowed to the
5632 * cpuset's cpus_allowed
5634 cpumask_copy(new_mask, cpus_allowed);
5639 free_cpumask_var(new_mask);
5640 out_free_cpus_allowed:
5641 free_cpumask_var(cpus_allowed);
5648 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5649 struct cpumask *new_mask)
5651 if (len < cpumask_size())
5652 cpumask_clear(new_mask);
5653 else if (len > cpumask_size())
5654 len = cpumask_size();
5656 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5660 * sys_sched_setaffinity - set the cpu affinity of a process
5661 * @pid: pid of the process
5662 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5663 * @user_mask_ptr: user-space pointer to the new cpu mask
5665 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5666 unsigned long __user *user_mask_ptr)
5668 cpumask_var_t new_mask;
5671 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5674 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5676 retval = sched_setaffinity(pid, new_mask);
5677 free_cpumask_var(new_mask);
5681 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5683 struct task_struct *p;
5687 read_lock(&tasklist_lock);
5690 p = find_process_by_pid(pid);
5694 retval = security_task_getscheduler(p);
5698 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5701 read_unlock(&tasklist_lock);
5708 * sys_sched_getaffinity - get the cpu affinity of a process
5709 * @pid: pid of the process
5710 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5711 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5713 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5714 unsigned long __user *user_mask_ptr)
5719 if (len < cpumask_size())
5722 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5725 ret = sched_getaffinity(pid, mask);
5727 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5730 ret = cpumask_size();
5732 free_cpumask_var(mask);
5738 * sys_sched_yield - yield the current processor to other threads.
5740 * This function yields the current CPU to other tasks. If there are no
5741 * other threads running on this CPU then this function will return.
5743 asmlinkage long sys_sched_yield(void)
5745 struct rq *rq = this_rq_lock();
5747 schedstat_inc(rq, yld_count);
5748 current->sched_class->yield_task(rq);
5751 * Since we are going to call schedule() anyway, there's
5752 * no need to preempt or enable interrupts:
5754 __release(rq->lock);
5755 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5756 _raw_spin_unlock(&rq->lock);
5757 preempt_enable_no_resched();
5764 static void __cond_resched(void)
5766 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5767 __might_sleep(__FILE__, __LINE__);
5770 * The BKS might be reacquired before we have dropped
5771 * PREEMPT_ACTIVE, which could trigger a second
5772 * cond_resched() call.
5775 add_preempt_count(PREEMPT_ACTIVE);
5777 sub_preempt_count(PREEMPT_ACTIVE);
5778 } while (need_resched());
5781 int __sched _cond_resched(void)
5783 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5784 system_state == SYSTEM_RUNNING) {
5790 EXPORT_SYMBOL(_cond_resched);
5793 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5794 * call schedule, and on return reacquire the lock.
5796 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5797 * operations here to prevent schedule() from being called twice (once via
5798 * spin_unlock(), once by hand).
5800 int cond_resched_lock(spinlock_t *lock)
5802 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5805 if (spin_needbreak(lock) || resched) {
5807 if (resched && need_resched())
5816 EXPORT_SYMBOL(cond_resched_lock);
5818 int __sched cond_resched_softirq(void)
5820 BUG_ON(!in_softirq());
5822 if (need_resched() && system_state == SYSTEM_RUNNING) {
5830 EXPORT_SYMBOL(cond_resched_softirq);
5833 * yield - yield the current processor to other threads.
5835 * This is a shortcut for kernel-space yielding - it marks the
5836 * thread runnable and calls sys_sched_yield().
5838 void __sched yield(void)
5840 set_current_state(TASK_RUNNING);
5843 EXPORT_SYMBOL(yield);
5846 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5847 * that process accounting knows that this is a task in IO wait state.
5849 * But don't do that if it is a deliberate, throttling IO wait (this task
5850 * has set its backing_dev_info: the queue against which it should throttle)
5852 void __sched io_schedule(void)
5854 struct rq *rq = &__raw_get_cpu_var(runqueues);
5856 delayacct_blkio_start();
5857 atomic_inc(&rq->nr_iowait);
5859 atomic_dec(&rq->nr_iowait);
5860 delayacct_blkio_end();
5862 EXPORT_SYMBOL(io_schedule);
5864 long __sched io_schedule_timeout(long timeout)
5866 struct rq *rq = &__raw_get_cpu_var(runqueues);
5869 delayacct_blkio_start();
5870 atomic_inc(&rq->nr_iowait);
5871 ret = schedule_timeout(timeout);
5872 atomic_dec(&rq->nr_iowait);
5873 delayacct_blkio_end();
5878 * sys_sched_get_priority_max - return maximum RT priority.
5879 * @policy: scheduling class.
5881 * this syscall returns the maximum rt_priority that can be used
5882 * by a given scheduling class.
5884 asmlinkage long sys_sched_get_priority_max(int policy)
5891 ret = MAX_USER_RT_PRIO-1;
5903 * sys_sched_get_priority_min - return minimum RT priority.
5904 * @policy: scheduling class.
5906 * this syscall returns the minimum rt_priority that can be used
5907 * by a given scheduling class.
5909 asmlinkage long sys_sched_get_priority_min(int policy)
5927 * sys_sched_rr_get_interval - return the default timeslice of a process.
5928 * @pid: pid of the process.
5929 * @interval: userspace pointer to the timeslice value.
5931 * this syscall writes the default timeslice value of a given process
5932 * into the user-space timespec buffer. A value of '0' means infinity.
5935 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5937 struct task_struct *p;
5938 unsigned int time_slice;
5946 read_lock(&tasklist_lock);
5947 p = find_process_by_pid(pid);
5951 retval = security_task_getscheduler(p);
5956 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5957 * tasks that are on an otherwise idle runqueue:
5960 if (p->policy == SCHED_RR) {
5961 time_slice = DEF_TIMESLICE;
5962 } else if (p->policy != SCHED_FIFO) {
5963 struct sched_entity *se = &p->se;
5964 unsigned long flags;
5967 rq = task_rq_lock(p, &flags);
5968 if (rq->cfs.load.weight)
5969 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5970 task_rq_unlock(rq, &flags);
5972 read_unlock(&tasklist_lock);
5973 jiffies_to_timespec(time_slice, &t);
5974 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5978 read_unlock(&tasklist_lock);
5982 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5984 void sched_show_task(struct task_struct *p)
5986 unsigned long free = 0;
5989 state = p->state ? __ffs(p->state) + 1 : 0;
5990 printk(KERN_INFO "%-13.13s %c", p->comm,
5991 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5992 #if BITS_PER_LONG == 32
5993 if (state == TASK_RUNNING)
5994 printk(KERN_CONT " running ");
5996 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5998 if (state == TASK_RUNNING)
5999 printk(KERN_CONT " running task ");
6001 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6003 #ifdef CONFIG_DEBUG_STACK_USAGE
6005 unsigned long *n = end_of_stack(p);
6008 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6011 printk(KERN_CONT "%5lu %5d %6d\n", free,
6012 task_pid_nr(p), task_pid_nr(p->real_parent));
6014 show_stack(p, NULL);
6017 void show_state_filter(unsigned long state_filter)
6019 struct task_struct *g, *p;
6021 #if BITS_PER_LONG == 32
6023 " task PC stack pid father\n");
6026 " task PC stack pid father\n");
6028 read_lock(&tasklist_lock);
6029 do_each_thread(g, p) {
6031 * reset the NMI-timeout, listing all files on a slow
6032 * console might take alot of time:
6034 touch_nmi_watchdog();
6035 if (!state_filter || (p->state & state_filter))
6037 } while_each_thread(g, p);
6039 touch_all_softlockup_watchdogs();
6041 #ifdef CONFIG_SCHED_DEBUG
6042 sysrq_sched_debug_show();
6044 read_unlock(&tasklist_lock);
6046 * Only show locks if all tasks are dumped:
6048 if (state_filter == -1)
6049 debug_show_all_locks();
6052 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6054 idle->sched_class = &idle_sched_class;
6058 * init_idle - set up an idle thread for a given CPU
6059 * @idle: task in question
6060 * @cpu: cpu the idle task belongs to
6062 * NOTE: this function does not set the idle thread's NEED_RESCHED
6063 * flag, to make booting more robust.
6065 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6067 struct rq *rq = cpu_rq(cpu);
6068 unsigned long flags;
6070 spin_lock_irqsave(&rq->lock, flags);
6073 idle->se.exec_start = sched_clock();
6075 idle->prio = idle->normal_prio = MAX_PRIO;
6076 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6077 __set_task_cpu(idle, cpu);
6079 rq->curr = rq->idle = idle;
6080 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6083 spin_unlock_irqrestore(&rq->lock, flags);
6085 /* Set the preempt count _outside_ the spinlocks! */
6086 #if defined(CONFIG_PREEMPT)
6087 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6089 task_thread_info(idle)->preempt_count = 0;
6092 * The idle tasks have their own, simple scheduling class:
6094 idle->sched_class = &idle_sched_class;
6095 ftrace_graph_init_task(idle);
6099 * In a system that switches off the HZ timer nohz_cpu_mask
6100 * indicates which cpus entered this state. This is used
6101 * in the rcu update to wait only for active cpus. For system
6102 * which do not switch off the HZ timer nohz_cpu_mask should
6103 * always be CPU_BITS_NONE.
6105 cpumask_var_t nohz_cpu_mask;
6108 * Increase the granularity value when there are more CPUs,
6109 * because with more CPUs the 'effective latency' as visible
6110 * to users decreases. But the relationship is not linear,
6111 * so pick a second-best guess by going with the log2 of the
6114 * This idea comes from the SD scheduler of Con Kolivas:
6116 static inline void sched_init_granularity(void)
6118 unsigned int factor = 1 + ilog2(num_online_cpus());
6119 const unsigned long limit = 200000000;
6121 sysctl_sched_min_granularity *= factor;
6122 if (sysctl_sched_min_granularity > limit)
6123 sysctl_sched_min_granularity = limit;
6125 sysctl_sched_latency *= factor;
6126 if (sysctl_sched_latency > limit)
6127 sysctl_sched_latency = limit;
6129 sysctl_sched_wakeup_granularity *= factor;
6131 sysctl_sched_shares_ratelimit *= factor;
6136 * This is how migration works:
6138 * 1) we queue a struct migration_req structure in the source CPU's
6139 * runqueue and wake up that CPU's migration thread.
6140 * 2) we down() the locked semaphore => thread blocks.
6141 * 3) migration thread wakes up (implicitly it forces the migrated
6142 * thread off the CPU)
6143 * 4) it gets the migration request and checks whether the migrated
6144 * task is still in the wrong runqueue.
6145 * 5) if it's in the wrong runqueue then the migration thread removes
6146 * it and puts it into the right queue.
6147 * 6) migration thread up()s the semaphore.
6148 * 7) we wake up and the migration is done.
6152 * Change a given task's CPU affinity. Migrate the thread to a
6153 * proper CPU and schedule it away if the CPU it's executing on
6154 * is removed from the allowed bitmask.
6156 * NOTE: the caller must have a valid reference to the task, the
6157 * task must not exit() & deallocate itself prematurely. The
6158 * call is not atomic; no spinlocks may be held.
6160 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6162 struct migration_req req;
6163 unsigned long flags;
6167 rq = task_rq_lock(p, &flags);
6168 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6173 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6174 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6179 if (p->sched_class->set_cpus_allowed)
6180 p->sched_class->set_cpus_allowed(p, new_mask);
6182 cpumask_copy(&p->cpus_allowed, new_mask);
6183 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6186 /* Can the task run on the task's current CPU? If so, we're done */
6187 if (cpumask_test_cpu(task_cpu(p), new_mask))
6190 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6191 /* Need help from migration thread: drop lock and wait. */
6192 task_rq_unlock(rq, &flags);
6193 wake_up_process(rq->migration_thread);
6194 wait_for_completion(&req.done);
6195 tlb_migrate_finish(p->mm);
6199 task_rq_unlock(rq, &flags);
6203 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6206 * Move (not current) task off this cpu, onto dest cpu. We're doing
6207 * this because either it can't run here any more (set_cpus_allowed()
6208 * away from this CPU, or CPU going down), or because we're
6209 * attempting to rebalance this task on exec (sched_exec).
6211 * So we race with normal scheduler movements, but that's OK, as long
6212 * as the task is no longer on this CPU.
6214 * Returns non-zero if task was successfully migrated.
6216 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6218 struct rq *rq_dest, *rq_src;
6221 if (unlikely(!cpu_active(dest_cpu)))
6224 rq_src = cpu_rq(src_cpu);
6225 rq_dest = cpu_rq(dest_cpu);
6227 double_rq_lock(rq_src, rq_dest);
6228 /* Already moved. */
6229 if (task_cpu(p) != src_cpu)
6231 /* Affinity changed (again). */
6232 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6235 on_rq = p->se.on_rq;
6237 deactivate_task(rq_src, p, 0);
6239 set_task_cpu(p, dest_cpu);
6241 activate_task(rq_dest, p, 0);
6242 check_preempt_curr(rq_dest, p, 0);
6247 double_rq_unlock(rq_src, rq_dest);
6252 * migration_thread - this is a highprio system thread that performs
6253 * thread migration by bumping thread off CPU then 'pushing' onto
6256 static int migration_thread(void *data)
6258 int cpu = (long)data;
6262 BUG_ON(rq->migration_thread != current);
6264 set_current_state(TASK_INTERRUPTIBLE);
6265 while (!kthread_should_stop()) {
6266 struct migration_req *req;
6267 struct list_head *head;
6269 spin_lock_irq(&rq->lock);
6271 if (cpu_is_offline(cpu)) {
6272 spin_unlock_irq(&rq->lock);
6276 if (rq->active_balance) {
6277 active_load_balance(rq, cpu);
6278 rq->active_balance = 0;
6281 head = &rq->migration_queue;
6283 if (list_empty(head)) {
6284 spin_unlock_irq(&rq->lock);
6286 set_current_state(TASK_INTERRUPTIBLE);
6289 req = list_entry(head->next, struct migration_req, list);
6290 list_del_init(head->next);
6292 spin_unlock(&rq->lock);
6293 __migrate_task(req->task, cpu, req->dest_cpu);
6296 complete(&req->done);
6298 __set_current_state(TASK_RUNNING);
6302 /* Wait for kthread_stop */
6303 set_current_state(TASK_INTERRUPTIBLE);
6304 while (!kthread_should_stop()) {
6306 set_current_state(TASK_INTERRUPTIBLE);
6308 __set_current_state(TASK_RUNNING);
6312 #ifdef CONFIG_HOTPLUG_CPU
6314 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6318 local_irq_disable();
6319 ret = __migrate_task(p, src_cpu, dest_cpu);
6325 * Figure out where task on dead CPU should go, use force if necessary.
6327 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6330 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6333 /* Look for allowed, online CPU in same node. */
6334 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6335 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6338 /* Any allowed, online CPU? */
6339 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6340 if (dest_cpu < nr_cpu_ids)
6343 /* No more Mr. Nice Guy. */
6344 if (dest_cpu >= nr_cpu_ids) {
6345 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6346 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6349 * Don't tell them about moving exiting tasks or
6350 * kernel threads (both mm NULL), since they never
6353 if (p->mm && printk_ratelimit()) {
6354 printk(KERN_INFO "process %d (%s) no "
6355 "longer affine to cpu%d\n",
6356 task_pid_nr(p), p->comm, dead_cpu);
6361 /* It can have affinity changed while we were choosing. */
6362 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6367 * While a dead CPU has no uninterruptible tasks queued at this point,
6368 * it might still have a nonzero ->nr_uninterruptible counter, because
6369 * for performance reasons the counter is not stricly tracking tasks to
6370 * their home CPUs. So we just add the counter to another CPU's counter,
6371 * to keep the global sum constant after CPU-down:
6373 static void migrate_nr_uninterruptible(struct rq *rq_src)
6375 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6376 unsigned long flags;
6378 local_irq_save(flags);
6379 double_rq_lock(rq_src, rq_dest);
6380 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6381 rq_src->nr_uninterruptible = 0;
6382 double_rq_unlock(rq_src, rq_dest);
6383 local_irq_restore(flags);
6386 /* Run through task list and migrate tasks from the dead cpu. */
6387 static void migrate_live_tasks(int src_cpu)
6389 struct task_struct *p, *t;
6391 read_lock(&tasklist_lock);
6393 do_each_thread(t, p) {
6397 if (task_cpu(p) == src_cpu)
6398 move_task_off_dead_cpu(src_cpu, p);
6399 } while_each_thread(t, p);
6401 read_unlock(&tasklist_lock);
6405 * Schedules idle task to be the next runnable task on current CPU.
6406 * It does so by boosting its priority to highest possible.
6407 * Used by CPU offline code.
6409 void sched_idle_next(void)
6411 int this_cpu = smp_processor_id();
6412 struct rq *rq = cpu_rq(this_cpu);
6413 struct task_struct *p = rq->idle;
6414 unsigned long flags;
6416 /* cpu has to be offline */
6417 BUG_ON(cpu_online(this_cpu));
6420 * Strictly not necessary since rest of the CPUs are stopped by now
6421 * and interrupts disabled on the current cpu.
6423 spin_lock_irqsave(&rq->lock, flags);
6425 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6427 update_rq_clock(rq);
6428 activate_task(rq, p, 0);
6430 spin_unlock_irqrestore(&rq->lock, flags);
6434 * Ensures that the idle task is using init_mm right before its cpu goes
6437 void idle_task_exit(void)
6439 struct mm_struct *mm = current->active_mm;
6441 BUG_ON(cpu_online(smp_processor_id()));
6444 switch_mm(mm, &init_mm, current);
6448 /* called under rq->lock with disabled interrupts */
6449 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6451 struct rq *rq = cpu_rq(dead_cpu);
6453 /* Must be exiting, otherwise would be on tasklist. */
6454 BUG_ON(!p->exit_state);
6456 /* Cannot have done final schedule yet: would have vanished. */
6457 BUG_ON(p->state == TASK_DEAD);
6462 * Drop lock around migration; if someone else moves it,
6463 * that's OK. No task can be added to this CPU, so iteration is
6466 spin_unlock_irq(&rq->lock);
6467 move_task_off_dead_cpu(dead_cpu, p);
6468 spin_lock_irq(&rq->lock);
6473 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6474 static void migrate_dead_tasks(unsigned int dead_cpu)
6476 struct rq *rq = cpu_rq(dead_cpu);
6477 struct task_struct *next;
6480 if (!rq->nr_running)
6482 update_rq_clock(rq);
6483 next = pick_next_task(rq, rq->curr);
6486 next->sched_class->put_prev_task(rq, next);
6487 migrate_dead(dead_cpu, next);
6491 #endif /* CONFIG_HOTPLUG_CPU */
6493 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6495 static struct ctl_table sd_ctl_dir[] = {
6497 .procname = "sched_domain",
6503 static struct ctl_table sd_ctl_root[] = {
6505 .ctl_name = CTL_KERN,
6506 .procname = "kernel",
6508 .child = sd_ctl_dir,
6513 static struct ctl_table *sd_alloc_ctl_entry(int n)
6515 struct ctl_table *entry =
6516 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6521 static void sd_free_ctl_entry(struct ctl_table **tablep)
6523 struct ctl_table *entry;
6526 * In the intermediate directories, both the child directory and
6527 * procname are dynamically allocated and could fail but the mode
6528 * will always be set. In the lowest directory the names are
6529 * static strings and all have proc handlers.
6531 for (entry = *tablep; entry->mode; entry++) {
6533 sd_free_ctl_entry(&entry->child);
6534 if (entry->proc_handler == NULL)
6535 kfree(entry->procname);
6543 set_table_entry(struct ctl_table *entry,
6544 const char *procname, void *data, int maxlen,
6545 mode_t mode, proc_handler *proc_handler)
6547 entry->procname = procname;
6549 entry->maxlen = maxlen;
6551 entry->proc_handler = proc_handler;
6554 static struct ctl_table *
6555 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6557 struct ctl_table *table = sd_alloc_ctl_entry(13);
6562 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6563 sizeof(long), 0644, proc_doulongvec_minmax);
6564 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6565 sizeof(long), 0644, proc_doulongvec_minmax);
6566 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6567 sizeof(int), 0644, proc_dointvec_minmax);
6568 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6569 sizeof(int), 0644, proc_dointvec_minmax);
6570 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6571 sizeof(int), 0644, proc_dointvec_minmax);
6572 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6573 sizeof(int), 0644, proc_dointvec_minmax);
6574 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6575 sizeof(int), 0644, proc_dointvec_minmax);
6576 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6577 sizeof(int), 0644, proc_dointvec_minmax);
6578 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6579 sizeof(int), 0644, proc_dointvec_minmax);
6580 set_table_entry(&table[9], "cache_nice_tries",
6581 &sd->cache_nice_tries,
6582 sizeof(int), 0644, proc_dointvec_minmax);
6583 set_table_entry(&table[10], "flags", &sd->flags,
6584 sizeof(int), 0644, proc_dointvec_minmax);
6585 set_table_entry(&table[11], "name", sd->name,
6586 CORENAME_MAX_SIZE, 0444, proc_dostring);
6587 /* &table[12] is terminator */
6592 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6594 struct ctl_table *entry, *table;
6595 struct sched_domain *sd;
6596 int domain_num = 0, i;
6599 for_each_domain(cpu, sd)
6601 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6606 for_each_domain(cpu, sd) {
6607 snprintf(buf, 32, "domain%d", i);
6608 entry->procname = kstrdup(buf, GFP_KERNEL);
6610 entry->child = sd_alloc_ctl_domain_table(sd);
6617 static struct ctl_table_header *sd_sysctl_header;
6618 static void register_sched_domain_sysctl(void)
6620 int i, cpu_num = num_online_cpus();
6621 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6624 WARN_ON(sd_ctl_dir[0].child);
6625 sd_ctl_dir[0].child = entry;
6630 for_each_online_cpu(i) {
6631 snprintf(buf, 32, "cpu%d", i);
6632 entry->procname = kstrdup(buf, GFP_KERNEL);
6634 entry->child = sd_alloc_ctl_cpu_table(i);
6638 WARN_ON(sd_sysctl_header);
6639 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6642 /* may be called multiple times per register */
6643 static void unregister_sched_domain_sysctl(void)
6645 if (sd_sysctl_header)
6646 unregister_sysctl_table(sd_sysctl_header);
6647 sd_sysctl_header = NULL;
6648 if (sd_ctl_dir[0].child)
6649 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6652 static void register_sched_domain_sysctl(void)
6655 static void unregister_sched_domain_sysctl(void)
6660 static void set_rq_online(struct rq *rq)
6663 const struct sched_class *class;
6665 cpumask_set_cpu(rq->cpu, rq->rd->online);
6668 for_each_class(class) {
6669 if (class->rq_online)
6670 class->rq_online(rq);
6675 static void set_rq_offline(struct rq *rq)
6678 const struct sched_class *class;
6680 for_each_class(class) {
6681 if (class->rq_offline)
6682 class->rq_offline(rq);
6685 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6691 * migration_call - callback that gets triggered when a CPU is added.
6692 * Here we can start up the necessary migration thread for the new CPU.
6694 static int __cpuinit
6695 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6697 struct task_struct *p;
6698 int cpu = (long)hcpu;
6699 unsigned long flags;
6704 case CPU_UP_PREPARE:
6705 case CPU_UP_PREPARE_FROZEN:
6706 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6709 kthread_bind(p, cpu);
6710 /* Must be high prio: stop_machine expects to yield to it. */
6711 rq = task_rq_lock(p, &flags);
6712 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6713 task_rq_unlock(rq, &flags);
6714 cpu_rq(cpu)->migration_thread = p;
6718 case CPU_ONLINE_FROZEN:
6719 /* Strictly unnecessary, as first user will wake it. */
6720 wake_up_process(cpu_rq(cpu)->migration_thread);
6722 /* Update our root-domain */
6724 spin_lock_irqsave(&rq->lock, flags);
6726 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6730 spin_unlock_irqrestore(&rq->lock, flags);
6733 #ifdef CONFIG_HOTPLUG_CPU
6734 case CPU_UP_CANCELED:
6735 case CPU_UP_CANCELED_FROZEN:
6736 if (!cpu_rq(cpu)->migration_thread)
6738 /* Unbind it from offline cpu so it can run. Fall thru. */
6739 kthread_bind(cpu_rq(cpu)->migration_thread,
6740 cpumask_any(cpu_online_mask));
6741 kthread_stop(cpu_rq(cpu)->migration_thread);
6742 cpu_rq(cpu)->migration_thread = NULL;
6746 case CPU_DEAD_FROZEN:
6747 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6748 migrate_live_tasks(cpu);
6750 kthread_stop(rq->migration_thread);
6751 rq->migration_thread = NULL;
6752 /* Idle task back to normal (off runqueue, low prio) */
6753 spin_lock_irq(&rq->lock);
6754 update_rq_clock(rq);
6755 deactivate_task(rq, rq->idle, 0);
6756 rq->idle->static_prio = MAX_PRIO;
6757 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6758 rq->idle->sched_class = &idle_sched_class;
6759 migrate_dead_tasks(cpu);
6760 spin_unlock_irq(&rq->lock);
6762 migrate_nr_uninterruptible(rq);
6763 BUG_ON(rq->nr_running != 0);
6766 * No need to migrate the tasks: it was best-effort if
6767 * they didn't take sched_hotcpu_mutex. Just wake up
6770 spin_lock_irq(&rq->lock);
6771 while (!list_empty(&rq->migration_queue)) {
6772 struct migration_req *req;
6774 req = list_entry(rq->migration_queue.next,
6775 struct migration_req, list);
6776 list_del_init(&req->list);
6777 spin_unlock_irq(&rq->lock);
6778 complete(&req->done);
6779 spin_lock_irq(&rq->lock);
6781 spin_unlock_irq(&rq->lock);
6785 case CPU_DYING_FROZEN:
6786 /* Update our root-domain */
6788 spin_lock_irqsave(&rq->lock, flags);
6790 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6793 spin_unlock_irqrestore(&rq->lock, flags);
6800 /* Register at highest priority so that task migration (migrate_all_tasks)
6801 * happens before everything else.
6803 static struct notifier_block __cpuinitdata migration_notifier = {
6804 .notifier_call = migration_call,
6808 static int __init migration_init(void)
6810 void *cpu = (void *)(long)smp_processor_id();
6813 /* Start one for the boot CPU: */
6814 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6815 BUG_ON(err == NOTIFY_BAD);
6816 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6817 register_cpu_notifier(&migration_notifier);
6821 early_initcall(migration_init);
6826 #ifdef CONFIG_SCHED_DEBUG
6828 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6829 struct cpumask *groupmask)
6831 struct sched_group *group = sd->groups;
6834 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6835 cpumask_clear(groupmask);
6837 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6839 if (!(sd->flags & SD_LOAD_BALANCE)) {
6840 printk("does not load-balance\n");
6842 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6847 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6849 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6850 printk(KERN_ERR "ERROR: domain->span does not contain "
6853 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6854 printk(KERN_ERR "ERROR: domain->groups does not contain"
6858 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6862 printk(KERN_ERR "ERROR: group is NULL\n");
6866 if (!group->__cpu_power) {
6867 printk(KERN_CONT "\n");
6868 printk(KERN_ERR "ERROR: domain->cpu_power not "
6873 if (!cpumask_weight(sched_group_cpus(group))) {
6874 printk(KERN_CONT "\n");
6875 printk(KERN_ERR "ERROR: empty group\n");
6879 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6880 printk(KERN_CONT "\n");
6881 printk(KERN_ERR "ERROR: repeated CPUs\n");
6885 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6887 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6888 printk(KERN_CONT " %s", str);
6890 group = group->next;
6891 } while (group != sd->groups);
6892 printk(KERN_CONT "\n");
6894 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6895 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6898 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6899 printk(KERN_ERR "ERROR: parent span is not a superset "
6900 "of domain->span\n");
6904 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6906 cpumask_var_t groupmask;
6910 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6914 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6916 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6917 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6922 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6929 free_cpumask_var(groupmask);
6931 #else /* !CONFIG_SCHED_DEBUG */
6932 # define sched_domain_debug(sd, cpu) do { } while (0)
6933 #endif /* CONFIG_SCHED_DEBUG */
6935 static int sd_degenerate(struct sched_domain *sd)
6937 if (cpumask_weight(sched_domain_span(sd)) == 1)
6940 /* Following flags need at least 2 groups */
6941 if (sd->flags & (SD_LOAD_BALANCE |
6942 SD_BALANCE_NEWIDLE |
6946 SD_SHARE_PKG_RESOURCES)) {
6947 if (sd->groups != sd->groups->next)
6951 /* Following flags don't use groups */
6952 if (sd->flags & (SD_WAKE_IDLE |
6961 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6963 unsigned long cflags = sd->flags, pflags = parent->flags;
6965 if (sd_degenerate(parent))
6968 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
6971 /* Does parent contain flags not in child? */
6972 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6973 if (cflags & SD_WAKE_AFFINE)
6974 pflags &= ~SD_WAKE_BALANCE;
6975 /* Flags needing groups don't count if only 1 group in parent */
6976 if (parent->groups == parent->groups->next) {
6977 pflags &= ~(SD_LOAD_BALANCE |
6978 SD_BALANCE_NEWIDLE |
6982 SD_SHARE_PKG_RESOURCES);
6983 if (nr_node_ids == 1)
6984 pflags &= ~SD_SERIALIZE;
6986 if (~cflags & pflags)
6992 static void free_rootdomain(struct root_domain *rd)
6994 cpupri_cleanup(&rd->cpupri);
6996 free_cpumask_var(rd->rto_mask);
6997 free_cpumask_var(rd->online);
6998 free_cpumask_var(rd->span);
7002 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7004 unsigned long flags;
7006 spin_lock_irqsave(&rq->lock, flags);
7009 struct root_domain *old_rd = rq->rd;
7011 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7014 cpumask_clear_cpu(rq->cpu, old_rd->span);
7016 if (atomic_dec_and_test(&old_rd->refcount))
7017 free_rootdomain(old_rd);
7020 atomic_inc(&rd->refcount);
7023 cpumask_set_cpu(rq->cpu, rd->span);
7024 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7027 spin_unlock_irqrestore(&rq->lock, flags);
7030 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7032 memset(rd, 0, sizeof(*rd));
7035 alloc_bootmem_cpumask_var(&def_root_domain.span);
7036 alloc_bootmem_cpumask_var(&def_root_domain.online);
7037 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7038 cpupri_init(&rd->cpupri, true);
7042 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7044 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7046 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7049 if (cpupri_init(&rd->cpupri, false) != 0)
7054 free_cpumask_var(rd->rto_mask);
7056 free_cpumask_var(rd->online);
7058 free_cpumask_var(rd->span);
7063 static void init_defrootdomain(void)
7065 init_rootdomain(&def_root_domain, true);
7067 atomic_set(&def_root_domain.refcount, 1);
7070 static struct root_domain *alloc_rootdomain(void)
7072 struct root_domain *rd;
7074 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7078 if (init_rootdomain(rd, false) != 0) {
7087 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7088 * hold the hotplug lock.
7091 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7093 struct rq *rq = cpu_rq(cpu);
7094 struct sched_domain *tmp;
7096 /* Remove the sched domains which do not contribute to scheduling. */
7097 for (tmp = sd; tmp; ) {
7098 struct sched_domain *parent = tmp->parent;
7102 if (sd_parent_degenerate(tmp, parent)) {
7103 tmp->parent = parent->parent;
7105 parent->parent->child = tmp;
7110 if (sd && sd_degenerate(sd)) {
7116 sched_domain_debug(sd, cpu);
7118 rq_attach_root(rq, rd);
7119 rcu_assign_pointer(rq->sd, sd);
7122 /* cpus with isolated domains */
7123 static cpumask_var_t cpu_isolated_map;
7125 /* Setup the mask of cpus configured for isolated domains */
7126 static int __init isolated_cpu_setup(char *str)
7128 cpulist_parse(str, cpu_isolated_map);
7132 __setup("isolcpus=", isolated_cpu_setup);
7135 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7136 * to a function which identifies what group(along with sched group) a CPU
7137 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7138 * (due to the fact that we keep track of groups covered with a struct cpumask).
7140 * init_sched_build_groups will build a circular linked list of the groups
7141 * covered by the given span, and will set each group's ->cpumask correctly,
7142 * and ->cpu_power to 0.
7145 init_sched_build_groups(const struct cpumask *span,
7146 const struct cpumask *cpu_map,
7147 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7148 struct sched_group **sg,
7149 struct cpumask *tmpmask),
7150 struct cpumask *covered, struct cpumask *tmpmask)
7152 struct sched_group *first = NULL, *last = NULL;
7155 cpumask_clear(covered);
7157 for_each_cpu(i, span) {
7158 struct sched_group *sg;
7159 int group = group_fn(i, cpu_map, &sg, tmpmask);
7162 if (cpumask_test_cpu(i, covered))
7165 cpumask_clear(sched_group_cpus(sg));
7166 sg->__cpu_power = 0;
7168 for_each_cpu(j, span) {
7169 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7172 cpumask_set_cpu(j, covered);
7173 cpumask_set_cpu(j, sched_group_cpus(sg));
7184 #define SD_NODES_PER_DOMAIN 16
7189 * find_next_best_node - find the next node to include in a sched_domain
7190 * @node: node whose sched_domain we're building
7191 * @used_nodes: nodes already in the sched_domain
7193 * Find the next node to include in a given scheduling domain. Simply
7194 * finds the closest node not already in the @used_nodes map.
7196 * Should use nodemask_t.
7198 static int find_next_best_node(int node, nodemask_t *used_nodes)
7200 int i, n, val, min_val, best_node = 0;
7204 for (i = 0; i < nr_node_ids; i++) {
7205 /* Start at @node */
7206 n = (node + i) % nr_node_ids;
7208 if (!nr_cpus_node(n))
7211 /* Skip already used nodes */
7212 if (node_isset(n, *used_nodes))
7215 /* Simple min distance search */
7216 val = node_distance(node, n);
7218 if (val < min_val) {
7224 node_set(best_node, *used_nodes);
7229 * sched_domain_node_span - get a cpumask for a node's sched_domain
7230 * @node: node whose cpumask we're constructing
7231 * @span: resulting cpumask
7233 * Given a node, construct a good cpumask for its sched_domain to span. It
7234 * should be one that prevents unnecessary balancing, but also spreads tasks
7237 static void sched_domain_node_span(int node, struct cpumask *span)
7239 nodemask_t used_nodes;
7242 cpumask_clear(span);
7243 nodes_clear(used_nodes);
7245 cpumask_or(span, span, cpumask_of_node(node));
7246 node_set(node, used_nodes);
7248 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7249 int next_node = find_next_best_node(node, &used_nodes);
7251 cpumask_or(span, span, cpumask_of_node(next_node));
7254 #endif /* CONFIG_NUMA */
7256 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7259 * The cpus mask in sched_group and sched_domain hangs off the end.
7260 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7261 * for nr_cpu_ids < CONFIG_NR_CPUS.
7263 struct static_sched_group {
7264 struct sched_group sg;
7265 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7268 struct static_sched_domain {
7269 struct sched_domain sd;
7270 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7274 * SMT sched-domains:
7276 #ifdef CONFIG_SCHED_SMT
7277 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7278 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7281 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7282 struct sched_group **sg, struct cpumask *unused)
7285 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7288 #endif /* CONFIG_SCHED_SMT */
7291 * multi-core sched-domains:
7293 #ifdef CONFIG_SCHED_MC
7294 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7295 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7296 #endif /* CONFIG_SCHED_MC */
7298 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7300 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7301 struct sched_group **sg, struct cpumask *mask)
7305 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7306 group = cpumask_first(mask);
7308 *sg = &per_cpu(sched_group_core, group).sg;
7311 #elif defined(CONFIG_SCHED_MC)
7313 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7314 struct sched_group **sg, struct cpumask *unused)
7317 *sg = &per_cpu(sched_group_core, cpu).sg;
7322 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7323 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7326 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7327 struct sched_group **sg, struct cpumask *mask)
7330 #ifdef CONFIG_SCHED_MC
7331 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7332 group = cpumask_first(mask);
7333 #elif defined(CONFIG_SCHED_SMT)
7334 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7335 group = cpumask_first(mask);
7340 *sg = &per_cpu(sched_group_phys, group).sg;
7346 * The init_sched_build_groups can't handle what we want to do with node
7347 * groups, so roll our own. Now each node has its own list of groups which
7348 * gets dynamically allocated.
7350 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7351 static struct sched_group ***sched_group_nodes_bycpu;
7353 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7354 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7356 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7357 struct sched_group **sg,
7358 struct cpumask *nodemask)
7362 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7363 group = cpumask_first(nodemask);
7366 *sg = &per_cpu(sched_group_allnodes, group).sg;
7370 static void init_numa_sched_groups_power(struct sched_group *group_head)
7372 struct sched_group *sg = group_head;
7378 for_each_cpu(j, sched_group_cpus(sg)) {
7379 struct sched_domain *sd;
7381 sd = &per_cpu(phys_domains, j).sd;
7382 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7384 * Only add "power" once for each
7390 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7393 } while (sg != group_head);
7395 #endif /* CONFIG_NUMA */
7398 /* Free memory allocated for various sched_group structures */
7399 static void free_sched_groups(const struct cpumask *cpu_map,
7400 struct cpumask *nodemask)
7404 for_each_cpu(cpu, cpu_map) {
7405 struct sched_group **sched_group_nodes
7406 = sched_group_nodes_bycpu[cpu];
7408 if (!sched_group_nodes)
7411 for (i = 0; i < nr_node_ids; i++) {
7412 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7414 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7415 if (cpumask_empty(nodemask))
7425 if (oldsg != sched_group_nodes[i])
7428 kfree(sched_group_nodes);
7429 sched_group_nodes_bycpu[cpu] = NULL;
7432 #else /* !CONFIG_NUMA */
7433 static void free_sched_groups(const struct cpumask *cpu_map,
7434 struct cpumask *nodemask)
7437 #endif /* CONFIG_NUMA */
7440 * Initialize sched groups cpu_power.
7442 * cpu_power indicates the capacity of sched group, which is used while
7443 * distributing the load between different sched groups in a sched domain.
7444 * Typically cpu_power for all the groups in a sched domain will be same unless
7445 * there are asymmetries in the topology. If there are asymmetries, group
7446 * having more cpu_power will pickup more load compared to the group having
7449 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7450 * the maximum number of tasks a group can handle in the presence of other idle
7451 * or lightly loaded groups in the same sched domain.
7453 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7455 struct sched_domain *child;
7456 struct sched_group *group;
7458 WARN_ON(!sd || !sd->groups);
7460 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7465 sd->groups->__cpu_power = 0;
7468 * For perf policy, if the groups in child domain share resources
7469 * (for example cores sharing some portions of the cache hierarchy
7470 * or SMT), then set this domain groups cpu_power such that each group
7471 * can handle only one task, when there are other idle groups in the
7472 * same sched domain.
7474 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7476 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7477 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7482 * add cpu_power of each child group to this groups cpu_power
7484 group = child->groups;
7486 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7487 group = group->next;
7488 } while (group != child->groups);
7492 * Initializers for schedule domains
7493 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7496 #ifdef CONFIG_SCHED_DEBUG
7497 # define SD_INIT_NAME(sd, type) sd->name = #type
7499 # define SD_INIT_NAME(sd, type) do { } while (0)
7502 #define SD_INIT(sd, type) sd_init_##type(sd)
7504 #define SD_INIT_FUNC(type) \
7505 static noinline void sd_init_##type(struct sched_domain *sd) \
7507 memset(sd, 0, sizeof(*sd)); \
7508 *sd = SD_##type##_INIT; \
7509 sd->level = SD_LV_##type; \
7510 SD_INIT_NAME(sd, type); \
7515 SD_INIT_FUNC(ALLNODES)
7518 #ifdef CONFIG_SCHED_SMT
7519 SD_INIT_FUNC(SIBLING)
7521 #ifdef CONFIG_SCHED_MC
7525 static int default_relax_domain_level = -1;
7527 static int __init setup_relax_domain_level(char *str)
7531 val = simple_strtoul(str, NULL, 0);
7532 if (val < SD_LV_MAX)
7533 default_relax_domain_level = val;
7537 __setup("relax_domain_level=", setup_relax_domain_level);
7539 static void set_domain_attribute(struct sched_domain *sd,
7540 struct sched_domain_attr *attr)
7544 if (!attr || attr->relax_domain_level < 0) {
7545 if (default_relax_domain_level < 0)
7548 request = default_relax_domain_level;
7550 request = attr->relax_domain_level;
7551 if (request < sd->level) {
7552 /* turn off idle balance on this domain */
7553 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7555 /* turn on idle balance on this domain */
7556 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7561 * Build sched domains for a given set of cpus and attach the sched domains
7562 * to the individual cpus
7564 static int __build_sched_domains(const struct cpumask *cpu_map,
7565 struct sched_domain_attr *attr)
7567 int i, err = -ENOMEM;
7568 struct root_domain *rd;
7569 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7572 cpumask_var_t domainspan, covered, notcovered;
7573 struct sched_group **sched_group_nodes = NULL;
7574 int sd_allnodes = 0;
7576 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7578 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7579 goto free_domainspan;
7580 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7584 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7585 goto free_notcovered;
7586 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7588 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7589 goto free_this_sibling_map;
7590 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7591 goto free_this_core_map;
7592 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7593 goto free_send_covered;
7597 * Allocate the per-node list of sched groups
7599 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7601 if (!sched_group_nodes) {
7602 printk(KERN_WARNING "Can not alloc sched group node list\n");
7607 rd = alloc_rootdomain();
7609 printk(KERN_WARNING "Cannot alloc root domain\n");
7610 goto free_sched_groups;
7614 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7618 * Set up domains for cpus specified by the cpu_map.
7620 for_each_cpu(i, cpu_map) {
7621 struct sched_domain *sd = NULL, *p;
7623 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7626 if (cpumask_weight(cpu_map) >
7627 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7628 sd = &per_cpu(allnodes_domains, i);
7629 SD_INIT(sd, ALLNODES);
7630 set_domain_attribute(sd, attr);
7631 cpumask_copy(sched_domain_span(sd), cpu_map);
7632 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7638 sd = &per_cpu(node_domains, i);
7640 set_domain_attribute(sd, attr);
7641 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7645 cpumask_and(sched_domain_span(sd),
7646 sched_domain_span(sd), cpu_map);
7650 sd = &per_cpu(phys_domains, i).sd;
7652 set_domain_attribute(sd, attr);
7653 cpumask_copy(sched_domain_span(sd), nodemask);
7657 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7659 #ifdef CONFIG_SCHED_MC
7661 sd = &per_cpu(core_domains, i).sd;
7663 set_domain_attribute(sd, attr);
7664 cpumask_and(sched_domain_span(sd), cpu_map,
7665 cpu_coregroup_mask(i));
7668 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7671 #ifdef CONFIG_SCHED_SMT
7673 sd = &per_cpu(cpu_domains, i).sd;
7674 SD_INIT(sd, SIBLING);
7675 set_domain_attribute(sd, attr);
7676 cpumask_and(sched_domain_span(sd),
7677 &per_cpu(cpu_sibling_map, i), cpu_map);
7680 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7684 #ifdef CONFIG_SCHED_SMT
7685 /* Set up CPU (sibling) groups */
7686 for_each_cpu(i, cpu_map) {
7687 cpumask_and(this_sibling_map,
7688 &per_cpu(cpu_sibling_map, i), cpu_map);
7689 if (i != cpumask_first(this_sibling_map))
7692 init_sched_build_groups(this_sibling_map, cpu_map,
7694 send_covered, tmpmask);
7698 #ifdef CONFIG_SCHED_MC
7699 /* Set up multi-core groups */
7700 for_each_cpu(i, cpu_map) {
7701 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7702 if (i != cpumask_first(this_core_map))
7705 init_sched_build_groups(this_core_map, cpu_map,
7707 send_covered, tmpmask);
7711 /* Set up physical groups */
7712 for (i = 0; i < nr_node_ids; i++) {
7713 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7714 if (cpumask_empty(nodemask))
7717 init_sched_build_groups(nodemask, cpu_map,
7719 send_covered, tmpmask);
7723 /* Set up node groups */
7725 init_sched_build_groups(cpu_map, cpu_map,
7726 &cpu_to_allnodes_group,
7727 send_covered, tmpmask);
7730 for (i = 0; i < nr_node_ids; i++) {
7731 /* Set up node groups */
7732 struct sched_group *sg, *prev;
7735 cpumask_clear(covered);
7736 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7737 if (cpumask_empty(nodemask)) {
7738 sched_group_nodes[i] = NULL;
7742 sched_domain_node_span(i, domainspan);
7743 cpumask_and(domainspan, domainspan, cpu_map);
7745 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7748 printk(KERN_WARNING "Can not alloc domain group for "
7752 sched_group_nodes[i] = sg;
7753 for_each_cpu(j, nodemask) {
7754 struct sched_domain *sd;
7756 sd = &per_cpu(node_domains, j);
7759 sg->__cpu_power = 0;
7760 cpumask_copy(sched_group_cpus(sg), nodemask);
7762 cpumask_or(covered, covered, nodemask);
7765 for (j = 0; j < nr_node_ids; j++) {
7766 int n = (i + j) % nr_node_ids;
7768 cpumask_complement(notcovered, covered);
7769 cpumask_and(tmpmask, notcovered, cpu_map);
7770 cpumask_and(tmpmask, tmpmask, domainspan);
7771 if (cpumask_empty(tmpmask))
7774 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7775 if (cpumask_empty(tmpmask))
7778 sg = kmalloc_node(sizeof(struct sched_group) +
7783 "Can not alloc domain group for node %d\n", j);
7786 sg->__cpu_power = 0;
7787 cpumask_copy(sched_group_cpus(sg), tmpmask);
7788 sg->next = prev->next;
7789 cpumask_or(covered, covered, tmpmask);
7796 /* Calculate CPU power for physical packages and nodes */
7797 #ifdef CONFIG_SCHED_SMT
7798 for_each_cpu(i, cpu_map) {
7799 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7801 init_sched_groups_power(i, sd);
7804 #ifdef CONFIG_SCHED_MC
7805 for_each_cpu(i, cpu_map) {
7806 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7808 init_sched_groups_power(i, sd);
7812 for_each_cpu(i, cpu_map) {
7813 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7815 init_sched_groups_power(i, sd);
7819 for (i = 0; i < nr_node_ids; i++)
7820 init_numa_sched_groups_power(sched_group_nodes[i]);
7823 struct sched_group *sg;
7825 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7827 init_numa_sched_groups_power(sg);
7831 /* Attach the domains */
7832 for_each_cpu(i, cpu_map) {
7833 struct sched_domain *sd;
7834 #ifdef CONFIG_SCHED_SMT
7835 sd = &per_cpu(cpu_domains, i).sd;
7836 #elif defined(CONFIG_SCHED_MC)
7837 sd = &per_cpu(core_domains, i).sd;
7839 sd = &per_cpu(phys_domains, i).sd;
7841 cpu_attach_domain(sd, rd, i);
7847 free_cpumask_var(tmpmask);
7849 free_cpumask_var(send_covered);
7851 free_cpumask_var(this_core_map);
7852 free_this_sibling_map:
7853 free_cpumask_var(this_sibling_map);
7855 free_cpumask_var(nodemask);
7858 free_cpumask_var(notcovered);
7860 free_cpumask_var(covered);
7862 free_cpumask_var(domainspan);
7869 kfree(sched_group_nodes);
7875 free_sched_groups(cpu_map, tmpmask);
7876 free_rootdomain(rd);
7881 static int build_sched_domains(const struct cpumask *cpu_map)
7883 return __build_sched_domains(cpu_map, NULL);
7886 static struct cpumask *doms_cur; /* current sched domains */
7887 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7888 static struct sched_domain_attr *dattr_cur;
7889 /* attribues of custom domains in 'doms_cur' */
7892 * Special case: If a kmalloc of a doms_cur partition (array of
7893 * cpumask) fails, then fallback to a single sched domain,
7894 * as determined by the single cpumask fallback_doms.
7896 static cpumask_var_t fallback_doms;
7899 * arch_update_cpu_topology lets virtualized architectures update the
7900 * cpu core maps. It is supposed to return 1 if the topology changed
7901 * or 0 if it stayed the same.
7903 int __attribute__((weak)) arch_update_cpu_topology(void)
7909 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7910 * For now this just excludes isolated cpus, but could be used to
7911 * exclude other special cases in the future.
7913 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7917 arch_update_cpu_topology();
7919 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
7921 doms_cur = fallback_doms;
7922 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
7924 err = build_sched_domains(doms_cur);
7925 register_sched_domain_sysctl();
7930 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
7931 struct cpumask *tmpmask)
7933 free_sched_groups(cpu_map, tmpmask);
7937 * Detach sched domains from a group of cpus specified in cpu_map
7938 * These cpus will now be attached to the NULL domain
7940 static void detach_destroy_domains(const struct cpumask *cpu_map)
7942 /* Save because hotplug lock held. */
7943 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
7946 for_each_cpu(i, cpu_map)
7947 cpu_attach_domain(NULL, &def_root_domain, i);
7948 synchronize_sched();
7949 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
7952 /* handle null as "default" */
7953 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7954 struct sched_domain_attr *new, int idx_new)
7956 struct sched_domain_attr tmp;
7963 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7964 new ? (new + idx_new) : &tmp,
7965 sizeof(struct sched_domain_attr));
7969 * Partition sched domains as specified by the 'ndoms_new'
7970 * cpumasks in the array doms_new[] of cpumasks. This compares
7971 * doms_new[] to the current sched domain partitioning, doms_cur[].
7972 * It destroys each deleted domain and builds each new domain.
7974 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
7975 * The masks don't intersect (don't overlap.) We should setup one
7976 * sched domain for each mask. CPUs not in any of the cpumasks will
7977 * not be load balanced. If the same cpumask appears both in the
7978 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7981 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7982 * ownership of it and will kfree it when done with it. If the caller
7983 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7984 * ndoms_new == 1, and partition_sched_domains() will fallback to
7985 * the single partition 'fallback_doms', it also forces the domains
7988 * If doms_new == NULL it will be replaced with cpu_online_mask.
7989 * ndoms_new == 0 is a special case for destroying existing domains,
7990 * and it will not create the default domain.
7992 * Call with hotplug lock held
7994 /* FIXME: Change to struct cpumask *doms_new[] */
7995 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
7996 struct sched_domain_attr *dattr_new)
8001 mutex_lock(&sched_domains_mutex);
8003 /* always unregister in case we don't destroy any domains */
8004 unregister_sched_domain_sysctl();
8006 /* Let architecture update cpu core mappings. */
8007 new_topology = arch_update_cpu_topology();
8009 n = doms_new ? ndoms_new : 0;
8011 /* Destroy deleted domains */
8012 for (i = 0; i < ndoms_cur; i++) {
8013 for (j = 0; j < n && !new_topology; j++) {
8014 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8015 && dattrs_equal(dattr_cur, i, dattr_new, j))
8018 /* no match - a current sched domain not in new doms_new[] */
8019 detach_destroy_domains(doms_cur + i);
8024 if (doms_new == NULL) {
8026 doms_new = fallback_doms;
8027 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8028 WARN_ON_ONCE(dattr_new);
8031 /* Build new domains */
8032 for (i = 0; i < ndoms_new; i++) {
8033 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8034 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8035 && dattrs_equal(dattr_new, i, dattr_cur, j))
8038 /* no match - add a new doms_new */
8039 __build_sched_domains(doms_new + i,
8040 dattr_new ? dattr_new + i : NULL);
8045 /* Remember the new sched domains */
8046 if (doms_cur != fallback_doms)
8048 kfree(dattr_cur); /* kfree(NULL) is safe */
8049 doms_cur = doms_new;
8050 dattr_cur = dattr_new;
8051 ndoms_cur = ndoms_new;
8053 register_sched_domain_sysctl();
8055 mutex_unlock(&sched_domains_mutex);
8058 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8059 static void arch_reinit_sched_domains(void)
8063 /* Destroy domains first to force the rebuild */
8064 partition_sched_domains(0, NULL, NULL);
8066 rebuild_sched_domains();
8070 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8072 unsigned int level = 0;
8074 if (sscanf(buf, "%u", &level) != 1)
8078 * level is always be positive so don't check for
8079 * level < POWERSAVINGS_BALANCE_NONE which is 0
8080 * What happens on 0 or 1 byte write,
8081 * need to check for count as well?
8084 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8088 sched_smt_power_savings = level;
8090 sched_mc_power_savings = level;
8092 arch_reinit_sched_domains();
8097 #ifdef CONFIG_SCHED_MC
8098 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8101 return sprintf(page, "%u\n", sched_mc_power_savings);
8103 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8104 const char *buf, size_t count)
8106 return sched_power_savings_store(buf, count, 0);
8108 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8109 sched_mc_power_savings_show,
8110 sched_mc_power_savings_store);
8113 #ifdef CONFIG_SCHED_SMT
8114 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8117 return sprintf(page, "%u\n", sched_smt_power_savings);
8119 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8120 const char *buf, size_t count)
8122 return sched_power_savings_store(buf, count, 1);
8124 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8125 sched_smt_power_savings_show,
8126 sched_smt_power_savings_store);
8129 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8133 #ifdef CONFIG_SCHED_SMT
8135 err = sysfs_create_file(&cls->kset.kobj,
8136 &attr_sched_smt_power_savings.attr);
8138 #ifdef CONFIG_SCHED_MC
8139 if (!err && mc_capable())
8140 err = sysfs_create_file(&cls->kset.kobj,
8141 &attr_sched_mc_power_savings.attr);
8145 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8147 #ifndef CONFIG_CPUSETS
8149 * Add online and remove offline CPUs from the scheduler domains.
8150 * When cpusets are enabled they take over this function.
8152 static int update_sched_domains(struct notifier_block *nfb,
8153 unsigned long action, void *hcpu)
8157 case CPU_ONLINE_FROZEN:
8159 case CPU_DEAD_FROZEN:
8160 partition_sched_domains(1, NULL, NULL);
8169 static int update_runtime(struct notifier_block *nfb,
8170 unsigned long action, void *hcpu)
8172 int cpu = (int)(long)hcpu;
8175 case CPU_DOWN_PREPARE:
8176 case CPU_DOWN_PREPARE_FROZEN:
8177 disable_runtime(cpu_rq(cpu));
8180 case CPU_DOWN_FAILED:
8181 case CPU_DOWN_FAILED_FROZEN:
8183 case CPU_ONLINE_FROZEN:
8184 enable_runtime(cpu_rq(cpu));
8192 void __init sched_init_smp(void)
8194 cpumask_var_t non_isolated_cpus;
8196 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8198 #if defined(CONFIG_NUMA)
8199 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8201 BUG_ON(sched_group_nodes_bycpu == NULL);
8204 mutex_lock(&sched_domains_mutex);
8205 arch_init_sched_domains(cpu_online_mask);
8206 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8207 if (cpumask_empty(non_isolated_cpus))
8208 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8209 mutex_unlock(&sched_domains_mutex);
8212 #ifndef CONFIG_CPUSETS
8213 /* XXX: Theoretical race here - CPU may be hotplugged now */
8214 hotcpu_notifier(update_sched_domains, 0);
8217 /* RT runtime code needs to handle some hotplug events */
8218 hotcpu_notifier(update_runtime, 0);
8222 /* Move init over to a non-isolated CPU */
8223 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8225 sched_init_granularity();
8226 free_cpumask_var(non_isolated_cpus);
8228 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8229 init_sched_rt_class();
8232 void __init sched_init_smp(void)
8234 sched_init_granularity();
8236 #endif /* CONFIG_SMP */
8238 int in_sched_functions(unsigned long addr)
8240 return in_lock_functions(addr) ||
8241 (addr >= (unsigned long)__sched_text_start
8242 && addr < (unsigned long)__sched_text_end);
8245 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8247 cfs_rq->tasks_timeline = RB_ROOT;
8248 INIT_LIST_HEAD(&cfs_rq->tasks);
8249 #ifdef CONFIG_FAIR_GROUP_SCHED
8252 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8255 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8257 struct rt_prio_array *array;
8260 array = &rt_rq->active;
8261 for (i = 0; i < MAX_RT_PRIO; i++) {
8262 INIT_LIST_HEAD(array->queue + i);
8263 __clear_bit(i, array->bitmap);
8265 /* delimiter for bitsearch: */
8266 __set_bit(MAX_RT_PRIO, array->bitmap);
8268 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8269 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8270 rt_rq->highest_prio.next = MAX_RT_PRIO;
8273 rt_rq->rt_nr_migratory = 0;
8274 rt_rq->overloaded = 0;
8275 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8279 rt_rq->rt_throttled = 0;
8280 rt_rq->rt_runtime = 0;
8281 spin_lock_init(&rt_rq->rt_runtime_lock);
8283 #ifdef CONFIG_RT_GROUP_SCHED
8284 rt_rq->rt_nr_boosted = 0;
8289 #ifdef CONFIG_FAIR_GROUP_SCHED
8290 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8291 struct sched_entity *se, int cpu, int add,
8292 struct sched_entity *parent)
8294 struct rq *rq = cpu_rq(cpu);
8295 tg->cfs_rq[cpu] = cfs_rq;
8296 init_cfs_rq(cfs_rq, rq);
8299 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8302 /* se could be NULL for init_task_group */
8307 se->cfs_rq = &rq->cfs;
8309 se->cfs_rq = parent->my_q;
8312 se->load.weight = tg->shares;
8313 se->load.inv_weight = 0;
8314 se->parent = parent;
8318 #ifdef CONFIG_RT_GROUP_SCHED
8319 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8320 struct sched_rt_entity *rt_se, int cpu, int add,
8321 struct sched_rt_entity *parent)
8323 struct rq *rq = cpu_rq(cpu);
8325 tg->rt_rq[cpu] = rt_rq;
8326 init_rt_rq(rt_rq, rq);
8328 rt_rq->rt_se = rt_se;
8329 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8331 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8333 tg->rt_se[cpu] = rt_se;
8338 rt_se->rt_rq = &rq->rt;
8340 rt_se->rt_rq = parent->my_q;
8342 rt_se->my_q = rt_rq;
8343 rt_se->parent = parent;
8344 INIT_LIST_HEAD(&rt_se->run_list);
8348 void __init sched_init(void)
8351 unsigned long alloc_size = 0, ptr;
8353 #ifdef CONFIG_FAIR_GROUP_SCHED
8354 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8356 #ifdef CONFIG_RT_GROUP_SCHED
8357 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8359 #ifdef CONFIG_USER_SCHED
8363 * As sched_init() is called before page_alloc is setup,
8364 * we use alloc_bootmem().
8367 ptr = (unsigned long)alloc_bootmem(alloc_size);
8369 #ifdef CONFIG_FAIR_GROUP_SCHED
8370 init_task_group.se = (struct sched_entity **)ptr;
8371 ptr += nr_cpu_ids * sizeof(void **);
8373 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8374 ptr += nr_cpu_ids * sizeof(void **);
8376 #ifdef CONFIG_USER_SCHED
8377 root_task_group.se = (struct sched_entity **)ptr;
8378 ptr += nr_cpu_ids * sizeof(void **);
8380 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8381 ptr += nr_cpu_ids * sizeof(void **);
8382 #endif /* CONFIG_USER_SCHED */
8383 #endif /* CONFIG_FAIR_GROUP_SCHED */
8384 #ifdef CONFIG_RT_GROUP_SCHED
8385 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8386 ptr += nr_cpu_ids * sizeof(void **);
8388 init_task_group.rt_rq = (struct rt_rq **)ptr;
8389 ptr += nr_cpu_ids * sizeof(void **);
8391 #ifdef CONFIG_USER_SCHED
8392 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8393 ptr += nr_cpu_ids * sizeof(void **);
8395 root_task_group.rt_rq = (struct rt_rq **)ptr;
8396 ptr += nr_cpu_ids * sizeof(void **);
8397 #endif /* CONFIG_USER_SCHED */
8398 #endif /* CONFIG_RT_GROUP_SCHED */
8402 init_defrootdomain();
8405 init_rt_bandwidth(&def_rt_bandwidth,
8406 global_rt_period(), global_rt_runtime());
8408 #ifdef CONFIG_RT_GROUP_SCHED
8409 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8410 global_rt_period(), global_rt_runtime());
8411 #ifdef CONFIG_USER_SCHED
8412 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8413 global_rt_period(), RUNTIME_INF);
8414 #endif /* CONFIG_USER_SCHED */
8415 #endif /* CONFIG_RT_GROUP_SCHED */
8417 #ifdef CONFIG_GROUP_SCHED
8418 list_add(&init_task_group.list, &task_groups);
8419 INIT_LIST_HEAD(&init_task_group.children);
8421 #ifdef CONFIG_USER_SCHED
8422 INIT_LIST_HEAD(&root_task_group.children);
8423 init_task_group.parent = &root_task_group;
8424 list_add(&init_task_group.siblings, &root_task_group.children);
8425 #endif /* CONFIG_USER_SCHED */
8426 #endif /* CONFIG_GROUP_SCHED */
8428 for_each_possible_cpu(i) {
8432 spin_lock_init(&rq->lock);
8434 init_cfs_rq(&rq->cfs, rq);
8435 init_rt_rq(&rq->rt, rq);
8436 #ifdef CONFIG_FAIR_GROUP_SCHED
8437 init_task_group.shares = init_task_group_load;
8438 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8439 #ifdef CONFIG_CGROUP_SCHED
8441 * How much cpu bandwidth does init_task_group get?
8443 * In case of task-groups formed thr' the cgroup filesystem, it
8444 * gets 100% of the cpu resources in the system. This overall
8445 * system cpu resource is divided among the tasks of
8446 * init_task_group and its child task-groups in a fair manner,
8447 * based on each entity's (task or task-group's) weight
8448 * (se->load.weight).
8450 * In other words, if init_task_group has 10 tasks of weight
8451 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8452 * then A0's share of the cpu resource is:
8454 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8456 * We achieve this by letting init_task_group's tasks sit
8457 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8459 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8460 #elif defined CONFIG_USER_SCHED
8461 root_task_group.shares = NICE_0_LOAD;
8462 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8464 * In case of task-groups formed thr' the user id of tasks,
8465 * init_task_group represents tasks belonging to root user.
8466 * Hence it forms a sibling of all subsequent groups formed.
8467 * In this case, init_task_group gets only a fraction of overall
8468 * system cpu resource, based on the weight assigned to root
8469 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8470 * by letting tasks of init_task_group sit in a separate cfs_rq
8471 * (init_cfs_rq) and having one entity represent this group of
8472 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8474 init_tg_cfs_entry(&init_task_group,
8475 &per_cpu(init_cfs_rq, i),
8476 &per_cpu(init_sched_entity, i), i, 1,
8477 root_task_group.se[i]);
8480 #endif /* CONFIG_FAIR_GROUP_SCHED */
8482 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8483 #ifdef CONFIG_RT_GROUP_SCHED
8484 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8485 #ifdef CONFIG_CGROUP_SCHED
8486 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8487 #elif defined CONFIG_USER_SCHED
8488 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8489 init_tg_rt_entry(&init_task_group,
8490 &per_cpu(init_rt_rq, i),
8491 &per_cpu(init_sched_rt_entity, i), i, 1,
8492 root_task_group.rt_se[i]);
8496 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8497 rq->cpu_load[j] = 0;
8501 rq->active_balance = 0;
8502 rq->next_balance = jiffies;
8506 rq->migration_thread = NULL;
8507 INIT_LIST_HEAD(&rq->migration_queue);
8508 rq_attach_root(rq, &def_root_domain);
8511 atomic_set(&rq->nr_iowait, 0);
8514 set_load_weight(&init_task);
8516 #ifdef CONFIG_PREEMPT_NOTIFIERS
8517 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8521 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8524 #ifdef CONFIG_RT_MUTEXES
8525 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8529 * The boot idle thread does lazy MMU switching as well:
8531 atomic_inc(&init_mm.mm_count);
8532 enter_lazy_tlb(&init_mm, current);
8535 * Make us the idle thread. Technically, schedule() should not be
8536 * called from this thread, however somewhere below it might be,
8537 * but because we are the idle thread, we just pick up running again
8538 * when this runqueue becomes "idle".
8540 init_idle(current, smp_processor_id());
8542 * During early bootup we pretend to be a normal task:
8544 current->sched_class = &fair_sched_class;
8546 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8547 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8550 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8552 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8555 scheduler_running = 1;
8558 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8559 void __might_sleep(char *file, int line)
8562 static unsigned long prev_jiffy; /* ratelimiting */
8564 if ((!in_atomic() && !irqs_disabled()) ||
8565 system_state != SYSTEM_RUNNING || oops_in_progress)
8567 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8569 prev_jiffy = jiffies;
8572 "BUG: sleeping function called from invalid context at %s:%d\n",
8575 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8576 in_atomic(), irqs_disabled(),
8577 current->pid, current->comm);
8579 debug_show_held_locks(current);
8580 if (irqs_disabled())
8581 print_irqtrace_events(current);
8585 EXPORT_SYMBOL(__might_sleep);
8588 #ifdef CONFIG_MAGIC_SYSRQ
8589 static void normalize_task(struct rq *rq, struct task_struct *p)
8593 update_rq_clock(rq);
8594 on_rq = p->se.on_rq;
8596 deactivate_task(rq, p, 0);
8597 __setscheduler(rq, p, SCHED_NORMAL, 0);
8599 activate_task(rq, p, 0);
8600 resched_task(rq->curr);
8604 void normalize_rt_tasks(void)
8606 struct task_struct *g, *p;
8607 unsigned long flags;
8610 read_lock_irqsave(&tasklist_lock, flags);
8611 do_each_thread(g, p) {
8613 * Only normalize user tasks:
8618 p->se.exec_start = 0;
8619 #ifdef CONFIG_SCHEDSTATS
8620 p->se.wait_start = 0;
8621 p->se.sleep_start = 0;
8622 p->se.block_start = 0;
8627 * Renice negative nice level userspace
8630 if (TASK_NICE(p) < 0 && p->mm)
8631 set_user_nice(p, 0);
8635 spin_lock(&p->pi_lock);
8636 rq = __task_rq_lock(p);
8638 normalize_task(rq, p);
8640 __task_rq_unlock(rq);
8641 spin_unlock(&p->pi_lock);
8642 } while_each_thread(g, p);
8644 read_unlock_irqrestore(&tasklist_lock, flags);
8647 #endif /* CONFIG_MAGIC_SYSRQ */
8651 * These functions are only useful for the IA64 MCA handling.
8653 * They can only be called when the whole system has been
8654 * stopped - every CPU needs to be quiescent, and no scheduling
8655 * activity can take place. Using them for anything else would
8656 * be a serious bug, and as a result, they aren't even visible
8657 * under any other configuration.
8661 * curr_task - return the current task for a given cpu.
8662 * @cpu: the processor in question.
8664 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8666 struct task_struct *curr_task(int cpu)
8668 return cpu_curr(cpu);
8672 * set_curr_task - set the current task for a given cpu.
8673 * @cpu: the processor in question.
8674 * @p: the task pointer to set.
8676 * Description: This function must only be used when non-maskable interrupts
8677 * are serviced on a separate stack. It allows the architecture to switch the
8678 * notion of the current task on a cpu in a non-blocking manner. This function
8679 * must be called with all CPU's synchronized, and interrupts disabled, the
8680 * and caller must save the original value of the current task (see
8681 * curr_task() above) and restore that value before reenabling interrupts and
8682 * re-starting the system.
8684 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8686 void set_curr_task(int cpu, struct task_struct *p)
8693 #ifdef CONFIG_FAIR_GROUP_SCHED
8694 static void free_fair_sched_group(struct task_group *tg)
8698 for_each_possible_cpu(i) {
8700 kfree(tg->cfs_rq[i]);
8710 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8712 struct cfs_rq *cfs_rq;
8713 struct sched_entity *se;
8717 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8720 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8724 tg->shares = NICE_0_LOAD;
8726 for_each_possible_cpu(i) {
8729 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8730 GFP_KERNEL, cpu_to_node(i));
8734 se = kzalloc_node(sizeof(struct sched_entity),
8735 GFP_KERNEL, cpu_to_node(i));
8739 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8748 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8750 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8751 &cpu_rq(cpu)->leaf_cfs_rq_list);
8754 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8756 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8758 #else /* !CONFG_FAIR_GROUP_SCHED */
8759 static inline void free_fair_sched_group(struct task_group *tg)
8764 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8769 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8773 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8776 #endif /* CONFIG_FAIR_GROUP_SCHED */
8778 #ifdef CONFIG_RT_GROUP_SCHED
8779 static void free_rt_sched_group(struct task_group *tg)
8783 destroy_rt_bandwidth(&tg->rt_bandwidth);
8785 for_each_possible_cpu(i) {
8787 kfree(tg->rt_rq[i]);
8789 kfree(tg->rt_se[i]);
8797 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8799 struct rt_rq *rt_rq;
8800 struct sched_rt_entity *rt_se;
8804 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8807 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8811 init_rt_bandwidth(&tg->rt_bandwidth,
8812 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8814 for_each_possible_cpu(i) {
8817 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8818 GFP_KERNEL, cpu_to_node(i));
8822 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8823 GFP_KERNEL, cpu_to_node(i));
8827 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8836 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8838 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8839 &cpu_rq(cpu)->leaf_rt_rq_list);
8842 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8844 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8846 #else /* !CONFIG_RT_GROUP_SCHED */
8847 static inline void free_rt_sched_group(struct task_group *tg)
8852 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8857 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8861 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8864 #endif /* CONFIG_RT_GROUP_SCHED */
8866 #ifdef CONFIG_GROUP_SCHED
8867 static void free_sched_group(struct task_group *tg)
8869 free_fair_sched_group(tg);
8870 free_rt_sched_group(tg);
8874 /* allocate runqueue etc for a new task group */
8875 struct task_group *sched_create_group(struct task_group *parent)
8877 struct task_group *tg;
8878 unsigned long flags;
8881 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8883 return ERR_PTR(-ENOMEM);
8885 if (!alloc_fair_sched_group(tg, parent))
8888 if (!alloc_rt_sched_group(tg, parent))
8891 spin_lock_irqsave(&task_group_lock, flags);
8892 for_each_possible_cpu(i) {
8893 register_fair_sched_group(tg, i);
8894 register_rt_sched_group(tg, i);
8896 list_add_rcu(&tg->list, &task_groups);
8898 WARN_ON(!parent); /* root should already exist */
8900 tg->parent = parent;
8901 INIT_LIST_HEAD(&tg->children);
8902 list_add_rcu(&tg->siblings, &parent->children);
8903 spin_unlock_irqrestore(&task_group_lock, flags);
8908 free_sched_group(tg);
8909 return ERR_PTR(-ENOMEM);
8912 /* rcu callback to free various structures associated with a task group */
8913 static void free_sched_group_rcu(struct rcu_head *rhp)
8915 /* now it should be safe to free those cfs_rqs */
8916 free_sched_group(container_of(rhp, struct task_group, rcu));
8919 /* Destroy runqueue etc associated with a task group */
8920 void sched_destroy_group(struct task_group *tg)
8922 unsigned long flags;
8925 spin_lock_irqsave(&task_group_lock, flags);
8926 for_each_possible_cpu(i) {
8927 unregister_fair_sched_group(tg, i);
8928 unregister_rt_sched_group(tg, i);
8930 list_del_rcu(&tg->list);
8931 list_del_rcu(&tg->siblings);
8932 spin_unlock_irqrestore(&task_group_lock, flags);
8934 /* wait for possible concurrent references to cfs_rqs complete */
8935 call_rcu(&tg->rcu, free_sched_group_rcu);
8938 /* change task's runqueue when it moves between groups.
8939 * The caller of this function should have put the task in its new group
8940 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8941 * reflect its new group.
8943 void sched_move_task(struct task_struct *tsk)
8946 unsigned long flags;
8949 rq = task_rq_lock(tsk, &flags);
8951 update_rq_clock(rq);
8953 running = task_current(rq, tsk);
8954 on_rq = tsk->se.on_rq;
8957 dequeue_task(rq, tsk, 0);
8958 if (unlikely(running))
8959 tsk->sched_class->put_prev_task(rq, tsk);
8961 set_task_rq(tsk, task_cpu(tsk));
8963 #ifdef CONFIG_FAIR_GROUP_SCHED
8964 if (tsk->sched_class->moved_group)
8965 tsk->sched_class->moved_group(tsk);
8968 if (unlikely(running))
8969 tsk->sched_class->set_curr_task(rq);
8971 enqueue_task(rq, tsk, 0);
8973 task_rq_unlock(rq, &flags);
8975 #endif /* CONFIG_GROUP_SCHED */
8977 #ifdef CONFIG_FAIR_GROUP_SCHED
8978 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8980 struct cfs_rq *cfs_rq = se->cfs_rq;
8985 dequeue_entity(cfs_rq, se, 0);
8987 se->load.weight = shares;
8988 se->load.inv_weight = 0;
8991 enqueue_entity(cfs_rq, se, 0);
8994 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8996 struct cfs_rq *cfs_rq = se->cfs_rq;
8997 struct rq *rq = cfs_rq->rq;
8998 unsigned long flags;
9000 spin_lock_irqsave(&rq->lock, flags);
9001 __set_se_shares(se, shares);
9002 spin_unlock_irqrestore(&rq->lock, flags);
9005 static DEFINE_MUTEX(shares_mutex);
9007 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9010 unsigned long flags;
9013 * We can't change the weight of the root cgroup.
9018 if (shares < MIN_SHARES)
9019 shares = MIN_SHARES;
9020 else if (shares > MAX_SHARES)
9021 shares = MAX_SHARES;
9023 mutex_lock(&shares_mutex);
9024 if (tg->shares == shares)
9027 spin_lock_irqsave(&task_group_lock, flags);
9028 for_each_possible_cpu(i)
9029 unregister_fair_sched_group(tg, i);
9030 list_del_rcu(&tg->siblings);
9031 spin_unlock_irqrestore(&task_group_lock, flags);
9033 /* wait for any ongoing reference to this group to finish */
9034 synchronize_sched();
9037 * Now we are free to modify the group's share on each cpu
9038 * w/o tripping rebalance_share or load_balance_fair.
9040 tg->shares = shares;
9041 for_each_possible_cpu(i) {
9045 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9046 set_se_shares(tg->se[i], shares);
9050 * Enable load balance activity on this group, by inserting it back on
9051 * each cpu's rq->leaf_cfs_rq_list.
9053 spin_lock_irqsave(&task_group_lock, flags);
9054 for_each_possible_cpu(i)
9055 register_fair_sched_group(tg, i);
9056 list_add_rcu(&tg->siblings, &tg->parent->children);
9057 spin_unlock_irqrestore(&task_group_lock, flags);
9059 mutex_unlock(&shares_mutex);
9063 unsigned long sched_group_shares(struct task_group *tg)
9069 #ifdef CONFIG_RT_GROUP_SCHED
9071 * Ensure that the real time constraints are schedulable.
9073 static DEFINE_MUTEX(rt_constraints_mutex);
9075 static unsigned long to_ratio(u64 period, u64 runtime)
9077 if (runtime == RUNTIME_INF)
9080 return div64_u64(runtime << 20, period);
9083 /* Must be called with tasklist_lock held */
9084 static inline int tg_has_rt_tasks(struct task_group *tg)
9086 struct task_struct *g, *p;
9088 do_each_thread(g, p) {
9089 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9091 } while_each_thread(g, p);
9096 struct rt_schedulable_data {
9097 struct task_group *tg;
9102 static int tg_schedulable(struct task_group *tg, void *data)
9104 struct rt_schedulable_data *d = data;
9105 struct task_group *child;
9106 unsigned long total, sum = 0;
9107 u64 period, runtime;
9109 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9110 runtime = tg->rt_bandwidth.rt_runtime;
9113 period = d->rt_period;
9114 runtime = d->rt_runtime;
9118 * Cannot have more runtime than the period.
9120 if (runtime > period && runtime != RUNTIME_INF)
9124 * Ensure we don't starve existing RT tasks.
9126 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9129 total = to_ratio(period, runtime);
9132 * Nobody can have more than the global setting allows.
9134 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9138 * The sum of our children's runtime should not exceed our own.
9140 list_for_each_entry_rcu(child, &tg->children, siblings) {
9141 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9142 runtime = child->rt_bandwidth.rt_runtime;
9144 if (child == d->tg) {
9145 period = d->rt_period;
9146 runtime = d->rt_runtime;
9149 sum += to_ratio(period, runtime);
9158 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9160 struct rt_schedulable_data data = {
9162 .rt_period = period,
9163 .rt_runtime = runtime,
9166 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9169 static int tg_set_bandwidth(struct task_group *tg,
9170 u64 rt_period, u64 rt_runtime)
9174 mutex_lock(&rt_constraints_mutex);
9175 read_lock(&tasklist_lock);
9176 err = __rt_schedulable(tg, rt_period, rt_runtime);
9180 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9181 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9182 tg->rt_bandwidth.rt_runtime = rt_runtime;
9184 for_each_possible_cpu(i) {
9185 struct rt_rq *rt_rq = tg->rt_rq[i];
9187 spin_lock(&rt_rq->rt_runtime_lock);
9188 rt_rq->rt_runtime = rt_runtime;
9189 spin_unlock(&rt_rq->rt_runtime_lock);
9191 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9193 read_unlock(&tasklist_lock);
9194 mutex_unlock(&rt_constraints_mutex);
9199 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9201 u64 rt_runtime, rt_period;
9203 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9204 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9205 if (rt_runtime_us < 0)
9206 rt_runtime = RUNTIME_INF;
9208 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9211 long sched_group_rt_runtime(struct task_group *tg)
9215 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9218 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9219 do_div(rt_runtime_us, NSEC_PER_USEC);
9220 return rt_runtime_us;
9223 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9225 u64 rt_runtime, rt_period;
9227 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9228 rt_runtime = tg->rt_bandwidth.rt_runtime;
9233 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9236 long sched_group_rt_period(struct task_group *tg)
9240 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9241 do_div(rt_period_us, NSEC_PER_USEC);
9242 return rt_period_us;
9245 static int sched_rt_global_constraints(void)
9247 u64 runtime, period;
9250 if (sysctl_sched_rt_period <= 0)
9253 runtime = global_rt_runtime();
9254 period = global_rt_period();
9257 * Sanity check on the sysctl variables.
9259 if (runtime > period && runtime != RUNTIME_INF)
9262 mutex_lock(&rt_constraints_mutex);
9263 read_lock(&tasklist_lock);
9264 ret = __rt_schedulable(NULL, 0, 0);
9265 read_unlock(&tasklist_lock);
9266 mutex_unlock(&rt_constraints_mutex);
9270 #else /* !CONFIG_RT_GROUP_SCHED */
9271 static int sched_rt_global_constraints(void)
9273 unsigned long flags;
9276 if (sysctl_sched_rt_period <= 0)
9279 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9280 for_each_possible_cpu(i) {
9281 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9283 spin_lock(&rt_rq->rt_runtime_lock);
9284 rt_rq->rt_runtime = global_rt_runtime();
9285 spin_unlock(&rt_rq->rt_runtime_lock);
9287 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9291 #endif /* CONFIG_RT_GROUP_SCHED */
9293 int sched_rt_handler(struct ctl_table *table, int write,
9294 struct file *filp, void __user *buffer, size_t *lenp,
9298 int old_period, old_runtime;
9299 static DEFINE_MUTEX(mutex);
9302 old_period = sysctl_sched_rt_period;
9303 old_runtime = sysctl_sched_rt_runtime;
9305 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9307 if (!ret && write) {
9308 ret = sched_rt_global_constraints();
9310 sysctl_sched_rt_period = old_period;
9311 sysctl_sched_rt_runtime = old_runtime;
9313 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9314 def_rt_bandwidth.rt_period =
9315 ns_to_ktime(global_rt_period());
9318 mutex_unlock(&mutex);
9323 #ifdef CONFIG_CGROUP_SCHED
9325 /* return corresponding task_group object of a cgroup */
9326 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9328 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9329 struct task_group, css);
9332 static struct cgroup_subsys_state *
9333 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9335 struct task_group *tg, *parent;
9337 if (!cgrp->parent) {
9338 /* This is early initialization for the top cgroup */
9339 return &init_task_group.css;
9342 parent = cgroup_tg(cgrp->parent);
9343 tg = sched_create_group(parent);
9345 return ERR_PTR(-ENOMEM);
9351 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9353 struct task_group *tg = cgroup_tg(cgrp);
9355 sched_destroy_group(tg);
9359 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9360 struct task_struct *tsk)
9362 #ifdef CONFIG_RT_GROUP_SCHED
9363 /* Don't accept realtime tasks when there is no way for them to run */
9364 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9367 /* We don't support RT-tasks being in separate groups */
9368 if (tsk->sched_class != &fair_sched_class)
9376 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9377 struct cgroup *old_cont, struct task_struct *tsk)
9379 sched_move_task(tsk);
9382 #ifdef CONFIG_FAIR_GROUP_SCHED
9383 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9386 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9389 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9391 struct task_group *tg = cgroup_tg(cgrp);
9393 return (u64) tg->shares;
9395 #endif /* CONFIG_FAIR_GROUP_SCHED */
9397 #ifdef CONFIG_RT_GROUP_SCHED
9398 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9401 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9404 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9406 return sched_group_rt_runtime(cgroup_tg(cgrp));
9409 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9412 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9415 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9417 return sched_group_rt_period(cgroup_tg(cgrp));
9419 #endif /* CONFIG_RT_GROUP_SCHED */
9421 static struct cftype cpu_files[] = {
9422 #ifdef CONFIG_FAIR_GROUP_SCHED
9425 .read_u64 = cpu_shares_read_u64,
9426 .write_u64 = cpu_shares_write_u64,
9429 #ifdef CONFIG_RT_GROUP_SCHED
9431 .name = "rt_runtime_us",
9432 .read_s64 = cpu_rt_runtime_read,
9433 .write_s64 = cpu_rt_runtime_write,
9436 .name = "rt_period_us",
9437 .read_u64 = cpu_rt_period_read_uint,
9438 .write_u64 = cpu_rt_period_write_uint,
9443 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9445 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9448 struct cgroup_subsys cpu_cgroup_subsys = {
9450 .create = cpu_cgroup_create,
9451 .destroy = cpu_cgroup_destroy,
9452 .can_attach = cpu_cgroup_can_attach,
9453 .attach = cpu_cgroup_attach,
9454 .populate = cpu_cgroup_populate,
9455 .subsys_id = cpu_cgroup_subsys_id,
9459 #endif /* CONFIG_CGROUP_SCHED */
9461 #ifdef CONFIG_CGROUP_CPUACCT
9464 * CPU accounting code for task groups.
9466 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9467 * (balbir@in.ibm.com).
9470 /* track cpu usage of a group of tasks and its child groups */
9472 struct cgroup_subsys_state css;
9473 /* cpuusage holds pointer to a u64-type object on every cpu */
9475 struct cpuacct *parent;
9478 struct cgroup_subsys cpuacct_subsys;
9480 /* return cpu accounting group corresponding to this container */
9481 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9483 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9484 struct cpuacct, css);
9487 /* return cpu accounting group to which this task belongs */
9488 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9490 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9491 struct cpuacct, css);
9494 /* create a new cpu accounting group */
9495 static struct cgroup_subsys_state *cpuacct_create(
9496 struct cgroup_subsys *ss, struct cgroup *cgrp)
9498 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9501 return ERR_PTR(-ENOMEM);
9503 ca->cpuusage = alloc_percpu(u64);
9504 if (!ca->cpuusage) {
9506 return ERR_PTR(-ENOMEM);
9510 ca->parent = cgroup_ca(cgrp->parent);
9515 /* destroy an existing cpu accounting group */
9517 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9519 struct cpuacct *ca = cgroup_ca(cgrp);
9521 free_percpu(ca->cpuusage);
9525 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9527 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9530 #ifndef CONFIG_64BIT
9532 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9534 spin_lock_irq(&cpu_rq(cpu)->lock);
9536 spin_unlock_irq(&cpu_rq(cpu)->lock);
9544 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9546 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9548 #ifndef CONFIG_64BIT
9550 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9552 spin_lock_irq(&cpu_rq(cpu)->lock);
9554 spin_unlock_irq(&cpu_rq(cpu)->lock);
9560 /* return total cpu usage (in nanoseconds) of a group */
9561 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9563 struct cpuacct *ca = cgroup_ca(cgrp);
9564 u64 totalcpuusage = 0;
9567 for_each_present_cpu(i)
9568 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9570 return totalcpuusage;
9573 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9576 struct cpuacct *ca = cgroup_ca(cgrp);
9585 for_each_present_cpu(i)
9586 cpuacct_cpuusage_write(ca, i, 0);
9592 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9595 struct cpuacct *ca = cgroup_ca(cgroup);
9599 for_each_present_cpu(i) {
9600 percpu = cpuacct_cpuusage_read(ca, i);
9601 seq_printf(m, "%llu ", (unsigned long long) percpu);
9603 seq_printf(m, "\n");
9607 static struct cftype files[] = {
9610 .read_u64 = cpuusage_read,
9611 .write_u64 = cpuusage_write,
9614 .name = "usage_percpu",
9615 .read_seq_string = cpuacct_percpu_seq_read,
9620 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9622 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9626 * charge this task's execution time to its accounting group.
9628 * called with rq->lock held.
9630 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9635 if (!cpuacct_subsys.active)
9638 cpu = task_cpu(tsk);
9641 for (; ca; ca = ca->parent) {
9642 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9643 *cpuusage += cputime;
9647 struct cgroup_subsys cpuacct_subsys = {
9649 .create = cpuacct_create,
9650 .destroy = cpuacct_destroy,
9651 .populate = cpuacct_populate,
9652 .subsys_id = cpuacct_subsys_id,
9654 #endif /* CONFIG_CGROUP_CPUACCT */