4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (!rt_bandwidth_enabled() || rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
237 if (hrtimer_active(&rt_b->rt_period_timer))
240 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
241 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
243 soft = hrtimer_get_softexpires(&rt_b->rt_period_timer);
244 hard = hrtimer_get_expires(&rt_b->rt_period_timer);
245 delta = ktime_to_ns(ktime_sub(hard, soft));
246 __hrtimer_start_range_ns(&rt_b->rt_period_timer, soft, delta,
247 HRTIMER_MODE_ABS, 0);
249 spin_unlock(&rt_b->rt_runtime_lock);
252 #ifdef CONFIG_RT_GROUP_SCHED
253 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
255 hrtimer_cancel(&rt_b->rt_period_timer);
260 * sched_domains_mutex serializes calls to arch_init_sched_domains,
261 * detach_destroy_domains and partition_sched_domains.
263 static DEFINE_MUTEX(sched_domains_mutex);
265 #ifdef CONFIG_GROUP_SCHED
267 #include <linux/cgroup.h>
271 static LIST_HEAD(task_groups);
273 /* task group related information */
275 #ifdef CONFIG_CGROUP_SCHED
276 struct cgroup_subsys_state css;
279 #ifdef CONFIG_USER_SCHED
283 #ifdef CONFIG_FAIR_GROUP_SCHED
284 /* schedulable entities of this group on each cpu */
285 struct sched_entity **se;
286 /* runqueue "owned" by this group on each cpu */
287 struct cfs_rq **cfs_rq;
288 unsigned long shares;
291 #ifdef CONFIG_RT_GROUP_SCHED
292 struct sched_rt_entity **rt_se;
293 struct rt_rq **rt_rq;
295 struct rt_bandwidth rt_bandwidth;
299 struct list_head list;
301 struct task_group *parent;
302 struct list_head siblings;
303 struct list_head children;
306 #ifdef CONFIG_USER_SCHED
308 /* Helper function to pass uid information to create_sched_user() */
309 void set_tg_uid(struct user_struct *user)
311 user->tg->uid = user->uid;
316 * Every UID task group (including init_task_group aka UID-0) will
317 * be a child to this group.
319 struct task_group root_task_group;
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 /* Default task group's sched entity on each cpu */
323 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
324 /* Default task group's cfs_rq on each cpu */
325 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 #ifdef CONFIG_RT_GROUP_SCHED
329 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
330 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
331 #endif /* CONFIG_RT_GROUP_SCHED */
332 #else /* !CONFIG_USER_SCHED */
333 #define root_task_group init_task_group
334 #endif /* CONFIG_USER_SCHED */
336 /* task_group_lock serializes add/remove of task groups and also changes to
337 * a task group's cpu shares.
339 static DEFINE_SPINLOCK(task_group_lock);
342 static int root_task_group_empty(void)
344 return list_empty(&root_task_group.children);
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 #ifdef CONFIG_USER_SCHED
350 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
351 #else /* !CONFIG_USER_SCHED */
352 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
353 #endif /* CONFIG_USER_SCHED */
356 * A weight of 0 or 1 can cause arithmetics problems.
357 * A weight of a cfs_rq is the sum of weights of which entities
358 * are queued on this cfs_rq, so a weight of a entity should not be
359 * too large, so as the shares value of a task group.
360 * (The default weight is 1024 - so there's no practical
361 * limitation from this.)
364 #define MAX_SHARES (1UL << 18)
366 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
369 /* Default task group.
370 * Every task in system belong to this group at bootup.
372 struct task_group init_task_group;
374 /* return group to which a task belongs */
375 static inline struct task_group *task_group(struct task_struct *p)
377 struct task_group *tg;
379 #ifdef CONFIG_USER_SCHED
381 tg = __task_cred(p)->user->tg;
383 #elif defined(CONFIG_CGROUP_SCHED)
384 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
385 struct task_group, css);
387 tg = &init_task_group;
392 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
393 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
395 #ifdef CONFIG_FAIR_GROUP_SCHED
396 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
397 p->se.parent = task_group(p)->se[cpu];
400 #ifdef CONFIG_RT_GROUP_SCHED
401 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
402 p->rt.parent = task_group(p)->rt_se[cpu];
409 static int root_task_group_empty(void)
415 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
416 static inline struct task_group *task_group(struct task_struct *p)
421 #endif /* CONFIG_GROUP_SCHED */
423 /* CFS-related fields in a runqueue */
425 struct load_weight load;
426 unsigned long nr_running;
431 struct rb_root tasks_timeline;
432 struct rb_node *rb_leftmost;
434 struct list_head tasks;
435 struct list_head *balance_iterator;
438 * 'curr' points to currently running entity on this cfs_rq.
439 * It is set to NULL otherwise (i.e when none are currently running).
441 struct sched_entity *curr, *next, *last;
443 unsigned int nr_spread_over;
445 #ifdef CONFIG_FAIR_GROUP_SCHED
446 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
449 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
450 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
451 * (like users, containers etc.)
453 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
454 * list is used during load balance.
456 struct list_head leaf_cfs_rq_list;
457 struct task_group *tg; /* group that "owns" this runqueue */
461 * the part of load.weight contributed by tasks
463 unsigned long task_weight;
466 * h_load = weight * f(tg)
468 * Where f(tg) is the recursive weight fraction assigned to
471 unsigned long h_load;
474 * this cpu's part of tg->shares
476 unsigned long shares;
479 * load.weight at the time we set shares
481 unsigned long rq_weight;
486 /* Real-Time classes' related field in a runqueue: */
488 struct rt_prio_array active;
489 unsigned long rt_nr_running;
490 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
492 int curr; /* highest queued rt task prio */
494 int next; /* next highest */
499 unsigned long rt_nr_migratory;
501 struct plist_head pushable_tasks;
506 /* Nests inside the rq lock: */
507 spinlock_t rt_runtime_lock;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 unsigned long rt_nr_boosted;
513 struct list_head leaf_rt_rq_list;
514 struct task_group *tg;
515 struct sched_rt_entity *rt_se;
522 * We add the notion of a root-domain which will be used to define per-domain
523 * variables. Each exclusive cpuset essentially defines an island domain by
524 * fully partitioning the member cpus from any other cpuset. Whenever a new
525 * exclusive cpuset is created, we also create and attach a new root-domain
532 cpumask_var_t online;
535 * The "RT overload" flag: it gets set if a CPU has more than
536 * one runnable RT task.
538 cpumask_var_t rto_mask;
541 struct cpupri cpupri;
543 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
545 * Preferred wake up cpu nominated by sched_mc balance that will be
546 * used when most cpus are idle in the system indicating overall very
547 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
549 unsigned int sched_mc_preferred_wakeup_cpu;
554 * By default the system creates a single root-domain with all cpus as
555 * members (mimicking the global state we have today).
557 static struct root_domain def_root_domain;
562 * This is the main, per-CPU runqueue data structure.
564 * Locking rule: those places that want to lock multiple runqueues
565 * (such as the load balancing or the thread migration code), lock
566 * acquire operations must be ordered by ascending &runqueue.
573 * nr_running and cpu_load should be in the same cacheline because
574 * remote CPUs use both these fields when doing load calculation.
576 unsigned long nr_running;
577 #define CPU_LOAD_IDX_MAX 5
578 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
580 unsigned long last_tick_seen;
581 unsigned char in_nohz_recently;
583 /* capture load from *all* tasks on this cpu: */
584 struct load_weight load;
585 unsigned long nr_load_updates;
591 #ifdef CONFIG_FAIR_GROUP_SCHED
592 /* list of leaf cfs_rq on this cpu: */
593 struct list_head leaf_cfs_rq_list;
595 #ifdef CONFIG_RT_GROUP_SCHED
596 struct list_head leaf_rt_rq_list;
600 * This is part of a global counter where only the total sum
601 * over all CPUs matters. A task can increase this counter on
602 * one CPU and if it got migrated afterwards it may decrease
603 * it on another CPU. Always updated under the runqueue lock:
605 unsigned long nr_uninterruptible;
607 struct task_struct *curr, *idle;
608 unsigned long next_balance;
609 struct mm_struct *prev_mm;
616 struct root_domain *rd;
617 struct sched_domain *sd;
619 unsigned char idle_at_tick;
620 /* For active balancing */
623 /* cpu of this runqueue: */
627 unsigned long avg_load_per_task;
629 struct task_struct *migration_thread;
630 struct list_head migration_queue;
633 /* calc_load related fields */
634 unsigned long calc_load_update;
635 long calc_load_active;
637 #ifdef CONFIG_SCHED_HRTICK
639 int hrtick_csd_pending;
640 struct call_single_data hrtick_csd;
642 struct hrtimer hrtick_timer;
645 #ifdef CONFIG_SCHEDSTATS
647 struct sched_info rq_sched_info;
648 unsigned long long rq_cpu_time;
649 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
651 /* sys_sched_yield() stats */
652 unsigned int yld_count;
654 /* schedule() stats */
655 unsigned int sched_switch;
656 unsigned int sched_count;
657 unsigned int sched_goidle;
659 /* try_to_wake_up() stats */
660 unsigned int ttwu_count;
661 unsigned int ttwu_local;
664 unsigned int bkl_count;
668 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
670 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
672 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
675 static inline int cpu_of(struct rq *rq)
685 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
686 * See detach_destroy_domains: synchronize_sched for details.
688 * The domain tree of any CPU may only be accessed from within
689 * preempt-disabled sections.
691 #define for_each_domain(cpu, __sd) \
692 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
694 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
695 #define this_rq() (&__get_cpu_var(runqueues))
696 #define task_rq(p) cpu_rq(task_cpu(p))
697 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
699 static inline void update_rq_clock(struct rq *rq)
701 rq->clock = sched_clock_cpu(cpu_of(rq));
705 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
707 #ifdef CONFIG_SCHED_DEBUG
708 # define const_debug __read_mostly
710 # define const_debug static const
716 * Returns true if the current cpu runqueue is locked.
717 * This interface allows printk to be called with the runqueue lock
718 * held and know whether or not it is OK to wake up the klogd.
720 int runqueue_is_locked(void)
723 struct rq *rq = cpu_rq(cpu);
726 ret = spin_is_locked(&rq->lock);
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug unsigned int sysctl_sched_features =
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly char *sched_feat_names[] = {
758 #include "sched_features.h"
764 static int sched_feat_show(struct seq_file *m, void *v)
768 for (i = 0; sched_feat_names[i]; i++) {
769 if (!(sysctl_sched_features & (1UL << i)))
771 seq_printf(m, "%s ", sched_feat_names[i]);
779 sched_feat_write(struct file *filp, const char __user *ubuf,
780 size_t cnt, loff_t *ppos)
790 if (copy_from_user(&buf, ubuf, cnt))
795 if (strncmp(buf, "NO_", 3) == 0) {
800 for (i = 0; sched_feat_names[i]; i++) {
801 int len = strlen(sched_feat_names[i]);
803 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
805 sysctl_sched_features &= ~(1UL << i);
807 sysctl_sched_features |= (1UL << i);
812 if (!sched_feat_names[i])
820 static int sched_feat_open(struct inode *inode, struct file *filp)
822 return single_open(filp, sched_feat_show, NULL);
825 static struct file_operations sched_feat_fops = {
826 .open = sched_feat_open,
827 .write = sched_feat_write,
830 .release = single_release,
833 static __init int sched_init_debug(void)
835 debugfs_create_file("sched_features", 0644, NULL, NULL,
840 late_initcall(sched_init_debug);
844 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
847 * Number of tasks to iterate in a single balance run.
848 * Limited because this is done with IRQs disabled.
850 const_debug unsigned int sysctl_sched_nr_migrate = 32;
853 * ratelimit for updating the group shares.
856 unsigned int sysctl_sched_shares_ratelimit = 250000;
859 * Inject some fuzzyness into changing the per-cpu group shares
860 * this avoids remote rq-locks at the expense of fairness.
863 unsigned int sysctl_sched_shares_thresh = 4;
866 * period over which we measure -rt task cpu usage in us.
869 unsigned int sysctl_sched_rt_period = 1000000;
871 static __read_mostly int scheduler_running;
874 * part of the period that we allow rt tasks to run in us.
877 int sysctl_sched_rt_runtime = 950000;
879 static inline u64 global_rt_period(void)
881 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
884 static inline u64 global_rt_runtime(void)
886 if (sysctl_sched_rt_runtime < 0)
889 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
892 #ifndef prepare_arch_switch
893 # define prepare_arch_switch(next) do { } while (0)
895 #ifndef finish_arch_switch
896 # define finish_arch_switch(prev) do { } while (0)
899 static inline int task_current(struct rq *rq, struct task_struct *p)
901 return rq->curr == p;
904 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
905 static inline int task_running(struct rq *rq, struct task_struct *p)
907 return task_current(rq, p);
910 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
914 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
916 #ifdef CONFIG_DEBUG_SPINLOCK
917 /* this is a valid case when another task releases the spinlock */
918 rq->lock.owner = current;
921 * If we are tracking spinlock dependencies then we have to
922 * fix up the runqueue lock - which gets 'carried over' from
925 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
927 spin_unlock_irq(&rq->lock);
930 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
931 static inline int task_running(struct rq *rq, struct task_struct *p)
936 return task_current(rq, p);
940 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
944 * We can optimise this out completely for !SMP, because the
945 * SMP rebalancing from interrupt is the only thing that cares
950 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
951 spin_unlock_irq(&rq->lock);
953 spin_unlock(&rq->lock);
957 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
961 * After ->oncpu is cleared, the task can be moved to a different CPU.
962 * We must ensure this doesn't happen until the switch is completely
968 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
972 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
975 * __task_rq_lock - lock the runqueue a given task resides on.
976 * Must be called interrupts disabled.
978 static inline struct rq *__task_rq_lock(struct task_struct *p)
982 struct rq *rq = task_rq(p);
983 spin_lock(&rq->lock);
984 if (likely(rq == task_rq(p)))
986 spin_unlock(&rq->lock);
991 * task_rq_lock - lock the runqueue a given task resides on and disable
992 * interrupts. Note the ordering: we can safely lookup the task_rq without
993 * explicitly disabling preemption.
995 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1001 local_irq_save(*flags);
1003 spin_lock(&rq->lock);
1004 if (likely(rq == task_rq(p)))
1006 spin_unlock_irqrestore(&rq->lock, *flags);
1010 void task_rq_unlock_wait(struct task_struct *p)
1012 struct rq *rq = task_rq(p);
1014 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1015 spin_unlock_wait(&rq->lock);
1018 static void __task_rq_unlock(struct rq *rq)
1019 __releases(rq->lock)
1021 spin_unlock(&rq->lock);
1024 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1025 __releases(rq->lock)
1027 spin_unlock_irqrestore(&rq->lock, *flags);
1031 * this_rq_lock - lock this runqueue and disable interrupts.
1033 static struct rq *this_rq_lock(void)
1034 __acquires(rq->lock)
1038 local_irq_disable();
1040 spin_lock(&rq->lock);
1045 #ifdef CONFIG_SCHED_HRTICK
1047 * Use HR-timers to deliver accurate preemption points.
1049 * Its all a bit involved since we cannot program an hrt while holding the
1050 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1053 * When we get rescheduled we reprogram the hrtick_timer outside of the
1059 * - enabled by features
1060 * - hrtimer is actually high res
1062 static inline int hrtick_enabled(struct rq *rq)
1064 if (!sched_feat(HRTICK))
1066 if (!cpu_active(cpu_of(rq)))
1068 return hrtimer_is_hres_active(&rq->hrtick_timer);
1071 static void hrtick_clear(struct rq *rq)
1073 if (hrtimer_active(&rq->hrtick_timer))
1074 hrtimer_cancel(&rq->hrtick_timer);
1078 * High-resolution timer tick.
1079 * Runs from hardirq context with interrupts disabled.
1081 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1083 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1085 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1087 spin_lock(&rq->lock);
1088 update_rq_clock(rq);
1089 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1090 spin_unlock(&rq->lock);
1092 return HRTIMER_NORESTART;
1097 * called from hardirq (IPI) context
1099 static void __hrtick_start(void *arg)
1101 struct rq *rq = arg;
1103 spin_lock(&rq->lock);
1104 hrtimer_restart(&rq->hrtick_timer);
1105 rq->hrtick_csd_pending = 0;
1106 spin_unlock(&rq->lock);
1110 * Called to set the hrtick timer state.
1112 * called with rq->lock held and irqs disabled
1114 static void hrtick_start(struct rq *rq, u64 delay)
1116 struct hrtimer *timer = &rq->hrtick_timer;
1117 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1119 hrtimer_set_expires(timer, time);
1121 if (rq == this_rq()) {
1122 hrtimer_restart(timer);
1123 } else if (!rq->hrtick_csd_pending) {
1124 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1125 rq->hrtick_csd_pending = 1;
1130 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1132 int cpu = (int)(long)hcpu;
1135 case CPU_UP_CANCELED:
1136 case CPU_UP_CANCELED_FROZEN:
1137 case CPU_DOWN_PREPARE:
1138 case CPU_DOWN_PREPARE_FROZEN:
1140 case CPU_DEAD_FROZEN:
1141 hrtick_clear(cpu_rq(cpu));
1148 static __init void init_hrtick(void)
1150 hotcpu_notifier(hotplug_hrtick, 0);
1154 * Called to set the hrtick timer state.
1156 * called with rq->lock held and irqs disabled
1158 static void hrtick_start(struct rq *rq, u64 delay)
1160 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1161 HRTIMER_MODE_REL, 0);
1164 static inline void init_hrtick(void)
1167 #endif /* CONFIG_SMP */
1169 static void init_rq_hrtick(struct rq *rq)
1172 rq->hrtick_csd_pending = 0;
1174 rq->hrtick_csd.flags = 0;
1175 rq->hrtick_csd.func = __hrtick_start;
1176 rq->hrtick_csd.info = rq;
1179 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1180 rq->hrtick_timer.function = hrtick;
1182 #else /* CONFIG_SCHED_HRTICK */
1183 static inline void hrtick_clear(struct rq *rq)
1187 static inline void init_rq_hrtick(struct rq *rq)
1191 static inline void init_hrtick(void)
1194 #endif /* CONFIG_SCHED_HRTICK */
1197 * resched_task - mark a task 'to be rescheduled now'.
1199 * On UP this means the setting of the need_resched flag, on SMP it
1200 * might also involve a cross-CPU call to trigger the scheduler on
1205 #ifndef tsk_is_polling
1206 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1209 static void resched_task(struct task_struct *p)
1213 assert_spin_locked(&task_rq(p)->lock);
1215 if (test_tsk_need_resched(p))
1218 set_tsk_need_resched(p);
1221 if (cpu == smp_processor_id())
1224 /* NEED_RESCHED must be visible before we test polling */
1226 if (!tsk_is_polling(p))
1227 smp_send_reschedule(cpu);
1230 static void resched_cpu(int cpu)
1232 struct rq *rq = cpu_rq(cpu);
1233 unsigned long flags;
1235 if (!spin_trylock_irqsave(&rq->lock, flags))
1237 resched_task(cpu_curr(cpu));
1238 spin_unlock_irqrestore(&rq->lock, flags);
1243 * When add_timer_on() enqueues a timer into the timer wheel of an
1244 * idle CPU then this timer might expire before the next timer event
1245 * which is scheduled to wake up that CPU. In case of a completely
1246 * idle system the next event might even be infinite time into the
1247 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1248 * leaves the inner idle loop so the newly added timer is taken into
1249 * account when the CPU goes back to idle and evaluates the timer
1250 * wheel for the next timer event.
1252 void wake_up_idle_cpu(int cpu)
1254 struct rq *rq = cpu_rq(cpu);
1256 if (cpu == smp_processor_id())
1260 * This is safe, as this function is called with the timer
1261 * wheel base lock of (cpu) held. When the CPU is on the way
1262 * to idle and has not yet set rq->curr to idle then it will
1263 * be serialized on the timer wheel base lock and take the new
1264 * timer into account automatically.
1266 if (rq->curr != rq->idle)
1270 * We can set TIF_RESCHED on the idle task of the other CPU
1271 * lockless. The worst case is that the other CPU runs the
1272 * idle task through an additional NOOP schedule()
1274 set_tsk_need_resched(rq->idle);
1276 /* NEED_RESCHED must be visible before we test polling */
1278 if (!tsk_is_polling(rq->idle))
1279 smp_send_reschedule(cpu);
1281 #endif /* CONFIG_NO_HZ */
1283 #else /* !CONFIG_SMP */
1284 static void resched_task(struct task_struct *p)
1286 assert_spin_locked(&task_rq(p)->lock);
1287 set_tsk_need_resched(p);
1289 #endif /* CONFIG_SMP */
1291 #if BITS_PER_LONG == 32
1292 # define WMULT_CONST (~0UL)
1294 # define WMULT_CONST (1UL << 32)
1297 #define WMULT_SHIFT 32
1300 * Shift right and round:
1302 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1305 * delta *= weight / lw
1307 static unsigned long
1308 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1309 struct load_weight *lw)
1313 if (!lw->inv_weight) {
1314 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1317 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1321 tmp = (u64)delta_exec * weight;
1323 * Check whether we'd overflow the 64-bit multiplication:
1325 if (unlikely(tmp > WMULT_CONST))
1326 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1329 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1331 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1334 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1340 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1347 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1348 * of tasks with abnormal "nice" values across CPUs the contribution that
1349 * each task makes to its run queue's load is weighted according to its
1350 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1351 * scaled version of the new time slice allocation that they receive on time
1355 #define WEIGHT_IDLEPRIO 3
1356 #define WMULT_IDLEPRIO 1431655765
1359 * Nice levels are multiplicative, with a gentle 10% change for every
1360 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1361 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1362 * that remained on nice 0.
1364 * The "10% effect" is relative and cumulative: from _any_ nice level,
1365 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1366 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1367 * If a task goes up by ~10% and another task goes down by ~10% then
1368 * the relative distance between them is ~25%.)
1370 static const int prio_to_weight[40] = {
1371 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1372 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1373 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1374 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1375 /* 0 */ 1024, 820, 655, 526, 423,
1376 /* 5 */ 335, 272, 215, 172, 137,
1377 /* 10 */ 110, 87, 70, 56, 45,
1378 /* 15 */ 36, 29, 23, 18, 15,
1382 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1384 * In cases where the weight does not change often, we can use the
1385 * precalculated inverse to speed up arithmetics by turning divisions
1386 * into multiplications:
1388 static const u32 prio_to_wmult[40] = {
1389 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1390 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1391 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1392 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1393 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1394 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1395 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1396 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1399 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1402 * runqueue iterator, to support SMP load-balancing between different
1403 * scheduling classes, without having to expose their internal data
1404 * structures to the load-balancing proper:
1406 struct rq_iterator {
1408 struct task_struct *(*start)(void *);
1409 struct task_struct *(*next)(void *);
1413 static unsigned long
1414 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1415 unsigned long max_load_move, struct sched_domain *sd,
1416 enum cpu_idle_type idle, int *all_pinned,
1417 int *this_best_prio, struct rq_iterator *iterator);
1420 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1421 struct sched_domain *sd, enum cpu_idle_type idle,
1422 struct rq_iterator *iterator);
1425 /* Time spent by the tasks of the cpu accounting group executing in ... */
1426 enum cpuacct_stat_index {
1427 CPUACCT_STAT_USER, /* ... user mode */
1428 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1430 CPUACCT_STAT_NSTATS,
1433 #ifdef CONFIG_CGROUP_CPUACCT
1434 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1435 static void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val);
1438 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1439 static inline void cpuacct_update_stats(struct task_struct *tsk,
1440 enum cpuacct_stat_index idx, cputime_t val) {}
1443 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1445 update_load_add(&rq->load, load);
1448 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1450 update_load_sub(&rq->load, load);
1453 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1454 typedef int (*tg_visitor)(struct task_group *, void *);
1457 * Iterate the full tree, calling @down when first entering a node and @up when
1458 * leaving it for the final time.
1460 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1462 struct task_group *parent, *child;
1466 parent = &root_task_group;
1468 ret = (*down)(parent, data);
1471 list_for_each_entry_rcu(child, &parent->children, siblings) {
1478 ret = (*up)(parent, data);
1483 parent = parent->parent;
1492 static int tg_nop(struct task_group *tg, void *data)
1499 static unsigned long source_load(int cpu, int type);
1500 static unsigned long target_load(int cpu, int type);
1501 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1503 static unsigned long cpu_avg_load_per_task(int cpu)
1505 struct rq *rq = cpu_rq(cpu);
1506 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1509 rq->avg_load_per_task = rq->load.weight / nr_running;
1511 rq->avg_load_per_task = 0;
1513 return rq->avg_load_per_task;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1521 * Calculate and set the cpu's group shares.
1524 update_group_shares_cpu(struct task_group *tg, int cpu,
1525 unsigned long sd_shares, unsigned long sd_rq_weight)
1527 unsigned long shares;
1528 unsigned long rq_weight;
1533 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1536 * \Sum shares * rq_weight
1537 * shares = -----------------------
1541 shares = (sd_shares * rq_weight) / sd_rq_weight;
1542 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1544 if (abs(shares - tg->se[cpu]->load.weight) >
1545 sysctl_sched_shares_thresh) {
1546 struct rq *rq = cpu_rq(cpu);
1547 unsigned long flags;
1549 spin_lock_irqsave(&rq->lock, flags);
1550 tg->cfs_rq[cpu]->shares = shares;
1552 __set_se_shares(tg->se[cpu], shares);
1553 spin_unlock_irqrestore(&rq->lock, flags);
1558 * Re-compute the task group their per cpu shares over the given domain.
1559 * This needs to be done in a bottom-up fashion because the rq weight of a
1560 * parent group depends on the shares of its child groups.
1562 static int tg_shares_up(struct task_group *tg, void *data)
1564 unsigned long weight, rq_weight = 0;
1565 unsigned long shares = 0;
1566 struct sched_domain *sd = data;
1569 for_each_cpu(i, sched_domain_span(sd)) {
1571 * If there are currently no tasks on the cpu pretend there
1572 * is one of average load so that when a new task gets to
1573 * run here it will not get delayed by group starvation.
1575 weight = tg->cfs_rq[i]->load.weight;
1577 weight = NICE_0_LOAD;
1579 tg->cfs_rq[i]->rq_weight = weight;
1580 rq_weight += weight;
1581 shares += tg->cfs_rq[i]->shares;
1584 if ((!shares && rq_weight) || shares > tg->shares)
1585 shares = tg->shares;
1587 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1588 shares = tg->shares;
1590 for_each_cpu(i, sched_domain_span(sd))
1591 update_group_shares_cpu(tg, i, shares, rq_weight);
1597 * Compute the cpu's hierarchical load factor for each task group.
1598 * This needs to be done in a top-down fashion because the load of a child
1599 * group is a fraction of its parents load.
1601 static int tg_load_down(struct task_group *tg, void *data)
1604 long cpu = (long)data;
1607 load = cpu_rq(cpu)->load.weight;
1609 load = tg->parent->cfs_rq[cpu]->h_load;
1610 load *= tg->cfs_rq[cpu]->shares;
1611 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1614 tg->cfs_rq[cpu]->h_load = load;
1619 static void update_shares(struct sched_domain *sd)
1621 u64 now = cpu_clock(raw_smp_processor_id());
1622 s64 elapsed = now - sd->last_update;
1624 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1625 sd->last_update = now;
1626 walk_tg_tree(tg_nop, tg_shares_up, sd);
1630 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1632 spin_unlock(&rq->lock);
1634 spin_lock(&rq->lock);
1637 static void update_h_load(long cpu)
1639 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1644 static inline void update_shares(struct sched_domain *sd)
1648 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1654 #ifdef CONFIG_PREEMPT
1657 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1658 * way at the expense of forcing extra atomic operations in all
1659 * invocations. This assures that the double_lock is acquired using the
1660 * same underlying policy as the spinlock_t on this architecture, which
1661 * reduces latency compared to the unfair variant below. However, it
1662 * also adds more overhead and therefore may reduce throughput.
1664 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1665 __releases(this_rq->lock)
1666 __acquires(busiest->lock)
1667 __acquires(this_rq->lock)
1669 spin_unlock(&this_rq->lock);
1670 double_rq_lock(this_rq, busiest);
1677 * Unfair double_lock_balance: Optimizes throughput at the expense of
1678 * latency by eliminating extra atomic operations when the locks are
1679 * already in proper order on entry. This favors lower cpu-ids and will
1680 * grant the double lock to lower cpus over higher ids under contention,
1681 * regardless of entry order into the function.
1683 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1684 __releases(this_rq->lock)
1685 __acquires(busiest->lock)
1686 __acquires(this_rq->lock)
1690 if (unlikely(!spin_trylock(&busiest->lock))) {
1691 if (busiest < this_rq) {
1692 spin_unlock(&this_rq->lock);
1693 spin_lock(&busiest->lock);
1694 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1697 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1702 #endif /* CONFIG_PREEMPT */
1705 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1707 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1709 if (unlikely(!irqs_disabled())) {
1710 /* printk() doesn't work good under rq->lock */
1711 spin_unlock(&this_rq->lock);
1715 return _double_lock_balance(this_rq, busiest);
1718 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1719 __releases(busiest->lock)
1721 spin_unlock(&busiest->lock);
1722 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1726 #ifdef CONFIG_FAIR_GROUP_SCHED
1727 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1730 cfs_rq->shares = shares;
1735 static void calc_load_account_active(struct rq *this_rq);
1737 #include "sched_stats.h"
1738 #include "sched_idletask.c"
1739 #include "sched_fair.c"
1740 #include "sched_rt.c"
1741 #ifdef CONFIG_SCHED_DEBUG
1742 # include "sched_debug.c"
1745 #define sched_class_highest (&rt_sched_class)
1746 #define for_each_class(class) \
1747 for (class = sched_class_highest; class; class = class->next)
1749 static void inc_nr_running(struct rq *rq)
1754 static void dec_nr_running(struct rq *rq)
1759 static void set_load_weight(struct task_struct *p)
1761 if (task_has_rt_policy(p)) {
1762 p->se.load.weight = prio_to_weight[0] * 2;
1763 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1768 * SCHED_IDLE tasks get minimal weight:
1770 if (p->policy == SCHED_IDLE) {
1771 p->se.load.weight = WEIGHT_IDLEPRIO;
1772 p->se.load.inv_weight = WMULT_IDLEPRIO;
1776 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1777 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1780 static void update_avg(u64 *avg, u64 sample)
1782 s64 diff = sample - *avg;
1786 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1789 p->se.start_runtime = p->se.sum_exec_runtime;
1791 sched_info_queued(p);
1792 p->sched_class->enqueue_task(rq, p, wakeup);
1796 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1799 if (p->se.last_wakeup) {
1800 update_avg(&p->se.avg_overlap,
1801 p->se.sum_exec_runtime - p->se.last_wakeup);
1802 p->se.last_wakeup = 0;
1804 update_avg(&p->se.avg_wakeup,
1805 sysctl_sched_wakeup_granularity);
1809 sched_info_dequeued(p);
1810 p->sched_class->dequeue_task(rq, p, sleep);
1815 * __normal_prio - return the priority that is based on the static prio
1817 static inline int __normal_prio(struct task_struct *p)
1819 return p->static_prio;
1823 * Calculate the expected normal priority: i.e. priority
1824 * without taking RT-inheritance into account. Might be
1825 * boosted by interactivity modifiers. Changes upon fork,
1826 * setprio syscalls, and whenever the interactivity
1827 * estimator recalculates.
1829 static inline int normal_prio(struct task_struct *p)
1833 if (task_has_rt_policy(p))
1834 prio = MAX_RT_PRIO-1 - p->rt_priority;
1836 prio = __normal_prio(p);
1841 * Calculate the current priority, i.e. the priority
1842 * taken into account by the scheduler. This value might
1843 * be boosted by RT tasks, or might be boosted by
1844 * interactivity modifiers. Will be RT if the task got
1845 * RT-boosted. If not then it returns p->normal_prio.
1847 static int effective_prio(struct task_struct *p)
1849 p->normal_prio = normal_prio(p);
1851 * If we are RT tasks or we were boosted to RT priority,
1852 * keep the priority unchanged. Otherwise, update priority
1853 * to the normal priority:
1855 if (!rt_prio(p->prio))
1856 return p->normal_prio;
1861 * activate_task - move a task to the runqueue.
1863 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1865 if (task_contributes_to_load(p))
1866 rq->nr_uninterruptible--;
1868 enqueue_task(rq, p, wakeup);
1873 * deactivate_task - remove a task from the runqueue.
1875 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1877 if (task_contributes_to_load(p))
1878 rq->nr_uninterruptible++;
1880 dequeue_task(rq, p, sleep);
1885 * task_curr - is this task currently executing on a CPU?
1886 * @p: the task in question.
1888 inline int task_curr(const struct task_struct *p)
1890 return cpu_curr(task_cpu(p)) == p;
1893 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1895 set_task_rq(p, cpu);
1898 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1899 * successfuly executed on another CPU. We must ensure that updates of
1900 * per-task data have been completed by this moment.
1903 task_thread_info(p)->cpu = cpu;
1907 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1908 const struct sched_class *prev_class,
1909 int oldprio, int running)
1911 if (prev_class != p->sched_class) {
1912 if (prev_class->switched_from)
1913 prev_class->switched_from(rq, p, running);
1914 p->sched_class->switched_to(rq, p, running);
1916 p->sched_class->prio_changed(rq, p, oldprio, running);
1921 /* Used instead of source_load when we know the type == 0 */
1922 static unsigned long weighted_cpuload(const int cpu)
1924 return cpu_rq(cpu)->load.weight;
1928 * Is this task likely cache-hot:
1931 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1936 * Buddy candidates are cache hot:
1938 if (sched_feat(CACHE_HOT_BUDDY) &&
1939 (&p->se == cfs_rq_of(&p->se)->next ||
1940 &p->se == cfs_rq_of(&p->se)->last))
1943 if (p->sched_class != &fair_sched_class)
1946 if (sysctl_sched_migration_cost == -1)
1948 if (sysctl_sched_migration_cost == 0)
1951 delta = now - p->se.exec_start;
1953 return delta < (s64)sysctl_sched_migration_cost;
1957 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1959 int old_cpu = task_cpu(p);
1960 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1961 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1962 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1965 clock_offset = old_rq->clock - new_rq->clock;
1967 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1969 #ifdef CONFIG_SCHEDSTATS
1970 if (p->se.wait_start)
1971 p->se.wait_start -= clock_offset;
1972 if (p->se.sleep_start)
1973 p->se.sleep_start -= clock_offset;
1974 if (p->se.block_start)
1975 p->se.block_start -= clock_offset;
1976 if (old_cpu != new_cpu) {
1977 schedstat_inc(p, se.nr_migrations);
1978 if (task_hot(p, old_rq->clock, NULL))
1979 schedstat_inc(p, se.nr_forced2_migrations);
1982 p->se.vruntime -= old_cfsrq->min_vruntime -
1983 new_cfsrq->min_vruntime;
1985 __set_task_cpu(p, new_cpu);
1988 struct migration_req {
1989 struct list_head list;
1991 struct task_struct *task;
1994 struct completion done;
1998 * The task's runqueue lock must be held.
1999 * Returns true if you have to wait for migration thread.
2002 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2004 struct rq *rq = task_rq(p);
2007 * If the task is not on a runqueue (and not running), then
2008 * it is sufficient to simply update the task's cpu field.
2010 if (!p->se.on_rq && !task_running(rq, p)) {
2011 set_task_cpu(p, dest_cpu);
2015 init_completion(&req->done);
2017 req->dest_cpu = dest_cpu;
2018 list_add(&req->list, &rq->migration_queue);
2024 * wait_task_inactive - wait for a thread to unschedule.
2026 * If @match_state is nonzero, it's the @p->state value just checked and
2027 * not expected to change. If it changes, i.e. @p might have woken up,
2028 * then return zero. When we succeed in waiting for @p to be off its CPU,
2029 * we return a positive number (its total switch count). If a second call
2030 * a short while later returns the same number, the caller can be sure that
2031 * @p has remained unscheduled the whole time.
2033 * The caller must ensure that the task *will* unschedule sometime soon,
2034 * else this function might spin for a *long* time. This function can't
2035 * be called with interrupts off, or it may introduce deadlock with
2036 * smp_call_function() if an IPI is sent by the same process we are
2037 * waiting to become inactive.
2039 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2041 unsigned long flags;
2048 * We do the initial early heuristics without holding
2049 * any task-queue locks at all. We'll only try to get
2050 * the runqueue lock when things look like they will
2056 * If the task is actively running on another CPU
2057 * still, just relax and busy-wait without holding
2060 * NOTE! Since we don't hold any locks, it's not
2061 * even sure that "rq" stays as the right runqueue!
2062 * But we don't care, since "task_running()" will
2063 * return false if the runqueue has changed and p
2064 * is actually now running somewhere else!
2066 while (task_running(rq, p)) {
2067 if (match_state && unlikely(p->state != match_state))
2073 * Ok, time to look more closely! We need the rq
2074 * lock now, to be *sure*. If we're wrong, we'll
2075 * just go back and repeat.
2077 rq = task_rq_lock(p, &flags);
2078 trace_sched_wait_task(rq, p);
2079 running = task_running(rq, p);
2080 on_rq = p->se.on_rq;
2082 if (!match_state || p->state == match_state)
2083 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2084 task_rq_unlock(rq, &flags);
2087 * If it changed from the expected state, bail out now.
2089 if (unlikely(!ncsw))
2093 * Was it really running after all now that we
2094 * checked with the proper locks actually held?
2096 * Oops. Go back and try again..
2098 if (unlikely(running)) {
2104 * It's not enough that it's not actively running,
2105 * it must be off the runqueue _entirely_, and not
2108 * So if it was still runnable (but just not actively
2109 * running right now), it's preempted, and we should
2110 * yield - it could be a while.
2112 if (unlikely(on_rq)) {
2113 schedule_timeout_uninterruptible(1);
2118 * Ahh, all good. It wasn't running, and it wasn't
2119 * runnable, which means that it will never become
2120 * running in the future either. We're all done!
2129 * kick_process - kick a running thread to enter/exit the kernel
2130 * @p: the to-be-kicked thread
2132 * Cause a process which is running on another CPU to enter
2133 * kernel-mode, without any delay. (to get signals handled.)
2135 * NOTE: this function doesnt have to take the runqueue lock,
2136 * because all it wants to ensure is that the remote task enters
2137 * the kernel. If the IPI races and the task has been migrated
2138 * to another CPU then no harm is done and the purpose has been
2141 void kick_process(struct task_struct *p)
2147 if ((cpu != smp_processor_id()) && task_curr(p))
2148 smp_send_reschedule(cpu);
2153 * Return a low guess at the load of a migration-source cpu weighted
2154 * according to the scheduling class and "nice" value.
2156 * We want to under-estimate the load of migration sources, to
2157 * balance conservatively.
2159 static unsigned long source_load(int cpu, int type)
2161 struct rq *rq = cpu_rq(cpu);
2162 unsigned long total = weighted_cpuload(cpu);
2164 if (type == 0 || !sched_feat(LB_BIAS))
2167 return min(rq->cpu_load[type-1], total);
2171 * Return a high guess at the load of a migration-target cpu weighted
2172 * according to the scheduling class and "nice" value.
2174 static unsigned long target_load(int cpu, int type)
2176 struct rq *rq = cpu_rq(cpu);
2177 unsigned long total = weighted_cpuload(cpu);
2179 if (type == 0 || !sched_feat(LB_BIAS))
2182 return max(rq->cpu_load[type-1], total);
2186 * find_idlest_group finds and returns the least busy CPU group within the
2189 static struct sched_group *
2190 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2192 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2193 unsigned long min_load = ULONG_MAX, this_load = 0;
2194 int load_idx = sd->forkexec_idx;
2195 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2198 unsigned long load, avg_load;
2202 /* Skip over this group if it has no CPUs allowed */
2203 if (!cpumask_intersects(sched_group_cpus(group),
2207 local_group = cpumask_test_cpu(this_cpu,
2208 sched_group_cpus(group));
2210 /* Tally up the load of all CPUs in the group */
2213 for_each_cpu(i, sched_group_cpus(group)) {
2214 /* Bias balancing toward cpus of our domain */
2216 load = source_load(i, load_idx);
2218 load = target_load(i, load_idx);
2223 /* Adjust by relative CPU power of the group */
2224 avg_load = sg_div_cpu_power(group,
2225 avg_load * SCHED_LOAD_SCALE);
2228 this_load = avg_load;
2230 } else if (avg_load < min_load) {
2231 min_load = avg_load;
2234 } while (group = group->next, group != sd->groups);
2236 if (!idlest || 100*this_load < imbalance*min_load)
2242 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2245 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2247 unsigned long load, min_load = ULONG_MAX;
2251 /* Traverse only the allowed CPUs */
2252 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2253 load = weighted_cpuload(i);
2255 if (load < min_load || (load == min_load && i == this_cpu)) {
2265 * sched_balance_self: balance the current task (running on cpu) in domains
2266 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2269 * Balance, ie. select the least loaded group.
2271 * Returns the target CPU number, or the same CPU if no balancing is needed.
2273 * preempt must be disabled.
2275 static int sched_balance_self(int cpu, int flag)
2277 struct task_struct *t = current;
2278 struct sched_domain *tmp, *sd = NULL;
2280 for_each_domain(cpu, tmp) {
2282 * If power savings logic is enabled for a domain, stop there.
2284 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2286 if (tmp->flags & flag)
2294 struct sched_group *group;
2295 int new_cpu, weight;
2297 if (!(sd->flags & flag)) {
2302 group = find_idlest_group(sd, t, cpu);
2308 new_cpu = find_idlest_cpu(group, t, cpu);
2309 if (new_cpu == -1 || new_cpu == cpu) {
2310 /* Now try balancing at a lower domain level of cpu */
2315 /* Now try balancing at a lower domain level of new_cpu */
2317 weight = cpumask_weight(sched_domain_span(sd));
2319 for_each_domain(cpu, tmp) {
2320 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2322 if (tmp->flags & flag)
2325 /* while loop will break here if sd == NULL */
2331 #endif /* CONFIG_SMP */
2334 * try_to_wake_up - wake up a thread
2335 * @p: the to-be-woken-up thread
2336 * @state: the mask of task states that can be woken
2337 * @sync: do a synchronous wakeup?
2339 * Put it on the run-queue if it's not already there. The "current"
2340 * thread is always on the run-queue (except when the actual
2341 * re-schedule is in progress), and as such you're allowed to do
2342 * the simpler "current->state = TASK_RUNNING" to mark yourself
2343 * runnable without the overhead of this.
2345 * returns failure only if the task is already active.
2347 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2349 int cpu, orig_cpu, this_cpu, success = 0;
2350 unsigned long flags;
2354 if (!sched_feat(SYNC_WAKEUPS))
2358 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2359 struct sched_domain *sd;
2361 this_cpu = raw_smp_processor_id();
2364 for_each_domain(this_cpu, sd) {
2365 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2374 rq = task_rq_lock(p, &flags);
2375 update_rq_clock(rq);
2376 old_state = p->state;
2377 if (!(old_state & state))
2385 this_cpu = smp_processor_id();
2388 if (unlikely(task_running(rq, p)))
2391 cpu = p->sched_class->select_task_rq(p, sync);
2392 if (cpu != orig_cpu) {
2393 set_task_cpu(p, cpu);
2394 task_rq_unlock(rq, &flags);
2395 /* might preempt at this point */
2396 rq = task_rq_lock(p, &flags);
2397 old_state = p->state;
2398 if (!(old_state & state))
2403 this_cpu = smp_processor_id();
2407 #ifdef CONFIG_SCHEDSTATS
2408 schedstat_inc(rq, ttwu_count);
2409 if (cpu == this_cpu)
2410 schedstat_inc(rq, ttwu_local);
2412 struct sched_domain *sd;
2413 for_each_domain(this_cpu, sd) {
2414 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2415 schedstat_inc(sd, ttwu_wake_remote);
2420 #endif /* CONFIG_SCHEDSTATS */
2423 #endif /* CONFIG_SMP */
2424 schedstat_inc(p, se.nr_wakeups);
2426 schedstat_inc(p, se.nr_wakeups_sync);
2427 if (orig_cpu != cpu)
2428 schedstat_inc(p, se.nr_wakeups_migrate);
2429 if (cpu == this_cpu)
2430 schedstat_inc(p, se.nr_wakeups_local);
2432 schedstat_inc(p, se.nr_wakeups_remote);
2433 activate_task(rq, p, 1);
2437 * Only attribute actual wakeups done by this task.
2439 if (!in_interrupt()) {
2440 struct sched_entity *se = ¤t->se;
2441 u64 sample = se->sum_exec_runtime;
2443 if (se->last_wakeup)
2444 sample -= se->last_wakeup;
2446 sample -= se->start_runtime;
2447 update_avg(&se->avg_wakeup, sample);
2449 se->last_wakeup = se->sum_exec_runtime;
2453 trace_sched_wakeup(rq, p, success);
2454 check_preempt_curr(rq, p, sync);
2456 p->state = TASK_RUNNING;
2458 if (p->sched_class->task_wake_up)
2459 p->sched_class->task_wake_up(rq, p);
2462 task_rq_unlock(rq, &flags);
2467 int wake_up_process(struct task_struct *p)
2469 return try_to_wake_up(p, TASK_ALL, 0);
2471 EXPORT_SYMBOL(wake_up_process);
2473 int wake_up_state(struct task_struct *p, unsigned int state)
2475 return try_to_wake_up(p, state, 0);
2479 * Perform scheduler related setup for a newly forked process p.
2480 * p is forked by current.
2482 * __sched_fork() is basic setup used by init_idle() too:
2484 static void __sched_fork(struct task_struct *p)
2486 p->se.exec_start = 0;
2487 p->se.sum_exec_runtime = 0;
2488 p->se.prev_sum_exec_runtime = 0;
2489 p->se.last_wakeup = 0;
2490 p->se.avg_overlap = 0;
2491 p->se.start_runtime = 0;
2492 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2494 #ifdef CONFIG_SCHEDSTATS
2495 p->se.wait_start = 0;
2496 p->se.sum_sleep_runtime = 0;
2497 p->se.sleep_start = 0;
2498 p->se.block_start = 0;
2499 p->se.sleep_max = 0;
2500 p->se.block_max = 0;
2502 p->se.slice_max = 0;
2506 INIT_LIST_HEAD(&p->rt.run_list);
2508 INIT_LIST_HEAD(&p->se.group_node);
2510 #ifdef CONFIG_PREEMPT_NOTIFIERS
2511 INIT_HLIST_HEAD(&p->preempt_notifiers);
2515 * We mark the process as running here, but have not actually
2516 * inserted it onto the runqueue yet. This guarantees that
2517 * nobody will actually run it, and a signal or other external
2518 * event cannot wake it up and insert it on the runqueue either.
2520 p->state = TASK_RUNNING;
2524 * fork()/clone()-time setup:
2526 void sched_fork(struct task_struct *p, int clone_flags)
2528 int cpu = get_cpu();
2533 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2535 set_task_cpu(p, cpu);
2538 * Make sure we do not leak PI boosting priority to the child:
2540 p->prio = current->normal_prio;
2541 if (!rt_prio(p->prio))
2542 p->sched_class = &fair_sched_class;
2544 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2545 if (likely(sched_info_on()))
2546 memset(&p->sched_info, 0, sizeof(p->sched_info));
2548 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2551 #ifdef CONFIG_PREEMPT
2552 /* Want to start with kernel preemption disabled. */
2553 task_thread_info(p)->preempt_count = 1;
2555 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2561 * wake_up_new_task - wake up a newly created task for the first time.
2563 * This function will do some initial scheduler statistics housekeeping
2564 * that must be done for every newly created context, then puts the task
2565 * on the runqueue and wakes it.
2567 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2569 unsigned long flags;
2572 rq = task_rq_lock(p, &flags);
2573 BUG_ON(p->state != TASK_RUNNING);
2574 update_rq_clock(rq);
2576 p->prio = effective_prio(p);
2578 if (!p->sched_class->task_new || !current->se.on_rq) {
2579 activate_task(rq, p, 0);
2582 * Let the scheduling class do new task startup
2583 * management (if any):
2585 p->sched_class->task_new(rq, p);
2588 trace_sched_wakeup_new(rq, p, 1);
2589 check_preempt_curr(rq, p, 0);
2591 if (p->sched_class->task_wake_up)
2592 p->sched_class->task_wake_up(rq, p);
2594 task_rq_unlock(rq, &flags);
2597 #ifdef CONFIG_PREEMPT_NOTIFIERS
2600 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2601 * @notifier: notifier struct to register
2603 void preempt_notifier_register(struct preempt_notifier *notifier)
2605 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2607 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2610 * preempt_notifier_unregister - no longer interested in preemption notifications
2611 * @notifier: notifier struct to unregister
2613 * This is safe to call from within a preemption notifier.
2615 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2617 hlist_del(¬ifier->link);
2619 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2621 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2623 struct preempt_notifier *notifier;
2624 struct hlist_node *node;
2626 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2627 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2631 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2632 struct task_struct *next)
2634 struct preempt_notifier *notifier;
2635 struct hlist_node *node;
2637 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2638 notifier->ops->sched_out(notifier, next);
2641 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2643 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2648 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2649 struct task_struct *next)
2653 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2656 * prepare_task_switch - prepare to switch tasks
2657 * @rq: the runqueue preparing to switch
2658 * @prev: the current task that is being switched out
2659 * @next: the task we are going to switch to.
2661 * This is called with the rq lock held and interrupts off. It must
2662 * be paired with a subsequent finish_task_switch after the context
2665 * prepare_task_switch sets up locking and calls architecture specific
2669 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2670 struct task_struct *next)
2672 fire_sched_out_preempt_notifiers(prev, next);
2673 prepare_lock_switch(rq, next);
2674 prepare_arch_switch(next);
2678 * finish_task_switch - clean up after a task-switch
2679 * @rq: runqueue associated with task-switch
2680 * @prev: the thread we just switched away from.
2682 * finish_task_switch must be called after the context switch, paired
2683 * with a prepare_task_switch call before the context switch.
2684 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2685 * and do any other architecture-specific cleanup actions.
2687 * Note that we may have delayed dropping an mm in context_switch(). If
2688 * so, we finish that here outside of the runqueue lock. (Doing it
2689 * with the lock held can cause deadlocks; see schedule() for
2692 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2693 __releases(rq->lock)
2695 struct mm_struct *mm = rq->prev_mm;
2698 int post_schedule = 0;
2700 if (current->sched_class->needs_post_schedule)
2701 post_schedule = current->sched_class->needs_post_schedule(rq);
2707 * A task struct has one reference for the use as "current".
2708 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2709 * schedule one last time. The schedule call will never return, and
2710 * the scheduled task must drop that reference.
2711 * The test for TASK_DEAD must occur while the runqueue locks are
2712 * still held, otherwise prev could be scheduled on another cpu, die
2713 * there before we look at prev->state, and then the reference would
2715 * Manfred Spraul <manfred@colorfullife.com>
2717 prev_state = prev->state;
2718 finish_arch_switch(prev);
2719 finish_lock_switch(rq, prev);
2722 current->sched_class->post_schedule(rq);
2725 fire_sched_in_preempt_notifiers(current);
2728 if (unlikely(prev_state == TASK_DEAD)) {
2730 * Remove function-return probe instances associated with this
2731 * task and put them back on the free list.
2733 kprobe_flush_task(prev);
2734 put_task_struct(prev);
2739 * schedule_tail - first thing a freshly forked thread must call.
2740 * @prev: the thread we just switched away from.
2742 asmlinkage void schedule_tail(struct task_struct *prev)
2743 __releases(rq->lock)
2745 struct rq *rq = this_rq();
2747 finish_task_switch(rq, prev);
2748 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2749 /* In this case, finish_task_switch does not reenable preemption */
2752 if (current->set_child_tid)
2753 put_user(task_pid_vnr(current), current->set_child_tid);
2757 * context_switch - switch to the new MM and the new
2758 * thread's register state.
2761 context_switch(struct rq *rq, struct task_struct *prev,
2762 struct task_struct *next)
2764 struct mm_struct *mm, *oldmm;
2766 prepare_task_switch(rq, prev, next);
2767 trace_sched_switch(rq, prev, next);
2769 oldmm = prev->active_mm;
2771 * For paravirt, this is coupled with an exit in switch_to to
2772 * combine the page table reload and the switch backend into
2775 arch_enter_lazy_cpu_mode();
2777 if (unlikely(!mm)) {
2778 next->active_mm = oldmm;
2779 atomic_inc(&oldmm->mm_count);
2780 enter_lazy_tlb(oldmm, next);
2782 switch_mm(oldmm, mm, next);
2784 if (unlikely(!prev->mm)) {
2785 prev->active_mm = NULL;
2786 rq->prev_mm = oldmm;
2789 * Since the runqueue lock will be released by the next
2790 * task (which is an invalid locking op but in the case
2791 * of the scheduler it's an obvious special-case), so we
2792 * do an early lockdep release here:
2794 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2795 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2798 /* Here we just switch the register state and the stack. */
2799 switch_to(prev, next, prev);
2803 * this_rq must be evaluated again because prev may have moved
2804 * CPUs since it called schedule(), thus the 'rq' on its stack
2805 * frame will be invalid.
2807 finish_task_switch(this_rq(), prev);
2811 * nr_running, nr_uninterruptible and nr_context_switches:
2813 * externally visible scheduler statistics: current number of runnable
2814 * threads, current number of uninterruptible-sleeping threads, total
2815 * number of context switches performed since bootup.
2817 unsigned long nr_running(void)
2819 unsigned long i, sum = 0;
2821 for_each_online_cpu(i)
2822 sum += cpu_rq(i)->nr_running;
2827 unsigned long nr_uninterruptible(void)
2829 unsigned long i, sum = 0;
2831 for_each_possible_cpu(i)
2832 sum += cpu_rq(i)->nr_uninterruptible;
2835 * Since we read the counters lockless, it might be slightly
2836 * inaccurate. Do not allow it to go below zero though:
2838 if (unlikely((long)sum < 0))
2844 unsigned long long nr_context_switches(void)
2847 unsigned long long sum = 0;
2849 for_each_possible_cpu(i)
2850 sum += cpu_rq(i)->nr_switches;
2855 unsigned long nr_iowait(void)
2857 unsigned long i, sum = 0;
2859 for_each_possible_cpu(i)
2860 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2865 /* Variables and functions for calc_load */
2866 static atomic_long_t calc_load_tasks;
2867 static unsigned long calc_load_update;
2868 unsigned long avenrun[3];
2869 EXPORT_SYMBOL(avenrun);
2872 * get_avenrun - get the load average array
2873 * @loads: pointer to dest load array
2874 * @offset: offset to add
2875 * @shift: shift count to shift the result left
2877 * These values are estimates at best, so no need for locking.
2879 void get_avenrun(unsigned long *loads, unsigned long offset, int shift)
2881 loads[0] = (avenrun[0] + offset) << shift;
2882 loads[1] = (avenrun[1] + offset) << shift;
2883 loads[2] = (avenrun[2] + offset) << shift;
2886 static unsigned long
2887 calc_load(unsigned long load, unsigned long exp, unsigned long active)
2890 load += active * (FIXED_1 - exp);
2891 return load >> FSHIFT;
2895 * calc_load - update the avenrun load estimates 10 ticks after the
2896 * CPUs have updated calc_load_tasks.
2898 void calc_global_load(void)
2900 unsigned long upd = calc_load_update + 10;
2903 if (time_before(jiffies, upd))
2906 active = atomic_long_read(&calc_load_tasks);
2907 active = active > 0 ? active * FIXED_1 : 0;
2909 avenrun[0] = calc_load(avenrun[0], EXP_1, active);
2910 avenrun[1] = calc_load(avenrun[1], EXP_5, active);
2911 avenrun[2] = calc_load(avenrun[2], EXP_15, active);
2913 calc_load_update += LOAD_FREQ;
2917 * Either called from update_cpu_load() or from a cpu going idle
2919 static void calc_load_account_active(struct rq *this_rq)
2921 long nr_active, delta;
2923 nr_active = this_rq->nr_running;
2924 nr_active += (long) this_rq->nr_uninterruptible;
2926 if (nr_active != this_rq->calc_load_active) {
2927 delta = nr_active - this_rq->calc_load_active;
2928 this_rq->calc_load_active = nr_active;
2929 atomic_long_add(delta, &calc_load_tasks);
2934 * Update rq->cpu_load[] statistics. This function is usually called every
2935 * scheduler tick (TICK_NSEC).
2937 static void update_cpu_load(struct rq *this_rq)
2939 unsigned long this_load = this_rq->load.weight;
2942 this_rq->nr_load_updates++;
2944 /* Update our load: */
2945 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2946 unsigned long old_load, new_load;
2948 /* scale is effectively 1 << i now, and >> i divides by scale */
2950 old_load = this_rq->cpu_load[i];
2951 new_load = this_load;
2953 * Round up the averaging division if load is increasing. This
2954 * prevents us from getting stuck on 9 if the load is 10, for
2957 if (new_load > old_load)
2958 new_load += scale-1;
2959 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2962 if (time_after_eq(jiffies, this_rq->calc_load_update)) {
2963 this_rq->calc_load_update += LOAD_FREQ;
2964 calc_load_account_active(this_rq);
2971 * double_rq_lock - safely lock two runqueues
2973 * Note this does not disable interrupts like task_rq_lock,
2974 * you need to do so manually before calling.
2976 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2977 __acquires(rq1->lock)
2978 __acquires(rq2->lock)
2980 BUG_ON(!irqs_disabled());
2982 spin_lock(&rq1->lock);
2983 __acquire(rq2->lock); /* Fake it out ;) */
2986 spin_lock(&rq1->lock);
2987 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2989 spin_lock(&rq2->lock);
2990 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2993 update_rq_clock(rq1);
2994 update_rq_clock(rq2);
2998 * double_rq_unlock - safely unlock two runqueues
3000 * Note this does not restore interrupts like task_rq_unlock,
3001 * you need to do so manually after calling.
3003 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3004 __releases(rq1->lock)
3005 __releases(rq2->lock)
3007 spin_unlock(&rq1->lock);
3009 spin_unlock(&rq2->lock);
3011 __release(rq2->lock);
3015 * If dest_cpu is allowed for this process, migrate the task to it.
3016 * This is accomplished by forcing the cpu_allowed mask to only
3017 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3018 * the cpu_allowed mask is restored.
3020 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3022 struct migration_req req;
3023 unsigned long flags;
3026 rq = task_rq_lock(p, &flags);
3027 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
3028 || unlikely(!cpu_active(dest_cpu)))
3031 /* force the process onto the specified CPU */
3032 if (migrate_task(p, dest_cpu, &req)) {
3033 /* Need to wait for migration thread (might exit: take ref). */
3034 struct task_struct *mt = rq->migration_thread;
3036 get_task_struct(mt);
3037 task_rq_unlock(rq, &flags);
3038 wake_up_process(mt);
3039 put_task_struct(mt);
3040 wait_for_completion(&req.done);
3045 task_rq_unlock(rq, &flags);
3049 * sched_exec - execve() is a valuable balancing opportunity, because at
3050 * this point the task has the smallest effective memory and cache footprint.
3052 void sched_exec(void)
3054 int new_cpu, this_cpu = get_cpu();
3055 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3057 if (new_cpu != this_cpu)
3058 sched_migrate_task(current, new_cpu);
3062 * pull_task - move a task from a remote runqueue to the local runqueue.
3063 * Both runqueues must be locked.
3065 static void pull_task(struct rq *src_rq, struct task_struct *p,
3066 struct rq *this_rq, int this_cpu)
3068 deactivate_task(src_rq, p, 0);
3069 set_task_cpu(p, this_cpu);
3070 activate_task(this_rq, p, 0);
3072 * Note that idle threads have a prio of MAX_PRIO, for this test
3073 * to be always true for them.
3075 check_preempt_curr(this_rq, p, 0);
3079 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3082 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3083 struct sched_domain *sd, enum cpu_idle_type idle,
3086 int tsk_cache_hot = 0;
3088 * We do not migrate tasks that are:
3089 * 1) running (obviously), or
3090 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3091 * 3) are cache-hot on their current CPU.
3093 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3094 schedstat_inc(p, se.nr_failed_migrations_affine);
3099 if (task_running(rq, p)) {
3100 schedstat_inc(p, se.nr_failed_migrations_running);
3105 * Aggressive migration if:
3106 * 1) task is cache cold, or
3107 * 2) too many balance attempts have failed.
3110 tsk_cache_hot = task_hot(p, rq->clock, sd);
3111 if (!tsk_cache_hot ||
3112 sd->nr_balance_failed > sd->cache_nice_tries) {
3113 #ifdef CONFIG_SCHEDSTATS
3114 if (tsk_cache_hot) {
3115 schedstat_inc(sd, lb_hot_gained[idle]);
3116 schedstat_inc(p, se.nr_forced_migrations);
3122 if (tsk_cache_hot) {
3123 schedstat_inc(p, se.nr_failed_migrations_hot);
3129 static unsigned long
3130 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3131 unsigned long max_load_move, struct sched_domain *sd,
3132 enum cpu_idle_type idle, int *all_pinned,
3133 int *this_best_prio, struct rq_iterator *iterator)
3135 int loops = 0, pulled = 0, pinned = 0;
3136 struct task_struct *p;
3137 long rem_load_move = max_load_move;
3139 if (max_load_move == 0)
3145 * Start the load-balancing iterator:
3147 p = iterator->start(iterator->arg);
3149 if (!p || loops++ > sysctl_sched_nr_migrate)
3152 if ((p->se.load.weight >> 1) > rem_load_move ||
3153 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3154 p = iterator->next(iterator->arg);
3158 pull_task(busiest, p, this_rq, this_cpu);
3160 rem_load_move -= p->se.load.weight;
3162 #ifdef CONFIG_PREEMPT
3164 * NEWIDLE balancing is a source of latency, so preemptible kernels
3165 * will stop after the first task is pulled to minimize the critical
3168 if (idle == CPU_NEWLY_IDLE)
3173 * We only want to steal up to the prescribed amount of weighted load.
3175 if (rem_load_move > 0) {
3176 if (p->prio < *this_best_prio)
3177 *this_best_prio = p->prio;
3178 p = iterator->next(iterator->arg);
3183 * Right now, this is one of only two places pull_task() is called,
3184 * so we can safely collect pull_task() stats here rather than
3185 * inside pull_task().
3187 schedstat_add(sd, lb_gained[idle], pulled);
3190 *all_pinned = pinned;
3192 return max_load_move - rem_load_move;
3196 * move_tasks tries to move up to max_load_move weighted load from busiest to
3197 * this_rq, as part of a balancing operation within domain "sd".
3198 * Returns 1 if successful and 0 otherwise.
3200 * Called with both runqueues locked.
3202 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3203 unsigned long max_load_move,
3204 struct sched_domain *sd, enum cpu_idle_type idle,
3207 const struct sched_class *class = sched_class_highest;
3208 unsigned long total_load_moved = 0;
3209 int this_best_prio = this_rq->curr->prio;
3213 class->load_balance(this_rq, this_cpu, busiest,
3214 max_load_move - total_load_moved,
3215 sd, idle, all_pinned, &this_best_prio);
3216 class = class->next;
3218 #ifdef CONFIG_PREEMPT
3220 * NEWIDLE balancing is a source of latency, so preemptible
3221 * kernels will stop after the first task is pulled to minimize
3222 * the critical section.
3224 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3227 } while (class && max_load_move > total_load_moved);
3229 return total_load_moved > 0;
3233 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3234 struct sched_domain *sd, enum cpu_idle_type idle,
3235 struct rq_iterator *iterator)
3237 struct task_struct *p = iterator->start(iterator->arg);
3241 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3242 pull_task(busiest, p, this_rq, this_cpu);
3244 * Right now, this is only the second place pull_task()
3245 * is called, so we can safely collect pull_task()
3246 * stats here rather than inside pull_task().
3248 schedstat_inc(sd, lb_gained[idle]);
3252 p = iterator->next(iterator->arg);
3259 * move_one_task tries to move exactly one task from busiest to this_rq, as
3260 * part of active balancing operations within "domain".
3261 * Returns 1 if successful and 0 otherwise.
3263 * Called with both runqueues locked.
3265 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3266 struct sched_domain *sd, enum cpu_idle_type idle)
3268 const struct sched_class *class;
3270 for (class = sched_class_highest; class; class = class->next)
3271 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3276 /********** Helpers for find_busiest_group ************************/
3278 * sd_lb_stats - Structure to store the statistics of a sched_domain
3279 * during load balancing.
3281 struct sd_lb_stats {
3282 struct sched_group *busiest; /* Busiest group in this sd */
3283 struct sched_group *this; /* Local group in this sd */
3284 unsigned long total_load; /* Total load of all groups in sd */
3285 unsigned long total_pwr; /* Total power of all groups in sd */
3286 unsigned long avg_load; /* Average load across all groups in sd */
3288 /** Statistics of this group */
3289 unsigned long this_load;
3290 unsigned long this_load_per_task;
3291 unsigned long this_nr_running;
3293 /* Statistics of the busiest group */
3294 unsigned long max_load;
3295 unsigned long busiest_load_per_task;
3296 unsigned long busiest_nr_running;
3298 int group_imb; /* Is there imbalance in this sd */
3299 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3300 int power_savings_balance; /* Is powersave balance needed for this sd */
3301 struct sched_group *group_min; /* Least loaded group in sd */
3302 struct sched_group *group_leader; /* Group which relieves group_min */
3303 unsigned long min_load_per_task; /* load_per_task in group_min */
3304 unsigned long leader_nr_running; /* Nr running of group_leader */
3305 unsigned long min_nr_running; /* Nr running of group_min */
3310 * sg_lb_stats - stats of a sched_group required for load_balancing
3312 struct sg_lb_stats {
3313 unsigned long avg_load; /*Avg load across the CPUs of the group */
3314 unsigned long group_load; /* Total load over the CPUs of the group */
3315 unsigned long sum_nr_running; /* Nr tasks running in the group */
3316 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3317 unsigned long group_capacity;
3318 int group_imb; /* Is there an imbalance in the group ? */
3322 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3323 * @group: The group whose first cpu is to be returned.
3325 static inline unsigned int group_first_cpu(struct sched_group *group)
3327 return cpumask_first(sched_group_cpus(group));
3331 * get_sd_load_idx - Obtain the load index for a given sched domain.
3332 * @sd: The sched_domain whose load_idx is to be obtained.
3333 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3335 static inline int get_sd_load_idx(struct sched_domain *sd,
3336 enum cpu_idle_type idle)
3342 load_idx = sd->busy_idx;
3345 case CPU_NEWLY_IDLE:
3346 load_idx = sd->newidle_idx;
3349 load_idx = sd->idle_idx;
3357 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3359 * init_sd_power_savings_stats - Initialize power savings statistics for
3360 * the given sched_domain, during load balancing.
3362 * @sd: Sched domain whose power-savings statistics are to be initialized.
3363 * @sds: Variable containing the statistics for sd.
3364 * @idle: Idle status of the CPU at which we're performing load-balancing.
3366 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3367 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3370 * Busy processors will not participate in power savings
3373 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3374 sds->power_savings_balance = 0;
3376 sds->power_savings_balance = 1;
3377 sds->min_nr_running = ULONG_MAX;
3378 sds->leader_nr_running = 0;
3383 * update_sd_power_savings_stats - Update the power saving stats for a
3384 * sched_domain while performing load balancing.
3386 * @group: sched_group belonging to the sched_domain under consideration.
3387 * @sds: Variable containing the statistics of the sched_domain
3388 * @local_group: Does group contain the CPU for which we're performing
3390 * @sgs: Variable containing the statistics of the group.
3392 static inline void update_sd_power_savings_stats(struct sched_group *group,
3393 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3396 if (!sds->power_savings_balance)
3400 * If the local group is idle or completely loaded
3401 * no need to do power savings balance at this domain
3403 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3404 !sds->this_nr_running))
3405 sds->power_savings_balance = 0;
3408 * If a group is already running at full capacity or idle,
3409 * don't include that group in power savings calculations
3411 if (!sds->power_savings_balance ||
3412 sgs->sum_nr_running >= sgs->group_capacity ||
3413 !sgs->sum_nr_running)
3417 * Calculate the group which has the least non-idle load.
3418 * This is the group from where we need to pick up the load
3421 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3422 (sgs->sum_nr_running == sds->min_nr_running &&
3423 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3424 sds->group_min = group;
3425 sds->min_nr_running = sgs->sum_nr_running;
3426 sds->min_load_per_task = sgs->sum_weighted_load /
3427 sgs->sum_nr_running;
3431 * Calculate the group which is almost near its
3432 * capacity but still has some space to pick up some load
3433 * from other group and save more power
3435 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3438 if (sgs->sum_nr_running > sds->leader_nr_running ||
3439 (sgs->sum_nr_running == sds->leader_nr_running &&
3440 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3441 sds->group_leader = group;
3442 sds->leader_nr_running = sgs->sum_nr_running;
3447 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3448 * @sds: Variable containing the statistics of the sched_domain
3449 * under consideration.
3450 * @this_cpu: Cpu at which we're currently performing load-balancing.
3451 * @imbalance: Variable to store the imbalance.
3454 * Check if we have potential to perform some power-savings balance.
3455 * If yes, set the busiest group to be the least loaded group in the
3456 * sched_domain, so that it's CPUs can be put to idle.
3458 * Returns 1 if there is potential to perform power-savings balance.
3461 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3462 int this_cpu, unsigned long *imbalance)
3464 if (!sds->power_savings_balance)
3467 if (sds->this != sds->group_leader ||
3468 sds->group_leader == sds->group_min)
3471 *imbalance = sds->min_load_per_task;
3472 sds->busiest = sds->group_min;
3474 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3475 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3476 group_first_cpu(sds->group_leader);
3482 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3483 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3484 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3489 static inline void update_sd_power_savings_stats(struct sched_group *group,
3490 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3495 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3496 int this_cpu, unsigned long *imbalance)
3500 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3504 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3505 * @group: sched_group whose statistics are to be updated.
3506 * @this_cpu: Cpu for which load balance is currently performed.
3507 * @idle: Idle status of this_cpu
3508 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3509 * @sd_idle: Idle status of the sched_domain containing group.
3510 * @local_group: Does group contain this_cpu.
3511 * @cpus: Set of cpus considered for load balancing.
3512 * @balance: Should we balance.
3513 * @sgs: variable to hold the statistics for this group.
3515 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3516 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3517 int local_group, const struct cpumask *cpus,
3518 int *balance, struct sg_lb_stats *sgs)
3520 unsigned long load, max_cpu_load, min_cpu_load;
3522 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3523 unsigned long sum_avg_load_per_task;
3524 unsigned long avg_load_per_task;
3527 balance_cpu = group_first_cpu(group);
3529 /* Tally up the load of all CPUs in the group */
3530 sum_avg_load_per_task = avg_load_per_task = 0;
3532 min_cpu_load = ~0UL;
3534 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3535 struct rq *rq = cpu_rq(i);
3537 if (*sd_idle && rq->nr_running)
3540 /* Bias balancing toward cpus of our domain */
3542 if (idle_cpu(i) && !first_idle_cpu) {
3547 load = target_load(i, load_idx);
3549 load = source_load(i, load_idx);
3550 if (load > max_cpu_load)
3551 max_cpu_load = load;
3552 if (min_cpu_load > load)
3553 min_cpu_load = load;
3556 sgs->group_load += load;
3557 sgs->sum_nr_running += rq->nr_running;
3558 sgs->sum_weighted_load += weighted_cpuload(i);
3560 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3564 * First idle cpu or the first cpu(busiest) in this sched group
3565 * is eligible for doing load balancing at this and above
3566 * domains. In the newly idle case, we will allow all the cpu's
3567 * to do the newly idle load balance.
3569 if (idle != CPU_NEWLY_IDLE && local_group &&
3570 balance_cpu != this_cpu && balance) {
3575 /* Adjust by relative CPU power of the group */
3576 sgs->avg_load = sg_div_cpu_power(group,
3577 sgs->group_load * SCHED_LOAD_SCALE);
3581 * Consider the group unbalanced when the imbalance is larger
3582 * than the average weight of two tasks.
3584 * APZ: with cgroup the avg task weight can vary wildly and
3585 * might not be a suitable number - should we keep a
3586 * normalized nr_running number somewhere that negates
3589 avg_load_per_task = sg_div_cpu_power(group,
3590 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3592 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3595 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3600 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3601 * @sd: sched_domain whose statistics are to be updated.
3602 * @this_cpu: Cpu for which load balance is currently performed.
3603 * @idle: Idle status of this_cpu
3604 * @sd_idle: Idle status of the sched_domain containing group.
3605 * @cpus: Set of cpus considered for load balancing.
3606 * @balance: Should we balance.
3607 * @sds: variable to hold the statistics for this sched_domain.
3609 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3610 enum cpu_idle_type idle, int *sd_idle,
3611 const struct cpumask *cpus, int *balance,
3612 struct sd_lb_stats *sds)
3614 struct sched_group *group = sd->groups;
3615 struct sg_lb_stats sgs;
3618 init_sd_power_savings_stats(sd, sds, idle);
3619 load_idx = get_sd_load_idx(sd, idle);
3624 local_group = cpumask_test_cpu(this_cpu,
3625 sched_group_cpus(group));
3626 memset(&sgs, 0, sizeof(sgs));
3627 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3628 local_group, cpus, balance, &sgs);
3630 if (local_group && balance && !(*balance))
3633 sds->total_load += sgs.group_load;
3634 sds->total_pwr += group->__cpu_power;
3637 sds->this_load = sgs.avg_load;
3639 sds->this_nr_running = sgs.sum_nr_running;
3640 sds->this_load_per_task = sgs.sum_weighted_load;
3641 } else if (sgs.avg_load > sds->max_load &&
3642 (sgs.sum_nr_running > sgs.group_capacity ||
3644 sds->max_load = sgs.avg_load;
3645 sds->busiest = group;
3646 sds->busiest_nr_running = sgs.sum_nr_running;
3647 sds->busiest_load_per_task = sgs.sum_weighted_load;
3648 sds->group_imb = sgs.group_imb;
3651 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3652 group = group->next;
3653 } while (group != sd->groups);
3658 * fix_small_imbalance - Calculate the minor imbalance that exists
3659 * amongst the groups of a sched_domain, during
3661 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3662 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3663 * @imbalance: Variable to store the imbalance.
3665 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3666 int this_cpu, unsigned long *imbalance)
3668 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3669 unsigned int imbn = 2;
3671 if (sds->this_nr_running) {
3672 sds->this_load_per_task /= sds->this_nr_running;
3673 if (sds->busiest_load_per_task >
3674 sds->this_load_per_task)
3677 sds->this_load_per_task =
3678 cpu_avg_load_per_task(this_cpu);
3680 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3681 sds->busiest_load_per_task * imbn) {
3682 *imbalance = sds->busiest_load_per_task;
3687 * OK, we don't have enough imbalance to justify moving tasks,
3688 * however we may be able to increase total CPU power used by
3692 pwr_now += sds->busiest->__cpu_power *
3693 min(sds->busiest_load_per_task, sds->max_load);
3694 pwr_now += sds->this->__cpu_power *
3695 min(sds->this_load_per_task, sds->this_load);
3696 pwr_now /= SCHED_LOAD_SCALE;
3698 /* Amount of load we'd subtract */
3699 tmp = sg_div_cpu_power(sds->busiest,
3700 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3701 if (sds->max_load > tmp)
3702 pwr_move += sds->busiest->__cpu_power *
3703 min(sds->busiest_load_per_task, sds->max_load - tmp);
3705 /* Amount of load we'd add */
3706 if (sds->max_load * sds->busiest->__cpu_power <
3707 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3708 tmp = sg_div_cpu_power(sds->this,
3709 sds->max_load * sds->busiest->__cpu_power);
3711 tmp = sg_div_cpu_power(sds->this,
3712 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3713 pwr_move += sds->this->__cpu_power *
3714 min(sds->this_load_per_task, sds->this_load + tmp);
3715 pwr_move /= SCHED_LOAD_SCALE;
3717 /* Move if we gain throughput */
3718 if (pwr_move > pwr_now)
3719 *imbalance = sds->busiest_load_per_task;
3723 * calculate_imbalance - Calculate the amount of imbalance present within the
3724 * groups of a given sched_domain during load balance.
3725 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3726 * @this_cpu: Cpu for which currently load balance is being performed.
3727 * @imbalance: The variable to store the imbalance.
3729 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3730 unsigned long *imbalance)
3732 unsigned long max_pull;
3734 * In the presence of smp nice balancing, certain scenarios can have
3735 * max load less than avg load(as we skip the groups at or below
3736 * its cpu_power, while calculating max_load..)
3738 if (sds->max_load < sds->avg_load) {
3740 return fix_small_imbalance(sds, this_cpu, imbalance);
3743 /* Don't want to pull so many tasks that a group would go idle */
3744 max_pull = min(sds->max_load - sds->avg_load,
3745 sds->max_load - sds->busiest_load_per_task);
3747 /* How much load to actually move to equalise the imbalance */
3748 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3749 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3753 * if *imbalance is less than the average load per runnable task
3754 * there is no gaurantee that any tasks will be moved so we'll have
3755 * a think about bumping its value to force at least one task to be
3758 if (*imbalance < sds->busiest_load_per_task)
3759 return fix_small_imbalance(sds, this_cpu, imbalance);
3762 /******* find_busiest_group() helpers end here *********************/
3765 * find_busiest_group - Returns the busiest group within the sched_domain
3766 * if there is an imbalance. If there isn't an imbalance, and
3767 * the user has opted for power-savings, it returns a group whose
3768 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3769 * such a group exists.
3771 * Also calculates the amount of weighted load which should be moved
3772 * to restore balance.
3774 * @sd: The sched_domain whose busiest group is to be returned.
3775 * @this_cpu: The cpu for which load balancing is currently being performed.
3776 * @imbalance: Variable which stores amount of weighted load which should
3777 * be moved to restore balance/put a group to idle.
3778 * @idle: The idle status of this_cpu.
3779 * @sd_idle: The idleness of sd
3780 * @cpus: The set of CPUs under consideration for load-balancing.
3781 * @balance: Pointer to a variable indicating if this_cpu
3782 * is the appropriate cpu to perform load balancing at this_level.
3784 * Returns: - the busiest group if imbalance exists.
3785 * - If no imbalance and user has opted for power-savings balance,
3786 * return the least loaded group whose CPUs can be
3787 * put to idle by rebalancing its tasks onto our group.
3789 static struct sched_group *
3790 find_busiest_group(struct sched_domain *sd, int this_cpu,
3791 unsigned long *imbalance, enum cpu_idle_type idle,
3792 int *sd_idle, const struct cpumask *cpus, int *balance)
3794 struct sd_lb_stats sds;
3796 memset(&sds, 0, sizeof(sds));
3799 * Compute the various statistics relavent for load balancing at
3802 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3805 /* Cases where imbalance does not exist from POV of this_cpu */
3806 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3808 * 2) There is no busy sibling group to pull from.
3809 * 3) This group is the busiest group.
3810 * 4) This group is more busy than the avg busieness at this
3812 * 5) The imbalance is within the specified limit.
3813 * 6) Any rebalance would lead to ping-pong
3815 if (balance && !(*balance))
3818 if (!sds.busiest || sds.busiest_nr_running == 0)
3821 if (sds.this_load >= sds.max_load)
3824 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3826 if (sds.this_load >= sds.avg_load)
3829 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3832 sds.busiest_load_per_task /= sds.busiest_nr_running;
3834 sds.busiest_load_per_task =
3835 min(sds.busiest_load_per_task, sds.avg_load);
3838 * We're trying to get all the cpus to the average_load, so we don't
3839 * want to push ourselves above the average load, nor do we wish to
3840 * reduce the max loaded cpu below the average load, as either of these
3841 * actions would just result in more rebalancing later, and ping-pong
3842 * tasks around. Thus we look for the minimum possible imbalance.
3843 * Negative imbalances (*we* are more loaded than anyone else) will
3844 * be counted as no imbalance for these purposes -- we can't fix that
3845 * by pulling tasks to us. Be careful of negative numbers as they'll
3846 * appear as very large values with unsigned longs.
3848 if (sds.max_load <= sds.busiest_load_per_task)
3851 /* Looks like there is an imbalance. Compute it */
3852 calculate_imbalance(&sds, this_cpu, imbalance);
3857 * There is no obvious imbalance. But check if we can do some balancing
3860 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3868 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3871 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3872 unsigned long imbalance, const struct cpumask *cpus)
3874 struct rq *busiest = NULL, *rq;
3875 unsigned long max_load = 0;
3878 for_each_cpu(i, sched_group_cpus(group)) {
3881 if (!cpumask_test_cpu(i, cpus))
3885 wl = weighted_cpuload(i);
3887 if (rq->nr_running == 1 && wl > imbalance)
3890 if (wl > max_load) {
3900 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3901 * so long as it is large enough.
3903 #define MAX_PINNED_INTERVAL 512
3905 /* Working cpumask for load_balance and load_balance_newidle. */
3906 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3909 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3910 * tasks if there is an imbalance.
3912 static int load_balance(int this_cpu, struct rq *this_rq,
3913 struct sched_domain *sd, enum cpu_idle_type idle,
3916 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3917 struct sched_group *group;
3918 unsigned long imbalance;
3920 unsigned long flags;
3921 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3923 cpumask_setall(cpus);
3926 * When power savings policy is enabled for the parent domain, idle
3927 * sibling can pick up load irrespective of busy siblings. In this case,
3928 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3929 * portraying it as CPU_NOT_IDLE.
3931 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3932 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3935 schedstat_inc(sd, lb_count[idle]);
3939 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3946 schedstat_inc(sd, lb_nobusyg[idle]);
3950 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3952 schedstat_inc(sd, lb_nobusyq[idle]);
3956 BUG_ON(busiest == this_rq);
3958 schedstat_add(sd, lb_imbalance[idle], imbalance);
3961 if (busiest->nr_running > 1) {
3963 * Attempt to move tasks. If find_busiest_group has found
3964 * an imbalance but busiest->nr_running <= 1, the group is
3965 * still unbalanced. ld_moved simply stays zero, so it is
3966 * correctly treated as an imbalance.
3968 local_irq_save(flags);
3969 double_rq_lock(this_rq, busiest);
3970 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3971 imbalance, sd, idle, &all_pinned);
3972 double_rq_unlock(this_rq, busiest);
3973 local_irq_restore(flags);
3976 * some other cpu did the load balance for us.
3978 if (ld_moved && this_cpu != smp_processor_id())
3979 resched_cpu(this_cpu);
3981 /* All tasks on this runqueue were pinned by CPU affinity */
3982 if (unlikely(all_pinned)) {
3983 cpumask_clear_cpu(cpu_of(busiest), cpus);
3984 if (!cpumask_empty(cpus))
3991 schedstat_inc(sd, lb_failed[idle]);
3992 sd->nr_balance_failed++;
3994 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3996 spin_lock_irqsave(&busiest->lock, flags);
3998 /* don't kick the migration_thread, if the curr
3999 * task on busiest cpu can't be moved to this_cpu
4001 if (!cpumask_test_cpu(this_cpu,
4002 &busiest->curr->cpus_allowed)) {
4003 spin_unlock_irqrestore(&busiest->lock, flags);
4005 goto out_one_pinned;
4008 if (!busiest->active_balance) {
4009 busiest->active_balance = 1;
4010 busiest->push_cpu = this_cpu;
4013 spin_unlock_irqrestore(&busiest->lock, flags);
4015 wake_up_process(busiest->migration_thread);
4018 * We've kicked active balancing, reset the failure
4021 sd->nr_balance_failed = sd->cache_nice_tries+1;
4024 sd->nr_balance_failed = 0;
4026 if (likely(!active_balance)) {
4027 /* We were unbalanced, so reset the balancing interval */
4028 sd->balance_interval = sd->min_interval;
4031 * If we've begun active balancing, start to back off. This
4032 * case may not be covered by the all_pinned logic if there
4033 * is only 1 task on the busy runqueue (because we don't call
4036 if (sd->balance_interval < sd->max_interval)
4037 sd->balance_interval *= 2;
4040 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4041 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4047 schedstat_inc(sd, lb_balanced[idle]);
4049 sd->nr_balance_failed = 0;
4052 /* tune up the balancing interval */
4053 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
4054 (sd->balance_interval < sd->max_interval))
4055 sd->balance_interval *= 2;
4057 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4058 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4069 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4070 * tasks if there is an imbalance.
4072 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4073 * this_rq is locked.
4076 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4078 struct sched_group *group;
4079 struct rq *busiest = NULL;
4080 unsigned long imbalance;
4084 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4086 cpumask_setall(cpus);
4089 * When power savings policy is enabled for the parent domain, idle
4090 * sibling can pick up load irrespective of busy siblings. In this case,
4091 * let the state of idle sibling percolate up as IDLE, instead of
4092 * portraying it as CPU_NOT_IDLE.
4094 if (sd->flags & SD_SHARE_CPUPOWER &&
4095 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4098 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4100 update_shares_locked(this_rq, sd);
4101 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4102 &sd_idle, cpus, NULL);
4104 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4108 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4110 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4114 BUG_ON(busiest == this_rq);
4116 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4119 if (busiest->nr_running > 1) {
4120 /* Attempt to move tasks */
4121 double_lock_balance(this_rq, busiest);
4122 /* this_rq->clock is already updated */
4123 update_rq_clock(busiest);
4124 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4125 imbalance, sd, CPU_NEWLY_IDLE,
4127 double_unlock_balance(this_rq, busiest);
4129 if (unlikely(all_pinned)) {
4130 cpumask_clear_cpu(cpu_of(busiest), cpus);
4131 if (!cpumask_empty(cpus))
4137 int active_balance = 0;
4139 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4140 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4141 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4144 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4147 if (sd->nr_balance_failed++ < 2)
4151 * The only task running in a non-idle cpu can be moved to this
4152 * cpu in an attempt to completely freeup the other CPU
4153 * package. The same method used to move task in load_balance()
4154 * have been extended for load_balance_newidle() to speedup
4155 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4157 * The package power saving logic comes from
4158 * find_busiest_group(). If there are no imbalance, then
4159 * f_b_g() will return NULL. However when sched_mc={1,2} then
4160 * f_b_g() will select a group from which a running task may be
4161 * pulled to this cpu in order to make the other package idle.
4162 * If there is no opportunity to make a package idle and if
4163 * there are no imbalance, then f_b_g() will return NULL and no
4164 * action will be taken in load_balance_newidle().
4166 * Under normal task pull operation due to imbalance, there
4167 * will be more than one task in the source run queue and
4168 * move_tasks() will succeed. ld_moved will be true and this
4169 * active balance code will not be triggered.
4172 /* Lock busiest in correct order while this_rq is held */
4173 double_lock_balance(this_rq, busiest);
4176 * don't kick the migration_thread, if the curr
4177 * task on busiest cpu can't be moved to this_cpu
4179 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4180 double_unlock_balance(this_rq, busiest);
4185 if (!busiest->active_balance) {
4186 busiest->active_balance = 1;
4187 busiest->push_cpu = this_cpu;
4191 double_unlock_balance(this_rq, busiest);
4193 * Should not call ttwu while holding a rq->lock
4195 spin_unlock(&this_rq->lock);
4197 wake_up_process(busiest->migration_thread);
4198 spin_lock(&this_rq->lock);
4201 sd->nr_balance_failed = 0;
4203 update_shares_locked(this_rq, sd);
4207 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4208 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4209 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4211 sd->nr_balance_failed = 0;
4217 * idle_balance is called by schedule() if this_cpu is about to become
4218 * idle. Attempts to pull tasks from other CPUs.
4220 static void idle_balance(int this_cpu, struct rq *this_rq)
4222 struct sched_domain *sd;
4223 int pulled_task = 0;
4224 unsigned long next_balance = jiffies + HZ;
4226 for_each_domain(this_cpu, sd) {
4227 unsigned long interval;
4229 if (!(sd->flags & SD_LOAD_BALANCE))
4232 if (sd->flags & SD_BALANCE_NEWIDLE)
4233 /* If we've pulled tasks over stop searching: */
4234 pulled_task = load_balance_newidle(this_cpu, this_rq,
4237 interval = msecs_to_jiffies(sd->balance_interval);
4238 if (time_after(next_balance, sd->last_balance + interval))
4239 next_balance = sd->last_balance + interval;
4243 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4245 * We are going idle. next_balance may be set based on
4246 * a busy processor. So reset next_balance.
4248 this_rq->next_balance = next_balance;
4253 * active_load_balance is run by migration threads. It pushes running tasks
4254 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4255 * running on each physical CPU where possible, and avoids physical /
4256 * logical imbalances.
4258 * Called with busiest_rq locked.
4260 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4262 int target_cpu = busiest_rq->push_cpu;
4263 struct sched_domain *sd;
4264 struct rq *target_rq;
4266 /* Is there any task to move? */
4267 if (busiest_rq->nr_running <= 1)
4270 target_rq = cpu_rq(target_cpu);
4273 * This condition is "impossible", if it occurs
4274 * we need to fix it. Originally reported by
4275 * Bjorn Helgaas on a 128-cpu setup.
4277 BUG_ON(busiest_rq == target_rq);
4279 /* move a task from busiest_rq to target_rq */
4280 double_lock_balance(busiest_rq, target_rq);
4281 update_rq_clock(busiest_rq);
4282 update_rq_clock(target_rq);
4284 /* Search for an sd spanning us and the target CPU. */
4285 for_each_domain(target_cpu, sd) {
4286 if ((sd->flags & SD_LOAD_BALANCE) &&
4287 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4292 schedstat_inc(sd, alb_count);
4294 if (move_one_task(target_rq, target_cpu, busiest_rq,
4296 schedstat_inc(sd, alb_pushed);
4298 schedstat_inc(sd, alb_failed);
4300 double_unlock_balance(busiest_rq, target_rq);
4305 atomic_t load_balancer;
4306 cpumask_var_t cpu_mask;
4307 cpumask_var_t ilb_grp_nohz_mask;
4308 } nohz ____cacheline_aligned = {
4309 .load_balancer = ATOMIC_INIT(-1),
4312 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4314 * lowest_flag_domain - Return lowest sched_domain containing flag.
4315 * @cpu: The cpu whose lowest level of sched domain is to
4317 * @flag: The flag to check for the lowest sched_domain
4318 * for the given cpu.
4320 * Returns the lowest sched_domain of a cpu which contains the given flag.
4322 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4324 struct sched_domain *sd;
4326 for_each_domain(cpu, sd)
4327 if (sd && (sd->flags & flag))
4334 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4335 * @cpu: The cpu whose domains we're iterating over.
4336 * @sd: variable holding the value of the power_savings_sd
4338 * @flag: The flag to filter the sched_domains to be iterated.
4340 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4341 * set, starting from the lowest sched_domain to the highest.
4343 #define for_each_flag_domain(cpu, sd, flag) \
4344 for (sd = lowest_flag_domain(cpu, flag); \
4345 (sd && (sd->flags & flag)); sd = sd->parent)
4348 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4349 * @ilb_group: group to be checked for semi-idleness
4351 * Returns: 1 if the group is semi-idle. 0 otherwise.
4353 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4354 * and atleast one non-idle CPU. This helper function checks if the given
4355 * sched_group is semi-idle or not.
4357 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4359 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4360 sched_group_cpus(ilb_group));
4363 * A sched_group is semi-idle when it has atleast one busy cpu
4364 * and atleast one idle cpu.
4366 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4369 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4375 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4376 * @cpu: The cpu which is nominating a new idle_load_balancer.
4378 * Returns: Returns the id of the idle load balancer if it exists,
4379 * Else, returns >= nr_cpu_ids.
4381 * This algorithm picks the idle load balancer such that it belongs to a
4382 * semi-idle powersavings sched_domain. The idea is to try and avoid
4383 * completely idle packages/cores just for the purpose of idle load balancing
4384 * when there are other idle cpu's which are better suited for that job.
4386 static int find_new_ilb(int cpu)
4388 struct sched_domain *sd;
4389 struct sched_group *ilb_group;
4392 * Have idle load balancer selection from semi-idle packages only
4393 * when power-aware load balancing is enabled
4395 if (!(sched_smt_power_savings || sched_mc_power_savings))
4399 * Optimize for the case when we have no idle CPUs or only one
4400 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4402 if (cpumask_weight(nohz.cpu_mask) < 2)
4405 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4406 ilb_group = sd->groups;
4409 if (is_semi_idle_group(ilb_group))
4410 return cpumask_first(nohz.ilb_grp_nohz_mask);
4412 ilb_group = ilb_group->next;
4414 } while (ilb_group != sd->groups);
4418 return cpumask_first(nohz.cpu_mask);
4420 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4421 static inline int find_new_ilb(int call_cpu)
4423 return cpumask_first(nohz.cpu_mask);
4428 * This routine will try to nominate the ilb (idle load balancing)
4429 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4430 * load balancing on behalf of all those cpus. If all the cpus in the system
4431 * go into this tickless mode, then there will be no ilb owner (as there is
4432 * no need for one) and all the cpus will sleep till the next wakeup event
4435 * For the ilb owner, tick is not stopped. And this tick will be used
4436 * for idle load balancing. ilb owner will still be part of
4439 * While stopping the tick, this cpu will become the ilb owner if there
4440 * is no other owner. And will be the owner till that cpu becomes busy
4441 * or if all cpus in the system stop their ticks at which point
4442 * there is no need for ilb owner.
4444 * When the ilb owner becomes busy, it nominates another owner, during the
4445 * next busy scheduler_tick()
4447 int select_nohz_load_balancer(int stop_tick)
4449 int cpu = smp_processor_id();
4452 cpu_rq(cpu)->in_nohz_recently = 1;
4454 if (!cpu_active(cpu)) {
4455 if (atomic_read(&nohz.load_balancer) != cpu)
4459 * If we are going offline and still the leader,
4462 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4468 cpumask_set_cpu(cpu, nohz.cpu_mask);
4470 /* time for ilb owner also to sleep */
4471 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4472 if (atomic_read(&nohz.load_balancer) == cpu)
4473 atomic_set(&nohz.load_balancer, -1);
4477 if (atomic_read(&nohz.load_balancer) == -1) {
4478 /* make me the ilb owner */
4479 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4481 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4484 if (!(sched_smt_power_savings ||
4485 sched_mc_power_savings))
4488 * Check to see if there is a more power-efficient
4491 new_ilb = find_new_ilb(cpu);
4492 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4493 atomic_set(&nohz.load_balancer, -1);
4494 resched_cpu(new_ilb);
4500 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4503 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4505 if (atomic_read(&nohz.load_balancer) == cpu)
4506 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4513 static DEFINE_SPINLOCK(balancing);
4516 * It checks each scheduling domain to see if it is due to be balanced,
4517 * and initiates a balancing operation if so.
4519 * Balancing parameters are set up in arch_init_sched_domains.
4521 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4524 struct rq *rq = cpu_rq(cpu);
4525 unsigned long interval;
4526 struct sched_domain *sd;
4527 /* Earliest time when we have to do rebalance again */
4528 unsigned long next_balance = jiffies + 60*HZ;
4529 int update_next_balance = 0;
4532 for_each_domain(cpu, sd) {
4533 if (!(sd->flags & SD_LOAD_BALANCE))
4536 interval = sd->balance_interval;
4537 if (idle != CPU_IDLE)
4538 interval *= sd->busy_factor;
4540 /* scale ms to jiffies */
4541 interval = msecs_to_jiffies(interval);
4542 if (unlikely(!interval))
4544 if (interval > HZ*NR_CPUS/10)
4545 interval = HZ*NR_CPUS/10;
4547 need_serialize = sd->flags & SD_SERIALIZE;
4549 if (need_serialize) {
4550 if (!spin_trylock(&balancing))
4554 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4555 if (load_balance(cpu, rq, sd, idle, &balance)) {
4557 * We've pulled tasks over so either we're no
4558 * longer idle, or one of our SMT siblings is
4561 idle = CPU_NOT_IDLE;
4563 sd->last_balance = jiffies;
4566 spin_unlock(&balancing);
4568 if (time_after(next_balance, sd->last_balance + interval)) {
4569 next_balance = sd->last_balance + interval;
4570 update_next_balance = 1;
4574 * Stop the load balance at this level. There is another
4575 * CPU in our sched group which is doing load balancing more
4583 * next_balance will be updated only when there is a need.
4584 * When the cpu is attached to null domain for ex, it will not be
4587 if (likely(update_next_balance))
4588 rq->next_balance = next_balance;
4592 * run_rebalance_domains is triggered when needed from the scheduler tick.
4593 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4594 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4596 static void run_rebalance_domains(struct softirq_action *h)
4598 int this_cpu = smp_processor_id();
4599 struct rq *this_rq = cpu_rq(this_cpu);
4600 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4601 CPU_IDLE : CPU_NOT_IDLE;
4603 rebalance_domains(this_cpu, idle);
4607 * If this cpu is the owner for idle load balancing, then do the
4608 * balancing on behalf of the other idle cpus whose ticks are
4611 if (this_rq->idle_at_tick &&
4612 atomic_read(&nohz.load_balancer) == this_cpu) {
4616 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4617 if (balance_cpu == this_cpu)
4621 * If this cpu gets work to do, stop the load balancing
4622 * work being done for other cpus. Next load
4623 * balancing owner will pick it up.
4628 rebalance_domains(balance_cpu, CPU_IDLE);
4630 rq = cpu_rq(balance_cpu);
4631 if (time_after(this_rq->next_balance, rq->next_balance))
4632 this_rq->next_balance = rq->next_balance;
4638 static inline int on_null_domain(int cpu)
4640 return !rcu_dereference(cpu_rq(cpu)->sd);
4644 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4646 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4647 * idle load balancing owner or decide to stop the periodic load balancing,
4648 * if the whole system is idle.
4650 static inline void trigger_load_balance(struct rq *rq, int cpu)
4654 * If we were in the nohz mode recently and busy at the current
4655 * scheduler tick, then check if we need to nominate new idle
4658 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4659 rq->in_nohz_recently = 0;
4661 if (atomic_read(&nohz.load_balancer) == cpu) {
4662 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4663 atomic_set(&nohz.load_balancer, -1);
4666 if (atomic_read(&nohz.load_balancer) == -1) {
4667 int ilb = find_new_ilb(cpu);
4669 if (ilb < nr_cpu_ids)
4675 * If this cpu is idle and doing idle load balancing for all the
4676 * cpus with ticks stopped, is it time for that to stop?
4678 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4679 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4685 * If this cpu is idle and the idle load balancing is done by
4686 * someone else, then no need raise the SCHED_SOFTIRQ
4688 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4689 cpumask_test_cpu(cpu, nohz.cpu_mask))
4692 /* Don't need to rebalance while attached to NULL domain */
4693 if (time_after_eq(jiffies, rq->next_balance) &&
4694 likely(!on_null_domain(cpu)))
4695 raise_softirq(SCHED_SOFTIRQ);
4698 #else /* CONFIG_SMP */
4701 * on UP we do not need to balance between CPUs:
4703 static inline void idle_balance(int cpu, struct rq *rq)
4709 DEFINE_PER_CPU(struct kernel_stat, kstat);
4711 EXPORT_PER_CPU_SYMBOL(kstat);
4714 * Return any ns on the sched_clock that have not yet been accounted in
4715 * @p in case that task is currently running.
4717 * Called with task_rq_lock() held on @rq.
4719 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4723 if (task_current(rq, p)) {
4724 update_rq_clock(rq);
4725 ns = rq->clock - p->se.exec_start;
4733 unsigned long long task_delta_exec(struct task_struct *p)
4735 unsigned long flags;
4739 rq = task_rq_lock(p, &flags);
4740 ns = do_task_delta_exec(p, rq);
4741 task_rq_unlock(rq, &flags);
4747 * Return accounted runtime for the task.
4748 * In case the task is currently running, return the runtime plus current's
4749 * pending runtime that have not been accounted yet.
4751 unsigned long long task_sched_runtime(struct task_struct *p)
4753 unsigned long flags;
4757 rq = task_rq_lock(p, &flags);
4758 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4759 task_rq_unlock(rq, &flags);
4765 * Return sum_exec_runtime for the thread group.
4766 * In case the task is currently running, return the sum plus current's
4767 * pending runtime that have not been accounted yet.
4769 * Note that the thread group might have other running tasks as well,
4770 * so the return value not includes other pending runtime that other
4771 * running tasks might have.
4773 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4775 struct task_cputime totals;
4776 unsigned long flags;
4780 rq = task_rq_lock(p, &flags);
4781 thread_group_cputime(p, &totals);
4782 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4783 task_rq_unlock(rq, &flags);
4789 * Account user cpu time to a process.
4790 * @p: the process that the cpu time gets accounted to
4791 * @cputime: the cpu time spent in user space since the last update
4792 * @cputime_scaled: cputime scaled by cpu frequency
4794 void account_user_time(struct task_struct *p, cputime_t cputime,
4795 cputime_t cputime_scaled)
4797 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4800 /* Add user time to process. */
4801 p->utime = cputime_add(p->utime, cputime);
4802 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4803 account_group_user_time(p, cputime);
4805 /* Add user time to cpustat. */
4806 tmp = cputime_to_cputime64(cputime);
4807 if (TASK_NICE(p) > 0)
4808 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4810 cpustat->user = cputime64_add(cpustat->user, tmp);
4812 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4813 /* Account for user time used */
4814 acct_update_integrals(p);
4818 * Account guest cpu time to a process.
4819 * @p: the process that the cpu time gets accounted to
4820 * @cputime: the cpu time spent in virtual machine since the last update
4821 * @cputime_scaled: cputime scaled by cpu frequency
4823 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4824 cputime_t cputime_scaled)
4827 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4829 tmp = cputime_to_cputime64(cputime);
4831 /* Add guest time to process. */
4832 p->utime = cputime_add(p->utime, cputime);
4833 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4834 account_group_user_time(p, cputime);
4835 p->gtime = cputime_add(p->gtime, cputime);
4837 /* Add guest time to cpustat. */
4838 cpustat->user = cputime64_add(cpustat->user, tmp);
4839 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4843 * Account system cpu time to a process.
4844 * @p: the process that the cpu time gets accounted to
4845 * @hardirq_offset: the offset to subtract from hardirq_count()
4846 * @cputime: the cpu time spent in kernel space since the last update
4847 * @cputime_scaled: cputime scaled by cpu frequency
4849 void account_system_time(struct task_struct *p, int hardirq_offset,
4850 cputime_t cputime, cputime_t cputime_scaled)
4852 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4855 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4856 account_guest_time(p, cputime, cputime_scaled);
4860 /* Add system time to process. */
4861 p->stime = cputime_add(p->stime, cputime);
4862 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4863 account_group_system_time(p, cputime);
4865 /* Add system time to cpustat. */
4866 tmp = cputime_to_cputime64(cputime);
4867 if (hardirq_count() - hardirq_offset)
4868 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4869 else if (softirq_count())
4870 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4872 cpustat->system = cputime64_add(cpustat->system, tmp);
4874 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4876 /* Account for system time used */
4877 acct_update_integrals(p);
4881 * Account for involuntary wait time.
4882 * @steal: the cpu time spent in involuntary wait
4884 void account_steal_time(cputime_t cputime)
4886 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4887 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4889 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4893 * Account for idle time.
4894 * @cputime: the cpu time spent in idle wait
4896 void account_idle_time(cputime_t cputime)
4898 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4899 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4900 struct rq *rq = this_rq();
4902 if (atomic_read(&rq->nr_iowait) > 0)
4903 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4905 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4908 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4911 * Account a single tick of cpu time.
4912 * @p: the process that the cpu time gets accounted to
4913 * @user_tick: indicates if the tick is a user or a system tick
4915 void account_process_tick(struct task_struct *p, int user_tick)
4917 cputime_t one_jiffy = jiffies_to_cputime(1);
4918 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4919 struct rq *rq = this_rq();
4922 account_user_time(p, one_jiffy, one_jiffy_scaled);
4923 else if ((p != rq->idle) || (irq_count() != HARDIRQ_OFFSET))
4924 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4927 account_idle_time(one_jiffy);
4931 * Account multiple ticks of steal time.
4932 * @p: the process from which the cpu time has been stolen
4933 * @ticks: number of stolen ticks
4935 void account_steal_ticks(unsigned long ticks)
4937 account_steal_time(jiffies_to_cputime(ticks));
4941 * Account multiple ticks of idle time.
4942 * @ticks: number of stolen ticks
4944 void account_idle_ticks(unsigned long ticks)
4946 account_idle_time(jiffies_to_cputime(ticks));
4952 * Use precise platform statistics if available:
4954 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4955 cputime_t task_utime(struct task_struct *p)
4960 cputime_t task_stime(struct task_struct *p)
4965 cputime_t task_utime(struct task_struct *p)
4967 clock_t utime = cputime_to_clock_t(p->utime),
4968 total = utime + cputime_to_clock_t(p->stime);
4972 * Use CFS's precise accounting:
4974 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4978 do_div(temp, total);
4980 utime = (clock_t)temp;
4982 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4983 return p->prev_utime;
4986 cputime_t task_stime(struct task_struct *p)
4991 * Use CFS's precise accounting. (we subtract utime from
4992 * the total, to make sure the total observed by userspace
4993 * grows monotonically - apps rely on that):
4995 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4996 cputime_to_clock_t(task_utime(p));
4999 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
5001 return p->prev_stime;
5005 inline cputime_t task_gtime(struct task_struct *p)
5011 * This function gets called by the timer code, with HZ frequency.
5012 * We call it with interrupts disabled.
5014 * It also gets called by the fork code, when changing the parent's
5017 void scheduler_tick(void)
5019 int cpu = smp_processor_id();
5020 struct rq *rq = cpu_rq(cpu);
5021 struct task_struct *curr = rq->curr;
5025 spin_lock(&rq->lock);
5026 update_rq_clock(rq);
5027 update_cpu_load(rq);
5028 curr->sched_class->task_tick(rq, curr, 0);
5029 spin_unlock(&rq->lock);
5032 rq->idle_at_tick = idle_cpu(cpu);
5033 trigger_load_balance(rq, cpu);
5037 notrace unsigned long get_parent_ip(unsigned long addr)
5039 if (in_lock_functions(addr)) {
5040 addr = CALLER_ADDR2;
5041 if (in_lock_functions(addr))
5042 addr = CALLER_ADDR3;
5047 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5048 defined(CONFIG_PREEMPT_TRACER))
5050 void __kprobes add_preempt_count(int val)
5052 #ifdef CONFIG_DEBUG_PREEMPT
5056 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5059 preempt_count() += val;
5060 #ifdef CONFIG_DEBUG_PREEMPT
5062 * Spinlock count overflowing soon?
5064 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5067 if (preempt_count() == val)
5068 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5070 EXPORT_SYMBOL(add_preempt_count);
5072 void __kprobes sub_preempt_count(int val)
5074 #ifdef CONFIG_DEBUG_PREEMPT
5078 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5081 * Is the spinlock portion underflowing?
5083 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5084 !(preempt_count() & PREEMPT_MASK)))
5088 if (preempt_count() == val)
5089 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5090 preempt_count() -= val;
5092 EXPORT_SYMBOL(sub_preempt_count);
5097 * Print scheduling while atomic bug:
5099 static noinline void __schedule_bug(struct task_struct *prev)
5101 struct pt_regs *regs = get_irq_regs();
5103 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5104 prev->comm, prev->pid, preempt_count());
5106 debug_show_held_locks(prev);
5108 if (irqs_disabled())
5109 print_irqtrace_events(prev);
5118 * Various schedule()-time debugging checks and statistics:
5120 static inline void schedule_debug(struct task_struct *prev)
5123 * Test if we are atomic. Since do_exit() needs to call into
5124 * schedule() atomically, we ignore that path for now.
5125 * Otherwise, whine if we are scheduling when we should not be.
5127 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5128 __schedule_bug(prev);
5130 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5132 schedstat_inc(this_rq(), sched_count);
5133 #ifdef CONFIG_SCHEDSTATS
5134 if (unlikely(prev->lock_depth >= 0)) {
5135 schedstat_inc(this_rq(), bkl_count);
5136 schedstat_inc(prev, sched_info.bkl_count);
5141 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5143 if (prev->state == TASK_RUNNING) {
5144 u64 runtime = prev->se.sum_exec_runtime;
5146 runtime -= prev->se.prev_sum_exec_runtime;
5147 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5150 * In order to avoid avg_overlap growing stale when we are
5151 * indeed overlapping and hence not getting put to sleep, grow
5152 * the avg_overlap on preemption.
5154 * We use the average preemption runtime because that
5155 * correlates to the amount of cache footprint a task can
5158 update_avg(&prev->se.avg_overlap, runtime);
5160 prev->sched_class->put_prev_task(rq, prev);
5164 * Pick up the highest-prio task:
5166 static inline struct task_struct *
5167 pick_next_task(struct rq *rq)
5169 const struct sched_class *class;
5170 struct task_struct *p;
5173 * Optimization: we know that if all tasks are in
5174 * the fair class we can call that function directly:
5176 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5177 p = fair_sched_class.pick_next_task(rq);
5182 class = sched_class_highest;
5184 p = class->pick_next_task(rq);
5188 * Will never be NULL as the idle class always
5189 * returns a non-NULL p:
5191 class = class->next;
5196 * schedule() is the main scheduler function.
5198 asmlinkage void __sched schedule(void)
5200 struct task_struct *prev, *next;
5201 unsigned long *switch_count;
5207 cpu = smp_processor_id();
5211 switch_count = &prev->nivcsw;
5213 release_kernel_lock(prev);
5214 need_resched_nonpreemptible:
5216 schedule_debug(prev);
5218 if (sched_feat(HRTICK))
5221 spin_lock_irq(&rq->lock);
5222 update_rq_clock(rq);
5223 clear_tsk_need_resched(prev);
5225 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5226 if (unlikely(signal_pending_state(prev->state, prev)))
5227 prev->state = TASK_RUNNING;
5229 deactivate_task(rq, prev, 1);
5230 switch_count = &prev->nvcsw;
5234 if (prev->sched_class->pre_schedule)
5235 prev->sched_class->pre_schedule(rq, prev);
5238 if (unlikely(!rq->nr_running))
5239 idle_balance(cpu, rq);
5241 put_prev_task(rq, prev);
5242 next = pick_next_task(rq);
5244 if (likely(prev != next)) {
5245 sched_info_switch(prev, next);
5251 context_switch(rq, prev, next); /* unlocks the rq */
5253 * the context switch might have flipped the stack from under
5254 * us, hence refresh the local variables.
5256 cpu = smp_processor_id();
5259 spin_unlock_irq(&rq->lock);
5261 if (unlikely(reacquire_kernel_lock(current) < 0))
5262 goto need_resched_nonpreemptible;
5264 preempt_enable_no_resched();
5268 EXPORT_SYMBOL(schedule);
5272 * Look out! "owner" is an entirely speculative pointer
5273 * access and not reliable.
5275 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5280 if (!sched_feat(OWNER_SPIN))
5283 #ifdef CONFIG_DEBUG_PAGEALLOC
5285 * Need to access the cpu field knowing that
5286 * DEBUG_PAGEALLOC could have unmapped it if
5287 * the mutex owner just released it and exited.
5289 if (probe_kernel_address(&owner->cpu, cpu))
5296 * Even if the access succeeded (likely case),
5297 * the cpu field may no longer be valid.
5299 if (cpu >= nr_cpumask_bits)
5303 * We need to validate that we can do a
5304 * get_cpu() and that we have the percpu area.
5306 if (!cpu_online(cpu))
5313 * Owner changed, break to re-assess state.
5315 if (lock->owner != owner)
5319 * Is that owner really running on that cpu?
5321 if (task_thread_info(rq->curr) != owner || need_resched())
5331 #ifdef CONFIG_PREEMPT
5333 * this is the entry point to schedule() from in-kernel preemption
5334 * off of preempt_enable. Kernel preemptions off return from interrupt
5335 * occur there and call schedule directly.
5337 asmlinkage void __sched preempt_schedule(void)
5339 struct thread_info *ti = current_thread_info();
5342 * If there is a non-zero preempt_count or interrupts are disabled,
5343 * we do not want to preempt the current task. Just return..
5345 if (likely(ti->preempt_count || irqs_disabled()))
5349 add_preempt_count(PREEMPT_ACTIVE);
5351 sub_preempt_count(PREEMPT_ACTIVE);
5354 * Check again in case we missed a preemption opportunity
5355 * between schedule and now.
5358 } while (need_resched());
5360 EXPORT_SYMBOL(preempt_schedule);
5363 * this is the entry point to schedule() from kernel preemption
5364 * off of irq context.
5365 * Note, that this is called and return with irqs disabled. This will
5366 * protect us against recursive calling from irq.
5368 asmlinkage void __sched preempt_schedule_irq(void)
5370 struct thread_info *ti = current_thread_info();
5372 /* Catch callers which need to be fixed */
5373 BUG_ON(ti->preempt_count || !irqs_disabled());
5376 add_preempt_count(PREEMPT_ACTIVE);
5379 local_irq_disable();
5380 sub_preempt_count(PREEMPT_ACTIVE);
5383 * Check again in case we missed a preemption opportunity
5384 * between schedule and now.
5387 } while (need_resched());
5390 #endif /* CONFIG_PREEMPT */
5392 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5395 return try_to_wake_up(curr->private, mode, sync);
5397 EXPORT_SYMBOL(default_wake_function);
5400 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5401 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5402 * number) then we wake all the non-exclusive tasks and one exclusive task.
5404 * There are circumstances in which we can try to wake a task which has already
5405 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5406 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5408 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5409 int nr_exclusive, int sync, void *key)
5411 wait_queue_t *curr, *next;
5413 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5414 unsigned flags = curr->flags;
5416 if (curr->func(curr, mode, sync, key) &&
5417 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5423 * __wake_up - wake up threads blocked on a waitqueue.
5425 * @mode: which threads
5426 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5427 * @key: is directly passed to the wakeup function
5429 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5430 int nr_exclusive, void *key)
5432 unsigned long flags;
5434 spin_lock_irqsave(&q->lock, flags);
5435 __wake_up_common(q, mode, nr_exclusive, 0, key);
5436 spin_unlock_irqrestore(&q->lock, flags);
5438 EXPORT_SYMBOL(__wake_up);
5441 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5443 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5445 __wake_up_common(q, mode, 1, 0, NULL);
5448 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5450 __wake_up_common(q, mode, 1, 0, key);
5454 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5456 * @mode: which threads
5457 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5458 * @key: opaque value to be passed to wakeup targets
5460 * The sync wakeup differs that the waker knows that it will schedule
5461 * away soon, so while the target thread will be woken up, it will not
5462 * be migrated to another CPU - ie. the two threads are 'synchronized'
5463 * with each other. This can prevent needless bouncing between CPUs.
5465 * On UP it can prevent extra preemption.
5467 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5468 int nr_exclusive, void *key)
5470 unsigned long flags;
5476 if (unlikely(!nr_exclusive))
5479 spin_lock_irqsave(&q->lock, flags);
5480 __wake_up_common(q, mode, nr_exclusive, sync, key);
5481 spin_unlock_irqrestore(&q->lock, flags);
5483 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5486 * __wake_up_sync - see __wake_up_sync_key()
5488 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5490 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5492 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5495 * complete: - signals a single thread waiting on this completion
5496 * @x: holds the state of this particular completion
5498 * This will wake up a single thread waiting on this completion. Threads will be
5499 * awakened in the same order in which they were queued.
5501 * See also complete_all(), wait_for_completion() and related routines.
5503 void complete(struct completion *x)
5505 unsigned long flags;
5507 spin_lock_irqsave(&x->wait.lock, flags);
5509 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5510 spin_unlock_irqrestore(&x->wait.lock, flags);
5512 EXPORT_SYMBOL(complete);
5515 * complete_all: - signals all threads waiting on this completion
5516 * @x: holds the state of this particular completion
5518 * This will wake up all threads waiting on this particular completion event.
5520 void complete_all(struct completion *x)
5522 unsigned long flags;
5524 spin_lock_irqsave(&x->wait.lock, flags);
5525 x->done += UINT_MAX/2;
5526 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5527 spin_unlock_irqrestore(&x->wait.lock, flags);
5529 EXPORT_SYMBOL(complete_all);
5531 static inline long __sched
5532 do_wait_for_common(struct completion *x, long timeout, int state)
5535 DECLARE_WAITQUEUE(wait, current);
5537 wait.flags |= WQ_FLAG_EXCLUSIVE;
5538 __add_wait_queue_tail(&x->wait, &wait);
5540 if (signal_pending_state(state, current)) {
5541 timeout = -ERESTARTSYS;
5544 __set_current_state(state);
5545 spin_unlock_irq(&x->wait.lock);
5546 timeout = schedule_timeout(timeout);
5547 spin_lock_irq(&x->wait.lock);
5548 } while (!x->done && timeout);
5549 __remove_wait_queue(&x->wait, &wait);
5554 return timeout ?: 1;
5558 wait_for_common(struct completion *x, long timeout, int state)
5562 spin_lock_irq(&x->wait.lock);
5563 timeout = do_wait_for_common(x, timeout, state);
5564 spin_unlock_irq(&x->wait.lock);
5569 * wait_for_completion: - waits for completion of a task
5570 * @x: holds the state of this particular completion
5572 * This waits to be signaled for completion of a specific task. It is NOT
5573 * interruptible and there is no timeout.
5575 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5576 * and interrupt capability. Also see complete().
5578 void __sched wait_for_completion(struct completion *x)
5580 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5582 EXPORT_SYMBOL(wait_for_completion);
5585 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5586 * @x: holds the state of this particular completion
5587 * @timeout: timeout value in jiffies
5589 * This waits for either a completion of a specific task to be signaled or for a
5590 * specified timeout to expire. The timeout is in jiffies. It is not
5593 unsigned long __sched
5594 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5596 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5598 EXPORT_SYMBOL(wait_for_completion_timeout);
5601 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5602 * @x: holds the state of this particular completion
5604 * This waits for completion of a specific task to be signaled. It is
5607 int __sched wait_for_completion_interruptible(struct completion *x)
5609 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5610 if (t == -ERESTARTSYS)
5614 EXPORT_SYMBOL(wait_for_completion_interruptible);
5617 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5618 * @x: holds the state of this particular completion
5619 * @timeout: timeout value in jiffies
5621 * This waits for either a completion of a specific task to be signaled or for a
5622 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5624 unsigned long __sched
5625 wait_for_completion_interruptible_timeout(struct completion *x,
5626 unsigned long timeout)
5628 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5630 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5633 * wait_for_completion_killable: - waits for completion of a task (killable)
5634 * @x: holds the state of this particular completion
5636 * This waits to be signaled for completion of a specific task. It can be
5637 * interrupted by a kill signal.
5639 int __sched wait_for_completion_killable(struct completion *x)
5641 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5642 if (t == -ERESTARTSYS)
5646 EXPORT_SYMBOL(wait_for_completion_killable);
5649 * try_wait_for_completion - try to decrement a completion without blocking
5650 * @x: completion structure
5652 * Returns: 0 if a decrement cannot be done without blocking
5653 * 1 if a decrement succeeded.
5655 * If a completion is being used as a counting completion,
5656 * attempt to decrement the counter without blocking. This
5657 * enables us to avoid waiting if the resource the completion
5658 * is protecting is not available.
5660 bool try_wait_for_completion(struct completion *x)
5664 spin_lock_irq(&x->wait.lock);
5669 spin_unlock_irq(&x->wait.lock);
5672 EXPORT_SYMBOL(try_wait_for_completion);
5675 * completion_done - Test to see if a completion has any waiters
5676 * @x: completion structure
5678 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5679 * 1 if there are no waiters.
5682 bool completion_done(struct completion *x)
5686 spin_lock_irq(&x->wait.lock);
5689 spin_unlock_irq(&x->wait.lock);
5692 EXPORT_SYMBOL(completion_done);
5695 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5697 unsigned long flags;
5700 init_waitqueue_entry(&wait, current);
5702 __set_current_state(state);
5704 spin_lock_irqsave(&q->lock, flags);
5705 __add_wait_queue(q, &wait);
5706 spin_unlock(&q->lock);
5707 timeout = schedule_timeout(timeout);
5708 spin_lock_irq(&q->lock);
5709 __remove_wait_queue(q, &wait);
5710 spin_unlock_irqrestore(&q->lock, flags);
5715 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5717 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5719 EXPORT_SYMBOL(interruptible_sleep_on);
5722 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5724 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5726 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5728 void __sched sleep_on(wait_queue_head_t *q)
5730 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5732 EXPORT_SYMBOL(sleep_on);
5734 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5736 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5738 EXPORT_SYMBOL(sleep_on_timeout);
5740 #ifdef CONFIG_RT_MUTEXES
5743 * rt_mutex_setprio - set the current priority of a task
5745 * @prio: prio value (kernel-internal form)
5747 * This function changes the 'effective' priority of a task. It does
5748 * not touch ->normal_prio like __setscheduler().
5750 * Used by the rt_mutex code to implement priority inheritance logic.
5752 void rt_mutex_setprio(struct task_struct *p, int prio)
5754 unsigned long flags;
5755 int oldprio, on_rq, running;
5757 const struct sched_class *prev_class = p->sched_class;
5759 BUG_ON(prio < 0 || prio > MAX_PRIO);
5761 rq = task_rq_lock(p, &flags);
5762 update_rq_clock(rq);
5765 on_rq = p->se.on_rq;
5766 running = task_current(rq, p);
5768 dequeue_task(rq, p, 0);
5770 p->sched_class->put_prev_task(rq, p);
5773 p->sched_class = &rt_sched_class;
5775 p->sched_class = &fair_sched_class;
5780 p->sched_class->set_curr_task(rq);
5782 enqueue_task(rq, p, 0);
5784 check_class_changed(rq, p, prev_class, oldprio, running);
5786 task_rq_unlock(rq, &flags);
5791 void set_user_nice(struct task_struct *p, long nice)
5793 int old_prio, delta, on_rq;
5794 unsigned long flags;
5797 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5800 * We have to be careful, if called from sys_setpriority(),
5801 * the task might be in the middle of scheduling on another CPU.
5803 rq = task_rq_lock(p, &flags);
5804 update_rq_clock(rq);
5806 * The RT priorities are set via sched_setscheduler(), but we still
5807 * allow the 'normal' nice value to be set - but as expected
5808 * it wont have any effect on scheduling until the task is
5809 * SCHED_FIFO/SCHED_RR:
5811 if (task_has_rt_policy(p)) {
5812 p->static_prio = NICE_TO_PRIO(nice);
5815 on_rq = p->se.on_rq;
5817 dequeue_task(rq, p, 0);
5819 p->static_prio = NICE_TO_PRIO(nice);
5822 p->prio = effective_prio(p);
5823 delta = p->prio - old_prio;
5826 enqueue_task(rq, p, 0);
5828 * If the task increased its priority or is running and
5829 * lowered its priority, then reschedule its CPU:
5831 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5832 resched_task(rq->curr);
5835 task_rq_unlock(rq, &flags);
5837 EXPORT_SYMBOL(set_user_nice);
5840 * can_nice - check if a task can reduce its nice value
5844 int can_nice(const struct task_struct *p, const int nice)
5846 /* convert nice value [19,-20] to rlimit style value [1,40] */
5847 int nice_rlim = 20 - nice;
5849 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5850 capable(CAP_SYS_NICE));
5853 #ifdef __ARCH_WANT_SYS_NICE
5856 * sys_nice - change the priority of the current process.
5857 * @increment: priority increment
5859 * sys_setpriority is a more generic, but much slower function that
5860 * does similar things.
5862 SYSCALL_DEFINE1(nice, int, increment)
5867 * Setpriority might change our priority at the same moment.
5868 * We don't have to worry. Conceptually one call occurs first
5869 * and we have a single winner.
5871 if (increment < -40)
5876 nice = TASK_NICE(current) + increment;
5882 if (increment < 0 && !can_nice(current, nice))
5885 retval = security_task_setnice(current, nice);
5889 set_user_nice(current, nice);
5896 * task_prio - return the priority value of a given task.
5897 * @p: the task in question.
5899 * This is the priority value as seen by users in /proc.
5900 * RT tasks are offset by -200. Normal tasks are centered
5901 * around 0, value goes from -16 to +15.
5903 int task_prio(const struct task_struct *p)
5905 return p->prio - MAX_RT_PRIO;
5909 * task_nice - return the nice value of a given task.
5910 * @p: the task in question.
5912 int task_nice(const struct task_struct *p)
5914 return TASK_NICE(p);
5916 EXPORT_SYMBOL(task_nice);
5919 * idle_cpu - is a given cpu idle currently?
5920 * @cpu: the processor in question.
5922 int idle_cpu(int cpu)
5924 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5928 * idle_task - return the idle task for a given cpu.
5929 * @cpu: the processor in question.
5931 struct task_struct *idle_task(int cpu)
5933 return cpu_rq(cpu)->idle;
5937 * find_process_by_pid - find a process with a matching PID value.
5938 * @pid: the pid in question.
5940 static struct task_struct *find_process_by_pid(pid_t pid)
5942 return pid ? find_task_by_vpid(pid) : current;
5945 /* Actually do priority change: must hold rq lock. */
5947 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5949 BUG_ON(p->se.on_rq);
5952 switch (p->policy) {
5956 p->sched_class = &fair_sched_class;
5960 p->sched_class = &rt_sched_class;
5964 p->rt_priority = prio;
5965 p->normal_prio = normal_prio(p);
5966 /* we are holding p->pi_lock already */
5967 p->prio = rt_mutex_getprio(p);
5972 * check the target process has a UID that matches the current process's
5974 static bool check_same_owner(struct task_struct *p)
5976 const struct cred *cred = current_cred(), *pcred;
5980 pcred = __task_cred(p);
5981 match = (cred->euid == pcred->euid ||
5982 cred->euid == pcred->uid);
5987 static int __sched_setscheduler(struct task_struct *p, int policy,
5988 struct sched_param *param, bool user)
5990 int retval, oldprio, oldpolicy = -1, on_rq, running;
5991 unsigned long flags;
5992 const struct sched_class *prev_class = p->sched_class;
5995 /* may grab non-irq protected spin_locks */
5996 BUG_ON(in_interrupt());
5998 /* double check policy once rq lock held */
6000 policy = oldpolicy = p->policy;
6001 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
6002 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
6003 policy != SCHED_IDLE)
6006 * Valid priorities for SCHED_FIFO and SCHED_RR are
6007 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6008 * SCHED_BATCH and SCHED_IDLE is 0.
6010 if (param->sched_priority < 0 ||
6011 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
6012 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
6014 if (rt_policy(policy) != (param->sched_priority != 0))
6018 * Allow unprivileged RT tasks to decrease priority:
6020 if (user && !capable(CAP_SYS_NICE)) {
6021 if (rt_policy(policy)) {
6022 unsigned long rlim_rtprio;
6024 if (!lock_task_sighand(p, &flags))
6026 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
6027 unlock_task_sighand(p, &flags);
6029 /* can't set/change the rt policy */
6030 if (policy != p->policy && !rlim_rtprio)
6033 /* can't increase priority */
6034 if (param->sched_priority > p->rt_priority &&
6035 param->sched_priority > rlim_rtprio)
6039 * Like positive nice levels, dont allow tasks to
6040 * move out of SCHED_IDLE either:
6042 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
6045 /* can't change other user's priorities */
6046 if (!check_same_owner(p))
6051 #ifdef CONFIG_RT_GROUP_SCHED
6053 * Do not allow realtime tasks into groups that have no runtime
6056 if (rt_bandwidth_enabled() && rt_policy(policy) &&
6057 task_group(p)->rt_bandwidth.rt_runtime == 0)
6061 retval = security_task_setscheduler(p, policy, param);
6067 * make sure no PI-waiters arrive (or leave) while we are
6068 * changing the priority of the task:
6070 spin_lock_irqsave(&p->pi_lock, flags);
6072 * To be able to change p->policy safely, the apropriate
6073 * runqueue lock must be held.
6075 rq = __task_rq_lock(p);
6076 /* recheck policy now with rq lock held */
6077 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6078 policy = oldpolicy = -1;
6079 __task_rq_unlock(rq);
6080 spin_unlock_irqrestore(&p->pi_lock, flags);
6083 update_rq_clock(rq);
6084 on_rq = p->se.on_rq;
6085 running = task_current(rq, p);
6087 deactivate_task(rq, p, 0);
6089 p->sched_class->put_prev_task(rq, p);
6092 __setscheduler(rq, p, policy, param->sched_priority);
6095 p->sched_class->set_curr_task(rq);
6097 activate_task(rq, p, 0);
6099 check_class_changed(rq, p, prev_class, oldprio, running);
6101 __task_rq_unlock(rq);
6102 spin_unlock_irqrestore(&p->pi_lock, flags);
6104 rt_mutex_adjust_pi(p);
6110 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6111 * @p: the task in question.
6112 * @policy: new policy.
6113 * @param: structure containing the new RT priority.
6115 * NOTE that the task may be already dead.
6117 int sched_setscheduler(struct task_struct *p, int policy,
6118 struct sched_param *param)
6120 return __sched_setscheduler(p, policy, param, true);
6122 EXPORT_SYMBOL_GPL(sched_setscheduler);
6125 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6126 * @p: the task in question.
6127 * @policy: new policy.
6128 * @param: structure containing the new RT priority.
6130 * Just like sched_setscheduler, only don't bother checking if the
6131 * current context has permission. For example, this is needed in
6132 * stop_machine(): we create temporary high priority worker threads,
6133 * but our caller might not have that capability.
6135 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6136 struct sched_param *param)
6138 return __sched_setscheduler(p, policy, param, false);
6142 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6144 struct sched_param lparam;
6145 struct task_struct *p;
6148 if (!param || pid < 0)
6150 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6155 p = find_process_by_pid(pid);
6157 retval = sched_setscheduler(p, policy, &lparam);
6164 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6165 * @pid: the pid in question.
6166 * @policy: new policy.
6167 * @param: structure containing the new RT priority.
6169 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6170 struct sched_param __user *, param)
6172 /* negative values for policy are not valid */
6176 return do_sched_setscheduler(pid, policy, param);
6180 * sys_sched_setparam - set/change the RT priority of a thread
6181 * @pid: the pid in question.
6182 * @param: structure containing the new RT priority.
6184 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6186 return do_sched_setscheduler(pid, -1, param);
6190 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6191 * @pid: the pid in question.
6193 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6195 struct task_struct *p;
6202 read_lock(&tasklist_lock);
6203 p = find_process_by_pid(pid);
6205 retval = security_task_getscheduler(p);
6209 read_unlock(&tasklist_lock);
6214 * sys_sched_getscheduler - get the RT priority of a thread
6215 * @pid: the pid in question.
6216 * @param: structure containing the RT priority.
6218 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6220 struct sched_param lp;
6221 struct task_struct *p;
6224 if (!param || pid < 0)
6227 read_lock(&tasklist_lock);
6228 p = find_process_by_pid(pid);
6233 retval = security_task_getscheduler(p);
6237 lp.sched_priority = p->rt_priority;
6238 read_unlock(&tasklist_lock);
6241 * This one might sleep, we cannot do it with a spinlock held ...
6243 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6248 read_unlock(&tasklist_lock);
6252 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6254 cpumask_var_t cpus_allowed, new_mask;
6255 struct task_struct *p;
6259 read_lock(&tasklist_lock);
6261 p = find_process_by_pid(pid);
6263 read_unlock(&tasklist_lock);
6269 * It is not safe to call set_cpus_allowed with the
6270 * tasklist_lock held. We will bump the task_struct's
6271 * usage count and then drop tasklist_lock.
6274 read_unlock(&tasklist_lock);
6276 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6280 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6282 goto out_free_cpus_allowed;
6285 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6288 retval = security_task_setscheduler(p, 0, NULL);
6292 cpuset_cpus_allowed(p, cpus_allowed);
6293 cpumask_and(new_mask, in_mask, cpus_allowed);
6295 retval = set_cpus_allowed_ptr(p, new_mask);
6298 cpuset_cpus_allowed(p, cpus_allowed);
6299 if (!cpumask_subset(new_mask, cpus_allowed)) {
6301 * We must have raced with a concurrent cpuset
6302 * update. Just reset the cpus_allowed to the
6303 * cpuset's cpus_allowed
6305 cpumask_copy(new_mask, cpus_allowed);
6310 free_cpumask_var(new_mask);
6311 out_free_cpus_allowed:
6312 free_cpumask_var(cpus_allowed);
6319 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6320 struct cpumask *new_mask)
6322 if (len < cpumask_size())
6323 cpumask_clear(new_mask);
6324 else if (len > cpumask_size())
6325 len = cpumask_size();
6327 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6331 * sys_sched_setaffinity - set the cpu affinity of a process
6332 * @pid: pid of the process
6333 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6334 * @user_mask_ptr: user-space pointer to the new cpu mask
6336 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6337 unsigned long __user *, user_mask_ptr)
6339 cpumask_var_t new_mask;
6342 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6345 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6347 retval = sched_setaffinity(pid, new_mask);
6348 free_cpumask_var(new_mask);
6352 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6354 struct task_struct *p;
6358 read_lock(&tasklist_lock);
6361 p = find_process_by_pid(pid);
6365 retval = security_task_getscheduler(p);
6369 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6372 read_unlock(&tasklist_lock);
6379 * sys_sched_getaffinity - get the cpu affinity of a process
6380 * @pid: pid of the process
6381 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6382 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6384 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6385 unsigned long __user *, user_mask_ptr)
6390 if (len < cpumask_size())
6393 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6396 ret = sched_getaffinity(pid, mask);
6398 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6401 ret = cpumask_size();
6403 free_cpumask_var(mask);
6409 * sys_sched_yield - yield the current processor to other threads.
6411 * This function yields the current CPU to other tasks. If there are no
6412 * other threads running on this CPU then this function will return.
6414 SYSCALL_DEFINE0(sched_yield)
6416 struct rq *rq = this_rq_lock();
6418 schedstat_inc(rq, yld_count);
6419 current->sched_class->yield_task(rq);
6422 * Since we are going to call schedule() anyway, there's
6423 * no need to preempt or enable interrupts:
6425 __release(rq->lock);
6426 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6427 _raw_spin_unlock(&rq->lock);
6428 preempt_enable_no_resched();
6435 static void __cond_resched(void)
6437 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6438 __might_sleep(__FILE__, __LINE__);
6441 * The BKS might be reacquired before we have dropped
6442 * PREEMPT_ACTIVE, which could trigger a second
6443 * cond_resched() call.
6446 add_preempt_count(PREEMPT_ACTIVE);
6448 sub_preempt_count(PREEMPT_ACTIVE);
6449 } while (need_resched());
6452 int __sched _cond_resched(void)
6454 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6455 system_state == SYSTEM_RUNNING) {
6461 EXPORT_SYMBOL(_cond_resched);
6464 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6465 * call schedule, and on return reacquire the lock.
6467 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6468 * operations here to prevent schedule() from being called twice (once via
6469 * spin_unlock(), once by hand).
6471 int cond_resched_lock(spinlock_t *lock)
6473 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6476 if (spin_needbreak(lock) || resched) {
6478 if (resched && need_resched())
6487 EXPORT_SYMBOL(cond_resched_lock);
6489 int __sched cond_resched_softirq(void)
6491 BUG_ON(!in_softirq());
6493 if (need_resched() && system_state == SYSTEM_RUNNING) {
6501 EXPORT_SYMBOL(cond_resched_softirq);
6504 * yield - yield the current processor to other threads.
6506 * This is a shortcut for kernel-space yielding - it marks the
6507 * thread runnable and calls sys_sched_yield().
6509 void __sched yield(void)
6511 set_current_state(TASK_RUNNING);
6514 EXPORT_SYMBOL(yield);
6517 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6518 * that process accounting knows that this is a task in IO wait state.
6520 * But don't do that if it is a deliberate, throttling IO wait (this task
6521 * has set its backing_dev_info: the queue against which it should throttle)
6523 void __sched io_schedule(void)
6525 struct rq *rq = &__raw_get_cpu_var(runqueues);
6527 delayacct_blkio_start();
6528 atomic_inc(&rq->nr_iowait);
6530 atomic_dec(&rq->nr_iowait);
6531 delayacct_blkio_end();
6533 EXPORT_SYMBOL(io_schedule);
6535 long __sched io_schedule_timeout(long timeout)
6537 struct rq *rq = &__raw_get_cpu_var(runqueues);
6540 delayacct_blkio_start();
6541 atomic_inc(&rq->nr_iowait);
6542 ret = schedule_timeout(timeout);
6543 atomic_dec(&rq->nr_iowait);
6544 delayacct_blkio_end();
6549 * sys_sched_get_priority_max - return maximum RT priority.
6550 * @policy: scheduling class.
6552 * this syscall returns the maximum rt_priority that can be used
6553 * by a given scheduling class.
6555 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6562 ret = MAX_USER_RT_PRIO-1;
6574 * sys_sched_get_priority_min - return minimum RT priority.
6575 * @policy: scheduling class.
6577 * this syscall returns the minimum rt_priority that can be used
6578 * by a given scheduling class.
6580 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6598 * sys_sched_rr_get_interval - return the default timeslice of a process.
6599 * @pid: pid of the process.
6600 * @interval: userspace pointer to the timeslice value.
6602 * this syscall writes the default timeslice value of a given process
6603 * into the user-space timespec buffer. A value of '0' means infinity.
6605 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6606 struct timespec __user *, interval)
6608 struct task_struct *p;
6609 unsigned int time_slice;
6617 read_lock(&tasklist_lock);
6618 p = find_process_by_pid(pid);
6622 retval = security_task_getscheduler(p);
6627 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6628 * tasks that are on an otherwise idle runqueue:
6631 if (p->policy == SCHED_RR) {
6632 time_slice = DEF_TIMESLICE;
6633 } else if (p->policy != SCHED_FIFO) {
6634 struct sched_entity *se = &p->se;
6635 unsigned long flags;
6638 rq = task_rq_lock(p, &flags);
6639 if (rq->cfs.load.weight)
6640 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6641 task_rq_unlock(rq, &flags);
6643 read_unlock(&tasklist_lock);
6644 jiffies_to_timespec(time_slice, &t);
6645 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6649 read_unlock(&tasklist_lock);
6653 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6655 void sched_show_task(struct task_struct *p)
6657 unsigned long free = 0;
6660 state = p->state ? __ffs(p->state) + 1 : 0;
6661 printk(KERN_INFO "%-13.13s %c", p->comm,
6662 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6663 #if BITS_PER_LONG == 32
6664 if (state == TASK_RUNNING)
6665 printk(KERN_CONT " running ");
6667 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6669 if (state == TASK_RUNNING)
6670 printk(KERN_CONT " running task ");
6672 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6674 #ifdef CONFIG_DEBUG_STACK_USAGE
6675 free = stack_not_used(p);
6677 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6678 task_pid_nr(p), task_pid_nr(p->real_parent),
6679 (unsigned long)task_thread_info(p)->flags);
6681 show_stack(p, NULL);
6684 void show_state_filter(unsigned long state_filter)
6686 struct task_struct *g, *p;
6688 #if BITS_PER_LONG == 32
6690 " task PC stack pid father\n");
6693 " task PC stack pid father\n");
6695 read_lock(&tasklist_lock);
6696 do_each_thread(g, p) {
6698 * reset the NMI-timeout, listing all files on a slow
6699 * console might take alot of time:
6701 touch_nmi_watchdog();
6702 if (!state_filter || (p->state & state_filter))
6704 } while_each_thread(g, p);
6706 touch_all_softlockup_watchdogs();
6708 #ifdef CONFIG_SCHED_DEBUG
6709 sysrq_sched_debug_show();
6711 read_unlock(&tasklist_lock);
6713 * Only show locks if all tasks are dumped:
6715 if (state_filter == -1)
6716 debug_show_all_locks();
6719 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6721 idle->sched_class = &idle_sched_class;
6725 * init_idle - set up an idle thread for a given CPU
6726 * @idle: task in question
6727 * @cpu: cpu the idle task belongs to
6729 * NOTE: this function does not set the idle thread's NEED_RESCHED
6730 * flag, to make booting more robust.
6732 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6734 struct rq *rq = cpu_rq(cpu);
6735 unsigned long flags;
6737 spin_lock_irqsave(&rq->lock, flags);
6740 idle->se.exec_start = sched_clock();
6742 idle->prio = idle->normal_prio = MAX_PRIO;
6743 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6744 __set_task_cpu(idle, cpu);
6746 rq->curr = rq->idle = idle;
6747 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6750 spin_unlock_irqrestore(&rq->lock, flags);
6752 /* Set the preempt count _outside_ the spinlocks! */
6753 #if defined(CONFIG_PREEMPT)
6754 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6756 task_thread_info(idle)->preempt_count = 0;
6759 * The idle tasks have their own, simple scheduling class:
6761 idle->sched_class = &idle_sched_class;
6762 ftrace_graph_init_task(idle);
6766 * In a system that switches off the HZ timer nohz_cpu_mask
6767 * indicates which cpus entered this state. This is used
6768 * in the rcu update to wait only for active cpus. For system
6769 * which do not switch off the HZ timer nohz_cpu_mask should
6770 * always be CPU_BITS_NONE.
6772 cpumask_var_t nohz_cpu_mask;
6775 * Increase the granularity value when there are more CPUs,
6776 * because with more CPUs the 'effective latency' as visible
6777 * to users decreases. But the relationship is not linear,
6778 * so pick a second-best guess by going with the log2 of the
6781 * This idea comes from the SD scheduler of Con Kolivas:
6783 static inline void sched_init_granularity(void)
6785 unsigned int factor = 1 + ilog2(num_online_cpus());
6786 const unsigned long limit = 200000000;
6788 sysctl_sched_min_granularity *= factor;
6789 if (sysctl_sched_min_granularity > limit)
6790 sysctl_sched_min_granularity = limit;
6792 sysctl_sched_latency *= factor;
6793 if (sysctl_sched_latency > limit)
6794 sysctl_sched_latency = limit;
6796 sysctl_sched_wakeup_granularity *= factor;
6798 sysctl_sched_shares_ratelimit *= factor;
6803 * This is how migration works:
6805 * 1) we queue a struct migration_req structure in the source CPU's
6806 * runqueue and wake up that CPU's migration thread.
6807 * 2) we down() the locked semaphore => thread blocks.
6808 * 3) migration thread wakes up (implicitly it forces the migrated
6809 * thread off the CPU)
6810 * 4) it gets the migration request and checks whether the migrated
6811 * task is still in the wrong runqueue.
6812 * 5) if it's in the wrong runqueue then the migration thread removes
6813 * it and puts it into the right queue.
6814 * 6) migration thread up()s the semaphore.
6815 * 7) we wake up and the migration is done.
6819 * Change a given task's CPU affinity. Migrate the thread to a
6820 * proper CPU and schedule it away if the CPU it's executing on
6821 * is removed from the allowed bitmask.
6823 * NOTE: the caller must have a valid reference to the task, the
6824 * task must not exit() & deallocate itself prematurely. The
6825 * call is not atomic; no spinlocks may be held.
6827 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6829 struct migration_req req;
6830 unsigned long flags;
6834 rq = task_rq_lock(p, &flags);
6835 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6840 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6841 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6846 if (p->sched_class->set_cpus_allowed)
6847 p->sched_class->set_cpus_allowed(p, new_mask);
6849 cpumask_copy(&p->cpus_allowed, new_mask);
6850 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6853 /* Can the task run on the task's current CPU? If so, we're done */
6854 if (cpumask_test_cpu(task_cpu(p), new_mask))
6857 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6858 /* Need help from migration thread: drop lock and wait. */
6859 task_rq_unlock(rq, &flags);
6860 wake_up_process(rq->migration_thread);
6861 wait_for_completion(&req.done);
6862 tlb_migrate_finish(p->mm);
6866 task_rq_unlock(rq, &flags);
6870 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6873 * Move (not current) task off this cpu, onto dest cpu. We're doing
6874 * this because either it can't run here any more (set_cpus_allowed()
6875 * away from this CPU, or CPU going down), or because we're
6876 * attempting to rebalance this task on exec (sched_exec).
6878 * So we race with normal scheduler movements, but that's OK, as long
6879 * as the task is no longer on this CPU.
6881 * Returns non-zero if task was successfully migrated.
6883 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6885 struct rq *rq_dest, *rq_src;
6888 if (unlikely(!cpu_active(dest_cpu)))
6891 rq_src = cpu_rq(src_cpu);
6892 rq_dest = cpu_rq(dest_cpu);
6894 double_rq_lock(rq_src, rq_dest);
6895 /* Already moved. */
6896 if (task_cpu(p) != src_cpu)
6898 /* Affinity changed (again). */
6899 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6902 on_rq = p->se.on_rq;
6904 deactivate_task(rq_src, p, 0);
6906 set_task_cpu(p, dest_cpu);
6908 activate_task(rq_dest, p, 0);
6909 check_preempt_curr(rq_dest, p, 0);
6914 double_rq_unlock(rq_src, rq_dest);
6919 * migration_thread - this is a highprio system thread that performs
6920 * thread migration by bumping thread off CPU then 'pushing' onto
6923 static int migration_thread(void *data)
6925 int cpu = (long)data;
6929 BUG_ON(rq->migration_thread != current);
6931 set_current_state(TASK_INTERRUPTIBLE);
6932 while (!kthread_should_stop()) {
6933 struct migration_req *req;
6934 struct list_head *head;
6936 spin_lock_irq(&rq->lock);
6938 if (cpu_is_offline(cpu)) {
6939 spin_unlock_irq(&rq->lock);
6943 if (rq->active_balance) {
6944 active_load_balance(rq, cpu);
6945 rq->active_balance = 0;
6948 head = &rq->migration_queue;
6950 if (list_empty(head)) {
6951 spin_unlock_irq(&rq->lock);
6953 set_current_state(TASK_INTERRUPTIBLE);
6956 req = list_entry(head->next, struct migration_req, list);
6957 list_del_init(head->next);
6959 spin_unlock(&rq->lock);
6960 __migrate_task(req->task, cpu, req->dest_cpu);
6963 complete(&req->done);
6965 __set_current_state(TASK_RUNNING);
6969 /* Wait for kthread_stop */
6970 set_current_state(TASK_INTERRUPTIBLE);
6971 while (!kthread_should_stop()) {
6973 set_current_state(TASK_INTERRUPTIBLE);
6975 __set_current_state(TASK_RUNNING);
6979 #ifdef CONFIG_HOTPLUG_CPU
6981 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6985 local_irq_disable();
6986 ret = __migrate_task(p, src_cpu, dest_cpu);
6992 * Figure out where task on dead CPU should go, use force if necessary.
6994 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6997 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
7000 /* Look for allowed, online CPU in same node. */
7001 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
7002 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
7005 /* Any allowed, online CPU? */
7006 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
7007 if (dest_cpu < nr_cpu_ids)
7010 /* No more Mr. Nice Guy. */
7011 if (dest_cpu >= nr_cpu_ids) {
7012 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
7013 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
7016 * Don't tell them about moving exiting tasks or
7017 * kernel threads (both mm NULL), since they never
7020 if (p->mm && printk_ratelimit()) {
7021 printk(KERN_INFO "process %d (%s) no "
7022 "longer affine to cpu%d\n",
7023 task_pid_nr(p), p->comm, dead_cpu);
7028 /* It can have affinity changed while we were choosing. */
7029 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
7034 * While a dead CPU has no uninterruptible tasks queued at this point,
7035 * it might still have a nonzero ->nr_uninterruptible counter, because
7036 * for performance reasons the counter is not stricly tracking tasks to
7037 * their home CPUs. So we just add the counter to another CPU's counter,
7038 * to keep the global sum constant after CPU-down:
7040 static void migrate_nr_uninterruptible(struct rq *rq_src)
7042 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
7043 unsigned long flags;
7045 local_irq_save(flags);
7046 double_rq_lock(rq_src, rq_dest);
7047 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
7048 rq_src->nr_uninterruptible = 0;
7049 double_rq_unlock(rq_src, rq_dest);
7050 local_irq_restore(flags);
7053 /* Run through task list and migrate tasks from the dead cpu. */
7054 static void migrate_live_tasks(int src_cpu)
7056 struct task_struct *p, *t;
7058 read_lock(&tasklist_lock);
7060 do_each_thread(t, p) {
7064 if (task_cpu(p) == src_cpu)
7065 move_task_off_dead_cpu(src_cpu, p);
7066 } while_each_thread(t, p);
7068 read_unlock(&tasklist_lock);
7072 * Schedules idle task to be the next runnable task on current CPU.
7073 * It does so by boosting its priority to highest possible.
7074 * Used by CPU offline code.
7076 void sched_idle_next(void)
7078 int this_cpu = smp_processor_id();
7079 struct rq *rq = cpu_rq(this_cpu);
7080 struct task_struct *p = rq->idle;
7081 unsigned long flags;
7083 /* cpu has to be offline */
7084 BUG_ON(cpu_online(this_cpu));
7087 * Strictly not necessary since rest of the CPUs are stopped by now
7088 * and interrupts disabled on the current cpu.
7090 spin_lock_irqsave(&rq->lock, flags);
7092 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7094 update_rq_clock(rq);
7095 activate_task(rq, p, 0);
7097 spin_unlock_irqrestore(&rq->lock, flags);
7101 * Ensures that the idle task is using init_mm right before its cpu goes
7104 void idle_task_exit(void)
7106 struct mm_struct *mm = current->active_mm;
7108 BUG_ON(cpu_online(smp_processor_id()));
7111 switch_mm(mm, &init_mm, current);
7115 /* called under rq->lock with disabled interrupts */
7116 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7118 struct rq *rq = cpu_rq(dead_cpu);
7120 /* Must be exiting, otherwise would be on tasklist. */
7121 BUG_ON(!p->exit_state);
7123 /* Cannot have done final schedule yet: would have vanished. */
7124 BUG_ON(p->state == TASK_DEAD);
7129 * Drop lock around migration; if someone else moves it,
7130 * that's OK. No task can be added to this CPU, so iteration is
7133 spin_unlock_irq(&rq->lock);
7134 move_task_off_dead_cpu(dead_cpu, p);
7135 spin_lock_irq(&rq->lock);
7140 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7141 static void migrate_dead_tasks(unsigned int dead_cpu)
7143 struct rq *rq = cpu_rq(dead_cpu);
7144 struct task_struct *next;
7147 if (!rq->nr_running)
7149 update_rq_clock(rq);
7150 next = pick_next_task(rq);
7153 next->sched_class->put_prev_task(rq, next);
7154 migrate_dead(dead_cpu, next);
7160 * remove the tasks which were accounted by rq from calc_load_tasks.
7162 static void calc_global_load_remove(struct rq *rq)
7164 atomic_long_sub(rq->calc_load_active, &calc_load_tasks);
7166 #endif /* CONFIG_HOTPLUG_CPU */
7168 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7170 static struct ctl_table sd_ctl_dir[] = {
7172 .procname = "sched_domain",
7178 static struct ctl_table sd_ctl_root[] = {
7180 .ctl_name = CTL_KERN,
7181 .procname = "kernel",
7183 .child = sd_ctl_dir,
7188 static struct ctl_table *sd_alloc_ctl_entry(int n)
7190 struct ctl_table *entry =
7191 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7196 static void sd_free_ctl_entry(struct ctl_table **tablep)
7198 struct ctl_table *entry;
7201 * In the intermediate directories, both the child directory and
7202 * procname are dynamically allocated and could fail but the mode
7203 * will always be set. In the lowest directory the names are
7204 * static strings and all have proc handlers.
7206 for (entry = *tablep; entry->mode; entry++) {
7208 sd_free_ctl_entry(&entry->child);
7209 if (entry->proc_handler == NULL)
7210 kfree(entry->procname);
7218 set_table_entry(struct ctl_table *entry,
7219 const char *procname, void *data, int maxlen,
7220 mode_t mode, proc_handler *proc_handler)
7222 entry->procname = procname;
7224 entry->maxlen = maxlen;
7226 entry->proc_handler = proc_handler;
7229 static struct ctl_table *
7230 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7232 struct ctl_table *table = sd_alloc_ctl_entry(13);
7237 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7238 sizeof(long), 0644, proc_doulongvec_minmax);
7239 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7240 sizeof(long), 0644, proc_doulongvec_minmax);
7241 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7242 sizeof(int), 0644, proc_dointvec_minmax);
7243 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7244 sizeof(int), 0644, proc_dointvec_minmax);
7245 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7246 sizeof(int), 0644, proc_dointvec_minmax);
7247 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7248 sizeof(int), 0644, proc_dointvec_minmax);
7249 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7250 sizeof(int), 0644, proc_dointvec_minmax);
7251 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7252 sizeof(int), 0644, proc_dointvec_minmax);
7253 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7254 sizeof(int), 0644, proc_dointvec_minmax);
7255 set_table_entry(&table[9], "cache_nice_tries",
7256 &sd->cache_nice_tries,
7257 sizeof(int), 0644, proc_dointvec_minmax);
7258 set_table_entry(&table[10], "flags", &sd->flags,
7259 sizeof(int), 0644, proc_dointvec_minmax);
7260 set_table_entry(&table[11], "name", sd->name,
7261 CORENAME_MAX_SIZE, 0444, proc_dostring);
7262 /* &table[12] is terminator */
7267 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7269 struct ctl_table *entry, *table;
7270 struct sched_domain *sd;
7271 int domain_num = 0, i;
7274 for_each_domain(cpu, sd)
7276 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7281 for_each_domain(cpu, sd) {
7282 snprintf(buf, 32, "domain%d", i);
7283 entry->procname = kstrdup(buf, GFP_KERNEL);
7285 entry->child = sd_alloc_ctl_domain_table(sd);
7292 static struct ctl_table_header *sd_sysctl_header;
7293 static void register_sched_domain_sysctl(void)
7295 int i, cpu_num = num_online_cpus();
7296 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7299 WARN_ON(sd_ctl_dir[0].child);
7300 sd_ctl_dir[0].child = entry;
7305 for_each_online_cpu(i) {
7306 snprintf(buf, 32, "cpu%d", i);
7307 entry->procname = kstrdup(buf, GFP_KERNEL);
7309 entry->child = sd_alloc_ctl_cpu_table(i);
7313 WARN_ON(sd_sysctl_header);
7314 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7317 /* may be called multiple times per register */
7318 static void unregister_sched_domain_sysctl(void)
7320 if (sd_sysctl_header)
7321 unregister_sysctl_table(sd_sysctl_header);
7322 sd_sysctl_header = NULL;
7323 if (sd_ctl_dir[0].child)
7324 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7327 static void register_sched_domain_sysctl(void)
7330 static void unregister_sched_domain_sysctl(void)
7335 static void set_rq_online(struct rq *rq)
7338 const struct sched_class *class;
7340 cpumask_set_cpu(rq->cpu, rq->rd->online);
7343 for_each_class(class) {
7344 if (class->rq_online)
7345 class->rq_online(rq);
7350 static void set_rq_offline(struct rq *rq)
7353 const struct sched_class *class;
7355 for_each_class(class) {
7356 if (class->rq_offline)
7357 class->rq_offline(rq);
7360 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7366 * migration_call - callback that gets triggered when a CPU is added.
7367 * Here we can start up the necessary migration thread for the new CPU.
7369 static int __cpuinit
7370 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7372 struct task_struct *p;
7373 int cpu = (long)hcpu;
7374 unsigned long flags;
7379 case CPU_UP_PREPARE:
7380 case CPU_UP_PREPARE_FROZEN:
7381 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7384 kthread_bind(p, cpu);
7385 /* Must be high prio: stop_machine expects to yield to it. */
7386 rq = task_rq_lock(p, &flags);
7387 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7388 task_rq_unlock(rq, &flags);
7389 cpu_rq(cpu)->migration_thread = p;
7393 case CPU_ONLINE_FROZEN:
7394 /* Strictly unnecessary, as first user will wake it. */
7395 wake_up_process(cpu_rq(cpu)->migration_thread);
7397 /* Update our root-domain */
7399 spin_lock_irqsave(&rq->lock, flags);
7400 rq->calc_load_update = calc_load_update;
7401 rq->calc_load_active = 0;
7403 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7407 spin_unlock_irqrestore(&rq->lock, flags);
7410 #ifdef CONFIG_HOTPLUG_CPU
7411 case CPU_UP_CANCELED:
7412 case CPU_UP_CANCELED_FROZEN:
7413 if (!cpu_rq(cpu)->migration_thread)
7415 /* Unbind it from offline cpu so it can run. Fall thru. */
7416 kthread_bind(cpu_rq(cpu)->migration_thread,
7417 cpumask_any(cpu_online_mask));
7418 kthread_stop(cpu_rq(cpu)->migration_thread);
7419 cpu_rq(cpu)->migration_thread = NULL;
7423 case CPU_DEAD_FROZEN:
7424 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7425 migrate_live_tasks(cpu);
7427 kthread_stop(rq->migration_thread);
7428 rq->migration_thread = NULL;
7429 /* Idle task back to normal (off runqueue, low prio) */
7430 spin_lock_irq(&rq->lock);
7431 update_rq_clock(rq);
7432 deactivate_task(rq, rq->idle, 0);
7433 rq->idle->static_prio = MAX_PRIO;
7434 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7435 rq->idle->sched_class = &idle_sched_class;
7436 migrate_dead_tasks(cpu);
7437 spin_unlock_irq(&rq->lock);
7439 migrate_nr_uninterruptible(rq);
7440 BUG_ON(rq->nr_running != 0);
7441 calc_global_load_remove(rq);
7443 * No need to migrate the tasks: it was best-effort if
7444 * they didn't take sched_hotcpu_mutex. Just wake up
7447 spin_lock_irq(&rq->lock);
7448 while (!list_empty(&rq->migration_queue)) {
7449 struct migration_req *req;
7451 req = list_entry(rq->migration_queue.next,
7452 struct migration_req, list);
7453 list_del_init(&req->list);
7454 spin_unlock_irq(&rq->lock);
7455 complete(&req->done);
7456 spin_lock_irq(&rq->lock);
7458 spin_unlock_irq(&rq->lock);
7462 case CPU_DYING_FROZEN:
7463 /* Update our root-domain */
7465 spin_lock_irqsave(&rq->lock, flags);
7467 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7470 spin_unlock_irqrestore(&rq->lock, flags);
7477 /* Register at highest priority so that task migration (migrate_all_tasks)
7478 * happens before everything else.
7480 static struct notifier_block __cpuinitdata migration_notifier = {
7481 .notifier_call = migration_call,
7485 static int __init migration_init(void)
7487 void *cpu = (void *)(long)smp_processor_id();
7490 /* Start one for the boot CPU: */
7491 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7492 BUG_ON(err == NOTIFY_BAD);
7493 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7494 register_cpu_notifier(&migration_notifier);
7498 early_initcall(migration_init);
7503 #ifdef CONFIG_SCHED_DEBUG
7505 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7506 struct cpumask *groupmask)
7508 struct sched_group *group = sd->groups;
7511 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7512 cpumask_clear(groupmask);
7514 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7516 if (!(sd->flags & SD_LOAD_BALANCE)) {
7517 printk("does not load-balance\n");
7519 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7524 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7526 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7527 printk(KERN_ERR "ERROR: domain->span does not contain "
7530 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7531 printk(KERN_ERR "ERROR: domain->groups does not contain"
7535 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7539 printk(KERN_ERR "ERROR: group is NULL\n");
7543 if (!group->__cpu_power) {
7544 printk(KERN_CONT "\n");
7545 printk(KERN_ERR "ERROR: domain->cpu_power not "
7550 if (!cpumask_weight(sched_group_cpus(group))) {
7551 printk(KERN_CONT "\n");
7552 printk(KERN_ERR "ERROR: empty group\n");
7556 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7557 printk(KERN_CONT "\n");
7558 printk(KERN_ERR "ERROR: repeated CPUs\n");
7562 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7564 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7566 printk(KERN_CONT " %s", str);
7567 if (group->__cpu_power != SCHED_LOAD_SCALE) {
7568 printk(KERN_CONT " (__cpu_power = %d)",
7569 group->__cpu_power);
7572 group = group->next;
7573 } while (group != sd->groups);
7574 printk(KERN_CONT "\n");
7576 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7577 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7580 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7581 printk(KERN_ERR "ERROR: parent span is not a superset "
7582 "of domain->span\n");
7586 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7588 cpumask_var_t groupmask;
7592 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7596 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7598 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7599 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7604 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7611 free_cpumask_var(groupmask);
7613 #else /* !CONFIG_SCHED_DEBUG */
7614 # define sched_domain_debug(sd, cpu) do { } while (0)
7615 #endif /* CONFIG_SCHED_DEBUG */
7617 static int sd_degenerate(struct sched_domain *sd)
7619 if (cpumask_weight(sched_domain_span(sd)) == 1)
7622 /* Following flags need at least 2 groups */
7623 if (sd->flags & (SD_LOAD_BALANCE |
7624 SD_BALANCE_NEWIDLE |
7628 SD_SHARE_PKG_RESOURCES)) {
7629 if (sd->groups != sd->groups->next)
7633 /* Following flags don't use groups */
7634 if (sd->flags & (SD_WAKE_IDLE |
7643 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7645 unsigned long cflags = sd->flags, pflags = parent->flags;
7647 if (sd_degenerate(parent))
7650 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7653 /* Does parent contain flags not in child? */
7654 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7655 if (cflags & SD_WAKE_AFFINE)
7656 pflags &= ~SD_WAKE_BALANCE;
7657 /* Flags needing groups don't count if only 1 group in parent */
7658 if (parent->groups == parent->groups->next) {
7659 pflags &= ~(SD_LOAD_BALANCE |
7660 SD_BALANCE_NEWIDLE |
7664 SD_SHARE_PKG_RESOURCES);
7665 if (nr_node_ids == 1)
7666 pflags &= ~SD_SERIALIZE;
7668 if (~cflags & pflags)
7674 static void free_rootdomain(struct root_domain *rd)
7676 cpupri_cleanup(&rd->cpupri);
7678 free_cpumask_var(rd->rto_mask);
7679 free_cpumask_var(rd->online);
7680 free_cpumask_var(rd->span);
7684 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7686 struct root_domain *old_rd = NULL;
7687 unsigned long flags;
7689 spin_lock_irqsave(&rq->lock, flags);
7694 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7697 cpumask_clear_cpu(rq->cpu, old_rd->span);
7700 * If we dont want to free the old_rt yet then
7701 * set old_rd to NULL to skip the freeing later
7704 if (!atomic_dec_and_test(&old_rd->refcount))
7708 atomic_inc(&rd->refcount);
7711 cpumask_set_cpu(rq->cpu, rd->span);
7712 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7715 spin_unlock_irqrestore(&rq->lock, flags);
7718 free_rootdomain(old_rd);
7721 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7723 memset(rd, 0, sizeof(*rd));
7726 alloc_bootmem_cpumask_var(&def_root_domain.span);
7727 alloc_bootmem_cpumask_var(&def_root_domain.online);
7728 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7729 cpupri_init(&rd->cpupri, true);
7733 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7735 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7737 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7740 if (cpupri_init(&rd->cpupri, false) != 0)
7745 free_cpumask_var(rd->rto_mask);
7747 free_cpumask_var(rd->online);
7749 free_cpumask_var(rd->span);
7754 static void init_defrootdomain(void)
7756 init_rootdomain(&def_root_domain, true);
7758 atomic_set(&def_root_domain.refcount, 1);
7761 static struct root_domain *alloc_rootdomain(void)
7763 struct root_domain *rd;
7765 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7769 if (init_rootdomain(rd, false) != 0) {
7778 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7779 * hold the hotplug lock.
7782 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7784 struct rq *rq = cpu_rq(cpu);
7785 struct sched_domain *tmp;
7787 /* Remove the sched domains which do not contribute to scheduling. */
7788 for (tmp = sd; tmp; ) {
7789 struct sched_domain *parent = tmp->parent;
7793 if (sd_parent_degenerate(tmp, parent)) {
7794 tmp->parent = parent->parent;
7796 parent->parent->child = tmp;
7801 if (sd && sd_degenerate(sd)) {
7807 sched_domain_debug(sd, cpu);
7809 rq_attach_root(rq, rd);
7810 rcu_assign_pointer(rq->sd, sd);
7813 /* cpus with isolated domains */
7814 static cpumask_var_t cpu_isolated_map;
7816 /* Setup the mask of cpus configured for isolated domains */
7817 static int __init isolated_cpu_setup(char *str)
7819 cpulist_parse(str, cpu_isolated_map);
7823 __setup("isolcpus=", isolated_cpu_setup);
7826 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7827 * to a function which identifies what group(along with sched group) a CPU
7828 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7829 * (due to the fact that we keep track of groups covered with a struct cpumask).
7831 * init_sched_build_groups will build a circular linked list of the groups
7832 * covered by the given span, and will set each group's ->cpumask correctly,
7833 * and ->cpu_power to 0.
7836 init_sched_build_groups(const struct cpumask *span,
7837 const struct cpumask *cpu_map,
7838 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7839 struct sched_group **sg,
7840 struct cpumask *tmpmask),
7841 struct cpumask *covered, struct cpumask *tmpmask)
7843 struct sched_group *first = NULL, *last = NULL;
7846 cpumask_clear(covered);
7848 for_each_cpu(i, span) {
7849 struct sched_group *sg;
7850 int group = group_fn(i, cpu_map, &sg, tmpmask);
7853 if (cpumask_test_cpu(i, covered))
7856 cpumask_clear(sched_group_cpus(sg));
7857 sg->__cpu_power = 0;
7859 for_each_cpu(j, span) {
7860 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7863 cpumask_set_cpu(j, covered);
7864 cpumask_set_cpu(j, sched_group_cpus(sg));
7875 #define SD_NODES_PER_DOMAIN 16
7880 * find_next_best_node - find the next node to include in a sched_domain
7881 * @node: node whose sched_domain we're building
7882 * @used_nodes: nodes already in the sched_domain
7884 * Find the next node to include in a given scheduling domain. Simply
7885 * finds the closest node not already in the @used_nodes map.
7887 * Should use nodemask_t.
7889 static int find_next_best_node(int node, nodemask_t *used_nodes)
7891 int i, n, val, min_val, best_node = 0;
7895 for (i = 0; i < nr_node_ids; i++) {
7896 /* Start at @node */
7897 n = (node + i) % nr_node_ids;
7899 if (!nr_cpus_node(n))
7902 /* Skip already used nodes */
7903 if (node_isset(n, *used_nodes))
7906 /* Simple min distance search */
7907 val = node_distance(node, n);
7909 if (val < min_val) {
7915 node_set(best_node, *used_nodes);
7920 * sched_domain_node_span - get a cpumask for a node's sched_domain
7921 * @node: node whose cpumask we're constructing
7922 * @span: resulting cpumask
7924 * Given a node, construct a good cpumask for its sched_domain to span. It
7925 * should be one that prevents unnecessary balancing, but also spreads tasks
7928 static void sched_domain_node_span(int node, struct cpumask *span)
7930 nodemask_t used_nodes;
7933 cpumask_clear(span);
7934 nodes_clear(used_nodes);
7936 cpumask_or(span, span, cpumask_of_node(node));
7937 node_set(node, used_nodes);
7939 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7940 int next_node = find_next_best_node(node, &used_nodes);
7942 cpumask_or(span, span, cpumask_of_node(next_node));
7945 #endif /* CONFIG_NUMA */
7947 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7950 * The cpus mask in sched_group and sched_domain hangs off the end.
7952 * ( See the the comments in include/linux/sched.h:struct sched_group
7953 * and struct sched_domain. )
7955 struct static_sched_group {
7956 struct sched_group sg;
7957 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7960 struct static_sched_domain {
7961 struct sched_domain sd;
7962 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7966 * SMT sched-domains:
7968 #ifdef CONFIG_SCHED_SMT
7969 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7970 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7973 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7974 struct sched_group **sg, struct cpumask *unused)
7977 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7980 #endif /* CONFIG_SCHED_SMT */
7983 * multi-core sched-domains:
7985 #ifdef CONFIG_SCHED_MC
7986 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7987 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7988 #endif /* CONFIG_SCHED_MC */
7990 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7992 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7993 struct sched_group **sg, struct cpumask *mask)
7997 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7998 group = cpumask_first(mask);
8000 *sg = &per_cpu(sched_group_core, group).sg;
8003 #elif defined(CONFIG_SCHED_MC)
8005 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
8006 struct sched_group **sg, struct cpumask *unused)
8009 *sg = &per_cpu(sched_group_core, cpu).sg;
8014 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
8015 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
8018 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
8019 struct sched_group **sg, struct cpumask *mask)
8022 #ifdef CONFIG_SCHED_MC
8023 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
8024 group = cpumask_first(mask);
8025 #elif defined(CONFIG_SCHED_SMT)
8026 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
8027 group = cpumask_first(mask);
8032 *sg = &per_cpu(sched_group_phys, group).sg;
8038 * The init_sched_build_groups can't handle what we want to do with node
8039 * groups, so roll our own. Now each node has its own list of groups which
8040 * gets dynamically allocated.
8042 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
8043 static struct sched_group ***sched_group_nodes_bycpu;
8045 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
8046 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
8048 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
8049 struct sched_group **sg,
8050 struct cpumask *nodemask)
8054 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
8055 group = cpumask_first(nodemask);
8058 *sg = &per_cpu(sched_group_allnodes, group).sg;
8062 static void init_numa_sched_groups_power(struct sched_group *group_head)
8064 struct sched_group *sg = group_head;
8070 for_each_cpu(j, sched_group_cpus(sg)) {
8071 struct sched_domain *sd;
8073 sd = &per_cpu(phys_domains, j).sd;
8074 if (j != group_first_cpu(sd->groups)) {
8076 * Only add "power" once for each
8082 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8085 } while (sg != group_head);
8087 #endif /* CONFIG_NUMA */
8090 /* Free memory allocated for various sched_group structures */
8091 static void free_sched_groups(const struct cpumask *cpu_map,
8092 struct cpumask *nodemask)
8096 for_each_cpu(cpu, cpu_map) {
8097 struct sched_group **sched_group_nodes
8098 = sched_group_nodes_bycpu[cpu];
8100 if (!sched_group_nodes)
8103 for (i = 0; i < nr_node_ids; i++) {
8104 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8106 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8107 if (cpumask_empty(nodemask))
8117 if (oldsg != sched_group_nodes[i])
8120 kfree(sched_group_nodes);
8121 sched_group_nodes_bycpu[cpu] = NULL;
8124 #else /* !CONFIG_NUMA */
8125 static void free_sched_groups(const struct cpumask *cpu_map,
8126 struct cpumask *nodemask)
8129 #endif /* CONFIG_NUMA */
8132 * Initialize sched groups cpu_power.
8134 * cpu_power indicates the capacity of sched group, which is used while
8135 * distributing the load between different sched groups in a sched domain.
8136 * Typically cpu_power for all the groups in a sched domain will be same unless
8137 * there are asymmetries in the topology. If there are asymmetries, group
8138 * having more cpu_power will pickup more load compared to the group having
8141 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8142 * the maximum number of tasks a group can handle in the presence of other idle
8143 * or lightly loaded groups in the same sched domain.
8145 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8147 struct sched_domain *child;
8148 struct sched_group *group;
8150 WARN_ON(!sd || !sd->groups);
8152 if (cpu != group_first_cpu(sd->groups))
8157 sd->groups->__cpu_power = 0;
8160 * For perf policy, if the groups in child domain share resources
8161 * (for example cores sharing some portions of the cache hierarchy
8162 * or SMT), then set this domain groups cpu_power such that each group
8163 * can handle only one task, when there are other idle groups in the
8164 * same sched domain.
8166 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8168 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8169 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8174 * add cpu_power of each child group to this groups cpu_power
8176 group = child->groups;
8178 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8179 group = group->next;
8180 } while (group != child->groups);
8184 * Initializers for schedule domains
8185 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8188 #ifdef CONFIG_SCHED_DEBUG
8189 # define SD_INIT_NAME(sd, type) sd->name = #type
8191 # define SD_INIT_NAME(sd, type) do { } while (0)
8194 #define SD_INIT(sd, type) sd_init_##type(sd)
8196 #define SD_INIT_FUNC(type) \
8197 static noinline void sd_init_##type(struct sched_domain *sd) \
8199 memset(sd, 0, sizeof(*sd)); \
8200 *sd = SD_##type##_INIT; \
8201 sd->level = SD_LV_##type; \
8202 SD_INIT_NAME(sd, type); \
8207 SD_INIT_FUNC(ALLNODES)
8210 #ifdef CONFIG_SCHED_SMT
8211 SD_INIT_FUNC(SIBLING)
8213 #ifdef CONFIG_SCHED_MC
8217 static int default_relax_domain_level = -1;
8219 static int __init setup_relax_domain_level(char *str)
8223 val = simple_strtoul(str, NULL, 0);
8224 if (val < SD_LV_MAX)
8225 default_relax_domain_level = val;
8229 __setup("relax_domain_level=", setup_relax_domain_level);
8231 static void set_domain_attribute(struct sched_domain *sd,
8232 struct sched_domain_attr *attr)
8236 if (!attr || attr->relax_domain_level < 0) {
8237 if (default_relax_domain_level < 0)
8240 request = default_relax_domain_level;
8242 request = attr->relax_domain_level;
8243 if (request < sd->level) {
8244 /* turn off idle balance on this domain */
8245 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8247 /* turn on idle balance on this domain */
8248 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8253 * Build sched domains for a given set of cpus and attach the sched domains
8254 * to the individual cpus
8256 static int __build_sched_domains(const struct cpumask *cpu_map,
8257 struct sched_domain_attr *attr)
8259 int i, err = -ENOMEM;
8260 struct root_domain *rd;
8261 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8264 cpumask_var_t domainspan, covered, notcovered;
8265 struct sched_group **sched_group_nodes = NULL;
8266 int sd_allnodes = 0;
8268 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8270 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8271 goto free_domainspan;
8272 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8276 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8277 goto free_notcovered;
8278 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8280 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8281 goto free_this_sibling_map;
8282 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8283 goto free_this_core_map;
8284 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8285 goto free_send_covered;
8289 * Allocate the per-node list of sched groups
8291 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8293 if (!sched_group_nodes) {
8294 printk(KERN_WARNING "Can not alloc sched group node list\n");
8299 rd = alloc_rootdomain();
8301 printk(KERN_WARNING "Cannot alloc root domain\n");
8302 goto free_sched_groups;
8306 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8310 * Set up domains for cpus specified by the cpu_map.
8312 for_each_cpu(i, cpu_map) {
8313 struct sched_domain *sd = NULL, *p;
8315 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8318 if (cpumask_weight(cpu_map) >
8319 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8320 sd = &per_cpu(allnodes_domains, i).sd;
8321 SD_INIT(sd, ALLNODES);
8322 set_domain_attribute(sd, attr);
8323 cpumask_copy(sched_domain_span(sd), cpu_map);
8324 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8330 sd = &per_cpu(node_domains, i).sd;
8332 set_domain_attribute(sd, attr);
8333 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8337 cpumask_and(sched_domain_span(sd),
8338 sched_domain_span(sd), cpu_map);
8342 sd = &per_cpu(phys_domains, i).sd;
8344 set_domain_attribute(sd, attr);
8345 cpumask_copy(sched_domain_span(sd), nodemask);
8349 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8351 #ifdef CONFIG_SCHED_MC
8353 sd = &per_cpu(core_domains, i).sd;
8355 set_domain_attribute(sd, attr);
8356 cpumask_and(sched_domain_span(sd), cpu_map,
8357 cpu_coregroup_mask(i));
8360 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8363 #ifdef CONFIG_SCHED_SMT
8365 sd = &per_cpu(cpu_domains, i).sd;
8366 SD_INIT(sd, SIBLING);
8367 set_domain_attribute(sd, attr);
8368 cpumask_and(sched_domain_span(sd),
8369 topology_thread_cpumask(i), cpu_map);
8372 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8376 #ifdef CONFIG_SCHED_SMT
8377 /* Set up CPU (sibling) groups */
8378 for_each_cpu(i, cpu_map) {
8379 cpumask_and(this_sibling_map,
8380 topology_thread_cpumask(i), cpu_map);
8381 if (i != cpumask_first(this_sibling_map))
8384 init_sched_build_groups(this_sibling_map, cpu_map,
8386 send_covered, tmpmask);
8390 #ifdef CONFIG_SCHED_MC
8391 /* Set up multi-core groups */
8392 for_each_cpu(i, cpu_map) {
8393 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8394 if (i != cpumask_first(this_core_map))
8397 init_sched_build_groups(this_core_map, cpu_map,
8399 send_covered, tmpmask);
8403 /* Set up physical groups */
8404 for (i = 0; i < nr_node_ids; i++) {
8405 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8406 if (cpumask_empty(nodemask))
8409 init_sched_build_groups(nodemask, cpu_map,
8411 send_covered, tmpmask);
8415 /* Set up node groups */
8417 init_sched_build_groups(cpu_map, cpu_map,
8418 &cpu_to_allnodes_group,
8419 send_covered, tmpmask);
8422 for (i = 0; i < nr_node_ids; i++) {
8423 /* Set up node groups */
8424 struct sched_group *sg, *prev;
8427 cpumask_clear(covered);
8428 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8429 if (cpumask_empty(nodemask)) {
8430 sched_group_nodes[i] = NULL;
8434 sched_domain_node_span(i, domainspan);
8435 cpumask_and(domainspan, domainspan, cpu_map);
8437 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8440 printk(KERN_WARNING "Can not alloc domain group for "
8444 sched_group_nodes[i] = sg;
8445 for_each_cpu(j, nodemask) {
8446 struct sched_domain *sd;
8448 sd = &per_cpu(node_domains, j).sd;
8451 sg->__cpu_power = 0;
8452 cpumask_copy(sched_group_cpus(sg), nodemask);
8454 cpumask_or(covered, covered, nodemask);
8457 for (j = 0; j < nr_node_ids; j++) {
8458 int n = (i + j) % nr_node_ids;
8460 cpumask_complement(notcovered, covered);
8461 cpumask_and(tmpmask, notcovered, cpu_map);
8462 cpumask_and(tmpmask, tmpmask, domainspan);
8463 if (cpumask_empty(tmpmask))
8466 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8467 if (cpumask_empty(tmpmask))
8470 sg = kmalloc_node(sizeof(struct sched_group) +
8475 "Can not alloc domain group for node %d\n", j);
8478 sg->__cpu_power = 0;
8479 cpumask_copy(sched_group_cpus(sg), tmpmask);
8480 sg->next = prev->next;
8481 cpumask_or(covered, covered, tmpmask);
8488 /* Calculate CPU power for physical packages and nodes */
8489 #ifdef CONFIG_SCHED_SMT
8490 for_each_cpu(i, cpu_map) {
8491 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8493 init_sched_groups_power(i, sd);
8496 #ifdef CONFIG_SCHED_MC
8497 for_each_cpu(i, cpu_map) {
8498 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8500 init_sched_groups_power(i, sd);
8504 for_each_cpu(i, cpu_map) {
8505 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8507 init_sched_groups_power(i, sd);
8511 for (i = 0; i < nr_node_ids; i++)
8512 init_numa_sched_groups_power(sched_group_nodes[i]);
8515 struct sched_group *sg;
8517 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8519 init_numa_sched_groups_power(sg);
8523 /* Attach the domains */
8524 for_each_cpu(i, cpu_map) {
8525 struct sched_domain *sd;
8526 #ifdef CONFIG_SCHED_SMT
8527 sd = &per_cpu(cpu_domains, i).sd;
8528 #elif defined(CONFIG_SCHED_MC)
8529 sd = &per_cpu(core_domains, i).sd;
8531 sd = &per_cpu(phys_domains, i).sd;
8533 cpu_attach_domain(sd, rd, i);
8539 free_cpumask_var(tmpmask);
8541 free_cpumask_var(send_covered);
8543 free_cpumask_var(this_core_map);
8544 free_this_sibling_map:
8545 free_cpumask_var(this_sibling_map);
8547 free_cpumask_var(nodemask);
8550 free_cpumask_var(notcovered);
8552 free_cpumask_var(covered);
8554 free_cpumask_var(domainspan);
8561 kfree(sched_group_nodes);
8567 free_sched_groups(cpu_map, tmpmask);
8568 free_rootdomain(rd);
8573 static int build_sched_domains(const struct cpumask *cpu_map)
8575 return __build_sched_domains(cpu_map, NULL);
8578 static struct cpumask *doms_cur; /* current sched domains */
8579 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8580 static struct sched_domain_attr *dattr_cur;
8581 /* attribues of custom domains in 'doms_cur' */
8584 * Special case: If a kmalloc of a doms_cur partition (array of
8585 * cpumask) fails, then fallback to a single sched domain,
8586 * as determined by the single cpumask fallback_doms.
8588 static cpumask_var_t fallback_doms;
8591 * arch_update_cpu_topology lets virtualized architectures update the
8592 * cpu core maps. It is supposed to return 1 if the topology changed
8593 * or 0 if it stayed the same.
8595 int __attribute__((weak)) arch_update_cpu_topology(void)
8601 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8602 * For now this just excludes isolated cpus, but could be used to
8603 * exclude other special cases in the future.
8605 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8609 arch_update_cpu_topology();
8611 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8613 doms_cur = fallback_doms;
8614 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8616 err = build_sched_domains(doms_cur);
8617 register_sched_domain_sysctl();
8622 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8623 struct cpumask *tmpmask)
8625 free_sched_groups(cpu_map, tmpmask);
8629 * Detach sched domains from a group of cpus specified in cpu_map
8630 * These cpus will now be attached to the NULL domain
8632 static void detach_destroy_domains(const struct cpumask *cpu_map)
8634 /* Save because hotplug lock held. */
8635 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8638 for_each_cpu(i, cpu_map)
8639 cpu_attach_domain(NULL, &def_root_domain, i);
8640 synchronize_sched();
8641 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8644 /* handle null as "default" */
8645 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8646 struct sched_domain_attr *new, int idx_new)
8648 struct sched_domain_attr tmp;
8655 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8656 new ? (new + idx_new) : &tmp,
8657 sizeof(struct sched_domain_attr));
8661 * Partition sched domains as specified by the 'ndoms_new'
8662 * cpumasks in the array doms_new[] of cpumasks. This compares
8663 * doms_new[] to the current sched domain partitioning, doms_cur[].
8664 * It destroys each deleted domain and builds each new domain.
8666 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8667 * The masks don't intersect (don't overlap.) We should setup one
8668 * sched domain for each mask. CPUs not in any of the cpumasks will
8669 * not be load balanced. If the same cpumask appears both in the
8670 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8673 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8674 * ownership of it and will kfree it when done with it. If the caller
8675 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8676 * ndoms_new == 1, and partition_sched_domains() will fallback to
8677 * the single partition 'fallback_doms', it also forces the domains
8680 * If doms_new == NULL it will be replaced with cpu_online_mask.
8681 * ndoms_new == 0 is a special case for destroying existing domains,
8682 * and it will not create the default domain.
8684 * Call with hotplug lock held
8686 /* FIXME: Change to struct cpumask *doms_new[] */
8687 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8688 struct sched_domain_attr *dattr_new)
8693 mutex_lock(&sched_domains_mutex);
8695 /* always unregister in case we don't destroy any domains */
8696 unregister_sched_domain_sysctl();
8698 /* Let architecture update cpu core mappings. */
8699 new_topology = arch_update_cpu_topology();
8701 n = doms_new ? ndoms_new : 0;
8703 /* Destroy deleted domains */
8704 for (i = 0; i < ndoms_cur; i++) {
8705 for (j = 0; j < n && !new_topology; j++) {
8706 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8707 && dattrs_equal(dattr_cur, i, dattr_new, j))
8710 /* no match - a current sched domain not in new doms_new[] */
8711 detach_destroy_domains(doms_cur + i);
8716 if (doms_new == NULL) {
8718 doms_new = fallback_doms;
8719 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8720 WARN_ON_ONCE(dattr_new);
8723 /* Build new domains */
8724 for (i = 0; i < ndoms_new; i++) {
8725 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8726 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8727 && dattrs_equal(dattr_new, i, dattr_cur, j))
8730 /* no match - add a new doms_new */
8731 __build_sched_domains(doms_new + i,
8732 dattr_new ? dattr_new + i : NULL);
8737 /* Remember the new sched domains */
8738 if (doms_cur != fallback_doms)
8740 kfree(dattr_cur); /* kfree(NULL) is safe */
8741 doms_cur = doms_new;
8742 dattr_cur = dattr_new;
8743 ndoms_cur = ndoms_new;
8745 register_sched_domain_sysctl();
8747 mutex_unlock(&sched_domains_mutex);
8750 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8751 static void arch_reinit_sched_domains(void)
8755 /* Destroy domains first to force the rebuild */
8756 partition_sched_domains(0, NULL, NULL);
8758 rebuild_sched_domains();
8762 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8764 unsigned int level = 0;
8766 if (sscanf(buf, "%u", &level) != 1)
8770 * level is always be positive so don't check for
8771 * level < POWERSAVINGS_BALANCE_NONE which is 0
8772 * What happens on 0 or 1 byte write,
8773 * need to check for count as well?
8776 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8780 sched_smt_power_savings = level;
8782 sched_mc_power_savings = level;
8784 arch_reinit_sched_domains();
8789 #ifdef CONFIG_SCHED_MC
8790 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8793 return sprintf(page, "%u\n", sched_mc_power_savings);
8795 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8796 const char *buf, size_t count)
8798 return sched_power_savings_store(buf, count, 0);
8800 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8801 sched_mc_power_savings_show,
8802 sched_mc_power_savings_store);
8805 #ifdef CONFIG_SCHED_SMT
8806 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8809 return sprintf(page, "%u\n", sched_smt_power_savings);
8811 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8812 const char *buf, size_t count)
8814 return sched_power_savings_store(buf, count, 1);
8816 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8817 sched_smt_power_savings_show,
8818 sched_smt_power_savings_store);
8821 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8825 #ifdef CONFIG_SCHED_SMT
8827 err = sysfs_create_file(&cls->kset.kobj,
8828 &attr_sched_smt_power_savings.attr);
8830 #ifdef CONFIG_SCHED_MC
8831 if (!err && mc_capable())
8832 err = sysfs_create_file(&cls->kset.kobj,
8833 &attr_sched_mc_power_savings.attr);
8837 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8839 #ifndef CONFIG_CPUSETS
8841 * Add online and remove offline CPUs from the scheduler domains.
8842 * When cpusets are enabled they take over this function.
8844 static int update_sched_domains(struct notifier_block *nfb,
8845 unsigned long action, void *hcpu)
8849 case CPU_ONLINE_FROZEN:
8851 case CPU_DEAD_FROZEN:
8852 partition_sched_domains(1, NULL, NULL);
8861 static int update_runtime(struct notifier_block *nfb,
8862 unsigned long action, void *hcpu)
8864 int cpu = (int)(long)hcpu;
8867 case CPU_DOWN_PREPARE:
8868 case CPU_DOWN_PREPARE_FROZEN:
8869 disable_runtime(cpu_rq(cpu));
8872 case CPU_DOWN_FAILED:
8873 case CPU_DOWN_FAILED_FROZEN:
8875 case CPU_ONLINE_FROZEN:
8876 enable_runtime(cpu_rq(cpu));
8884 void __init sched_init_smp(void)
8886 cpumask_var_t non_isolated_cpus;
8888 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8890 #if defined(CONFIG_NUMA)
8891 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8893 BUG_ON(sched_group_nodes_bycpu == NULL);
8896 mutex_lock(&sched_domains_mutex);
8897 arch_init_sched_domains(cpu_online_mask);
8898 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8899 if (cpumask_empty(non_isolated_cpus))
8900 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8901 mutex_unlock(&sched_domains_mutex);
8904 #ifndef CONFIG_CPUSETS
8905 /* XXX: Theoretical race here - CPU may be hotplugged now */
8906 hotcpu_notifier(update_sched_domains, 0);
8909 /* RT runtime code needs to handle some hotplug events */
8910 hotcpu_notifier(update_runtime, 0);
8914 /* Move init over to a non-isolated CPU */
8915 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8917 sched_init_granularity();
8918 free_cpumask_var(non_isolated_cpus);
8920 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8921 init_sched_rt_class();
8924 void __init sched_init_smp(void)
8926 sched_init_granularity();
8928 #endif /* CONFIG_SMP */
8930 int in_sched_functions(unsigned long addr)
8932 return in_lock_functions(addr) ||
8933 (addr >= (unsigned long)__sched_text_start
8934 && addr < (unsigned long)__sched_text_end);
8937 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8939 cfs_rq->tasks_timeline = RB_ROOT;
8940 INIT_LIST_HEAD(&cfs_rq->tasks);
8941 #ifdef CONFIG_FAIR_GROUP_SCHED
8944 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8947 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8949 struct rt_prio_array *array;
8952 array = &rt_rq->active;
8953 for (i = 0; i < MAX_RT_PRIO; i++) {
8954 INIT_LIST_HEAD(array->queue + i);
8955 __clear_bit(i, array->bitmap);
8957 /* delimiter for bitsearch: */
8958 __set_bit(MAX_RT_PRIO, array->bitmap);
8960 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8961 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8963 rt_rq->highest_prio.next = MAX_RT_PRIO;
8967 rt_rq->rt_nr_migratory = 0;
8968 rt_rq->overloaded = 0;
8969 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8973 rt_rq->rt_throttled = 0;
8974 rt_rq->rt_runtime = 0;
8975 spin_lock_init(&rt_rq->rt_runtime_lock);
8977 #ifdef CONFIG_RT_GROUP_SCHED
8978 rt_rq->rt_nr_boosted = 0;
8983 #ifdef CONFIG_FAIR_GROUP_SCHED
8984 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8985 struct sched_entity *se, int cpu, int add,
8986 struct sched_entity *parent)
8988 struct rq *rq = cpu_rq(cpu);
8989 tg->cfs_rq[cpu] = cfs_rq;
8990 init_cfs_rq(cfs_rq, rq);
8993 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8996 /* se could be NULL for init_task_group */
9001 se->cfs_rq = &rq->cfs;
9003 se->cfs_rq = parent->my_q;
9006 se->load.weight = tg->shares;
9007 se->load.inv_weight = 0;
9008 se->parent = parent;
9012 #ifdef CONFIG_RT_GROUP_SCHED
9013 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
9014 struct sched_rt_entity *rt_se, int cpu, int add,
9015 struct sched_rt_entity *parent)
9017 struct rq *rq = cpu_rq(cpu);
9019 tg->rt_rq[cpu] = rt_rq;
9020 init_rt_rq(rt_rq, rq);
9022 rt_rq->rt_se = rt_se;
9023 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
9025 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
9027 tg->rt_se[cpu] = rt_se;
9032 rt_se->rt_rq = &rq->rt;
9034 rt_se->rt_rq = parent->my_q;
9036 rt_se->my_q = rt_rq;
9037 rt_se->parent = parent;
9038 INIT_LIST_HEAD(&rt_se->run_list);
9042 void __init sched_init(void)
9045 unsigned long alloc_size = 0, ptr;
9047 #ifdef CONFIG_FAIR_GROUP_SCHED
9048 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9050 #ifdef CONFIG_RT_GROUP_SCHED
9051 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
9053 #ifdef CONFIG_USER_SCHED
9056 #ifdef CONFIG_CPUMASK_OFFSTACK
9057 alloc_size += num_possible_cpus() * cpumask_size();
9060 * As sched_init() is called before page_alloc is setup,
9061 * we use alloc_bootmem().
9064 ptr = (unsigned long)alloc_bootmem(alloc_size);
9066 #ifdef CONFIG_FAIR_GROUP_SCHED
9067 init_task_group.se = (struct sched_entity **)ptr;
9068 ptr += nr_cpu_ids * sizeof(void **);
9070 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
9071 ptr += nr_cpu_ids * sizeof(void **);
9073 #ifdef CONFIG_USER_SCHED
9074 root_task_group.se = (struct sched_entity **)ptr;
9075 ptr += nr_cpu_ids * sizeof(void **);
9077 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
9078 ptr += nr_cpu_ids * sizeof(void **);
9079 #endif /* CONFIG_USER_SCHED */
9080 #endif /* CONFIG_FAIR_GROUP_SCHED */
9081 #ifdef CONFIG_RT_GROUP_SCHED
9082 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9083 ptr += nr_cpu_ids * sizeof(void **);
9085 init_task_group.rt_rq = (struct rt_rq **)ptr;
9086 ptr += nr_cpu_ids * sizeof(void **);
9088 #ifdef CONFIG_USER_SCHED
9089 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9090 ptr += nr_cpu_ids * sizeof(void **);
9092 root_task_group.rt_rq = (struct rt_rq **)ptr;
9093 ptr += nr_cpu_ids * sizeof(void **);
9094 #endif /* CONFIG_USER_SCHED */
9095 #endif /* CONFIG_RT_GROUP_SCHED */
9096 #ifdef CONFIG_CPUMASK_OFFSTACK
9097 for_each_possible_cpu(i) {
9098 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9099 ptr += cpumask_size();
9101 #endif /* CONFIG_CPUMASK_OFFSTACK */
9105 init_defrootdomain();
9108 init_rt_bandwidth(&def_rt_bandwidth,
9109 global_rt_period(), global_rt_runtime());
9111 #ifdef CONFIG_RT_GROUP_SCHED
9112 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9113 global_rt_period(), global_rt_runtime());
9114 #ifdef CONFIG_USER_SCHED
9115 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9116 global_rt_period(), RUNTIME_INF);
9117 #endif /* CONFIG_USER_SCHED */
9118 #endif /* CONFIG_RT_GROUP_SCHED */
9120 #ifdef CONFIG_GROUP_SCHED
9121 list_add(&init_task_group.list, &task_groups);
9122 INIT_LIST_HEAD(&init_task_group.children);
9124 #ifdef CONFIG_USER_SCHED
9125 INIT_LIST_HEAD(&root_task_group.children);
9126 init_task_group.parent = &root_task_group;
9127 list_add(&init_task_group.siblings, &root_task_group.children);
9128 #endif /* CONFIG_USER_SCHED */
9129 #endif /* CONFIG_GROUP_SCHED */
9131 for_each_possible_cpu(i) {
9135 spin_lock_init(&rq->lock);
9137 rq->calc_load_active = 0;
9138 rq->calc_load_update = jiffies + LOAD_FREQ;
9139 init_cfs_rq(&rq->cfs, rq);
9140 init_rt_rq(&rq->rt, rq);
9141 #ifdef CONFIG_FAIR_GROUP_SCHED
9142 init_task_group.shares = init_task_group_load;
9143 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9144 #ifdef CONFIG_CGROUP_SCHED
9146 * How much cpu bandwidth does init_task_group get?
9148 * In case of task-groups formed thr' the cgroup filesystem, it
9149 * gets 100% of the cpu resources in the system. This overall
9150 * system cpu resource is divided among the tasks of
9151 * init_task_group and its child task-groups in a fair manner,
9152 * based on each entity's (task or task-group's) weight
9153 * (se->load.weight).
9155 * In other words, if init_task_group has 10 tasks of weight
9156 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9157 * then A0's share of the cpu resource is:
9159 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9161 * We achieve this by letting init_task_group's tasks sit
9162 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9164 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9165 #elif defined CONFIG_USER_SCHED
9166 root_task_group.shares = NICE_0_LOAD;
9167 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9169 * In case of task-groups formed thr' the user id of tasks,
9170 * init_task_group represents tasks belonging to root user.
9171 * Hence it forms a sibling of all subsequent groups formed.
9172 * In this case, init_task_group gets only a fraction of overall
9173 * system cpu resource, based on the weight assigned to root
9174 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9175 * by letting tasks of init_task_group sit in a separate cfs_rq
9176 * (init_cfs_rq) and having one entity represent this group of
9177 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9179 init_tg_cfs_entry(&init_task_group,
9180 &per_cpu(init_cfs_rq, i),
9181 &per_cpu(init_sched_entity, i), i, 1,
9182 root_task_group.se[i]);
9185 #endif /* CONFIG_FAIR_GROUP_SCHED */
9187 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9188 #ifdef CONFIG_RT_GROUP_SCHED
9189 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9190 #ifdef CONFIG_CGROUP_SCHED
9191 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9192 #elif defined CONFIG_USER_SCHED
9193 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9194 init_tg_rt_entry(&init_task_group,
9195 &per_cpu(init_rt_rq, i),
9196 &per_cpu(init_sched_rt_entity, i), i, 1,
9197 root_task_group.rt_se[i]);
9201 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9202 rq->cpu_load[j] = 0;
9206 rq->active_balance = 0;
9207 rq->next_balance = jiffies;
9211 rq->migration_thread = NULL;
9212 INIT_LIST_HEAD(&rq->migration_queue);
9213 rq_attach_root(rq, &def_root_domain);
9216 atomic_set(&rq->nr_iowait, 0);
9219 set_load_weight(&init_task);
9221 #ifdef CONFIG_PREEMPT_NOTIFIERS
9222 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9226 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9229 #ifdef CONFIG_RT_MUTEXES
9230 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9234 * The boot idle thread does lazy MMU switching as well:
9236 atomic_inc(&init_mm.mm_count);
9237 enter_lazy_tlb(&init_mm, current);
9240 * Make us the idle thread. Technically, schedule() should not be
9241 * called from this thread, however somewhere below it might be,
9242 * but because we are the idle thread, we just pick up running again
9243 * when this runqueue becomes "idle".
9245 init_idle(current, smp_processor_id());
9247 calc_load_update = jiffies + LOAD_FREQ;
9250 * During early bootup we pretend to be a normal task:
9252 current->sched_class = &fair_sched_class;
9254 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9255 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
9258 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
9259 alloc_bootmem_cpumask_var(&nohz.ilb_grp_nohz_mask);
9261 alloc_bootmem_cpumask_var(&cpu_isolated_map);
9264 scheduler_running = 1;
9267 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9268 void __might_sleep(char *file, int line)
9271 static unsigned long prev_jiffy; /* ratelimiting */
9273 if ((!in_atomic() && !irqs_disabled()) ||
9274 system_state != SYSTEM_RUNNING || oops_in_progress)
9276 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9278 prev_jiffy = jiffies;
9281 "BUG: sleeping function called from invalid context at %s:%d\n",
9284 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9285 in_atomic(), irqs_disabled(),
9286 current->pid, current->comm);
9288 debug_show_held_locks(current);
9289 if (irqs_disabled())
9290 print_irqtrace_events(current);
9294 EXPORT_SYMBOL(__might_sleep);
9297 #ifdef CONFIG_MAGIC_SYSRQ
9298 static void normalize_task(struct rq *rq, struct task_struct *p)
9302 update_rq_clock(rq);
9303 on_rq = p->se.on_rq;
9305 deactivate_task(rq, p, 0);
9306 __setscheduler(rq, p, SCHED_NORMAL, 0);
9308 activate_task(rq, p, 0);
9309 resched_task(rq->curr);
9313 void normalize_rt_tasks(void)
9315 struct task_struct *g, *p;
9316 unsigned long flags;
9319 read_lock_irqsave(&tasklist_lock, flags);
9320 do_each_thread(g, p) {
9322 * Only normalize user tasks:
9327 p->se.exec_start = 0;
9328 #ifdef CONFIG_SCHEDSTATS
9329 p->se.wait_start = 0;
9330 p->se.sleep_start = 0;
9331 p->se.block_start = 0;
9336 * Renice negative nice level userspace
9339 if (TASK_NICE(p) < 0 && p->mm)
9340 set_user_nice(p, 0);
9344 spin_lock(&p->pi_lock);
9345 rq = __task_rq_lock(p);
9347 normalize_task(rq, p);
9349 __task_rq_unlock(rq);
9350 spin_unlock(&p->pi_lock);
9351 } while_each_thread(g, p);
9353 read_unlock_irqrestore(&tasklist_lock, flags);
9356 #endif /* CONFIG_MAGIC_SYSRQ */
9360 * These functions are only useful for the IA64 MCA handling.
9362 * They can only be called when the whole system has been
9363 * stopped - every CPU needs to be quiescent, and no scheduling
9364 * activity can take place. Using them for anything else would
9365 * be a serious bug, and as a result, they aren't even visible
9366 * under any other configuration.
9370 * curr_task - return the current task for a given cpu.
9371 * @cpu: the processor in question.
9373 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9375 struct task_struct *curr_task(int cpu)
9377 return cpu_curr(cpu);
9381 * set_curr_task - set the current task for a given cpu.
9382 * @cpu: the processor in question.
9383 * @p: the task pointer to set.
9385 * Description: This function must only be used when non-maskable interrupts
9386 * are serviced on a separate stack. It allows the architecture to switch the
9387 * notion of the current task on a cpu in a non-blocking manner. This function
9388 * must be called with all CPU's synchronized, and interrupts disabled, the
9389 * and caller must save the original value of the current task (see
9390 * curr_task() above) and restore that value before reenabling interrupts and
9391 * re-starting the system.
9393 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9395 void set_curr_task(int cpu, struct task_struct *p)
9402 #ifdef CONFIG_FAIR_GROUP_SCHED
9403 static void free_fair_sched_group(struct task_group *tg)
9407 for_each_possible_cpu(i) {
9409 kfree(tg->cfs_rq[i]);
9419 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9421 struct cfs_rq *cfs_rq;
9422 struct sched_entity *se;
9426 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9429 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9433 tg->shares = NICE_0_LOAD;
9435 for_each_possible_cpu(i) {
9438 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9439 GFP_KERNEL, cpu_to_node(i));
9443 se = kzalloc_node(sizeof(struct sched_entity),
9444 GFP_KERNEL, cpu_to_node(i));
9448 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9457 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9459 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9460 &cpu_rq(cpu)->leaf_cfs_rq_list);
9463 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9465 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9467 #else /* !CONFG_FAIR_GROUP_SCHED */
9468 static inline void free_fair_sched_group(struct task_group *tg)
9473 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9478 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9482 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9485 #endif /* CONFIG_FAIR_GROUP_SCHED */
9487 #ifdef CONFIG_RT_GROUP_SCHED
9488 static void free_rt_sched_group(struct task_group *tg)
9492 destroy_rt_bandwidth(&tg->rt_bandwidth);
9494 for_each_possible_cpu(i) {
9496 kfree(tg->rt_rq[i]);
9498 kfree(tg->rt_se[i]);
9506 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9508 struct rt_rq *rt_rq;
9509 struct sched_rt_entity *rt_se;
9513 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9516 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9520 init_rt_bandwidth(&tg->rt_bandwidth,
9521 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9523 for_each_possible_cpu(i) {
9526 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9527 GFP_KERNEL, cpu_to_node(i));
9531 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9532 GFP_KERNEL, cpu_to_node(i));
9536 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9545 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9547 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9548 &cpu_rq(cpu)->leaf_rt_rq_list);
9551 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9553 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9555 #else /* !CONFIG_RT_GROUP_SCHED */
9556 static inline void free_rt_sched_group(struct task_group *tg)
9561 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9566 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9570 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9573 #endif /* CONFIG_RT_GROUP_SCHED */
9575 #ifdef CONFIG_GROUP_SCHED
9576 static void free_sched_group(struct task_group *tg)
9578 free_fair_sched_group(tg);
9579 free_rt_sched_group(tg);
9583 /* allocate runqueue etc for a new task group */
9584 struct task_group *sched_create_group(struct task_group *parent)
9586 struct task_group *tg;
9587 unsigned long flags;
9590 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9592 return ERR_PTR(-ENOMEM);
9594 if (!alloc_fair_sched_group(tg, parent))
9597 if (!alloc_rt_sched_group(tg, parent))
9600 spin_lock_irqsave(&task_group_lock, flags);
9601 for_each_possible_cpu(i) {
9602 register_fair_sched_group(tg, i);
9603 register_rt_sched_group(tg, i);
9605 list_add_rcu(&tg->list, &task_groups);
9607 WARN_ON(!parent); /* root should already exist */
9609 tg->parent = parent;
9610 INIT_LIST_HEAD(&tg->children);
9611 list_add_rcu(&tg->siblings, &parent->children);
9612 spin_unlock_irqrestore(&task_group_lock, flags);
9617 free_sched_group(tg);
9618 return ERR_PTR(-ENOMEM);
9621 /* rcu callback to free various structures associated with a task group */
9622 static void free_sched_group_rcu(struct rcu_head *rhp)
9624 /* now it should be safe to free those cfs_rqs */
9625 free_sched_group(container_of(rhp, struct task_group, rcu));
9628 /* Destroy runqueue etc associated with a task group */
9629 void sched_destroy_group(struct task_group *tg)
9631 unsigned long flags;
9634 spin_lock_irqsave(&task_group_lock, flags);
9635 for_each_possible_cpu(i) {
9636 unregister_fair_sched_group(tg, i);
9637 unregister_rt_sched_group(tg, i);
9639 list_del_rcu(&tg->list);
9640 list_del_rcu(&tg->siblings);
9641 spin_unlock_irqrestore(&task_group_lock, flags);
9643 /* wait for possible concurrent references to cfs_rqs complete */
9644 call_rcu(&tg->rcu, free_sched_group_rcu);
9647 /* change task's runqueue when it moves between groups.
9648 * The caller of this function should have put the task in its new group
9649 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9650 * reflect its new group.
9652 void sched_move_task(struct task_struct *tsk)
9655 unsigned long flags;
9658 rq = task_rq_lock(tsk, &flags);
9660 update_rq_clock(rq);
9662 running = task_current(rq, tsk);
9663 on_rq = tsk->se.on_rq;
9666 dequeue_task(rq, tsk, 0);
9667 if (unlikely(running))
9668 tsk->sched_class->put_prev_task(rq, tsk);
9670 set_task_rq(tsk, task_cpu(tsk));
9672 #ifdef CONFIG_FAIR_GROUP_SCHED
9673 if (tsk->sched_class->moved_group)
9674 tsk->sched_class->moved_group(tsk);
9677 if (unlikely(running))
9678 tsk->sched_class->set_curr_task(rq);
9680 enqueue_task(rq, tsk, 0);
9682 task_rq_unlock(rq, &flags);
9684 #endif /* CONFIG_GROUP_SCHED */
9686 #ifdef CONFIG_FAIR_GROUP_SCHED
9687 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9689 struct cfs_rq *cfs_rq = se->cfs_rq;
9694 dequeue_entity(cfs_rq, se, 0);
9696 se->load.weight = shares;
9697 se->load.inv_weight = 0;
9700 enqueue_entity(cfs_rq, se, 0);
9703 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9705 struct cfs_rq *cfs_rq = se->cfs_rq;
9706 struct rq *rq = cfs_rq->rq;
9707 unsigned long flags;
9709 spin_lock_irqsave(&rq->lock, flags);
9710 __set_se_shares(se, shares);
9711 spin_unlock_irqrestore(&rq->lock, flags);
9714 static DEFINE_MUTEX(shares_mutex);
9716 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9719 unsigned long flags;
9722 * We can't change the weight of the root cgroup.
9727 if (shares < MIN_SHARES)
9728 shares = MIN_SHARES;
9729 else if (shares > MAX_SHARES)
9730 shares = MAX_SHARES;
9732 mutex_lock(&shares_mutex);
9733 if (tg->shares == shares)
9736 spin_lock_irqsave(&task_group_lock, flags);
9737 for_each_possible_cpu(i)
9738 unregister_fair_sched_group(tg, i);
9739 list_del_rcu(&tg->siblings);
9740 spin_unlock_irqrestore(&task_group_lock, flags);
9742 /* wait for any ongoing reference to this group to finish */
9743 synchronize_sched();
9746 * Now we are free to modify the group's share on each cpu
9747 * w/o tripping rebalance_share or load_balance_fair.
9749 tg->shares = shares;
9750 for_each_possible_cpu(i) {
9754 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9755 set_se_shares(tg->se[i], shares);
9759 * Enable load balance activity on this group, by inserting it back on
9760 * each cpu's rq->leaf_cfs_rq_list.
9762 spin_lock_irqsave(&task_group_lock, flags);
9763 for_each_possible_cpu(i)
9764 register_fair_sched_group(tg, i);
9765 list_add_rcu(&tg->siblings, &tg->parent->children);
9766 spin_unlock_irqrestore(&task_group_lock, flags);
9768 mutex_unlock(&shares_mutex);
9772 unsigned long sched_group_shares(struct task_group *tg)
9778 #ifdef CONFIG_RT_GROUP_SCHED
9780 * Ensure that the real time constraints are schedulable.
9782 static DEFINE_MUTEX(rt_constraints_mutex);
9784 static unsigned long to_ratio(u64 period, u64 runtime)
9786 if (runtime == RUNTIME_INF)
9789 return div64_u64(runtime << 20, period);
9792 /* Must be called with tasklist_lock held */
9793 static inline int tg_has_rt_tasks(struct task_group *tg)
9795 struct task_struct *g, *p;
9797 do_each_thread(g, p) {
9798 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9800 } while_each_thread(g, p);
9805 struct rt_schedulable_data {
9806 struct task_group *tg;
9811 static int tg_schedulable(struct task_group *tg, void *data)
9813 struct rt_schedulable_data *d = data;
9814 struct task_group *child;
9815 unsigned long total, sum = 0;
9816 u64 period, runtime;
9818 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9819 runtime = tg->rt_bandwidth.rt_runtime;
9822 period = d->rt_period;
9823 runtime = d->rt_runtime;
9826 #ifdef CONFIG_USER_SCHED
9827 if (tg == &root_task_group) {
9828 period = global_rt_period();
9829 runtime = global_rt_runtime();
9834 * Cannot have more runtime than the period.
9836 if (runtime > period && runtime != RUNTIME_INF)
9840 * Ensure we don't starve existing RT tasks.
9842 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9845 total = to_ratio(period, runtime);
9848 * Nobody can have more than the global setting allows.
9850 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9854 * The sum of our children's runtime should not exceed our own.
9856 list_for_each_entry_rcu(child, &tg->children, siblings) {
9857 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9858 runtime = child->rt_bandwidth.rt_runtime;
9860 if (child == d->tg) {
9861 period = d->rt_period;
9862 runtime = d->rt_runtime;
9865 sum += to_ratio(period, runtime);
9874 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9876 struct rt_schedulable_data data = {
9878 .rt_period = period,
9879 .rt_runtime = runtime,
9882 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9885 static int tg_set_bandwidth(struct task_group *tg,
9886 u64 rt_period, u64 rt_runtime)
9890 mutex_lock(&rt_constraints_mutex);
9891 read_lock(&tasklist_lock);
9892 err = __rt_schedulable(tg, rt_period, rt_runtime);
9896 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9897 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9898 tg->rt_bandwidth.rt_runtime = rt_runtime;
9900 for_each_possible_cpu(i) {
9901 struct rt_rq *rt_rq = tg->rt_rq[i];
9903 spin_lock(&rt_rq->rt_runtime_lock);
9904 rt_rq->rt_runtime = rt_runtime;
9905 spin_unlock(&rt_rq->rt_runtime_lock);
9907 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9909 read_unlock(&tasklist_lock);
9910 mutex_unlock(&rt_constraints_mutex);
9915 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9917 u64 rt_runtime, rt_period;
9919 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9920 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9921 if (rt_runtime_us < 0)
9922 rt_runtime = RUNTIME_INF;
9924 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9927 long sched_group_rt_runtime(struct task_group *tg)
9931 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9934 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9935 do_div(rt_runtime_us, NSEC_PER_USEC);
9936 return rt_runtime_us;
9939 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9941 u64 rt_runtime, rt_period;
9943 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9944 rt_runtime = tg->rt_bandwidth.rt_runtime;
9949 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9952 long sched_group_rt_period(struct task_group *tg)
9956 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9957 do_div(rt_period_us, NSEC_PER_USEC);
9958 return rt_period_us;
9961 static int sched_rt_global_constraints(void)
9963 u64 runtime, period;
9966 if (sysctl_sched_rt_period <= 0)
9969 runtime = global_rt_runtime();
9970 period = global_rt_period();
9973 * Sanity check on the sysctl variables.
9975 if (runtime > period && runtime != RUNTIME_INF)
9978 mutex_lock(&rt_constraints_mutex);
9979 read_lock(&tasklist_lock);
9980 ret = __rt_schedulable(NULL, 0, 0);
9981 read_unlock(&tasklist_lock);
9982 mutex_unlock(&rt_constraints_mutex);
9987 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9989 /* Don't accept realtime tasks when there is no way for them to run */
9990 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9996 #else /* !CONFIG_RT_GROUP_SCHED */
9997 static int sched_rt_global_constraints(void)
9999 unsigned long flags;
10002 if (sysctl_sched_rt_period <= 0)
10006 * There's always some RT tasks in the root group
10007 * -- migration, kstopmachine etc..
10009 if (sysctl_sched_rt_runtime == 0)
10012 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
10013 for_each_possible_cpu(i) {
10014 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
10016 spin_lock(&rt_rq->rt_runtime_lock);
10017 rt_rq->rt_runtime = global_rt_runtime();
10018 spin_unlock(&rt_rq->rt_runtime_lock);
10020 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
10024 #endif /* CONFIG_RT_GROUP_SCHED */
10026 int sched_rt_handler(struct ctl_table *table, int write,
10027 struct file *filp, void __user *buffer, size_t *lenp,
10031 int old_period, old_runtime;
10032 static DEFINE_MUTEX(mutex);
10034 mutex_lock(&mutex);
10035 old_period = sysctl_sched_rt_period;
10036 old_runtime = sysctl_sched_rt_runtime;
10038 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
10040 if (!ret && write) {
10041 ret = sched_rt_global_constraints();
10043 sysctl_sched_rt_period = old_period;
10044 sysctl_sched_rt_runtime = old_runtime;
10046 def_rt_bandwidth.rt_runtime = global_rt_runtime();
10047 def_rt_bandwidth.rt_period =
10048 ns_to_ktime(global_rt_period());
10051 mutex_unlock(&mutex);
10056 #ifdef CONFIG_CGROUP_SCHED
10058 /* return corresponding task_group object of a cgroup */
10059 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
10061 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
10062 struct task_group, css);
10065 static struct cgroup_subsys_state *
10066 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
10068 struct task_group *tg, *parent;
10070 if (!cgrp->parent) {
10071 /* This is early initialization for the top cgroup */
10072 return &init_task_group.css;
10075 parent = cgroup_tg(cgrp->parent);
10076 tg = sched_create_group(parent);
10078 return ERR_PTR(-ENOMEM);
10084 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10086 struct task_group *tg = cgroup_tg(cgrp);
10088 sched_destroy_group(tg);
10092 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10093 struct task_struct *tsk)
10095 #ifdef CONFIG_RT_GROUP_SCHED
10096 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10099 /* We don't support RT-tasks being in separate groups */
10100 if (tsk->sched_class != &fair_sched_class)
10108 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10109 struct cgroup *old_cont, struct task_struct *tsk)
10111 sched_move_task(tsk);
10114 #ifdef CONFIG_FAIR_GROUP_SCHED
10115 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10118 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10121 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10123 struct task_group *tg = cgroup_tg(cgrp);
10125 return (u64) tg->shares;
10127 #endif /* CONFIG_FAIR_GROUP_SCHED */
10129 #ifdef CONFIG_RT_GROUP_SCHED
10130 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10133 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10136 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10138 return sched_group_rt_runtime(cgroup_tg(cgrp));
10141 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10144 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10147 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10149 return sched_group_rt_period(cgroup_tg(cgrp));
10151 #endif /* CONFIG_RT_GROUP_SCHED */
10153 static struct cftype cpu_files[] = {
10154 #ifdef CONFIG_FAIR_GROUP_SCHED
10157 .read_u64 = cpu_shares_read_u64,
10158 .write_u64 = cpu_shares_write_u64,
10161 #ifdef CONFIG_RT_GROUP_SCHED
10163 .name = "rt_runtime_us",
10164 .read_s64 = cpu_rt_runtime_read,
10165 .write_s64 = cpu_rt_runtime_write,
10168 .name = "rt_period_us",
10169 .read_u64 = cpu_rt_period_read_uint,
10170 .write_u64 = cpu_rt_period_write_uint,
10175 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10177 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10180 struct cgroup_subsys cpu_cgroup_subsys = {
10182 .create = cpu_cgroup_create,
10183 .destroy = cpu_cgroup_destroy,
10184 .can_attach = cpu_cgroup_can_attach,
10185 .attach = cpu_cgroup_attach,
10186 .populate = cpu_cgroup_populate,
10187 .subsys_id = cpu_cgroup_subsys_id,
10191 #endif /* CONFIG_CGROUP_SCHED */
10193 #ifdef CONFIG_CGROUP_CPUACCT
10196 * CPU accounting code for task groups.
10198 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10199 * (balbir@in.ibm.com).
10202 /* track cpu usage of a group of tasks and its child groups */
10204 struct cgroup_subsys_state css;
10205 /* cpuusage holds pointer to a u64-type object on every cpu */
10207 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10208 struct cpuacct *parent;
10211 struct cgroup_subsys cpuacct_subsys;
10213 /* return cpu accounting group corresponding to this container */
10214 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10216 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10217 struct cpuacct, css);
10220 /* return cpu accounting group to which this task belongs */
10221 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10223 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10224 struct cpuacct, css);
10227 /* create a new cpu accounting group */
10228 static struct cgroup_subsys_state *cpuacct_create(
10229 struct cgroup_subsys *ss, struct cgroup *cgrp)
10231 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10237 ca->cpuusage = alloc_percpu(u64);
10241 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10242 if (percpu_counter_init(&ca->cpustat[i], 0))
10243 goto out_free_counters;
10246 ca->parent = cgroup_ca(cgrp->parent);
10252 percpu_counter_destroy(&ca->cpustat[i]);
10253 free_percpu(ca->cpuusage);
10257 return ERR_PTR(-ENOMEM);
10260 /* destroy an existing cpu accounting group */
10262 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10264 struct cpuacct *ca = cgroup_ca(cgrp);
10267 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10268 percpu_counter_destroy(&ca->cpustat[i]);
10269 free_percpu(ca->cpuusage);
10273 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10275 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10278 #ifndef CONFIG_64BIT
10280 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10282 spin_lock_irq(&cpu_rq(cpu)->lock);
10284 spin_unlock_irq(&cpu_rq(cpu)->lock);
10292 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10294 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10296 #ifndef CONFIG_64BIT
10298 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10300 spin_lock_irq(&cpu_rq(cpu)->lock);
10302 spin_unlock_irq(&cpu_rq(cpu)->lock);
10308 /* return total cpu usage (in nanoseconds) of a group */
10309 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10311 struct cpuacct *ca = cgroup_ca(cgrp);
10312 u64 totalcpuusage = 0;
10315 for_each_present_cpu(i)
10316 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10318 return totalcpuusage;
10321 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10324 struct cpuacct *ca = cgroup_ca(cgrp);
10333 for_each_present_cpu(i)
10334 cpuacct_cpuusage_write(ca, i, 0);
10340 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10341 struct seq_file *m)
10343 struct cpuacct *ca = cgroup_ca(cgroup);
10347 for_each_present_cpu(i) {
10348 percpu = cpuacct_cpuusage_read(ca, i);
10349 seq_printf(m, "%llu ", (unsigned long long) percpu);
10351 seq_printf(m, "\n");
10355 static const char *cpuacct_stat_desc[] = {
10356 [CPUACCT_STAT_USER] = "user",
10357 [CPUACCT_STAT_SYSTEM] = "system",
10360 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10361 struct cgroup_map_cb *cb)
10363 struct cpuacct *ca = cgroup_ca(cgrp);
10366 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10367 s64 val = percpu_counter_read(&ca->cpustat[i]);
10368 val = cputime64_to_clock_t(val);
10369 cb->fill(cb, cpuacct_stat_desc[i], val);
10374 static struct cftype files[] = {
10377 .read_u64 = cpuusage_read,
10378 .write_u64 = cpuusage_write,
10381 .name = "usage_percpu",
10382 .read_seq_string = cpuacct_percpu_seq_read,
10386 .read_map = cpuacct_stats_show,
10390 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10392 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10396 * charge this task's execution time to its accounting group.
10398 * called with rq->lock held.
10400 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10402 struct cpuacct *ca;
10405 if (unlikely(!cpuacct_subsys.active))
10408 cpu = task_cpu(tsk);
10414 for (; ca; ca = ca->parent) {
10415 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10416 *cpuusage += cputime;
10423 * Charge the system/user time to the task's accounting group.
10425 static void cpuacct_update_stats(struct task_struct *tsk,
10426 enum cpuacct_stat_index idx, cputime_t val)
10428 struct cpuacct *ca;
10430 if (unlikely(!cpuacct_subsys.active))
10437 percpu_counter_add(&ca->cpustat[idx], val);
10443 struct cgroup_subsys cpuacct_subsys = {
10445 .create = cpuacct_create,
10446 .destroy = cpuacct_destroy,
10447 .populate = cpuacct_populate,
10448 .subsys_id = cpuacct_subsys_id,
10450 #endif /* CONFIG_CGROUP_CPUACCT */