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 #ifdef CONFIG_SCHED_HRTICK
635 int hrtick_csd_pending;
636 struct call_single_data hrtick_csd;
638 struct hrtimer hrtick_timer;
641 #ifdef CONFIG_SCHEDSTATS
643 struct sched_info rq_sched_info;
644 unsigned long long rq_cpu_time;
645 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
647 /* sys_sched_yield() stats */
648 unsigned int yld_count;
650 /* schedule() stats */
651 unsigned int sched_switch;
652 unsigned int sched_count;
653 unsigned int sched_goidle;
655 /* try_to_wake_up() stats */
656 unsigned int ttwu_count;
657 unsigned int ttwu_local;
660 unsigned int bkl_count;
664 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
666 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
668 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
671 static inline int cpu_of(struct rq *rq)
681 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
682 * See detach_destroy_domains: synchronize_sched for details.
684 * The domain tree of any CPU may only be accessed from within
685 * preempt-disabled sections.
687 #define for_each_domain(cpu, __sd) \
688 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
690 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
691 #define this_rq() (&__get_cpu_var(runqueues))
692 #define task_rq(p) cpu_rq(task_cpu(p))
693 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
695 static inline void update_rq_clock(struct rq *rq)
697 rq->clock = sched_clock_cpu(cpu_of(rq));
701 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
703 #ifdef CONFIG_SCHED_DEBUG
704 # define const_debug __read_mostly
706 # define const_debug static const
712 * Returns true if the current cpu runqueue is locked.
713 * This interface allows printk to be called with the runqueue lock
714 * held and know whether or not it is OK to wake up the klogd.
716 int runqueue_is_locked(void)
719 struct rq *rq = cpu_rq(cpu);
722 ret = spin_is_locked(&rq->lock);
728 * Debugging: various feature bits
731 #define SCHED_FEAT(name, enabled) \
732 __SCHED_FEAT_##name ,
735 #include "sched_features.h"
740 #define SCHED_FEAT(name, enabled) \
741 (1UL << __SCHED_FEAT_##name) * enabled |
743 const_debug unsigned int sysctl_sched_features =
744 #include "sched_features.h"
749 #ifdef CONFIG_SCHED_DEBUG
750 #define SCHED_FEAT(name, enabled) \
753 static __read_mostly char *sched_feat_names[] = {
754 #include "sched_features.h"
760 static int sched_feat_show(struct seq_file *m, void *v)
764 for (i = 0; sched_feat_names[i]; i++) {
765 if (!(sysctl_sched_features & (1UL << i)))
767 seq_printf(m, "%s ", sched_feat_names[i]);
775 sched_feat_write(struct file *filp, const char __user *ubuf,
776 size_t cnt, loff_t *ppos)
786 if (copy_from_user(&buf, ubuf, cnt))
791 if (strncmp(buf, "NO_", 3) == 0) {
796 for (i = 0; sched_feat_names[i]; i++) {
797 int len = strlen(sched_feat_names[i]);
799 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
801 sysctl_sched_features &= ~(1UL << i);
803 sysctl_sched_features |= (1UL << i);
808 if (!sched_feat_names[i])
816 static int sched_feat_open(struct inode *inode, struct file *filp)
818 return single_open(filp, sched_feat_show, NULL);
821 static struct file_operations sched_feat_fops = {
822 .open = sched_feat_open,
823 .write = sched_feat_write,
826 .release = single_release,
829 static __init int sched_init_debug(void)
831 debugfs_create_file("sched_features", 0644, NULL, NULL,
836 late_initcall(sched_init_debug);
840 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
843 * Number of tasks to iterate in a single balance run.
844 * Limited because this is done with IRQs disabled.
846 const_debug unsigned int sysctl_sched_nr_migrate = 32;
849 * ratelimit for updating the group shares.
852 unsigned int sysctl_sched_shares_ratelimit = 250000;
855 * Inject some fuzzyness into changing the per-cpu group shares
856 * this avoids remote rq-locks at the expense of fairness.
859 unsigned int sysctl_sched_shares_thresh = 4;
862 * period over which we measure -rt task cpu usage in us.
865 unsigned int sysctl_sched_rt_period = 1000000;
867 static __read_mostly int scheduler_running;
870 * part of the period that we allow rt tasks to run in us.
873 int sysctl_sched_rt_runtime = 950000;
875 static inline u64 global_rt_period(void)
877 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
880 static inline u64 global_rt_runtime(void)
882 if (sysctl_sched_rt_runtime < 0)
885 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
888 #ifndef prepare_arch_switch
889 # define prepare_arch_switch(next) do { } while (0)
891 #ifndef finish_arch_switch
892 # define finish_arch_switch(prev) do { } while (0)
895 static inline int task_current(struct rq *rq, struct task_struct *p)
897 return rq->curr == p;
900 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
901 static inline int task_running(struct rq *rq, struct task_struct *p)
903 return task_current(rq, p);
906 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
910 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
912 #ifdef CONFIG_DEBUG_SPINLOCK
913 /* this is a valid case when another task releases the spinlock */
914 rq->lock.owner = current;
917 * If we are tracking spinlock dependencies then we have to
918 * fix up the runqueue lock - which gets 'carried over' from
921 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
923 spin_unlock_irq(&rq->lock);
926 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
927 static inline int task_running(struct rq *rq, struct task_struct *p)
932 return task_current(rq, p);
936 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
940 * We can optimise this out completely for !SMP, because the
941 * SMP rebalancing from interrupt is the only thing that cares
946 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
947 spin_unlock_irq(&rq->lock);
949 spin_unlock(&rq->lock);
953 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
957 * After ->oncpu is cleared, the task can be moved to a different CPU.
958 * We must ensure this doesn't happen until the switch is completely
964 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
968 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
971 * __task_rq_lock - lock the runqueue a given task resides on.
972 * Must be called interrupts disabled.
974 static inline struct rq *__task_rq_lock(struct task_struct *p)
978 struct rq *rq = task_rq(p);
979 spin_lock(&rq->lock);
980 if (likely(rq == task_rq(p)))
982 spin_unlock(&rq->lock);
987 * task_rq_lock - lock the runqueue a given task resides on and disable
988 * interrupts. Note the ordering: we can safely lookup the task_rq without
989 * explicitly disabling preemption.
991 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
997 local_irq_save(*flags);
999 spin_lock(&rq->lock);
1000 if (likely(rq == task_rq(p)))
1002 spin_unlock_irqrestore(&rq->lock, *flags);
1006 void task_rq_unlock_wait(struct task_struct *p)
1008 struct rq *rq = task_rq(p);
1010 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1011 spin_unlock_wait(&rq->lock);
1014 static void __task_rq_unlock(struct rq *rq)
1015 __releases(rq->lock)
1017 spin_unlock(&rq->lock);
1020 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1021 __releases(rq->lock)
1023 spin_unlock_irqrestore(&rq->lock, *flags);
1027 * this_rq_lock - lock this runqueue and disable interrupts.
1029 static struct rq *this_rq_lock(void)
1030 __acquires(rq->lock)
1034 local_irq_disable();
1036 spin_lock(&rq->lock);
1041 #ifdef CONFIG_SCHED_HRTICK
1043 * Use HR-timers to deliver accurate preemption points.
1045 * Its all a bit involved since we cannot program an hrt while holding the
1046 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1049 * When we get rescheduled we reprogram the hrtick_timer outside of the
1055 * - enabled by features
1056 * - hrtimer is actually high res
1058 static inline int hrtick_enabled(struct rq *rq)
1060 if (!sched_feat(HRTICK))
1062 if (!cpu_active(cpu_of(rq)))
1064 return hrtimer_is_hres_active(&rq->hrtick_timer);
1067 static void hrtick_clear(struct rq *rq)
1069 if (hrtimer_active(&rq->hrtick_timer))
1070 hrtimer_cancel(&rq->hrtick_timer);
1074 * High-resolution timer tick.
1075 * Runs from hardirq context with interrupts disabled.
1077 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1079 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1081 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1083 spin_lock(&rq->lock);
1084 update_rq_clock(rq);
1085 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1086 spin_unlock(&rq->lock);
1088 return HRTIMER_NORESTART;
1093 * called from hardirq (IPI) context
1095 static void __hrtick_start(void *arg)
1097 struct rq *rq = arg;
1099 spin_lock(&rq->lock);
1100 hrtimer_restart(&rq->hrtick_timer);
1101 rq->hrtick_csd_pending = 0;
1102 spin_unlock(&rq->lock);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq *rq, u64 delay)
1112 struct hrtimer *timer = &rq->hrtick_timer;
1113 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1115 hrtimer_set_expires(timer, time);
1117 if (rq == this_rq()) {
1118 hrtimer_restart(timer);
1119 } else if (!rq->hrtick_csd_pending) {
1120 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd, 0);
1121 rq->hrtick_csd_pending = 1;
1126 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1128 int cpu = (int)(long)hcpu;
1131 case CPU_UP_CANCELED:
1132 case CPU_UP_CANCELED_FROZEN:
1133 case CPU_DOWN_PREPARE:
1134 case CPU_DOWN_PREPARE_FROZEN:
1136 case CPU_DEAD_FROZEN:
1137 hrtick_clear(cpu_rq(cpu));
1144 static __init void init_hrtick(void)
1146 hotcpu_notifier(hotplug_hrtick, 0);
1150 * Called to set the hrtick timer state.
1152 * called with rq->lock held and irqs disabled
1154 static void hrtick_start(struct rq *rq, u64 delay)
1156 __hrtimer_start_range_ns(&rq->hrtick_timer, ns_to_ktime(delay), 0,
1157 HRTIMER_MODE_REL, 0);
1160 static inline void init_hrtick(void)
1163 #endif /* CONFIG_SMP */
1165 static void init_rq_hrtick(struct rq *rq)
1168 rq->hrtick_csd_pending = 0;
1170 rq->hrtick_csd.flags = 0;
1171 rq->hrtick_csd.func = __hrtick_start;
1172 rq->hrtick_csd.info = rq;
1175 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1176 rq->hrtick_timer.function = hrtick;
1178 #else /* CONFIG_SCHED_HRTICK */
1179 static inline void hrtick_clear(struct rq *rq)
1183 static inline void init_rq_hrtick(struct rq *rq)
1187 static inline void init_hrtick(void)
1190 #endif /* CONFIG_SCHED_HRTICK */
1193 * resched_task - mark a task 'to be rescheduled now'.
1195 * On UP this means the setting of the need_resched flag, on SMP it
1196 * might also involve a cross-CPU call to trigger the scheduler on
1201 #ifndef tsk_is_polling
1202 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1205 static void resched_task(struct task_struct *p)
1209 assert_spin_locked(&task_rq(p)->lock);
1211 if (test_tsk_need_resched(p))
1214 set_tsk_need_resched(p);
1217 if (cpu == smp_processor_id())
1220 /* NEED_RESCHED must be visible before we test polling */
1222 if (!tsk_is_polling(p))
1223 smp_send_reschedule(cpu);
1226 static void resched_cpu(int cpu)
1228 struct rq *rq = cpu_rq(cpu);
1229 unsigned long flags;
1231 if (!spin_trylock_irqsave(&rq->lock, flags))
1233 resched_task(cpu_curr(cpu));
1234 spin_unlock_irqrestore(&rq->lock, flags);
1239 * When add_timer_on() enqueues a timer into the timer wheel of an
1240 * idle CPU then this timer might expire before the next timer event
1241 * which is scheduled to wake up that CPU. In case of a completely
1242 * idle system the next event might even be infinite time into the
1243 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1244 * leaves the inner idle loop so the newly added timer is taken into
1245 * account when the CPU goes back to idle and evaluates the timer
1246 * wheel for the next timer event.
1248 void wake_up_idle_cpu(int cpu)
1250 struct rq *rq = cpu_rq(cpu);
1252 if (cpu == smp_processor_id())
1256 * This is safe, as this function is called with the timer
1257 * wheel base lock of (cpu) held. When the CPU is on the way
1258 * to idle and has not yet set rq->curr to idle then it will
1259 * be serialized on the timer wheel base lock and take the new
1260 * timer into account automatically.
1262 if (rq->curr != rq->idle)
1266 * We can set TIF_RESCHED on the idle task of the other CPU
1267 * lockless. The worst case is that the other CPU runs the
1268 * idle task through an additional NOOP schedule()
1270 set_tsk_need_resched(rq->idle);
1272 /* NEED_RESCHED must be visible before we test polling */
1274 if (!tsk_is_polling(rq->idle))
1275 smp_send_reschedule(cpu);
1277 #endif /* CONFIG_NO_HZ */
1279 #else /* !CONFIG_SMP */
1280 static void resched_task(struct task_struct *p)
1282 assert_spin_locked(&task_rq(p)->lock);
1283 set_tsk_need_resched(p);
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1305 struct load_weight *lw)
1309 if (!lw->inv_weight) {
1310 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1313 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1317 tmp = (u64)delta_exec * weight;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp > WMULT_CONST))
1322 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1325 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1327 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1330 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1336 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator {
1404 struct task_struct *(*start)(void *);
1405 struct task_struct *(*next)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1411 unsigned long max_load_move, struct sched_domain *sd,
1412 enum cpu_idle_type idle, int *all_pinned,
1413 int *this_best_prio, struct rq_iterator *iterator);
1416 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1417 struct sched_domain *sd, enum cpu_idle_type idle,
1418 struct rq_iterator *iterator);
1421 /* Time spent by the tasks of the cpu accounting group executing in ... */
1422 enum cpuacct_stat_index {
1423 CPUACCT_STAT_USER, /* ... user mode */
1424 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1426 CPUACCT_STAT_NSTATS,
1429 #ifdef CONFIG_CGROUP_CPUACCT
1430 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1431 static void cpuacct_update_stats(struct task_struct *tsk,
1432 enum cpuacct_stat_index idx, cputime_t val);
1434 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1435 static inline void cpuacct_update_stats(struct task_struct *tsk,
1436 enum cpuacct_stat_index idx, cputime_t val) {}
1439 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1441 update_load_add(&rq->load, load);
1444 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1446 update_load_sub(&rq->load, load);
1449 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1450 typedef int (*tg_visitor)(struct task_group *, void *);
1453 * Iterate the full tree, calling @down when first entering a node and @up when
1454 * leaving it for the final time.
1456 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1458 struct task_group *parent, *child;
1462 parent = &root_task_group;
1464 ret = (*down)(parent, data);
1467 list_for_each_entry_rcu(child, &parent->children, siblings) {
1474 ret = (*up)(parent, data);
1479 parent = parent->parent;
1488 static int tg_nop(struct task_group *tg, void *data)
1495 static unsigned long source_load(int cpu, int type);
1496 static unsigned long target_load(int cpu, int type);
1497 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1499 static unsigned long cpu_avg_load_per_task(int cpu)
1501 struct rq *rq = cpu_rq(cpu);
1502 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1505 rq->avg_load_per_task = rq->load.weight / nr_running;
1507 rq->avg_load_per_task = 0;
1509 return rq->avg_load_per_task;
1512 #ifdef CONFIG_FAIR_GROUP_SCHED
1514 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1517 * Calculate and set the cpu's group shares.
1520 update_group_shares_cpu(struct task_group *tg, int cpu,
1521 unsigned long sd_shares, unsigned long sd_rq_weight)
1523 unsigned long shares;
1524 unsigned long rq_weight;
1529 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1532 * \Sum shares * rq_weight
1533 * shares = -----------------------
1537 shares = (sd_shares * rq_weight) / sd_rq_weight;
1538 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1540 if (abs(shares - tg->se[cpu]->load.weight) >
1541 sysctl_sched_shares_thresh) {
1542 struct rq *rq = cpu_rq(cpu);
1543 unsigned long flags;
1545 spin_lock_irqsave(&rq->lock, flags);
1546 tg->cfs_rq[cpu]->shares = shares;
1548 __set_se_shares(tg->se[cpu], shares);
1549 spin_unlock_irqrestore(&rq->lock, flags);
1554 * Re-compute the task group their per cpu shares over the given domain.
1555 * This needs to be done in a bottom-up fashion because the rq weight of a
1556 * parent group depends on the shares of its child groups.
1558 static int tg_shares_up(struct task_group *tg, void *data)
1560 unsigned long weight, rq_weight = 0;
1561 unsigned long shares = 0;
1562 struct sched_domain *sd = data;
1565 for_each_cpu(i, sched_domain_span(sd)) {
1567 * If there are currently no tasks on the cpu pretend there
1568 * is one of average load so that when a new task gets to
1569 * run here it will not get delayed by group starvation.
1571 weight = tg->cfs_rq[i]->load.weight;
1573 weight = NICE_0_LOAD;
1575 tg->cfs_rq[i]->rq_weight = weight;
1576 rq_weight += weight;
1577 shares += tg->cfs_rq[i]->shares;
1580 if ((!shares && rq_weight) || shares > tg->shares)
1581 shares = tg->shares;
1583 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1584 shares = tg->shares;
1586 for_each_cpu(i, sched_domain_span(sd))
1587 update_group_shares_cpu(tg, i, shares, rq_weight);
1593 * Compute the cpu's hierarchical load factor for each task group.
1594 * This needs to be done in a top-down fashion because the load of a child
1595 * group is a fraction of its parents load.
1597 static int tg_load_down(struct task_group *tg, void *data)
1600 long cpu = (long)data;
1603 load = cpu_rq(cpu)->load.weight;
1605 load = tg->parent->cfs_rq[cpu]->h_load;
1606 load *= tg->cfs_rq[cpu]->shares;
1607 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1610 tg->cfs_rq[cpu]->h_load = load;
1615 static void update_shares(struct sched_domain *sd)
1617 u64 now = cpu_clock(raw_smp_processor_id());
1618 s64 elapsed = now - sd->last_update;
1620 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1621 sd->last_update = now;
1622 walk_tg_tree(tg_nop, tg_shares_up, sd);
1626 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1628 spin_unlock(&rq->lock);
1630 spin_lock(&rq->lock);
1633 static void update_h_load(long cpu)
1635 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1640 static inline void update_shares(struct sched_domain *sd)
1644 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1650 #ifdef CONFIG_PREEMPT
1653 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1654 * way at the expense of forcing extra atomic operations in all
1655 * invocations. This assures that the double_lock is acquired using the
1656 * same underlying policy as the spinlock_t on this architecture, which
1657 * reduces latency compared to the unfair variant below. However, it
1658 * also adds more overhead and therefore may reduce throughput.
1660 static inline int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1661 __releases(this_rq->lock)
1662 __acquires(busiest->lock)
1663 __acquires(this_rq->lock)
1665 spin_unlock(&this_rq->lock);
1666 double_rq_lock(this_rq, busiest);
1673 * Unfair double_lock_balance: Optimizes throughput at the expense of
1674 * latency by eliminating extra atomic operations when the locks are
1675 * already in proper order on entry. This favors lower cpu-ids and will
1676 * grant the double lock to lower cpus over higher ids under contention,
1677 * regardless of entry order into the function.
1679 static int _double_lock_balance(struct rq *this_rq, struct rq *busiest)
1680 __releases(this_rq->lock)
1681 __acquires(busiest->lock)
1682 __acquires(this_rq->lock)
1686 if (unlikely(!spin_trylock(&busiest->lock))) {
1687 if (busiest < this_rq) {
1688 spin_unlock(&this_rq->lock);
1689 spin_lock(&busiest->lock);
1690 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1693 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1698 #endif /* CONFIG_PREEMPT */
1701 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1703 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1705 if (unlikely(!irqs_disabled())) {
1706 /* printk() doesn't work good under rq->lock */
1707 spin_unlock(&this_rq->lock);
1711 return _double_lock_balance(this_rq, busiest);
1714 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1715 __releases(busiest->lock)
1717 spin_unlock(&busiest->lock);
1718 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1722 #ifdef CONFIG_FAIR_GROUP_SCHED
1723 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1726 cfs_rq->shares = shares;
1731 #include "sched_stats.h"
1732 #include "sched_idletask.c"
1733 #include "sched_fair.c"
1734 #include "sched_rt.c"
1735 #ifdef CONFIG_SCHED_DEBUG
1736 # include "sched_debug.c"
1739 #define sched_class_highest (&rt_sched_class)
1740 #define for_each_class(class) \
1741 for (class = sched_class_highest; class; class = class->next)
1743 static void inc_nr_running(struct rq *rq)
1748 static void dec_nr_running(struct rq *rq)
1753 static void set_load_weight(struct task_struct *p)
1755 if (task_has_rt_policy(p)) {
1756 p->se.load.weight = prio_to_weight[0] * 2;
1757 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1762 * SCHED_IDLE tasks get minimal weight:
1764 if (p->policy == SCHED_IDLE) {
1765 p->se.load.weight = WEIGHT_IDLEPRIO;
1766 p->se.load.inv_weight = WMULT_IDLEPRIO;
1770 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1771 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1774 static void update_avg(u64 *avg, u64 sample)
1776 s64 diff = sample - *avg;
1780 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1783 p->se.start_runtime = p->se.sum_exec_runtime;
1785 sched_info_queued(p);
1786 p->sched_class->enqueue_task(rq, p, wakeup);
1790 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1793 if (p->se.last_wakeup) {
1794 update_avg(&p->se.avg_overlap,
1795 p->se.sum_exec_runtime - p->se.last_wakeup);
1796 p->se.last_wakeup = 0;
1798 update_avg(&p->se.avg_wakeup,
1799 sysctl_sched_wakeup_granularity);
1803 sched_info_dequeued(p);
1804 p->sched_class->dequeue_task(rq, p, sleep);
1809 * __normal_prio - return the priority that is based on the static prio
1811 static inline int __normal_prio(struct task_struct *p)
1813 return p->static_prio;
1817 * Calculate the expected normal priority: i.e. priority
1818 * without taking RT-inheritance into account. Might be
1819 * boosted by interactivity modifiers. Changes upon fork,
1820 * setprio syscalls, and whenever the interactivity
1821 * estimator recalculates.
1823 static inline int normal_prio(struct task_struct *p)
1827 if (task_has_rt_policy(p))
1828 prio = MAX_RT_PRIO-1 - p->rt_priority;
1830 prio = __normal_prio(p);
1835 * Calculate the current priority, i.e. the priority
1836 * taken into account by the scheduler. This value might
1837 * be boosted by RT tasks, or might be boosted by
1838 * interactivity modifiers. Will be RT if the task got
1839 * RT-boosted. If not then it returns p->normal_prio.
1841 static int effective_prio(struct task_struct *p)
1843 p->normal_prio = normal_prio(p);
1845 * If we are RT tasks or we were boosted to RT priority,
1846 * keep the priority unchanged. Otherwise, update priority
1847 * to the normal priority:
1849 if (!rt_prio(p->prio))
1850 return p->normal_prio;
1855 * activate_task - move a task to the runqueue.
1857 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1859 if (task_contributes_to_load(p))
1860 rq->nr_uninterruptible--;
1862 enqueue_task(rq, p, wakeup);
1867 * deactivate_task - remove a task from the runqueue.
1869 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1871 if (task_contributes_to_load(p))
1872 rq->nr_uninterruptible++;
1874 dequeue_task(rq, p, sleep);
1879 * task_curr - is this task currently executing on a CPU?
1880 * @p: the task in question.
1882 inline int task_curr(const struct task_struct *p)
1884 return cpu_curr(task_cpu(p)) == p;
1887 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1889 set_task_rq(p, cpu);
1892 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1893 * successfuly executed on another CPU. We must ensure that updates of
1894 * per-task data have been completed by this moment.
1897 task_thread_info(p)->cpu = cpu;
1901 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1902 const struct sched_class *prev_class,
1903 int oldprio, int running)
1905 if (prev_class != p->sched_class) {
1906 if (prev_class->switched_from)
1907 prev_class->switched_from(rq, p, running);
1908 p->sched_class->switched_to(rq, p, running);
1910 p->sched_class->prio_changed(rq, p, oldprio, running);
1915 /* Used instead of source_load when we know the type == 0 */
1916 static unsigned long weighted_cpuload(const int cpu)
1918 return cpu_rq(cpu)->load.weight;
1922 * Is this task likely cache-hot:
1925 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1930 * Buddy candidates are cache hot:
1932 if (sched_feat(CACHE_HOT_BUDDY) &&
1933 (&p->se == cfs_rq_of(&p->se)->next ||
1934 &p->se == cfs_rq_of(&p->se)->last))
1937 if (p->sched_class != &fair_sched_class)
1940 if (sysctl_sched_migration_cost == -1)
1942 if (sysctl_sched_migration_cost == 0)
1945 delta = now - p->se.exec_start;
1947 return delta < (s64)sysctl_sched_migration_cost;
1951 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1953 int old_cpu = task_cpu(p);
1954 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1955 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1956 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1959 clock_offset = old_rq->clock - new_rq->clock;
1961 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1963 #ifdef CONFIG_SCHEDSTATS
1964 if (p->se.wait_start)
1965 p->se.wait_start -= clock_offset;
1966 if (p->se.sleep_start)
1967 p->se.sleep_start -= clock_offset;
1968 if (p->se.block_start)
1969 p->se.block_start -= clock_offset;
1970 if (old_cpu != new_cpu) {
1971 schedstat_inc(p, se.nr_migrations);
1972 if (task_hot(p, old_rq->clock, NULL))
1973 schedstat_inc(p, se.nr_forced2_migrations);
1976 p->se.vruntime -= old_cfsrq->min_vruntime -
1977 new_cfsrq->min_vruntime;
1979 __set_task_cpu(p, new_cpu);
1982 struct migration_req {
1983 struct list_head list;
1985 struct task_struct *task;
1988 struct completion done;
1992 * The task's runqueue lock must be held.
1993 * Returns true if you have to wait for migration thread.
1996 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1998 struct rq *rq = task_rq(p);
2001 * If the task is not on a runqueue (and not running), then
2002 * it is sufficient to simply update the task's cpu field.
2004 if (!p->se.on_rq && !task_running(rq, p)) {
2005 set_task_cpu(p, dest_cpu);
2009 init_completion(&req->done);
2011 req->dest_cpu = dest_cpu;
2012 list_add(&req->list, &rq->migration_queue);
2018 * wait_task_inactive - wait for a thread to unschedule.
2020 * If @match_state is nonzero, it's the @p->state value just checked and
2021 * not expected to change. If it changes, i.e. @p might have woken up,
2022 * then return zero. When we succeed in waiting for @p to be off its CPU,
2023 * we return a positive number (its total switch count). If a second call
2024 * a short while later returns the same number, the caller can be sure that
2025 * @p has remained unscheduled the whole time.
2027 * The caller must ensure that the task *will* unschedule sometime soon,
2028 * else this function might spin for a *long* time. This function can't
2029 * be called with interrupts off, or it may introduce deadlock with
2030 * smp_call_function() if an IPI is sent by the same process we are
2031 * waiting to become inactive.
2033 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
2035 unsigned long flags;
2042 * We do the initial early heuristics without holding
2043 * any task-queue locks at all. We'll only try to get
2044 * the runqueue lock when things look like they will
2050 * If the task is actively running on another CPU
2051 * still, just relax and busy-wait without holding
2054 * NOTE! Since we don't hold any locks, it's not
2055 * even sure that "rq" stays as the right runqueue!
2056 * But we don't care, since "task_running()" will
2057 * return false if the runqueue has changed and p
2058 * is actually now running somewhere else!
2060 while (task_running(rq, p)) {
2061 if (match_state && unlikely(p->state != match_state))
2067 * Ok, time to look more closely! We need the rq
2068 * lock now, to be *sure*. If we're wrong, we'll
2069 * just go back and repeat.
2071 rq = task_rq_lock(p, &flags);
2072 trace_sched_wait_task(rq, p);
2073 running = task_running(rq, p);
2074 on_rq = p->se.on_rq;
2076 if (!match_state || p->state == match_state)
2077 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2078 task_rq_unlock(rq, &flags);
2081 * If it changed from the expected state, bail out now.
2083 if (unlikely(!ncsw))
2087 * Was it really running after all now that we
2088 * checked with the proper locks actually held?
2090 * Oops. Go back and try again..
2092 if (unlikely(running)) {
2098 * It's not enough that it's not actively running,
2099 * it must be off the runqueue _entirely_, and not
2102 * So if it was still runnable (but just not actively
2103 * running right now), it's preempted, and we should
2104 * yield - it could be a while.
2106 if (unlikely(on_rq)) {
2107 schedule_timeout_uninterruptible(1);
2112 * Ahh, all good. It wasn't running, and it wasn't
2113 * runnable, which means that it will never become
2114 * running in the future either. We're all done!
2123 * kick_process - kick a running thread to enter/exit the kernel
2124 * @p: the to-be-kicked thread
2126 * Cause a process which is running on another CPU to enter
2127 * kernel-mode, without any delay. (to get signals handled.)
2129 * NOTE: this function doesnt have to take the runqueue lock,
2130 * because all it wants to ensure is that the remote task enters
2131 * the kernel. If the IPI races and the task has been migrated
2132 * to another CPU then no harm is done and the purpose has been
2135 void kick_process(struct task_struct *p)
2141 if ((cpu != smp_processor_id()) && task_curr(p))
2142 smp_send_reschedule(cpu);
2147 * Return a low guess at the load of a migration-source cpu weighted
2148 * according to the scheduling class and "nice" value.
2150 * We want to under-estimate the load of migration sources, to
2151 * balance conservatively.
2153 static unsigned long source_load(int cpu, int type)
2155 struct rq *rq = cpu_rq(cpu);
2156 unsigned long total = weighted_cpuload(cpu);
2158 if (type == 0 || !sched_feat(LB_BIAS))
2161 return min(rq->cpu_load[type-1], total);
2165 * Return a high guess at the load of a migration-target cpu weighted
2166 * according to the scheduling class and "nice" value.
2168 static unsigned long target_load(int cpu, int type)
2170 struct rq *rq = cpu_rq(cpu);
2171 unsigned long total = weighted_cpuload(cpu);
2173 if (type == 0 || !sched_feat(LB_BIAS))
2176 return max(rq->cpu_load[type-1], total);
2180 * find_idlest_group finds and returns the least busy CPU group within the
2183 static struct sched_group *
2184 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2186 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2187 unsigned long min_load = ULONG_MAX, this_load = 0;
2188 int load_idx = sd->forkexec_idx;
2189 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2192 unsigned long load, avg_load;
2196 /* Skip over this group if it has no CPUs allowed */
2197 if (!cpumask_intersects(sched_group_cpus(group),
2201 local_group = cpumask_test_cpu(this_cpu,
2202 sched_group_cpus(group));
2204 /* Tally up the load of all CPUs in the group */
2207 for_each_cpu(i, sched_group_cpus(group)) {
2208 /* Bias balancing toward cpus of our domain */
2210 load = source_load(i, load_idx);
2212 load = target_load(i, load_idx);
2217 /* Adjust by relative CPU power of the group */
2218 avg_load = sg_div_cpu_power(group,
2219 avg_load * SCHED_LOAD_SCALE);
2222 this_load = avg_load;
2224 } else if (avg_load < min_load) {
2225 min_load = avg_load;
2228 } while (group = group->next, group != sd->groups);
2230 if (!idlest || 100*this_load < imbalance*min_load)
2236 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2239 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2241 unsigned long load, min_load = ULONG_MAX;
2245 /* Traverse only the allowed CPUs */
2246 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2247 load = weighted_cpuload(i);
2249 if (load < min_load || (load == min_load && i == this_cpu)) {
2259 * sched_balance_self: balance the current task (running on cpu) in domains
2260 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2263 * Balance, ie. select the least loaded group.
2265 * Returns the target CPU number, or the same CPU if no balancing is needed.
2267 * preempt must be disabled.
2269 static int sched_balance_self(int cpu, int flag)
2271 struct task_struct *t = current;
2272 struct sched_domain *tmp, *sd = NULL;
2274 for_each_domain(cpu, tmp) {
2276 * If power savings logic is enabled for a domain, stop there.
2278 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2280 if (tmp->flags & flag)
2288 struct sched_group *group;
2289 int new_cpu, weight;
2291 if (!(sd->flags & flag)) {
2296 group = find_idlest_group(sd, t, cpu);
2302 new_cpu = find_idlest_cpu(group, t, cpu);
2303 if (new_cpu == -1 || new_cpu == cpu) {
2304 /* Now try balancing at a lower domain level of cpu */
2309 /* Now try balancing at a lower domain level of new_cpu */
2311 weight = cpumask_weight(sched_domain_span(sd));
2313 for_each_domain(cpu, tmp) {
2314 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2316 if (tmp->flags & flag)
2319 /* while loop will break here if sd == NULL */
2325 #endif /* CONFIG_SMP */
2328 * try_to_wake_up - wake up a thread
2329 * @p: the to-be-woken-up thread
2330 * @state: the mask of task states that can be woken
2331 * @sync: do a synchronous wakeup?
2333 * Put it on the run-queue if it's not already there. The "current"
2334 * thread is always on the run-queue (except when the actual
2335 * re-schedule is in progress), and as such you're allowed to do
2336 * the simpler "current->state = TASK_RUNNING" to mark yourself
2337 * runnable without the overhead of this.
2339 * returns failure only if the task is already active.
2341 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2343 int cpu, orig_cpu, this_cpu, success = 0;
2344 unsigned long flags;
2348 if (!sched_feat(SYNC_WAKEUPS))
2352 if (sched_feat(LB_WAKEUP_UPDATE) && !root_task_group_empty()) {
2353 struct sched_domain *sd;
2355 this_cpu = raw_smp_processor_id();
2358 for_each_domain(this_cpu, sd) {
2359 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2368 rq = task_rq_lock(p, &flags);
2369 update_rq_clock(rq);
2370 old_state = p->state;
2371 if (!(old_state & state))
2379 this_cpu = smp_processor_id();
2382 if (unlikely(task_running(rq, p)))
2385 cpu = p->sched_class->select_task_rq(p, sync);
2386 if (cpu != orig_cpu) {
2387 set_task_cpu(p, cpu);
2388 task_rq_unlock(rq, &flags);
2389 /* might preempt at this point */
2390 rq = task_rq_lock(p, &flags);
2391 old_state = p->state;
2392 if (!(old_state & state))
2397 this_cpu = smp_processor_id();
2401 #ifdef CONFIG_SCHEDSTATS
2402 schedstat_inc(rq, ttwu_count);
2403 if (cpu == this_cpu)
2404 schedstat_inc(rq, ttwu_local);
2406 struct sched_domain *sd;
2407 for_each_domain(this_cpu, sd) {
2408 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2409 schedstat_inc(sd, ttwu_wake_remote);
2414 #endif /* CONFIG_SCHEDSTATS */
2417 #endif /* CONFIG_SMP */
2418 schedstat_inc(p, se.nr_wakeups);
2420 schedstat_inc(p, se.nr_wakeups_sync);
2421 if (orig_cpu != cpu)
2422 schedstat_inc(p, se.nr_wakeups_migrate);
2423 if (cpu == this_cpu)
2424 schedstat_inc(p, se.nr_wakeups_local);
2426 schedstat_inc(p, se.nr_wakeups_remote);
2427 activate_task(rq, p, 1);
2431 * Only attribute actual wakeups done by this task.
2433 if (!in_interrupt()) {
2434 struct sched_entity *se = ¤t->se;
2435 u64 sample = se->sum_exec_runtime;
2437 if (se->last_wakeup)
2438 sample -= se->last_wakeup;
2440 sample -= se->start_runtime;
2441 update_avg(&se->avg_wakeup, sample);
2443 se->last_wakeup = se->sum_exec_runtime;
2447 trace_sched_wakeup(rq, p, success);
2448 check_preempt_curr(rq, p, sync);
2450 p->state = TASK_RUNNING;
2452 if (p->sched_class->task_wake_up)
2453 p->sched_class->task_wake_up(rq, p);
2456 task_rq_unlock(rq, &flags);
2461 int wake_up_process(struct task_struct *p)
2463 return try_to_wake_up(p, TASK_ALL, 0);
2465 EXPORT_SYMBOL(wake_up_process);
2467 int wake_up_state(struct task_struct *p, unsigned int state)
2469 return try_to_wake_up(p, state, 0);
2473 * Perform scheduler related setup for a newly forked process p.
2474 * p is forked by current.
2476 * __sched_fork() is basic setup used by init_idle() too:
2478 static void __sched_fork(struct task_struct *p)
2480 p->se.exec_start = 0;
2481 p->se.sum_exec_runtime = 0;
2482 p->se.prev_sum_exec_runtime = 0;
2483 p->se.last_wakeup = 0;
2484 p->se.avg_overlap = 0;
2485 p->se.start_runtime = 0;
2486 p->se.avg_wakeup = sysctl_sched_wakeup_granularity;
2488 #ifdef CONFIG_SCHEDSTATS
2489 p->se.wait_start = 0;
2490 p->se.sum_sleep_runtime = 0;
2491 p->se.sleep_start = 0;
2492 p->se.block_start = 0;
2493 p->se.sleep_max = 0;
2494 p->se.block_max = 0;
2496 p->se.slice_max = 0;
2500 INIT_LIST_HEAD(&p->rt.run_list);
2502 INIT_LIST_HEAD(&p->se.group_node);
2504 #ifdef CONFIG_PREEMPT_NOTIFIERS
2505 INIT_HLIST_HEAD(&p->preempt_notifiers);
2509 * We mark the process as running here, but have not actually
2510 * inserted it onto the runqueue yet. This guarantees that
2511 * nobody will actually run it, and a signal or other external
2512 * event cannot wake it up and insert it on the runqueue either.
2514 p->state = TASK_RUNNING;
2518 * fork()/clone()-time setup:
2520 void sched_fork(struct task_struct *p, int clone_flags)
2522 int cpu = get_cpu();
2527 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2529 set_task_cpu(p, cpu);
2532 * Make sure we do not leak PI boosting priority to the child:
2534 p->prio = current->normal_prio;
2535 if (!rt_prio(p->prio))
2536 p->sched_class = &fair_sched_class;
2538 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2539 if (likely(sched_info_on()))
2540 memset(&p->sched_info, 0, sizeof(p->sched_info));
2542 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2545 #ifdef CONFIG_PREEMPT
2546 /* Want to start with kernel preemption disabled. */
2547 task_thread_info(p)->preempt_count = 1;
2549 plist_node_init(&p->pushable_tasks, MAX_PRIO);
2555 * wake_up_new_task - wake up a newly created task for the first time.
2557 * This function will do some initial scheduler statistics housekeeping
2558 * that must be done for every newly created context, then puts the task
2559 * on the runqueue and wakes it.
2561 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2563 unsigned long flags;
2566 rq = task_rq_lock(p, &flags);
2567 BUG_ON(p->state != TASK_RUNNING);
2568 update_rq_clock(rq);
2570 p->prio = effective_prio(p);
2572 if (!p->sched_class->task_new || !current->se.on_rq) {
2573 activate_task(rq, p, 0);
2576 * Let the scheduling class do new task startup
2577 * management (if any):
2579 p->sched_class->task_new(rq, p);
2582 trace_sched_wakeup_new(rq, p, 1);
2583 check_preempt_curr(rq, p, 0);
2585 if (p->sched_class->task_wake_up)
2586 p->sched_class->task_wake_up(rq, p);
2588 task_rq_unlock(rq, &flags);
2591 #ifdef CONFIG_PREEMPT_NOTIFIERS
2594 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2595 * @notifier: notifier struct to register
2597 void preempt_notifier_register(struct preempt_notifier *notifier)
2599 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2601 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2604 * preempt_notifier_unregister - no longer interested in preemption notifications
2605 * @notifier: notifier struct to unregister
2607 * This is safe to call from within a preemption notifier.
2609 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2611 hlist_del(¬ifier->link);
2613 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2615 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2617 struct preempt_notifier *notifier;
2618 struct hlist_node *node;
2620 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2621 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2625 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2626 struct task_struct *next)
2628 struct preempt_notifier *notifier;
2629 struct hlist_node *node;
2631 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2632 notifier->ops->sched_out(notifier, next);
2635 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2637 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2642 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2643 struct task_struct *next)
2647 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2650 * prepare_task_switch - prepare to switch tasks
2651 * @rq: the runqueue preparing to switch
2652 * @prev: the current task that is being switched out
2653 * @next: the task we are going to switch to.
2655 * This is called with the rq lock held and interrupts off. It must
2656 * be paired with a subsequent finish_task_switch after the context
2659 * prepare_task_switch sets up locking and calls architecture specific
2663 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2664 struct task_struct *next)
2666 fire_sched_out_preempt_notifiers(prev, next);
2667 prepare_lock_switch(rq, next);
2668 prepare_arch_switch(next);
2672 * finish_task_switch - clean up after a task-switch
2673 * @rq: runqueue associated with task-switch
2674 * @prev: the thread we just switched away from.
2676 * finish_task_switch must be called after the context switch, paired
2677 * with a prepare_task_switch call before the context switch.
2678 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2679 * and do any other architecture-specific cleanup actions.
2681 * Note that we may have delayed dropping an mm in context_switch(). If
2682 * so, we finish that here outside of the runqueue lock. (Doing it
2683 * with the lock held can cause deadlocks; see schedule() for
2686 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2687 __releases(rq->lock)
2689 struct mm_struct *mm = rq->prev_mm;
2692 int post_schedule = 0;
2694 if (current->sched_class->needs_post_schedule)
2695 post_schedule = current->sched_class->needs_post_schedule(rq);
2701 * A task struct has one reference for the use as "current".
2702 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2703 * schedule one last time. The schedule call will never return, and
2704 * the scheduled task must drop that reference.
2705 * The test for TASK_DEAD must occur while the runqueue locks are
2706 * still held, otherwise prev could be scheduled on another cpu, die
2707 * there before we look at prev->state, and then the reference would
2709 * Manfred Spraul <manfred@colorfullife.com>
2711 prev_state = prev->state;
2712 finish_arch_switch(prev);
2713 finish_lock_switch(rq, prev);
2716 current->sched_class->post_schedule(rq);
2719 fire_sched_in_preempt_notifiers(current);
2722 if (unlikely(prev_state == TASK_DEAD)) {
2724 * Remove function-return probe instances associated with this
2725 * task and put them back on the free list.
2727 kprobe_flush_task(prev);
2728 put_task_struct(prev);
2733 * schedule_tail - first thing a freshly forked thread must call.
2734 * @prev: the thread we just switched away from.
2736 asmlinkage void schedule_tail(struct task_struct *prev)
2737 __releases(rq->lock)
2739 struct rq *rq = this_rq();
2741 finish_task_switch(rq, prev);
2742 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2743 /* In this case, finish_task_switch does not reenable preemption */
2746 if (current->set_child_tid)
2747 put_user(task_pid_vnr(current), current->set_child_tid);
2751 * context_switch - switch to the new MM and the new
2752 * thread's register state.
2755 context_switch(struct rq *rq, struct task_struct *prev,
2756 struct task_struct *next)
2758 struct mm_struct *mm, *oldmm;
2760 prepare_task_switch(rq, prev, next);
2761 trace_sched_switch(rq, prev, next);
2763 oldmm = prev->active_mm;
2765 * For paravirt, this is coupled with an exit in switch_to to
2766 * combine the page table reload and the switch backend into
2769 arch_enter_lazy_cpu_mode();
2771 if (unlikely(!mm)) {
2772 next->active_mm = oldmm;
2773 atomic_inc(&oldmm->mm_count);
2774 enter_lazy_tlb(oldmm, next);
2776 switch_mm(oldmm, mm, next);
2778 if (unlikely(!prev->mm)) {
2779 prev->active_mm = NULL;
2780 rq->prev_mm = oldmm;
2783 * Since the runqueue lock will be released by the next
2784 * task (which is an invalid locking op but in the case
2785 * of the scheduler it's an obvious special-case), so we
2786 * do an early lockdep release here:
2788 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2789 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2792 /* Here we just switch the register state and the stack. */
2793 switch_to(prev, next, prev);
2797 * this_rq must be evaluated again because prev may have moved
2798 * CPUs since it called schedule(), thus the 'rq' on its stack
2799 * frame will be invalid.
2801 finish_task_switch(this_rq(), prev);
2805 * nr_running, nr_uninterruptible and nr_context_switches:
2807 * externally visible scheduler statistics: current number of runnable
2808 * threads, current number of uninterruptible-sleeping threads, total
2809 * number of context switches performed since bootup.
2811 unsigned long nr_running(void)
2813 unsigned long i, sum = 0;
2815 for_each_online_cpu(i)
2816 sum += cpu_rq(i)->nr_running;
2821 unsigned long nr_uninterruptible(void)
2823 unsigned long i, sum = 0;
2825 for_each_possible_cpu(i)
2826 sum += cpu_rq(i)->nr_uninterruptible;
2829 * Since we read the counters lockless, it might be slightly
2830 * inaccurate. Do not allow it to go below zero though:
2832 if (unlikely((long)sum < 0))
2838 unsigned long long nr_context_switches(void)
2841 unsigned long long sum = 0;
2843 for_each_possible_cpu(i)
2844 sum += cpu_rq(i)->nr_switches;
2849 unsigned long nr_iowait(void)
2851 unsigned long i, sum = 0;
2853 for_each_possible_cpu(i)
2854 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2859 unsigned long nr_active(void)
2861 unsigned long i, running = 0, uninterruptible = 0;
2863 for_each_online_cpu(i) {
2864 running += cpu_rq(i)->nr_running;
2865 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2868 if (unlikely((long)uninterruptible < 0))
2869 uninterruptible = 0;
2871 return running + uninterruptible;
2875 * Update rq->cpu_load[] statistics. This function is usually called every
2876 * scheduler tick (TICK_NSEC).
2878 static void update_cpu_load(struct rq *this_rq)
2880 unsigned long this_load = this_rq->load.weight;
2883 this_rq->nr_load_updates++;
2885 /* Update our load: */
2886 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2887 unsigned long old_load, new_load;
2889 /* scale is effectively 1 << i now, and >> i divides by scale */
2891 old_load = this_rq->cpu_load[i];
2892 new_load = this_load;
2894 * Round up the averaging division if load is increasing. This
2895 * prevents us from getting stuck on 9 if the load is 10, for
2898 if (new_load > old_load)
2899 new_load += scale-1;
2900 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2907 * double_rq_lock - safely lock two runqueues
2909 * Note this does not disable interrupts like task_rq_lock,
2910 * you need to do so manually before calling.
2912 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2913 __acquires(rq1->lock)
2914 __acquires(rq2->lock)
2916 BUG_ON(!irqs_disabled());
2918 spin_lock(&rq1->lock);
2919 __acquire(rq2->lock); /* Fake it out ;) */
2922 spin_lock(&rq1->lock);
2923 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2925 spin_lock(&rq2->lock);
2926 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2929 update_rq_clock(rq1);
2930 update_rq_clock(rq2);
2934 * double_rq_unlock - safely unlock two runqueues
2936 * Note this does not restore interrupts like task_rq_unlock,
2937 * you need to do so manually after calling.
2939 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2940 __releases(rq1->lock)
2941 __releases(rq2->lock)
2943 spin_unlock(&rq1->lock);
2945 spin_unlock(&rq2->lock);
2947 __release(rq2->lock);
2951 * If dest_cpu is allowed for this process, migrate the task to it.
2952 * This is accomplished by forcing the cpu_allowed mask to only
2953 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2954 * the cpu_allowed mask is restored.
2956 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2958 struct migration_req req;
2959 unsigned long flags;
2962 rq = task_rq_lock(p, &flags);
2963 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2964 || unlikely(!cpu_active(dest_cpu)))
2967 /* force the process onto the specified CPU */
2968 if (migrate_task(p, dest_cpu, &req)) {
2969 /* Need to wait for migration thread (might exit: take ref). */
2970 struct task_struct *mt = rq->migration_thread;
2972 get_task_struct(mt);
2973 task_rq_unlock(rq, &flags);
2974 wake_up_process(mt);
2975 put_task_struct(mt);
2976 wait_for_completion(&req.done);
2981 task_rq_unlock(rq, &flags);
2985 * sched_exec - execve() is a valuable balancing opportunity, because at
2986 * this point the task has the smallest effective memory and cache footprint.
2988 void sched_exec(void)
2990 int new_cpu, this_cpu = get_cpu();
2991 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2993 if (new_cpu != this_cpu)
2994 sched_migrate_task(current, new_cpu);
2998 * pull_task - move a task from a remote runqueue to the local runqueue.
2999 * Both runqueues must be locked.
3001 static void pull_task(struct rq *src_rq, struct task_struct *p,
3002 struct rq *this_rq, int this_cpu)
3004 deactivate_task(src_rq, p, 0);
3005 set_task_cpu(p, this_cpu);
3006 activate_task(this_rq, p, 0);
3008 * Note that idle threads have a prio of MAX_PRIO, for this test
3009 * to be always true for them.
3011 check_preempt_curr(this_rq, p, 0);
3015 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3018 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3019 struct sched_domain *sd, enum cpu_idle_type idle,
3022 int tsk_cache_hot = 0;
3024 * We do not migrate tasks that are:
3025 * 1) running (obviously), or
3026 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3027 * 3) are cache-hot on their current CPU.
3029 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
3030 schedstat_inc(p, se.nr_failed_migrations_affine);
3035 if (task_running(rq, p)) {
3036 schedstat_inc(p, se.nr_failed_migrations_running);
3041 * Aggressive migration if:
3042 * 1) task is cache cold, or
3043 * 2) too many balance attempts have failed.
3046 tsk_cache_hot = task_hot(p, rq->clock, sd);
3047 if (!tsk_cache_hot ||
3048 sd->nr_balance_failed > sd->cache_nice_tries) {
3049 #ifdef CONFIG_SCHEDSTATS
3050 if (tsk_cache_hot) {
3051 schedstat_inc(sd, lb_hot_gained[idle]);
3052 schedstat_inc(p, se.nr_forced_migrations);
3058 if (tsk_cache_hot) {
3059 schedstat_inc(p, se.nr_failed_migrations_hot);
3065 static unsigned long
3066 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3067 unsigned long max_load_move, struct sched_domain *sd,
3068 enum cpu_idle_type idle, int *all_pinned,
3069 int *this_best_prio, struct rq_iterator *iterator)
3071 int loops = 0, pulled = 0, pinned = 0;
3072 struct task_struct *p;
3073 long rem_load_move = max_load_move;
3075 if (max_load_move == 0)
3081 * Start the load-balancing iterator:
3083 p = iterator->start(iterator->arg);
3085 if (!p || loops++ > sysctl_sched_nr_migrate)
3088 if ((p->se.load.weight >> 1) > rem_load_move ||
3089 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3090 p = iterator->next(iterator->arg);
3094 pull_task(busiest, p, this_rq, this_cpu);
3096 rem_load_move -= p->se.load.weight;
3098 #ifdef CONFIG_PREEMPT
3100 * NEWIDLE balancing is a source of latency, so preemptible kernels
3101 * will stop after the first task is pulled to minimize the critical
3104 if (idle == CPU_NEWLY_IDLE)
3109 * We only want to steal up to the prescribed amount of weighted load.
3111 if (rem_load_move > 0) {
3112 if (p->prio < *this_best_prio)
3113 *this_best_prio = p->prio;
3114 p = iterator->next(iterator->arg);
3119 * Right now, this is one of only two places pull_task() is called,
3120 * so we can safely collect pull_task() stats here rather than
3121 * inside pull_task().
3123 schedstat_add(sd, lb_gained[idle], pulled);
3126 *all_pinned = pinned;
3128 return max_load_move - rem_load_move;
3132 * move_tasks tries to move up to max_load_move weighted load from busiest to
3133 * this_rq, as part of a balancing operation within domain "sd".
3134 * Returns 1 if successful and 0 otherwise.
3136 * Called with both runqueues locked.
3138 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3139 unsigned long max_load_move,
3140 struct sched_domain *sd, enum cpu_idle_type idle,
3143 const struct sched_class *class = sched_class_highest;
3144 unsigned long total_load_moved = 0;
3145 int this_best_prio = this_rq->curr->prio;
3149 class->load_balance(this_rq, this_cpu, busiest,
3150 max_load_move - total_load_moved,
3151 sd, idle, all_pinned, &this_best_prio);
3152 class = class->next;
3154 #ifdef CONFIG_PREEMPT
3156 * NEWIDLE balancing is a source of latency, so preemptible
3157 * kernels will stop after the first task is pulled to minimize
3158 * the critical section.
3160 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3163 } while (class && max_load_move > total_load_moved);
3165 return total_load_moved > 0;
3169 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3170 struct sched_domain *sd, enum cpu_idle_type idle,
3171 struct rq_iterator *iterator)
3173 struct task_struct *p = iterator->start(iterator->arg);
3177 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3178 pull_task(busiest, p, this_rq, this_cpu);
3180 * Right now, this is only the second place pull_task()
3181 * is called, so we can safely collect pull_task()
3182 * stats here rather than inside pull_task().
3184 schedstat_inc(sd, lb_gained[idle]);
3188 p = iterator->next(iterator->arg);
3195 * move_one_task tries to move exactly one task from busiest to this_rq, as
3196 * part of active balancing operations within "domain".
3197 * Returns 1 if successful and 0 otherwise.
3199 * Called with both runqueues locked.
3201 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3202 struct sched_domain *sd, enum cpu_idle_type idle)
3204 const struct sched_class *class;
3206 for (class = sched_class_highest; class; class = class->next)
3207 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3212 /********** Helpers for find_busiest_group ************************/
3214 * sd_lb_stats - Structure to store the statistics of a sched_domain
3215 * during load balancing.
3217 struct sd_lb_stats {
3218 struct sched_group *busiest; /* Busiest group in this sd */
3219 struct sched_group *this; /* Local group in this sd */
3220 unsigned long total_load; /* Total load of all groups in sd */
3221 unsigned long total_pwr; /* Total power of all groups in sd */
3222 unsigned long avg_load; /* Average load across all groups in sd */
3224 /** Statistics of this group */
3225 unsigned long this_load;
3226 unsigned long this_load_per_task;
3227 unsigned long this_nr_running;
3229 /* Statistics of the busiest group */
3230 unsigned long max_load;
3231 unsigned long busiest_load_per_task;
3232 unsigned long busiest_nr_running;
3234 int group_imb; /* Is there imbalance in this sd */
3235 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3236 int power_savings_balance; /* Is powersave balance needed for this sd */
3237 struct sched_group *group_min; /* Least loaded group in sd */
3238 struct sched_group *group_leader; /* Group which relieves group_min */
3239 unsigned long min_load_per_task; /* load_per_task in group_min */
3240 unsigned long leader_nr_running; /* Nr running of group_leader */
3241 unsigned long min_nr_running; /* Nr running of group_min */
3246 * sg_lb_stats - stats of a sched_group required for load_balancing
3248 struct sg_lb_stats {
3249 unsigned long avg_load; /*Avg load across the CPUs of the group */
3250 unsigned long group_load; /* Total load over the CPUs of the group */
3251 unsigned long sum_nr_running; /* Nr tasks running in the group */
3252 unsigned long sum_weighted_load; /* Weighted load of group's tasks */
3253 unsigned long group_capacity;
3254 int group_imb; /* Is there an imbalance in the group ? */
3258 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3259 * @group: The group whose first cpu is to be returned.
3261 static inline unsigned int group_first_cpu(struct sched_group *group)
3263 return cpumask_first(sched_group_cpus(group));
3267 * get_sd_load_idx - Obtain the load index for a given sched domain.
3268 * @sd: The sched_domain whose load_idx is to be obtained.
3269 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3271 static inline int get_sd_load_idx(struct sched_domain *sd,
3272 enum cpu_idle_type idle)
3278 load_idx = sd->busy_idx;
3281 case CPU_NEWLY_IDLE:
3282 load_idx = sd->newidle_idx;
3285 load_idx = sd->idle_idx;
3293 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3295 * init_sd_power_savings_stats - Initialize power savings statistics for
3296 * the given sched_domain, during load balancing.
3298 * @sd: Sched domain whose power-savings statistics are to be initialized.
3299 * @sds: Variable containing the statistics for sd.
3300 * @idle: Idle status of the CPU at which we're performing load-balancing.
3302 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3303 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3306 * Busy processors will not participate in power savings
3309 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3310 sds->power_savings_balance = 0;
3312 sds->power_savings_balance = 1;
3313 sds->min_nr_running = ULONG_MAX;
3314 sds->leader_nr_running = 0;
3319 * update_sd_power_savings_stats - Update the power saving stats for a
3320 * sched_domain while performing load balancing.
3322 * @group: sched_group belonging to the sched_domain under consideration.
3323 * @sds: Variable containing the statistics of the sched_domain
3324 * @local_group: Does group contain the CPU for which we're performing
3326 * @sgs: Variable containing the statistics of the group.
3328 static inline void update_sd_power_savings_stats(struct sched_group *group,
3329 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3332 if (!sds->power_savings_balance)
3336 * If the local group is idle or completely loaded
3337 * no need to do power savings balance at this domain
3339 if (local_group && (sds->this_nr_running >= sgs->group_capacity ||
3340 !sds->this_nr_running))
3341 sds->power_savings_balance = 0;
3344 * If a group is already running at full capacity or idle,
3345 * don't include that group in power savings calculations
3347 if (!sds->power_savings_balance ||
3348 sgs->sum_nr_running >= sgs->group_capacity ||
3349 !sgs->sum_nr_running)
3353 * Calculate the group which has the least non-idle load.
3354 * This is the group from where we need to pick up the load
3357 if ((sgs->sum_nr_running < sds->min_nr_running) ||
3358 (sgs->sum_nr_running == sds->min_nr_running &&
3359 group_first_cpu(group) > group_first_cpu(sds->group_min))) {
3360 sds->group_min = group;
3361 sds->min_nr_running = sgs->sum_nr_running;
3362 sds->min_load_per_task = sgs->sum_weighted_load /
3363 sgs->sum_nr_running;
3367 * Calculate the group which is almost near its
3368 * capacity but still has some space to pick up some load
3369 * from other group and save more power
3371 if (sgs->sum_nr_running > sgs->group_capacity - 1)
3374 if (sgs->sum_nr_running > sds->leader_nr_running ||
3375 (sgs->sum_nr_running == sds->leader_nr_running &&
3376 group_first_cpu(group) < group_first_cpu(sds->group_leader))) {
3377 sds->group_leader = group;
3378 sds->leader_nr_running = sgs->sum_nr_running;
3383 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3384 * @sds: Variable containing the statistics of the sched_domain
3385 * under consideration.
3386 * @this_cpu: Cpu at which we're currently performing load-balancing.
3387 * @imbalance: Variable to store the imbalance.
3390 * Check if we have potential to perform some power-savings balance.
3391 * If yes, set the busiest group to be the least loaded group in the
3392 * sched_domain, so that it's CPUs can be put to idle.
3394 * Returns 1 if there is potential to perform power-savings balance.
3397 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3398 int this_cpu, unsigned long *imbalance)
3400 if (!sds->power_savings_balance)
3403 if (sds->this != sds->group_leader ||
3404 sds->group_leader == sds->group_min)
3407 *imbalance = sds->min_load_per_task;
3408 sds->busiest = sds->group_min;
3410 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3411 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3412 group_first_cpu(sds->group_leader);
3418 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3419 static inline void init_sd_power_savings_stats(struct sched_domain *sd,
3420 struct sd_lb_stats *sds, enum cpu_idle_type idle)
3425 static inline void update_sd_power_savings_stats(struct sched_group *group,
3426 struct sd_lb_stats *sds, int local_group, struct sg_lb_stats *sgs)
3431 static inline int check_power_save_busiest_group(struct sd_lb_stats *sds,
3432 int this_cpu, unsigned long *imbalance)
3436 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3440 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3441 * @group: sched_group whose statistics are to be updated.
3442 * @this_cpu: Cpu for which load balance is currently performed.
3443 * @idle: Idle status of this_cpu
3444 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3445 * @sd_idle: Idle status of the sched_domain containing group.
3446 * @local_group: Does group contain this_cpu.
3447 * @cpus: Set of cpus considered for load balancing.
3448 * @balance: Should we balance.
3449 * @sgs: variable to hold the statistics for this group.
3451 static inline void update_sg_lb_stats(struct sched_group *group, int this_cpu,
3452 enum cpu_idle_type idle, int load_idx, int *sd_idle,
3453 int local_group, const struct cpumask *cpus,
3454 int *balance, struct sg_lb_stats *sgs)
3456 unsigned long load, max_cpu_load, min_cpu_load;
3458 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3459 unsigned long sum_avg_load_per_task;
3460 unsigned long avg_load_per_task;
3463 balance_cpu = group_first_cpu(group);
3465 /* Tally up the load of all CPUs in the group */
3466 sum_avg_load_per_task = avg_load_per_task = 0;
3468 min_cpu_load = ~0UL;
3470 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3471 struct rq *rq = cpu_rq(i);
3473 if (*sd_idle && rq->nr_running)
3476 /* Bias balancing toward cpus of our domain */
3478 if (idle_cpu(i) && !first_idle_cpu) {
3483 load = target_load(i, load_idx);
3485 load = source_load(i, load_idx);
3486 if (load > max_cpu_load)
3487 max_cpu_load = load;
3488 if (min_cpu_load > load)
3489 min_cpu_load = load;
3492 sgs->group_load += load;
3493 sgs->sum_nr_running += rq->nr_running;
3494 sgs->sum_weighted_load += weighted_cpuload(i);
3496 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3500 * First idle cpu or the first cpu(busiest) in this sched group
3501 * is eligible for doing load balancing at this and above
3502 * domains. In the newly idle case, we will allow all the cpu's
3503 * to do the newly idle load balance.
3505 if (idle != CPU_NEWLY_IDLE && local_group &&
3506 balance_cpu != this_cpu && balance) {
3511 /* Adjust by relative CPU power of the group */
3512 sgs->avg_load = sg_div_cpu_power(group,
3513 sgs->group_load * SCHED_LOAD_SCALE);
3517 * Consider the group unbalanced when the imbalance is larger
3518 * than the average weight of two tasks.
3520 * APZ: with cgroup the avg task weight can vary wildly and
3521 * might not be a suitable number - should we keep a
3522 * normalized nr_running number somewhere that negates
3525 avg_load_per_task = sg_div_cpu_power(group,
3526 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3528 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3531 sgs->group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3536 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3537 * @sd: sched_domain whose statistics are to be updated.
3538 * @this_cpu: Cpu for which load balance is currently performed.
3539 * @idle: Idle status of this_cpu
3540 * @sd_idle: Idle status of the sched_domain containing group.
3541 * @cpus: Set of cpus considered for load balancing.
3542 * @balance: Should we balance.
3543 * @sds: variable to hold the statistics for this sched_domain.
3545 static inline void update_sd_lb_stats(struct sched_domain *sd, int this_cpu,
3546 enum cpu_idle_type idle, int *sd_idle,
3547 const struct cpumask *cpus, int *balance,
3548 struct sd_lb_stats *sds)
3550 struct sched_group *group = sd->groups;
3551 struct sg_lb_stats sgs;
3554 init_sd_power_savings_stats(sd, sds, idle);
3555 load_idx = get_sd_load_idx(sd, idle);
3560 local_group = cpumask_test_cpu(this_cpu,
3561 sched_group_cpus(group));
3562 memset(&sgs, 0, sizeof(sgs));
3563 update_sg_lb_stats(group, this_cpu, idle, load_idx, sd_idle,
3564 local_group, cpus, balance, &sgs);
3566 if (local_group && balance && !(*balance))
3569 sds->total_load += sgs.group_load;
3570 sds->total_pwr += group->__cpu_power;
3573 sds->this_load = sgs.avg_load;
3575 sds->this_nr_running = sgs.sum_nr_running;
3576 sds->this_load_per_task = sgs.sum_weighted_load;
3577 } else if (sgs.avg_load > sds->max_load &&
3578 (sgs.sum_nr_running > sgs.group_capacity ||
3580 sds->max_load = sgs.avg_load;
3581 sds->busiest = group;
3582 sds->busiest_nr_running = sgs.sum_nr_running;
3583 sds->busiest_load_per_task = sgs.sum_weighted_load;
3584 sds->group_imb = sgs.group_imb;
3587 update_sd_power_savings_stats(group, sds, local_group, &sgs);
3588 group = group->next;
3589 } while (group != sd->groups);
3594 * fix_small_imbalance - Calculate the minor imbalance that exists
3595 * amongst the groups of a sched_domain, during
3597 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3598 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3599 * @imbalance: Variable to store the imbalance.
3601 static inline void fix_small_imbalance(struct sd_lb_stats *sds,
3602 int this_cpu, unsigned long *imbalance)
3604 unsigned long tmp, pwr_now = 0, pwr_move = 0;
3605 unsigned int imbn = 2;
3607 if (sds->this_nr_running) {
3608 sds->this_load_per_task /= sds->this_nr_running;
3609 if (sds->busiest_load_per_task >
3610 sds->this_load_per_task)
3613 sds->this_load_per_task =
3614 cpu_avg_load_per_task(this_cpu);
3616 if (sds->max_load - sds->this_load + sds->busiest_load_per_task >=
3617 sds->busiest_load_per_task * imbn) {
3618 *imbalance = sds->busiest_load_per_task;
3623 * OK, we don't have enough imbalance to justify moving tasks,
3624 * however we may be able to increase total CPU power used by
3628 pwr_now += sds->busiest->__cpu_power *
3629 min(sds->busiest_load_per_task, sds->max_load);
3630 pwr_now += sds->this->__cpu_power *
3631 min(sds->this_load_per_task, sds->this_load);
3632 pwr_now /= SCHED_LOAD_SCALE;
3634 /* Amount of load we'd subtract */
3635 tmp = sg_div_cpu_power(sds->busiest,
3636 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3637 if (sds->max_load > tmp)
3638 pwr_move += sds->busiest->__cpu_power *
3639 min(sds->busiest_load_per_task, sds->max_load - tmp);
3641 /* Amount of load we'd add */
3642 if (sds->max_load * sds->busiest->__cpu_power <
3643 sds->busiest_load_per_task * SCHED_LOAD_SCALE)
3644 tmp = sg_div_cpu_power(sds->this,
3645 sds->max_load * sds->busiest->__cpu_power);
3647 tmp = sg_div_cpu_power(sds->this,
3648 sds->busiest_load_per_task * SCHED_LOAD_SCALE);
3649 pwr_move += sds->this->__cpu_power *
3650 min(sds->this_load_per_task, sds->this_load + tmp);
3651 pwr_move /= SCHED_LOAD_SCALE;
3653 /* Move if we gain throughput */
3654 if (pwr_move > pwr_now)
3655 *imbalance = sds->busiest_load_per_task;
3659 * calculate_imbalance - Calculate the amount of imbalance present within the
3660 * groups of a given sched_domain during load balance.
3661 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3662 * @this_cpu: Cpu for which currently load balance is being performed.
3663 * @imbalance: The variable to store the imbalance.
3665 static inline void calculate_imbalance(struct sd_lb_stats *sds, int this_cpu,
3666 unsigned long *imbalance)
3668 unsigned long max_pull;
3670 * In the presence of smp nice balancing, certain scenarios can have
3671 * max load less than avg load(as we skip the groups at or below
3672 * its cpu_power, while calculating max_load..)
3674 if (sds->max_load < sds->avg_load) {
3676 return fix_small_imbalance(sds, this_cpu, imbalance);
3679 /* Don't want to pull so many tasks that a group would go idle */
3680 max_pull = min(sds->max_load - sds->avg_load,
3681 sds->max_load - sds->busiest_load_per_task);
3683 /* How much load to actually move to equalise the imbalance */
3684 *imbalance = min(max_pull * sds->busiest->__cpu_power,
3685 (sds->avg_load - sds->this_load) * sds->this->__cpu_power)
3689 * if *imbalance is less than the average load per runnable task
3690 * there is no gaurantee that any tasks will be moved so we'll have
3691 * a think about bumping its value to force at least one task to be
3694 if (*imbalance < sds->busiest_load_per_task)
3695 return fix_small_imbalance(sds, this_cpu, imbalance);
3698 /******* find_busiest_group() helpers end here *********************/
3701 * find_busiest_group - Returns the busiest group within the sched_domain
3702 * if there is an imbalance. If there isn't an imbalance, and
3703 * the user has opted for power-savings, it returns a group whose
3704 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3705 * such a group exists.
3707 * Also calculates the amount of weighted load which should be moved
3708 * to restore balance.
3710 * @sd: The sched_domain whose busiest group is to be returned.
3711 * @this_cpu: The cpu for which load balancing is currently being performed.
3712 * @imbalance: Variable which stores amount of weighted load which should
3713 * be moved to restore balance/put a group to idle.
3714 * @idle: The idle status of this_cpu.
3715 * @sd_idle: The idleness of sd
3716 * @cpus: The set of CPUs under consideration for load-balancing.
3717 * @balance: Pointer to a variable indicating if this_cpu
3718 * is the appropriate cpu to perform load balancing at this_level.
3720 * Returns: - the busiest group if imbalance exists.
3721 * - If no imbalance and user has opted for power-savings balance,
3722 * return the least loaded group whose CPUs can be
3723 * put to idle by rebalancing its tasks onto our group.
3725 static struct sched_group *
3726 find_busiest_group(struct sched_domain *sd, int this_cpu,
3727 unsigned long *imbalance, enum cpu_idle_type idle,
3728 int *sd_idle, const struct cpumask *cpus, int *balance)
3730 struct sd_lb_stats sds;
3732 memset(&sds, 0, sizeof(sds));
3735 * Compute the various statistics relavent for load balancing at
3738 update_sd_lb_stats(sd, this_cpu, idle, sd_idle, cpus,
3741 /* Cases where imbalance does not exist from POV of this_cpu */
3742 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3744 * 2) There is no busy sibling group to pull from.
3745 * 3) This group is the busiest group.
3746 * 4) This group is more busy than the avg busieness at this
3748 * 5) The imbalance is within the specified limit.
3749 * 6) Any rebalance would lead to ping-pong
3751 if (balance && !(*balance))
3754 if (!sds.busiest || sds.busiest_nr_running == 0)
3757 if (sds.this_load >= sds.max_load)
3760 sds.avg_load = (SCHED_LOAD_SCALE * sds.total_load) / sds.total_pwr;
3762 if (sds.this_load >= sds.avg_load)
3765 if (100 * sds.max_load <= sd->imbalance_pct * sds.this_load)
3768 sds.busiest_load_per_task /= sds.busiest_nr_running;
3770 sds.busiest_load_per_task =
3771 min(sds.busiest_load_per_task, sds.avg_load);
3774 * We're trying to get all the cpus to the average_load, so we don't
3775 * want to push ourselves above the average load, nor do we wish to
3776 * reduce the max loaded cpu below the average load, as either of these
3777 * actions would just result in more rebalancing later, and ping-pong
3778 * tasks around. Thus we look for the minimum possible imbalance.
3779 * Negative imbalances (*we* are more loaded than anyone else) will
3780 * be counted as no imbalance for these purposes -- we can't fix that
3781 * by pulling tasks to us. Be careful of negative numbers as they'll
3782 * appear as very large values with unsigned longs.
3784 if (sds.max_load <= sds.busiest_load_per_task)
3787 /* Looks like there is an imbalance. Compute it */
3788 calculate_imbalance(&sds, this_cpu, imbalance);
3793 * There is no obvious imbalance. But check if we can do some balancing
3796 if (check_power_save_busiest_group(&sds, this_cpu, imbalance))
3804 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3807 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3808 unsigned long imbalance, const struct cpumask *cpus)
3810 struct rq *busiest = NULL, *rq;
3811 unsigned long max_load = 0;
3814 for_each_cpu(i, sched_group_cpus(group)) {
3817 if (!cpumask_test_cpu(i, cpus))
3821 wl = weighted_cpuload(i);
3823 if (rq->nr_running == 1 && wl > imbalance)
3826 if (wl > max_load) {
3836 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3837 * so long as it is large enough.
3839 #define MAX_PINNED_INTERVAL 512
3841 /* Working cpumask for load_balance and load_balance_newidle. */
3842 static DEFINE_PER_CPU(cpumask_var_t, load_balance_tmpmask);
3845 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3846 * tasks if there is an imbalance.
3848 static int load_balance(int this_cpu, struct rq *this_rq,
3849 struct sched_domain *sd, enum cpu_idle_type idle,
3852 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3853 struct sched_group *group;
3854 unsigned long imbalance;
3856 unsigned long flags;
3857 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
3859 cpumask_setall(cpus);
3862 * When power savings policy is enabled for the parent domain, idle
3863 * sibling can pick up load irrespective of busy siblings. In this case,
3864 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3865 * portraying it as CPU_NOT_IDLE.
3867 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3868 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3871 schedstat_inc(sd, lb_count[idle]);
3875 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3882 schedstat_inc(sd, lb_nobusyg[idle]);
3886 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3888 schedstat_inc(sd, lb_nobusyq[idle]);
3892 BUG_ON(busiest == this_rq);
3894 schedstat_add(sd, lb_imbalance[idle], imbalance);
3897 if (busiest->nr_running > 1) {
3899 * Attempt to move tasks. If find_busiest_group has found
3900 * an imbalance but busiest->nr_running <= 1, the group is
3901 * still unbalanced. ld_moved simply stays zero, so it is
3902 * correctly treated as an imbalance.
3904 local_irq_save(flags);
3905 double_rq_lock(this_rq, busiest);
3906 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3907 imbalance, sd, idle, &all_pinned);
3908 double_rq_unlock(this_rq, busiest);
3909 local_irq_restore(flags);
3912 * some other cpu did the load balance for us.
3914 if (ld_moved && this_cpu != smp_processor_id())
3915 resched_cpu(this_cpu);
3917 /* All tasks on this runqueue were pinned by CPU affinity */
3918 if (unlikely(all_pinned)) {
3919 cpumask_clear_cpu(cpu_of(busiest), cpus);
3920 if (!cpumask_empty(cpus))
3927 schedstat_inc(sd, lb_failed[idle]);
3928 sd->nr_balance_failed++;
3930 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3932 spin_lock_irqsave(&busiest->lock, flags);
3934 /* don't kick the migration_thread, if the curr
3935 * task on busiest cpu can't be moved to this_cpu
3937 if (!cpumask_test_cpu(this_cpu,
3938 &busiest->curr->cpus_allowed)) {
3939 spin_unlock_irqrestore(&busiest->lock, flags);
3941 goto out_one_pinned;
3944 if (!busiest->active_balance) {
3945 busiest->active_balance = 1;
3946 busiest->push_cpu = this_cpu;
3949 spin_unlock_irqrestore(&busiest->lock, flags);
3951 wake_up_process(busiest->migration_thread);
3954 * We've kicked active balancing, reset the failure
3957 sd->nr_balance_failed = sd->cache_nice_tries+1;
3960 sd->nr_balance_failed = 0;
3962 if (likely(!active_balance)) {
3963 /* We were unbalanced, so reset the balancing interval */
3964 sd->balance_interval = sd->min_interval;
3967 * If we've begun active balancing, start to back off. This
3968 * case may not be covered by the all_pinned logic if there
3969 * is only 1 task on the busy runqueue (because we don't call
3972 if (sd->balance_interval < sd->max_interval)
3973 sd->balance_interval *= 2;
3976 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3977 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3983 schedstat_inc(sd, lb_balanced[idle]);
3985 sd->nr_balance_failed = 0;
3988 /* tune up the balancing interval */
3989 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3990 (sd->balance_interval < sd->max_interval))
3991 sd->balance_interval *= 2;
3993 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3994 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4005 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4006 * tasks if there is an imbalance.
4008 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4009 * this_rq is locked.
4012 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd)
4014 struct sched_group *group;
4015 struct rq *busiest = NULL;
4016 unsigned long imbalance;
4020 struct cpumask *cpus = __get_cpu_var(load_balance_tmpmask);
4022 cpumask_setall(cpus);
4025 * When power savings policy is enabled for the parent domain, idle
4026 * sibling can pick up load irrespective of busy siblings. In this case,
4027 * let the state of idle sibling percolate up as IDLE, instead of
4028 * portraying it as CPU_NOT_IDLE.
4030 if (sd->flags & SD_SHARE_CPUPOWER &&
4031 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4034 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
4036 update_shares_locked(this_rq, sd);
4037 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
4038 &sd_idle, cpus, NULL);
4040 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
4044 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
4046 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
4050 BUG_ON(busiest == this_rq);
4052 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
4055 if (busiest->nr_running > 1) {
4056 /* Attempt to move tasks */
4057 double_lock_balance(this_rq, busiest);
4058 /* this_rq->clock is already updated */
4059 update_rq_clock(busiest);
4060 ld_moved = move_tasks(this_rq, this_cpu, busiest,
4061 imbalance, sd, CPU_NEWLY_IDLE,
4063 double_unlock_balance(this_rq, busiest);
4065 if (unlikely(all_pinned)) {
4066 cpumask_clear_cpu(cpu_of(busiest), cpus);
4067 if (!cpumask_empty(cpus))
4073 int active_balance = 0;
4075 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
4076 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4077 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4080 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
4083 if (sd->nr_balance_failed++ < 2)
4087 * The only task running in a non-idle cpu can be moved to this
4088 * cpu in an attempt to completely freeup the other CPU
4089 * package. The same method used to move task in load_balance()
4090 * have been extended for load_balance_newidle() to speedup
4091 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4093 * The package power saving logic comes from
4094 * find_busiest_group(). If there are no imbalance, then
4095 * f_b_g() will return NULL. However when sched_mc={1,2} then
4096 * f_b_g() will select a group from which a running task may be
4097 * pulled to this cpu in order to make the other package idle.
4098 * If there is no opportunity to make a package idle and if
4099 * there are no imbalance, then f_b_g() will return NULL and no
4100 * action will be taken in load_balance_newidle().
4102 * Under normal task pull operation due to imbalance, there
4103 * will be more than one task in the source run queue and
4104 * move_tasks() will succeed. ld_moved will be true and this
4105 * active balance code will not be triggered.
4108 /* Lock busiest in correct order while this_rq is held */
4109 double_lock_balance(this_rq, busiest);
4112 * don't kick the migration_thread, if the curr
4113 * task on busiest cpu can't be moved to this_cpu
4115 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
4116 double_unlock_balance(this_rq, busiest);
4121 if (!busiest->active_balance) {
4122 busiest->active_balance = 1;
4123 busiest->push_cpu = this_cpu;
4127 double_unlock_balance(this_rq, busiest);
4129 * Should not call ttwu while holding a rq->lock
4131 spin_unlock(&this_rq->lock);
4133 wake_up_process(busiest->migration_thread);
4134 spin_lock(&this_rq->lock);
4137 sd->nr_balance_failed = 0;
4139 update_shares_locked(this_rq, sd);
4143 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
4144 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
4145 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
4147 sd->nr_balance_failed = 0;
4153 * idle_balance is called by schedule() if this_cpu is about to become
4154 * idle. Attempts to pull tasks from other CPUs.
4156 static void idle_balance(int this_cpu, struct rq *this_rq)
4158 struct sched_domain *sd;
4159 int pulled_task = 0;
4160 unsigned long next_balance = jiffies + HZ;
4162 for_each_domain(this_cpu, sd) {
4163 unsigned long interval;
4165 if (!(sd->flags & SD_LOAD_BALANCE))
4168 if (sd->flags & SD_BALANCE_NEWIDLE)
4169 /* If we've pulled tasks over stop searching: */
4170 pulled_task = load_balance_newidle(this_cpu, this_rq,
4173 interval = msecs_to_jiffies(sd->balance_interval);
4174 if (time_after(next_balance, sd->last_balance + interval))
4175 next_balance = sd->last_balance + interval;
4179 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
4181 * We are going idle. next_balance may be set based on
4182 * a busy processor. So reset next_balance.
4184 this_rq->next_balance = next_balance;
4189 * active_load_balance is run by migration threads. It pushes running tasks
4190 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4191 * running on each physical CPU where possible, and avoids physical /
4192 * logical imbalances.
4194 * Called with busiest_rq locked.
4196 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
4198 int target_cpu = busiest_rq->push_cpu;
4199 struct sched_domain *sd;
4200 struct rq *target_rq;
4202 /* Is there any task to move? */
4203 if (busiest_rq->nr_running <= 1)
4206 target_rq = cpu_rq(target_cpu);
4209 * This condition is "impossible", if it occurs
4210 * we need to fix it. Originally reported by
4211 * Bjorn Helgaas on a 128-cpu setup.
4213 BUG_ON(busiest_rq == target_rq);
4215 /* move a task from busiest_rq to target_rq */
4216 double_lock_balance(busiest_rq, target_rq);
4217 update_rq_clock(busiest_rq);
4218 update_rq_clock(target_rq);
4220 /* Search for an sd spanning us and the target CPU. */
4221 for_each_domain(target_cpu, sd) {
4222 if ((sd->flags & SD_LOAD_BALANCE) &&
4223 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
4228 schedstat_inc(sd, alb_count);
4230 if (move_one_task(target_rq, target_cpu, busiest_rq,
4232 schedstat_inc(sd, alb_pushed);
4234 schedstat_inc(sd, alb_failed);
4236 double_unlock_balance(busiest_rq, target_rq);
4241 atomic_t load_balancer;
4242 cpumask_var_t cpu_mask;
4243 cpumask_var_t ilb_grp_nohz_mask;
4244 } nohz ____cacheline_aligned = {
4245 .load_balancer = ATOMIC_INIT(-1),
4248 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4250 * lowest_flag_domain - Return lowest sched_domain containing flag.
4251 * @cpu: The cpu whose lowest level of sched domain is to
4253 * @flag: The flag to check for the lowest sched_domain
4254 * for the given cpu.
4256 * Returns the lowest sched_domain of a cpu which contains the given flag.
4258 static inline struct sched_domain *lowest_flag_domain(int cpu, int flag)
4260 struct sched_domain *sd;
4262 for_each_domain(cpu, sd)
4263 if (sd && (sd->flags & flag))
4270 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4271 * @cpu: The cpu whose domains we're iterating over.
4272 * @sd: variable holding the value of the power_savings_sd
4274 * @flag: The flag to filter the sched_domains to be iterated.
4276 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4277 * set, starting from the lowest sched_domain to the highest.
4279 #define for_each_flag_domain(cpu, sd, flag) \
4280 for (sd = lowest_flag_domain(cpu, flag); \
4281 (sd && (sd->flags & flag)); sd = sd->parent)
4284 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4285 * @ilb_group: group to be checked for semi-idleness
4287 * Returns: 1 if the group is semi-idle. 0 otherwise.
4289 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4290 * and atleast one non-idle CPU. This helper function checks if the given
4291 * sched_group is semi-idle or not.
4293 static inline int is_semi_idle_group(struct sched_group *ilb_group)
4295 cpumask_and(nohz.ilb_grp_nohz_mask, nohz.cpu_mask,
4296 sched_group_cpus(ilb_group));
4299 * A sched_group is semi-idle when it has atleast one busy cpu
4300 * and atleast one idle cpu.
4302 if (cpumask_empty(nohz.ilb_grp_nohz_mask))
4305 if (cpumask_equal(nohz.ilb_grp_nohz_mask, sched_group_cpus(ilb_group)))
4311 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4312 * @cpu: The cpu which is nominating a new idle_load_balancer.
4314 * Returns: Returns the id of the idle load balancer if it exists,
4315 * Else, returns >= nr_cpu_ids.
4317 * This algorithm picks the idle load balancer such that it belongs to a
4318 * semi-idle powersavings sched_domain. The idea is to try and avoid
4319 * completely idle packages/cores just for the purpose of idle load balancing
4320 * when there are other idle cpu's which are better suited for that job.
4322 static int find_new_ilb(int cpu)
4324 struct sched_domain *sd;
4325 struct sched_group *ilb_group;
4328 * Have idle load balancer selection from semi-idle packages only
4329 * when power-aware load balancing is enabled
4331 if (!(sched_smt_power_savings || sched_mc_power_savings))
4335 * Optimize for the case when we have no idle CPUs or only one
4336 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4338 if (cpumask_weight(nohz.cpu_mask) < 2)
4341 for_each_flag_domain(cpu, sd, SD_POWERSAVINGS_BALANCE) {
4342 ilb_group = sd->groups;
4345 if (is_semi_idle_group(ilb_group))
4346 return cpumask_first(nohz.ilb_grp_nohz_mask);
4348 ilb_group = ilb_group->next;
4350 } while (ilb_group != sd->groups);
4354 return cpumask_first(nohz.cpu_mask);
4356 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4357 static inline int find_new_ilb(int call_cpu)
4359 return cpumask_first(nohz.cpu_mask);
4364 * This routine will try to nominate the ilb (idle load balancing)
4365 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4366 * load balancing on behalf of all those cpus. If all the cpus in the system
4367 * go into this tickless mode, then there will be no ilb owner (as there is
4368 * no need for one) and all the cpus will sleep till the next wakeup event
4371 * For the ilb owner, tick is not stopped. And this tick will be used
4372 * for idle load balancing. ilb owner will still be part of
4375 * While stopping the tick, this cpu will become the ilb owner if there
4376 * is no other owner. And will be the owner till that cpu becomes busy
4377 * or if all cpus in the system stop their ticks at which point
4378 * there is no need for ilb owner.
4380 * When the ilb owner becomes busy, it nominates another owner, during the
4381 * next busy scheduler_tick()
4383 int select_nohz_load_balancer(int stop_tick)
4385 int cpu = smp_processor_id();
4388 cpu_rq(cpu)->in_nohz_recently = 1;
4390 if (!cpu_active(cpu)) {
4391 if (atomic_read(&nohz.load_balancer) != cpu)
4395 * If we are going offline and still the leader,
4398 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4404 cpumask_set_cpu(cpu, nohz.cpu_mask);
4406 /* time for ilb owner also to sleep */
4407 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4408 if (atomic_read(&nohz.load_balancer) == cpu)
4409 atomic_set(&nohz.load_balancer, -1);
4413 if (atomic_read(&nohz.load_balancer) == -1) {
4414 /* make me the ilb owner */
4415 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4417 } else if (atomic_read(&nohz.load_balancer) == cpu) {
4420 if (!(sched_smt_power_savings ||
4421 sched_mc_power_savings))
4424 * Check to see if there is a more power-efficient
4427 new_ilb = find_new_ilb(cpu);
4428 if (new_ilb < nr_cpu_ids && new_ilb != cpu) {
4429 atomic_set(&nohz.load_balancer, -1);
4430 resched_cpu(new_ilb);
4436 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
4439 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4441 if (atomic_read(&nohz.load_balancer) == cpu)
4442 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4449 static DEFINE_SPINLOCK(balancing);
4452 * It checks each scheduling domain to see if it is due to be balanced,
4453 * and initiates a balancing operation if so.
4455 * Balancing parameters are set up in arch_init_sched_domains.
4457 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4460 struct rq *rq = cpu_rq(cpu);
4461 unsigned long interval;
4462 struct sched_domain *sd;
4463 /* Earliest time when we have to do rebalance again */
4464 unsigned long next_balance = jiffies + 60*HZ;
4465 int update_next_balance = 0;
4468 for_each_domain(cpu, sd) {
4469 if (!(sd->flags & SD_LOAD_BALANCE))
4472 interval = sd->balance_interval;
4473 if (idle != CPU_IDLE)
4474 interval *= sd->busy_factor;
4476 /* scale ms to jiffies */
4477 interval = msecs_to_jiffies(interval);
4478 if (unlikely(!interval))
4480 if (interval > HZ*NR_CPUS/10)
4481 interval = HZ*NR_CPUS/10;
4483 need_serialize = sd->flags & SD_SERIALIZE;
4485 if (need_serialize) {
4486 if (!spin_trylock(&balancing))
4490 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4491 if (load_balance(cpu, rq, sd, idle, &balance)) {
4493 * We've pulled tasks over so either we're no
4494 * longer idle, or one of our SMT siblings is
4497 idle = CPU_NOT_IDLE;
4499 sd->last_balance = jiffies;
4502 spin_unlock(&balancing);
4504 if (time_after(next_balance, sd->last_balance + interval)) {
4505 next_balance = sd->last_balance + interval;
4506 update_next_balance = 1;
4510 * Stop the load balance at this level. There is another
4511 * CPU in our sched group which is doing load balancing more
4519 * next_balance will be updated only when there is a need.
4520 * When the cpu is attached to null domain for ex, it will not be
4523 if (likely(update_next_balance))
4524 rq->next_balance = next_balance;
4528 * run_rebalance_domains is triggered when needed from the scheduler tick.
4529 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4530 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4532 static void run_rebalance_domains(struct softirq_action *h)
4534 int this_cpu = smp_processor_id();
4535 struct rq *this_rq = cpu_rq(this_cpu);
4536 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4537 CPU_IDLE : CPU_NOT_IDLE;
4539 rebalance_domains(this_cpu, idle);
4543 * If this cpu is the owner for idle load balancing, then do the
4544 * balancing on behalf of the other idle cpus whose ticks are
4547 if (this_rq->idle_at_tick &&
4548 atomic_read(&nohz.load_balancer) == this_cpu) {
4552 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4553 if (balance_cpu == this_cpu)
4557 * If this cpu gets work to do, stop the load balancing
4558 * work being done for other cpus. Next load
4559 * balancing owner will pick it up.
4564 rebalance_domains(balance_cpu, CPU_IDLE);
4566 rq = cpu_rq(balance_cpu);
4567 if (time_after(this_rq->next_balance, rq->next_balance))
4568 this_rq->next_balance = rq->next_balance;
4574 static inline int on_null_domain(int cpu)
4576 return !rcu_dereference(cpu_rq(cpu)->sd);
4580 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4582 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4583 * idle load balancing owner or decide to stop the periodic load balancing,
4584 * if the whole system is idle.
4586 static inline void trigger_load_balance(struct rq *rq, int cpu)
4590 * If we were in the nohz mode recently and busy at the current
4591 * scheduler tick, then check if we need to nominate new idle
4594 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4595 rq->in_nohz_recently = 0;
4597 if (atomic_read(&nohz.load_balancer) == cpu) {
4598 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4599 atomic_set(&nohz.load_balancer, -1);
4602 if (atomic_read(&nohz.load_balancer) == -1) {
4603 int ilb = find_new_ilb(cpu);
4605 if (ilb < nr_cpu_ids)
4611 * If this cpu is idle and doing idle load balancing for all the
4612 * cpus with ticks stopped, is it time for that to stop?
4614 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4615 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4621 * If this cpu is idle and the idle load balancing is done by
4622 * someone else, then no need raise the SCHED_SOFTIRQ
4624 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4625 cpumask_test_cpu(cpu, nohz.cpu_mask))
4628 /* Don't need to rebalance while attached to NULL domain */
4629 if (time_after_eq(jiffies, rq->next_balance) &&
4630 likely(!on_null_domain(cpu)))
4631 raise_softirq(SCHED_SOFTIRQ);
4634 #else /* CONFIG_SMP */
4637 * on UP we do not need to balance between CPUs:
4639 static inline void idle_balance(int cpu, struct rq *rq)
4645 DEFINE_PER_CPU(struct kernel_stat, kstat);
4647 EXPORT_PER_CPU_SYMBOL(kstat);
4650 * Return any ns on the sched_clock that have not yet been accounted in
4651 * @p in case that task is currently running.
4653 * Called with task_rq_lock() held on @rq.
4655 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4659 if (task_current(rq, p)) {
4660 update_rq_clock(rq);
4661 ns = rq->clock - p->se.exec_start;
4669 unsigned long long task_delta_exec(struct task_struct *p)
4671 unsigned long flags;
4675 rq = task_rq_lock(p, &flags);
4676 ns = do_task_delta_exec(p, rq);
4677 task_rq_unlock(rq, &flags);
4683 * Return accounted runtime for the task.
4684 * In case the task is currently running, return the runtime plus current's
4685 * pending runtime that have not been accounted yet.
4687 unsigned long long task_sched_runtime(struct task_struct *p)
4689 unsigned long flags;
4693 rq = task_rq_lock(p, &flags);
4694 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4695 task_rq_unlock(rq, &flags);
4701 * Return sum_exec_runtime for the thread group.
4702 * In case the task is currently running, return the sum plus current's
4703 * pending runtime that have not been accounted yet.
4705 * Note that the thread group might have other running tasks as well,
4706 * so the return value not includes other pending runtime that other
4707 * running tasks might have.
4709 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4711 struct task_cputime totals;
4712 unsigned long flags;
4716 rq = task_rq_lock(p, &flags);
4717 thread_group_cputime(p, &totals);
4718 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4719 task_rq_unlock(rq, &flags);
4725 * Account user cpu time to a process.
4726 * @p: the process that the cpu time gets accounted to
4727 * @cputime: the cpu time spent in user space since the last update
4728 * @cputime_scaled: cputime scaled by cpu frequency
4730 void account_user_time(struct task_struct *p, cputime_t cputime,
4731 cputime_t cputime_scaled)
4733 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4736 /* Add user time to process. */
4737 p->utime = cputime_add(p->utime, cputime);
4738 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4739 account_group_user_time(p, cputime);
4741 /* Add user time to cpustat. */
4742 tmp = cputime_to_cputime64(cputime);
4743 if (TASK_NICE(p) > 0)
4744 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4746 cpustat->user = cputime64_add(cpustat->user, tmp);
4748 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4749 /* Account for user time used */
4750 acct_update_integrals(p);
4754 * Account guest cpu time to a process.
4755 * @p: the process that the cpu time gets accounted to
4756 * @cputime: the cpu time spent in virtual machine since the last update
4757 * @cputime_scaled: cputime scaled by cpu frequency
4759 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4760 cputime_t cputime_scaled)
4763 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4765 tmp = cputime_to_cputime64(cputime);
4767 /* Add guest time to process. */
4768 p->utime = cputime_add(p->utime, cputime);
4769 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4770 account_group_user_time(p, cputime);
4771 p->gtime = cputime_add(p->gtime, cputime);
4773 /* Add guest time to cpustat. */
4774 cpustat->user = cputime64_add(cpustat->user, tmp);
4775 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4779 * Account system cpu time to a process.
4780 * @p: the process that the cpu time gets accounted to
4781 * @hardirq_offset: the offset to subtract from hardirq_count()
4782 * @cputime: the cpu time spent in kernel space since the last update
4783 * @cputime_scaled: cputime scaled by cpu frequency
4785 void account_system_time(struct task_struct *p, int hardirq_offset,
4786 cputime_t cputime, cputime_t cputime_scaled)
4788 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4791 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4792 account_guest_time(p, cputime, cputime_scaled);
4796 /* Add system time to process. */
4797 p->stime = cputime_add(p->stime, cputime);
4798 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4799 account_group_system_time(p, cputime);
4801 /* Add system time to cpustat. */
4802 tmp = cputime_to_cputime64(cputime);
4803 if (hardirq_count() - hardirq_offset)
4804 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4805 else if (softirq_count())
4806 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4808 cpustat->system = cputime64_add(cpustat->system, tmp);
4810 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4812 /* Account for system time used */
4813 acct_update_integrals(p);
4817 * Account for involuntary wait time.
4818 * @steal: the cpu time spent in involuntary wait
4820 void account_steal_time(cputime_t cputime)
4822 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4823 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4825 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4829 * Account for idle time.
4830 * @cputime: the cpu time spent in idle wait
4832 void account_idle_time(cputime_t cputime)
4834 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4835 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4836 struct rq *rq = this_rq();
4838 if (atomic_read(&rq->nr_iowait) > 0)
4839 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4841 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4844 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4847 * Account a single tick of cpu time.
4848 * @p: the process that the cpu time gets accounted to
4849 * @user_tick: indicates if the tick is a user or a system tick
4851 void account_process_tick(struct task_struct *p, int user_tick)
4853 cputime_t one_jiffy = jiffies_to_cputime(1);
4854 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4855 struct rq *rq = this_rq();
4858 account_user_time(p, one_jiffy, one_jiffy_scaled);
4859 else if (p != rq->idle)
4860 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4863 account_idle_time(one_jiffy);
4867 * Account multiple ticks of steal time.
4868 * @p: the process from which the cpu time has been stolen
4869 * @ticks: number of stolen ticks
4871 void account_steal_ticks(unsigned long ticks)
4873 account_steal_time(jiffies_to_cputime(ticks));
4877 * Account multiple ticks of idle time.
4878 * @ticks: number of stolen ticks
4880 void account_idle_ticks(unsigned long ticks)
4882 account_idle_time(jiffies_to_cputime(ticks));
4888 * Use precise platform statistics if available:
4890 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4891 cputime_t task_utime(struct task_struct *p)
4896 cputime_t task_stime(struct task_struct *p)
4901 cputime_t task_utime(struct task_struct *p)
4903 clock_t utime = cputime_to_clock_t(p->utime),
4904 total = utime + cputime_to_clock_t(p->stime);
4908 * Use CFS's precise accounting:
4910 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4914 do_div(temp, total);
4916 utime = (clock_t)temp;
4918 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4919 return p->prev_utime;
4922 cputime_t task_stime(struct task_struct *p)
4927 * Use CFS's precise accounting. (we subtract utime from
4928 * the total, to make sure the total observed by userspace
4929 * grows monotonically - apps rely on that):
4931 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4932 cputime_to_clock_t(task_utime(p));
4935 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4937 return p->prev_stime;
4941 inline cputime_t task_gtime(struct task_struct *p)
4947 * This function gets called by the timer code, with HZ frequency.
4948 * We call it with interrupts disabled.
4950 * It also gets called by the fork code, when changing the parent's
4953 void scheduler_tick(void)
4955 int cpu = smp_processor_id();
4956 struct rq *rq = cpu_rq(cpu);
4957 struct task_struct *curr = rq->curr;
4961 spin_lock(&rq->lock);
4962 update_rq_clock(rq);
4963 update_cpu_load(rq);
4964 curr->sched_class->task_tick(rq, curr, 0);
4965 spin_unlock(&rq->lock);
4968 rq->idle_at_tick = idle_cpu(cpu);
4969 trigger_load_balance(rq, cpu);
4973 unsigned long get_parent_ip(unsigned long addr)
4975 if (in_lock_functions(addr)) {
4976 addr = CALLER_ADDR2;
4977 if (in_lock_functions(addr))
4978 addr = CALLER_ADDR3;
4983 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4984 defined(CONFIG_PREEMPT_TRACER))
4986 void __kprobes add_preempt_count(int val)
4988 #ifdef CONFIG_DEBUG_PREEMPT
4992 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4995 preempt_count() += val;
4996 #ifdef CONFIG_DEBUG_PREEMPT
4998 * Spinlock count overflowing soon?
5000 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
5003 if (preempt_count() == val)
5004 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5006 EXPORT_SYMBOL(add_preempt_count);
5008 void __kprobes sub_preempt_count(int val)
5010 #ifdef CONFIG_DEBUG_PREEMPT
5014 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
5017 * Is the spinlock portion underflowing?
5019 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
5020 !(preempt_count() & PREEMPT_MASK)))
5024 if (preempt_count() == val)
5025 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
5026 preempt_count() -= val;
5028 EXPORT_SYMBOL(sub_preempt_count);
5033 * Print scheduling while atomic bug:
5035 static noinline void __schedule_bug(struct task_struct *prev)
5037 struct pt_regs *regs = get_irq_regs();
5039 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
5040 prev->comm, prev->pid, preempt_count());
5042 debug_show_held_locks(prev);
5044 if (irqs_disabled())
5045 print_irqtrace_events(prev);
5054 * Various schedule()-time debugging checks and statistics:
5056 static inline void schedule_debug(struct task_struct *prev)
5059 * Test if we are atomic. Since do_exit() needs to call into
5060 * schedule() atomically, we ignore that path for now.
5061 * Otherwise, whine if we are scheduling when we should not be.
5063 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
5064 __schedule_bug(prev);
5066 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
5068 schedstat_inc(this_rq(), sched_count);
5069 #ifdef CONFIG_SCHEDSTATS
5070 if (unlikely(prev->lock_depth >= 0)) {
5071 schedstat_inc(this_rq(), bkl_count);
5072 schedstat_inc(prev, sched_info.bkl_count);
5077 static void put_prev_task(struct rq *rq, struct task_struct *prev)
5079 if (prev->state == TASK_RUNNING) {
5080 u64 runtime = prev->se.sum_exec_runtime;
5082 runtime -= prev->se.prev_sum_exec_runtime;
5083 runtime = min_t(u64, runtime, 2*sysctl_sched_migration_cost);
5086 * In order to avoid avg_overlap growing stale when we are
5087 * indeed overlapping and hence not getting put to sleep, grow
5088 * the avg_overlap on preemption.
5090 * We use the average preemption runtime because that
5091 * correlates to the amount of cache footprint a task can
5094 update_avg(&prev->se.avg_overlap, runtime);
5096 prev->sched_class->put_prev_task(rq, prev);
5100 * Pick up the highest-prio task:
5102 static inline struct task_struct *
5103 pick_next_task(struct rq *rq)
5105 const struct sched_class *class;
5106 struct task_struct *p;
5109 * Optimization: we know that if all tasks are in
5110 * the fair class we can call that function directly:
5112 if (likely(rq->nr_running == rq->cfs.nr_running)) {
5113 p = fair_sched_class.pick_next_task(rq);
5118 class = sched_class_highest;
5120 p = class->pick_next_task(rq);
5124 * Will never be NULL as the idle class always
5125 * returns a non-NULL p:
5127 class = class->next;
5132 * schedule() is the main scheduler function.
5134 asmlinkage void __sched schedule(void)
5136 struct task_struct *prev, *next;
5137 unsigned long *switch_count;
5143 cpu = smp_processor_id();
5147 switch_count = &prev->nivcsw;
5149 release_kernel_lock(prev);
5150 need_resched_nonpreemptible:
5152 schedule_debug(prev);
5154 if (sched_feat(HRTICK))
5157 spin_lock_irq(&rq->lock);
5158 update_rq_clock(rq);
5159 clear_tsk_need_resched(prev);
5161 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
5162 if (unlikely(signal_pending_state(prev->state, prev)))
5163 prev->state = TASK_RUNNING;
5165 deactivate_task(rq, prev, 1);
5166 switch_count = &prev->nvcsw;
5170 if (prev->sched_class->pre_schedule)
5171 prev->sched_class->pre_schedule(rq, prev);
5174 if (unlikely(!rq->nr_running))
5175 idle_balance(cpu, rq);
5177 put_prev_task(rq, prev);
5178 next = pick_next_task(rq);
5180 if (likely(prev != next)) {
5181 sched_info_switch(prev, next);
5187 context_switch(rq, prev, next); /* unlocks the rq */
5189 * the context switch might have flipped the stack from under
5190 * us, hence refresh the local variables.
5192 cpu = smp_processor_id();
5195 spin_unlock_irq(&rq->lock);
5197 if (unlikely(reacquire_kernel_lock(current) < 0))
5198 goto need_resched_nonpreemptible;
5200 preempt_enable_no_resched();
5204 EXPORT_SYMBOL(schedule);
5208 * Look out! "owner" is an entirely speculative pointer
5209 * access and not reliable.
5211 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
5216 if (!sched_feat(OWNER_SPIN))
5219 #ifdef CONFIG_DEBUG_PAGEALLOC
5221 * Need to access the cpu field knowing that
5222 * DEBUG_PAGEALLOC could have unmapped it if
5223 * the mutex owner just released it and exited.
5225 if (probe_kernel_address(&owner->cpu, cpu))
5232 * Even if the access succeeded (likely case),
5233 * the cpu field may no longer be valid.
5235 if (cpu >= nr_cpumask_bits)
5239 * We need to validate that we can do a
5240 * get_cpu() and that we have the percpu area.
5242 if (!cpu_online(cpu))
5249 * Owner changed, break to re-assess state.
5251 if (lock->owner != owner)
5255 * Is that owner really running on that cpu?
5257 if (task_thread_info(rq->curr) != owner || need_resched())
5267 #ifdef CONFIG_PREEMPT
5269 * this is the entry point to schedule() from in-kernel preemption
5270 * off of preempt_enable. Kernel preemptions off return from interrupt
5271 * occur there and call schedule directly.
5273 asmlinkage void __sched preempt_schedule(void)
5275 struct thread_info *ti = current_thread_info();
5278 * If there is a non-zero preempt_count or interrupts are disabled,
5279 * we do not want to preempt the current task. Just return..
5281 if (likely(ti->preempt_count || irqs_disabled()))
5285 add_preempt_count(PREEMPT_ACTIVE);
5287 sub_preempt_count(PREEMPT_ACTIVE);
5290 * Check again in case we missed a preemption opportunity
5291 * between schedule and now.
5294 } while (need_resched());
5296 EXPORT_SYMBOL(preempt_schedule);
5299 * this is the entry point to schedule() from kernel preemption
5300 * off of irq context.
5301 * Note, that this is called and return with irqs disabled. This will
5302 * protect us against recursive calling from irq.
5304 asmlinkage void __sched preempt_schedule_irq(void)
5306 struct thread_info *ti = current_thread_info();
5308 /* Catch callers which need to be fixed */
5309 BUG_ON(ti->preempt_count || !irqs_disabled());
5312 add_preempt_count(PREEMPT_ACTIVE);
5315 local_irq_disable();
5316 sub_preempt_count(PREEMPT_ACTIVE);
5319 * Check again in case we missed a preemption opportunity
5320 * between schedule and now.
5323 } while (need_resched());
5326 #endif /* CONFIG_PREEMPT */
5328 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
5331 return try_to_wake_up(curr->private, mode, sync);
5333 EXPORT_SYMBOL(default_wake_function);
5336 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5337 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5338 * number) then we wake all the non-exclusive tasks and one exclusive task.
5340 * There are circumstances in which we can try to wake a task which has already
5341 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5342 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5344 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
5345 int nr_exclusive, int sync, void *key)
5347 wait_queue_t *curr, *next;
5349 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
5350 unsigned flags = curr->flags;
5352 if (curr->func(curr, mode, sync, key) &&
5353 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
5359 * __wake_up - wake up threads blocked on a waitqueue.
5361 * @mode: which threads
5362 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5363 * @key: is directly passed to the wakeup function
5365 void __wake_up(wait_queue_head_t *q, unsigned int mode,
5366 int nr_exclusive, void *key)
5368 unsigned long flags;
5370 spin_lock_irqsave(&q->lock, flags);
5371 __wake_up_common(q, mode, nr_exclusive, 0, key);
5372 spin_unlock_irqrestore(&q->lock, flags);
5374 EXPORT_SYMBOL(__wake_up);
5377 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5379 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
5381 __wake_up_common(q, mode, 1, 0, NULL);
5384 void __wake_up_locked_key(wait_queue_head_t *q, unsigned int mode, void *key)
5386 __wake_up_common(q, mode, 1, 0, key);
5390 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5392 * @mode: which threads
5393 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5394 * @key: opaque value to be passed to wakeup targets
5396 * The sync wakeup differs that the waker knows that it will schedule
5397 * away soon, so while the target thread will be woken up, it will not
5398 * be migrated to another CPU - ie. the two threads are 'synchronized'
5399 * with each other. This can prevent needless bouncing between CPUs.
5401 * On UP it can prevent extra preemption.
5403 void __wake_up_sync_key(wait_queue_head_t *q, unsigned int mode,
5404 int nr_exclusive, void *key)
5406 unsigned long flags;
5412 if (unlikely(!nr_exclusive))
5415 spin_lock_irqsave(&q->lock, flags);
5416 __wake_up_common(q, mode, nr_exclusive, sync, key);
5417 spin_unlock_irqrestore(&q->lock, flags);
5419 EXPORT_SYMBOL_GPL(__wake_up_sync_key);
5422 * __wake_up_sync - see __wake_up_sync_key()
5424 void __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
5426 __wake_up_sync_key(q, mode, nr_exclusive, NULL);
5428 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
5431 * complete: - signals a single thread waiting on this completion
5432 * @x: holds the state of this particular completion
5434 * This will wake up a single thread waiting on this completion. Threads will be
5435 * awakened in the same order in which they were queued.
5437 * See also complete_all(), wait_for_completion() and related routines.
5439 void complete(struct completion *x)
5441 unsigned long flags;
5443 spin_lock_irqsave(&x->wait.lock, flags);
5445 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
5446 spin_unlock_irqrestore(&x->wait.lock, flags);
5448 EXPORT_SYMBOL(complete);
5451 * complete_all: - signals all threads waiting on this completion
5452 * @x: holds the state of this particular completion
5454 * This will wake up all threads waiting on this particular completion event.
5456 void complete_all(struct completion *x)
5458 unsigned long flags;
5460 spin_lock_irqsave(&x->wait.lock, flags);
5461 x->done += UINT_MAX/2;
5462 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
5463 spin_unlock_irqrestore(&x->wait.lock, flags);
5465 EXPORT_SYMBOL(complete_all);
5467 static inline long __sched
5468 do_wait_for_common(struct completion *x, long timeout, int state)
5471 DECLARE_WAITQUEUE(wait, current);
5473 wait.flags |= WQ_FLAG_EXCLUSIVE;
5474 __add_wait_queue_tail(&x->wait, &wait);
5476 if (signal_pending_state(state, current)) {
5477 timeout = -ERESTARTSYS;
5480 __set_current_state(state);
5481 spin_unlock_irq(&x->wait.lock);
5482 timeout = schedule_timeout(timeout);
5483 spin_lock_irq(&x->wait.lock);
5484 } while (!x->done && timeout);
5485 __remove_wait_queue(&x->wait, &wait);
5490 return timeout ?: 1;
5494 wait_for_common(struct completion *x, long timeout, int state)
5498 spin_lock_irq(&x->wait.lock);
5499 timeout = do_wait_for_common(x, timeout, state);
5500 spin_unlock_irq(&x->wait.lock);
5505 * wait_for_completion: - waits for completion of a task
5506 * @x: holds the state of this particular completion
5508 * This waits to be signaled for completion of a specific task. It is NOT
5509 * interruptible and there is no timeout.
5511 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5512 * and interrupt capability. Also see complete().
5514 void __sched wait_for_completion(struct completion *x)
5516 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
5518 EXPORT_SYMBOL(wait_for_completion);
5521 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5522 * @x: holds the state of this particular completion
5523 * @timeout: timeout value in jiffies
5525 * This waits for either a completion of a specific task to be signaled or for a
5526 * specified timeout to expire. The timeout is in jiffies. It is not
5529 unsigned long __sched
5530 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5532 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5534 EXPORT_SYMBOL(wait_for_completion_timeout);
5537 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5538 * @x: holds the state of this particular completion
5540 * This waits for completion of a specific task to be signaled. It is
5543 int __sched wait_for_completion_interruptible(struct completion *x)
5545 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5546 if (t == -ERESTARTSYS)
5550 EXPORT_SYMBOL(wait_for_completion_interruptible);
5553 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5554 * @x: holds the state of this particular completion
5555 * @timeout: timeout value in jiffies
5557 * This waits for either a completion of a specific task to be signaled or for a
5558 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5560 unsigned long __sched
5561 wait_for_completion_interruptible_timeout(struct completion *x,
5562 unsigned long timeout)
5564 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5566 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5569 * wait_for_completion_killable: - waits for completion of a task (killable)
5570 * @x: holds the state of this particular completion
5572 * This waits to be signaled for completion of a specific task. It can be
5573 * interrupted by a kill signal.
5575 int __sched wait_for_completion_killable(struct completion *x)
5577 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5578 if (t == -ERESTARTSYS)
5582 EXPORT_SYMBOL(wait_for_completion_killable);
5585 * try_wait_for_completion - try to decrement a completion without blocking
5586 * @x: completion structure
5588 * Returns: 0 if a decrement cannot be done without blocking
5589 * 1 if a decrement succeeded.
5591 * If a completion is being used as a counting completion,
5592 * attempt to decrement the counter without blocking. This
5593 * enables us to avoid waiting if the resource the completion
5594 * is protecting is not available.
5596 bool try_wait_for_completion(struct completion *x)
5600 spin_lock_irq(&x->wait.lock);
5605 spin_unlock_irq(&x->wait.lock);
5608 EXPORT_SYMBOL(try_wait_for_completion);
5611 * completion_done - Test to see if a completion has any waiters
5612 * @x: completion structure
5614 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5615 * 1 if there are no waiters.
5618 bool completion_done(struct completion *x)
5622 spin_lock_irq(&x->wait.lock);
5625 spin_unlock_irq(&x->wait.lock);
5628 EXPORT_SYMBOL(completion_done);
5631 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5633 unsigned long flags;
5636 init_waitqueue_entry(&wait, current);
5638 __set_current_state(state);
5640 spin_lock_irqsave(&q->lock, flags);
5641 __add_wait_queue(q, &wait);
5642 spin_unlock(&q->lock);
5643 timeout = schedule_timeout(timeout);
5644 spin_lock_irq(&q->lock);
5645 __remove_wait_queue(q, &wait);
5646 spin_unlock_irqrestore(&q->lock, flags);
5651 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5653 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5655 EXPORT_SYMBOL(interruptible_sleep_on);
5658 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5660 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5662 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5664 void __sched sleep_on(wait_queue_head_t *q)
5666 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5668 EXPORT_SYMBOL(sleep_on);
5670 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5672 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5674 EXPORT_SYMBOL(sleep_on_timeout);
5676 #ifdef CONFIG_RT_MUTEXES
5679 * rt_mutex_setprio - set the current priority of a task
5681 * @prio: prio value (kernel-internal form)
5683 * This function changes the 'effective' priority of a task. It does
5684 * not touch ->normal_prio like __setscheduler().
5686 * Used by the rt_mutex code to implement priority inheritance logic.
5688 void rt_mutex_setprio(struct task_struct *p, int prio)
5690 unsigned long flags;
5691 int oldprio, on_rq, running;
5693 const struct sched_class *prev_class = p->sched_class;
5695 BUG_ON(prio < 0 || prio > MAX_PRIO);
5697 rq = task_rq_lock(p, &flags);
5698 update_rq_clock(rq);
5701 on_rq = p->se.on_rq;
5702 running = task_current(rq, p);
5704 dequeue_task(rq, p, 0);
5706 p->sched_class->put_prev_task(rq, p);
5709 p->sched_class = &rt_sched_class;
5711 p->sched_class = &fair_sched_class;
5716 p->sched_class->set_curr_task(rq);
5718 enqueue_task(rq, p, 0);
5720 check_class_changed(rq, p, prev_class, oldprio, running);
5722 task_rq_unlock(rq, &flags);
5727 void set_user_nice(struct task_struct *p, long nice)
5729 int old_prio, delta, on_rq;
5730 unsigned long flags;
5733 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5736 * We have to be careful, if called from sys_setpriority(),
5737 * the task might be in the middle of scheduling on another CPU.
5739 rq = task_rq_lock(p, &flags);
5740 update_rq_clock(rq);
5742 * The RT priorities are set via sched_setscheduler(), but we still
5743 * allow the 'normal' nice value to be set - but as expected
5744 * it wont have any effect on scheduling until the task is
5745 * SCHED_FIFO/SCHED_RR:
5747 if (task_has_rt_policy(p)) {
5748 p->static_prio = NICE_TO_PRIO(nice);
5751 on_rq = p->se.on_rq;
5753 dequeue_task(rq, p, 0);
5755 p->static_prio = NICE_TO_PRIO(nice);
5758 p->prio = effective_prio(p);
5759 delta = p->prio - old_prio;
5762 enqueue_task(rq, p, 0);
5764 * If the task increased its priority or is running and
5765 * lowered its priority, then reschedule its CPU:
5767 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5768 resched_task(rq->curr);
5771 task_rq_unlock(rq, &flags);
5773 EXPORT_SYMBOL(set_user_nice);
5776 * can_nice - check if a task can reduce its nice value
5780 int can_nice(const struct task_struct *p, const int nice)
5782 /* convert nice value [19,-20] to rlimit style value [1,40] */
5783 int nice_rlim = 20 - nice;
5785 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5786 capable(CAP_SYS_NICE));
5789 #ifdef __ARCH_WANT_SYS_NICE
5792 * sys_nice - change the priority of the current process.
5793 * @increment: priority increment
5795 * sys_setpriority is a more generic, but much slower function that
5796 * does similar things.
5798 SYSCALL_DEFINE1(nice, int, increment)
5803 * Setpriority might change our priority at the same moment.
5804 * We don't have to worry. Conceptually one call occurs first
5805 * and we have a single winner.
5807 if (increment < -40)
5812 nice = TASK_NICE(current) + increment;
5818 if (increment < 0 && !can_nice(current, nice))
5821 retval = security_task_setnice(current, nice);
5825 set_user_nice(current, nice);
5832 * task_prio - return the priority value of a given task.
5833 * @p: the task in question.
5835 * This is the priority value as seen by users in /proc.
5836 * RT tasks are offset by -200. Normal tasks are centered
5837 * around 0, value goes from -16 to +15.
5839 int task_prio(const struct task_struct *p)
5841 return p->prio - MAX_RT_PRIO;
5845 * task_nice - return the nice value of a given task.
5846 * @p: the task in question.
5848 int task_nice(const struct task_struct *p)
5850 return TASK_NICE(p);
5852 EXPORT_SYMBOL(task_nice);
5855 * idle_cpu - is a given cpu idle currently?
5856 * @cpu: the processor in question.
5858 int idle_cpu(int cpu)
5860 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5864 * idle_task - return the idle task for a given cpu.
5865 * @cpu: the processor in question.
5867 struct task_struct *idle_task(int cpu)
5869 return cpu_rq(cpu)->idle;
5873 * find_process_by_pid - find a process with a matching PID value.
5874 * @pid: the pid in question.
5876 static struct task_struct *find_process_by_pid(pid_t pid)
5878 return pid ? find_task_by_vpid(pid) : current;
5881 /* Actually do priority change: must hold rq lock. */
5883 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5885 BUG_ON(p->se.on_rq);
5888 switch (p->policy) {
5892 p->sched_class = &fair_sched_class;
5896 p->sched_class = &rt_sched_class;
5900 p->rt_priority = prio;
5901 p->normal_prio = normal_prio(p);
5902 /* we are holding p->pi_lock already */
5903 p->prio = rt_mutex_getprio(p);
5908 * check the target process has a UID that matches the current process's
5910 static bool check_same_owner(struct task_struct *p)
5912 const struct cred *cred = current_cred(), *pcred;
5916 pcred = __task_cred(p);
5917 match = (cred->euid == pcred->euid ||
5918 cred->euid == pcred->uid);
5923 static int __sched_setscheduler(struct task_struct *p, int policy,
5924 struct sched_param *param, bool user)
5926 int retval, oldprio, oldpolicy = -1, on_rq, running;
5927 unsigned long flags;
5928 const struct sched_class *prev_class = p->sched_class;
5931 /* may grab non-irq protected spin_locks */
5932 BUG_ON(in_interrupt());
5934 /* double check policy once rq lock held */
5936 policy = oldpolicy = p->policy;
5937 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5938 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5939 policy != SCHED_IDLE)
5942 * Valid priorities for SCHED_FIFO and SCHED_RR are
5943 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5944 * SCHED_BATCH and SCHED_IDLE is 0.
5946 if (param->sched_priority < 0 ||
5947 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5948 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5950 if (rt_policy(policy) != (param->sched_priority != 0))
5954 * Allow unprivileged RT tasks to decrease priority:
5956 if (user && !capable(CAP_SYS_NICE)) {
5957 if (rt_policy(policy)) {
5958 unsigned long rlim_rtprio;
5960 if (!lock_task_sighand(p, &flags))
5962 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5963 unlock_task_sighand(p, &flags);
5965 /* can't set/change the rt policy */
5966 if (policy != p->policy && !rlim_rtprio)
5969 /* can't increase priority */
5970 if (param->sched_priority > p->rt_priority &&
5971 param->sched_priority > rlim_rtprio)
5975 * Like positive nice levels, dont allow tasks to
5976 * move out of SCHED_IDLE either:
5978 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5981 /* can't change other user's priorities */
5982 if (!check_same_owner(p))
5987 #ifdef CONFIG_RT_GROUP_SCHED
5989 * Do not allow realtime tasks into groups that have no runtime
5992 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5993 task_group(p)->rt_bandwidth.rt_runtime == 0)
5997 retval = security_task_setscheduler(p, policy, param);
6003 * make sure no PI-waiters arrive (or leave) while we are
6004 * changing the priority of the task:
6006 spin_lock_irqsave(&p->pi_lock, flags);
6008 * To be able to change p->policy safely, the apropriate
6009 * runqueue lock must be held.
6011 rq = __task_rq_lock(p);
6012 /* recheck policy now with rq lock held */
6013 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
6014 policy = oldpolicy = -1;
6015 __task_rq_unlock(rq);
6016 spin_unlock_irqrestore(&p->pi_lock, flags);
6019 update_rq_clock(rq);
6020 on_rq = p->se.on_rq;
6021 running = task_current(rq, p);
6023 deactivate_task(rq, p, 0);
6025 p->sched_class->put_prev_task(rq, p);
6028 __setscheduler(rq, p, policy, param->sched_priority);
6031 p->sched_class->set_curr_task(rq);
6033 activate_task(rq, p, 0);
6035 check_class_changed(rq, p, prev_class, oldprio, running);
6037 __task_rq_unlock(rq);
6038 spin_unlock_irqrestore(&p->pi_lock, flags);
6040 rt_mutex_adjust_pi(p);
6046 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6047 * @p: the task in question.
6048 * @policy: new policy.
6049 * @param: structure containing the new RT priority.
6051 * NOTE that the task may be already dead.
6053 int sched_setscheduler(struct task_struct *p, int policy,
6054 struct sched_param *param)
6056 return __sched_setscheduler(p, policy, param, true);
6058 EXPORT_SYMBOL_GPL(sched_setscheduler);
6061 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6062 * @p: the task in question.
6063 * @policy: new policy.
6064 * @param: structure containing the new RT priority.
6066 * Just like sched_setscheduler, only don't bother checking if the
6067 * current context has permission. For example, this is needed in
6068 * stop_machine(): we create temporary high priority worker threads,
6069 * but our caller might not have that capability.
6071 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
6072 struct sched_param *param)
6074 return __sched_setscheduler(p, policy, param, false);
6078 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
6080 struct sched_param lparam;
6081 struct task_struct *p;
6084 if (!param || pid < 0)
6086 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
6091 p = find_process_by_pid(pid);
6093 retval = sched_setscheduler(p, policy, &lparam);
6100 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6101 * @pid: the pid in question.
6102 * @policy: new policy.
6103 * @param: structure containing the new RT priority.
6105 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
6106 struct sched_param __user *, param)
6108 /* negative values for policy are not valid */
6112 return do_sched_setscheduler(pid, policy, param);
6116 * sys_sched_setparam - set/change the RT priority of a thread
6117 * @pid: the pid in question.
6118 * @param: structure containing the new RT priority.
6120 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
6122 return do_sched_setscheduler(pid, -1, param);
6126 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6127 * @pid: the pid in question.
6129 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
6131 struct task_struct *p;
6138 read_lock(&tasklist_lock);
6139 p = find_process_by_pid(pid);
6141 retval = security_task_getscheduler(p);
6145 read_unlock(&tasklist_lock);
6150 * sys_sched_getscheduler - get the RT priority of a thread
6151 * @pid: the pid in question.
6152 * @param: structure containing the RT priority.
6154 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
6156 struct sched_param lp;
6157 struct task_struct *p;
6160 if (!param || pid < 0)
6163 read_lock(&tasklist_lock);
6164 p = find_process_by_pid(pid);
6169 retval = security_task_getscheduler(p);
6173 lp.sched_priority = p->rt_priority;
6174 read_unlock(&tasklist_lock);
6177 * This one might sleep, we cannot do it with a spinlock held ...
6179 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
6184 read_unlock(&tasklist_lock);
6188 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
6190 cpumask_var_t cpus_allowed, new_mask;
6191 struct task_struct *p;
6195 read_lock(&tasklist_lock);
6197 p = find_process_by_pid(pid);
6199 read_unlock(&tasklist_lock);
6205 * It is not safe to call set_cpus_allowed with the
6206 * tasklist_lock held. We will bump the task_struct's
6207 * usage count and then drop tasklist_lock.
6210 read_unlock(&tasklist_lock);
6212 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
6216 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
6218 goto out_free_cpus_allowed;
6221 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
6224 retval = security_task_setscheduler(p, 0, NULL);
6228 cpuset_cpus_allowed(p, cpus_allowed);
6229 cpumask_and(new_mask, in_mask, cpus_allowed);
6231 retval = set_cpus_allowed_ptr(p, new_mask);
6234 cpuset_cpus_allowed(p, cpus_allowed);
6235 if (!cpumask_subset(new_mask, cpus_allowed)) {
6237 * We must have raced with a concurrent cpuset
6238 * update. Just reset the cpus_allowed to the
6239 * cpuset's cpus_allowed
6241 cpumask_copy(new_mask, cpus_allowed);
6246 free_cpumask_var(new_mask);
6247 out_free_cpus_allowed:
6248 free_cpumask_var(cpus_allowed);
6255 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
6256 struct cpumask *new_mask)
6258 if (len < cpumask_size())
6259 cpumask_clear(new_mask);
6260 else if (len > cpumask_size())
6261 len = cpumask_size();
6263 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
6267 * sys_sched_setaffinity - set the cpu affinity of a process
6268 * @pid: pid of the process
6269 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6270 * @user_mask_ptr: user-space pointer to the new cpu mask
6272 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
6273 unsigned long __user *, user_mask_ptr)
6275 cpumask_var_t new_mask;
6278 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
6281 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
6283 retval = sched_setaffinity(pid, new_mask);
6284 free_cpumask_var(new_mask);
6288 long sched_getaffinity(pid_t pid, struct cpumask *mask)
6290 struct task_struct *p;
6294 read_lock(&tasklist_lock);
6297 p = find_process_by_pid(pid);
6301 retval = security_task_getscheduler(p);
6305 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
6308 read_unlock(&tasklist_lock);
6315 * sys_sched_getaffinity - get the cpu affinity of a process
6316 * @pid: pid of the process
6317 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6318 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6320 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
6321 unsigned long __user *, user_mask_ptr)
6326 if (len < cpumask_size())
6329 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
6332 ret = sched_getaffinity(pid, mask);
6334 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
6337 ret = cpumask_size();
6339 free_cpumask_var(mask);
6345 * sys_sched_yield - yield the current processor to other threads.
6347 * This function yields the current CPU to other tasks. If there are no
6348 * other threads running on this CPU then this function will return.
6350 SYSCALL_DEFINE0(sched_yield)
6352 struct rq *rq = this_rq_lock();
6354 schedstat_inc(rq, yld_count);
6355 current->sched_class->yield_task(rq);
6358 * Since we are going to call schedule() anyway, there's
6359 * no need to preempt or enable interrupts:
6361 __release(rq->lock);
6362 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
6363 _raw_spin_unlock(&rq->lock);
6364 preempt_enable_no_resched();
6371 static void __cond_resched(void)
6373 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6374 __might_sleep(__FILE__, __LINE__);
6377 * The BKS might be reacquired before we have dropped
6378 * PREEMPT_ACTIVE, which could trigger a second
6379 * cond_resched() call.
6382 add_preempt_count(PREEMPT_ACTIVE);
6384 sub_preempt_count(PREEMPT_ACTIVE);
6385 } while (need_resched());
6388 int __sched _cond_resched(void)
6390 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
6391 system_state == SYSTEM_RUNNING) {
6397 EXPORT_SYMBOL(_cond_resched);
6400 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6401 * call schedule, and on return reacquire the lock.
6403 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6404 * operations here to prevent schedule() from being called twice (once via
6405 * spin_unlock(), once by hand).
6407 int cond_resched_lock(spinlock_t *lock)
6409 int resched = need_resched() && system_state == SYSTEM_RUNNING;
6412 if (spin_needbreak(lock) || resched) {
6414 if (resched && need_resched())
6423 EXPORT_SYMBOL(cond_resched_lock);
6425 int __sched cond_resched_softirq(void)
6427 BUG_ON(!in_softirq());
6429 if (need_resched() && system_state == SYSTEM_RUNNING) {
6437 EXPORT_SYMBOL(cond_resched_softirq);
6440 * yield - yield the current processor to other threads.
6442 * This is a shortcut for kernel-space yielding - it marks the
6443 * thread runnable and calls sys_sched_yield().
6445 void __sched yield(void)
6447 set_current_state(TASK_RUNNING);
6450 EXPORT_SYMBOL(yield);
6453 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6454 * that process accounting knows that this is a task in IO wait state.
6456 * But don't do that if it is a deliberate, throttling IO wait (this task
6457 * has set its backing_dev_info: the queue against which it should throttle)
6459 void __sched io_schedule(void)
6461 struct rq *rq = &__raw_get_cpu_var(runqueues);
6463 delayacct_blkio_start();
6464 atomic_inc(&rq->nr_iowait);
6466 atomic_dec(&rq->nr_iowait);
6467 delayacct_blkio_end();
6469 EXPORT_SYMBOL(io_schedule);
6471 long __sched io_schedule_timeout(long timeout)
6473 struct rq *rq = &__raw_get_cpu_var(runqueues);
6476 delayacct_blkio_start();
6477 atomic_inc(&rq->nr_iowait);
6478 ret = schedule_timeout(timeout);
6479 atomic_dec(&rq->nr_iowait);
6480 delayacct_blkio_end();
6485 * sys_sched_get_priority_max - return maximum RT priority.
6486 * @policy: scheduling class.
6488 * this syscall returns the maximum rt_priority that can be used
6489 * by a given scheduling class.
6491 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
6498 ret = MAX_USER_RT_PRIO-1;
6510 * sys_sched_get_priority_min - return minimum RT priority.
6511 * @policy: scheduling class.
6513 * this syscall returns the minimum rt_priority that can be used
6514 * by a given scheduling class.
6516 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6534 * sys_sched_rr_get_interval - return the default timeslice of a process.
6535 * @pid: pid of the process.
6536 * @interval: userspace pointer to the timeslice value.
6538 * this syscall writes the default timeslice value of a given process
6539 * into the user-space timespec buffer. A value of '0' means infinity.
6541 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6542 struct timespec __user *, interval)
6544 struct task_struct *p;
6545 unsigned int time_slice;
6553 read_lock(&tasklist_lock);
6554 p = find_process_by_pid(pid);
6558 retval = security_task_getscheduler(p);
6563 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6564 * tasks that are on an otherwise idle runqueue:
6567 if (p->policy == SCHED_RR) {
6568 time_slice = DEF_TIMESLICE;
6569 } else if (p->policy != SCHED_FIFO) {
6570 struct sched_entity *se = &p->se;
6571 unsigned long flags;
6574 rq = task_rq_lock(p, &flags);
6575 if (rq->cfs.load.weight)
6576 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6577 task_rq_unlock(rq, &flags);
6579 read_unlock(&tasklist_lock);
6580 jiffies_to_timespec(time_slice, &t);
6581 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6585 read_unlock(&tasklist_lock);
6589 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6591 void sched_show_task(struct task_struct *p)
6593 unsigned long free = 0;
6596 state = p->state ? __ffs(p->state) + 1 : 0;
6597 printk(KERN_INFO "%-13.13s %c", p->comm,
6598 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6599 #if BITS_PER_LONG == 32
6600 if (state == TASK_RUNNING)
6601 printk(KERN_CONT " running ");
6603 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6605 if (state == TASK_RUNNING)
6606 printk(KERN_CONT " running task ");
6608 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6610 #ifdef CONFIG_DEBUG_STACK_USAGE
6611 free = stack_not_used(p);
6613 printk(KERN_CONT "%5lu %5d %6d 0x%08lx\n", free,
6614 task_pid_nr(p), task_pid_nr(p->real_parent),
6615 (unsigned long)task_thread_info(p)->flags);
6617 show_stack(p, NULL);
6620 void show_state_filter(unsigned long state_filter)
6622 struct task_struct *g, *p;
6624 #if BITS_PER_LONG == 32
6626 " task PC stack pid father\n");
6629 " task PC stack pid father\n");
6631 read_lock(&tasklist_lock);
6632 do_each_thread(g, p) {
6634 * reset the NMI-timeout, listing all files on a slow
6635 * console might take alot of time:
6637 touch_nmi_watchdog();
6638 if (!state_filter || (p->state & state_filter))
6640 } while_each_thread(g, p);
6642 touch_all_softlockup_watchdogs();
6644 #ifdef CONFIG_SCHED_DEBUG
6645 sysrq_sched_debug_show();
6647 read_unlock(&tasklist_lock);
6649 * Only show locks if all tasks are dumped:
6651 if (state_filter == -1)
6652 debug_show_all_locks();
6655 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6657 idle->sched_class = &idle_sched_class;
6661 * init_idle - set up an idle thread for a given CPU
6662 * @idle: task in question
6663 * @cpu: cpu the idle task belongs to
6665 * NOTE: this function does not set the idle thread's NEED_RESCHED
6666 * flag, to make booting more robust.
6668 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6670 struct rq *rq = cpu_rq(cpu);
6671 unsigned long flags;
6673 spin_lock_irqsave(&rq->lock, flags);
6676 idle->se.exec_start = sched_clock();
6678 idle->prio = idle->normal_prio = MAX_PRIO;
6679 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6680 __set_task_cpu(idle, cpu);
6682 rq->curr = rq->idle = idle;
6683 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6686 spin_unlock_irqrestore(&rq->lock, flags);
6688 /* Set the preempt count _outside_ the spinlocks! */
6689 #if defined(CONFIG_PREEMPT)
6690 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6692 task_thread_info(idle)->preempt_count = 0;
6695 * The idle tasks have their own, simple scheduling class:
6697 idle->sched_class = &idle_sched_class;
6698 ftrace_graph_init_task(idle);
6702 * In a system that switches off the HZ timer nohz_cpu_mask
6703 * indicates which cpus entered this state. This is used
6704 * in the rcu update to wait only for active cpus. For system
6705 * which do not switch off the HZ timer nohz_cpu_mask should
6706 * always be CPU_BITS_NONE.
6708 cpumask_var_t nohz_cpu_mask;
6711 * Increase the granularity value when there are more CPUs,
6712 * because with more CPUs the 'effective latency' as visible
6713 * to users decreases. But the relationship is not linear,
6714 * so pick a second-best guess by going with the log2 of the
6717 * This idea comes from the SD scheduler of Con Kolivas:
6719 static inline void sched_init_granularity(void)
6721 unsigned int factor = 1 + ilog2(num_online_cpus());
6722 const unsigned long limit = 200000000;
6724 sysctl_sched_min_granularity *= factor;
6725 if (sysctl_sched_min_granularity > limit)
6726 sysctl_sched_min_granularity = limit;
6728 sysctl_sched_latency *= factor;
6729 if (sysctl_sched_latency > limit)
6730 sysctl_sched_latency = limit;
6732 sysctl_sched_wakeup_granularity *= factor;
6734 sysctl_sched_shares_ratelimit *= factor;
6739 * This is how migration works:
6741 * 1) we queue a struct migration_req structure in the source CPU's
6742 * runqueue and wake up that CPU's migration thread.
6743 * 2) we down() the locked semaphore => thread blocks.
6744 * 3) migration thread wakes up (implicitly it forces the migrated
6745 * thread off the CPU)
6746 * 4) it gets the migration request and checks whether the migrated
6747 * task is still in the wrong runqueue.
6748 * 5) if it's in the wrong runqueue then the migration thread removes
6749 * it and puts it into the right queue.
6750 * 6) migration thread up()s the semaphore.
6751 * 7) we wake up and the migration is done.
6755 * Change a given task's CPU affinity. Migrate the thread to a
6756 * proper CPU and schedule it away if the CPU it's executing on
6757 * is removed from the allowed bitmask.
6759 * NOTE: the caller must have a valid reference to the task, the
6760 * task must not exit() & deallocate itself prematurely. The
6761 * call is not atomic; no spinlocks may be held.
6763 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6765 struct migration_req req;
6766 unsigned long flags;
6770 rq = task_rq_lock(p, &flags);
6771 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6776 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6777 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6782 if (p->sched_class->set_cpus_allowed)
6783 p->sched_class->set_cpus_allowed(p, new_mask);
6785 cpumask_copy(&p->cpus_allowed, new_mask);
6786 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6789 /* Can the task run on the task's current CPU? If so, we're done */
6790 if (cpumask_test_cpu(task_cpu(p), new_mask))
6793 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6794 /* Need help from migration thread: drop lock and wait. */
6795 task_rq_unlock(rq, &flags);
6796 wake_up_process(rq->migration_thread);
6797 wait_for_completion(&req.done);
6798 tlb_migrate_finish(p->mm);
6802 task_rq_unlock(rq, &flags);
6806 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6809 * Move (not current) task off this cpu, onto dest cpu. We're doing
6810 * this because either it can't run here any more (set_cpus_allowed()
6811 * away from this CPU, or CPU going down), or because we're
6812 * attempting to rebalance this task on exec (sched_exec).
6814 * So we race with normal scheduler movements, but that's OK, as long
6815 * as the task is no longer on this CPU.
6817 * Returns non-zero if task was successfully migrated.
6819 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6821 struct rq *rq_dest, *rq_src;
6824 if (unlikely(!cpu_active(dest_cpu)))
6827 rq_src = cpu_rq(src_cpu);
6828 rq_dest = cpu_rq(dest_cpu);
6830 double_rq_lock(rq_src, rq_dest);
6831 /* Already moved. */
6832 if (task_cpu(p) != src_cpu)
6834 /* Affinity changed (again). */
6835 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6838 on_rq = p->se.on_rq;
6840 deactivate_task(rq_src, p, 0);
6842 set_task_cpu(p, dest_cpu);
6844 activate_task(rq_dest, p, 0);
6845 check_preempt_curr(rq_dest, p, 0);
6850 double_rq_unlock(rq_src, rq_dest);
6855 * migration_thread - this is a highprio system thread that performs
6856 * thread migration by bumping thread off CPU then 'pushing' onto
6859 static int migration_thread(void *data)
6861 int cpu = (long)data;
6865 BUG_ON(rq->migration_thread != current);
6867 set_current_state(TASK_INTERRUPTIBLE);
6868 while (!kthread_should_stop()) {
6869 struct migration_req *req;
6870 struct list_head *head;
6872 spin_lock_irq(&rq->lock);
6874 if (cpu_is_offline(cpu)) {
6875 spin_unlock_irq(&rq->lock);
6879 if (rq->active_balance) {
6880 active_load_balance(rq, cpu);
6881 rq->active_balance = 0;
6884 head = &rq->migration_queue;
6886 if (list_empty(head)) {
6887 spin_unlock_irq(&rq->lock);
6889 set_current_state(TASK_INTERRUPTIBLE);
6892 req = list_entry(head->next, struct migration_req, list);
6893 list_del_init(head->next);
6895 spin_unlock(&rq->lock);
6896 __migrate_task(req->task, cpu, req->dest_cpu);
6899 complete(&req->done);
6901 __set_current_state(TASK_RUNNING);
6905 /* Wait for kthread_stop */
6906 set_current_state(TASK_INTERRUPTIBLE);
6907 while (!kthread_should_stop()) {
6909 set_current_state(TASK_INTERRUPTIBLE);
6911 __set_current_state(TASK_RUNNING);
6915 #ifdef CONFIG_HOTPLUG_CPU
6917 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6921 local_irq_disable();
6922 ret = __migrate_task(p, src_cpu, dest_cpu);
6928 * Figure out where task on dead CPU should go, use force if necessary.
6930 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6933 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6936 /* Look for allowed, online CPU in same node. */
6937 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6938 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6941 /* Any allowed, online CPU? */
6942 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6943 if (dest_cpu < nr_cpu_ids)
6946 /* No more Mr. Nice Guy. */
6947 if (dest_cpu >= nr_cpu_ids) {
6948 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6949 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6952 * Don't tell them about moving exiting tasks or
6953 * kernel threads (both mm NULL), since they never
6956 if (p->mm && printk_ratelimit()) {
6957 printk(KERN_INFO "process %d (%s) no "
6958 "longer affine to cpu%d\n",
6959 task_pid_nr(p), p->comm, dead_cpu);
6964 /* It can have affinity changed while we were choosing. */
6965 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6970 * While a dead CPU has no uninterruptible tasks queued at this point,
6971 * it might still have a nonzero ->nr_uninterruptible counter, because
6972 * for performance reasons the counter is not stricly tracking tasks to
6973 * their home CPUs. So we just add the counter to another CPU's counter,
6974 * to keep the global sum constant after CPU-down:
6976 static void migrate_nr_uninterruptible(struct rq *rq_src)
6978 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6979 unsigned long flags;
6981 local_irq_save(flags);
6982 double_rq_lock(rq_src, rq_dest);
6983 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6984 rq_src->nr_uninterruptible = 0;
6985 double_rq_unlock(rq_src, rq_dest);
6986 local_irq_restore(flags);
6989 /* Run through task list and migrate tasks from the dead cpu. */
6990 static void migrate_live_tasks(int src_cpu)
6992 struct task_struct *p, *t;
6994 read_lock(&tasklist_lock);
6996 do_each_thread(t, p) {
7000 if (task_cpu(p) == src_cpu)
7001 move_task_off_dead_cpu(src_cpu, p);
7002 } while_each_thread(t, p);
7004 read_unlock(&tasklist_lock);
7008 * Schedules idle task to be the next runnable task on current CPU.
7009 * It does so by boosting its priority to highest possible.
7010 * Used by CPU offline code.
7012 void sched_idle_next(void)
7014 int this_cpu = smp_processor_id();
7015 struct rq *rq = cpu_rq(this_cpu);
7016 struct task_struct *p = rq->idle;
7017 unsigned long flags;
7019 /* cpu has to be offline */
7020 BUG_ON(cpu_online(this_cpu));
7023 * Strictly not necessary since rest of the CPUs are stopped by now
7024 * and interrupts disabled on the current cpu.
7026 spin_lock_irqsave(&rq->lock, flags);
7028 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7030 update_rq_clock(rq);
7031 activate_task(rq, p, 0);
7033 spin_unlock_irqrestore(&rq->lock, flags);
7037 * Ensures that the idle task is using init_mm right before its cpu goes
7040 void idle_task_exit(void)
7042 struct mm_struct *mm = current->active_mm;
7044 BUG_ON(cpu_online(smp_processor_id()));
7047 switch_mm(mm, &init_mm, current);
7051 /* called under rq->lock with disabled interrupts */
7052 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
7054 struct rq *rq = cpu_rq(dead_cpu);
7056 /* Must be exiting, otherwise would be on tasklist. */
7057 BUG_ON(!p->exit_state);
7059 /* Cannot have done final schedule yet: would have vanished. */
7060 BUG_ON(p->state == TASK_DEAD);
7065 * Drop lock around migration; if someone else moves it,
7066 * that's OK. No task can be added to this CPU, so iteration is
7069 spin_unlock_irq(&rq->lock);
7070 move_task_off_dead_cpu(dead_cpu, p);
7071 spin_lock_irq(&rq->lock);
7076 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7077 static void migrate_dead_tasks(unsigned int dead_cpu)
7079 struct rq *rq = cpu_rq(dead_cpu);
7080 struct task_struct *next;
7083 if (!rq->nr_running)
7085 update_rq_clock(rq);
7086 next = pick_next_task(rq);
7089 next->sched_class->put_prev_task(rq, next);
7090 migrate_dead(dead_cpu, next);
7094 #endif /* CONFIG_HOTPLUG_CPU */
7096 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7098 static struct ctl_table sd_ctl_dir[] = {
7100 .procname = "sched_domain",
7106 static struct ctl_table sd_ctl_root[] = {
7108 .ctl_name = CTL_KERN,
7109 .procname = "kernel",
7111 .child = sd_ctl_dir,
7116 static struct ctl_table *sd_alloc_ctl_entry(int n)
7118 struct ctl_table *entry =
7119 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
7124 static void sd_free_ctl_entry(struct ctl_table **tablep)
7126 struct ctl_table *entry;
7129 * In the intermediate directories, both the child directory and
7130 * procname are dynamically allocated and could fail but the mode
7131 * will always be set. In the lowest directory the names are
7132 * static strings and all have proc handlers.
7134 for (entry = *tablep; entry->mode; entry++) {
7136 sd_free_ctl_entry(&entry->child);
7137 if (entry->proc_handler == NULL)
7138 kfree(entry->procname);
7146 set_table_entry(struct ctl_table *entry,
7147 const char *procname, void *data, int maxlen,
7148 mode_t mode, proc_handler *proc_handler)
7150 entry->procname = procname;
7152 entry->maxlen = maxlen;
7154 entry->proc_handler = proc_handler;
7157 static struct ctl_table *
7158 sd_alloc_ctl_domain_table(struct sched_domain *sd)
7160 struct ctl_table *table = sd_alloc_ctl_entry(13);
7165 set_table_entry(&table[0], "min_interval", &sd->min_interval,
7166 sizeof(long), 0644, proc_doulongvec_minmax);
7167 set_table_entry(&table[1], "max_interval", &sd->max_interval,
7168 sizeof(long), 0644, proc_doulongvec_minmax);
7169 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
7170 sizeof(int), 0644, proc_dointvec_minmax);
7171 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
7172 sizeof(int), 0644, proc_dointvec_minmax);
7173 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
7174 sizeof(int), 0644, proc_dointvec_minmax);
7175 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
7176 sizeof(int), 0644, proc_dointvec_minmax);
7177 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
7178 sizeof(int), 0644, proc_dointvec_minmax);
7179 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
7180 sizeof(int), 0644, proc_dointvec_minmax);
7181 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
7182 sizeof(int), 0644, proc_dointvec_minmax);
7183 set_table_entry(&table[9], "cache_nice_tries",
7184 &sd->cache_nice_tries,
7185 sizeof(int), 0644, proc_dointvec_minmax);
7186 set_table_entry(&table[10], "flags", &sd->flags,
7187 sizeof(int), 0644, proc_dointvec_minmax);
7188 set_table_entry(&table[11], "name", sd->name,
7189 CORENAME_MAX_SIZE, 0444, proc_dostring);
7190 /* &table[12] is terminator */
7195 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
7197 struct ctl_table *entry, *table;
7198 struct sched_domain *sd;
7199 int domain_num = 0, i;
7202 for_each_domain(cpu, sd)
7204 entry = table = sd_alloc_ctl_entry(domain_num + 1);
7209 for_each_domain(cpu, sd) {
7210 snprintf(buf, 32, "domain%d", i);
7211 entry->procname = kstrdup(buf, GFP_KERNEL);
7213 entry->child = sd_alloc_ctl_domain_table(sd);
7220 static struct ctl_table_header *sd_sysctl_header;
7221 static void register_sched_domain_sysctl(void)
7223 int i, cpu_num = num_online_cpus();
7224 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
7227 WARN_ON(sd_ctl_dir[0].child);
7228 sd_ctl_dir[0].child = entry;
7233 for_each_online_cpu(i) {
7234 snprintf(buf, 32, "cpu%d", i);
7235 entry->procname = kstrdup(buf, GFP_KERNEL);
7237 entry->child = sd_alloc_ctl_cpu_table(i);
7241 WARN_ON(sd_sysctl_header);
7242 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
7245 /* may be called multiple times per register */
7246 static void unregister_sched_domain_sysctl(void)
7248 if (sd_sysctl_header)
7249 unregister_sysctl_table(sd_sysctl_header);
7250 sd_sysctl_header = NULL;
7251 if (sd_ctl_dir[0].child)
7252 sd_free_ctl_entry(&sd_ctl_dir[0].child);
7255 static void register_sched_domain_sysctl(void)
7258 static void unregister_sched_domain_sysctl(void)
7263 static void set_rq_online(struct rq *rq)
7266 const struct sched_class *class;
7268 cpumask_set_cpu(rq->cpu, rq->rd->online);
7271 for_each_class(class) {
7272 if (class->rq_online)
7273 class->rq_online(rq);
7278 static void set_rq_offline(struct rq *rq)
7281 const struct sched_class *class;
7283 for_each_class(class) {
7284 if (class->rq_offline)
7285 class->rq_offline(rq);
7288 cpumask_clear_cpu(rq->cpu, rq->rd->online);
7294 * migration_call - callback that gets triggered when a CPU is added.
7295 * Here we can start up the necessary migration thread for the new CPU.
7297 static int __cpuinit
7298 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
7300 struct task_struct *p;
7301 int cpu = (long)hcpu;
7302 unsigned long flags;
7307 case CPU_UP_PREPARE:
7308 case CPU_UP_PREPARE_FROZEN:
7309 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
7312 kthread_bind(p, cpu);
7313 /* Must be high prio: stop_machine expects to yield to it. */
7314 rq = task_rq_lock(p, &flags);
7315 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
7316 task_rq_unlock(rq, &flags);
7317 cpu_rq(cpu)->migration_thread = p;
7321 case CPU_ONLINE_FROZEN:
7322 /* Strictly unnecessary, as first user will wake it. */
7323 wake_up_process(cpu_rq(cpu)->migration_thread);
7325 /* Update our root-domain */
7327 spin_lock_irqsave(&rq->lock, flags);
7329 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7333 spin_unlock_irqrestore(&rq->lock, flags);
7336 #ifdef CONFIG_HOTPLUG_CPU
7337 case CPU_UP_CANCELED:
7338 case CPU_UP_CANCELED_FROZEN:
7339 if (!cpu_rq(cpu)->migration_thread)
7341 /* Unbind it from offline cpu so it can run. Fall thru. */
7342 kthread_bind(cpu_rq(cpu)->migration_thread,
7343 cpumask_any(cpu_online_mask));
7344 kthread_stop(cpu_rq(cpu)->migration_thread);
7345 cpu_rq(cpu)->migration_thread = NULL;
7349 case CPU_DEAD_FROZEN:
7350 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7351 migrate_live_tasks(cpu);
7353 kthread_stop(rq->migration_thread);
7354 rq->migration_thread = NULL;
7355 /* Idle task back to normal (off runqueue, low prio) */
7356 spin_lock_irq(&rq->lock);
7357 update_rq_clock(rq);
7358 deactivate_task(rq, rq->idle, 0);
7359 rq->idle->static_prio = MAX_PRIO;
7360 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
7361 rq->idle->sched_class = &idle_sched_class;
7362 migrate_dead_tasks(cpu);
7363 spin_unlock_irq(&rq->lock);
7365 migrate_nr_uninterruptible(rq);
7366 BUG_ON(rq->nr_running != 0);
7369 * No need to migrate the tasks: it was best-effort if
7370 * they didn't take sched_hotcpu_mutex. Just wake up
7373 spin_lock_irq(&rq->lock);
7374 while (!list_empty(&rq->migration_queue)) {
7375 struct migration_req *req;
7377 req = list_entry(rq->migration_queue.next,
7378 struct migration_req, list);
7379 list_del_init(&req->list);
7380 spin_unlock_irq(&rq->lock);
7381 complete(&req->done);
7382 spin_lock_irq(&rq->lock);
7384 spin_unlock_irq(&rq->lock);
7388 case CPU_DYING_FROZEN:
7389 /* Update our root-domain */
7391 spin_lock_irqsave(&rq->lock, flags);
7393 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
7396 spin_unlock_irqrestore(&rq->lock, flags);
7403 /* Register at highest priority so that task migration (migrate_all_tasks)
7404 * happens before everything else.
7406 static struct notifier_block __cpuinitdata migration_notifier = {
7407 .notifier_call = migration_call,
7411 static int __init migration_init(void)
7413 void *cpu = (void *)(long)smp_processor_id();
7416 /* Start one for the boot CPU: */
7417 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
7418 BUG_ON(err == NOTIFY_BAD);
7419 migration_call(&migration_notifier, CPU_ONLINE, cpu);
7420 register_cpu_notifier(&migration_notifier);
7424 early_initcall(migration_init);
7429 #ifdef CONFIG_SCHED_DEBUG
7431 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
7432 struct cpumask *groupmask)
7434 struct sched_group *group = sd->groups;
7437 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
7438 cpumask_clear(groupmask);
7440 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
7442 if (!(sd->flags & SD_LOAD_BALANCE)) {
7443 printk("does not load-balance\n");
7445 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
7450 printk(KERN_CONT "span %s level %s\n", str, sd->name);
7452 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
7453 printk(KERN_ERR "ERROR: domain->span does not contain "
7456 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
7457 printk(KERN_ERR "ERROR: domain->groups does not contain"
7461 printk(KERN_DEBUG "%*s groups:", level + 1, "");
7465 printk(KERN_ERR "ERROR: group is NULL\n");
7469 if (!group->__cpu_power) {
7470 printk(KERN_CONT "\n");
7471 printk(KERN_ERR "ERROR: domain->cpu_power not "
7476 if (!cpumask_weight(sched_group_cpus(group))) {
7477 printk(KERN_CONT "\n");
7478 printk(KERN_ERR "ERROR: empty group\n");
7482 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
7483 printk(KERN_CONT "\n");
7484 printk(KERN_ERR "ERROR: repeated CPUs\n");
7488 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
7490 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
7491 printk(KERN_CONT " %s (__cpu_power = %d)", str,
7492 group->__cpu_power);
7494 group = group->next;
7495 } while (group != sd->groups);
7496 printk(KERN_CONT "\n");
7498 if (!cpumask_equal(sched_domain_span(sd), groupmask))
7499 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
7502 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
7503 printk(KERN_ERR "ERROR: parent span is not a superset "
7504 "of domain->span\n");
7508 static void sched_domain_debug(struct sched_domain *sd, int cpu)
7510 cpumask_var_t groupmask;
7514 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
7518 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
7520 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
7521 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7526 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7533 free_cpumask_var(groupmask);
7535 #else /* !CONFIG_SCHED_DEBUG */
7536 # define sched_domain_debug(sd, cpu) do { } while (0)
7537 #endif /* CONFIG_SCHED_DEBUG */
7539 static int sd_degenerate(struct sched_domain *sd)
7541 if (cpumask_weight(sched_domain_span(sd)) == 1)
7544 /* Following flags need at least 2 groups */
7545 if (sd->flags & (SD_LOAD_BALANCE |
7546 SD_BALANCE_NEWIDLE |
7550 SD_SHARE_PKG_RESOURCES)) {
7551 if (sd->groups != sd->groups->next)
7555 /* Following flags don't use groups */
7556 if (sd->flags & (SD_WAKE_IDLE |
7565 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7567 unsigned long cflags = sd->flags, pflags = parent->flags;
7569 if (sd_degenerate(parent))
7572 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7575 /* Does parent contain flags not in child? */
7576 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7577 if (cflags & SD_WAKE_AFFINE)
7578 pflags &= ~SD_WAKE_BALANCE;
7579 /* Flags needing groups don't count if only 1 group in parent */
7580 if (parent->groups == parent->groups->next) {
7581 pflags &= ~(SD_LOAD_BALANCE |
7582 SD_BALANCE_NEWIDLE |
7586 SD_SHARE_PKG_RESOURCES);
7587 if (nr_node_ids == 1)
7588 pflags &= ~SD_SERIALIZE;
7590 if (~cflags & pflags)
7596 static void free_rootdomain(struct root_domain *rd)
7598 cpupri_cleanup(&rd->cpupri);
7600 free_cpumask_var(rd->rto_mask);
7601 free_cpumask_var(rd->online);
7602 free_cpumask_var(rd->span);
7606 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7608 struct root_domain *old_rd = NULL;
7609 unsigned long flags;
7611 spin_lock_irqsave(&rq->lock, flags);
7616 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7619 cpumask_clear_cpu(rq->cpu, old_rd->span);
7622 * If we dont want to free the old_rt yet then
7623 * set old_rd to NULL to skip the freeing later
7626 if (!atomic_dec_and_test(&old_rd->refcount))
7630 atomic_inc(&rd->refcount);
7633 cpumask_set_cpu(rq->cpu, rd->span);
7634 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7637 spin_unlock_irqrestore(&rq->lock, flags);
7640 free_rootdomain(old_rd);
7643 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7645 memset(rd, 0, sizeof(*rd));
7648 alloc_bootmem_cpumask_var(&def_root_domain.span);
7649 alloc_bootmem_cpumask_var(&def_root_domain.online);
7650 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7651 cpupri_init(&rd->cpupri, true);
7655 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7657 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7659 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7662 if (cpupri_init(&rd->cpupri, false) != 0)
7667 free_cpumask_var(rd->rto_mask);
7669 free_cpumask_var(rd->online);
7671 free_cpumask_var(rd->span);
7676 static void init_defrootdomain(void)
7678 init_rootdomain(&def_root_domain, true);
7680 atomic_set(&def_root_domain.refcount, 1);
7683 static struct root_domain *alloc_rootdomain(void)
7685 struct root_domain *rd;
7687 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7691 if (init_rootdomain(rd, false) != 0) {
7700 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7701 * hold the hotplug lock.
7704 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7706 struct rq *rq = cpu_rq(cpu);
7707 struct sched_domain *tmp;
7709 /* Remove the sched domains which do not contribute to scheduling. */
7710 for (tmp = sd; tmp; ) {
7711 struct sched_domain *parent = tmp->parent;
7715 if (sd_parent_degenerate(tmp, parent)) {
7716 tmp->parent = parent->parent;
7718 parent->parent->child = tmp;
7723 if (sd && sd_degenerate(sd)) {
7729 sched_domain_debug(sd, cpu);
7731 rq_attach_root(rq, rd);
7732 rcu_assign_pointer(rq->sd, sd);
7735 /* cpus with isolated domains */
7736 static cpumask_var_t cpu_isolated_map;
7738 /* Setup the mask of cpus configured for isolated domains */
7739 static int __init isolated_cpu_setup(char *str)
7741 cpulist_parse(str, cpu_isolated_map);
7745 __setup("isolcpus=", isolated_cpu_setup);
7748 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7749 * to a function which identifies what group(along with sched group) a CPU
7750 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7751 * (due to the fact that we keep track of groups covered with a struct cpumask).
7753 * init_sched_build_groups will build a circular linked list of the groups
7754 * covered by the given span, and will set each group's ->cpumask correctly,
7755 * and ->cpu_power to 0.
7758 init_sched_build_groups(const struct cpumask *span,
7759 const struct cpumask *cpu_map,
7760 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7761 struct sched_group **sg,
7762 struct cpumask *tmpmask),
7763 struct cpumask *covered, struct cpumask *tmpmask)
7765 struct sched_group *first = NULL, *last = NULL;
7768 cpumask_clear(covered);
7770 for_each_cpu(i, span) {
7771 struct sched_group *sg;
7772 int group = group_fn(i, cpu_map, &sg, tmpmask);
7775 if (cpumask_test_cpu(i, covered))
7778 cpumask_clear(sched_group_cpus(sg));
7779 sg->__cpu_power = 0;
7781 for_each_cpu(j, span) {
7782 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7785 cpumask_set_cpu(j, covered);
7786 cpumask_set_cpu(j, sched_group_cpus(sg));
7797 #define SD_NODES_PER_DOMAIN 16
7802 * find_next_best_node - find the next node to include in a sched_domain
7803 * @node: node whose sched_domain we're building
7804 * @used_nodes: nodes already in the sched_domain
7806 * Find the next node to include in a given scheduling domain. Simply
7807 * finds the closest node not already in the @used_nodes map.
7809 * Should use nodemask_t.
7811 static int find_next_best_node(int node, nodemask_t *used_nodes)
7813 int i, n, val, min_val, best_node = 0;
7817 for (i = 0; i < nr_node_ids; i++) {
7818 /* Start at @node */
7819 n = (node + i) % nr_node_ids;
7821 if (!nr_cpus_node(n))
7824 /* Skip already used nodes */
7825 if (node_isset(n, *used_nodes))
7828 /* Simple min distance search */
7829 val = node_distance(node, n);
7831 if (val < min_val) {
7837 node_set(best_node, *used_nodes);
7842 * sched_domain_node_span - get a cpumask for a node's sched_domain
7843 * @node: node whose cpumask we're constructing
7844 * @span: resulting cpumask
7846 * Given a node, construct a good cpumask for its sched_domain to span. It
7847 * should be one that prevents unnecessary balancing, but also spreads tasks
7850 static void sched_domain_node_span(int node, struct cpumask *span)
7852 nodemask_t used_nodes;
7855 cpumask_clear(span);
7856 nodes_clear(used_nodes);
7858 cpumask_or(span, span, cpumask_of_node(node));
7859 node_set(node, used_nodes);
7861 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7862 int next_node = find_next_best_node(node, &used_nodes);
7864 cpumask_or(span, span, cpumask_of_node(next_node));
7867 #endif /* CONFIG_NUMA */
7869 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7872 * The cpus mask in sched_group and sched_domain hangs off the end.
7873 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7874 * for nr_cpu_ids < CONFIG_NR_CPUS.
7876 struct static_sched_group {
7877 struct sched_group sg;
7878 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7881 struct static_sched_domain {
7882 struct sched_domain sd;
7883 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7887 * SMT sched-domains:
7889 #ifdef CONFIG_SCHED_SMT
7890 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7891 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7894 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7895 struct sched_group **sg, struct cpumask *unused)
7898 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7901 #endif /* CONFIG_SCHED_SMT */
7904 * multi-core sched-domains:
7906 #ifdef CONFIG_SCHED_MC
7907 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7908 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7909 #endif /* CONFIG_SCHED_MC */
7911 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7913 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7914 struct sched_group **sg, struct cpumask *mask)
7918 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7919 group = cpumask_first(mask);
7921 *sg = &per_cpu(sched_group_core, group).sg;
7924 #elif defined(CONFIG_SCHED_MC)
7926 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7927 struct sched_group **sg, struct cpumask *unused)
7930 *sg = &per_cpu(sched_group_core, cpu).sg;
7935 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7936 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7939 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7940 struct sched_group **sg, struct cpumask *mask)
7943 #ifdef CONFIG_SCHED_MC
7944 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7945 group = cpumask_first(mask);
7946 #elif defined(CONFIG_SCHED_SMT)
7947 cpumask_and(mask, topology_thread_cpumask(cpu), cpu_map);
7948 group = cpumask_first(mask);
7953 *sg = &per_cpu(sched_group_phys, group).sg;
7959 * The init_sched_build_groups can't handle what we want to do with node
7960 * groups, so roll our own. Now each node has its own list of groups which
7961 * gets dynamically allocated.
7963 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7964 static struct sched_group ***sched_group_nodes_bycpu;
7966 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7967 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7969 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7970 struct sched_group **sg,
7971 struct cpumask *nodemask)
7975 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7976 group = cpumask_first(nodemask);
7979 *sg = &per_cpu(sched_group_allnodes, group).sg;
7983 static void init_numa_sched_groups_power(struct sched_group *group_head)
7985 struct sched_group *sg = group_head;
7991 for_each_cpu(j, sched_group_cpus(sg)) {
7992 struct sched_domain *sd;
7994 sd = &per_cpu(phys_domains, j).sd;
7995 if (j != group_first_cpu(sd->groups)) {
7997 * Only add "power" once for each
8003 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
8006 } while (sg != group_head);
8008 #endif /* CONFIG_NUMA */
8011 /* Free memory allocated for various sched_group structures */
8012 static void free_sched_groups(const struct cpumask *cpu_map,
8013 struct cpumask *nodemask)
8017 for_each_cpu(cpu, cpu_map) {
8018 struct sched_group **sched_group_nodes
8019 = sched_group_nodes_bycpu[cpu];
8021 if (!sched_group_nodes)
8024 for (i = 0; i < nr_node_ids; i++) {
8025 struct sched_group *oldsg, *sg = sched_group_nodes[i];
8027 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8028 if (cpumask_empty(nodemask))
8038 if (oldsg != sched_group_nodes[i])
8041 kfree(sched_group_nodes);
8042 sched_group_nodes_bycpu[cpu] = NULL;
8045 #else /* !CONFIG_NUMA */
8046 static void free_sched_groups(const struct cpumask *cpu_map,
8047 struct cpumask *nodemask)
8050 #endif /* CONFIG_NUMA */
8053 * Initialize sched groups cpu_power.
8055 * cpu_power indicates the capacity of sched group, which is used while
8056 * distributing the load between different sched groups in a sched domain.
8057 * Typically cpu_power for all the groups in a sched domain will be same unless
8058 * there are asymmetries in the topology. If there are asymmetries, group
8059 * having more cpu_power will pickup more load compared to the group having
8062 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8063 * the maximum number of tasks a group can handle in the presence of other idle
8064 * or lightly loaded groups in the same sched domain.
8066 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
8068 struct sched_domain *child;
8069 struct sched_group *group;
8071 WARN_ON(!sd || !sd->groups);
8073 if (cpu != group_first_cpu(sd->groups))
8078 sd->groups->__cpu_power = 0;
8081 * For perf policy, if the groups in child domain share resources
8082 * (for example cores sharing some portions of the cache hierarchy
8083 * or SMT), then set this domain groups cpu_power such that each group
8084 * can handle only one task, when there are other idle groups in the
8085 * same sched domain.
8087 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
8089 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
8090 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
8095 * add cpu_power of each child group to this groups cpu_power
8097 group = child->groups;
8099 sg_inc_cpu_power(sd->groups, group->__cpu_power);
8100 group = group->next;
8101 } while (group != child->groups);
8105 * Initializers for schedule domains
8106 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8109 #ifdef CONFIG_SCHED_DEBUG
8110 # define SD_INIT_NAME(sd, type) sd->name = #type
8112 # define SD_INIT_NAME(sd, type) do { } while (0)
8115 #define SD_INIT(sd, type) sd_init_##type(sd)
8117 #define SD_INIT_FUNC(type) \
8118 static noinline void sd_init_##type(struct sched_domain *sd) \
8120 memset(sd, 0, sizeof(*sd)); \
8121 *sd = SD_##type##_INIT; \
8122 sd->level = SD_LV_##type; \
8123 SD_INIT_NAME(sd, type); \
8128 SD_INIT_FUNC(ALLNODES)
8131 #ifdef CONFIG_SCHED_SMT
8132 SD_INIT_FUNC(SIBLING)
8134 #ifdef CONFIG_SCHED_MC
8138 static int default_relax_domain_level = -1;
8140 static int __init setup_relax_domain_level(char *str)
8144 val = simple_strtoul(str, NULL, 0);
8145 if (val < SD_LV_MAX)
8146 default_relax_domain_level = val;
8150 __setup("relax_domain_level=", setup_relax_domain_level);
8152 static void set_domain_attribute(struct sched_domain *sd,
8153 struct sched_domain_attr *attr)
8157 if (!attr || attr->relax_domain_level < 0) {
8158 if (default_relax_domain_level < 0)
8161 request = default_relax_domain_level;
8163 request = attr->relax_domain_level;
8164 if (request < sd->level) {
8165 /* turn off idle balance on this domain */
8166 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
8168 /* turn on idle balance on this domain */
8169 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
8174 * Build sched domains for a given set of cpus and attach the sched domains
8175 * to the individual cpus
8177 static int __build_sched_domains(const struct cpumask *cpu_map,
8178 struct sched_domain_attr *attr)
8180 int i, err = -ENOMEM;
8181 struct root_domain *rd;
8182 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
8185 cpumask_var_t domainspan, covered, notcovered;
8186 struct sched_group **sched_group_nodes = NULL;
8187 int sd_allnodes = 0;
8189 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
8191 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
8192 goto free_domainspan;
8193 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
8197 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
8198 goto free_notcovered;
8199 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
8201 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
8202 goto free_this_sibling_map;
8203 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
8204 goto free_this_core_map;
8205 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
8206 goto free_send_covered;
8210 * Allocate the per-node list of sched groups
8212 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
8214 if (!sched_group_nodes) {
8215 printk(KERN_WARNING "Can not alloc sched group node list\n");
8220 rd = alloc_rootdomain();
8222 printk(KERN_WARNING "Cannot alloc root domain\n");
8223 goto free_sched_groups;
8227 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
8231 * Set up domains for cpus specified by the cpu_map.
8233 for_each_cpu(i, cpu_map) {
8234 struct sched_domain *sd = NULL, *p;
8236 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
8239 if (cpumask_weight(cpu_map) >
8240 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
8241 sd = &per_cpu(allnodes_domains, i).sd;
8242 SD_INIT(sd, ALLNODES);
8243 set_domain_attribute(sd, attr);
8244 cpumask_copy(sched_domain_span(sd), cpu_map);
8245 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
8251 sd = &per_cpu(node_domains, i).sd;
8253 set_domain_attribute(sd, attr);
8254 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
8258 cpumask_and(sched_domain_span(sd),
8259 sched_domain_span(sd), cpu_map);
8263 sd = &per_cpu(phys_domains, i).sd;
8265 set_domain_attribute(sd, attr);
8266 cpumask_copy(sched_domain_span(sd), nodemask);
8270 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
8272 #ifdef CONFIG_SCHED_MC
8274 sd = &per_cpu(core_domains, i).sd;
8276 set_domain_attribute(sd, attr);
8277 cpumask_and(sched_domain_span(sd), cpu_map,
8278 cpu_coregroup_mask(i));
8281 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
8284 #ifdef CONFIG_SCHED_SMT
8286 sd = &per_cpu(cpu_domains, i).sd;
8287 SD_INIT(sd, SIBLING);
8288 set_domain_attribute(sd, attr);
8289 cpumask_and(sched_domain_span(sd),
8290 topology_thread_cpumask(i), cpu_map);
8293 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
8297 #ifdef CONFIG_SCHED_SMT
8298 /* Set up CPU (sibling) groups */
8299 for_each_cpu(i, cpu_map) {
8300 cpumask_and(this_sibling_map,
8301 topology_thread_cpumask(i), cpu_map);
8302 if (i != cpumask_first(this_sibling_map))
8305 init_sched_build_groups(this_sibling_map, cpu_map,
8307 send_covered, tmpmask);
8311 #ifdef CONFIG_SCHED_MC
8312 /* Set up multi-core groups */
8313 for_each_cpu(i, cpu_map) {
8314 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
8315 if (i != cpumask_first(this_core_map))
8318 init_sched_build_groups(this_core_map, cpu_map,
8320 send_covered, tmpmask);
8324 /* Set up physical groups */
8325 for (i = 0; i < nr_node_ids; i++) {
8326 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8327 if (cpumask_empty(nodemask))
8330 init_sched_build_groups(nodemask, cpu_map,
8332 send_covered, tmpmask);
8336 /* Set up node groups */
8338 init_sched_build_groups(cpu_map, cpu_map,
8339 &cpu_to_allnodes_group,
8340 send_covered, tmpmask);
8343 for (i = 0; i < nr_node_ids; i++) {
8344 /* Set up node groups */
8345 struct sched_group *sg, *prev;
8348 cpumask_clear(covered);
8349 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
8350 if (cpumask_empty(nodemask)) {
8351 sched_group_nodes[i] = NULL;
8355 sched_domain_node_span(i, domainspan);
8356 cpumask_and(domainspan, domainspan, cpu_map);
8358 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
8361 printk(KERN_WARNING "Can not alloc domain group for "
8365 sched_group_nodes[i] = sg;
8366 for_each_cpu(j, nodemask) {
8367 struct sched_domain *sd;
8369 sd = &per_cpu(node_domains, j).sd;
8372 sg->__cpu_power = 0;
8373 cpumask_copy(sched_group_cpus(sg), nodemask);
8375 cpumask_or(covered, covered, nodemask);
8378 for (j = 0; j < nr_node_ids; j++) {
8379 int n = (i + j) % nr_node_ids;
8381 cpumask_complement(notcovered, covered);
8382 cpumask_and(tmpmask, notcovered, cpu_map);
8383 cpumask_and(tmpmask, tmpmask, domainspan);
8384 if (cpumask_empty(tmpmask))
8387 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
8388 if (cpumask_empty(tmpmask))
8391 sg = kmalloc_node(sizeof(struct sched_group) +
8396 "Can not alloc domain group for node %d\n", j);
8399 sg->__cpu_power = 0;
8400 cpumask_copy(sched_group_cpus(sg), tmpmask);
8401 sg->next = prev->next;
8402 cpumask_or(covered, covered, tmpmask);
8409 /* Calculate CPU power for physical packages and nodes */
8410 #ifdef CONFIG_SCHED_SMT
8411 for_each_cpu(i, cpu_map) {
8412 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
8414 init_sched_groups_power(i, sd);
8417 #ifdef CONFIG_SCHED_MC
8418 for_each_cpu(i, cpu_map) {
8419 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
8421 init_sched_groups_power(i, sd);
8425 for_each_cpu(i, cpu_map) {
8426 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
8428 init_sched_groups_power(i, sd);
8432 for (i = 0; i < nr_node_ids; i++)
8433 init_numa_sched_groups_power(sched_group_nodes[i]);
8436 struct sched_group *sg;
8438 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
8440 init_numa_sched_groups_power(sg);
8444 /* Attach the domains */
8445 for_each_cpu(i, cpu_map) {
8446 struct sched_domain *sd;
8447 #ifdef CONFIG_SCHED_SMT
8448 sd = &per_cpu(cpu_domains, i).sd;
8449 #elif defined(CONFIG_SCHED_MC)
8450 sd = &per_cpu(core_domains, i).sd;
8452 sd = &per_cpu(phys_domains, i).sd;
8454 cpu_attach_domain(sd, rd, i);
8460 free_cpumask_var(tmpmask);
8462 free_cpumask_var(send_covered);
8464 free_cpumask_var(this_core_map);
8465 free_this_sibling_map:
8466 free_cpumask_var(this_sibling_map);
8468 free_cpumask_var(nodemask);
8471 free_cpumask_var(notcovered);
8473 free_cpumask_var(covered);
8475 free_cpumask_var(domainspan);
8482 kfree(sched_group_nodes);
8488 free_sched_groups(cpu_map, tmpmask);
8489 free_rootdomain(rd);
8494 static int build_sched_domains(const struct cpumask *cpu_map)
8496 return __build_sched_domains(cpu_map, NULL);
8499 static struct cpumask *doms_cur; /* current sched domains */
8500 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
8501 static struct sched_domain_attr *dattr_cur;
8502 /* attribues of custom domains in 'doms_cur' */
8505 * Special case: If a kmalloc of a doms_cur partition (array of
8506 * cpumask) fails, then fallback to a single sched domain,
8507 * as determined by the single cpumask fallback_doms.
8509 static cpumask_var_t fallback_doms;
8512 * arch_update_cpu_topology lets virtualized architectures update the
8513 * cpu core maps. It is supposed to return 1 if the topology changed
8514 * or 0 if it stayed the same.
8516 int __attribute__((weak)) arch_update_cpu_topology(void)
8522 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8523 * For now this just excludes isolated cpus, but could be used to
8524 * exclude other special cases in the future.
8526 static int arch_init_sched_domains(const struct cpumask *cpu_map)
8530 arch_update_cpu_topology();
8532 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8534 doms_cur = fallback_doms;
8535 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8537 err = build_sched_domains(doms_cur);
8538 register_sched_domain_sysctl();
8543 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8544 struct cpumask *tmpmask)
8546 free_sched_groups(cpu_map, tmpmask);
8550 * Detach sched domains from a group of cpus specified in cpu_map
8551 * These cpus will now be attached to the NULL domain
8553 static void detach_destroy_domains(const struct cpumask *cpu_map)
8555 /* Save because hotplug lock held. */
8556 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8559 for_each_cpu(i, cpu_map)
8560 cpu_attach_domain(NULL, &def_root_domain, i);
8561 synchronize_sched();
8562 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8565 /* handle null as "default" */
8566 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8567 struct sched_domain_attr *new, int idx_new)
8569 struct sched_domain_attr tmp;
8576 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8577 new ? (new + idx_new) : &tmp,
8578 sizeof(struct sched_domain_attr));
8582 * Partition sched domains as specified by the 'ndoms_new'
8583 * cpumasks in the array doms_new[] of cpumasks. This compares
8584 * doms_new[] to the current sched domain partitioning, doms_cur[].
8585 * It destroys each deleted domain and builds each new domain.
8587 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8588 * The masks don't intersect (don't overlap.) We should setup one
8589 * sched domain for each mask. CPUs not in any of the cpumasks will
8590 * not be load balanced. If the same cpumask appears both in the
8591 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8594 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8595 * ownership of it and will kfree it when done with it. If the caller
8596 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8597 * ndoms_new == 1, and partition_sched_domains() will fallback to
8598 * the single partition 'fallback_doms', it also forces the domains
8601 * If doms_new == NULL it will be replaced with cpu_online_mask.
8602 * ndoms_new == 0 is a special case for destroying existing domains,
8603 * and it will not create the default domain.
8605 * Call with hotplug lock held
8607 /* FIXME: Change to struct cpumask *doms_new[] */
8608 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8609 struct sched_domain_attr *dattr_new)
8614 mutex_lock(&sched_domains_mutex);
8616 /* always unregister in case we don't destroy any domains */
8617 unregister_sched_domain_sysctl();
8619 /* Let architecture update cpu core mappings. */
8620 new_topology = arch_update_cpu_topology();
8622 n = doms_new ? ndoms_new : 0;
8624 /* Destroy deleted domains */
8625 for (i = 0; i < ndoms_cur; i++) {
8626 for (j = 0; j < n && !new_topology; j++) {
8627 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8628 && dattrs_equal(dattr_cur, i, dattr_new, j))
8631 /* no match - a current sched domain not in new doms_new[] */
8632 detach_destroy_domains(doms_cur + i);
8637 if (doms_new == NULL) {
8639 doms_new = fallback_doms;
8640 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8641 WARN_ON_ONCE(dattr_new);
8644 /* Build new domains */
8645 for (i = 0; i < ndoms_new; i++) {
8646 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8647 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8648 && dattrs_equal(dattr_new, i, dattr_cur, j))
8651 /* no match - add a new doms_new */
8652 __build_sched_domains(doms_new + i,
8653 dattr_new ? dattr_new + i : NULL);
8658 /* Remember the new sched domains */
8659 if (doms_cur != fallback_doms)
8661 kfree(dattr_cur); /* kfree(NULL) is safe */
8662 doms_cur = doms_new;
8663 dattr_cur = dattr_new;
8664 ndoms_cur = ndoms_new;
8666 register_sched_domain_sysctl();
8668 mutex_unlock(&sched_domains_mutex);
8671 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8672 static void arch_reinit_sched_domains(void)
8676 /* Destroy domains first to force the rebuild */
8677 partition_sched_domains(0, NULL, NULL);
8679 rebuild_sched_domains();
8683 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8685 unsigned int level = 0;
8687 if (sscanf(buf, "%u", &level) != 1)
8691 * level is always be positive so don't check for
8692 * level < POWERSAVINGS_BALANCE_NONE which is 0
8693 * What happens on 0 or 1 byte write,
8694 * need to check for count as well?
8697 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8701 sched_smt_power_savings = level;
8703 sched_mc_power_savings = level;
8705 arch_reinit_sched_domains();
8710 #ifdef CONFIG_SCHED_MC
8711 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8714 return sprintf(page, "%u\n", sched_mc_power_savings);
8716 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8717 const char *buf, size_t count)
8719 return sched_power_savings_store(buf, count, 0);
8721 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8722 sched_mc_power_savings_show,
8723 sched_mc_power_savings_store);
8726 #ifdef CONFIG_SCHED_SMT
8727 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8730 return sprintf(page, "%u\n", sched_smt_power_savings);
8732 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8733 const char *buf, size_t count)
8735 return sched_power_savings_store(buf, count, 1);
8737 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8738 sched_smt_power_savings_show,
8739 sched_smt_power_savings_store);
8742 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8746 #ifdef CONFIG_SCHED_SMT
8748 err = sysfs_create_file(&cls->kset.kobj,
8749 &attr_sched_smt_power_savings.attr);
8751 #ifdef CONFIG_SCHED_MC
8752 if (!err && mc_capable())
8753 err = sysfs_create_file(&cls->kset.kobj,
8754 &attr_sched_mc_power_savings.attr);
8758 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8760 #ifndef CONFIG_CPUSETS
8762 * Add online and remove offline CPUs from the scheduler domains.
8763 * When cpusets are enabled they take over this function.
8765 static int update_sched_domains(struct notifier_block *nfb,
8766 unsigned long action, void *hcpu)
8770 case CPU_ONLINE_FROZEN:
8772 case CPU_DEAD_FROZEN:
8773 partition_sched_domains(1, NULL, NULL);
8782 static int update_runtime(struct notifier_block *nfb,
8783 unsigned long action, void *hcpu)
8785 int cpu = (int)(long)hcpu;
8788 case CPU_DOWN_PREPARE:
8789 case CPU_DOWN_PREPARE_FROZEN:
8790 disable_runtime(cpu_rq(cpu));
8793 case CPU_DOWN_FAILED:
8794 case CPU_DOWN_FAILED_FROZEN:
8796 case CPU_ONLINE_FROZEN:
8797 enable_runtime(cpu_rq(cpu));
8805 void __init sched_init_smp(void)
8807 cpumask_var_t non_isolated_cpus;
8809 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8811 #if defined(CONFIG_NUMA)
8812 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8814 BUG_ON(sched_group_nodes_bycpu == NULL);
8817 mutex_lock(&sched_domains_mutex);
8818 arch_init_sched_domains(cpu_online_mask);
8819 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8820 if (cpumask_empty(non_isolated_cpus))
8821 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8822 mutex_unlock(&sched_domains_mutex);
8825 #ifndef CONFIG_CPUSETS
8826 /* XXX: Theoretical race here - CPU may be hotplugged now */
8827 hotcpu_notifier(update_sched_domains, 0);
8830 /* RT runtime code needs to handle some hotplug events */
8831 hotcpu_notifier(update_runtime, 0);
8835 /* Move init over to a non-isolated CPU */
8836 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8838 sched_init_granularity();
8839 free_cpumask_var(non_isolated_cpus);
8841 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8842 init_sched_rt_class();
8845 void __init sched_init_smp(void)
8847 sched_init_granularity();
8849 #endif /* CONFIG_SMP */
8851 int in_sched_functions(unsigned long addr)
8853 return in_lock_functions(addr) ||
8854 (addr >= (unsigned long)__sched_text_start
8855 && addr < (unsigned long)__sched_text_end);
8858 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8860 cfs_rq->tasks_timeline = RB_ROOT;
8861 INIT_LIST_HEAD(&cfs_rq->tasks);
8862 #ifdef CONFIG_FAIR_GROUP_SCHED
8865 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8868 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8870 struct rt_prio_array *array;
8873 array = &rt_rq->active;
8874 for (i = 0; i < MAX_RT_PRIO; i++) {
8875 INIT_LIST_HEAD(array->queue + i);
8876 __clear_bit(i, array->bitmap);
8878 /* delimiter for bitsearch: */
8879 __set_bit(MAX_RT_PRIO, array->bitmap);
8881 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8882 rt_rq->highest_prio.curr = MAX_RT_PRIO;
8884 rt_rq->highest_prio.next = MAX_RT_PRIO;
8888 rt_rq->rt_nr_migratory = 0;
8889 rt_rq->overloaded = 0;
8890 plist_head_init(&rq->rt.pushable_tasks, &rq->lock);
8894 rt_rq->rt_throttled = 0;
8895 rt_rq->rt_runtime = 0;
8896 spin_lock_init(&rt_rq->rt_runtime_lock);
8898 #ifdef CONFIG_RT_GROUP_SCHED
8899 rt_rq->rt_nr_boosted = 0;
8904 #ifdef CONFIG_FAIR_GROUP_SCHED
8905 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8906 struct sched_entity *se, int cpu, int add,
8907 struct sched_entity *parent)
8909 struct rq *rq = cpu_rq(cpu);
8910 tg->cfs_rq[cpu] = cfs_rq;
8911 init_cfs_rq(cfs_rq, rq);
8914 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8917 /* se could be NULL for init_task_group */
8922 se->cfs_rq = &rq->cfs;
8924 se->cfs_rq = parent->my_q;
8927 se->load.weight = tg->shares;
8928 se->load.inv_weight = 0;
8929 se->parent = parent;
8933 #ifdef CONFIG_RT_GROUP_SCHED
8934 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8935 struct sched_rt_entity *rt_se, int cpu, int add,
8936 struct sched_rt_entity *parent)
8938 struct rq *rq = cpu_rq(cpu);
8940 tg->rt_rq[cpu] = rt_rq;
8941 init_rt_rq(rt_rq, rq);
8943 rt_rq->rt_se = rt_se;
8944 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8946 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8948 tg->rt_se[cpu] = rt_se;
8953 rt_se->rt_rq = &rq->rt;
8955 rt_se->rt_rq = parent->my_q;
8957 rt_se->my_q = rt_rq;
8958 rt_se->parent = parent;
8959 INIT_LIST_HEAD(&rt_se->run_list);
8963 void __init sched_init(void)
8966 unsigned long alloc_size = 0, ptr;
8968 #ifdef CONFIG_FAIR_GROUP_SCHED
8969 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8971 #ifdef CONFIG_RT_GROUP_SCHED
8972 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8974 #ifdef CONFIG_USER_SCHED
8977 #ifdef CONFIG_CPUMASK_OFFSTACK
8978 alloc_size += num_possible_cpus() * cpumask_size();
8981 * As sched_init() is called before page_alloc is setup,
8982 * we use alloc_bootmem().
8985 ptr = (unsigned long)alloc_bootmem(alloc_size);
8987 #ifdef CONFIG_FAIR_GROUP_SCHED
8988 init_task_group.se = (struct sched_entity **)ptr;
8989 ptr += nr_cpu_ids * sizeof(void **);
8991 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8992 ptr += nr_cpu_ids * sizeof(void **);
8994 #ifdef CONFIG_USER_SCHED
8995 root_task_group.se = (struct sched_entity **)ptr;
8996 ptr += nr_cpu_ids * sizeof(void **);
8998 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8999 ptr += nr_cpu_ids * sizeof(void **);
9000 #endif /* CONFIG_USER_SCHED */
9001 #endif /* CONFIG_FAIR_GROUP_SCHED */
9002 #ifdef CONFIG_RT_GROUP_SCHED
9003 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
9004 ptr += nr_cpu_ids * sizeof(void **);
9006 init_task_group.rt_rq = (struct rt_rq **)ptr;
9007 ptr += nr_cpu_ids * sizeof(void **);
9009 #ifdef CONFIG_USER_SCHED
9010 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
9011 ptr += nr_cpu_ids * sizeof(void **);
9013 root_task_group.rt_rq = (struct rt_rq **)ptr;
9014 ptr += nr_cpu_ids * sizeof(void **);
9015 #endif /* CONFIG_USER_SCHED */
9016 #endif /* CONFIG_RT_GROUP_SCHED */
9017 #ifdef CONFIG_CPUMASK_OFFSTACK
9018 for_each_possible_cpu(i) {
9019 per_cpu(load_balance_tmpmask, i) = (void *)ptr;
9020 ptr += cpumask_size();
9022 #endif /* CONFIG_CPUMASK_OFFSTACK */
9026 init_defrootdomain();
9029 init_rt_bandwidth(&def_rt_bandwidth,
9030 global_rt_period(), global_rt_runtime());
9032 #ifdef CONFIG_RT_GROUP_SCHED
9033 init_rt_bandwidth(&init_task_group.rt_bandwidth,
9034 global_rt_period(), global_rt_runtime());
9035 #ifdef CONFIG_USER_SCHED
9036 init_rt_bandwidth(&root_task_group.rt_bandwidth,
9037 global_rt_period(), RUNTIME_INF);
9038 #endif /* CONFIG_USER_SCHED */
9039 #endif /* CONFIG_RT_GROUP_SCHED */
9041 #ifdef CONFIG_GROUP_SCHED
9042 list_add(&init_task_group.list, &task_groups);
9043 INIT_LIST_HEAD(&init_task_group.children);
9045 #ifdef CONFIG_USER_SCHED
9046 INIT_LIST_HEAD(&root_task_group.children);
9047 init_task_group.parent = &root_task_group;
9048 list_add(&init_task_group.siblings, &root_task_group.children);
9049 #endif /* CONFIG_USER_SCHED */
9050 #endif /* CONFIG_GROUP_SCHED */
9052 for_each_possible_cpu(i) {
9056 spin_lock_init(&rq->lock);
9058 init_cfs_rq(&rq->cfs, rq);
9059 init_rt_rq(&rq->rt, rq);
9060 #ifdef CONFIG_FAIR_GROUP_SCHED
9061 init_task_group.shares = init_task_group_load;
9062 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
9063 #ifdef CONFIG_CGROUP_SCHED
9065 * How much cpu bandwidth does init_task_group get?
9067 * In case of task-groups formed thr' the cgroup filesystem, it
9068 * gets 100% of the cpu resources in the system. This overall
9069 * system cpu resource is divided among the tasks of
9070 * init_task_group and its child task-groups in a fair manner,
9071 * based on each entity's (task or task-group's) weight
9072 * (se->load.weight).
9074 * In other words, if init_task_group has 10 tasks of weight
9075 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9076 * then A0's share of the cpu resource is:
9078 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9080 * We achieve this by letting init_task_group's tasks sit
9081 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9083 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
9084 #elif defined CONFIG_USER_SCHED
9085 root_task_group.shares = NICE_0_LOAD;
9086 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
9088 * In case of task-groups formed thr' the user id of tasks,
9089 * init_task_group represents tasks belonging to root user.
9090 * Hence it forms a sibling of all subsequent groups formed.
9091 * In this case, init_task_group gets only a fraction of overall
9092 * system cpu resource, based on the weight assigned to root
9093 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9094 * by letting tasks of init_task_group sit in a separate cfs_rq
9095 * (init_cfs_rq) and having one entity represent this group of
9096 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9098 init_tg_cfs_entry(&init_task_group,
9099 &per_cpu(init_cfs_rq, i),
9100 &per_cpu(init_sched_entity, i), i, 1,
9101 root_task_group.se[i]);
9104 #endif /* CONFIG_FAIR_GROUP_SCHED */
9106 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
9107 #ifdef CONFIG_RT_GROUP_SCHED
9108 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
9109 #ifdef CONFIG_CGROUP_SCHED
9110 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
9111 #elif defined CONFIG_USER_SCHED
9112 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
9113 init_tg_rt_entry(&init_task_group,
9114 &per_cpu(init_rt_rq, i),
9115 &per_cpu(init_sched_rt_entity, i), i, 1,
9116 root_task_group.rt_se[i]);
9120 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
9121 rq->cpu_load[j] = 0;
9125 rq->active_balance = 0;
9126 rq->next_balance = jiffies;
9130 rq->migration_thread = NULL;
9131 INIT_LIST_HEAD(&rq->migration_queue);
9132 rq_attach_root(rq, &def_root_domain);
9135 atomic_set(&rq->nr_iowait, 0);
9138 set_load_weight(&init_task);
9140 #ifdef CONFIG_PREEMPT_NOTIFIERS
9141 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
9145 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
9148 #ifdef CONFIG_RT_MUTEXES
9149 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
9153 * The boot idle thread does lazy MMU switching as well:
9155 atomic_inc(&init_mm.mm_count);
9156 enter_lazy_tlb(&init_mm, current);
9159 * Make us the idle thread. Technically, schedule() should not be
9160 * called from this thread, however somewhere below it might be,
9161 * but because we are the idle thread, we just pick up running again
9162 * when this runqueue becomes "idle".
9164 init_idle(current, smp_processor_id());
9166 * During early bootup we pretend to be a normal task:
9168 current->sched_class = &fair_sched_class;
9170 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9171 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
9174 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
9175 alloc_bootmem_cpumask_var(&nohz.ilb_grp_nohz_mask);
9177 alloc_bootmem_cpumask_var(&cpu_isolated_map);
9180 scheduler_running = 1;
9183 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9184 void __might_sleep(char *file, int line)
9187 static unsigned long prev_jiffy; /* ratelimiting */
9189 if ((!in_atomic() && !irqs_disabled()) ||
9190 system_state != SYSTEM_RUNNING || oops_in_progress)
9192 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
9194 prev_jiffy = jiffies;
9197 "BUG: sleeping function called from invalid context at %s:%d\n",
9200 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9201 in_atomic(), irqs_disabled(),
9202 current->pid, current->comm);
9204 debug_show_held_locks(current);
9205 if (irqs_disabled())
9206 print_irqtrace_events(current);
9210 EXPORT_SYMBOL(__might_sleep);
9213 #ifdef CONFIG_MAGIC_SYSRQ
9214 static void normalize_task(struct rq *rq, struct task_struct *p)
9218 update_rq_clock(rq);
9219 on_rq = p->se.on_rq;
9221 deactivate_task(rq, p, 0);
9222 __setscheduler(rq, p, SCHED_NORMAL, 0);
9224 activate_task(rq, p, 0);
9225 resched_task(rq->curr);
9229 void normalize_rt_tasks(void)
9231 struct task_struct *g, *p;
9232 unsigned long flags;
9235 read_lock_irqsave(&tasklist_lock, flags);
9236 do_each_thread(g, p) {
9238 * Only normalize user tasks:
9243 p->se.exec_start = 0;
9244 #ifdef CONFIG_SCHEDSTATS
9245 p->se.wait_start = 0;
9246 p->se.sleep_start = 0;
9247 p->se.block_start = 0;
9252 * Renice negative nice level userspace
9255 if (TASK_NICE(p) < 0 && p->mm)
9256 set_user_nice(p, 0);
9260 spin_lock(&p->pi_lock);
9261 rq = __task_rq_lock(p);
9263 normalize_task(rq, p);
9265 __task_rq_unlock(rq);
9266 spin_unlock(&p->pi_lock);
9267 } while_each_thread(g, p);
9269 read_unlock_irqrestore(&tasklist_lock, flags);
9272 #endif /* CONFIG_MAGIC_SYSRQ */
9276 * These functions are only useful for the IA64 MCA handling.
9278 * They can only be called when the whole system has been
9279 * stopped - every CPU needs to be quiescent, and no scheduling
9280 * activity can take place. Using them for anything else would
9281 * be a serious bug, and as a result, they aren't even visible
9282 * under any other configuration.
9286 * curr_task - return the current task for a given cpu.
9287 * @cpu: the processor in question.
9289 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9291 struct task_struct *curr_task(int cpu)
9293 return cpu_curr(cpu);
9297 * set_curr_task - set the current task for a given cpu.
9298 * @cpu: the processor in question.
9299 * @p: the task pointer to set.
9301 * Description: This function must only be used when non-maskable interrupts
9302 * are serviced on a separate stack. It allows the architecture to switch the
9303 * notion of the current task on a cpu in a non-blocking manner. This function
9304 * must be called with all CPU's synchronized, and interrupts disabled, the
9305 * and caller must save the original value of the current task (see
9306 * curr_task() above) and restore that value before reenabling interrupts and
9307 * re-starting the system.
9309 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9311 void set_curr_task(int cpu, struct task_struct *p)
9318 #ifdef CONFIG_FAIR_GROUP_SCHED
9319 static void free_fair_sched_group(struct task_group *tg)
9323 for_each_possible_cpu(i) {
9325 kfree(tg->cfs_rq[i]);
9335 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9337 struct cfs_rq *cfs_rq;
9338 struct sched_entity *se;
9342 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
9345 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
9349 tg->shares = NICE_0_LOAD;
9351 for_each_possible_cpu(i) {
9354 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
9355 GFP_KERNEL, cpu_to_node(i));
9359 se = kzalloc_node(sizeof(struct sched_entity),
9360 GFP_KERNEL, cpu_to_node(i));
9364 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
9373 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9375 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
9376 &cpu_rq(cpu)->leaf_cfs_rq_list);
9379 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9381 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
9383 #else /* !CONFG_FAIR_GROUP_SCHED */
9384 static inline void free_fair_sched_group(struct task_group *tg)
9389 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
9394 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
9398 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
9401 #endif /* CONFIG_FAIR_GROUP_SCHED */
9403 #ifdef CONFIG_RT_GROUP_SCHED
9404 static void free_rt_sched_group(struct task_group *tg)
9408 destroy_rt_bandwidth(&tg->rt_bandwidth);
9410 for_each_possible_cpu(i) {
9412 kfree(tg->rt_rq[i]);
9414 kfree(tg->rt_se[i]);
9422 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9424 struct rt_rq *rt_rq;
9425 struct sched_rt_entity *rt_se;
9429 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
9432 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
9436 init_rt_bandwidth(&tg->rt_bandwidth,
9437 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
9439 for_each_possible_cpu(i) {
9442 rt_rq = kzalloc_node(sizeof(struct rt_rq),
9443 GFP_KERNEL, cpu_to_node(i));
9447 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
9448 GFP_KERNEL, cpu_to_node(i));
9452 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
9461 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9463 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
9464 &cpu_rq(cpu)->leaf_rt_rq_list);
9467 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9469 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
9471 #else /* !CONFIG_RT_GROUP_SCHED */
9472 static inline void free_rt_sched_group(struct task_group *tg)
9477 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
9482 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
9486 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
9489 #endif /* CONFIG_RT_GROUP_SCHED */
9491 #ifdef CONFIG_GROUP_SCHED
9492 static void free_sched_group(struct task_group *tg)
9494 free_fair_sched_group(tg);
9495 free_rt_sched_group(tg);
9499 /* allocate runqueue etc for a new task group */
9500 struct task_group *sched_create_group(struct task_group *parent)
9502 struct task_group *tg;
9503 unsigned long flags;
9506 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
9508 return ERR_PTR(-ENOMEM);
9510 if (!alloc_fair_sched_group(tg, parent))
9513 if (!alloc_rt_sched_group(tg, parent))
9516 spin_lock_irqsave(&task_group_lock, flags);
9517 for_each_possible_cpu(i) {
9518 register_fair_sched_group(tg, i);
9519 register_rt_sched_group(tg, i);
9521 list_add_rcu(&tg->list, &task_groups);
9523 WARN_ON(!parent); /* root should already exist */
9525 tg->parent = parent;
9526 INIT_LIST_HEAD(&tg->children);
9527 list_add_rcu(&tg->siblings, &parent->children);
9528 spin_unlock_irqrestore(&task_group_lock, flags);
9533 free_sched_group(tg);
9534 return ERR_PTR(-ENOMEM);
9537 /* rcu callback to free various structures associated with a task group */
9538 static void free_sched_group_rcu(struct rcu_head *rhp)
9540 /* now it should be safe to free those cfs_rqs */
9541 free_sched_group(container_of(rhp, struct task_group, rcu));
9544 /* Destroy runqueue etc associated with a task group */
9545 void sched_destroy_group(struct task_group *tg)
9547 unsigned long flags;
9550 spin_lock_irqsave(&task_group_lock, flags);
9551 for_each_possible_cpu(i) {
9552 unregister_fair_sched_group(tg, i);
9553 unregister_rt_sched_group(tg, i);
9555 list_del_rcu(&tg->list);
9556 list_del_rcu(&tg->siblings);
9557 spin_unlock_irqrestore(&task_group_lock, flags);
9559 /* wait for possible concurrent references to cfs_rqs complete */
9560 call_rcu(&tg->rcu, free_sched_group_rcu);
9563 /* change task's runqueue when it moves between groups.
9564 * The caller of this function should have put the task in its new group
9565 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9566 * reflect its new group.
9568 void sched_move_task(struct task_struct *tsk)
9571 unsigned long flags;
9574 rq = task_rq_lock(tsk, &flags);
9576 update_rq_clock(rq);
9578 running = task_current(rq, tsk);
9579 on_rq = tsk->se.on_rq;
9582 dequeue_task(rq, tsk, 0);
9583 if (unlikely(running))
9584 tsk->sched_class->put_prev_task(rq, tsk);
9586 set_task_rq(tsk, task_cpu(tsk));
9588 #ifdef CONFIG_FAIR_GROUP_SCHED
9589 if (tsk->sched_class->moved_group)
9590 tsk->sched_class->moved_group(tsk);
9593 if (unlikely(running))
9594 tsk->sched_class->set_curr_task(rq);
9596 enqueue_task(rq, tsk, 0);
9598 task_rq_unlock(rq, &flags);
9600 #endif /* CONFIG_GROUP_SCHED */
9602 #ifdef CONFIG_FAIR_GROUP_SCHED
9603 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9605 struct cfs_rq *cfs_rq = se->cfs_rq;
9610 dequeue_entity(cfs_rq, se, 0);
9612 se->load.weight = shares;
9613 se->load.inv_weight = 0;
9616 enqueue_entity(cfs_rq, se, 0);
9619 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9621 struct cfs_rq *cfs_rq = se->cfs_rq;
9622 struct rq *rq = cfs_rq->rq;
9623 unsigned long flags;
9625 spin_lock_irqsave(&rq->lock, flags);
9626 __set_se_shares(se, shares);
9627 spin_unlock_irqrestore(&rq->lock, flags);
9630 static DEFINE_MUTEX(shares_mutex);
9632 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9635 unsigned long flags;
9638 * We can't change the weight of the root cgroup.
9643 if (shares < MIN_SHARES)
9644 shares = MIN_SHARES;
9645 else if (shares > MAX_SHARES)
9646 shares = MAX_SHARES;
9648 mutex_lock(&shares_mutex);
9649 if (tg->shares == shares)
9652 spin_lock_irqsave(&task_group_lock, flags);
9653 for_each_possible_cpu(i)
9654 unregister_fair_sched_group(tg, i);
9655 list_del_rcu(&tg->siblings);
9656 spin_unlock_irqrestore(&task_group_lock, flags);
9658 /* wait for any ongoing reference to this group to finish */
9659 synchronize_sched();
9662 * Now we are free to modify the group's share on each cpu
9663 * w/o tripping rebalance_share or load_balance_fair.
9665 tg->shares = shares;
9666 for_each_possible_cpu(i) {
9670 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9671 set_se_shares(tg->se[i], shares);
9675 * Enable load balance activity on this group, by inserting it back on
9676 * each cpu's rq->leaf_cfs_rq_list.
9678 spin_lock_irqsave(&task_group_lock, flags);
9679 for_each_possible_cpu(i)
9680 register_fair_sched_group(tg, i);
9681 list_add_rcu(&tg->siblings, &tg->parent->children);
9682 spin_unlock_irqrestore(&task_group_lock, flags);
9684 mutex_unlock(&shares_mutex);
9688 unsigned long sched_group_shares(struct task_group *tg)
9694 #ifdef CONFIG_RT_GROUP_SCHED
9696 * Ensure that the real time constraints are schedulable.
9698 static DEFINE_MUTEX(rt_constraints_mutex);
9700 static unsigned long to_ratio(u64 period, u64 runtime)
9702 if (runtime == RUNTIME_INF)
9705 return div64_u64(runtime << 20, period);
9708 /* Must be called with tasklist_lock held */
9709 static inline int tg_has_rt_tasks(struct task_group *tg)
9711 struct task_struct *g, *p;
9713 do_each_thread(g, p) {
9714 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9716 } while_each_thread(g, p);
9721 struct rt_schedulable_data {
9722 struct task_group *tg;
9727 static int tg_schedulable(struct task_group *tg, void *data)
9729 struct rt_schedulable_data *d = data;
9730 struct task_group *child;
9731 unsigned long total, sum = 0;
9732 u64 period, runtime;
9734 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9735 runtime = tg->rt_bandwidth.rt_runtime;
9738 period = d->rt_period;
9739 runtime = d->rt_runtime;
9742 #ifdef CONFIG_USER_SCHED
9743 if (tg == &root_task_group) {
9744 period = global_rt_period();
9745 runtime = global_rt_runtime();
9750 * Cannot have more runtime than the period.
9752 if (runtime > period && runtime != RUNTIME_INF)
9756 * Ensure we don't starve existing RT tasks.
9758 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9761 total = to_ratio(period, runtime);
9764 * Nobody can have more than the global setting allows.
9766 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9770 * The sum of our children's runtime should not exceed our own.
9772 list_for_each_entry_rcu(child, &tg->children, siblings) {
9773 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9774 runtime = child->rt_bandwidth.rt_runtime;
9776 if (child == d->tg) {
9777 period = d->rt_period;
9778 runtime = d->rt_runtime;
9781 sum += to_ratio(period, runtime);
9790 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9792 struct rt_schedulable_data data = {
9794 .rt_period = period,
9795 .rt_runtime = runtime,
9798 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9801 static int tg_set_bandwidth(struct task_group *tg,
9802 u64 rt_period, u64 rt_runtime)
9806 mutex_lock(&rt_constraints_mutex);
9807 read_lock(&tasklist_lock);
9808 err = __rt_schedulable(tg, rt_period, rt_runtime);
9812 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9813 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9814 tg->rt_bandwidth.rt_runtime = rt_runtime;
9816 for_each_possible_cpu(i) {
9817 struct rt_rq *rt_rq = tg->rt_rq[i];
9819 spin_lock(&rt_rq->rt_runtime_lock);
9820 rt_rq->rt_runtime = rt_runtime;
9821 spin_unlock(&rt_rq->rt_runtime_lock);
9823 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9825 read_unlock(&tasklist_lock);
9826 mutex_unlock(&rt_constraints_mutex);
9831 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9833 u64 rt_runtime, rt_period;
9835 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9836 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9837 if (rt_runtime_us < 0)
9838 rt_runtime = RUNTIME_INF;
9840 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9843 long sched_group_rt_runtime(struct task_group *tg)
9847 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9850 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9851 do_div(rt_runtime_us, NSEC_PER_USEC);
9852 return rt_runtime_us;
9855 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9857 u64 rt_runtime, rt_period;
9859 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9860 rt_runtime = tg->rt_bandwidth.rt_runtime;
9865 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9868 long sched_group_rt_period(struct task_group *tg)
9872 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9873 do_div(rt_period_us, NSEC_PER_USEC);
9874 return rt_period_us;
9877 static int sched_rt_global_constraints(void)
9879 u64 runtime, period;
9882 if (sysctl_sched_rt_period <= 0)
9885 runtime = global_rt_runtime();
9886 period = global_rt_period();
9889 * Sanity check on the sysctl variables.
9891 if (runtime > period && runtime != RUNTIME_INF)
9894 mutex_lock(&rt_constraints_mutex);
9895 read_lock(&tasklist_lock);
9896 ret = __rt_schedulable(NULL, 0, 0);
9897 read_unlock(&tasklist_lock);
9898 mutex_unlock(&rt_constraints_mutex);
9903 int sched_rt_can_attach(struct task_group *tg, struct task_struct *tsk)
9905 /* Don't accept realtime tasks when there is no way for them to run */
9906 if (rt_task(tsk) && tg->rt_bandwidth.rt_runtime == 0)
9912 #else /* !CONFIG_RT_GROUP_SCHED */
9913 static int sched_rt_global_constraints(void)
9915 unsigned long flags;
9918 if (sysctl_sched_rt_period <= 0)
9922 * There's always some RT tasks in the root group
9923 * -- migration, kstopmachine etc..
9925 if (sysctl_sched_rt_runtime == 0)
9928 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9929 for_each_possible_cpu(i) {
9930 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9932 spin_lock(&rt_rq->rt_runtime_lock);
9933 rt_rq->rt_runtime = global_rt_runtime();
9934 spin_unlock(&rt_rq->rt_runtime_lock);
9936 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9940 #endif /* CONFIG_RT_GROUP_SCHED */
9942 int sched_rt_handler(struct ctl_table *table, int write,
9943 struct file *filp, void __user *buffer, size_t *lenp,
9947 int old_period, old_runtime;
9948 static DEFINE_MUTEX(mutex);
9951 old_period = sysctl_sched_rt_period;
9952 old_runtime = sysctl_sched_rt_runtime;
9954 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9956 if (!ret && write) {
9957 ret = sched_rt_global_constraints();
9959 sysctl_sched_rt_period = old_period;
9960 sysctl_sched_rt_runtime = old_runtime;
9962 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9963 def_rt_bandwidth.rt_period =
9964 ns_to_ktime(global_rt_period());
9967 mutex_unlock(&mutex);
9972 #ifdef CONFIG_CGROUP_SCHED
9974 /* return corresponding task_group object of a cgroup */
9975 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9977 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9978 struct task_group, css);
9981 static struct cgroup_subsys_state *
9982 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9984 struct task_group *tg, *parent;
9986 if (!cgrp->parent) {
9987 /* This is early initialization for the top cgroup */
9988 return &init_task_group.css;
9991 parent = cgroup_tg(cgrp->parent);
9992 tg = sched_create_group(parent);
9994 return ERR_PTR(-ENOMEM);
10000 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10002 struct task_group *tg = cgroup_tg(cgrp);
10004 sched_destroy_group(tg);
10008 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10009 struct task_struct *tsk)
10011 #ifdef CONFIG_RT_GROUP_SCHED
10012 if (!sched_rt_can_attach(cgroup_tg(cgrp), tsk))
10015 /* We don't support RT-tasks being in separate groups */
10016 if (tsk->sched_class != &fair_sched_class)
10024 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
10025 struct cgroup *old_cont, struct task_struct *tsk)
10027 sched_move_task(tsk);
10030 #ifdef CONFIG_FAIR_GROUP_SCHED
10031 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
10034 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
10037 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
10039 struct task_group *tg = cgroup_tg(cgrp);
10041 return (u64) tg->shares;
10043 #endif /* CONFIG_FAIR_GROUP_SCHED */
10045 #ifdef CONFIG_RT_GROUP_SCHED
10046 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
10049 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
10052 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
10054 return sched_group_rt_runtime(cgroup_tg(cgrp));
10057 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
10060 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
10063 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
10065 return sched_group_rt_period(cgroup_tg(cgrp));
10067 #endif /* CONFIG_RT_GROUP_SCHED */
10069 static struct cftype cpu_files[] = {
10070 #ifdef CONFIG_FAIR_GROUP_SCHED
10073 .read_u64 = cpu_shares_read_u64,
10074 .write_u64 = cpu_shares_write_u64,
10077 #ifdef CONFIG_RT_GROUP_SCHED
10079 .name = "rt_runtime_us",
10080 .read_s64 = cpu_rt_runtime_read,
10081 .write_s64 = cpu_rt_runtime_write,
10084 .name = "rt_period_us",
10085 .read_u64 = cpu_rt_period_read_uint,
10086 .write_u64 = cpu_rt_period_write_uint,
10091 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
10093 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
10096 struct cgroup_subsys cpu_cgroup_subsys = {
10098 .create = cpu_cgroup_create,
10099 .destroy = cpu_cgroup_destroy,
10100 .can_attach = cpu_cgroup_can_attach,
10101 .attach = cpu_cgroup_attach,
10102 .populate = cpu_cgroup_populate,
10103 .subsys_id = cpu_cgroup_subsys_id,
10107 #endif /* CONFIG_CGROUP_SCHED */
10109 #ifdef CONFIG_CGROUP_CPUACCT
10112 * CPU accounting code for task groups.
10114 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10115 * (balbir@in.ibm.com).
10118 /* track cpu usage of a group of tasks and its child groups */
10120 struct cgroup_subsys_state css;
10121 /* cpuusage holds pointer to a u64-type object on every cpu */
10123 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
10124 struct cpuacct *parent;
10127 struct cgroup_subsys cpuacct_subsys;
10129 /* return cpu accounting group corresponding to this container */
10130 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
10132 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
10133 struct cpuacct, css);
10136 /* return cpu accounting group to which this task belongs */
10137 static inline struct cpuacct *task_ca(struct task_struct *tsk)
10139 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
10140 struct cpuacct, css);
10143 /* create a new cpu accounting group */
10144 static struct cgroup_subsys_state *cpuacct_create(
10145 struct cgroup_subsys *ss, struct cgroup *cgrp)
10147 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
10153 ca->cpuusage = alloc_percpu(u64);
10157 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10158 if (percpu_counter_init(&ca->cpustat[i], 0))
10159 goto out_free_counters;
10162 ca->parent = cgroup_ca(cgrp->parent);
10168 percpu_counter_destroy(&ca->cpustat[i]);
10169 free_percpu(ca->cpuusage);
10173 return ERR_PTR(-ENOMEM);
10176 /* destroy an existing cpu accounting group */
10178 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
10180 struct cpuacct *ca = cgroup_ca(cgrp);
10183 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
10184 percpu_counter_destroy(&ca->cpustat[i]);
10185 free_percpu(ca->cpuusage);
10189 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
10191 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10194 #ifndef CONFIG_64BIT
10196 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10198 spin_lock_irq(&cpu_rq(cpu)->lock);
10200 spin_unlock_irq(&cpu_rq(cpu)->lock);
10208 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
10210 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10212 #ifndef CONFIG_64BIT
10214 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10216 spin_lock_irq(&cpu_rq(cpu)->lock);
10218 spin_unlock_irq(&cpu_rq(cpu)->lock);
10224 /* return total cpu usage (in nanoseconds) of a group */
10225 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
10227 struct cpuacct *ca = cgroup_ca(cgrp);
10228 u64 totalcpuusage = 0;
10231 for_each_present_cpu(i)
10232 totalcpuusage += cpuacct_cpuusage_read(ca, i);
10234 return totalcpuusage;
10237 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
10240 struct cpuacct *ca = cgroup_ca(cgrp);
10249 for_each_present_cpu(i)
10250 cpuacct_cpuusage_write(ca, i, 0);
10256 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
10257 struct seq_file *m)
10259 struct cpuacct *ca = cgroup_ca(cgroup);
10263 for_each_present_cpu(i) {
10264 percpu = cpuacct_cpuusage_read(ca, i);
10265 seq_printf(m, "%llu ", (unsigned long long) percpu);
10267 seq_printf(m, "\n");
10271 static const char *cpuacct_stat_desc[] = {
10272 [CPUACCT_STAT_USER] = "user",
10273 [CPUACCT_STAT_SYSTEM] = "system",
10276 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
10277 struct cgroup_map_cb *cb)
10279 struct cpuacct *ca = cgroup_ca(cgrp);
10282 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
10283 s64 val = percpu_counter_read(&ca->cpustat[i]);
10284 val = cputime64_to_clock_t(val);
10285 cb->fill(cb, cpuacct_stat_desc[i], val);
10290 static struct cftype files[] = {
10293 .read_u64 = cpuusage_read,
10294 .write_u64 = cpuusage_write,
10297 .name = "usage_percpu",
10298 .read_seq_string = cpuacct_percpu_seq_read,
10302 .read_map = cpuacct_stats_show,
10306 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
10308 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
10312 * charge this task's execution time to its accounting group.
10314 * called with rq->lock held.
10316 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
10318 struct cpuacct *ca;
10321 if (unlikely(!cpuacct_subsys.active))
10324 cpu = task_cpu(tsk);
10330 for (; ca; ca = ca->parent) {
10331 u64 *cpuusage = per_cpu_ptr(ca->cpuusage, cpu);
10332 *cpuusage += cputime;
10339 * Charge the system/user time to the task's accounting group.
10341 static void cpuacct_update_stats(struct task_struct *tsk,
10342 enum cpuacct_stat_index idx, cputime_t val)
10344 struct cpuacct *ca;
10346 if (unlikely(!cpuacct_subsys.active))
10353 percpu_counter_add(&ca->cpustat[idx], val);
10359 struct cgroup_subsys cpuacct_subsys = {
10361 .create = cpuacct_create,
10362 .destroy = cpuacct_destroy,
10363 .populate = cpuacct_populate,
10364 .subsys_id = cpuacct_subsys_id,
10366 #endif /* CONFIG_CGROUP_CPUACCT */