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)
123 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
124 * Since cpu_power is a 'constant', we can use a reciprocal divide.
126 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
128 return reciprocal_divide(load, sg->reciprocal_cpu_power);
132 * Each time a sched group cpu_power is changed,
133 * we must compute its reciprocal value
135 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
137 sg->__cpu_power += val;
138 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
142 static inline int rt_policy(int policy)
144 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
149 static inline int task_has_rt_policy(struct task_struct *p)
151 return rt_policy(p->policy);
155 * This is the priority-queue data structure of the RT scheduling class:
157 struct rt_prio_array {
158 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
159 struct list_head queue[MAX_RT_PRIO];
162 struct rt_bandwidth {
163 /* nests inside the rq lock: */
164 spinlock_t rt_runtime_lock;
167 struct hrtimer rt_period_timer;
170 static struct rt_bandwidth def_rt_bandwidth;
172 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
174 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
176 struct rt_bandwidth *rt_b =
177 container_of(timer, struct rt_bandwidth, rt_period_timer);
183 now = hrtimer_cb_get_time(timer);
184 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
189 idle = do_sched_rt_period_timer(rt_b, overrun);
192 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
196 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
198 rt_b->rt_period = ns_to_ktime(period);
199 rt_b->rt_runtime = runtime;
201 spin_lock_init(&rt_b->rt_runtime_lock);
203 hrtimer_init(&rt_b->rt_period_timer,
204 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
205 rt_b->rt_period_timer.function = sched_rt_period_timer;
208 static inline int rt_bandwidth_enabled(void)
210 return sysctl_sched_rt_runtime >= 0;
213 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
217 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
220 if (hrtimer_active(&rt_b->rt_period_timer))
223 spin_lock(&rt_b->rt_runtime_lock);
225 if (hrtimer_active(&rt_b->rt_period_timer))
228 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
229 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
230 hrtimer_start_expires(&rt_b->rt_period_timer,
233 spin_unlock(&rt_b->rt_runtime_lock);
236 #ifdef CONFIG_RT_GROUP_SCHED
237 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
239 hrtimer_cancel(&rt_b->rt_period_timer);
244 * sched_domains_mutex serializes calls to arch_init_sched_domains,
245 * detach_destroy_domains and partition_sched_domains.
247 static DEFINE_MUTEX(sched_domains_mutex);
249 #ifdef CONFIG_GROUP_SCHED
251 #include <linux/cgroup.h>
255 static LIST_HEAD(task_groups);
257 /* task group related information */
259 #ifdef CONFIG_CGROUP_SCHED
260 struct cgroup_subsys_state css;
263 #ifdef CONFIG_FAIR_GROUP_SCHED
264 /* schedulable entities of this group on each cpu */
265 struct sched_entity **se;
266 /* runqueue "owned" by this group on each cpu */
267 struct cfs_rq **cfs_rq;
268 unsigned long shares;
271 #ifdef CONFIG_RT_GROUP_SCHED
272 struct sched_rt_entity **rt_se;
273 struct rt_rq **rt_rq;
275 struct rt_bandwidth rt_bandwidth;
279 struct list_head list;
281 struct task_group *parent;
282 struct list_head siblings;
283 struct list_head children;
286 #ifdef CONFIG_USER_SCHED
290 * Every UID task group (including init_task_group aka UID-0) will
291 * be a child to this group.
293 struct task_group root_task_group;
295 #ifdef CONFIG_FAIR_GROUP_SCHED
296 /* Default task group's sched entity on each cpu */
297 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
298 /* Default task group's cfs_rq on each cpu */
299 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
300 #endif /* CONFIG_FAIR_GROUP_SCHED */
302 #ifdef CONFIG_RT_GROUP_SCHED
303 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
304 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
305 #endif /* CONFIG_RT_GROUP_SCHED */
306 #else /* !CONFIG_USER_SCHED */
307 #define root_task_group init_task_group
308 #endif /* CONFIG_USER_SCHED */
310 /* task_group_lock serializes add/remove of task groups and also changes to
311 * a task group's cpu shares.
313 static DEFINE_SPINLOCK(task_group_lock);
315 #ifdef CONFIG_FAIR_GROUP_SCHED
316 #ifdef CONFIG_USER_SCHED
317 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
318 #else /* !CONFIG_USER_SCHED */
319 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
320 #endif /* CONFIG_USER_SCHED */
323 * A weight of 0 or 1 can cause arithmetics problems.
324 * A weight of a cfs_rq is the sum of weights of which entities
325 * are queued on this cfs_rq, so a weight of a entity should not be
326 * too large, so as the shares value of a task group.
327 * (The default weight is 1024 - so there's no practical
328 * limitation from this.)
331 #define MAX_SHARES (1UL << 18)
333 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
336 /* Default task group.
337 * Every task in system belong to this group at bootup.
339 struct task_group init_task_group;
341 /* return group to which a task belongs */
342 static inline struct task_group *task_group(struct task_struct *p)
344 struct task_group *tg;
346 #ifdef CONFIG_USER_SCHED
348 #elif defined(CONFIG_CGROUP_SCHED)
349 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
350 struct task_group, css);
352 tg = &init_task_group;
357 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
358 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
360 #ifdef CONFIG_FAIR_GROUP_SCHED
361 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
362 p->se.parent = task_group(p)->se[cpu];
365 #ifdef CONFIG_RT_GROUP_SCHED
366 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
367 p->rt.parent = task_group(p)->rt_se[cpu];
373 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
374 static inline struct task_group *task_group(struct task_struct *p)
379 #endif /* CONFIG_GROUP_SCHED */
381 /* CFS-related fields in a runqueue */
383 struct load_weight load;
384 unsigned long nr_running;
389 struct rb_root tasks_timeline;
390 struct rb_node *rb_leftmost;
392 struct list_head tasks;
393 struct list_head *balance_iterator;
396 * 'curr' points to currently running entity on this cfs_rq.
397 * It is set to NULL otherwise (i.e when none are currently running).
399 struct sched_entity *curr, *next, *last;
401 unsigned int nr_spread_over;
403 #ifdef CONFIG_FAIR_GROUP_SCHED
404 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
407 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
408 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
409 * (like users, containers etc.)
411 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
412 * list is used during load balance.
414 struct list_head leaf_cfs_rq_list;
415 struct task_group *tg; /* group that "owns" this runqueue */
419 * the part of load.weight contributed by tasks
421 unsigned long task_weight;
424 * h_load = weight * f(tg)
426 * Where f(tg) is the recursive weight fraction assigned to
429 unsigned long h_load;
432 * this cpu's part of tg->shares
434 unsigned long shares;
437 * load.weight at the time we set shares
439 unsigned long rq_weight;
444 /* Real-Time classes' related field in a runqueue: */
446 struct rt_prio_array active;
447 unsigned long rt_nr_running;
448 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
449 int highest_prio; /* highest queued rt task prio */
452 unsigned long rt_nr_migratory;
458 /* Nests inside the rq lock: */
459 spinlock_t rt_runtime_lock;
461 #ifdef CONFIG_RT_GROUP_SCHED
462 unsigned long rt_nr_boosted;
465 struct list_head leaf_rt_rq_list;
466 struct task_group *tg;
467 struct sched_rt_entity *rt_se;
474 * We add the notion of a root-domain which will be used to define per-domain
475 * variables. Each exclusive cpuset essentially defines an island domain by
476 * fully partitioning the member cpus from any other cpuset. Whenever a new
477 * exclusive cpuset is created, we also create and attach a new root-domain
487 * The "RT overload" flag: it gets set if a CPU has more than
488 * one runnable RT task.
493 struct cpupri cpupri;
498 * By default the system creates a single root-domain with all cpus as
499 * members (mimicking the global state we have today).
501 static struct root_domain def_root_domain;
506 * This is the main, per-CPU runqueue data structure.
508 * Locking rule: those places that want to lock multiple runqueues
509 * (such as the load balancing or the thread migration code), lock
510 * acquire operations must be ordered by ascending &runqueue.
517 * nr_running and cpu_load should be in the same cacheline because
518 * remote CPUs use both these fields when doing load calculation.
520 unsigned long nr_running;
521 #define CPU_LOAD_IDX_MAX 5
522 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
523 unsigned char idle_at_tick;
525 unsigned long last_tick_seen;
526 unsigned char in_nohz_recently;
528 /* capture load from *all* tasks on this cpu: */
529 struct load_weight load;
530 unsigned long nr_load_updates;
536 #ifdef CONFIG_FAIR_GROUP_SCHED
537 /* list of leaf cfs_rq on this cpu: */
538 struct list_head leaf_cfs_rq_list;
540 #ifdef CONFIG_RT_GROUP_SCHED
541 struct list_head leaf_rt_rq_list;
545 * This is part of a global counter where only the total sum
546 * over all CPUs matters. A task can increase this counter on
547 * one CPU and if it got migrated afterwards it may decrease
548 * it on another CPU. Always updated under the runqueue lock:
550 unsigned long nr_uninterruptible;
552 struct task_struct *curr, *idle;
553 unsigned long next_balance;
554 struct mm_struct *prev_mm;
561 struct root_domain *rd;
562 struct sched_domain *sd;
564 /* For active balancing */
567 /* cpu of this runqueue: */
571 unsigned long avg_load_per_task;
573 struct task_struct *migration_thread;
574 struct list_head migration_queue;
577 #ifdef CONFIG_SCHED_HRTICK
579 int hrtick_csd_pending;
580 struct call_single_data hrtick_csd;
582 struct hrtimer hrtick_timer;
585 #ifdef CONFIG_SCHEDSTATS
587 struct sched_info rq_sched_info;
589 /* sys_sched_yield() stats */
590 unsigned int yld_exp_empty;
591 unsigned int yld_act_empty;
592 unsigned int yld_both_empty;
593 unsigned int yld_count;
595 /* schedule() stats */
596 unsigned int sched_switch;
597 unsigned int sched_count;
598 unsigned int sched_goidle;
600 /* try_to_wake_up() stats */
601 unsigned int ttwu_count;
602 unsigned int ttwu_local;
605 unsigned int bkl_count;
609 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
611 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
613 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
616 static inline int cpu_of(struct rq *rq)
626 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
627 * See detach_destroy_domains: synchronize_sched for details.
629 * The domain tree of any CPU may only be accessed from within
630 * preempt-disabled sections.
632 #define for_each_domain(cpu, __sd) \
633 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
635 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
636 #define this_rq() (&__get_cpu_var(runqueues))
637 #define task_rq(p) cpu_rq(task_cpu(p))
638 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
640 static inline void update_rq_clock(struct rq *rq)
642 rq->clock = sched_clock_cpu(cpu_of(rq));
646 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
648 #ifdef CONFIG_SCHED_DEBUG
649 # define const_debug __read_mostly
651 # define const_debug static const
657 * Returns true if the current cpu runqueue is locked.
658 * This interface allows printk to be called with the runqueue lock
659 * held and know whether or not it is OK to wake up the klogd.
661 int runqueue_is_locked(void)
664 struct rq *rq = cpu_rq(cpu);
667 ret = spin_is_locked(&rq->lock);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug unsigned int sysctl_sched_features =
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly char *sched_feat_names[] = {
699 #include "sched_features.h"
705 static int sched_feat_open(struct inode *inode, struct file *filp)
707 filp->private_data = inode->i_private;
712 sched_feat_read(struct file *filp, char __user *ubuf,
713 size_t cnt, loff_t *ppos)
720 for (i = 0; sched_feat_names[i]; i++) {
721 len += strlen(sched_feat_names[i]);
725 buf = kmalloc(len + 2, GFP_KERNEL);
729 for (i = 0; sched_feat_names[i]; i++) {
730 if (sysctl_sched_features & (1UL << i))
731 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
733 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
736 r += sprintf(buf + r, "\n");
737 WARN_ON(r >= len + 2);
739 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
747 sched_feat_write(struct file *filp, const char __user *ubuf,
748 size_t cnt, loff_t *ppos)
758 if (copy_from_user(&buf, ubuf, cnt))
763 if (strncmp(buf, "NO_", 3) == 0) {
768 for (i = 0; sched_feat_names[i]; i++) {
769 int len = strlen(sched_feat_names[i]);
771 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
773 sysctl_sched_features &= ~(1UL << i);
775 sysctl_sched_features |= (1UL << i);
780 if (!sched_feat_names[i])
788 static struct file_operations sched_feat_fops = {
789 .open = sched_feat_open,
790 .read = sched_feat_read,
791 .write = sched_feat_write,
794 static __init int sched_init_debug(void)
796 debugfs_create_file("sched_features", 0644, NULL, NULL,
801 late_initcall(sched_init_debug);
805 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
808 * Number of tasks to iterate in a single balance run.
809 * Limited because this is done with IRQs disabled.
811 const_debug unsigned int sysctl_sched_nr_migrate = 32;
814 * ratelimit for updating the group shares.
817 unsigned int sysctl_sched_shares_ratelimit = 250000;
820 * Inject some fuzzyness into changing the per-cpu group shares
821 * this avoids remote rq-locks at the expense of fairness.
824 unsigned int sysctl_sched_shares_thresh = 4;
827 * period over which we measure -rt task cpu usage in us.
830 unsigned int sysctl_sched_rt_period = 1000000;
832 static __read_mostly int scheduler_running;
835 * part of the period that we allow rt tasks to run in us.
838 int sysctl_sched_rt_runtime = 950000;
840 static inline u64 global_rt_period(void)
842 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
845 static inline u64 global_rt_runtime(void)
847 if (sysctl_sched_rt_runtime < 0)
850 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
853 #ifndef prepare_arch_switch
854 # define prepare_arch_switch(next) do { } while (0)
856 #ifndef finish_arch_switch
857 # define finish_arch_switch(prev) do { } while (0)
860 static inline int task_current(struct rq *rq, struct task_struct *p)
862 return rq->curr == p;
865 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
866 static inline int task_running(struct rq *rq, struct task_struct *p)
868 return task_current(rq, p);
871 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
875 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
877 #ifdef CONFIG_DEBUG_SPINLOCK
878 /* this is a valid case when another task releases the spinlock */
879 rq->lock.owner = current;
882 * If we are tracking spinlock dependencies then we have to
883 * fix up the runqueue lock - which gets 'carried over' from
886 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
888 spin_unlock_irq(&rq->lock);
891 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
892 static inline int task_running(struct rq *rq, struct task_struct *p)
897 return task_current(rq, p);
901 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
905 * We can optimise this out completely for !SMP, because the
906 * SMP rebalancing from interrupt is the only thing that cares
911 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
912 spin_unlock_irq(&rq->lock);
914 spin_unlock(&rq->lock);
918 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
922 * After ->oncpu is cleared, the task can be moved to a different CPU.
923 * We must ensure this doesn't happen until the switch is completely
929 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
933 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
936 * __task_rq_lock - lock the runqueue a given task resides on.
937 * Must be called interrupts disabled.
939 static inline struct rq *__task_rq_lock(struct task_struct *p)
943 struct rq *rq = task_rq(p);
944 spin_lock(&rq->lock);
945 if (likely(rq == task_rq(p)))
947 spin_unlock(&rq->lock);
952 * task_rq_lock - lock the runqueue a given task resides on and disable
953 * interrupts. Note the ordering: we can safely lookup the task_rq without
954 * explicitly disabling preemption.
956 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
962 local_irq_save(*flags);
964 spin_lock(&rq->lock);
965 if (likely(rq == task_rq(p)))
967 spin_unlock_irqrestore(&rq->lock, *flags);
971 void task_rq_unlock_wait(struct task_struct *p)
973 struct rq *rq = task_rq(p);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 spin_unlock_wait(&rq->lock);
979 static void __task_rq_unlock(struct rq *rq)
982 spin_unlock(&rq->lock);
985 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
988 spin_unlock_irqrestore(&rq->lock, *flags);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq *this_rq_lock(void)
1001 spin_lock(&rq->lock);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq *rq)
1025 if (!sched_feat(HRTICK))
1027 if (!cpu_active(cpu_of(rq)))
1029 return hrtimer_is_hres_active(&rq->hrtick_timer);
1032 static void hrtick_clear(struct rq *rq)
1034 if (hrtimer_active(&rq->hrtick_timer))
1035 hrtimer_cancel(&rq->hrtick_timer);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1044 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1046 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1048 spin_lock(&rq->lock);
1049 update_rq_clock(rq);
1050 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1051 spin_unlock(&rq->lock);
1053 return HRTIMER_NORESTART;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg)
1062 struct rq *rq = arg;
1064 spin_lock(&rq->lock);
1065 hrtimer_restart(&rq->hrtick_timer);
1066 rq->hrtick_csd_pending = 0;
1067 spin_unlock(&rq->lock);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq *rq, u64 delay)
1077 struct hrtimer *timer = &rq->hrtick_timer;
1078 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1080 hrtimer_set_expires(timer, time);
1082 if (rq == this_rq()) {
1083 hrtimer_restart(timer);
1084 } else if (!rq->hrtick_csd_pending) {
1085 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1086 rq->hrtick_csd_pending = 1;
1091 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1093 int cpu = (int)(long)hcpu;
1096 case CPU_UP_CANCELED:
1097 case CPU_UP_CANCELED_FROZEN:
1098 case CPU_DOWN_PREPARE:
1099 case CPU_DOWN_PREPARE_FROZEN:
1101 case CPU_DEAD_FROZEN:
1102 hrtick_clear(cpu_rq(cpu));
1109 static __init void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq *rq, u64 delay)
1121 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1124 static inline void init_hrtick(void)
1127 #endif /* CONFIG_SMP */
1129 static void init_rq_hrtick(struct rq *rq)
1132 rq->hrtick_csd_pending = 0;
1134 rq->hrtick_csd.flags = 0;
1135 rq->hrtick_csd.func = __hrtick_start;
1136 rq->hrtick_csd.info = rq;
1139 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1140 rq->hrtick_timer.function = hrtick;
1142 #else /* CONFIG_SCHED_HRTICK */
1143 static inline void hrtick_clear(struct rq *rq)
1147 static inline void init_rq_hrtick(struct rq *rq)
1151 static inline void init_hrtick(void)
1154 #endif /* CONFIG_SCHED_HRTICK */
1157 * resched_task - mark a task 'to be rescheduled now'.
1159 * On UP this means the setting of the need_resched flag, on SMP it
1160 * might also involve a cross-CPU call to trigger the scheduler on
1165 #ifndef tsk_is_polling
1166 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1169 static void resched_task(struct task_struct *p)
1173 assert_spin_locked(&task_rq(p)->lock);
1175 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1178 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1181 if (cpu == smp_processor_id())
1184 /* NEED_RESCHED must be visible before we test polling */
1186 if (!tsk_is_polling(p))
1187 smp_send_reschedule(cpu);
1190 static void resched_cpu(int cpu)
1192 struct rq *rq = cpu_rq(cpu);
1193 unsigned long flags;
1195 if (!spin_trylock_irqsave(&rq->lock, flags))
1197 resched_task(cpu_curr(cpu));
1198 spin_unlock_irqrestore(&rq->lock, flags);
1203 * When add_timer_on() enqueues a timer into the timer wheel of an
1204 * idle CPU then this timer might expire before the next timer event
1205 * which is scheduled to wake up that CPU. In case of a completely
1206 * idle system the next event might even be infinite time into the
1207 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1208 * leaves the inner idle loop so the newly added timer is taken into
1209 * account when the CPU goes back to idle and evaluates the timer
1210 * wheel for the next timer event.
1212 void wake_up_idle_cpu(int cpu)
1214 struct rq *rq = cpu_rq(cpu);
1216 if (cpu == smp_processor_id())
1220 * This is safe, as this function is called with the timer
1221 * wheel base lock of (cpu) held. When the CPU is on the way
1222 * to idle and has not yet set rq->curr to idle then it will
1223 * be serialized on the timer wheel base lock and take the new
1224 * timer into account automatically.
1226 if (rq->curr != rq->idle)
1230 * We can set TIF_RESCHED on the idle task of the other CPU
1231 * lockless. The worst case is that the other CPU runs the
1232 * idle task through an additional NOOP schedule()
1234 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1236 /* NEED_RESCHED must be visible before we test polling */
1238 if (!tsk_is_polling(rq->idle))
1239 smp_send_reschedule(cpu);
1241 #endif /* CONFIG_NO_HZ */
1243 #else /* !CONFIG_SMP */
1244 static void resched_task(struct task_struct *p)
1246 assert_spin_locked(&task_rq(p)->lock);
1247 set_tsk_need_resched(p);
1249 #endif /* CONFIG_SMP */
1251 #if BITS_PER_LONG == 32
1252 # define WMULT_CONST (~0UL)
1254 # define WMULT_CONST (1UL << 32)
1257 #define WMULT_SHIFT 32
1260 * Shift right and round:
1262 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1265 * delta *= weight / lw
1267 static unsigned long
1268 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1269 struct load_weight *lw)
1273 if (!lw->inv_weight) {
1274 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1277 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1281 tmp = (u64)delta_exec * weight;
1283 * Check whether we'd overflow the 64-bit multiplication:
1285 if (unlikely(tmp > WMULT_CONST))
1286 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1289 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1291 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1294 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1300 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1307 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1308 * of tasks with abnormal "nice" values across CPUs the contribution that
1309 * each task makes to its run queue's load is weighted according to its
1310 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1311 * scaled version of the new time slice allocation that they receive on time
1315 #define WEIGHT_IDLEPRIO 2
1316 #define WMULT_IDLEPRIO (1 << 31)
1319 * Nice levels are multiplicative, with a gentle 10% change for every
1320 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1321 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1322 * that remained on nice 0.
1324 * The "10% effect" is relative and cumulative: from _any_ nice level,
1325 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1326 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1327 * If a task goes up by ~10% and another task goes down by ~10% then
1328 * the relative distance between them is ~25%.)
1330 static const int prio_to_weight[40] = {
1331 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1332 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1333 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1334 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1335 /* 0 */ 1024, 820, 655, 526, 423,
1336 /* 5 */ 335, 272, 215, 172, 137,
1337 /* 10 */ 110, 87, 70, 56, 45,
1338 /* 15 */ 36, 29, 23, 18, 15,
1342 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1344 * In cases where the weight does not change often, we can use the
1345 * precalculated inverse to speed up arithmetics by turning divisions
1346 * into multiplications:
1348 static const u32 prio_to_wmult[40] = {
1349 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1350 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1351 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1352 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1353 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1354 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1355 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1356 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1359 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1362 * runqueue iterator, to support SMP load-balancing between different
1363 * scheduling classes, without having to expose their internal data
1364 * structures to the load-balancing proper:
1366 struct rq_iterator {
1368 struct task_struct *(*start)(void *);
1369 struct task_struct *(*next)(void *);
1373 static unsigned long
1374 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1375 unsigned long max_load_move, struct sched_domain *sd,
1376 enum cpu_idle_type idle, int *all_pinned,
1377 int *this_best_prio, struct rq_iterator *iterator);
1380 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1381 struct sched_domain *sd, enum cpu_idle_type idle,
1382 struct rq_iterator *iterator);
1385 #ifdef CONFIG_CGROUP_CPUACCT
1386 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1388 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1391 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1393 update_load_add(&rq->load, load);
1396 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1398 update_load_sub(&rq->load, load);
1401 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1402 typedef int (*tg_visitor)(struct task_group *, void *);
1405 * Iterate the full tree, calling @down when first entering a node and @up when
1406 * leaving it for the final time.
1408 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1410 struct task_group *parent, *child;
1414 parent = &root_task_group;
1416 ret = (*down)(parent, data);
1419 list_for_each_entry_rcu(child, &parent->children, siblings) {
1426 ret = (*up)(parent, data);
1431 parent = parent->parent;
1440 static int tg_nop(struct task_group *tg, void *data)
1447 static unsigned long source_load(int cpu, int type);
1448 static unsigned long target_load(int cpu, int type);
1449 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1451 static unsigned long cpu_avg_load_per_task(int cpu)
1453 struct rq *rq = cpu_rq(cpu);
1454 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1457 rq->avg_load_per_task = rq->load.weight / nr_running;
1459 rq->avg_load_per_task = 0;
1461 return rq->avg_load_per_task;
1464 #ifdef CONFIG_FAIR_GROUP_SCHED
1466 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1469 * Calculate and set the cpu's group shares.
1472 update_group_shares_cpu(struct task_group *tg, int cpu,
1473 unsigned long sd_shares, unsigned long sd_rq_weight)
1476 unsigned long shares;
1477 unsigned long rq_weight;
1482 rq_weight = tg->cfs_rq[cpu]->load.weight;
1485 * If there are currently no tasks on the cpu pretend there is one of
1486 * average load so that when a new task gets to run here it will not
1487 * get delayed by group starvation.
1491 rq_weight = NICE_0_LOAD;
1494 if (unlikely(rq_weight > sd_rq_weight))
1495 rq_weight = sd_rq_weight;
1498 * \Sum shares * rq_weight
1499 * shares = -----------------------
1503 shares = (sd_shares * rq_weight) / (sd_rq_weight + 1);
1504 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1506 if (abs(shares - tg->se[cpu]->load.weight) >
1507 sysctl_sched_shares_thresh) {
1508 struct rq *rq = cpu_rq(cpu);
1509 unsigned long flags;
1511 spin_lock_irqsave(&rq->lock, flags);
1513 * record the actual number of shares, not the boosted amount.
1515 tg->cfs_rq[cpu]->shares = boost ? 0 : shares;
1516 tg->cfs_rq[cpu]->rq_weight = rq_weight;
1518 __set_se_shares(tg->se[cpu], shares);
1519 spin_unlock_irqrestore(&rq->lock, flags);
1524 * Re-compute the task group their per cpu shares over the given domain.
1525 * This needs to be done in a bottom-up fashion because the rq weight of a
1526 * parent group depends on the shares of its child groups.
1528 static int tg_shares_up(struct task_group *tg, void *data)
1530 unsigned long rq_weight = 0;
1531 unsigned long shares = 0;
1532 struct sched_domain *sd = data;
1535 for_each_cpu_mask(i, sd->span) {
1536 rq_weight += tg->cfs_rq[i]->load.weight;
1537 shares += tg->cfs_rq[i]->shares;
1540 if ((!shares && rq_weight) || shares > tg->shares)
1541 shares = tg->shares;
1543 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1544 shares = tg->shares;
1547 rq_weight = cpus_weight(sd->span) * NICE_0_LOAD;
1549 for_each_cpu_mask(i, sd->span)
1550 update_group_shares_cpu(tg, i, shares, rq_weight);
1556 * Compute the cpu's hierarchical load factor for each task group.
1557 * This needs to be done in a top-down fashion because the load of a child
1558 * group is a fraction of its parents load.
1560 static int tg_load_down(struct task_group *tg, void *data)
1563 long cpu = (long)data;
1566 load = cpu_rq(cpu)->load.weight;
1568 load = tg->parent->cfs_rq[cpu]->h_load;
1569 load *= tg->cfs_rq[cpu]->shares;
1570 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1573 tg->cfs_rq[cpu]->h_load = load;
1578 static void update_shares(struct sched_domain *sd)
1580 u64 now = cpu_clock(raw_smp_processor_id());
1581 s64 elapsed = now - sd->last_update;
1583 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1584 sd->last_update = now;
1585 walk_tg_tree(tg_nop, tg_shares_up, sd);
1589 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1591 spin_unlock(&rq->lock);
1593 spin_lock(&rq->lock);
1596 static void update_h_load(long cpu)
1598 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1603 static inline void update_shares(struct sched_domain *sd)
1607 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1615 #ifdef CONFIG_FAIR_GROUP_SCHED
1616 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1619 cfs_rq->shares = shares;
1624 #include "sched_stats.h"
1625 #include "sched_idletask.c"
1626 #include "sched_fair.c"
1627 #include "sched_rt.c"
1628 #ifdef CONFIG_SCHED_DEBUG
1629 # include "sched_debug.c"
1632 #define sched_class_highest (&rt_sched_class)
1633 #define for_each_class(class) \
1634 for (class = sched_class_highest; class; class = class->next)
1636 static void inc_nr_running(struct rq *rq)
1641 static void dec_nr_running(struct rq *rq)
1646 static void set_load_weight(struct task_struct *p)
1648 if (task_has_rt_policy(p)) {
1649 p->se.load.weight = prio_to_weight[0] * 2;
1650 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1655 * SCHED_IDLE tasks get minimal weight:
1657 if (p->policy == SCHED_IDLE) {
1658 p->se.load.weight = WEIGHT_IDLEPRIO;
1659 p->se.load.inv_weight = WMULT_IDLEPRIO;
1663 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1664 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1667 static void update_avg(u64 *avg, u64 sample)
1669 s64 diff = sample - *avg;
1673 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1675 sched_info_queued(p);
1676 p->sched_class->enqueue_task(rq, p, wakeup);
1680 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1682 if (sleep && p->se.last_wakeup) {
1683 update_avg(&p->se.avg_overlap,
1684 p->se.sum_exec_runtime - p->se.last_wakeup);
1685 p->se.last_wakeup = 0;
1688 sched_info_dequeued(p);
1689 p->sched_class->dequeue_task(rq, p, sleep);
1694 * __normal_prio - return the priority that is based on the static prio
1696 static inline int __normal_prio(struct task_struct *p)
1698 return p->static_prio;
1702 * Calculate the expected normal priority: i.e. priority
1703 * without taking RT-inheritance into account. Might be
1704 * boosted by interactivity modifiers. Changes upon fork,
1705 * setprio syscalls, and whenever the interactivity
1706 * estimator recalculates.
1708 static inline int normal_prio(struct task_struct *p)
1712 if (task_has_rt_policy(p))
1713 prio = MAX_RT_PRIO-1 - p->rt_priority;
1715 prio = __normal_prio(p);
1720 * Calculate the current priority, i.e. the priority
1721 * taken into account by the scheduler. This value might
1722 * be boosted by RT tasks, or might be boosted by
1723 * interactivity modifiers. Will be RT if the task got
1724 * RT-boosted. If not then it returns p->normal_prio.
1726 static int effective_prio(struct task_struct *p)
1728 p->normal_prio = normal_prio(p);
1730 * If we are RT tasks or we were boosted to RT priority,
1731 * keep the priority unchanged. Otherwise, update priority
1732 * to the normal priority:
1734 if (!rt_prio(p->prio))
1735 return p->normal_prio;
1740 * activate_task - move a task to the runqueue.
1742 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1744 if (task_contributes_to_load(p))
1745 rq->nr_uninterruptible--;
1747 enqueue_task(rq, p, wakeup);
1752 * deactivate_task - remove a task from the runqueue.
1754 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1756 if (task_contributes_to_load(p))
1757 rq->nr_uninterruptible++;
1759 dequeue_task(rq, p, sleep);
1764 * task_curr - is this task currently executing on a CPU?
1765 * @p: the task in question.
1767 inline int task_curr(const struct task_struct *p)
1769 return cpu_curr(task_cpu(p)) == p;
1772 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1774 set_task_rq(p, cpu);
1777 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1778 * successfuly executed on another CPU. We must ensure that updates of
1779 * per-task data have been completed by this moment.
1782 task_thread_info(p)->cpu = cpu;
1786 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1787 const struct sched_class *prev_class,
1788 int oldprio, int running)
1790 if (prev_class != p->sched_class) {
1791 if (prev_class->switched_from)
1792 prev_class->switched_from(rq, p, running);
1793 p->sched_class->switched_to(rq, p, running);
1795 p->sched_class->prio_changed(rq, p, oldprio, running);
1800 /* Used instead of source_load when we know the type == 0 */
1801 static unsigned long weighted_cpuload(const int cpu)
1803 return cpu_rq(cpu)->load.weight;
1807 * Is this task likely cache-hot:
1810 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1815 * Buddy candidates are cache hot:
1817 if (sched_feat(CACHE_HOT_BUDDY) &&
1818 (&p->se == cfs_rq_of(&p->se)->next ||
1819 &p->se == cfs_rq_of(&p->se)->last))
1822 if (p->sched_class != &fair_sched_class)
1825 if (sysctl_sched_migration_cost == -1)
1827 if (sysctl_sched_migration_cost == 0)
1830 delta = now - p->se.exec_start;
1832 return delta < (s64)sysctl_sched_migration_cost;
1836 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1838 int old_cpu = task_cpu(p);
1839 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1840 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1841 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1844 clock_offset = old_rq->clock - new_rq->clock;
1846 #ifdef CONFIG_SCHEDSTATS
1847 if (p->se.wait_start)
1848 p->se.wait_start -= clock_offset;
1849 if (p->se.sleep_start)
1850 p->se.sleep_start -= clock_offset;
1851 if (p->se.block_start)
1852 p->se.block_start -= clock_offset;
1853 if (old_cpu != new_cpu) {
1854 schedstat_inc(p, se.nr_migrations);
1855 if (task_hot(p, old_rq->clock, NULL))
1856 schedstat_inc(p, se.nr_forced2_migrations);
1859 p->se.vruntime -= old_cfsrq->min_vruntime -
1860 new_cfsrq->min_vruntime;
1862 __set_task_cpu(p, new_cpu);
1865 struct migration_req {
1866 struct list_head list;
1868 struct task_struct *task;
1871 struct completion done;
1875 * The task's runqueue lock must be held.
1876 * Returns true if you have to wait for migration thread.
1879 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1881 struct rq *rq = task_rq(p);
1884 * If the task is not on a runqueue (and not running), then
1885 * it is sufficient to simply update the task's cpu field.
1887 if (!p->se.on_rq && !task_running(rq, p)) {
1888 set_task_cpu(p, dest_cpu);
1892 init_completion(&req->done);
1894 req->dest_cpu = dest_cpu;
1895 list_add(&req->list, &rq->migration_queue);
1901 * wait_task_inactive - wait for a thread to unschedule.
1903 * If @match_state is nonzero, it's the @p->state value just checked and
1904 * not expected to change. If it changes, i.e. @p might have woken up,
1905 * then return zero. When we succeed in waiting for @p to be off its CPU,
1906 * we return a positive number (its total switch count). If a second call
1907 * a short while later returns the same number, the caller can be sure that
1908 * @p has remained unscheduled the whole time.
1910 * The caller must ensure that the task *will* unschedule sometime soon,
1911 * else this function might spin for a *long* time. This function can't
1912 * be called with interrupts off, or it may introduce deadlock with
1913 * smp_call_function() if an IPI is sent by the same process we are
1914 * waiting to become inactive.
1916 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1918 unsigned long flags;
1925 * We do the initial early heuristics without holding
1926 * any task-queue locks at all. We'll only try to get
1927 * the runqueue lock when things look like they will
1933 * If the task is actively running on another CPU
1934 * still, just relax and busy-wait without holding
1937 * NOTE! Since we don't hold any locks, it's not
1938 * even sure that "rq" stays as the right runqueue!
1939 * But we don't care, since "task_running()" will
1940 * return false if the runqueue has changed and p
1941 * is actually now running somewhere else!
1943 while (task_running(rq, p)) {
1944 if (match_state && unlikely(p->state != match_state))
1950 * Ok, time to look more closely! We need the rq
1951 * lock now, to be *sure*. If we're wrong, we'll
1952 * just go back and repeat.
1954 rq = task_rq_lock(p, &flags);
1955 trace_sched_wait_task(rq, p);
1956 running = task_running(rq, p);
1957 on_rq = p->se.on_rq;
1959 if (!match_state || p->state == match_state)
1960 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
1961 task_rq_unlock(rq, &flags);
1964 * If it changed from the expected state, bail out now.
1966 if (unlikely(!ncsw))
1970 * Was it really running after all now that we
1971 * checked with the proper locks actually held?
1973 * Oops. Go back and try again..
1975 if (unlikely(running)) {
1981 * It's not enough that it's not actively running,
1982 * it must be off the runqueue _entirely_, and not
1985 * So if it wa still runnable (but just not actively
1986 * running right now), it's preempted, and we should
1987 * yield - it could be a while.
1989 if (unlikely(on_rq)) {
1990 schedule_timeout_uninterruptible(1);
1995 * Ahh, all good. It wasn't running, and it wasn't
1996 * runnable, which means that it will never become
1997 * running in the future either. We're all done!
2006 * kick_process - kick a running thread to enter/exit the kernel
2007 * @p: the to-be-kicked thread
2009 * Cause a process which is running on another CPU to enter
2010 * kernel-mode, without any delay. (to get signals handled.)
2012 * NOTE: this function doesnt have to take the runqueue lock,
2013 * because all it wants to ensure is that the remote task enters
2014 * the kernel. If the IPI races and the task has been migrated
2015 * to another CPU then no harm is done and the purpose has been
2018 void kick_process(struct task_struct *p)
2024 if ((cpu != smp_processor_id()) && task_curr(p))
2025 smp_send_reschedule(cpu);
2030 * Return a low guess at the load of a migration-source cpu weighted
2031 * according to the scheduling class and "nice" value.
2033 * We want to under-estimate the load of migration sources, to
2034 * balance conservatively.
2036 static unsigned long source_load(int cpu, int type)
2038 struct rq *rq = cpu_rq(cpu);
2039 unsigned long total = weighted_cpuload(cpu);
2041 if (type == 0 || !sched_feat(LB_BIAS))
2044 return min(rq->cpu_load[type-1], total);
2048 * Return a high guess at the load of a migration-target cpu weighted
2049 * according to the scheduling class and "nice" value.
2051 static unsigned long target_load(int cpu, int type)
2053 struct rq *rq = cpu_rq(cpu);
2054 unsigned long total = weighted_cpuload(cpu);
2056 if (type == 0 || !sched_feat(LB_BIAS))
2059 return max(rq->cpu_load[type-1], total);
2063 * find_idlest_group finds and returns the least busy CPU group within the
2066 static struct sched_group *
2067 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2069 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2070 unsigned long min_load = ULONG_MAX, this_load = 0;
2071 int load_idx = sd->forkexec_idx;
2072 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2075 unsigned long load, avg_load;
2079 /* Skip over this group if it has no CPUs allowed */
2080 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2083 local_group = cpu_isset(this_cpu, group->cpumask);
2085 /* Tally up the load of all CPUs in the group */
2088 for_each_cpu_mask_nr(i, group->cpumask) {
2089 /* Bias balancing toward cpus of our domain */
2091 load = source_load(i, load_idx);
2093 load = target_load(i, load_idx);
2098 /* Adjust by relative CPU power of the group */
2099 avg_load = sg_div_cpu_power(group,
2100 avg_load * SCHED_LOAD_SCALE);
2103 this_load = avg_load;
2105 } else if (avg_load < min_load) {
2106 min_load = avg_load;
2109 } while (group = group->next, group != sd->groups);
2111 if (!idlest || 100*this_load < imbalance*min_load)
2117 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2120 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2123 unsigned long load, min_load = ULONG_MAX;
2127 /* Traverse only the allowed CPUs */
2128 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2130 for_each_cpu_mask_nr(i, *tmp) {
2131 load = weighted_cpuload(i);
2133 if (load < min_load || (load == min_load && i == this_cpu)) {
2143 * sched_balance_self: balance the current task (running on cpu) in domains
2144 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2147 * Balance, ie. select the least loaded group.
2149 * Returns the target CPU number, or the same CPU if no balancing is needed.
2151 * preempt must be disabled.
2153 static int sched_balance_self(int cpu, int flag)
2155 struct task_struct *t = current;
2156 struct sched_domain *tmp, *sd = NULL;
2158 for_each_domain(cpu, tmp) {
2160 * If power savings logic is enabled for a domain, stop there.
2162 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2164 if (tmp->flags & flag)
2172 cpumask_t span, tmpmask;
2173 struct sched_group *group;
2174 int new_cpu, weight;
2176 if (!(sd->flags & flag)) {
2182 group = find_idlest_group(sd, t, cpu);
2188 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2189 if (new_cpu == -1 || new_cpu == cpu) {
2190 /* Now try balancing at a lower domain level of cpu */
2195 /* Now try balancing at a lower domain level of new_cpu */
2198 weight = cpus_weight(span);
2199 for_each_domain(cpu, tmp) {
2200 if (weight <= cpus_weight(tmp->span))
2202 if (tmp->flags & flag)
2205 /* while loop will break here if sd == NULL */
2211 #endif /* CONFIG_SMP */
2214 * try_to_wake_up - wake up a thread
2215 * @p: the to-be-woken-up thread
2216 * @state: the mask of task states that can be woken
2217 * @sync: do a synchronous wakeup?
2219 * Put it on the run-queue if it's not already there. The "current"
2220 * thread is always on the run-queue (except when the actual
2221 * re-schedule is in progress), and as such you're allowed to do
2222 * the simpler "current->state = TASK_RUNNING" to mark yourself
2223 * runnable without the overhead of this.
2225 * returns failure only if the task is already active.
2227 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2229 int cpu, orig_cpu, this_cpu, success = 0;
2230 unsigned long flags;
2234 if (!sched_feat(SYNC_WAKEUPS))
2238 if (sched_feat(LB_WAKEUP_UPDATE)) {
2239 struct sched_domain *sd;
2241 this_cpu = raw_smp_processor_id();
2244 for_each_domain(this_cpu, sd) {
2245 if (cpu_isset(cpu, sd->span)) {
2254 rq = task_rq_lock(p, &flags);
2255 old_state = p->state;
2256 if (!(old_state & state))
2264 this_cpu = smp_processor_id();
2267 if (unlikely(task_running(rq, p)))
2270 cpu = p->sched_class->select_task_rq(p, sync);
2271 if (cpu != orig_cpu) {
2272 set_task_cpu(p, cpu);
2273 task_rq_unlock(rq, &flags);
2274 /* might preempt at this point */
2275 rq = task_rq_lock(p, &flags);
2276 old_state = p->state;
2277 if (!(old_state & state))
2282 this_cpu = smp_processor_id();
2286 #ifdef CONFIG_SCHEDSTATS
2287 schedstat_inc(rq, ttwu_count);
2288 if (cpu == this_cpu)
2289 schedstat_inc(rq, ttwu_local);
2291 struct sched_domain *sd;
2292 for_each_domain(this_cpu, sd) {
2293 if (cpu_isset(cpu, sd->span)) {
2294 schedstat_inc(sd, ttwu_wake_remote);
2299 #endif /* CONFIG_SCHEDSTATS */
2302 #endif /* CONFIG_SMP */
2303 schedstat_inc(p, se.nr_wakeups);
2305 schedstat_inc(p, se.nr_wakeups_sync);
2306 if (orig_cpu != cpu)
2307 schedstat_inc(p, se.nr_wakeups_migrate);
2308 if (cpu == this_cpu)
2309 schedstat_inc(p, se.nr_wakeups_local);
2311 schedstat_inc(p, se.nr_wakeups_remote);
2312 update_rq_clock(rq);
2313 activate_task(rq, p, 1);
2317 trace_sched_wakeup(rq, p);
2318 check_preempt_curr(rq, p, sync);
2320 p->state = TASK_RUNNING;
2322 if (p->sched_class->task_wake_up)
2323 p->sched_class->task_wake_up(rq, p);
2326 current->se.last_wakeup = current->se.sum_exec_runtime;
2328 task_rq_unlock(rq, &flags);
2333 int wake_up_process(struct task_struct *p)
2335 return try_to_wake_up(p, TASK_ALL, 0);
2337 EXPORT_SYMBOL(wake_up_process);
2339 int wake_up_state(struct task_struct *p, unsigned int state)
2341 return try_to_wake_up(p, state, 0);
2345 * Perform scheduler related setup for a newly forked process p.
2346 * p is forked by current.
2348 * __sched_fork() is basic setup used by init_idle() too:
2350 static void __sched_fork(struct task_struct *p)
2352 p->se.exec_start = 0;
2353 p->se.sum_exec_runtime = 0;
2354 p->se.prev_sum_exec_runtime = 0;
2355 p->se.last_wakeup = 0;
2356 p->se.avg_overlap = 0;
2358 #ifdef CONFIG_SCHEDSTATS
2359 p->se.wait_start = 0;
2360 p->se.sum_sleep_runtime = 0;
2361 p->se.sleep_start = 0;
2362 p->se.block_start = 0;
2363 p->se.sleep_max = 0;
2364 p->se.block_max = 0;
2366 p->se.slice_max = 0;
2370 INIT_LIST_HEAD(&p->rt.run_list);
2372 INIT_LIST_HEAD(&p->se.group_node);
2374 #ifdef CONFIG_PREEMPT_NOTIFIERS
2375 INIT_HLIST_HEAD(&p->preempt_notifiers);
2379 * We mark the process as running here, but have not actually
2380 * inserted it onto the runqueue yet. This guarantees that
2381 * nobody will actually run it, and a signal or other external
2382 * event cannot wake it up and insert it on the runqueue either.
2384 p->state = TASK_RUNNING;
2388 * fork()/clone()-time setup:
2390 void sched_fork(struct task_struct *p, int clone_flags)
2392 int cpu = get_cpu();
2397 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2399 set_task_cpu(p, cpu);
2402 * Make sure we do not leak PI boosting priority to the child:
2404 p->prio = current->normal_prio;
2405 if (!rt_prio(p->prio))
2406 p->sched_class = &fair_sched_class;
2408 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2409 if (likely(sched_info_on()))
2410 memset(&p->sched_info, 0, sizeof(p->sched_info));
2412 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2415 #ifdef CONFIG_PREEMPT
2416 /* Want to start with kernel preemption disabled. */
2417 task_thread_info(p)->preempt_count = 1;
2423 * wake_up_new_task - wake up a newly created task for the first time.
2425 * This function will do some initial scheduler statistics housekeeping
2426 * that must be done for every newly created context, then puts the task
2427 * on the runqueue and wakes it.
2429 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2431 unsigned long flags;
2434 rq = task_rq_lock(p, &flags);
2435 BUG_ON(p->state != TASK_RUNNING);
2436 update_rq_clock(rq);
2438 p->prio = effective_prio(p);
2440 if (!p->sched_class->task_new || !current->se.on_rq) {
2441 activate_task(rq, p, 0);
2444 * Let the scheduling class do new task startup
2445 * management (if any):
2447 p->sched_class->task_new(rq, p);
2450 trace_sched_wakeup_new(rq, p);
2451 check_preempt_curr(rq, p, 0);
2453 if (p->sched_class->task_wake_up)
2454 p->sched_class->task_wake_up(rq, p);
2456 task_rq_unlock(rq, &flags);
2459 #ifdef CONFIG_PREEMPT_NOTIFIERS
2462 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2463 * @notifier: notifier struct to register
2465 void preempt_notifier_register(struct preempt_notifier *notifier)
2467 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2469 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2472 * preempt_notifier_unregister - no longer interested in preemption notifications
2473 * @notifier: notifier struct to unregister
2475 * This is safe to call from within a preemption notifier.
2477 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2479 hlist_del(¬ifier->link);
2481 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2483 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2485 struct preempt_notifier *notifier;
2486 struct hlist_node *node;
2488 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2489 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2493 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2494 struct task_struct *next)
2496 struct preempt_notifier *notifier;
2497 struct hlist_node *node;
2499 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2500 notifier->ops->sched_out(notifier, next);
2503 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2505 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2510 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2511 struct task_struct *next)
2515 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2518 * prepare_task_switch - prepare to switch tasks
2519 * @rq: the runqueue preparing to switch
2520 * @prev: the current task that is being switched out
2521 * @next: the task we are going to switch to.
2523 * This is called with the rq lock held and interrupts off. It must
2524 * be paired with a subsequent finish_task_switch after the context
2527 * prepare_task_switch sets up locking and calls architecture specific
2531 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2532 struct task_struct *next)
2534 fire_sched_out_preempt_notifiers(prev, next);
2535 prepare_lock_switch(rq, next);
2536 prepare_arch_switch(next);
2540 * finish_task_switch - clean up after a task-switch
2541 * @rq: runqueue associated with task-switch
2542 * @prev: the thread we just switched away from.
2544 * finish_task_switch must be called after the context switch, paired
2545 * with a prepare_task_switch call before the context switch.
2546 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2547 * and do any other architecture-specific cleanup actions.
2549 * Note that we may have delayed dropping an mm in context_switch(). If
2550 * so, we finish that here outside of the runqueue lock. (Doing it
2551 * with the lock held can cause deadlocks; see schedule() for
2554 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2555 __releases(rq->lock)
2557 struct mm_struct *mm = rq->prev_mm;
2563 * A task struct has one reference for the use as "current".
2564 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2565 * schedule one last time. The schedule call will never return, and
2566 * the scheduled task must drop that reference.
2567 * The test for TASK_DEAD must occur while the runqueue locks are
2568 * still held, otherwise prev could be scheduled on another cpu, die
2569 * there before we look at prev->state, and then the reference would
2571 * Manfred Spraul <manfred@colorfullife.com>
2573 prev_state = prev->state;
2574 finish_arch_switch(prev);
2575 finish_lock_switch(rq, prev);
2577 if (current->sched_class->post_schedule)
2578 current->sched_class->post_schedule(rq);
2581 fire_sched_in_preempt_notifiers(current);
2584 if (unlikely(prev_state == TASK_DEAD)) {
2586 * Remove function-return probe instances associated with this
2587 * task and put them back on the free list.
2589 kprobe_flush_task(prev);
2590 put_task_struct(prev);
2595 * schedule_tail - first thing a freshly forked thread must call.
2596 * @prev: the thread we just switched away from.
2598 asmlinkage void schedule_tail(struct task_struct *prev)
2599 __releases(rq->lock)
2601 struct rq *rq = this_rq();
2603 finish_task_switch(rq, prev);
2604 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2605 /* In this case, finish_task_switch does not reenable preemption */
2608 if (current->set_child_tid)
2609 put_user(task_pid_vnr(current), current->set_child_tid);
2613 * context_switch - switch to the new MM and the new
2614 * thread's register state.
2617 context_switch(struct rq *rq, struct task_struct *prev,
2618 struct task_struct *next)
2620 struct mm_struct *mm, *oldmm;
2622 prepare_task_switch(rq, prev, next);
2623 trace_sched_switch(rq, prev, next);
2625 oldmm = prev->active_mm;
2627 * For paravirt, this is coupled with an exit in switch_to to
2628 * combine the page table reload and the switch backend into
2631 arch_enter_lazy_cpu_mode();
2633 if (unlikely(!mm)) {
2634 next->active_mm = oldmm;
2635 atomic_inc(&oldmm->mm_count);
2636 enter_lazy_tlb(oldmm, next);
2638 switch_mm(oldmm, mm, next);
2640 if (unlikely(!prev->mm)) {
2641 prev->active_mm = NULL;
2642 rq->prev_mm = oldmm;
2645 * Since the runqueue lock will be released by the next
2646 * task (which is an invalid locking op but in the case
2647 * of the scheduler it's an obvious special-case), so we
2648 * do an early lockdep release here:
2650 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2651 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2654 /* Here we just switch the register state and the stack. */
2655 switch_to(prev, next, prev);
2659 * this_rq must be evaluated again because prev may have moved
2660 * CPUs since it called schedule(), thus the 'rq' on its stack
2661 * frame will be invalid.
2663 finish_task_switch(this_rq(), prev);
2667 * nr_running, nr_uninterruptible and nr_context_switches:
2669 * externally visible scheduler statistics: current number of runnable
2670 * threads, current number of uninterruptible-sleeping threads, total
2671 * number of context switches performed since bootup.
2673 unsigned long nr_running(void)
2675 unsigned long i, sum = 0;
2677 for_each_online_cpu(i)
2678 sum += cpu_rq(i)->nr_running;
2683 unsigned long nr_uninterruptible(void)
2685 unsigned long i, sum = 0;
2687 for_each_possible_cpu(i)
2688 sum += cpu_rq(i)->nr_uninterruptible;
2691 * Since we read the counters lockless, it might be slightly
2692 * inaccurate. Do not allow it to go below zero though:
2694 if (unlikely((long)sum < 0))
2700 unsigned long long nr_context_switches(void)
2703 unsigned long long sum = 0;
2705 for_each_possible_cpu(i)
2706 sum += cpu_rq(i)->nr_switches;
2711 unsigned long nr_iowait(void)
2713 unsigned long i, sum = 0;
2715 for_each_possible_cpu(i)
2716 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2721 unsigned long nr_active(void)
2723 unsigned long i, running = 0, uninterruptible = 0;
2725 for_each_online_cpu(i) {
2726 running += cpu_rq(i)->nr_running;
2727 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2730 if (unlikely((long)uninterruptible < 0))
2731 uninterruptible = 0;
2733 return running + uninterruptible;
2737 * Update rq->cpu_load[] statistics. This function is usually called every
2738 * scheduler tick (TICK_NSEC).
2740 static void update_cpu_load(struct rq *this_rq)
2742 unsigned long this_load = this_rq->load.weight;
2745 this_rq->nr_load_updates++;
2747 /* Update our load: */
2748 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2749 unsigned long old_load, new_load;
2751 /* scale is effectively 1 << i now, and >> i divides by scale */
2753 old_load = this_rq->cpu_load[i];
2754 new_load = this_load;
2756 * Round up the averaging division if load is increasing. This
2757 * prevents us from getting stuck on 9 if the load is 10, for
2760 if (new_load > old_load)
2761 new_load += scale-1;
2762 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2769 * double_rq_lock - safely lock two runqueues
2771 * Note this does not disable interrupts like task_rq_lock,
2772 * you need to do so manually before calling.
2774 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2775 __acquires(rq1->lock)
2776 __acquires(rq2->lock)
2778 BUG_ON(!irqs_disabled());
2780 spin_lock(&rq1->lock);
2781 __acquire(rq2->lock); /* Fake it out ;) */
2784 spin_lock(&rq1->lock);
2785 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2787 spin_lock(&rq2->lock);
2788 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2791 update_rq_clock(rq1);
2792 update_rq_clock(rq2);
2796 * double_rq_unlock - safely unlock two runqueues
2798 * Note this does not restore interrupts like task_rq_unlock,
2799 * you need to do so manually after calling.
2801 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2802 __releases(rq1->lock)
2803 __releases(rq2->lock)
2805 spin_unlock(&rq1->lock);
2807 spin_unlock(&rq2->lock);
2809 __release(rq2->lock);
2813 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2815 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2816 __releases(this_rq->lock)
2817 __acquires(busiest->lock)
2818 __acquires(this_rq->lock)
2822 if (unlikely(!irqs_disabled())) {
2823 /* printk() doesn't work good under rq->lock */
2824 spin_unlock(&this_rq->lock);
2827 if (unlikely(!spin_trylock(&busiest->lock))) {
2828 if (busiest < this_rq) {
2829 spin_unlock(&this_rq->lock);
2830 spin_lock(&busiest->lock);
2831 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
2834 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
2839 static void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
2840 __releases(busiest->lock)
2842 spin_unlock(&busiest->lock);
2843 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
2847 * If dest_cpu is allowed for this process, migrate the task to it.
2848 * This is accomplished by forcing the cpu_allowed mask to only
2849 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2850 * the cpu_allowed mask is restored.
2852 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2854 struct migration_req req;
2855 unsigned long flags;
2858 rq = task_rq_lock(p, &flags);
2859 if (!cpu_isset(dest_cpu, p->cpus_allowed)
2860 || unlikely(!cpu_active(dest_cpu)))
2863 trace_sched_migrate_task(rq, p, dest_cpu);
2864 /* force the process onto the specified CPU */
2865 if (migrate_task(p, dest_cpu, &req)) {
2866 /* Need to wait for migration thread (might exit: take ref). */
2867 struct task_struct *mt = rq->migration_thread;
2869 get_task_struct(mt);
2870 task_rq_unlock(rq, &flags);
2871 wake_up_process(mt);
2872 put_task_struct(mt);
2873 wait_for_completion(&req.done);
2878 task_rq_unlock(rq, &flags);
2882 * sched_exec - execve() is a valuable balancing opportunity, because at
2883 * this point the task has the smallest effective memory and cache footprint.
2885 void sched_exec(void)
2887 int new_cpu, this_cpu = get_cpu();
2888 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2890 if (new_cpu != this_cpu)
2891 sched_migrate_task(current, new_cpu);
2895 * pull_task - move a task from a remote runqueue to the local runqueue.
2896 * Both runqueues must be locked.
2898 static void pull_task(struct rq *src_rq, struct task_struct *p,
2899 struct rq *this_rq, int this_cpu)
2901 deactivate_task(src_rq, p, 0);
2902 set_task_cpu(p, this_cpu);
2903 activate_task(this_rq, p, 0);
2905 * Note that idle threads have a prio of MAX_PRIO, for this test
2906 * to be always true for them.
2908 check_preempt_curr(this_rq, p, 0);
2912 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2915 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2916 struct sched_domain *sd, enum cpu_idle_type idle,
2920 * We do not migrate tasks that are:
2921 * 1) running (obviously), or
2922 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2923 * 3) are cache-hot on their current CPU.
2925 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
2926 schedstat_inc(p, se.nr_failed_migrations_affine);
2931 if (task_running(rq, p)) {
2932 schedstat_inc(p, se.nr_failed_migrations_running);
2937 * Aggressive migration if:
2938 * 1) task is cache cold, or
2939 * 2) too many balance attempts have failed.
2942 if (!task_hot(p, rq->clock, sd) ||
2943 sd->nr_balance_failed > sd->cache_nice_tries) {
2944 #ifdef CONFIG_SCHEDSTATS
2945 if (task_hot(p, rq->clock, sd)) {
2946 schedstat_inc(sd, lb_hot_gained[idle]);
2947 schedstat_inc(p, se.nr_forced_migrations);
2953 if (task_hot(p, rq->clock, sd)) {
2954 schedstat_inc(p, se.nr_failed_migrations_hot);
2960 static unsigned long
2961 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2962 unsigned long max_load_move, struct sched_domain *sd,
2963 enum cpu_idle_type idle, int *all_pinned,
2964 int *this_best_prio, struct rq_iterator *iterator)
2966 int loops = 0, pulled = 0, pinned = 0;
2967 struct task_struct *p;
2968 long rem_load_move = max_load_move;
2970 if (max_load_move == 0)
2976 * Start the load-balancing iterator:
2978 p = iterator->start(iterator->arg);
2980 if (!p || loops++ > sysctl_sched_nr_migrate)
2983 if ((p->se.load.weight >> 1) > rem_load_move ||
2984 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
2985 p = iterator->next(iterator->arg);
2989 pull_task(busiest, p, this_rq, this_cpu);
2991 rem_load_move -= p->se.load.weight;
2994 * We only want to steal up to the prescribed amount of weighted load.
2996 if (rem_load_move > 0) {
2997 if (p->prio < *this_best_prio)
2998 *this_best_prio = p->prio;
2999 p = iterator->next(iterator->arg);
3004 * Right now, this is one of only two places pull_task() is called,
3005 * so we can safely collect pull_task() stats here rather than
3006 * inside pull_task().
3008 schedstat_add(sd, lb_gained[idle], pulled);
3011 *all_pinned = pinned;
3013 return max_load_move - rem_load_move;
3017 * move_tasks tries to move up to max_load_move weighted load from busiest to
3018 * this_rq, as part of a balancing operation within domain "sd".
3019 * Returns 1 if successful and 0 otherwise.
3021 * Called with both runqueues locked.
3023 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3024 unsigned long max_load_move,
3025 struct sched_domain *sd, enum cpu_idle_type idle,
3028 const struct sched_class *class = sched_class_highest;
3029 unsigned long total_load_moved = 0;
3030 int this_best_prio = this_rq->curr->prio;
3034 class->load_balance(this_rq, this_cpu, busiest,
3035 max_load_move - total_load_moved,
3036 sd, idle, all_pinned, &this_best_prio);
3037 class = class->next;
3039 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3042 } while (class && max_load_move > total_load_moved);
3044 return total_load_moved > 0;
3048 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3049 struct sched_domain *sd, enum cpu_idle_type idle,
3050 struct rq_iterator *iterator)
3052 struct task_struct *p = iterator->start(iterator->arg);
3056 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3057 pull_task(busiest, p, this_rq, this_cpu);
3059 * Right now, this is only the second place pull_task()
3060 * is called, so we can safely collect pull_task()
3061 * stats here rather than inside pull_task().
3063 schedstat_inc(sd, lb_gained[idle]);
3067 p = iterator->next(iterator->arg);
3074 * move_one_task tries to move exactly one task from busiest to this_rq, as
3075 * part of active balancing operations within "domain".
3076 * Returns 1 if successful and 0 otherwise.
3078 * Called with both runqueues locked.
3080 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3081 struct sched_domain *sd, enum cpu_idle_type idle)
3083 const struct sched_class *class;
3085 for (class = sched_class_highest; class; class = class->next)
3086 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3093 * find_busiest_group finds and returns the busiest CPU group within the
3094 * domain. It calculates and returns the amount of weighted load which
3095 * should be moved to restore balance via the imbalance parameter.
3097 static struct sched_group *
3098 find_busiest_group(struct sched_domain *sd, int this_cpu,
3099 unsigned long *imbalance, enum cpu_idle_type idle,
3100 int *sd_idle, const cpumask_t *cpus, int *balance)
3102 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3103 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3104 unsigned long max_pull;
3105 unsigned long busiest_load_per_task, busiest_nr_running;
3106 unsigned long this_load_per_task, this_nr_running;
3107 int load_idx, group_imb = 0;
3108 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3109 int power_savings_balance = 1;
3110 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3111 unsigned long min_nr_running = ULONG_MAX;
3112 struct sched_group *group_min = NULL, *group_leader = NULL;
3115 max_load = this_load = total_load = total_pwr = 0;
3116 busiest_load_per_task = busiest_nr_running = 0;
3117 this_load_per_task = this_nr_running = 0;
3119 if (idle == CPU_NOT_IDLE)
3120 load_idx = sd->busy_idx;
3121 else if (idle == CPU_NEWLY_IDLE)
3122 load_idx = sd->newidle_idx;
3124 load_idx = sd->idle_idx;
3127 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3130 int __group_imb = 0;
3131 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3132 unsigned long sum_nr_running, sum_weighted_load;
3133 unsigned long sum_avg_load_per_task;
3134 unsigned long avg_load_per_task;
3136 local_group = cpu_isset(this_cpu, group->cpumask);
3139 balance_cpu = first_cpu(group->cpumask);
3141 /* Tally up the load of all CPUs in the group */
3142 sum_weighted_load = sum_nr_running = avg_load = 0;
3143 sum_avg_load_per_task = avg_load_per_task = 0;
3146 min_cpu_load = ~0UL;
3148 for_each_cpu_mask_nr(i, group->cpumask) {
3151 if (!cpu_isset(i, *cpus))
3156 if (*sd_idle && rq->nr_running)
3159 /* Bias balancing toward cpus of our domain */
3161 if (idle_cpu(i) && !first_idle_cpu) {
3166 load = target_load(i, load_idx);
3168 load = source_load(i, load_idx);
3169 if (load > max_cpu_load)
3170 max_cpu_load = load;
3171 if (min_cpu_load > load)
3172 min_cpu_load = load;
3176 sum_nr_running += rq->nr_running;
3177 sum_weighted_load += weighted_cpuload(i);
3179 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3183 * First idle cpu or the first cpu(busiest) in this sched group
3184 * is eligible for doing load balancing at this and above
3185 * domains. In the newly idle case, we will allow all the cpu's
3186 * to do the newly idle load balance.
3188 if (idle != CPU_NEWLY_IDLE && local_group &&
3189 balance_cpu != this_cpu && balance) {
3194 total_load += avg_load;
3195 total_pwr += group->__cpu_power;
3197 /* Adjust by relative CPU power of the group */
3198 avg_load = sg_div_cpu_power(group,
3199 avg_load * SCHED_LOAD_SCALE);
3203 * Consider the group unbalanced when the imbalance is larger
3204 * than the average weight of two tasks.
3206 * APZ: with cgroup the avg task weight can vary wildly and
3207 * might not be a suitable number - should we keep a
3208 * normalized nr_running number somewhere that negates
3211 avg_load_per_task = sg_div_cpu_power(group,
3212 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3214 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3217 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3220 this_load = avg_load;
3222 this_nr_running = sum_nr_running;
3223 this_load_per_task = sum_weighted_load;
3224 } else if (avg_load > max_load &&
3225 (sum_nr_running > group_capacity || __group_imb)) {
3226 max_load = avg_load;
3228 busiest_nr_running = sum_nr_running;
3229 busiest_load_per_task = sum_weighted_load;
3230 group_imb = __group_imb;
3233 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3235 * Busy processors will not participate in power savings
3238 if (idle == CPU_NOT_IDLE ||
3239 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3243 * If the local group is idle or completely loaded
3244 * no need to do power savings balance at this domain
3246 if (local_group && (this_nr_running >= group_capacity ||
3248 power_savings_balance = 0;
3251 * If a group is already running at full capacity or idle,
3252 * don't include that group in power savings calculations
3254 if (!power_savings_balance || sum_nr_running >= group_capacity
3259 * Calculate the group which has the least non-idle load.
3260 * This is the group from where we need to pick up the load
3263 if ((sum_nr_running < min_nr_running) ||
3264 (sum_nr_running == min_nr_running &&
3265 first_cpu(group->cpumask) <
3266 first_cpu(group_min->cpumask))) {
3268 min_nr_running = sum_nr_running;
3269 min_load_per_task = sum_weighted_load /
3274 * Calculate the group which is almost near its
3275 * capacity but still has some space to pick up some load
3276 * from other group and save more power
3278 if (sum_nr_running <= group_capacity - 1) {
3279 if (sum_nr_running > leader_nr_running ||
3280 (sum_nr_running == leader_nr_running &&
3281 first_cpu(group->cpumask) >
3282 first_cpu(group_leader->cpumask))) {
3283 group_leader = group;
3284 leader_nr_running = sum_nr_running;
3289 group = group->next;
3290 } while (group != sd->groups);
3292 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3295 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3297 if (this_load >= avg_load ||
3298 100*max_load <= sd->imbalance_pct*this_load)
3301 busiest_load_per_task /= busiest_nr_running;
3303 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3306 * We're trying to get all the cpus to the average_load, so we don't
3307 * want to push ourselves above the average load, nor do we wish to
3308 * reduce the max loaded cpu below the average load, as either of these
3309 * actions would just result in more rebalancing later, and ping-pong
3310 * tasks around. Thus we look for the minimum possible imbalance.
3311 * Negative imbalances (*we* are more loaded than anyone else) will
3312 * be counted as no imbalance for these purposes -- we can't fix that
3313 * by pulling tasks to us. Be careful of negative numbers as they'll
3314 * appear as very large values with unsigned longs.
3316 if (max_load <= busiest_load_per_task)
3320 * In the presence of smp nice balancing, certain scenarios can have
3321 * max load less than avg load(as we skip the groups at or below
3322 * its cpu_power, while calculating max_load..)
3324 if (max_load < avg_load) {
3326 goto small_imbalance;
3329 /* Don't want to pull so many tasks that a group would go idle */
3330 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3332 /* How much load to actually move to equalise the imbalance */
3333 *imbalance = min(max_pull * busiest->__cpu_power,
3334 (avg_load - this_load) * this->__cpu_power)
3338 * if *imbalance is less than the average load per runnable task
3339 * there is no gaurantee that any tasks will be moved so we'll have
3340 * a think about bumping its value to force at least one task to be
3343 if (*imbalance < busiest_load_per_task) {
3344 unsigned long tmp, pwr_now, pwr_move;
3348 pwr_move = pwr_now = 0;
3350 if (this_nr_running) {
3351 this_load_per_task /= this_nr_running;
3352 if (busiest_load_per_task > this_load_per_task)
3355 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3357 if (max_load - this_load + busiest_load_per_task >=
3358 busiest_load_per_task * imbn) {
3359 *imbalance = busiest_load_per_task;
3364 * OK, we don't have enough imbalance to justify moving tasks,
3365 * however we may be able to increase total CPU power used by
3369 pwr_now += busiest->__cpu_power *
3370 min(busiest_load_per_task, max_load);
3371 pwr_now += this->__cpu_power *
3372 min(this_load_per_task, this_load);
3373 pwr_now /= SCHED_LOAD_SCALE;
3375 /* Amount of load we'd subtract */
3376 tmp = sg_div_cpu_power(busiest,
3377 busiest_load_per_task * SCHED_LOAD_SCALE);
3379 pwr_move += busiest->__cpu_power *
3380 min(busiest_load_per_task, max_load - tmp);
3382 /* Amount of load we'd add */
3383 if (max_load * busiest->__cpu_power <
3384 busiest_load_per_task * SCHED_LOAD_SCALE)
3385 tmp = sg_div_cpu_power(this,
3386 max_load * busiest->__cpu_power);
3388 tmp = sg_div_cpu_power(this,
3389 busiest_load_per_task * SCHED_LOAD_SCALE);
3390 pwr_move += this->__cpu_power *
3391 min(this_load_per_task, this_load + tmp);
3392 pwr_move /= SCHED_LOAD_SCALE;
3394 /* Move if we gain throughput */
3395 if (pwr_move > pwr_now)
3396 *imbalance = busiest_load_per_task;
3402 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3403 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3406 if (this == group_leader && group_leader != group_min) {
3407 *imbalance = min_load_per_task;
3417 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3420 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3421 unsigned long imbalance, const cpumask_t *cpus)
3423 struct rq *busiest = NULL, *rq;
3424 unsigned long max_load = 0;
3427 for_each_cpu_mask_nr(i, group->cpumask) {
3430 if (!cpu_isset(i, *cpus))
3434 wl = weighted_cpuload(i);
3436 if (rq->nr_running == 1 && wl > imbalance)
3439 if (wl > max_load) {
3449 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3450 * so long as it is large enough.
3452 #define MAX_PINNED_INTERVAL 512
3455 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3456 * tasks if there is an imbalance.
3458 static int load_balance(int this_cpu, struct rq *this_rq,
3459 struct sched_domain *sd, enum cpu_idle_type idle,
3460 int *balance, cpumask_t *cpus)
3462 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3463 struct sched_group *group;
3464 unsigned long imbalance;
3466 unsigned long flags;
3471 * When power savings policy is enabled for the parent domain, idle
3472 * sibling can pick up load irrespective of busy siblings. In this case,
3473 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3474 * portraying it as CPU_NOT_IDLE.
3476 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3477 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3480 schedstat_inc(sd, lb_count[idle]);
3484 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3491 schedstat_inc(sd, lb_nobusyg[idle]);
3495 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3497 schedstat_inc(sd, lb_nobusyq[idle]);
3501 BUG_ON(busiest == this_rq);
3503 schedstat_add(sd, lb_imbalance[idle], imbalance);
3506 if (busiest->nr_running > 1) {
3508 * Attempt to move tasks. If find_busiest_group has found
3509 * an imbalance but busiest->nr_running <= 1, the group is
3510 * still unbalanced. ld_moved simply stays zero, so it is
3511 * correctly treated as an imbalance.
3513 local_irq_save(flags);
3514 double_rq_lock(this_rq, busiest);
3515 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3516 imbalance, sd, idle, &all_pinned);
3517 double_rq_unlock(this_rq, busiest);
3518 local_irq_restore(flags);
3521 * some other cpu did the load balance for us.
3523 if (ld_moved && this_cpu != smp_processor_id())
3524 resched_cpu(this_cpu);
3526 /* All tasks on this runqueue were pinned by CPU affinity */
3527 if (unlikely(all_pinned)) {
3528 cpu_clear(cpu_of(busiest), *cpus);
3529 if (!cpus_empty(*cpus))
3536 schedstat_inc(sd, lb_failed[idle]);
3537 sd->nr_balance_failed++;
3539 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3541 spin_lock_irqsave(&busiest->lock, flags);
3543 /* don't kick the migration_thread, if the curr
3544 * task on busiest cpu can't be moved to this_cpu
3546 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3547 spin_unlock_irqrestore(&busiest->lock, flags);
3549 goto out_one_pinned;
3552 if (!busiest->active_balance) {
3553 busiest->active_balance = 1;
3554 busiest->push_cpu = this_cpu;
3557 spin_unlock_irqrestore(&busiest->lock, flags);
3559 wake_up_process(busiest->migration_thread);
3562 * We've kicked active balancing, reset the failure
3565 sd->nr_balance_failed = sd->cache_nice_tries+1;
3568 sd->nr_balance_failed = 0;
3570 if (likely(!active_balance)) {
3571 /* We were unbalanced, so reset the balancing interval */
3572 sd->balance_interval = sd->min_interval;
3575 * If we've begun active balancing, start to back off. This
3576 * case may not be covered by the all_pinned logic if there
3577 * is only 1 task on the busy runqueue (because we don't call
3580 if (sd->balance_interval < sd->max_interval)
3581 sd->balance_interval *= 2;
3584 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3585 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3591 schedstat_inc(sd, lb_balanced[idle]);
3593 sd->nr_balance_failed = 0;
3596 /* tune up the balancing interval */
3597 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3598 (sd->balance_interval < sd->max_interval))
3599 sd->balance_interval *= 2;
3601 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3602 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3613 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3614 * tasks if there is an imbalance.
3616 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3617 * this_rq is locked.
3620 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3623 struct sched_group *group;
3624 struct rq *busiest = NULL;
3625 unsigned long imbalance;
3633 * When power savings policy is enabled for the parent domain, idle
3634 * sibling can pick up load irrespective of busy siblings. In this case,
3635 * let the state of idle sibling percolate up as IDLE, instead of
3636 * portraying it as CPU_NOT_IDLE.
3638 if (sd->flags & SD_SHARE_CPUPOWER &&
3639 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3642 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3644 update_shares_locked(this_rq, sd);
3645 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3646 &sd_idle, cpus, NULL);
3648 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3652 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3654 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3658 BUG_ON(busiest == this_rq);
3660 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3663 if (busiest->nr_running > 1) {
3664 /* Attempt to move tasks */
3665 double_lock_balance(this_rq, busiest);
3666 /* this_rq->clock is already updated */
3667 update_rq_clock(busiest);
3668 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3669 imbalance, sd, CPU_NEWLY_IDLE,
3671 double_unlock_balance(this_rq, busiest);
3673 if (unlikely(all_pinned)) {
3674 cpu_clear(cpu_of(busiest), *cpus);
3675 if (!cpus_empty(*cpus))
3681 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3682 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3683 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3686 sd->nr_balance_failed = 0;
3688 update_shares_locked(this_rq, sd);
3692 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3693 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3694 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3696 sd->nr_balance_failed = 0;
3702 * idle_balance is called by schedule() if this_cpu is about to become
3703 * idle. Attempts to pull tasks from other CPUs.
3705 static void idle_balance(int this_cpu, struct rq *this_rq)
3707 struct sched_domain *sd;
3708 int pulled_task = -1;
3709 unsigned long next_balance = jiffies + HZ;
3712 for_each_domain(this_cpu, sd) {
3713 unsigned long interval;
3715 if (!(sd->flags & SD_LOAD_BALANCE))
3718 if (sd->flags & SD_BALANCE_NEWIDLE)
3719 /* If we've pulled tasks over stop searching: */
3720 pulled_task = load_balance_newidle(this_cpu, this_rq,
3723 interval = msecs_to_jiffies(sd->balance_interval);
3724 if (time_after(next_balance, sd->last_balance + interval))
3725 next_balance = sd->last_balance + interval;
3729 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3731 * We are going idle. next_balance may be set based on
3732 * a busy processor. So reset next_balance.
3734 this_rq->next_balance = next_balance;
3739 * active_load_balance is run by migration threads. It pushes running tasks
3740 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3741 * running on each physical CPU where possible, and avoids physical /
3742 * logical imbalances.
3744 * Called with busiest_rq locked.
3746 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3748 int target_cpu = busiest_rq->push_cpu;
3749 struct sched_domain *sd;
3750 struct rq *target_rq;
3752 /* Is there any task to move? */
3753 if (busiest_rq->nr_running <= 1)
3756 target_rq = cpu_rq(target_cpu);
3759 * This condition is "impossible", if it occurs
3760 * we need to fix it. Originally reported by
3761 * Bjorn Helgaas on a 128-cpu setup.
3763 BUG_ON(busiest_rq == target_rq);
3765 /* move a task from busiest_rq to target_rq */
3766 double_lock_balance(busiest_rq, target_rq);
3767 update_rq_clock(busiest_rq);
3768 update_rq_clock(target_rq);
3770 /* Search for an sd spanning us and the target CPU. */
3771 for_each_domain(target_cpu, sd) {
3772 if ((sd->flags & SD_LOAD_BALANCE) &&
3773 cpu_isset(busiest_cpu, sd->span))
3778 schedstat_inc(sd, alb_count);
3780 if (move_one_task(target_rq, target_cpu, busiest_rq,
3782 schedstat_inc(sd, alb_pushed);
3784 schedstat_inc(sd, alb_failed);
3786 double_unlock_balance(busiest_rq, target_rq);
3791 atomic_t load_balancer;
3793 } nohz ____cacheline_aligned = {
3794 .load_balancer = ATOMIC_INIT(-1),
3795 .cpu_mask = CPU_MASK_NONE,
3799 * This routine will try to nominate the ilb (idle load balancing)
3800 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3801 * load balancing on behalf of all those cpus. If all the cpus in the system
3802 * go into this tickless mode, then there will be no ilb owner (as there is
3803 * no need for one) and all the cpus will sleep till the next wakeup event
3806 * For the ilb owner, tick is not stopped. And this tick will be used
3807 * for idle load balancing. ilb owner will still be part of
3810 * While stopping the tick, this cpu will become the ilb owner if there
3811 * is no other owner. And will be the owner till that cpu becomes busy
3812 * or if all cpus in the system stop their ticks at which point
3813 * there is no need for ilb owner.
3815 * When the ilb owner becomes busy, it nominates another owner, during the
3816 * next busy scheduler_tick()
3818 int select_nohz_load_balancer(int stop_tick)
3820 int cpu = smp_processor_id();
3823 cpu_set(cpu, nohz.cpu_mask);
3824 cpu_rq(cpu)->in_nohz_recently = 1;
3827 * If we are going offline and still the leader, give up!
3829 if (!cpu_active(cpu) &&
3830 atomic_read(&nohz.load_balancer) == cpu) {
3831 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3836 /* time for ilb owner also to sleep */
3837 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3838 if (atomic_read(&nohz.load_balancer) == cpu)
3839 atomic_set(&nohz.load_balancer, -1);
3843 if (atomic_read(&nohz.load_balancer) == -1) {
3844 /* make me the ilb owner */
3845 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3847 } else if (atomic_read(&nohz.load_balancer) == cpu)
3850 if (!cpu_isset(cpu, nohz.cpu_mask))
3853 cpu_clear(cpu, nohz.cpu_mask);
3855 if (atomic_read(&nohz.load_balancer) == cpu)
3856 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3863 static DEFINE_SPINLOCK(balancing);
3866 * It checks each scheduling domain to see if it is due to be balanced,
3867 * and initiates a balancing operation if so.
3869 * Balancing parameters are set up in arch_init_sched_domains.
3871 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3874 struct rq *rq = cpu_rq(cpu);
3875 unsigned long interval;
3876 struct sched_domain *sd;
3877 /* Earliest time when we have to do rebalance again */
3878 unsigned long next_balance = jiffies + 60*HZ;
3879 int update_next_balance = 0;
3883 for_each_domain(cpu, sd) {
3884 if (!(sd->flags & SD_LOAD_BALANCE))
3887 interval = sd->balance_interval;
3888 if (idle != CPU_IDLE)
3889 interval *= sd->busy_factor;
3891 /* scale ms to jiffies */
3892 interval = msecs_to_jiffies(interval);
3893 if (unlikely(!interval))
3895 if (interval > HZ*NR_CPUS/10)
3896 interval = HZ*NR_CPUS/10;
3898 need_serialize = sd->flags & SD_SERIALIZE;
3900 if (need_serialize) {
3901 if (!spin_trylock(&balancing))
3905 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3906 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
3908 * We've pulled tasks over so either we're no
3909 * longer idle, or one of our SMT siblings is
3912 idle = CPU_NOT_IDLE;
3914 sd->last_balance = jiffies;
3917 spin_unlock(&balancing);
3919 if (time_after(next_balance, sd->last_balance + interval)) {
3920 next_balance = sd->last_balance + interval;
3921 update_next_balance = 1;
3925 * Stop the load balance at this level. There is another
3926 * CPU in our sched group which is doing load balancing more
3934 * next_balance will be updated only when there is a need.
3935 * When the cpu is attached to null domain for ex, it will not be
3938 if (likely(update_next_balance))
3939 rq->next_balance = next_balance;
3943 * run_rebalance_domains is triggered when needed from the scheduler tick.
3944 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3945 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3947 static void run_rebalance_domains(struct softirq_action *h)
3949 int this_cpu = smp_processor_id();
3950 struct rq *this_rq = cpu_rq(this_cpu);
3951 enum cpu_idle_type idle = this_rq->idle_at_tick ?
3952 CPU_IDLE : CPU_NOT_IDLE;
3954 rebalance_domains(this_cpu, idle);
3958 * If this cpu is the owner for idle load balancing, then do the
3959 * balancing on behalf of the other idle cpus whose ticks are
3962 if (this_rq->idle_at_tick &&
3963 atomic_read(&nohz.load_balancer) == this_cpu) {
3964 cpumask_t cpus = nohz.cpu_mask;
3968 cpu_clear(this_cpu, cpus);
3969 for_each_cpu_mask_nr(balance_cpu, cpus) {
3971 * If this cpu gets work to do, stop the load balancing
3972 * work being done for other cpus. Next load
3973 * balancing owner will pick it up.
3978 rebalance_domains(balance_cpu, CPU_IDLE);
3980 rq = cpu_rq(balance_cpu);
3981 if (time_after(this_rq->next_balance, rq->next_balance))
3982 this_rq->next_balance = rq->next_balance;
3989 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3991 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3992 * idle load balancing owner or decide to stop the periodic load balancing,
3993 * if the whole system is idle.
3995 static inline void trigger_load_balance(struct rq *rq, int cpu)
3999 * If we were in the nohz mode recently and busy at the current
4000 * scheduler tick, then check if we need to nominate new idle
4003 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4004 rq->in_nohz_recently = 0;
4006 if (atomic_read(&nohz.load_balancer) == cpu) {
4007 cpu_clear(cpu, nohz.cpu_mask);
4008 atomic_set(&nohz.load_balancer, -1);
4011 if (atomic_read(&nohz.load_balancer) == -1) {
4013 * simple selection for now: Nominate the
4014 * first cpu in the nohz list to be the next
4017 * TBD: Traverse the sched domains and nominate
4018 * the nearest cpu in the nohz.cpu_mask.
4020 int ilb = first_cpu(nohz.cpu_mask);
4022 if (ilb < nr_cpu_ids)
4028 * If this cpu is idle and doing idle load balancing for all the
4029 * cpus with ticks stopped, is it time for that to stop?
4031 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4032 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4038 * If this cpu is idle and the idle load balancing is done by
4039 * someone else, then no need raise the SCHED_SOFTIRQ
4041 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4042 cpu_isset(cpu, nohz.cpu_mask))
4045 if (time_after_eq(jiffies, rq->next_balance))
4046 raise_softirq(SCHED_SOFTIRQ);
4049 #else /* CONFIG_SMP */
4052 * on UP we do not need to balance between CPUs:
4054 static inline void idle_balance(int cpu, struct rq *rq)
4060 DEFINE_PER_CPU(struct kernel_stat, kstat);
4062 EXPORT_PER_CPU_SYMBOL(kstat);
4065 * Return any ns on the sched_clock that have not yet been banked in
4066 * @p in case that task is currently running.
4068 unsigned long long task_delta_exec(struct task_struct *p)
4070 unsigned long flags;
4074 rq = task_rq_lock(p, &flags);
4076 if (task_current(rq, p)) {
4079 update_rq_clock(rq);
4080 delta_exec = rq->clock - p->se.exec_start;
4081 if ((s64)delta_exec > 0)
4085 task_rq_unlock(rq, &flags);
4091 * Account user cpu time to a process.
4092 * @p: the process that the cpu time gets accounted to
4093 * @cputime: the cpu time spent in user space since the last update
4095 void account_user_time(struct task_struct *p, cputime_t cputime)
4097 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4100 p->utime = cputime_add(p->utime, cputime);
4101 account_group_user_time(p, cputime);
4103 /* Add user time to cpustat. */
4104 tmp = cputime_to_cputime64(cputime);
4105 if (TASK_NICE(p) > 0)
4106 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4108 cpustat->user = cputime64_add(cpustat->user, tmp);
4109 /* Account for user time used */
4110 acct_update_integrals(p);
4114 * Account guest cpu time to a process.
4115 * @p: the process that the cpu time gets accounted to
4116 * @cputime: the cpu time spent in virtual machine since the last update
4118 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4121 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4123 tmp = cputime_to_cputime64(cputime);
4125 p->utime = cputime_add(p->utime, cputime);
4126 account_group_user_time(p, cputime);
4127 p->gtime = cputime_add(p->gtime, cputime);
4129 cpustat->user = cputime64_add(cpustat->user, tmp);
4130 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4134 * Account scaled user cpu time to a process.
4135 * @p: the process that the cpu time gets accounted to
4136 * @cputime: the cpu time spent in user space since the last update
4138 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4140 p->utimescaled = cputime_add(p->utimescaled, cputime);
4144 * Account system cpu time to a process.
4145 * @p: the process that the cpu time gets accounted to
4146 * @hardirq_offset: the offset to subtract from hardirq_count()
4147 * @cputime: the cpu time spent in kernel space since the last update
4149 void account_system_time(struct task_struct *p, int hardirq_offset,
4152 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4153 struct rq *rq = this_rq();
4156 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4157 account_guest_time(p, cputime);
4161 p->stime = cputime_add(p->stime, cputime);
4162 account_group_system_time(p, cputime);
4164 /* Add system time to cpustat. */
4165 tmp = cputime_to_cputime64(cputime);
4166 if (hardirq_count() - hardirq_offset)
4167 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4168 else if (softirq_count())
4169 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4170 else if (p != rq->idle)
4171 cpustat->system = cputime64_add(cpustat->system, tmp);
4172 else if (atomic_read(&rq->nr_iowait) > 0)
4173 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4175 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4176 /* Account for system time used */
4177 acct_update_integrals(p);
4181 * Account scaled system cpu time to a process.
4182 * @p: the process that the cpu time gets accounted to
4183 * @hardirq_offset: the offset to subtract from hardirq_count()
4184 * @cputime: the cpu time spent in kernel space since the last update
4186 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4188 p->stimescaled = cputime_add(p->stimescaled, cputime);
4192 * Account for involuntary wait time.
4193 * @p: the process from which the cpu time has been stolen
4194 * @steal: the cpu time spent in involuntary wait
4196 void account_steal_time(struct task_struct *p, cputime_t steal)
4198 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4199 cputime64_t tmp = cputime_to_cputime64(steal);
4200 struct rq *rq = this_rq();
4202 if (p == rq->idle) {
4203 p->stime = cputime_add(p->stime, steal);
4204 account_group_system_time(p, steal);
4205 if (atomic_read(&rq->nr_iowait) > 0)
4206 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4208 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4210 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4214 * Use precise platform statistics if available:
4216 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4217 cputime_t task_utime(struct task_struct *p)
4222 cputime_t task_stime(struct task_struct *p)
4227 cputime_t task_utime(struct task_struct *p)
4229 clock_t utime = cputime_to_clock_t(p->utime),
4230 total = utime + cputime_to_clock_t(p->stime);
4234 * Use CFS's precise accounting:
4236 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4240 do_div(temp, total);
4242 utime = (clock_t)temp;
4244 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4245 return p->prev_utime;
4248 cputime_t task_stime(struct task_struct *p)
4253 * Use CFS's precise accounting. (we subtract utime from
4254 * the total, to make sure the total observed by userspace
4255 * grows monotonically - apps rely on that):
4257 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4258 cputime_to_clock_t(task_utime(p));
4261 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4263 return p->prev_stime;
4267 inline cputime_t task_gtime(struct task_struct *p)
4273 * This function gets called by the timer code, with HZ frequency.
4274 * We call it with interrupts disabled.
4276 * It also gets called by the fork code, when changing the parent's
4279 void scheduler_tick(void)
4281 int cpu = smp_processor_id();
4282 struct rq *rq = cpu_rq(cpu);
4283 struct task_struct *curr = rq->curr;
4287 spin_lock(&rq->lock);
4288 update_rq_clock(rq);
4289 update_cpu_load(rq);
4290 curr->sched_class->task_tick(rq, curr, 0);
4291 spin_unlock(&rq->lock);
4294 rq->idle_at_tick = idle_cpu(cpu);
4295 trigger_load_balance(rq, cpu);
4299 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4300 defined(CONFIG_PREEMPT_TRACER))
4302 static inline unsigned long get_parent_ip(unsigned long addr)
4304 if (in_lock_functions(addr)) {
4305 addr = CALLER_ADDR2;
4306 if (in_lock_functions(addr))
4307 addr = CALLER_ADDR3;
4312 void __kprobes add_preempt_count(int val)
4314 #ifdef CONFIG_DEBUG_PREEMPT
4318 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4321 preempt_count() += val;
4322 #ifdef CONFIG_DEBUG_PREEMPT
4324 * Spinlock count overflowing soon?
4326 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4329 if (preempt_count() == val)
4330 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4332 EXPORT_SYMBOL(add_preempt_count);
4334 void __kprobes sub_preempt_count(int val)
4336 #ifdef CONFIG_DEBUG_PREEMPT
4340 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4343 * Is the spinlock portion underflowing?
4345 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4346 !(preempt_count() & PREEMPT_MASK)))
4350 if (preempt_count() == val)
4351 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4352 preempt_count() -= val;
4354 EXPORT_SYMBOL(sub_preempt_count);
4359 * Print scheduling while atomic bug:
4361 static noinline void __schedule_bug(struct task_struct *prev)
4363 struct pt_regs *regs = get_irq_regs();
4365 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4366 prev->comm, prev->pid, preempt_count());
4368 debug_show_held_locks(prev);
4370 if (irqs_disabled())
4371 print_irqtrace_events(prev);
4380 * Various schedule()-time debugging checks and statistics:
4382 static inline void schedule_debug(struct task_struct *prev)
4385 * Test if we are atomic. Since do_exit() needs to call into
4386 * schedule() atomically, we ignore that path for now.
4387 * Otherwise, whine if we are scheduling when we should not be.
4389 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4390 __schedule_bug(prev);
4392 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4394 schedstat_inc(this_rq(), sched_count);
4395 #ifdef CONFIG_SCHEDSTATS
4396 if (unlikely(prev->lock_depth >= 0)) {
4397 schedstat_inc(this_rq(), bkl_count);
4398 schedstat_inc(prev, sched_info.bkl_count);
4404 * Pick up the highest-prio task:
4406 static inline struct task_struct *
4407 pick_next_task(struct rq *rq, struct task_struct *prev)
4409 const struct sched_class *class;
4410 struct task_struct *p;
4413 * Optimization: we know that if all tasks are in
4414 * the fair class we can call that function directly:
4416 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4417 p = fair_sched_class.pick_next_task(rq);
4422 class = sched_class_highest;
4424 p = class->pick_next_task(rq);
4428 * Will never be NULL as the idle class always
4429 * returns a non-NULL p:
4431 class = class->next;
4436 * schedule() is the main scheduler function.
4438 asmlinkage void __sched schedule(void)
4440 struct task_struct *prev, *next;
4441 unsigned long *switch_count;
4447 cpu = smp_processor_id();
4451 switch_count = &prev->nivcsw;
4453 release_kernel_lock(prev);
4454 need_resched_nonpreemptible:
4456 schedule_debug(prev);
4458 if (sched_feat(HRTICK))
4461 spin_lock_irq(&rq->lock);
4462 update_rq_clock(rq);
4463 clear_tsk_need_resched(prev);
4465 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4466 if (unlikely(signal_pending_state(prev->state, prev)))
4467 prev->state = TASK_RUNNING;
4469 deactivate_task(rq, prev, 1);
4470 switch_count = &prev->nvcsw;
4474 if (prev->sched_class->pre_schedule)
4475 prev->sched_class->pre_schedule(rq, prev);
4478 if (unlikely(!rq->nr_running))
4479 idle_balance(cpu, rq);
4481 prev->sched_class->put_prev_task(rq, prev);
4482 next = pick_next_task(rq, prev);
4484 if (likely(prev != next)) {
4485 sched_info_switch(prev, next);
4491 context_switch(rq, prev, next); /* unlocks the rq */
4493 * the context switch might have flipped the stack from under
4494 * us, hence refresh the local variables.
4496 cpu = smp_processor_id();
4499 spin_unlock_irq(&rq->lock);
4501 if (unlikely(reacquire_kernel_lock(current) < 0))
4502 goto need_resched_nonpreemptible;
4504 preempt_enable_no_resched();
4505 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4508 EXPORT_SYMBOL(schedule);
4510 #ifdef CONFIG_PREEMPT
4512 * this is the entry point to schedule() from in-kernel preemption
4513 * off of preempt_enable. Kernel preemptions off return from interrupt
4514 * occur there and call schedule directly.
4516 asmlinkage void __sched preempt_schedule(void)
4518 struct thread_info *ti = current_thread_info();
4521 * If there is a non-zero preempt_count or interrupts are disabled,
4522 * we do not want to preempt the current task. Just return..
4524 if (likely(ti->preempt_count || irqs_disabled()))
4528 add_preempt_count(PREEMPT_ACTIVE);
4530 sub_preempt_count(PREEMPT_ACTIVE);
4533 * Check again in case we missed a preemption opportunity
4534 * between schedule and now.
4537 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4539 EXPORT_SYMBOL(preempt_schedule);
4542 * this is the entry point to schedule() from kernel preemption
4543 * off of irq context.
4544 * Note, that this is called and return with irqs disabled. This will
4545 * protect us against recursive calling from irq.
4547 asmlinkage void __sched preempt_schedule_irq(void)
4549 struct thread_info *ti = current_thread_info();
4551 /* Catch callers which need to be fixed */
4552 BUG_ON(ti->preempt_count || !irqs_disabled());
4555 add_preempt_count(PREEMPT_ACTIVE);
4558 local_irq_disable();
4559 sub_preempt_count(PREEMPT_ACTIVE);
4562 * Check again in case we missed a preemption opportunity
4563 * between schedule and now.
4566 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4569 #endif /* CONFIG_PREEMPT */
4571 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4574 return try_to_wake_up(curr->private, mode, sync);
4576 EXPORT_SYMBOL(default_wake_function);
4579 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4580 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4581 * number) then we wake all the non-exclusive tasks and one exclusive task.
4583 * There are circumstances in which we can try to wake a task which has already
4584 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4585 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4587 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4588 int nr_exclusive, int sync, void *key)
4590 wait_queue_t *curr, *next;
4592 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4593 unsigned flags = curr->flags;
4595 if (curr->func(curr, mode, sync, key) &&
4596 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4602 * __wake_up - wake up threads blocked on a waitqueue.
4604 * @mode: which threads
4605 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4606 * @key: is directly passed to the wakeup function
4608 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4609 int nr_exclusive, void *key)
4611 unsigned long flags;
4613 spin_lock_irqsave(&q->lock, flags);
4614 __wake_up_common(q, mode, nr_exclusive, 0, key);
4615 spin_unlock_irqrestore(&q->lock, flags);
4617 EXPORT_SYMBOL(__wake_up);
4620 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4622 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4624 __wake_up_common(q, mode, 1, 0, NULL);
4628 * __wake_up_sync - wake up threads blocked on a waitqueue.
4630 * @mode: which threads
4631 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4633 * The sync wakeup differs that the waker knows that it will schedule
4634 * away soon, so while the target thread will be woken up, it will not
4635 * be migrated to another CPU - ie. the two threads are 'synchronized'
4636 * with each other. This can prevent needless bouncing between CPUs.
4638 * On UP it can prevent extra preemption.
4641 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4643 unsigned long flags;
4649 if (unlikely(!nr_exclusive))
4652 spin_lock_irqsave(&q->lock, flags);
4653 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4654 spin_unlock_irqrestore(&q->lock, flags);
4656 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4659 * complete: - signals a single thread waiting on this completion
4660 * @x: holds the state of this particular completion
4662 * This will wake up a single thread waiting on this completion. Threads will be
4663 * awakened in the same order in which they were queued.
4665 * See also complete_all(), wait_for_completion() and related routines.
4667 void complete(struct completion *x)
4669 unsigned long flags;
4671 spin_lock_irqsave(&x->wait.lock, flags);
4673 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4674 spin_unlock_irqrestore(&x->wait.lock, flags);
4676 EXPORT_SYMBOL(complete);
4679 * complete_all: - signals all threads waiting on this completion
4680 * @x: holds the state of this particular completion
4682 * This will wake up all threads waiting on this particular completion event.
4684 void complete_all(struct completion *x)
4686 unsigned long flags;
4688 spin_lock_irqsave(&x->wait.lock, flags);
4689 x->done += UINT_MAX/2;
4690 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4691 spin_unlock_irqrestore(&x->wait.lock, flags);
4693 EXPORT_SYMBOL(complete_all);
4695 static inline long __sched
4696 do_wait_for_common(struct completion *x, long timeout, int state)
4699 DECLARE_WAITQUEUE(wait, current);
4701 wait.flags |= WQ_FLAG_EXCLUSIVE;
4702 __add_wait_queue_tail(&x->wait, &wait);
4704 if (signal_pending_state(state, current)) {
4705 timeout = -ERESTARTSYS;
4708 __set_current_state(state);
4709 spin_unlock_irq(&x->wait.lock);
4710 timeout = schedule_timeout(timeout);
4711 spin_lock_irq(&x->wait.lock);
4712 } while (!x->done && timeout);
4713 __remove_wait_queue(&x->wait, &wait);
4718 return timeout ?: 1;
4722 wait_for_common(struct completion *x, long timeout, int state)
4726 spin_lock_irq(&x->wait.lock);
4727 timeout = do_wait_for_common(x, timeout, state);
4728 spin_unlock_irq(&x->wait.lock);
4733 * wait_for_completion: - waits for completion of a task
4734 * @x: holds the state of this particular completion
4736 * This waits to be signaled for completion of a specific task. It is NOT
4737 * interruptible and there is no timeout.
4739 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4740 * and interrupt capability. Also see complete().
4742 void __sched wait_for_completion(struct completion *x)
4744 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4746 EXPORT_SYMBOL(wait_for_completion);
4749 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4750 * @x: holds the state of this particular completion
4751 * @timeout: timeout value in jiffies
4753 * This waits for either a completion of a specific task to be signaled or for a
4754 * specified timeout to expire. The timeout is in jiffies. It is not
4757 unsigned long __sched
4758 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4760 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4762 EXPORT_SYMBOL(wait_for_completion_timeout);
4765 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4766 * @x: holds the state of this particular completion
4768 * This waits for completion of a specific task to be signaled. It is
4771 int __sched wait_for_completion_interruptible(struct completion *x)
4773 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4774 if (t == -ERESTARTSYS)
4778 EXPORT_SYMBOL(wait_for_completion_interruptible);
4781 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4782 * @x: holds the state of this particular completion
4783 * @timeout: timeout value in jiffies
4785 * This waits for either a completion of a specific task to be signaled or for a
4786 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4788 unsigned long __sched
4789 wait_for_completion_interruptible_timeout(struct completion *x,
4790 unsigned long timeout)
4792 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4794 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4797 * wait_for_completion_killable: - waits for completion of a task (killable)
4798 * @x: holds the state of this particular completion
4800 * This waits to be signaled for completion of a specific task. It can be
4801 * interrupted by a kill signal.
4803 int __sched wait_for_completion_killable(struct completion *x)
4805 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4806 if (t == -ERESTARTSYS)
4810 EXPORT_SYMBOL(wait_for_completion_killable);
4813 * try_wait_for_completion - try to decrement a completion without blocking
4814 * @x: completion structure
4816 * Returns: 0 if a decrement cannot be done without blocking
4817 * 1 if a decrement succeeded.
4819 * If a completion is being used as a counting completion,
4820 * attempt to decrement the counter without blocking. This
4821 * enables us to avoid waiting if the resource the completion
4822 * is protecting is not available.
4824 bool try_wait_for_completion(struct completion *x)
4828 spin_lock_irq(&x->wait.lock);
4833 spin_unlock_irq(&x->wait.lock);
4836 EXPORT_SYMBOL(try_wait_for_completion);
4839 * completion_done - Test to see if a completion has any waiters
4840 * @x: completion structure
4842 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4843 * 1 if there are no waiters.
4846 bool completion_done(struct completion *x)
4850 spin_lock_irq(&x->wait.lock);
4853 spin_unlock_irq(&x->wait.lock);
4856 EXPORT_SYMBOL(completion_done);
4859 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4861 unsigned long flags;
4864 init_waitqueue_entry(&wait, current);
4866 __set_current_state(state);
4868 spin_lock_irqsave(&q->lock, flags);
4869 __add_wait_queue(q, &wait);
4870 spin_unlock(&q->lock);
4871 timeout = schedule_timeout(timeout);
4872 spin_lock_irq(&q->lock);
4873 __remove_wait_queue(q, &wait);
4874 spin_unlock_irqrestore(&q->lock, flags);
4879 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4881 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4883 EXPORT_SYMBOL(interruptible_sleep_on);
4886 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4888 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4890 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4892 void __sched sleep_on(wait_queue_head_t *q)
4894 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4896 EXPORT_SYMBOL(sleep_on);
4898 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4900 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4902 EXPORT_SYMBOL(sleep_on_timeout);
4904 #ifdef CONFIG_RT_MUTEXES
4907 * rt_mutex_setprio - set the current priority of a task
4909 * @prio: prio value (kernel-internal form)
4911 * This function changes the 'effective' priority of a task. It does
4912 * not touch ->normal_prio like __setscheduler().
4914 * Used by the rt_mutex code to implement priority inheritance logic.
4916 void rt_mutex_setprio(struct task_struct *p, int prio)
4918 unsigned long flags;
4919 int oldprio, on_rq, running;
4921 const struct sched_class *prev_class = p->sched_class;
4923 BUG_ON(prio < 0 || prio > MAX_PRIO);
4925 rq = task_rq_lock(p, &flags);
4926 update_rq_clock(rq);
4929 on_rq = p->se.on_rq;
4930 running = task_current(rq, p);
4932 dequeue_task(rq, p, 0);
4934 p->sched_class->put_prev_task(rq, p);
4937 p->sched_class = &rt_sched_class;
4939 p->sched_class = &fair_sched_class;
4944 p->sched_class->set_curr_task(rq);
4946 enqueue_task(rq, p, 0);
4948 check_class_changed(rq, p, prev_class, oldprio, running);
4950 task_rq_unlock(rq, &flags);
4955 void set_user_nice(struct task_struct *p, long nice)
4957 int old_prio, delta, on_rq;
4958 unsigned long flags;
4961 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4964 * We have to be careful, if called from sys_setpriority(),
4965 * the task might be in the middle of scheduling on another CPU.
4967 rq = task_rq_lock(p, &flags);
4968 update_rq_clock(rq);
4970 * The RT priorities are set via sched_setscheduler(), but we still
4971 * allow the 'normal' nice value to be set - but as expected
4972 * it wont have any effect on scheduling until the task is
4973 * SCHED_FIFO/SCHED_RR:
4975 if (task_has_rt_policy(p)) {
4976 p->static_prio = NICE_TO_PRIO(nice);
4979 on_rq = p->se.on_rq;
4981 dequeue_task(rq, p, 0);
4983 p->static_prio = NICE_TO_PRIO(nice);
4986 p->prio = effective_prio(p);
4987 delta = p->prio - old_prio;
4990 enqueue_task(rq, p, 0);
4992 * If the task increased its priority or is running and
4993 * lowered its priority, then reschedule its CPU:
4995 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4996 resched_task(rq->curr);
4999 task_rq_unlock(rq, &flags);
5001 EXPORT_SYMBOL(set_user_nice);
5004 * can_nice - check if a task can reduce its nice value
5008 int can_nice(const struct task_struct *p, const int nice)
5010 /* convert nice value [19,-20] to rlimit style value [1,40] */
5011 int nice_rlim = 20 - nice;
5013 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5014 capable(CAP_SYS_NICE));
5017 #ifdef __ARCH_WANT_SYS_NICE
5020 * sys_nice - change the priority of the current process.
5021 * @increment: priority increment
5023 * sys_setpriority is a more generic, but much slower function that
5024 * does similar things.
5026 asmlinkage long sys_nice(int increment)
5031 * Setpriority might change our priority at the same moment.
5032 * We don't have to worry. Conceptually one call occurs first
5033 * and we have a single winner.
5035 if (increment < -40)
5040 nice = PRIO_TO_NICE(current->static_prio) + increment;
5046 if (increment < 0 && !can_nice(current, nice))
5049 retval = security_task_setnice(current, nice);
5053 set_user_nice(current, nice);
5060 * task_prio - return the priority value of a given task.
5061 * @p: the task in question.
5063 * This is the priority value as seen by users in /proc.
5064 * RT tasks are offset by -200. Normal tasks are centered
5065 * around 0, value goes from -16 to +15.
5067 int task_prio(const struct task_struct *p)
5069 return p->prio - MAX_RT_PRIO;
5073 * task_nice - return the nice value of a given task.
5074 * @p: the task in question.
5076 int task_nice(const struct task_struct *p)
5078 return TASK_NICE(p);
5080 EXPORT_SYMBOL(task_nice);
5083 * idle_cpu - is a given cpu idle currently?
5084 * @cpu: the processor in question.
5086 int idle_cpu(int cpu)
5088 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5092 * idle_task - return the idle task for a given cpu.
5093 * @cpu: the processor in question.
5095 struct task_struct *idle_task(int cpu)
5097 return cpu_rq(cpu)->idle;
5101 * find_process_by_pid - find a process with a matching PID value.
5102 * @pid: the pid in question.
5104 static struct task_struct *find_process_by_pid(pid_t pid)
5106 return pid ? find_task_by_vpid(pid) : current;
5109 /* Actually do priority change: must hold rq lock. */
5111 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5113 BUG_ON(p->se.on_rq);
5116 switch (p->policy) {
5120 p->sched_class = &fair_sched_class;
5124 p->sched_class = &rt_sched_class;
5128 p->rt_priority = prio;
5129 p->normal_prio = normal_prio(p);
5130 /* we are holding p->pi_lock already */
5131 p->prio = rt_mutex_getprio(p);
5135 static int __sched_setscheduler(struct task_struct *p, int policy,
5136 struct sched_param *param, bool user)
5138 int retval, oldprio, oldpolicy = -1, on_rq, running;
5139 unsigned long flags;
5140 const struct sched_class *prev_class = p->sched_class;
5143 /* may grab non-irq protected spin_locks */
5144 BUG_ON(in_interrupt());
5146 /* double check policy once rq lock held */
5148 policy = oldpolicy = p->policy;
5149 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5150 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5151 policy != SCHED_IDLE)
5154 * Valid priorities for SCHED_FIFO and SCHED_RR are
5155 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5156 * SCHED_BATCH and SCHED_IDLE is 0.
5158 if (param->sched_priority < 0 ||
5159 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5160 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5162 if (rt_policy(policy) != (param->sched_priority != 0))
5166 * Allow unprivileged RT tasks to decrease priority:
5168 if (user && !capable(CAP_SYS_NICE)) {
5169 if (rt_policy(policy)) {
5170 unsigned long rlim_rtprio;
5172 if (!lock_task_sighand(p, &flags))
5174 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5175 unlock_task_sighand(p, &flags);
5177 /* can't set/change the rt policy */
5178 if (policy != p->policy && !rlim_rtprio)
5181 /* can't increase priority */
5182 if (param->sched_priority > p->rt_priority &&
5183 param->sched_priority > rlim_rtprio)
5187 * Like positive nice levels, dont allow tasks to
5188 * move out of SCHED_IDLE either:
5190 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5193 /* can't change other user's priorities */
5194 if ((current->euid != p->euid) &&
5195 (current->euid != p->uid))
5200 #ifdef CONFIG_RT_GROUP_SCHED
5202 * Do not allow realtime tasks into groups that have no runtime
5205 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5206 task_group(p)->rt_bandwidth.rt_runtime == 0)
5210 retval = security_task_setscheduler(p, policy, param);
5216 * make sure no PI-waiters arrive (or leave) while we are
5217 * changing the priority of the task:
5219 spin_lock_irqsave(&p->pi_lock, flags);
5221 * To be able to change p->policy safely, the apropriate
5222 * runqueue lock must be held.
5224 rq = __task_rq_lock(p);
5225 /* recheck policy now with rq lock held */
5226 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5227 policy = oldpolicy = -1;
5228 __task_rq_unlock(rq);
5229 spin_unlock_irqrestore(&p->pi_lock, flags);
5232 update_rq_clock(rq);
5233 on_rq = p->se.on_rq;
5234 running = task_current(rq, p);
5236 deactivate_task(rq, p, 0);
5238 p->sched_class->put_prev_task(rq, p);
5241 __setscheduler(rq, p, policy, param->sched_priority);
5244 p->sched_class->set_curr_task(rq);
5246 activate_task(rq, p, 0);
5248 check_class_changed(rq, p, prev_class, oldprio, running);
5250 __task_rq_unlock(rq);
5251 spin_unlock_irqrestore(&p->pi_lock, flags);
5253 rt_mutex_adjust_pi(p);
5259 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5260 * @p: the task in question.
5261 * @policy: new policy.
5262 * @param: structure containing the new RT priority.
5264 * NOTE that the task may be already dead.
5266 int sched_setscheduler(struct task_struct *p, int policy,
5267 struct sched_param *param)
5269 return __sched_setscheduler(p, policy, param, true);
5271 EXPORT_SYMBOL_GPL(sched_setscheduler);
5274 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5275 * @p: the task in question.
5276 * @policy: new policy.
5277 * @param: structure containing the new RT priority.
5279 * Just like sched_setscheduler, only don't bother checking if the
5280 * current context has permission. For example, this is needed in
5281 * stop_machine(): we create temporary high priority worker threads,
5282 * but our caller might not have that capability.
5284 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5285 struct sched_param *param)
5287 return __sched_setscheduler(p, policy, param, false);
5291 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5293 struct sched_param lparam;
5294 struct task_struct *p;
5297 if (!param || pid < 0)
5299 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5304 p = find_process_by_pid(pid);
5306 retval = sched_setscheduler(p, policy, &lparam);
5313 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5314 * @pid: the pid in question.
5315 * @policy: new policy.
5316 * @param: structure containing the new RT priority.
5319 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5321 /* negative values for policy are not valid */
5325 return do_sched_setscheduler(pid, policy, param);
5329 * sys_sched_setparam - set/change the RT priority of a thread
5330 * @pid: the pid in question.
5331 * @param: structure containing the new RT priority.
5333 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5335 return do_sched_setscheduler(pid, -1, param);
5339 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5340 * @pid: the pid in question.
5342 asmlinkage long sys_sched_getscheduler(pid_t pid)
5344 struct task_struct *p;
5351 read_lock(&tasklist_lock);
5352 p = find_process_by_pid(pid);
5354 retval = security_task_getscheduler(p);
5358 read_unlock(&tasklist_lock);
5363 * sys_sched_getscheduler - get the RT priority of a thread
5364 * @pid: the pid in question.
5365 * @param: structure containing the RT priority.
5367 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5369 struct sched_param lp;
5370 struct task_struct *p;
5373 if (!param || pid < 0)
5376 read_lock(&tasklist_lock);
5377 p = find_process_by_pid(pid);
5382 retval = security_task_getscheduler(p);
5386 lp.sched_priority = p->rt_priority;
5387 read_unlock(&tasklist_lock);
5390 * This one might sleep, we cannot do it with a spinlock held ...
5392 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5397 read_unlock(&tasklist_lock);
5401 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5403 cpumask_t cpus_allowed;
5404 cpumask_t new_mask = *in_mask;
5405 struct task_struct *p;
5409 read_lock(&tasklist_lock);
5411 p = find_process_by_pid(pid);
5413 read_unlock(&tasklist_lock);
5419 * It is not safe to call set_cpus_allowed with the
5420 * tasklist_lock held. We will bump the task_struct's
5421 * usage count and then drop tasklist_lock.
5424 read_unlock(&tasklist_lock);
5427 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5428 !capable(CAP_SYS_NICE))
5431 retval = security_task_setscheduler(p, 0, NULL);
5435 cpuset_cpus_allowed(p, &cpus_allowed);
5436 cpus_and(new_mask, new_mask, cpus_allowed);
5438 retval = set_cpus_allowed_ptr(p, &new_mask);
5441 cpuset_cpus_allowed(p, &cpus_allowed);
5442 if (!cpus_subset(new_mask, cpus_allowed)) {
5444 * We must have raced with a concurrent cpuset
5445 * update. Just reset the cpus_allowed to the
5446 * cpuset's cpus_allowed
5448 new_mask = cpus_allowed;
5458 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5459 cpumask_t *new_mask)
5461 if (len < sizeof(cpumask_t)) {
5462 memset(new_mask, 0, sizeof(cpumask_t));
5463 } else if (len > sizeof(cpumask_t)) {
5464 len = sizeof(cpumask_t);
5466 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5470 * sys_sched_setaffinity - set the cpu affinity of a process
5471 * @pid: pid of the process
5472 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5473 * @user_mask_ptr: user-space pointer to the new cpu mask
5475 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5476 unsigned long __user *user_mask_ptr)
5481 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5485 return sched_setaffinity(pid, &new_mask);
5488 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5490 struct task_struct *p;
5494 read_lock(&tasklist_lock);
5497 p = find_process_by_pid(pid);
5501 retval = security_task_getscheduler(p);
5505 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5508 read_unlock(&tasklist_lock);
5515 * sys_sched_getaffinity - get the cpu affinity of a process
5516 * @pid: pid of the process
5517 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5518 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5520 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5521 unsigned long __user *user_mask_ptr)
5526 if (len < sizeof(cpumask_t))
5529 ret = sched_getaffinity(pid, &mask);
5533 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5536 return sizeof(cpumask_t);
5540 * sys_sched_yield - yield the current processor to other threads.
5542 * This function yields the current CPU to other tasks. If there are no
5543 * other threads running on this CPU then this function will return.
5545 asmlinkage long sys_sched_yield(void)
5547 struct rq *rq = this_rq_lock();
5549 schedstat_inc(rq, yld_count);
5550 current->sched_class->yield_task(rq);
5553 * Since we are going to call schedule() anyway, there's
5554 * no need to preempt or enable interrupts:
5556 __release(rq->lock);
5557 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5558 _raw_spin_unlock(&rq->lock);
5559 preempt_enable_no_resched();
5566 static void __cond_resched(void)
5568 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5569 __might_sleep(__FILE__, __LINE__);
5572 * The BKS might be reacquired before we have dropped
5573 * PREEMPT_ACTIVE, which could trigger a second
5574 * cond_resched() call.
5577 add_preempt_count(PREEMPT_ACTIVE);
5579 sub_preempt_count(PREEMPT_ACTIVE);
5580 } while (need_resched());
5583 int __sched _cond_resched(void)
5585 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5586 system_state == SYSTEM_RUNNING) {
5592 EXPORT_SYMBOL(_cond_resched);
5595 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5596 * call schedule, and on return reacquire the lock.
5598 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5599 * operations here to prevent schedule() from being called twice (once via
5600 * spin_unlock(), once by hand).
5602 int cond_resched_lock(spinlock_t *lock)
5604 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5607 if (spin_needbreak(lock) || resched) {
5609 if (resched && need_resched())
5618 EXPORT_SYMBOL(cond_resched_lock);
5620 int __sched cond_resched_softirq(void)
5622 BUG_ON(!in_softirq());
5624 if (need_resched() && system_state == SYSTEM_RUNNING) {
5632 EXPORT_SYMBOL(cond_resched_softirq);
5635 * yield - yield the current processor to other threads.
5637 * This is a shortcut for kernel-space yielding - it marks the
5638 * thread runnable and calls sys_sched_yield().
5640 void __sched yield(void)
5642 set_current_state(TASK_RUNNING);
5645 EXPORT_SYMBOL(yield);
5648 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5649 * that process accounting knows that this is a task in IO wait state.
5651 * But don't do that if it is a deliberate, throttling IO wait (this task
5652 * has set its backing_dev_info: the queue against which it should throttle)
5654 void __sched io_schedule(void)
5656 struct rq *rq = &__raw_get_cpu_var(runqueues);
5658 delayacct_blkio_start();
5659 atomic_inc(&rq->nr_iowait);
5661 atomic_dec(&rq->nr_iowait);
5662 delayacct_blkio_end();
5664 EXPORT_SYMBOL(io_schedule);
5666 long __sched io_schedule_timeout(long timeout)
5668 struct rq *rq = &__raw_get_cpu_var(runqueues);
5671 delayacct_blkio_start();
5672 atomic_inc(&rq->nr_iowait);
5673 ret = schedule_timeout(timeout);
5674 atomic_dec(&rq->nr_iowait);
5675 delayacct_blkio_end();
5680 * sys_sched_get_priority_max - return maximum RT priority.
5681 * @policy: scheduling class.
5683 * this syscall returns the maximum rt_priority that can be used
5684 * by a given scheduling class.
5686 asmlinkage long sys_sched_get_priority_max(int policy)
5693 ret = MAX_USER_RT_PRIO-1;
5705 * sys_sched_get_priority_min - return minimum RT priority.
5706 * @policy: scheduling class.
5708 * this syscall returns the minimum rt_priority that can be used
5709 * by a given scheduling class.
5711 asmlinkage long sys_sched_get_priority_min(int policy)
5729 * sys_sched_rr_get_interval - return the default timeslice of a process.
5730 * @pid: pid of the process.
5731 * @interval: userspace pointer to the timeslice value.
5733 * this syscall writes the default timeslice value of a given process
5734 * into the user-space timespec buffer. A value of '0' means infinity.
5737 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5739 struct task_struct *p;
5740 unsigned int time_slice;
5748 read_lock(&tasklist_lock);
5749 p = find_process_by_pid(pid);
5753 retval = security_task_getscheduler(p);
5758 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5759 * tasks that are on an otherwise idle runqueue:
5762 if (p->policy == SCHED_RR) {
5763 time_slice = DEF_TIMESLICE;
5764 } else if (p->policy != SCHED_FIFO) {
5765 struct sched_entity *se = &p->se;
5766 unsigned long flags;
5769 rq = task_rq_lock(p, &flags);
5770 if (rq->cfs.load.weight)
5771 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5772 task_rq_unlock(rq, &flags);
5774 read_unlock(&tasklist_lock);
5775 jiffies_to_timespec(time_slice, &t);
5776 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5780 read_unlock(&tasklist_lock);
5784 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5786 void sched_show_task(struct task_struct *p)
5788 unsigned long free = 0;
5791 state = p->state ? __ffs(p->state) + 1 : 0;
5792 printk(KERN_INFO "%-13.13s %c", p->comm,
5793 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5794 #if BITS_PER_LONG == 32
5795 if (state == TASK_RUNNING)
5796 printk(KERN_CONT " running ");
5798 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5800 if (state == TASK_RUNNING)
5801 printk(KERN_CONT " running task ");
5803 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5805 #ifdef CONFIG_DEBUG_STACK_USAGE
5807 unsigned long *n = end_of_stack(p);
5810 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5813 printk(KERN_CONT "%5lu %5d %6d\n", free,
5814 task_pid_nr(p), task_pid_nr(p->real_parent));
5816 show_stack(p, NULL);
5819 void show_state_filter(unsigned long state_filter)
5821 struct task_struct *g, *p;
5823 #if BITS_PER_LONG == 32
5825 " task PC stack pid father\n");
5828 " task PC stack pid father\n");
5830 read_lock(&tasklist_lock);
5831 do_each_thread(g, p) {
5833 * reset the NMI-timeout, listing all files on a slow
5834 * console might take alot of time:
5836 touch_nmi_watchdog();
5837 if (!state_filter || (p->state & state_filter))
5839 } while_each_thread(g, p);
5841 touch_all_softlockup_watchdogs();
5843 #ifdef CONFIG_SCHED_DEBUG
5844 sysrq_sched_debug_show();
5846 read_unlock(&tasklist_lock);
5848 * Only show locks if all tasks are dumped:
5850 if (state_filter == -1)
5851 debug_show_all_locks();
5854 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5856 idle->sched_class = &idle_sched_class;
5860 * init_idle - set up an idle thread for a given CPU
5861 * @idle: task in question
5862 * @cpu: cpu the idle task belongs to
5864 * NOTE: this function does not set the idle thread's NEED_RESCHED
5865 * flag, to make booting more robust.
5867 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5869 struct rq *rq = cpu_rq(cpu);
5870 unsigned long flags;
5872 spin_lock_irqsave(&rq->lock, flags);
5875 idle->se.exec_start = sched_clock();
5877 idle->prio = idle->normal_prio = MAX_PRIO;
5878 idle->cpus_allowed = cpumask_of_cpu(cpu);
5879 __set_task_cpu(idle, cpu);
5881 rq->curr = rq->idle = idle;
5882 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5885 spin_unlock_irqrestore(&rq->lock, flags);
5887 /* Set the preempt count _outside_ the spinlocks! */
5888 #if defined(CONFIG_PREEMPT)
5889 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5891 task_thread_info(idle)->preempt_count = 0;
5894 * The idle tasks have their own, simple scheduling class:
5896 idle->sched_class = &idle_sched_class;
5900 * In a system that switches off the HZ timer nohz_cpu_mask
5901 * indicates which cpus entered this state. This is used
5902 * in the rcu update to wait only for active cpus. For system
5903 * which do not switch off the HZ timer nohz_cpu_mask should
5904 * always be CPU_MASK_NONE.
5906 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5909 * Increase the granularity value when there are more CPUs,
5910 * because with more CPUs the 'effective latency' as visible
5911 * to users decreases. But the relationship is not linear,
5912 * so pick a second-best guess by going with the log2 of the
5915 * This idea comes from the SD scheduler of Con Kolivas:
5917 static inline void sched_init_granularity(void)
5919 unsigned int factor = 1 + ilog2(num_online_cpus());
5920 const unsigned long limit = 200000000;
5922 sysctl_sched_min_granularity *= factor;
5923 if (sysctl_sched_min_granularity > limit)
5924 sysctl_sched_min_granularity = limit;
5926 sysctl_sched_latency *= factor;
5927 if (sysctl_sched_latency > limit)
5928 sysctl_sched_latency = limit;
5930 sysctl_sched_wakeup_granularity *= factor;
5932 sysctl_sched_shares_ratelimit *= factor;
5937 * This is how migration works:
5939 * 1) we queue a struct migration_req structure in the source CPU's
5940 * runqueue and wake up that CPU's migration thread.
5941 * 2) we down() the locked semaphore => thread blocks.
5942 * 3) migration thread wakes up (implicitly it forces the migrated
5943 * thread off the CPU)
5944 * 4) it gets the migration request and checks whether the migrated
5945 * task is still in the wrong runqueue.
5946 * 5) if it's in the wrong runqueue then the migration thread removes
5947 * it and puts it into the right queue.
5948 * 6) migration thread up()s the semaphore.
5949 * 7) we wake up and the migration is done.
5953 * Change a given task's CPU affinity. Migrate the thread to a
5954 * proper CPU and schedule it away if the CPU it's executing on
5955 * is removed from the allowed bitmask.
5957 * NOTE: the caller must have a valid reference to the task, the
5958 * task must not exit() & deallocate itself prematurely. The
5959 * call is not atomic; no spinlocks may be held.
5961 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5963 struct migration_req req;
5964 unsigned long flags;
5968 rq = task_rq_lock(p, &flags);
5969 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5974 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5975 !cpus_equal(p->cpus_allowed, *new_mask))) {
5980 if (p->sched_class->set_cpus_allowed)
5981 p->sched_class->set_cpus_allowed(p, new_mask);
5983 p->cpus_allowed = *new_mask;
5984 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5987 /* Can the task run on the task's current CPU? If so, we're done */
5988 if (cpu_isset(task_cpu(p), *new_mask))
5991 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5992 /* Need help from migration thread: drop lock and wait. */
5993 task_rq_unlock(rq, &flags);
5994 wake_up_process(rq->migration_thread);
5995 wait_for_completion(&req.done);
5996 tlb_migrate_finish(p->mm);
6000 task_rq_unlock(rq, &flags);
6004 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6007 * Move (not current) task off this cpu, onto dest cpu. We're doing
6008 * this because either it can't run here any more (set_cpus_allowed()
6009 * away from this CPU, or CPU going down), or because we're
6010 * attempting to rebalance this task on exec (sched_exec).
6012 * So we race with normal scheduler movements, but that's OK, as long
6013 * as the task is no longer on this CPU.
6015 * Returns non-zero if task was successfully migrated.
6017 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6019 struct rq *rq_dest, *rq_src;
6022 if (unlikely(!cpu_active(dest_cpu)))
6025 rq_src = cpu_rq(src_cpu);
6026 rq_dest = cpu_rq(dest_cpu);
6028 double_rq_lock(rq_src, rq_dest);
6029 /* Already moved. */
6030 if (task_cpu(p) != src_cpu)
6032 /* Affinity changed (again). */
6033 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6036 on_rq = p->se.on_rq;
6038 deactivate_task(rq_src, p, 0);
6040 set_task_cpu(p, dest_cpu);
6042 activate_task(rq_dest, p, 0);
6043 check_preempt_curr(rq_dest, p, 0);
6048 double_rq_unlock(rq_src, rq_dest);
6053 * migration_thread - this is a highprio system thread that performs
6054 * thread migration by bumping thread off CPU then 'pushing' onto
6057 static int migration_thread(void *data)
6059 int cpu = (long)data;
6063 BUG_ON(rq->migration_thread != current);
6065 set_current_state(TASK_INTERRUPTIBLE);
6066 while (!kthread_should_stop()) {
6067 struct migration_req *req;
6068 struct list_head *head;
6070 spin_lock_irq(&rq->lock);
6072 if (cpu_is_offline(cpu)) {
6073 spin_unlock_irq(&rq->lock);
6077 if (rq->active_balance) {
6078 active_load_balance(rq, cpu);
6079 rq->active_balance = 0;
6082 head = &rq->migration_queue;
6084 if (list_empty(head)) {
6085 spin_unlock_irq(&rq->lock);
6087 set_current_state(TASK_INTERRUPTIBLE);
6090 req = list_entry(head->next, struct migration_req, list);
6091 list_del_init(head->next);
6093 spin_unlock(&rq->lock);
6094 __migrate_task(req->task, cpu, req->dest_cpu);
6097 complete(&req->done);
6099 __set_current_state(TASK_RUNNING);
6103 /* Wait for kthread_stop */
6104 set_current_state(TASK_INTERRUPTIBLE);
6105 while (!kthread_should_stop()) {
6107 set_current_state(TASK_INTERRUPTIBLE);
6109 __set_current_state(TASK_RUNNING);
6113 #ifdef CONFIG_HOTPLUG_CPU
6115 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6119 local_irq_disable();
6120 ret = __migrate_task(p, src_cpu, dest_cpu);
6126 * Figure out where task on dead CPU should go, use force if necessary.
6127 * NOTE: interrupts should be disabled by the caller
6129 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6131 unsigned long flags;
6138 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6139 cpus_and(mask, mask, p->cpus_allowed);
6140 dest_cpu = any_online_cpu(mask);
6142 /* On any allowed CPU? */
6143 if (dest_cpu >= nr_cpu_ids)
6144 dest_cpu = any_online_cpu(p->cpus_allowed);
6146 /* No more Mr. Nice Guy. */
6147 if (dest_cpu >= nr_cpu_ids) {
6148 cpumask_t cpus_allowed;
6150 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6152 * Try to stay on the same cpuset, where the
6153 * current cpuset may be a subset of all cpus.
6154 * The cpuset_cpus_allowed_locked() variant of
6155 * cpuset_cpus_allowed() will not block. It must be
6156 * called within calls to cpuset_lock/cpuset_unlock.
6158 rq = task_rq_lock(p, &flags);
6159 p->cpus_allowed = cpus_allowed;
6160 dest_cpu = any_online_cpu(p->cpus_allowed);
6161 task_rq_unlock(rq, &flags);
6164 * Don't tell them about moving exiting tasks or
6165 * kernel threads (both mm NULL), since they never
6168 if (p->mm && printk_ratelimit()) {
6169 printk(KERN_INFO "process %d (%s) no "
6170 "longer affine to cpu%d\n",
6171 task_pid_nr(p), p->comm, dead_cpu);
6174 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6178 * While a dead CPU has no uninterruptible tasks queued at this point,
6179 * it might still have a nonzero ->nr_uninterruptible counter, because
6180 * for performance reasons the counter is not stricly tracking tasks to
6181 * their home CPUs. So we just add the counter to another CPU's counter,
6182 * to keep the global sum constant after CPU-down:
6184 static void migrate_nr_uninterruptible(struct rq *rq_src)
6186 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6187 unsigned long flags;
6189 local_irq_save(flags);
6190 double_rq_lock(rq_src, rq_dest);
6191 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6192 rq_src->nr_uninterruptible = 0;
6193 double_rq_unlock(rq_src, rq_dest);
6194 local_irq_restore(flags);
6197 /* Run through task list and migrate tasks from the dead cpu. */
6198 static void migrate_live_tasks(int src_cpu)
6200 struct task_struct *p, *t;
6202 read_lock(&tasklist_lock);
6204 do_each_thread(t, p) {
6208 if (task_cpu(p) == src_cpu)
6209 move_task_off_dead_cpu(src_cpu, p);
6210 } while_each_thread(t, p);
6212 read_unlock(&tasklist_lock);
6216 * Schedules idle task to be the next runnable task on current CPU.
6217 * It does so by boosting its priority to highest possible.
6218 * Used by CPU offline code.
6220 void sched_idle_next(void)
6222 int this_cpu = smp_processor_id();
6223 struct rq *rq = cpu_rq(this_cpu);
6224 struct task_struct *p = rq->idle;
6225 unsigned long flags;
6227 /* cpu has to be offline */
6228 BUG_ON(cpu_online(this_cpu));
6231 * Strictly not necessary since rest of the CPUs are stopped by now
6232 * and interrupts disabled on the current cpu.
6234 spin_lock_irqsave(&rq->lock, flags);
6236 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6238 update_rq_clock(rq);
6239 activate_task(rq, p, 0);
6241 spin_unlock_irqrestore(&rq->lock, flags);
6245 * Ensures that the idle task is using init_mm right before its cpu goes
6248 void idle_task_exit(void)
6250 struct mm_struct *mm = current->active_mm;
6252 BUG_ON(cpu_online(smp_processor_id()));
6255 switch_mm(mm, &init_mm, current);
6259 /* called under rq->lock with disabled interrupts */
6260 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6262 struct rq *rq = cpu_rq(dead_cpu);
6264 /* Must be exiting, otherwise would be on tasklist. */
6265 BUG_ON(!p->exit_state);
6267 /* Cannot have done final schedule yet: would have vanished. */
6268 BUG_ON(p->state == TASK_DEAD);
6273 * Drop lock around migration; if someone else moves it,
6274 * that's OK. No task can be added to this CPU, so iteration is
6277 spin_unlock_irq(&rq->lock);
6278 move_task_off_dead_cpu(dead_cpu, p);
6279 spin_lock_irq(&rq->lock);
6284 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6285 static void migrate_dead_tasks(unsigned int dead_cpu)
6287 struct rq *rq = cpu_rq(dead_cpu);
6288 struct task_struct *next;
6291 if (!rq->nr_running)
6293 update_rq_clock(rq);
6294 next = pick_next_task(rq, rq->curr);
6297 next->sched_class->put_prev_task(rq, next);
6298 migrate_dead(dead_cpu, next);
6302 #endif /* CONFIG_HOTPLUG_CPU */
6304 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6306 static struct ctl_table sd_ctl_dir[] = {
6308 .procname = "sched_domain",
6314 static struct ctl_table sd_ctl_root[] = {
6316 .ctl_name = CTL_KERN,
6317 .procname = "kernel",
6319 .child = sd_ctl_dir,
6324 static struct ctl_table *sd_alloc_ctl_entry(int n)
6326 struct ctl_table *entry =
6327 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6332 static void sd_free_ctl_entry(struct ctl_table **tablep)
6334 struct ctl_table *entry;
6337 * In the intermediate directories, both the child directory and
6338 * procname are dynamically allocated and could fail but the mode
6339 * will always be set. In the lowest directory the names are
6340 * static strings and all have proc handlers.
6342 for (entry = *tablep; entry->mode; entry++) {
6344 sd_free_ctl_entry(&entry->child);
6345 if (entry->proc_handler == NULL)
6346 kfree(entry->procname);
6354 set_table_entry(struct ctl_table *entry,
6355 const char *procname, void *data, int maxlen,
6356 mode_t mode, proc_handler *proc_handler)
6358 entry->procname = procname;
6360 entry->maxlen = maxlen;
6362 entry->proc_handler = proc_handler;
6365 static struct ctl_table *
6366 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6368 struct ctl_table *table = sd_alloc_ctl_entry(13);
6373 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6374 sizeof(long), 0644, proc_doulongvec_minmax);
6375 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6376 sizeof(long), 0644, proc_doulongvec_minmax);
6377 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6378 sizeof(int), 0644, proc_dointvec_minmax);
6379 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6380 sizeof(int), 0644, proc_dointvec_minmax);
6381 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6382 sizeof(int), 0644, proc_dointvec_minmax);
6383 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6384 sizeof(int), 0644, proc_dointvec_minmax);
6385 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6386 sizeof(int), 0644, proc_dointvec_minmax);
6387 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6388 sizeof(int), 0644, proc_dointvec_minmax);
6389 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6390 sizeof(int), 0644, proc_dointvec_minmax);
6391 set_table_entry(&table[9], "cache_nice_tries",
6392 &sd->cache_nice_tries,
6393 sizeof(int), 0644, proc_dointvec_minmax);
6394 set_table_entry(&table[10], "flags", &sd->flags,
6395 sizeof(int), 0644, proc_dointvec_minmax);
6396 set_table_entry(&table[11], "name", sd->name,
6397 CORENAME_MAX_SIZE, 0444, proc_dostring);
6398 /* &table[12] is terminator */
6403 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6405 struct ctl_table *entry, *table;
6406 struct sched_domain *sd;
6407 int domain_num = 0, i;
6410 for_each_domain(cpu, sd)
6412 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6417 for_each_domain(cpu, sd) {
6418 snprintf(buf, 32, "domain%d", i);
6419 entry->procname = kstrdup(buf, GFP_KERNEL);
6421 entry->child = sd_alloc_ctl_domain_table(sd);
6428 static struct ctl_table_header *sd_sysctl_header;
6429 static void register_sched_domain_sysctl(void)
6431 int i, cpu_num = num_online_cpus();
6432 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6435 WARN_ON(sd_ctl_dir[0].child);
6436 sd_ctl_dir[0].child = entry;
6441 for_each_online_cpu(i) {
6442 snprintf(buf, 32, "cpu%d", i);
6443 entry->procname = kstrdup(buf, GFP_KERNEL);
6445 entry->child = sd_alloc_ctl_cpu_table(i);
6449 WARN_ON(sd_sysctl_header);
6450 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6453 /* may be called multiple times per register */
6454 static void unregister_sched_domain_sysctl(void)
6456 if (sd_sysctl_header)
6457 unregister_sysctl_table(sd_sysctl_header);
6458 sd_sysctl_header = NULL;
6459 if (sd_ctl_dir[0].child)
6460 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6463 static void register_sched_domain_sysctl(void)
6466 static void unregister_sched_domain_sysctl(void)
6471 static void set_rq_online(struct rq *rq)
6474 const struct sched_class *class;
6476 cpu_set(rq->cpu, rq->rd->online);
6479 for_each_class(class) {
6480 if (class->rq_online)
6481 class->rq_online(rq);
6486 static void set_rq_offline(struct rq *rq)
6489 const struct sched_class *class;
6491 for_each_class(class) {
6492 if (class->rq_offline)
6493 class->rq_offline(rq);
6496 cpu_clear(rq->cpu, rq->rd->online);
6502 * migration_call - callback that gets triggered when a CPU is added.
6503 * Here we can start up the necessary migration thread for the new CPU.
6505 static int __cpuinit
6506 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6508 struct task_struct *p;
6509 int cpu = (long)hcpu;
6510 unsigned long flags;
6515 case CPU_UP_PREPARE:
6516 case CPU_UP_PREPARE_FROZEN:
6517 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6520 kthread_bind(p, cpu);
6521 /* Must be high prio: stop_machine expects to yield to it. */
6522 rq = task_rq_lock(p, &flags);
6523 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6524 task_rq_unlock(rq, &flags);
6525 cpu_rq(cpu)->migration_thread = p;
6529 case CPU_ONLINE_FROZEN:
6530 /* Strictly unnecessary, as first user will wake it. */
6531 wake_up_process(cpu_rq(cpu)->migration_thread);
6533 /* Update our root-domain */
6535 spin_lock_irqsave(&rq->lock, flags);
6537 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6541 spin_unlock_irqrestore(&rq->lock, flags);
6544 #ifdef CONFIG_HOTPLUG_CPU
6545 case CPU_UP_CANCELED:
6546 case CPU_UP_CANCELED_FROZEN:
6547 if (!cpu_rq(cpu)->migration_thread)
6549 /* Unbind it from offline cpu so it can run. Fall thru. */
6550 kthread_bind(cpu_rq(cpu)->migration_thread,
6551 any_online_cpu(cpu_online_map));
6552 kthread_stop(cpu_rq(cpu)->migration_thread);
6553 cpu_rq(cpu)->migration_thread = NULL;
6557 case CPU_DEAD_FROZEN:
6558 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6559 migrate_live_tasks(cpu);
6561 kthread_stop(rq->migration_thread);
6562 rq->migration_thread = NULL;
6563 /* Idle task back to normal (off runqueue, low prio) */
6564 spin_lock_irq(&rq->lock);
6565 update_rq_clock(rq);
6566 deactivate_task(rq, rq->idle, 0);
6567 rq->idle->static_prio = MAX_PRIO;
6568 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6569 rq->idle->sched_class = &idle_sched_class;
6570 migrate_dead_tasks(cpu);
6571 spin_unlock_irq(&rq->lock);
6573 migrate_nr_uninterruptible(rq);
6574 BUG_ON(rq->nr_running != 0);
6577 * No need to migrate the tasks: it was best-effort if
6578 * they didn't take sched_hotcpu_mutex. Just wake up
6581 spin_lock_irq(&rq->lock);
6582 while (!list_empty(&rq->migration_queue)) {
6583 struct migration_req *req;
6585 req = list_entry(rq->migration_queue.next,
6586 struct migration_req, list);
6587 list_del_init(&req->list);
6588 spin_unlock_irq(&rq->lock);
6589 complete(&req->done);
6590 spin_lock_irq(&rq->lock);
6592 spin_unlock_irq(&rq->lock);
6596 case CPU_DYING_FROZEN:
6597 /* Update our root-domain */
6599 spin_lock_irqsave(&rq->lock, flags);
6601 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6604 spin_unlock_irqrestore(&rq->lock, flags);
6611 /* Register at highest priority so that task migration (migrate_all_tasks)
6612 * happens before everything else.
6614 static struct notifier_block __cpuinitdata migration_notifier = {
6615 .notifier_call = migration_call,
6619 static int __init migration_init(void)
6621 void *cpu = (void *)(long)smp_processor_id();
6624 /* Start one for the boot CPU: */
6625 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6626 BUG_ON(err == NOTIFY_BAD);
6627 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6628 register_cpu_notifier(&migration_notifier);
6632 early_initcall(migration_init);
6637 #ifdef CONFIG_SCHED_DEBUG
6639 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6652 case SD_LV_ALLNODES:
6661 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6662 cpumask_t *groupmask)
6664 struct sched_group *group = sd->groups;
6667 cpulist_scnprintf(str, sizeof(str), sd->span);
6668 cpus_clear(*groupmask);
6670 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6672 if (!(sd->flags & SD_LOAD_BALANCE)) {
6673 printk("does not load-balance\n");
6675 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6680 printk(KERN_CONT "span %s level %s\n",
6681 str, sd_level_to_string(sd->level));
6683 if (!cpu_isset(cpu, sd->span)) {
6684 printk(KERN_ERR "ERROR: domain->span does not contain "
6687 if (!cpu_isset(cpu, group->cpumask)) {
6688 printk(KERN_ERR "ERROR: domain->groups does not contain"
6692 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6696 printk(KERN_ERR "ERROR: group is NULL\n");
6700 if (!group->__cpu_power) {
6701 printk(KERN_CONT "\n");
6702 printk(KERN_ERR "ERROR: domain->cpu_power not "
6707 if (!cpus_weight(group->cpumask)) {
6708 printk(KERN_CONT "\n");
6709 printk(KERN_ERR "ERROR: empty group\n");
6713 if (cpus_intersects(*groupmask, group->cpumask)) {
6714 printk(KERN_CONT "\n");
6715 printk(KERN_ERR "ERROR: repeated CPUs\n");
6719 cpus_or(*groupmask, *groupmask, group->cpumask);
6721 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6722 printk(KERN_CONT " %s", str);
6724 group = group->next;
6725 } while (group != sd->groups);
6726 printk(KERN_CONT "\n");
6728 if (!cpus_equal(sd->span, *groupmask))
6729 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6731 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6732 printk(KERN_ERR "ERROR: parent span is not a superset "
6733 "of domain->span\n");
6737 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6739 cpumask_t *groupmask;
6743 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6747 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6749 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6751 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6756 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6765 #else /* !CONFIG_SCHED_DEBUG */
6766 # define sched_domain_debug(sd, cpu) do { } while (0)
6767 #endif /* CONFIG_SCHED_DEBUG */
6769 static int sd_degenerate(struct sched_domain *sd)
6771 if (cpus_weight(sd->span) == 1)
6774 /* Following flags need at least 2 groups */
6775 if (sd->flags & (SD_LOAD_BALANCE |
6776 SD_BALANCE_NEWIDLE |
6780 SD_SHARE_PKG_RESOURCES)) {
6781 if (sd->groups != sd->groups->next)
6785 /* Following flags don't use groups */
6786 if (sd->flags & (SD_WAKE_IDLE |
6795 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6797 unsigned long cflags = sd->flags, pflags = parent->flags;
6799 if (sd_degenerate(parent))
6802 if (!cpus_equal(sd->span, parent->span))
6805 /* Does parent contain flags not in child? */
6806 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6807 if (cflags & SD_WAKE_AFFINE)
6808 pflags &= ~SD_WAKE_BALANCE;
6809 /* Flags needing groups don't count if only 1 group in parent */
6810 if (parent->groups == parent->groups->next) {
6811 pflags &= ~(SD_LOAD_BALANCE |
6812 SD_BALANCE_NEWIDLE |
6816 SD_SHARE_PKG_RESOURCES);
6818 if (~cflags & pflags)
6824 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6826 unsigned long flags;
6828 spin_lock_irqsave(&rq->lock, flags);
6831 struct root_domain *old_rd = rq->rd;
6833 if (cpu_isset(rq->cpu, old_rd->online))
6836 cpu_clear(rq->cpu, old_rd->span);
6838 if (atomic_dec_and_test(&old_rd->refcount))
6842 atomic_inc(&rd->refcount);
6845 cpu_set(rq->cpu, rd->span);
6846 if (cpu_isset(rq->cpu, cpu_online_map))
6849 spin_unlock_irqrestore(&rq->lock, flags);
6852 static void init_rootdomain(struct root_domain *rd)
6854 memset(rd, 0, sizeof(*rd));
6856 cpus_clear(rd->span);
6857 cpus_clear(rd->online);
6859 cpupri_init(&rd->cpupri);
6862 static void init_defrootdomain(void)
6864 init_rootdomain(&def_root_domain);
6865 atomic_set(&def_root_domain.refcount, 1);
6868 static struct root_domain *alloc_rootdomain(void)
6870 struct root_domain *rd;
6872 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6876 init_rootdomain(rd);
6882 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6883 * hold the hotplug lock.
6886 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6888 struct rq *rq = cpu_rq(cpu);
6889 struct sched_domain *tmp;
6891 /* Remove the sched domains which do not contribute to scheduling. */
6892 for (tmp = sd; tmp; ) {
6893 struct sched_domain *parent = tmp->parent;
6897 if (sd_parent_degenerate(tmp, parent)) {
6898 tmp->parent = parent->parent;
6900 parent->parent->child = tmp;
6905 if (sd && sd_degenerate(sd)) {
6911 sched_domain_debug(sd, cpu);
6913 rq_attach_root(rq, rd);
6914 rcu_assign_pointer(rq->sd, sd);
6917 /* cpus with isolated domains */
6918 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6920 /* Setup the mask of cpus configured for isolated domains */
6921 static int __init isolated_cpu_setup(char *str)
6923 static int __initdata ints[NR_CPUS];
6926 str = get_options(str, ARRAY_SIZE(ints), ints);
6927 cpus_clear(cpu_isolated_map);
6928 for (i = 1; i <= ints[0]; i++)
6929 if (ints[i] < NR_CPUS)
6930 cpu_set(ints[i], cpu_isolated_map);
6934 __setup("isolcpus=", isolated_cpu_setup);
6937 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6938 * to a function which identifies what group(along with sched group) a CPU
6939 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6940 * (due to the fact that we keep track of groups covered with a cpumask_t).
6942 * init_sched_build_groups will build a circular linked list of the groups
6943 * covered by the given span, and will set each group's ->cpumask correctly,
6944 * and ->cpu_power to 0.
6947 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6948 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6949 struct sched_group **sg,
6950 cpumask_t *tmpmask),
6951 cpumask_t *covered, cpumask_t *tmpmask)
6953 struct sched_group *first = NULL, *last = NULL;
6956 cpus_clear(*covered);
6958 for_each_cpu_mask_nr(i, *span) {
6959 struct sched_group *sg;
6960 int group = group_fn(i, cpu_map, &sg, tmpmask);
6963 if (cpu_isset(i, *covered))
6966 cpus_clear(sg->cpumask);
6967 sg->__cpu_power = 0;
6969 for_each_cpu_mask_nr(j, *span) {
6970 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6973 cpu_set(j, *covered);
6974 cpu_set(j, sg->cpumask);
6985 #define SD_NODES_PER_DOMAIN 16
6990 * find_next_best_node - find the next node to include in a sched_domain
6991 * @node: node whose sched_domain we're building
6992 * @used_nodes: nodes already in the sched_domain
6994 * Find the next node to include in a given scheduling domain. Simply
6995 * finds the closest node not already in the @used_nodes map.
6997 * Should use nodemask_t.
6999 static int find_next_best_node(int node, nodemask_t *used_nodes)
7001 int i, n, val, min_val, best_node = 0;
7005 for (i = 0; i < nr_node_ids; i++) {
7006 /* Start at @node */
7007 n = (node + i) % nr_node_ids;
7009 if (!nr_cpus_node(n))
7012 /* Skip already used nodes */
7013 if (node_isset(n, *used_nodes))
7016 /* Simple min distance search */
7017 val = node_distance(node, n);
7019 if (val < min_val) {
7025 node_set(best_node, *used_nodes);
7030 * sched_domain_node_span - get a cpumask for a node's sched_domain
7031 * @node: node whose cpumask we're constructing
7032 * @span: resulting cpumask
7034 * Given a node, construct a good cpumask for its sched_domain to span. It
7035 * should be one that prevents unnecessary balancing, but also spreads tasks
7038 static void sched_domain_node_span(int node, cpumask_t *span)
7040 nodemask_t used_nodes;
7041 node_to_cpumask_ptr(nodemask, node);
7045 nodes_clear(used_nodes);
7047 cpus_or(*span, *span, *nodemask);
7048 node_set(node, used_nodes);
7050 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7051 int next_node = find_next_best_node(node, &used_nodes);
7053 node_to_cpumask_ptr_next(nodemask, next_node);
7054 cpus_or(*span, *span, *nodemask);
7057 #endif /* CONFIG_NUMA */
7059 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7062 * SMT sched-domains:
7064 #ifdef CONFIG_SCHED_SMT
7065 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7066 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7069 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7073 *sg = &per_cpu(sched_group_cpus, cpu);
7076 #endif /* CONFIG_SCHED_SMT */
7079 * multi-core sched-domains:
7081 #ifdef CONFIG_SCHED_MC
7082 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7083 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7084 #endif /* CONFIG_SCHED_MC */
7086 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7088 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7093 *mask = per_cpu(cpu_sibling_map, cpu);
7094 cpus_and(*mask, *mask, *cpu_map);
7095 group = first_cpu(*mask);
7097 *sg = &per_cpu(sched_group_core, group);
7100 #elif defined(CONFIG_SCHED_MC)
7102 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7106 *sg = &per_cpu(sched_group_core, cpu);
7111 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7112 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7115 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7119 #ifdef CONFIG_SCHED_MC
7120 *mask = cpu_coregroup_map(cpu);
7121 cpus_and(*mask, *mask, *cpu_map);
7122 group = first_cpu(*mask);
7123 #elif defined(CONFIG_SCHED_SMT)
7124 *mask = per_cpu(cpu_sibling_map, cpu);
7125 cpus_and(*mask, *mask, *cpu_map);
7126 group = first_cpu(*mask);
7131 *sg = &per_cpu(sched_group_phys, group);
7137 * The init_sched_build_groups can't handle what we want to do with node
7138 * groups, so roll our own. Now each node has its own list of groups which
7139 * gets dynamically allocated.
7141 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7142 static struct sched_group ***sched_group_nodes_bycpu;
7144 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7145 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7147 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7148 struct sched_group **sg, cpumask_t *nodemask)
7152 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7153 cpus_and(*nodemask, *nodemask, *cpu_map);
7154 group = first_cpu(*nodemask);
7157 *sg = &per_cpu(sched_group_allnodes, group);
7161 static void init_numa_sched_groups_power(struct sched_group *group_head)
7163 struct sched_group *sg = group_head;
7169 for_each_cpu_mask_nr(j, sg->cpumask) {
7170 struct sched_domain *sd;
7172 sd = &per_cpu(phys_domains, j);
7173 if (j != first_cpu(sd->groups->cpumask)) {
7175 * Only add "power" once for each
7181 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7184 } while (sg != group_head);
7186 #endif /* CONFIG_NUMA */
7189 /* Free memory allocated for various sched_group structures */
7190 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7194 for_each_cpu_mask_nr(cpu, *cpu_map) {
7195 struct sched_group **sched_group_nodes
7196 = sched_group_nodes_bycpu[cpu];
7198 if (!sched_group_nodes)
7201 for (i = 0; i < nr_node_ids; i++) {
7202 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7204 *nodemask = node_to_cpumask(i);
7205 cpus_and(*nodemask, *nodemask, *cpu_map);
7206 if (cpus_empty(*nodemask))
7216 if (oldsg != sched_group_nodes[i])
7219 kfree(sched_group_nodes);
7220 sched_group_nodes_bycpu[cpu] = NULL;
7223 #else /* !CONFIG_NUMA */
7224 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7227 #endif /* CONFIG_NUMA */
7230 * Initialize sched groups cpu_power.
7232 * cpu_power indicates the capacity of sched group, which is used while
7233 * distributing the load between different sched groups in a sched domain.
7234 * Typically cpu_power for all the groups in a sched domain will be same unless
7235 * there are asymmetries in the topology. If there are asymmetries, group
7236 * having more cpu_power will pickup more load compared to the group having
7239 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7240 * the maximum number of tasks a group can handle in the presence of other idle
7241 * or lightly loaded groups in the same sched domain.
7243 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7245 struct sched_domain *child;
7246 struct sched_group *group;
7248 WARN_ON(!sd || !sd->groups);
7250 if (cpu != first_cpu(sd->groups->cpumask))
7255 sd->groups->__cpu_power = 0;
7258 * For perf policy, if the groups in child domain share resources
7259 * (for example cores sharing some portions of the cache hierarchy
7260 * or SMT), then set this domain groups cpu_power such that each group
7261 * can handle only one task, when there are other idle groups in the
7262 * same sched domain.
7264 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7266 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7267 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7272 * add cpu_power of each child group to this groups cpu_power
7274 group = child->groups;
7276 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7277 group = group->next;
7278 } while (group != child->groups);
7282 * Initializers for schedule domains
7283 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7286 #ifdef CONFIG_SCHED_DEBUG
7287 # define SD_INIT_NAME(sd, type) sd->name = #type
7289 # define SD_INIT_NAME(sd, type) do { } while (0)
7292 #define SD_INIT(sd, type) sd_init_##type(sd)
7294 #define SD_INIT_FUNC(type) \
7295 static noinline void sd_init_##type(struct sched_domain *sd) \
7297 memset(sd, 0, sizeof(*sd)); \
7298 *sd = SD_##type##_INIT; \
7299 sd->level = SD_LV_##type; \
7300 SD_INIT_NAME(sd, type); \
7305 SD_INIT_FUNC(ALLNODES)
7308 #ifdef CONFIG_SCHED_SMT
7309 SD_INIT_FUNC(SIBLING)
7311 #ifdef CONFIG_SCHED_MC
7316 * To minimize stack usage kmalloc room for cpumasks and share the
7317 * space as the usage in build_sched_domains() dictates. Used only
7318 * if the amount of space is significant.
7321 cpumask_t tmpmask; /* make this one first */
7324 cpumask_t this_sibling_map;
7325 cpumask_t this_core_map;
7327 cpumask_t send_covered;
7330 cpumask_t domainspan;
7332 cpumask_t notcovered;
7337 #define SCHED_CPUMASK_ALLOC 1
7338 #define SCHED_CPUMASK_FREE(v) kfree(v)
7339 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7341 #define SCHED_CPUMASK_ALLOC 0
7342 #define SCHED_CPUMASK_FREE(v)
7343 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7346 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7347 ((unsigned long)(a) + offsetof(struct allmasks, v))
7349 static int default_relax_domain_level = -1;
7351 static int __init setup_relax_domain_level(char *str)
7355 val = simple_strtoul(str, NULL, 0);
7356 if (val < SD_LV_MAX)
7357 default_relax_domain_level = val;
7361 __setup("relax_domain_level=", setup_relax_domain_level);
7363 static void set_domain_attribute(struct sched_domain *sd,
7364 struct sched_domain_attr *attr)
7368 if (!attr || attr->relax_domain_level < 0) {
7369 if (default_relax_domain_level < 0)
7372 request = default_relax_domain_level;
7374 request = attr->relax_domain_level;
7375 if (request < sd->level) {
7376 /* turn off idle balance on this domain */
7377 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7379 /* turn on idle balance on this domain */
7380 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7385 * Build sched domains for a given set of cpus and attach the sched domains
7386 * to the individual cpus
7388 static int __build_sched_domains(const cpumask_t *cpu_map,
7389 struct sched_domain_attr *attr)
7392 struct root_domain *rd;
7393 SCHED_CPUMASK_DECLARE(allmasks);
7396 struct sched_group **sched_group_nodes = NULL;
7397 int sd_allnodes = 0;
7400 * Allocate the per-node list of sched groups
7402 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7404 if (!sched_group_nodes) {
7405 printk(KERN_WARNING "Can not alloc sched group node list\n");
7410 rd = alloc_rootdomain();
7412 printk(KERN_WARNING "Cannot alloc root domain\n");
7414 kfree(sched_group_nodes);
7419 #if SCHED_CPUMASK_ALLOC
7420 /* get space for all scratch cpumask variables */
7421 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7423 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7426 kfree(sched_group_nodes);
7431 tmpmask = (cpumask_t *)allmasks;
7435 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7439 * Set up domains for cpus specified by the cpu_map.
7441 for_each_cpu_mask_nr(i, *cpu_map) {
7442 struct sched_domain *sd = NULL, *p;
7443 SCHED_CPUMASK_VAR(nodemask, allmasks);
7445 *nodemask = node_to_cpumask(cpu_to_node(i));
7446 cpus_and(*nodemask, *nodemask, *cpu_map);
7449 if (cpus_weight(*cpu_map) >
7450 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7451 sd = &per_cpu(allnodes_domains, i);
7452 SD_INIT(sd, ALLNODES);
7453 set_domain_attribute(sd, attr);
7454 sd->span = *cpu_map;
7455 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7461 sd = &per_cpu(node_domains, i);
7463 set_domain_attribute(sd, attr);
7464 sched_domain_node_span(cpu_to_node(i), &sd->span);
7468 cpus_and(sd->span, sd->span, *cpu_map);
7472 sd = &per_cpu(phys_domains, i);
7474 set_domain_attribute(sd, attr);
7475 sd->span = *nodemask;
7479 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7481 #ifdef CONFIG_SCHED_MC
7483 sd = &per_cpu(core_domains, i);
7485 set_domain_attribute(sd, attr);
7486 sd->span = cpu_coregroup_map(i);
7487 cpus_and(sd->span, sd->span, *cpu_map);
7490 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7493 #ifdef CONFIG_SCHED_SMT
7495 sd = &per_cpu(cpu_domains, i);
7496 SD_INIT(sd, SIBLING);
7497 set_domain_attribute(sd, attr);
7498 sd->span = per_cpu(cpu_sibling_map, i);
7499 cpus_and(sd->span, sd->span, *cpu_map);
7502 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7506 #ifdef CONFIG_SCHED_SMT
7507 /* Set up CPU (sibling) groups */
7508 for_each_cpu_mask_nr(i, *cpu_map) {
7509 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7510 SCHED_CPUMASK_VAR(send_covered, allmasks);
7512 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7513 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7514 if (i != first_cpu(*this_sibling_map))
7517 init_sched_build_groups(this_sibling_map, cpu_map,
7519 send_covered, tmpmask);
7523 #ifdef CONFIG_SCHED_MC
7524 /* Set up multi-core groups */
7525 for_each_cpu_mask_nr(i, *cpu_map) {
7526 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7527 SCHED_CPUMASK_VAR(send_covered, allmasks);
7529 *this_core_map = cpu_coregroup_map(i);
7530 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7531 if (i != first_cpu(*this_core_map))
7534 init_sched_build_groups(this_core_map, cpu_map,
7536 send_covered, tmpmask);
7540 /* Set up physical groups */
7541 for (i = 0; i < nr_node_ids; i++) {
7542 SCHED_CPUMASK_VAR(nodemask, allmasks);
7543 SCHED_CPUMASK_VAR(send_covered, allmasks);
7545 *nodemask = node_to_cpumask(i);
7546 cpus_and(*nodemask, *nodemask, *cpu_map);
7547 if (cpus_empty(*nodemask))
7550 init_sched_build_groups(nodemask, cpu_map,
7552 send_covered, tmpmask);
7556 /* Set up node groups */
7558 SCHED_CPUMASK_VAR(send_covered, allmasks);
7560 init_sched_build_groups(cpu_map, cpu_map,
7561 &cpu_to_allnodes_group,
7562 send_covered, tmpmask);
7565 for (i = 0; i < nr_node_ids; i++) {
7566 /* Set up node groups */
7567 struct sched_group *sg, *prev;
7568 SCHED_CPUMASK_VAR(nodemask, allmasks);
7569 SCHED_CPUMASK_VAR(domainspan, allmasks);
7570 SCHED_CPUMASK_VAR(covered, allmasks);
7573 *nodemask = node_to_cpumask(i);
7574 cpus_clear(*covered);
7576 cpus_and(*nodemask, *nodemask, *cpu_map);
7577 if (cpus_empty(*nodemask)) {
7578 sched_group_nodes[i] = NULL;
7582 sched_domain_node_span(i, domainspan);
7583 cpus_and(*domainspan, *domainspan, *cpu_map);
7585 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7587 printk(KERN_WARNING "Can not alloc domain group for "
7591 sched_group_nodes[i] = sg;
7592 for_each_cpu_mask_nr(j, *nodemask) {
7593 struct sched_domain *sd;
7595 sd = &per_cpu(node_domains, j);
7598 sg->__cpu_power = 0;
7599 sg->cpumask = *nodemask;
7601 cpus_or(*covered, *covered, *nodemask);
7604 for (j = 0; j < nr_node_ids; j++) {
7605 SCHED_CPUMASK_VAR(notcovered, allmasks);
7606 int n = (i + j) % nr_node_ids;
7607 node_to_cpumask_ptr(pnodemask, n);
7609 cpus_complement(*notcovered, *covered);
7610 cpus_and(*tmpmask, *notcovered, *cpu_map);
7611 cpus_and(*tmpmask, *tmpmask, *domainspan);
7612 if (cpus_empty(*tmpmask))
7615 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7616 if (cpus_empty(*tmpmask))
7619 sg = kmalloc_node(sizeof(struct sched_group),
7623 "Can not alloc domain group for node %d\n", j);
7626 sg->__cpu_power = 0;
7627 sg->cpumask = *tmpmask;
7628 sg->next = prev->next;
7629 cpus_or(*covered, *covered, *tmpmask);
7636 /* Calculate CPU power for physical packages and nodes */
7637 #ifdef CONFIG_SCHED_SMT
7638 for_each_cpu_mask_nr(i, *cpu_map) {
7639 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7641 init_sched_groups_power(i, sd);
7644 #ifdef CONFIG_SCHED_MC
7645 for_each_cpu_mask_nr(i, *cpu_map) {
7646 struct sched_domain *sd = &per_cpu(core_domains, i);
7648 init_sched_groups_power(i, sd);
7652 for_each_cpu_mask_nr(i, *cpu_map) {
7653 struct sched_domain *sd = &per_cpu(phys_domains, i);
7655 init_sched_groups_power(i, sd);
7659 for (i = 0; i < nr_node_ids; i++)
7660 init_numa_sched_groups_power(sched_group_nodes[i]);
7663 struct sched_group *sg;
7665 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7667 init_numa_sched_groups_power(sg);
7671 /* Attach the domains */
7672 for_each_cpu_mask_nr(i, *cpu_map) {
7673 struct sched_domain *sd;
7674 #ifdef CONFIG_SCHED_SMT
7675 sd = &per_cpu(cpu_domains, i);
7676 #elif defined(CONFIG_SCHED_MC)
7677 sd = &per_cpu(core_domains, i);
7679 sd = &per_cpu(phys_domains, i);
7681 cpu_attach_domain(sd, rd, i);
7684 SCHED_CPUMASK_FREE((void *)allmasks);
7689 free_sched_groups(cpu_map, tmpmask);
7690 SCHED_CPUMASK_FREE((void *)allmasks);
7696 static int build_sched_domains(const cpumask_t *cpu_map)
7698 return __build_sched_domains(cpu_map, NULL);
7701 static cpumask_t *doms_cur; /* current sched domains */
7702 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7703 static struct sched_domain_attr *dattr_cur;
7704 /* attribues of custom domains in 'doms_cur' */
7707 * Special case: If a kmalloc of a doms_cur partition (array of
7708 * cpumask_t) fails, then fallback to a single sched domain,
7709 * as determined by the single cpumask_t fallback_doms.
7711 static cpumask_t fallback_doms;
7713 void __attribute__((weak)) arch_update_cpu_topology(void)
7718 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7719 * For now this just excludes isolated cpus, but could be used to
7720 * exclude other special cases in the future.
7722 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7726 arch_update_cpu_topology();
7728 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7730 doms_cur = &fallback_doms;
7731 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7733 err = build_sched_domains(doms_cur);
7734 register_sched_domain_sysctl();
7739 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7742 free_sched_groups(cpu_map, tmpmask);
7746 * Detach sched domains from a group of cpus specified in cpu_map
7747 * These cpus will now be attached to the NULL domain
7749 static void detach_destroy_domains(const cpumask_t *cpu_map)
7754 unregister_sched_domain_sysctl();
7756 for_each_cpu_mask_nr(i, *cpu_map)
7757 cpu_attach_domain(NULL, &def_root_domain, i);
7758 synchronize_sched();
7759 arch_destroy_sched_domains(cpu_map, &tmpmask);
7762 /* handle null as "default" */
7763 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7764 struct sched_domain_attr *new, int idx_new)
7766 struct sched_domain_attr tmp;
7773 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7774 new ? (new + idx_new) : &tmp,
7775 sizeof(struct sched_domain_attr));
7779 * Partition sched domains as specified by the 'ndoms_new'
7780 * cpumasks in the array doms_new[] of cpumasks. This compares
7781 * doms_new[] to the current sched domain partitioning, doms_cur[].
7782 * It destroys each deleted domain and builds each new domain.
7784 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7785 * The masks don't intersect (don't overlap.) We should setup one
7786 * sched domain for each mask. CPUs not in any of the cpumasks will
7787 * not be load balanced. If the same cpumask appears both in the
7788 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7791 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7792 * ownership of it and will kfree it when done with it. If the caller
7793 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7794 * ndoms_new == 1, and partition_sched_domains() will fallback to
7795 * the single partition 'fallback_doms', it also forces the domains
7798 * If doms_new == NULL it will be replaced with cpu_online_map.
7799 * ndoms_new == 0 is a special case for destroying existing domains,
7800 * and it will not create the default domain.
7802 * Call with hotplug lock held
7804 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7805 struct sched_domain_attr *dattr_new)
7809 mutex_lock(&sched_domains_mutex);
7811 /* always unregister in case we don't destroy any domains */
7812 unregister_sched_domain_sysctl();
7814 n = doms_new ? ndoms_new : 0;
7816 /* Destroy deleted domains */
7817 for (i = 0; i < ndoms_cur; i++) {
7818 for (j = 0; j < n; j++) {
7819 if (cpus_equal(doms_cur[i], doms_new[j])
7820 && dattrs_equal(dattr_cur, i, dattr_new, j))
7823 /* no match - a current sched domain not in new doms_new[] */
7824 detach_destroy_domains(doms_cur + i);
7829 if (doms_new == NULL) {
7831 doms_new = &fallback_doms;
7832 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7836 /* Build new domains */
7837 for (i = 0; i < ndoms_new; i++) {
7838 for (j = 0; j < ndoms_cur; j++) {
7839 if (cpus_equal(doms_new[i], doms_cur[j])
7840 && dattrs_equal(dattr_new, i, dattr_cur, j))
7843 /* no match - add a new doms_new */
7844 __build_sched_domains(doms_new + i,
7845 dattr_new ? dattr_new + i : NULL);
7850 /* Remember the new sched domains */
7851 if (doms_cur != &fallback_doms)
7853 kfree(dattr_cur); /* kfree(NULL) is safe */
7854 doms_cur = doms_new;
7855 dattr_cur = dattr_new;
7856 ndoms_cur = ndoms_new;
7858 register_sched_domain_sysctl();
7860 mutex_unlock(&sched_domains_mutex);
7863 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7864 int arch_reinit_sched_domains(void)
7868 /* Destroy domains first to force the rebuild */
7869 partition_sched_domains(0, NULL, NULL);
7871 rebuild_sched_domains();
7877 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7881 if (buf[0] != '0' && buf[0] != '1')
7885 sched_smt_power_savings = (buf[0] == '1');
7887 sched_mc_power_savings = (buf[0] == '1');
7889 ret = arch_reinit_sched_domains();
7891 return ret ? ret : count;
7894 #ifdef CONFIG_SCHED_MC
7895 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
7898 return sprintf(page, "%u\n", sched_mc_power_savings);
7900 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
7901 const char *buf, size_t count)
7903 return sched_power_savings_store(buf, count, 0);
7905 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
7906 sched_mc_power_savings_show,
7907 sched_mc_power_savings_store);
7910 #ifdef CONFIG_SCHED_SMT
7911 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
7914 return sprintf(page, "%u\n", sched_smt_power_savings);
7916 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
7917 const char *buf, size_t count)
7919 return sched_power_savings_store(buf, count, 1);
7921 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
7922 sched_smt_power_savings_show,
7923 sched_smt_power_savings_store);
7926 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7930 #ifdef CONFIG_SCHED_SMT
7932 err = sysfs_create_file(&cls->kset.kobj,
7933 &attr_sched_smt_power_savings.attr);
7935 #ifdef CONFIG_SCHED_MC
7936 if (!err && mc_capable())
7937 err = sysfs_create_file(&cls->kset.kobj,
7938 &attr_sched_mc_power_savings.attr);
7942 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7944 #ifndef CONFIG_CPUSETS
7946 * Add online and remove offline CPUs from the scheduler domains.
7947 * When cpusets are enabled they take over this function.
7949 static int update_sched_domains(struct notifier_block *nfb,
7950 unsigned long action, void *hcpu)
7954 case CPU_ONLINE_FROZEN:
7956 case CPU_DEAD_FROZEN:
7957 partition_sched_domains(1, NULL, NULL);
7966 static int update_runtime(struct notifier_block *nfb,
7967 unsigned long action, void *hcpu)
7969 int cpu = (int)(long)hcpu;
7972 case CPU_DOWN_PREPARE:
7973 case CPU_DOWN_PREPARE_FROZEN:
7974 disable_runtime(cpu_rq(cpu));
7977 case CPU_DOWN_FAILED:
7978 case CPU_DOWN_FAILED_FROZEN:
7980 case CPU_ONLINE_FROZEN:
7981 enable_runtime(cpu_rq(cpu));
7989 void __init sched_init_smp(void)
7991 cpumask_t non_isolated_cpus;
7993 #if defined(CONFIG_NUMA)
7994 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7996 BUG_ON(sched_group_nodes_bycpu == NULL);
7999 mutex_lock(&sched_domains_mutex);
8000 arch_init_sched_domains(&cpu_online_map);
8001 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
8002 if (cpus_empty(non_isolated_cpus))
8003 cpu_set(smp_processor_id(), non_isolated_cpus);
8004 mutex_unlock(&sched_domains_mutex);
8007 #ifndef CONFIG_CPUSETS
8008 /* XXX: Theoretical race here - CPU may be hotplugged now */
8009 hotcpu_notifier(update_sched_domains, 0);
8012 /* RT runtime code needs to handle some hotplug events */
8013 hotcpu_notifier(update_runtime, 0);
8017 /* Move init over to a non-isolated CPU */
8018 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
8020 sched_init_granularity();
8023 void __init sched_init_smp(void)
8025 sched_init_granularity();
8027 #endif /* CONFIG_SMP */
8029 int in_sched_functions(unsigned long addr)
8031 return in_lock_functions(addr) ||
8032 (addr >= (unsigned long)__sched_text_start
8033 && addr < (unsigned long)__sched_text_end);
8036 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8038 cfs_rq->tasks_timeline = RB_ROOT;
8039 INIT_LIST_HEAD(&cfs_rq->tasks);
8040 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8046 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8048 struct rt_prio_array *array;
8051 array = &rt_rq->active;
8052 for (i = 0; i < MAX_RT_PRIO; i++) {
8053 INIT_LIST_HEAD(array->queue + i);
8054 __clear_bit(i, array->bitmap);
8056 /* delimiter for bitsearch: */
8057 __set_bit(MAX_RT_PRIO, array->bitmap);
8059 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8060 rt_rq->highest_prio = MAX_RT_PRIO;
8063 rt_rq->rt_nr_migratory = 0;
8064 rt_rq->overloaded = 0;
8068 rt_rq->rt_throttled = 0;
8069 rt_rq->rt_runtime = 0;
8070 spin_lock_init(&rt_rq->rt_runtime_lock);
8072 #ifdef CONFIG_RT_GROUP_SCHED
8073 rt_rq->rt_nr_boosted = 0;
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8080 struct sched_entity *se, int cpu, int add,
8081 struct sched_entity *parent)
8083 struct rq *rq = cpu_rq(cpu);
8084 tg->cfs_rq[cpu] = cfs_rq;
8085 init_cfs_rq(cfs_rq, rq);
8088 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8091 /* se could be NULL for init_task_group */
8096 se->cfs_rq = &rq->cfs;
8098 se->cfs_rq = parent->my_q;
8101 se->load.weight = tg->shares;
8102 se->load.inv_weight = 0;
8103 se->parent = parent;
8107 #ifdef CONFIG_RT_GROUP_SCHED
8108 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8109 struct sched_rt_entity *rt_se, int cpu, int add,
8110 struct sched_rt_entity *parent)
8112 struct rq *rq = cpu_rq(cpu);
8114 tg->rt_rq[cpu] = rt_rq;
8115 init_rt_rq(rt_rq, rq);
8117 rt_rq->rt_se = rt_se;
8118 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8120 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8122 tg->rt_se[cpu] = rt_se;
8127 rt_se->rt_rq = &rq->rt;
8129 rt_se->rt_rq = parent->my_q;
8131 rt_se->my_q = rt_rq;
8132 rt_se->parent = parent;
8133 INIT_LIST_HEAD(&rt_se->run_list);
8137 void __init sched_init(void)
8140 unsigned long alloc_size = 0, ptr;
8142 #ifdef CONFIG_FAIR_GROUP_SCHED
8143 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8145 #ifdef CONFIG_RT_GROUP_SCHED
8146 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8148 #ifdef CONFIG_USER_SCHED
8152 * As sched_init() is called before page_alloc is setup,
8153 * we use alloc_bootmem().
8156 ptr = (unsigned long)alloc_bootmem(alloc_size);
8158 #ifdef CONFIG_FAIR_GROUP_SCHED
8159 init_task_group.se = (struct sched_entity **)ptr;
8160 ptr += nr_cpu_ids * sizeof(void **);
8162 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8163 ptr += nr_cpu_ids * sizeof(void **);
8165 #ifdef CONFIG_USER_SCHED
8166 root_task_group.se = (struct sched_entity **)ptr;
8167 ptr += nr_cpu_ids * sizeof(void **);
8169 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8170 ptr += nr_cpu_ids * sizeof(void **);
8171 #endif /* CONFIG_USER_SCHED */
8172 #endif /* CONFIG_FAIR_GROUP_SCHED */
8173 #ifdef CONFIG_RT_GROUP_SCHED
8174 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8175 ptr += nr_cpu_ids * sizeof(void **);
8177 init_task_group.rt_rq = (struct rt_rq **)ptr;
8178 ptr += nr_cpu_ids * sizeof(void **);
8180 #ifdef CONFIG_USER_SCHED
8181 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8182 ptr += nr_cpu_ids * sizeof(void **);
8184 root_task_group.rt_rq = (struct rt_rq **)ptr;
8185 ptr += nr_cpu_ids * sizeof(void **);
8186 #endif /* CONFIG_USER_SCHED */
8187 #endif /* CONFIG_RT_GROUP_SCHED */
8191 init_defrootdomain();
8194 init_rt_bandwidth(&def_rt_bandwidth,
8195 global_rt_period(), global_rt_runtime());
8197 #ifdef CONFIG_RT_GROUP_SCHED
8198 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8199 global_rt_period(), global_rt_runtime());
8200 #ifdef CONFIG_USER_SCHED
8201 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8202 global_rt_period(), RUNTIME_INF);
8203 #endif /* CONFIG_USER_SCHED */
8204 #endif /* CONFIG_RT_GROUP_SCHED */
8206 #ifdef CONFIG_GROUP_SCHED
8207 list_add(&init_task_group.list, &task_groups);
8208 INIT_LIST_HEAD(&init_task_group.children);
8210 #ifdef CONFIG_USER_SCHED
8211 INIT_LIST_HEAD(&root_task_group.children);
8212 init_task_group.parent = &root_task_group;
8213 list_add(&init_task_group.siblings, &root_task_group.children);
8214 #endif /* CONFIG_USER_SCHED */
8215 #endif /* CONFIG_GROUP_SCHED */
8217 for_each_possible_cpu(i) {
8221 spin_lock_init(&rq->lock);
8223 init_cfs_rq(&rq->cfs, rq);
8224 init_rt_rq(&rq->rt, rq);
8225 #ifdef CONFIG_FAIR_GROUP_SCHED
8226 init_task_group.shares = init_task_group_load;
8227 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8228 #ifdef CONFIG_CGROUP_SCHED
8230 * How much cpu bandwidth does init_task_group get?
8232 * In case of task-groups formed thr' the cgroup filesystem, it
8233 * gets 100% of the cpu resources in the system. This overall
8234 * system cpu resource is divided among the tasks of
8235 * init_task_group and its child task-groups in a fair manner,
8236 * based on each entity's (task or task-group's) weight
8237 * (se->load.weight).
8239 * In other words, if init_task_group has 10 tasks of weight
8240 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8241 * then A0's share of the cpu resource is:
8243 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8245 * We achieve this by letting init_task_group's tasks sit
8246 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8248 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8249 #elif defined CONFIG_USER_SCHED
8250 root_task_group.shares = NICE_0_LOAD;
8251 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8253 * In case of task-groups formed thr' the user id of tasks,
8254 * init_task_group represents tasks belonging to root user.
8255 * Hence it forms a sibling of all subsequent groups formed.
8256 * In this case, init_task_group gets only a fraction of overall
8257 * system cpu resource, based on the weight assigned to root
8258 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8259 * by letting tasks of init_task_group sit in a separate cfs_rq
8260 * (init_cfs_rq) and having one entity represent this group of
8261 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8263 init_tg_cfs_entry(&init_task_group,
8264 &per_cpu(init_cfs_rq, i),
8265 &per_cpu(init_sched_entity, i), i, 1,
8266 root_task_group.se[i]);
8269 #endif /* CONFIG_FAIR_GROUP_SCHED */
8271 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8272 #ifdef CONFIG_RT_GROUP_SCHED
8273 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8274 #ifdef CONFIG_CGROUP_SCHED
8275 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8276 #elif defined CONFIG_USER_SCHED
8277 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8278 init_tg_rt_entry(&init_task_group,
8279 &per_cpu(init_rt_rq, i),
8280 &per_cpu(init_sched_rt_entity, i), i, 1,
8281 root_task_group.rt_se[i]);
8285 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8286 rq->cpu_load[j] = 0;
8290 rq->active_balance = 0;
8291 rq->next_balance = jiffies;
8295 rq->migration_thread = NULL;
8296 INIT_LIST_HEAD(&rq->migration_queue);
8297 rq_attach_root(rq, &def_root_domain);
8300 atomic_set(&rq->nr_iowait, 0);
8303 set_load_weight(&init_task);
8305 #ifdef CONFIG_PREEMPT_NOTIFIERS
8306 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8310 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8313 #ifdef CONFIG_RT_MUTEXES
8314 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8318 * The boot idle thread does lazy MMU switching as well:
8320 atomic_inc(&init_mm.mm_count);
8321 enter_lazy_tlb(&init_mm, current);
8324 * Make us the idle thread. Technically, schedule() should not be
8325 * called from this thread, however somewhere below it might be,
8326 * but because we are the idle thread, we just pick up running again
8327 * when this runqueue becomes "idle".
8329 init_idle(current, smp_processor_id());
8331 * During early bootup we pretend to be a normal task:
8333 current->sched_class = &fair_sched_class;
8335 scheduler_running = 1;
8338 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8339 void __might_sleep(char *file, int line)
8342 static unsigned long prev_jiffy; /* ratelimiting */
8344 if ((!in_atomic() && !irqs_disabled()) ||
8345 system_state != SYSTEM_RUNNING || oops_in_progress)
8347 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8349 prev_jiffy = jiffies;
8352 "BUG: sleeping function called from invalid context at %s:%d\n",
8355 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8356 in_atomic(), irqs_disabled(),
8357 current->pid, current->comm);
8359 debug_show_held_locks(current);
8360 if (irqs_disabled())
8361 print_irqtrace_events(current);
8365 EXPORT_SYMBOL(__might_sleep);
8368 #ifdef CONFIG_MAGIC_SYSRQ
8369 static void normalize_task(struct rq *rq, struct task_struct *p)
8373 update_rq_clock(rq);
8374 on_rq = p->se.on_rq;
8376 deactivate_task(rq, p, 0);
8377 __setscheduler(rq, p, SCHED_NORMAL, 0);
8379 activate_task(rq, p, 0);
8380 resched_task(rq->curr);
8384 void normalize_rt_tasks(void)
8386 struct task_struct *g, *p;
8387 unsigned long flags;
8390 read_lock_irqsave(&tasklist_lock, flags);
8391 do_each_thread(g, p) {
8393 * Only normalize user tasks:
8398 p->se.exec_start = 0;
8399 #ifdef CONFIG_SCHEDSTATS
8400 p->se.wait_start = 0;
8401 p->se.sleep_start = 0;
8402 p->se.block_start = 0;
8407 * Renice negative nice level userspace
8410 if (TASK_NICE(p) < 0 && p->mm)
8411 set_user_nice(p, 0);
8415 spin_lock(&p->pi_lock);
8416 rq = __task_rq_lock(p);
8418 normalize_task(rq, p);
8420 __task_rq_unlock(rq);
8421 spin_unlock(&p->pi_lock);
8422 } while_each_thread(g, p);
8424 read_unlock_irqrestore(&tasklist_lock, flags);
8427 #endif /* CONFIG_MAGIC_SYSRQ */
8431 * These functions are only useful for the IA64 MCA handling.
8433 * They can only be called when the whole system has been
8434 * stopped - every CPU needs to be quiescent, and no scheduling
8435 * activity can take place. Using them for anything else would
8436 * be a serious bug, and as a result, they aren't even visible
8437 * under any other configuration.
8441 * curr_task - return the current task for a given cpu.
8442 * @cpu: the processor in question.
8444 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8446 struct task_struct *curr_task(int cpu)
8448 return cpu_curr(cpu);
8452 * set_curr_task - set the current task for a given cpu.
8453 * @cpu: the processor in question.
8454 * @p: the task pointer to set.
8456 * Description: This function must only be used when non-maskable interrupts
8457 * are serviced on a separate stack. It allows the architecture to switch the
8458 * notion of the current task on a cpu in a non-blocking manner. This function
8459 * must be called with all CPU's synchronized, and interrupts disabled, the
8460 * and caller must save the original value of the current task (see
8461 * curr_task() above) and restore that value before reenabling interrupts and
8462 * re-starting the system.
8464 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8466 void set_curr_task(int cpu, struct task_struct *p)
8473 #ifdef CONFIG_FAIR_GROUP_SCHED
8474 static void free_fair_sched_group(struct task_group *tg)
8478 for_each_possible_cpu(i) {
8480 kfree(tg->cfs_rq[i]);
8490 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8492 struct cfs_rq *cfs_rq;
8493 struct sched_entity *se, *parent_se;
8497 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8500 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8504 tg->shares = NICE_0_LOAD;
8506 for_each_possible_cpu(i) {
8509 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8510 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8514 se = kmalloc_node(sizeof(struct sched_entity),
8515 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8519 parent_se = parent ? parent->se[i] : NULL;
8520 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8529 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8531 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8532 &cpu_rq(cpu)->leaf_cfs_rq_list);
8535 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8537 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8539 #else /* !CONFG_FAIR_GROUP_SCHED */
8540 static inline void free_fair_sched_group(struct task_group *tg)
8545 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8550 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8554 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8557 #endif /* CONFIG_FAIR_GROUP_SCHED */
8559 #ifdef CONFIG_RT_GROUP_SCHED
8560 static void free_rt_sched_group(struct task_group *tg)
8564 destroy_rt_bandwidth(&tg->rt_bandwidth);
8566 for_each_possible_cpu(i) {
8568 kfree(tg->rt_rq[i]);
8570 kfree(tg->rt_se[i]);
8578 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8580 struct rt_rq *rt_rq;
8581 struct sched_rt_entity *rt_se, *parent_se;
8585 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8588 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8592 init_rt_bandwidth(&tg->rt_bandwidth,
8593 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8595 for_each_possible_cpu(i) {
8598 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8599 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8603 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8604 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8608 parent_se = parent ? parent->rt_se[i] : NULL;
8609 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8618 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8620 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8621 &cpu_rq(cpu)->leaf_rt_rq_list);
8624 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8626 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8628 #else /* !CONFIG_RT_GROUP_SCHED */
8629 static inline void free_rt_sched_group(struct task_group *tg)
8634 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8639 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8643 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8646 #endif /* CONFIG_RT_GROUP_SCHED */
8648 #ifdef CONFIG_GROUP_SCHED
8649 static void free_sched_group(struct task_group *tg)
8651 free_fair_sched_group(tg);
8652 free_rt_sched_group(tg);
8656 /* allocate runqueue etc for a new task group */
8657 struct task_group *sched_create_group(struct task_group *parent)
8659 struct task_group *tg;
8660 unsigned long flags;
8663 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8665 return ERR_PTR(-ENOMEM);
8667 if (!alloc_fair_sched_group(tg, parent))
8670 if (!alloc_rt_sched_group(tg, parent))
8673 spin_lock_irqsave(&task_group_lock, flags);
8674 for_each_possible_cpu(i) {
8675 register_fair_sched_group(tg, i);
8676 register_rt_sched_group(tg, i);
8678 list_add_rcu(&tg->list, &task_groups);
8680 WARN_ON(!parent); /* root should already exist */
8682 tg->parent = parent;
8683 INIT_LIST_HEAD(&tg->children);
8684 list_add_rcu(&tg->siblings, &parent->children);
8685 spin_unlock_irqrestore(&task_group_lock, flags);
8690 free_sched_group(tg);
8691 return ERR_PTR(-ENOMEM);
8694 /* rcu callback to free various structures associated with a task group */
8695 static void free_sched_group_rcu(struct rcu_head *rhp)
8697 /* now it should be safe to free those cfs_rqs */
8698 free_sched_group(container_of(rhp, struct task_group, rcu));
8701 /* Destroy runqueue etc associated with a task group */
8702 void sched_destroy_group(struct task_group *tg)
8704 unsigned long flags;
8707 spin_lock_irqsave(&task_group_lock, flags);
8708 for_each_possible_cpu(i) {
8709 unregister_fair_sched_group(tg, i);
8710 unregister_rt_sched_group(tg, i);
8712 list_del_rcu(&tg->list);
8713 list_del_rcu(&tg->siblings);
8714 spin_unlock_irqrestore(&task_group_lock, flags);
8716 /* wait for possible concurrent references to cfs_rqs complete */
8717 call_rcu(&tg->rcu, free_sched_group_rcu);
8720 /* change task's runqueue when it moves between groups.
8721 * The caller of this function should have put the task in its new group
8722 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8723 * reflect its new group.
8725 void sched_move_task(struct task_struct *tsk)
8728 unsigned long flags;
8731 rq = task_rq_lock(tsk, &flags);
8733 update_rq_clock(rq);
8735 running = task_current(rq, tsk);
8736 on_rq = tsk->se.on_rq;
8739 dequeue_task(rq, tsk, 0);
8740 if (unlikely(running))
8741 tsk->sched_class->put_prev_task(rq, tsk);
8743 set_task_rq(tsk, task_cpu(tsk));
8745 #ifdef CONFIG_FAIR_GROUP_SCHED
8746 if (tsk->sched_class->moved_group)
8747 tsk->sched_class->moved_group(tsk);
8750 if (unlikely(running))
8751 tsk->sched_class->set_curr_task(rq);
8753 enqueue_task(rq, tsk, 0);
8755 task_rq_unlock(rq, &flags);
8757 #endif /* CONFIG_GROUP_SCHED */
8759 #ifdef CONFIG_FAIR_GROUP_SCHED
8760 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8762 struct cfs_rq *cfs_rq = se->cfs_rq;
8767 dequeue_entity(cfs_rq, se, 0);
8769 se->load.weight = shares;
8770 se->load.inv_weight = 0;
8773 enqueue_entity(cfs_rq, se, 0);
8776 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8778 struct cfs_rq *cfs_rq = se->cfs_rq;
8779 struct rq *rq = cfs_rq->rq;
8780 unsigned long flags;
8782 spin_lock_irqsave(&rq->lock, flags);
8783 __set_se_shares(se, shares);
8784 spin_unlock_irqrestore(&rq->lock, flags);
8787 static DEFINE_MUTEX(shares_mutex);
8789 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8792 unsigned long flags;
8795 * We can't change the weight of the root cgroup.
8800 if (shares < MIN_SHARES)
8801 shares = MIN_SHARES;
8802 else if (shares > MAX_SHARES)
8803 shares = MAX_SHARES;
8805 mutex_lock(&shares_mutex);
8806 if (tg->shares == shares)
8809 spin_lock_irqsave(&task_group_lock, flags);
8810 for_each_possible_cpu(i)
8811 unregister_fair_sched_group(tg, i);
8812 list_del_rcu(&tg->siblings);
8813 spin_unlock_irqrestore(&task_group_lock, flags);
8815 /* wait for any ongoing reference to this group to finish */
8816 synchronize_sched();
8819 * Now we are free to modify the group's share on each cpu
8820 * w/o tripping rebalance_share or load_balance_fair.
8822 tg->shares = shares;
8823 for_each_possible_cpu(i) {
8827 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8828 set_se_shares(tg->se[i], shares);
8832 * Enable load balance activity on this group, by inserting it back on
8833 * each cpu's rq->leaf_cfs_rq_list.
8835 spin_lock_irqsave(&task_group_lock, flags);
8836 for_each_possible_cpu(i)
8837 register_fair_sched_group(tg, i);
8838 list_add_rcu(&tg->siblings, &tg->parent->children);
8839 spin_unlock_irqrestore(&task_group_lock, flags);
8841 mutex_unlock(&shares_mutex);
8845 unsigned long sched_group_shares(struct task_group *tg)
8851 #ifdef CONFIG_RT_GROUP_SCHED
8853 * Ensure that the real time constraints are schedulable.
8855 static DEFINE_MUTEX(rt_constraints_mutex);
8857 static unsigned long to_ratio(u64 period, u64 runtime)
8859 if (runtime == RUNTIME_INF)
8862 return div64_u64(runtime << 20, period);
8865 /* Must be called with tasklist_lock held */
8866 static inline int tg_has_rt_tasks(struct task_group *tg)
8868 struct task_struct *g, *p;
8870 do_each_thread(g, p) {
8871 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8873 } while_each_thread(g, p);
8878 struct rt_schedulable_data {
8879 struct task_group *tg;
8884 static int tg_schedulable(struct task_group *tg, void *data)
8886 struct rt_schedulable_data *d = data;
8887 struct task_group *child;
8888 unsigned long total, sum = 0;
8889 u64 period, runtime;
8891 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8892 runtime = tg->rt_bandwidth.rt_runtime;
8895 period = d->rt_period;
8896 runtime = d->rt_runtime;
8900 * Cannot have more runtime than the period.
8902 if (runtime > period && runtime != RUNTIME_INF)
8906 * Ensure we don't starve existing RT tasks.
8908 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
8911 total = to_ratio(period, runtime);
8914 * Nobody can have more than the global setting allows.
8916 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
8920 * The sum of our children's runtime should not exceed our own.
8922 list_for_each_entry_rcu(child, &tg->children, siblings) {
8923 period = ktime_to_ns(child->rt_bandwidth.rt_period);
8924 runtime = child->rt_bandwidth.rt_runtime;
8926 if (child == d->tg) {
8927 period = d->rt_period;
8928 runtime = d->rt_runtime;
8931 sum += to_ratio(period, runtime);
8940 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8942 struct rt_schedulable_data data = {
8944 .rt_period = period,
8945 .rt_runtime = runtime,
8948 return walk_tg_tree(tg_schedulable, tg_nop, &data);
8951 static int tg_set_bandwidth(struct task_group *tg,
8952 u64 rt_period, u64 rt_runtime)
8956 mutex_lock(&rt_constraints_mutex);
8957 read_lock(&tasklist_lock);
8958 err = __rt_schedulable(tg, rt_period, rt_runtime);
8962 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8963 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8964 tg->rt_bandwidth.rt_runtime = rt_runtime;
8966 for_each_possible_cpu(i) {
8967 struct rt_rq *rt_rq = tg->rt_rq[i];
8969 spin_lock(&rt_rq->rt_runtime_lock);
8970 rt_rq->rt_runtime = rt_runtime;
8971 spin_unlock(&rt_rq->rt_runtime_lock);
8973 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8975 read_unlock(&tasklist_lock);
8976 mutex_unlock(&rt_constraints_mutex);
8981 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8983 u64 rt_runtime, rt_period;
8985 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8986 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8987 if (rt_runtime_us < 0)
8988 rt_runtime = RUNTIME_INF;
8990 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8993 long sched_group_rt_runtime(struct task_group *tg)
8997 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9000 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9001 do_div(rt_runtime_us, NSEC_PER_USEC);
9002 return rt_runtime_us;
9005 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9007 u64 rt_runtime, rt_period;
9009 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9010 rt_runtime = tg->rt_bandwidth.rt_runtime;
9015 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9018 long sched_group_rt_period(struct task_group *tg)
9022 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9023 do_div(rt_period_us, NSEC_PER_USEC);
9024 return rt_period_us;
9027 static int sched_rt_global_constraints(void)
9029 u64 runtime, period;
9032 if (sysctl_sched_rt_period <= 0)
9035 runtime = global_rt_runtime();
9036 period = global_rt_period();
9039 * Sanity check on the sysctl variables.
9041 if (runtime > period && runtime != RUNTIME_INF)
9044 mutex_lock(&rt_constraints_mutex);
9045 read_lock(&tasklist_lock);
9046 ret = __rt_schedulable(NULL, 0, 0);
9047 read_unlock(&tasklist_lock);
9048 mutex_unlock(&rt_constraints_mutex);
9052 #else /* !CONFIG_RT_GROUP_SCHED */
9053 static int sched_rt_global_constraints(void)
9055 unsigned long flags;
9058 if (sysctl_sched_rt_period <= 0)
9061 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9062 for_each_possible_cpu(i) {
9063 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9065 spin_lock(&rt_rq->rt_runtime_lock);
9066 rt_rq->rt_runtime = global_rt_runtime();
9067 spin_unlock(&rt_rq->rt_runtime_lock);
9069 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9073 #endif /* CONFIG_RT_GROUP_SCHED */
9075 int sched_rt_handler(struct ctl_table *table, int write,
9076 struct file *filp, void __user *buffer, size_t *lenp,
9080 int old_period, old_runtime;
9081 static DEFINE_MUTEX(mutex);
9084 old_period = sysctl_sched_rt_period;
9085 old_runtime = sysctl_sched_rt_runtime;
9087 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9089 if (!ret && write) {
9090 ret = sched_rt_global_constraints();
9092 sysctl_sched_rt_period = old_period;
9093 sysctl_sched_rt_runtime = old_runtime;
9095 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9096 def_rt_bandwidth.rt_period =
9097 ns_to_ktime(global_rt_period());
9100 mutex_unlock(&mutex);
9105 #ifdef CONFIG_CGROUP_SCHED
9107 /* return corresponding task_group object of a cgroup */
9108 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9110 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9111 struct task_group, css);
9114 static struct cgroup_subsys_state *
9115 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9117 struct task_group *tg, *parent;
9119 if (!cgrp->parent) {
9120 /* This is early initialization for the top cgroup */
9121 return &init_task_group.css;
9124 parent = cgroup_tg(cgrp->parent);
9125 tg = sched_create_group(parent);
9127 return ERR_PTR(-ENOMEM);
9133 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9135 struct task_group *tg = cgroup_tg(cgrp);
9137 sched_destroy_group(tg);
9141 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9142 struct task_struct *tsk)
9144 #ifdef CONFIG_RT_GROUP_SCHED
9145 /* Don't accept realtime tasks when there is no way for them to run */
9146 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9149 /* We don't support RT-tasks being in separate groups */
9150 if (tsk->sched_class != &fair_sched_class)
9158 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9159 struct cgroup *old_cont, struct task_struct *tsk)
9161 sched_move_task(tsk);
9164 #ifdef CONFIG_FAIR_GROUP_SCHED
9165 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9168 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9171 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9173 struct task_group *tg = cgroup_tg(cgrp);
9175 return (u64) tg->shares;
9177 #endif /* CONFIG_FAIR_GROUP_SCHED */
9179 #ifdef CONFIG_RT_GROUP_SCHED
9180 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9183 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9186 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9188 return sched_group_rt_runtime(cgroup_tg(cgrp));
9191 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9194 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9197 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9199 return sched_group_rt_period(cgroup_tg(cgrp));
9201 #endif /* CONFIG_RT_GROUP_SCHED */
9203 static struct cftype cpu_files[] = {
9204 #ifdef CONFIG_FAIR_GROUP_SCHED
9207 .read_u64 = cpu_shares_read_u64,
9208 .write_u64 = cpu_shares_write_u64,
9211 #ifdef CONFIG_RT_GROUP_SCHED
9213 .name = "rt_runtime_us",
9214 .read_s64 = cpu_rt_runtime_read,
9215 .write_s64 = cpu_rt_runtime_write,
9218 .name = "rt_period_us",
9219 .read_u64 = cpu_rt_period_read_uint,
9220 .write_u64 = cpu_rt_period_write_uint,
9225 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9227 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9230 struct cgroup_subsys cpu_cgroup_subsys = {
9232 .create = cpu_cgroup_create,
9233 .destroy = cpu_cgroup_destroy,
9234 .can_attach = cpu_cgroup_can_attach,
9235 .attach = cpu_cgroup_attach,
9236 .populate = cpu_cgroup_populate,
9237 .subsys_id = cpu_cgroup_subsys_id,
9241 #endif /* CONFIG_CGROUP_SCHED */
9243 #ifdef CONFIG_CGROUP_CPUACCT
9246 * CPU accounting code for task groups.
9248 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9249 * (balbir@in.ibm.com).
9252 /* track cpu usage of a group of tasks */
9254 struct cgroup_subsys_state css;
9255 /* cpuusage holds pointer to a u64-type object on every cpu */
9259 struct cgroup_subsys cpuacct_subsys;
9261 /* return cpu accounting group corresponding to this container */
9262 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9264 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9265 struct cpuacct, css);
9268 /* return cpu accounting group to which this task belongs */
9269 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9271 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9272 struct cpuacct, css);
9275 /* create a new cpu accounting group */
9276 static struct cgroup_subsys_state *cpuacct_create(
9277 struct cgroup_subsys *ss, struct cgroup *cgrp)
9279 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9282 return ERR_PTR(-ENOMEM);
9284 ca->cpuusage = alloc_percpu(u64);
9285 if (!ca->cpuusage) {
9287 return ERR_PTR(-ENOMEM);
9293 /* destroy an existing cpu accounting group */
9295 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9297 struct cpuacct *ca = cgroup_ca(cgrp);
9299 free_percpu(ca->cpuusage);
9303 /* return total cpu usage (in nanoseconds) of a group */
9304 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9306 struct cpuacct *ca = cgroup_ca(cgrp);
9307 u64 totalcpuusage = 0;
9310 for_each_possible_cpu(i) {
9311 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9314 * Take rq->lock to make 64-bit addition safe on 32-bit
9317 spin_lock_irq(&cpu_rq(i)->lock);
9318 totalcpuusage += *cpuusage;
9319 spin_unlock_irq(&cpu_rq(i)->lock);
9322 return totalcpuusage;
9325 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9328 struct cpuacct *ca = cgroup_ca(cgrp);
9337 for_each_possible_cpu(i) {
9338 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9340 spin_lock_irq(&cpu_rq(i)->lock);
9342 spin_unlock_irq(&cpu_rq(i)->lock);
9348 static struct cftype files[] = {
9351 .read_u64 = cpuusage_read,
9352 .write_u64 = cpuusage_write,
9356 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9358 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9362 * charge this task's execution time to its accounting group.
9364 * called with rq->lock held.
9366 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9370 if (!cpuacct_subsys.active)
9375 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9377 *cpuusage += cputime;
9381 struct cgroup_subsys cpuacct_subsys = {
9383 .create = cpuacct_create,
9384 .destroy = cpuacct_destroy,
9385 .populate = cpuacct_populate,
9386 .subsys_id = cpuacct_subsys_id,
9388 #endif /* CONFIG_CGROUP_CPUACCT */