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/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
108 * Timeslices get refilled after they expire.
110 #define DEF_TIMESLICE (100 * HZ / 1000)
113 * single value that denotes runtime == period, ie unlimited time.
115 #define RUNTIME_INF ((u64)~0ULL)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 struct rt_bandwidth {
159 /* nests inside the rq lock: */
160 spinlock_t rt_runtime_lock;
163 struct hrtimer rt_period_timer;
166 static struct rt_bandwidth def_rt_bandwidth;
168 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
170 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
172 struct rt_bandwidth *rt_b =
173 container_of(timer, struct rt_bandwidth, rt_period_timer);
179 now = hrtimer_cb_get_time(timer);
180 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
185 idle = do_sched_rt_period_timer(rt_b, overrun);
188 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
192 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
194 rt_b->rt_period = ns_to_ktime(period);
195 rt_b->rt_runtime = runtime;
197 spin_lock_init(&rt_b->rt_runtime_lock);
199 hrtimer_init(&rt_b->rt_period_timer,
200 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
201 rt_b->rt_period_timer.function = sched_rt_period_timer;
202 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
205 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
209 if (rt_b->rt_runtime == RUNTIME_INF)
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 spin_lock(&rt_b->rt_runtime_lock);
217 if (hrtimer_active(&rt_b->rt_period_timer))
220 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
221 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
222 hrtimer_start(&rt_b->rt_period_timer,
223 rt_b->rt_period_timer.expires,
226 spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to arch_init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_GROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 #ifdef CONFIG_CGROUP_SCHED
253 struct cgroup_subsys_state css;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
283 * Every UID task group (including init_task_group aka UID-0) will
284 * be a child to this group.
286 struct task_group root_task_group;
288 #ifdef CONFIG_FAIR_GROUP_SCHED
289 /* Default task group's sched entity on each cpu */
290 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
291 /* Default task group's cfs_rq on each cpu */
292 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
295 #ifdef CONFIG_RT_GROUP_SCHED
296 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
297 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
300 #define root_task_group init_task_group
303 /* task_group_lock serializes add/remove of task groups and also changes to
304 * a task group's cpu shares.
306 static DEFINE_SPINLOCK(task_group_lock);
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 #ifdef CONFIG_USER_SCHED
310 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
316 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
317 * (The default weight is 1024 - so there's no practical
318 * limitation from this.)
321 #define MAX_SHARES (ULONG_MAX - 1)
323 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
326 /* Default task group.
327 * Every task in system belong to this group at bootup.
329 struct task_group init_task_group;
331 /* return group to which a task belongs */
332 static inline struct task_group *task_group(struct task_struct *p)
334 struct task_group *tg;
336 #ifdef CONFIG_USER_SCHED
338 #elif defined(CONFIG_CGROUP_SCHED)
339 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
340 struct task_group, css);
342 tg = &init_task_group;
347 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
348 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
350 #ifdef CONFIG_FAIR_GROUP_SCHED
351 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
352 p->se.parent = task_group(p)->se[cpu];
355 #ifdef CONFIG_RT_GROUP_SCHED
356 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
357 p->rt.parent = task_group(p)->rt_se[cpu];
363 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
365 #endif /* CONFIG_GROUP_SCHED */
367 /* CFS-related fields in a runqueue */
369 struct load_weight load;
370 unsigned long nr_running;
375 struct rb_root tasks_timeline;
376 struct rb_node *rb_leftmost;
378 struct list_head tasks;
379 struct list_head *balance_iterator;
382 * 'curr' points to currently running entity on this cfs_rq.
383 * It is set to NULL otherwise (i.e when none are currently running).
385 struct sched_entity *curr, *next;
387 unsigned long nr_spread_over;
389 #ifdef CONFIG_FAIR_GROUP_SCHED
390 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
393 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
394 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
395 * (like users, containers etc.)
397 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
398 * list is used during load balance.
400 struct list_head leaf_cfs_rq_list;
401 struct task_group *tg; /* group that "owns" this runqueue */
404 unsigned long task_weight;
405 unsigned long shares;
407 * We need space to build a sched_domain wide view of the full task
408 * group tree, in order to avoid depending on dynamic memory allocation
409 * during the load balancing we place this in the per cpu task group
410 * hierarchy. This limits the load balancing to one instance per cpu,
411 * but more should not be needed anyway.
413 struct aggregate_struct {
415 * load = weight(cpus) * f(tg)
417 * Where f(tg) is the recursive weight fraction assigned to
423 * part of the group weight distributed to this span.
425 unsigned long shares;
428 * The sum of all runqueue weights within this span.
430 unsigned long rq_weight;
433 * Weight contributed by tasks; this is the part we can
434 * influence by moving tasks around.
436 unsigned long task_weight;
442 /* Real-Time classes' related field in a runqueue: */
444 struct rt_prio_array active;
445 unsigned long rt_nr_running;
446 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
447 int highest_prio; /* highest queued rt task prio */
450 unsigned long rt_nr_migratory;
456 /* Nests inside the rq lock: */
457 spinlock_t rt_runtime_lock;
459 #ifdef CONFIG_RT_GROUP_SCHED
460 unsigned long rt_nr_boosted;
463 struct list_head leaf_rt_rq_list;
464 struct task_group *tg;
465 struct sched_rt_entity *rt_se;
472 * We add the notion of a root-domain which will be used to define per-domain
473 * variables. Each exclusive cpuset essentially defines an island domain by
474 * fully partitioning the member cpus from any other cpuset. Whenever a new
475 * exclusive cpuset is created, we also create and attach a new root-domain
485 * The "RT overload" flag: it gets set if a CPU has more than
486 * one runnable RT task.
493 * By default the system creates a single root-domain with all cpus as
494 * members (mimicking the global state we have today).
496 static struct root_domain def_root_domain;
501 * This is the main, per-CPU runqueue data structure.
503 * Locking rule: those places that want to lock multiple runqueues
504 * (such as the load balancing or the thread migration code), lock
505 * acquire operations must be ordered by ascending &runqueue.
512 * nr_running and cpu_load should be in the same cacheline because
513 * remote CPUs use both these fields when doing load calculation.
515 unsigned long nr_running;
516 #define CPU_LOAD_IDX_MAX 5
517 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
518 unsigned char idle_at_tick;
520 unsigned long last_tick_seen;
521 unsigned char in_nohz_recently;
523 /* capture load from *all* tasks on this cpu: */
524 struct load_weight load;
525 unsigned long nr_load_updates;
531 #ifdef CONFIG_FAIR_GROUP_SCHED
532 /* list of leaf cfs_rq on this cpu: */
533 struct list_head leaf_cfs_rq_list;
535 #ifdef CONFIG_RT_GROUP_SCHED
536 struct list_head leaf_rt_rq_list;
540 * This is part of a global counter where only the total sum
541 * over all CPUs matters. A task can increase this counter on
542 * one CPU and if it got migrated afterwards it may decrease
543 * it on another CPU. Always updated under the runqueue lock:
545 unsigned long nr_uninterruptible;
547 struct task_struct *curr, *idle;
548 unsigned long next_balance;
549 struct mm_struct *prev_mm;
556 struct root_domain *rd;
557 struct sched_domain *sd;
559 /* For active balancing */
562 /* cpu of this runqueue: */
565 struct task_struct *migration_thread;
566 struct list_head migration_queue;
569 #ifdef CONFIG_SCHED_HRTICK
570 unsigned long hrtick_flags;
571 ktime_t hrtick_expire;
572 struct hrtimer hrtick_timer;
575 #ifdef CONFIG_SCHEDSTATS
577 struct sched_info rq_sched_info;
579 /* sys_sched_yield() stats */
580 unsigned int yld_exp_empty;
581 unsigned int yld_act_empty;
582 unsigned int yld_both_empty;
583 unsigned int yld_count;
585 /* schedule() stats */
586 unsigned int sched_switch;
587 unsigned int sched_count;
588 unsigned int sched_goidle;
590 /* try_to_wake_up() stats */
591 unsigned int ttwu_count;
592 unsigned int ttwu_local;
595 unsigned int bkl_count;
597 struct lock_class_key rq_lock_key;
600 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
602 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
604 rq->curr->sched_class->check_preempt_curr(rq, p);
607 static inline int cpu_of(struct rq *rq)
617 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
618 * See detach_destroy_domains: synchronize_sched for details.
620 * The domain tree of any CPU may only be accessed from within
621 * preempt-disabled sections.
623 #define for_each_domain(cpu, __sd) \
624 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
626 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
627 #define this_rq() (&__get_cpu_var(runqueues))
628 #define task_rq(p) cpu_rq(task_cpu(p))
629 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
631 static inline void update_rq_clock(struct rq *rq)
633 rq->clock = sched_clock_cpu(cpu_of(rq));
637 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
639 #ifdef CONFIG_SCHED_DEBUG
640 # define const_debug __read_mostly
642 # define const_debug static const
648 * Returns true if the current cpu runqueue is locked.
649 * This interface allows printk to be called with the runqueue lock
650 * held and know whether or not it is OK to wake up the klogd.
652 int runqueue_is_locked(void)
655 struct rq *rq = cpu_rq(cpu);
658 ret = spin_is_locked(&rq->lock);
664 * Debugging: various feature bits
667 #define SCHED_FEAT(name, enabled) \
668 __SCHED_FEAT_##name ,
671 #include "sched_features.h"
676 #define SCHED_FEAT(name, enabled) \
677 (1UL << __SCHED_FEAT_##name) * enabled |
679 const_debug unsigned int sysctl_sched_features =
680 #include "sched_features.h"
685 #ifdef CONFIG_SCHED_DEBUG
686 #define SCHED_FEAT(name, enabled) \
689 static __read_mostly char *sched_feat_names[] = {
690 #include "sched_features.h"
696 static int sched_feat_open(struct inode *inode, struct file *filp)
698 filp->private_data = inode->i_private;
703 sched_feat_read(struct file *filp, char __user *ubuf,
704 size_t cnt, loff_t *ppos)
711 for (i = 0; sched_feat_names[i]; i++) {
712 len += strlen(sched_feat_names[i]);
716 buf = kmalloc(len + 2, GFP_KERNEL);
720 for (i = 0; sched_feat_names[i]; i++) {
721 if (sysctl_sched_features & (1UL << i))
722 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
724 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
727 r += sprintf(buf + r, "\n");
728 WARN_ON(r >= len + 2);
730 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
738 sched_feat_write(struct file *filp, const char __user *ubuf,
739 size_t cnt, loff_t *ppos)
749 if (copy_from_user(&buf, ubuf, cnt))
754 if (strncmp(buf, "NO_", 3) == 0) {
759 for (i = 0; sched_feat_names[i]; i++) {
760 int len = strlen(sched_feat_names[i]);
762 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
764 sysctl_sched_features &= ~(1UL << i);
766 sysctl_sched_features |= (1UL << i);
771 if (!sched_feat_names[i])
779 static struct file_operations sched_feat_fops = {
780 .open = sched_feat_open,
781 .read = sched_feat_read,
782 .write = sched_feat_write,
785 static __init int sched_init_debug(void)
787 debugfs_create_file("sched_features", 0644, NULL, NULL,
792 late_initcall(sched_init_debug);
796 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
799 * Number of tasks to iterate in a single balance run.
800 * Limited because this is done with IRQs disabled.
802 const_debug unsigned int sysctl_sched_nr_migrate = 32;
805 * period over which we measure -rt task cpu usage in us.
808 unsigned int sysctl_sched_rt_period = 1000000;
810 static __read_mostly int scheduler_running;
813 * part of the period that we allow rt tasks to run in us.
816 int sysctl_sched_rt_runtime = 950000;
818 static inline u64 global_rt_period(void)
820 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
823 static inline u64 global_rt_runtime(void)
825 if (sysctl_sched_rt_period < 0)
828 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
831 unsigned long long time_sync_thresh = 100000;
833 static DEFINE_PER_CPU(unsigned long long, time_offset);
834 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
837 * Global lock which we take every now and then to synchronize
838 * the CPUs time. This method is not warp-safe, but it's good
839 * enough to synchronize slowly diverging time sources and thus
840 * it's good enough for tracing:
842 static DEFINE_SPINLOCK(time_sync_lock);
843 static unsigned long long prev_global_time;
845 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
848 * We want this inlined, to not get tracer function calls
849 * in this critical section:
851 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
852 __raw_spin_lock(&time_sync_lock.raw_lock);
854 if (time < prev_global_time) {
855 per_cpu(time_offset, cpu) += prev_global_time - time;
856 time = prev_global_time;
858 prev_global_time = time;
861 __raw_spin_unlock(&time_sync_lock.raw_lock);
862 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
867 static unsigned long long __cpu_clock(int cpu)
869 unsigned long long now;
872 * Only call sched_clock() if the scheduler has already been
873 * initialized (some code might call cpu_clock() very early):
875 if (unlikely(!scheduler_running))
878 now = sched_clock_cpu(cpu);
884 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
885 * clock constructed from sched_clock():
887 unsigned long long cpu_clock(int cpu)
889 unsigned long long prev_cpu_time, time, delta_time;
892 local_irq_save(flags);
893 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
894 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
895 delta_time = time-prev_cpu_time;
897 if (unlikely(delta_time > time_sync_thresh)) {
898 time = __sync_cpu_clock(time, cpu);
899 per_cpu(prev_cpu_time, cpu) = time;
901 local_irq_restore(flags);
905 EXPORT_SYMBOL_GPL(cpu_clock);
907 #ifndef prepare_arch_switch
908 # define prepare_arch_switch(next) do { } while (0)
910 #ifndef finish_arch_switch
911 # define finish_arch_switch(prev) do { } while (0)
914 static inline int task_current(struct rq *rq, struct task_struct *p)
916 return rq->curr == p;
919 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
920 static inline int task_running(struct rq *rq, struct task_struct *p)
922 return task_current(rq, p);
925 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
929 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
931 #ifdef CONFIG_DEBUG_SPINLOCK
932 /* this is a valid case when another task releases the spinlock */
933 rq->lock.owner = current;
936 * If we are tracking spinlock dependencies then we have to
937 * fix up the runqueue lock - which gets 'carried over' from
940 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
942 spin_unlock_irq(&rq->lock);
945 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
946 static inline int task_running(struct rq *rq, struct task_struct *p)
951 return task_current(rq, p);
955 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
959 * We can optimise this out completely for !SMP, because the
960 * SMP rebalancing from interrupt is the only thing that cares
965 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
966 spin_unlock_irq(&rq->lock);
968 spin_unlock(&rq->lock);
972 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
976 * After ->oncpu is cleared, the task can be moved to a different CPU.
977 * We must ensure this doesn't happen until the switch is completely
983 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
987 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
990 * __task_rq_lock - lock the runqueue a given task resides on.
991 * Must be called interrupts disabled.
993 static inline struct rq *__task_rq_lock(struct task_struct *p)
997 struct rq *rq = task_rq(p);
998 spin_lock(&rq->lock);
999 if (likely(rq == task_rq(p)))
1001 spin_unlock(&rq->lock);
1006 * task_rq_lock - lock the runqueue a given task resides on and disable
1007 * interrupts. Note the ordering: we can safely lookup the task_rq without
1008 * explicitly disabling preemption.
1010 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1011 __acquires(rq->lock)
1016 local_irq_save(*flags);
1018 spin_lock(&rq->lock);
1019 if (likely(rq == task_rq(p)))
1021 spin_unlock_irqrestore(&rq->lock, *flags);
1025 static void __task_rq_unlock(struct rq *rq)
1026 __releases(rq->lock)
1028 spin_unlock(&rq->lock);
1031 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1032 __releases(rq->lock)
1034 spin_unlock_irqrestore(&rq->lock, *flags);
1038 * this_rq_lock - lock this runqueue and disable interrupts.
1040 static struct rq *this_rq_lock(void)
1041 __acquires(rq->lock)
1045 local_irq_disable();
1047 spin_lock(&rq->lock);
1052 static void __resched_task(struct task_struct *p, int tif_bit);
1054 static inline void resched_task(struct task_struct *p)
1056 __resched_task(p, TIF_NEED_RESCHED);
1059 #ifdef CONFIG_SCHED_HRTICK
1061 * Use HR-timers to deliver accurate preemption points.
1063 * Its all a bit involved since we cannot program an hrt while holding the
1064 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1067 * When we get rescheduled we reprogram the hrtick_timer outside of the
1070 static inline void resched_hrt(struct task_struct *p)
1072 __resched_task(p, TIF_HRTICK_RESCHED);
1075 static inline void resched_rq(struct rq *rq)
1077 unsigned long flags;
1079 spin_lock_irqsave(&rq->lock, flags);
1080 resched_task(rq->curr);
1081 spin_unlock_irqrestore(&rq->lock, flags);
1085 HRTICK_SET, /* re-programm hrtick_timer */
1086 HRTICK_RESET, /* not a new slice */
1087 HRTICK_BLOCK, /* stop hrtick operations */
1092 * - enabled by features
1093 * - hrtimer is actually high res
1095 static inline int hrtick_enabled(struct rq *rq)
1097 if (!sched_feat(HRTICK))
1099 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1101 return hrtimer_is_hres_active(&rq->hrtick_timer);
1105 * Called to set the hrtick timer state.
1107 * called with rq->lock held and irqs disabled
1109 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1111 assert_spin_locked(&rq->lock);
1114 * preempt at: now + delay
1117 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1119 * indicate we need to program the timer
1121 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1123 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1126 * New slices are called from the schedule path and don't need a
1127 * forced reschedule.
1130 resched_hrt(rq->curr);
1133 static void hrtick_clear(struct rq *rq)
1135 if (hrtimer_active(&rq->hrtick_timer))
1136 hrtimer_cancel(&rq->hrtick_timer);
1140 * Update the timer from the possible pending state.
1142 static void hrtick_set(struct rq *rq)
1146 unsigned long flags;
1148 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1150 spin_lock_irqsave(&rq->lock, flags);
1151 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1152 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1153 time = rq->hrtick_expire;
1154 clear_thread_flag(TIF_HRTICK_RESCHED);
1155 spin_unlock_irqrestore(&rq->lock, flags);
1158 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1159 if (reset && !hrtimer_active(&rq->hrtick_timer))
1166 * High-resolution timer tick.
1167 * Runs from hardirq context with interrupts disabled.
1169 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1171 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1173 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1175 spin_lock(&rq->lock);
1176 update_rq_clock(rq);
1177 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1178 spin_unlock(&rq->lock);
1180 return HRTIMER_NORESTART;
1183 static void hotplug_hrtick_disable(int cpu)
1185 struct rq *rq = cpu_rq(cpu);
1186 unsigned long flags;
1188 spin_lock_irqsave(&rq->lock, flags);
1189 rq->hrtick_flags = 0;
1190 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1191 spin_unlock_irqrestore(&rq->lock, flags);
1196 static void hotplug_hrtick_enable(int cpu)
1198 struct rq *rq = cpu_rq(cpu);
1199 unsigned long flags;
1201 spin_lock_irqsave(&rq->lock, flags);
1202 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1203 spin_unlock_irqrestore(&rq->lock, flags);
1207 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1209 int cpu = (int)(long)hcpu;
1212 case CPU_UP_CANCELED:
1213 case CPU_UP_CANCELED_FROZEN:
1214 case CPU_DOWN_PREPARE:
1215 case CPU_DOWN_PREPARE_FROZEN:
1217 case CPU_DEAD_FROZEN:
1218 hotplug_hrtick_disable(cpu);
1221 case CPU_UP_PREPARE:
1222 case CPU_UP_PREPARE_FROZEN:
1223 case CPU_DOWN_FAILED:
1224 case CPU_DOWN_FAILED_FROZEN:
1226 case CPU_ONLINE_FROZEN:
1227 hotplug_hrtick_enable(cpu);
1234 static void init_hrtick(void)
1236 hotcpu_notifier(hotplug_hrtick, 0);
1239 static void init_rq_hrtick(struct rq *rq)
1241 rq->hrtick_flags = 0;
1242 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1243 rq->hrtick_timer.function = hrtick;
1244 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1247 void hrtick_resched(void)
1250 unsigned long flags;
1252 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1255 local_irq_save(flags);
1256 rq = cpu_rq(smp_processor_id());
1258 local_irq_restore(flags);
1261 static inline void hrtick_clear(struct rq *rq)
1265 static inline void hrtick_set(struct rq *rq)
1269 static inline void init_rq_hrtick(struct rq *rq)
1273 void hrtick_resched(void)
1277 static inline void init_hrtick(void)
1283 * resched_task - mark a task 'to be rescheduled now'.
1285 * On UP this means the setting of the need_resched flag, on SMP it
1286 * might also involve a cross-CPU call to trigger the scheduler on
1291 #ifndef tsk_is_polling
1292 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1295 static void __resched_task(struct task_struct *p, int tif_bit)
1299 assert_spin_locked(&task_rq(p)->lock);
1301 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1304 set_tsk_thread_flag(p, tif_bit);
1307 if (cpu == smp_processor_id())
1310 /* NEED_RESCHED must be visible before we test polling */
1312 if (!tsk_is_polling(p))
1313 smp_send_reschedule(cpu);
1316 static void resched_cpu(int cpu)
1318 struct rq *rq = cpu_rq(cpu);
1319 unsigned long flags;
1321 if (!spin_trylock_irqsave(&rq->lock, flags))
1323 resched_task(cpu_curr(cpu));
1324 spin_unlock_irqrestore(&rq->lock, flags);
1329 * When add_timer_on() enqueues a timer into the timer wheel of an
1330 * idle CPU then this timer might expire before the next timer event
1331 * which is scheduled to wake up that CPU. In case of a completely
1332 * idle system the next event might even be infinite time into the
1333 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1334 * leaves the inner idle loop so the newly added timer is taken into
1335 * account when the CPU goes back to idle and evaluates the timer
1336 * wheel for the next timer event.
1338 void wake_up_idle_cpu(int cpu)
1340 struct rq *rq = cpu_rq(cpu);
1342 if (cpu == smp_processor_id())
1346 * This is safe, as this function is called with the timer
1347 * wheel base lock of (cpu) held. When the CPU is on the way
1348 * to idle and has not yet set rq->curr to idle then it will
1349 * be serialized on the timer wheel base lock and take the new
1350 * timer into account automatically.
1352 if (rq->curr != rq->idle)
1356 * We can set TIF_RESCHED on the idle task of the other CPU
1357 * lockless. The worst case is that the other CPU runs the
1358 * idle task through an additional NOOP schedule()
1360 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1362 /* NEED_RESCHED must be visible before we test polling */
1364 if (!tsk_is_polling(rq->idle))
1365 smp_send_reschedule(cpu);
1370 static void __resched_task(struct task_struct *p, int tif_bit)
1372 assert_spin_locked(&task_rq(p)->lock);
1373 set_tsk_thread_flag(p, tif_bit);
1377 #if BITS_PER_LONG == 32
1378 # define WMULT_CONST (~0UL)
1380 # define WMULT_CONST (1UL << 32)
1383 #define WMULT_SHIFT 32
1386 * Shift right and round:
1388 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1391 * delta *= weight / lw
1393 static unsigned long
1394 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1395 struct load_weight *lw)
1399 if (!lw->inv_weight)
1400 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1402 tmp = (u64)delta_exec * weight;
1404 * Check whether we'd overflow the 64-bit multiplication:
1406 if (unlikely(tmp > WMULT_CONST))
1407 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1410 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1412 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1415 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1421 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1428 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1429 * of tasks with abnormal "nice" values across CPUs the contribution that
1430 * each task makes to its run queue's load is weighted according to its
1431 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1432 * scaled version of the new time slice allocation that they receive on time
1436 #define WEIGHT_IDLEPRIO 2
1437 #define WMULT_IDLEPRIO (1 << 31)
1440 * Nice levels are multiplicative, with a gentle 10% change for every
1441 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1442 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1443 * that remained on nice 0.
1445 * The "10% effect" is relative and cumulative: from _any_ nice level,
1446 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1447 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1448 * If a task goes up by ~10% and another task goes down by ~10% then
1449 * the relative distance between them is ~25%.)
1451 static const int prio_to_weight[40] = {
1452 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1453 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1454 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1455 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1456 /* 0 */ 1024, 820, 655, 526, 423,
1457 /* 5 */ 335, 272, 215, 172, 137,
1458 /* 10 */ 110, 87, 70, 56, 45,
1459 /* 15 */ 36, 29, 23, 18, 15,
1463 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1465 * In cases where the weight does not change often, we can use the
1466 * precalculated inverse to speed up arithmetics by turning divisions
1467 * into multiplications:
1469 static const u32 prio_to_wmult[40] = {
1470 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1471 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1472 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1473 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1474 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1475 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1476 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1477 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1480 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1483 * runqueue iterator, to support SMP load-balancing between different
1484 * scheduling classes, without having to expose their internal data
1485 * structures to the load-balancing proper:
1487 struct rq_iterator {
1489 struct task_struct *(*start)(void *);
1490 struct task_struct *(*next)(void *);
1494 static unsigned long
1495 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1496 unsigned long max_load_move, struct sched_domain *sd,
1497 enum cpu_idle_type idle, int *all_pinned,
1498 int *this_best_prio, struct rq_iterator *iterator);
1501 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1502 struct sched_domain *sd, enum cpu_idle_type idle,
1503 struct rq_iterator *iterator);
1506 #ifdef CONFIG_CGROUP_CPUACCT
1507 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1509 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1512 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1514 update_load_add(&rq->load, load);
1517 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1519 update_load_sub(&rq->load, load);
1523 static unsigned long source_load(int cpu, int type);
1524 static unsigned long target_load(int cpu, int type);
1525 static unsigned long cpu_avg_load_per_task(int cpu);
1526 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1528 #ifdef CONFIG_FAIR_GROUP_SCHED
1531 * Group load balancing.
1533 * We calculate a few balance domain wide aggregate numbers; load and weight.
1534 * Given the pictures below, and assuming each item has equal weight:
1545 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1546 * which equals 1/9-th of the total load.
1549 * The weight of this group on the selected cpus.
1552 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1556 * Part of the rq_weight contributed by tasks; all groups except B would
1560 static inline struct aggregate_struct *
1561 aggregate(struct task_group *tg, struct sched_domain *sd)
1563 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1566 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1569 * Iterate the full tree, calling @down when first entering a node and @up when
1570 * leaving it for the final time.
1573 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1574 struct sched_domain *sd)
1576 struct task_group *parent, *child;
1579 parent = &root_task_group;
1581 (*down)(parent, sd);
1582 list_for_each_entry_rcu(child, &parent->children, siblings) {
1592 parent = parent->parent;
1599 * Calculate the aggregate runqueue weight.
1602 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1604 unsigned long rq_weight = 0;
1605 unsigned long task_weight = 0;
1608 for_each_cpu_mask(i, sd->span) {
1609 rq_weight += tg->cfs_rq[i]->load.weight;
1610 task_weight += tg->cfs_rq[i]->task_weight;
1613 aggregate(tg, sd)->rq_weight = rq_weight;
1614 aggregate(tg, sd)->task_weight = task_weight;
1618 * Compute the weight of this group on the given cpus.
1621 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1623 unsigned long shares = 0;
1626 for_each_cpu_mask(i, sd->span)
1627 shares += tg->cfs_rq[i]->shares;
1629 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1630 shares = tg->shares;
1632 aggregate(tg, sd)->shares = shares;
1636 * Compute the load fraction assigned to this group, relies on the aggregate
1637 * weight and this group's parent's load, i.e. top-down.
1640 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1648 for_each_cpu_mask(i, sd->span)
1649 load += cpu_rq(i)->load.weight;
1652 load = aggregate(tg->parent, sd)->load;
1655 * shares is our weight in the parent's rq so
1656 * shares/parent->rq_weight gives our fraction of the load
1658 load *= aggregate(tg, sd)->shares;
1659 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1662 aggregate(tg, sd)->load = load;
1665 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1668 * Calculate and set the cpu's group shares.
1671 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1675 unsigned long shares;
1676 unsigned long rq_weight;
1681 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1684 * If there are currently no tasks on the cpu pretend there is one of
1685 * average load so that when a new task gets to run here it will not
1686 * get delayed by group starvation.
1690 rq_weight = NICE_0_LOAD;
1694 * \Sum shares * rq_weight
1695 * shares = -----------------------
1699 shares = aggregate(tg, sd)->shares * rq_weight;
1700 shares /= aggregate(tg, sd)->rq_weight + 1;
1703 * record the actual number of shares, not the boosted amount.
1705 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1707 if (shares < MIN_SHARES)
1708 shares = MIN_SHARES;
1709 else if (shares > MAX_SHARES)
1710 shares = MAX_SHARES;
1712 __set_se_shares(tg->se[tcpu], shares);
1716 * Re-adjust the weights on the cpu the task came from and on the cpu the
1720 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1723 unsigned long shares;
1725 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1727 __update_group_shares_cpu(tg, sd, scpu);
1728 __update_group_shares_cpu(tg, sd, dcpu);
1731 * ensure we never loose shares due to rounding errors in the
1732 * above redistribution.
1734 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1736 tg->cfs_rq[dcpu]->shares += shares;
1740 * Because changing a group's shares changes the weight of the super-group
1741 * we need to walk up the tree and change all shares until we hit the root.
1744 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1748 __move_group_shares(tg, sd, scpu, dcpu);
1754 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1756 unsigned long shares = aggregate(tg, sd)->shares;
1759 for_each_cpu_mask(i, sd->span) {
1760 struct rq *rq = cpu_rq(i);
1761 unsigned long flags;
1763 spin_lock_irqsave(&rq->lock, flags);
1764 __update_group_shares_cpu(tg, sd, i);
1765 spin_unlock_irqrestore(&rq->lock, flags);
1768 aggregate_group_shares(tg, sd);
1771 * ensure we never loose shares due to rounding errors in the
1772 * above redistribution.
1774 shares -= aggregate(tg, sd)->shares;
1776 tg->cfs_rq[sd->first_cpu]->shares += shares;
1777 aggregate(tg, sd)->shares += shares;
1782 * Calculate the accumulative weight and recursive load of each task group
1783 * while walking down the tree.
1786 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1788 aggregate_group_weight(tg, sd);
1789 aggregate_group_shares(tg, sd);
1790 aggregate_group_load(tg, sd);
1794 * Rebalance the cpu shares while walking back up the tree.
1797 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1799 aggregate_group_set_shares(tg, sd);
1802 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1804 static void __init init_aggregate(void)
1808 for_each_possible_cpu(i)
1809 spin_lock_init(&per_cpu(aggregate_lock, i));
1812 static int get_aggregate(struct sched_domain *sd)
1814 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1817 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1821 static void put_aggregate(struct sched_domain *sd)
1823 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1826 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1828 cfs_rq->shares = shares;
1833 static inline void init_aggregate(void)
1837 static inline int get_aggregate(struct sched_domain *sd)
1842 static inline void put_aggregate(struct sched_domain *sd)
1847 #else /* CONFIG_SMP */
1849 #ifdef CONFIG_FAIR_GROUP_SCHED
1850 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1855 #endif /* CONFIG_SMP */
1857 #include "sched_stats.h"
1858 #include "sched_idletask.c"
1859 #include "sched_fair.c"
1860 #include "sched_rt.c"
1861 #ifdef CONFIG_SCHED_DEBUG
1862 # include "sched_debug.c"
1865 #define sched_class_highest (&rt_sched_class)
1867 static void inc_nr_running(struct rq *rq)
1872 static void dec_nr_running(struct rq *rq)
1877 static void set_load_weight(struct task_struct *p)
1879 if (task_has_rt_policy(p)) {
1880 p->se.load.weight = prio_to_weight[0] * 2;
1881 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1886 * SCHED_IDLE tasks get minimal weight:
1888 if (p->policy == SCHED_IDLE) {
1889 p->se.load.weight = WEIGHT_IDLEPRIO;
1890 p->se.load.inv_weight = WMULT_IDLEPRIO;
1894 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1895 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1898 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1900 sched_info_queued(p);
1901 p->sched_class->enqueue_task(rq, p, wakeup);
1905 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1907 p->sched_class->dequeue_task(rq, p, sleep);
1912 * __normal_prio - return the priority that is based on the static prio
1914 static inline int __normal_prio(struct task_struct *p)
1916 return p->static_prio;
1920 * Calculate the expected normal priority: i.e. priority
1921 * without taking RT-inheritance into account. Might be
1922 * boosted by interactivity modifiers. Changes upon fork,
1923 * setprio syscalls, and whenever the interactivity
1924 * estimator recalculates.
1926 static inline int normal_prio(struct task_struct *p)
1930 if (task_has_rt_policy(p))
1931 prio = MAX_RT_PRIO-1 - p->rt_priority;
1933 prio = __normal_prio(p);
1938 * Calculate the current priority, i.e. the priority
1939 * taken into account by the scheduler. This value might
1940 * be boosted by RT tasks, or might be boosted by
1941 * interactivity modifiers. Will be RT if the task got
1942 * RT-boosted. If not then it returns p->normal_prio.
1944 static int effective_prio(struct task_struct *p)
1946 p->normal_prio = normal_prio(p);
1948 * If we are RT tasks or we were boosted to RT priority,
1949 * keep the priority unchanged. Otherwise, update priority
1950 * to the normal priority:
1952 if (!rt_prio(p->prio))
1953 return p->normal_prio;
1958 * activate_task - move a task to the runqueue.
1960 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1962 if (task_contributes_to_load(p))
1963 rq->nr_uninterruptible--;
1965 enqueue_task(rq, p, wakeup);
1970 * deactivate_task - remove a task from the runqueue.
1972 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1974 if (task_contributes_to_load(p))
1975 rq->nr_uninterruptible++;
1977 dequeue_task(rq, p, sleep);
1982 * task_curr - is this task currently executing on a CPU?
1983 * @p: the task in question.
1985 inline int task_curr(const struct task_struct *p)
1987 return cpu_curr(task_cpu(p)) == p;
1990 /* Used instead of source_load when we know the type == 0 */
1991 unsigned long weighted_cpuload(const int cpu)
1993 return cpu_rq(cpu)->load.weight;
1996 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1998 set_task_rq(p, cpu);
2001 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2002 * successfuly executed on another CPU. We must ensure that updates of
2003 * per-task data have been completed by this moment.
2006 task_thread_info(p)->cpu = cpu;
2010 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
2011 const struct sched_class *prev_class,
2012 int oldprio, int running)
2014 if (prev_class != p->sched_class) {
2015 if (prev_class->switched_from)
2016 prev_class->switched_from(rq, p, running);
2017 p->sched_class->switched_to(rq, p, running);
2019 p->sched_class->prio_changed(rq, p, oldprio, running);
2025 * Is this task likely cache-hot:
2028 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2033 * Buddy candidates are cache hot:
2035 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2038 if (p->sched_class != &fair_sched_class)
2041 if (sysctl_sched_migration_cost == -1)
2043 if (sysctl_sched_migration_cost == 0)
2046 delta = now - p->se.exec_start;
2048 return delta < (s64)sysctl_sched_migration_cost;
2052 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2054 int old_cpu = task_cpu(p);
2055 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2056 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2057 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2060 clock_offset = old_rq->clock - new_rq->clock;
2062 #ifdef CONFIG_SCHEDSTATS
2063 if (p->se.wait_start)
2064 p->se.wait_start -= clock_offset;
2065 if (p->se.sleep_start)
2066 p->se.sleep_start -= clock_offset;
2067 if (p->se.block_start)
2068 p->se.block_start -= clock_offset;
2069 if (old_cpu != new_cpu) {
2070 schedstat_inc(p, se.nr_migrations);
2071 if (task_hot(p, old_rq->clock, NULL))
2072 schedstat_inc(p, se.nr_forced2_migrations);
2075 p->se.vruntime -= old_cfsrq->min_vruntime -
2076 new_cfsrq->min_vruntime;
2078 __set_task_cpu(p, new_cpu);
2081 struct migration_req {
2082 struct list_head list;
2084 struct task_struct *task;
2087 struct completion done;
2091 * The task's runqueue lock must be held.
2092 * Returns true if you have to wait for migration thread.
2095 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2097 struct rq *rq = task_rq(p);
2100 * If the task is not on a runqueue (and not running), then
2101 * it is sufficient to simply update the task's cpu field.
2103 if (!p->se.on_rq && !task_running(rq, p)) {
2104 set_task_cpu(p, dest_cpu);
2108 init_completion(&req->done);
2110 req->dest_cpu = dest_cpu;
2111 list_add(&req->list, &rq->migration_queue);
2117 * wait_task_inactive - wait for a thread to unschedule.
2119 * The caller must ensure that the task *will* unschedule sometime soon,
2120 * else this function might spin for a *long* time. This function can't
2121 * be called with interrupts off, or it may introduce deadlock with
2122 * smp_call_function() if an IPI is sent by the same process we are
2123 * waiting to become inactive.
2125 void wait_task_inactive(struct task_struct *p)
2127 unsigned long flags;
2133 * We do the initial early heuristics without holding
2134 * any task-queue locks at all. We'll only try to get
2135 * the runqueue lock when things look like they will
2141 * If the task is actively running on another CPU
2142 * still, just relax and busy-wait without holding
2145 * NOTE! Since we don't hold any locks, it's not
2146 * even sure that "rq" stays as the right runqueue!
2147 * But we don't care, since "task_running()" will
2148 * return false if the runqueue has changed and p
2149 * is actually now running somewhere else!
2151 while (task_running(rq, p))
2155 * Ok, time to look more closely! We need the rq
2156 * lock now, to be *sure*. If we're wrong, we'll
2157 * just go back and repeat.
2159 rq = task_rq_lock(p, &flags);
2160 running = task_running(rq, p);
2161 on_rq = p->se.on_rq;
2162 task_rq_unlock(rq, &flags);
2165 * Was it really running after all now that we
2166 * checked with the proper locks actually held?
2168 * Oops. Go back and try again..
2170 if (unlikely(running)) {
2176 * It's not enough that it's not actively running,
2177 * it must be off the runqueue _entirely_, and not
2180 * So if it wa still runnable (but just not actively
2181 * running right now), it's preempted, and we should
2182 * yield - it could be a while.
2184 if (unlikely(on_rq)) {
2185 schedule_timeout_uninterruptible(1);
2190 * Ahh, all good. It wasn't running, and it wasn't
2191 * runnable, which means that it will never become
2192 * running in the future either. We're all done!
2199 * kick_process - kick a running thread to enter/exit the kernel
2200 * @p: the to-be-kicked thread
2202 * Cause a process which is running on another CPU to enter
2203 * kernel-mode, without any delay. (to get signals handled.)
2205 * NOTE: this function doesnt have to take the runqueue lock,
2206 * because all it wants to ensure is that the remote task enters
2207 * the kernel. If the IPI races and the task has been migrated
2208 * to another CPU then no harm is done and the purpose has been
2211 void kick_process(struct task_struct *p)
2217 if ((cpu != smp_processor_id()) && task_curr(p))
2218 smp_send_reschedule(cpu);
2223 * Return a low guess at the load of a migration-source cpu weighted
2224 * according to the scheduling class and "nice" value.
2226 * We want to under-estimate the load of migration sources, to
2227 * balance conservatively.
2229 static unsigned long source_load(int cpu, int type)
2231 struct rq *rq = cpu_rq(cpu);
2232 unsigned long total = weighted_cpuload(cpu);
2237 return min(rq->cpu_load[type-1], total);
2241 * Return a high guess at the load of a migration-target cpu weighted
2242 * according to the scheduling class and "nice" value.
2244 static unsigned long target_load(int cpu, int type)
2246 struct rq *rq = cpu_rq(cpu);
2247 unsigned long total = weighted_cpuload(cpu);
2252 return max(rq->cpu_load[type-1], total);
2256 * Return the average load per task on the cpu's run queue
2258 static unsigned long cpu_avg_load_per_task(int cpu)
2260 struct rq *rq = cpu_rq(cpu);
2261 unsigned long total = weighted_cpuload(cpu);
2262 unsigned long n = rq->nr_running;
2264 return n ? total / n : SCHED_LOAD_SCALE;
2268 * find_idlest_group finds and returns the least busy CPU group within the
2271 static struct sched_group *
2272 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2274 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2275 unsigned long min_load = ULONG_MAX, this_load = 0;
2276 int load_idx = sd->forkexec_idx;
2277 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2280 unsigned long load, avg_load;
2284 /* Skip over this group if it has no CPUs allowed */
2285 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2288 local_group = cpu_isset(this_cpu, group->cpumask);
2290 /* Tally up the load of all CPUs in the group */
2293 for_each_cpu_mask(i, group->cpumask) {
2294 /* Bias balancing toward cpus of our domain */
2296 load = source_load(i, load_idx);
2298 load = target_load(i, load_idx);
2303 /* Adjust by relative CPU power of the group */
2304 avg_load = sg_div_cpu_power(group,
2305 avg_load * SCHED_LOAD_SCALE);
2308 this_load = avg_load;
2310 } else if (avg_load < min_load) {
2311 min_load = avg_load;
2314 } while (group = group->next, group != sd->groups);
2316 if (!idlest || 100*this_load < imbalance*min_load)
2322 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2325 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2328 unsigned long load, min_load = ULONG_MAX;
2332 /* Traverse only the allowed CPUs */
2333 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2335 for_each_cpu_mask(i, *tmp) {
2336 load = weighted_cpuload(i);
2338 if (load < min_load || (load == min_load && i == this_cpu)) {
2348 * sched_balance_self: balance the current task (running on cpu) in domains
2349 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2352 * Balance, ie. select the least loaded group.
2354 * Returns the target CPU number, or the same CPU if no balancing is needed.
2356 * preempt must be disabled.
2358 static int sched_balance_self(int cpu, int flag)
2360 struct task_struct *t = current;
2361 struct sched_domain *tmp, *sd = NULL;
2363 for_each_domain(cpu, tmp) {
2365 * If power savings logic is enabled for a domain, stop there.
2367 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2369 if (tmp->flags & flag)
2374 cpumask_t span, tmpmask;
2375 struct sched_group *group;
2376 int new_cpu, weight;
2378 if (!(sd->flags & flag)) {
2384 group = find_idlest_group(sd, t, cpu);
2390 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2391 if (new_cpu == -1 || new_cpu == cpu) {
2392 /* Now try balancing at a lower domain level of cpu */
2397 /* Now try balancing at a lower domain level of new_cpu */
2400 weight = cpus_weight(span);
2401 for_each_domain(cpu, tmp) {
2402 if (weight <= cpus_weight(tmp->span))
2404 if (tmp->flags & flag)
2407 /* while loop will break here if sd == NULL */
2413 #endif /* CONFIG_SMP */
2415 #ifdef CONFIG_CONTEXT_SWITCH_TRACER
2417 void ftrace_task(struct task_struct *p, void *__tr, void *__data)
2421 * trace timeline tree
2423 __trace_special(__tr, __data,
2424 p->pid, p->se.vruntime, p->se.sum_exec_runtime);
2427 * trace balance metrics
2429 __trace_special(__tr, __data,
2430 p->pid, p->se.avg_overlap, 0);
2434 void ftrace_all_fair_tasks(void *__rq, void *__tr, void *__data)
2436 struct task_struct *p;
2437 struct sched_entity *se;
2438 struct rb_node *curr;
2439 struct rq *rq = __rq;
2442 p = task_of(rq->cfs.curr);
2443 ftrace_task(p, __tr, __data);
2446 p = task_of(rq->cfs.next);
2447 ftrace_task(p, __tr, __data);
2450 for (curr = first_fair(&rq->cfs); curr; curr = rb_next(curr)) {
2451 se = rb_entry(curr, struct sched_entity, run_node);
2452 if (!entity_is_task(se))
2456 ftrace_task(p, __tr, __data);
2463 * try_to_wake_up - wake up a thread
2464 * @p: the to-be-woken-up thread
2465 * @state: the mask of task states that can be woken
2466 * @sync: do a synchronous wakeup?
2468 * Put it on the run-queue if it's not already there. The "current"
2469 * thread is always on the run-queue (except when the actual
2470 * re-schedule is in progress), and as such you're allowed to do
2471 * the simpler "current->state = TASK_RUNNING" to mark yourself
2472 * runnable without the overhead of this.
2474 * returns failure only if the task is already active.
2476 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2478 int cpu, orig_cpu, this_cpu, success = 0;
2479 unsigned long flags;
2483 if (!sched_feat(SYNC_WAKEUPS))
2487 rq = task_rq_lock(p, &flags);
2488 old_state = p->state;
2489 if (!(old_state & state))
2497 this_cpu = smp_processor_id();
2500 if (unlikely(task_running(rq, p)))
2503 cpu = p->sched_class->select_task_rq(p, sync);
2504 if (cpu != orig_cpu) {
2505 set_task_cpu(p, cpu);
2506 task_rq_unlock(rq, &flags);
2507 /* might preempt at this point */
2508 rq = task_rq_lock(p, &flags);
2509 old_state = p->state;
2510 if (!(old_state & state))
2515 this_cpu = smp_processor_id();
2519 #ifdef CONFIG_SCHEDSTATS
2520 schedstat_inc(rq, ttwu_count);
2521 if (cpu == this_cpu)
2522 schedstat_inc(rq, ttwu_local);
2524 struct sched_domain *sd;
2525 for_each_domain(this_cpu, sd) {
2526 if (cpu_isset(cpu, sd->span)) {
2527 schedstat_inc(sd, ttwu_wake_remote);
2535 #endif /* CONFIG_SMP */
2536 schedstat_inc(p, se.nr_wakeups);
2538 schedstat_inc(p, se.nr_wakeups_sync);
2539 if (orig_cpu != cpu)
2540 schedstat_inc(p, se.nr_wakeups_migrate);
2541 if (cpu == this_cpu)
2542 schedstat_inc(p, se.nr_wakeups_local);
2544 schedstat_inc(p, se.nr_wakeups_remote);
2545 update_rq_clock(rq);
2546 activate_task(rq, p, 1);
2550 ftrace_wake_up_task(rq, p, rq->curr);
2551 check_preempt_curr(rq, p);
2553 p->state = TASK_RUNNING;
2555 if (p->sched_class->task_wake_up)
2556 p->sched_class->task_wake_up(rq, p);
2559 task_rq_unlock(rq, &flags);
2564 int wake_up_process(struct task_struct *p)
2566 return try_to_wake_up(p, TASK_ALL, 0);
2568 EXPORT_SYMBOL(wake_up_process);
2570 int wake_up_state(struct task_struct *p, unsigned int state)
2572 return try_to_wake_up(p, state, 0);
2576 * Perform scheduler related setup for a newly forked process p.
2577 * p is forked by current.
2579 * __sched_fork() is basic setup used by init_idle() too:
2581 static void __sched_fork(struct task_struct *p)
2583 p->se.exec_start = 0;
2584 p->se.sum_exec_runtime = 0;
2585 p->se.prev_sum_exec_runtime = 0;
2586 p->se.last_wakeup = 0;
2587 p->se.avg_overlap = 0;
2589 #ifdef CONFIG_SCHEDSTATS
2590 p->se.wait_start = 0;
2591 p->se.sum_sleep_runtime = 0;
2592 p->se.sleep_start = 0;
2593 p->se.block_start = 0;
2594 p->se.sleep_max = 0;
2595 p->se.block_max = 0;
2597 p->se.slice_max = 0;
2601 INIT_LIST_HEAD(&p->rt.run_list);
2603 INIT_LIST_HEAD(&p->se.group_node);
2605 #ifdef CONFIG_PREEMPT_NOTIFIERS
2606 INIT_HLIST_HEAD(&p->preempt_notifiers);
2610 * We mark the process as running here, but have not actually
2611 * inserted it onto the runqueue yet. This guarantees that
2612 * nobody will actually run it, and a signal or other external
2613 * event cannot wake it up and insert it on the runqueue either.
2615 p->state = TASK_RUNNING;
2619 * fork()/clone()-time setup:
2621 void sched_fork(struct task_struct *p, int clone_flags)
2623 int cpu = get_cpu();
2628 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2630 set_task_cpu(p, cpu);
2633 * Make sure we do not leak PI boosting priority to the child:
2635 p->prio = current->normal_prio;
2636 if (!rt_prio(p->prio))
2637 p->sched_class = &fair_sched_class;
2639 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2640 if (likely(sched_info_on()))
2641 memset(&p->sched_info, 0, sizeof(p->sched_info));
2643 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2646 #ifdef CONFIG_PREEMPT
2647 /* Want to start with kernel preemption disabled. */
2648 task_thread_info(p)->preempt_count = 1;
2654 * wake_up_new_task - wake up a newly created task for the first time.
2656 * This function will do some initial scheduler statistics housekeeping
2657 * that must be done for every newly created context, then puts the task
2658 * on the runqueue and wakes it.
2660 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2662 unsigned long flags;
2665 rq = task_rq_lock(p, &flags);
2666 BUG_ON(p->state != TASK_RUNNING);
2667 update_rq_clock(rq);
2669 p->prio = effective_prio(p);
2671 if (!p->sched_class->task_new || !current->se.on_rq) {
2672 activate_task(rq, p, 0);
2675 * Let the scheduling class do new task startup
2676 * management (if any):
2678 p->sched_class->task_new(rq, p);
2681 ftrace_wake_up_task(rq, p, rq->curr);
2682 check_preempt_curr(rq, p);
2684 if (p->sched_class->task_wake_up)
2685 p->sched_class->task_wake_up(rq, p);
2687 task_rq_unlock(rq, &flags);
2690 #ifdef CONFIG_PREEMPT_NOTIFIERS
2693 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2694 * @notifier: notifier struct to register
2696 void preempt_notifier_register(struct preempt_notifier *notifier)
2698 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2700 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2703 * preempt_notifier_unregister - no longer interested in preemption notifications
2704 * @notifier: notifier struct to unregister
2706 * This is safe to call from within a preemption notifier.
2708 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2710 hlist_del(¬ifier->link);
2712 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2714 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2716 struct preempt_notifier *notifier;
2717 struct hlist_node *node;
2719 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2720 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2724 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2725 struct task_struct *next)
2727 struct preempt_notifier *notifier;
2728 struct hlist_node *node;
2730 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2731 notifier->ops->sched_out(notifier, next);
2736 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2741 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2742 struct task_struct *next)
2749 * prepare_task_switch - prepare to switch tasks
2750 * @rq: the runqueue preparing to switch
2751 * @prev: the current task that is being switched out
2752 * @next: the task we are going to switch to.
2754 * This is called with the rq lock held and interrupts off. It must
2755 * be paired with a subsequent finish_task_switch after the context
2758 * prepare_task_switch sets up locking and calls architecture specific
2762 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2763 struct task_struct *next)
2765 fire_sched_out_preempt_notifiers(prev, next);
2766 prepare_lock_switch(rq, next);
2767 prepare_arch_switch(next);
2771 * finish_task_switch - clean up after a task-switch
2772 * @rq: runqueue associated with task-switch
2773 * @prev: the thread we just switched away from.
2775 * finish_task_switch must be called after the context switch, paired
2776 * with a prepare_task_switch call before the context switch.
2777 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2778 * and do any other architecture-specific cleanup actions.
2780 * Note that we may have delayed dropping an mm in context_switch(). If
2781 * so, we finish that here outside of the runqueue lock. (Doing it
2782 * with the lock held can cause deadlocks; see schedule() for
2785 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2786 __releases(rq->lock)
2788 struct mm_struct *mm = rq->prev_mm;
2794 * A task struct has one reference for the use as "current".
2795 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2796 * schedule one last time. The schedule call will never return, and
2797 * the scheduled task must drop that reference.
2798 * The test for TASK_DEAD must occur while the runqueue locks are
2799 * still held, otherwise prev could be scheduled on another cpu, die
2800 * there before we look at prev->state, and then the reference would
2802 * Manfred Spraul <manfred@colorfullife.com>
2804 prev_state = prev->state;
2805 finish_arch_switch(prev);
2806 finish_lock_switch(rq, prev);
2808 if (current->sched_class->post_schedule)
2809 current->sched_class->post_schedule(rq);
2812 fire_sched_in_preempt_notifiers(current);
2815 if (unlikely(prev_state == TASK_DEAD)) {
2817 * Remove function-return probe instances associated with this
2818 * task and put them back on the free list.
2820 kprobe_flush_task(prev);
2821 put_task_struct(prev);
2826 * schedule_tail - first thing a freshly forked thread must call.
2827 * @prev: the thread we just switched away from.
2829 asmlinkage void schedule_tail(struct task_struct *prev)
2830 __releases(rq->lock)
2832 struct rq *rq = this_rq();
2834 finish_task_switch(rq, prev);
2835 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2836 /* In this case, finish_task_switch does not reenable preemption */
2839 if (current->set_child_tid)
2840 put_user(task_pid_vnr(current), current->set_child_tid);
2844 * context_switch - switch to the new MM and the new
2845 * thread's register state.
2848 context_switch(struct rq *rq, struct task_struct *prev,
2849 struct task_struct *next)
2851 struct mm_struct *mm, *oldmm;
2853 prepare_task_switch(rq, prev, next);
2854 ftrace_ctx_switch(rq, prev, next);
2856 oldmm = prev->active_mm;
2858 * For paravirt, this is coupled with an exit in switch_to to
2859 * combine the page table reload and the switch backend into
2862 arch_enter_lazy_cpu_mode();
2864 if (unlikely(!mm)) {
2865 next->active_mm = oldmm;
2866 atomic_inc(&oldmm->mm_count);
2867 enter_lazy_tlb(oldmm, next);
2869 switch_mm(oldmm, mm, next);
2871 if (unlikely(!prev->mm)) {
2872 prev->active_mm = NULL;
2873 rq->prev_mm = oldmm;
2876 * Since the runqueue lock will be released by the next
2877 * task (which is an invalid locking op but in the case
2878 * of the scheduler it's an obvious special-case), so we
2879 * do an early lockdep release here:
2881 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2882 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2885 /* Here we just switch the register state and the stack. */
2886 switch_to(prev, next, prev);
2890 * this_rq must be evaluated again because prev may have moved
2891 * CPUs since it called schedule(), thus the 'rq' on its stack
2892 * frame will be invalid.
2894 finish_task_switch(this_rq(), prev);
2898 * nr_running, nr_uninterruptible and nr_context_switches:
2900 * externally visible scheduler statistics: current number of runnable
2901 * threads, current number of uninterruptible-sleeping threads, total
2902 * number of context switches performed since bootup.
2904 unsigned long nr_running(void)
2906 unsigned long i, sum = 0;
2908 for_each_online_cpu(i)
2909 sum += cpu_rq(i)->nr_running;
2914 unsigned long nr_uninterruptible(void)
2916 unsigned long i, sum = 0;
2918 for_each_possible_cpu(i)
2919 sum += cpu_rq(i)->nr_uninterruptible;
2922 * Since we read the counters lockless, it might be slightly
2923 * inaccurate. Do not allow it to go below zero though:
2925 if (unlikely((long)sum < 0))
2931 unsigned long long nr_context_switches(void)
2934 unsigned long long sum = 0;
2936 for_each_possible_cpu(i)
2937 sum += cpu_rq(i)->nr_switches;
2942 unsigned long nr_iowait(void)
2944 unsigned long i, sum = 0;
2946 for_each_possible_cpu(i)
2947 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2952 unsigned long nr_active(void)
2954 unsigned long i, running = 0, uninterruptible = 0;
2956 for_each_online_cpu(i) {
2957 running += cpu_rq(i)->nr_running;
2958 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2961 if (unlikely((long)uninterruptible < 0))
2962 uninterruptible = 0;
2964 return running + uninterruptible;
2968 * Update rq->cpu_load[] statistics. This function is usually called every
2969 * scheduler tick (TICK_NSEC).
2971 static void update_cpu_load(struct rq *this_rq)
2973 unsigned long this_load = this_rq->load.weight;
2976 this_rq->nr_load_updates++;
2978 /* Update our load: */
2979 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2980 unsigned long old_load, new_load;
2982 /* scale is effectively 1 << i now, and >> i divides by scale */
2984 old_load = this_rq->cpu_load[i];
2985 new_load = this_load;
2987 * Round up the averaging division if load is increasing. This
2988 * prevents us from getting stuck on 9 if the load is 10, for
2991 if (new_load > old_load)
2992 new_load += scale-1;
2993 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
3000 * double_rq_lock - safely lock two runqueues
3002 * Note this does not disable interrupts like task_rq_lock,
3003 * you need to do so manually before calling.
3005 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
3006 __acquires(rq1->lock)
3007 __acquires(rq2->lock)
3009 BUG_ON(!irqs_disabled());
3011 spin_lock(&rq1->lock);
3012 __acquire(rq2->lock); /* Fake it out ;) */
3015 spin_lock(&rq1->lock);
3016 spin_lock(&rq2->lock);
3018 spin_lock(&rq2->lock);
3019 spin_lock(&rq1->lock);
3022 update_rq_clock(rq1);
3023 update_rq_clock(rq2);
3027 * double_rq_unlock - safely unlock two runqueues
3029 * Note this does not restore interrupts like task_rq_unlock,
3030 * you need to do so manually after calling.
3032 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
3033 __releases(rq1->lock)
3034 __releases(rq2->lock)
3036 spin_unlock(&rq1->lock);
3038 spin_unlock(&rq2->lock);
3040 __release(rq2->lock);
3044 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3046 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
3047 __releases(this_rq->lock)
3048 __acquires(busiest->lock)
3049 __acquires(this_rq->lock)
3053 if (unlikely(!irqs_disabled())) {
3054 /* printk() doesn't work good under rq->lock */
3055 spin_unlock(&this_rq->lock);
3058 if (unlikely(!spin_trylock(&busiest->lock))) {
3059 if (busiest < this_rq) {
3060 spin_unlock(&this_rq->lock);
3061 spin_lock(&busiest->lock);
3062 spin_lock(&this_rq->lock);
3065 spin_lock(&busiest->lock);
3071 * If dest_cpu is allowed for this process, migrate the task to it.
3072 * This is accomplished by forcing the cpu_allowed mask to only
3073 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3074 * the cpu_allowed mask is restored.
3076 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3078 struct migration_req req;
3079 unsigned long flags;
3082 rq = task_rq_lock(p, &flags);
3083 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3084 || unlikely(cpu_is_offline(dest_cpu)))
3087 /* force the process onto the specified CPU */
3088 if (migrate_task(p, dest_cpu, &req)) {
3089 /* Need to wait for migration thread (might exit: take ref). */
3090 struct task_struct *mt = rq->migration_thread;
3092 get_task_struct(mt);
3093 task_rq_unlock(rq, &flags);
3094 wake_up_process(mt);
3095 put_task_struct(mt);
3096 wait_for_completion(&req.done);
3101 task_rq_unlock(rq, &flags);
3105 * sched_exec - execve() is a valuable balancing opportunity, because at
3106 * this point the task has the smallest effective memory and cache footprint.
3108 void sched_exec(void)
3110 int new_cpu, this_cpu = get_cpu();
3111 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3113 if (new_cpu != this_cpu)
3114 sched_migrate_task(current, new_cpu);
3118 * pull_task - move a task from a remote runqueue to the local runqueue.
3119 * Both runqueues must be locked.
3121 static void pull_task(struct rq *src_rq, struct task_struct *p,
3122 struct rq *this_rq, int this_cpu)
3124 deactivate_task(src_rq, p, 0);
3125 set_task_cpu(p, this_cpu);
3126 activate_task(this_rq, p, 0);
3128 * Note that idle threads have a prio of MAX_PRIO, for this test
3129 * to be always true for them.
3131 check_preempt_curr(this_rq, p);
3135 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3138 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3139 struct sched_domain *sd, enum cpu_idle_type idle,
3143 * We do not migrate tasks that are:
3144 * 1) running (obviously), or
3145 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3146 * 3) are cache-hot on their current CPU.
3148 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3149 schedstat_inc(p, se.nr_failed_migrations_affine);
3154 if (task_running(rq, p)) {
3155 schedstat_inc(p, se.nr_failed_migrations_running);
3160 * Aggressive migration if:
3161 * 1) task is cache cold, or
3162 * 2) too many balance attempts have failed.
3165 if (!task_hot(p, rq->clock, sd) ||
3166 sd->nr_balance_failed > sd->cache_nice_tries) {
3167 #ifdef CONFIG_SCHEDSTATS
3168 if (task_hot(p, rq->clock, sd)) {
3169 schedstat_inc(sd, lb_hot_gained[idle]);
3170 schedstat_inc(p, se.nr_forced_migrations);
3176 if (task_hot(p, rq->clock, sd)) {
3177 schedstat_inc(p, se.nr_failed_migrations_hot);
3183 static unsigned long
3184 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3185 unsigned long max_load_move, struct sched_domain *sd,
3186 enum cpu_idle_type idle, int *all_pinned,
3187 int *this_best_prio, struct rq_iterator *iterator)
3189 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3190 struct task_struct *p;
3191 long rem_load_move = max_load_move;
3193 if (max_load_move == 0)
3199 * Start the load-balancing iterator:
3201 p = iterator->start(iterator->arg);
3203 if (!p || loops++ > sysctl_sched_nr_migrate)
3206 * To help distribute high priority tasks across CPUs we don't
3207 * skip a task if it will be the highest priority task (i.e. smallest
3208 * prio value) on its new queue regardless of its load weight
3210 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3211 SCHED_LOAD_SCALE_FUZZ;
3212 if ((skip_for_load && p->prio >= *this_best_prio) ||
3213 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3214 p = iterator->next(iterator->arg);
3218 pull_task(busiest, p, this_rq, this_cpu);
3220 rem_load_move -= p->se.load.weight;
3223 * We only want to steal up to the prescribed amount of weighted load.
3225 if (rem_load_move > 0) {
3226 if (p->prio < *this_best_prio)
3227 *this_best_prio = p->prio;
3228 p = iterator->next(iterator->arg);
3233 * Right now, this is one of only two places pull_task() is called,
3234 * so we can safely collect pull_task() stats here rather than
3235 * inside pull_task().
3237 schedstat_add(sd, lb_gained[idle], pulled);
3240 *all_pinned = pinned;
3242 return max_load_move - rem_load_move;
3246 * move_tasks tries to move up to max_load_move weighted load from busiest to
3247 * this_rq, as part of a balancing operation within domain "sd".
3248 * Returns 1 if successful and 0 otherwise.
3250 * Called with both runqueues locked.
3252 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3253 unsigned long max_load_move,
3254 struct sched_domain *sd, enum cpu_idle_type idle,
3257 const struct sched_class *class = sched_class_highest;
3258 unsigned long total_load_moved = 0;
3259 int this_best_prio = this_rq->curr->prio;
3263 class->load_balance(this_rq, this_cpu, busiest,
3264 max_load_move - total_load_moved,
3265 sd, idle, all_pinned, &this_best_prio);
3266 class = class->next;
3267 } while (class && max_load_move > total_load_moved);
3269 return total_load_moved > 0;
3273 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3274 struct sched_domain *sd, enum cpu_idle_type idle,
3275 struct rq_iterator *iterator)
3277 struct task_struct *p = iterator->start(iterator->arg);
3281 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3282 pull_task(busiest, p, this_rq, this_cpu);
3284 * Right now, this is only the second place pull_task()
3285 * is called, so we can safely collect pull_task()
3286 * stats here rather than inside pull_task().
3288 schedstat_inc(sd, lb_gained[idle]);
3292 p = iterator->next(iterator->arg);
3299 * move_one_task tries to move exactly one task from busiest to this_rq, as
3300 * part of active balancing operations within "domain".
3301 * Returns 1 if successful and 0 otherwise.
3303 * Called with both runqueues locked.
3305 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3306 struct sched_domain *sd, enum cpu_idle_type idle)
3308 const struct sched_class *class;
3310 for (class = sched_class_highest; class; class = class->next)
3311 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3318 * find_busiest_group finds and returns the busiest CPU group within the
3319 * domain. It calculates and returns the amount of weighted load which
3320 * should be moved to restore balance via the imbalance parameter.
3322 static struct sched_group *
3323 find_busiest_group(struct sched_domain *sd, int this_cpu,
3324 unsigned long *imbalance, enum cpu_idle_type idle,
3325 int *sd_idle, const cpumask_t *cpus, int *balance)
3327 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3328 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3329 unsigned long max_pull;
3330 unsigned long busiest_load_per_task, busiest_nr_running;
3331 unsigned long this_load_per_task, this_nr_running;
3332 int load_idx, group_imb = 0;
3333 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3334 int power_savings_balance = 1;
3335 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3336 unsigned long min_nr_running = ULONG_MAX;
3337 struct sched_group *group_min = NULL, *group_leader = NULL;
3340 max_load = this_load = total_load = total_pwr = 0;
3341 busiest_load_per_task = busiest_nr_running = 0;
3342 this_load_per_task = this_nr_running = 0;
3343 if (idle == CPU_NOT_IDLE)
3344 load_idx = sd->busy_idx;
3345 else if (idle == CPU_NEWLY_IDLE)
3346 load_idx = sd->newidle_idx;
3348 load_idx = sd->idle_idx;
3351 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3354 int __group_imb = 0;
3355 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3356 unsigned long sum_nr_running, sum_weighted_load;
3358 local_group = cpu_isset(this_cpu, group->cpumask);
3361 balance_cpu = first_cpu(group->cpumask);
3363 /* Tally up the load of all CPUs in the group */
3364 sum_weighted_load = sum_nr_running = avg_load = 0;
3366 min_cpu_load = ~0UL;
3368 for_each_cpu_mask(i, group->cpumask) {
3371 if (!cpu_isset(i, *cpus))
3376 if (*sd_idle && rq->nr_running)
3379 /* Bias balancing toward cpus of our domain */
3381 if (idle_cpu(i) && !first_idle_cpu) {
3386 load = target_load(i, load_idx);
3388 load = source_load(i, load_idx);
3389 if (load > max_cpu_load)
3390 max_cpu_load = load;
3391 if (min_cpu_load > load)
3392 min_cpu_load = load;
3396 sum_nr_running += rq->nr_running;
3397 sum_weighted_load += weighted_cpuload(i);
3401 * First idle cpu or the first cpu(busiest) in this sched group
3402 * is eligible for doing load balancing at this and above
3403 * domains. In the newly idle case, we will allow all the cpu's
3404 * to do the newly idle load balance.
3406 if (idle != CPU_NEWLY_IDLE && local_group &&
3407 balance_cpu != this_cpu && balance) {
3412 total_load += avg_load;
3413 total_pwr += group->__cpu_power;
3415 /* Adjust by relative CPU power of the group */
3416 avg_load = sg_div_cpu_power(group,
3417 avg_load * SCHED_LOAD_SCALE);
3419 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3422 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3425 this_load = avg_load;
3427 this_nr_running = sum_nr_running;
3428 this_load_per_task = sum_weighted_load;
3429 } else if (avg_load > max_load &&
3430 (sum_nr_running > group_capacity || __group_imb)) {
3431 max_load = avg_load;
3433 busiest_nr_running = sum_nr_running;
3434 busiest_load_per_task = sum_weighted_load;
3435 group_imb = __group_imb;
3438 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3440 * Busy processors will not participate in power savings
3443 if (idle == CPU_NOT_IDLE ||
3444 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3448 * If the local group is idle or completely loaded
3449 * no need to do power savings balance at this domain
3451 if (local_group && (this_nr_running >= group_capacity ||
3453 power_savings_balance = 0;
3456 * If a group is already running at full capacity or idle,
3457 * don't include that group in power savings calculations
3459 if (!power_savings_balance || sum_nr_running >= group_capacity
3464 * Calculate the group which has the least non-idle load.
3465 * This is the group from where we need to pick up the load
3468 if ((sum_nr_running < min_nr_running) ||
3469 (sum_nr_running == min_nr_running &&
3470 first_cpu(group->cpumask) <
3471 first_cpu(group_min->cpumask))) {
3473 min_nr_running = sum_nr_running;
3474 min_load_per_task = sum_weighted_load /
3479 * Calculate the group which is almost near its
3480 * capacity but still has some space to pick up some load
3481 * from other group and save more power
3483 if (sum_nr_running <= group_capacity - 1) {
3484 if (sum_nr_running > leader_nr_running ||
3485 (sum_nr_running == leader_nr_running &&
3486 first_cpu(group->cpumask) >
3487 first_cpu(group_leader->cpumask))) {
3488 group_leader = group;
3489 leader_nr_running = sum_nr_running;
3494 group = group->next;
3495 } while (group != sd->groups);
3497 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3500 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3502 if (this_load >= avg_load ||
3503 100*max_load <= sd->imbalance_pct*this_load)
3506 busiest_load_per_task /= busiest_nr_running;
3508 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3511 * We're trying to get all the cpus to the average_load, so we don't
3512 * want to push ourselves above the average load, nor do we wish to
3513 * reduce the max loaded cpu below the average load, as either of these
3514 * actions would just result in more rebalancing later, and ping-pong
3515 * tasks around. Thus we look for the minimum possible imbalance.
3516 * Negative imbalances (*we* are more loaded than anyone else) will
3517 * be counted as no imbalance for these purposes -- we can't fix that
3518 * by pulling tasks to us. Be careful of negative numbers as they'll
3519 * appear as very large values with unsigned longs.
3521 if (max_load <= busiest_load_per_task)
3525 * In the presence of smp nice balancing, certain scenarios can have
3526 * max load less than avg load(as we skip the groups at or below
3527 * its cpu_power, while calculating max_load..)
3529 if (max_load < avg_load) {
3531 goto small_imbalance;
3534 /* Don't want to pull so many tasks that a group would go idle */
3535 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3537 /* How much load to actually move to equalise the imbalance */
3538 *imbalance = min(max_pull * busiest->__cpu_power,
3539 (avg_load - this_load) * this->__cpu_power)
3543 * if *imbalance is less than the average load per runnable task
3544 * there is no gaurantee that any tasks will be moved so we'll have
3545 * a think about bumping its value to force at least one task to be
3548 if (*imbalance < busiest_load_per_task) {
3549 unsigned long tmp, pwr_now, pwr_move;
3553 pwr_move = pwr_now = 0;
3555 if (this_nr_running) {
3556 this_load_per_task /= this_nr_running;
3557 if (busiest_load_per_task > this_load_per_task)
3560 this_load_per_task = SCHED_LOAD_SCALE;
3562 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3563 busiest_load_per_task * imbn) {
3564 *imbalance = busiest_load_per_task;
3569 * OK, we don't have enough imbalance to justify moving tasks,
3570 * however we may be able to increase total CPU power used by
3574 pwr_now += busiest->__cpu_power *
3575 min(busiest_load_per_task, max_load);
3576 pwr_now += this->__cpu_power *
3577 min(this_load_per_task, this_load);
3578 pwr_now /= SCHED_LOAD_SCALE;
3580 /* Amount of load we'd subtract */
3581 tmp = sg_div_cpu_power(busiest,
3582 busiest_load_per_task * SCHED_LOAD_SCALE);
3584 pwr_move += busiest->__cpu_power *
3585 min(busiest_load_per_task, max_load - tmp);
3587 /* Amount of load we'd add */
3588 if (max_load * busiest->__cpu_power <
3589 busiest_load_per_task * SCHED_LOAD_SCALE)
3590 tmp = sg_div_cpu_power(this,
3591 max_load * busiest->__cpu_power);
3593 tmp = sg_div_cpu_power(this,
3594 busiest_load_per_task * SCHED_LOAD_SCALE);
3595 pwr_move += this->__cpu_power *
3596 min(this_load_per_task, this_load + tmp);
3597 pwr_move /= SCHED_LOAD_SCALE;
3599 /* Move if we gain throughput */
3600 if (pwr_move > pwr_now)
3601 *imbalance = busiest_load_per_task;
3607 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3608 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3611 if (this == group_leader && group_leader != group_min) {
3612 *imbalance = min_load_per_task;
3622 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3625 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3626 unsigned long imbalance, const cpumask_t *cpus)
3628 struct rq *busiest = NULL, *rq;
3629 unsigned long max_load = 0;
3632 for_each_cpu_mask(i, group->cpumask) {
3635 if (!cpu_isset(i, *cpus))
3639 wl = weighted_cpuload(i);
3641 if (rq->nr_running == 1 && wl > imbalance)
3644 if (wl > max_load) {
3654 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3655 * so long as it is large enough.
3657 #define MAX_PINNED_INTERVAL 512
3660 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3661 * tasks if there is an imbalance.
3663 static int load_balance(int this_cpu, struct rq *this_rq,
3664 struct sched_domain *sd, enum cpu_idle_type idle,
3665 int *balance, cpumask_t *cpus)
3667 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3668 struct sched_group *group;
3669 unsigned long imbalance;
3671 unsigned long flags;
3672 int unlock_aggregate;
3676 unlock_aggregate = get_aggregate(sd);
3679 * When power savings policy is enabled for the parent domain, idle
3680 * sibling can pick up load irrespective of busy siblings. In this case,
3681 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3682 * portraying it as CPU_NOT_IDLE.
3684 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3685 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3688 schedstat_inc(sd, lb_count[idle]);
3691 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3698 schedstat_inc(sd, lb_nobusyg[idle]);
3702 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3704 schedstat_inc(sd, lb_nobusyq[idle]);
3708 BUG_ON(busiest == this_rq);
3710 schedstat_add(sd, lb_imbalance[idle], imbalance);
3713 if (busiest->nr_running > 1) {
3715 * Attempt to move tasks. If find_busiest_group has found
3716 * an imbalance but busiest->nr_running <= 1, the group is
3717 * still unbalanced. ld_moved simply stays zero, so it is
3718 * correctly treated as an imbalance.
3720 local_irq_save(flags);
3721 double_rq_lock(this_rq, busiest);
3722 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3723 imbalance, sd, idle, &all_pinned);
3724 double_rq_unlock(this_rq, busiest);
3725 local_irq_restore(flags);
3728 * some other cpu did the load balance for us.
3730 if (ld_moved && this_cpu != smp_processor_id())
3731 resched_cpu(this_cpu);
3733 /* All tasks on this runqueue were pinned by CPU affinity */
3734 if (unlikely(all_pinned)) {
3735 cpu_clear(cpu_of(busiest), *cpus);
3736 if (!cpus_empty(*cpus))
3743 schedstat_inc(sd, lb_failed[idle]);
3744 sd->nr_balance_failed++;
3746 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3748 spin_lock_irqsave(&busiest->lock, flags);
3750 /* don't kick the migration_thread, if the curr
3751 * task on busiest cpu can't be moved to this_cpu
3753 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3754 spin_unlock_irqrestore(&busiest->lock, flags);
3756 goto out_one_pinned;
3759 if (!busiest->active_balance) {
3760 busiest->active_balance = 1;
3761 busiest->push_cpu = this_cpu;
3764 spin_unlock_irqrestore(&busiest->lock, flags);
3766 wake_up_process(busiest->migration_thread);
3769 * We've kicked active balancing, reset the failure
3772 sd->nr_balance_failed = sd->cache_nice_tries+1;
3775 sd->nr_balance_failed = 0;
3777 if (likely(!active_balance)) {
3778 /* We were unbalanced, so reset the balancing interval */
3779 sd->balance_interval = sd->min_interval;
3782 * If we've begun active balancing, start to back off. This
3783 * case may not be covered by the all_pinned logic if there
3784 * is only 1 task on the busy runqueue (because we don't call
3787 if (sd->balance_interval < sd->max_interval)
3788 sd->balance_interval *= 2;
3791 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3792 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3798 schedstat_inc(sd, lb_balanced[idle]);
3800 sd->nr_balance_failed = 0;
3803 /* tune up the balancing interval */
3804 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3805 (sd->balance_interval < sd->max_interval))
3806 sd->balance_interval *= 2;
3808 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3809 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3814 if (unlock_aggregate)
3820 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3821 * tasks if there is an imbalance.
3823 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3824 * this_rq is locked.
3827 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3830 struct sched_group *group;
3831 struct rq *busiest = NULL;
3832 unsigned long imbalance;
3840 * When power savings policy is enabled for the parent domain, idle
3841 * sibling can pick up load irrespective of busy siblings. In this case,
3842 * let the state of idle sibling percolate up as IDLE, instead of
3843 * portraying it as CPU_NOT_IDLE.
3845 if (sd->flags & SD_SHARE_CPUPOWER &&
3846 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3849 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3851 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3852 &sd_idle, cpus, NULL);
3854 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3858 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3860 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3864 BUG_ON(busiest == this_rq);
3866 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3869 if (busiest->nr_running > 1) {
3870 /* Attempt to move tasks */
3871 double_lock_balance(this_rq, busiest);
3872 /* this_rq->clock is already updated */
3873 update_rq_clock(busiest);
3874 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3875 imbalance, sd, CPU_NEWLY_IDLE,
3877 spin_unlock(&busiest->lock);
3879 if (unlikely(all_pinned)) {
3880 cpu_clear(cpu_of(busiest), *cpus);
3881 if (!cpus_empty(*cpus))
3887 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3888 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3889 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3892 sd->nr_balance_failed = 0;
3897 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3898 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3899 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3901 sd->nr_balance_failed = 0;
3907 * idle_balance is called by schedule() if this_cpu is about to become
3908 * idle. Attempts to pull tasks from other CPUs.
3910 static void idle_balance(int this_cpu, struct rq *this_rq)
3912 struct sched_domain *sd;
3913 int pulled_task = -1;
3914 unsigned long next_balance = jiffies + HZ;
3917 for_each_domain(this_cpu, sd) {
3918 unsigned long interval;
3920 if (!(sd->flags & SD_LOAD_BALANCE))
3923 if (sd->flags & SD_BALANCE_NEWIDLE)
3924 /* If we've pulled tasks over stop searching: */
3925 pulled_task = load_balance_newidle(this_cpu, this_rq,
3928 interval = msecs_to_jiffies(sd->balance_interval);
3929 if (time_after(next_balance, sd->last_balance + interval))
3930 next_balance = sd->last_balance + interval;
3934 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3936 * We are going idle. next_balance may be set based on
3937 * a busy processor. So reset next_balance.
3939 this_rq->next_balance = next_balance;
3944 * active_load_balance is run by migration threads. It pushes running tasks
3945 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3946 * running on each physical CPU where possible, and avoids physical /
3947 * logical imbalances.
3949 * Called with busiest_rq locked.
3951 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3953 int target_cpu = busiest_rq->push_cpu;
3954 struct sched_domain *sd;
3955 struct rq *target_rq;
3957 /* Is there any task to move? */
3958 if (busiest_rq->nr_running <= 1)
3961 target_rq = cpu_rq(target_cpu);
3964 * This condition is "impossible", if it occurs
3965 * we need to fix it. Originally reported by
3966 * Bjorn Helgaas on a 128-cpu setup.
3968 BUG_ON(busiest_rq == target_rq);
3970 /* move a task from busiest_rq to target_rq */
3971 double_lock_balance(busiest_rq, target_rq);
3972 update_rq_clock(busiest_rq);
3973 update_rq_clock(target_rq);
3975 /* Search for an sd spanning us and the target CPU. */
3976 for_each_domain(target_cpu, sd) {
3977 if ((sd->flags & SD_LOAD_BALANCE) &&
3978 cpu_isset(busiest_cpu, sd->span))
3983 schedstat_inc(sd, alb_count);
3985 if (move_one_task(target_rq, target_cpu, busiest_rq,
3987 schedstat_inc(sd, alb_pushed);
3989 schedstat_inc(sd, alb_failed);
3991 spin_unlock(&target_rq->lock);
3996 atomic_t load_balancer;
3998 } nohz ____cacheline_aligned = {
3999 .load_balancer = ATOMIC_INIT(-1),
4000 .cpu_mask = CPU_MASK_NONE,
4004 * This routine will try to nominate the ilb (idle load balancing)
4005 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4006 * load balancing on behalf of all those cpus. If all the cpus in the system
4007 * go into this tickless mode, then there will be no ilb owner (as there is
4008 * no need for one) and all the cpus will sleep till the next wakeup event
4011 * For the ilb owner, tick is not stopped. And this tick will be used
4012 * for idle load balancing. ilb owner will still be part of
4015 * While stopping the tick, this cpu will become the ilb owner if there
4016 * is no other owner. And will be the owner till that cpu becomes busy
4017 * or if all cpus in the system stop their ticks at which point
4018 * there is no need for ilb owner.
4020 * When the ilb owner becomes busy, it nominates another owner, during the
4021 * next busy scheduler_tick()
4023 int select_nohz_load_balancer(int stop_tick)
4025 int cpu = smp_processor_id();
4028 cpu_set(cpu, nohz.cpu_mask);
4029 cpu_rq(cpu)->in_nohz_recently = 1;
4032 * If we are going offline and still the leader, give up!
4034 if (cpu_is_offline(cpu) &&
4035 atomic_read(&nohz.load_balancer) == cpu) {
4036 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4041 /* time for ilb owner also to sleep */
4042 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4043 if (atomic_read(&nohz.load_balancer) == cpu)
4044 atomic_set(&nohz.load_balancer, -1);
4048 if (atomic_read(&nohz.load_balancer) == -1) {
4049 /* make me the ilb owner */
4050 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4052 } else if (atomic_read(&nohz.load_balancer) == cpu)
4055 if (!cpu_isset(cpu, nohz.cpu_mask))
4058 cpu_clear(cpu, nohz.cpu_mask);
4060 if (atomic_read(&nohz.load_balancer) == cpu)
4061 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4068 static DEFINE_SPINLOCK(balancing);
4071 * It checks each scheduling domain to see if it is due to be balanced,
4072 * and initiates a balancing operation if so.
4074 * Balancing parameters are set up in arch_init_sched_domains.
4076 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4079 struct rq *rq = cpu_rq(cpu);
4080 unsigned long interval;
4081 struct sched_domain *sd;
4082 /* Earliest time when we have to do rebalance again */
4083 unsigned long next_balance = jiffies + 60*HZ;
4084 int update_next_balance = 0;
4087 for_each_domain(cpu, sd) {
4088 if (!(sd->flags & SD_LOAD_BALANCE))
4091 interval = sd->balance_interval;
4092 if (idle != CPU_IDLE)
4093 interval *= sd->busy_factor;
4095 /* scale ms to jiffies */
4096 interval = msecs_to_jiffies(interval);
4097 if (unlikely(!interval))
4099 if (interval > HZ*NR_CPUS/10)
4100 interval = HZ*NR_CPUS/10;
4103 if (sd->flags & SD_SERIALIZE) {
4104 if (!spin_trylock(&balancing))
4108 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4109 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4111 * We've pulled tasks over so either we're no
4112 * longer idle, or one of our SMT siblings is
4115 idle = CPU_NOT_IDLE;
4117 sd->last_balance = jiffies;
4119 if (sd->flags & SD_SERIALIZE)
4120 spin_unlock(&balancing);
4122 if (time_after(next_balance, sd->last_balance + interval)) {
4123 next_balance = sd->last_balance + interval;
4124 update_next_balance = 1;
4128 * Stop the load balance at this level. There is another
4129 * CPU in our sched group which is doing load balancing more
4137 * next_balance will be updated only when there is a need.
4138 * When the cpu is attached to null domain for ex, it will not be
4141 if (likely(update_next_balance))
4142 rq->next_balance = next_balance;
4146 * run_rebalance_domains is triggered when needed from the scheduler tick.
4147 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4148 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4150 static void run_rebalance_domains(struct softirq_action *h)
4152 int this_cpu = smp_processor_id();
4153 struct rq *this_rq = cpu_rq(this_cpu);
4154 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4155 CPU_IDLE : CPU_NOT_IDLE;
4157 rebalance_domains(this_cpu, idle);
4161 * If this cpu is the owner for idle load balancing, then do the
4162 * balancing on behalf of the other idle cpus whose ticks are
4165 if (this_rq->idle_at_tick &&
4166 atomic_read(&nohz.load_balancer) == this_cpu) {
4167 cpumask_t cpus = nohz.cpu_mask;
4171 cpu_clear(this_cpu, cpus);
4172 for_each_cpu_mask(balance_cpu, cpus) {
4174 * If this cpu gets work to do, stop the load balancing
4175 * work being done for other cpus. Next load
4176 * balancing owner will pick it up.
4181 rebalance_domains(balance_cpu, CPU_IDLE);
4183 rq = cpu_rq(balance_cpu);
4184 if (time_after(this_rq->next_balance, rq->next_balance))
4185 this_rq->next_balance = rq->next_balance;
4192 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4194 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4195 * idle load balancing owner or decide to stop the periodic load balancing,
4196 * if the whole system is idle.
4198 static inline void trigger_load_balance(struct rq *rq, int cpu)
4202 * If we were in the nohz mode recently and busy at the current
4203 * scheduler tick, then check if we need to nominate new idle
4206 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4207 rq->in_nohz_recently = 0;
4209 if (atomic_read(&nohz.load_balancer) == cpu) {
4210 cpu_clear(cpu, nohz.cpu_mask);
4211 atomic_set(&nohz.load_balancer, -1);
4214 if (atomic_read(&nohz.load_balancer) == -1) {
4216 * simple selection for now: Nominate the
4217 * first cpu in the nohz list to be the next
4220 * TBD: Traverse the sched domains and nominate
4221 * the nearest cpu in the nohz.cpu_mask.
4223 int ilb = first_cpu(nohz.cpu_mask);
4225 if (ilb < nr_cpu_ids)
4231 * If this cpu is idle and doing idle load balancing for all the
4232 * cpus with ticks stopped, is it time for that to stop?
4234 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4235 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4241 * If this cpu is idle and the idle load balancing is done by
4242 * someone else, then no need raise the SCHED_SOFTIRQ
4244 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4245 cpu_isset(cpu, nohz.cpu_mask))
4248 if (time_after_eq(jiffies, rq->next_balance))
4249 raise_softirq(SCHED_SOFTIRQ);
4252 #else /* CONFIG_SMP */
4255 * on UP we do not need to balance between CPUs:
4257 static inline void idle_balance(int cpu, struct rq *rq)
4263 DEFINE_PER_CPU(struct kernel_stat, kstat);
4265 EXPORT_PER_CPU_SYMBOL(kstat);
4268 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4269 * that have not yet been banked in case the task is currently running.
4271 unsigned long long task_sched_runtime(struct task_struct *p)
4273 unsigned long flags;
4277 rq = task_rq_lock(p, &flags);
4278 ns = p->se.sum_exec_runtime;
4279 if (task_current(rq, p)) {
4280 update_rq_clock(rq);
4281 delta_exec = rq->clock - p->se.exec_start;
4282 if ((s64)delta_exec > 0)
4285 task_rq_unlock(rq, &flags);
4291 * Account user cpu time to a process.
4292 * @p: the process that the cpu time gets accounted to
4293 * @cputime: the cpu time spent in user space since the last update
4295 void account_user_time(struct task_struct *p, cputime_t cputime)
4297 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4300 p->utime = cputime_add(p->utime, cputime);
4302 /* Add user time to cpustat. */
4303 tmp = cputime_to_cputime64(cputime);
4304 if (TASK_NICE(p) > 0)
4305 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4307 cpustat->user = cputime64_add(cpustat->user, tmp);
4311 * Account guest cpu time to a process.
4312 * @p: the process that the cpu time gets accounted to
4313 * @cputime: the cpu time spent in virtual machine since the last update
4315 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4318 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4320 tmp = cputime_to_cputime64(cputime);
4322 p->utime = cputime_add(p->utime, cputime);
4323 p->gtime = cputime_add(p->gtime, cputime);
4325 cpustat->user = cputime64_add(cpustat->user, tmp);
4326 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4330 * Account scaled user cpu time to a process.
4331 * @p: the process that the cpu time gets accounted to
4332 * @cputime: the cpu time spent in user space since the last update
4334 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4336 p->utimescaled = cputime_add(p->utimescaled, cputime);
4340 * Account system cpu time to a process.
4341 * @p: the process that the cpu time gets accounted to
4342 * @hardirq_offset: the offset to subtract from hardirq_count()
4343 * @cputime: the cpu time spent in kernel space since the last update
4345 void account_system_time(struct task_struct *p, int hardirq_offset,
4348 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4349 struct rq *rq = this_rq();
4352 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4353 account_guest_time(p, cputime);
4357 p->stime = cputime_add(p->stime, cputime);
4359 /* Add system time to cpustat. */
4360 tmp = cputime_to_cputime64(cputime);
4361 if (hardirq_count() - hardirq_offset)
4362 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4363 else if (softirq_count())
4364 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4365 else if (p != rq->idle)
4366 cpustat->system = cputime64_add(cpustat->system, tmp);
4367 else if (atomic_read(&rq->nr_iowait) > 0)
4368 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4370 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4371 /* Account for system time used */
4372 acct_update_integrals(p);
4376 * Account scaled system cpu time to a process.
4377 * @p: the process that the cpu time gets accounted to
4378 * @hardirq_offset: the offset to subtract from hardirq_count()
4379 * @cputime: the cpu time spent in kernel space since the last update
4381 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4383 p->stimescaled = cputime_add(p->stimescaled, cputime);
4387 * Account for involuntary wait time.
4388 * @p: the process from which the cpu time has been stolen
4389 * @steal: the cpu time spent in involuntary wait
4391 void account_steal_time(struct task_struct *p, cputime_t steal)
4393 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4394 cputime64_t tmp = cputime_to_cputime64(steal);
4395 struct rq *rq = this_rq();
4397 if (p == rq->idle) {
4398 p->stime = cputime_add(p->stime, steal);
4399 if (atomic_read(&rq->nr_iowait) > 0)
4400 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4402 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4404 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4408 * This function gets called by the timer code, with HZ frequency.
4409 * We call it with interrupts disabled.
4411 * It also gets called by the fork code, when changing the parent's
4414 void scheduler_tick(void)
4416 int cpu = smp_processor_id();
4417 struct rq *rq = cpu_rq(cpu);
4418 struct task_struct *curr = rq->curr;
4422 spin_lock(&rq->lock);
4423 update_rq_clock(rq);
4424 update_cpu_load(rq);
4425 curr->sched_class->task_tick(rq, curr, 0);
4426 spin_unlock(&rq->lock);
4429 rq->idle_at_tick = idle_cpu(cpu);
4430 trigger_load_balance(rq, cpu);
4434 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4435 defined(CONFIG_PREEMPT_TRACER))
4437 static inline unsigned long get_parent_ip(unsigned long addr)
4439 if (in_lock_functions(addr)) {
4440 addr = CALLER_ADDR2;
4441 if (in_lock_functions(addr))
4442 addr = CALLER_ADDR3;
4447 void __kprobes add_preempt_count(int val)
4449 #ifdef CONFIG_DEBUG_PREEMPT
4453 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4456 preempt_count() += val;
4457 #ifdef CONFIG_DEBUG_PREEMPT
4459 * Spinlock count overflowing soon?
4461 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4464 if (preempt_count() == val)
4465 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4467 EXPORT_SYMBOL(add_preempt_count);
4469 void __kprobes sub_preempt_count(int val)
4471 #ifdef CONFIG_DEBUG_PREEMPT
4475 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4478 * Is the spinlock portion underflowing?
4480 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4481 !(preempt_count() & PREEMPT_MASK)))
4485 if (preempt_count() == val)
4486 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4487 preempt_count() -= val;
4489 EXPORT_SYMBOL(sub_preempt_count);
4494 * Print scheduling while atomic bug:
4496 static noinline void __schedule_bug(struct task_struct *prev)
4498 struct pt_regs *regs = get_irq_regs();
4500 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4501 prev->comm, prev->pid, preempt_count());
4503 debug_show_held_locks(prev);
4504 if (irqs_disabled())
4505 print_irqtrace_events(prev);
4514 * Various schedule()-time debugging checks and statistics:
4516 static inline void schedule_debug(struct task_struct *prev)
4519 * Test if we are atomic. Since do_exit() needs to call into
4520 * schedule() atomically, we ignore that path for now.
4521 * Otherwise, whine if we are scheduling when we should not be.
4523 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4524 __schedule_bug(prev);
4526 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4528 schedstat_inc(this_rq(), sched_count);
4529 #ifdef CONFIG_SCHEDSTATS
4530 if (unlikely(prev->lock_depth >= 0)) {
4531 schedstat_inc(this_rq(), bkl_count);
4532 schedstat_inc(prev, sched_info.bkl_count);
4538 * Pick up the highest-prio task:
4540 static inline struct task_struct *
4541 pick_next_task(struct rq *rq, struct task_struct *prev)
4543 const struct sched_class *class;
4544 struct task_struct *p;
4547 * Optimization: we know that if all tasks are in
4548 * the fair class we can call that function directly:
4550 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4551 p = fair_sched_class.pick_next_task(rq);
4556 class = sched_class_highest;
4558 p = class->pick_next_task(rq);
4562 * Will never be NULL as the idle class always
4563 * returns a non-NULL p:
4565 class = class->next;
4570 * schedule() is the main scheduler function.
4572 asmlinkage void __sched schedule(void)
4574 struct task_struct *prev, *next;
4575 unsigned long *switch_count;
4581 cpu = smp_processor_id();
4585 switch_count = &prev->nivcsw;
4587 release_kernel_lock(prev);
4588 need_resched_nonpreemptible:
4590 schedule_debug(prev);
4595 * Do the rq-clock update outside the rq lock:
4597 local_irq_disable();
4598 update_rq_clock(rq);
4599 spin_lock(&rq->lock);
4600 clear_tsk_need_resched(prev);
4602 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4603 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4604 signal_pending(prev))) {
4605 prev->state = TASK_RUNNING;
4607 deactivate_task(rq, prev, 1);
4609 switch_count = &prev->nvcsw;
4613 if (prev->sched_class->pre_schedule)
4614 prev->sched_class->pre_schedule(rq, prev);
4617 if (unlikely(!rq->nr_running))
4618 idle_balance(cpu, rq);
4620 prev->sched_class->put_prev_task(rq, prev);
4621 next = pick_next_task(rq, prev);
4623 if (likely(prev != next)) {
4624 sched_info_switch(prev, next);
4630 context_switch(rq, prev, next); /* unlocks the rq */
4632 * the context switch might have flipped the stack from under
4633 * us, hence refresh the local variables.
4635 cpu = smp_processor_id();
4638 spin_unlock_irq(&rq->lock);
4642 if (unlikely(reacquire_kernel_lock(current) < 0))
4643 goto need_resched_nonpreemptible;
4645 preempt_enable_no_resched();
4646 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4649 EXPORT_SYMBOL(schedule);
4651 #ifdef CONFIG_PREEMPT
4653 * this is the entry point to schedule() from in-kernel preemption
4654 * off of preempt_enable. Kernel preemptions off return from interrupt
4655 * occur there and call schedule directly.
4657 asmlinkage void __sched preempt_schedule(void)
4659 struct thread_info *ti = current_thread_info();
4662 * If there is a non-zero preempt_count or interrupts are disabled,
4663 * we do not want to preempt the current task. Just return..
4665 if (likely(ti->preempt_count || irqs_disabled()))
4669 add_preempt_count(PREEMPT_ACTIVE);
4671 sub_preempt_count(PREEMPT_ACTIVE);
4674 * Check again in case we missed a preemption opportunity
4675 * between schedule and now.
4678 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4680 EXPORT_SYMBOL(preempt_schedule);
4683 * this is the entry point to schedule() from kernel preemption
4684 * off of irq context.
4685 * Note, that this is called and return with irqs disabled. This will
4686 * protect us against recursive calling from irq.
4688 asmlinkage void __sched preempt_schedule_irq(void)
4690 struct thread_info *ti = current_thread_info();
4692 /* Catch callers which need to be fixed */
4693 BUG_ON(ti->preempt_count || !irqs_disabled());
4696 add_preempt_count(PREEMPT_ACTIVE);
4699 local_irq_disable();
4700 sub_preempt_count(PREEMPT_ACTIVE);
4703 * Check again in case we missed a preemption opportunity
4704 * between schedule and now.
4707 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4710 #endif /* CONFIG_PREEMPT */
4712 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4715 return try_to_wake_up(curr->private, mode, sync);
4717 EXPORT_SYMBOL(default_wake_function);
4720 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4721 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4722 * number) then we wake all the non-exclusive tasks and one exclusive task.
4724 * There are circumstances in which we can try to wake a task which has already
4725 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4726 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4728 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4729 int nr_exclusive, int sync, void *key)
4731 wait_queue_t *curr, *next;
4733 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4734 unsigned flags = curr->flags;
4736 if (curr->func(curr, mode, sync, key) &&
4737 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4743 * __wake_up - wake up threads blocked on a waitqueue.
4745 * @mode: which threads
4746 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4747 * @key: is directly passed to the wakeup function
4749 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4750 int nr_exclusive, void *key)
4752 unsigned long flags;
4754 spin_lock_irqsave(&q->lock, flags);
4755 __wake_up_common(q, mode, nr_exclusive, 0, key);
4756 spin_unlock_irqrestore(&q->lock, flags);
4758 EXPORT_SYMBOL(__wake_up);
4761 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4763 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4765 __wake_up_common(q, mode, 1, 0, NULL);
4769 * __wake_up_sync - wake up threads blocked on a waitqueue.
4771 * @mode: which threads
4772 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4774 * The sync wakeup differs that the waker knows that it will schedule
4775 * away soon, so while the target thread will be woken up, it will not
4776 * be migrated to another CPU - ie. the two threads are 'synchronized'
4777 * with each other. This can prevent needless bouncing between CPUs.
4779 * On UP it can prevent extra preemption.
4782 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4784 unsigned long flags;
4790 if (unlikely(!nr_exclusive))
4793 spin_lock_irqsave(&q->lock, flags);
4794 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4795 spin_unlock_irqrestore(&q->lock, flags);
4797 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4799 void complete(struct completion *x)
4801 unsigned long flags;
4803 spin_lock_irqsave(&x->wait.lock, flags);
4805 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4806 spin_unlock_irqrestore(&x->wait.lock, flags);
4808 EXPORT_SYMBOL(complete);
4810 void complete_all(struct completion *x)
4812 unsigned long flags;
4814 spin_lock_irqsave(&x->wait.lock, flags);
4815 x->done += UINT_MAX/2;
4816 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4817 spin_unlock_irqrestore(&x->wait.lock, flags);
4819 EXPORT_SYMBOL(complete_all);
4821 static inline long __sched
4822 do_wait_for_common(struct completion *x, long timeout, int state)
4825 DECLARE_WAITQUEUE(wait, current);
4827 wait.flags |= WQ_FLAG_EXCLUSIVE;
4828 __add_wait_queue_tail(&x->wait, &wait);
4830 if ((state == TASK_INTERRUPTIBLE &&
4831 signal_pending(current)) ||
4832 (state == TASK_KILLABLE &&
4833 fatal_signal_pending(current))) {
4834 __remove_wait_queue(&x->wait, &wait);
4835 return -ERESTARTSYS;
4837 __set_current_state(state);
4838 spin_unlock_irq(&x->wait.lock);
4839 timeout = schedule_timeout(timeout);
4840 spin_lock_irq(&x->wait.lock);
4842 __remove_wait_queue(&x->wait, &wait);
4846 __remove_wait_queue(&x->wait, &wait);
4853 wait_for_common(struct completion *x, long timeout, int state)
4857 spin_lock_irq(&x->wait.lock);
4858 timeout = do_wait_for_common(x, timeout, state);
4859 spin_unlock_irq(&x->wait.lock);
4863 void __sched wait_for_completion(struct completion *x)
4865 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4867 EXPORT_SYMBOL(wait_for_completion);
4869 unsigned long __sched
4870 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4872 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4874 EXPORT_SYMBOL(wait_for_completion_timeout);
4876 int __sched wait_for_completion_interruptible(struct completion *x)
4878 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4879 if (t == -ERESTARTSYS)
4883 EXPORT_SYMBOL(wait_for_completion_interruptible);
4885 unsigned long __sched
4886 wait_for_completion_interruptible_timeout(struct completion *x,
4887 unsigned long timeout)
4889 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4891 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4893 int __sched wait_for_completion_killable(struct completion *x)
4895 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4896 if (t == -ERESTARTSYS)
4900 EXPORT_SYMBOL(wait_for_completion_killable);
4903 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4905 unsigned long flags;
4908 init_waitqueue_entry(&wait, current);
4910 __set_current_state(state);
4912 spin_lock_irqsave(&q->lock, flags);
4913 __add_wait_queue(q, &wait);
4914 spin_unlock(&q->lock);
4915 timeout = schedule_timeout(timeout);
4916 spin_lock_irq(&q->lock);
4917 __remove_wait_queue(q, &wait);
4918 spin_unlock_irqrestore(&q->lock, flags);
4923 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4925 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4927 EXPORT_SYMBOL(interruptible_sleep_on);
4930 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4932 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4934 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4936 void __sched sleep_on(wait_queue_head_t *q)
4938 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4940 EXPORT_SYMBOL(sleep_on);
4942 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4944 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4946 EXPORT_SYMBOL(sleep_on_timeout);
4948 #ifdef CONFIG_RT_MUTEXES
4951 * rt_mutex_setprio - set the current priority of a task
4953 * @prio: prio value (kernel-internal form)
4955 * This function changes the 'effective' priority of a task. It does
4956 * not touch ->normal_prio like __setscheduler().
4958 * Used by the rt_mutex code to implement priority inheritance logic.
4960 void rt_mutex_setprio(struct task_struct *p, int prio)
4962 unsigned long flags;
4963 int oldprio, on_rq, running;
4965 const struct sched_class *prev_class = p->sched_class;
4967 BUG_ON(prio < 0 || prio > MAX_PRIO);
4969 rq = task_rq_lock(p, &flags);
4970 update_rq_clock(rq);
4973 on_rq = p->se.on_rq;
4974 running = task_current(rq, p);
4976 dequeue_task(rq, p, 0);
4978 p->sched_class->put_prev_task(rq, p);
4981 p->sched_class = &rt_sched_class;
4983 p->sched_class = &fair_sched_class;
4988 p->sched_class->set_curr_task(rq);
4990 enqueue_task(rq, p, 0);
4992 check_class_changed(rq, p, prev_class, oldprio, running);
4994 task_rq_unlock(rq, &flags);
4999 void set_user_nice(struct task_struct *p, long nice)
5001 int old_prio, delta, on_rq;
5002 unsigned long flags;
5005 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5008 * We have to be careful, if called from sys_setpriority(),
5009 * the task might be in the middle of scheduling on another CPU.
5011 rq = task_rq_lock(p, &flags);
5012 update_rq_clock(rq);
5014 * The RT priorities are set via sched_setscheduler(), but we still
5015 * allow the 'normal' nice value to be set - but as expected
5016 * it wont have any effect on scheduling until the task is
5017 * SCHED_FIFO/SCHED_RR:
5019 if (task_has_rt_policy(p)) {
5020 p->static_prio = NICE_TO_PRIO(nice);
5023 on_rq = p->se.on_rq;
5025 dequeue_task(rq, p, 0);
5027 p->static_prio = NICE_TO_PRIO(nice);
5030 p->prio = effective_prio(p);
5031 delta = p->prio - old_prio;
5034 enqueue_task(rq, p, 0);
5036 * If the task increased its priority or is running and
5037 * lowered its priority, then reschedule its CPU:
5039 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5040 resched_task(rq->curr);
5043 task_rq_unlock(rq, &flags);
5045 EXPORT_SYMBOL(set_user_nice);
5048 * can_nice - check if a task can reduce its nice value
5052 int can_nice(const struct task_struct *p, const int nice)
5054 /* convert nice value [19,-20] to rlimit style value [1,40] */
5055 int nice_rlim = 20 - nice;
5057 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5058 capable(CAP_SYS_NICE));
5061 #ifdef __ARCH_WANT_SYS_NICE
5064 * sys_nice - change the priority of the current process.
5065 * @increment: priority increment
5067 * sys_setpriority is a more generic, but much slower function that
5068 * does similar things.
5070 asmlinkage long sys_nice(int increment)
5075 * Setpriority might change our priority at the same moment.
5076 * We don't have to worry. Conceptually one call occurs first
5077 * and we have a single winner.
5079 if (increment < -40)
5084 nice = PRIO_TO_NICE(current->static_prio) + increment;
5090 if (increment < 0 && !can_nice(current, nice))
5093 retval = security_task_setnice(current, nice);
5097 set_user_nice(current, nice);
5104 * task_prio - return the priority value of a given task.
5105 * @p: the task in question.
5107 * This is the priority value as seen by users in /proc.
5108 * RT tasks are offset by -200. Normal tasks are centered
5109 * around 0, value goes from -16 to +15.
5111 int task_prio(const struct task_struct *p)
5113 return p->prio - MAX_RT_PRIO;
5117 * task_nice - return the nice value of a given task.
5118 * @p: the task in question.
5120 int task_nice(const struct task_struct *p)
5122 return TASK_NICE(p);
5124 EXPORT_SYMBOL(task_nice);
5127 * idle_cpu - is a given cpu idle currently?
5128 * @cpu: the processor in question.
5130 int idle_cpu(int cpu)
5132 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5136 * idle_task - return the idle task for a given cpu.
5137 * @cpu: the processor in question.
5139 struct task_struct *idle_task(int cpu)
5141 return cpu_rq(cpu)->idle;
5145 * find_process_by_pid - find a process with a matching PID value.
5146 * @pid: the pid in question.
5148 static struct task_struct *find_process_by_pid(pid_t pid)
5150 return pid ? find_task_by_vpid(pid) : current;
5153 /* Actually do priority change: must hold rq lock. */
5155 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5157 BUG_ON(p->se.on_rq);
5160 switch (p->policy) {
5164 p->sched_class = &fair_sched_class;
5168 p->sched_class = &rt_sched_class;
5172 p->rt_priority = prio;
5173 p->normal_prio = normal_prio(p);
5174 /* we are holding p->pi_lock already */
5175 p->prio = rt_mutex_getprio(p);
5180 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5181 * @p: the task in question.
5182 * @policy: new policy.
5183 * @param: structure containing the new RT priority.
5185 * NOTE that the task may be already dead.
5187 int sched_setscheduler(struct task_struct *p, int policy,
5188 struct sched_param *param)
5190 int retval, oldprio, oldpolicy = -1, on_rq, running;
5191 unsigned long flags;
5192 const struct sched_class *prev_class = p->sched_class;
5195 /* may grab non-irq protected spin_locks */
5196 BUG_ON(in_interrupt());
5198 /* double check policy once rq lock held */
5200 policy = oldpolicy = p->policy;
5201 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5202 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5203 policy != SCHED_IDLE)
5206 * Valid priorities for SCHED_FIFO and SCHED_RR are
5207 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5208 * SCHED_BATCH and SCHED_IDLE is 0.
5210 if (param->sched_priority < 0 ||
5211 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5212 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5214 if (rt_policy(policy) != (param->sched_priority != 0))
5218 * Allow unprivileged RT tasks to decrease priority:
5220 if (!capable(CAP_SYS_NICE)) {
5221 if (rt_policy(policy)) {
5222 unsigned long rlim_rtprio;
5224 if (!lock_task_sighand(p, &flags))
5226 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5227 unlock_task_sighand(p, &flags);
5229 /* can't set/change the rt policy */
5230 if (policy != p->policy && !rlim_rtprio)
5233 /* can't increase priority */
5234 if (param->sched_priority > p->rt_priority &&
5235 param->sched_priority > rlim_rtprio)
5239 * Like positive nice levels, dont allow tasks to
5240 * move out of SCHED_IDLE either:
5242 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5245 /* can't change other user's priorities */
5246 if ((current->euid != p->euid) &&
5247 (current->euid != p->uid))
5251 #ifdef CONFIG_RT_GROUP_SCHED
5253 * Do not allow realtime tasks into groups that have no runtime
5256 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5260 retval = security_task_setscheduler(p, policy, param);
5264 * make sure no PI-waiters arrive (or leave) while we are
5265 * changing the priority of the task:
5267 spin_lock_irqsave(&p->pi_lock, flags);
5269 * To be able to change p->policy safely, the apropriate
5270 * runqueue lock must be held.
5272 rq = __task_rq_lock(p);
5273 /* recheck policy now with rq lock held */
5274 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5275 policy = oldpolicy = -1;
5276 __task_rq_unlock(rq);
5277 spin_unlock_irqrestore(&p->pi_lock, flags);
5280 update_rq_clock(rq);
5281 on_rq = p->se.on_rq;
5282 running = task_current(rq, p);
5284 deactivate_task(rq, p, 0);
5286 p->sched_class->put_prev_task(rq, p);
5289 __setscheduler(rq, p, policy, param->sched_priority);
5292 p->sched_class->set_curr_task(rq);
5294 activate_task(rq, p, 0);
5296 check_class_changed(rq, p, prev_class, oldprio, running);
5298 __task_rq_unlock(rq);
5299 spin_unlock_irqrestore(&p->pi_lock, flags);
5301 rt_mutex_adjust_pi(p);
5305 EXPORT_SYMBOL_GPL(sched_setscheduler);
5308 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5310 struct sched_param lparam;
5311 struct task_struct *p;
5314 if (!param || pid < 0)
5316 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5321 p = find_process_by_pid(pid);
5323 retval = sched_setscheduler(p, policy, &lparam);
5330 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5331 * @pid: the pid in question.
5332 * @policy: new policy.
5333 * @param: structure containing the new RT priority.
5336 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5338 /* negative values for policy are not valid */
5342 return do_sched_setscheduler(pid, policy, param);
5346 * sys_sched_setparam - set/change the RT priority of a thread
5347 * @pid: the pid in question.
5348 * @param: structure containing the new RT priority.
5350 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5352 return do_sched_setscheduler(pid, -1, param);
5356 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5357 * @pid: the pid in question.
5359 asmlinkage long sys_sched_getscheduler(pid_t pid)
5361 struct task_struct *p;
5368 read_lock(&tasklist_lock);
5369 p = find_process_by_pid(pid);
5371 retval = security_task_getscheduler(p);
5375 read_unlock(&tasklist_lock);
5380 * sys_sched_getscheduler - get the RT priority of a thread
5381 * @pid: the pid in question.
5382 * @param: structure containing the RT priority.
5384 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5386 struct sched_param lp;
5387 struct task_struct *p;
5390 if (!param || pid < 0)
5393 read_lock(&tasklist_lock);
5394 p = find_process_by_pid(pid);
5399 retval = security_task_getscheduler(p);
5403 lp.sched_priority = p->rt_priority;
5404 read_unlock(&tasklist_lock);
5407 * This one might sleep, we cannot do it with a spinlock held ...
5409 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5414 read_unlock(&tasklist_lock);
5418 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5420 cpumask_t cpus_allowed;
5421 cpumask_t new_mask = *in_mask;
5422 struct task_struct *p;
5426 read_lock(&tasklist_lock);
5428 p = find_process_by_pid(pid);
5430 read_unlock(&tasklist_lock);
5436 * It is not safe to call set_cpus_allowed with the
5437 * tasklist_lock held. We will bump the task_struct's
5438 * usage count and then drop tasklist_lock.
5441 read_unlock(&tasklist_lock);
5444 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5445 !capable(CAP_SYS_NICE))
5448 retval = security_task_setscheduler(p, 0, NULL);
5452 cpuset_cpus_allowed(p, &cpus_allowed);
5453 cpus_and(new_mask, new_mask, cpus_allowed);
5455 retval = set_cpus_allowed_ptr(p, &new_mask);
5458 cpuset_cpus_allowed(p, &cpus_allowed);
5459 if (!cpus_subset(new_mask, cpus_allowed)) {
5461 * We must have raced with a concurrent cpuset
5462 * update. Just reset the cpus_allowed to the
5463 * cpuset's cpus_allowed
5465 new_mask = cpus_allowed;
5475 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5476 cpumask_t *new_mask)
5478 if (len < sizeof(cpumask_t)) {
5479 memset(new_mask, 0, sizeof(cpumask_t));
5480 } else if (len > sizeof(cpumask_t)) {
5481 len = sizeof(cpumask_t);
5483 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5487 * sys_sched_setaffinity - set the cpu affinity of a process
5488 * @pid: pid of the process
5489 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5490 * @user_mask_ptr: user-space pointer to the new cpu mask
5492 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5493 unsigned long __user *user_mask_ptr)
5498 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5502 return sched_setaffinity(pid, &new_mask);
5506 * Represents all cpu's present in the system
5507 * In systems capable of hotplug, this map could dynamically grow
5508 * as new cpu's are detected in the system via any platform specific
5509 * method, such as ACPI for e.g.
5512 cpumask_t cpu_present_map __read_mostly;
5513 EXPORT_SYMBOL(cpu_present_map);
5516 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5517 EXPORT_SYMBOL(cpu_online_map);
5519 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5520 EXPORT_SYMBOL(cpu_possible_map);
5523 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5525 struct task_struct *p;
5529 read_lock(&tasklist_lock);
5532 p = find_process_by_pid(pid);
5536 retval = security_task_getscheduler(p);
5540 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5543 read_unlock(&tasklist_lock);
5550 * sys_sched_getaffinity - get the cpu affinity of a process
5551 * @pid: pid of the process
5552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5553 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5555 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5556 unsigned long __user *user_mask_ptr)
5561 if (len < sizeof(cpumask_t))
5564 ret = sched_getaffinity(pid, &mask);
5568 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5571 return sizeof(cpumask_t);
5575 * sys_sched_yield - yield the current processor to other threads.
5577 * This function yields the current CPU to other tasks. If there are no
5578 * other threads running on this CPU then this function will return.
5580 asmlinkage long sys_sched_yield(void)
5582 struct rq *rq = this_rq_lock();
5584 schedstat_inc(rq, yld_count);
5585 current->sched_class->yield_task(rq);
5588 * Since we are going to call schedule() anyway, there's
5589 * no need to preempt or enable interrupts:
5591 __release(rq->lock);
5592 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5593 _raw_spin_unlock(&rq->lock);
5594 preempt_enable_no_resched();
5601 static void __cond_resched(void)
5603 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5604 __might_sleep(__FILE__, __LINE__);
5607 * The BKS might be reacquired before we have dropped
5608 * PREEMPT_ACTIVE, which could trigger a second
5609 * cond_resched() call.
5612 add_preempt_count(PREEMPT_ACTIVE);
5614 sub_preempt_count(PREEMPT_ACTIVE);
5615 } while (need_resched());
5618 int __sched _cond_resched(void)
5620 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5621 system_state == SYSTEM_RUNNING) {
5627 EXPORT_SYMBOL(_cond_resched);
5630 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5631 * call schedule, and on return reacquire the lock.
5633 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5634 * operations here to prevent schedule() from being called twice (once via
5635 * spin_unlock(), once by hand).
5637 int cond_resched_lock(spinlock_t *lock)
5639 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5642 if (spin_needbreak(lock) || resched) {
5644 if (resched && need_resched())
5653 EXPORT_SYMBOL(cond_resched_lock);
5655 int __sched cond_resched_softirq(void)
5657 BUG_ON(!in_softirq());
5659 if (need_resched() && system_state == SYSTEM_RUNNING) {
5667 EXPORT_SYMBOL(cond_resched_softirq);
5670 * yield - yield the current processor to other threads.
5672 * This is a shortcut for kernel-space yielding - it marks the
5673 * thread runnable and calls sys_sched_yield().
5675 void __sched yield(void)
5677 set_current_state(TASK_RUNNING);
5680 EXPORT_SYMBOL(yield);
5683 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5684 * that process accounting knows that this is a task in IO wait state.
5686 * But don't do that if it is a deliberate, throttling IO wait (this task
5687 * has set its backing_dev_info: the queue against which it should throttle)
5689 void __sched io_schedule(void)
5691 struct rq *rq = &__raw_get_cpu_var(runqueues);
5693 delayacct_blkio_start();
5694 atomic_inc(&rq->nr_iowait);
5696 atomic_dec(&rq->nr_iowait);
5697 delayacct_blkio_end();
5699 EXPORT_SYMBOL(io_schedule);
5701 long __sched io_schedule_timeout(long timeout)
5703 struct rq *rq = &__raw_get_cpu_var(runqueues);
5706 delayacct_blkio_start();
5707 atomic_inc(&rq->nr_iowait);
5708 ret = schedule_timeout(timeout);
5709 atomic_dec(&rq->nr_iowait);
5710 delayacct_blkio_end();
5715 * sys_sched_get_priority_max - return maximum RT priority.
5716 * @policy: scheduling class.
5718 * this syscall returns the maximum rt_priority that can be used
5719 * by a given scheduling class.
5721 asmlinkage long sys_sched_get_priority_max(int policy)
5728 ret = MAX_USER_RT_PRIO-1;
5740 * sys_sched_get_priority_min - return minimum RT priority.
5741 * @policy: scheduling class.
5743 * this syscall returns the minimum rt_priority that can be used
5744 * by a given scheduling class.
5746 asmlinkage long sys_sched_get_priority_min(int policy)
5764 * sys_sched_rr_get_interval - return the default timeslice of a process.
5765 * @pid: pid of the process.
5766 * @interval: userspace pointer to the timeslice value.
5768 * this syscall writes the default timeslice value of a given process
5769 * into the user-space timespec buffer. A value of '0' means infinity.
5772 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5774 struct task_struct *p;
5775 unsigned int time_slice;
5783 read_lock(&tasklist_lock);
5784 p = find_process_by_pid(pid);
5788 retval = security_task_getscheduler(p);
5793 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5794 * tasks that are on an otherwise idle runqueue:
5797 if (p->policy == SCHED_RR) {
5798 time_slice = DEF_TIMESLICE;
5799 } else if (p->policy != SCHED_FIFO) {
5800 struct sched_entity *se = &p->se;
5801 unsigned long flags;
5804 rq = task_rq_lock(p, &flags);
5805 if (rq->cfs.load.weight)
5806 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5807 task_rq_unlock(rq, &flags);
5809 read_unlock(&tasklist_lock);
5810 jiffies_to_timespec(time_slice, &t);
5811 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5815 read_unlock(&tasklist_lock);
5819 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5821 void sched_show_task(struct task_struct *p)
5823 unsigned long free = 0;
5826 state = p->state ? __ffs(p->state) + 1 : 0;
5827 printk(KERN_INFO "%-13.13s %c", p->comm,
5828 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5829 #if BITS_PER_LONG == 32
5830 if (state == TASK_RUNNING)
5831 printk(KERN_CONT " running ");
5833 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5835 if (state == TASK_RUNNING)
5836 printk(KERN_CONT " running task ");
5838 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5840 #ifdef CONFIG_DEBUG_STACK_USAGE
5842 unsigned long *n = end_of_stack(p);
5845 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5848 printk(KERN_CONT "%5lu %5d %6d\n", free,
5849 task_pid_nr(p), task_pid_nr(p->real_parent));
5851 show_stack(p, NULL);
5854 void show_state_filter(unsigned long state_filter)
5856 struct task_struct *g, *p;
5858 #if BITS_PER_LONG == 32
5860 " task PC stack pid father\n");
5863 " task PC stack pid father\n");
5865 read_lock(&tasklist_lock);
5866 do_each_thread(g, p) {
5868 * reset the NMI-timeout, listing all files on a slow
5869 * console might take alot of time:
5871 touch_nmi_watchdog();
5872 if (!state_filter || (p->state & state_filter))
5874 } while_each_thread(g, p);
5876 touch_all_softlockup_watchdogs();
5878 #ifdef CONFIG_SCHED_DEBUG
5879 sysrq_sched_debug_show();
5881 read_unlock(&tasklist_lock);
5883 * Only show locks if all tasks are dumped:
5885 if (state_filter == -1)
5886 debug_show_all_locks();
5889 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5891 idle->sched_class = &idle_sched_class;
5895 * init_idle - set up an idle thread for a given CPU
5896 * @idle: task in question
5897 * @cpu: cpu the idle task belongs to
5899 * NOTE: this function does not set the idle thread's NEED_RESCHED
5900 * flag, to make booting more robust.
5902 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5904 struct rq *rq = cpu_rq(cpu);
5905 unsigned long flags;
5908 idle->se.exec_start = sched_clock();
5910 idle->prio = idle->normal_prio = MAX_PRIO;
5911 idle->cpus_allowed = cpumask_of_cpu(cpu);
5912 __set_task_cpu(idle, cpu);
5914 spin_lock_irqsave(&rq->lock, flags);
5915 rq->curr = rq->idle = idle;
5916 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5919 spin_unlock_irqrestore(&rq->lock, flags);
5921 /* Set the preempt count _outside_ the spinlocks! */
5922 #if defined(CONFIG_PREEMPT)
5923 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5925 task_thread_info(idle)->preempt_count = 0;
5928 * The idle tasks have their own, simple scheduling class:
5930 idle->sched_class = &idle_sched_class;
5934 * In a system that switches off the HZ timer nohz_cpu_mask
5935 * indicates which cpus entered this state. This is used
5936 * in the rcu update to wait only for active cpus. For system
5937 * which do not switch off the HZ timer nohz_cpu_mask should
5938 * always be CPU_MASK_NONE.
5940 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5943 * Increase the granularity value when there are more CPUs,
5944 * because with more CPUs the 'effective latency' as visible
5945 * to users decreases. But the relationship is not linear,
5946 * so pick a second-best guess by going with the log2 of the
5949 * This idea comes from the SD scheduler of Con Kolivas:
5951 static inline void sched_init_granularity(void)
5953 unsigned int factor = 1 + ilog2(num_online_cpus());
5954 const unsigned long limit = 200000000;
5956 sysctl_sched_min_granularity *= factor;
5957 if (sysctl_sched_min_granularity > limit)
5958 sysctl_sched_min_granularity = limit;
5960 sysctl_sched_latency *= factor;
5961 if (sysctl_sched_latency > limit)
5962 sysctl_sched_latency = limit;
5964 sysctl_sched_wakeup_granularity *= factor;
5969 * This is how migration works:
5971 * 1) we queue a struct migration_req structure in the source CPU's
5972 * runqueue and wake up that CPU's migration thread.
5973 * 2) we down() the locked semaphore => thread blocks.
5974 * 3) migration thread wakes up (implicitly it forces the migrated
5975 * thread off the CPU)
5976 * 4) it gets the migration request and checks whether the migrated
5977 * task is still in the wrong runqueue.
5978 * 5) if it's in the wrong runqueue then the migration thread removes
5979 * it and puts it into the right queue.
5980 * 6) migration thread up()s the semaphore.
5981 * 7) we wake up and the migration is done.
5985 * Change a given task's CPU affinity. Migrate the thread to a
5986 * proper CPU and schedule it away if the CPU it's executing on
5987 * is removed from the allowed bitmask.
5989 * NOTE: the caller must have a valid reference to the task, the
5990 * task must not exit() & deallocate itself prematurely. The
5991 * call is not atomic; no spinlocks may be held.
5993 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5995 struct migration_req req;
5996 unsigned long flags;
6000 rq = task_rq_lock(p, &flags);
6001 if (!cpus_intersects(*new_mask, cpu_online_map)) {
6006 if (p->sched_class->set_cpus_allowed)
6007 p->sched_class->set_cpus_allowed(p, new_mask);
6009 p->cpus_allowed = *new_mask;
6010 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
6013 /* Can the task run on the task's current CPU? If so, we're done */
6014 if (cpu_isset(task_cpu(p), *new_mask))
6017 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
6018 /* Need help from migration thread: drop lock and wait. */
6019 task_rq_unlock(rq, &flags);
6020 wake_up_process(rq->migration_thread);
6021 wait_for_completion(&req.done);
6022 tlb_migrate_finish(p->mm);
6026 task_rq_unlock(rq, &flags);
6030 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6033 * Move (not current) task off this cpu, onto dest cpu. We're doing
6034 * this because either it can't run here any more (set_cpus_allowed()
6035 * away from this CPU, or CPU going down), or because we're
6036 * attempting to rebalance this task on exec (sched_exec).
6038 * So we race with normal scheduler movements, but that's OK, as long
6039 * as the task is no longer on this CPU.
6041 * Returns non-zero if task was successfully migrated.
6043 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6045 struct rq *rq_dest, *rq_src;
6048 if (unlikely(cpu_is_offline(dest_cpu)))
6051 rq_src = cpu_rq(src_cpu);
6052 rq_dest = cpu_rq(dest_cpu);
6054 double_rq_lock(rq_src, rq_dest);
6055 /* Already moved. */
6056 if (task_cpu(p) != src_cpu)
6058 /* Affinity changed (again). */
6059 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6062 on_rq = p->se.on_rq;
6064 deactivate_task(rq_src, p, 0);
6066 set_task_cpu(p, dest_cpu);
6068 activate_task(rq_dest, p, 0);
6069 check_preempt_curr(rq_dest, p);
6073 double_rq_unlock(rq_src, rq_dest);
6078 * migration_thread - this is a highprio system thread that performs
6079 * thread migration by bumping thread off CPU then 'pushing' onto
6082 static int migration_thread(void *data)
6084 int cpu = (long)data;
6088 BUG_ON(rq->migration_thread != current);
6090 set_current_state(TASK_INTERRUPTIBLE);
6091 while (!kthread_should_stop()) {
6092 struct migration_req *req;
6093 struct list_head *head;
6095 spin_lock_irq(&rq->lock);
6097 if (cpu_is_offline(cpu)) {
6098 spin_unlock_irq(&rq->lock);
6102 if (rq->active_balance) {
6103 active_load_balance(rq, cpu);
6104 rq->active_balance = 0;
6107 head = &rq->migration_queue;
6109 if (list_empty(head)) {
6110 spin_unlock_irq(&rq->lock);
6112 set_current_state(TASK_INTERRUPTIBLE);
6115 req = list_entry(head->next, struct migration_req, list);
6116 list_del_init(head->next);
6118 spin_unlock(&rq->lock);
6119 __migrate_task(req->task, cpu, req->dest_cpu);
6122 complete(&req->done);
6124 __set_current_state(TASK_RUNNING);
6128 /* Wait for kthread_stop */
6129 set_current_state(TASK_INTERRUPTIBLE);
6130 while (!kthread_should_stop()) {
6132 set_current_state(TASK_INTERRUPTIBLE);
6134 __set_current_state(TASK_RUNNING);
6138 #ifdef CONFIG_HOTPLUG_CPU
6140 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6144 local_irq_disable();
6145 ret = __migrate_task(p, src_cpu, dest_cpu);
6151 * Figure out where task on dead CPU should go, use force if necessary.
6152 * NOTE: interrupts should be disabled by the caller
6154 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6156 unsigned long flags;
6163 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6164 cpus_and(mask, mask, p->cpus_allowed);
6165 dest_cpu = any_online_cpu(mask);
6167 /* On any allowed CPU? */
6168 if (dest_cpu >= nr_cpu_ids)
6169 dest_cpu = any_online_cpu(p->cpus_allowed);
6171 /* No more Mr. Nice Guy. */
6172 if (dest_cpu >= nr_cpu_ids) {
6173 cpumask_t cpus_allowed;
6175 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6177 * Try to stay on the same cpuset, where the
6178 * current cpuset may be a subset of all cpus.
6179 * The cpuset_cpus_allowed_locked() variant of
6180 * cpuset_cpus_allowed() will not block. It must be
6181 * called within calls to cpuset_lock/cpuset_unlock.
6183 rq = task_rq_lock(p, &flags);
6184 p->cpus_allowed = cpus_allowed;
6185 dest_cpu = any_online_cpu(p->cpus_allowed);
6186 task_rq_unlock(rq, &flags);
6189 * Don't tell them about moving exiting tasks or
6190 * kernel threads (both mm NULL), since they never
6193 if (p->mm && printk_ratelimit()) {
6194 printk(KERN_INFO "process %d (%s) no "
6195 "longer affine to cpu%d\n",
6196 task_pid_nr(p), p->comm, dead_cpu);
6199 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6203 * While a dead CPU has no uninterruptible tasks queued at this point,
6204 * it might still have a nonzero ->nr_uninterruptible counter, because
6205 * for performance reasons the counter is not stricly tracking tasks to
6206 * their home CPUs. So we just add the counter to another CPU's counter,
6207 * to keep the global sum constant after CPU-down:
6209 static void migrate_nr_uninterruptible(struct rq *rq_src)
6211 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6212 unsigned long flags;
6214 local_irq_save(flags);
6215 double_rq_lock(rq_src, rq_dest);
6216 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6217 rq_src->nr_uninterruptible = 0;
6218 double_rq_unlock(rq_src, rq_dest);
6219 local_irq_restore(flags);
6222 /* Run through task list and migrate tasks from the dead cpu. */
6223 static void migrate_live_tasks(int src_cpu)
6225 struct task_struct *p, *t;
6227 read_lock(&tasklist_lock);
6229 do_each_thread(t, p) {
6233 if (task_cpu(p) == src_cpu)
6234 move_task_off_dead_cpu(src_cpu, p);
6235 } while_each_thread(t, p);
6237 read_unlock(&tasklist_lock);
6241 * Schedules idle task to be the next runnable task on current CPU.
6242 * It does so by boosting its priority to highest possible.
6243 * Used by CPU offline code.
6245 void sched_idle_next(void)
6247 int this_cpu = smp_processor_id();
6248 struct rq *rq = cpu_rq(this_cpu);
6249 struct task_struct *p = rq->idle;
6250 unsigned long flags;
6252 /* cpu has to be offline */
6253 BUG_ON(cpu_online(this_cpu));
6256 * Strictly not necessary since rest of the CPUs are stopped by now
6257 * and interrupts disabled on the current cpu.
6259 spin_lock_irqsave(&rq->lock, flags);
6261 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6263 update_rq_clock(rq);
6264 activate_task(rq, p, 0);
6266 spin_unlock_irqrestore(&rq->lock, flags);
6270 * Ensures that the idle task is using init_mm right before its cpu goes
6273 void idle_task_exit(void)
6275 struct mm_struct *mm = current->active_mm;
6277 BUG_ON(cpu_online(smp_processor_id()));
6280 switch_mm(mm, &init_mm, current);
6284 /* called under rq->lock with disabled interrupts */
6285 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6287 struct rq *rq = cpu_rq(dead_cpu);
6289 /* Must be exiting, otherwise would be on tasklist. */
6290 BUG_ON(!p->exit_state);
6292 /* Cannot have done final schedule yet: would have vanished. */
6293 BUG_ON(p->state == TASK_DEAD);
6298 * Drop lock around migration; if someone else moves it,
6299 * that's OK. No task can be added to this CPU, so iteration is
6302 spin_unlock_irq(&rq->lock);
6303 move_task_off_dead_cpu(dead_cpu, p);
6304 spin_lock_irq(&rq->lock);
6309 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6310 static void migrate_dead_tasks(unsigned int dead_cpu)
6312 struct rq *rq = cpu_rq(dead_cpu);
6313 struct task_struct *next;
6316 if (!rq->nr_running)
6318 update_rq_clock(rq);
6319 next = pick_next_task(rq, rq->curr);
6322 migrate_dead(dead_cpu, next);
6326 #endif /* CONFIG_HOTPLUG_CPU */
6328 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6330 static struct ctl_table sd_ctl_dir[] = {
6332 .procname = "sched_domain",
6338 static struct ctl_table sd_ctl_root[] = {
6340 .ctl_name = CTL_KERN,
6341 .procname = "kernel",
6343 .child = sd_ctl_dir,
6348 static struct ctl_table *sd_alloc_ctl_entry(int n)
6350 struct ctl_table *entry =
6351 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6356 static void sd_free_ctl_entry(struct ctl_table **tablep)
6358 struct ctl_table *entry;
6361 * In the intermediate directories, both the child directory and
6362 * procname are dynamically allocated and could fail but the mode
6363 * will always be set. In the lowest directory the names are
6364 * static strings and all have proc handlers.
6366 for (entry = *tablep; entry->mode; entry++) {
6368 sd_free_ctl_entry(&entry->child);
6369 if (entry->proc_handler == NULL)
6370 kfree(entry->procname);
6378 set_table_entry(struct ctl_table *entry,
6379 const char *procname, void *data, int maxlen,
6380 mode_t mode, proc_handler *proc_handler)
6382 entry->procname = procname;
6384 entry->maxlen = maxlen;
6386 entry->proc_handler = proc_handler;
6389 static struct ctl_table *
6390 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6392 struct ctl_table *table = sd_alloc_ctl_entry(12);
6397 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6398 sizeof(long), 0644, proc_doulongvec_minmax);
6399 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6400 sizeof(long), 0644, proc_doulongvec_minmax);
6401 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6402 sizeof(int), 0644, proc_dointvec_minmax);
6403 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6404 sizeof(int), 0644, proc_dointvec_minmax);
6405 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6406 sizeof(int), 0644, proc_dointvec_minmax);
6407 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6408 sizeof(int), 0644, proc_dointvec_minmax);
6409 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6410 sizeof(int), 0644, proc_dointvec_minmax);
6411 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6412 sizeof(int), 0644, proc_dointvec_minmax);
6413 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6414 sizeof(int), 0644, proc_dointvec_minmax);
6415 set_table_entry(&table[9], "cache_nice_tries",
6416 &sd->cache_nice_tries,
6417 sizeof(int), 0644, proc_dointvec_minmax);
6418 set_table_entry(&table[10], "flags", &sd->flags,
6419 sizeof(int), 0644, proc_dointvec_minmax);
6420 /* &table[11] is terminator */
6425 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6427 struct ctl_table *entry, *table;
6428 struct sched_domain *sd;
6429 int domain_num = 0, i;
6432 for_each_domain(cpu, sd)
6434 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6439 for_each_domain(cpu, sd) {
6440 snprintf(buf, 32, "domain%d", i);
6441 entry->procname = kstrdup(buf, GFP_KERNEL);
6443 entry->child = sd_alloc_ctl_domain_table(sd);
6450 static struct ctl_table_header *sd_sysctl_header;
6451 static void register_sched_domain_sysctl(void)
6453 int i, cpu_num = num_online_cpus();
6454 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6457 WARN_ON(sd_ctl_dir[0].child);
6458 sd_ctl_dir[0].child = entry;
6463 for_each_online_cpu(i) {
6464 snprintf(buf, 32, "cpu%d", i);
6465 entry->procname = kstrdup(buf, GFP_KERNEL);
6467 entry->child = sd_alloc_ctl_cpu_table(i);
6471 WARN_ON(sd_sysctl_header);
6472 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6475 /* may be called multiple times per register */
6476 static void unregister_sched_domain_sysctl(void)
6478 if (sd_sysctl_header)
6479 unregister_sysctl_table(sd_sysctl_header);
6480 sd_sysctl_header = NULL;
6481 if (sd_ctl_dir[0].child)
6482 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6485 static void register_sched_domain_sysctl(void)
6488 static void unregister_sched_domain_sysctl(void)
6494 * migration_call - callback that gets triggered when a CPU is added.
6495 * Here we can start up the necessary migration thread for the new CPU.
6497 static int __cpuinit
6498 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6500 struct task_struct *p;
6501 int cpu = (long)hcpu;
6502 unsigned long flags;
6507 case CPU_UP_PREPARE:
6508 case CPU_UP_PREPARE_FROZEN:
6509 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6512 kthread_bind(p, cpu);
6513 /* Must be high prio: stop_machine expects to yield to it. */
6514 rq = task_rq_lock(p, &flags);
6515 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6516 task_rq_unlock(rq, &flags);
6517 cpu_rq(cpu)->migration_thread = p;
6521 case CPU_ONLINE_FROZEN:
6522 /* Strictly unnecessary, as first user will wake it. */
6523 wake_up_process(cpu_rq(cpu)->migration_thread);
6525 /* Update our root-domain */
6527 spin_lock_irqsave(&rq->lock, flags);
6529 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6530 cpu_set(cpu, rq->rd->online);
6532 spin_unlock_irqrestore(&rq->lock, flags);
6535 #ifdef CONFIG_HOTPLUG_CPU
6536 case CPU_UP_CANCELED:
6537 case CPU_UP_CANCELED_FROZEN:
6538 if (!cpu_rq(cpu)->migration_thread)
6540 /* Unbind it from offline cpu so it can run. Fall thru. */
6541 kthread_bind(cpu_rq(cpu)->migration_thread,
6542 any_online_cpu(cpu_online_map));
6543 kthread_stop(cpu_rq(cpu)->migration_thread);
6544 cpu_rq(cpu)->migration_thread = NULL;
6548 case CPU_DEAD_FROZEN:
6549 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6550 migrate_live_tasks(cpu);
6552 kthread_stop(rq->migration_thread);
6553 rq->migration_thread = NULL;
6554 /* Idle task back to normal (off runqueue, low prio) */
6555 spin_lock_irq(&rq->lock);
6556 update_rq_clock(rq);
6557 deactivate_task(rq, rq->idle, 0);
6558 rq->idle->static_prio = MAX_PRIO;
6559 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6560 rq->idle->sched_class = &idle_sched_class;
6561 migrate_dead_tasks(cpu);
6562 spin_unlock_irq(&rq->lock);
6564 migrate_nr_uninterruptible(rq);
6565 BUG_ON(rq->nr_running != 0);
6568 * No need to migrate the tasks: it was best-effort if
6569 * they didn't take sched_hotcpu_mutex. Just wake up
6572 spin_lock_irq(&rq->lock);
6573 while (!list_empty(&rq->migration_queue)) {
6574 struct migration_req *req;
6576 req = list_entry(rq->migration_queue.next,
6577 struct migration_req, list);
6578 list_del_init(&req->list);
6579 complete(&req->done);
6581 spin_unlock_irq(&rq->lock);
6585 case CPU_DYING_FROZEN:
6586 /* Update our root-domain */
6588 spin_lock_irqsave(&rq->lock, flags);
6590 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6591 cpu_clear(cpu, rq->rd->online);
6593 spin_unlock_irqrestore(&rq->lock, flags);
6600 /* Register at highest priority so that task migration (migrate_all_tasks)
6601 * happens before everything else.
6603 static struct notifier_block __cpuinitdata migration_notifier = {
6604 .notifier_call = migration_call,
6608 void __init migration_init(void)
6610 void *cpu = (void *)(long)smp_processor_id();
6613 /* Start one for the boot CPU: */
6614 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6615 BUG_ON(err == NOTIFY_BAD);
6616 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6617 register_cpu_notifier(&migration_notifier);
6623 #ifdef CONFIG_SCHED_DEBUG
6625 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6626 cpumask_t *groupmask)
6628 struct sched_group *group = sd->groups;
6631 cpulist_scnprintf(str, sizeof(str), sd->span);
6632 cpus_clear(*groupmask);
6634 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6636 if (!(sd->flags & SD_LOAD_BALANCE)) {
6637 printk("does not load-balance\n");
6639 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6644 printk(KERN_CONT "span %s\n", str);
6646 if (!cpu_isset(cpu, sd->span)) {
6647 printk(KERN_ERR "ERROR: domain->span does not contain "
6650 if (!cpu_isset(cpu, group->cpumask)) {
6651 printk(KERN_ERR "ERROR: domain->groups does not contain"
6655 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6659 printk(KERN_ERR "ERROR: group is NULL\n");
6663 if (!group->__cpu_power) {
6664 printk(KERN_CONT "\n");
6665 printk(KERN_ERR "ERROR: domain->cpu_power not "
6670 if (!cpus_weight(group->cpumask)) {
6671 printk(KERN_CONT "\n");
6672 printk(KERN_ERR "ERROR: empty group\n");
6676 if (cpus_intersects(*groupmask, group->cpumask)) {
6677 printk(KERN_CONT "\n");
6678 printk(KERN_ERR "ERROR: repeated CPUs\n");
6682 cpus_or(*groupmask, *groupmask, group->cpumask);
6684 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6685 printk(KERN_CONT " %s", str);
6687 group = group->next;
6688 } while (group != sd->groups);
6689 printk(KERN_CONT "\n");
6691 if (!cpus_equal(sd->span, *groupmask))
6692 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6694 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6695 printk(KERN_ERR "ERROR: parent span is not a superset "
6696 "of domain->span\n");
6700 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6702 cpumask_t *groupmask;
6706 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6710 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6712 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6714 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6719 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6729 # define sched_domain_debug(sd, cpu) do { } while (0)
6732 static int sd_degenerate(struct sched_domain *sd)
6734 if (cpus_weight(sd->span) == 1)
6737 /* Following flags need at least 2 groups */
6738 if (sd->flags & (SD_LOAD_BALANCE |
6739 SD_BALANCE_NEWIDLE |
6743 SD_SHARE_PKG_RESOURCES)) {
6744 if (sd->groups != sd->groups->next)
6748 /* Following flags don't use groups */
6749 if (sd->flags & (SD_WAKE_IDLE |
6758 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6760 unsigned long cflags = sd->flags, pflags = parent->flags;
6762 if (sd_degenerate(parent))
6765 if (!cpus_equal(sd->span, parent->span))
6768 /* Does parent contain flags not in child? */
6769 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6770 if (cflags & SD_WAKE_AFFINE)
6771 pflags &= ~SD_WAKE_BALANCE;
6772 /* Flags needing groups don't count if only 1 group in parent */
6773 if (parent->groups == parent->groups->next) {
6774 pflags &= ~(SD_LOAD_BALANCE |
6775 SD_BALANCE_NEWIDLE |
6779 SD_SHARE_PKG_RESOURCES);
6781 if (~cflags & pflags)
6787 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6789 unsigned long flags;
6790 const struct sched_class *class;
6792 spin_lock_irqsave(&rq->lock, flags);
6795 struct root_domain *old_rd = rq->rd;
6797 for (class = sched_class_highest; class; class = class->next) {
6798 if (class->leave_domain)
6799 class->leave_domain(rq);
6802 cpu_clear(rq->cpu, old_rd->span);
6803 cpu_clear(rq->cpu, old_rd->online);
6805 if (atomic_dec_and_test(&old_rd->refcount))
6809 atomic_inc(&rd->refcount);
6812 cpu_set(rq->cpu, rd->span);
6813 if (cpu_isset(rq->cpu, cpu_online_map))
6814 cpu_set(rq->cpu, rd->online);
6816 for (class = sched_class_highest; class; class = class->next) {
6817 if (class->join_domain)
6818 class->join_domain(rq);
6821 spin_unlock_irqrestore(&rq->lock, flags);
6824 static void init_rootdomain(struct root_domain *rd)
6826 memset(rd, 0, sizeof(*rd));
6828 cpus_clear(rd->span);
6829 cpus_clear(rd->online);
6832 static void init_defrootdomain(void)
6834 init_rootdomain(&def_root_domain);
6835 atomic_set(&def_root_domain.refcount, 1);
6838 static struct root_domain *alloc_rootdomain(void)
6840 struct root_domain *rd;
6842 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6846 init_rootdomain(rd);
6852 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6853 * hold the hotplug lock.
6856 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6858 struct rq *rq = cpu_rq(cpu);
6859 struct sched_domain *tmp;
6861 /* Remove the sched domains which do not contribute to scheduling. */
6862 for (tmp = sd; tmp; tmp = tmp->parent) {
6863 struct sched_domain *parent = tmp->parent;
6866 if (sd_parent_degenerate(tmp, parent)) {
6867 tmp->parent = parent->parent;
6869 parent->parent->child = tmp;
6873 if (sd && sd_degenerate(sd)) {
6879 sched_domain_debug(sd, cpu);
6881 rq_attach_root(rq, rd);
6882 rcu_assign_pointer(rq->sd, sd);
6885 /* cpus with isolated domains */
6886 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6888 /* Setup the mask of cpus configured for isolated domains */
6889 static int __init isolated_cpu_setup(char *str)
6891 int ints[NR_CPUS], i;
6893 str = get_options(str, ARRAY_SIZE(ints), ints);
6894 cpus_clear(cpu_isolated_map);
6895 for (i = 1; i <= ints[0]; i++)
6896 if (ints[i] < NR_CPUS)
6897 cpu_set(ints[i], cpu_isolated_map);
6901 __setup("isolcpus=", isolated_cpu_setup);
6904 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6905 * to a function which identifies what group(along with sched group) a CPU
6906 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6907 * (due to the fact that we keep track of groups covered with a cpumask_t).
6909 * init_sched_build_groups will build a circular linked list of the groups
6910 * covered by the given span, and will set each group's ->cpumask correctly,
6911 * and ->cpu_power to 0.
6914 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6915 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6916 struct sched_group **sg,
6917 cpumask_t *tmpmask),
6918 cpumask_t *covered, cpumask_t *tmpmask)
6920 struct sched_group *first = NULL, *last = NULL;
6923 cpus_clear(*covered);
6925 for_each_cpu_mask(i, *span) {
6926 struct sched_group *sg;
6927 int group = group_fn(i, cpu_map, &sg, tmpmask);
6930 if (cpu_isset(i, *covered))
6933 cpus_clear(sg->cpumask);
6934 sg->__cpu_power = 0;
6936 for_each_cpu_mask(j, *span) {
6937 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6940 cpu_set(j, *covered);
6941 cpu_set(j, sg->cpumask);
6952 #define SD_NODES_PER_DOMAIN 16
6957 * find_next_best_node - find the next node to include in a sched_domain
6958 * @node: node whose sched_domain we're building
6959 * @used_nodes: nodes already in the sched_domain
6961 * Find the next node to include in a given scheduling domain. Simply
6962 * finds the closest node not already in the @used_nodes map.
6964 * Should use nodemask_t.
6966 static int find_next_best_node(int node, nodemask_t *used_nodes)
6968 int i, n, val, min_val, best_node = 0;
6972 for (i = 0; i < MAX_NUMNODES; i++) {
6973 /* Start at @node */
6974 n = (node + i) % MAX_NUMNODES;
6976 if (!nr_cpus_node(n))
6979 /* Skip already used nodes */
6980 if (node_isset(n, *used_nodes))
6983 /* Simple min distance search */
6984 val = node_distance(node, n);
6986 if (val < min_val) {
6992 node_set(best_node, *used_nodes);
6997 * sched_domain_node_span - get a cpumask for a node's sched_domain
6998 * @node: node whose cpumask we're constructing
6999 * @span: resulting cpumask
7001 * Given a node, construct a good cpumask for its sched_domain to span. It
7002 * should be one that prevents unnecessary balancing, but also spreads tasks
7005 static void sched_domain_node_span(int node, cpumask_t *span)
7007 nodemask_t used_nodes;
7008 node_to_cpumask_ptr(nodemask, node);
7012 nodes_clear(used_nodes);
7014 cpus_or(*span, *span, *nodemask);
7015 node_set(node, used_nodes);
7017 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7018 int next_node = find_next_best_node(node, &used_nodes);
7020 node_to_cpumask_ptr_next(nodemask, next_node);
7021 cpus_or(*span, *span, *nodemask);
7026 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7029 * SMT sched-domains:
7031 #ifdef CONFIG_SCHED_SMT
7032 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
7033 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7036 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7040 *sg = &per_cpu(sched_group_cpus, cpu);
7046 * multi-core sched-domains:
7048 #ifdef CONFIG_SCHED_MC
7049 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7050 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7053 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7055 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7060 *mask = per_cpu(cpu_sibling_map, cpu);
7061 cpus_and(*mask, *mask, *cpu_map);
7062 group = first_cpu(*mask);
7064 *sg = &per_cpu(sched_group_core, group);
7067 #elif defined(CONFIG_SCHED_MC)
7069 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7073 *sg = &per_cpu(sched_group_core, cpu);
7078 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7079 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7082 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7086 #ifdef CONFIG_SCHED_MC
7087 *mask = cpu_coregroup_map(cpu);
7088 cpus_and(*mask, *mask, *cpu_map);
7089 group = first_cpu(*mask);
7090 #elif defined(CONFIG_SCHED_SMT)
7091 *mask = per_cpu(cpu_sibling_map, cpu);
7092 cpus_and(*mask, *mask, *cpu_map);
7093 group = first_cpu(*mask);
7098 *sg = &per_cpu(sched_group_phys, group);
7104 * The init_sched_build_groups can't handle what we want to do with node
7105 * groups, so roll our own. Now each node has its own list of groups which
7106 * gets dynamically allocated.
7108 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7109 static struct sched_group ***sched_group_nodes_bycpu;
7111 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7112 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7114 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7115 struct sched_group **sg, cpumask_t *nodemask)
7119 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7120 cpus_and(*nodemask, *nodemask, *cpu_map);
7121 group = first_cpu(*nodemask);
7124 *sg = &per_cpu(sched_group_allnodes, group);
7128 static void init_numa_sched_groups_power(struct sched_group *group_head)
7130 struct sched_group *sg = group_head;
7136 for_each_cpu_mask(j, sg->cpumask) {
7137 struct sched_domain *sd;
7139 sd = &per_cpu(phys_domains, j);
7140 if (j != first_cpu(sd->groups->cpumask)) {
7142 * Only add "power" once for each
7148 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7151 } while (sg != group_head);
7156 /* Free memory allocated for various sched_group structures */
7157 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7161 for_each_cpu_mask(cpu, *cpu_map) {
7162 struct sched_group **sched_group_nodes
7163 = sched_group_nodes_bycpu[cpu];
7165 if (!sched_group_nodes)
7168 for (i = 0; i < MAX_NUMNODES; i++) {
7169 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7171 *nodemask = node_to_cpumask(i);
7172 cpus_and(*nodemask, *nodemask, *cpu_map);
7173 if (cpus_empty(*nodemask))
7183 if (oldsg != sched_group_nodes[i])
7186 kfree(sched_group_nodes);
7187 sched_group_nodes_bycpu[cpu] = NULL;
7191 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7197 * Initialize sched groups cpu_power.
7199 * cpu_power indicates the capacity of sched group, which is used while
7200 * distributing the load between different sched groups in a sched domain.
7201 * Typically cpu_power for all the groups in a sched domain will be same unless
7202 * there are asymmetries in the topology. If there are asymmetries, group
7203 * having more cpu_power will pickup more load compared to the group having
7206 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7207 * the maximum number of tasks a group can handle in the presence of other idle
7208 * or lightly loaded groups in the same sched domain.
7210 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7212 struct sched_domain *child;
7213 struct sched_group *group;
7215 WARN_ON(!sd || !sd->groups);
7217 if (cpu != first_cpu(sd->groups->cpumask))
7222 sd->groups->__cpu_power = 0;
7225 * For perf policy, if the groups in child domain share resources
7226 * (for example cores sharing some portions of the cache hierarchy
7227 * or SMT), then set this domain groups cpu_power such that each group
7228 * can handle only one task, when there are other idle groups in the
7229 * same sched domain.
7231 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7233 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7234 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7239 * add cpu_power of each child group to this groups cpu_power
7241 group = child->groups;
7243 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7244 group = group->next;
7245 } while (group != child->groups);
7249 * Initializers for schedule domains
7250 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7253 #define SD_INIT(sd, type) sd_init_##type(sd)
7254 #define SD_INIT_FUNC(type) \
7255 static noinline void sd_init_##type(struct sched_domain *sd) \
7257 memset(sd, 0, sizeof(*sd)); \
7258 *sd = SD_##type##_INIT; \
7259 sd->level = SD_LV_##type; \
7264 SD_INIT_FUNC(ALLNODES)
7267 #ifdef CONFIG_SCHED_SMT
7268 SD_INIT_FUNC(SIBLING)
7270 #ifdef CONFIG_SCHED_MC
7275 * To minimize stack usage kmalloc room for cpumasks and share the
7276 * space as the usage in build_sched_domains() dictates. Used only
7277 * if the amount of space is significant.
7280 cpumask_t tmpmask; /* make this one first */
7283 cpumask_t this_sibling_map;
7284 cpumask_t this_core_map;
7286 cpumask_t send_covered;
7289 cpumask_t domainspan;
7291 cpumask_t notcovered;
7296 #define SCHED_CPUMASK_ALLOC 1
7297 #define SCHED_CPUMASK_FREE(v) kfree(v)
7298 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7300 #define SCHED_CPUMASK_ALLOC 0
7301 #define SCHED_CPUMASK_FREE(v)
7302 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7305 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7306 ((unsigned long)(a) + offsetof(struct allmasks, v))
7308 static int default_relax_domain_level = -1;
7310 static int __init setup_relax_domain_level(char *str)
7312 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7315 __setup("relax_domain_level=", setup_relax_domain_level);
7317 static void set_domain_attribute(struct sched_domain *sd,
7318 struct sched_domain_attr *attr)
7322 if (!attr || attr->relax_domain_level < 0) {
7323 if (default_relax_domain_level < 0)
7326 request = default_relax_domain_level;
7328 request = attr->relax_domain_level;
7329 if (request < sd->level) {
7330 /* turn off idle balance on this domain */
7331 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7333 /* turn on idle balance on this domain */
7334 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7339 * Build sched domains for a given set of cpus and attach the sched domains
7340 * to the individual cpus
7342 static int __build_sched_domains(const cpumask_t *cpu_map,
7343 struct sched_domain_attr *attr)
7346 struct root_domain *rd;
7347 SCHED_CPUMASK_DECLARE(allmasks);
7350 struct sched_group **sched_group_nodes = NULL;
7351 int sd_allnodes = 0;
7354 * Allocate the per-node list of sched groups
7356 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7358 if (!sched_group_nodes) {
7359 printk(KERN_WARNING "Can not alloc sched group node list\n");
7364 rd = alloc_rootdomain();
7366 printk(KERN_WARNING "Cannot alloc root domain\n");
7368 kfree(sched_group_nodes);
7373 #if SCHED_CPUMASK_ALLOC
7374 /* get space for all scratch cpumask variables */
7375 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7377 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7380 kfree(sched_group_nodes);
7385 tmpmask = (cpumask_t *)allmasks;
7389 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7393 * Set up domains for cpus specified by the cpu_map.
7395 for_each_cpu_mask(i, *cpu_map) {
7396 struct sched_domain *sd = NULL, *p;
7397 SCHED_CPUMASK_VAR(nodemask, allmasks);
7399 *nodemask = node_to_cpumask(cpu_to_node(i));
7400 cpus_and(*nodemask, *nodemask, *cpu_map);
7403 if (cpus_weight(*cpu_map) >
7404 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7405 sd = &per_cpu(allnodes_domains, i);
7406 SD_INIT(sd, ALLNODES);
7407 set_domain_attribute(sd, attr);
7408 sd->span = *cpu_map;
7409 sd->first_cpu = first_cpu(sd->span);
7410 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7416 sd = &per_cpu(node_domains, i);
7418 set_domain_attribute(sd, attr);
7419 sched_domain_node_span(cpu_to_node(i), &sd->span);
7420 sd->first_cpu = first_cpu(sd->span);
7424 cpus_and(sd->span, sd->span, *cpu_map);
7428 sd = &per_cpu(phys_domains, i);
7430 set_domain_attribute(sd, attr);
7431 sd->span = *nodemask;
7432 sd->first_cpu = first_cpu(sd->span);
7436 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7438 #ifdef CONFIG_SCHED_MC
7440 sd = &per_cpu(core_domains, i);
7442 set_domain_attribute(sd, attr);
7443 sd->span = cpu_coregroup_map(i);
7444 sd->first_cpu = first_cpu(sd->span);
7445 cpus_and(sd->span, sd->span, *cpu_map);
7448 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7451 #ifdef CONFIG_SCHED_SMT
7453 sd = &per_cpu(cpu_domains, i);
7454 SD_INIT(sd, SIBLING);
7455 set_domain_attribute(sd, attr);
7456 sd->span = per_cpu(cpu_sibling_map, i);
7457 sd->first_cpu = first_cpu(sd->span);
7458 cpus_and(sd->span, sd->span, *cpu_map);
7461 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7465 #ifdef CONFIG_SCHED_SMT
7466 /* Set up CPU (sibling) groups */
7467 for_each_cpu_mask(i, *cpu_map) {
7468 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7469 SCHED_CPUMASK_VAR(send_covered, allmasks);
7471 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7472 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7473 if (i != first_cpu(*this_sibling_map))
7476 init_sched_build_groups(this_sibling_map, cpu_map,
7478 send_covered, tmpmask);
7482 #ifdef CONFIG_SCHED_MC
7483 /* Set up multi-core groups */
7484 for_each_cpu_mask(i, *cpu_map) {
7485 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7486 SCHED_CPUMASK_VAR(send_covered, allmasks);
7488 *this_core_map = cpu_coregroup_map(i);
7489 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7490 if (i != first_cpu(*this_core_map))
7493 init_sched_build_groups(this_core_map, cpu_map,
7495 send_covered, tmpmask);
7499 /* Set up physical groups */
7500 for (i = 0; i < MAX_NUMNODES; i++) {
7501 SCHED_CPUMASK_VAR(nodemask, allmasks);
7502 SCHED_CPUMASK_VAR(send_covered, allmasks);
7504 *nodemask = node_to_cpumask(i);
7505 cpus_and(*nodemask, *nodemask, *cpu_map);
7506 if (cpus_empty(*nodemask))
7509 init_sched_build_groups(nodemask, cpu_map,
7511 send_covered, tmpmask);
7515 /* Set up node groups */
7517 SCHED_CPUMASK_VAR(send_covered, allmasks);
7519 init_sched_build_groups(cpu_map, cpu_map,
7520 &cpu_to_allnodes_group,
7521 send_covered, tmpmask);
7524 for (i = 0; i < MAX_NUMNODES; i++) {
7525 /* Set up node groups */
7526 struct sched_group *sg, *prev;
7527 SCHED_CPUMASK_VAR(nodemask, allmasks);
7528 SCHED_CPUMASK_VAR(domainspan, allmasks);
7529 SCHED_CPUMASK_VAR(covered, allmasks);
7532 *nodemask = node_to_cpumask(i);
7533 cpus_clear(*covered);
7535 cpus_and(*nodemask, *nodemask, *cpu_map);
7536 if (cpus_empty(*nodemask)) {
7537 sched_group_nodes[i] = NULL;
7541 sched_domain_node_span(i, domainspan);
7542 cpus_and(*domainspan, *domainspan, *cpu_map);
7544 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7546 printk(KERN_WARNING "Can not alloc domain group for "
7550 sched_group_nodes[i] = sg;
7551 for_each_cpu_mask(j, *nodemask) {
7552 struct sched_domain *sd;
7554 sd = &per_cpu(node_domains, j);
7557 sg->__cpu_power = 0;
7558 sg->cpumask = *nodemask;
7560 cpus_or(*covered, *covered, *nodemask);
7563 for (j = 0; j < MAX_NUMNODES; j++) {
7564 SCHED_CPUMASK_VAR(notcovered, allmasks);
7565 int n = (i + j) % MAX_NUMNODES;
7566 node_to_cpumask_ptr(pnodemask, n);
7568 cpus_complement(*notcovered, *covered);
7569 cpus_and(*tmpmask, *notcovered, *cpu_map);
7570 cpus_and(*tmpmask, *tmpmask, *domainspan);
7571 if (cpus_empty(*tmpmask))
7574 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7575 if (cpus_empty(*tmpmask))
7578 sg = kmalloc_node(sizeof(struct sched_group),
7582 "Can not alloc domain group for node %d\n", j);
7585 sg->__cpu_power = 0;
7586 sg->cpumask = *tmpmask;
7587 sg->next = prev->next;
7588 cpus_or(*covered, *covered, *tmpmask);
7595 /* Calculate CPU power for physical packages and nodes */
7596 #ifdef CONFIG_SCHED_SMT
7597 for_each_cpu_mask(i, *cpu_map) {
7598 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7600 init_sched_groups_power(i, sd);
7603 #ifdef CONFIG_SCHED_MC
7604 for_each_cpu_mask(i, *cpu_map) {
7605 struct sched_domain *sd = &per_cpu(core_domains, i);
7607 init_sched_groups_power(i, sd);
7611 for_each_cpu_mask(i, *cpu_map) {
7612 struct sched_domain *sd = &per_cpu(phys_domains, i);
7614 init_sched_groups_power(i, sd);
7618 for (i = 0; i < MAX_NUMNODES; i++)
7619 init_numa_sched_groups_power(sched_group_nodes[i]);
7622 struct sched_group *sg;
7624 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7626 init_numa_sched_groups_power(sg);
7630 /* Attach the domains */
7631 for_each_cpu_mask(i, *cpu_map) {
7632 struct sched_domain *sd;
7633 #ifdef CONFIG_SCHED_SMT
7634 sd = &per_cpu(cpu_domains, i);
7635 #elif defined(CONFIG_SCHED_MC)
7636 sd = &per_cpu(core_domains, i);
7638 sd = &per_cpu(phys_domains, i);
7640 cpu_attach_domain(sd, rd, i);
7643 SCHED_CPUMASK_FREE((void *)allmasks);
7648 free_sched_groups(cpu_map, tmpmask);
7649 SCHED_CPUMASK_FREE((void *)allmasks);
7654 static int build_sched_domains(const cpumask_t *cpu_map)
7656 return __build_sched_domains(cpu_map, NULL);
7659 static cpumask_t *doms_cur; /* current sched domains */
7660 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7661 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7665 * Special case: If a kmalloc of a doms_cur partition (array of
7666 * cpumask_t) fails, then fallback to a single sched domain,
7667 * as determined by the single cpumask_t fallback_doms.
7669 static cpumask_t fallback_doms;
7671 void __attribute__((weak)) arch_update_cpu_topology(void)
7676 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7677 * For now this just excludes isolated cpus, but could be used to
7678 * exclude other special cases in the future.
7680 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7684 arch_update_cpu_topology();
7686 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7688 doms_cur = &fallback_doms;
7689 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7691 err = build_sched_domains(doms_cur);
7692 register_sched_domain_sysctl();
7697 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7700 free_sched_groups(cpu_map, tmpmask);
7704 * Detach sched domains from a group of cpus specified in cpu_map
7705 * These cpus will now be attached to the NULL domain
7707 static void detach_destroy_domains(const cpumask_t *cpu_map)
7712 unregister_sched_domain_sysctl();
7714 for_each_cpu_mask(i, *cpu_map)
7715 cpu_attach_domain(NULL, &def_root_domain, i);
7716 synchronize_sched();
7717 arch_destroy_sched_domains(cpu_map, &tmpmask);
7720 /* handle null as "default" */
7721 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7722 struct sched_domain_attr *new, int idx_new)
7724 struct sched_domain_attr tmp;
7731 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7732 new ? (new + idx_new) : &tmp,
7733 sizeof(struct sched_domain_attr));
7737 * Partition sched domains as specified by the 'ndoms_new'
7738 * cpumasks in the array doms_new[] of cpumasks. This compares
7739 * doms_new[] to the current sched domain partitioning, doms_cur[].
7740 * It destroys each deleted domain and builds each new domain.
7742 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7743 * The masks don't intersect (don't overlap.) We should setup one
7744 * sched domain for each mask. CPUs not in any of the cpumasks will
7745 * not be load balanced. If the same cpumask appears both in the
7746 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7749 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7750 * ownership of it and will kfree it when done with it. If the caller
7751 * failed the kmalloc call, then it can pass in doms_new == NULL,
7752 * and partition_sched_domains() will fallback to the single partition
7755 * Call with hotplug lock held
7757 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7758 struct sched_domain_attr *dattr_new)
7762 mutex_lock(&sched_domains_mutex);
7764 /* always unregister in case we don't destroy any domains */
7765 unregister_sched_domain_sysctl();
7767 if (doms_new == NULL) {
7769 doms_new = &fallback_doms;
7770 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7774 /* Destroy deleted domains */
7775 for (i = 0; i < ndoms_cur; i++) {
7776 for (j = 0; j < ndoms_new; j++) {
7777 if (cpus_equal(doms_cur[i], doms_new[j])
7778 && dattrs_equal(dattr_cur, i, dattr_new, j))
7781 /* no match - a current sched domain not in new doms_new[] */
7782 detach_destroy_domains(doms_cur + i);
7787 /* Build new domains */
7788 for (i = 0; i < ndoms_new; i++) {
7789 for (j = 0; j < ndoms_cur; j++) {
7790 if (cpus_equal(doms_new[i], doms_cur[j])
7791 && dattrs_equal(dattr_new, i, dattr_cur, j))
7794 /* no match - add a new doms_new */
7795 __build_sched_domains(doms_new + i,
7796 dattr_new ? dattr_new + i : NULL);
7801 /* Remember the new sched domains */
7802 if (doms_cur != &fallback_doms)
7804 kfree(dattr_cur); /* kfree(NULL) is safe */
7805 doms_cur = doms_new;
7806 dattr_cur = dattr_new;
7807 ndoms_cur = ndoms_new;
7809 register_sched_domain_sysctl();
7811 mutex_unlock(&sched_domains_mutex);
7814 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7815 int arch_reinit_sched_domains(void)
7820 mutex_lock(&sched_domains_mutex);
7821 detach_destroy_domains(&cpu_online_map);
7822 err = arch_init_sched_domains(&cpu_online_map);
7823 mutex_unlock(&sched_domains_mutex);
7829 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7833 if (buf[0] != '0' && buf[0] != '1')
7837 sched_smt_power_savings = (buf[0] == '1');
7839 sched_mc_power_savings = (buf[0] == '1');
7841 ret = arch_reinit_sched_domains();
7843 return ret ? ret : count;
7846 #ifdef CONFIG_SCHED_MC
7847 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7849 return sprintf(page, "%u\n", sched_mc_power_savings);
7851 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7852 const char *buf, size_t count)
7854 return sched_power_savings_store(buf, count, 0);
7856 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7857 sched_mc_power_savings_store);
7860 #ifdef CONFIG_SCHED_SMT
7861 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7863 return sprintf(page, "%u\n", sched_smt_power_savings);
7865 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7866 const char *buf, size_t count)
7868 return sched_power_savings_store(buf, count, 1);
7870 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7871 sched_smt_power_savings_store);
7874 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7878 #ifdef CONFIG_SCHED_SMT
7880 err = sysfs_create_file(&cls->kset.kobj,
7881 &attr_sched_smt_power_savings.attr);
7883 #ifdef CONFIG_SCHED_MC
7884 if (!err && mc_capable())
7885 err = sysfs_create_file(&cls->kset.kobj,
7886 &attr_sched_mc_power_savings.attr);
7893 * Force a reinitialization of the sched domains hierarchy. The domains
7894 * and groups cannot be updated in place without racing with the balancing
7895 * code, so we temporarily attach all running cpus to the NULL domain
7896 * which will prevent rebalancing while the sched domains are recalculated.
7898 static int update_sched_domains(struct notifier_block *nfb,
7899 unsigned long action, void *hcpu)
7902 case CPU_UP_PREPARE:
7903 case CPU_UP_PREPARE_FROZEN:
7904 case CPU_DOWN_PREPARE:
7905 case CPU_DOWN_PREPARE_FROZEN:
7906 detach_destroy_domains(&cpu_online_map);
7909 case CPU_UP_CANCELED:
7910 case CPU_UP_CANCELED_FROZEN:
7911 case CPU_DOWN_FAILED:
7912 case CPU_DOWN_FAILED_FROZEN:
7914 case CPU_ONLINE_FROZEN:
7916 case CPU_DEAD_FROZEN:
7918 * Fall through and re-initialise the domains.
7925 /* The hotplug lock is already held by cpu_up/cpu_down */
7926 arch_init_sched_domains(&cpu_online_map);
7931 void __init sched_init_smp(void)
7933 cpumask_t non_isolated_cpus;
7935 #if defined(CONFIG_NUMA)
7936 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7938 BUG_ON(sched_group_nodes_bycpu == NULL);
7941 mutex_lock(&sched_domains_mutex);
7942 arch_init_sched_domains(&cpu_online_map);
7943 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7944 if (cpus_empty(non_isolated_cpus))
7945 cpu_set(smp_processor_id(), non_isolated_cpus);
7946 mutex_unlock(&sched_domains_mutex);
7948 /* XXX: Theoretical race here - CPU may be hotplugged now */
7949 hotcpu_notifier(update_sched_domains, 0);
7952 /* Move init over to a non-isolated CPU */
7953 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7955 sched_init_granularity();
7958 void __init sched_init_smp(void)
7960 sched_init_granularity();
7962 #endif /* CONFIG_SMP */
7964 int in_sched_functions(unsigned long addr)
7966 return in_lock_functions(addr) ||
7967 (addr >= (unsigned long)__sched_text_start
7968 && addr < (unsigned long)__sched_text_end);
7971 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7973 cfs_rq->tasks_timeline = RB_ROOT;
7974 INIT_LIST_HEAD(&cfs_rq->tasks);
7975 #ifdef CONFIG_FAIR_GROUP_SCHED
7978 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7981 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7983 struct rt_prio_array *array;
7986 array = &rt_rq->active;
7987 for (i = 0; i < MAX_RT_PRIO; i++) {
7988 INIT_LIST_HEAD(array->queue + i);
7989 __clear_bit(i, array->bitmap);
7991 /* delimiter for bitsearch: */
7992 __set_bit(MAX_RT_PRIO, array->bitmap);
7994 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7995 rt_rq->highest_prio = MAX_RT_PRIO;
7998 rt_rq->rt_nr_migratory = 0;
7999 rt_rq->overloaded = 0;
8003 rt_rq->rt_throttled = 0;
8004 rt_rq->rt_runtime = 0;
8005 spin_lock_init(&rt_rq->rt_runtime_lock);
8007 #ifdef CONFIG_RT_GROUP_SCHED
8008 rt_rq->rt_nr_boosted = 0;
8013 #ifdef CONFIG_FAIR_GROUP_SCHED
8014 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8015 struct sched_entity *se, int cpu, int add,
8016 struct sched_entity *parent)
8018 struct rq *rq = cpu_rq(cpu);
8019 tg->cfs_rq[cpu] = cfs_rq;
8020 init_cfs_rq(cfs_rq, rq);
8023 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8026 /* se could be NULL for init_task_group */
8031 se->cfs_rq = &rq->cfs;
8033 se->cfs_rq = parent->my_q;
8036 se->load.weight = tg->shares;
8037 se->load.inv_weight = 0;
8038 se->parent = parent;
8042 #ifdef CONFIG_RT_GROUP_SCHED
8043 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8044 struct sched_rt_entity *rt_se, int cpu, int add,
8045 struct sched_rt_entity *parent)
8047 struct rq *rq = cpu_rq(cpu);
8049 tg->rt_rq[cpu] = rt_rq;
8050 init_rt_rq(rt_rq, rq);
8052 rt_rq->rt_se = rt_se;
8053 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8055 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8057 tg->rt_se[cpu] = rt_se;
8062 rt_se->rt_rq = &rq->rt;
8064 rt_se->rt_rq = parent->my_q;
8066 rt_se->rt_rq = &rq->rt;
8067 rt_se->my_q = rt_rq;
8068 rt_se->parent = parent;
8069 INIT_LIST_HEAD(&rt_se->run_list);
8073 void __init sched_init(void)
8076 unsigned long alloc_size = 0, ptr;
8078 #ifdef CONFIG_FAIR_GROUP_SCHED
8079 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8081 #ifdef CONFIG_RT_GROUP_SCHED
8082 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8084 #ifdef CONFIG_USER_SCHED
8088 * As sched_init() is called before page_alloc is setup,
8089 * we use alloc_bootmem().
8092 ptr = (unsigned long)alloc_bootmem(alloc_size);
8094 #ifdef CONFIG_FAIR_GROUP_SCHED
8095 init_task_group.se = (struct sched_entity **)ptr;
8096 ptr += nr_cpu_ids * sizeof(void **);
8098 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8099 ptr += nr_cpu_ids * sizeof(void **);
8101 #ifdef CONFIG_USER_SCHED
8102 root_task_group.se = (struct sched_entity **)ptr;
8103 ptr += nr_cpu_ids * sizeof(void **);
8105 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8106 ptr += nr_cpu_ids * sizeof(void **);
8109 #ifdef CONFIG_RT_GROUP_SCHED
8110 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8111 ptr += nr_cpu_ids * sizeof(void **);
8113 init_task_group.rt_rq = (struct rt_rq **)ptr;
8114 ptr += nr_cpu_ids * sizeof(void **);
8116 #ifdef CONFIG_USER_SCHED
8117 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8118 ptr += nr_cpu_ids * sizeof(void **);
8120 root_task_group.rt_rq = (struct rt_rq **)ptr;
8121 ptr += nr_cpu_ids * sizeof(void **);
8128 init_defrootdomain();
8131 init_rt_bandwidth(&def_rt_bandwidth,
8132 global_rt_period(), global_rt_runtime());
8134 #ifdef CONFIG_RT_GROUP_SCHED
8135 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8136 global_rt_period(), global_rt_runtime());
8137 #ifdef CONFIG_USER_SCHED
8138 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8139 global_rt_period(), RUNTIME_INF);
8143 #ifdef CONFIG_GROUP_SCHED
8144 list_add(&init_task_group.list, &task_groups);
8145 INIT_LIST_HEAD(&init_task_group.children);
8147 #ifdef CONFIG_USER_SCHED
8148 INIT_LIST_HEAD(&root_task_group.children);
8149 init_task_group.parent = &root_task_group;
8150 list_add(&init_task_group.siblings, &root_task_group.children);
8154 for_each_possible_cpu(i) {
8158 spin_lock_init(&rq->lock);
8159 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8161 init_cfs_rq(&rq->cfs, rq);
8162 init_rt_rq(&rq->rt, rq);
8163 #ifdef CONFIG_FAIR_GROUP_SCHED
8164 init_task_group.shares = init_task_group_load;
8165 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8166 #ifdef CONFIG_CGROUP_SCHED
8168 * How much cpu bandwidth does init_task_group get?
8170 * In case of task-groups formed thr' the cgroup filesystem, it
8171 * gets 100% of the cpu resources in the system. This overall
8172 * system cpu resource is divided among the tasks of
8173 * init_task_group and its child task-groups in a fair manner,
8174 * based on each entity's (task or task-group's) weight
8175 * (se->load.weight).
8177 * In other words, if init_task_group has 10 tasks of weight
8178 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8179 * then A0's share of the cpu resource is:
8181 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8183 * We achieve this by letting init_task_group's tasks sit
8184 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8186 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8187 #elif defined CONFIG_USER_SCHED
8188 root_task_group.shares = NICE_0_LOAD;
8189 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8191 * In case of task-groups formed thr' the user id of tasks,
8192 * init_task_group represents tasks belonging to root user.
8193 * Hence it forms a sibling of all subsequent groups formed.
8194 * In this case, init_task_group gets only a fraction of overall
8195 * system cpu resource, based on the weight assigned to root
8196 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8197 * by letting tasks of init_task_group sit in a separate cfs_rq
8198 * (init_cfs_rq) and having one entity represent this group of
8199 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8201 init_tg_cfs_entry(&init_task_group,
8202 &per_cpu(init_cfs_rq, i),
8203 &per_cpu(init_sched_entity, i), i, 1,
8204 root_task_group.se[i]);
8207 #endif /* CONFIG_FAIR_GROUP_SCHED */
8209 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8210 #ifdef CONFIG_RT_GROUP_SCHED
8211 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8212 #ifdef CONFIG_CGROUP_SCHED
8213 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8214 #elif defined CONFIG_USER_SCHED
8215 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8216 init_tg_rt_entry(&init_task_group,
8217 &per_cpu(init_rt_rq, i),
8218 &per_cpu(init_sched_rt_entity, i), i, 1,
8219 root_task_group.rt_se[i]);
8223 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8224 rq->cpu_load[j] = 0;
8228 rq->active_balance = 0;
8229 rq->next_balance = jiffies;
8232 rq->migration_thread = NULL;
8233 INIT_LIST_HEAD(&rq->migration_queue);
8234 rq_attach_root(rq, &def_root_domain);
8237 atomic_set(&rq->nr_iowait, 0);
8240 set_load_weight(&init_task);
8242 #ifdef CONFIG_PREEMPT_NOTIFIERS
8243 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8247 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8250 #ifdef CONFIG_RT_MUTEXES
8251 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8255 * The boot idle thread does lazy MMU switching as well:
8257 atomic_inc(&init_mm.mm_count);
8258 enter_lazy_tlb(&init_mm, current);
8261 * Make us the idle thread. Technically, schedule() should not be
8262 * called from this thread, however somewhere below it might be,
8263 * but because we are the idle thread, we just pick up running again
8264 * when this runqueue becomes "idle".
8266 init_idle(current, smp_processor_id());
8268 * During early bootup we pretend to be a normal task:
8270 current->sched_class = &fair_sched_class;
8272 scheduler_running = 1;
8275 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8276 void __might_sleep(char *file, int line)
8279 static unsigned long prev_jiffy; /* ratelimiting */
8281 if ((in_atomic() || irqs_disabled()) &&
8282 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8283 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8285 prev_jiffy = jiffies;
8286 printk(KERN_ERR "BUG: sleeping function called from invalid"
8287 " context at %s:%d\n", file, line);
8288 printk("in_atomic():%d, irqs_disabled():%d\n",
8289 in_atomic(), irqs_disabled());
8290 debug_show_held_locks(current);
8291 if (irqs_disabled())
8292 print_irqtrace_events(current);
8297 EXPORT_SYMBOL(__might_sleep);
8300 #ifdef CONFIG_MAGIC_SYSRQ
8301 static void normalize_task(struct rq *rq, struct task_struct *p)
8305 update_rq_clock(rq);
8306 on_rq = p->se.on_rq;
8308 deactivate_task(rq, p, 0);
8309 __setscheduler(rq, p, SCHED_NORMAL, 0);
8311 activate_task(rq, p, 0);
8312 resched_task(rq->curr);
8316 void normalize_rt_tasks(void)
8318 struct task_struct *g, *p;
8319 unsigned long flags;
8322 read_lock_irqsave(&tasklist_lock, flags);
8323 do_each_thread(g, p) {
8325 * Only normalize user tasks:
8330 p->se.exec_start = 0;
8331 #ifdef CONFIG_SCHEDSTATS
8332 p->se.wait_start = 0;
8333 p->se.sleep_start = 0;
8334 p->se.block_start = 0;
8339 * Renice negative nice level userspace
8342 if (TASK_NICE(p) < 0 && p->mm)
8343 set_user_nice(p, 0);
8347 spin_lock(&p->pi_lock);
8348 rq = __task_rq_lock(p);
8350 normalize_task(rq, p);
8352 __task_rq_unlock(rq);
8353 spin_unlock(&p->pi_lock);
8354 } while_each_thread(g, p);
8356 read_unlock_irqrestore(&tasklist_lock, flags);
8359 #endif /* CONFIG_MAGIC_SYSRQ */
8363 * These functions are only useful for the IA64 MCA handling.
8365 * They can only be called when the whole system has been
8366 * stopped - every CPU needs to be quiescent, and no scheduling
8367 * activity can take place. Using them for anything else would
8368 * be a serious bug, and as a result, they aren't even visible
8369 * under any other configuration.
8373 * curr_task - return the current task for a given cpu.
8374 * @cpu: the processor in question.
8376 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8378 struct task_struct *curr_task(int cpu)
8380 return cpu_curr(cpu);
8384 * set_curr_task - set the current task for a given cpu.
8385 * @cpu: the processor in question.
8386 * @p: the task pointer to set.
8388 * Description: This function must only be used when non-maskable interrupts
8389 * are serviced on a separate stack. It allows the architecture to switch the
8390 * notion of the current task on a cpu in a non-blocking manner. This function
8391 * must be called with all CPU's synchronized, and interrupts disabled, the
8392 * and caller must save the original value of the current task (see
8393 * curr_task() above) and restore that value before reenabling interrupts and
8394 * re-starting the system.
8396 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8398 void set_curr_task(int cpu, struct task_struct *p)
8405 #ifdef CONFIG_FAIR_GROUP_SCHED
8406 static void free_fair_sched_group(struct task_group *tg)
8410 for_each_possible_cpu(i) {
8412 kfree(tg->cfs_rq[i]);
8422 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8424 struct cfs_rq *cfs_rq;
8425 struct sched_entity *se, *parent_se;
8429 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8432 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8436 tg->shares = NICE_0_LOAD;
8438 for_each_possible_cpu(i) {
8441 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8442 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8446 se = kmalloc_node(sizeof(struct sched_entity),
8447 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8451 parent_se = parent ? parent->se[i] : NULL;
8452 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8461 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8463 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8464 &cpu_rq(cpu)->leaf_cfs_rq_list);
8467 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8469 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8472 static inline void free_fair_sched_group(struct task_group *tg)
8477 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8482 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8486 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8491 #ifdef CONFIG_RT_GROUP_SCHED
8492 static void free_rt_sched_group(struct task_group *tg)
8496 destroy_rt_bandwidth(&tg->rt_bandwidth);
8498 for_each_possible_cpu(i) {
8500 kfree(tg->rt_rq[i]);
8502 kfree(tg->rt_se[i]);
8510 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8512 struct rt_rq *rt_rq;
8513 struct sched_rt_entity *rt_se, *parent_se;
8517 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8520 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8524 init_rt_bandwidth(&tg->rt_bandwidth,
8525 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8527 for_each_possible_cpu(i) {
8530 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8531 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8535 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8536 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8540 parent_se = parent ? parent->rt_se[i] : NULL;
8541 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8550 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8552 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8553 &cpu_rq(cpu)->leaf_rt_rq_list);
8556 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8558 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8561 static inline void free_rt_sched_group(struct task_group *tg)
8566 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8571 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8575 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8580 #ifdef CONFIG_GROUP_SCHED
8581 static void free_sched_group(struct task_group *tg)
8583 free_fair_sched_group(tg);
8584 free_rt_sched_group(tg);
8588 /* allocate runqueue etc for a new task group */
8589 struct task_group *sched_create_group(struct task_group *parent)
8591 struct task_group *tg;
8592 unsigned long flags;
8595 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8597 return ERR_PTR(-ENOMEM);
8599 if (!alloc_fair_sched_group(tg, parent))
8602 if (!alloc_rt_sched_group(tg, parent))
8605 spin_lock_irqsave(&task_group_lock, flags);
8606 for_each_possible_cpu(i) {
8607 register_fair_sched_group(tg, i);
8608 register_rt_sched_group(tg, i);
8610 list_add_rcu(&tg->list, &task_groups);
8612 WARN_ON(!parent); /* root should already exist */
8614 tg->parent = parent;
8615 list_add_rcu(&tg->siblings, &parent->children);
8616 INIT_LIST_HEAD(&tg->children);
8617 spin_unlock_irqrestore(&task_group_lock, flags);
8622 free_sched_group(tg);
8623 return ERR_PTR(-ENOMEM);
8626 /* rcu callback to free various structures associated with a task group */
8627 static void free_sched_group_rcu(struct rcu_head *rhp)
8629 /* now it should be safe to free those cfs_rqs */
8630 free_sched_group(container_of(rhp, struct task_group, rcu));
8633 /* Destroy runqueue etc associated with a task group */
8634 void sched_destroy_group(struct task_group *tg)
8636 unsigned long flags;
8639 spin_lock_irqsave(&task_group_lock, flags);
8640 for_each_possible_cpu(i) {
8641 unregister_fair_sched_group(tg, i);
8642 unregister_rt_sched_group(tg, i);
8644 list_del_rcu(&tg->list);
8645 list_del_rcu(&tg->siblings);
8646 spin_unlock_irqrestore(&task_group_lock, flags);
8648 /* wait for possible concurrent references to cfs_rqs complete */
8649 call_rcu(&tg->rcu, free_sched_group_rcu);
8652 /* change task's runqueue when it moves between groups.
8653 * The caller of this function should have put the task in its new group
8654 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8655 * reflect its new group.
8657 void sched_move_task(struct task_struct *tsk)
8660 unsigned long flags;
8663 rq = task_rq_lock(tsk, &flags);
8665 update_rq_clock(rq);
8667 running = task_current(rq, tsk);
8668 on_rq = tsk->se.on_rq;
8671 dequeue_task(rq, tsk, 0);
8672 if (unlikely(running))
8673 tsk->sched_class->put_prev_task(rq, tsk);
8675 set_task_rq(tsk, task_cpu(tsk));
8677 #ifdef CONFIG_FAIR_GROUP_SCHED
8678 if (tsk->sched_class->moved_group)
8679 tsk->sched_class->moved_group(tsk);
8682 if (unlikely(running))
8683 tsk->sched_class->set_curr_task(rq);
8685 enqueue_task(rq, tsk, 0);
8687 task_rq_unlock(rq, &flags);
8691 #ifdef CONFIG_FAIR_GROUP_SCHED
8692 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8694 struct cfs_rq *cfs_rq = se->cfs_rq;
8699 dequeue_entity(cfs_rq, se, 0);
8701 se->load.weight = shares;
8702 se->load.inv_weight = 0;
8705 enqueue_entity(cfs_rq, se, 0);
8708 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8710 struct cfs_rq *cfs_rq = se->cfs_rq;
8711 struct rq *rq = cfs_rq->rq;
8712 unsigned long flags;
8714 spin_lock_irqsave(&rq->lock, flags);
8715 __set_se_shares(se, shares);
8716 spin_unlock_irqrestore(&rq->lock, flags);
8719 static DEFINE_MUTEX(shares_mutex);
8721 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8724 unsigned long flags;
8727 * We can't change the weight of the root cgroup.
8732 if (shares < MIN_SHARES)
8733 shares = MIN_SHARES;
8734 else if (shares > MAX_SHARES)
8735 shares = MAX_SHARES;
8737 mutex_lock(&shares_mutex);
8738 if (tg->shares == shares)
8741 spin_lock_irqsave(&task_group_lock, flags);
8742 for_each_possible_cpu(i)
8743 unregister_fair_sched_group(tg, i);
8744 list_del_rcu(&tg->siblings);
8745 spin_unlock_irqrestore(&task_group_lock, flags);
8747 /* wait for any ongoing reference to this group to finish */
8748 synchronize_sched();
8751 * Now we are free to modify the group's share on each cpu
8752 * w/o tripping rebalance_share or load_balance_fair.
8754 tg->shares = shares;
8755 for_each_possible_cpu(i) {
8759 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8760 set_se_shares(tg->se[i], shares);
8764 * Enable load balance activity on this group, by inserting it back on
8765 * each cpu's rq->leaf_cfs_rq_list.
8767 spin_lock_irqsave(&task_group_lock, flags);
8768 for_each_possible_cpu(i)
8769 register_fair_sched_group(tg, i);
8770 list_add_rcu(&tg->siblings, &tg->parent->children);
8771 spin_unlock_irqrestore(&task_group_lock, flags);
8773 mutex_unlock(&shares_mutex);
8777 unsigned long sched_group_shares(struct task_group *tg)
8783 #ifdef CONFIG_RT_GROUP_SCHED
8785 * Ensure that the real time constraints are schedulable.
8787 static DEFINE_MUTEX(rt_constraints_mutex);
8789 static unsigned long to_ratio(u64 period, u64 runtime)
8791 if (runtime == RUNTIME_INF)
8794 return div64_u64(runtime << 16, period);
8797 #ifdef CONFIG_CGROUP_SCHED
8798 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8800 struct task_group *tgi, *parent = tg->parent;
8801 unsigned long total = 0;
8804 if (global_rt_period() < period)
8807 return to_ratio(period, runtime) <
8808 to_ratio(global_rt_period(), global_rt_runtime());
8811 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8815 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8819 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8820 tgi->rt_bandwidth.rt_runtime);
8824 return total + to_ratio(period, runtime) <
8825 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8826 parent->rt_bandwidth.rt_runtime);
8828 #elif defined CONFIG_USER_SCHED
8829 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8831 struct task_group *tgi;
8832 unsigned long total = 0;
8833 unsigned long global_ratio =
8834 to_ratio(global_rt_period(), global_rt_runtime());
8837 list_for_each_entry_rcu(tgi, &task_groups, list) {
8841 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8842 tgi->rt_bandwidth.rt_runtime);
8846 return total + to_ratio(period, runtime) < global_ratio;
8850 /* Must be called with tasklist_lock held */
8851 static inline int tg_has_rt_tasks(struct task_group *tg)
8853 struct task_struct *g, *p;
8854 do_each_thread(g, p) {
8855 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8857 } while_each_thread(g, p);
8861 static int tg_set_bandwidth(struct task_group *tg,
8862 u64 rt_period, u64 rt_runtime)
8866 mutex_lock(&rt_constraints_mutex);
8867 read_lock(&tasklist_lock);
8868 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8872 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8877 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8878 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8879 tg->rt_bandwidth.rt_runtime = rt_runtime;
8881 for_each_possible_cpu(i) {
8882 struct rt_rq *rt_rq = tg->rt_rq[i];
8884 spin_lock(&rt_rq->rt_runtime_lock);
8885 rt_rq->rt_runtime = rt_runtime;
8886 spin_unlock(&rt_rq->rt_runtime_lock);
8888 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8890 read_unlock(&tasklist_lock);
8891 mutex_unlock(&rt_constraints_mutex);
8896 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8898 u64 rt_runtime, rt_period;
8900 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8901 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8902 if (rt_runtime_us < 0)
8903 rt_runtime = RUNTIME_INF;
8905 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8908 long sched_group_rt_runtime(struct task_group *tg)
8912 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8915 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8916 do_div(rt_runtime_us, NSEC_PER_USEC);
8917 return rt_runtime_us;
8920 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8922 u64 rt_runtime, rt_period;
8924 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8925 rt_runtime = tg->rt_bandwidth.rt_runtime;
8927 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8930 long sched_group_rt_period(struct task_group *tg)
8934 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8935 do_div(rt_period_us, NSEC_PER_USEC);
8936 return rt_period_us;
8939 static int sched_rt_global_constraints(void)
8943 mutex_lock(&rt_constraints_mutex);
8944 if (!__rt_schedulable(NULL, 1, 0))
8946 mutex_unlock(&rt_constraints_mutex);
8951 static int sched_rt_global_constraints(void)
8953 unsigned long flags;
8956 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8957 for_each_possible_cpu(i) {
8958 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8960 spin_lock(&rt_rq->rt_runtime_lock);
8961 rt_rq->rt_runtime = global_rt_runtime();
8962 spin_unlock(&rt_rq->rt_runtime_lock);
8964 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8970 int sched_rt_handler(struct ctl_table *table, int write,
8971 struct file *filp, void __user *buffer, size_t *lenp,
8975 int old_period, old_runtime;
8976 static DEFINE_MUTEX(mutex);
8979 old_period = sysctl_sched_rt_period;
8980 old_runtime = sysctl_sched_rt_runtime;
8982 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8984 if (!ret && write) {
8985 ret = sched_rt_global_constraints();
8987 sysctl_sched_rt_period = old_period;
8988 sysctl_sched_rt_runtime = old_runtime;
8990 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8991 def_rt_bandwidth.rt_period =
8992 ns_to_ktime(global_rt_period());
8995 mutex_unlock(&mutex);
9000 #ifdef CONFIG_CGROUP_SCHED
9002 /* return corresponding task_group object of a cgroup */
9003 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9005 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9006 struct task_group, css);
9009 static struct cgroup_subsys_state *
9010 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9012 struct task_group *tg, *parent;
9014 if (!cgrp->parent) {
9015 /* This is early initialization for the top cgroup */
9016 init_task_group.css.cgroup = cgrp;
9017 return &init_task_group.css;
9020 parent = cgroup_tg(cgrp->parent);
9021 tg = sched_create_group(parent);
9023 return ERR_PTR(-ENOMEM);
9025 /* Bind the cgroup to task_group object we just created */
9026 tg->css.cgroup = cgrp;
9032 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9034 struct task_group *tg = cgroup_tg(cgrp);
9036 sched_destroy_group(tg);
9040 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9041 struct task_struct *tsk)
9043 #ifdef CONFIG_RT_GROUP_SCHED
9044 /* Don't accept realtime tasks when there is no way for them to run */
9045 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9048 /* We don't support RT-tasks being in separate groups */
9049 if (tsk->sched_class != &fair_sched_class)
9057 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9058 struct cgroup *old_cont, struct task_struct *tsk)
9060 sched_move_task(tsk);
9063 #ifdef CONFIG_FAIR_GROUP_SCHED
9064 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9067 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9070 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9072 struct task_group *tg = cgroup_tg(cgrp);
9074 return (u64) tg->shares;
9078 #ifdef CONFIG_RT_GROUP_SCHED
9079 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9082 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9085 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9087 return sched_group_rt_runtime(cgroup_tg(cgrp));
9090 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9093 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9096 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9098 return sched_group_rt_period(cgroup_tg(cgrp));
9102 static struct cftype cpu_files[] = {
9103 #ifdef CONFIG_FAIR_GROUP_SCHED
9106 .read_u64 = cpu_shares_read_u64,
9107 .write_u64 = cpu_shares_write_u64,
9110 #ifdef CONFIG_RT_GROUP_SCHED
9112 .name = "rt_runtime_us",
9113 .read_s64 = cpu_rt_runtime_read,
9114 .write_s64 = cpu_rt_runtime_write,
9117 .name = "rt_period_us",
9118 .read_u64 = cpu_rt_period_read_uint,
9119 .write_u64 = cpu_rt_period_write_uint,
9124 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9126 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9129 struct cgroup_subsys cpu_cgroup_subsys = {
9131 .create = cpu_cgroup_create,
9132 .destroy = cpu_cgroup_destroy,
9133 .can_attach = cpu_cgroup_can_attach,
9134 .attach = cpu_cgroup_attach,
9135 .populate = cpu_cgroup_populate,
9136 .subsys_id = cpu_cgroup_subsys_id,
9140 #endif /* CONFIG_CGROUP_SCHED */
9142 #ifdef CONFIG_CGROUP_CPUACCT
9145 * CPU accounting code for task groups.
9147 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9148 * (balbir@in.ibm.com).
9151 /* track cpu usage of a group of tasks */
9153 struct cgroup_subsys_state css;
9154 /* cpuusage holds pointer to a u64-type object on every cpu */
9158 struct cgroup_subsys cpuacct_subsys;
9160 /* return cpu accounting group corresponding to this container */
9161 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9163 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9164 struct cpuacct, css);
9167 /* return cpu accounting group to which this task belongs */
9168 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9170 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9171 struct cpuacct, css);
9174 /* create a new cpu accounting group */
9175 static struct cgroup_subsys_state *cpuacct_create(
9176 struct cgroup_subsys *ss, struct cgroup *cgrp)
9178 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9181 return ERR_PTR(-ENOMEM);
9183 ca->cpuusage = alloc_percpu(u64);
9184 if (!ca->cpuusage) {
9186 return ERR_PTR(-ENOMEM);
9192 /* destroy an existing cpu accounting group */
9194 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9196 struct cpuacct *ca = cgroup_ca(cgrp);
9198 free_percpu(ca->cpuusage);
9202 /* return total cpu usage (in nanoseconds) of a group */
9203 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9205 struct cpuacct *ca = cgroup_ca(cgrp);
9206 u64 totalcpuusage = 0;
9209 for_each_possible_cpu(i) {
9210 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9213 * Take rq->lock to make 64-bit addition safe on 32-bit
9216 spin_lock_irq(&cpu_rq(i)->lock);
9217 totalcpuusage += *cpuusage;
9218 spin_unlock_irq(&cpu_rq(i)->lock);
9221 return totalcpuusage;
9224 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9227 struct cpuacct *ca = cgroup_ca(cgrp);
9236 for_each_possible_cpu(i) {
9237 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9239 spin_lock_irq(&cpu_rq(i)->lock);
9241 spin_unlock_irq(&cpu_rq(i)->lock);
9247 static struct cftype files[] = {
9250 .read_u64 = cpuusage_read,
9251 .write_u64 = cpuusage_write,
9255 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9257 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9261 * charge this task's execution time to its accounting group.
9263 * called with rq->lock held.
9265 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9269 if (!cpuacct_subsys.active)
9274 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9276 *cpuusage += cputime;
9280 struct cgroup_subsys cpuacct_subsys = {
9282 .create = cpuacct_create,
9283 .destroy = cpuacct_destroy,
9284 .populate = cpuacct_populate,
9285 .subsys_id = cpuacct_subsys_id,
9287 #endif /* CONFIG_CGROUP_CPUACCT */