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>
75 #include <asm/irq_regs.h>
77 #include "sched_cpupri.h"
80 * Convert user-nice values [ -20 ... 0 ... 19 ]
81 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
84 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
85 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
86 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
89 * 'User priority' is the nice value converted to something we
90 * can work with better when scaling various scheduler parameters,
91 * it's a [ 0 ... 39 ] range.
93 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
94 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
95 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
98 * Helpers for converting nanosecond timing to jiffy resolution
100 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
102 #define NICE_0_LOAD SCHED_LOAD_SCALE
103 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
106 * These are the 'tuning knobs' of the scheduler:
108 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
109 * Timeslices get refilled after they expire.
111 #define DEF_TIMESLICE (100 * HZ / 1000)
114 * single value that denotes runtime == period, ie unlimited time.
116 #define RUNTIME_INF ((u64)~0ULL)
120 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
121 * Since cpu_power is a 'constant', we can use a reciprocal divide.
123 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
125 return reciprocal_divide(load, sg->reciprocal_cpu_power);
129 * Each time a sched group cpu_power is changed,
130 * we must compute its reciprocal value
132 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
134 sg->__cpu_power += val;
135 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
139 static inline int rt_policy(int policy)
141 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
146 static inline int task_has_rt_policy(struct task_struct *p)
148 return rt_policy(p->policy);
152 * This is the priority-queue data structure of the RT scheduling class:
154 struct rt_prio_array {
155 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
156 struct list_head queue[MAX_RT_PRIO];
159 struct rt_bandwidth {
160 /* nests inside the rq lock: */
161 spinlock_t rt_runtime_lock;
164 struct hrtimer rt_period_timer;
167 static struct rt_bandwidth def_rt_bandwidth;
169 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
171 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
173 struct rt_bandwidth *rt_b =
174 container_of(timer, struct rt_bandwidth, rt_period_timer);
180 now = hrtimer_cb_get_time(timer);
181 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
186 idle = do_sched_rt_period_timer(rt_b, overrun);
189 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
193 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
195 rt_b->rt_period = ns_to_ktime(period);
196 rt_b->rt_runtime = runtime;
198 spin_lock_init(&rt_b->rt_runtime_lock);
200 hrtimer_init(&rt_b->rt_period_timer,
201 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
202 rt_b->rt_period_timer.function = sched_rt_period_timer;
203 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
206 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
210 if (rt_b->rt_runtime == RUNTIME_INF)
213 if (hrtimer_active(&rt_b->rt_period_timer))
216 spin_lock(&rt_b->rt_runtime_lock);
218 if (hrtimer_active(&rt_b->rt_period_timer))
221 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
222 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
223 hrtimer_start(&rt_b->rt_period_timer,
224 rt_b->rt_period_timer.expires,
227 spin_unlock(&rt_b->rt_runtime_lock);
230 #ifdef CONFIG_RT_GROUP_SCHED
231 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
233 hrtimer_cancel(&rt_b->rt_period_timer);
238 * sched_domains_mutex serializes calls to arch_init_sched_domains,
239 * detach_destroy_domains and partition_sched_domains.
241 static DEFINE_MUTEX(sched_domains_mutex);
243 #ifdef CONFIG_GROUP_SCHED
245 #include <linux/cgroup.h>
249 static LIST_HEAD(task_groups);
251 /* task group related information */
253 #ifdef CONFIG_CGROUP_SCHED
254 struct cgroup_subsys_state css;
257 #ifdef CONFIG_FAIR_GROUP_SCHED
258 /* schedulable entities of this group on each cpu */
259 struct sched_entity **se;
260 /* runqueue "owned" by this group on each cpu */
261 struct cfs_rq **cfs_rq;
262 unsigned long shares;
265 #ifdef CONFIG_RT_GROUP_SCHED
266 struct sched_rt_entity **rt_se;
267 struct rt_rq **rt_rq;
269 struct rt_bandwidth rt_bandwidth;
273 struct list_head list;
275 struct task_group *parent;
276 struct list_head siblings;
277 struct list_head children;
280 #ifdef CONFIG_USER_SCHED
284 * Every UID task group (including init_task_group aka UID-0) will
285 * be a child to this group.
287 struct task_group root_task_group;
289 #ifdef CONFIG_FAIR_GROUP_SCHED
290 /* Default task group's sched entity on each cpu */
291 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
292 /* Default task group's cfs_rq on each cpu */
293 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
294 #endif /* CONFIG_FAIR_GROUP_SCHED */
296 #ifdef CONFIG_RT_GROUP_SCHED
297 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
298 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
299 #endif /* CONFIG_RT_GROUP_SCHED */
300 #else /* !CONFIG_FAIR_GROUP_SCHED */
301 #define root_task_group init_task_group
302 #endif /* CONFIG_FAIR_GROUP_SCHED */
304 /* task_group_lock serializes add/remove of task groups and also changes to
305 * a task group's cpu shares.
307 static DEFINE_SPINLOCK(task_group_lock);
309 #ifdef CONFIG_FAIR_GROUP_SCHED
310 #ifdef CONFIG_USER_SCHED
311 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 #else /* !CONFIG_USER_SCHED */
313 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
314 #endif /* CONFIG_USER_SCHED */
317 * A weight of 0 or 1 can cause arithmetics problems.
318 * A weight of a cfs_rq is the sum of weights of which entities
319 * are queued on this cfs_rq, so a weight of a entity should not be
320 * too large, so as the shares value of a task group.
321 * (The default weight is 1024 - so there's no practical
322 * limitation from this.)
325 #define MAX_SHARES (1UL << 18)
327 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
330 /* Default task group.
331 * Every task in system belong to this group at bootup.
333 struct task_group init_task_group;
335 /* return group to which a task belongs */
336 static inline struct task_group *task_group(struct task_struct *p)
338 struct task_group *tg;
340 #ifdef CONFIG_USER_SCHED
342 #elif defined(CONFIG_CGROUP_SCHED)
343 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
344 struct task_group, css);
346 tg = &init_task_group;
351 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
352 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
354 #ifdef CONFIG_FAIR_GROUP_SCHED
355 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
356 p->se.parent = task_group(p)->se[cpu];
359 #ifdef CONFIG_RT_GROUP_SCHED
360 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
361 p->rt.parent = task_group(p)->rt_se[cpu];
367 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
369 #endif /* CONFIG_GROUP_SCHED */
371 /* CFS-related fields in a runqueue */
373 struct load_weight load;
374 unsigned long nr_running;
379 struct rb_root tasks_timeline;
380 struct rb_node *rb_leftmost;
382 struct list_head tasks;
383 struct list_head *balance_iterator;
386 * 'curr' points to currently running entity on this cfs_rq.
387 * It is set to NULL otherwise (i.e when none are currently running).
389 struct sched_entity *curr, *next;
391 unsigned long nr_spread_over;
393 #ifdef CONFIG_FAIR_GROUP_SCHED
394 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
397 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
398 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
399 * (like users, containers etc.)
401 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
402 * list is used during load balance.
404 struct list_head leaf_cfs_rq_list;
405 struct task_group *tg; /* group that "owns" this runqueue */
408 unsigned long task_weight;
409 unsigned long shares;
411 * We need space to build a sched_domain wide view of the full task
412 * group tree, in order to avoid depending on dynamic memory allocation
413 * during the load balancing we place this in the per cpu task group
414 * hierarchy. This limits the load balancing to one instance per cpu,
415 * but more should not be needed anyway.
417 struct aggregate_struct {
419 * load = weight(cpus) * f(tg)
421 * Where f(tg) is the recursive weight fraction assigned to
427 * part of the group weight distributed to this span.
429 unsigned long shares;
432 * The sum of all runqueue weights within this span.
434 unsigned long rq_weight;
437 * Weight contributed by tasks; this is the part we can
438 * influence by moving tasks around.
440 unsigned long task_weight;
446 /* Real-Time classes' related field in a runqueue: */
448 struct rt_prio_array active;
449 unsigned long rt_nr_running;
450 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
451 int highest_prio; /* highest queued rt task prio */
454 unsigned long rt_nr_migratory;
460 /* Nests inside the rq lock: */
461 spinlock_t rt_runtime_lock;
463 #ifdef CONFIG_RT_GROUP_SCHED
464 unsigned long rt_nr_boosted;
467 struct list_head leaf_rt_rq_list;
468 struct task_group *tg;
469 struct sched_rt_entity *rt_se;
476 * We add the notion of a root-domain which will be used to define per-domain
477 * variables. Each exclusive cpuset essentially defines an island domain by
478 * fully partitioning the member cpus from any other cpuset. Whenever a new
479 * exclusive cpuset is created, we also create and attach a new root-domain
489 * The "RT overload" flag: it gets set if a CPU has more than
490 * one runnable RT task.
495 struct cpupri cpupri;
500 * By default the system creates a single root-domain with all cpus as
501 * members (mimicking the global state we have today).
503 static struct root_domain def_root_domain;
508 * This is the main, per-CPU runqueue data structure.
510 * Locking rule: those places that want to lock multiple runqueues
511 * (such as the load balancing or the thread migration code), lock
512 * acquire operations must be ordered by ascending &runqueue.
519 * nr_running and cpu_load should be in the same cacheline because
520 * remote CPUs use both these fields when doing load calculation.
522 unsigned long nr_running;
523 #define CPU_LOAD_IDX_MAX 5
524 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
525 unsigned char idle_at_tick;
527 unsigned long last_tick_seen;
528 unsigned char in_nohz_recently;
530 /* capture load from *all* tasks on this cpu: */
531 struct load_weight load;
532 unsigned long nr_load_updates;
538 #ifdef CONFIG_FAIR_GROUP_SCHED
539 /* list of leaf cfs_rq on this cpu: */
540 struct list_head leaf_cfs_rq_list;
542 #ifdef CONFIG_RT_GROUP_SCHED
543 struct list_head leaf_rt_rq_list;
547 * This is part of a global counter where only the total sum
548 * over all CPUs matters. A task can increase this counter on
549 * one CPU and if it got migrated afterwards it may decrease
550 * it on another CPU. Always updated under the runqueue lock:
552 unsigned long nr_uninterruptible;
554 struct task_struct *curr, *idle;
555 unsigned long next_balance;
556 struct mm_struct *prev_mm;
563 struct root_domain *rd;
564 struct sched_domain *sd;
566 /* For active balancing */
569 /* cpu of this runqueue: */
573 struct task_struct *migration_thread;
574 struct list_head migration_queue;
577 #ifdef CONFIG_SCHED_HRTICK
578 unsigned long hrtick_flags;
579 ktime_t hrtick_expire;
580 struct hrtimer hrtick_timer;
583 #ifdef CONFIG_SCHEDSTATS
585 struct sched_info rq_sched_info;
587 /* sys_sched_yield() stats */
588 unsigned int yld_exp_empty;
589 unsigned int yld_act_empty;
590 unsigned int yld_both_empty;
591 unsigned int yld_count;
593 /* schedule() stats */
594 unsigned int sched_switch;
595 unsigned int sched_count;
596 unsigned int sched_goidle;
598 /* try_to_wake_up() stats */
599 unsigned int ttwu_count;
600 unsigned int ttwu_local;
603 unsigned int bkl_count;
605 struct lock_class_key rq_lock_key;
608 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
610 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
612 rq->curr->sched_class->check_preempt_curr(rq, p);
615 static inline int cpu_of(struct rq *rq)
625 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
626 * See detach_destroy_domains: synchronize_sched for details.
628 * The domain tree of any CPU may only be accessed from within
629 * preempt-disabled sections.
631 #define for_each_domain(cpu, __sd) \
632 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
634 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
635 #define this_rq() (&__get_cpu_var(runqueues))
636 #define task_rq(p) cpu_rq(task_cpu(p))
637 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
639 static inline void update_rq_clock(struct rq *rq)
641 rq->clock = sched_clock_cpu(cpu_of(rq));
645 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
647 #ifdef CONFIG_SCHED_DEBUG
648 # define const_debug __read_mostly
650 # define const_debug static const
654 * Debugging: various feature bits
657 #define SCHED_FEAT(name, enabled) \
658 __SCHED_FEAT_##name ,
661 #include "sched_features.h"
666 #define SCHED_FEAT(name, enabled) \
667 (1UL << __SCHED_FEAT_##name) * enabled |
669 const_debug unsigned int sysctl_sched_features =
670 #include "sched_features.h"
675 #ifdef CONFIG_SCHED_DEBUG
676 #define SCHED_FEAT(name, enabled) \
679 static __read_mostly char *sched_feat_names[] = {
680 #include "sched_features.h"
686 static int sched_feat_open(struct inode *inode, struct file *filp)
688 filp->private_data = inode->i_private;
693 sched_feat_read(struct file *filp, char __user *ubuf,
694 size_t cnt, loff_t *ppos)
701 for (i = 0; sched_feat_names[i]; i++) {
702 len += strlen(sched_feat_names[i]);
706 buf = kmalloc(len + 2, GFP_KERNEL);
710 for (i = 0; sched_feat_names[i]; i++) {
711 if (sysctl_sched_features & (1UL << i))
712 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
714 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
717 r += sprintf(buf + r, "\n");
718 WARN_ON(r >= len + 2);
720 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
728 sched_feat_write(struct file *filp, const char __user *ubuf,
729 size_t cnt, loff_t *ppos)
739 if (copy_from_user(&buf, ubuf, cnt))
744 if (strncmp(buf, "NO_", 3) == 0) {
749 for (i = 0; sched_feat_names[i]; i++) {
750 int len = strlen(sched_feat_names[i]);
752 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
754 sysctl_sched_features &= ~(1UL << i);
756 sysctl_sched_features |= (1UL << i);
761 if (!sched_feat_names[i])
769 static struct file_operations sched_feat_fops = {
770 .open = sched_feat_open,
771 .read = sched_feat_read,
772 .write = sched_feat_write,
775 static __init int sched_init_debug(void)
777 debugfs_create_file("sched_features", 0644, NULL, NULL,
782 late_initcall(sched_init_debug);
786 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
789 * Number of tasks to iterate in a single balance run.
790 * Limited because this is done with IRQs disabled.
792 const_debug unsigned int sysctl_sched_nr_migrate = 32;
795 * period over which we measure -rt task cpu usage in us.
798 unsigned int sysctl_sched_rt_period = 1000000;
800 static __read_mostly int scheduler_running;
803 * part of the period that we allow rt tasks to run in us.
806 int sysctl_sched_rt_runtime = 950000;
808 static inline u64 global_rt_period(void)
810 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
813 static inline u64 global_rt_runtime(void)
815 if (sysctl_sched_rt_period < 0)
818 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
821 unsigned long long time_sync_thresh = 100000;
823 static DEFINE_PER_CPU(unsigned long long, time_offset);
824 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
827 * Global lock which we take every now and then to synchronize
828 * the CPUs time. This method is not warp-safe, but it's good
829 * enough to synchronize slowly diverging time sources and thus
830 * it's good enough for tracing:
832 static DEFINE_SPINLOCK(time_sync_lock);
833 static unsigned long long prev_global_time;
835 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
838 * We want this inlined, to not get tracer function calls
839 * in this critical section:
841 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
842 __raw_spin_lock(&time_sync_lock.raw_lock);
844 if (time < prev_global_time) {
845 per_cpu(time_offset, cpu) += prev_global_time - time;
846 time = prev_global_time;
848 prev_global_time = time;
851 __raw_spin_unlock(&time_sync_lock.raw_lock);
852 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
857 static unsigned long long __cpu_clock(int cpu)
859 unsigned long long now;
862 * Only call sched_clock() if the scheduler has already been
863 * initialized (some code might call cpu_clock() very early):
865 if (unlikely(!scheduler_running))
868 now = sched_clock_cpu(cpu);
874 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
875 * clock constructed from sched_clock():
877 unsigned long long cpu_clock(int cpu)
879 unsigned long long prev_cpu_time, time, delta_time;
882 local_irq_save(flags);
883 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
884 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
885 delta_time = time-prev_cpu_time;
887 if (unlikely(delta_time > time_sync_thresh)) {
888 time = __sync_cpu_clock(time, cpu);
889 per_cpu(prev_cpu_time, cpu) = time;
891 local_irq_restore(flags);
895 EXPORT_SYMBOL_GPL(cpu_clock);
897 #ifndef prepare_arch_switch
898 # define prepare_arch_switch(next) do { } while (0)
900 #ifndef finish_arch_switch
901 # define finish_arch_switch(prev) do { } while (0)
904 static inline int task_current(struct rq *rq, struct task_struct *p)
906 return rq->curr == p;
909 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
910 static inline int task_running(struct rq *rq, struct task_struct *p)
912 return task_current(rq, p);
915 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
919 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
921 #ifdef CONFIG_DEBUG_SPINLOCK
922 /* this is a valid case when another task releases the spinlock */
923 rq->lock.owner = current;
926 * If we are tracking spinlock dependencies then we have to
927 * fix up the runqueue lock - which gets 'carried over' from
930 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
932 spin_unlock_irq(&rq->lock);
935 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
936 static inline int task_running(struct rq *rq, struct task_struct *p)
941 return task_current(rq, p);
945 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
949 * We can optimise this out completely for !SMP, because the
950 * SMP rebalancing from interrupt is the only thing that cares
955 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
956 spin_unlock_irq(&rq->lock);
958 spin_unlock(&rq->lock);
962 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
966 * After ->oncpu is cleared, the task can be moved to a different CPU.
967 * We must ensure this doesn't happen until the switch is completely
973 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
977 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
980 * __task_rq_lock - lock the runqueue a given task resides on.
981 * Must be called interrupts disabled.
983 static inline struct rq *__task_rq_lock(struct task_struct *p)
987 struct rq *rq = task_rq(p);
988 spin_lock(&rq->lock);
989 if (likely(rq == task_rq(p)))
991 spin_unlock(&rq->lock);
996 * task_rq_lock - lock the runqueue a given task resides on and disable
997 * interrupts. Note the ordering: we can safely lookup the task_rq without
998 * explicitly disabling preemption.
1000 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
1001 __acquires(rq->lock)
1006 local_irq_save(*flags);
1008 spin_lock(&rq->lock);
1009 if (likely(rq == task_rq(p)))
1011 spin_unlock_irqrestore(&rq->lock, *flags);
1015 static void __task_rq_unlock(struct rq *rq)
1016 __releases(rq->lock)
1018 spin_unlock(&rq->lock);
1021 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1022 __releases(rq->lock)
1024 spin_unlock_irqrestore(&rq->lock, *flags);
1028 * this_rq_lock - lock this runqueue and disable interrupts.
1030 static struct rq *this_rq_lock(void)
1031 __acquires(rq->lock)
1035 local_irq_disable();
1037 spin_lock(&rq->lock);
1042 static void __resched_task(struct task_struct *p, int tif_bit);
1044 static inline void resched_task(struct task_struct *p)
1046 __resched_task(p, TIF_NEED_RESCHED);
1049 #ifdef CONFIG_SCHED_HRTICK
1051 * Use HR-timers to deliver accurate preemption points.
1053 * Its all a bit involved since we cannot program an hrt while holding the
1054 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1057 * When we get rescheduled we reprogram the hrtick_timer outside of the
1060 static inline void resched_hrt(struct task_struct *p)
1062 __resched_task(p, TIF_HRTICK_RESCHED);
1065 static inline void resched_rq(struct rq *rq)
1067 unsigned long flags;
1069 spin_lock_irqsave(&rq->lock, flags);
1070 resched_task(rq->curr);
1071 spin_unlock_irqrestore(&rq->lock, flags);
1075 HRTICK_SET, /* re-programm hrtick_timer */
1076 HRTICK_RESET, /* not a new slice */
1077 HRTICK_BLOCK, /* stop hrtick operations */
1082 * - enabled by features
1083 * - hrtimer is actually high res
1085 static inline int hrtick_enabled(struct rq *rq)
1087 if (!sched_feat(HRTICK))
1089 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1091 return hrtimer_is_hres_active(&rq->hrtick_timer);
1095 * Called to set the hrtick timer state.
1097 * called with rq->lock held and irqs disabled
1099 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1101 assert_spin_locked(&rq->lock);
1104 * preempt at: now + delay
1107 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1109 * indicate we need to program the timer
1111 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1113 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1116 * New slices are called from the schedule path and don't need a
1117 * forced reschedule.
1120 resched_hrt(rq->curr);
1123 static void hrtick_clear(struct rq *rq)
1125 if (hrtimer_active(&rq->hrtick_timer))
1126 hrtimer_cancel(&rq->hrtick_timer);
1130 * Update the timer from the possible pending state.
1132 static void hrtick_set(struct rq *rq)
1136 unsigned long flags;
1138 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1140 spin_lock_irqsave(&rq->lock, flags);
1141 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1142 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1143 time = rq->hrtick_expire;
1144 clear_thread_flag(TIF_HRTICK_RESCHED);
1145 spin_unlock_irqrestore(&rq->lock, flags);
1148 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1149 if (reset && !hrtimer_active(&rq->hrtick_timer))
1156 * High-resolution timer tick.
1157 * Runs from hardirq context with interrupts disabled.
1159 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1161 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1163 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1165 spin_lock(&rq->lock);
1166 update_rq_clock(rq);
1167 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1168 spin_unlock(&rq->lock);
1170 return HRTIMER_NORESTART;
1174 static void hotplug_hrtick_disable(int cpu)
1176 struct rq *rq = cpu_rq(cpu);
1177 unsigned long flags;
1179 spin_lock_irqsave(&rq->lock, flags);
1180 rq->hrtick_flags = 0;
1181 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1182 spin_unlock_irqrestore(&rq->lock, flags);
1187 static void hotplug_hrtick_enable(int cpu)
1189 struct rq *rq = cpu_rq(cpu);
1190 unsigned long flags;
1192 spin_lock_irqsave(&rq->lock, flags);
1193 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1194 spin_unlock_irqrestore(&rq->lock, flags);
1198 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1200 int cpu = (int)(long)hcpu;
1203 case CPU_UP_CANCELED:
1204 case CPU_UP_CANCELED_FROZEN:
1205 case CPU_DOWN_PREPARE:
1206 case CPU_DOWN_PREPARE_FROZEN:
1208 case CPU_DEAD_FROZEN:
1209 hotplug_hrtick_disable(cpu);
1212 case CPU_UP_PREPARE:
1213 case CPU_UP_PREPARE_FROZEN:
1214 case CPU_DOWN_FAILED:
1215 case CPU_DOWN_FAILED_FROZEN:
1217 case CPU_ONLINE_FROZEN:
1218 hotplug_hrtick_enable(cpu);
1225 static void init_hrtick(void)
1227 hotcpu_notifier(hotplug_hrtick, 0);
1229 #endif /* CONFIG_SMP */
1231 static void init_rq_hrtick(struct rq *rq)
1233 rq->hrtick_flags = 0;
1234 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1235 rq->hrtick_timer.function = hrtick;
1236 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1239 void hrtick_resched(void)
1242 unsigned long flags;
1244 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1247 local_irq_save(flags);
1248 rq = cpu_rq(smp_processor_id());
1250 local_irq_restore(flags);
1253 static inline void hrtick_clear(struct rq *rq)
1257 static inline void hrtick_set(struct rq *rq)
1261 static inline void init_rq_hrtick(struct rq *rq)
1265 void hrtick_resched(void)
1269 static inline void init_hrtick(void)
1275 * resched_task - mark a task 'to be rescheduled now'.
1277 * On UP this means the setting of the need_resched flag, on SMP it
1278 * might also involve a cross-CPU call to trigger the scheduler on
1283 #ifndef tsk_is_polling
1284 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1287 static void __resched_task(struct task_struct *p, int tif_bit)
1291 assert_spin_locked(&task_rq(p)->lock);
1293 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1296 set_tsk_thread_flag(p, tif_bit);
1299 if (cpu == smp_processor_id())
1302 /* NEED_RESCHED must be visible before we test polling */
1304 if (!tsk_is_polling(p))
1305 smp_send_reschedule(cpu);
1308 static void resched_cpu(int cpu)
1310 struct rq *rq = cpu_rq(cpu);
1311 unsigned long flags;
1313 if (!spin_trylock_irqsave(&rq->lock, flags))
1315 resched_task(cpu_curr(cpu));
1316 spin_unlock_irqrestore(&rq->lock, flags);
1321 * When add_timer_on() enqueues a timer into the timer wheel of an
1322 * idle CPU then this timer might expire before the next timer event
1323 * which is scheduled to wake up that CPU. In case of a completely
1324 * idle system the next event might even be infinite time into the
1325 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1326 * leaves the inner idle loop so the newly added timer is taken into
1327 * account when the CPU goes back to idle and evaluates the timer
1328 * wheel for the next timer event.
1330 void wake_up_idle_cpu(int cpu)
1332 struct rq *rq = cpu_rq(cpu);
1334 if (cpu == smp_processor_id())
1338 * This is safe, as this function is called with the timer
1339 * wheel base lock of (cpu) held. When the CPU is on the way
1340 * to idle and has not yet set rq->curr to idle then it will
1341 * be serialized on the timer wheel base lock and take the new
1342 * timer into account automatically.
1344 if (rq->curr != rq->idle)
1348 * We can set TIF_RESCHED on the idle task of the other CPU
1349 * lockless. The worst case is that the other CPU runs the
1350 * idle task through an additional NOOP schedule()
1352 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1354 /* NEED_RESCHED must be visible before we test polling */
1356 if (!tsk_is_polling(rq->idle))
1357 smp_send_reschedule(cpu);
1359 #endif /* CONFIG_NO_HZ */
1361 #else /* !CONFIG_SMP */
1362 static void __resched_task(struct task_struct *p, int tif_bit)
1364 assert_spin_locked(&task_rq(p)->lock);
1365 set_tsk_thread_flag(p, tif_bit);
1367 #endif /* CONFIG_SMP */
1369 #if BITS_PER_LONG == 32
1370 # define WMULT_CONST (~0UL)
1372 # define WMULT_CONST (1UL << 32)
1375 #define WMULT_SHIFT 32
1378 * Shift right and round:
1380 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1383 * delta *= weight / lw
1385 static unsigned long
1386 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1387 struct load_weight *lw)
1391 if (!lw->inv_weight) {
1392 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1395 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1399 tmp = (u64)delta_exec * weight;
1401 * Check whether we'd overflow the 64-bit multiplication:
1403 if (unlikely(tmp > WMULT_CONST))
1404 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1407 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1409 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1412 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1418 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1425 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1426 * of tasks with abnormal "nice" values across CPUs the contribution that
1427 * each task makes to its run queue's load is weighted according to its
1428 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1429 * scaled version of the new time slice allocation that they receive on time
1433 #define WEIGHT_IDLEPRIO 2
1434 #define WMULT_IDLEPRIO (1 << 31)
1437 * Nice levels are multiplicative, with a gentle 10% change for every
1438 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1439 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1440 * that remained on nice 0.
1442 * The "10% effect" is relative and cumulative: from _any_ nice level,
1443 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1444 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1445 * If a task goes up by ~10% and another task goes down by ~10% then
1446 * the relative distance between them is ~25%.)
1448 static const int prio_to_weight[40] = {
1449 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1450 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1451 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1452 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1453 /* 0 */ 1024, 820, 655, 526, 423,
1454 /* 5 */ 335, 272, 215, 172, 137,
1455 /* 10 */ 110, 87, 70, 56, 45,
1456 /* 15 */ 36, 29, 23, 18, 15,
1460 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1462 * In cases where the weight does not change often, we can use the
1463 * precalculated inverse to speed up arithmetics by turning divisions
1464 * into multiplications:
1466 static const u32 prio_to_wmult[40] = {
1467 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1468 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1469 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1470 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1471 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1472 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1473 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1474 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1477 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1480 * runqueue iterator, to support SMP load-balancing between different
1481 * scheduling classes, without having to expose their internal data
1482 * structures to the load-balancing proper:
1484 struct rq_iterator {
1486 struct task_struct *(*start)(void *);
1487 struct task_struct *(*next)(void *);
1491 static unsigned long
1492 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1493 unsigned long max_load_move, struct sched_domain *sd,
1494 enum cpu_idle_type idle, int *all_pinned,
1495 int *this_best_prio, struct rq_iterator *iterator);
1498 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1499 struct sched_domain *sd, enum cpu_idle_type idle,
1500 struct rq_iterator *iterator);
1503 #ifdef CONFIG_CGROUP_CPUACCT
1504 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1506 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1509 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1511 update_load_add(&rq->load, load);
1514 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1516 update_load_sub(&rq->load, load);
1520 static unsigned long source_load(int cpu, int type);
1521 static unsigned long target_load(int cpu, int type);
1522 static unsigned long cpu_avg_load_per_task(int cpu);
1523 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1525 #ifdef CONFIG_FAIR_GROUP_SCHED
1528 * Group load balancing.
1530 * We calculate a few balance domain wide aggregate numbers; load and weight.
1531 * Given the pictures below, and assuming each item has equal weight:
1542 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1543 * which equals 1/9-th of the total load.
1546 * The weight of this group on the selected cpus.
1549 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1553 * Part of the rq_weight contributed by tasks; all groups except B would
1557 static inline struct aggregate_struct *
1558 aggregate(struct task_group *tg, struct sched_domain *sd)
1560 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1563 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1566 * Iterate the full tree, calling @down when first entering a node and @up when
1567 * leaving it for the final time.
1570 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1571 struct sched_domain *sd)
1573 struct task_group *parent, *child;
1576 parent = &root_task_group;
1578 (*down)(parent, sd);
1579 list_for_each_entry_rcu(child, &parent->children, siblings) {
1589 parent = parent->parent;
1596 * Calculate the aggregate runqueue weight.
1599 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1601 unsigned long rq_weight = 0;
1602 unsigned long task_weight = 0;
1605 for_each_cpu_mask(i, sd->span) {
1606 rq_weight += tg->cfs_rq[i]->load.weight;
1607 task_weight += tg->cfs_rq[i]->task_weight;
1610 aggregate(tg, sd)->rq_weight = rq_weight;
1611 aggregate(tg, sd)->task_weight = task_weight;
1615 * Compute the weight of this group on the given cpus.
1618 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1620 unsigned long shares = 0;
1623 for_each_cpu_mask(i, sd->span)
1624 shares += tg->cfs_rq[i]->shares;
1626 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1627 shares = tg->shares;
1629 aggregate(tg, sd)->shares = shares;
1633 * Compute the load fraction assigned to this group, relies on the aggregate
1634 * weight and this group's parent's load, i.e. top-down.
1637 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1645 for_each_cpu_mask(i, sd->span)
1646 load += cpu_rq(i)->load.weight;
1649 load = aggregate(tg->parent, sd)->load;
1652 * shares is our weight in the parent's rq so
1653 * shares/parent->rq_weight gives our fraction of the load
1655 load *= aggregate(tg, sd)->shares;
1656 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1659 aggregate(tg, sd)->load = load;
1662 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1665 * Calculate and set the cpu's group shares.
1668 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1672 unsigned long shares;
1673 unsigned long rq_weight;
1678 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1681 * If there are currently no tasks on the cpu pretend there is one of
1682 * average load so that when a new task gets to run here it will not
1683 * get delayed by group starvation.
1687 rq_weight = NICE_0_LOAD;
1691 * \Sum shares * rq_weight
1692 * shares = -----------------------
1696 shares = aggregate(tg, sd)->shares * rq_weight;
1697 shares /= aggregate(tg, sd)->rq_weight + 1;
1700 * record the actual number of shares, not the boosted amount.
1702 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1704 if (shares < MIN_SHARES)
1705 shares = MIN_SHARES;
1706 else if (shares > MAX_SHARES)
1707 shares = MAX_SHARES;
1709 __set_se_shares(tg->se[tcpu], shares);
1713 * Re-adjust the weights on the cpu the task came from and on the cpu the
1717 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1720 unsigned long shares;
1722 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1724 __update_group_shares_cpu(tg, sd, scpu);
1725 __update_group_shares_cpu(tg, sd, dcpu);
1728 * ensure we never loose shares due to rounding errors in the
1729 * above redistribution.
1731 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1733 tg->cfs_rq[dcpu]->shares += shares;
1737 * Because changing a group's shares changes the weight of the super-group
1738 * we need to walk up the tree and change all shares until we hit the root.
1741 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1745 __move_group_shares(tg, sd, scpu, dcpu);
1751 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1753 unsigned long shares = aggregate(tg, sd)->shares;
1756 for_each_cpu_mask(i, sd->span) {
1757 struct rq *rq = cpu_rq(i);
1758 unsigned long flags;
1760 spin_lock_irqsave(&rq->lock, flags);
1761 __update_group_shares_cpu(tg, sd, i);
1762 spin_unlock_irqrestore(&rq->lock, flags);
1765 aggregate_group_shares(tg, sd);
1768 * ensure we never loose shares due to rounding errors in the
1769 * above redistribution.
1771 shares -= aggregate(tg, sd)->shares;
1773 tg->cfs_rq[sd->first_cpu]->shares += shares;
1774 aggregate(tg, sd)->shares += shares;
1779 * Calculate the accumulative weight and recursive load of each task group
1780 * while walking down the tree.
1783 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1785 aggregate_group_weight(tg, sd);
1786 aggregate_group_shares(tg, sd);
1787 aggregate_group_load(tg, sd);
1791 * Rebalance the cpu shares while walking back up the tree.
1794 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1796 aggregate_group_set_shares(tg, sd);
1799 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1801 static void __init init_aggregate(void)
1805 for_each_possible_cpu(i)
1806 spin_lock_init(&per_cpu(aggregate_lock, i));
1809 static int get_aggregate(struct sched_domain *sd)
1811 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1814 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1818 static void put_aggregate(struct sched_domain *sd)
1820 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1823 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1825 cfs_rq->shares = shares;
1830 static inline void init_aggregate(void)
1834 static inline int get_aggregate(struct sched_domain *sd)
1839 static inline void put_aggregate(struct sched_domain *sd)
1846 #include "sched_stats.h"
1847 #include "sched_idletask.c"
1848 #include "sched_fair.c"
1849 #include "sched_rt.c"
1850 #ifdef CONFIG_SCHED_DEBUG
1851 # include "sched_debug.c"
1854 #define sched_class_highest (&rt_sched_class)
1855 #define for_each_class(class) \
1856 for (class = sched_class_highest; class; class = class->next)
1858 static void inc_nr_running(struct rq *rq)
1863 static void dec_nr_running(struct rq *rq)
1868 static void set_load_weight(struct task_struct *p)
1870 if (task_has_rt_policy(p)) {
1871 p->se.load.weight = prio_to_weight[0] * 2;
1872 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1877 * SCHED_IDLE tasks get minimal weight:
1879 if (p->policy == SCHED_IDLE) {
1880 p->se.load.weight = WEIGHT_IDLEPRIO;
1881 p->se.load.inv_weight = WMULT_IDLEPRIO;
1885 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1886 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1889 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1891 sched_info_queued(p);
1892 p->sched_class->enqueue_task(rq, p, wakeup);
1896 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1898 p->sched_class->dequeue_task(rq, p, sleep);
1903 * __normal_prio - return the priority that is based on the static prio
1905 static inline int __normal_prio(struct task_struct *p)
1907 return p->static_prio;
1911 * Calculate the expected normal priority: i.e. priority
1912 * without taking RT-inheritance into account. Might be
1913 * boosted by interactivity modifiers. Changes upon fork,
1914 * setprio syscalls, and whenever the interactivity
1915 * estimator recalculates.
1917 static inline int normal_prio(struct task_struct *p)
1921 if (task_has_rt_policy(p))
1922 prio = MAX_RT_PRIO-1 - p->rt_priority;
1924 prio = __normal_prio(p);
1929 * Calculate the current priority, i.e. the priority
1930 * taken into account by the scheduler. This value might
1931 * be boosted by RT tasks, or might be boosted by
1932 * interactivity modifiers. Will be RT if the task got
1933 * RT-boosted. If not then it returns p->normal_prio.
1935 static int effective_prio(struct task_struct *p)
1937 p->normal_prio = normal_prio(p);
1939 * If we are RT tasks or we were boosted to RT priority,
1940 * keep the priority unchanged. Otherwise, update priority
1941 * to the normal priority:
1943 if (!rt_prio(p->prio))
1944 return p->normal_prio;
1949 * activate_task - move a task to the runqueue.
1951 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1953 if (task_contributes_to_load(p))
1954 rq->nr_uninterruptible--;
1956 enqueue_task(rq, p, wakeup);
1961 * deactivate_task - remove a task from the runqueue.
1963 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1965 if (task_contributes_to_load(p))
1966 rq->nr_uninterruptible++;
1968 dequeue_task(rq, p, sleep);
1973 * task_curr - is this task currently executing on a CPU?
1974 * @p: the task in question.
1976 inline int task_curr(const struct task_struct *p)
1978 return cpu_curr(task_cpu(p)) == p;
1981 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1983 set_task_rq(p, cpu);
1986 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1987 * successfuly executed on another CPU. We must ensure that updates of
1988 * per-task data have been completed by this moment.
1991 task_thread_info(p)->cpu = cpu;
1995 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1996 const struct sched_class *prev_class,
1997 int oldprio, int running)
1999 if (prev_class != p->sched_class) {
2000 if (prev_class->switched_from)
2001 prev_class->switched_from(rq, p, running);
2002 p->sched_class->switched_to(rq, p, running);
2004 p->sched_class->prio_changed(rq, p, oldprio, running);
2009 /* Used instead of source_load when we know the type == 0 */
2010 static unsigned long weighted_cpuload(const int cpu)
2012 return cpu_rq(cpu)->load.weight;
2016 * Is this task likely cache-hot:
2019 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2024 * Buddy candidates are cache hot:
2026 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2029 if (p->sched_class != &fair_sched_class)
2032 if (sysctl_sched_migration_cost == -1)
2034 if (sysctl_sched_migration_cost == 0)
2037 delta = now - p->se.exec_start;
2039 return delta < (s64)sysctl_sched_migration_cost;
2043 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2045 int old_cpu = task_cpu(p);
2046 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2047 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2048 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2051 clock_offset = old_rq->clock - new_rq->clock;
2053 #ifdef CONFIG_SCHEDSTATS
2054 if (p->se.wait_start)
2055 p->se.wait_start -= clock_offset;
2056 if (p->se.sleep_start)
2057 p->se.sleep_start -= clock_offset;
2058 if (p->se.block_start)
2059 p->se.block_start -= clock_offset;
2060 if (old_cpu != new_cpu) {
2061 schedstat_inc(p, se.nr_migrations);
2062 if (task_hot(p, old_rq->clock, NULL))
2063 schedstat_inc(p, se.nr_forced2_migrations);
2066 p->se.vruntime -= old_cfsrq->min_vruntime -
2067 new_cfsrq->min_vruntime;
2069 __set_task_cpu(p, new_cpu);
2072 struct migration_req {
2073 struct list_head list;
2075 struct task_struct *task;
2078 struct completion done;
2082 * The task's runqueue lock must be held.
2083 * Returns true if you have to wait for migration thread.
2086 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2088 struct rq *rq = task_rq(p);
2091 * If the task is not on a runqueue (and not running), then
2092 * it is sufficient to simply update the task's cpu field.
2094 if (!p->se.on_rq && !task_running(rq, p)) {
2095 set_task_cpu(p, dest_cpu);
2099 init_completion(&req->done);
2101 req->dest_cpu = dest_cpu;
2102 list_add(&req->list, &rq->migration_queue);
2108 * wait_task_inactive - wait for a thread to unschedule.
2110 * The caller must ensure that the task *will* unschedule sometime soon,
2111 * else this function might spin for a *long* time. This function can't
2112 * be called with interrupts off, or it may introduce deadlock with
2113 * smp_call_function() if an IPI is sent by the same process we are
2114 * waiting to become inactive.
2116 void wait_task_inactive(struct task_struct *p)
2118 unsigned long flags;
2124 * We do the initial early heuristics without holding
2125 * any task-queue locks at all. We'll only try to get
2126 * the runqueue lock when things look like they will
2132 * If the task is actively running on another CPU
2133 * still, just relax and busy-wait without holding
2136 * NOTE! Since we don't hold any locks, it's not
2137 * even sure that "rq" stays as the right runqueue!
2138 * But we don't care, since "task_running()" will
2139 * return false if the runqueue has changed and p
2140 * is actually now running somewhere else!
2142 while (task_running(rq, p))
2146 * Ok, time to look more closely! We need the rq
2147 * lock now, to be *sure*. If we're wrong, we'll
2148 * just go back and repeat.
2150 rq = task_rq_lock(p, &flags);
2151 running = task_running(rq, p);
2152 on_rq = p->se.on_rq;
2153 task_rq_unlock(rq, &flags);
2156 * Was it really running after all now that we
2157 * checked with the proper locks actually held?
2159 * Oops. Go back and try again..
2161 if (unlikely(running)) {
2167 * It's not enough that it's not actively running,
2168 * it must be off the runqueue _entirely_, and not
2171 * So if it wa still runnable (but just not actively
2172 * running right now), it's preempted, and we should
2173 * yield - it could be a while.
2175 if (unlikely(on_rq)) {
2176 schedule_timeout_uninterruptible(1);
2181 * Ahh, all good. It wasn't running, and it wasn't
2182 * runnable, which means that it will never become
2183 * running in the future either. We're all done!
2190 * kick_process - kick a running thread to enter/exit the kernel
2191 * @p: the to-be-kicked thread
2193 * Cause a process which is running on another CPU to enter
2194 * kernel-mode, without any delay. (to get signals handled.)
2196 * NOTE: this function doesnt have to take the runqueue lock,
2197 * because all it wants to ensure is that the remote task enters
2198 * the kernel. If the IPI races and the task has been migrated
2199 * to another CPU then no harm is done and the purpose has been
2202 void kick_process(struct task_struct *p)
2208 if ((cpu != smp_processor_id()) && task_curr(p))
2209 smp_send_reschedule(cpu);
2214 * Return a low guess at the load of a migration-source cpu weighted
2215 * according to the scheduling class and "nice" value.
2217 * We want to under-estimate the load of migration sources, to
2218 * balance conservatively.
2220 static unsigned long source_load(int cpu, int type)
2222 struct rq *rq = cpu_rq(cpu);
2223 unsigned long total = weighted_cpuload(cpu);
2228 return min(rq->cpu_load[type-1], total);
2232 * Return a high guess at the load of a migration-target cpu weighted
2233 * according to the scheduling class and "nice" value.
2235 static unsigned long target_load(int cpu, int type)
2237 struct rq *rq = cpu_rq(cpu);
2238 unsigned long total = weighted_cpuload(cpu);
2243 return max(rq->cpu_load[type-1], total);
2247 * Return the average load per task on the cpu's run queue
2249 static unsigned long cpu_avg_load_per_task(int cpu)
2251 struct rq *rq = cpu_rq(cpu);
2252 unsigned long total = weighted_cpuload(cpu);
2253 unsigned long n = rq->nr_running;
2255 return n ? total / n : SCHED_LOAD_SCALE;
2259 * find_idlest_group finds and returns the least busy CPU group within the
2262 static struct sched_group *
2263 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2265 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2266 unsigned long min_load = ULONG_MAX, this_load = 0;
2267 int load_idx = sd->forkexec_idx;
2268 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2271 unsigned long load, avg_load;
2275 /* Skip over this group if it has no CPUs allowed */
2276 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2279 local_group = cpu_isset(this_cpu, group->cpumask);
2281 /* Tally up the load of all CPUs in the group */
2284 for_each_cpu_mask(i, group->cpumask) {
2285 /* Bias balancing toward cpus of our domain */
2287 load = source_load(i, load_idx);
2289 load = target_load(i, load_idx);
2294 /* Adjust by relative CPU power of the group */
2295 avg_load = sg_div_cpu_power(group,
2296 avg_load * SCHED_LOAD_SCALE);
2299 this_load = avg_load;
2301 } else if (avg_load < min_load) {
2302 min_load = avg_load;
2305 } while (group = group->next, group != sd->groups);
2307 if (!idlest || 100*this_load < imbalance*min_load)
2313 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2316 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2319 unsigned long load, min_load = ULONG_MAX;
2323 /* Traverse only the allowed CPUs */
2324 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2326 for_each_cpu_mask(i, *tmp) {
2327 load = weighted_cpuload(i);
2329 if (load < min_load || (load == min_load && i == this_cpu)) {
2339 * sched_balance_self: balance the current task (running on cpu) in domains
2340 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2343 * Balance, ie. select the least loaded group.
2345 * Returns the target CPU number, or the same CPU if no balancing is needed.
2347 * preempt must be disabled.
2349 static int sched_balance_self(int cpu, int flag)
2351 struct task_struct *t = current;
2352 struct sched_domain *tmp, *sd = NULL;
2354 for_each_domain(cpu, tmp) {
2356 * If power savings logic is enabled for a domain, stop there.
2358 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2360 if (tmp->flags & flag)
2365 cpumask_t span, tmpmask;
2366 struct sched_group *group;
2367 int new_cpu, weight;
2369 if (!(sd->flags & flag)) {
2375 group = find_idlest_group(sd, t, cpu);
2381 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2382 if (new_cpu == -1 || new_cpu == cpu) {
2383 /* Now try balancing at a lower domain level of cpu */
2388 /* Now try balancing at a lower domain level of new_cpu */
2391 weight = cpus_weight(span);
2392 for_each_domain(cpu, tmp) {
2393 if (weight <= cpus_weight(tmp->span))
2395 if (tmp->flags & flag)
2398 /* while loop will break here if sd == NULL */
2404 #endif /* CONFIG_SMP */
2407 * try_to_wake_up - wake up a thread
2408 * @p: the to-be-woken-up thread
2409 * @state: the mask of task states that can be woken
2410 * @sync: do a synchronous wakeup?
2412 * Put it on the run-queue if it's not already there. The "current"
2413 * thread is always on the run-queue (except when the actual
2414 * re-schedule is in progress), and as such you're allowed to do
2415 * the simpler "current->state = TASK_RUNNING" to mark yourself
2416 * runnable without the overhead of this.
2418 * returns failure only if the task is already active.
2420 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2422 int cpu, orig_cpu, this_cpu, success = 0;
2423 unsigned long flags;
2427 if (!sched_feat(SYNC_WAKEUPS))
2431 rq = task_rq_lock(p, &flags);
2432 old_state = p->state;
2433 if (!(old_state & state))
2441 this_cpu = smp_processor_id();
2444 if (unlikely(task_running(rq, p)))
2447 cpu = p->sched_class->select_task_rq(p, sync);
2448 if (cpu != orig_cpu) {
2449 set_task_cpu(p, cpu);
2450 task_rq_unlock(rq, &flags);
2451 /* might preempt at this point */
2452 rq = task_rq_lock(p, &flags);
2453 old_state = p->state;
2454 if (!(old_state & state))
2459 this_cpu = smp_processor_id();
2463 #ifdef CONFIG_SCHEDSTATS
2464 schedstat_inc(rq, ttwu_count);
2465 if (cpu == this_cpu)
2466 schedstat_inc(rq, ttwu_local);
2468 struct sched_domain *sd;
2469 for_each_domain(this_cpu, sd) {
2470 if (cpu_isset(cpu, sd->span)) {
2471 schedstat_inc(sd, ttwu_wake_remote);
2476 #endif /* CONFIG_SCHEDSTATS */
2479 #endif /* CONFIG_SMP */
2480 schedstat_inc(p, se.nr_wakeups);
2482 schedstat_inc(p, se.nr_wakeups_sync);
2483 if (orig_cpu != cpu)
2484 schedstat_inc(p, se.nr_wakeups_migrate);
2485 if (cpu == this_cpu)
2486 schedstat_inc(p, se.nr_wakeups_local);
2488 schedstat_inc(p, se.nr_wakeups_remote);
2489 update_rq_clock(rq);
2490 activate_task(rq, p, 1);
2494 check_preempt_curr(rq, p);
2496 p->state = TASK_RUNNING;
2498 if (p->sched_class->task_wake_up)
2499 p->sched_class->task_wake_up(rq, p);
2502 task_rq_unlock(rq, &flags);
2507 int wake_up_process(struct task_struct *p)
2509 return try_to_wake_up(p, TASK_ALL, 0);
2511 EXPORT_SYMBOL(wake_up_process);
2513 int wake_up_state(struct task_struct *p, unsigned int state)
2515 return try_to_wake_up(p, state, 0);
2519 * Perform scheduler related setup for a newly forked process p.
2520 * p is forked by current.
2522 * __sched_fork() is basic setup used by init_idle() too:
2524 static void __sched_fork(struct task_struct *p)
2526 p->se.exec_start = 0;
2527 p->se.sum_exec_runtime = 0;
2528 p->se.prev_sum_exec_runtime = 0;
2529 p->se.last_wakeup = 0;
2530 p->se.avg_overlap = 0;
2532 #ifdef CONFIG_SCHEDSTATS
2533 p->se.wait_start = 0;
2534 p->se.sum_sleep_runtime = 0;
2535 p->se.sleep_start = 0;
2536 p->se.block_start = 0;
2537 p->se.sleep_max = 0;
2538 p->se.block_max = 0;
2540 p->se.slice_max = 0;
2544 INIT_LIST_HEAD(&p->rt.run_list);
2546 INIT_LIST_HEAD(&p->se.group_node);
2548 #ifdef CONFIG_PREEMPT_NOTIFIERS
2549 INIT_HLIST_HEAD(&p->preempt_notifiers);
2553 * We mark the process as running here, but have not actually
2554 * inserted it onto the runqueue yet. This guarantees that
2555 * nobody will actually run it, and a signal or other external
2556 * event cannot wake it up and insert it on the runqueue either.
2558 p->state = TASK_RUNNING;
2562 * fork()/clone()-time setup:
2564 void sched_fork(struct task_struct *p, int clone_flags)
2566 int cpu = get_cpu();
2571 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2573 set_task_cpu(p, cpu);
2576 * Make sure we do not leak PI boosting priority to the child:
2578 p->prio = current->normal_prio;
2579 if (!rt_prio(p->prio))
2580 p->sched_class = &fair_sched_class;
2582 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2583 if (likely(sched_info_on()))
2584 memset(&p->sched_info, 0, sizeof(p->sched_info));
2586 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2589 #ifdef CONFIG_PREEMPT
2590 /* Want to start with kernel preemption disabled. */
2591 task_thread_info(p)->preempt_count = 1;
2597 * wake_up_new_task - wake up a newly created task for the first time.
2599 * This function will do some initial scheduler statistics housekeeping
2600 * that must be done for every newly created context, then puts the task
2601 * on the runqueue and wakes it.
2603 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2605 unsigned long flags;
2608 rq = task_rq_lock(p, &flags);
2609 BUG_ON(p->state != TASK_RUNNING);
2610 update_rq_clock(rq);
2612 p->prio = effective_prio(p);
2614 if (!p->sched_class->task_new || !current->se.on_rq) {
2615 activate_task(rq, p, 0);
2618 * Let the scheduling class do new task startup
2619 * management (if any):
2621 p->sched_class->task_new(rq, p);
2624 check_preempt_curr(rq, p);
2626 if (p->sched_class->task_wake_up)
2627 p->sched_class->task_wake_up(rq, p);
2629 task_rq_unlock(rq, &flags);
2632 #ifdef CONFIG_PREEMPT_NOTIFIERS
2635 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2636 * @notifier: notifier struct to register
2638 void preempt_notifier_register(struct preempt_notifier *notifier)
2640 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2642 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2645 * preempt_notifier_unregister - no longer interested in preemption notifications
2646 * @notifier: notifier struct to unregister
2648 * This is safe to call from within a preemption notifier.
2650 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2652 hlist_del(¬ifier->link);
2654 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2656 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2658 struct preempt_notifier *notifier;
2659 struct hlist_node *node;
2661 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2662 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2666 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2667 struct task_struct *next)
2669 struct preempt_notifier *notifier;
2670 struct hlist_node *node;
2672 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2673 notifier->ops->sched_out(notifier, next);
2676 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2678 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2683 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2684 struct task_struct *next)
2688 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2691 * prepare_task_switch - prepare to switch tasks
2692 * @rq: the runqueue preparing to switch
2693 * @prev: the current task that is being switched out
2694 * @next: the task we are going to switch to.
2696 * This is called with the rq lock held and interrupts off. It must
2697 * be paired with a subsequent finish_task_switch after the context
2700 * prepare_task_switch sets up locking and calls architecture specific
2704 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2705 struct task_struct *next)
2707 fire_sched_out_preempt_notifiers(prev, next);
2708 prepare_lock_switch(rq, next);
2709 prepare_arch_switch(next);
2713 * finish_task_switch - clean up after a task-switch
2714 * @rq: runqueue associated with task-switch
2715 * @prev: the thread we just switched away from.
2717 * finish_task_switch must be called after the context switch, paired
2718 * with a prepare_task_switch call before the context switch.
2719 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2720 * and do any other architecture-specific cleanup actions.
2722 * Note that we may have delayed dropping an mm in context_switch(). If
2723 * so, we finish that here outside of the runqueue lock. (Doing it
2724 * with the lock held can cause deadlocks; see schedule() for
2727 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2728 __releases(rq->lock)
2730 struct mm_struct *mm = rq->prev_mm;
2736 * A task struct has one reference for the use as "current".
2737 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2738 * schedule one last time. The schedule call will never return, and
2739 * the scheduled task must drop that reference.
2740 * The test for TASK_DEAD must occur while the runqueue locks are
2741 * still held, otherwise prev could be scheduled on another cpu, die
2742 * there before we look at prev->state, and then the reference would
2744 * Manfred Spraul <manfred@colorfullife.com>
2746 prev_state = prev->state;
2747 finish_arch_switch(prev);
2748 finish_lock_switch(rq, prev);
2750 if (current->sched_class->post_schedule)
2751 current->sched_class->post_schedule(rq);
2754 fire_sched_in_preempt_notifiers(current);
2757 if (unlikely(prev_state == TASK_DEAD)) {
2759 * Remove function-return probe instances associated with this
2760 * task and put them back on the free list.
2762 kprobe_flush_task(prev);
2763 put_task_struct(prev);
2768 * schedule_tail - first thing a freshly forked thread must call.
2769 * @prev: the thread we just switched away from.
2771 asmlinkage void schedule_tail(struct task_struct *prev)
2772 __releases(rq->lock)
2774 struct rq *rq = this_rq();
2776 finish_task_switch(rq, prev);
2777 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2778 /* In this case, finish_task_switch does not reenable preemption */
2781 if (current->set_child_tid)
2782 put_user(task_pid_vnr(current), current->set_child_tid);
2786 * context_switch - switch to the new MM and the new
2787 * thread's register state.
2790 context_switch(struct rq *rq, struct task_struct *prev,
2791 struct task_struct *next)
2793 struct mm_struct *mm, *oldmm;
2795 prepare_task_switch(rq, prev, next);
2797 oldmm = prev->active_mm;
2799 * For paravirt, this is coupled with an exit in switch_to to
2800 * combine the page table reload and the switch backend into
2803 arch_enter_lazy_cpu_mode();
2805 if (unlikely(!mm)) {
2806 next->active_mm = oldmm;
2807 atomic_inc(&oldmm->mm_count);
2808 enter_lazy_tlb(oldmm, next);
2810 switch_mm(oldmm, mm, next);
2812 if (unlikely(!prev->mm)) {
2813 prev->active_mm = NULL;
2814 rq->prev_mm = oldmm;
2817 * Since the runqueue lock will be released by the next
2818 * task (which is an invalid locking op but in the case
2819 * of the scheduler it's an obvious special-case), so we
2820 * do an early lockdep release here:
2822 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2823 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2826 /* Here we just switch the register state and the stack. */
2827 switch_to(prev, next, prev);
2831 * this_rq must be evaluated again because prev may have moved
2832 * CPUs since it called schedule(), thus the 'rq' on its stack
2833 * frame will be invalid.
2835 finish_task_switch(this_rq(), prev);
2839 * nr_running, nr_uninterruptible and nr_context_switches:
2841 * externally visible scheduler statistics: current number of runnable
2842 * threads, current number of uninterruptible-sleeping threads, total
2843 * number of context switches performed since bootup.
2845 unsigned long nr_running(void)
2847 unsigned long i, sum = 0;
2849 for_each_online_cpu(i)
2850 sum += cpu_rq(i)->nr_running;
2855 unsigned long nr_uninterruptible(void)
2857 unsigned long i, sum = 0;
2859 for_each_possible_cpu(i)
2860 sum += cpu_rq(i)->nr_uninterruptible;
2863 * Since we read the counters lockless, it might be slightly
2864 * inaccurate. Do not allow it to go below zero though:
2866 if (unlikely((long)sum < 0))
2872 unsigned long long nr_context_switches(void)
2875 unsigned long long sum = 0;
2877 for_each_possible_cpu(i)
2878 sum += cpu_rq(i)->nr_switches;
2883 unsigned long nr_iowait(void)
2885 unsigned long i, sum = 0;
2887 for_each_possible_cpu(i)
2888 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2893 unsigned long nr_active(void)
2895 unsigned long i, running = 0, uninterruptible = 0;
2897 for_each_online_cpu(i) {
2898 running += cpu_rq(i)->nr_running;
2899 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2902 if (unlikely((long)uninterruptible < 0))
2903 uninterruptible = 0;
2905 return running + uninterruptible;
2909 * Update rq->cpu_load[] statistics. This function is usually called every
2910 * scheduler tick (TICK_NSEC).
2912 static void update_cpu_load(struct rq *this_rq)
2914 unsigned long this_load = this_rq->load.weight;
2917 this_rq->nr_load_updates++;
2919 /* Update our load: */
2920 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2921 unsigned long old_load, new_load;
2923 /* scale is effectively 1 << i now, and >> i divides by scale */
2925 old_load = this_rq->cpu_load[i];
2926 new_load = this_load;
2928 * Round up the averaging division if load is increasing. This
2929 * prevents us from getting stuck on 9 if the load is 10, for
2932 if (new_load > old_load)
2933 new_load += scale-1;
2934 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2941 * double_rq_lock - safely lock two runqueues
2943 * Note this does not disable interrupts like task_rq_lock,
2944 * you need to do so manually before calling.
2946 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2947 __acquires(rq1->lock)
2948 __acquires(rq2->lock)
2950 BUG_ON(!irqs_disabled());
2952 spin_lock(&rq1->lock);
2953 __acquire(rq2->lock); /* Fake it out ;) */
2956 spin_lock(&rq1->lock);
2957 spin_lock(&rq2->lock);
2959 spin_lock(&rq2->lock);
2960 spin_lock(&rq1->lock);
2963 update_rq_clock(rq1);
2964 update_rq_clock(rq2);
2968 * double_rq_unlock - safely unlock two runqueues
2970 * Note this does not restore interrupts like task_rq_unlock,
2971 * you need to do so manually after calling.
2973 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2974 __releases(rq1->lock)
2975 __releases(rq2->lock)
2977 spin_unlock(&rq1->lock);
2979 spin_unlock(&rq2->lock);
2981 __release(rq2->lock);
2985 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2987 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
2988 __releases(this_rq->lock)
2989 __acquires(busiest->lock)
2990 __acquires(this_rq->lock)
2994 if (unlikely(!irqs_disabled())) {
2995 /* printk() doesn't work good under rq->lock */
2996 spin_unlock(&this_rq->lock);
2999 if (unlikely(!spin_trylock(&busiest->lock))) {
3000 if (busiest < this_rq) {
3001 spin_unlock(&this_rq->lock);
3002 spin_lock(&busiest->lock);
3003 spin_lock(&this_rq->lock);
3006 spin_lock(&busiest->lock);
3012 * If dest_cpu is allowed for this process, migrate the task to it.
3013 * This is accomplished by forcing the cpu_allowed mask to only
3014 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3015 * the cpu_allowed mask is restored.
3017 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3019 struct migration_req req;
3020 unsigned long flags;
3023 rq = task_rq_lock(p, &flags);
3024 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3025 || unlikely(cpu_is_offline(dest_cpu)))
3028 /* force the process onto the specified CPU */
3029 if (migrate_task(p, dest_cpu, &req)) {
3030 /* Need to wait for migration thread (might exit: take ref). */
3031 struct task_struct *mt = rq->migration_thread;
3033 get_task_struct(mt);
3034 task_rq_unlock(rq, &flags);
3035 wake_up_process(mt);
3036 put_task_struct(mt);
3037 wait_for_completion(&req.done);
3042 task_rq_unlock(rq, &flags);
3046 * sched_exec - execve() is a valuable balancing opportunity, because at
3047 * this point the task has the smallest effective memory and cache footprint.
3049 void sched_exec(void)
3051 int new_cpu, this_cpu = get_cpu();
3052 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3054 if (new_cpu != this_cpu)
3055 sched_migrate_task(current, new_cpu);
3059 * pull_task - move a task from a remote runqueue to the local runqueue.
3060 * Both runqueues must be locked.
3062 static void pull_task(struct rq *src_rq, struct task_struct *p,
3063 struct rq *this_rq, int this_cpu)
3065 deactivate_task(src_rq, p, 0);
3066 set_task_cpu(p, this_cpu);
3067 activate_task(this_rq, p, 0);
3069 * Note that idle threads have a prio of MAX_PRIO, for this test
3070 * to be always true for them.
3072 check_preempt_curr(this_rq, p);
3076 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3079 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3080 struct sched_domain *sd, enum cpu_idle_type idle,
3084 * We do not migrate tasks that are:
3085 * 1) running (obviously), or
3086 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3087 * 3) are cache-hot on their current CPU.
3089 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3090 schedstat_inc(p, se.nr_failed_migrations_affine);
3095 if (task_running(rq, p)) {
3096 schedstat_inc(p, se.nr_failed_migrations_running);
3101 * Aggressive migration if:
3102 * 1) task is cache cold, or
3103 * 2) too many balance attempts have failed.
3106 if (!task_hot(p, rq->clock, sd) ||
3107 sd->nr_balance_failed > sd->cache_nice_tries) {
3108 #ifdef CONFIG_SCHEDSTATS
3109 if (task_hot(p, rq->clock, sd)) {
3110 schedstat_inc(sd, lb_hot_gained[idle]);
3111 schedstat_inc(p, se.nr_forced_migrations);
3117 if (task_hot(p, rq->clock, sd)) {
3118 schedstat_inc(p, se.nr_failed_migrations_hot);
3124 static unsigned long
3125 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3126 unsigned long max_load_move, struct sched_domain *sd,
3127 enum cpu_idle_type idle, int *all_pinned,
3128 int *this_best_prio, struct rq_iterator *iterator)
3130 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3131 struct task_struct *p;
3132 long rem_load_move = max_load_move;
3134 if (max_load_move == 0)
3140 * Start the load-balancing iterator:
3142 p = iterator->start(iterator->arg);
3144 if (!p || loops++ > sysctl_sched_nr_migrate)
3147 * To help distribute high priority tasks across CPUs we don't
3148 * skip a task if it will be the highest priority task (i.e. smallest
3149 * prio value) on its new queue regardless of its load weight
3151 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3152 SCHED_LOAD_SCALE_FUZZ;
3153 if ((skip_for_load && p->prio >= *this_best_prio) ||
3154 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3155 p = iterator->next(iterator->arg);
3159 pull_task(busiest, p, this_rq, this_cpu);
3161 rem_load_move -= p->se.load.weight;
3164 * We only want to steal up to the prescribed amount of weighted load.
3166 if (rem_load_move > 0) {
3167 if (p->prio < *this_best_prio)
3168 *this_best_prio = p->prio;
3169 p = iterator->next(iterator->arg);
3174 * Right now, this is one of only two places pull_task() is called,
3175 * so we can safely collect pull_task() stats here rather than
3176 * inside pull_task().
3178 schedstat_add(sd, lb_gained[idle], pulled);
3181 *all_pinned = pinned;
3183 return max_load_move - rem_load_move;
3187 * move_tasks tries to move up to max_load_move weighted load from busiest to
3188 * this_rq, as part of a balancing operation within domain "sd".
3189 * Returns 1 if successful and 0 otherwise.
3191 * Called with both runqueues locked.
3193 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3194 unsigned long max_load_move,
3195 struct sched_domain *sd, enum cpu_idle_type idle,
3198 const struct sched_class *class = sched_class_highest;
3199 unsigned long total_load_moved = 0;
3200 int this_best_prio = this_rq->curr->prio;
3204 class->load_balance(this_rq, this_cpu, busiest,
3205 max_load_move - total_load_moved,
3206 sd, idle, all_pinned, &this_best_prio);
3207 class = class->next;
3208 } while (class && max_load_move > total_load_moved);
3210 return total_load_moved > 0;
3214 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3215 struct sched_domain *sd, enum cpu_idle_type idle,
3216 struct rq_iterator *iterator)
3218 struct task_struct *p = iterator->start(iterator->arg);
3222 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3223 pull_task(busiest, p, this_rq, this_cpu);
3225 * Right now, this is only the second place pull_task()
3226 * is called, so we can safely collect pull_task()
3227 * stats here rather than inside pull_task().
3229 schedstat_inc(sd, lb_gained[idle]);
3233 p = iterator->next(iterator->arg);
3240 * move_one_task tries to move exactly one task from busiest to this_rq, as
3241 * part of active balancing operations within "domain".
3242 * Returns 1 if successful and 0 otherwise.
3244 * Called with both runqueues locked.
3246 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3247 struct sched_domain *sd, enum cpu_idle_type idle)
3249 const struct sched_class *class;
3251 for (class = sched_class_highest; class; class = class->next)
3252 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3259 * find_busiest_group finds and returns the busiest CPU group within the
3260 * domain. It calculates and returns the amount of weighted load which
3261 * should be moved to restore balance via the imbalance parameter.
3263 static struct sched_group *
3264 find_busiest_group(struct sched_domain *sd, int this_cpu,
3265 unsigned long *imbalance, enum cpu_idle_type idle,
3266 int *sd_idle, const cpumask_t *cpus, int *balance)
3268 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3269 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3270 unsigned long max_pull;
3271 unsigned long busiest_load_per_task, busiest_nr_running;
3272 unsigned long this_load_per_task, this_nr_running;
3273 int load_idx, group_imb = 0;
3274 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3275 int power_savings_balance = 1;
3276 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3277 unsigned long min_nr_running = ULONG_MAX;
3278 struct sched_group *group_min = NULL, *group_leader = NULL;
3281 max_load = this_load = total_load = total_pwr = 0;
3282 busiest_load_per_task = busiest_nr_running = 0;
3283 this_load_per_task = this_nr_running = 0;
3284 if (idle == CPU_NOT_IDLE)
3285 load_idx = sd->busy_idx;
3286 else if (idle == CPU_NEWLY_IDLE)
3287 load_idx = sd->newidle_idx;
3289 load_idx = sd->idle_idx;
3292 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3295 int __group_imb = 0;
3296 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3297 unsigned long sum_nr_running, sum_weighted_load;
3299 local_group = cpu_isset(this_cpu, group->cpumask);
3302 balance_cpu = first_cpu(group->cpumask);
3304 /* Tally up the load of all CPUs in the group */
3305 sum_weighted_load = sum_nr_running = avg_load = 0;
3307 min_cpu_load = ~0UL;
3309 for_each_cpu_mask(i, group->cpumask) {
3312 if (!cpu_isset(i, *cpus))
3317 if (*sd_idle && rq->nr_running)
3320 /* Bias balancing toward cpus of our domain */
3322 if (idle_cpu(i) && !first_idle_cpu) {
3327 load = target_load(i, load_idx);
3329 load = source_load(i, load_idx);
3330 if (load > max_cpu_load)
3331 max_cpu_load = load;
3332 if (min_cpu_load > load)
3333 min_cpu_load = load;
3337 sum_nr_running += rq->nr_running;
3338 sum_weighted_load += weighted_cpuload(i);
3342 * First idle cpu or the first cpu(busiest) in this sched group
3343 * is eligible for doing load balancing at this and above
3344 * domains. In the newly idle case, we will allow all the cpu's
3345 * to do the newly idle load balance.
3347 if (idle != CPU_NEWLY_IDLE && local_group &&
3348 balance_cpu != this_cpu && balance) {
3353 total_load += avg_load;
3354 total_pwr += group->__cpu_power;
3356 /* Adjust by relative CPU power of the group */
3357 avg_load = sg_div_cpu_power(group,
3358 avg_load * SCHED_LOAD_SCALE);
3360 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3363 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3366 this_load = avg_load;
3368 this_nr_running = sum_nr_running;
3369 this_load_per_task = sum_weighted_load;
3370 } else if (avg_load > max_load &&
3371 (sum_nr_running > group_capacity || __group_imb)) {
3372 max_load = avg_load;
3374 busiest_nr_running = sum_nr_running;
3375 busiest_load_per_task = sum_weighted_load;
3376 group_imb = __group_imb;
3379 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3381 * Busy processors will not participate in power savings
3384 if (idle == CPU_NOT_IDLE ||
3385 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3389 * If the local group is idle or completely loaded
3390 * no need to do power savings balance at this domain
3392 if (local_group && (this_nr_running >= group_capacity ||
3394 power_savings_balance = 0;
3397 * If a group is already running at full capacity or idle,
3398 * don't include that group in power savings calculations
3400 if (!power_savings_balance || sum_nr_running >= group_capacity
3405 * Calculate the group which has the least non-idle load.
3406 * This is the group from where we need to pick up the load
3409 if ((sum_nr_running < min_nr_running) ||
3410 (sum_nr_running == min_nr_running &&
3411 first_cpu(group->cpumask) <
3412 first_cpu(group_min->cpumask))) {
3414 min_nr_running = sum_nr_running;
3415 min_load_per_task = sum_weighted_load /
3420 * Calculate the group which is almost near its
3421 * capacity but still has some space to pick up some load
3422 * from other group and save more power
3424 if (sum_nr_running <= group_capacity - 1) {
3425 if (sum_nr_running > leader_nr_running ||
3426 (sum_nr_running == leader_nr_running &&
3427 first_cpu(group->cpumask) >
3428 first_cpu(group_leader->cpumask))) {
3429 group_leader = group;
3430 leader_nr_running = sum_nr_running;
3435 group = group->next;
3436 } while (group != sd->groups);
3438 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3441 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3443 if (this_load >= avg_load ||
3444 100*max_load <= sd->imbalance_pct*this_load)
3447 busiest_load_per_task /= busiest_nr_running;
3449 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3452 * We're trying to get all the cpus to the average_load, so we don't
3453 * want to push ourselves above the average load, nor do we wish to
3454 * reduce the max loaded cpu below the average load, as either of these
3455 * actions would just result in more rebalancing later, and ping-pong
3456 * tasks around. Thus we look for the minimum possible imbalance.
3457 * Negative imbalances (*we* are more loaded than anyone else) will
3458 * be counted as no imbalance for these purposes -- we can't fix that
3459 * by pulling tasks to us. Be careful of negative numbers as they'll
3460 * appear as very large values with unsigned longs.
3462 if (max_load <= busiest_load_per_task)
3466 * In the presence of smp nice balancing, certain scenarios can have
3467 * max load less than avg load(as we skip the groups at or below
3468 * its cpu_power, while calculating max_load..)
3470 if (max_load < avg_load) {
3472 goto small_imbalance;
3475 /* Don't want to pull so many tasks that a group would go idle */
3476 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3478 /* How much load to actually move to equalise the imbalance */
3479 *imbalance = min(max_pull * busiest->__cpu_power,
3480 (avg_load - this_load) * this->__cpu_power)
3484 * if *imbalance is less than the average load per runnable task
3485 * there is no gaurantee that any tasks will be moved so we'll have
3486 * a think about bumping its value to force at least one task to be
3489 if (*imbalance < busiest_load_per_task) {
3490 unsigned long tmp, pwr_now, pwr_move;
3494 pwr_move = pwr_now = 0;
3496 if (this_nr_running) {
3497 this_load_per_task /= this_nr_running;
3498 if (busiest_load_per_task > this_load_per_task)
3501 this_load_per_task = SCHED_LOAD_SCALE;
3503 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3504 busiest_load_per_task * imbn) {
3505 *imbalance = busiest_load_per_task;
3510 * OK, we don't have enough imbalance to justify moving tasks,
3511 * however we may be able to increase total CPU power used by
3515 pwr_now += busiest->__cpu_power *
3516 min(busiest_load_per_task, max_load);
3517 pwr_now += this->__cpu_power *
3518 min(this_load_per_task, this_load);
3519 pwr_now /= SCHED_LOAD_SCALE;
3521 /* Amount of load we'd subtract */
3522 tmp = sg_div_cpu_power(busiest,
3523 busiest_load_per_task * SCHED_LOAD_SCALE);
3525 pwr_move += busiest->__cpu_power *
3526 min(busiest_load_per_task, max_load - tmp);
3528 /* Amount of load we'd add */
3529 if (max_load * busiest->__cpu_power <
3530 busiest_load_per_task * SCHED_LOAD_SCALE)
3531 tmp = sg_div_cpu_power(this,
3532 max_load * busiest->__cpu_power);
3534 tmp = sg_div_cpu_power(this,
3535 busiest_load_per_task * SCHED_LOAD_SCALE);
3536 pwr_move += this->__cpu_power *
3537 min(this_load_per_task, this_load + tmp);
3538 pwr_move /= SCHED_LOAD_SCALE;
3540 /* Move if we gain throughput */
3541 if (pwr_move > pwr_now)
3542 *imbalance = busiest_load_per_task;
3548 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3549 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3552 if (this == group_leader && group_leader != group_min) {
3553 *imbalance = min_load_per_task;
3563 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3566 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3567 unsigned long imbalance, const cpumask_t *cpus)
3569 struct rq *busiest = NULL, *rq;
3570 unsigned long max_load = 0;
3573 for_each_cpu_mask(i, group->cpumask) {
3576 if (!cpu_isset(i, *cpus))
3580 wl = weighted_cpuload(i);
3582 if (rq->nr_running == 1 && wl > imbalance)
3585 if (wl > max_load) {
3595 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3596 * so long as it is large enough.
3598 #define MAX_PINNED_INTERVAL 512
3601 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3602 * tasks if there is an imbalance.
3604 static int load_balance(int this_cpu, struct rq *this_rq,
3605 struct sched_domain *sd, enum cpu_idle_type idle,
3606 int *balance, cpumask_t *cpus)
3608 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3609 struct sched_group *group;
3610 unsigned long imbalance;
3612 unsigned long flags;
3613 int unlock_aggregate;
3617 unlock_aggregate = get_aggregate(sd);
3620 * When power savings policy is enabled for the parent domain, idle
3621 * sibling can pick up load irrespective of busy siblings. In this case,
3622 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3623 * portraying it as CPU_NOT_IDLE.
3625 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3626 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3629 schedstat_inc(sd, lb_count[idle]);
3632 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3639 schedstat_inc(sd, lb_nobusyg[idle]);
3643 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3645 schedstat_inc(sd, lb_nobusyq[idle]);
3649 BUG_ON(busiest == this_rq);
3651 schedstat_add(sd, lb_imbalance[idle], imbalance);
3654 if (busiest->nr_running > 1) {
3656 * Attempt to move tasks. If find_busiest_group has found
3657 * an imbalance but busiest->nr_running <= 1, the group is
3658 * still unbalanced. ld_moved simply stays zero, so it is
3659 * correctly treated as an imbalance.
3661 local_irq_save(flags);
3662 double_rq_lock(this_rq, busiest);
3663 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3664 imbalance, sd, idle, &all_pinned);
3665 double_rq_unlock(this_rq, busiest);
3666 local_irq_restore(flags);
3669 * some other cpu did the load balance for us.
3671 if (ld_moved && this_cpu != smp_processor_id())
3672 resched_cpu(this_cpu);
3674 /* All tasks on this runqueue were pinned by CPU affinity */
3675 if (unlikely(all_pinned)) {
3676 cpu_clear(cpu_of(busiest), *cpus);
3677 if (!cpus_empty(*cpus))
3684 schedstat_inc(sd, lb_failed[idle]);
3685 sd->nr_balance_failed++;
3687 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3689 spin_lock_irqsave(&busiest->lock, flags);
3691 /* don't kick the migration_thread, if the curr
3692 * task on busiest cpu can't be moved to this_cpu
3694 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3695 spin_unlock_irqrestore(&busiest->lock, flags);
3697 goto out_one_pinned;
3700 if (!busiest->active_balance) {
3701 busiest->active_balance = 1;
3702 busiest->push_cpu = this_cpu;
3705 spin_unlock_irqrestore(&busiest->lock, flags);
3707 wake_up_process(busiest->migration_thread);
3710 * We've kicked active balancing, reset the failure
3713 sd->nr_balance_failed = sd->cache_nice_tries+1;
3716 sd->nr_balance_failed = 0;
3718 if (likely(!active_balance)) {
3719 /* We were unbalanced, so reset the balancing interval */
3720 sd->balance_interval = sd->min_interval;
3723 * If we've begun active balancing, start to back off. This
3724 * case may not be covered by the all_pinned logic if there
3725 * is only 1 task on the busy runqueue (because we don't call
3728 if (sd->balance_interval < sd->max_interval)
3729 sd->balance_interval *= 2;
3732 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3733 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3739 schedstat_inc(sd, lb_balanced[idle]);
3741 sd->nr_balance_failed = 0;
3744 /* tune up the balancing interval */
3745 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3746 (sd->balance_interval < sd->max_interval))
3747 sd->balance_interval *= 2;
3749 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3750 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3755 if (unlock_aggregate)
3761 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3762 * tasks if there is an imbalance.
3764 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3765 * this_rq is locked.
3768 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3771 struct sched_group *group;
3772 struct rq *busiest = NULL;
3773 unsigned long imbalance;
3781 * When power savings policy is enabled for the parent domain, idle
3782 * sibling can pick up load irrespective of busy siblings. In this case,
3783 * let the state of idle sibling percolate up as IDLE, instead of
3784 * portraying it as CPU_NOT_IDLE.
3786 if (sd->flags & SD_SHARE_CPUPOWER &&
3787 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3790 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3792 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3793 &sd_idle, cpus, NULL);
3795 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3799 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3801 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3805 BUG_ON(busiest == this_rq);
3807 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3810 if (busiest->nr_running > 1) {
3811 /* Attempt to move tasks */
3812 double_lock_balance(this_rq, busiest);
3813 /* this_rq->clock is already updated */
3814 update_rq_clock(busiest);
3815 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3816 imbalance, sd, CPU_NEWLY_IDLE,
3818 spin_unlock(&busiest->lock);
3820 if (unlikely(all_pinned)) {
3821 cpu_clear(cpu_of(busiest), *cpus);
3822 if (!cpus_empty(*cpus))
3828 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3829 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3830 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3833 sd->nr_balance_failed = 0;
3838 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3839 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3840 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3842 sd->nr_balance_failed = 0;
3848 * idle_balance is called by schedule() if this_cpu is about to become
3849 * idle. Attempts to pull tasks from other CPUs.
3851 static void idle_balance(int this_cpu, struct rq *this_rq)
3853 struct sched_domain *sd;
3854 int pulled_task = -1;
3855 unsigned long next_balance = jiffies + HZ;
3858 for_each_domain(this_cpu, sd) {
3859 unsigned long interval;
3861 if (!(sd->flags & SD_LOAD_BALANCE))
3864 if (sd->flags & SD_BALANCE_NEWIDLE)
3865 /* If we've pulled tasks over stop searching: */
3866 pulled_task = load_balance_newidle(this_cpu, this_rq,
3869 interval = msecs_to_jiffies(sd->balance_interval);
3870 if (time_after(next_balance, sd->last_balance + interval))
3871 next_balance = sd->last_balance + interval;
3875 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3877 * We are going idle. next_balance may be set based on
3878 * a busy processor. So reset next_balance.
3880 this_rq->next_balance = next_balance;
3885 * active_load_balance is run by migration threads. It pushes running tasks
3886 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3887 * running on each physical CPU where possible, and avoids physical /
3888 * logical imbalances.
3890 * Called with busiest_rq locked.
3892 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3894 int target_cpu = busiest_rq->push_cpu;
3895 struct sched_domain *sd;
3896 struct rq *target_rq;
3898 /* Is there any task to move? */
3899 if (busiest_rq->nr_running <= 1)
3902 target_rq = cpu_rq(target_cpu);
3905 * This condition is "impossible", if it occurs
3906 * we need to fix it. Originally reported by
3907 * Bjorn Helgaas on a 128-cpu setup.
3909 BUG_ON(busiest_rq == target_rq);
3911 /* move a task from busiest_rq to target_rq */
3912 double_lock_balance(busiest_rq, target_rq);
3913 update_rq_clock(busiest_rq);
3914 update_rq_clock(target_rq);
3916 /* Search for an sd spanning us and the target CPU. */
3917 for_each_domain(target_cpu, sd) {
3918 if ((sd->flags & SD_LOAD_BALANCE) &&
3919 cpu_isset(busiest_cpu, sd->span))
3924 schedstat_inc(sd, alb_count);
3926 if (move_one_task(target_rq, target_cpu, busiest_rq,
3928 schedstat_inc(sd, alb_pushed);
3930 schedstat_inc(sd, alb_failed);
3932 spin_unlock(&target_rq->lock);
3937 atomic_t load_balancer;
3939 } nohz ____cacheline_aligned = {
3940 .load_balancer = ATOMIC_INIT(-1),
3941 .cpu_mask = CPU_MASK_NONE,
3945 * This routine will try to nominate the ilb (idle load balancing)
3946 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3947 * load balancing on behalf of all those cpus. If all the cpus in the system
3948 * go into this tickless mode, then there will be no ilb owner (as there is
3949 * no need for one) and all the cpus will sleep till the next wakeup event
3952 * For the ilb owner, tick is not stopped. And this tick will be used
3953 * for idle load balancing. ilb owner will still be part of
3956 * While stopping the tick, this cpu will become the ilb owner if there
3957 * is no other owner. And will be the owner till that cpu becomes busy
3958 * or if all cpus in the system stop their ticks at which point
3959 * there is no need for ilb owner.
3961 * When the ilb owner becomes busy, it nominates another owner, during the
3962 * next busy scheduler_tick()
3964 int select_nohz_load_balancer(int stop_tick)
3966 int cpu = smp_processor_id();
3969 cpu_set(cpu, nohz.cpu_mask);
3970 cpu_rq(cpu)->in_nohz_recently = 1;
3973 * If we are going offline and still the leader, give up!
3975 if (cpu_is_offline(cpu) &&
3976 atomic_read(&nohz.load_balancer) == cpu) {
3977 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3982 /* time for ilb owner also to sleep */
3983 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
3984 if (atomic_read(&nohz.load_balancer) == cpu)
3985 atomic_set(&nohz.load_balancer, -1);
3989 if (atomic_read(&nohz.load_balancer) == -1) {
3990 /* make me the ilb owner */
3991 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3993 } else if (atomic_read(&nohz.load_balancer) == cpu)
3996 if (!cpu_isset(cpu, nohz.cpu_mask))
3999 cpu_clear(cpu, nohz.cpu_mask);
4001 if (atomic_read(&nohz.load_balancer) == cpu)
4002 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4009 static DEFINE_SPINLOCK(balancing);
4012 * It checks each scheduling domain to see if it is due to be balanced,
4013 * and initiates a balancing operation if so.
4015 * Balancing parameters are set up in arch_init_sched_domains.
4017 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4020 struct rq *rq = cpu_rq(cpu);
4021 unsigned long interval;
4022 struct sched_domain *sd;
4023 /* Earliest time when we have to do rebalance again */
4024 unsigned long next_balance = jiffies + 60*HZ;
4025 int update_next_balance = 0;
4029 for_each_domain(cpu, sd) {
4030 if (!(sd->flags & SD_LOAD_BALANCE))
4033 interval = sd->balance_interval;
4034 if (idle != CPU_IDLE)
4035 interval *= sd->busy_factor;
4037 /* scale ms to jiffies */
4038 interval = msecs_to_jiffies(interval);
4039 if (unlikely(!interval))
4041 if (interval > HZ*NR_CPUS/10)
4042 interval = HZ*NR_CPUS/10;
4044 need_serialize = sd->flags & SD_SERIALIZE;
4046 if (need_serialize) {
4047 if (!spin_trylock(&balancing))
4051 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4052 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4054 * We've pulled tasks over so either we're no
4055 * longer idle, or one of our SMT siblings is
4058 idle = CPU_NOT_IDLE;
4060 sd->last_balance = jiffies;
4063 spin_unlock(&balancing);
4065 if (time_after(next_balance, sd->last_balance + interval)) {
4066 next_balance = sd->last_balance + interval;
4067 update_next_balance = 1;
4071 * Stop the load balance at this level. There is another
4072 * CPU in our sched group which is doing load balancing more
4080 * next_balance will be updated only when there is a need.
4081 * When the cpu is attached to null domain for ex, it will not be
4084 if (likely(update_next_balance))
4085 rq->next_balance = next_balance;
4089 * run_rebalance_domains is triggered when needed from the scheduler tick.
4090 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4091 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4093 static void run_rebalance_domains(struct softirq_action *h)
4095 int this_cpu = smp_processor_id();
4096 struct rq *this_rq = cpu_rq(this_cpu);
4097 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4098 CPU_IDLE : CPU_NOT_IDLE;
4100 rebalance_domains(this_cpu, idle);
4104 * If this cpu is the owner for idle load balancing, then do the
4105 * balancing on behalf of the other idle cpus whose ticks are
4108 if (this_rq->idle_at_tick &&
4109 atomic_read(&nohz.load_balancer) == this_cpu) {
4110 cpumask_t cpus = nohz.cpu_mask;
4114 cpu_clear(this_cpu, cpus);
4115 for_each_cpu_mask(balance_cpu, cpus) {
4117 * If this cpu gets work to do, stop the load balancing
4118 * work being done for other cpus. Next load
4119 * balancing owner will pick it up.
4124 rebalance_domains(balance_cpu, CPU_IDLE);
4126 rq = cpu_rq(balance_cpu);
4127 if (time_after(this_rq->next_balance, rq->next_balance))
4128 this_rq->next_balance = rq->next_balance;
4135 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4137 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4138 * idle load balancing owner or decide to stop the periodic load balancing,
4139 * if the whole system is idle.
4141 static inline void trigger_load_balance(struct rq *rq, int cpu)
4145 * If we were in the nohz mode recently and busy at the current
4146 * scheduler tick, then check if we need to nominate new idle
4149 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4150 rq->in_nohz_recently = 0;
4152 if (atomic_read(&nohz.load_balancer) == cpu) {
4153 cpu_clear(cpu, nohz.cpu_mask);
4154 atomic_set(&nohz.load_balancer, -1);
4157 if (atomic_read(&nohz.load_balancer) == -1) {
4159 * simple selection for now: Nominate the
4160 * first cpu in the nohz list to be the next
4163 * TBD: Traverse the sched domains and nominate
4164 * the nearest cpu in the nohz.cpu_mask.
4166 int ilb = first_cpu(nohz.cpu_mask);
4168 if (ilb < nr_cpu_ids)
4174 * If this cpu is idle and doing idle load balancing for all the
4175 * cpus with ticks stopped, is it time for that to stop?
4177 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4178 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4184 * If this cpu is idle and the idle load balancing is done by
4185 * someone else, then no need raise the SCHED_SOFTIRQ
4187 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4188 cpu_isset(cpu, nohz.cpu_mask))
4191 if (time_after_eq(jiffies, rq->next_balance))
4192 raise_softirq(SCHED_SOFTIRQ);
4195 #else /* CONFIG_SMP */
4198 * on UP we do not need to balance between CPUs:
4200 static inline void idle_balance(int cpu, struct rq *rq)
4206 DEFINE_PER_CPU(struct kernel_stat, kstat);
4208 EXPORT_PER_CPU_SYMBOL(kstat);
4211 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4212 * that have not yet been banked in case the task is currently running.
4214 unsigned long long task_sched_runtime(struct task_struct *p)
4216 unsigned long flags;
4220 rq = task_rq_lock(p, &flags);
4221 ns = p->se.sum_exec_runtime;
4222 if (task_current(rq, p)) {
4223 update_rq_clock(rq);
4224 delta_exec = rq->clock - p->se.exec_start;
4225 if ((s64)delta_exec > 0)
4228 task_rq_unlock(rq, &flags);
4234 * Account user cpu time to a process.
4235 * @p: the process that the cpu time gets accounted to
4236 * @cputime: the cpu time spent in user space since the last update
4238 void account_user_time(struct task_struct *p, cputime_t cputime)
4240 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4243 p->utime = cputime_add(p->utime, cputime);
4245 /* Add user time to cpustat. */
4246 tmp = cputime_to_cputime64(cputime);
4247 if (TASK_NICE(p) > 0)
4248 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4250 cpustat->user = cputime64_add(cpustat->user, tmp);
4254 * Account guest cpu time to a process.
4255 * @p: the process that the cpu time gets accounted to
4256 * @cputime: the cpu time spent in virtual machine since the last update
4258 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4261 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4263 tmp = cputime_to_cputime64(cputime);
4265 p->utime = cputime_add(p->utime, cputime);
4266 p->gtime = cputime_add(p->gtime, cputime);
4268 cpustat->user = cputime64_add(cpustat->user, tmp);
4269 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4273 * Account scaled user cpu time to a process.
4274 * @p: the process that the cpu time gets accounted to
4275 * @cputime: the cpu time spent in user space since the last update
4277 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4279 p->utimescaled = cputime_add(p->utimescaled, cputime);
4283 * Account system cpu time to a process.
4284 * @p: the process that the cpu time gets accounted to
4285 * @hardirq_offset: the offset to subtract from hardirq_count()
4286 * @cputime: the cpu time spent in kernel space since the last update
4288 void account_system_time(struct task_struct *p, int hardirq_offset,
4291 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4292 struct rq *rq = this_rq();
4295 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4296 account_guest_time(p, cputime);
4300 p->stime = cputime_add(p->stime, cputime);
4302 /* Add system time to cpustat. */
4303 tmp = cputime_to_cputime64(cputime);
4304 if (hardirq_count() - hardirq_offset)
4305 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4306 else if (softirq_count())
4307 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4308 else if (p != rq->idle)
4309 cpustat->system = cputime64_add(cpustat->system, tmp);
4310 else if (atomic_read(&rq->nr_iowait) > 0)
4311 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4313 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4314 /* Account for system time used */
4315 acct_update_integrals(p);
4319 * Account scaled system cpu time to a process.
4320 * @p: the process that the cpu time gets accounted to
4321 * @hardirq_offset: the offset to subtract from hardirq_count()
4322 * @cputime: the cpu time spent in kernel space since the last update
4324 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4326 p->stimescaled = cputime_add(p->stimescaled, cputime);
4330 * Account for involuntary wait time.
4331 * @p: the process from which the cpu time has been stolen
4332 * @steal: the cpu time spent in involuntary wait
4334 void account_steal_time(struct task_struct *p, cputime_t steal)
4336 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4337 cputime64_t tmp = cputime_to_cputime64(steal);
4338 struct rq *rq = this_rq();
4340 if (p == rq->idle) {
4341 p->stime = cputime_add(p->stime, steal);
4342 if (atomic_read(&rq->nr_iowait) > 0)
4343 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4345 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4347 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4351 * This function gets called by the timer code, with HZ frequency.
4352 * We call it with interrupts disabled.
4354 * It also gets called by the fork code, when changing the parent's
4357 void scheduler_tick(void)
4359 int cpu = smp_processor_id();
4360 struct rq *rq = cpu_rq(cpu);
4361 struct task_struct *curr = rq->curr;
4365 spin_lock(&rq->lock);
4366 update_rq_clock(rq);
4367 update_cpu_load(rq);
4368 curr->sched_class->task_tick(rq, curr, 0);
4369 spin_unlock(&rq->lock);
4372 rq->idle_at_tick = idle_cpu(cpu);
4373 trigger_load_balance(rq, cpu);
4377 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4379 void __kprobes add_preempt_count(int val)
4384 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4386 preempt_count() += val;
4388 * Spinlock count overflowing soon?
4390 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4393 EXPORT_SYMBOL(add_preempt_count);
4395 void __kprobes sub_preempt_count(int val)
4400 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4403 * Is the spinlock portion underflowing?
4405 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4406 !(preempt_count() & PREEMPT_MASK)))
4409 preempt_count() -= val;
4411 EXPORT_SYMBOL(sub_preempt_count);
4416 * Print scheduling while atomic bug:
4418 static noinline void __schedule_bug(struct task_struct *prev)
4420 struct pt_regs *regs = get_irq_regs();
4422 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4423 prev->comm, prev->pid, preempt_count());
4425 debug_show_held_locks(prev);
4427 if (irqs_disabled())
4428 print_irqtrace_events(prev);
4437 * Various schedule()-time debugging checks and statistics:
4439 static inline void schedule_debug(struct task_struct *prev)
4442 * Test if we are atomic. Since do_exit() needs to call into
4443 * schedule() atomically, we ignore that path for now.
4444 * Otherwise, whine if we are scheduling when we should not be.
4446 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4447 __schedule_bug(prev);
4449 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4451 schedstat_inc(this_rq(), sched_count);
4452 #ifdef CONFIG_SCHEDSTATS
4453 if (unlikely(prev->lock_depth >= 0)) {
4454 schedstat_inc(this_rq(), bkl_count);
4455 schedstat_inc(prev, sched_info.bkl_count);
4461 * Pick up the highest-prio task:
4463 static inline struct task_struct *
4464 pick_next_task(struct rq *rq, struct task_struct *prev)
4466 const struct sched_class *class;
4467 struct task_struct *p;
4470 * Optimization: we know that if all tasks are in
4471 * the fair class we can call that function directly:
4473 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4474 p = fair_sched_class.pick_next_task(rq);
4479 class = sched_class_highest;
4481 p = class->pick_next_task(rq);
4485 * Will never be NULL as the idle class always
4486 * returns a non-NULL p:
4488 class = class->next;
4493 * schedule() is the main scheduler function.
4495 asmlinkage void __sched schedule(void)
4497 struct task_struct *prev, *next;
4498 unsigned long *switch_count;
4500 int cpu, hrtick = sched_feat(HRTICK);
4504 cpu = smp_processor_id();
4508 switch_count = &prev->nivcsw;
4510 release_kernel_lock(prev);
4511 need_resched_nonpreemptible:
4513 schedule_debug(prev);
4519 * Do the rq-clock update outside the rq lock:
4521 local_irq_disable();
4522 update_rq_clock(rq);
4523 spin_lock(&rq->lock);
4524 clear_tsk_need_resched(prev);
4526 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4527 if (unlikely(signal_pending_state(prev->state, prev)))
4528 prev->state = TASK_RUNNING;
4530 deactivate_task(rq, prev, 1);
4531 switch_count = &prev->nvcsw;
4535 if (prev->sched_class->pre_schedule)
4536 prev->sched_class->pre_schedule(rq, prev);
4539 if (unlikely(!rq->nr_running))
4540 idle_balance(cpu, rq);
4542 prev->sched_class->put_prev_task(rq, prev);
4543 next = pick_next_task(rq, prev);
4545 if (likely(prev != next)) {
4546 sched_info_switch(prev, next);
4552 context_switch(rq, prev, next); /* unlocks the rq */
4554 * the context switch might have flipped the stack from under
4555 * us, hence refresh the local variables.
4557 cpu = smp_processor_id();
4560 spin_unlock_irq(&rq->lock);
4565 if (unlikely(reacquire_kernel_lock(current) < 0))
4566 goto need_resched_nonpreemptible;
4568 preempt_enable_no_resched();
4569 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4572 EXPORT_SYMBOL(schedule);
4574 #ifdef CONFIG_PREEMPT
4576 * this is the entry point to schedule() from in-kernel preemption
4577 * off of preempt_enable. Kernel preemptions off return from interrupt
4578 * occur there and call schedule directly.
4580 asmlinkage void __sched preempt_schedule(void)
4582 struct thread_info *ti = current_thread_info();
4585 * If there is a non-zero preempt_count or interrupts are disabled,
4586 * we do not want to preempt the current task. Just return..
4588 if (likely(ti->preempt_count || irqs_disabled()))
4592 add_preempt_count(PREEMPT_ACTIVE);
4594 sub_preempt_count(PREEMPT_ACTIVE);
4597 * Check again in case we missed a preemption opportunity
4598 * between schedule and now.
4601 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4603 EXPORT_SYMBOL(preempt_schedule);
4606 * this is the entry point to schedule() from kernel preemption
4607 * off of irq context.
4608 * Note, that this is called and return with irqs disabled. This will
4609 * protect us against recursive calling from irq.
4611 asmlinkage void __sched preempt_schedule_irq(void)
4613 struct thread_info *ti = current_thread_info();
4615 /* Catch callers which need to be fixed */
4616 BUG_ON(ti->preempt_count || !irqs_disabled());
4619 add_preempt_count(PREEMPT_ACTIVE);
4622 local_irq_disable();
4623 sub_preempt_count(PREEMPT_ACTIVE);
4626 * Check again in case we missed a preemption opportunity
4627 * between schedule and now.
4630 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4633 #endif /* CONFIG_PREEMPT */
4635 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4638 return try_to_wake_up(curr->private, mode, sync);
4640 EXPORT_SYMBOL(default_wake_function);
4643 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4644 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4645 * number) then we wake all the non-exclusive tasks and one exclusive task.
4647 * There are circumstances in which we can try to wake a task which has already
4648 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4649 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4651 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4652 int nr_exclusive, int sync, void *key)
4654 wait_queue_t *curr, *next;
4656 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4657 unsigned flags = curr->flags;
4659 if (curr->func(curr, mode, sync, key) &&
4660 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4666 * __wake_up - wake up threads blocked on a waitqueue.
4668 * @mode: which threads
4669 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4670 * @key: is directly passed to the wakeup function
4672 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4673 int nr_exclusive, void *key)
4675 unsigned long flags;
4677 spin_lock_irqsave(&q->lock, flags);
4678 __wake_up_common(q, mode, nr_exclusive, 0, key);
4679 spin_unlock_irqrestore(&q->lock, flags);
4681 EXPORT_SYMBOL(__wake_up);
4684 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4686 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4688 __wake_up_common(q, mode, 1, 0, NULL);
4692 * __wake_up_sync - wake up threads blocked on a waitqueue.
4694 * @mode: which threads
4695 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4697 * The sync wakeup differs that the waker knows that it will schedule
4698 * away soon, so while the target thread will be woken up, it will not
4699 * be migrated to another CPU - ie. the two threads are 'synchronized'
4700 * with each other. This can prevent needless bouncing between CPUs.
4702 * On UP it can prevent extra preemption.
4705 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4707 unsigned long flags;
4713 if (unlikely(!nr_exclusive))
4716 spin_lock_irqsave(&q->lock, flags);
4717 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4718 spin_unlock_irqrestore(&q->lock, flags);
4720 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4722 void complete(struct completion *x)
4724 unsigned long flags;
4726 spin_lock_irqsave(&x->wait.lock, flags);
4728 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4729 spin_unlock_irqrestore(&x->wait.lock, flags);
4731 EXPORT_SYMBOL(complete);
4733 void complete_all(struct completion *x)
4735 unsigned long flags;
4737 spin_lock_irqsave(&x->wait.lock, flags);
4738 x->done += UINT_MAX/2;
4739 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4740 spin_unlock_irqrestore(&x->wait.lock, flags);
4742 EXPORT_SYMBOL(complete_all);
4744 static inline long __sched
4745 do_wait_for_common(struct completion *x, long timeout, int state)
4748 DECLARE_WAITQUEUE(wait, current);
4750 wait.flags |= WQ_FLAG_EXCLUSIVE;
4751 __add_wait_queue_tail(&x->wait, &wait);
4753 if ((state == TASK_INTERRUPTIBLE &&
4754 signal_pending(current)) ||
4755 (state == TASK_KILLABLE &&
4756 fatal_signal_pending(current))) {
4757 timeout = -ERESTARTSYS;
4760 __set_current_state(state);
4761 spin_unlock_irq(&x->wait.lock);
4762 timeout = schedule_timeout(timeout);
4763 spin_lock_irq(&x->wait.lock);
4764 } while (!x->done && timeout);
4765 __remove_wait_queue(&x->wait, &wait);
4770 return timeout ?: 1;
4774 wait_for_common(struct completion *x, long timeout, int state)
4778 spin_lock_irq(&x->wait.lock);
4779 timeout = do_wait_for_common(x, timeout, state);
4780 spin_unlock_irq(&x->wait.lock);
4784 void __sched wait_for_completion(struct completion *x)
4786 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4788 EXPORT_SYMBOL(wait_for_completion);
4790 unsigned long __sched
4791 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4793 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4795 EXPORT_SYMBOL(wait_for_completion_timeout);
4797 int __sched wait_for_completion_interruptible(struct completion *x)
4799 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4800 if (t == -ERESTARTSYS)
4804 EXPORT_SYMBOL(wait_for_completion_interruptible);
4806 unsigned long __sched
4807 wait_for_completion_interruptible_timeout(struct completion *x,
4808 unsigned long timeout)
4810 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4812 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4814 int __sched wait_for_completion_killable(struct completion *x)
4816 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4817 if (t == -ERESTARTSYS)
4821 EXPORT_SYMBOL(wait_for_completion_killable);
4824 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4826 unsigned long flags;
4829 init_waitqueue_entry(&wait, current);
4831 __set_current_state(state);
4833 spin_lock_irqsave(&q->lock, flags);
4834 __add_wait_queue(q, &wait);
4835 spin_unlock(&q->lock);
4836 timeout = schedule_timeout(timeout);
4837 spin_lock_irq(&q->lock);
4838 __remove_wait_queue(q, &wait);
4839 spin_unlock_irqrestore(&q->lock, flags);
4844 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4846 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4848 EXPORT_SYMBOL(interruptible_sleep_on);
4851 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4853 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4855 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4857 void __sched sleep_on(wait_queue_head_t *q)
4859 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4861 EXPORT_SYMBOL(sleep_on);
4863 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4865 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4867 EXPORT_SYMBOL(sleep_on_timeout);
4869 #ifdef CONFIG_RT_MUTEXES
4872 * rt_mutex_setprio - set the current priority of a task
4874 * @prio: prio value (kernel-internal form)
4876 * This function changes the 'effective' priority of a task. It does
4877 * not touch ->normal_prio like __setscheduler().
4879 * Used by the rt_mutex code to implement priority inheritance logic.
4881 void rt_mutex_setprio(struct task_struct *p, int prio)
4883 unsigned long flags;
4884 int oldprio, on_rq, running;
4886 const struct sched_class *prev_class = p->sched_class;
4888 BUG_ON(prio < 0 || prio > MAX_PRIO);
4890 rq = task_rq_lock(p, &flags);
4891 update_rq_clock(rq);
4894 on_rq = p->se.on_rq;
4895 running = task_current(rq, p);
4897 dequeue_task(rq, p, 0);
4899 p->sched_class->put_prev_task(rq, p);
4902 p->sched_class = &rt_sched_class;
4904 p->sched_class = &fair_sched_class;
4909 p->sched_class->set_curr_task(rq);
4911 enqueue_task(rq, p, 0);
4913 check_class_changed(rq, p, prev_class, oldprio, running);
4915 task_rq_unlock(rq, &flags);
4920 void set_user_nice(struct task_struct *p, long nice)
4922 int old_prio, delta, on_rq;
4923 unsigned long flags;
4926 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4929 * We have to be careful, if called from sys_setpriority(),
4930 * the task might be in the middle of scheduling on another CPU.
4932 rq = task_rq_lock(p, &flags);
4933 update_rq_clock(rq);
4935 * The RT priorities are set via sched_setscheduler(), but we still
4936 * allow the 'normal' nice value to be set - but as expected
4937 * it wont have any effect on scheduling until the task is
4938 * SCHED_FIFO/SCHED_RR:
4940 if (task_has_rt_policy(p)) {
4941 p->static_prio = NICE_TO_PRIO(nice);
4944 on_rq = p->se.on_rq;
4946 dequeue_task(rq, p, 0);
4948 p->static_prio = NICE_TO_PRIO(nice);
4951 p->prio = effective_prio(p);
4952 delta = p->prio - old_prio;
4955 enqueue_task(rq, p, 0);
4957 * If the task increased its priority or is running and
4958 * lowered its priority, then reschedule its CPU:
4960 if (delta < 0 || (delta > 0 && task_running(rq, p)))
4961 resched_task(rq->curr);
4964 task_rq_unlock(rq, &flags);
4966 EXPORT_SYMBOL(set_user_nice);
4969 * can_nice - check if a task can reduce its nice value
4973 int can_nice(const struct task_struct *p, const int nice)
4975 /* convert nice value [19,-20] to rlimit style value [1,40] */
4976 int nice_rlim = 20 - nice;
4978 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
4979 capable(CAP_SYS_NICE));
4982 #ifdef __ARCH_WANT_SYS_NICE
4985 * sys_nice - change the priority of the current process.
4986 * @increment: priority increment
4988 * sys_setpriority is a more generic, but much slower function that
4989 * does similar things.
4991 asmlinkage long sys_nice(int increment)
4996 * Setpriority might change our priority at the same moment.
4997 * We don't have to worry. Conceptually one call occurs first
4998 * and we have a single winner.
5000 if (increment < -40)
5005 nice = PRIO_TO_NICE(current->static_prio) + increment;
5011 if (increment < 0 && !can_nice(current, nice))
5014 retval = security_task_setnice(current, nice);
5018 set_user_nice(current, nice);
5025 * task_prio - return the priority value of a given task.
5026 * @p: the task in question.
5028 * This is the priority value as seen by users in /proc.
5029 * RT tasks are offset by -200. Normal tasks are centered
5030 * around 0, value goes from -16 to +15.
5032 int task_prio(const struct task_struct *p)
5034 return p->prio - MAX_RT_PRIO;
5038 * task_nice - return the nice value of a given task.
5039 * @p: the task in question.
5041 int task_nice(const struct task_struct *p)
5043 return TASK_NICE(p);
5045 EXPORT_SYMBOL(task_nice);
5048 * idle_cpu - is a given cpu idle currently?
5049 * @cpu: the processor in question.
5051 int idle_cpu(int cpu)
5053 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5057 * idle_task - return the idle task for a given cpu.
5058 * @cpu: the processor in question.
5060 struct task_struct *idle_task(int cpu)
5062 return cpu_rq(cpu)->idle;
5066 * find_process_by_pid - find a process with a matching PID value.
5067 * @pid: the pid in question.
5069 static struct task_struct *find_process_by_pid(pid_t pid)
5071 return pid ? find_task_by_vpid(pid) : current;
5074 /* Actually do priority change: must hold rq lock. */
5076 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5078 BUG_ON(p->se.on_rq);
5081 switch (p->policy) {
5085 p->sched_class = &fair_sched_class;
5089 p->sched_class = &rt_sched_class;
5093 p->rt_priority = prio;
5094 p->normal_prio = normal_prio(p);
5095 /* we are holding p->pi_lock already */
5096 p->prio = rt_mutex_getprio(p);
5101 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5102 * @p: the task in question.
5103 * @policy: new policy.
5104 * @param: structure containing the new RT priority.
5106 * NOTE that the task may be already dead.
5108 int sched_setscheduler(struct task_struct *p, int policy,
5109 struct sched_param *param)
5111 int retval, oldprio, oldpolicy = -1, on_rq, running;
5112 unsigned long flags;
5113 const struct sched_class *prev_class = p->sched_class;
5116 /* may grab non-irq protected spin_locks */
5117 BUG_ON(in_interrupt());
5119 /* double check policy once rq lock held */
5121 policy = oldpolicy = p->policy;
5122 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5123 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5124 policy != SCHED_IDLE)
5127 * Valid priorities for SCHED_FIFO and SCHED_RR are
5128 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5129 * SCHED_BATCH and SCHED_IDLE is 0.
5131 if (param->sched_priority < 0 ||
5132 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5133 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5135 if (rt_policy(policy) != (param->sched_priority != 0))
5139 * Allow unprivileged RT tasks to decrease priority:
5141 if (!capable(CAP_SYS_NICE)) {
5142 if (rt_policy(policy)) {
5143 unsigned long rlim_rtprio;
5145 if (!lock_task_sighand(p, &flags))
5147 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5148 unlock_task_sighand(p, &flags);
5150 /* can't set/change the rt policy */
5151 if (policy != p->policy && !rlim_rtprio)
5154 /* can't increase priority */
5155 if (param->sched_priority > p->rt_priority &&
5156 param->sched_priority > rlim_rtprio)
5160 * Like positive nice levels, dont allow tasks to
5161 * move out of SCHED_IDLE either:
5163 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5166 /* can't change other user's priorities */
5167 if ((current->euid != p->euid) &&
5168 (current->euid != p->uid))
5172 #ifdef CONFIG_RT_GROUP_SCHED
5174 * Do not allow realtime tasks into groups that have no runtime
5177 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5181 retval = security_task_setscheduler(p, policy, param);
5185 * make sure no PI-waiters arrive (or leave) while we are
5186 * changing the priority of the task:
5188 spin_lock_irqsave(&p->pi_lock, flags);
5190 * To be able to change p->policy safely, the apropriate
5191 * runqueue lock must be held.
5193 rq = __task_rq_lock(p);
5194 /* recheck policy now with rq lock held */
5195 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5196 policy = oldpolicy = -1;
5197 __task_rq_unlock(rq);
5198 spin_unlock_irqrestore(&p->pi_lock, flags);
5201 update_rq_clock(rq);
5202 on_rq = p->se.on_rq;
5203 running = task_current(rq, p);
5205 deactivate_task(rq, p, 0);
5207 p->sched_class->put_prev_task(rq, p);
5210 __setscheduler(rq, p, policy, param->sched_priority);
5213 p->sched_class->set_curr_task(rq);
5215 activate_task(rq, p, 0);
5217 check_class_changed(rq, p, prev_class, oldprio, running);
5219 __task_rq_unlock(rq);
5220 spin_unlock_irqrestore(&p->pi_lock, flags);
5222 rt_mutex_adjust_pi(p);
5226 EXPORT_SYMBOL_GPL(sched_setscheduler);
5229 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5231 struct sched_param lparam;
5232 struct task_struct *p;
5235 if (!param || pid < 0)
5237 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5242 p = find_process_by_pid(pid);
5244 retval = sched_setscheduler(p, policy, &lparam);
5251 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5252 * @pid: the pid in question.
5253 * @policy: new policy.
5254 * @param: structure containing the new RT priority.
5257 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5259 /* negative values for policy are not valid */
5263 return do_sched_setscheduler(pid, policy, param);
5267 * sys_sched_setparam - set/change the RT priority of a thread
5268 * @pid: the pid in question.
5269 * @param: structure containing the new RT priority.
5271 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5273 return do_sched_setscheduler(pid, -1, param);
5277 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5278 * @pid: the pid in question.
5280 asmlinkage long sys_sched_getscheduler(pid_t pid)
5282 struct task_struct *p;
5289 read_lock(&tasklist_lock);
5290 p = find_process_by_pid(pid);
5292 retval = security_task_getscheduler(p);
5296 read_unlock(&tasklist_lock);
5301 * sys_sched_getscheduler - get the RT priority of a thread
5302 * @pid: the pid in question.
5303 * @param: structure containing the RT priority.
5305 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5307 struct sched_param lp;
5308 struct task_struct *p;
5311 if (!param || pid < 0)
5314 read_lock(&tasklist_lock);
5315 p = find_process_by_pid(pid);
5320 retval = security_task_getscheduler(p);
5324 lp.sched_priority = p->rt_priority;
5325 read_unlock(&tasklist_lock);
5328 * This one might sleep, we cannot do it with a spinlock held ...
5330 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5335 read_unlock(&tasklist_lock);
5339 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5341 cpumask_t cpus_allowed;
5342 cpumask_t new_mask = *in_mask;
5343 struct task_struct *p;
5347 read_lock(&tasklist_lock);
5349 p = find_process_by_pid(pid);
5351 read_unlock(&tasklist_lock);
5357 * It is not safe to call set_cpus_allowed with the
5358 * tasklist_lock held. We will bump the task_struct's
5359 * usage count and then drop tasklist_lock.
5362 read_unlock(&tasklist_lock);
5365 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5366 !capable(CAP_SYS_NICE))
5369 retval = security_task_setscheduler(p, 0, NULL);
5373 cpuset_cpus_allowed(p, &cpus_allowed);
5374 cpus_and(new_mask, new_mask, cpus_allowed);
5376 retval = set_cpus_allowed_ptr(p, &new_mask);
5379 cpuset_cpus_allowed(p, &cpus_allowed);
5380 if (!cpus_subset(new_mask, cpus_allowed)) {
5382 * We must have raced with a concurrent cpuset
5383 * update. Just reset the cpus_allowed to the
5384 * cpuset's cpus_allowed
5386 new_mask = cpus_allowed;
5396 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5397 cpumask_t *new_mask)
5399 if (len < sizeof(cpumask_t)) {
5400 memset(new_mask, 0, sizeof(cpumask_t));
5401 } else if (len > sizeof(cpumask_t)) {
5402 len = sizeof(cpumask_t);
5404 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5408 * sys_sched_setaffinity - set the cpu affinity of a process
5409 * @pid: pid of the process
5410 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5411 * @user_mask_ptr: user-space pointer to the new cpu mask
5413 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5414 unsigned long __user *user_mask_ptr)
5419 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5423 return sched_setaffinity(pid, &new_mask);
5426 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5428 struct task_struct *p;
5432 read_lock(&tasklist_lock);
5435 p = find_process_by_pid(pid);
5439 retval = security_task_getscheduler(p);
5443 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5446 read_unlock(&tasklist_lock);
5453 * sys_sched_getaffinity - get the cpu affinity of a process
5454 * @pid: pid of the process
5455 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5456 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5458 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5459 unsigned long __user *user_mask_ptr)
5464 if (len < sizeof(cpumask_t))
5467 ret = sched_getaffinity(pid, &mask);
5471 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5474 return sizeof(cpumask_t);
5478 * sys_sched_yield - yield the current processor to other threads.
5480 * This function yields the current CPU to other tasks. If there are no
5481 * other threads running on this CPU then this function will return.
5483 asmlinkage long sys_sched_yield(void)
5485 struct rq *rq = this_rq_lock();
5487 schedstat_inc(rq, yld_count);
5488 current->sched_class->yield_task(rq);
5491 * Since we are going to call schedule() anyway, there's
5492 * no need to preempt or enable interrupts:
5494 __release(rq->lock);
5495 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5496 _raw_spin_unlock(&rq->lock);
5497 preempt_enable_no_resched();
5504 static void __cond_resched(void)
5506 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5507 __might_sleep(__FILE__, __LINE__);
5510 * The BKS might be reacquired before we have dropped
5511 * PREEMPT_ACTIVE, which could trigger a second
5512 * cond_resched() call.
5515 add_preempt_count(PREEMPT_ACTIVE);
5517 sub_preempt_count(PREEMPT_ACTIVE);
5518 } while (need_resched());
5521 int __sched _cond_resched(void)
5523 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5524 system_state == SYSTEM_RUNNING) {
5530 EXPORT_SYMBOL(_cond_resched);
5533 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5534 * call schedule, and on return reacquire the lock.
5536 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5537 * operations here to prevent schedule() from being called twice (once via
5538 * spin_unlock(), once by hand).
5540 int cond_resched_lock(spinlock_t *lock)
5542 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5545 if (spin_needbreak(lock) || resched) {
5547 if (resched && need_resched())
5556 EXPORT_SYMBOL(cond_resched_lock);
5558 int __sched cond_resched_softirq(void)
5560 BUG_ON(!in_softirq());
5562 if (need_resched() && system_state == SYSTEM_RUNNING) {
5570 EXPORT_SYMBOL(cond_resched_softirq);
5573 * yield - yield the current processor to other threads.
5575 * This is a shortcut for kernel-space yielding - it marks the
5576 * thread runnable and calls sys_sched_yield().
5578 void __sched yield(void)
5580 set_current_state(TASK_RUNNING);
5583 EXPORT_SYMBOL(yield);
5586 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5587 * that process accounting knows that this is a task in IO wait state.
5589 * But don't do that if it is a deliberate, throttling IO wait (this task
5590 * has set its backing_dev_info: the queue against which it should throttle)
5592 void __sched io_schedule(void)
5594 struct rq *rq = &__raw_get_cpu_var(runqueues);
5596 delayacct_blkio_start();
5597 atomic_inc(&rq->nr_iowait);
5599 atomic_dec(&rq->nr_iowait);
5600 delayacct_blkio_end();
5602 EXPORT_SYMBOL(io_schedule);
5604 long __sched io_schedule_timeout(long timeout)
5606 struct rq *rq = &__raw_get_cpu_var(runqueues);
5609 delayacct_blkio_start();
5610 atomic_inc(&rq->nr_iowait);
5611 ret = schedule_timeout(timeout);
5612 atomic_dec(&rq->nr_iowait);
5613 delayacct_blkio_end();
5618 * sys_sched_get_priority_max - return maximum RT priority.
5619 * @policy: scheduling class.
5621 * this syscall returns the maximum rt_priority that can be used
5622 * by a given scheduling class.
5624 asmlinkage long sys_sched_get_priority_max(int policy)
5631 ret = MAX_USER_RT_PRIO-1;
5643 * sys_sched_get_priority_min - return minimum RT priority.
5644 * @policy: scheduling class.
5646 * this syscall returns the minimum rt_priority that can be used
5647 * by a given scheduling class.
5649 asmlinkage long sys_sched_get_priority_min(int policy)
5667 * sys_sched_rr_get_interval - return the default timeslice of a process.
5668 * @pid: pid of the process.
5669 * @interval: userspace pointer to the timeslice value.
5671 * this syscall writes the default timeslice value of a given process
5672 * into the user-space timespec buffer. A value of '0' means infinity.
5675 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5677 struct task_struct *p;
5678 unsigned int time_slice;
5686 read_lock(&tasklist_lock);
5687 p = find_process_by_pid(pid);
5691 retval = security_task_getscheduler(p);
5696 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5697 * tasks that are on an otherwise idle runqueue:
5700 if (p->policy == SCHED_RR) {
5701 time_slice = DEF_TIMESLICE;
5702 } else if (p->policy != SCHED_FIFO) {
5703 struct sched_entity *se = &p->se;
5704 unsigned long flags;
5707 rq = task_rq_lock(p, &flags);
5708 if (rq->cfs.load.weight)
5709 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5710 task_rq_unlock(rq, &flags);
5712 read_unlock(&tasklist_lock);
5713 jiffies_to_timespec(time_slice, &t);
5714 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5718 read_unlock(&tasklist_lock);
5722 static const char stat_nam[] = "RSDTtZX";
5724 void sched_show_task(struct task_struct *p)
5726 unsigned long free = 0;
5729 state = p->state ? __ffs(p->state) + 1 : 0;
5730 printk(KERN_INFO "%-13.13s %c", p->comm,
5731 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5732 #if BITS_PER_LONG == 32
5733 if (state == TASK_RUNNING)
5734 printk(KERN_CONT " running ");
5736 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5738 if (state == TASK_RUNNING)
5739 printk(KERN_CONT " running task ");
5741 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5743 #ifdef CONFIG_DEBUG_STACK_USAGE
5745 unsigned long *n = end_of_stack(p);
5748 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5751 printk(KERN_CONT "%5lu %5d %6d\n", free,
5752 task_pid_nr(p), task_pid_nr(p->real_parent));
5754 show_stack(p, NULL);
5757 void show_state_filter(unsigned long state_filter)
5759 struct task_struct *g, *p;
5761 #if BITS_PER_LONG == 32
5763 " task PC stack pid father\n");
5766 " task PC stack pid father\n");
5768 read_lock(&tasklist_lock);
5769 do_each_thread(g, p) {
5771 * reset the NMI-timeout, listing all files on a slow
5772 * console might take alot of time:
5774 touch_nmi_watchdog();
5775 if (!state_filter || (p->state & state_filter))
5777 } while_each_thread(g, p);
5779 touch_all_softlockup_watchdogs();
5781 #ifdef CONFIG_SCHED_DEBUG
5782 sysrq_sched_debug_show();
5784 read_unlock(&tasklist_lock);
5786 * Only show locks if all tasks are dumped:
5788 if (state_filter == -1)
5789 debug_show_all_locks();
5792 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5794 idle->sched_class = &idle_sched_class;
5798 * init_idle - set up an idle thread for a given CPU
5799 * @idle: task in question
5800 * @cpu: cpu the idle task belongs to
5802 * NOTE: this function does not set the idle thread's NEED_RESCHED
5803 * flag, to make booting more robust.
5805 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5807 struct rq *rq = cpu_rq(cpu);
5808 unsigned long flags;
5811 idle->se.exec_start = sched_clock();
5813 idle->prio = idle->normal_prio = MAX_PRIO;
5814 idle->cpus_allowed = cpumask_of_cpu(cpu);
5815 __set_task_cpu(idle, cpu);
5817 spin_lock_irqsave(&rq->lock, flags);
5818 rq->curr = rq->idle = idle;
5819 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5822 spin_unlock_irqrestore(&rq->lock, flags);
5824 /* Set the preempt count _outside_ the spinlocks! */
5825 #if defined(CONFIG_PREEMPT)
5826 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5828 task_thread_info(idle)->preempt_count = 0;
5831 * The idle tasks have their own, simple scheduling class:
5833 idle->sched_class = &idle_sched_class;
5837 * In a system that switches off the HZ timer nohz_cpu_mask
5838 * indicates which cpus entered this state. This is used
5839 * in the rcu update to wait only for active cpus. For system
5840 * which do not switch off the HZ timer nohz_cpu_mask should
5841 * always be CPU_MASK_NONE.
5843 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5846 * Increase the granularity value when there are more CPUs,
5847 * because with more CPUs the 'effective latency' as visible
5848 * to users decreases. But the relationship is not linear,
5849 * so pick a second-best guess by going with the log2 of the
5852 * This idea comes from the SD scheduler of Con Kolivas:
5854 static inline void sched_init_granularity(void)
5856 unsigned int factor = 1 + ilog2(num_online_cpus());
5857 const unsigned long limit = 200000000;
5859 sysctl_sched_min_granularity *= factor;
5860 if (sysctl_sched_min_granularity > limit)
5861 sysctl_sched_min_granularity = limit;
5863 sysctl_sched_latency *= factor;
5864 if (sysctl_sched_latency > limit)
5865 sysctl_sched_latency = limit;
5867 sysctl_sched_wakeup_granularity *= factor;
5872 * This is how migration works:
5874 * 1) we queue a struct migration_req structure in the source CPU's
5875 * runqueue and wake up that CPU's migration thread.
5876 * 2) we down() the locked semaphore => thread blocks.
5877 * 3) migration thread wakes up (implicitly it forces the migrated
5878 * thread off the CPU)
5879 * 4) it gets the migration request and checks whether the migrated
5880 * task is still in the wrong runqueue.
5881 * 5) if it's in the wrong runqueue then the migration thread removes
5882 * it and puts it into the right queue.
5883 * 6) migration thread up()s the semaphore.
5884 * 7) we wake up and the migration is done.
5888 * Change a given task's CPU affinity. Migrate the thread to a
5889 * proper CPU and schedule it away if the CPU it's executing on
5890 * is removed from the allowed bitmask.
5892 * NOTE: the caller must have a valid reference to the task, the
5893 * task must not exit() & deallocate itself prematurely. The
5894 * call is not atomic; no spinlocks may be held.
5896 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5898 struct migration_req req;
5899 unsigned long flags;
5903 rq = task_rq_lock(p, &flags);
5904 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5909 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
5910 !cpus_equal(p->cpus_allowed, *new_mask))) {
5915 if (p->sched_class->set_cpus_allowed)
5916 p->sched_class->set_cpus_allowed(p, new_mask);
5918 p->cpus_allowed = *new_mask;
5919 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5922 /* Can the task run on the task's current CPU? If so, we're done */
5923 if (cpu_isset(task_cpu(p), *new_mask))
5926 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5927 /* Need help from migration thread: drop lock and wait. */
5928 task_rq_unlock(rq, &flags);
5929 wake_up_process(rq->migration_thread);
5930 wait_for_completion(&req.done);
5931 tlb_migrate_finish(p->mm);
5935 task_rq_unlock(rq, &flags);
5939 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5942 * Move (not current) task off this cpu, onto dest cpu. We're doing
5943 * this because either it can't run here any more (set_cpus_allowed()
5944 * away from this CPU, or CPU going down), or because we're
5945 * attempting to rebalance this task on exec (sched_exec).
5947 * So we race with normal scheduler movements, but that's OK, as long
5948 * as the task is no longer on this CPU.
5950 * Returns non-zero if task was successfully migrated.
5952 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
5954 struct rq *rq_dest, *rq_src;
5957 if (unlikely(cpu_is_offline(dest_cpu)))
5960 rq_src = cpu_rq(src_cpu);
5961 rq_dest = cpu_rq(dest_cpu);
5963 double_rq_lock(rq_src, rq_dest);
5964 /* Already moved. */
5965 if (task_cpu(p) != src_cpu)
5967 /* Affinity changed (again). */
5968 if (!cpu_isset(dest_cpu, p->cpus_allowed))
5971 on_rq = p->se.on_rq;
5973 deactivate_task(rq_src, p, 0);
5975 set_task_cpu(p, dest_cpu);
5977 activate_task(rq_dest, p, 0);
5978 check_preempt_curr(rq_dest, p);
5982 double_rq_unlock(rq_src, rq_dest);
5987 * migration_thread - this is a highprio system thread that performs
5988 * thread migration by bumping thread off CPU then 'pushing' onto
5991 static int migration_thread(void *data)
5993 int cpu = (long)data;
5997 BUG_ON(rq->migration_thread != current);
5999 set_current_state(TASK_INTERRUPTIBLE);
6000 while (!kthread_should_stop()) {
6001 struct migration_req *req;
6002 struct list_head *head;
6004 spin_lock_irq(&rq->lock);
6006 if (cpu_is_offline(cpu)) {
6007 spin_unlock_irq(&rq->lock);
6011 if (rq->active_balance) {
6012 active_load_balance(rq, cpu);
6013 rq->active_balance = 0;
6016 head = &rq->migration_queue;
6018 if (list_empty(head)) {
6019 spin_unlock_irq(&rq->lock);
6021 set_current_state(TASK_INTERRUPTIBLE);
6024 req = list_entry(head->next, struct migration_req, list);
6025 list_del_init(head->next);
6027 spin_unlock(&rq->lock);
6028 __migrate_task(req->task, cpu, req->dest_cpu);
6031 complete(&req->done);
6033 __set_current_state(TASK_RUNNING);
6037 /* Wait for kthread_stop */
6038 set_current_state(TASK_INTERRUPTIBLE);
6039 while (!kthread_should_stop()) {
6041 set_current_state(TASK_INTERRUPTIBLE);
6043 __set_current_state(TASK_RUNNING);
6047 #ifdef CONFIG_HOTPLUG_CPU
6049 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6053 local_irq_disable();
6054 ret = __migrate_task(p, src_cpu, dest_cpu);
6060 * Figure out where task on dead CPU should go, use force if necessary.
6061 * NOTE: interrupts should be disabled by the caller
6063 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6065 unsigned long flags;
6072 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6073 cpus_and(mask, mask, p->cpus_allowed);
6074 dest_cpu = any_online_cpu(mask);
6076 /* On any allowed CPU? */
6077 if (dest_cpu >= nr_cpu_ids)
6078 dest_cpu = any_online_cpu(p->cpus_allowed);
6080 /* No more Mr. Nice Guy. */
6081 if (dest_cpu >= nr_cpu_ids) {
6082 cpumask_t cpus_allowed;
6084 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6086 * Try to stay on the same cpuset, where the
6087 * current cpuset may be a subset of all cpus.
6088 * The cpuset_cpus_allowed_locked() variant of
6089 * cpuset_cpus_allowed() will not block. It must be
6090 * called within calls to cpuset_lock/cpuset_unlock.
6092 rq = task_rq_lock(p, &flags);
6093 p->cpus_allowed = cpus_allowed;
6094 dest_cpu = any_online_cpu(p->cpus_allowed);
6095 task_rq_unlock(rq, &flags);
6098 * Don't tell them about moving exiting tasks or
6099 * kernel threads (both mm NULL), since they never
6102 if (p->mm && printk_ratelimit()) {
6103 printk(KERN_INFO "process %d (%s) no "
6104 "longer affine to cpu%d\n",
6105 task_pid_nr(p), p->comm, dead_cpu);
6108 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6112 * While a dead CPU has no uninterruptible tasks queued at this point,
6113 * it might still have a nonzero ->nr_uninterruptible counter, because
6114 * for performance reasons the counter is not stricly tracking tasks to
6115 * their home CPUs. So we just add the counter to another CPU's counter,
6116 * to keep the global sum constant after CPU-down:
6118 static void migrate_nr_uninterruptible(struct rq *rq_src)
6120 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6121 unsigned long flags;
6123 local_irq_save(flags);
6124 double_rq_lock(rq_src, rq_dest);
6125 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6126 rq_src->nr_uninterruptible = 0;
6127 double_rq_unlock(rq_src, rq_dest);
6128 local_irq_restore(flags);
6131 /* Run through task list and migrate tasks from the dead cpu. */
6132 static void migrate_live_tasks(int src_cpu)
6134 struct task_struct *p, *t;
6136 read_lock(&tasklist_lock);
6138 do_each_thread(t, p) {
6142 if (task_cpu(p) == src_cpu)
6143 move_task_off_dead_cpu(src_cpu, p);
6144 } while_each_thread(t, p);
6146 read_unlock(&tasklist_lock);
6150 * Schedules idle task to be the next runnable task on current CPU.
6151 * It does so by boosting its priority to highest possible.
6152 * Used by CPU offline code.
6154 void sched_idle_next(void)
6156 int this_cpu = smp_processor_id();
6157 struct rq *rq = cpu_rq(this_cpu);
6158 struct task_struct *p = rq->idle;
6159 unsigned long flags;
6161 /* cpu has to be offline */
6162 BUG_ON(cpu_online(this_cpu));
6165 * Strictly not necessary since rest of the CPUs are stopped by now
6166 * and interrupts disabled on the current cpu.
6168 spin_lock_irqsave(&rq->lock, flags);
6170 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6172 update_rq_clock(rq);
6173 activate_task(rq, p, 0);
6175 spin_unlock_irqrestore(&rq->lock, flags);
6179 * Ensures that the idle task is using init_mm right before its cpu goes
6182 void idle_task_exit(void)
6184 struct mm_struct *mm = current->active_mm;
6186 BUG_ON(cpu_online(smp_processor_id()));
6189 switch_mm(mm, &init_mm, current);
6193 /* called under rq->lock with disabled interrupts */
6194 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6196 struct rq *rq = cpu_rq(dead_cpu);
6198 /* Must be exiting, otherwise would be on tasklist. */
6199 BUG_ON(!p->exit_state);
6201 /* Cannot have done final schedule yet: would have vanished. */
6202 BUG_ON(p->state == TASK_DEAD);
6207 * Drop lock around migration; if someone else moves it,
6208 * that's OK. No task can be added to this CPU, so iteration is
6211 spin_unlock_irq(&rq->lock);
6212 move_task_off_dead_cpu(dead_cpu, p);
6213 spin_lock_irq(&rq->lock);
6218 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6219 static void migrate_dead_tasks(unsigned int dead_cpu)
6221 struct rq *rq = cpu_rq(dead_cpu);
6222 struct task_struct *next;
6225 if (!rq->nr_running)
6227 update_rq_clock(rq);
6228 next = pick_next_task(rq, rq->curr);
6231 migrate_dead(dead_cpu, next);
6235 #endif /* CONFIG_HOTPLUG_CPU */
6237 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6239 static struct ctl_table sd_ctl_dir[] = {
6241 .procname = "sched_domain",
6247 static struct ctl_table sd_ctl_root[] = {
6249 .ctl_name = CTL_KERN,
6250 .procname = "kernel",
6252 .child = sd_ctl_dir,
6257 static struct ctl_table *sd_alloc_ctl_entry(int n)
6259 struct ctl_table *entry =
6260 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6265 static void sd_free_ctl_entry(struct ctl_table **tablep)
6267 struct ctl_table *entry;
6270 * In the intermediate directories, both the child directory and
6271 * procname are dynamically allocated and could fail but the mode
6272 * will always be set. In the lowest directory the names are
6273 * static strings and all have proc handlers.
6275 for (entry = *tablep; entry->mode; entry++) {
6277 sd_free_ctl_entry(&entry->child);
6278 if (entry->proc_handler == NULL)
6279 kfree(entry->procname);
6287 set_table_entry(struct ctl_table *entry,
6288 const char *procname, void *data, int maxlen,
6289 mode_t mode, proc_handler *proc_handler)
6291 entry->procname = procname;
6293 entry->maxlen = maxlen;
6295 entry->proc_handler = proc_handler;
6298 static struct ctl_table *
6299 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6301 struct ctl_table *table = sd_alloc_ctl_entry(12);
6306 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6307 sizeof(long), 0644, proc_doulongvec_minmax);
6308 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6309 sizeof(long), 0644, proc_doulongvec_minmax);
6310 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6311 sizeof(int), 0644, proc_dointvec_minmax);
6312 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6313 sizeof(int), 0644, proc_dointvec_minmax);
6314 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6315 sizeof(int), 0644, proc_dointvec_minmax);
6316 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6317 sizeof(int), 0644, proc_dointvec_minmax);
6318 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6319 sizeof(int), 0644, proc_dointvec_minmax);
6320 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6321 sizeof(int), 0644, proc_dointvec_minmax);
6322 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6323 sizeof(int), 0644, proc_dointvec_minmax);
6324 set_table_entry(&table[9], "cache_nice_tries",
6325 &sd->cache_nice_tries,
6326 sizeof(int), 0644, proc_dointvec_minmax);
6327 set_table_entry(&table[10], "flags", &sd->flags,
6328 sizeof(int), 0644, proc_dointvec_minmax);
6329 /* &table[11] is terminator */
6334 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6336 struct ctl_table *entry, *table;
6337 struct sched_domain *sd;
6338 int domain_num = 0, i;
6341 for_each_domain(cpu, sd)
6343 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6348 for_each_domain(cpu, sd) {
6349 snprintf(buf, 32, "domain%d", i);
6350 entry->procname = kstrdup(buf, GFP_KERNEL);
6352 entry->child = sd_alloc_ctl_domain_table(sd);
6359 static struct ctl_table_header *sd_sysctl_header;
6360 static void register_sched_domain_sysctl(void)
6362 int i, cpu_num = num_online_cpus();
6363 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6366 WARN_ON(sd_ctl_dir[0].child);
6367 sd_ctl_dir[0].child = entry;
6372 for_each_online_cpu(i) {
6373 snprintf(buf, 32, "cpu%d", i);
6374 entry->procname = kstrdup(buf, GFP_KERNEL);
6376 entry->child = sd_alloc_ctl_cpu_table(i);
6380 WARN_ON(sd_sysctl_header);
6381 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6384 /* may be called multiple times per register */
6385 static void unregister_sched_domain_sysctl(void)
6387 if (sd_sysctl_header)
6388 unregister_sysctl_table(sd_sysctl_header);
6389 sd_sysctl_header = NULL;
6390 if (sd_ctl_dir[0].child)
6391 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6394 static void register_sched_domain_sysctl(void)
6397 static void unregister_sched_domain_sysctl(void)
6402 static void set_rq_online(struct rq *rq)
6405 const struct sched_class *class;
6407 cpu_set(rq->cpu, rq->rd->online);
6410 for_each_class(class) {
6411 if (class->rq_online)
6412 class->rq_online(rq);
6417 static void set_rq_offline(struct rq *rq)
6420 const struct sched_class *class;
6422 for_each_class(class) {
6423 if (class->rq_offline)
6424 class->rq_offline(rq);
6427 cpu_clear(rq->cpu, rq->rd->online);
6433 * migration_call - callback that gets triggered when a CPU is added.
6434 * Here we can start up the necessary migration thread for the new CPU.
6436 static int __cpuinit
6437 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6439 struct task_struct *p;
6440 int cpu = (long)hcpu;
6441 unsigned long flags;
6446 case CPU_UP_PREPARE:
6447 case CPU_UP_PREPARE_FROZEN:
6448 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6451 kthread_bind(p, cpu);
6452 /* Must be high prio: stop_machine expects to yield to it. */
6453 rq = task_rq_lock(p, &flags);
6454 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6455 task_rq_unlock(rq, &flags);
6456 cpu_rq(cpu)->migration_thread = p;
6460 case CPU_ONLINE_FROZEN:
6461 /* Strictly unnecessary, as first user will wake it. */
6462 wake_up_process(cpu_rq(cpu)->migration_thread);
6464 /* Update our root-domain */
6466 spin_lock_irqsave(&rq->lock, flags);
6468 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6472 spin_unlock_irqrestore(&rq->lock, flags);
6475 #ifdef CONFIG_HOTPLUG_CPU
6476 case CPU_UP_CANCELED:
6477 case CPU_UP_CANCELED_FROZEN:
6478 if (!cpu_rq(cpu)->migration_thread)
6480 /* Unbind it from offline cpu so it can run. Fall thru. */
6481 kthread_bind(cpu_rq(cpu)->migration_thread,
6482 any_online_cpu(cpu_online_map));
6483 kthread_stop(cpu_rq(cpu)->migration_thread);
6484 cpu_rq(cpu)->migration_thread = NULL;
6488 case CPU_DEAD_FROZEN:
6489 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6490 migrate_live_tasks(cpu);
6492 kthread_stop(rq->migration_thread);
6493 rq->migration_thread = NULL;
6494 /* Idle task back to normal (off runqueue, low prio) */
6495 spin_lock_irq(&rq->lock);
6496 update_rq_clock(rq);
6497 deactivate_task(rq, rq->idle, 0);
6498 rq->idle->static_prio = MAX_PRIO;
6499 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6500 rq->idle->sched_class = &idle_sched_class;
6501 migrate_dead_tasks(cpu);
6502 spin_unlock_irq(&rq->lock);
6504 migrate_nr_uninterruptible(rq);
6505 BUG_ON(rq->nr_running != 0);
6508 * No need to migrate the tasks: it was best-effort if
6509 * they didn't take sched_hotcpu_mutex. Just wake up
6512 spin_lock_irq(&rq->lock);
6513 while (!list_empty(&rq->migration_queue)) {
6514 struct migration_req *req;
6516 req = list_entry(rq->migration_queue.next,
6517 struct migration_req, list);
6518 list_del_init(&req->list);
6519 complete(&req->done);
6521 spin_unlock_irq(&rq->lock);
6525 case CPU_DYING_FROZEN:
6526 /* Update our root-domain */
6528 spin_lock_irqsave(&rq->lock, flags);
6530 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6533 spin_unlock_irqrestore(&rq->lock, flags);
6540 /* Register at highest priority so that task migration (migrate_all_tasks)
6541 * happens before everything else.
6543 static struct notifier_block __cpuinitdata migration_notifier = {
6544 .notifier_call = migration_call,
6548 void __init migration_init(void)
6550 void *cpu = (void *)(long)smp_processor_id();
6553 /* Start one for the boot CPU: */
6554 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6555 BUG_ON(err == NOTIFY_BAD);
6556 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6557 register_cpu_notifier(&migration_notifier);
6563 #ifdef CONFIG_SCHED_DEBUG
6565 static inline const char *sd_level_to_string(enum sched_domain_level lvl)
6578 case SD_LV_ALLNODES:
6587 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6588 cpumask_t *groupmask)
6590 struct sched_group *group = sd->groups;
6593 cpulist_scnprintf(str, sizeof(str), sd->span);
6594 cpus_clear(*groupmask);
6596 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6598 if (!(sd->flags & SD_LOAD_BALANCE)) {
6599 printk("does not load-balance\n");
6601 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6606 printk(KERN_CONT "span %s level %s\n",
6607 str, sd_level_to_string(sd->level));
6609 if (!cpu_isset(cpu, sd->span)) {
6610 printk(KERN_ERR "ERROR: domain->span does not contain "
6613 if (!cpu_isset(cpu, group->cpumask)) {
6614 printk(KERN_ERR "ERROR: domain->groups does not contain"
6618 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6622 printk(KERN_ERR "ERROR: group is NULL\n");
6626 if (!group->__cpu_power) {
6627 printk(KERN_CONT "\n");
6628 printk(KERN_ERR "ERROR: domain->cpu_power not "
6633 if (!cpus_weight(group->cpumask)) {
6634 printk(KERN_CONT "\n");
6635 printk(KERN_ERR "ERROR: empty group\n");
6639 if (cpus_intersects(*groupmask, group->cpumask)) {
6640 printk(KERN_CONT "\n");
6641 printk(KERN_ERR "ERROR: repeated CPUs\n");
6645 cpus_or(*groupmask, *groupmask, group->cpumask);
6647 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6648 printk(KERN_CONT " %s", str);
6650 group = group->next;
6651 } while (group != sd->groups);
6652 printk(KERN_CONT "\n");
6654 if (!cpus_equal(sd->span, *groupmask))
6655 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6657 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6658 printk(KERN_ERR "ERROR: parent span is not a superset "
6659 "of domain->span\n");
6663 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6665 cpumask_t *groupmask;
6669 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6673 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6675 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6677 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6682 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6691 #else /* !CONFIG_SCHED_DEBUG */
6692 # define sched_domain_debug(sd, cpu) do { } while (0)
6693 #endif /* CONFIG_SCHED_DEBUG */
6695 static int sd_degenerate(struct sched_domain *sd)
6697 if (cpus_weight(sd->span) == 1)
6700 /* Following flags need at least 2 groups */
6701 if (sd->flags & (SD_LOAD_BALANCE |
6702 SD_BALANCE_NEWIDLE |
6706 SD_SHARE_PKG_RESOURCES)) {
6707 if (sd->groups != sd->groups->next)
6711 /* Following flags don't use groups */
6712 if (sd->flags & (SD_WAKE_IDLE |
6721 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6723 unsigned long cflags = sd->flags, pflags = parent->flags;
6725 if (sd_degenerate(parent))
6728 if (!cpus_equal(sd->span, parent->span))
6731 /* Does parent contain flags not in child? */
6732 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6733 if (cflags & SD_WAKE_AFFINE)
6734 pflags &= ~SD_WAKE_BALANCE;
6735 /* Flags needing groups don't count if only 1 group in parent */
6736 if (parent->groups == parent->groups->next) {
6737 pflags &= ~(SD_LOAD_BALANCE |
6738 SD_BALANCE_NEWIDLE |
6742 SD_SHARE_PKG_RESOURCES);
6744 if (~cflags & pflags)
6750 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6752 unsigned long flags;
6754 spin_lock_irqsave(&rq->lock, flags);
6757 struct root_domain *old_rd = rq->rd;
6759 if (cpu_isset(rq->cpu, old_rd->online))
6762 cpu_clear(rq->cpu, old_rd->span);
6764 if (atomic_dec_and_test(&old_rd->refcount))
6768 atomic_inc(&rd->refcount);
6771 cpu_set(rq->cpu, rd->span);
6772 if (cpu_isset(rq->cpu, cpu_online_map))
6775 spin_unlock_irqrestore(&rq->lock, flags);
6778 static void init_rootdomain(struct root_domain *rd)
6780 memset(rd, 0, sizeof(*rd));
6782 cpus_clear(rd->span);
6783 cpus_clear(rd->online);
6785 cpupri_init(&rd->cpupri);
6788 static void init_defrootdomain(void)
6790 init_rootdomain(&def_root_domain);
6791 atomic_set(&def_root_domain.refcount, 1);
6794 static struct root_domain *alloc_rootdomain(void)
6796 struct root_domain *rd;
6798 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6802 init_rootdomain(rd);
6808 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6809 * hold the hotplug lock.
6812 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6814 struct rq *rq = cpu_rq(cpu);
6815 struct sched_domain *tmp;
6817 /* Remove the sched domains which do not contribute to scheduling. */
6818 for (tmp = sd; tmp; tmp = tmp->parent) {
6819 struct sched_domain *parent = tmp->parent;
6822 if (sd_parent_degenerate(tmp, parent)) {
6823 tmp->parent = parent->parent;
6825 parent->parent->child = tmp;
6829 if (sd && sd_degenerate(sd)) {
6835 sched_domain_debug(sd, cpu);
6837 rq_attach_root(rq, rd);
6838 rcu_assign_pointer(rq->sd, sd);
6841 /* cpus with isolated domains */
6842 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6844 /* Setup the mask of cpus configured for isolated domains */
6845 static int __init isolated_cpu_setup(char *str)
6847 int ints[NR_CPUS], i;
6849 str = get_options(str, ARRAY_SIZE(ints), ints);
6850 cpus_clear(cpu_isolated_map);
6851 for (i = 1; i <= ints[0]; i++)
6852 if (ints[i] < NR_CPUS)
6853 cpu_set(ints[i], cpu_isolated_map);
6857 __setup("isolcpus=", isolated_cpu_setup);
6860 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6861 * to a function which identifies what group(along with sched group) a CPU
6862 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6863 * (due to the fact that we keep track of groups covered with a cpumask_t).
6865 * init_sched_build_groups will build a circular linked list of the groups
6866 * covered by the given span, and will set each group's ->cpumask correctly,
6867 * and ->cpu_power to 0.
6870 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6871 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6872 struct sched_group **sg,
6873 cpumask_t *tmpmask),
6874 cpumask_t *covered, cpumask_t *tmpmask)
6876 struct sched_group *first = NULL, *last = NULL;
6879 cpus_clear(*covered);
6881 for_each_cpu_mask(i, *span) {
6882 struct sched_group *sg;
6883 int group = group_fn(i, cpu_map, &sg, tmpmask);
6886 if (cpu_isset(i, *covered))
6889 cpus_clear(sg->cpumask);
6890 sg->__cpu_power = 0;
6892 for_each_cpu_mask(j, *span) {
6893 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6896 cpu_set(j, *covered);
6897 cpu_set(j, sg->cpumask);
6908 #define SD_NODES_PER_DOMAIN 16
6913 * find_next_best_node - find the next node to include in a sched_domain
6914 * @node: node whose sched_domain we're building
6915 * @used_nodes: nodes already in the sched_domain
6917 * Find the next node to include in a given scheduling domain. Simply
6918 * finds the closest node not already in the @used_nodes map.
6920 * Should use nodemask_t.
6922 static int find_next_best_node(int node, nodemask_t *used_nodes)
6924 int i, n, val, min_val, best_node = 0;
6928 for (i = 0; i < MAX_NUMNODES; i++) {
6929 /* Start at @node */
6930 n = (node + i) % MAX_NUMNODES;
6932 if (!nr_cpus_node(n))
6935 /* Skip already used nodes */
6936 if (node_isset(n, *used_nodes))
6939 /* Simple min distance search */
6940 val = node_distance(node, n);
6942 if (val < min_val) {
6948 node_set(best_node, *used_nodes);
6953 * sched_domain_node_span - get a cpumask for a node's sched_domain
6954 * @node: node whose cpumask we're constructing
6955 * @span: resulting cpumask
6957 * Given a node, construct a good cpumask for its sched_domain to span. It
6958 * should be one that prevents unnecessary balancing, but also spreads tasks
6961 static void sched_domain_node_span(int node, cpumask_t *span)
6963 nodemask_t used_nodes;
6964 node_to_cpumask_ptr(nodemask, node);
6968 nodes_clear(used_nodes);
6970 cpus_or(*span, *span, *nodemask);
6971 node_set(node, used_nodes);
6973 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6974 int next_node = find_next_best_node(node, &used_nodes);
6976 node_to_cpumask_ptr_next(nodemask, next_node);
6977 cpus_or(*span, *span, *nodemask);
6980 #endif /* CONFIG_NUMA */
6982 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6985 * SMT sched-domains:
6987 #ifdef CONFIG_SCHED_SMT
6988 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6989 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
6992 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
6996 *sg = &per_cpu(sched_group_cpus, cpu);
6999 #endif /* CONFIG_SCHED_SMT */
7002 * multi-core sched-domains:
7004 #ifdef CONFIG_SCHED_MC
7005 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7006 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7007 #endif /* CONFIG_SCHED_MC */
7009 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7011 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7016 *mask = per_cpu(cpu_sibling_map, cpu);
7017 cpus_and(*mask, *mask, *cpu_map);
7018 group = first_cpu(*mask);
7020 *sg = &per_cpu(sched_group_core, group);
7023 #elif defined(CONFIG_SCHED_MC)
7025 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7029 *sg = &per_cpu(sched_group_core, cpu);
7034 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7035 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7038 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7042 #ifdef CONFIG_SCHED_MC
7043 *mask = cpu_coregroup_map(cpu);
7044 cpus_and(*mask, *mask, *cpu_map);
7045 group = first_cpu(*mask);
7046 #elif defined(CONFIG_SCHED_SMT)
7047 *mask = per_cpu(cpu_sibling_map, cpu);
7048 cpus_and(*mask, *mask, *cpu_map);
7049 group = first_cpu(*mask);
7054 *sg = &per_cpu(sched_group_phys, group);
7060 * The init_sched_build_groups can't handle what we want to do with node
7061 * groups, so roll our own. Now each node has its own list of groups which
7062 * gets dynamically allocated.
7064 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7065 static struct sched_group ***sched_group_nodes_bycpu;
7067 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7068 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7070 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7071 struct sched_group **sg, cpumask_t *nodemask)
7075 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7076 cpus_and(*nodemask, *nodemask, *cpu_map);
7077 group = first_cpu(*nodemask);
7080 *sg = &per_cpu(sched_group_allnodes, group);
7084 static void init_numa_sched_groups_power(struct sched_group *group_head)
7086 struct sched_group *sg = group_head;
7092 for_each_cpu_mask(j, sg->cpumask) {
7093 struct sched_domain *sd;
7095 sd = &per_cpu(phys_domains, j);
7096 if (j != first_cpu(sd->groups->cpumask)) {
7098 * Only add "power" once for each
7104 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7107 } while (sg != group_head);
7109 #endif /* CONFIG_NUMA */
7112 /* Free memory allocated for various sched_group structures */
7113 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7117 for_each_cpu_mask(cpu, *cpu_map) {
7118 struct sched_group **sched_group_nodes
7119 = sched_group_nodes_bycpu[cpu];
7121 if (!sched_group_nodes)
7124 for (i = 0; i < MAX_NUMNODES; i++) {
7125 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7127 *nodemask = node_to_cpumask(i);
7128 cpus_and(*nodemask, *nodemask, *cpu_map);
7129 if (cpus_empty(*nodemask))
7139 if (oldsg != sched_group_nodes[i])
7142 kfree(sched_group_nodes);
7143 sched_group_nodes_bycpu[cpu] = NULL;
7146 #else /* !CONFIG_NUMA */
7147 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7150 #endif /* CONFIG_NUMA */
7153 * Initialize sched groups cpu_power.
7155 * cpu_power indicates the capacity of sched group, which is used while
7156 * distributing the load between different sched groups in a sched domain.
7157 * Typically cpu_power for all the groups in a sched domain will be same unless
7158 * there are asymmetries in the topology. If there are asymmetries, group
7159 * having more cpu_power will pickup more load compared to the group having
7162 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7163 * the maximum number of tasks a group can handle in the presence of other idle
7164 * or lightly loaded groups in the same sched domain.
7166 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7168 struct sched_domain *child;
7169 struct sched_group *group;
7171 WARN_ON(!sd || !sd->groups);
7173 if (cpu != first_cpu(sd->groups->cpumask))
7178 sd->groups->__cpu_power = 0;
7181 * For perf policy, if the groups in child domain share resources
7182 * (for example cores sharing some portions of the cache hierarchy
7183 * or SMT), then set this domain groups cpu_power such that each group
7184 * can handle only one task, when there are other idle groups in the
7185 * same sched domain.
7187 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7189 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7190 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7195 * add cpu_power of each child group to this groups cpu_power
7197 group = child->groups;
7199 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7200 group = group->next;
7201 } while (group != child->groups);
7205 * Initializers for schedule domains
7206 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7209 #define SD_INIT(sd, type) sd_init_##type(sd)
7210 #define SD_INIT_FUNC(type) \
7211 static noinline void sd_init_##type(struct sched_domain *sd) \
7213 memset(sd, 0, sizeof(*sd)); \
7214 *sd = SD_##type##_INIT; \
7215 sd->level = SD_LV_##type; \
7220 SD_INIT_FUNC(ALLNODES)
7223 #ifdef CONFIG_SCHED_SMT
7224 SD_INIT_FUNC(SIBLING)
7226 #ifdef CONFIG_SCHED_MC
7231 * To minimize stack usage kmalloc room for cpumasks and share the
7232 * space as the usage in build_sched_domains() dictates. Used only
7233 * if the amount of space is significant.
7236 cpumask_t tmpmask; /* make this one first */
7239 cpumask_t this_sibling_map;
7240 cpumask_t this_core_map;
7242 cpumask_t send_covered;
7245 cpumask_t domainspan;
7247 cpumask_t notcovered;
7252 #define SCHED_CPUMASK_ALLOC 1
7253 #define SCHED_CPUMASK_FREE(v) kfree(v)
7254 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7256 #define SCHED_CPUMASK_ALLOC 0
7257 #define SCHED_CPUMASK_FREE(v)
7258 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7261 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7262 ((unsigned long)(a) + offsetof(struct allmasks, v))
7264 static int default_relax_domain_level = -1;
7266 static int __init setup_relax_domain_level(char *str)
7270 val = simple_strtoul(str, NULL, 0);
7271 if (val < SD_LV_MAX)
7272 default_relax_domain_level = val;
7276 __setup("relax_domain_level=", setup_relax_domain_level);
7278 static void set_domain_attribute(struct sched_domain *sd,
7279 struct sched_domain_attr *attr)
7283 if (!attr || attr->relax_domain_level < 0) {
7284 if (default_relax_domain_level < 0)
7287 request = default_relax_domain_level;
7289 request = attr->relax_domain_level;
7290 if (request < sd->level) {
7291 /* turn off idle balance on this domain */
7292 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7294 /* turn on idle balance on this domain */
7295 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7300 * Build sched domains for a given set of cpus and attach the sched domains
7301 * to the individual cpus
7303 static int __build_sched_domains(const cpumask_t *cpu_map,
7304 struct sched_domain_attr *attr)
7307 struct root_domain *rd;
7308 SCHED_CPUMASK_DECLARE(allmasks);
7311 struct sched_group **sched_group_nodes = NULL;
7312 int sd_allnodes = 0;
7315 * Allocate the per-node list of sched groups
7317 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7319 if (!sched_group_nodes) {
7320 printk(KERN_WARNING "Can not alloc sched group node list\n");
7325 rd = alloc_rootdomain();
7327 printk(KERN_WARNING "Cannot alloc root domain\n");
7329 kfree(sched_group_nodes);
7334 #if SCHED_CPUMASK_ALLOC
7335 /* get space for all scratch cpumask variables */
7336 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7338 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7341 kfree(sched_group_nodes);
7346 tmpmask = (cpumask_t *)allmasks;
7350 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7354 * Set up domains for cpus specified by the cpu_map.
7356 for_each_cpu_mask(i, *cpu_map) {
7357 struct sched_domain *sd = NULL, *p;
7358 SCHED_CPUMASK_VAR(nodemask, allmasks);
7360 *nodemask = node_to_cpumask(cpu_to_node(i));
7361 cpus_and(*nodemask, *nodemask, *cpu_map);
7364 if (cpus_weight(*cpu_map) >
7365 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7366 sd = &per_cpu(allnodes_domains, i);
7367 SD_INIT(sd, ALLNODES);
7368 set_domain_attribute(sd, attr);
7369 sd->span = *cpu_map;
7370 sd->first_cpu = first_cpu(sd->span);
7371 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7377 sd = &per_cpu(node_domains, i);
7379 set_domain_attribute(sd, attr);
7380 sched_domain_node_span(cpu_to_node(i), &sd->span);
7381 sd->first_cpu = first_cpu(sd->span);
7385 cpus_and(sd->span, sd->span, *cpu_map);
7389 sd = &per_cpu(phys_domains, i);
7391 set_domain_attribute(sd, attr);
7392 sd->span = *nodemask;
7393 sd->first_cpu = first_cpu(sd->span);
7397 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7399 #ifdef CONFIG_SCHED_MC
7401 sd = &per_cpu(core_domains, i);
7403 set_domain_attribute(sd, attr);
7404 sd->span = cpu_coregroup_map(i);
7405 sd->first_cpu = first_cpu(sd->span);
7406 cpus_and(sd->span, sd->span, *cpu_map);
7409 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7412 #ifdef CONFIG_SCHED_SMT
7414 sd = &per_cpu(cpu_domains, i);
7415 SD_INIT(sd, SIBLING);
7416 set_domain_attribute(sd, attr);
7417 sd->span = per_cpu(cpu_sibling_map, i);
7418 sd->first_cpu = first_cpu(sd->span);
7419 cpus_and(sd->span, sd->span, *cpu_map);
7422 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7426 #ifdef CONFIG_SCHED_SMT
7427 /* Set up CPU (sibling) groups */
7428 for_each_cpu_mask(i, *cpu_map) {
7429 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7430 SCHED_CPUMASK_VAR(send_covered, allmasks);
7432 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7433 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7434 if (i != first_cpu(*this_sibling_map))
7437 init_sched_build_groups(this_sibling_map, cpu_map,
7439 send_covered, tmpmask);
7443 #ifdef CONFIG_SCHED_MC
7444 /* Set up multi-core groups */
7445 for_each_cpu_mask(i, *cpu_map) {
7446 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7447 SCHED_CPUMASK_VAR(send_covered, allmasks);
7449 *this_core_map = cpu_coregroup_map(i);
7450 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7451 if (i != first_cpu(*this_core_map))
7454 init_sched_build_groups(this_core_map, cpu_map,
7456 send_covered, tmpmask);
7460 /* Set up physical groups */
7461 for (i = 0; i < MAX_NUMNODES; i++) {
7462 SCHED_CPUMASK_VAR(nodemask, allmasks);
7463 SCHED_CPUMASK_VAR(send_covered, allmasks);
7465 *nodemask = node_to_cpumask(i);
7466 cpus_and(*nodemask, *nodemask, *cpu_map);
7467 if (cpus_empty(*nodemask))
7470 init_sched_build_groups(nodemask, cpu_map,
7472 send_covered, tmpmask);
7476 /* Set up node groups */
7478 SCHED_CPUMASK_VAR(send_covered, allmasks);
7480 init_sched_build_groups(cpu_map, cpu_map,
7481 &cpu_to_allnodes_group,
7482 send_covered, tmpmask);
7485 for (i = 0; i < MAX_NUMNODES; i++) {
7486 /* Set up node groups */
7487 struct sched_group *sg, *prev;
7488 SCHED_CPUMASK_VAR(nodemask, allmasks);
7489 SCHED_CPUMASK_VAR(domainspan, allmasks);
7490 SCHED_CPUMASK_VAR(covered, allmasks);
7493 *nodemask = node_to_cpumask(i);
7494 cpus_clear(*covered);
7496 cpus_and(*nodemask, *nodemask, *cpu_map);
7497 if (cpus_empty(*nodemask)) {
7498 sched_group_nodes[i] = NULL;
7502 sched_domain_node_span(i, domainspan);
7503 cpus_and(*domainspan, *domainspan, *cpu_map);
7505 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7507 printk(KERN_WARNING "Can not alloc domain group for "
7511 sched_group_nodes[i] = sg;
7512 for_each_cpu_mask(j, *nodemask) {
7513 struct sched_domain *sd;
7515 sd = &per_cpu(node_domains, j);
7518 sg->__cpu_power = 0;
7519 sg->cpumask = *nodemask;
7521 cpus_or(*covered, *covered, *nodemask);
7524 for (j = 0; j < MAX_NUMNODES; j++) {
7525 SCHED_CPUMASK_VAR(notcovered, allmasks);
7526 int n = (i + j) % MAX_NUMNODES;
7527 node_to_cpumask_ptr(pnodemask, n);
7529 cpus_complement(*notcovered, *covered);
7530 cpus_and(*tmpmask, *notcovered, *cpu_map);
7531 cpus_and(*tmpmask, *tmpmask, *domainspan);
7532 if (cpus_empty(*tmpmask))
7535 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7536 if (cpus_empty(*tmpmask))
7539 sg = kmalloc_node(sizeof(struct sched_group),
7543 "Can not alloc domain group for node %d\n", j);
7546 sg->__cpu_power = 0;
7547 sg->cpumask = *tmpmask;
7548 sg->next = prev->next;
7549 cpus_or(*covered, *covered, *tmpmask);
7556 /* Calculate CPU power for physical packages and nodes */
7557 #ifdef CONFIG_SCHED_SMT
7558 for_each_cpu_mask(i, *cpu_map) {
7559 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7561 init_sched_groups_power(i, sd);
7564 #ifdef CONFIG_SCHED_MC
7565 for_each_cpu_mask(i, *cpu_map) {
7566 struct sched_domain *sd = &per_cpu(core_domains, i);
7568 init_sched_groups_power(i, sd);
7572 for_each_cpu_mask(i, *cpu_map) {
7573 struct sched_domain *sd = &per_cpu(phys_domains, i);
7575 init_sched_groups_power(i, sd);
7579 for (i = 0; i < MAX_NUMNODES; i++)
7580 init_numa_sched_groups_power(sched_group_nodes[i]);
7583 struct sched_group *sg;
7585 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7587 init_numa_sched_groups_power(sg);
7591 /* Attach the domains */
7592 for_each_cpu_mask(i, *cpu_map) {
7593 struct sched_domain *sd;
7594 #ifdef CONFIG_SCHED_SMT
7595 sd = &per_cpu(cpu_domains, i);
7596 #elif defined(CONFIG_SCHED_MC)
7597 sd = &per_cpu(core_domains, i);
7599 sd = &per_cpu(phys_domains, i);
7601 cpu_attach_domain(sd, rd, i);
7604 SCHED_CPUMASK_FREE((void *)allmasks);
7609 free_sched_groups(cpu_map, tmpmask);
7610 SCHED_CPUMASK_FREE((void *)allmasks);
7615 static int build_sched_domains(const cpumask_t *cpu_map)
7617 return __build_sched_domains(cpu_map, NULL);
7620 static cpumask_t *doms_cur; /* current sched domains */
7621 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7622 static struct sched_domain_attr *dattr_cur;
7623 /* attribues of custom domains in 'doms_cur' */
7626 * Special case: If a kmalloc of a doms_cur partition (array of
7627 * cpumask_t) fails, then fallback to a single sched domain,
7628 * as determined by the single cpumask_t fallback_doms.
7630 static cpumask_t fallback_doms;
7632 void __attribute__((weak)) arch_update_cpu_topology(void)
7637 * Free current domain masks.
7638 * Called after all cpus are attached to NULL domain.
7640 static void free_sched_domains(void)
7643 if (doms_cur != &fallback_doms)
7645 doms_cur = &fallback_doms;
7649 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7650 * For now this just excludes isolated cpus, but could be used to
7651 * exclude other special cases in the future.
7653 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7657 arch_update_cpu_topology();
7659 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7661 doms_cur = &fallback_doms;
7662 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7664 err = build_sched_domains(doms_cur);
7665 register_sched_domain_sysctl();
7670 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7673 free_sched_groups(cpu_map, tmpmask);
7677 * Detach sched domains from a group of cpus specified in cpu_map
7678 * These cpus will now be attached to the NULL domain
7680 static void detach_destroy_domains(const cpumask_t *cpu_map)
7685 unregister_sched_domain_sysctl();
7687 for_each_cpu_mask(i, *cpu_map)
7688 cpu_attach_domain(NULL, &def_root_domain, i);
7689 synchronize_sched();
7690 arch_destroy_sched_domains(cpu_map, &tmpmask);
7693 /* handle null as "default" */
7694 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7695 struct sched_domain_attr *new, int idx_new)
7697 struct sched_domain_attr tmp;
7704 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7705 new ? (new + idx_new) : &tmp,
7706 sizeof(struct sched_domain_attr));
7710 * Partition sched domains as specified by the 'ndoms_new'
7711 * cpumasks in the array doms_new[] of cpumasks. This compares
7712 * doms_new[] to the current sched domain partitioning, doms_cur[].
7713 * It destroys each deleted domain and builds each new domain.
7715 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7716 * The masks don't intersect (don't overlap.) We should setup one
7717 * sched domain for each mask. CPUs not in any of the cpumasks will
7718 * not be load balanced. If the same cpumask appears both in the
7719 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7722 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7723 * ownership of it and will kfree it when done with it. If the caller
7724 * failed the kmalloc call, then it can pass in doms_new == NULL,
7725 * and partition_sched_domains() will fallback to the single partition
7728 * Call with hotplug lock held
7730 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7731 struct sched_domain_attr *dattr_new)
7735 mutex_lock(&sched_domains_mutex);
7737 /* always unregister in case we don't destroy any domains */
7738 unregister_sched_domain_sysctl();
7740 if (doms_new == NULL) {
7742 doms_new = &fallback_doms;
7743 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7747 /* Destroy deleted domains */
7748 for (i = 0; i < ndoms_cur; i++) {
7749 for (j = 0; j < ndoms_new; j++) {
7750 if (cpus_equal(doms_cur[i], doms_new[j])
7751 && dattrs_equal(dattr_cur, i, dattr_new, j))
7754 /* no match - a current sched domain not in new doms_new[] */
7755 detach_destroy_domains(doms_cur + i);
7760 /* Build new domains */
7761 for (i = 0; i < ndoms_new; i++) {
7762 for (j = 0; j < ndoms_cur; j++) {
7763 if (cpus_equal(doms_new[i], doms_cur[j])
7764 && dattrs_equal(dattr_new, i, dattr_cur, j))
7767 /* no match - add a new doms_new */
7768 __build_sched_domains(doms_new + i,
7769 dattr_new ? dattr_new + i : NULL);
7774 /* Remember the new sched domains */
7775 if (doms_cur != &fallback_doms)
7777 kfree(dattr_cur); /* kfree(NULL) is safe */
7778 doms_cur = doms_new;
7779 dattr_cur = dattr_new;
7780 ndoms_cur = ndoms_new;
7782 register_sched_domain_sysctl();
7784 mutex_unlock(&sched_domains_mutex);
7787 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7788 int arch_reinit_sched_domains(void)
7793 mutex_lock(&sched_domains_mutex);
7794 detach_destroy_domains(&cpu_online_map);
7795 free_sched_domains();
7796 err = arch_init_sched_domains(&cpu_online_map);
7797 mutex_unlock(&sched_domains_mutex);
7803 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7807 if (buf[0] != '0' && buf[0] != '1')
7811 sched_smt_power_savings = (buf[0] == '1');
7813 sched_mc_power_savings = (buf[0] == '1');
7815 ret = arch_reinit_sched_domains();
7817 return ret ? ret : count;
7820 #ifdef CONFIG_SCHED_MC
7821 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7823 return sprintf(page, "%u\n", sched_mc_power_savings);
7825 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7826 const char *buf, size_t count)
7828 return sched_power_savings_store(buf, count, 0);
7830 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7831 sched_mc_power_savings_store);
7834 #ifdef CONFIG_SCHED_SMT
7835 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7837 return sprintf(page, "%u\n", sched_smt_power_savings);
7839 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7840 const char *buf, size_t count)
7842 return sched_power_savings_store(buf, count, 1);
7844 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7845 sched_smt_power_savings_store);
7848 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7852 #ifdef CONFIG_SCHED_SMT
7854 err = sysfs_create_file(&cls->kset.kobj,
7855 &attr_sched_smt_power_savings.attr);
7857 #ifdef CONFIG_SCHED_MC
7858 if (!err && mc_capable())
7859 err = sysfs_create_file(&cls->kset.kobj,
7860 &attr_sched_mc_power_savings.attr);
7864 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7867 * Force a reinitialization of the sched domains hierarchy. The domains
7868 * and groups cannot be updated in place without racing with the balancing
7869 * code, so we temporarily attach all running cpus to the NULL domain
7870 * which will prevent rebalancing while the sched domains are recalculated.
7872 static int update_sched_domains(struct notifier_block *nfb,
7873 unsigned long action, void *hcpu)
7875 int cpu = (int)(long)hcpu;
7878 case CPU_DOWN_PREPARE:
7879 case CPU_DOWN_PREPARE_FROZEN:
7880 disable_runtime(cpu_rq(cpu));
7882 case CPU_UP_PREPARE:
7883 case CPU_UP_PREPARE_FROZEN:
7884 detach_destroy_domains(&cpu_online_map);
7885 free_sched_domains();
7889 case CPU_DOWN_FAILED:
7890 case CPU_DOWN_FAILED_FROZEN:
7892 case CPU_ONLINE_FROZEN:
7893 enable_runtime(cpu_rq(cpu));
7895 case CPU_UP_CANCELED:
7896 case CPU_UP_CANCELED_FROZEN:
7898 case CPU_DEAD_FROZEN:
7900 * Fall through and re-initialise the domains.
7907 #ifndef CONFIG_CPUSETS
7909 * Create default domain partitioning if cpusets are disabled.
7910 * Otherwise we let cpusets rebuild the domains based on the
7914 /* The hotplug lock is already held by cpu_up/cpu_down */
7915 arch_init_sched_domains(&cpu_online_map);
7921 void __init sched_init_smp(void)
7923 cpumask_t non_isolated_cpus;
7925 #if defined(CONFIG_NUMA)
7926 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7928 BUG_ON(sched_group_nodes_bycpu == NULL);
7931 mutex_lock(&sched_domains_mutex);
7932 arch_init_sched_domains(&cpu_online_map);
7933 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7934 if (cpus_empty(non_isolated_cpus))
7935 cpu_set(smp_processor_id(), non_isolated_cpus);
7936 mutex_unlock(&sched_domains_mutex);
7938 /* XXX: Theoretical race here - CPU may be hotplugged now */
7939 hotcpu_notifier(update_sched_domains, 0);
7942 /* Move init over to a non-isolated CPU */
7943 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7945 sched_init_granularity();
7948 void __init sched_init_smp(void)
7950 sched_init_granularity();
7952 #endif /* CONFIG_SMP */
7954 int in_sched_functions(unsigned long addr)
7956 return in_lock_functions(addr) ||
7957 (addr >= (unsigned long)__sched_text_start
7958 && addr < (unsigned long)__sched_text_end);
7961 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7963 cfs_rq->tasks_timeline = RB_ROOT;
7964 INIT_LIST_HEAD(&cfs_rq->tasks);
7965 #ifdef CONFIG_FAIR_GROUP_SCHED
7968 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7971 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7973 struct rt_prio_array *array;
7976 array = &rt_rq->active;
7977 for (i = 0; i < MAX_RT_PRIO; i++) {
7978 INIT_LIST_HEAD(array->queue + i);
7979 __clear_bit(i, array->bitmap);
7981 /* delimiter for bitsearch: */
7982 __set_bit(MAX_RT_PRIO, array->bitmap);
7984 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7985 rt_rq->highest_prio = MAX_RT_PRIO;
7988 rt_rq->rt_nr_migratory = 0;
7989 rt_rq->overloaded = 0;
7993 rt_rq->rt_throttled = 0;
7994 rt_rq->rt_runtime = 0;
7995 spin_lock_init(&rt_rq->rt_runtime_lock);
7997 #ifdef CONFIG_RT_GROUP_SCHED
7998 rt_rq->rt_nr_boosted = 0;
8003 #ifdef CONFIG_FAIR_GROUP_SCHED
8004 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8005 struct sched_entity *se, int cpu, int add,
8006 struct sched_entity *parent)
8008 struct rq *rq = cpu_rq(cpu);
8009 tg->cfs_rq[cpu] = cfs_rq;
8010 init_cfs_rq(cfs_rq, rq);
8013 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8016 /* se could be NULL for init_task_group */
8021 se->cfs_rq = &rq->cfs;
8023 se->cfs_rq = parent->my_q;
8026 se->load.weight = tg->shares;
8027 se->load.inv_weight = 0;
8028 se->parent = parent;
8032 #ifdef CONFIG_RT_GROUP_SCHED
8033 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8034 struct sched_rt_entity *rt_se, int cpu, int add,
8035 struct sched_rt_entity *parent)
8037 struct rq *rq = cpu_rq(cpu);
8039 tg->rt_rq[cpu] = rt_rq;
8040 init_rt_rq(rt_rq, rq);
8042 rt_rq->rt_se = rt_se;
8043 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8045 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8047 tg->rt_se[cpu] = rt_se;
8052 rt_se->rt_rq = &rq->rt;
8054 rt_se->rt_rq = parent->my_q;
8056 rt_se->my_q = rt_rq;
8057 rt_se->parent = parent;
8058 INIT_LIST_HEAD(&rt_se->run_list);
8062 void __init sched_init(void)
8065 unsigned long alloc_size = 0, ptr;
8067 #ifdef CONFIG_FAIR_GROUP_SCHED
8068 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8070 #ifdef CONFIG_RT_GROUP_SCHED
8071 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8073 #ifdef CONFIG_USER_SCHED
8077 * As sched_init() is called before page_alloc is setup,
8078 * we use alloc_bootmem().
8081 ptr = (unsigned long)alloc_bootmem(alloc_size);
8083 #ifdef CONFIG_FAIR_GROUP_SCHED
8084 init_task_group.se = (struct sched_entity **)ptr;
8085 ptr += nr_cpu_ids * sizeof(void **);
8087 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8088 ptr += nr_cpu_ids * sizeof(void **);
8090 #ifdef CONFIG_USER_SCHED
8091 root_task_group.se = (struct sched_entity **)ptr;
8092 ptr += nr_cpu_ids * sizeof(void **);
8094 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8095 ptr += nr_cpu_ids * sizeof(void **);
8096 #endif /* CONFIG_USER_SCHED */
8097 #endif /* CONFIG_FAIR_GROUP_SCHED */
8098 #ifdef CONFIG_RT_GROUP_SCHED
8099 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8100 ptr += nr_cpu_ids * sizeof(void **);
8102 init_task_group.rt_rq = (struct rt_rq **)ptr;
8103 ptr += nr_cpu_ids * sizeof(void **);
8105 #ifdef CONFIG_USER_SCHED
8106 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8107 ptr += nr_cpu_ids * sizeof(void **);
8109 root_task_group.rt_rq = (struct rt_rq **)ptr;
8110 ptr += nr_cpu_ids * sizeof(void **);
8111 #endif /* CONFIG_USER_SCHED */
8112 #endif /* CONFIG_RT_GROUP_SCHED */
8117 init_defrootdomain();
8120 init_rt_bandwidth(&def_rt_bandwidth,
8121 global_rt_period(), global_rt_runtime());
8123 #ifdef CONFIG_RT_GROUP_SCHED
8124 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8125 global_rt_period(), global_rt_runtime());
8126 #ifdef CONFIG_USER_SCHED
8127 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8128 global_rt_period(), RUNTIME_INF);
8129 #endif /* CONFIG_USER_SCHED */
8130 #endif /* CONFIG_RT_GROUP_SCHED */
8132 #ifdef CONFIG_GROUP_SCHED
8133 list_add(&init_task_group.list, &task_groups);
8134 INIT_LIST_HEAD(&init_task_group.children);
8136 #ifdef CONFIG_USER_SCHED
8137 INIT_LIST_HEAD(&root_task_group.children);
8138 init_task_group.parent = &root_task_group;
8139 list_add(&init_task_group.siblings, &root_task_group.children);
8140 #endif /* CONFIG_USER_SCHED */
8141 #endif /* CONFIG_GROUP_SCHED */
8143 for_each_possible_cpu(i) {
8147 spin_lock_init(&rq->lock);
8148 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8150 init_cfs_rq(&rq->cfs, rq);
8151 init_rt_rq(&rq->rt, rq);
8152 #ifdef CONFIG_FAIR_GROUP_SCHED
8153 init_task_group.shares = init_task_group_load;
8154 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8155 #ifdef CONFIG_CGROUP_SCHED
8157 * How much cpu bandwidth does init_task_group get?
8159 * In case of task-groups formed thr' the cgroup filesystem, it
8160 * gets 100% of the cpu resources in the system. This overall
8161 * system cpu resource is divided among the tasks of
8162 * init_task_group and its child task-groups in a fair manner,
8163 * based on each entity's (task or task-group's) weight
8164 * (se->load.weight).
8166 * In other words, if init_task_group has 10 tasks of weight
8167 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8168 * then A0's share of the cpu resource is:
8170 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8172 * We achieve this by letting init_task_group's tasks sit
8173 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8175 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8176 #elif defined CONFIG_USER_SCHED
8177 root_task_group.shares = NICE_0_LOAD;
8178 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8180 * In case of task-groups formed thr' the user id of tasks,
8181 * init_task_group represents tasks belonging to root user.
8182 * Hence it forms a sibling of all subsequent groups formed.
8183 * In this case, init_task_group gets only a fraction of overall
8184 * system cpu resource, based on the weight assigned to root
8185 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8186 * by letting tasks of init_task_group sit in a separate cfs_rq
8187 * (init_cfs_rq) and having one entity represent this group of
8188 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8190 init_tg_cfs_entry(&init_task_group,
8191 &per_cpu(init_cfs_rq, i),
8192 &per_cpu(init_sched_entity, i), i, 1,
8193 root_task_group.se[i]);
8196 #endif /* CONFIG_FAIR_GROUP_SCHED */
8198 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8199 #ifdef CONFIG_RT_GROUP_SCHED
8200 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8201 #ifdef CONFIG_CGROUP_SCHED
8202 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8203 #elif defined CONFIG_USER_SCHED
8204 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8205 init_tg_rt_entry(&init_task_group,
8206 &per_cpu(init_rt_rq, i),
8207 &per_cpu(init_sched_rt_entity, i), i, 1,
8208 root_task_group.rt_se[i]);
8212 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8213 rq->cpu_load[j] = 0;
8217 rq->active_balance = 0;
8218 rq->next_balance = jiffies;
8222 rq->migration_thread = NULL;
8223 INIT_LIST_HEAD(&rq->migration_queue);
8224 rq_attach_root(rq, &def_root_domain);
8227 atomic_set(&rq->nr_iowait, 0);
8230 set_load_weight(&init_task);
8232 #ifdef CONFIG_PREEMPT_NOTIFIERS
8233 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8237 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8240 #ifdef CONFIG_RT_MUTEXES
8241 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8245 * The boot idle thread does lazy MMU switching as well:
8247 atomic_inc(&init_mm.mm_count);
8248 enter_lazy_tlb(&init_mm, current);
8251 * Make us the idle thread. Technically, schedule() should not be
8252 * called from this thread, however somewhere below it might be,
8253 * but because we are the idle thread, we just pick up running again
8254 * when this runqueue becomes "idle".
8256 init_idle(current, smp_processor_id());
8258 * During early bootup we pretend to be a normal task:
8260 current->sched_class = &fair_sched_class;
8262 scheduler_running = 1;
8265 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8266 void __might_sleep(char *file, int line)
8269 static unsigned long prev_jiffy; /* ratelimiting */
8271 if ((in_atomic() || irqs_disabled()) &&
8272 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8273 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8275 prev_jiffy = jiffies;
8276 printk(KERN_ERR "BUG: sleeping function called from invalid"
8277 " context at %s:%d\n", file, line);
8278 printk("in_atomic():%d, irqs_disabled():%d\n",
8279 in_atomic(), irqs_disabled());
8280 debug_show_held_locks(current);
8281 if (irqs_disabled())
8282 print_irqtrace_events(current);
8287 EXPORT_SYMBOL(__might_sleep);
8290 #ifdef CONFIG_MAGIC_SYSRQ
8291 static void normalize_task(struct rq *rq, struct task_struct *p)
8295 update_rq_clock(rq);
8296 on_rq = p->se.on_rq;
8298 deactivate_task(rq, p, 0);
8299 __setscheduler(rq, p, SCHED_NORMAL, 0);
8301 activate_task(rq, p, 0);
8302 resched_task(rq->curr);
8306 void normalize_rt_tasks(void)
8308 struct task_struct *g, *p;
8309 unsigned long flags;
8312 read_lock_irqsave(&tasklist_lock, flags);
8313 do_each_thread(g, p) {
8315 * Only normalize user tasks:
8320 p->se.exec_start = 0;
8321 #ifdef CONFIG_SCHEDSTATS
8322 p->se.wait_start = 0;
8323 p->se.sleep_start = 0;
8324 p->se.block_start = 0;
8329 * Renice negative nice level userspace
8332 if (TASK_NICE(p) < 0 && p->mm)
8333 set_user_nice(p, 0);
8337 spin_lock(&p->pi_lock);
8338 rq = __task_rq_lock(p);
8340 normalize_task(rq, p);
8342 __task_rq_unlock(rq);
8343 spin_unlock(&p->pi_lock);
8344 } while_each_thread(g, p);
8346 read_unlock_irqrestore(&tasklist_lock, flags);
8349 #endif /* CONFIG_MAGIC_SYSRQ */
8353 * These functions are only useful for the IA64 MCA handling.
8355 * They can only be called when the whole system has been
8356 * stopped - every CPU needs to be quiescent, and no scheduling
8357 * activity can take place. Using them for anything else would
8358 * be a serious bug, and as a result, they aren't even visible
8359 * under any other configuration.
8363 * curr_task - return the current task for a given cpu.
8364 * @cpu: the processor in question.
8366 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8368 struct task_struct *curr_task(int cpu)
8370 return cpu_curr(cpu);
8374 * set_curr_task - set the current task for a given cpu.
8375 * @cpu: the processor in question.
8376 * @p: the task pointer to set.
8378 * Description: This function must only be used when non-maskable interrupts
8379 * are serviced on a separate stack. It allows the architecture to switch the
8380 * notion of the current task on a cpu in a non-blocking manner. This function
8381 * must be called with all CPU's synchronized, and interrupts disabled, the
8382 * and caller must save the original value of the current task (see
8383 * curr_task() above) and restore that value before reenabling interrupts and
8384 * re-starting the system.
8386 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8388 void set_curr_task(int cpu, struct task_struct *p)
8395 #ifdef CONFIG_FAIR_GROUP_SCHED
8396 static void free_fair_sched_group(struct task_group *tg)
8400 for_each_possible_cpu(i) {
8402 kfree(tg->cfs_rq[i]);
8412 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8414 struct cfs_rq *cfs_rq;
8415 struct sched_entity *se, *parent_se;
8419 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8422 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8426 tg->shares = NICE_0_LOAD;
8428 for_each_possible_cpu(i) {
8431 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8432 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8436 se = kmalloc_node(sizeof(struct sched_entity),
8437 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8441 parent_se = parent ? parent->se[i] : NULL;
8442 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8451 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8453 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8454 &cpu_rq(cpu)->leaf_cfs_rq_list);
8457 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8459 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8461 #else /* !CONFG_FAIR_GROUP_SCHED */
8462 static inline void free_fair_sched_group(struct task_group *tg)
8467 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8472 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8476 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8479 #endif /* CONFIG_FAIR_GROUP_SCHED */
8481 #ifdef CONFIG_RT_GROUP_SCHED
8482 static void free_rt_sched_group(struct task_group *tg)
8486 destroy_rt_bandwidth(&tg->rt_bandwidth);
8488 for_each_possible_cpu(i) {
8490 kfree(tg->rt_rq[i]);
8492 kfree(tg->rt_se[i]);
8500 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8502 struct rt_rq *rt_rq;
8503 struct sched_rt_entity *rt_se, *parent_se;
8507 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8510 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8514 init_rt_bandwidth(&tg->rt_bandwidth,
8515 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8517 for_each_possible_cpu(i) {
8520 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8521 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8525 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8526 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8530 parent_se = parent ? parent->rt_se[i] : NULL;
8531 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8540 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8542 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8543 &cpu_rq(cpu)->leaf_rt_rq_list);
8546 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8548 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8550 #else /* !CONFIG_RT_GROUP_SCHED */
8551 static inline void free_rt_sched_group(struct task_group *tg)
8556 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8561 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8565 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8568 #endif /* CONFIG_RT_GROUP_SCHED */
8570 #ifdef CONFIG_GROUP_SCHED
8571 static void free_sched_group(struct task_group *tg)
8573 free_fair_sched_group(tg);
8574 free_rt_sched_group(tg);
8578 /* allocate runqueue etc for a new task group */
8579 struct task_group *sched_create_group(struct task_group *parent)
8581 struct task_group *tg;
8582 unsigned long flags;
8585 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8587 return ERR_PTR(-ENOMEM);
8589 if (!alloc_fair_sched_group(tg, parent))
8592 if (!alloc_rt_sched_group(tg, parent))
8595 spin_lock_irqsave(&task_group_lock, flags);
8596 for_each_possible_cpu(i) {
8597 register_fair_sched_group(tg, i);
8598 register_rt_sched_group(tg, i);
8600 list_add_rcu(&tg->list, &task_groups);
8602 WARN_ON(!parent); /* root should already exist */
8604 tg->parent = parent;
8605 list_add_rcu(&tg->siblings, &parent->children);
8606 INIT_LIST_HEAD(&tg->children);
8607 spin_unlock_irqrestore(&task_group_lock, flags);
8612 free_sched_group(tg);
8613 return ERR_PTR(-ENOMEM);
8616 /* rcu callback to free various structures associated with a task group */
8617 static void free_sched_group_rcu(struct rcu_head *rhp)
8619 /* now it should be safe to free those cfs_rqs */
8620 free_sched_group(container_of(rhp, struct task_group, rcu));
8623 /* Destroy runqueue etc associated with a task group */
8624 void sched_destroy_group(struct task_group *tg)
8626 unsigned long flags;
8629 spin_lock_irqsave(&task_group_lock, flags);
8630 for_each_possible_cpu(i) {
8631 unregister_fair_sched_group(tg, i);
8632 unregister_rt_sched_group(tg, i);
8634 list_del_rcu(&tg->list);
8635 list_del_rcu(&tg->siblings);
8636 spin_unlock_irqrestore(&task_group_lock, flags);
8638 /* wait for possible concurrent references to cfs_rqs complete */
8639 call_rcu(&tg->rcu, free_sched_group_rcu);
8642 /* change task's runqueue when it moves between groups.
8643 * The caller of this function should have put the task in its new group
8644 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8645 * reflect its new group.
8647 void sched_move_task(struct task_struct *tsk)
8650 unsigned long flags;
8653 rq = task_rq_lock(tsk, &flags);
8655 update_rq_clock(rq);
8657 running = task_current(rq, tsk);
8658 on_rq = tsk->se.on_rq;
8661 dequeue_task(rq, tsk, 0);
8662 if (unlikely(running))
8663 tsk->sched_class->put_prev_task(rq, tsk);
8665 set_task_rq(tsk, task_cpu(tsk));
8667 #ifdef CONFIG_FAIR_GROUP_SCHED
8668 if (tsk->sched_class->moved_group)
8669 tsk->sched_class->moved_group(tsk);
8672 if (unlikely(running))
8673 tsk->sched_class->set_curr_task(rq);
8675 enqueue_task(rq, tsk, 0);
8677 task_rq_unlock(rq, &flags);
8679 #endif /* CONFIG_GROUP_SCHED */
8681 #ifdef CONFIG_FAIR_GROUP_SCHED
8682 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8684 struct cfs_rq *cfs_rq = se->cfs_rq;
8689 dequeue_entity(cfs_rq, se, 0);
8691 se->load.weight = shares;
8692 se->load.inv_weight = 0;
8695 enqueue_entity(cfs_rq, se, 0);
8698 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8700 struct cfs_rq *cfs_rq = se->cfs_rq;
8701 struct rq *rq = cfs_rq->rq;
8702 unsigned long flags;
8704 spin_lock_irqsave(&rq->lock, flags);
8705 __set_se_shares(se, shares);
8706 spin_unlock_irqrestore(&rq->lock, flags);
8709 static DEFINE_MUTEX(shares_mutex);
8711 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8714 unsigned long flags;
8717 * We can't change the weight of the root cgroup.
8722 if (shares < MIN_SHARES)
8723 shares = MIN_SHARES;
8724 else if (shares > MAX_SHARES)
8725 shares = MAX_SHARES;
8727 mutex_lock(&shares_mutex);
8728 if (tg->shares == shares)
8731 spin_lock_irqsave(&task_group_lock, flags);
8732 for_each_possible_cpu(i)
8733 unregister_fair_sched_group(tg, i);
8734 list_del_rcu(&tg->siblings);
8735 spin_unlock_irqrestore(&task_group_lock, flags);
8737 /* wait for any ongoing reference to this group to finish */
8738 synchronize_sched();
8741 * Now we are free to modify the group's share on each cpu
8742 * w/o tripping rebalance_share or load_balance_fair.
8744 tg->shares = shares;
8745 for_each_possible_cpu(i) {
8749 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8750 set_se_shares(tg->se[i], shares);
8754 * Enable load balance activity on this group, by inserting it back on
8755 * each cpu's rq->leaf_cfs_rq_list.
8757 spin_lock_irqsave(&task_group_lock, flags);
8758 for_each_possible_cpu(i)
8759 register_fair_sched_group(tg, i);
8760 list_add_rcu(&tg->siblings, &tg->parent->children);
8761 spin_unlock_irqrestore(&task_group_lock, flags);
8763 mutex_unlock(&shares_mutex);
8767 unsigned long sched_group_shares(struct task_group *tg)
8773 #ifdef CONFIG_RT_GROUP_SCHED
8775 * Ensure that the real time constraints are schedulable.
8777 static DEFINE_MUTEX(rt_constraints_mutex);
8779 static unsigned long to_ratio(u64 period, u64 runtime)
8781 if (runtime == RUNTIME_INF)
8784 return div64_u64(runtime << 16, period);
8787 #ifdef CONFIG_CGROUP_SCHED
8788 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8790 struct task_group *tgi, *parent = tg->parent;
8791 unsigned long total = 0;
8794 if (global_rt_period() < period)
8797 return to_ratio(period, runtime) <
8798 to_ratio(global_rt_period(), global_rt_runtime());
8801 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8805 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8809 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8810 tgi->rt_bandwidth.rt_runtime);
8814 return total + to_ratio(period, runtime) <=
8815 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8816 parent->rt_bandwidth.rt_runtime);
8818 #elif defined CONFIG_USER_SCHED
8819 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8821 struct task_group *tgi;
8822 unsigned long total = 0;
8823 unsigned long global_ratio =
8824 to_ratio(global_rt_period(), global_rt_runtime());
8827 list_for_each_entry_rcu(tgi, &task_groups, list) {
8831 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8832 tgi->rt_bandwidth.rt_runtime);
8836 return total + to_ratio(period, runtime) < global_ratio;
8840 /* Must be called with tasklist_lock held */
8841 static inline int tg_has_rt_tasks(struct task_group *tg)
8843 struct task_struct *g, *p;
8844 do_each_thread(g, p) {
8845 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8847 } while_each_thread(g, p);
8851 static int tg_set_bandwidth(struct task_group *tg,
8852 u64 rt_period, u64 rt_runtime)
8856 mutex_lock(&rt_constraints_mutex);
8857 read_lock(&tasklist_lock);
8858 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8862 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8867 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8868 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8869 tg->rt_bandwidth.rt_runtime = rt_runtime;
8871 for_each_possible_cpu(i) {
8872 struct rt_rq *rt_rq = tg->rt_rq[i];
8874 spin_lock(&rt_rq->rt_runtime_lock);
8875 rt_rq->rt_runtime = rt_runtime;
8876 spin_unlock(&rt_rq->rt_runtime_lock);
8878 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8880 read_unlock(&tasklist_lock);
8881 mutex_unlock(&rt_constraints_mutex);
8886 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8888 u64 rt_runtime, rt_period;
8890 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8891 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8892 if (rt_runtime_us < 0)
8893 rt_runtime = RUNTIME_INF;
8895 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8898 long sched_group_rt_runtime(struct task_group *tg)
8902 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8905 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8906 do_div(rt_runtime_us, NSEC_PER_USEC);
8907 return rt_runtime_us;
8910 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8912 u64 rt_runtime, rt_period;
8914 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8915 rt_runtime = tg->rt_bandwidth.rt_runtime;
8917 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8920 long sched_group_rt_period(struct task_group *tg)
8924 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8925 do_div(rt_period_us, NSEC_PER_USEC);
8926 return rt_period_us;
8929 static int sched_rt_global_constraints(void)
8931 struct task_group *tg = &root_task_group;
8932 u64 rt_runtime, rt_period;
8935 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8936 rt_runtime = tg->rt_bandwidth.rt_runtime;
8938 mutex_lock(&rt_constraints_mutex);
8939 if (!__rt_schedulable(tg, rt_period, rt_runtime))
8941 mutex_unlock(&rt_constraints_mutex);
8945 #else /* !CONFIG_RT_GROUP_SCHED */
8946 static int sched_rt_global_constraints(void)
8948 unsigned long flags;
8951 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8952 for_each_possible_cpu(i) {
8953 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8955 spin_lock(&rt_rq->rt_runtime_lock);
8956 rt_rq->rt_runtime = global_rt_runtime();
8957 spin_unlock(&rt_rq->rt_runtime_lock);
8959 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8963 #endif /* CONFIG_RT_GROUP_SCHED */
8965 int sched_rt_handler(struct ctl_table *table, int write,
8966 struct file *filp, void __user *buffer, size_t *lenp,
8970 int old_period, old_runtime;
8971 static DEFINE_MUTEX(mutex);
8974 old_period = sysctl_sched_rt_period;
8975 old_runtime = sysctl_sched_rt_runtime;
8977 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8979 if (!ret && write) {
8980 ret = sched_rt_global_constraints();
8982 sysctl_sched_rt_period = old_period;
8983 sysctl_sched_rt_runtime = old_runtime;
8985 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8986 def_rt_bandwidth.rt_period =
8987 ns_to_ktime(global_rt_period());
8990 mutex_unlock(&mutex);
8995 #ifdef CONFIG_CGROUP_SCHED
8997 /* return corresponding task_group object of a cgroup */
8998 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9000 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9001 struct task_group, css);
9004 static struct cgroup_subsys_state *
9005 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9007 struct task_group *tg, *parent;
9009 if (!cgrp->parent) {
9010 /* This is early initialization for the top cgroup */
9011 init_task_group.css.cgroup = cgrp;
9012 return &init_task_group.css;
9015 parent = cgroup_tg(cgrp->parent);
9016 tg = sched_create_group(parent);
9018 return ERR_PTR(-ENOMEM);
9020 /* Bind the cgroup to task_group object we just created */
9021 tg->css.cgroup = cgrp;
9027 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9029 struct task_group *tg = cgroup_tg(cgrp);
9031 sched_destroy_group(tg);
9035 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9036 struct task_struct *tsk)
9038 #ifdef CONFIG_RT_GROUP_SCHED
9039 /* Don't accept realtime tasks when there is no way for them to run */
9040 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9043 /* We don't support RT-tasks being in separate groups */
9044 if (tsk->sched_class != &fair_sched_class)
9052 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9053 struct cgroup *old_cont, struct task_struct *tsk)
9055 sched_move_task(tsk);
9058 #ifdef CONFIG_FAIR_GROUP_SCHED
9059 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9062 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9065 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9067 struct task_group *tg = cgroup_tg(cgrp);
9069 return (u64) tg->shares;
9071 #endif /* CONFIG_FAIR_GROUP_SCHED */
9073 #ifdef CONFIG_RT_GROUP_SCHED
9074 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9077 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9080 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9082 return sched_group_rt_runtime(cgroup_tg(cgrp));
9085 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9088 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9091 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9093 return sched_group_rt_period(cgroup_tg(cgrp));
9095 #endif /* CONFIG_RT_GROUP_SCHED */
9097 static struct cftype cpu_files[] = {
9098 #ifdef CONFIG_FAIR_GROUP_SCHED
9101 .read_u64 = cpu_shares_read_u64,
9102 .write_u64 = cpu_shares_write_u64,
9105 #ifdef CONFIG_RT_GROUP_SCHED
9107 .name = "rt_runtime_us",
9108 .read_s64 = cpu_rt_runtime_read,
9109 .write_s64 = cpu_rt_runtime_write,
9112 .name = "rt_period_us",
9113 .read_u64 = cpu_rt_period_read_uint,
9114 .write_u64 = cpu_rt_period_write_uint,
9119 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9121 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9124 struct cgroup_subsys cpu_cgroup_subsys = {
9126 .create = cpu_cgroup_create,
9127 .destroy = cpu_cgroup_destroy,
9128 .can_attach = cpu_cgroup_can_attach,
9129 .attach = cpu_cgroup_attach,
9130 .populate = cpu_cgroup_populate,
9131 .subsys_id = cpu_cgroup_subsys_id,
9135 #endif /* CONFIG_CGROUP_SCHED */
9137 #ifdef CONFIG_CGROUP_CPUACCT
9140 * CPU accounting code for task groups.
9142 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9143 * (balbir@in.ibm.com).
9146 /* track cpu usage of a group of tasks */
9148 struct cgroup_subsys_state css;
9149 /* cpuusage holds pointer to a u64-type object on every cpu */
9153 struct cgroup_subsys cpuacct_subsys;
9155 /* return cpu accounting group corresponding to this container */
9156 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9158 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9159 struct cpuacct, css);
9162 /* return cpu accounting group to which this task belongs */
9163 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9165 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9166 struct cpuacct, css);
9169 /* create a new cpu accounting group */
9170 static struct cgroup_subsys_state *cpuacct_create(
9171 struct cgroup_subsys *ss, struct cgroup *cgrp)
9173 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9176 return ERR_PTR(-ENOMEM);
9178 ca->cpuusage = alloc_percpu(u64);
9179 if (!ca->cpuusage) {
9181 return ERR_PTR(-ENOMEM);
9187 /* destroy an existing cpu accounting group */
9189 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9191 struct cpuacct *ca = cgroup_ca(cgrp);
9193 free_percpu(ca->cpuusage);
9197 /* return total cpu usage (in nanoseconds) of a group */
9198 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9200 struct cpuacct *ca = cgroup_ca(cgrp);
9201 u64 totalcpuusage = 0;
9204 for_each_possible_cpu(i) {
9205 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9208 * Take rq->lock to make 64-bit addition safe on 32-bit
9211 spin_lock_irq(&cpu_rq(i)->lock);
9212 totalcpuusage += *cpuusage;
9213 spin_unlock_irq(&cpu_rq(i)->lock);
9216 return totalcpuusage;
9219 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9222 struct cpuacct *ca = cgroup_ca(cgrp);
9231 for_each_possible_cpu(i) {
9232 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9234 spin_lock_irq(&cpu_rq(i)->lock);
9236 spin_unlock_irq(&cpu_rq(i)->lock);
9242 static struct cftype files[] = {
9245 .read_u64 = cpuusage_read,
9246 .write_u64 = cpuusage_write,
9250 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9252 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9256 * charge this task's execution time to its accounting group.
9258 * called with rq->lock held.
9260 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9264 if (!cpuacct_subsys.active)
9269 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9271 *cpuusage += cputime;
9275 struct cgroup_subsys cpuacct_subsys = {
9277 .create = cpuacct_create,
9278 .destroy = cpuacct_destroy,
9279 .populate = cpuacct_populate,
9280 .subsys_id = cpuacct_subsys_id,
9282 #endif /* CONFIG_CGROUP_CPUACCT */