4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task);
122 DEFINE_TRACE(sched_wakeup);
123 DEFINE_TRACE(sched_wakeup_new);
124 DEFINE_TRACE(sched_switch);
125 DEFINE_TRACE(sched_migrate_task);
129 static void double_rq_lock(struct rq *rq1, struct rq *rq2);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
137 return reciprocal_divide(load, sg->reciprocal_cpu_power);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
146 sg->__cpu_power += val;
147 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
151 static inline int rt_policy(int policy)
153 if (unlikely(policy == SCHED_FIFO || policy == SCHED_RR))
158 static inline int task_has_rt_policy(struct task_struct *p)
160 return rt_policy(p->policy);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array {
167 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
168 struct list_head queue[MAX_RT_PRIO];
171 struct rt_bandwidth {
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock;
176 struct hrtimer rt_period_timer;
179 static struct rt_bandwidth def_rt_bandwidth;
181 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
183 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
185 struct rt_bandwidth *rt_b =
186 container_of(timer, struct rt_bandwidth, rt_period_timer);
192 now = hrtimer_cb_get_time(timer);
193 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
198 idle = do_sched_rt_period_timer(rt_b, overrun);
201 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
205 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
207 rt_b->rt_period = ns_to_ktime(period);
208 rt_b->rt_runtime = runtime;
210 spin_lock_init(&rt_b->rt_runtime_lock);
212 hrtimer_init(&rt_b->rt_period_timer,
213 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
214 rt_b->rt_period_timer.function = sched_rt_period_timer;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime >= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
226 if (rt_bandwidth_enabled() && rt_b->rt_runtime == RUNTIME_INF)
229 if (hrtimer_active(&rt_b->rt_period_timer))
232 spin_lock(&rt_b->rt_runtime_lock);
234 if (hrtimer_active(&rt_b->rt_period_timer))
237 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
238 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
239 hrtimer_start_expires(&rt_b->rt_period_timer,
242 spin_unlock(&rt_b->rt_runtime_lock);
245 #ifdef CONFIG_RT_GROUP_SCHED
246 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
248 hrtimer_cancel(&rt_b->rt_period_timer);
253 * sched_domains_mutex serializes calls to arch_init_sched_domains,
254 * detach_destroy_domains and partition_sched_domains.
256 static DEFINE_MUTEX(sched_domains_mutex);
258 #ifdef CONFIG_GROUP_SCHED
260 #include <linux/cgroup.h>
264 static LIST_HEAD(task_groups);
266 /* task group related information */
268 #ifdef CONFIG_CGROUP_SCHED
269 struct cgroup_subsys_state css;
272 #ifdef CONFIG_USER_SCHED
276 #ifdef CONFIG_FAIR_GROUP_SCHED
277 /* schedulable entities of this group on each cpu */
278 struct sched_entity **se;
279 /* runqueue "owned" by this group on each cpu */
280 struct cfs_rq **cfs_rq;
281 unsigned long shares;
284 #ifdef CONFIG_RT_GROUP_SCHED
285 struct sched_rt_entity **rt_se;
286 struct rt_rq **rt_rq;
288 struct rt_bandwidth rt_bandwidth;
292 struct list_head list;
294 struct task_group *parent;
295 struct list_head siblings;
296 struct list_head children;
299 #ifdef CONFIG_USER_SCHED
301 /* Helper function to pass uid information to create_sched_user() */
302 void set_tg_uid(struct user_struct *user)
304 user->tg->uid = user->uid;
309 * Every UID task group (including init_task_group aka UID-0) will
310 * be a child to this group.
312 struct task_group root_task_group;
314 #ifdef CONFIG_FAIR_GROUP_SCHED
315 /* Default task group's sched entity on each cpu */
316 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
317 /* Default task group's cfs_rq on each cpu */
318 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
319 #endif /* CONFIG_FAIR_GROUP_SCHED */
321 #ifdef CONFIG_RT_GROUP_SCHED
322 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
323 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
324 #endif /* CONFIG_RT_GROUP_SCHED */
325 #else /* !CONFIG_USER_SCHED */
326 #define root_task_group init_task_group
327 #endif /* CONFIG_USER_SCHED */
329 /* task_group_lock serializes add/remove of task groups and also changes to
330 * a task group's cpu shares.
332 static DEFINE_SPINLOCK(task_group_lock);
334 #ifdef CONFIG_FAIR_GROUP_SCHED
335 #ifdef CONFIG_USER_SCHED
336 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
337 #else /* !CONFIG_USER_SCHED */
338 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
339 #endif /* CONFIG_USER_SCHED */
342 * A weight of 0 or 1 can cause arithmetics problems.
343 * A weight of a cfs_rq is the sum of weights of which entities
344 * are queued on this cfs_rq, so a weight of a entity should not be
345 * too large, so as the shares value of a task group.
346 * (The default weight is 1024 - so there's no practical
347 * limitation from this.)
350 #define MAX_SHARES (1UL << 18)
352 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
355 /* Default task group.
356 * Every task in system belong to this group at bootup.
358 struct task_group init_task_group;
360 /* return group to which a task belongs */
361 static inline struct task_group *task_group(struct task_struct *p)
363 struct task_group *tg;
365 #ifdef CONFIG_USER_SCHED
367 tg = __task_cred(p)->user->tg;
369 #elif defined(CONFIG_CGROUP_SCHED)
370 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
371 struct task_group, css);
373 tg = &init_task_group;
378 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
379 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
381 #ifdef CONFIG_FAIR_GROUP_SCHED
382 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
383 p->se.parent = task_group(p)->se[cpu];
386 #ifdef CONFIG_RT_GROUP_SCHED
387 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
388 p->rt.parent = task_group(p)->rt_se[cpu];
394 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
395 static inline struct task_group *task_group(struct task_struct *p)
400 #endif /* CONFIG_GROUP_SCHED */
402 /* CFS-related fields in a runqueue */
404 struct load_weight load;
405 unsigned long nr_running;
410 struct rb_root tasks_timeline;
411 struct rb_node *rb_leftmost;
413 struct list_head tasks;
414 struct list_head *balance_iterator;
417 * 'curr' points to currently running entity on this cfs_rq.
418 * It is set to NULL otherwise (i.e when none are currently running).
420 struct sched_entity *curr, *next, *last;
422 unsigned int nr_spread_over;
424 #ifdef CONFIG_FAIR_GROUP_SCHED
425 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
428 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
429 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
430 * (like users, containers etc.)
432 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
433 * list is used during load balance.
435 struct list_head leaf_cfs_rq_list;
436 struct task_group *tg; /* group that "owns" this runqueue */
440 * the part of load.weight contributed by tasks
442 unsigned long task_weight;
445 * h_load = weight * f(tg)
447 * Where f(tg) is the recursive weight fraction assigned to
450 unsigned long h_load;
453 * this cpu's part of tg->shares
455 unsigned long shares;
458 * load.weight at the time we set shares
460 unsigned long rq_weight;
465 /* Real-Time classes' related field in a runqueue: */
467 struct rt_prio_array active;
468 unsigned long rt_nr_running;
469 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
470 int highest_prio; /* highest queued rt task prio */
473 unsigned long rt_nr_migratory;
479 /* Nests inside the rq lock: */
480 spinlock_t rt_runtime_lock;
482 #ifdef CONFIG_RT_GROUP_SCHED
483 unsigned long rt_nr_boosted;
486 struct list_head leaf_rt_rq_list;
487 struct task_group *tg;
488 struct sched_rt_entity *rt_se;
495 * We add the notion of a root-domain which will be used to define per-domain
496 * variables. Each exclusive cpuset essentially defines an island domain by
497 * fully partitioning the member cpus from any other cpuset. Whenever a new
498 * exclusive cpuset is created, we also create and attach a new root-domain
505 cpumask_var_t online;
508 * The "RT overload" flag: it gets set if a CPU has more than
509 * one runnable RT task.
511 cpumask_var_t rto_mask;
514 struct cpupri cpupri;
516 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
518 * Preferred wake up cpu nominated by sched_mc balance that will be
519 * used when most cpus are idle in the system indicating overall very
520 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
522 unsigned int sched_mc_preferred_wakeup_cpu;
527 * By default the system creates a single root-domain with all cpus as
528 * members (mimicking the global state we have today).
530 static struct root_domain def_root_domain;
535 * This is the main, per-CPU runqueue data structure.
537 * Locking rule: those places that want to lock multiple runqueues
538 * (such as the load balancing or the thread migration code), lock
539 * acquire operations must be ordered by ascending &runqueue.
546 * nr_running and cpu_load should be in the same cacheline because
547 * remote CPUs use both these fields when doing load calculation.
549 unsigned long nr_running;
550 #define CPU_LOAD_IDX_MAX 5
551 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
552 unsigned char idle_at_tick;
554 unsigned long last_tick_seen;
555 unsigned char in_nohz_recently;
557 /* capture load from *all* tasks on this cpu: */
558 struct load_weight load;
559 unsigned long nr_load_updates;
565 #ifdef CONFIG_FAIR_GROUP_SCHED
566 /* list of leaf cfs_rq on this cpu: */
567 struct list_head leaf_cfs_rq_list;
569 #ifdef CONFIG_RT_GROUP_SCHED
570 struct list_head leaf_rt_rq_list;
574 * This is part of a global counter where only the total sum
575 * over all CPUs matters. A task can increase this counter on
576 * one CPU and if it got migrated afterwards it may decrease
577 * it on another CPU. Always updated under the runqueue lock:
579 unsigned long nr_uninterruptible;
581 struct task_struct *curr, *idle;
582 unsigned long next_balance;
583 struct mm_struct *prev_mm;
590 struct root_domain *rd;
591 struct sched_domain *sd;
593 /* For active balancing */
596 /* cpu of this runqueue: */
600 unsigned long avg_load_per_task;
602 struct task_struct *migration_thread;
603 struct list_head migration_queue;
606 #ifdef CONFIG_SCHED_HRTICK
608 int hrtick_csd_pending;
609 struct call_single_data hrtick_csd;
611 struct hrtimer hrtick_timer;
614 #ifdef CONFIG_SCHEDSTATS
616 struct sched_info rq_sched_info;
617 unsigned long long rq_cpu_time;
618 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
620 /* sys_sched_yield() stats */
621 unsigned int yld_exp_empty;
622 unsigned int yld_act_empty;
623 unsigned int yld_both_empty;
624 unsigned int yld_count;
626 /* schedule() stats */
627 unsigned int sched_switch;
628 unsigned int sched_count;
629 unsigned int sched_goidle;
631 /* try_to_wake_up() stats */
632 unsigned int ttwu_count;
633 unsigned int ttwu_local;
636 unsigned int bkl_count;
640 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
642 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p, int sync)
644 rq->curr->sched_class->check_preempt_curr(rq, p, sync);
647 static inline int cpu_of(struct rq *rq)
657 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
658 * See detach_destroy_domains: synchronize_sched for details.
660 * The domain tree of any CPU may only be accessed from within
661 * preempt-disabled sections.
663 #define for_each_domain(cpu, __sd) \
664 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
666 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
667 #define this_rq() (&__get_cpu_var(runqueues))
668 #define task_rq(p) cpu_rq(task_cpu(p))
669 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
671 static inline void update_rq_clock(struct rq *rq)
673 rq->clock = sched_clock_cpu(cpu_of(rq));
677 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
679 #ifdef CONFIG_SCHED_DEBUG
680 # define const_debug __read_mostly
682 # define const_debug static const
688 * Returns true if the current cpu runqueue is locked.
689 * This interface allows printk to be called with the runqueue lock
690 * held and know whether or not it is OK to wake up the klogd.
692 int runqueue_is_locked(void)
695 struct rq *rq = cpu_rq(cpu);
698 ret = spin_is_locked(&rq->lock);
704 * Debugging: various feature bits
707 #define SCHED_FEAT(name, enabled) \
708 __SCHED_FEAT_##name ,
711 #include "sched_features.h"
716 #define SCHED_FEAT(name, enabled) \
717 (1UL << __SCHED_FEAT_##name) * enabled |
719 const_debug unsigned int sysctl_sched_features =
720 #include "sched_features.h"
725 #ifdef CONFIG_SCHED_DEBUG
726 #define SCHED_FEAT(name, enabled) \
729 static __read_mostly char *sched_feat_names[] = {
730 #include "sched_features.h"
736 static int sched_feat_show(struct seq_file *m, void *v)
740 for (i = 0; sched_feat_names[i]; i++) {
741 if (!(sysctl_sched_features & (1UL << i)))
743 seq_printf(m, "%s ", sched_feat_names[i]);
751 sched_feat_write(struct file *filp, const char __user *ubuf,
752 size_t cnt, loff_t *ppos)
762 if (copy_from_user(&buf, ubuf, cnt))
767 if (strncmp(buf, "NO_", 3) == 0) {
772 for (i = 0; sched_feat_names[i]; i++) {
773 int len = strlen(sched_feat_names[i]);
775 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
777 sysctl_sched_features &= ~(1UL << i);
779 sysctl_sched_features |= (1UL << i);
784 if (!sched_feat_names[i])
792 static int sched_feat_open(struct inode *inode, struct file *filp)
794 return single_open(filp, sched_feat_show, NULL);
797 static struct file_operations sched_feat_fops = {
798 .open = sched_feat_open,
799 .write = sched_feat_write,
802 .release = single_release,
805 static __init int sched_init_debug(void)
807 debugfs_create_file("sched_features", 0644, NULL, NULL,
812 late_initcall(sched_init_debug);
816 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
819 * Number of tasks to iterate in a single balance run.
820 * Limited because this is done with IRQs disabled.
822 const_debug unsigned int sysctl_sched_nr_migrate = 32;
825 * ratelimit for updating the group shares.
828 unsigned int sysctl_sched_shares_ratelimit = 250000;
831 * Inject some fuzzyness into changing the per-cpu group shares
832 * this avoids remote rq-locks at the expense of fairness.
835 unsigned int sysctl_sched_shares_thresh = 4;
838 * period over which we measure -rt task cpu usage in us.
841 unsigned int sysctl_sched_rt_period = 1000000;
843 static __read_mostly int scheduler_running;
846 * part of the period that we allow rt tasks to run in us.
849 int sysctl_sched_rt_runtime = 950000;
851 static inline u64 global_rt_period(void)
853 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
856 static inline u64 global_rt_runtime(void)
858 if (sysctl_sched_rt_runtime < 0)
861 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
864 #ifndef prepare_arch_switch
865 # define prepare_arch_switch(next) do { } while (0)
867 #ifndef finish_arch_switch
868 # define finish_arch_switch(prev) do { } while (0)
871 static inline int task_current(struct rq *rq, struct task_struct *p)
873 return rq->curr == p;
876 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
877 static inline int task_running(struct rq *rq, struct task_struct *p)
879 return task_current(rq, p);
882 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
886 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
888 #ifdef CONFIG_DEBUG_SPINLOCK
889 /* this is a valid case when another task releases the spinlock */
890 rq->lock.owner = current;
893 * If we are tracking spinlock dependencies then we have to
894 * fix up the runqueue lock - which gets 'carried over' from
897 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
899 spin_unlock_irq(&rq->lock);
902 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
903 static inline int task_running(struct rq *rq, struct task_struct *p)
908 return task_current(rq, p);
912 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
916 * We can optimise this out completely for !SMP, because the
917 * SMP rebalancing from interrupt is the only thing that cares
922 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
923 spin_unlock_irq(&rq->lock);
925 spin_unlock(&rq->lock);
929 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
933 * After ->oncpu is cleared, the task can be moved to a different CPU.
934 * We must ensure this doesn't happen until the switch is completely
940 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
944 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
947 * __task_rq_lock - lock the runqueue a given task resides on.
948 * Must be called interrupts disabled.
950 static inline struct rq *__task_rq_lock(struct task_struct *p)
954 struct rq *rq = task_rq(p);
955 spin_lock(&rq->lock);
956 if (likely(rq == task_rq(p)))
958 spin_unlock(&rq->lock);
963 * task_rq_lock - lock the runqueue a given task resides on and disable
964 * interrupts. Note the ordering: we can safely lookup the task_rq without
965 * explicitly disabling preemption.
967 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
973 local_irq_save(*flags);
975 spin_lock(&rq->lock);
976 if (likely(rq == task_rq(p)))
978 spin_unlock_irqrestore(&rq->lock, *flags);
982 void task_rq_unlock_wait(struct task_struct *p)
984 struct rq *rq = task_rq(p);
986 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
987 spin_unlock_wait(&rq->lock);
990 static void __task_rq_unlock(struct rq *rq)
993 spin_unlock(&rq->lock);
996 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
999 spin_unlock_irqrestore(&rq->lock, *flags);
1003 * this_rq_lock - lock this runqueue and disable interrupts.
1005 static struct rq *this_rq_lock(void)
1006 __acquires(rq->lock)
1010 local_irq_disable();
1012 spin_lock(&rq->lock);
1017 #ifdef CONFIG_SCHED_HRTICK
1019 * Use HR-timers to deliver accurate preemption points.
1021 * Its all a bit involved since we cannot program an hrt while holding the
1022 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1025 * When we get rescheduled we reprogram the hrtick_timer outside of the
1031 * - enabled by features
1032 * - hrtimer is actually high res
1034 static inline int hrtick_enabled(struct rq *rq)
1036 if (!sched_feat(HRTICK))
1038 if (!cpu_active(cpu_of(rq)))
1040 return hrtimer_is_hres_active(&rq->hrtick_timer);
1043 static void hrtick_clear(struct rq *rq)
1045 if (hrtimer_active(&rq->hrtick_timer))
1046 hrtimer_cancel(&rq->hrtick_timer);
1050 * High-resolution timer tick.
1051 * Runs from hardirq context with interrupts disabled.
1053 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1055 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1057 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1059 spin_lock(&rq->lock);
1060 update_rq_clock(rq);
1061 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1062 spin_unlock(&rq->lock);
1064 return HRTIMER_NORESTART;
1069 * called from hardirq (IPI) context
1071 static void __hrtick_start(void *arg)
1073 struct rq *rq = arg;
1075 spin_lock(&rq->lock);
1076 hrtimer_restart(&rq->hrtick_timer);
1077 rq->hrtick_csd_pending = 0;
1078 spin_unlock(&rq->lock);
1082 * Called to set the hrtick timer state.
1084 * called with rq->lock held and irqs disabled
1086 static void hrtick_start(struct rq *rq, u64 delay)
1088 struct hrtimer *timer = &rq->hrtick_timer;
1089 ktime_t time = ktime_add_ns(timer->base->get_time(), delay);
1091 hrtimer_set_expires(timer, time);
1093 if (rq == this_rq()) {
1094 hrtimer_restart(timer);
1095 } else if (!rq->hrtick_csd_pending) {
1096 __smp_call_function_single(cpu_of(rq), &rq->hrtick_csd);
1097 rq->hrtick_csd_pending = 1;
1102 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1104 int cpu = (int)(long)hcpu;
1107 case CPU_UP_CANCELED:
1108 case CPU_UP_CANCELED_FROZEN:
1109 case CPU_DOWN_PREPARE:
1110 case CPU_DOWN_PREPARE_FROZEN:
1112 case CPU_DEAD_FROZEN:
1113 hrtick_clear(cpu_rq(cpu));
1120 static __init void init_hrtick(void)
1122 hotcpu_notifier(hotplug_hrtick, 0);
1126 * Called to set the hrtick timer state.
1128 * called with rq->lock held and irqs disabled
1130 static void hrtick_start(struct rq *rq, u64 delay)
1132 hrtimer_start(&rq->hrtick_timer, ns_to_ktime(delay), HRTIMER_MODE_REL);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq *rq)
1143 rq->hrtick_csd_pending = 0;
1145 rq->hrtick_csd.flags = 0;
1146 rq->hrtick_csd.func = __hrtick_start;
1147 rq->hrtick_csd.info = rq;
1150 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1151 rq->hrtick_timer.function = hrtick;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq *rq)
1158 static inline void init_rq_hrtick(struct rq *rq)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct *p)
1184 assert_spin_locked(&task_rq(p)->lock);
1186 if (unlikely(test_tsk_thread_flag(p, TIF_NEED_RESCHED)))
1189 set_tsk_thread_flag(p, TIF_NEED_RESCHED);
1192 if (cpu == smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p))
1198 smp_send_reschedule(cpu);
1201 static void resched_cpu(int cpu)
1203 struct rq *rq = cpu_rq(cpu);
1204 unsigned long flags;
1206 if (!spin_trylock_irqsave(&rq->lock, flags))
1208 resched_task(cpu_curr(cpu));
1209 spin_unlock_irqrestore(&rq->lock, flags);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu)
1225 struct rq *rq = cpu_rq(cpu);
1227 if (cpu == smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq->curr != rq->idle)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq->idle))
1250 smp_send_reschedule(cpu);
1252 #endif /* CONFIG_NO_HZ */
1254 #else /* !CONFIG_SMP */
1255 static void resched_task(struct task_struct *p)
1257 assert_spin_locked(&task_rq(p)->lock);
1258 set_tsk_need_resched(p);
1260 #endif /* CONFIG_SMP */
1262 #if BITS_PER_LONG == 32
1263 # define WMULT_CONST (~0UL)
1265 # define WMULT_CONST (1UL << 32)
1268 #define WMULT_SHIFT 32
1271 * Shift right and round:
1273 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1276 * delta *= weight / lw
1278 static unsigned long
1279 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1280 struct load_weight *lw)
1284 if (!lw->inv_weight) {
1285 if (BITS_PER_LONG > 32 && unlikely(lw->weight >= WMULT_CONST))
1288 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)
1292 tmp = (u64)delta_exec * weight;
1294 * Check whether we'd overflow the 64-bit multiplication:
1296 if (unlikely(tmp > WMULT_CONST))
1297 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1300 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1302 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1305 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1311 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1318 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1319 * of tasks with abnormal "nice" values across CPUs the contribution that
1320 * each task makes to its run queue's load is weighted according to its
1321 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1322 * scaled version of the new time slice allocation that they receive on time
1326 #define WEIGHT_IDLEPRIO 3
1327 #define WMULT_IDLEPRIO 1431655765
1330 * Nice levels are multiplicative, with a gentle 10% change for every
1331 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1332 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1333 * that remained on nice 0.
1335 * The "10% effect" is relative and cumulative: from _any_ nice level,
1336 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1337 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1338 * If a task goes up by ~10% and another task goes down by ~10% then
1339 * the relative distance between them is ~25%.)
1341 static const int prio_to_weight[40] = {
1342 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1343 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1344 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1345 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1346 /* 0 */ 1024, 820, 655, 526, 423,
1347 /* 5 */ 335, 272, 215, 172, 137,
1348 /* 10 */ 110, 87, 70, 56, 45,
1349 /* 15 */ 36, 29, 23, 18, 15,
1353 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1355 * In cases where the weight does not change often, we can use the
1356 * precalculated inverse to speed up arithmetics by turning divisions
1357 * into multiplications:
1359 static const u32 prio_to_wmult[40] = {
1360 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1361 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1362 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1363 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1364 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1365 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1366 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1367 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1370 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1373 * runqueue iterator, to support SMP load-balancing between different
1374 * scheduling classes, without having to expose their internal data
1375 * structures to the load-balancing proper:
1377 struct rq_iterator {
1379 struct task_struct *(*start)(void *);
1380 struct task_struct *(*next)(void *);
1384 static unsigned long
1385 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1386 unsigned long max_load_move, struct sched_domain *sd,
1387 enum cpu_idle_type idle, int *all_pinned,
1388 int *this_best_prio, struct rq_iterator *iterator);
1391 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1392 struct sched_domain *sd, enum cpu_idle_type idle,
1393 struct rq_iterator *iterator);
1396 /* Time spent by the tasks of the cpu accounting group executing in ... */
1397 enum cpuacct_stat_index {
1398 CPUACCT_STAT_USER, /* ... user mode */
1399 CPUACCT_STAT_SYSTEM, /* ... kernel mode */
1401 CPUACCT_STAT_NSTATS,
1404 #ifdef CONFIG_CGROUP_CPUACCT
1405 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1406 static void cpuacct_update_stats(struct task_struct *tsk,
1407 enum cpuacct_stat_index idx, cputime_t val);
1409 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1410 static inline void cpuacct_update_stats(struct task_struct *tsk,
1411 enum cpuacct_stat_index idx, cputime_t val) {}
1414 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1416 update_load_add(&rq->load, load);
1419 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1421 update_load_sub(&rq->load, load);
1424 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1425 typedef int (*tg_visitor)(struct task_group *, void *);
1428 * Iterate the full tree, calling @down when first entering a node and @up when
1429 * leaving it for the final time.
1431 static int walk_tg_tree(tg_visitor down, tg_visitor up, void *data)
1433 struct task_group *parent, *child;
1437 parent = &root_task_group;
1439 ret = (*down)(parent, data);
1442 list_for_each_entry_rcu(child, &parent->children, siblings) {
1449 ret = (*up)(parent, data);
1454 parent = parent->parent;
1463 static int tg_nop(struct task_group *tg, void *data)
1470 static unsigned long source_load(int cpu, int type);
1471 static unsigned long target_load(int cpu, int type);
1472 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1474 static unsigned long cpu_avg_load_per_task(int cpu)
1476 struct rq *rq = cpu_rq(cpu);
1477 unsigned long nr_running = ACCESS_ONCE(rq->nr_running);
1480 rq->avg_load_per_task = rq->load.weight / nr_running;
1482 rq->avg_load_per_task = 0;
1484 return rq->avg_load_per_task;
1487 #ifdef CONFIG_FAIR_GROUP_SCHED
1489 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1492 * Calculate and set the cpu's group shares.
1495 update_group_shares_cpu(struct task_group *tg, int cpu,
1496 unsigned long sd_shares, unsigned long sd_rq_weight)
1498 unsigned long shares;
1499 unsigned long rq_weight;
1504 rq_weight = tg->cfs_rq[cpu]->rq_weight;
1507 * \Sum shares * rq_weight
1508 * shares = -----------------------
1512 shares = (sd_shares * rq_weight) / sd_rq_weight;
1513 shares = clamp_t(unsigned long, shares, MIN_SHARES, MAX_SHARES);
1515 if (abs(shares - tg->se[cpu]->load.weight) >
1516 sysctl_sched_shares_thresh) {
1517 struct rq *rq = cpu_rq(cpu);
1518 unsigned long flags;
1520 spin_lock_irqsave(&rq->lock, flags);
1521 tg->cfs_rq[cpu]->shares = shares;
1523 __set_se_shares(tg->se[cpu], shares);
1524 spin_unlock_irqrestore(&rq->lock, flags);
1529 * Re-compute the task group their per cpu shares over the given domain.
1530 * This needs to be done in a bottom-up fashion because the rq weight of a
1531 * parent group depends on the shares of its child groups.
1533 static int tg_shares_up(struct task_group *tg, void *data)
1535 unsigned long weight, rq_weight = 0;
1536 unsigned long shares = 0;
1537 struct sched_domain *sd = data;
1540 for_each_cpu(i, sched_domain_span(sd)) {
1542 * If there are currently no tasks on the cpu pretend there
1543 * is one of average load so that when a new task gets to
1544 * run here it will not get delayed by group starvation.
1546 weight = tg->cfs_rq[i]->load.weight;
1548 weight = NICE_0_LOAD;
1550 tg->cfs_rq[i]->rq_weight = weight;
1551 rq_weight += weight;
1552 shares += tg->cfs_rq[i]->shares;
1555 if ((!shares && rq_weight) || shares > tg->shares)
1556 shares = tg->shares;
1558 if (!sd->parent || !(sd->parent->flags & SD_LOAD_BALANCE))
1559 shares = tg->shares;
1561 for_each_cpu(i, sched_domain_span(sd))
1562 update_group_shares_cpu(tg, i, shares, rq_weight);
1568 * Compute the cpu's hierarchical load factor for each task group.
1569 * This needs to be done in a top-down fashion because the load of a child
1570 * group is a fraction of its parents load.
1572 static int tg_load_down(struct task_group *tg, void *data)
1575 long cpu = (long)data;
1578 load = cpu_rq(cpu)->load.weight;
1580 load = tg->parent->cfs_rq[cpu]->h_load;
1581 load *= tg->cfs_rq[cpu]->shares;
1582 load /= tg->parent->cfs_rq[cpu]->load.weight + 1;
1585 tg->cfs_rq[cpu]->h_load = load;
1590 static void update_shares(struct sched_domain *sd)
1592 u64 now = cpu_clock(raw_smp_processor_id());
1593 s64 elapsed = now - sd->last_update;
1595 if (elapsed >= (s64)(u64)sysctl_sched_shares_ratelimit) {
1596 sd->last_update = now;
1597 walk_tg_tree(tg_nop, tg_shares_up, sd);
1601 static void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1603 spin_unlock(&rq->lock);
1605 spin_lock(&rq->lock);
1608 static void update_h_load(long cpu)
1610 walk_tg_tree(tg_load_down, tg_nop, (void *)cpu);
1615 static inline void update_shares(struct sched_domain *sd)
1619 static inline void update_shares_locked(struct rq *rq, struct sched_domain *sd)
1626 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1628 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
1629 __releases(this_rq->lock)
1630 __acquires(busiest->lock)
1631 __acquires(this_rq->lock)
1635 if (unlikely(!irqs_disabled())) {
1636 /* printk() doesn't work good under rq->lock */
1637 spin_unlock(&this_rq->lock);
1640 if (unlikely(!spin_trylock(&busiest->lock))) {
1641 if (busiest < this_rq) {
1642 spin_unlock(&this_rq->lock);
1643 spin_lock(&busiest->lock);
1644 spin_lock_nested(&this_rq->lock, SINGLE_DEPTH_NESTING);
1647 spin_lock_nested(&busiest->lock, SINGLE_DEPTH_NESTING);
1652 static inline void double_unlock_balance(struct rq *this_rq, struct rq *busiest)
1653 __releases(busiest->lock)
1655 spin_unlock(&busiest->lock);
1656 lock_set_subclass(&this_rq->lock.dep_map, 0, _RET_IP_);
1660 #ifdef CONFIG_FAIR_GROUP_SCHED
1661 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1664 cfs_rq->shares = shares;
1669 #include "sched_stats.h"
1670 #include "sched_idletask.c"
1671 #include "sched_fair.c"
1672 #include "sched_rt.c"
1673 #ifdef CONFIG_SCHED_DEBUG
1674 # include "sched_debug.c"
1677 #define sched_class_highest (&rt_sched_class)
1678 #define for_each_class(class) \
1679 for (class = sched_class_highest; class; class = class->next)
1681 static void inc_nr_running(struct rq *rq)
1686 static void dec_nr_running(struct rq *rq)
1691 static void set_load_weight(struct task_struct *p)
1693 if (task_has_rt_policy(p)) {
1694 p->se.load.weight = prio_to_weight[0] * 2;
1695 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1700 * SCHED_IDLE tasks get minimal weight:
1702 if (p->policy == SCHED_IDLE) {
1703 p->se.load.weight = WEIGHT_IDLEPRIO;
1704 p->se.load.inv_weight = WMULT_IDLEPRIO;
1708 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1709 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1712 static void update_avg(u64 *avg, u64 sample)
1714 s64 diff = sample - *avg;
1718 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1720 sched_info_queued(p);
1721 p->sched_class->enqueue_task(rq, p, wakeup);
1725 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1727 if (sleep && p->se.last_wakeup) {
1728 update_avg(&p->se.avg_overlap,
1729 p->se.sum_exec_runtime - p->se.last_wakeup);
1730 p->se.last_wakeup = 0;
1733 sched_info_dequeued(p);
1734 p->sched_class->dequeue_task(rq, p, sleep);
1739 * __normal_prio - return the priority that is based on the static prio
1741 static inline int __normal_prio(struct task_struct *p)
1743 return p->static_prio;
1747 * Calculate the expected normal priority: i.e. priority
1748 * without taking RT-inheritance into account. Might be
1749 * boosted by interactivity modifiers. Changes upon fork,
1750 * setprio syscalls, and whenever the interactivity
1751 * estimator recalculates.
1753 static inline int normal_prio(struct task_struct *p)
1757 if (task_has_rt_policy(p))
1758 prio = MAX_RT_PRIO-1 - p->rt_priority;
1760 prio = __normal_prio(p);
1765 * Calculate the current priority, i.e. the priority
1766 * taken into account by the scheduler. This value might
1767 * be boosted by RT tasks, or might be boosted by
1768 * interactivity modifiers. Will be RT if the task got
1769 * RT-boosted. If not then it returns p->normal_prio.
1771 static int effective_prio(struct task_struct *p)
1773 p->normal_prio = normal_prio(p);
1775 * If we are RT tasks or we were boosted to RT priority,
1776 * keep the priority unchanged. Otherwise, update priority
1777 * to the normal priority:
1779 if (!rt_prio(p->prio))
1780 return p->normal_prio;
1785 * activate_task - move a task to the runqueue.
1787 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1789 if (task_contributes_to_load(p))
1790 rq->nr_uninterruptible--;
1792 enqueue_task(rq, p, wakeup);
1797 * deactivate_task - remove a task from the runqueue.
1799 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1801 if (task_contributes_to_load(p))
1802 rq->nr_uninterruptible++;
1804 dequeue_task(rq, p, sleep);
1809 * task_curr - is this task currently executing on a CPU?
1810 * @p: the task in question.
1812 inline int task_curr(const struct task_struct *p)
1814 return cpu_curr(task_cpu(p)) == p;
1817 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1819 set_task_rq(p, cpu);
1822 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1823 * successfuly executed on another CPU. We must ensure that updates of
1824 * per-task data have been completed by this moment.
1827 task_thread_info(p)->cpu = cpu;
1831 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1832 const struct sched_class *prev_class,
1833 int oldprio, int running)
1835 if (prev_class != p->sched_class) {
1836 if (prev_class->switched_from)
1837 prev_class->switched_from(rq, p, running);
1838 p->sched_class->switched_to(rq, p, running);
1840 p->sched_class->prio_changed(rq, p, oldprio, running);
1845 /* Used instead of source_load when we know the type == 0 */
1846 static unsigned long weighted_cpuload(const int cpu)
1848 return cpu_rq(cpu)->load.weight;
1852 * Is this task likely cache-hot:
1855 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
1860 * Buddy candidates are cache hot:
1862 if (sched_feat(CACHE_HOT_BUDDY) &&
1863 (&p->se == cfs_rq_of(&p->se)->next ||
1864 &p->se == cfs_rq_of(&p->se)->last))
1867 if (p->sched_class != &fair_sched_class)
1870 if (sysctl_sched_migration_cost == -1)
1872 if (sysctl_sched_migration_cost == 0)
1875 delta = now - p->se.exec_start;
1877 return delta < (s64)sysctl_sched_migration_cost;
1881 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
1883 int old_cpu = task_cpu(p);
1884 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
1885 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
1886 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
1889 clock_offset = old_rq->clock - new_rq->clock;
1891 trace_sched_migrate_task(p, task_cpu(p), new_cpu);
1893 #ifdef CONFIG_SCHEDSTATS
1894 if (p->se.wait_start)
1895 p->se.wait_start -= clock_offset;
1896 if (p->se.sleep_start)
1897 p->se.sleep_start -= clock_offset;
1898 if (p->se.block_start)
1899 p->se.block_start -= clock_offset;
1900 if (old_cpu != new_cpu) {
1901 schedstat_inc(p, se.nr_migrations);
1902 if (task_hot(p, old_rq->clock, NULL))
1903 schedstat_inc(p, se.nr_forced2_migrations);
1906 p->se.vruntime -= old_cfsrq->min_vruntime -
1907 new_cfsrq->min_vruntime;
1909 __set_task_cpu(p, new_cpu);
1912 struct migration_req {
1913 struct list_head list;
1915 struct task_struct *task;
1918 struct completion done;
1922 * The task's runqueue lock must be held.
1923 * Returns true if you have to wait for migration thread.
1926 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
1928 struct rq *rq = task_rq(p);
1931 * If the task is not on a runqueue (and not running), then
1932 * it is sufficient to simply update the task's cpu field.
1934 if (!p->se.on_rq && !task_running(rq, p)) {
1935 set_task_cpu(p, dest_cpu);
1939 init_completion(&req->done);
1941 req->dest_cpu = dest_cpu;
1942 list_add(&req->list, &rq->migration_queue);
1948 * wait_task_inactive - wait for a thread to unschedule.
1950 * If @match_state is nonzero, it's the @p->state value just checked and
1951 * not expected to change. If it changes, i.e. @p might have woken up,
1952 * then return zero. When we succeed in waiting for @p to be off its CPU,
1953 * we return a positive number (its total switch count). If a second call
1954 * a short while later returns the same number, the caller can be sure that
1955 * @p has remained unscheduled the whole time.
1957 * The caller must ensure that the task *will* unschedule sometime soon,
1958 * else this function might spin for a *long* time. This function can't
1959 * be called with interrupts off, or it may introduce deadlock with
1960 * smp_call_function() if an IPI is sent by the same process we are
1961 * waiting to become inactive.
1963 unsigned long wait_task_inactive(struct task_struct *p, long match_state)
1965 unsigned long flags;
1972 * We do the initial early heuristics without holding
1973 * any task-queue locks at all. We'll only try to get
1974 * the runqueue lock when things look like they will
1980 * If the task is actively running on another CPU
1981 * still, just relax and busy-wait without holding
1984 * NOTE! Since we don't hold any locks, it's not
1985 * even sure that "rq" stays as the right runqueue!
1986 * But we don't care, since "task_running()" will
1987 * return false if the runqueue has changed and p
1988 * is actually now running somewhere else!
1990 while (task_running(rq, p)) {
1991 if (match_state && unlikely(p->state != match_state))
1997 * Ok, time to look more closely! We need the rq
1998 * lock now, to be *sure*. If we're wrong, we'll
1999 * just go back and repeat.
2001 rq = task_rq_lock(p, &flags);
2002 trace_sched_wait_task(rq, p);
2003 running = task_running(rq, p);
2004 on_rq = p->se.on_rq;
2006 if (!match_state || p->state == match_state)
2007 ncsw = p->nvcsw | LONG_MIN; /* sets MSB */
2008 task_rq_unlock(rq, &flags);
2011 * If it changed from the expected state, bail out now.
2013 if (unlikely(!ncsw))
2017 * Was it really running after all now that we
2018 * checked with the proper locks actually held?
2020 * Oops. Go back and try again..
2022 if (unlikely(running)) {
2028 * It's not enough that it's not actively running,
2029 * it must be off the runqueue _entirely_, and not
2032 * So if it wa still runnable (but just not actively
2033 * running right now), it's preempted, and we should
2034 * yield - it could be a while.
2036 if (unlikely(on_rq)) {
2037 schedule_timeout_uninterruptible(1);
2042 * Ahh, all good. It wasn't running, and it wasn't
2043 * runnable, which means that it will never become
2044 * running in the future either. We're all done!
2053 * kick_process - kick a running thread to enter/exit the kernel
2054 * @p: the to-be-kicked thread
2056 * Cause a process which is running on another CPU to enter
2057 * kernel-mode, without any delay. (to get signals handled.)
2059 * NOTE: this function doesnt have to take the runqueue lock,
2060 * because all it wants to ensure is that the remote task enters
2061 * the kernel. If the IPI races and the task has been migrated
2062 * to another CPU then no harm is done and the purpose has been
2065 void kick_process(struct task_struct *p)
2071 if ((cpu != smp_processor_id()) && task_curr(p))
2072 smp_send_reschedule(cpu);
2077 * Return a low guess at the load of a migration-source cpu weighted
2078 * according to the scheduling class and "nice" value.
2080 * We want to under-estimate the load of migration sources, to
2081 * balance conservatively.
2083 static unsigned long source_load(int cpu, int type)
2085 struct rq *rq = cpu_rq(cpu);
2086 unsigned long total = weighted_cpuload(cpu);
2088 if (type == 0 || !sched_feat(LB_BIAS))
2091 return min(rq->cpu_load[type-1], total);
2095 * Return a high guess at the load of a migration-target cpu weighted
2096 * according to the scheduling class and "nice" value.
2098 static unsigned long target_load(int cpu, int type)
2100 struct rq *rq = cpu_rq(cpu);
2101 unsigned long total = weighted_cpuload(cpu);
2103 if (type == 0 || !sched_feat(LB_BIAS))
2106 return max(rq->cpu_load[type-1], total);
2110 * find_idlest_group finds and returns the least busy CPU group within the
2113 static struct sched_group *
2114 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2116 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2117 unsigned long min_load = ULONG_MAX, this_load = 0;
2118 int load_idx = sd->forkexec_idx;
2119 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2122 unsigned long load, avg_load;
2126 /* Skip over this group if it has no CPUs allowed */
2127 if (!cpumask_intersects(sched_group_cpus(group),
2131 local_group = cpumask_test_cpu(this_cpu,
2132 sched_group_cpus(group));
2134 /* Tally up the load of all CPUs in the group */
2137 for_each_cpu(i, sched_group_cpus(group)) {
2138 /* Bias balancing toward cpus of our domain */
2140 load = source_load(i, load_idx);
2142 load = target_load(i, load_idx);
2147 /* Adjust by relative CPU power of the group */
2148 avg_load = sg_div_cpu_power(group,
2149 avg_load * SCHED_LOAD_SCALE);
2152 this_load = avg_load;
2154 } else if (avg_load < min_load) {
2155 min_load = avg_load;
2158 } while (group = group->next, group != sd->groups);
2160 if (!idlest || 100*this_load < imbalance*min_load)
2166 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2169 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu)
2171 unsigned long load, min_load = ULONG_MAX;
2175 /* Traverse only the allowed CPUs */
2176 for_each_cpu_and(i, sched_group_cpus(group), &p->cpus_allowed) {
2177 load = weighted_cpuload(i);
2179 if (load < min_load || (load == min_load && i == this_cpu)) {
2189 * sched_balance_self: balance the current task (running on cpu) in domains
2190 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2193 * Balance, ie. select the least loaded group.
2195 * Returns the target CPU number, or the same CPU if no balancing is needed.
2197 * preempt must be disabled.
2199 static int sched_balance_self(int cpu, int flag)
2201 struct task_struct *t = current;
2202 struct sched_domain *tmp, *sd = NULL;
2204 for_each_domain(cpu, tmp) {
2206 * If power savings logic is enabled for a domain, stop there.
2208 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2210 if (tmp->flags & flag)
2218 struct sched_group *group;
2219 int new_cpu, weight;
2221 if (!(sd->flags & flag)) {
2226 group = find_idlest_group(sd, t, cpu);
2232 new_cpu = find_idlest_cpu(group, t, cpu);
2233 if (new_cpu == -1 || new_cpu == cpu) {
2234 /* Now try balancing at a lower domain level of cpu */
2239 /* Now try balancing at a lower domain level of new_cpu */
2241 weight = cpumask_weight(sched_domain_span(sd));
2243 for_each_domain(cpu, tmp) {
2244 if (weight <= cpumask_weight(sched_domain_span(tmp)))
2246 if (tmp->flags & flag)
2249 /* while loop will break here if sd == NULL */
2255 #endif /* CONFIG_SMP */
2258 * try_to_wake_up - wake up a thread
2259 * @p: the to-be-woken-up thread
2260 * @state: the mask of task states that can be woken
2261 * @sync: do a synchronous wakeup?
2263 * Put it on the run-queue if it's not already there. The "current"
2264 * thread is always on the run-queue (except when the actual
2265 * re-schedule is in progress), and as such you're allowed to do
2266 * the simpler "current->state = TASK_RUNNING" to mark yourself
2267 * runnable without the overhead of this.
2269 * returns failure only if the task is already active.
2271 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2273 int cpu, orig_cpu, this_cpu, success = 0;
2274 unsigned long flags;
2278 if (!sched_feat(SYNC_WAKEUPS))
2282 if (current->se.avg_overlap < sysctl_sched_migration_cost &&
2283 p->se.avg_overlap < sysctl_sched_migration_cost)
2286 if (current->se.avg_overlap >= sysctl_sched_migration_cost ||
2287 p->se.avg_overlap >= sysctl_sched_migration_cost)
2292 if (sched_feat(LB_WAKEUP_UPDATE)) {
2293 struct sched_domain *sd;
2295 this_cpu = raw_smp_processor_id();
2298 for_each_domain(this_cpu, sd) {
2299 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2308 rq = task_rq_lock(p, &flags);
2309 update_rq_clock(rq);
2310 old_state = p->state;
2311 if (!(old_state & state))
2319 this_cpu = smp_processor_id();
2322 if (unlikely(task_running(rq, p)))
2325 cpu = p->sched_class->select_task_rq(p, sync);
2326 if (cpu != orig_cpu) {
2327 set_task_cpu(p, cpu);
2328 task_rq_unlock(rq, &flags);
2329 /* might preempt at this point */
2330 rq = task_rq_lock(p, &flags);
2331 old_state = p->state;
2332 if (!(old_state & state))
2337 this_cpu = smp_processor_id();
2341 #ifdef CONFIG_SCHEDSTATS
2342 schedstat_inc(rq, ttwu_count);
2343 if (cpu == this_cpu)
2344 schedstat_inc(rq, ttwu_local);
2346 struct sched_domain *sd;
2347 for_each_domain(this_cpu, sd) {
2348 if (cpumask_test_cpu(cpu, sched_domain_span(sd))) {
2349 schedstat_inc(sd, ttwu_wake_remote);
2354 #endif /* CONFIG_SCHEDSTATS */
2357 #endif /* CONFIG_SMP */
2358 schedstat_inc(p, se.nr_wakeups);
2360 schedstat_inc(p, se.nr_wakeups_sync);
2361 if (orig_cpu != cpu)
2362 schedstat_inc(p, se.nr_wakeups_migrate);
2363 if (cpu == this_cpu)
2364 schedstat_inc(p, se.nr_wakeups_local);
2366 schedstat_inc(p, se.nr_wakeups_remote);
2367 activate_task(rq, p, 1);
2371 trace_sched_wakeup(rq, p, success);
2372 check_preempt_curr(rq, p, sync);
2374 p->state = TASK_RUNNING;
2376 if (p->sched_class->task_wake_up)
2377 p->sched_class->task_wake_up(rq, p);
2380 current->se.last_wakeup = current->se.sum_exec_runtime;
2382 task_rq_unlock(rq, &flags);
2387 int wake_up_process(struct task_struct *p)
2389 return try_to_wake_up(p, TASK_ALL, 0);
2391 EXPORT_SYMBOL(wake_up_process);
2393 int wake_up_state(struct task_struct *p, unsigned int state)
2395 return try_to_wake_up(p, state, 0);
2399 * Perform scheduler related setup for a newly forked process p.
2400 * p is forked by current.
2402 * __sched_fork() is basic setup used by init_idle() too:
2404 static void __sched_fork(struct task_struct *p)
2406 p->se.exec_start = 0;
2407 p->se.sum_exec_runtime = 0;
2408 p->se.prev_sum_exec_runtime = 0;
2409 p->se.last_wakeup = 0;
2410 p->se.avg_overlap = 0;
2412 #ifdef CONFIG_SCHEDSTATS
2413 p->se.wait_start = 0;
2414 p->se.sum_sleep_runtime = 0;
2415 p->se.sleep_start = 0;
2416 p->se.block_start = 0;
2417 p->se.sleep_max = 0;
2418 p->se.block_max = 0;
2420 p->se.slice_max = 0;
2424 INIT_LIST_HEAD(&p->rt.run_list);
2426 INIT_LIST_HEAD(&p->se.group_node);
2428 #ifdef CONFIG_PREEMPT_NOTIFIERS
2429 INIT_HLIST_HEAD(&p->preempt_notifiers);
2433 * We mark the process as running here, but have not actually
2434 * inserted it onto the runqueue yet. This guarantees that
2435 * nobody will actually run it, and a signal or other external
2436 * event cannot wake it up and insert it on the runqueue either.
2438 p->state = TASK_RUNNING;
2442 * fork()/clone()-time setup:
2444 void sched_fork(struct task_struct *p, int clone_flags)
2446 int cpu = get_cpu();
2451 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2453 set_task_cpu(p, cpu);
2456 * Make sure we do not leak PI boosting priority to the child:
2458 p->prio = current->normal_prio;
2459 if (!rt_prio(p->prio))
2460 p->sched_class = &fair_sched_class;
2462 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2463 if (likely(sched_info_on()))
2464 memset(&p->sched_info, 0, sizeof(p->sched_info));
2466 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2469 #ifdef CONFIG_PREEMPT
2470 /* Want to start with kernel preemption disabled. */
2471 task_thread_info(p)->preempt_count = 1;
2477 * wake_up_new_task - wake up a newly created task for the first time.
2479 * This function will do some initial scheduler statistics housekeeping
2480 * that must be done for every newly created context, then puts the task
2481 * on the runqueue and wakes it.
2483 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2485 unsigned long flags;
2488 rq = task_rq_lock(p, &flags);
2489 BUG_ON(p->state != TASK_RUNNING);
2490 update_rq_clock(rq);
2492 p->prio = effective_prio(p);
2494 if (!p->sched_class->task_new || !current->se.on_rq) {
2495 activate_task(rq, p, 0);
2498 * Let the scheduling class do new task startup
2499 * management (if any):
2501 p->sched_class->task_new(rq, p);
2504 trace_sched_wakeup_new(rq, p, 1);
2505 check_preempt_curr(rq, p, 0);
2507 if (p->sched_class->task_wake_up)
2508 p->sched_class->task_wake_up(rq, p);
2510 task_rq_unlock(rq, &flags);
2513 #ifdef CONFIG_PREEMPT_NOTIFIERS
2516 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2517 * @notifier: notifier struct to register
2519 void preempt_notifier_register(struct preempt_notifier *notifier)
2521 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2523 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2526 * preempt_notifier_unregister - no longer interested in preemption notifications
2527 * @notifier: notifier struct to unregister
2529 * This is safe to call from within a preemption notifier.
2531 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2533 hlist_del(¬ifier->link);
2535 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2537 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2539 struct preempt_notifier *notifier;
2540 struct hlist_node *node;
2542 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2543 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2547 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2548 struct task_struct *next)
2550 struct preempt_notifier *notifier;
2551 struct hlist_node *node;
2553 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2554 notifier->ops->sched_out(notifier, next);
2557 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2559 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2564 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2565 struct task_struct *next)
2569 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2572 * prepare_task_switch - prepare to switch tasks
2573 * @rq: the runqueue preparing to switch
2574 * @prev: the current task that is being switched out
2575 * @next: the task we are going to switch to.
2577 * This is called with the rq lock held and interrupts off. It must
2578 * be paired with a subsequent finish_task_switch after the context
2581 * prepare_task_switch sets up locking and calls architecture specific
2585 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2586 struct task_struct *next)
2588 fire_sched_out_preempt_notifiers(prev, next);
2589 prepare_lock_switch(rq, next);
2590 prepare_arch_switch(next);
2594 * finish_task_switch - clean up after a task-switch
2595 * @rq: runqueue associated with task-switch
2596 * @prev: the thread we just switched away from.
2598 * finish_task_switch must be called after the context switch, paired
2599 * with a prepare_task_switch call before the context switch.
2600 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2601 * and do any other architecture-specific cleanup actions.
2603 * Note that we may have delayed dropping an mm in context_switch(). If
2604 * so, we finish that here outside of the runqueue lock. (Doing it
2605 * with the lock held can cause deadlocks; see schedule() for
2608 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2609 __releases(rq->lock)
2611 struct mm_struct *mm = rq->prev_mm;
2617 * A task struct has one reference for the use as "current".
2618 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2619 * schedule one last time. The schedule call will never return, and
2620 * the scheduled task must drop that reference.
2621 * The test for TASK_DEAD must occur while the runqueue locks are
2622 * still held, otherwise prev could be scheduled on another cpu, die
2623 * there before we look at prev->state, and then the reference would
2625 * Manfred Spraul <manfred@colorfullife.com>
2627 prev_state = prev->state;
2628 finish_arch_switch(prev);
2629 finish_lock_switch(rq, prev);
2631 if (current->sched_class->post_schedule)
2632 current->sched_class->post_schedule(rq);
2635 fire_sched_in_preempt_notifiers(current);
2638 if (unlikely(prev_state == TASK_DEAD)) {
2640 * Remove function-return probe instances associated with this
2641 * task and put them back on the free list.
2643 kprobe_flush_task(prev);
2644 put_task_struct(prev);
2649 * schedule_tail - first thing a freshly forked thread must call.
2650 * @prev: the thread we just switched away from.
2652 asmlinkage void schedule_tail(struct task_struct *prev)
2653 __releases(rq->lock)
2655 struct rq *rq = this_rq();
2657 finish_task_switch(rq, prev);
2658 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2659 /* In this case, finish_task_switch does not reenable preemption */
2662 if (current->set_child_tid)
2663 put_user(task_pid_vnr(current), current->set_child_tid);
2667 * context_switch - switch to the new MM and the new
2668 * thread's register state.
2671 context_switch(struct rq *rq, struct task_struct *prev,
2672 struct task_struct *next)
2674 struct mm_struct *mm, *oldmm;
2676 prepare_task_switch(rq, prev, next);
2677 trace_sched_switch(rq, prev, next);
2679 oldmm = prev->active_mm;
2681 * For paravirt, this is coupled with an exit in switch_to to
2682 * combine the page table reload and the switch backend into
2685 arch_enter_lazy_cpu_mode();
2687 if (unlikely(!mm)) {
2688 next->active_mm = oldmm;
2689 atomic_inc(&oldmm->mm_count);
2690 enter_lazy_tlb(oldmm, next);
2692 switch_mm(oldmm, mm, next);
2694 if (unlikely(!prev->mm)) {
2695 prev->active_mm = NULL;
2696 rq->prev_mm = oldmm;
2699 * Since the runqueue lock will be released by the next
2700 * task (which is an invalid locking op but in the case
2701 * of the scheduler it's an obvious special-case), so we
2702 * do an early lockdep release here:
2704 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2705 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2708 /* Here we just switch the register state and the stack. */
2709 switch_to(prev, next, prev);
2713 * this_rq must be evaluated again because prev may have moved
2714 * CPUs since it called schedule(), thus the 'rq' on its stack
2715 * frame will be invalid.
2717 finish_task_switch(this_rq(), prev);
2721 * nr_running, nr_uninterruptible and nr_context_switches:
2723 * externally visible scheduler statistics: current number of runnable
2724 * threads, current number of uninterruptible-sleeping threads, total
2725 * number of context switches performed since bootup.
2727 unsigned long nr_running(void)
2729 unsigned long i, sum = 0;
2731 for_each_online_cpu(i)
2732 sum += cpu_rq(i)->nr_running;
2737 unsigned long nr_uninterruptible(void)
2739 unsigned long i, sum = 0;
2741 for_each_possible_cpu(i)
2742 sum += cpu_rq(i)->nr_uninterruptible;
2745 * Since we read the counters lockless, it might be slightly
2746 * inaccurate. Do not allow it to go below zero though:
2748 if (unlikely((long)sum < 0))
2754 unsigned long long nr_context_switches(void)
2757 unsigned long long sum = 0;
2759 for_each_possible_cpu(i)
2760 sum += cpu_rq(i)->nr_switches;
2765 unsigned long nr_iowait(void)
2767 unsigned long i, sum = 0;
2769 for_each_possible_cpu(i)
2770 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2775 unsigned long nr_active(void)
2777 unsigned long i, running = 0, uninterruptible = 0;
2779 for_each_online_cpu(i) {
2780 running += cpu_rq(i)->nr_running;
2781 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2784 if (unlikely((long)uninterruptible < 0))
2785 uninterruptible = 0;
2787 return running + uninterruptible;
2791 * Update rq->cpu_load[] statistics. This function is usually called every
2792 * scheduler tick (TICK_NSEC).
2794 static void update_cpu_load(struct rq *this_rq)
2796 unsigned long this_load = this_rq->load.weight;
2799 this_rq->nr_load_updates++;
2801 /* Update our load: */
2802 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2803 unsigned long old_load, new_load;
2805 /* scale is effectively 1 << i now, and >> i divides by scale */
2807 old_load = this_rq->cpu_load[i];
2808 new_load = this_load;
2810 * Round up the averaging division if load is increasing. This
2811 * prevents us from getting stuck on 9 if the load is 10, for
2814 if (new_load > old_load)
2815 new_load += scale-1;
2816 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2823 * double_rq_lock - safely lock two runqueues
2825 * Note this does not disable interrupts like task_rq_lock,
2826 * you need to do so manually before calling.
2828 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2829 __acquires(rq1->lock)
2830 __acquires(rq2->lock)
2832 BUG_ON(!irqs_disabled());
2834 spin_lock(&rq1->lock);
2835 __acquire(rq2->lock); /* Fake it out ;) */
2838 spin_lock(&rq1->lock);
2839 spin_lock_nested(&rq2->lock, SINGLE_DEPTH_NESTING);
2841 spin_lock(&rq2->lock);
2842 spin_lock_nested(&rq1->lock, SINGLE_DEPTH_NESTING);
2845 update_rq_clock(rq1);
2846 update_rq_clock(rq2);
2850 * double_rq_unlock - safely unlock two runqueues
2852 * Note this does not restore interrupts like task_rq_unlock,
2853 * you need to do so manually after calling.
2855 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2856 __releases(rq1->lock)
2857 __releases(rq2->lock)
2859 spin_unlock(&rq1->lock);
2861 spin_unlock(&rq2->lock);
2863 __release(rq2->lock);
2867 * If dest_cpu is allowed for this process, migrate the task to it.
2868 * This is accomplished by forcing the cpu_allowed mask to only
2869 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2870 * the cpu_allowed mask is restored.
2872 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
2874 struct migration_req req;
2875 unsigned long flags;
2878 rq = task_rq_lock(p, &flags);
2879 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed)
2880 || unlikely(!cpu_active(dest_cpu)))
2883 /* force the process onto the specified CPU */
2884 if (migrate_task(p, dest_cpu, &req)) {
2885 /* Need to wait for migration thread (might exit: take ref). */
2886 struct task_struct *mt = rq->migration_thread;
2888 get_task_struct(mt);
2889 task_rq_unlock(rq, &flags);
2890 wake_up_process(mt);
2891 put_task_struct(mt);
2892 wait_for_completion(&req.done);
2897 task_rq_unlock(rq, &flags);
2901 * sched_exec - execve() is a valuable balancing opportunity, because at
2902 * this point the task has the smallest effective memory and cache footprint.
2904 void sched_exec(void)
2906 int new_cpu, this_cpu = get_cpu();
2907 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
2909 if (new_cpu != this_cpu)
2910 sched_migrate_task(current, new_cpu);
2914 * pull_task - move a task from a remote runqueue to the local runqueue.
2915 * Both runqueues must be locked.
2917 static void pull_task(struct rq *src_rq, struct task_struct *p,
2918 struct rq *this_rq, int this_cpu)
2920 deactivate_task(src_rq, p, 0);
2921 set_task_cpu(p, this_cpu);
2922 activate_task(this_rq, p, 0);
2924 * Note that idle threads have a prio of MAX_PRIO, for this test
2925 * to be always true for them.
2927 check_preempt_curr(this_rq, p, 0);
2931 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2934 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
2935 struct sched_domain *sd, enum cpu_idle_type idle,
2939 * We do not migrate tasks that are:
2940 * 1) running (obviously), or
2941 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2942 * 3) are cache-hot on their current CPU.
2944 if (!cpumask_test_cpu(this_cpu, &p->cpus_allowed)) {
2945 schedstat_inc(p, se.nr_failed_migrations_affine);
2950 if (task_running(rq, p)) {
2951 schedstat_inc(p, se.nr_failed_migrations_running);
2956 * Aggressive migration if:
2957 * 1) task is cache cold, or
2958 * 2) too many balance attempts have failed.
2961 if (!task_hot(p, rq->clock, sd) ||
2962 sd->nr_balance_failed > sd->cache_nice_tries) {
2963 #ifdef CONFIG_SCHEDSTATS
2964 if (task_hot(p, rq->clock, sd)) {
2965 schedstat_inc(sd, lb_hot_gained[idle]);
2966 schedstat_inc(p, se.nr_forced_migrations);
2972 if (task_hot(p, rq->clock, sd)) {
2973 schedstat_inc(p, se.nr_failed_migrations_hot);
2979 static unsigned long
2980 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
2981 unsigned long max_load_move, struct sched_domain *sd,
2982 enum cpu_idle_type idle, int *all_pinned,
2983 int *this_best_prio, struct rq_iterator *iterator)
2985 int loops = 0, pulled = 0, pinned = 0;
2986 struct task_struct *p;
2987 long rem_load_move = max_load_move;
2989 if (max_load_move == 0)
2995 * Start the load-balancing iterator:
2997 p = iterator->start(iterator->arg);
2999 if (!p || loops++ > sysctl_sched_nr_migrate)
3002 if ((p->se.load.weight >> 1) > rem_load_move ||
3003 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3004 p = iterator->next(iterator->arg);
3008 pull_task(busiest, p, this_rq, this_cpu);
3010 rem_load_move -= p->se.load.weight;
3013 * We only want to steal up to the prescribed amount of weighted load.
3015 if (rem_load_move > 0) {
3016 if (p->prio < *this_best_prio)
3017 *this_best_prio = p->prio;
3018 p = iterator->next(iterator->arg);
3023 * Right now, this is one of only two places pull_task() is called,
3024 * so we can safely collect pull_task() stats here rather than
3025 * inside pull_task().
3027 schedstat_add(sd, lb_gained[idle], pulled);
3030 *all_pinned = pinned;
3032 return max_load_move - rem_load_move;
3036 * move_tasks tries to move up to max_load_move weighted load from busiest to
3037 * this_rq, as part of a balancing operation within domain "sd".
3038 * Returns 1 if successful and 0 otherwise.
3040 * Called with both runqueues locked.
3042 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3043 unsigned long max_load_move,
3044 struct sched_domain *sd, enum cpu_idle_type idle,
3047 const struct sched_class *class = sched_class_highest;
3048 unsigned long total_load_moved = 0;
3049 int this_best_prio = this_rq->curr->prio;
3053 class->load_balance(this_rq, this_cpu, busiest,
3054 max_load_move - total_load_moved,
3055 sd, idle, all_pinned, &this_best_prio);
3056 class = class->next;
3058 if (idle == CPU_NEWLY_IDLE && this_rq->nr_running)
3061 } while (class && max_load_move > total_load_moved);
3063 return total_load_moved > 0;
3067 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3068 struct sched_domain *sd, enum cpu_idle_type idle,
3069 struct rq_iterator *iterator)
3071 struct task_struct *p = iterator->start(iterator->arg);
3075 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3076 pull_task(busiest, p, this_rq, this_cpu);
3078 * Right now, this is only the second place pull_task()
3079 * is called, so we can safely collect pull_task()
3080 * stats here rather than inside pull_task().
3082 schedstat_inc(sd, lb_gained[idle]);
3086 p = iterator->next(iterator->arg);
3093 * move_one_task tries to move exactly one task from busiest to this_rq, as
3094 * part of active balancing operations within "domain".
3095 * Returns 1 if successful and 0 otherwise.
3097 * Called with both runqueues locked.
3099 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3100 struct sched_domain *sd, enum cpu_idle_type idle)
3102 const struct sched_class *class;
3104 for (class = sched_class_highest; class; class = class->next)
3105 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3112 * find_busiest_group finds and returns the busiest CPU group within the
3113 * domain. It calculates and returns the amount of weighted load which
3114 * should be moved to restore balance via the imbalance parameter.
3116 static struct sched_group *
3117 find_busiest_group(struct sched_domain *sd, int this_cpu,
3118 unsigned long *imbalance, enum cpu_idle_type idle,
3119 int *sd_idle, const struct cpumask *cpus, int *balance)
3121 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3122 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3123 unsigned long max_pull;
3124 unsigned long busiest_load_per_task, busiest_nr_running;
3125 unsigned long this_load_per_task, this_nr_running;
3126 int load_idx, group_imb = 0;
3127 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3128 int power_savings_balance = 1;
3129 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3130 unsigned long min_nr_running = ULONG_MAX;
3131 struct sched_group *group_min = NULL, *group_leader = NULL;
3134 max_load = this_load = total_load = total_pwr = 0;
3135 busiest_load_per_task = busiest_nr_running = 0;
3136 this_load_per_task = this_nr_running = 0;
3138 if (idle == CPU_NOT_IDLE)
3139 load_idx = sd->busy_idx;
3140 else if (idle == CPU_NEWLY_IDLE)
3141 load_idx = sd->newidle_idx;
3143 load_idx = sd->idle_idx;
3146 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3149 int __group_imb = 0;
3150 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3151 unsigned long sum_nr_running, sum_weighted_load;
3152 unsigned long sum_avg_load_per_task;
3153 unsigned long avg_load_per_task;
3155 local_group = cpumask_test_cpu(this_cpu,
3156 sched_group_cpus(group));
3159 balance_cpu = cpumask_first(sched_group_cpus(group));
3161 /* Tally up the load of all CPUs in the group */
3162 sum_weighted_load = sum_nr_running = avg_load = 0;
3163 sum_avg_load_per_task = avg_load_per_task = 0;
3166 min_cpu_load = ~0UL;
3168 for_each_cpu_and(i, sched_group_cpus(group), cpus) {
3169 struct rq *rq = cpu_rq(i);
3171 if (*sd_idle && rq->nr_running)
3174 /* Bias balancing toward cpus of our domain */
3176 if (idle_cpu(i) && !first_idle_cpu) {
3181 load = target_load(i, load_idx);
3183 load = source_load(i, load_idx);
3184 if (load > max_cpu_load)
3185 max_cpu_load = load;
3186 if (min_cpu_load > load)
3187 min_cpu_load = load;
3191 sum_nr_running += rq->nr_running;
3192 sum_weighted_load += weighted_cpuload(i);
3194 sum_avg_load_per_task += cpu_avg_load_per_task(i);
3198 * First idle cpu or the first cpu(busiest) in this sched group
3199 * is eligible for doing load balancing at this and above
3200 * domains. In the newly idle case, we will allow all the cpu's
3201 * to do the newly idle load balance.
3203 if (idle != CPU_NEWLY_IDLE && local_group &&
3204 balance_cpu != this_cpu && balance) {
3209 total_load += avg_load;
3210 total_pwr += group->__cpu_power;
3212 /* Adjust by relative CPU power of the group */
3213 avg_load = sg_div_cpu_power(group,
3214 avg_load * SCHED_LOAD_SCALE);
3218 * Consider the group unbalanced when the imbalance is larger
3219 * than the average weight of two tasks.
3221 * APZ: with cgroup the avg task weight can vary wildly and
3222 * might not be a suitable number - should we keep a
3223 * normalized nr_running number somewhere that negates
3226 avg_load_per_task = sg_div_cpu_power(group,
3227 sum_avg_load_per_task * SCHED_LOAD_SCALE);
3229 if ((max_cpu_load - min_cpu_load) > 2*avg_load_per_task)
3232 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3235 this_load = avg_load;
3237 this_nr_running = sum_nr_running;
3238 this_load_per_task = sum_weighted_load;
3239 } else if (avg_load > max_load &&
3240 (sum_nr_running > group_capacity || __group_imb)) {
3241 max_load = avg_load;
3243 busiest_nr_running = sum_nr_running;
3244 busiest_load_per_task = sum_weighted_load;
3245 group_imb = __group_imb;
3248 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3250 * Busy processors will not participate in power savings
3253 if (idle == CPU_NOT_IDLE ||
3254 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3258 * If the local group is idle or completely loaded
3259 * no need to do power savings balance at this domain
3261 if (local_group && (this_nr_running >= group_capacity ||
3263 power_savings_balance = 0;
3266 * If a group is already running at full capacity or idle,
3267 * don't include that group in power savings calculations
3269 if (!power_savings_balance || sum_nr_running >= group_capacity
3274 * Calculate the group which has the least non-idle load.
3275 * This is the group from where we need to pick up the load
3278 if ((sum_nr_running < min_nr_running) ||
3279 (sum_nr_running == min_nr_running &&
3280 cpumask_first(sched_group_cpus(group)) >
3281 cpumask_first(sched_group_cpus(group_min)))) {
3283 min_nr_running = sum_nr_running;
3284 min_load_per_task = sum_weighted_load /
3289 * Calculate the group which is almost near its
3290 * capacity but still has some space to pick up some load
3291 * from other group and save more power
3293 if (sum_nr_running <= group_capacity - 1) {
3294 if (sum_nr_running > leader_nr_running ||
3295 (sum_nr_running == leader_nr_running &&
3296 cpumask_first(sched_group_cpus(group)) <
3297 cpumask_first(sched_group_cpus(group_leader)))) {
3298 group_leader = group;
3299 leader_nr_running = sum_nr_running;
3304 group = group->next;
3305 } while (group != sd->groups);
3307 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3310 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3312 if (this_load >= avg_load ||
3313 100*max_load <= sd->imbalance_pct*this_load)
3316 busiest_load_per_task /= busiest_nr_running;
3318 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3321 * We're trying to get all the cpus to the average_load, so we don't
3322 * want to push ourselves above the average load, nor do we wish to
3323 * reduce the max loaded cpu below the average load, as either of these
3324 * actions would just result in more rebalancing later, and ping-pong
3325 * tasks around. Thus we look for the minimum possible imbalance.
3326 * Negative imbalances (*we* are more loaded than anyone else) will
3327 * be counted as no imbalance for these purposes -- we can't fix that
3328 * by pulling tasks to us. Be careful of negative numbers as they'll
3329 * appear as very large values with unsigned longs.
3331 if (max_load <= busiest_load_per_task)
3335 * In the presence of smp nice balancing, certain scenarios can have
3336 * max load less than avg load(as we skip the groups at or below
3337 * its cpu_power, while calculating max_load..)
3339 if (max_load < avg_load) {
3341 goto small_imbalance;
3344 /* Don't want to pull so many tasks that a group would go idle */
3345 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3347 /* How much load to actually move to equalise the imbalance */
3348 *imbalance = min(max_pull * busiest->__cpu_power,
3349 (avg_load - this_load) * this->__cpu_power)
3353 * if *imbalance is less than the average load per runnable task
3354 * there is no gaurantee that any tasks will be moved so we'll have
3355 * a think about bumping its value to force at least one task to be
3358 if (*imbalance < busiest_load_per_task) {
3359 unsigned long tmp, pwr_now, pwr_move;
3363 pwr_move = pwr_now = 0;
3365 if (this_nr_running) {
3366 this_load_per_task /= this_nr_running;
3367 if (busiest_load_per_task > this_load_per_task)
3370 this_load_per_task = cpu_avg_load_per_task(this_cpu);
3372 if (max_load - this_load + busiest_load_per_task >=
3373 busiest_load_per_task * imbn) {
3374 *imbalance = busiest_load_per_task;
3379 * OK, we don't have enough imbalance to justify moving tasks,
3380 * however we may be able to increase total CPU power used by
3384 pwr_now += busiest->__cpu_power *
3385 min(busiest_load_per_task, max_load);
3386 pwr_now += this->__cpu_power *
3387 min(this_load_per_task, this_load);
3388 pwr_now /= SCHED_LOAD_SCALE;
3390 /* Amount of load we'd subtract */
3391 tmp = sg_div_cpu_power(busiest,
3392 busiest_load_per_task * SCHED_LOAD_SCALE);
3394 pwr_move += busiest->__cpu_power *
3395 min(busiest_load_per_task, max_load - tmp);
3397 /* Amount of load we'd add */
3398 if (max_load * busiest->__cpu_power <
3399 busiest_load_per_task * SCHED_LOAD_SCALE)
3400 tmp = sg_div_cpu_power(this,
3401 max_load * busiest->__cpu_power);
3403 tmp = sg_div_cpu_power(this,
3404 busiest_load_per_task * SCHED_LOAD_SCALE);
3405 pwr_move += this->__cpu_power *
3406 min(this_load_per_task, this_load + tmp);
3407 pwr_move /= SCHED_LOAD_SCALE;
3409 /* Move if we gain throughput */
3410 if (pwr_move > pwr_now)
3411 *imbalance = busiest_load_per_task;
3417 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3418 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3421 if (this == group_leader && group_leader != group_min) {
3422 *imbalance = min_load_per_task;
3423 if (sched_mc_power_savings >= POWERSAVINGS_BALANCE_WAKEUP) {
3424 cpu_rq(this_cpu)->rd->sched_mc_preferred_wakeup_cpu =
3425 cpumask_first(sched_group_cpus(group_leader));
3436 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3439 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3440 unsigned long imbalance, const struct cpumask *cpus)
3442 struct rq *busiest = NULL, *rq;
3443 unsigned long max_load = 0;
3446 for_each_cpu(i, sched_group_cpus(group)) {
3449 if (!cpumask_test_cpu(i, cpus))
3453 wl = weighted_cpuload(i);
3455 if (rq->nr_running == 1 && wl > imbalance)
3458 if (wl > max_load) {
3468 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3469 * so long as it is large enough.
3471 #define MAX_PINNED_INTERVAL 512
3474 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3475 * tasks if there is an imbalance.
3477 static int load_balance(int this_cpu, struct rq *this_rq,
3478 struct sched_domain *sd, enum cpu_idle_type idle,
3479 int *balance, struct cpumask *cpus)
3481 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3482 struct sched_group *group;
3483 unsigned long imbalance;
3485 unsigned long flags;
3487 cpumask_setall(cpus);
3490 * When power savings policy is enabled for the parent domain, idle
3491 * sibling can pick up load irrespective of busy siblings. In this case,
3492 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3493 * portraying it as CPU_NOT_IDLE.
3495 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3496 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3499 schedstat_inc(sd, lb_count[idle]);
3503 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3510 schedstat_inc(sd, lb_nobusyg[idle]);
3514 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3516 schedstat_inc(sd, lb_nobusyq[idle]);
3520 BUG_ON(busiest == this_rq);
3522 schedstat_add(sd, lb_imbalance[idle], imbalance);
3525 if (busiest->nr_running > 1) {
3527 * Attempt to move tasks. If find_busiest_group has found
3528 * an imbalance but busiest->nr_running <= 1, the group is
3529 * still unbalanced. ld_moved simply stays zero, so it is
3530 * correctly treated as an imbalance.
3532 local_irq_save(flags);
3533 double_rq_lock(this_rq, busiest);
3534 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3535 imbalance, sd, idle, &all_pinned);
3536 double_rq_unlock(this_rq, busiest);
3537 local_irq_restore(flags);
3540 * some other cpu did the load balance for us.
3542 if (ld_moved && this_cpu != smp_processor_id())
3543 resched_cpu(this_cpu);
3545 /* All tasks on this runqueue were pinned by CPU affinity */
3546 if (unlikely(all_pinned)) {
3547 cpumask_clear_cpu(cpu_of(busiest), cpus);
3548 if (!cpumask_empty(cpus))
3555 schedstat_inc(sd, lb_failed[idle]);
3556 sd->nr_balance_failed++;
3558 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3560 spin_lock_irqsave(&busiest->lock, flags);
3562 /* don't kick the migration_thread, if the curr
3563 * task on busiest cpu can't be moved to this_cpu
3565 if (!cpumask_test_cpu(this_cpu,
3566 &busiest->curr->cpus_allowed)) {
3567 spin_unlock_irqrestore(&busiest->lock, flags);
3569 goto out_one_pinned;
3572 if (!busiest->active_balance) {
3573 busiest->active_balance = 1;
3574 busiest->push_cpu = this_cpu;
3577 spin_unlock_irqrestore(&busiest->lock, flags);
3579 wake_up_process(busiest->migration_thread);
3582 * We've kicked active balancing, reset the failure
3585 sd->nr_balance_failed = sd->cache_nice_tries+1;
3588 sd->nr_balance_failed = 0;
3590 if (likely(!active_balance)) {
3591 /* We were unbalanced, so reset the balancing interval */
3592 sd->balance_interval = sd->min_interval;
3595 * If we've begun active balancing, start to back off. This
3596 * case may not be covered by the all_pinned logic if there
3597 * is only 1 task on the busy runqueue (because we don't call
3600 if (sd->balance_interval < sd->max_interval)
3601 sd->balance_interval *= 2;
3604 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3605 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3611 schedstat_inc(sd, lb_balanced[idle]);
3613 sd->nr_balance_failed = 0;
3616 /* tune up the balancing interval */
3617 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3618 (sd->balance_interval < sd->max_interval))
3619 sd->balance_interval *= 2;
3621 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3622 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3633 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3634 * tasks if there is an imbalance.
3636 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3637 * this_rq is locked.
3640 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3641 struct cpumask *cpus)
3643 struct sched_group *group;
3644 struct rq *busiest = NULL;
3645 unsigned long imbalance;
3650 cpumask_setall(cpus);
3653 * When power savings policy is enabled for the parent domain, idle
3654 * sibling can pick up load irrespective of busy siblings. In this case,
3655 * let the state of idle sibling percolate up as IDLE, instead of
3656 * portraying it as CPU_NOT_IDLE.
3658 if (sd->flags & SD_SHARE_CPUPOWER &&
3659 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3662 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3664 update_shares_locked(this_rq, sd);
3665 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3666 &sd_idle, cpus, NULL);
3668 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3672 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3674 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3678 BUG_ON(busiest == this_rq);
3680 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3683 if (busiest->nr_running > 1) {
3684 /* Attempt to move tasks */
3685 double_lock_balance(this_rq, busiest);
3686 /* this_rq->clock is already updated */
3687 update_rq_clock(busiest);
3688 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3689 imbalance, sd, CPU_NEWLY_IDLE,
3691 double_unlock_balance(this_rq, busiest);
3693 if (unlikely(all_pinned)) {
3694 cpumask_clear_cpu(cpu_of(busiest), cpus);
3695 if (!cpumask_empty(cpus))
3701 int active_balance = 0;
3703 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3704 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3705 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3708 if (sched_mc_power_savings < POWERSAVINGS_BALANCE_WAKEUP)
3711 if (sd->nr_balance_failed++ < 2)
3715 * The only task running in a non-idle cpu can be moved to this
3716 * cpu in an attempt to completely freeup the other CPU
3717 * package. The same method used to move task in load_balance()
3718 * have been extended for load_balance_newidle() to speedup
3719 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
3721 * The package power saving logic comes from
3722 * find_busiest_group(). If there are no imbalance, then
3723 * f_b_g() will return NULL. However when sched_mc={1,2} then
3724 * f_b_g() will select a group from which a running task may be
3725 * pulled to this cpu in order to make the other package idle.
3726 * If there is no opportunity to make a package idle and if
3727 * there are no imbalance, then f_b_g() will return NULL and no
3728 * action will be taken in load_balance_newidle().
3730 * Under normal task pull operation due to imbalance, there
3731 * will be more than one task in the source run queue and
3732 * move_tasks() will succeed. ld_moved will be true and this
3733 * active balance code will not be triggered.
3736 /* Lock busiest in correct order while this_rq is held */
3737 double_lock_balance(this_rq, busiest);
3740 * don't kick the migration_thread, if the curr
3741 * task on busiest cpu can't be moved to this_cpu
3743 if (!cpumask_test_cpu(this_cpu, &busiest->curr->cpus_allowed)) {
3744 double_unlock_balance(this_rq, busiest);
3749 if (!busiest->active_balance) {
3750 busiest->active_balance = 1;
3751 busiest->push_cpu = this_cpu;
3755 double_unlock_balance(this_rq, busiest);
3757 * Should not call ttwu while holding a rq->lock
3759 spin_unlock(&this_rq->lock);
3761 wake_up_process(busiest->migration_thread);
3762 spin_lock(&this_rq->lock);
3765 sd->nr_balance_failed = 0;
3767 update_shares_locked(this_rq, sd);
3771 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3772 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3773 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3775 sd->nr_balance_failed = 0;
3781 * idle_balance is called by schedule() if this_cpu is about to become
3782 * idle. Attempts to pull tasks from other CPUs.
3784 static void idle_balance(int this_cpu, struct rq *this_rq)
3786 struct sched_domain *sd;
3787 int pulled_task = 0;
3788 unsigned long next_balance = jiffies + HZ;
3789 cpumask_var_t tmpmask;
3791 if (!alloc_cpumask_var(&tmpmask, GFP_ATOMIC))
3794 for_each_domain(this_cpu, sd) {
3795 unsigned long interval;
3797 if (!(sd->flags & SD_LOAD_BALANCE))
3800 if (sd->flags & SD_BALANCE_NEWIDLE)
3801 /* If we've pulled tasks over stop searching: */
3802 pulled_task = load_balance_newidle(this_cpu, this_rq,
3805 interval = msecs_to_jiffies(sd->balance_interval);
3806 if (time_after(next_balance, sd->last_balance + interval))
3807 next_balance = sd->last_balance + interval;
3811 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3813 * We are going idle. next_balance may be set based on
3814 * a busy processor. So reset next_balance.
3816 this_rq->next_balance = next_balance;
3818 free_cpumask_var(tmpmask);
3822 * active_load_balance is run by migration threads. It pushes running tasks
3823 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3824 * running on each physical CPU where possible, and avoids physical /
3825 * logical imbalances.
3827 * Called with busiest_rq locked.
3829 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3831 int target_cpu = busiest_rq->push_cpu;
3832 struct sched_domain *sd;
3833 struct rq *target_rq;
3835 /* Is there any task to move? */
3836 if (busiest_rq->nr_running <= 1)
3839 target_rq = cpu_rq(target_cpu);
3842 * This condition is "impossible", if it occurs
3843 * we need to fix it. Originally reported by
3844 * Bjorn Helgaas on a 128-cpu setup.
3846 BUG_ON(busiest_rq == target_rq);
3848 /* move a task from busiest_rq to target_rq */
3849 double_lock_balance(busiest_rq, target_rq);
3850 update_rq_clock(busiest_rq);
3851 update_rq_clock(target_rq);
3853 /* Search for an sd spanning us and the target CPU. */
3854 for_each_domain(target_cpu, sd) {
3855 if ((sd->flags & SD_LOAD_BALANCE) &&
3856 cpumask_test_cpu(busiest_cpu, sched_domain_span(sd)))
3861 schedstat_inc(sd, alb_count);
3863 if (move_one_task(target_rq, target_cpu, busiest_rq,
3865 schedstat_inc(sd, alb_pushed);
3867 schedstat_inc(sd, alb_failed);
3869 double_unlock_balance(busiest_rq, target_rq);
3874 atomic_t load_balancer;
3875 cpumask_var_t cpu_mask;
3876 } nohz ____cacheline_aligned = {
3877 .load_balancer = ATOMIC_INIT(-1),
3881 * This routine will try to nominate the ilb (idle load balancing)
3882 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3883 * load balancing on behalf of all those cpus. If all the cpus in the system
3884 * go into this tickless mode, then there will be no ilb owner (as there is
3885 * no need for one) and all the cpus will sleep till the next wakeup event
3888 * For the ilb owner, tick is not stopped. And this tick will be used
3889 * for idle load balancing. ilb owner will still be part of
3892 * While stopping the tick, this cpu will become the ilb owner if there
3893 * is no other owner. And will be the owner till that cpu becomes busy
3894 * or if all cpus in the system stop their ticks at which point
3895 * there is no need for ilb owner.
3897 * When the ilb owner becomes busy, it nominates another owner, during the
3898 * next busy scheduler_tick()
3900 int select_nohz_load_balancer(int stop_tick)
3902 int cpu = smp_processor_id();
3905 cpumask_set_cpu(cpu, nohz.cpu_mask);
3906 cpu_rq(cpu)->in_nohz_recently = 1;
3909 * If we are going offline and still the leader, give up!
3911 if (!cpu_active(cpu) &&
3912 atomic_read(&nohz.load_balancer) == cpu) {
3913 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3918 /* time for ilb owner also to sleep */
3919 if (cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
3920 if (atomic_read(&nohz.load_balancer) == cpu)
3921 atomic_set(&nohz.load_balancer, -1);
3925 if (atomic_read(&nohz.load_balancer) == -1) {
3926 /* make me the ilb owner */
3927 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
3929 } else if (atomic_read(&nohz.load_balancer) == cpu)
3932 if (!cpumask_test_cpu(cpu, nohz.cpu_mask))
3935 cpumask_clear_cpu(cpu, nohz.cpu_mask);
3937 if (atomic_read(&nohz.load_balancer) == cpu)
3938 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
3945 static DEFINE_SPINLOCK(balancing);
3948 * It checks each scheduling domain to see if it is due to be balanced,
3949 * and initiates a balancing operation if so.
3951 * Balancing parameters are set up in arch_init_sched_domains.
3953 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
3956 struct rq *rq = cpu_rq(cpu);
3957 unsigned long interval;
3958 struct sched_domain *sd;
3959 /* Earliest time when we have to do rebalance again */
3960 unsigned long next_balance = jiffies + 60*HZ;
3961 int update_next_balance = 0;
3965 /* Fails alloc? Rebalancing probably not a priority right now. */
3966 if (!alloc_cpumask_var(&tmp, GFP_ATOMIC))
3969 for_each_domain(cpu, sd) {
3970 if (!(sd->flags & SD_LOAD_BALANCE))
3973 interval = sd->balance_interval;
3974 if (idle != CPU_IDLE)
3975 interval *= sd->busy_factor;
3977 /* scale ms to jiffies */
3978 interval = msecs_to_jiffies(interval);
3979 if (unlikely(!interval))
3981 if (interval > HZ*NR_CPUS/10)
3982 interval = HZ*NR_CPUS/10;
3984 need_serialize = sd->flags & SD_SERIALIZE;
3986 if (need_serialize) {
3987 if (!spin_trylock(&balancing))
3991 if (time_after_eq(jiffies, sd->last_balance + interval)) {
3992 if (load_balance(cpu, rq, sd, idle, &balance, tmp)) {
3994 * We've pulled tasks over so either we're no
3995 * longer idle, or one of our SMT siblings is
3998 idle = CPU_NOT_IDLE;
4000 sd->last_balance = jiffies;
4003 spin_unlock(&balancing);
4005 if (time_after(next_balance, sd->last_balance + interval)) {
4006 next_balance = sd->last_balance + interval;
4007 update_next_balance = 1;
4011 * Stop the load balance at this level. There is another
4012 * CPU in our sched group which is doing load balancing more
4020 * next_balance will be updated only when there is a need.
4021 * When the cpu is attached to null domain for ex, it will not be
4024 if (likely(update_next_balance))
4025 rq->next_balance = next_balance;
4027 free_cpumask_var(tmp);
4031 * run_rebalance_domains is triggered when needed from the scheduler tick.
4032 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4033 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4035 static void run_rebalance_domains(struct softirq_action *h)
4037 int this_cpu = smp_processor_id();
4038 struct rq *this_rq = cpu_rq(this_cpu);
4039 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4040 CPU_IDLE : CPU_NOT_IDLE;
4042 rebalance_domains(this_cpu, idle);
4046 * If this cpu is the owner for idle load balancing, then do the
4047 * balancing on behalf of the other idle cpus whose ticks are
4050 if (this_rq->idle_at_tick &&
4051 atomic_read(&nohz.load_balancer) == this_cpu) {
4055 for_each_cpu(balance_cpu, nohz.cpu_mask) {
4056 if (balance_cpu == this_cpu)
4060 * If this cpu gets work to do, stop the load balancing
4061 * work being done for other cpus. Next load
4062 * balancing owner will pick it up.
4067 rebalance_domains(balance_cpu, CPU_IDLE);
4069 rq = cpu_rq(balance_cpu);
4070 if (time_after(this_rq->next_balance, rq->next_balance))
4071 this_rq->next_balance = rq->next_balance;
4078 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4080 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4081 * idle load balancing owner or decide to stop the periodic load balancing,
4082 * if the whole system is idle.
4084 static inline void trigger_load_balance(struct rq *rq, int cpu)
4088 * If we were in the nohz mode recently and busy at the current
4089 * scheduler tick, then check if we need to nominate new idle
4092 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4093 rq->in_nohz_recently = 0;
4095 if (atomic_read(&nohz.load_balancer) == cpu) {
4096 cpumask_clear_cpu(cpu, nohz.cpu_mask);
4097 atomic_set(&nohz.load_balancer, -1);
4100 if (atomic_read(&nohz.load_balancer) == -1) {
4102 * simple selection for now: Nominate the
4103 * first cpu in the nohz list to be the next
4106 * TBD: Traverse the sched domains and nominate
4107 * the nearest cpu in the nohz.cpu_mask.
4109 int ilb = cpumask_first(nohz.cpu_mask);
4111 if (ilb < nr_cpu_ids)
4117 * If this cpu is idle and doing idle load balancing for all the
4118 * cpus with ticks stopped, is it time for that to stop?
4120 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4121 cpumask_weight(nohz.cpu_mask) == num_online_cpus()) {
4127 * If this cpu is idle and the idle load balancing is done by
4128 * someone else, then no need raise the SCHED_SOFTIRQ
4130 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4131 cpumask_test_cpu(cpu, nohz.cpu_mask))
4134 if (time_after_eq(jiffies, rq->next_balance))
4135 raise_softirq(SCHED_SOFTIRQ);
4138 #else /* CONFIG_SMP */
4141 * on UP we do not need to balance between CPUs:
4143 static inline void idle_balance(int cpu, struct rq *rq)
4149 DEFINE_PER_CPU(struct kernel_stat, kstat);
4151 EXPORT_PER_CPU_SYMBOL(kstat);
4154 * Return any ns on the sched_clock that have not yet been accounted in
4155 * @p in case that task is currently running.
4157 * Called with task_rq_lock() held on @rq.
4159 static u64 do_task_delta_exec(struct task_struct *p, struct rq *rq)
4163 if (task_current(rq, p)) {
4164 update_rq_clock(rq);
4165 ns = rq->clock - p->se.exec_start;
4173 unsigned long long task_delta_exec(struct task_struct *p)
4175 unsigned long flags;
4179 rq = task_rq_lock(p, &flags);
4180 ns = do_task_delta_exec(p, rq);
4181 task_rq_unlock(rq, &flags);
4187 * Return accounted runtime for the task.
4188 * In case the task is currently running, return the runtime plus current's
4189 * pending runtime that have not been accounted yet.
4191 unsigned long long task_sched_runtime(struct task_struct *p)
4193 unsigned long flags;
4197 rq = task_rq_lock(p, &flags);
4198 ns = p->se.sum_exec_runtime + do_task_delta_exec(p, rq);
4199 task_rq_unlock(rq, &flags);
4205 * Return sum_exec_runtime for the thread group.
4206 * In case the task is currently running, return the sum plus current's
4207 * pending runtime that have not been accounted yet.
4209 * Note that the thread group might have other running tasks as well,
4210 * so the return value not includes other pending runtime that other
4211 * running tasks might have.
4213 unsigned long long thread_group_sched_runtime(struct task_struct *p)
4215 struct task_cputime totals;
4216 unsigned long flags;
4220 rq = task_rq_lock(p, &flags);
4221 thread_group_cputime(p, &totals);
4222 ns = totals.sum_exec_runtime + do_task_delta_exec(p, rq);
4223 task_rq_unlock(rq, &flags);
4229 * Account user cpu time to a process.
4230 * @p: the process that the cpu time gets accounted to
4231 * @cputime: the cpu time spent in user space since the last update
4232 * @cputime_scaled: cputime scaled by cpu frequency
4234 void account_user_time(struct task_struct *p, cputime_t cputime,
4235 cputime_t cputime_scaled)
4237 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4240 /* Add user time to process. */
4241 p->utime = cputime_add(p->utime, cputime);
4242 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4243 account_group_user_time(p, 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);
4252 cpuacct_update_stats(p, CPUACCT_STAT_USER, cputime);
4253 /* Account for user time used */
4254 acct_update_integrals(p);
4258 * Account guest cpu time to a process.
4259 * @p: the process that the cpu time gets accounted to
4260 * @cputime: the cpu time spent in virtual machine since the last update
4261 * @cputime_scaled: cputime scaled by cpu frequency
4263 static void account_guest_time(struct task_struct *p, cputime_t cputime,
4264 cputime_t cputime_scaled)
4267 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4269 tmp = cputime_to_cputime64(cputime);
4271 /* Add guest time to process. */
4272 p->utime = cputime_add(p->utime, cputime);
4273 p->utimescaled = cputime_add(p->utimescaled, cputime_scaled);
4274 account_group_user_time(p, cputime);
4275 p->gtime = cputime_add(p->gtime, cputime);
4277 /* Add guest time to cpustat. */
4278 cpustat->user = cputime64_add(cpustat->user, tmp);
4279 cpustat->guest = cputime64_add(cpustat->guest, tmp);
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
4287 * @cputime_scaled: cputime scaled by cpu frequency
4289 void account_system_time(struct task_struct *p, int hardirq_offset,
4290 cputime_t cputime, cputime_t cputime_scaled)
4292 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4295 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4296 account_guest_time(p, cputime, cputime_scaled);
4300 /* Add system time to process. */
4301 p->stime = cputime_add(p->stime, cputime);
4302 p->stimescaled = cputime_add(p->stimescaled, cputime_scaled);
4303 account_group_system_time(p, cputime);
4305 /* Add system time to cpustat. */
4306 tmp = cputime_to_cputime64(cputime);
4307 if (hardirq_count() - hardirq_offset)
4308 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4309 else if (softirq_count())
4310 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4312 cpustat->system = cputime64_add(cpustat->system, tmp);
4314 cpuacct_update_stats(p, CPUACCT_STAT_SYSTEM, cputime);
4316 /* Account for system time used */
4317 acct_update_integrals(p);
4321 * Account for involuntary wait time.
4322 * @steal: the cpu time spent in involuntary wait
4324 void account_steal_time(cputime_t cputime)
4326 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4327 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4329 cpustat->steal = cputime64_add(cpustat->steal, cputime64);
4333 * Account for idle time.
4334 * @cputime: the cpu time spent in idle wait
4336 void account_idle_time(cputime_t cputime)
4338 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4339 cputime64_t cputime64 = cputime_to_cputime64(cputime);
4340 struct rq *rq = this_rq();
4342 if (atomic_read(&rq->nr_iowait) > 0)
4343 cpustat->iowait = cputime64_add(cpustat->iowait, cputime64);
4345 cpustat->idle = cputime64_add(cpustat->idle, cputime64);
4348 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4351 * Account a single tick of cpu time.
4352 * @p: the process that the cpu time gets accounted to
4353 * @user_tick: indicates if the tick is a user or a system tick
4355 void account_process_tick(struct task_struct *p, int user_tick)
4357 cputime_t one_jiffy = jiffies_to_cputime(1);
4358 cputime_t one_jiffy_scaled = cputime_to_scaled(one_jiffy);
4359 struct rq *rq = this_rq();
4362 account_user_time(p, one_jiffy, one_jiffy_scaled);
4363 else if (p != rq->idle)
4364 account_system_time(p, HARDIRQ_OFFSET, one_jiffy,
4367 account_idle_time(one_jiffy);
4371 * Account multiple ticks of steal time.
4372 * @p: the process from which the cpu time has been stolen
4373 * @ticks: number of stolen ticks
4375 void account_steal_ticks(unsigned long ticks)
4377 account_steal_time(jiffies_to_cputime(ticks));
4381 * Account multiple ticks of idle time.
4382 * @ticks: number of stolen ticks
4384 void account_idle_ticks(unsigned long ticks)
4386 account_idle_time(jiffies_to_cputime(ticks));
4392 * Use precise platform statistics if available:
4394 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4395 cputime_t task_utime(struct task_struct *p)
4400 cputime_t task_stime(struct task_struct *p)
4405 cputime_t task_utime(struct task_struct *p)
4407 clock_t utime = cputime_to_clock_t(p->utime),
4408 total = utime + cputime_to_clock_t(p->stime);
4412 * Use CFS's precise accounting:
4414 temp = (u64)nsec_to_clock_t(p->se.sum_exec_runtime);
4418 do_div(temp, total);
4420 utime = (clock_t)temp;
4422 p->prev_utime = max(p->prev_utime, clock_t_to_cputime(utime));
4423 return p->prev_utime;
4426 cputime_t task_stime(struct task_struct *p)
4431 * Use CFS's precise accounting. (we subtract utime from
4432 * the total, to make sure the total observed by userspace
4433 * grows monotonically - apps rely on that):
4435 stime = nsec_to_clock_t(p->se.sum_exec_runtime) -
4436 cputime_to_clock_t(task_utime(p));
4439 p->prev_stime = max(p->prev_stime, clock_t_to_cputime(stime));
4441 return p->prev_stime;
4445 inline cputime_t task_gtime(struct task_struct *p)
4451 * This function gets called by the timer code, with HZ frequency.
4452 * We call it with interrupts disabled.
4454 * It also gets called by the fork code, when changing the parent's
4457 void scheduler_tick(void)
4459 int cpu = smp_processor_id();
4460 struct rq *rq = cpu_rq(cpu);
4461 struct task_struct *curr = rq->curr;
4465 spin_lock(&rq->lock);
4466 update_rq_clock(rq);
4467 update_cpu_load(rq);
4468 curr->sched_class->task_tick(rq, curr, 0);
4469 spin_unlock(&rq->lock);
4472 rq->idle_at_tick = idle_cpu(cpu);
4473 trigger_load_balance(rq, cpu);
4477 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4478 defined(CONFIG_PREEMPT_TRACER))
4480 static inline unsigned long get_parent_ip(unsigned long addr)
4482 if (in_lock_functions(addr)) {
4483 addr = CALLER_ADDR2;
4484 if (in_lock_functions(addr))
4485 addr = CALLER_ADDR3;
4490 void __kprobes add_preempt_count(int val)
4492 #ifdef CONFIG_DEBUG_PREEMPT
4496 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4499 preempt_count() += val;
4500 #ifdef CONFIG_DEBUG_PREEMPT
4502 * Spinlock count overflowing soon?
4504 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4507 if (preempt_count() == val)
4508 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4510 EXPORT_SYMBOL(add_preempt_count);
4512 void __kprobes sub_preempt_count(int val)
4514 #ifdef CONFIG_DEBUG_PREEMPT
4518 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4521 * Is the spinlock portion underflowing?
4523 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4524 !(preempt_count() & PREEMPT_MASK)))
4528 if (preempt_count() == val)
4529 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4530 preempt_count() -= val;
4532 EXPORT_SYMBOL(sub_preempt_count);
4537 * Print scheduling while atomic bug:
4539 static noinline void __schedule_bug(struct task_struct *prev)
4541 struct pt_regs *regs = get_irq_regs();
4543 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4544 prev->comm, prev->pid, preempt_count());
4546 debug_show_held_locks(prev);
4548 if (irqs_disabled())
4549 print_irqtrace_events(prev);
4558 * Various schedule()-time debugging checks and statistics:
4560 static inline void schedule_debug(struct task_struct *prev)
4563 * Test if we are atomic. Since do_exit() needs to call into
4564 * schedule() atomically, we ignore that path for now.
4565 * Otherwise, whine if we are scheduling when we should not be.
4567 if (unlikely(in_atomic_preempt_off() && !prev->exit_state))
4568 __schedule_bug(prev);
4570 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4572 schedstat_inc(this_rq(), sched_count);
4573 #ifdef CONFIG_SCHEDSTATS
4574 if (unlikely(prev->lock_depth >= 0)) {
4575 schedstat_inc(this_rq(), bkl_count);
4576 schedstat_inc(prev, sched_info.bkl_count);
4582 * Pick up the highest-prio task:
4584 static inline struct task_struct *
4585 pick_next_task(struct rq *rq, struct task_struct *prev)
4587 const struct sched_class *class;
4588 struct task_struct *p;
4591 * Optimization: we know that if all tasks are in
4592 * the fair class we can call that function directly:
4594 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4595 p = fair_sched_class.pick_next_task(rq);
4600 class = sched_class_highest;
4602 p = class->pick_next_task(rq);
4606 * Will never be NULL as the idle class always
4607 * returns a non-NULL p:
4609 class = class->next;
4614 * schedule() is the main scheduler function.
4616 asmlinkage void __sched __schedule(void)
4618 struct task_struct *prev, *next;
4619 unsigned long *switch_count;
4623 cpu = smp_processor_id();
4627 switch_count = &prev->nivcsw;
4629 release_kernel_lock(prev);
4630 need_resched_nonpreemptible:
4632 schedule_debug(prev);
4634 if (sched_feat(HRTICK))
4637 spin_lock_irq(&rq->lock);
4638 update_rq_clock(rq);
4639 clear_tsk_need_resched(prev);
4641 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4642 if (unlikely(signal_pending_state(prev->state, prev)))
4643 prev->state = TASK_RUNNING;
4645 deactivate_task(rq, prev, 1);
4646 switch_count = &prev->nvcsw;
4650 if (prev->sched_class->pre_schedule)
4651 prev->sched_class->pre_schedule(rq, prev);
4654 if (unlikely(!rq->nr_running))
4655 idle_balance(cpu, rq);
4657 prev->sched_class->put_prev_task(rq, prev);
4658 next = pick_next_task(rq, prev);
4660 if (likely(prev != next)) {
4661 sched_info_switch(prev, next);
4667 context_switch(rq, prev, next); /* unlocks the rq */
4669 * the context switch might have flipped the stack from under
4670 * us, hence refresh the local variables.
4672 cpu = smp_processor_id();
4675 spin_unlock_irq(&rq->lock);
4677 if (unlikely(reacquire_kernel_lock(current) < 0))
4678 goto need_resched_nonpreemptible;
4681 asmlinkage void __sched schedule(void)
4686 preempt_enable_no_resched();
4687 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4690 EXPORT_SYMBOL(schedule);
4694 * Look out! "owner" is an entirely speculative pointer
4695 * access and not reliable.
4697 int mutex_spin_on_owner(struct mutex *lock, struct thread_info *owner)
4702 if (!sched_feat(OWNER_SPIN))
4705 #ifdef CONFIG_DEBUG_PAGEALLOC
4707 * Need to access the cpu field knowing that
4708 * DEBUG_PAGEALLOC could have unmapped it if
4709 * the mutex owner just released it and exited.
4711 if (probe_kernel_address(&owner->cpu, cpu))
4718 * Even if the access succeeded (likely case),
4719 * the cpu field may no longer be valid.
4721 if (cpu >= nr_cpumask_bits)
4725 * We need to validate that we can do a
4726 * get_cpu() and that we have the percpu area.
4728 if (!cpu_online(cpu))
4735 * Owner changed, break to re-assess state.
4737 if (lock->owner != owner)
4741 * Is that owner really running on that cpu?
4743 if (task_thread_info(rq->curr) != owner || need_resched())
4753 #ifdef CONFIG_PREEMPT
4755 * this is the entry point to schedule() from in-kernel preemption
4756 * off of preempt_enable. Kernel preemptions off return from interrupt
4757 * occur there and call schedule directly.
4759 asmlinkage void __sched preempt_schedule(void)
4761 struct thread_info *ti = current_thread_info();
4764 * If there is a non-zero preempt_count or interrupts are disabled,
4765 * we do not want to preempt the current task. Just return..
4767 if (likely(ti->preempt_count || irqs_disabled()))
4771 add_preempt_count(PREEMPT_ACTIVE);
4773 sub_preempt_count(PREEMPT_ACTIVE);
4776 * Check again in case we missed a preemption opportunity
4777 * between schedule and now.
4780 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4782 EXPORT_SYMBOL(preempt_schedule);
4785 * this is the entry point to schedule() from kernel preemption
4786 * off of irq context.
4787 * Note, that this is called and return with irqs disabled. This will
4788 * protect us against recursive calling from irq.
4790 asmlinkage void __sched preempt_schedule_irq(void)
4792 struct thread_info *ti = current_thread_info();
4794 /* Catch callers which need to be fixed */
4795 BUG_ON(ti->preempt_count || !irqs_disabled());
4798 add_preempt_count(PREEMPT_ACTIVE);
4801 local_irq_disable();
4802 sub_preempt_count(PREEMPT_ACTIVE);
4805 * Check again in case we missed a preemption opportunity
4806 * between schedule and now.
4809 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4812 #endif /* CONFIG_PREEMPT */
4814 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4817 return try_to_wake_up(curr->private, mode, sync);
4819 EXPORT_SYMBOL(default_wake_function);
4822 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4823 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4824 * number) then we wake all the non-exclusive tasks and one exclusive task.
4826 * There are circumstances in which we can try to wake a task which has already
4827 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4828 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4830 void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4831 int nr_exclusive, int sync, void *key)
4833 wait_queue_t *curr, *next;
4835 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4836 unsigned flags = curr->flags;
4838 if (curr->func(curr, mode, sync, key) &&
4839 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4845 * __wake_up - wake up threads blocked on a waitqueue.
4847 * @mode: which threads
4848 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4849 * @key: is directly passed to the wakeup function
4851 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4852 int nr_exclusive, void *key)
4854 unsigned long flags;
4856 spin_lock_irqsave(&q->lock, flags);
4857 __wake_up_common(q, mode, nr_exclusive, 0, key);
4858 spin_unlock_irqrestore(&q->lock, flags);
4860 EXPORT_SYMBOL(__wake_up);
4863 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4865 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4867 __wake_up_common(q, mode, 1, 0, NULL);
4871 * __wake_up_sync - wake up threads blocked on a waitqueue.
4873 * @mode: which threads
4874 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4876 * The sync wakeup differs that the waker knows that it will schedule
4877 * away soon, so while the target thread will be woken up, it will not
4878 * be migrated to another CPU - ie. the two threads are 'synchronized'
4879 * with each other. This can prevent needless bouncing between CPUs.
4881 * On UP it can prevent extra preemption.
4884 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4886 unsigned long flags;
4892 if (unlikely(!nr_exclusive))
4895 spin_lock_irqsave(&q->lock, flags);
4896 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4897 spin_unlock_irqrestore(&q->lock, flags);
4899 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4902 * complete: - signals a single thread waiting on this completion
4903 * @x: holds the state of this particular completion
4905 * This will wake up a single thread waiting on this completion. Threads will be
4906 * awakened in the same order in which they were queued.
4908 * See also complete_all(), wait_for_completion() and related routines.
4910 void complete(struct completion *x)
4912 unsigned long flags;
4914 spin_lock_irqsave(&x->wait.lock, flags);
4916 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4917 spin_unlock_irqrestore(&x->wait.lock, flags);
4919 EXPORT_SYMBOL(complete);
4922 * complete_all: - signals all threads waiting on this completion
4923 * @x: holds the state of this particular completion
4925 * This will wake up all threads waiting on this particular completion event.
4927 void complete_all(struct completion *x)
4929 unsigned long flags;
4931 spin_lock_irqsave(&x->wait.lock, flags);
4932 x->done += UINT_MAX/2;
4933 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4934 spin_unlock_irqrestore(&x->wait.lock, flags);
4936 EXPORT_SYMBOL(complete_all);
4938 static inline long __sched
4939 do_wait_for_common(struct completion *x, long timeout, int state)
4942 DECLARE_WAITQUEUE(wait, current);
4944 wait.flags |= WQ_FLAG_EXCLUSIVE;
4945 __add_wait_queue_tail(&x->wait, &wait);
4947 if (signal_pending_state(state, current)) {
4948 timeout = -ERESTARTSYS;
4951 __set_current_state(state);
4952 spin_unlock_irq(&x->wait.lock);
4953 timeout = schedule_timeout(timeout);
4954 spin_lock_irq(&x->wait.lock);
4955 } while (!x->done && timeout);
4956 __remove_wait_queue(&x->wait, &wait);
4961 return timeout ?: 1;
4965 wait_for_common(struct completion *x, long timeout, int state)
4969 spin_lock_irq(&x->wait.lock);
4970 timeout = do_wait_for_common(x, timeout, state);
4971 spin_unlock_irq(&x->wait.lock);
4976 * wait_for_completion: - waits for completion of a task
4977 * @x: holds the state of this particular completion
4979 * This waits to be signaled for completion of a specific task. It is NOT
4980 * interruptible and there is no timeout.
4982 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4983 * and interrupt capability. Also see complete().
4985 void __sched wait_for_completion(struct completion *x)
4987 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4989 EXPORT_SYMBOL(wait_for_completion);
4992 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4993 * @x: holds the state of this particular completion
4994 * @timeout: timeout value in jiffies
4996 * This waits for either a completion of a specific task to be signaled or for a
4997 * specified timeout to expire. The timeout is in jiffies. It is not
5000 unsigned long __sched
5001 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
5003 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
5005 EXPORT_SYMBOL(wait_for_completion_timeout);
5008 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5009 * @x: holds the state of this particular completion
5011 * This waits for completion of a specific task to be signaled. It is
5014 int __sched wait_for_completion_interruptible(struct completion *x)
5016 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
5017 if (t == -ERESTARTSYS)
5021 EXPORT_SYMBOL(wait_for_completion_interruptible);
5024 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5025 * @x: holds the state of this particular completion
5026 * @timeout: timeout value in jiffies
5028 * This waits for either a completion of a specific task to be signaled or for a
5029 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5031 unsigned long __sched
5032 wait_for_completion_interruptible_timeout(struct completion *x,
5033 unsigned long timeout)
5035 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
5037 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
5040 * wait_for_completion_killable: - waits for completion of a task (killable)
5041 * @x: holds the state of this particular completion
5043 * This waits to be signaled for completion of a specific task. It can be
5044 * interrupted by a kill signal.
5046 int __sched wait_for_completion_killable(struct completion *x)
5048 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
5049 if (t == -ERESTARTSYS)
5053 EXPORT_SYMBOL(wait_for_completion_killable);
5056 * try_wait_for_completion - try to decrement a completion without blocking
5057 * @x: completion structure
5059 * Returns: 0 if a decrement cannot be done without blocking
5060 * 1 if a decrement succeeded.
5062 * If a completion is being used as a counting completion,
5063 * attempt to decrement the counter without blocking. This
5064 * enables us to avoid waiting if the resource the completion
5065 * is protecting is not available.
5067 bool try_wait_for_completion(struct completion *x)
5071 spin_lock_irq(&x->wait.lock);
5076 spin_unlock_irq(&x->wait.lock);
5079 EXPORT_SYMBOL(try_wait_for_completion);
5082 * completion_done - Test to see if a completion has any waiters
5083 * @x: completion structure
5085 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5086 * 1 if there are no waiters.
5089 bool completion_done(struct completion *x)
5093 spin_lock_irq(&x->wait.lock);
5096 spin_unlock_irq(&x->wait.lock);
5099 EXPORT_SYMBOL(completion_done);
5102 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
5104 unsigned long flags;
5107 init_waitqueue_entry(&wait, current);
5109 __set_current_state(state);
5111 spin_lock_irqsave(&q->lock, flags);
5112 __add_wait_queue(q, &wait);
5113 spin_unlock(&q->lock);
5114 timeout = schedule_timeout(timeout);
5115 spin_lock_irq(&q->lock);
5116 __remove_wait_queue(q, &wait);
5117 spin_unlock_irqrestore(&q->lock, flags);
5122 void __sched interruptible_sleep_on(wait_queue_head_t *q)
5124 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5126 EXPORT_SYMBOL(interruptible_sleep_on);
5129 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
5131 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
5133 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
5135 void __sched sleep_on(wait_queue_head_t *q)
5137 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
5139 EXPORT_SYMBOL(sleep_on);
5141 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
5143 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
5145 EXPORT_SYMBOL(sleep_on_timeout);
5147 #ifdef CONFIG_RT_MUTEXES
5150 * rt_mutex_setprio - set the current priority of a task
5152 * @prio: prio value (kernel-internal form)
5154 * This function changes the 'effective' priority of a task. It does
5155 * not touch ->normal_prio like __setscheduler().
5157 * Used by the rt_mutex code to implement priority inheritance logic.
5159 void rt_mutex_setprio(struct task_struct *p, int prio)
5161 unsigned long flags;
5162 int oldprio, on_rq, running;
5164 const struct sched_class *prev_class = p->sched_class;
5166 BUG_ON(prio < 0 || prio > MAX_PRIO);
5168 rq = task_rq_lock(p, &flags);
5169 update_rq_clock(rq);
5172 on_rq = p->se.on_rq;
5173 running = task_current(rq, p);
5175 dequeue_task(rq, p, 0);
5177 p->sched_class->put_prev_task(rq, p);
5180 p->sched_class = &rt_sched_class;
5182 p->sched_class = &fair_sched_class;
5187 p->sched_class->set_curr_task(rq);
5189 enqueue_task(rq, p, 0);
5191 check_class_changed(rq, p, prev_class, oldprio, running);
5193 task_rq_unlock(rq, &flags);
5198 void set_user_nice(struct task_struct *p, long nice)
5200 int old_prio, delta, on_rq;
5201 unsigned long flags;
5204 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
5207 * We have to be careful, if called from sys_setpriority(),
5208 * the task might be in the middle of scheduling on another CPU.
5210 rq = task_rq_lock(p, &flags);
5211 update_rq_clock(rq);
5213 * The RT priorities are set via sched_setscheduler(), but we still
5214 * allow the 'normal' nice value to be set - but as expected
5215 * it wont have any effect on scheduling until the task is
5216 * SCHED_FIFO/SCHED_RR:
5218 if (task_has_rt_policy(p)) {
5219 p->static_prio = NICE_TO_PRIO(nice);
5222 on_rq = p->se.on_rq;
5224 dequeue_task(rq, p, 0);
5226 p->static_prio = NICE_TO_PRIO(nice);
5229 p->prio = effective_prio(p);
5230 delta = p->prio - old_prio;
5233 enqueue_task(rq, p, 0);
5235 * If the task increased its priority or is running and
5236 * lowered its priority, then reschedule its CPU:
5238 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5239 resched_task(rq->curr);
5242 task_rq_unlock(rq, &flags);
5244 EXPORT_SYMBOL(set_user_nice);
5247 * can_nice - check if a task can reduce its nice value
5251 int can_nice(const struct task_struct *p, const int nice)
5253 /* convert nice value [19,-20] to rlimit style value [1,40] */
5254 int nice_rlim = 20 - nice;
5256 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5257 capable(CAP_SYS_NICE));
5260 #ifdef __ARCH_WANT_SYS_NICE
5263 * sys_nice - change the priority of the current process.
5264 * @increment: priority increment
5266 * sys_setpriority is a more generic, but much slower function that
5267 * does similar things.
5269 SYSCALL_DEFINE1(nice, int, increment)
5274 * Setpriority might change our priority at the same moment.
5275 * We don't have to worry. Conceptually one call occurs first
5276 * and we have a single winner.
5278 if (increment < -40)
5283 nice = PRIO_TO_NICE(current->static_prio) + increment;
5289 if (increment < 0 && !can_nice(current, nice))
5292 retval = security_task_setnice(current, nice);
5296 set_user_nice(current, nice);
5303 * task_prio - return the priority value of a given task.
5304 * @p: the task in question.
5306 * This is the priority value as seen by users in /proc.
5307 * RT tasks are offset by -200. Normal tasks are centered
5308 * around 0, value goes from -16 to +15.
5310 int task_prio(const struct task_struct *p)
5312 return p->prio - MAX_RT_PRIO;
5316 * task_nice - return the nice value of a given task.
5317 * @p: the task in question.
5319 int task_nice(const struct task_struct *p)
5321 return TASK_NICE(p);
5323 EXPORT_SYMBOL(task_nice);
5326 * idle_cpu - is a given cpu idle currently?
5327 * @cpu: the processor in question.
5329 int idle_cpu(int cpu)
5331 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5335 * idle_task - return the idle task for a given cpu.
5336 * @cpu: the processor in question.
5338 struct task_struct *idle_task(int cpu)
5340 return cpu_rq(cpu)->idle;
5344 * find_process_by_pid - find a process with a matching PID value.
5345 * @pid: the pid in question.
5347 static struct task_struct *find_process_by_pid(pid_t pid)
5349 return pid ? find_task_by_vpid(pid) : current;
5352 /* Actually do priority change: must hold rq lock. */
5354 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5356 BUG_ON(p->se.on_rq);
5359 switch (p->policy) {
5363 p->sched_class = &fair_sched_class;
5367 p->sched_class = &rt_sched_class;
5371 p->rt_priority = prio;
5372 p->normal_prio = normal_prio(p);
5373 /* we are holding p->pi_lock already */
5374 p->prio = rt_mutex_getprio(p);
5379 * check the target process has a UID that matches the current process's
5381 static bool check_same_owner(struct task_struct *p)
5383 const struct cred *cred = current_cred(), *pcred;
5387 pcred = __task_cred(p);
5388 match = (cred->euid == pcred->euid ||
5389 cred->euid == pcred->uid);
5394 static int __sched_setscheduler(struct task_struct *p, int policy,
5395 struct sched_param *param, bool user)
5397 int retval, oldprio, oldpolicy = -1, on_rq, running;
5398 unsigned long flags;
5399 const struct sched_class *prev_class = p->sched_class;
5402 /* may grab non-irq protected spin_locks */
5403 BUG_ON(in_interrupt());
5405 /* double check policy once rq lock held */
5407 policy = oldpolicy = p->policy;
5408 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5409 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5410 policy != SCHED_IDLE)
5413 * Valid priorities for SCHED_FIFO and SCHED_RR are
5414 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5415 * SCHED_BATCH and SCHED_IDLE is 0.
5417 if (param->sched_priority < 0 ||
5418 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5419 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5421 if (rt_policy(policy) != (param->sched_priority != 0))
5425 * Allow unprivileged RT tasks to decrease priority:
5427 if (user && !capable(CAP_SYS_NICE)) {
5428 if (rt_policy(policy)) {
5429 unsigned long rlim_rtprio;
5431 if (!lock_task_sighand(p, &flags))
5433 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5434 unlock_task_sighand(p, &flags);
5436 /* can't set/change the rt policy */
5437 if (policy != p->policy && !rlim_rtprio)
5440 /* can't increase priority */
5441 if (param->sched_priority > p->rt_priority &&
5442 param->sched_priority > rlim_rtprio)
5446 * Like positive nice levels, dont allow tasks to
5447 * move out of SCHED_IDLE either:
5449 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5452 /* can't change other user's priorities */
5453 if (!check_same_owner(p))
5458 #ifdef CONFIG_RT_GROUP_SCHED
5460 * Do not allow realtime tasks into groups that have no runtime
5463 if (rt_bandwidth_enabled() && rt_policy(policy) &&
5464 task_group(p)->rt_bandwidth.rt_runtime == 0)
5468 retval = security_task_setscheduler(p, policy, param);
5474 * make sure no PI-waiters arrive (or leave) while we are
5475 * changing the priority of the task:
5477 spin_lock_irqsave(&p->pi_lock, flags);
5479 * To be able to change p->policy safely, the apropriate
5480 * runqueue lock must be held.
5482 rq = __task_rq_lock(p);
5483 /* recheck policy now with rq lock held */
5484 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5485 policy = oldpolicy = -1;
5486 __task_rq_unlock(rq);
5487 spin_unlock_irqrestore(&p->pi_lock, flags);
5490 update_rq_clock(rq);
5491 on_rq = p->se.on_rq;
5492 running = task_current(rq, p);
5494 deactivate_task(rq, p, 0);
5496 p->sched_class->put_prev_task(rq, p);
5499 __setscheduler(rq, p, policy, param->sched_priority);
5502 p->sched_class->set_curr_task(rq);
5504 activate_task(rq, p, 0);
5506 check_class_changed(rq, p, prev_class, oldprio, running);
5508 __task_rq_unlock(rq);
5509 spin_unlock_irqrestore(&p->pi_lock, flags);
5511 rt_mutex_adjust_pi(p);
5517 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5518 * @p: the task in question.
5519 * @policy: new policy.
5520 * @param: structure containing the new RT priority.
5522 * NOTE that the task may be already dead.
5524 int sched_setscheduler(struct task_struct *p, int policy,
5525 struct sched_param *param)
5527 return __sched_setscheduler(p, policy, param, true);
5529 EXPORT_SYMBOL_GPL(sched_setscheduler);
5532 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5533 * @p: the task in question.
5534 * @policy: new policy.
5535 * @param: structure containing the new RT priority.
5537 * Just like sched_setscheduler, only don't bother checking if the
5538 * current context has permission. For example, this is needed in
5539 * stop_machine(): we create temporary high priority worker threads,
5540 * but our caller might not have that capability.
5542 int sched_setscheduler_nocheck(struct task_struct *p, int policy,
5543 struct sched_param *param)
5545 return __sched_setscheduler(p, policy, param, false);
5549 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5551 struct sched_param lparam;
5552 struct task_struct *p;
5555 if (!param || pid < 0)
5557 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5562 p = find_process_by_pid(pid);
5564 retval = sched_setscheduler(p, policy, &lparam);
5571 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5572 * @pid: the pid in question.
5573 * @policy: new policy.
5574 * @param: structure containing the new RT priority.
5576 SYSCALL_DEFINE3(sched_setscheduler, pid_t, pid, int, policy,
5577 struct sched_param __user *, param)
5579 /* negative values for policy are not valid */
5583 return do_sched_setscheduler(pid, policy, param);
5587 * sys_sched_setparam - set/change the RT priority of a thread
5588 * @pid: the pid in question.
5589 * @param: structure containing the new RT priority.
5591 SYSCALL_DEFINE2(sched_setparam, pid_t, pid, struct sched_param __user *, param)
5593 return do_sched_setscheduler(pid, -1, param);
5597 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5598 * @pid: the pid in question.
5600 SYSCALL_DEFINE1(sched_getscheduler, pid_t, pid)
5602 struct task_struct *p;
5609 read_lock(&tasklist_lock);
5610 p = find_process_by_pid(pid);
5612 retval = security_task_getscheduler(p);
5616 read_unlock(&tasklist_lock);
5621 * sys_sched_getscheduler - get the RT priority of a thread
5622 * @pid: the pid in question.
5623 * @param: structure containing the RT priority.
5625 SYSCALL_DEFINE2(sched_getparam, pid_t, pid, struct sched_param __user *, param)
5627 struct sched_param lp;
5628 struct task_struct *p;
5631 if (!param || pid < 0)
5634 read_lock(&tasklist_lock);
5635 p = find_process_by_pid(pid);
5640 retval = security_task_getscheduler(p);
5644 lp.sched_priority = p->rt_priority;
5645 read_unlock(&tasklist_lock);
5648 * This one might sleep, we cannot do it with a spinlock held ...
5650 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5655 read_unlock(&tasklist_lock);
5659 long sched_setaffinity(pid_t pid, const struct cpumask *in_mask)
5661 cpumask_var_t cpus_allowed, new_mask;
5662 struct task_struct *p;
5666 read_lock(&tasklist_lock);
5668 p = find_process_by_pid(pid);
5670 read_unlock(&tasklist_lock);
5676 * It is not safe to call set_cpus_allowed with the
5677 * tasklist_lock held. We will bump the task_struct's
5678 * usage count and then drop tasklist_lock.
5681 read_unlock(&tasklist_lock);
5683 if (!alloc_cpumask_var(&cpus_allowed, GFP_KERNEL)) {
5687 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL)) {
5689 goto out_free_cpus_allowed;
5692 if (!check_same_owner(p) && !capable(CAP_SYS_NICE))
5695 retval = security_task_setscheduler(p, 0, NULL);
5699 cpuset_cpus_allowed(p, cpus_allowed);
5700 cpumask_and(new_mask, in_mask, cpus_allowed);
5702 retval = set_cpus_allowed_ptr(p, new_mask);
5705 cpuset_cpus_allowed(p, cpus_allowed);
5706 if (!cpumask_subset(new_mask, cpus_allowed)) {
5708 * We must have raced with a concurrent cpuset
5709 * update. Just reset the cpus_allowed to the
5710 * cpuset's cpus_allowed
5712 cpumask_copy(new_mask, cpus_allowed);
5717 free_cpumask_var(new_mask);
5718 out_free_cpus_allowed:
5719 free_cpumask_var(cpus_allowed);
5726 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5727 struct cpumask *new_mask)
5729 if (len < cpumask_size())
5730 cpumask_clear(new_mask);
5731 else if (len > cpumask_size())
5732 len = cpumask_size();
5734 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5738 * sys_sched_setaffinity - set the cpu affinity of a process
5739 * @pid: pid of the process
5740 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5741 * @user_mask_ptr: user-space pointer to the new cpu mask
5743 SYSCALL_DEFINE3(sched_setaffinity, pid_t, pid, unsigned int, len,
5744 unsigned long __user *, user_mask_ptr)
5746 cpumask_var_t new_mask;
5749 if (!alloc_cpumask_var(&new_mask, GFP_KERNEL))
5752 retval = get_user_cpu_mask(user_mask_ptr, len, new_mask);
5754 retval = sched_setaffinity(pid, new_mask);
5755 free_cpumask_var(new_mask);
5759 long sched_getaffinity(pid_t pid, struct cpumask *mask)
5761 struct task_struct *p;
5765 read_lock(&tasklist_lock);
5768 p = find_process_by_pid(pid);
5772 retval = security_task_getscheduler(p);
5776 cpumask_and(mask, &p->cpus_allowed, cpu_online_mask);
5779 read_unlock(&tasklist_lock);
5786 * sys_sched_getaffinity - get the cpu affinity of a process
5787 * @pid: pid of the process
5788 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5789 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5791 SYSCALL_DEFINE3(sched_getaffinity, pid_t, pid, unsigned int, len,
5792 unsigned long __user *, user_mask_ptr)
5797 if (len < cpumask_size())
5800 if (!alloc_cpumask_var(&mask, GFP_KERNEL))
5803 ret = sched_getaffinity(pid, mask);
5805 if (copy_to_user(user_mask_ptr, mask, cpumask_size()))
5808 ret = cpumask_size();
5810 free_cpumask_var(mask);
5816 * sys_sched_yield - yield the current processor to other threads.
5818 * This function yields the current CPU to other tasks. If there are no
5819 * other threads running on this CPU then this function will return.
5821 SYSCALL_DEFINE0(sched_yield)
5823 struct rq *rq = this_rq_lock();
5825 schedstat_inc(rq, yld_count);
5826 current->sched_class->yield_task(rq);
5829 * Since we are going to call schedule() anyway, there's
5830 * no need to preempt or enable interrupts:
5832 __release(rq->lock);
5833 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5834 _raw_spin_unlock(&rq->lock);
5835 preempt_enable_no_resched();
5842 static void __cond_resched(void)
5844 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5845 __might_sleep(__FILE__, __LINE__);
5848 * The BKS might be reacquired before we have dropped
5849 * PREEMPT_ACTIVE, which could trigger a second
5850 * cond_resched() call.
5853 add_preempt_count(PREEMPT_ACTIVE);
5855 sub_preempt_count(PREEMPT_ACTIVE);
5856 } while (need_resched());
5859 int __sched _cond_resched(void)
5861 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5862 system_state == SYSTEM_RUNNING) {
5868 EXPORT_SYMBOL(_cond_resched);
5871 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5872 * call schedule, and on return reacquire the lock.
5874 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5875 * operations here to prevent schedule() from being called twice (once via
5876 * spin_unlock(), once by hand).
5878 int cond_resched_lock(spinlock_t *lock)
5880 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5883 if (spin_needbreak(lock) || resched) {
5885 if (resched && need_resched())
5894 EXPORT_SYMBOL(cond_resched_lock);
5896 int __sched cond_resched_softirq(void)
5898 BUG_ON(!in_softirq());
5900 if (need_resched() && system_state == SYSTEM_RUNNING) {
5908 EXPORT_SYMBOL(cond_resched_softirq);
5911 * yield - yield the current processor to other threads.
5913 * This is a shortcut for kernel-space yielding - it marks the
5914 * thread runnable and calls sys_sched_yield().
5916 void __sched yield(void)
5918 set_current_state(TASK_RUNNING);
5921 EXPORT_SYMBOL(yield);
5924 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5925 * that process accounting knows that this is a task in IO wait state.
5927 * But don't do that if it is a deliberate, throttling IO wait (this task
5928 * has set its backing_dev_info: the queue against which it should throttle)
5930 void __sched io_schedule(void)
5932 struct rq *rq = &__raw_get_cpu_var(runqueues);
5934 delayacct_blkio_start();
5935 atomic_inc(&rq->nr_iowait);
5937 atomic_dec(&rq->nr_iowait);
5938 delayacct_blkio_end();
5940 EXPORT_SYMBOL(io_schedule);
5942 long __sched io_schedule_timeout(long timeout)
5944 struct rq *rq = &__raw_get_cpu_var(runqueues);
5947 delayacct_blkio_start();
5948 atomic_inc(&rq->nr_iowait);
5949 ret = schedule_timeout(timeout);
5950 atomic_dec(&rq->nr_iowait);
5951 delayacct_blkio_end();
5956 * sys_sched_get_priority_max - return maximum RT priority.
5957 * @policy: scheduling class.
5959 * this syscall returns the maximum rt_priority that can be used
5960 * by a given scheduling class.
5962 SYSCALL_DEFINE1(sched_get_priority_max, int, policy)
5969 ret = MAX_USER_RT_PRIO-1;
5981 * sys_sched_get_priority_min - return minimum RT priority.
5982 * @policy: scheduling class.
5984 * this syscall returns the minimum rt_priority that can be used
5985 * by a given scheduling class.
5987 SYSCALL_DEFINE1(sched_get_priority_min, int, policy)
6005 * sys_sched_rr_get_interval - return the default timeslice of a process.
6006 * @pid: pid of the process.
6007 * @interval: userspace pointer to the timeslice value.
6009 * this syscall writes the default timeslice value of a given process
6010 * into the user-space timespec buffer. A value of '0' means infinity.
6012 SYSCALL_DEFINE2(sched_rr_get_interval, pid_t, pid,
6013 struct timespec __user *, interval)
6015 struct task_struct *p;
6016 unsigned int time_slice;
6024 read_lock(&tasklist_lock);
6025 p = find_process_by_pid(pid);
6029 retval = security_task_getscheduler(p);
6034 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6035 * tasks that are on an otherwise idle runqueue:
6038 if (p->policy == SCHED_RR) {
6039 time_slice = DEF_TIMESLICE;
6040 } else if (p->policy != SCHED_FIFO) {
6041 struct sched_entity *se = &p->se;
6042 unsigned long flags;
6045 rq = task_rq_lock(p, &flags);
6046 if (rq->cfs.load.weight)
6047 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
6048 task_rq_unlock(rq, &flags);
6050 read_unlock(&tasklist_lock);
6051 jiffies_to_timespec(time_slice, &t);
6052 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
6056 read_unlock(&tasklist_lock);
6060 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
6062 void sched_show_task(struct task_struct *p)
6064 unsigned long free = 0;
6067 state = p->state ? __ffs(p->state) + 1 : 0;
6068 printk(KERN_INFO "%-13.13s %c", p->comm,
6069 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
6070 #if BITS_PER_LONG == 32
6071 if (state == TASK_RUNNING)
6072 printk(KERN_CONT " running ");
6074 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
6076 if (state == TASK_RUNNING)
6077 printk(KERN_CONT " running task ");
6079 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
6081 #ifdef CONFIG_DEBUG_STACK_USAGE
6083 unsigned long *n = end_of_stack(p);
6086 free = (unsigned long)n - (unsigned long)end_of_stack(p);
6089 printk(KERN_CONT "%5lu %5d %6d\n", free,
6090 task_pid_nr(p), task_pid_nr(p->real_parent));
6092 show_stack(p, NULL);
6095 void show_state_filter(unsigned long state_filter)
6097 struct task_struct *g, *p;
6099 #if BITS_PER_LONG == 32
6101 " task PC stack pid father\n");
6104 " task PC stack pid father\n");
6106 read_lock(&tasklist_lock);
6107 do_each_thread(g, p) {
6109 * reset the NMI-timeout, listing all files on a slow
6110 * console might take alot of time:
6112 touch_nmi_watchdog();
6113 if (!state_filter || (p->state & state_filter))
6115 } while_each_thread(g, p);
6117 touch_all_softlockup_watchdogs();
6119 #ifdef CONFIG_SCHED_DEBUG
6120 sysrq_sched_debug_show();
6122 read_unlock(&tasklist_lock);
6124 * Only show locks if all tasks are dumped:
6126 if (state_filter == -1)
6127 debug_show_all_locks();
6130 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
6132 idle->sched_class = &idle_sched_class;
6136 * init_idle - set up an idle thread for a given CPU
6137 * @idle: task in question
6138 * @cpu: cpu the idle task belongs to
6140 * NOTE: this function does not set the idle thread's NEED_RESCHED
6141 * flag, to make booting more robust.
6143 void __cpuinit init_idle(struct task_struct *idle, int cpu)
6145 struct rq *rq = cpu_rq(cpu);
6146 unsigned long flags;
6148 spin_lock_irqsave(&rq->lock, flags);
6151 idle->se.exec_start = sched_clock();
6153 idle->prio = idle->normal_prio = MAX_PRIO;
6154 cpumask_copy(&idle->cpus_allowed, cpumask_of(cpu));
6155 __set_task_cpu(idle, cpu);
6157 rq->curr = rq->idle = idle;
6158 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6161 spin_unlock_irqrestore(&rq->lock, flags);
6163 /* Set the preempt count _outside_ the spinlocks! */
6164 #if defined(CONFIG_PREEMPT)
6165 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
6167 task_thread_info(idle)->preempt_count = 0;
6170 * The idle tasks have their own, simple scheduling class:
6172 idle->sched_class = &idle_sched_class;
6173 ftrace_graph_init_task(idle);
6177 * In a system that switches off the HZ timer nohz_cpu_mask
6178 * indicates which cpus entered this state. This is used
6179 * in the rcu update to wait only for active cpus. For system
6180 * which do not switch off the HZ timer nohz_cpu_mask should
6181 * always be CPU_BITS_NONE.
6183 cpumask_var_t nohz_cpu_mask;
6186 * Increase the granularity value when there are more CPUs,
6187 * because with more CPUs the 'effective latency' as visible
6188 * to users decreases. But the relationship is not linear,
6189 * so pick a second-best guess by going with the log2 of the
6192 * This idea comes from the SD scheduler of Con Kolivas:
6194 static inline void sched_init_granularity(void)
6196 unsigned int factor = 1 + ilog2(num_online_cpus());
6197 const unsigned long limit = 200000000;
6199 sysctl_sched_min_granularity *= factor;
6200 if (sysctl_sched_min_granularity > limit)
6201 sysctl_sched_min_granularity = limit;
6203 sysctl_sched_latency *= factor;
6204 if (sysctl_sched_latency > limit)
6205 sysctl_sched_latency = limit;
6207 sysctl_sched_wakeup_granularity *= factor;
6209 sysctl_sched_shares_ratelimit *= factor;
6214 * This is how migration works:
6216 * 1) we queue a struct migration_req structure in the source CPU's
6217 * runqueue and wake up that CPU's migration thread.
6218 * 2) we down() the locked semaphore => thread blocks.
6219 * 3) migration thread wakes up (implicitly it forces the migrated
6220 * thread off the CPU)
6221 * 4) it gets the migration request and checks whether the migrated
6222 * task is still in the wrong runqueue.
6223 * 5) if it's in the wrong runqueue then the migration thread removes
6224 * it and puts it into the right queue.
6225 * 6) migration thread up()s the semaphore.
6226 * 7) we wake up and the migration is done.
6230 * Change a given task's CPU affinity. Migrate the thread to a
6231 * proper CPU and schedule it away if the CPU it's executing on
6232 * is removed from the allowed bitmask.
6234 * NOTE: the caller must have a valid reference to the task, the
6235 * task must not exit() & deallocate itself prematurely. The
6236 * call is not atomic; no spinlocks may be held.
6238 int set_cpus_allowed_ptr(struct task_struct *p, const struct cpumask *new_mask)
6240 struct migration_req req;
6241 unsigned long flags;
6245 rq = task_rq_lock(p, &flags);
6246 if (!cpumask_intersects(new_mask, cpu_online_mask)) {
6251 if (unlikely((p->flags & PF_THREAD_BOUND) && p != current &&
6252 !cpumask_equal(&p->cpus_allowed, new_mask))) {
6257 if (p->sched_class->set_cpus_allowed)
6258 p->sched_class->set_cpus_allowed(p, new_mask);
6260 cpumask_copy(&p->cpus_allowed, new_mask);
6261 p->rt.nr_cpus_allowed = cpumask_weight(new_mask);
6264 /* Can the task run on the task's current CPU? If so, we're done */
6265 if (cpumask_test_cpu(task_cpu(p), new_mask))
6268 if (migrate_task(p, cpumask_any_and(cpu_online_mask, new_mask), &req)) {
6269 /* Need help from migration thread: drop lock and wait. */
6270 task_rq_unlock(rq, &flags);
6271 wake_up_process(rq->migration_thread);
6272 wait_for_completion(&req.done);
6273 tlb_migrate_finish(p->mm);
6277 task_rq_unlock(rq, &flags);
6281 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
6284 * Move (not current) task off this cpu, onto dest cpu. We're doing
6285 * this because either it can't run here any more (set_cpus_allowed()
6286 * away from this CPU, or CPU going down), or because we're
6287 * attempting to rebalance this task on exec (sched_exec).
6289 * So we race with normal scheduler movements, but that's OK, as long
6290 * as the task is no longer on this CPU.
6292 * Returns non-zero if task was successfully migrated.
6294 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6296 struct rq *rq_dest, *rq_src;
6299 if (unlikely(!cpu_active(dest_cpu)))
6302 rq_src = cpu_rq(src_cpu);
6303 rq_dest = cpu_rq(dest_cpu);
6305 double_rq_lock(rq_src, rq_dest);
6306 /* Already moved. */
6307 if (task_cpu(p) != src_cpu)
6309 /* Affinity changed (again). */
6310 if (!cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6313 on_rq = p->se.on_rq;
6315 deactivate_task(rq_src, p, 0);
6317 set_task_cpu(p, dest_cpu);
6319 activate_task(rq_dest, p, 0);
6320 check_preempt_curr(rq_dest, p, 0);
6325 double_rq_unlock(rq_src, rq_dest);
6330 * migration_thread - this is a highprio system thread that performs
6331 * thread migration by bumping thread off CPU then 'pushing' onto
6334 static int migration_thread(void *data)
6336 int cpu = (long)data;
6340 BUG_ON(rq->migration_thread != current);
6342 set_current_state(TASK_INTERRUPTIBLE);
6343 while (!kthread_should_stop()) {
6344 struct migration_req *req;
6345 struct list_head *head;
6347 spin_lock_irq(&rq->lock);
6349 if (cpu_is_offline(cpu)) {
6350 spin_unlock_irq(&rq->lock);
6354 if (rq->active_balance) {
6355 active_load_balance(rq, cpu);
6356 rq->active_balance = 0;
6359 head = &rq->migration_queue;
6361 if (list_empty(head)) {
6362 spin_unlock_irq(&rq->lock);
6364 set_current_state(TASK_INTERRUPTIBLE);
6367 req = list_entry(head->next, struct migration_req, list);
6368 list_del_init(head->next);
6370 spin_unlock(&rq->lock);
6371 __migrate_task(req->task, cpu, req->dest_cpu);
6374 complete(&req->done);
6376 __set_current_state(TASK_RUNNING);
6380 /* Wait for kthread_stop */
6381 set_current_state(TASK_INTERRUPTIBLE);
6382 while (!kthread_should_stop()) {
6384 set_current_state(TASK_INTERRUPTIBLE);
6386 __set_current_state(TASK_RUNNING);
6390 #ifdef CONFIG_HOTPLUG_CPU
6392 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6396 local_irq_disable();
6397 ret = __migrate_task(p, src_cpu, dest_cpu);
6403 * Figure out where task on dead CPU should go, use force if necessary.
6405 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6408 const struct cpumask *nodemask = cpumask_of_node(cpu_to_node(dead_cpu));
6411 /* Look for allowed, online CPU in same node. */
6412 for_each_cpu_and(dest_cpu, nodemask, cpu_online_mask)
6413 if (cpumask_test_cpu(dest_cpu, &p->cpus_allowed))
6416 /* Any allowed, online CPU? */
6417 dest_cpu = cpumask_any_and(&p->cpus_allowed, cpu_online_mask);
6418 if (dest_cpu < nr_cpu_ids)
6421 /* No more Mr. Nice Guy. */
6422 if (dest_cpu >= nr_cpu_ids) {
6423 cpuset_cpus_allowed_locked(p, &p->cpus_allowed);
6424 dest_cpu = cpumask_any_and(cpu_online_mask, &p->cpus_allowed);
6427 * Don't tell them about moving exiting tasks or
6428 * kernel threads (both mm NULL), since they never
6431 if (p->mm && printk_ratelimit()) {
6432 printk(KERN_INFO "process %d (%s) no "
6433 "longer affine to cpu%d\n",
6434 task_pid_nr(p), p->comm, dead_cpu);
6439 /* It can have affinity changed while we were choosing. */
6440 if (unlikely(!__migrate_task_irq(p, dead_cpu, dest_cpu)))
6445 * While a dead CPU has no uninterruptible tasks queued at this point,
6446 * it might still have a nonzero ->nr_uninterruptible counter, because
6447 * for performance reasons the counter is not stricly tracking tasks to
6448 * their home CPUs. So we just add the counter to another CPU's counter,
6449 * to keep the global sum constant after CPU-down:
6451 static void migrate_nr_uninterruptible(struct rq *rq_src)
6453 struct rq *rq_dest = cpu_rq(cpumask_any(cpu_online_mask));
6454 unsigned long flags;
6456 local_irq_save(flags);
6457 double_rq_lock(rq_src, rq_dest);
6458 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6459 rq_src->nr_uninterruptible = 0;
6460 double_rq_unlock(rq_src, rq_dest);
6461 local_irq_restore(flags);
6464 /* Run through task list and migrate tasks from the dead cpu. */
6465 static void migrate_live_tasks(int src_cpu)
6467 struct task_struct *p, *t;
6469 read_lock(&tasklist_lock);
6471 do_each_thread(t, p) {
6475 if (task_cpu(p) == src_cpu)
6476 move_task_off_dead_cpu(src_cpu, p);
6477 } while_each_thread(t, p);
6479 read_unlock(&tasklist_lock);
6483 * Schedules idle task to be the next runnable task on current CPU.
6484 * It does so by boosting its priority to highest possible.
6485 * Used by CPU offline code.
6487 void sched_idle_next(void)
6489 int this_cpu = smp_processor_id();
6490 struct rq *rq = cpu_rq(this_cpu);
6491 struct task_struct *p = rq->idle;
6492 unsigned long flags;
6494 /* cpu has to be offline */
6495 BUG_ON(cpu_online(this_cpu));
6498 * Strictly not necessary since rest of the CPUs are stopped by now
6499 * and interrupts disabled on the current cpu.
6501 spin_lock_irqsave(&rq->lock, flags);
6503 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6505 update_rq_clock(rq);
6506 activate_task(rq, p, 0);
6508 spin_unlock_irqrestore(&rq->lock, flags);
6512 * Ensures that the idle task is using init_mm right before its cpu goes
6515 void idle_task_exit(void)
6517 struct mm_struct *mm = current->active_mm;
6519 BUG_ON(cpu_online(smp_processor_id()));
6522 switch_mm(mm, &init_mm, current);
6526 /* called under rq->lock with disabled interrupts */
6527 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6529 struct rq *rq = cpu_rq(dead_cpu);
6531 /* Must be exiting, otherwise would be on tasklist. */
6532 BUG_ON(!p->exit_state);
6534 /* Cannot have done final schedule yet: would have vanished. */
6535 BUG_ON(p->state == TASK_DEAD);
6540 * Drop lock around migration; if someone else moves it,
6541 * that's OK. No task can be added to this CPU, so iteration is
6544 spin_unlock_irq(&rq->lock);
6545 move_task_off_dead_cpu(dead_cpu, p);
6546 spin_lock_irq(&rq->lock);
6551 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6552 static void migrate_dead_tasks(unsigned int dead_cpu)
6554 struct rq *rq = cpu_rq(dead_cpu);
6555 struct task_struct *next;
6558 if (!rq->nr_running)
6560 update_rq_clock(rq);
6561 next = pick_next_task(rq, rq->curr);
6564 next->sched_class->put_prev_task(rq, next);
6565 migrate_dead(dead_cpu, next);
6569 #endif /* CONFIG_HOTPLUG_CPU */
6571 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6573 static struct ctl_table sd_ctl_dir[] = {
6575 .procname = "sched_domain",
6581 static struct ctl_table sd_ctl_root[] = {
6583 .ctl_name = CTL_KERN,
6584 .procname = "kernel",
6586 .child = sd_ctl_dir,
6591 static struct ctl_table *sd_alloc_ctl_entry(int n)
6593 struct ctl_table *entry =
6594 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6599 static void sd_free_ctl_entry(struct ctl_table **tablep)
6601 struct ctl_table *entry;
6604 * In the intermediate directories, both the child directory and
6605 * procname are dynamically allocated and could fail but the mode
6606 * will always be set. In the lowest directory the names are
6607 * static strings and all have proc handlers.
6609 for (entry = *tablep; entry->mode; entry++) {
6611 sd_free_ctl_entry(&entry->child);
6612 if (entry->proc_handler == NULL)
6613 kfree(entry->procname);
6621 set_table_entry(struct ctl_table *entry,
6622 const char *procname, void *data, int maxlen,
6623 mode_t mode, proc_handler *proc_handler)
6625 entry->procname = procname;
6627 entry->maxlen = maxlen;
6629 entry->proc_handler = proc_handler;
6632 static struct ctl_table *
6633 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6635 struct ctl_table *table = sd_alloc_ctl_entry(13);
6640 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6641 sizeof(long), 0644, proc_doulongvec_minmax);
6642 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6643 sizeof(long), 0644, proc_doulongvec_minmax);
6644 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6645 sizeof(int), 0644, proc_dointvec_minmax);
6646 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6647 sizeof(int), 0644, proc_dointvec_minmax);
6648 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6649 sizeof(int), 0644, proc_dointvec_minmax);
6650 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6651 sizeof(int), 0644, proc_dointvec_minmax);
6652 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6653 sizeof(int), 0644, proc_dointvec_minmax);
6654 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6655 sizeof(int), 0644, proc_dointvec_minmax);
6656 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6657 sizeof(int), 0644, proc_dointvec_minmax);
6658 set_table_entry(&table[9], "cache_nice_tries",
6659 &sd->cache_nice_tries,
6660 sizeof(int), 0644, proc_dointvec_minmax);
6661 set_table_entry(&table[10], "flags", &sd->flags,
6662 sizeof(int), 0644, proc_dointvec_minmax);
6663 set_table_entry(&table[11], "name", sd->name,
6664 CORENAME_MAX_SIZE, 0444, proc_dostring);
6665 /* &table[12] is terminator */
6670 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6672 struct ctl_table *entry, *table;
6673 struct sched_domain *sd;
6674 int domain_num = 0, i;
6677 for_each_domain(cpu, sd)
6679 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6684 for_each_domain(cpu, sd) {
6685 snprintf(buf, 32, "domain%d", i);
6686 entry->procname = kstrdup(buf, GFP_KERNEL);
6688 entry->child = sd_alloc_ctl_domain_table(sd);
6695 static struct ctl_table_header *sd_sysctl_header;
6696 static void register_sched_domain_sysctl(void)
6698 int i, cpu_num = num_online_cpus();
6699 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6702 WARN_ON(sd_ctl_dir[0].child);
6703 sd_ctl_dir[0].child = entry;
6708 for_each_online_cpu(i) {
6709 snprintf(buf, 32, "cpu%d", i);
6710 entry->procname = kstrdup(buf, GFP_KERNEL);
6712 entry->child = sd_alloc_ctl_cpu_table(i);
6716 WARN_ON(sd_sysctl_header);
6717 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6720 /* may be called multiple times per register */
6721 static void unregister_sched_domain_sysctl(void)
6723 if (sd_sysctl_header)
6724 unregister_sysctl_table(sd_sysctl_header);
6725 sd_sysctl_header = NULL;
6726 if (sd_ctl_dir[0].child)
6727 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6730 static void register_sched_domain_sysctl(void)
6733 static void unregister_sched_domain_sysctl(void)
6738 static void set_rq_online(struct rq *rq)
6741 const struct sched_class *class;
6743 cpumask_set_cpu(rq->cpu, rq->rd->online);
6746 for_each_class(class) {
6747 if (class->rq_online)
6748 class->rq_online(rq);
6753 static void set_rq_offline(struct rq *rq)
6756 const struct sched_class *class;
6758 for_each_class(class) {
6759 if (class->rq_offline)
6760 class->rq_offline(rq);
6763 cpumask_clear_cpu(rq->cpu, rq->rd->online);
6769 * migration_call - callback that gets triggered when a CPU is added.
6770 * Here we can start up the necessary migration thread for the new CPU.
6772 static int __cpuinit
6773 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6775 struct task_struct *p;
6776 int cpu = (long)hcpu;
6777 unsigned long flags;
6782 case CPU_UP_PREPARE:
6783 case CPU_UP_PREPARE_FROZEN:
6784 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6787 kthread_bind(p, cpu);
6788 /* Must be high prio: stop_machine expects to yield to it. */
6789 rq = task_rq_lock(p, &flags);
6790 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6791 task_rq_unlock(rq, &flags);
6792 cpu_rq(cpu)->migration_thread = p;
6796 case CPU_ONLINE_FROZEN:
6797 /* Strictly unnecessary, as first user will wake it. */
6798 wake_up_process(cpu_rq(cpu)->migration_thread);
6800 /* Update our root-domain */
6802 spin_lock_irqsave(&rq->lock, flags);
6804 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6808 spin_unlock_irqrestore(&rq->lock, flags);
6811 #ifdef CONFIG_HOTPLUG_CPU
6812 case CPU_UP_CANCELED:
6813 case CPU_UP_CANCELED_FROZEN:
6814 if (!cpu_rq(cpu)->migration_thread)
6816 /* Unbind it from offline cpu so it can run. Fall thru. */
6817 kthread_bind(cpu_rq(cpu)->migration_thread,
6818 cpumask_any(cpu_online_mask));
6819 kthread_stop(cpu_rq(cpu)->migration_thread);
6820 cpu_rq(cpu)->migration_thread = NULL;
6824 case CPU_DEAD_FROZEN:
6825 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6826 migrate_live_tasks(cpu);
6828 kthread_stop(rq->migration_thread);
6829 rq->migration_thread = NULL;
6830 /* Idle task back to normal (off runqueue, low prio) */
6831 spin_lock_irq(&rq->lock);
6832 update_rq_clock(rq);
6833 deactivate_task(rq, rq->idle, 0);
6834 rq->idle->static_prio = MAX_PRIO;
6835 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6836 rq->idle->sched_class = &idle_sched_class;
6837 migrate_dead_tasks(cpu);
6838 spin_unlock_irq(&rq->lock);
6840 migrate_nr_uninterruptible(rq);
6841 BUG_ON(rq->nr_running != 0);
6844 * No need to migrate the tasks: it was best-effort if
6845 * they didn't take sched_hotcpu_mutex. Just wake up
6848 spin_lock_irq(&rq->lock);
6849 while (!list_empty(&rq->migration_queue)) {
6850 struct migration_req *req;
6852 req = list_entry(rq->migration_queue.next,
6853 struct migration_req, list);
6854 list_del_init(&req->list);
6855 spin_unlock_irq(&rq->lock);
6856 complete(&req->done);
6857 spin_lock_irq(&rq->lock);
6859 spin_unlock_irq(&rq->lock);
6863 case CPU_DYING_FROZEN:
6864 /* Update our root-domain */
6866 spin_lock_irqsave(&rq->lock, flags);
6868 BUG_ON(!cpumask_test_cpu(cpu, rq->rd->span));
6871 spin_unlock_irqrestore(&rq->lock, flags);
6878 /* Register at highest priority so that task migration (migrate_all_tasks)
6879 * happens before everything else.
6881 static struct notifier_block __cpuinitdata migration_notifier = {
6882 .notifier_call = migration_call,
6886 static int __init migration_init(void)
6888 void *cpu = (void *)(long)smp_processor_id();
6891 /* Start one for the boot CPU: */
6892 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6893 BUG_ON(err == NOTIFY_BAD);
6894 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6895 register_cpu_notifier(&migration_notifier);
6899 early_initcall(migration_init);
6904 #ifdef CONFIG_SCHED_DEBUG
6906 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6907 struct cpumask *groupmask)
6909 struct sched_group *group = sd->groups;
6912 cpulist_scnprintf(str, sizeof(str), sched_domain_span(sd));
6913 cpumask_clear(groupmask);
6915 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6917 if (!(sd->flags & SD_LOAD_BALANCE)) {
6918 printk("does not load-balance\n");
6920 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6925 printk(KERN_CONT "span %s level %s\n", str, sd->name);
6927 if (!cpumask_test_cpu(cpu, sched_domain_span(sd))) {
6928 printk(KERN_ERR "ERROR: domain->span does not contain "
6931 if (!cpumask_test_cpu(cpu, sched_group_cpus(group))) {
6932 printk(KERN_ERR "ERROR: domain->groups does not contain"
6936 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6940 printk(KERN_ERR "ERROR: group is NULL\n");
6944 if (!group->__cpu_power) {
6945 printk(KERN_CONT "\n");
6946 printk(KERN_ERR "ERROR: domain->cpu_power not "
6951 if (!cpumask_weight(sched_group_cpus(group))) {
6952 printk(KERN_CONT "\n");
6953 printk(KERN_ERR "ERROR: empty group\n");
6957 if (cpumask_intersects(groupmask, sched_group_cpus(group))) {
6958 printk(KERN_CONT "\n");
6959 printk(KERN_ERR "ERROR: repeated CPUs\n");
6963 cpumask_or(groupmask, groupmask, sched_group_cpus(group));
6965 cpulist_scnprintf(str, sizeof(str), sched_group_cpus(group));
6966 printk(KERN_CONT " %s (__cpu_power = %d)", str,
6967 group->__cpu_power);
6969 group = group->next;
6970 } while (group != sd->groups);
6971 printk(KERN_CONT "\n");
6973 if (!cpumask_equal(sched_domain_span(sd), groupmask))
6974 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6977 !cpumask_subset(groupmask, sched_domain_span(sd->parent)))
6978 printk(KERN_ERR "ERROR: parent span is not a superset "
6979 "of domain->span\n");
6983 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6985 cpumask_var_t groupmask;
6989 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6993 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6995 if (!alloc_cpumask_var(&groupmask, GFP_KERNEL)) {
6996 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
7001 if (sched_domain_debug_one(sd, cpu, level, groupmask))
7008 free_cpumask_var(groupmask);
7010 #else /* !CONFIG_SCHED_DEBUG */
7011 # define sched_domain_debug(sd, cpu) do { } while (0)
7012 #endif /* CONFIG_SCHED_DEBUG */
7014 static int sd_degenerate(struct sched_domain *sd)
7016 if (cpumask_weight(sched_domain_span(sd)) == 1)
7019 /* Following flags need at least 2 groups */
7020 if (sd->flags & (SD_LOAD_BALANCE |
7021 SD_BALANCE_NEWIDLE |
7025 SD_SHARE_PKG_RESOURCES)) {
7026 if (sd->groups != sd->groups->next)
7030 /* Following flags don't use groups */
7031 if (sd->flags & (SD_WAKE_IDLE |
7040 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
7042 unsigned long cflags = sd->flags, pflags = parent->flags;
7044 if (sd_degenerate(parent))
7047 if (!cpumask_equal(sched_domain_span(sd), sched_domain_span(parent)))
7050 /* Does parent contain flags not in child? */
7051 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7052 if (cflags & SD_WAKE_AFFINE)
7053 pflags &= ~SD_WAKE_BALANCE;
7054 /* Flags needing groups don't count if only 1 group in parent */
7055 if (parent->groups == parent->groups->next) {
7056 pflags &= ~(SD_LOAD_BALANCE |
7057 SD_BALANCE_NEWIDLE |
7061 SD_SHARE_PKG_RESOURCES);
7062 if (nr_node_ids == 1)
7063 pflags &= ~SD_SERIALIZE;
7065 if (~cflags & pflags)
7071 static void free_rootdomain(struct root_domain *rd)
7073 cpupri_cleanup(&rd->cpupri);
7075 free_cpumask_var(rd->rto_mask);
7076 free_cpumask_var(rd->online);
7077 free_cpumask_var(rd->span);
7081 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
7083 unsigned long flags;
7085 spin_lock_irqsave(&rq->lock, flags);
7088 struct root_domain *old_rd = rq->rd;
7090 if (cpumask_test_cpu(rq->cpu, old_rd->online))
7093 cpumask_clear_cpu(rq->cpu, old_rd->span);
7095 if (atomic_dec_and_test(&old_rd->refcount))
7096 free_rootdomain(old_rd);
7099 atomic_inc(&rd->refcount);
7102 cpumask_set_cpu(rq->cpu, rd->span);
7103 if (cpumask_test_cpu(rq->cpu, cpu_online_mask))
7106 spin_unlock_irqrestore(&rq->lock, flags);
7109 static int __init_refok init_rootdomain(struct root_domain *rd, bool bootmem)
7111 memset(rd, 0, sizeof(*rd));
7114 alloc_bootmem_cpumask_var(&def_root_domain.span);
7115 alloc_bootmem_cpumask_var(&def_root_domain.online);
7116 alloc_bootmem_cpumask_var(&def_root_domain.rto_mask);
7117 cpupri_init(&rd->cpupri, true);
7121 if (!alloc_cpumask_var(&rd->span, GFP_KERNEL))
7123 if (!alloc_cpumask_var(&rd->online, GFP_KERNEL))
7125 if (!alloc_cpumask_var(&rd->rto_mask, GFP_KERNEL))
7128 if (cpupri_init(&rd->cpupri, false) != 0)
7133 free_cpumask_var(rd->rto_mask);
7135 free_cpumask_var(rd->online);
7137 free_cpumask_var(rd->span);
7142 static void init_defrootdomain(void)
7144 init_rootdomain(&def_root_domain, true);
7146 atomic_set(&def_root_domain.refcount, 1);
7149 static struct root_domain *alloc_rootdomain(void)
7151 struct root_domain *rd;
7153 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
7157 if (init_rootdomain(rd, false) != 0) {
7166 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7167 * hold the hotplug lock.
7170 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
7172 struct rq *rq = cpu_rq(cpu);
7173 struct sched_domain *tmp;
7175 /* Remove the sched domains which do not contribute to scheduling. */
7176 for (tmp = sd; tmp; ) {
7177 struct sched_domain *parent = tmp->parent;
7181 if (sd_parent_degenerate(tmp, parent)) {
7182 tmp->parent = parent->parent;
7184 parent->parent->child = tmp;
7189 if (sd && sd_degenerate(sd)) {
7195 sched_domain_debug(sd, cpu);
7197 rq_attach_root(rq, rd);
7198 rcu_assign_pointer(rq->sd, sd);
7201 /* cpus with isolated domains */
7202 static cpumask_var_t cpu_isolated_map;
7204 /* Setup the mask of cpus configured for isolated domains */
7205 static int __init isolated_cpu_setup(char *str)
7207 cpulist_parse(str, cpu_isolated_map);
7211 __setup("isolcpus=", isolated_cpu_setup);
7214 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7215 * to a function which identifies what group(along with sched group) a CPU
7216 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7217 * (due to the fact that we keep track of groups covered with a struct cpumask).
7219 * init_sched_build_groups will build a circular linked list of the groups
7220 * covered by the given span, and will set each group's ->cpumask correctly,
7221 * and ->cpu_power to 0.
7224 init_sched_build_groups(const struct cpumask *span,
7225 const struct cpumask *cpu_map,
7226 int (*group_fn)(int cpu, const struct cpumask *cpu_map,
7227 struct sched_group **sg,
7228 struct cpumask *tmpmask),
7229 struct cpumask *covered, struct cpumask *tmpmask)
7231 struct sched_group *first = NULL, *last = NULL;
7234 cpumask_clear(covered);
7236 for_each_cpu(i, span) {
7237 struct sched_group *sg;
7238 int group = group_fn(i, cpu_map, &sg, tmpmask);
7241 if (cpumask_test_cpu(i, covered))
7244 cpumask_clear(sched_group_cpus(sg));
7245 sg->__cpu_power = 0;
7247 for_each_cpu(j, span) {
7248 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
7251 cpumask_set_cpu(j, covered);
7252 cpumask_set_cpu(j, sched_group_cpus(sg));
7263 #define SD_NODES_PER_DOMAIN 16
7268 * find_next_best_node - find the next node to include in a sched_domain
7269 * @node: node whose sched_domain we're building
7270 * @used_nodes: nodes already in the sched_domain
7272 * Find the next node to include in a given scheduling domain. Simply
7273 * finds the closest node not already in the @used_nodes map.
7275 * Should use nodemask_t.
7277 static int find_next_best_node(int node, nodemask_t *used_nodes)
7279 int i, n, val, min_val, best_node = 0;
7283 for (i = 0; i < nr_node_ids; i++) {
7284 /* Start at @node */
7285 n = (node + i) % nr_node_ids;
7287 if (!nr_cpus_node(n))
7290 /* Skip already used nodes */
7291 if (node_isset(n, *used_nodes))
7294 /* Simple min distance search */
7295 val = node_distance(node, n);
7297 if (val < min_val) {
7303 node_set(best_node, *used_nodes);
7308 * sched_domain_node_span - get a cpumask for a node's sched_domain
7309 * @node: node whose cpumask we're constructing
7310 * @span: resulting cpumask
7312 * Given a node, construct a good cpumask for its sched_domain to span. It
7313 * should be one that prevents unnecessary balancing, but also spreads tasks
7316 static void sched_domain_node_span(int node, struct cpumask *span)
7318 nodemask_t used_nodes;
7321 cpumask_clear(span);
7322 nodes_clear(used_nodes);
7324 cpumask_or(span, span, cpumask_of_node(node));
7325 node_set(node, used_nodes);
7327 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
7328 int next_node = find_next_best_node(node, &used_nodes);
7330 cpumask_or(span, span, cpumask_of_node(next_node));
7333 #endif /* CONFIG_NUMA */
7335 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
7338 * The cpus mask in sched_group and sched_domain hangs off the end.
7339 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7340 * for nr_cpu_ids < CONFIG_NR_CPUS.
7342 struct static_sched_group {
7343 struct sched_group sg;
7344 DECLARE_BITMAP(cpus, CONFIG_NR_CPUS);
7347 struct static_sched_domain {
7348 struct sched_domain sd;
7349 DECLARE_BITMAP(span, CONFIG_NR_CPUS);
7353 * SMT sched-domains:
7355 #ifdef CONFIG_SCHED_SMT
7356 static DEFINE_PER_CPU(struct static_sched_domain, cpu_domains);
7357 static DEFINE_PER_CPU(struct static_sched_group, sched_group_cpus);
7360 cpu_to_cpu_group(int cpu, const struct cpumask *cpu_map,
7361 struct sched_group **sg, struct cpumask *unused)
7364 *sg = &per_cpu(sched_group_cpus, cpu).sg;
7367 #endif /* CONFIG_SCHED_SMT */
7370 * multi-core sched-domains:
7372 #ifdef CONFIG_SCHED_MC
7373 static DEFINE_PER_CPU(struct static_sched_domain, core_domains);
7374 static DEFINE_PER_CPU(struct static_sched_group, sched_group_core);
7375 #endif /* CONFIG_SCHED_MC */
7377 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7379 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7380 struct sched_group **sg, struct cpumask *mask)
7384 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7385 group = cpumask_first(mask);
7387 *sg = &per_cpu(sched_group_core, group).sg;
7390 #elif defined(CONFIG_SCHED_MC)
7392 cpu_to_core_group(int cpu, const struct cpumask *cpu_map,
7393 struct sched_group **sg, struct cpumask *unused)
7396 *sg = &per_cpu(sched_group_core, cpu).sg;
7401 static DEFINE_PER_CPU(struct static_sched_domain, phys_domains);
7402 static DEFINE_PER_CPU(struct static_sched_group, sched_group_phys);
7405 cpu_to_phys_group(int cpu, const struct cpumask *cpu_map,
7406 struct sched_group **sg, struct cpumask *mask)
7409 #ifdef CONFIG_SCHED_MC
7410 cpumask_and(mask, cpu_coregroup_mask(cpu), cpu_map);
7411 group = cpumask_first(mask);
7412 #elif defined(CONFIG_SCHED_SMT)
7413 cpumask_and(mask, &per_cpu(cpu_sibling_map, cpu), cpu_map);
7414 group = cpumask_first(mask);
7419 *sg = &per_cpu(sched_group_phys, group).sg;
7425 * The init_sched_build_groups can't handle what we want to do with node
7426 * groups, so roll our own. Now each node has its own list of groups which
7427 * gets dynamically allocated.
7429 static DEFINE_PER_CPU(struct static_sched_domain, node_domains);
7430 static struct sched_group ***sched_group_nodes_bycpu;
7432 static DEFINE_PER_CPU(struct static_sched_domain, allnodes_domains);
7433 static DEFINE_PER_CPU(struct static_sched_group, sched_group_allnodes);
7435 static int cpu_to_allnodes_group(int cpu, const struct cpumask *cpu_map,
7436 struct sched_group **sg,
7437 struct cpumask *nodemask)
7441 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(cpu)), cpu_map);
7442 group = cpumask_first(nodemask);
7445 *sg = &per_cpu(sched_group_allnodes, group).sg;
7449 static void init_numa_sched_groups_power(struct sched_group *group_head)
7451 struct sched_group *sg = group_head;
7457 for_each_cpu(j, sched_group_cpus(sg)) {
7458 struct sched_domain *sd;
7460 sd = &per_cpu(phys_domains, j).sd;
7461 if (j != cpumask_first(sched_group_cpus(sd->groups))) {
7463 * Only add "power" once for each
7469 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7472 } while (sg != group_head);
7474 #endif /* CONFIG_NUMA */
7477 /* Free memory allocated for various sched_group structures */
7478 static void free_sched_groups(const struct cpumask *cpu_map,
7479 struct cpumask *nodemask)
7483 for_each_cpu(cpu, cpu_map) {
7484 struct sched_group **sched_group_nodes
7485 = sched_group_nodes_bycpu[cpu];
7487 if (!sched_group_nodes)
7490 for (i = 0; i < nr_node_ids; i++) {
7491 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7493 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7494 if (cpumask_empty(nodemask))
7504 if (oldsg != sched_group_nodes[i])
7507 kfree(sched_group_nodes);
7508 sched_group_nodes_bycpu[cpu] = NULL;
7511 #else /* !CONFIG_NUMA */
7512 static void free_sched_groups(const struct cpumask *cpu_map,
7513 struct cpumask *nodemask)
7516 #endif /* CONFIG_NUMA */
7519 * Initialize sched groups cpu_power.
7521 * cpu_power indicates the capacity of sched group, which is used while
7522 * distributing the load between different sched groups in a sched domain.
7523 * Typically cpu_power for all the groups in a sched domain will be same unless
7524 * there are asymmetries in the topology. If there are asymmetries, group
7525 * having more cpu_power will pickup more load compared to the group having
7528 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7529 * the maximum number of tasks a group can handle in the presence of other idle
7530 * or lightly loaded groups in the same sched domain.
7532 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7534 struct sched_domain *child;
7535 struct sched_group *group;
7537 WARN_ON(!sd || !sd->groups);
7539 if (cpu != cpumask_first(sched_group_cpus(sd->groups)))
7544 sd->groups->__cpu_power = 0;
7547 * For perf policy, if the groups in child domain share resources
7548 * (for example cores sharing some portions of the cache hierarchy
7549 * or SMT), then set this domain groups cpu_power such that each group
7550 * can handle only one task, when there are other idle groups in the
7551 * same sched domain.
7553 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7555 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7556 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7561 * add cpu_power of each child group to this groups cpu_power
7563 group = child->groups;
7565 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7566 group = group->next;
7567 } while (group != child->groups);
7571 * Initializers for schedule domains
7572 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7575 #ifdef CONFIG_SCHED_DEBUG
7576 # define SD_INIT_NAME(sd, type) sd->name = #type
7578 # define SD_INIT_NAME(sd, type) do { } while (0)
7581 #define SD_INIT(sd, type) sd_init_##type(sd)
7583 #define SD_INIT_FUNC(type) \
7584 static noinline void sd_init_##type(struct sched_domain *sd) \
7586 memset(sd, 0, sizeof(*sd)); \
7587 *sd = SD_##type##_INIT; \
7588 sd->level = SD_LV_##type; \
7589 SD_INIT_NAME(sd, type); \
7594 SD_INIT_FUNC(ALLNODES)
7597 #ifdef CONFIG_SCHED_SMT
7598 SD_INIT_FUNC(SIBLING)
7600 #ifdef CONFIG_SCHED_MC
7604 static int default_relax_domain_level = -1;
7606 static int __init setup_relax_domain_level(char *str)
7610 val = simple_strtoul(str, NULL, 0);
7611 if (val < SD_LV_MAX)
7612 default_relax_domain_level = val;
7616 __setup("relax_domain_level=", setup_relax_domain_level);
7618 static void set_domain_attribute(struct sched_domain *sd,
7619 struct sched_domain_attr *attr)
7623 if (!attr || attr->relax_domain_level < 0) {
7624 if (default_relax_domain_level < 0)
7627 request = default_relax_domain_level;
7629 request = attr->relax_domain_level;
7630 if (request < sd->level) {
7631 /* turn off idle balance on this domain */
7632 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7634 /* turn on idle balance on this domain */
7635 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7640 * Build sched domains for a given set of cpus and attach the sched domains
7641 * to the individual cpus
7643 static int __build_sched_domains(const struct cpumask *cpu_map,
7644 struct sched_domain_attr *attr)
7646 int i, err = -ENOMEM;
7647 struct root_domain *rd;
7648 cpumask_var_t nodemask, this_sibling_map, this_core_map, send_covered,
7651 cpumask_var_t domainspan, covered, notcovered;
7652 struct sched_group **sched_group_nodes = NULL;
7653 int sd_allnodes = 0;
7655 if (!alloc_cpumask_var(&domainspan, GFP_KERNEL))
7657 if (!alloc_cpumask_var(&covered, GFP_KERNEL))
7658 goto free_domainspan;
7659 if (!alloc_cpumask_var(¬covered, GFP_KERNEL))
7663 if (!alloc_cpumask_var(&nodemask, GFP_KERNEL))
7664 goto free_notcovered;
7665 if (!alloc_cpumask_var(&this_sibling_map, GFP_KERNEL))
7667 if (!alloc_cpumask_var(&this_core_map, GFP_KERNEL))
7668 goto free_this_sibling_map;
7669 if (!alloc_cpumask_var(&send_covered, GFP_KERNEL))
7670 goto free_this_core_map;
7671 if (!alloc_cpumask_var(&tmpmask, GFP_KERNEL))
7672 goto free_send_covered;
7676 * Allocate the per-node list of sched groups
7678 sched_group_nodes = kcalloc(nr_node_ids, sizeof(struct sched_group *),
7680 if (!sched_group_nodes) {
7681 printk(KERN_WARNING "Can not alloc sched group node list\n");
7686 rd = alloc_rootdomain();
7688 printk(KERN_WARNING "Cannot alloc root domain\n");
7689 goto free_sched_groups;
7693 sched_group_nodes_bycpu[cpumask_first(cpu_map)] = sched_group_nodes;
7697 * Set up domains for cpus specified by the cpu_map.
7699 for_each_cpu(i, cpu_map) {
7700 struct sched_domain *sd = NULL, *p;
7702 cpumask_and(nodemask, cpumask_of_node(cpu_to_node(i)), cpu_map);
7705 if (cpumask_weight(cpu_map) >
7706 SD_NODES_PER_DOMAIN*cpumask_weight(nodemask)) {
7707 sd = &per_cpu(allnodes_domains, i).sd;
7708 SD_INIT(sd, ALLNODES);
7709 set_domain_attribute(sd, attr);
7710 cpumask_copy(sched_domain_span(sd), cpu_map);
7711 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7717 sd = &per_cpu(node_domains, i).sd;
7719 set_domain_attribute(sd, attr);
7720 sched_domain_node_span(cpu_to_node(i), sched_domain_span(sd));
7724 cpumask_and(sched_domain_span(sd),
7725 sched_domain_span(sd), cpu_map);
7729 sd = &per_cpu(phys_domains, i).sd;
7731 set_domain_attribute(sd, attr);
7732 cpumask_copy(sched_domain_span(sd), nodemask);
7736 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7738 #ifdef CONFIG_SCHED_MC
7740 sd = &per_cpu(core_domains, i).sd;
7742 set_domain_attribute(sd, attr);
7743 cpumask_and(sched_domain_span(sd), cpu_map,
7744 cpu_coregroup_mask(i));
7747 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7750 #ifdef CONFIG_SCHED_SMT
7752 sd = &per_cpu(cpu_domains, i).sd;
7753 SD_INIT(sd, SIBLING);
7754 set_domain_attribute(sd, attr);
7755 cpumask_and(sched_domain_span(sd),
7756 &per_cpu(cpu_sibling_map, i), cpu_map);
7759 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7763 #ifdef CONFIG_SCHED_SMT
7764 /* Set up CPU (sibling) groups */
7765 for_each_cpu(i, cpu_map) {
7766 cpumask_and(this_sibling_map,
7767 &per_cpu(cpu_sibling_map, i), cpu_map);
7768 if (i != cpumask_first(this_sibling_map))
7771 init_sched_build_groups(this_sibling_map, cpu_map,
7773 send_covered, tmpmask);
7777 #ifdef CONFIG_SCHED_MC
7778 /* Set up multi-core groups */
7779 for_each_cpu(i, cpu_map) {
7780 cpumask_and(this_core_map, cpu_coregroup_mask(i), cpu_map);
7781 if (i != cpumask_first(this_core_map))
7784 init_sched_build_groups(this_core_map, cpu_map,
7786 send_covered, tmpmask);
7790 /* Set up physical groups */
7791 for (i = 0; i < nr_node_ids; i++) {
7792 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7793 if (cpumask_empty(nodemask))
7796 init_sched_build_groups(nodemask, cpu_map,
7798 send_covered, tmpmask);
7802 /* Set up node groups */
7804 init_sched_build_groups(cpu_map, cpu_map,
7805 &cpu_to_allnodes_group,
7806 send_covered, tmpmask);
7809 for (i = 0; i < nr_node_ids; i++) {
7810 /* Set up node groups */
7811 struct sched_group *sg, *prev;
7814 cpumask_clear(covered);
7815 cpumask_and(nodemask, cpumask_of_node(i), cpu_map);
7816 if (cpumask_empty(nodemask)) {
7817 sched_group_nodes[i] = NULL;
7821 sched_domain_node_span(i, domainspan);
7822 cpumask_and(domainspan, domainspan, cpu_map);
7824 sg = kmalloc_node(sizeof(struct sched_group) + cpumask_size(),
7827 printk(KERN_WARNING "Can not alloc domain group for "
7831 sched_group_nodes[i] = sg;
7832 for_each_cpu(j, nodemask) {
7833 struct sched_domain *sd;
7835 sd = &per_cpu(node_domains, j).sd;
7838 sg->__cpu_power = 0;
7839 cpumask_copy(sched_group_cpus(sg), nodemask);
7841 cpumask_or(covered, covered, nodemask);
7844 for (j = 0; j < nr_node_ids; j++) {
7845 int n = (i + j) % nr_node_ids;
7847 cpumask_complement(notcovered, covered);
7848 cpumask_and(tmpmask, notcovered, cpu_map);
7849 cpumask_and(tmpmask, tmpmask, domainspan);
7850 if (cpumask_empty(tmpmask))
7853 cpumask_and(tmpmask, tmpmask, cpumask_of_node(n));
7854 if (cpumask_empty(tmpmask))
7857 sg = kmalloc_node(sizeof(struct sched_group) +
7862 "Can not alloc domain group for node %d\n", j);
7865 sg->__cpu_power = 0;
7866 cpumask_copy(sched_group_cpus(sg), tmpmask);
7867 sg->next = prev->next;
7868 cpumask_or(covered, covered, tmpmask);
7875 /* Calculate CPU power for physical packages and nodes */
7876 #ifdef CONFIG_SCHED_SMT
7877 for_each_cpu(i, cpu_map) {
7878 struct sched_domain *sd = &per_cpu(cpu_domains, i).sd;
7880 init_sched_groups_power(i, sd);
7883 #ifdef CONFIG_SCHED_MC
7884 for_each_cpu(i, cpu_map) {
7885 struct sched_domain *sd = &per_cpu(core_domains, i).sd;
7887 init_sched_groups_power(i, sd);
7891 for_each_cpu(i, cpu_map) {
7892 struct sched_domain *sd = &per_cpu(phys_domains, i).sd;
7894 init_sched_groups_power(i, sd);
7898 for (i = 0; i < nr_node_ids; i++)
7899 init_numa_sched_groups_power(sched_group_nodes[i]);
7902 struct sched_group *sg;
7904 cpu_to_allnodes_group(cpumask_first(cpu_map), cpu_map, &sg,
7906 init_numa_sched_groups_power(sg);
7910 /* Attach the domains */
7911 for_each_cpu(i, cpu_map) {
7912 struct sched_domain *sd;
7913 #ifdef CONFIG_SCHED_SMT
7914 sd = &per_cpu(cpu_domains, i).sd;
7915 #elif defined(CONFIG_SCHED_MC)
7916 sd = &per_cpu(core_domains, i).sd;
7918 sd = &per_cpu(phys_domains, i).sd;
7920 cpu_attach_domain(sd, rd, i);
7926 free_cpumask_var(tmpmask);
7928 free_cpumask_var(send_covered);
7930 free_cpumask_var(this_core_map);
7931 free_this_sibling_map:
7932 free_cpumask_var(this_sibling_map);
7934 free_cpumask_var(nodemask);
7937 free_cpumask_var(notcovered);
7939 free_cpumask_var(covered);
7941 free_cpumask_var(domainspan);
7948 kfree(sched_group_nodes);
7954 free_sched_groups(cpu_map, tmpmask);
7955 free_rootdomain(rd);
7960 static int build_sched_domains(const struct cpumask *cpu_map)
7962 return __build_sched_domains(cpu_map, NULL);
7965 static struct cpumask *doms_cur; /* current sched domains */
7966 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7967 static struct sched_domain_attr *dattr_cur;
7968 /* attribues of custom domains in 'doms_cur' */
7971 * Special case: If a kmalloc of a doms_cur partition (array of
7972 * cpumask) fails, then fallback to a single sched domain,
7973 * as determined by the single cpumask fallback_doms.
7975 static cpumask_var_t fallback_doms;
7978 * arch_update_cpu_topology lets virtualized architectures update the
7979 * cpu core maps. It is supposed to return 1 if the topology changed
7980 * or 0 if it stayed the same.
7982 int __attribute__((weak)) arch_update_cpu_topology(void)
7988 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7989 * For now this just excludes isolated cpus, but could be used to
7990 * exclude other special cases in the future.
7992 static int arch_init_sched_domains(const struct cpumask *cpu_map)
7996 arch_update_cpu_topology();
7998 doms_cur = kmalloc(cpumask_size(), GFP_KERNEL);
8000 doms_cur = fallback_doms;
8001 cpumask_andnot(doms_cur, cpu_map, cpu_isolated_map);
8003 err = build_sched_domains(doms_cur);
8004 register_sched_domain_sysctl();
8009 static void arch_destroy_sched_domains(const struct cpumask *cpu_map,
8010 struct cpumask *tmpmask)
8012 free_sched_groups(cpu_map, tmpmask);
8016 * Detach sched domains from a group of cpus specified in cpu_map
8017 * These cpus will now be attached to the NULL domain
8019 static void detach_destroy_domains(const struct cpumask *cpu_map)
8021 /* Save because hotplug lock held. */
8022 static DECLARE_BITMAP(tmpmask, CONFIG_NR_CPUS);
8025 for_each_cpu(i, cpu_map)
8026 cpu_attach_domain(NULL, &def_root_domain, i);
8027 synchronize_sched();
8028 arch_destroy_sched_domains(cpu_map, to_cpumask(tmpmask));
8031 /* handle null as "default" */
8032 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
8033 struct sched_domain_attr *new, int idx_new)
8035 struct sched_domain_attr tmp;
8042 return !memcmp(cur ? (cur + idx_cur) : &tmp,
8043 new ? (new + idx_new) : &tmp,
8044 sizeof(struct sched_domain_attr));
8048 * Partition sched domains as specified by the 'ndoms_new'
8049 * cpumasks in the array doms_new[] of cpumasks. This compares
8050 * doms_new[] to the current sched domain partitioning, doms_cur[].
8051 * It destroys each deleted domain and builds each new domain.
8053 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8054 * The masks don't intersect (don't overlap.) We should setup one
8055 * sched domain for each mask. CPUs not in any of the cpumasks will
8056 * not be load balanced. If the same cpumask appears both in the
8057 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8060 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8061 * ownership of it and will kfree it when done with it. If the caller
8062 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8063 * ndoms_new == 1, and partition_sched_domains() will fallback to
8064 * the single partition 'fallback_doms', it also forces the domains
8067 * If doms_new == NULL it will be replaced with cpu_online_mask.
8068 * ndoms_new == 0 is a special case for destroying existing domains,
8069 * and it will not create the default domain.
8071 * Call with hotplug lock held
8073 /* FIXME: Change to struct cpumask *doms_new[] */
8074 void partition_sched_domains(int ndoms_new, struct cpumask *doms_new,
8075 struct sched_domain_attr *dattr_new)
8080 mutex_lock(&sched_domains_mutex);
8082 /* always unregister in case we don't destroy any domains */
8083 unregister_sched_domain_sysctl();
8085 /* Let architecture update cpu core mappings. */
8086 new_topology = arch_update_cpu_topology();
8088 n = doms_new ? ndoms_new : 0;
8090 /* Destroy deleted domains */
8091 for (i = 0; i < ndoms_cur; i++) {
8092 for (j = 0; j < n && !new_topology; j++) {
8093 if (cpumask_equal(&doms_cur[i], &doms_new[j])
8094 && dattrs_equal(dattr_cur, i, dattr_new, j))
8097 /* no match - a current sched domain not in new doms_new[] */
8098 detach_destroy_domains(doms_cur + i);
8103 if (doms_new == NULL) {
8105 doms_new = fallback_doms;
8106 cpumask_andnot(&doms_new[0], cpu_online_mask, cpu_isolated_map);
8107 WARN_ON_ONCE(dattr_new);
8110 /* Build new domains */
8111 for (i = 0; i < ndoms_new; i++) {
8112 for (j = 0; j < ndoms_cur && !new_topology; j++) {
8113 if (cpumask_equal(&doms_new[i], &doms_cur[j])
8114 && dattrs_equal(dattr_new, i, dattr_cur, j))
8117 /* no match - add a new doms_new */
8118 __build_sched_domains(doms_new + i,
8119 dattr_new ? dattr_new + i : NULL);
8124 /* Remember the new sched domains */
8125 if (doms_cur != fallback_doms)
8127 kfree(dattr_cur); /* kfree(NULL) is safe */
8128 doms_cur = doms_new;
8129 dattr_cur = dattr_new;
8130 ndoms_cur = ndoms_new;
8132 register_sched_domain_sysctl();
8134 mutex_unlock(&sched_domains_mutex);
8137 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8138 static void arch_reinit_sched_domains(void)
8142 /* Destroy domains first to force the rebuild */
8143 partition_sched_domains(0, NULL, NULL);
8145 rebuild_sched_domains();
8149 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
8151 unsigned int level = 0;
8153 if (sscanf(buf, "%u", &level) != 1)
8157 * level is always be positive so don't check for
8158 * level < POWERSAVINGS_BALANCE_NONE which is 0
8159 * What happens on 0 or 1 byte write,
8160 * need to check for count as well?
8163 if (level >= MAX_POWERSAVINGS_BALANCE_LEVELS)
8167 sched_smt_power_savings = level;
8169 sched_mc_power_savings = level;
8171 arch_reinit_sched_domains();
8176 #ifdef CONFIG_SCHED_MC
8177 static ssize_t sched_mc_power_savings_show(struct sysdev_class *class,
8180 return sprintf(page, "%u\n", sched_mc_power_savings);
8182 static ssize_t sched_mc_power_savings_store(struct sysdev_class *class,
8183 const char *buf, size_t count)
8185 return sched_power_savings_store(buf, count, 0);
8187 static SYSDEV_CLASS_ATTR(sched_mc_power_savings, 0644,
8188 sched_mc_power_savings_show,
8189 sched_mc_power_savings_store);
8192 #ifdef CONFIG_SCHED_SMT
8193 static ssize_t sched_smt_power_savings_show(struct sysdev_class *dev,
8196 return sprintf(page, "%u\n", sched_smt_power_savings);
8198 static ssize_t sched_smt_power_savings_store(struct sysdev_class *dev,
8199 const char *buf, size_t count)
8201 return sched_power_savings_store(buf, count, 1);
8203 static SYSDEV_CLASS_ATTR(sched_smt_power_savings, 0644,
8204 sched_smt_power_savings_show,
8205 sched_smt_power_savings_store);
8208 int __init sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
8212 #ifdef CONFIG_SCHED_SMT
8214 err = sysfs_create_file(&cls->kset.kobj,
8215 &attr_sched_smt_power_savings.attr);
8217 #ifdef CONFIG_SCHED_MC
8218 if (!err && mc_capable())
8219 err = sysfs_create_file(&cls->kset.kobj,
8220 &attr_sched_mc_power_savings.attr);
8224 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8226 #ifndef CONFIG_CPUSETS
8228 * Add online and remove offline CPUs from the scheduler domains.
8229 * When cpusets are enabled they take over this function.
8231 static int update_sched_domains(struct notifier_block *nfb,
8232 unsigned long action, void *hcpu)
8236 case CPU_ONLINE_FROZEN:
8238 case CPU_DEAD_FROZEN:
8239 partition_sched_domains(1, NULL, NULL);
8248 static int update_runtime(struct notifier_block *nfb,
8249 unsigned long action, void *hcpu)
8251 int cpu = (int)(long)hcpu;
8254 case CPU_DOWN_PREPARE:
8255 case CPU_DOWN_PREPARE_FROZEN:
8256 disable_runtime(cpu_rq(cpu));
8259 case CPU_DOWN_FAILED:
8260 case CPU_DOWN_FAILED_FROZEN:
8262 case CPU_ONLINE_FROZEN:
8263 enable_runtime(cpu_rq(cpu));
8271 void __init sched_init_smp(void)
8273 cpumask_var_t non_isolated_cpus;
8275 alloc_cpumask_var(&non_isolated_cpus, GFP_KERNEL);
8277 #if defined(CONFIG_NUMA)
8278 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
8280 BUG_ON(sched_group_nodes_bycpu == NULL);
8283 mutex_lock(&sched_domains_mutex);
8284 arch_init_sched_domains(cpu_online_mask);
8285 cpumask_andnot(non_isolated_cpus, cpu_possible_mask, cpu_isolated_map);
8286 if (cpumask_empty(non_isolated_cpus))
8287 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus);
8288 mutex_unlock(&sched_domains_mutex);
8291 #ifndef CONFIG_CPUSETS
8292 /* XXX: Theoretical race here - CPU may be hotplugged now */
8293 hotcpu_notifier(update_sched_domains, 0);
8296 /* RT runtime code needs to handle some hotplug events */
8297 hotcpu_notifier(update_runtime, 0);
8301 /* Move init over to a non-isolated CPU */
8302 if (set_cpus_allowed_ptr(current, non_isolated_cpus) < 0)
8304 sched_init_granularity();
8305 free_cpumask_var(non_isolated_cpus);
8307 alloc_cpumask_var(&fallback_doms, GFP_KERNEL);
8308 init_sched_rt_class();
8311 void __init sched_init_smp(void)
8313 sched_init_granularity();
8315 #endif /* CONFIG_SMP */
8317 int in_sched_functions(unsigned long addr)
8319 return in_lock_functions(addr) ||
8320 (addr >= (unsigned long)__sched_text_start
8321 && addr < (unsigned long)__sched_text_end);
8324 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
8326 cfs_rq->tasks_timeline = RB_ROOT;
8327 INIT_LIST_HEAD(&cfs_rq->tasks);
8328 #ifdef CONFIG_FAIR_GROUP_SCHED
8331 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
8334 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
8336 struct rt_prio_array *array;
8339 array = &rt_rq->active;
8340 for (i = 0; i < MAX_RT_PRIO; i++) {
8341 INIT_LIST_HEAD(array->queue + i);
8342 __clear_bit(i, array->bitmap);
8344 /* delimiter for bitsearch: */
8345 __set_bit(MAX_RT_PRIO, array->bitmap);
8347 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8348 rt_rq->highest_prio = MAX_RT_PRIO;
8351 rt_rq->rt_nr_migratory = 0;
8352 rt_rq->overloaded = 0;
8356 rt_rq->rt_throttled = 0;
8357 rt_rq->rt_runtime = 0;
8358 spin_lock_init(&rt_rq->rt_runtime_lock);
8360 #ifdef CONFIG_RT_GROUP_SCHED
8361 rt_rq->rt_nr_boosted = 0;
8366 #ifdef CONFIG_FAIR_GROUP_SCHED
8367 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
8368 struct sched_entity *se, int cpu, int add,
8369 struct sched_entity *parent)
8371 struct rq *rq = cpu_rq(cpu);
8372 tg->cfs_rq[cpu] = cfs_rq;
8373 init_cfs_rq(cfs_rq, rq);
8376 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
8379 /* se could be NULL for init_task_group */
8384 se->cfs_rq = &rq->cfs;
8386 se->cfs_rq = parent->my_q;
8389 se->load.weight = tg->shares;
8390 se->load.inv_weight = 0;
8391 se->parent = parent;
8395 #ifdef CONFIG_RT_GROUP_SCHED
8396 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8397 struct sched_rt_entity *rt_se, int cpu, int add,
8398 struct sched_rt_entity *parent)
8400 struct rq *rq = cpu_rq(cpu);
8402 tg->rt_rq[cpu] = rt_rq;
8403 init_rt_rq(rt_rq, rq);
8405 rt_rq->rt_se = rt_se;
8406 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8408 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8410 tg->rt_se[cpu] = rt_se;
8415 rt_se->rt_rq = &rq->rt;
8417 rt_se->rt_rq = parent->my_q;
8419 rt_se->my_q = rt_rq;
8420 rt_se->parent = parent;
8421 INIT_LIST_HEAD(&rt_se->run_list);
8425 void __init sched_init(void)
8428 unsigned long alloc_size = 0, ptr;
8430 #ifdef CONFIG_FAIR_GROUP_SCHED
8431 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8433 #ifdef CONFIG_RT_GROUP_SCHED
8434 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8436 #ifdef CONFIG_USER_SCHED
8440 * As sched_init() is called before page_alloc is setup,
8441 * we use alloc_bootmem().
8444 ptr = (unsigned long)alloc_bootmem(alloc_size);
8446 #ifdef CONFIG_FAIR_GROUP_SCHED
8447 init_task_group.se = (struct sched_entity **)ptr;
8448 ptr += nr_cpu_ids * sizeof(void **);
8450 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8451 ptr += nr_cpu_ids * sizeof(void **);
8453 #ifdef CONFIG_USER_SCHED
8454 root_task_group.se = (struct sched_entity **)ptr;
8455 ptr += nr_cpu_ids * sizeof(void **);
8457 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8458 ptr += nr_cpu_ids * sizeof(void **);
8459 #endif /* CONFIG_USER_SCHED */
8460 #endif /* CONFIG_FAIR_GROUP_SCHED */
8461 #ifdef CONFIG_RT_GROUP_SCHED
8462 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8463 ptr += nr_cpu_ids * sizeof(void **);
8465 init_task_group.rt_rq = (struct rt_rq **)ptr;
8466 ptr += nr_cpu_ids * sizeof(void **);
8468 #ifdef CONFIG_USER_SCHED
8469 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8470 ptr += nr_cpu_ids * sizeof(void **);
8472 root_task_group.rt_rq = (struct rt_rq **)ptr;
8473 ptr += nr_cpu_ids * sizeof(void **);
8474 #endif /* CONFIG_USER_SCHED */
8475 #endif /* CONFIG_RT_GROUP_SCHED */
8479 init_defrootdomain();
8482 init_rt_bandwidth(&def_rt_bandwidth,
8483 global_rt_period(), global_rt_runtime());
8485 #ifdef CONFIG_RT_GROUP_SCHED
8486 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8487 global_rt_period(), global_rt_runtime());
8488 #ifdef CONFIG_USER_SCHED
8489 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8490 global_rt_period(), RUNTIME_INF);
8491 #endif /* CONFIG_USER_SCHED */
8492 #endif /* CONFIG_RT_GROUP_SCHED */
8494 #ifdef CONFIG_GROUP_SCHED
8495 list_add(&init_task_group.list, &task_groups);
8496 INIT_LIST_HEAD(&init_task_group.children);
8498 #ifdef CONFIG_USER_SCHED
8499 INIT_LIST_HEAD(&root_task_group.children);
8500 init_task_group.parent = &root_task_group;
8501 list_add(&init_task_group.siblings, &root_task_group.children);
8502 #endif /* CONFIG_USER_SCHED */
8503 #endif /* CONFIG_GROUP_SCHED */
8505 for_each_possible_cpu(i) {
8509 spin_lock_init(&rq->lock);
8511 init_cfs_rq(&rq->cfs, rq);
8512 init_rt_rq(&rq->rt, rq);
8513 #ifdef CONFIG_FAIR_GROUP_SCHED
8514 init_task_group.shares = init_task_group_load;
8515 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8516 #ifdef CONFIG_CGROUP_SCHED
8518 * How much cpu bandwidth does init_task_group get?
8520 * In case of task-groups formed thr' the cgroup filesystem, it
8521 * gets 100% of the cpu resources in the system. This overall
8522 * system cpu resource is divided among the tasks of
8523 * init_task_group and its child task-groups in a fair manner,
8524 * based on each entity's (task or task-group's) weight
8525 * (se->load.weight).
8527 * In other words, if init_task_group has 10 tasks of weight
8528 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8529 * then A0's share of the cpu resource is:
8531 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8533 * We achieve this by letting init_task_group's tasks sit
8534 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8536 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8537 #elif defined CONFIG_USER_SCHED
8538 root_task_group.shares = NICE_0_LOAD;
8539 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8541 * In case of task-groups formed thr' the user id of tasks,
8542 * init_task_group represents tasks belonging to root user.
8543 * Hence it forms a sibling of all subsequent groups formed.
8544 * In this case, init_task_group gets only a fraction of overall
8545 * system cpu resource, based on the weight assigned to root
8546 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8547 * by letting tasks of init_task_group sit in a separate cfs_rq
8548 * (init_cfs_rq) and having one entity represent this group of
8549 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8551 init_tg_cfs_entry(&init_task_group,
8552 &per_cpu(init_cfs_rq, i),
8553 &per_cpu(init_sched_entity, i), i, 1,
8554 root_task_group.se[i]);
8557 #endif /* CONFIG_FAIR_GROUP_SCHED */
8559 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8560 #ifdef CONFIG_RT_GROUP_SCHED
8561 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8562 #ifdef CONFIG_CGROUP_SCHED
8563 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8564 #elif defined CONFIG_USER_SCHED
8565 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8566 init_tg_rt_entry(&init_task_group,
8567 &per_cpu(init_rt_rq, i),
8568 &per_cpu(init_sched_rt_entity, i), i, 1,
8569 root_task_group.rt_se[i]);
8573 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8574 rq->cpu_load[j] = 0;
8578 rq->active_balance = 0;
8579 rq->next_balance = jiffies;
8583 rq->migration_thread = NULL;
8584 INIT_LIST_HEAD(&rq->migration_queue);
8585 rq_attach_root(rq, &def_root_domain);
8588 atomic_set(&rq->nr_iowait, 0);
8591 set_load_weight(&init_task);
8593 #ifdef CONFIG_PREEMPT_NOTIFIERS
8594 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8598 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains);
8601 #ifdef CONFIG_RT_MUTEXES
8602 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8606 * The boot idle thread does lazy MMU switching as well:
8608 atomic_inc(&init_mm.mm_count);
8609 enter_lazy_tlb(&init_mm, current);
8612 * Make us the idle thread. Technically, schedule() should not be
8613 * called from this thread, however somewhere below it might be,
8614 * but because we are the idle thread, we just pick up running again
8615 * when this runqueue becomes "idle".
8617 init_idle(current, smp_processor_id());
8619 * During early bootup we pretend to be a normal task:
8621 current->sched_class = &fair_sched_class;
8623 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8624 alloc_bootmem_cpumask_var(&nohz_cpu_mask);
8627 alloc_bootmem_cpumask_var(&nohz.cpu_mask);
8629 alloc_bootmem_cpumask_var(&cpu_isolated_map);
8632 scheduler_running = 1;
8635 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8636 void __might_sleep(char *file, int line)
8639 static unsigned long prev_jiffy; /* ratelimiting */
8641 if ((!in_atomic() && !irqs_disabled()) ||
8642 system_state != SYSTEM_RUNNING || oops_in_progress)
8644 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8646 prev_jiffy = jiffies;
8649 "BUG: sleeping function called from invalid context at %s:%d\n",
8652 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8653 in_atomic(), irqs_disabled(),
8654 current->pid, current->comm);
8656 debug_show_held_locks(current);
8657 if (irqs_disabled())
8658 print_irqtrace_events(current);
8662 EXPORT_SYMBOL(__might_sleep);
8665 #ifdef CONFIG_MAGIC_SYSRQ
8666 static void normalize_task(struct rq *rq, struct task_struct *p)
8670 update_rq_clock(rq);
8671 on_rq = p->se.on_rq;
8673 deactivate_task(rq, p, 0);
8674 __setscheduler(rq, p, SCHED_NORMAL, 0);
8676 activate_task(rq, p, 0);
8677 resched_task(rq->curr);
8681 void normalize_rt_tasks(void)
8683 struct task_struct *g, *p;
8684 unsigned long flags;
8687 read_lock_irqsave(&tasklist_lock, flags);
8688 do_each_thread(g, p) {
8690 * Only normalize user tasks:
8695 p->se.exec_start = 0;
8696 #ifdef CONFIG_SCHEDSTATS
8697 p->se.wait_start = 0;
8698 p->se.sleep_start = 0;
8699 p->se.block_start = 0;
8704 * Renice negative nice level userspace
8707 if (TASK_NICE(p) < 0 && p->mm)
8708 set_user_nice(p, 0);
8712 spin_lock(&p->pi_lock);
8713 rq = __task_rq_lock(p);
8715 normalize_task(rq, p);
8717 __task_rq_unlock(rq);
8718 spin_unlock(&p->pi_lock);
8719 } while_each_thread(g, p);
8721 read_unlock_irqrestore(&tasklist_lock, flags);
8724 #endif /* CONFIG_MAGIC_SYSRQ */
8728 * These functions are only useful for the IA64 MCA handling.
8730 * They can only be called when the whole system has been
8731 * stopped - every CPU needs to be quiescent, and no scheduling
8732 * activity can take place. Using them for anything else would
8733 * be a serious bug, and as a result, they aren't even visible
8734 * under any other configuration.
8738 * curr_task - return the current task for a given cpu.
8739 * @cpu: the processor in question.
8741 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8743 struct task_struct *curr_task(int cpu)
8745 return cpu_curr(cpu);
8749 * set_curr_task - set the current task for a given cpu.
8750 * @cpu: the processor in question.
8751 * @p: the task pointer to set.
8753 * Description: This function must only be used when non-maskable interrupts
8754 * are serviced on a separate stack. It allows the architecture to switch the
8755 * notion of the current task on a cpu in a non-blocking manner. This function
8756 * must be called with all CPU's synchronized, and interrupts disabled, the
8757 * and caller must save the original value of the current task (see
8758 * curr_task() above) and restore that value before reenabling interrupts and
8759 * re-starting the system.
8761 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8763 void set_curr_task(int cpu, struct task_struct *p)
8770 #ifdef CONFIG_FAIR_GROUP_SCHED
8771 static void free_fair_sched_group(struct task_group *tg)
8775 for_each_possible_cpu(i) {
8777 kfree(tg->cfs_rq[i]);
8787 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8789 struct cfs_rq *cfs_rq;
8790 struct sched_entity *se;
8794 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8797 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8801 tg->shares = NICE_0_LOAD;
8803 for_each_possible_cpu(i) {
8806 cfs_rq = kzalloc_node(sizeof(struct cfs_rq),
8807 GFP_KERNEL, cpu_to_node(i));
8811 se = kzalloc_node(sizeof(struct sched_entity),
8812 GFP_KERNEL, cpu_to_node(i));
8816 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent->se[i]);
8825 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8827 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8828 &cpu_rq(cpu)->leaf_cfs_rq_list);
8831 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8833 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8835 #else /* !CONFG_FAIR_GROUP_SCHED */
8836 static inline void free_fair_sched_group(struct task_group *tg)
8841 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8846 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8850 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8853 #endif /* CONFIG_FAIR_GROUP_SCHED */
8855 #ifdef CONFIG_RT_GROUP_SCHED
8856 static void free_rt_sched_group(struct task_group *tg)
8860 destroy_rt_bandwidth(&tg->rt_bandwidth);
8862 for_each_possible_cpu(i) {
8864 kfree(tg->rt_rq[i]);
8866 kfree(tg->rt_se[i]);
8874 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8876 struct rt_rq *rt_rq;
8877 struct sched_rt_entity *rt_se;
8881 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8884 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8888 init_rt_bandwidth(&tg->rt_bandwidth,
8889 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8891 for_each_possible_cpu(i) {
8894 rt_rq = kzalloc_node(sizeof(struct rt_rq),
8895 GFP_KERNEL, cpu_to_node(i));
8899 rt_se = kzalloc_node(sizeof(struct sched_rt_entity),
8900 GFP_KERNEL, cpu_to_node(i));
8904 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent->rt_se[i]);
8913 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8915 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8916 &cpu_rq(cpu)->leaf_rt_rq_list);
8919 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8921 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8923 #else /* !CONFIG_RT_GROUP_SCHED */
8924 static inline void free_rt_sched_group(struct task_group *tg)
8929 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8934 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8938 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8941 #endif /* CONFIG_RT_GROUP_SCHED */
8943 #ifdef CONFIG_GROUP_SCHED
8944 static void free_sched_group(struct task_group *tg)
8946 free_fair_sched_group(tg);
8947 free_rt_sched_group(tg);
8951 /* allocate runqueue etc for a new task group */
8952 struct task_group *sched_create_group(struct task_group *parent)
8954 struct task_group *tg;
8955 unsigned long flags;
8958 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8960 return ERR_PTR(-ENOMEM);
8962 if (!alloc_fair_sched_group(tg, parent))
8965 if (!alloc_rt_sched_group(tg, parent))
8968 spin_lock_irqsave(&task_group_lock, flags);
8969 for_each_possible_cpu(i) {
8970 register_fair_sched_group(tg, i);
8971 register_rt_sched_group(tg, i);
8973 list_add_rcu(&tg->list, &task_groups);
8975 WARN_ON(!parent); /* root should already exist */
8977 tg->parent = parent;
8978 INIT_LIST_HEAD(&tg->children);
8979 list_add_rcu(&tg->siblings, &parent->children);
8980 spin_unlock_irqrestore(&task_group_lock, flags);
8985 free_sched_group(tg);
8986 return ERR_PTR(-ENOMEM);
8989 /* rcu callback to free various structures associated with a task group */
8990 static void free_sched_group_rcu(struct rcu_head *rhp)
8992 /* now it should be safe to free those cfs_rqs */
8993 free_sched_group(container_of(rhp, struct task_group, rcu));
8996 /* Destroy runqueue etc associated with a task group */
8997 void sched_destroy_group(struct task_group *tg)
8999 unsigned long flags;
9002 spin_lock_irqsave(&task_group_lock, flags);
9003 for_each_possible_cpu(i) {
9004 unregister_fair_sched_group(tg, i);
9005 unregister_rt_sched_group(tg, i);
9007 list_del_rcu(&tg->list);
9008 list_del_rcu(&tg->siblings);
9009 spin_unlock_irqrestore(&task_group_lock, flags);
9011 /* wait for possible concurrent references to cfs_rqs complete */
9012 call_rcu(&tg->rcu, free_sched_group_rcu);
9015 /* change task's runqueue when it moves between groups.
9016 * The caller of this function should have put the task in its new group
9017 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9018 * reflect its new group.
9020 void sched_move_task(struct task_struct *tsk)
9023 unsigned long flags;
9026 rq = task_rq_lock(tsk, &flags);
9028 update_rq_clock(rq);
9030 running = task_current(rq, tsk);
9031 on_rq = tsk->se.on_rq;
9034 dequeue_task(rq, tsk, 0);
9035 if (unlikely(running))
9036 tsk->sched_class->put_prev_task(rq, tsk);
9038 set_task_rq(tsk, task_cpu(tsk));
9040 #ifdef CONFIG_FAIR_GROUP_SCHED
9041 if (tsk->sched_class->moved_group)
9042 tsk->sched_class->moved_group(tsk);
9045 if (unlikely(running))
9046 tsk->sched_class->set_curr_task(rq);
9048 enqueue_task(rq, tsk, 0);
9050 task_rq_unlock(rq, &flags);
9052 #endif /* CONFIG_GROUP_SCHED */
9054 #ifdef CONFIG_FAIR_GROUP_SCHED
9055 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
9057 struct cfs_rq *cfs_rq = se->cfs_rq;
9062 dequeue_entity(cfs_rq, se, 0);
9064 se->load.weight = shares;
9065 se->load.inv_weight = 0;
9068 enqueue_entity(cfs_rq, se, 0);
9071 static void set_se_shares(struct sched_entity *se, unsigned long shares)
9073 struct cfs_rq *cfs_rq = se->cfs_rq;
9074 struct rq *rq = cfs_rq->rq;
9075 unsigned long flags;
9077 spin_lock_irqsave(&rq->lock, flags);
9078 __set_se_shares(se, shares);
9079 spin_unlock_irqrestore(&rq->lock, flags);
9082 static DEFINE_MUTEX(shares_mutex);
9084 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
9087 unsigned long flags;
9090 * We can't change the weight of the root cgroup.
9095 if (shares < MIN_SHARES)
9096 shares = MIN_SHARES;
9097 else if (shares > MAX_SHARES)
9098 shares = MAX_SHARES;
9100 mutex_lock(&shares_mutex);
9101 if (tg->shares == shares)
9104 spin_lock_irqsave(&task_group_lock, flags);
9105 for_each_possible_cpu(i)
9106 unregister_fair_sched_group(tg, i);
9107 list_del_rcu(&tg->siblings);
9108 spin_unlock_irqrestore(&task_group_lock, flags);
9110 /* wait for any ongoing reference to this group to finish */
9111 synchronize_sched();
9114 * Now we are free to modify the group's share on each cpu
9115 * w/o tripping rebalance_share or load_balance_fair.
9117 tg->shares = shares;
9118 for_each_possible_cpu(i) {
9122 cfs_rq_set_shares(tg->cfs_rq[i], 0);
9123 set_se_shares(tg->se[i], shares);
9127 * Enable load balance activity on this group, by inserting it back on
9128 * each cpu's rq->leaf_cfs_rq_list.
9130 spin_lock_irqsave(&task_group_lock, flags);
9131 for_each_possible_cpu(i)
9132 register_fair_sched_group(tg, i);
9133 list_add_rcu(&tg->siblings, &tg->parent->children);
9134 spin_unlock_irqrestore(&task_group_lock, flags);
9136 mutex_unlock(&shares_mutex);
9140 unsigned long sched_group_shares(struct task_group *tg)
9146 #ifdef CONFIG_RT_GROUP_SCHED
9148 * Ensure that the real time constraints are schedulable.
9150 static DEFINE_MUTEX(rt_constraints_mutex);
9152 static unsigned long to_ratio(u64 period, u64 runtime)
9154 if (runtime == RUNTIME_INF)
9157 return div64_u64(runtime << 20, period);
9160 /* Must be called with tasklist_lock held */
9161 static inline int tg_has_rt_tasks(struct task_group *tg)
9163 struct task_struct *g, *p;
9165 do_each_thread(g, p) {
9166 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
9168 } while_each_thread(g, p);
9173 struct rt_schedulable_data {
9174 struct task_group *tg;
9179 static int tg_schedulable(struct task_group *tg, void *data)
9181 struct rt_schedulable_data *d = data;
9182 struct task_group *child;
9183 unsigned long total, sum = 0;
9184 u64 period, runtime;
9186 period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9187 runtime = tg->rt_bandwidth.rt_runtime;
9190 period = d->rt_period;
9191 runtime = d->rt_runtime;
9194 #ifdef CONFIG_USER_SCHED
9195 if (tg == &root_task_group) {
9196 period = global_rt_period();
9197 runtime = global_rt_runtime();
9202 * Cannot have more runtime than the period.
9204 if (runtime > period && runtime != RUNTIME_INF)
9208 * Ensure we don't starve existing RT tasks.
9210 if (rt_bandwidth_enabled() && !runtime && tg_has_rt_tasks(tg))
9213 total = to_ratio(period, runtime);
9216 * Nobody can have more than the global setting allows.
9218 if (total > to_ratio(global_rt_period(), global_rt_runtime()))
9222 * The sum of our children's runtime should not exceed our own.
9224 list_for_each_entry_rcu(child, &tg->children, siblings) {
9225 period = ktime_to_ns(child->rt_bandwidth.rt_period);
9226 runtime = child->rt_bandwidth.rt_runtime;
9228 if (child == d->tg) {
9229 period = d->rt_period;
9230 runtime = d->rt_runtime;
9233 sum += to_ratio(period, runtime);
9242 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
9244 struct rt_schedulable_data data = {
9246 .rt_period = period,
9247 .rt_runtime = runtime,
9250 return walk_tg_tree(tg_schedulable, tg_nop, &data);
9253 static int tg_set_bandwidth(struct task_group *tg,
9254 u64 rt_period, u64 rt_runtime)
9258 mutex_lock(&rt_constraints_mutex);
9259 read_lock(&tasklist_lock);
9260 err = __rt_schedulable(tg, rt_period, rt_runtime);
9264 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9265 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
9266 tg->rt_bandwidth.rt_runtime = rt_runtime;
9268 for_each_possible_cpu(i) {
9269 struct rt_rq *rt_rq = tg->rt_rq[i];
9271 spin_lock(&rt_rq->rt_runtime_lock);
9272 rt_rq->rt_runtime = rt_runtime;
9273 spin_unlock(&rt_rq->rt_runtime_lock);
9275 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
9277 read_unlock(&tasklist_lock);
9278 mutex_unlock(&rt_constraints_mutex);
9283 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
9285 u64 rt_runtime, rt_period;
9287 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
9288 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
9289 if (rt_runtime_us < 0)
9290 rt_runtime = RUNTIME_INF;
9292 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9295 long sched_group_rt_runtime(struct task_group *tg)
9299 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
9302 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
9303 do_div(rt_runtime_us, NSEC_PER_USEC);
9304 return rt_runtime_us;
9307 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
9309 u64 rt_runtime, rt_period;
9311 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
9312 rt_runtime = tg->rt_bandwidth.rt_runtime;
9317 return tg_set_bandwidth(tg, rt_period, rt_runtime);
9320 long sched_group_rt_period(struct task_group *tg)
9324 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
9325 do_div(rt_period_us, NSEC_PER_USEC);
9326 return rt_period_us;
9329 static int sched_rt_global_constraints(void)
9331 u64 runtime, period;
9334 if (sysctl_sched_rt_period <= 0)
9337 runtime = global_rt_runtime();
9338 period = global_rt_period();
9341 * Sanity check on the sysctl variables.
9343 if (runtime > period && runtime != RUNTIME_INF)
9346 mutex_lock(&rt_constraints_mutex);
9347 read_lock(&tasklist_lock);
9348 ret = __rt_schedulable(NULL, 0, 0);
9349 read_unlock(&tasklist_lock);
9350 mutex_unlock(&rt_constraints_mutex);
9354 #else /* !CONFIG_RT_GROUP_SCHED */
9355 static int sched_rt_global_constraints(void)
9357 unsigned long flags;
9360 if (sysctl_sched_rt_period <= 0)
9363 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
9364 for_each_possible_cpu(i) {
9365 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
9367 spin_lock(&rt_rq->rt_runtime_lock);
9368 rt_rq->rt_runtime = global_rt_runtime();
9369 spin_unlock(&rt_rq->rt_runtime_lock);
9371 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
9375 #endif /* CONFIG_RT_GROUP_SCHED */
9377 int sched_rt_handler(struct ctl_table *table, int write,
9378 struct file *filp, void __user *buffer, size_t *lenp,
9382 int old_period, old_runtime;
9383 static DEFINE_MUTEX(mutex);
9386 old_period = sysctl_sched_rt_period;
9387 old_runtime = sysctl_sched_rt_runtime;
9389 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
9391 if (!ret && write) {
9392 ret = sched_rt_global_constraints();
9394 sysctl_sched_rt_period = old_period;
9395 sysctl_sched_rt_runtime = old_runtime;
9397 def_rt_bandwidth.rt_runtime = global_rt_runtime();
9398 def_rt_bandwidth.rt_period =
9399 ns_to_ktime(global_rt_period());
9402 mutex_unlock(&mutex);
9407 #ifdef CONFIG_CGROUP_SCHED
9409 /* return corresponding task_group object of a cgroup */
9410 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
9412 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
9413 struct task_group, css);
9416 static struct cgroup_subsys_state *
9417 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
9419 struct task_group *tg, *parent;
9421 if (!cgrp->parent) {
9422 /* This is early initialization for the top cgroup */
9423 return &init_task_group.css;
9426 parent = cgroup_tg(cgrp->parent);
9427 tg = sched_create_group(parent);
9429 return ERR_PTR(-ENOMEM);
9435 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9437 struct task_group *tg = cgroup_tg(cgrp);
9439 sched_destroy_group(tg);
9443 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9444 struct task_struct *tsk)
9446 #ifdef CONFIG_RT_GROUP_SCHED
9447 /* Don't accept realtime tasks when there is no way for them to run */
9448 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9451 /* We don't support RT-tasks being in separate groups */
9452 if (tsk->sched_class != &fair_sched_class)
9460 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9461 struct cgroup *old_cont, struct task_struct *tsk)
9463 sched_move_task(tsk);
9466 #ifdef CONFIG_FAIR_GROUP_SCHED
9467 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9470 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9473 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9475 struct task_group *tg = cgroup_tg(cgrp);
9477 return (u64) tg->shares;
9479 #endif /* CONFIG_FAIR_GROUP_SCHED */
9481 #ifdef CONFIG_RT_GROUP_SCHED
9482 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9485 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9488 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9490 return sched_group_rt_runtime(cgroup_tg(cgrp));
9493 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9496 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9499 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9501 return sched_group_rt_period(cgroup_tg(cgrp));
9503 #endif /* CONFIG_RT_GROUP_SCHED */
9505 static struct cftype cpu_files[] = {
9506 #ifdef CONFIG_FAIR_GROUP_SCHED
9509 .read_u64 = cpu_shares_read_u64,
9510 .write_u64 = cpu_shares_write_u64,
9513 #ifdef CONFIG_RT_GROUP_SCHED
9515 .name = "rt_runtime_us",
9516 .read_s64 = cpu_rt_runtime_read,
9517 .write_s64 = cpu_rt_runtime_write,
9520 .name = "rt_period_us",
9521 .read_u64 = cpu_rt_period_read_uint,
9522 .write_u64 = cpu_rt_period_write_uint,
9527 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9529 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9532 struct cgroup_subsys cpu_cgroup_subsys = {
9534 .create = cpu_cgroup_create,
9535 .destroy = cpu_cgroup_destroy,
9536 .can_attach = cpu_cgroup_can_attach,
9537 .attach = cpu_cgroup_attach,
9538 .populate = cpu_cgroup_populate,
9539 .subsys_id = cpu_cgroup_subsys_id,
9543 #endif /* CONFIG_CGROUP_SCHED */
9545 #ifdef CONFIG_CGROUP_CPUACCT
9548 * CPU accounting code for task groups.
9550 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9551 * (balbir@in.ibm.com).
9554 /* track cpu usage of a group of tasks and its child groups */
9556 struct cgroup_subsys_state css;
9557 /* cpuusage holds pointer to a u64-type object on every cpu */
9559 struct percpu_counter cpustat[CPUACCT_STAT_NSTATS];
9560 struct cpuacct *parent;
9563 struct cgroup_subsys cpuacct_subsys;
9565 /* return cpu accounting group corresponding to this container */
9566 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9568 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9569 struct cpuacct, css);
9572 /* return cpu accounting group to which this task belongs */
9573 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9575 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9576 struct cpuacct, css);
9579 /* create a new cpu accounting group */
9580 static struct cgroup_subsys_state *cpuacct_create(
9581 struct cgroup_subsys *ss, struct cgroup *cgrp)
9583 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9589 ca->cpuusage = alloc_percpu(u64);
9593 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9594 if (percpu_counter_init(&ca->cpustat[i], 0))
9595 goto out_free_counters;
9598 ca->parent = cgroup_ca(cgrp->parent);
9604 percpu_counter_destroy(&ca->cpustat[i]);
9605 free_percpu(ca->cpuusage);
9609 return ERR_PTR(-ENOMEM);
9612 /* destroy an existing cpu accounting group */
9614 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9616 struct cpuacct *ca = cgroup_ca(cgrp);
9619 for (i = 0; i < CPUACCT_STAT_NSTATS; i++)
9620 percpu_counter_destroy(&ca->cpustat[i]);
9621 free_percpu(ca->cpuusage);
9625 static u64 cpuacct_cpuusage_read(struct cpuacct *ca, int cpu)
9627 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9630 #ifndef CONFIG_64BIT
9632 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9634 spin_lock_irq(&cpu_rq(cpu)->lock);
9636 spin_unlock_irq(&cpu_rq(cpu)->lock);
9644 static void cpuacct_cpuusage_write(struct cpuacct *ca, int cpu, u64 val)
9646 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9648 #ifndef CONFIG_64BIT
9650 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9652 spin_lock_irq(&cpu_rq(cpu)->lock);
9654 spin_unlock_irq(&cpu_rq(cpu)->lock);
9660 /* return total cpu usage (in nanoseconds) of a group */
9661 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9663 struct cpuacct *ca = cgroup_ca(cgrp);
9664 u64 totalcpuusage = 0;
9667 for_each_present_cpu(i)
9668 totalcpuusage += cpuacct_cpuusage_read(ca, i);
9670 return totalcpuusage;
9673 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9676 struct cpuacct *ca = cgroup_ca(cgrp);
9685 for_each_present_cpu(i)
9686 cpuacct_cpuusage_write(ca, i, 0);
9692 static int cpuacct_percpu_seq_read(struct cgroup *cgroup, struct cftype *cft,
9695 struct cpuacct *ca = cgroup_ca(cgroup);
9699 for_each_present_cpu(i) {
9700 percpu = cpuacct_cpuusage_read(ca, i);
9701 seq_printf(m, "%llu ", (unsigned long long) percpu);
9703 seq_printf(m, "\n");
9707 static const char *cpuacct_stat_desc[] = {
9708 [CPUACCT_STAT_USER] = "user",
9709 [CPUACCT_STAT_SYSTEM] = "system",
9712 static int cpuacct_stats_show(struct cgroup *cgrp, struct cftype *cft,
9713 struct cgroup_map_cb *cb)
9715 struct cpuacct *ca = cgroup_ca(cgrp);
9718 for (i = 0; i < CPUACCT_STAT_NSTATS; i++) {
9719 s64 val = percpu_counter_read(&ca->cpustat[i]);
9720 val = cputime64_to_clock_t(val);
9721 cb->fill(cb, cpuacct_stat_desc[i], val);
9726 static struct cftype files[] = {
9729 .read_u64 = cpuusage_read,
9730 .write_u64 = cpuusage_write,
9733 .name = "usage_percpu",
9734 .read_seq_string = cpuacct_percpu_seq_read,
9738 .read_map = cpuacct_stats_show,
9742 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9744 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9748 * charge this task's execution time to its accounting group.
9750 * called with rq->lock held.
9752 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9757 if (!cpuacct_subsys.active)
9760 cpu = task_cpu(tsk);
9766 for (; ca; ca = ca->parent) {
9767 u64 *cpuusage = percpu_ptr(ca->cpuusage, cpu);
9768 *cpuusage += cputime;
9775 * Charge the system/user time to the task's accounting group.
9777 static void cpuacct_update_stats(struct task_struct *tsk,
9778 enum cpuacct_stat_index idx, cputime_t val)
9782 if (unlikely(!cpuacct_subsys.active))
9789 percpu_counter_add(&ca->cpustat[idx], val);
9795 struct cgroup_subsys cpuacct_subsys = {
9797 .create = cpuacct_create,
9798 .destroy = cpuacct_destroy,
9799 .populate = cpuacct_populate,
9800 .subsys_id = cpuacct_subsys_id,
9802 #endif /* CONFIG_CGROUP_CPUACCT */