4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
108 * Timeslices get refilled after they expire.
110 #define DEF_TIMESLICE (100 * HZ / 1000)
113 * single value that denotes runtime == period, ie unlimited time.
115 #define RUNTIME_INF ((u64)~0ULL)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32 sg_div_cpu_power(const struct sched_group *sg, u32 load)
124 return reciprocal_divide(load, sg->reciprocal_cpu_power);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group *sg, u32 val)
133 sg->__cpu_power += val;
134 sg->reciprocal_cpu_power = reciprocal_value(sg->__cpu_power);
138 static inline int rt_policy(int policy)
140 if (unlikely(policy == SCHED_FIFO) || unlikely(policy == SCHED_RR))
145 static inline int task_has_rt_policy(struct task_struct *p)
147 return rt_policy(p->policy);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array {
154 DECLARE_BITMAP(bitmap, MAX_RT_PRIO+1); /* include 1 bit for delimiter */
155 struct list_head queue[MAX_RT_PRIO];
158 struct rt_bandwidth {
159 /* nests inside the rq lock: */
160 spinlock_t rt_runtime_lock;
163 struct hrtimer rt_period_timer;
166 static struct rt_bandwidth def_rt_bandwidth;
168 static int do_sched_rt_period_timer(struct rt_bandwidth *rt_b, int overrun);
170 static enum hrtimer_restart sched_rt_period_timer(struct hrtimer *timer)
172 struct rt_bandwidth *rt_b =
173 container_of(timer, struct rt_bandwidth, rt_period_timer);
179 now = hrtimer_cb_get_time(timer);
180 overrun = hrtimer_forward(timer, now, rt_b->rt_period);
185 idle = do_sched_rt_period_timer(rt_b, overrun);
188 return idle ? HRTIMER_NORESTART : HRTIMER_RESTART;
192 void init_rt_bandwidth(struct rt_bandwidth *rt_b, u64 period, u64 runtime)
194 rt_b->rt_period = ns_to_ktime(period);
195 rt_b->rt_runtime = runtime;
197 spin_lock_init(&rt_b->rt_runtime_lock);
199 hrtimer_init(&rt_b->rt_period_timer,
200 CLOCK_MONOTONIC, HRTIMER_MODE_REL);
201 rt_b->rt_period_timer.function = sched_rt_period_timer;
202 rt_b->rt_period_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
205 static void start_rt_bandwidth(struct rt_bandwidth *rt_b)
209 if (rt_b->rt_runtime == RUNTIME_INF)
212 if (hrtimer_active(&rt_b->rt_period_timer))
215 spin_lock(&rt_b->rt_runtime_lock);
217 if (hrtimer_active(&rt_b->rt_period_timer))
220 now = hrtimer_cb_get_time(&rt_b->rt_period_timer);
221 hrtimer_forward(&rt_b->rt_period_timer, now, rt_b->rt_period);
222 hrtimer_start(&rt_b->rt_period_timer,
223 rt_b->rt_period_timer.expires,
226 spin_unlock(&rt_b->rt_runtime_lock);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth *rt_b)
232 hrtimer_cancel(&rt_b->rt_period_timer);
237 * sched_domains_mutex serializes calls to arch_init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex);
242 #ifdef CONFIG_GROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups);
250 /* task group related information */
252 #ifdef CONFIG_CGROUP_SCHED
253 struct cgroup_subsys_state css;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity **se;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq **cfs_rq;
261 unsigned long shares;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity **rt_se;
266 struct rt_rq **rt_rq;
268 struct rt_bandwidth rt_bandwidth;
272 struct list_head list;
274 struct task_group *parent;
275 struct list_head siblings;
276 struct list_head children;
279 #ifdef CONFIG_USER_SCHED
283 * Every UID task group (including init_task_group aka UID-0) will
284 * be a child to this group.
286 struct task_group root_task_group;
288 #ifdef CONFIG_FAIR_GROUP_SCHED
289 /* Default task group's sched entity on each cpu */
290 static DEFINE_PER_CPU(struct sched_entity, init_sched_entity);
291 /* Default task group's cfs_rq on each cpu */
292 static DEFINE_PER_CPU(struct cfs_rq, init_cfs_rq) ____cacheline_aligned_in_smp;
295 #ifdef CONFIG_RT_GROUP_SCHED
296 static DEFINE_PER_CPU(struct sched_rt_entity, init_sched_rt_entity);
297 static DEFINE_PER_CPU(struct rt_rq, init_rt_rq) ____cacheline_aligned_in_smp;
300 #define root_task_group init_task_group
303 /* task_group_lock serializes add/remove of task groups and also changes to
304 * a task group's cpu shares.
306 static DEFINE_SPINLOCK(task_group_lock);
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 #ifdef CONFIG_USER_SCHED
310 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
316 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
317 * (The default weight is 1024 - so there's no practical
318 * limitation from this.)
321 #define MAX_SHARES (ULONG_MAX - 1)
323 static int init_task_group_load = INIT_TASK_GROUP_LOAD;
326 /* Default task group.
327 * Every task in system belong to this group at bootup.
329 struct task_group init_task_group;
331 /* return group to which a task belongs */
332 static inline struct task_group *task_group(struct task_struct *p)
334 struct task_group *tg;
336 #ifdef CONFIG_USER_SCHED
338 #elif defined(CONFIG_CGROUP_SCHED)
339 tg = container_of(task_subsys_state(p, cpu_cgroup_subsys_id),
340 struct task_group, css);
342 tg = &init_task_group;
347 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
348 static inline void set_task_rq(struct task_struct *p, unsigned int cpu)
350 #ifdef CONFIG_FAIR_GROUP_SCHED
351 p->se.cfs_rq = task_group(p)->cfs_rq[cpu];
352 p->se.parent = task_group(p)->se[cpu];
355 #ifdef CONFIG_RT_GROUP_SCHED
356 p->rt.rt_rq = task_group(p)->rt_rq[cpu];
357 p->rt.parent = task_group(p)->rt_se[cpu];
363 static inline void set_task_rq(struct task_struct *p, unsigned int cpu) { }
365 #endif /* CONFIG_GROUP_SCHED */
367 /* CFS-related fields in a runqueue */
369 struct load_weight load;
370 unsigned long nr_running;
375 struct rb_root tasks_timeline;
376 struct rb_node *rb_leftmost;
378 struct list_head tasks;
379 struct list_head *balance_iterator;
382 * 'curr' points to currently running entity on this cfs_rq.
383 * It is set to NULL otherwise (i.e when none are currently running).
385 struct sched_entity *curr, *next;
387 unsigned long nr_spread_over;
389 #ifdef CONFIG_FAIR_GROUP_SCHED
390 struct rq *rq; /* cpu runqueue to which this cfs_rq is attached */
393 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
394 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
395 * (like users, containers etc.)
397 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
398 * list is used during load balance.
400 struct list_head leaf_cfs_rq_list;
401 struct task_group *tg; /* group that "owns" this runqueue */
404 unsigned long task_weight;
405 unsigned long shares;
407 * We need space to build a sched_domain wide view of the full task
408 * group tree, in order to avoid depending on dynamic memory allocation
409 * during the load balancing we place this in the per cpu task group
410 * hierarchy. This limits the load balancing to one instance per cpu,
411 * but more should not be needed anyway.
413 struct aggregate_struct {
415 * load = weight(cpus) * f(tg)
417 * Where f(tg) is the recursive weight fraction assigned to
423 * part of the group weight distributed to this span.
425 unsigned long shares;
428 * The sum of all runqueue weights within this span.
430 unsigned long rq_weight;
433 * Weight contributed by tasks; this is the part we can
434 * influence by moving tasks around.
436 unsigned long task_weight;
442 /* Real-Time classes' related field in a runqueue: */
444 struct rt_prio_array active;
445 unsigned long rt_nr_running;
446 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
447 int highest_prio; /* highest queued rt task prio */
450 unsigned long rt_nr_migratory;
456 /* Nests inside the rq lock: */
457 spinlock_t rt_runtime_lock;
459 #ifdef CONFIG_RT_GROUP_SCHED
460 unsigned long rt_nr_boosted;
463 struct list_head leaf_rt_rq_list;
464 struct task_group *tg;
465 struct sched_rt_entity *rt_se;
472 * We add the notion of a root-domain which will be used to define per-domain
473 * variables. Each exclusive cpuset essentially defines an island domain by
474 * fully partitioning the member cpus from any other cpuset. Whenever a new
475 * exclusive cpuset is created, we also create and attach a new root-domain
485 * The "RT overload" flag: it gets set if a CPU has more than
486 * one runnable RT task.
493 * By default the system creates a single root-domain with all cpus as
494 * members (mimicking the global state we have today).
496 static struct root_domain def_root_domain;
501 * This is the main, per-CPU runqueue data structure.
503 * Locking rule: those places that want to lock multiple runqueues
504 * (such as the load balancing or the thread migration code), lock
505 * acquire operations must be ordered by ascending &runqueue.
512 * nr_running and cpu_load should be in the same cacheline because
513 * remote CPUs use both these fields when doing load calculation.
515 unsigned long nr_running;
516 #define CPU_LOAD_IDX_MAX 5
517 unsigned long cpu_load[CPU_LOAD_IDX_MAX];
518 unsigned char idle_at_tick;
520 unsigned long last_tick_seen;
521 unsigned char in_nohz_recently;
523 /* capture load from *all* tasks on this cpu: */
524 struct load_weight load;
525 unsigned long nr_load_updates;
531 #ifdef CONFIG_FAIR_GROUP_SCHED
532 /* list of leaf cfs_rq on this cpu: */
533 struct list_head leaf_cfs_rq_list;
535 #ifdef CONFIG_RT_GROUP_SCHED
536 struct list_head leaf_rt_rq_list;
540 * This is part of a global counter where only the total sum
541 * over all CPUs matters. A task can increase this counter on
542 * one CPU and if it got migrated afterwards it may decrease
543 * it on another CPU. Always updated under the runqueue lock:
545 unsigned long nr_uninterruptible;
547 struct task_struct *curr, *idle;
548 unsigned long next_balance;
549 struct mm_struct *prev_mm;
556 struct root_domain *rd;
557 struct sched_domain *sd;
559 /* For active balancing */
562 /* cpu of this runqueue: */
565 struct task_struct *migration_thread;
566 struct list_head migration_queue;
569 #ifdef CONFIG_SCHED_HRTICK
570 unsigned long hrtick_flags;
571 ktime_t hrtick_expire;
572 struct hrtimer hrtick_timer;
575 #ifdef CONFIG_SCHEDSTATS
577 struct sched_info rq_sched_info;
579 /* sys_sched_yield() stats */
580 unsigned int yld_exp_empty;
581 unsigned int yld_act_empty;
582 unsigned int yld_both_empty;
583 unsigned int yld_count;
585 /* schedule() stats */
586 unsigned int sched_switch;
587 unsigned int sched_count;
588 unsigned int sched_goidle;
590 /* try_to_wake_up() stats */
591 unsigned int ttwu_count;
592 unsigned int ttwu_local;
595 unsigned int bkl_count;
597 struct lock_class_key rq_lock_key;
600 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq, runqueues);
602 static inline void check_preempt_curr(struct rq *rq, struct task_struct *p)
604 rq->curr->sched_class->check_preempt_curr(rq, p);
607 static inline int cpu_of(struct rq *rq)
617 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
618 * See detach_destroy_domains: synchronize_sched for details.
620 * The domain tree of any CPU may only be accessed from within
621 * preempt-disabled sections.
623 #define for_each_domain(cpu, __sd) \
624 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
626 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
627 #define this_rq() (&__get_cpu_var(runqueues))
628 #define task_rq(p) cpu_rq(task_cpu(p))
629 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
631 static inline void update_rq_clock(struct rq *rq)
633 rq->clock = sched_clock_cpu(cpu_of(rq));
637 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
639 #ifdef CONFIG_SCHED_DEBUG
640 # define const_debug __read_mostly
642 # define const_debug static const
646 * Debugging: various feature bits
649 #define SCHED_FEAT(name, enabled) \
650 __SCHED_FEAT_##name ,
653 #include "sched_features.h"
658 #define SCHED_FEAT(name, enabled) \
659 (1UL << __SCHED_FEAT_##name) * enabled |
661 const_debug unsigned int sysctl_sched_features =
662 #include "sched_features.h"
667 #ifdef CONFIG_SCHED_DEBUG
668 #define SCHED_FEAT(name, enabled) \
671 static __read_mostly char *sched_feat_names[] = {
672 #include "sched_features.h"
678 static int sched_feat_open(struct inode *inode, struct file *filp)
680 filp->private_data = inode->i_private;
685 sched_feat_read(struct file *filp, char __user *ubuf,
686 size_t cnt, loff_t *ppos)
693 for (i = 0; sched_feat_names[i]; i++) {
694 len += strlen(sched_feat_names[i]);
698 buf = kmalloc(len + 2, GFP_KERNEL);
702 for (i = 0; sched_feat_names[i]; i++) {
703 if (sysctl_sched_features & (1UL << i))
704 r += sprintf(buf + r, "%s ", sched_feat_names[i]);
706 r += sprintf(buf + r, "NO_%s ", sched_feat_names[i]);
709 r += sprintf(buf + r, "\n");
710 WARN_ON(r >= len + 2);
712 r = simple_read_from_buffer(ubuf, cnt, ppos, buf, r);
720 sched_feat_write(struct file *filp, const char __user *ubuf,
721 size_t cnt, loff_t *ppos)
731 if (copy_from_user(&buf, ubuf, cnt))
736 if (strncmp(buf, "NO_", 3) == 0) {
741 for (i = 0; sched_feat_names[i]; i++) {
742 int len = strlen(sched_feat_names[i]);
744 if (strncmp(cmp, sched_feat_names[i], len) == 0) {
746 sysctl_sched_features &= ~(1UL << i);
748 sysctl_sched_features |= (1UL << i);
753 if (!sched_feat_names[i])
761 static struct file_operations sched_feat_fops = {
762 .open = sched_feat_open,
763 .read = sched_feat_read,
764 .write = sched_feat_write,
767 static __init int sched_init_debug(void)
769 debugfs_create_file("sched_features", 0644, NULL, NULL,
774 late_initcall(sched_init_debug);
778 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
781 * Number of tasks to iterate in a single balance run.
782 * Limited because this is done with IRQs disabled.
784 const_debug unsigned int sysctl_sched_nr_migrate = 32;
787 * period over which we measure -rt task cpu usage in us.
790 unsigned int sysctl_sched_rt_period = 1000000;
792 static __read_mostly int scheduler_running;
795 * part of the period that we allow rt tasks to run in us.
798 int sysctl_sched_rt_runtime = 950000;
800 static inline u64 global_rt_period(void)
802 return (u64)sysctl_sched_rt_period * NSEC_PER_USEC;
805 static inline u64 global_rt_runtime(void)
807 if (sysctl_sched_rt_period < 0)
810 return (u64)sysctl_sched_rt_runtime * NSEC_PER_USEC;
813 unsigned long long time_sync_thresh = 100000;
815 static DEFINE_PER_CPU(unsigned long long, time_offset);
816 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time);
819 * Global lock which we take every now and then to synchronize
820 * the CPUs time. This method is not warp-safe, but it's good
821 * enough to synchronize slowly diverging time sources and thus
822 * it's good enough for tracing:
824 static DEFINE_SPINLOCK(time_sync_lock);
825 static unsigned long long prev_global_time;
827 static unsigned long long __sync_cpu_clock(unsigned long long time, int cpu)
830 * We want this inlined, to not get tracer function calls
831 * in this critical section:
833 spin_acquire(&time_sync_lock.dep_map, 0, 0, _THIS_IP_);
834 __raw_spin_lock(&time_sync_lock.raw_lock);
836 if (time < prev_global_time) {
837 per_cpu(time_offset, cpu) += prev_global_time - time;
838 time = prev_global_time;
840 prev_global_time = time;
843 __raw_spin_unlock(&time_sync_lock.raw_lock);
844 spin_release(&time_sync_lock.dep_map, 1, _THIS_IP_);
849 static unsigned long long __cpu_clock(int cpu)
851 unsigned long long now;
854 * Only call sched_clock() if the scheduler has already been
855 * initialized (some code might call cpu_clock() very early):
857 if (unlikely(!scheduler_running))
860 now = sched_clock_cpu(cpu);
866 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
867 * clock constructed from sched_clock():
869 unsigned long long cpu_clock(int cpu)
871 unsigned long long prev_cpu_time, time, delta_time;
874 local_irq_save(flags);
875 prev_cpu_time = per_cpu(prev_cpu_time, cpu);
876 time = __cpu_clock(cpu) + per_cpu(time_offset, cpu);
877 delta_time = time-prev_cpu_time;
879 if (unlikely(delta_time > time_sync_thresh)) {
880 time = __sync_cpu_clock(time, cpu);
881 per_cpu(prev_cpu_time, cpu) = time;
883 local_irq_restore(flags);
887 EXPORT_SYMBOL_GPL(cpu_clock);
889 #ifndef prepare_arch_switch
890 # define prepare_arch_switch(next) do { } while (0)
892 #ifndef finish_arch_switch
893 # define finish_arch_switch(prev) do { } while (0)
896 static inline int task_current(struct rq *rq, struct task_struct *p)
898 return rq->curr == p;
901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
902 static inline int task_running(struct rq *rq, struct task_struct *p)
904 return task_current(rq, p);
907 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
911 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
913 #ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq->lock.owner = current;
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
922 spin_acquire(&rq->lock.dep_map, 0, 0, _THIS_IP_);
924 spin_unlock_irq(&rq->lock);
927 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
928 static inline int task_running(struct rq *rq, struct task_struct *p)
933 return task_current(rq, p);
937 static inline void prepare_lock_switch(struct rq *rq, struct task_struct *next)
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
947 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq->lock);
950 spin_unlock(&rq->lock);
954 static inline void finish_lock_switch(struct rq *rq, struct task_struct *prev)
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
965 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
969 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
975 static inline struct rq *__task_rq_lock(struct task_struct *p)
979 struct rq *rq = task_rq(p);
980 spin_lock(&rq->lock);
981 if (likely(rq == task_rq(p)))
983 spin_unlock(&rq->lock);
988 * task_rq_lock - lock the runqueue a given task resides on and disable
989 * interrupts. Note the ordering: we can safely lookup the task_rq without
990 * explicitly disabling preemption.
992 static struct rq *task_rq_lock(struct task_struct *p, unsigned long *flags)
998 local_irq_save(*flags);
1000 spin_lock(&rq->lock);
1001 if (likely(rq == task_rq(p)))
1003 spin_unlock_irqrestore(&rq->lock, *flags);
1007 static void __task_rq_unlock(struct rq *rq)
1008 __releases(rq->lock)
1010 spin_unlock(&rq->lock);
1013 static inline void task_rq_unlock(struct rq *rq, unsigned long *flags)
1014 __releases(rq->lock)
1016 spin_unlock_irqrestore(&rq->lock, *flags);
1020 * this_rq_lock - lock this runqueue and disable interrupts.
1022 static struct rq *this_rq_lock(void)
1023 __acquires(rq->lock)
1027 local_irq_disable();
1029 spin_lock(&rq->lock);
1034 static void __resched_task(struct task_struct *p, int tif_bit);
1036 static inline void resched_task(struct task_struct *p)
1038 __resched_task(p, TIF_NEED_RESCHED);
1041 #ifdef CONFIG_SCHED_HRTICK
1043 * Use HR-timers to deliver accurate preemption points.
1045 * Its all a bit involved since we cannot program an hrt while holding the
1046 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1049 * When we get rescheduled we reprogram the hrtick_timer outside of the
1052 static inline void resched_hrt(struct task_struct *p)
1054 __resched_task(p, TIF_HRTICK_RESCHED);
1057 static inline void resched_rq(struct rq *rq)
1059 unsigned long flags;
1061 spin_lock_irqsave(&rq->lock, flags);
1062 resched_task(rq->curr);
1063 spin_unlock_irqrestore(&rq->lock, flags);
1067 HRTICK_SET, /* re-programm hrtick_timer */
1068 HRTICK_RESET, /* not a new slice */
1069 HRTICK_BLOCK, /* stop hrtick operations */
1074 * - enabled by features
1075 * - hrtimer is actually high res
1077 static inline int hrtick_enabled(struct rq *rq)
1079 if (!sched_feat(HRTICK))
1081 if (unlikely(test_bit(HRTICK_BLOCK, &rq->hrtick_flags)))
1083 return hrtimer_is_hres_active(&rq->hrtick_timer);
1087 * Called to set the hrtick timer state.
1089 * called with rq->lock held and irqs disabled
1091 static void hrtick_start(struct rq *rq, u64 delay, int reset)
1093 assert_spin_locked(&rq->lock);
1096 * preempt at: now + delay
1099 ktime_add_ns(rq->hrtick_timer.base->get_time(), delay);
1101 * indicate we need to program the timer
1103 __set_bit(HRTICK_SET, &rq->hrtick_flags);
1105 __set_bit(HRTICK_RESET, &rq->hrtick_flags);
1108 * New slices are called from the schedule path and don't need a
1109 * forced reschedule.
1112 resched_hrt(rq->curr);
1115 static void hrtick_clear(struct rq *rq)
1117 if (hrtimer_active(&rq->hrtick_timer))
1118 hrtimer_cancel(&rq->hrtick_timer);
1122 * Update the timer from the possible pending state.
1124 static void hrtick_set(struct rq *rq)
1128 unsigned long flags;
1130 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1132 spin_lock_irqsave(&rq->lock, flags);
1133 set = __test_and_clear_bit(HRTICK_SET, &rq->hrtick_flags);
1134 reset = __test_and_clear_bit(HRTICK_RESET, &rq->hrtick_flags);
1135 time = rq->hrtick_expire;
1136 clear_thread_flag(TIF_HRTICK_RESCHED);
1137 spin_unlock_irqrestore(&rq->lock, flags);
1140 hrtimer_start(&rq->hrtick_timer, time, HRTIMER_MODE_ABS);
1141 if (reset && !hrtimer_active(&rq->hrtick_timer))
1148 * High-resolution timer tick.
1149 * Runs from hardirq context with interrupts disabled.
1151 static enum hrtimer_restart hrtick(struct hrtimer *timer)
1153 struct rq *rq = container_of(timer, struct rq, hrtick_timer);
1155 WARN_ON_ONCE(cpu_of(rq) != smp_processor_id());
1157 spin_lock(&rq->lock);
1158 update_rq_clock(rq);
1159 rq->curr->sched_class->task_tick(rq, rq->curr, 1);
1160 spin_unlock(&rq->lock);
1162 return HRTIMER_NORESTART;
1165 static void hotplug_hrtick_disable(int cpu)
1167 struct rq *rq = cpu_rq(cpu);
1168 unsigned long flags;
1170 spin_lock_irqsave(&rq->lock, flags);
1171 rq->hrtick_flags = 0;
1172 __set_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1173 spin_unlock_irqrestore(&rq->lock, flags);
1178 static void hotplug_hrtick_enable(int cpu)
1180 struct rq *rq = cpu_rq(cpu);
1181 unsigned long flags;
1183 spin_lock_irqsave(&rq->lock, flags);
1184 __clear_bit(HRTICK_BLOCK, &rq->hrtick_flags);
1185 spin_unlock_irqrestore(&rq->lock, flags);
1189 hotplug_hrtick(struct notifier_block *nfb, unsigned long action, void *hcpu)
1191 int cpu = (int)(long)hcpu;
1194 case CPU_UP_CANCELED:
1195 case CPU_UP_CANCELED_FROZEN:
1196 case CPU_DOWN_PREPARE:
1197 case CPU_DOWN_PREPARE_FROZEN:
1199 case CPU_DEAD_FROZEN:
1200 hotplug_hrtick_disable(cpu);
1203 case CPU_UP_PREPARE:
1204 case CPU_UP_PREPARE_FROZEN:
1205 case CPU_DOWN_FAILED:
1206 case CPU_DOWN_FAILED_FROZEN:
1208 case CPU_ONLINE_FROZEN:
1209 hotplug_hrtick_enable(cpu);
1216 static void init_hrtick(void)
1218 hotcpu_notifier(hotplug_hrtick, 0);
1221 static void init_rq_hrtick(struct rq *rq)
1223 rq->hrtick_flags = 0;
1224 hrtimer_init(&rq->hrtick_timer, CLOCK_MONOTONIC, HRTIMER_MODE_REL);
1225 rq->hrtick_timer.function = hrtick;
1226 rq->hrtick_timer.cb_mode = HRTIMER_CB_IRQSAFE_NO_SOFTIRQ;
1229 void hrtick_resched(void)
1232 unsigned long flags;
1234 if (!test_thread_flag(TIF_HRTICK_RESCHED))
1237 local_irq_save(flags);
1238 rq = cpu_rq(smp_processor_id());
1240 local_irq_restore(flags);
1243 static inline void hrtick_clear(struct rq *rq)
1247 static inline void hrtick_set(struct rq *rq)
1251 static inline void init_rq_hrtick(struct rq *rq)
1255 void hrtick_resched(void)
1259 static inline void init_hrtick(void)
1265 * resched_task - mark a task 'to be rescheduled now'.
1267 * On UP this means the setting of the need_resched flag, on SMP it
1268 * might also involve a cross-CPU call to trigger the scheduler on
1273 #ifndef tsk_is_polling
1274 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1277 static void __resched_task(struct task_struct *p, int tif_bit)
1281 assert_spin_locked(&task_rq(p)->lock);
1283 if (unlikely(test_tsk_thread_flag(p, tif_bit)))
1286 set_tsk_thread_flag(p, tif_bit);
1289 if (cpu == smp_processor_id())
1292 /* NEED_RESCHED must be visible before we test polling */
1294 if (!tsk_is_polling(p))
1295 smp_send_reschedule(cpu);
1298 static void resched_cpu(int cpu)
1300 struct rq *rq = cpu_rq(cpu);
1301 unsigned long flags;
1303 if (!spin_trylock_irqsave(&rq->lock, flags))
1305 resched_task(cpu_curr(cpu));
1306 spin_unlock_irqrestore(&rq->lock, flags);
1311 * When add_timer_on() enqueues a timer into the timer wheel of an
1312 * idle CPU then this timer might expire before the next timer event
1313 * which is scheduled to wake up that CPU. In case of a completely
1314 * idle system the next event might even be infinite time into the
1315 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1316 * leaves the inner idle loop so the newly added timer is taken into
1317 * account when the CPU goes back to idle and evaluates the timer
1318 * wheel for the next timer event.
1320 void wake_up_idle_cpu(int cpu)
1322 struct rq *rq = cpu_rq(cpu);
1324 if (cpu == smp_processor_id())
1328 * This is safe, as this function is called with the timer
1329 * wheel base lock of (cpu) held. When the CPU is on the way
1330 * to idle and has not yet set rq->curr to idle then it will
1331 * be serialized on the timer wheel base lock and take the new
1332 * timer into account automatically.
1334 if (rq->curr != rq->idle)
1338 * We can set TIF_RESCHED on the idle task of the other CPU
1339 * lockless. The worst case is that the other CPU runs the
1340 * idle task through an additional NOOP schedule()
1342 set_tsk_thread_flag(rq->idle, TIF_NEED_RESCHED);
1344 /* NEED_RESCHED must be visible before we test polling */
1346 if (!tsk_is_polling(rq->idle))
1347 smp_send_reschedule(cpu);
1352 static void __resched_task(struct task_struct *p, int tif_bit)
1354 assert_spin_locked(&task_rq(p)->lock);
1355 set_tsk_thread_flag(p, tif_bit);
1359 #if BITS_PER_LONG == 32
1360 # define WMULT_CONST (~0UL)
1362 # define WMULT_CONST (1UL << 32)
1365 #define WMULT_SHIFT 32
1368 * Shift right and round:
1370 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1373 * delta *= weight / lw
1375 static unsigned long
1376 calc_delta_mine(unsigned long delta_exec, unsigned long weight,
1377 struct load_weight *lw)
1381 if (!lw->inv_weight)
1382 lw->inv_weight = 1 + (WMULT_CONST-lw->weight/2)/(lw->weight+1);
1384 tmp = (u64)delta_exec * weight;
1386 * Check whether we'd overflow the 64-bit multiplication:
1388 if (unlikely(tmp > WMULT_CONST))
1389 tmp = SRR(SRR(tmp, WMULT_SHIFT/2) * lw->inv_weight,
1392 tmp = SRR(tmp * lw->inv_weight, WMULT_SHIFT);
1394 return (unsigned long)min(tmp, (u64)(unsigned long)LONG_MAX);
1397 static inline void update_load_add(struct load_weight *lw, unsigned long inc)
1403 static inline void update_load_sub(struct load_weight *lw, unsigned long dec)
1410 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1411 * of tasks with abnormal "nice" values across CPUs the contribution that
1412 * each task makes to its run queue's load is weighted according to its
1413 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1414 * scaled version of the new time slice allocation that they receive on time
1418 #define WEIGHT_IDLEPRIO 2
1419 #define WMULT_IDLEPRIO (1 << 31)
1422 * Nice levels are multiplicative, with a gentle 10% change for every
1423 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1424 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1425 * that remained on nice 0.
1427 * The "10% effect" is relative and cumulative: from _any_ nice level,
1428 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1429 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1430 * If a task goes up by ~10% and another task goes down by ~10% then
1431 * the relative distance between them is ~25%.)
1433 static const int prio_to_weight[40] = {
1434 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1435 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1436 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1437 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1438 /* 0 */ 1024, 820, 655, 526, 423,
1439 /* 5 */ 335, 272, 215, 172, 137,
1440 /* 10 */ 110, 87, 70, 56, 45,
1441 /* 15 */ 36, 29, 23, 18, 15,
1445 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1447 * In cases where the weight does not change often, we can use the
1448 * precalculated inverse to speed up arithmetics by turning divisions
1449 * into multiplications:
1451 static const u32 prio_to_wmult[40] = {
1452 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1453 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1454 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1455 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1456 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1457 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1458 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1459 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1462 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup);
1465 * runqueue iterator, to support SMP load-balancing between different
1466 * scheduling classes, without having to expose their internal data
1467 * structures to the load-balancing proper:
1469 struct rq_iterator {
1471 struct task_struct *(*start)(void *);
1472 struct task_struct *(*next)(void *);
1476 static unsigned long
1477 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
1478 unsigned long max_load_move, struct sched_domain *sd,
1479 enum cpu_idle_type idle, int *all_pinned,
1480 int *this_best_prio, struct rq_iterator *iterator);
1483 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
1484 struct sched_domain *sd, enum cpu_idle_type idle,
1485 struct rq_iterator *iterator);
1488 #ifdef CONFIG_CGROUP_CPUACCT
1489 static void cpuacct_charge(struct task_struct *tsk, u64 cputime);
1491 static inline void cpuacct_charge(struct task_struct *tsk, u64 cputime) {}
1494 static inline void inc_cpu_load(struct rq *rq, unsigned long load)
1496 update_load_add(&rq->load, load);
1499 static inline void dec_cpu_load(struct rq *rq, unsigned long load)
1501 update_load_sub(&rq->load, load);
1505 static unsigned long source_load(int cpu, int type);
1506 static unsigned long target_load(int cpu, int type);
1507 static unsigned long cpu_avg_load_per_task(int cpu);
1508 static int task_hot(struct task_struct *p, u64 now, struct sched_domain *sd);
1510 #ifdef CONFIG_FAIR_GROUP_SCHED
1513 * Group load balancing.
1515 * We calculate a few balance domain wide aggregate numbers; load and weight.
1516 * Given the pictures below, and assuming each item has equal weight:
1527 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1528 * which equals 1/9-th of the total load.
1531 * The weight of this group on the selected cpus.
1534 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1538 * Part of the rq_weight contributed by tasks; all groups except B would
1542 static inline struct aggregate_struct *
1543 aggregate(struct task_group *tg, struct sched_domain *sd)
1545 return &tg->cfs_rq[sd->first_cpu]->aggregate;
1548 typedef void (*aggregate_func)(struct task_group *, struct sched_domain *);
1551 * Iterate the full tree, calling @down when first entering a node and @up when
1552 * leaving it for the final time.
1555 void aggregate_walk_tree(aggregate_func down, aggregate_func up,
1556 struct sched_domain *sd)
1558 struct task_group *parent, *child;
1561 parent = &root_task_group;
1563 (*down)(parent, sd);
1564 list_for_each_entry_rcu(child, &parent->children, siblings) {
1574 parent = parent->parent;
1581 * Calculate the aggregate runqueue weight.
1584 void aggregate_group_weight(struct task_group *tg, struct sched_domain *sd)
1586 unsigned long rq_weight = 0;
1587 unsigned long task_weight = 0;
1590 for_each_cpu_mask(i, sd->span) {
1591 rq_weight += tg->cfs_rq[i]->load.weight;
1592 task_weight += tg->cfs_rq[i]->task_weight;
1595 aggregate(tg, sd)->rq_weight = rq_weight;
1596 aggregate(tg, sd)->task_weight = task_weight;
1600 * Compute the weight of this group on the given cpus.
1603 void aggregate_group_shares(struct task_group *tg, struct sched_domain *sd)
1605 unsigned long shares = 0;
1608 for_each_cpu_mask(i, sd->span)
1609 shares += tg->cfs_rq[i]->shares;
1611 if ((!shares && aggregate(tg, sd)->rq_weight) || shares > tg->shares)
1612 shares = tg->shares;
1614 aggregate(tg, sd)->shares = shares;
1618 * Compute the load fraction assigned to this group, relies on the aggregate
1619 * weight and this group's parent's load, i.e. top-down.
1622 void aggregate_group_load(struct task_group *tg, struct sched_domain *sd)
1630 for_each_cpu_mask(i, sd->span)
1631 load += cpu_rq(i)->load.weight;
1634 load = aggregate(tg->parent, sd)->load;
1637 * shares is our weight in the parent's rq so
1638 * shares/parent->rq_weight gives our fraction of the load
1640 load *= aggregate(tg, sd)->shares;
1641 load /= aggregate(tg->parent, sd)->rq_weight + 1;
1644 aggregate(tg, sd)->load = load;
1647 static void __set_se_shares(struct sched_entity *se, unsigned long shares);
1650 * Calculate and set the cpu's group shares.
1653 __update_group_shares_cpu(struct task_group *tg, struct sched_domain *sd,
1657 unsigned long shares;
1658 unsigned long rq_weight;
1663 rq_weight = tg->cfs_rq[tcpu]->load.weight;
1666 * If there are currently no tasks on the cpu pretend there is one of
1667 * average load so that when a new task gets to run here it will not
1668 * get delayed by group starvation.
1672 rq_weight = NICE_0_LOAD;
1676 * \Sum shares * rq_weight
1677 * shares = -----------------------
1681 shares = aggregate(tg, sd)->shares * rq_weight;
1682 shares /= aggregate(tg, sd)->rq_weight + 1;
1685 * record the actual number of shares, not the boosted amount.
1687 tg->cfs_rq[tcpu]->shares = boost ? 0 : shares;
1689 if (shares < MIN_SHARES)
1690 shares = MIN_SHARES;
1691 else if (shares > MAX_SHARES)
1692 shares = MAX_SHARES;
1694 __set_se_shares(tg->se[tcpu], shares);
1698 * Re-adjust the weights on the cpu the task came from and on the cpu the
1702 __move_group_shares(struct task_group *tg, struct sched_domain *sd,
1705 unsigned long shares;
1707 shares = tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1709 __update_group_shares_cpu(tg, sd, scpu);
1710 __update_group_shares_cpu(tg, sd, dcpu);
1713 * ensure we never loose shares due to rounding errors in the
1714 * above redistribution.
1716 shares -= tg->cfs_rq[scpu]->shares + tg->cfs_rq[dcpu]->shares;
1718 tg->cfs_rq[dcpu]->shares += shares;
1722 * Because changing a group's shares changes the weight of the super-group
1723 * we need to walk up the tree and change all shares until we hit the root.
1726 move_group_shares(struct task_group *tg, struct sched_domain *sd,
1730 __move_group_shares(tg, sd, scpu, dcpu);
1736 void aggregate_group_set_shares(struct task_group *tg, struct sched_domain *sd)
1738 unsigned long shares = aggregate(tg, sd)->shares;
1741 for_each_cpu_mask(i, sd->span) {
1742 struct rq *rq = cpu_rq(i);
1743 unsigned long flags;
1745 spin_lock_irqsave(&rq->lock, flags);
1746 __update_group_shares_cpu(tg, sd, i);
1747 spin_unlock_irqrestore(&rq->lock, flags);
1750 aggregate_group_shares(tg, sd);
1753 * ensure we never loose shares due to rounding errors in the
1754 * above redistribution.
1756 shares -= aggregate(tg, sd)->shares;
1758 tg->cfs_rq[sd->first_cpu]->shares += shares;
1759 aggregate(tg, sd)->shares += shares;
1764 * Calculate the accumulative weight and recursive load of each task group
1765 * while walking down the tree.
1768 void aggregate_get_down(struct task_group *tg, struct sched_domain *sd)
1770 aggregate_group_weight(tg, sd);
1771 aggregate_group_shares(tg, sd);
1772 aggregate_group_load(tg, sd);
1776 * Rebalance the cpu shares while walking back up the tree.
1779 void aggregate_get_up(struct task_group *tg, struct sched_domain *sd)
1781 aggregate_group_set_shares(tg, sd);
1784 static DEFINE_PER_CPU(spinlock_t, aggregate_lock);
1786 static void __init init_aggregate(void)
1790 for_each_possible_cpu(i)
1791 spin_lock_init(&per_cpu(aggregate_lock, i));
1794 static int get_aggregate(struct sched_domain *sd)
1796 if (!spin_trylock(&per_cpu(aggregate_lock, sd->first_cpu)))
1799 aggregate_walk_tree(aggregate_get_down, aggregate_get_up, sd);
1803 static void put_aggregate(struct sched_domain *sd)
1805 spin_unlock(&per_cpu(aggregate_lock, sd->first_cpu));
1808 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1810 cfs_rq->shares = shares;
1815 static inline void init_aggregate(void)
1819 static inline int get_aggregate(struct sched_domain *sd)
1824 static inline void put_aggregate(struct sched_domain *sd)
1829 #else /* CONFIG_SMP */
1831 #ifdef CONFIG_FAIR_GROUP_SCHED
1832 static void cfs_rq_set_shares(struct cfs_rq *cfs_rq, unsigned long shares)
1837 #endif /* CONFIG_SMP */
1839 #include "sched_stats.h"
1840 #include "sched_idletask.c"
1841 #include "sched_fair.c"
1842 #include "sched_rt.c"
1843 #ifdef CONFIG_SCHED_DEBUG
1844 # include "sched_debug.c"
1847 #define sched_class_highest (&rt_sched_class)
1849 static void inc_nr_running(struct rq *rq)
1854 static void dec_nr_running(struct rq *rq)
1859 static void set_load_weight(struct task_struct *p)
1861 if (task_has_rt_policy(p)) {
1862 p->se.load.weight = prio_to_weight[0] * 2;
1863 p->se.load.inv_weight = prio_to_wmult[0] >> 1;
1868 * SCHED_IDLE tasks get minimal weight:
1870 if (p->policy == SCHED_IDLE) {
1871 p->se.load.weight = WEIGHT_IDLEPRIO;
1872 p->se.load.inv_weight = WMULT_IDLEPRIO;
1876 p->se.load.weight = prio_to_weight[p->static_prio - MAX_RT_PRIO];
1877 p->se.load.inv_weight = prio_to_wmult[p->static_prio - MAX_RT_PRIO];
1880 static void enqueue_task(struct rq *rq, struct task_struct *p, int wakeup)
1882 sched_info_queued(p);
1883 p->sched_class->enqueue_task(rq, p, wakeup);
1887 static void dequeue_task(struct rq *rq, struct task_struct *p, int sleep)
1889 p->sched_class->dequeue_task(rq, p, sleep);
1894 * __normal_prio - return the priority that is based on the static prio
1896 static inline int __normal_prio(struct task_struct *p)
1898 return p->static_prio;
1902 * Calculate the expected normal priority: i.e. priority
1903 * without taking RT-inheritance into account. Might be
1904 * boosted by interactivity modifiers. Changes upon fork,
1905 * setprio syscalls, and whenever the interactivity
1906 * estimator recalculates.
1908 static inline int normal_prio(struct task_struct *p)
1912 if (task_has_rt_policy(p))
1913 prio = MAX_RT_PRIO-1 - p->rt_priority;
1915 prio = __normal_prio(p);
1920 * Calculate the current priority, i.e. the priority
1921 * taken into account by the scheduler. This value might
1922 * be boosted by RT tasks, or might be boosted by
1923 * interactivity modifiers. Will be RT if the task got
1924 * RT-boosted. If not then it returns p->normal_prio.
1926 static int effective_prio(struct task_struct *p)
1928 p->normal_prio = normal_prio(p);
1930 * If we are RT tasks or we were boosted to RT priority,
1931 * keep the priority unchanged. Otherwise, update priority
1932 * to the normal priority:
1934 if (!rt_prio(p->prio))
1935 return p->normal_prio;
1940 * activate_task - move a task to the runqueue.
1942 static void activate_task(struct rq *rq, struct task_struct *p, int wakeup)
1944 if (task_contributes_to_load(p))
1945 rq->nr_uninterruptible--;
1947 enqueue_task(rq, p, wakeup);
1952 * deactivate_task - remove a task from the runqueue.
1954 static void deactivate_task(struct rq *rq, struct task_struct *p, int sleep)
1956 if (task_contributes_to_load(p))
1957 rq->nr_uninterruptible++;
1959 dequeue_task(rq, p, sleep);
1964 * task_curr - is this task currently executing on a CPU?
1965 * @p: the task in question.
1967 inline int task_curr(const struct task_struct *p)
1969 return cpu_curr(task_cpu(p)) == p;
1972 /* Used instead of source_load when we know the type == 0 */
1973 unsigned long weighted_cpuload(const int cpu)
1975 return cpu_rq(cpu)->load.weight;
1978 static inline void __set_task_cpu(struct task_struct *p, unsigned int cpu)
1980 set_task_rq(p, cpu);
1983 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1984 * successfuly executed on another CPU. We must ensure that updates of
1985 * per-task data have been completed by this moment.
1988 task_thread_info(p)->cpu = cpu;
1992 static inline void check_class_changed(struct rq *rq, struct task_struct *p,
1993 const struct sched_class *prev_class,
1994 int oldprio, int running)
1996 if (prev_class != p->sched_class) {
1997 if (prev_class->switched_from)
1998 prev_class->switched_from(rq, p, running);
1999 p->sched_class->switched_to(rq, p, running);
2001 p->sched_class->prio_changed(rq, p, oldprio, running);
2007 * Is this task likely cache-hot:
2010 task_hot(struct task_struct *p, u64 now, struct sched_domain *sd)
2015 * Buddy candidates are cache hot:
2017 if (sched_feat(CACHE_HOT_BUDDY) && (&p->se == cfs_rq_of(&p->se)->next))
2020 if (p->sched_class != &fair_sched_class)
2023 if (sysctl_sched_migration_cost == -1)
2025 if (sysctl_sched_migration_cost == 0)
2028 delta = now - p->se.exec_start;
2030 return delta < (s64)sysctl_sched_migration_cost;
2034 void set_task_cpu(struct task_struct *p, unsigned int new_cpu)
2036 int old_cpu = task_cpu(p);
2037 struct rq *old_rq = cpu_rq(old_cpu), *new_rq = cpu_rq(new_cpu);
2038 struct cfs_rq *old_cfsrq = task_cfs_rq(p),
2039 *new_cfsrq = cpu_cfs_rq(old_cfsrq, new_cpu);
2042 clock_offset = old_rq->clock - new_rq->clock;
2044 #ifdef CONFIG_SCHEDSTATS
2045 if (p->se.wait_start)
2046 p->se.wait_start -= clock_offset;
2047 if (p->se.sleep_start)
2048 p->se.sleep_start -= clock_offset;
2049 if (p->se.block_start)
2050 p->se.block_start -= clock_offset;
2051 if (old_cpu != new_cpu) {
2052 schedstat_inc(p, se.nr_migrations);
2053 if (task_hot(p, old_rq->clock, NULL))
2054 schedstat_inc(p, se.nr_forced2_migrations);
2057 p->se.vruntime -= old_cfsrq->min_vruntime -
2058 new_cfsrq->min_vruntime;
2060 __set_task_cpu(p, new_cpu);
2063 struct migration_req {
2064 struct list_head list;
2066 struct task_struct *task;
2069 struct completion done;
2073 * The task's runqueue lock must be held.
2074 * Returns true if you have to wait for migration thread.
2077 migrate_task(struct task_struct *p, int dest_cpu, struct migration_req *req)
2079 struct rq *rq = task_rq(p);
2082 * If the task is not on a runqueue (and not running), then
2083 * it is sufficient to simply update the task's cpu field.
2085 if (!p->se.on_rq && !task_running(rq, p)) {
2086 set_task_cpu(p, dest_cpu);
2090 init_completion(&req->done);
2092 req->dest_cpu = dest_cpu;
2093 list_add(&req->list, &rq->migration_queue);
2099 * wait_task_inactive - wait for a thread to unschedule.
2101 * The caller must ensure that the task *will* unschedule sometime soon,
2102 * else this function might spin for a *long* time. This function can't
2103 * be called with interrupts off, or it may introduce deadlock with
2104 * smp_call_function() if an IPI is sent by the same process we are
2105 * waiting to become inactive.
2107 void wait_task_inactive(struct task_struct *p)
2109 unsigned long flags;
2115 * We do the initial early heuristics without holding
2116 * any task-queue locks at all. We'll only try to get
2117 * the runqueue lock when things look like they will
2123 * If the task is actively running on another CPU
2124 * still, just relax and busy-wait without holding
2127 * NOTE! Since we don't hold any locks, it's not
2128 * even sure that "rq" stays as the right runqueue!
2129 * But we don't care, since "task_running()" will
2130 * return false if the runqueue has changed and p
2131 * is actually now running somewhere else!
2133 while (task_running(rq, p))
2137 * Ok, time to look more closely! We need the rq
2138 * lock now, to be *sure*. If we're wrong, we'll
2139 * just go back and repeat.
2141 rq = task_rq_lock(p, &flags);
2142 running = task_running(rq, p);
2143 on_rq = p->se.on_rq;
2144 task_rq_unlock(rq, &flags);
2147 * Was it really running after all now that we
2148 * checked with the proper locks actually held?
2150 * Oops. Go back and try again..
2152 if (unlikely(running)) {
2158 * It's not enough that it's not actively running,
2159 * it must be off the runqueue _entirely_, and not
2162 * So if it wa still runnable (but just not actively
2163 * running right now), it's preempted, and we should
2164 * yield - it could be a while.
2166 if (unlikely(on_rq)) {
2167 schedule_timeout_uninterruptible(1);
2172 * Ahh, all good. It wasn't running, and it wasn't
2173 * runnable, which means that it will never become
2174 * running in the future either. We're all done!
2181 * kick_process - kick a running thread to enter/exit the kernel
2182 * @p: the to-be-kicked thread
2184 * Cause a process which is running on another CPU to enter
2185 * kernel-mode, without any delay. (to get signals handled.)
2187 * NOTE: this function doesnt have to take the runqueue lock,
2188 * because all it wants to ensure is that the remote task enters
2189 * the kernel. If the IPI races and the task has been migrated
2190 * to another CPU then no harm is done and the purpose has been
2193 void kick_process(struct task_struct *p)
2199 if ((cpu != smp_processor_id()) && task_curr(p))
2200 smp_send_reschedule(cpu);
2205 * Return a low guess at the load of a migration-source cpu weighted
2206 * according to the scheduling class and "nice" value.
2208 * We want to under-estimate the load of migration sources, to
2209 * balance conservatively.
2211 static unsigned long source_load(int cpu, int type)
2213 struct rq *rq = cpu_rq(cpu);
2214 unsigned long total = weighted_cpuload(cpu);
2219 return min(rq->cpu_load[type-1], total);
2223 * Return a high guess at the load of a migration-target cpu weighted
2224 * according to the scheduling class and "nice" value.
2226 static unsigned long target_load(int cpu, int type)
2228 struct rq *rq = cpu_rq(cpu);
2229 unsigned long total = weighted_cpuload(cpu);
2234 return max(rq->cpu_load[type-1], total);
2238 * Return the average load per task on the cpu's run queue
2240 static unsigned long cpu_avg_load_per_task(int cpu)
2242 struct rq *rq = cpu_rq(cpu);
2243 unsigned long total = weighted_cpuload(cpu);
2244 unsigned long n = rq->nr_running;
2246 return n ? total / n : SCHED_LOAD_SCALE;
2250 * find_idlest_group finds and returns the least busy CPU group within the
2253 static struct sched_group *
2254 find_idlest_group(struct sched_domain *sd, struct task_struct *p, int this_cpu)
2256 struct sched_group *idlest = NULL, *this = NULL, *group = sd->groups;
2257 unsigned long min_load = ULONG_MAX, this_load = 0;
2258 int load_idx = sd->forkexec_idx;
2259 int imbalance = 100 + (sd->imbalance_pct-100)/2;
2262 unsigned long load, avg_load;
2266 /* Skip over this group if it has no CPUs allowed */
2267 if (!cpus_intersects(group->cpumask, p->cpus_allowed))
2270 local_group = cpu_isset(this_cpu, group->cpumask);
2272 /* Tally up the load of all CPUs in the group */
2275 for_each_cpu_mask(i, group->cpumask) {
2276 /* Bias balancing toward cpus of our domain */
2278 load = source_load(i, load_idx);
2280 load = target_load(i, load_idx);
2285 /* Adjust by relative CPU power of the group */
2286 avg_load = sg_div_cpu_power(group,
2287 avg_load * SCHED_LOAD_SCALE);
2290 this_load = avg_load;
2292 } else if (avg_load < min_load) {
2293 min_load = avg_load;
2296 } while (group = group->next, group != sd->groups);
2298 if (!idlest || 100*this_load < imbalance*min_load)
2304 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2307 find_idlest_cpu(struct sched_group *group, struct task_struct *p, int this_cpu,
2310 unsigned long load, min_load = ULONG_MAX;
2314 /* Traverse only the allowed CPUs */
2315 cpus_and(*tmp, group->cpumask, p->cpus_allowed);
2317 for_each_cpu_mask(i, *tmp) {
2318 load = weighted_cpuload(i);
2320 if (load < min_load || (load == min_load && i == this_cpu)) {
2330 * sched_balance_self: balance the current task (running on cpu) in domains
2331 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2334 * Balance, ie. select the least loaded group.
2336 * Returns the target CPU number, or the same CPU if no balancing is needed.
2338 * preempt must be disabled.
2340 static int sched_balance_self(int cpu, int flag)
2342 struct task_struct *t = current;
2343 struct sched_domain *tmp, *sd = NULL;
2345 for_each_domain(cpu, tmp) {
2347 * If power savings logic is enabled for a domain, stop there.
2349 if (tmp->flags & SD_POWERSAVINGS_BALANCE)
2351 if (tmp->flags & flag)
2356 cpumask_t span, tmpmask;
2357 struct sched_group *group;
2358 int new_cpu, weight;
2360 if (!(sd->flags & flag)) {
2366 group = find_idlest_group(sd, t, cpu);
2372 new_cpu = find_idlest_cpu(group, t, cpu, &tmpmask);
2373 if (new_cpu == -1 || new_cpu == cpu) {
2374 /* Now try balancing at a lower domain level of cpu */
2379 /* Now try balancing at a lower domain level of new_cpu */
2382 weight = cpus_weight(span);
2383 for_each_domain(cpu, tmp) {
2384 if (weight <= cpus_weight(tmp->span))
2386 if (tmp->flags & flag)
2389 /* while loop will break here if sd == NULL */
2395 #endif /* CONFIG_SMP */
2397 #ifdef CONFIG_CONTEXT_SWITCH_TRACER
2399 void ftrace_all_fair_tasks(void *__rq, void *__tr, void *__data)
2401 struct sched_entity *se;
2402 struct task_struct *p;
2403 struct rb_node *curr;
2404 struct rq *rq = __rq;
2406 curr = first_fair(&rq->cfs);
2411 se = rb_entry(curr, struct sched_entity, run_node);
2412 if (!entity_is_task(se))
2417 __trace_special(__tr, __data,
2418 p->pid, p->se.vruntime, p->se.sum_exec_runtime);
2420 curr = rb_next(curr);
2427 * try_to_wake_up - wake up a thread
2428 * @p: the to-be-woken-up thread
2429 * @state: the mask of task states that can be woken
2430 * @sync: do a synchronous wakeup?
2432 * Put it on the run-queue if it's not already there. The "current"
2433 * thread is always on the run-queue (except when the actual
2434 * re-schedule is in progress), and as such you're allowed to do
2435 * the simpler "current->state = TASK_RUNNING" to mark yourself
2436 * runnable without the overhead of this.
2438 * returns failure only if the task is already active.
2440 static int try_to_wake_up(struct task_struct *p, unsigned int state, int sync)
2442 int cpu, orig_cpu, this_cpu, success = 0;
2443 unsigned long flags;
2447 if (!sched_feat(SYNC_WAKEUPS))
2451 rq = task_rq_lock(p, &flags);
2452 old_state = p->state;
2453 if (!(old_state & state))
2461 this_cpu = smp_processor_id();
2464 if (unlikely(task_running(rq, p)))
2467 cpu = p->sched_class->select_task_rq(p, sync);
2468 if (cpu != orig_cpu) {
2469 set_task_cpu(p, cpu);
2470 task_rq_unlock(rq, &flags);
2471 /* might preempt at this point */
2472 rq = task_rq_lock(p, &flags);
2473 old_state = p->state;
2474 if (!(old_state & state))
2479 this_cpu = smp_processor_id();
2483 #ifdef CONFIG_SCHEDSTATS
2484 schedstat_inc(rq, ttwu_count);
2485 if (cpu == this_cpu)
2486 schedstat_inc(rq, ttwu_local);
2488 struct sched_domain *sd;
2489 for_each_domain(this_cpu, sd) {
2490 if (cpu_isset(cpu, sd->span)) {
2491 schedstat_inc(sd, ttwu_wake_remote);
2499 #endif /* CONFIG_SMP */
2500 ftrace_wake_up_task(rq, p, rq->curr);
2501 schedstat_inc(p, se.nr_wakeups);
2503 schedstat_inc(p, se.nr_wakeups_sync);
2504 if (orig_cpu != cpu)
2505 schedstat_inc(p, se.nr_wakeups_migrate);
2506 if (cpu == this_cpu)
2507 schedstat_inc(p, se.nr_wakeups_local);
2509 schedstat_inc(p, se.nr_wakeups_remote);
2510 update_rq_clock(rq);
2511 activate_task(rq, p, 1);
2515 check_preempt_curr(rq, p);
2517 p->state = TASK_RUNNING;
2519 if (p->sched_class->task_wake_up)
2520 p->sched_class->task_wake_up(rq, p);
2523 task_rq_unlock(rq, &flags);
2528 int wake_up_process(struct task_struct *p)
2530 return try_to_wake_up(p, TASK_ALL, 0);
2532 EXPORT_SYMBOL(wake_up_process);
2534 int wake_up_state(struct task_struct *p, unsigned int state)
2536 return try_to_wake_up(p, state, 0);
2540 * Perform scheduler related setup for a newly forked process p.
2541 * p is forked by current.
2543 * __sched_fork() is basic setup used by init_idle() too:
2545 static void __sched_fork(struct task_struct *p)
2547 p->se.exec_start = 0;
2548 p->se.sum_exec_runtime = 0;
2549 p->se.prev_sum_exec_runtime = 0;
2550 p->se.last_wakeup = 0;
2551 p->se.avg_overlap = 0;
2553 #ifdef CONFIG_SCHEDSTATS
2554 p->se.wait_start = 0;
2555 p->se.sum_sleep_runtime = 0;
2556 p->se.sleep_start = 0;
2557 p->se.block_start = 0;
2558 p->se.sleep_max = 0;
2559 p->se.block_max = 0;
2561 p->se.slice_max = 0;
2565 INIT_LIST_HEAD(&p->rt.run_list);
2567 INIT_LIST_HEAD(&p->se.group_node);
2569 #ifdef CONFIG_PREEMPT_NOTIFIERS
2570 INIT_HLIST_HEAD(&p->preempt_notifiers);
2574 * We mark the process as running here, but have not actually
2575 * inserted it onto the runqueue yet. This guarantees that
2576 * nobody will actually run it, and a signal or other external
2577 * event cannot wake it up and insert it on the runqueue either.
2579 p->state = TASK_RUNNING;
2583 * fork()/clone()-time setup:
2585 void sched_fork(struct task_struct *p, int clone_flags)
2587 int cpu = get_cpu();
2592 cpu = sched_balance_self(cpu, SD_BALANCE_FORK);
2594 set_task_cpu(p, cpu);
2597 * Make sure we do not leak PI boosting priority to the child:
2599 p->prio = current->normal_prio;
2600 if (!rt_prio(p->prio))
2601 p->sched_class = &fair_sched_class;
2603 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2604 if (likely(sched_info_on()))
2605 memset(&p->sched_info, 0, sizeof(p->sched_info));
2607 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2610 #ifdef CONFIG_PREEMPT
2611 /* Want to start with kernel preemption disabled. */
2612 task_thread_info(p)->preempt_count = 1;
2618 * wake_up_new_task - wake up a newly created task for the first time.
2620 * This function will do some initial scheduler statistics housekeeping
2621 * that must be done for every newly created context, then puts the task
2622 * on the runqueue and wakes it.
2624 void wake_up_new_task(struct task_struct *p, unsigned long clone_flags)
2626 unsigned long flags;
2629 rq = task_rq_lock(p, &flags);
2630 BUG_ON(p->state != TASK_RUNNING);
2631 update_rq_clock(rq);
2633 p->prio = effective_prio(p);
2635 if (!p->sched_class->task_new || !current->se.on_rq) {
2636 activate_task(rq, p, 0);
2639 * Let the scheduling class do new task startup
2640 * management (if any):
2642 p->sched_class->task_new(rq, p);
2645 ftrace_wake_up_task(rq, p, rq->curr);
2646 check_preempt_curr(rq, p);
2648 if (p->sched_class->task_wake_up)
2649 p->sched_class->task_wake_up(rq, p);
2651 task_rq_unlock(rq, &flags);
2654 #ifdef CONFIG_PREEMPT_NOTIFIERS
2657 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2658 * @notifier: notifier struct to register
2660 void preempt_notifier_register(struct preempt_notifier *notifier)
2662 hlist_add_head(¬ifier->link, ¤t->preempt_notifiers);
2664 EXPORT_SYMBOL_GPL(preempt_notifier_register);
2667 * preempt_notifier_unregister - no longer interested in preemption notifications
2668 * @notifier: notifier struct to unregister
2670 * This is safe to call from within a preemption notifier.
2672 void preempt_notifier_unregister(struct preempt_notifier *notifier)
2674 hlist_del(¬ifier->link);
2676 EXPORT_SYMBOL_GPL(preempt_notifier_unregister);
2678 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2680 struct preempt_notifier *notifier;
2681 struct hlist_node *node;
2683 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2684 notifier->ops->sched_in(notifier, raw_smp_processor_id());
2688 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2689 struct task_struct *next)
2691 struct preempt_notifier *notifier;
2692 struct hlist_node *node;
2694 hlist_for_each_entry(notifier, node, &curr->preempt_notifiers, link)
2695 notifier->ops->sched_out(notifier, next);
2700 static void fire_sched_in_preempt_notifiers(struct task_struct *curr)
2705 fire_sched_out_preempt_notifiers(struct task_struct *curr,
2706 struct task_struct *next)
2713 * prepare_task_switch - prepare to switch tasks
2714 * @rq: the runqueue preparing to switch
2715 * @prev: the current task that is being switched out
2716 * @next: the task we are going to switch to.
2718 * This is called with the rq lock held and interrupts off. It must
2719 * be paired with a subsequent finish_task_switch after the context
2722 * prepare_task_switch sets up locking and calls architecture specific
2726 prepare_task_switch(struct rq *rq, struct task_struct *prev,
2727 struct task_struct *next)
2729 fire_sched_out_preempt_notifiers(prev, next);
2730 prepare_lock_switch(rq, next);
2731 prepare_arch_switch(next);
2735 * finish_task_switch - clean up after a task-switch
2736 * @rq: runqueue associated with task-switch
2737 * @prev: the thread we just switched away from.
2739 * finish_task_switch must be called after the context switch, paired
2740 * with a prepare_task_switch call before the context switch.
2741 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2742 * and do any other architecture-specific cleanup actions.
2744 * Note that we may have delayed dropping an mm in context_switch(). If
2745 * so, we finish that here outside of the runqueue lock. (Doing it
2746 * with the lock held can cause deadlocks; see schedule() for
2749 static void finish_task_switch(struct rq *rq, struct task_struct *prev)
2750 __releases(rq->lock)
2752 struct mm_struct *mm = rq->prev_mm;
2758 * A task struct has one reference for the use as "current".
2759 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2760 * schedule one last time. The schedule call will never return, and
2761 * the scheduled task must drop that reference.
2762 * The test for TASK_DEAD must occur while the runqueue locks are
2763 * still held, otherwise prev could be scheduled on another cpu, die
2764 * there before we look at prev->state, and then the reference would
2766 * Manfred Spraul <manfred@colorfullife.com>
2768 prev_state = prev->state;
2769 finish_arch_switch(prev);
2770 finish_lock_switch(rq, prev);
2772 if (current->sched_class->post_schedule)
2773 current->sched_class->post_schedule(rq);
2776 fire_sched_in_preempt_notifiers(current);
2779 if (unlikely(prev_state == TASK_DEAD)) {
2781 * Remove function-return probe instances associated with this
2782 * task and put them back on the free list.
2784 kprobe_flush_task(prev);
2785 put_task_struct(prev);
2790 * schedule_tail - first thing a freshly forked thread must call.
2791 * @prev: the thread we just switched away from.
2793 asmlinkage void schedule_tail(struct task_struct *prev)
2794 __releases(rq->lock)
2796 struct rq *rq = this_rq();
2798 finish_task_switch(rq, prev);
2799 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2800 /* In this case, finish_task_switch does not reenable preemption */
2803 if (current->set_child_tid)
2804 put_user(task_pid_vnr(current), current->set_child_tid);
2808 * context_switch - switch to the new MM and the new
2809 * thread's register state.
2812 context_switch(struct rq *rq, struct task_struct *prev,
2813 struct task_struct *next)
2815 struct mm_struct *mm, *oldmm;
2817 prepare_task_switch(rq, prev, next);
2818 ftrace_ctx_switch(rq, prev, next);
2820 oldmm = prev->active_mm;
2822 * For paravirt, this is coupled with an exit in switch_to to
2823 * combine the page table reload and the switch backend into
2826 arch_enter_lazy_cpu_mode();
2828 if (unlikely(!mm)) {
2829 next->active_mm = oldmm;
2830 atomic_inc(&oldmm->mm_count);
2831 enter_lazy_tlb(oldmm, next);
2833 switch_mm(oldmm, mm, next);
2835 if (unlikely(!prev->mm)) {
2836 prev->active_mm = NULL;
2837 rq->prev_mm = oldmm;
2840 * Since the runqueue lock will be released by the next
2841 * task (which is an invalid locking op but in the case
2842 * of the scheduler it's an obvious special-case), so we
2843 * do an early lockdep release here:
2845 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2846 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
2849 /* Here we just switch the register state and the stack. */
2850 switch_to(prev, next, prev);
2854 * this_rq must be evaluated again because prev may have moved
2855 * CPUs since it called schedule(), thus the 'rq' on its stack
2856 * frame will be invalid.
2858 finish_task_switch(this_rq(), prev);
2862 * nr_running, nr_uninterruptible and nr_context_switches:
2864 * externally visible scheduler statistics: current number of runnable
2865 * threads, current number of uninterruptible-sleeping threads, total
2866 * number of context switches performed since bootup.
2868 unsigned long nr_running(void)
2870 unsigned long i, sum = 0;
2872 for_each_online_cpu(i)
2873 sum += cpu_rq(i)->nr_running;
2878 unsigned long nr_uninterruptible(void)
2880 unsigned long i, sum = 0;
2882 for_each_possible_cpu(i)
2883 sum += cpu_rq(i)->nr_uninterruptible;
2886 * Since we read the counters lockless, it might be slightly
2887 * inaccurate. Do not allow it to go below zero though:
2889 if (unlikely((long)sum < 0))
2895 unsigned long long nr_context_switches(void)
2898 unsigned long long sum = 0;
2900 for_each_possible_cpu(i)
2901 sum += cpu_rq(i)->nr_switches;
2906 unsigned long nr_iowait(void)
2908 unsigned long i, sum = 0;
2910 for_each_possible_cpu(i)
2911 sum += atomic_read(&cpu_rq(i)->nr_iowait);
2916 unsigned long nr_active(void)
2918 unsigned long i, running = 0, uninterruptible = 0;
2920 for_each_online_cpu(i) {
2921 running += cpu_rq(i)->nr_running;
2922 uninterruptible += cpu_rq(i)->nr_uninterruptible;
2925 if (unlikely((long)uninterruptible < 0))
2926 uninterruptible = 0;
2928 return running + uninterruptible;
2932 * Update rq->cpu_load[] statistics. This function is usually called every
2933 * scheduler tick (TICK_NSEC).
2935 static void update_cpu_load(struct rq *this_rq)
2937 unsigned long this_load = this_rq->load.weight;
2940 this_rq->nr_load_updates++;
2942 /* Update our load: */
2943 for (i = 0, scale = 1; i < CPU_LOAD_IDX_MAX; i++, scale += scale) {
2944 unsigned long old_load, new_load;
2946 /* scale is effectively 1 << i now, and >> i divides by scale */
2948 old_load = this_rq->cpu_load[i];
2949 new_load = this_load;
2951 * Round up the averaging division if load is increasing. This
2952 * prevents us from getting stuck on 9 if the load is 10, for
2955 if (new_load > old_load)
2956 new_load += scale-1;
2957 this_rq->cpu_load[i] = (old_load*(scale-1) + new_load) >> i;
2964 * double_rq_lock - safely lock two runqueues
2966 * Note this does not disable interrupts like task_rq_lock,
2967 * you need to do so manually before calling.
2969 static void double_rq_lock(struct rq *rq1, struct rq *rq2)
2970 __acquires(rq1->lock)
2971 __acquires(rq2->lock)
2973 BUG_ON(!irqs_disabled());
2975 spin_lock(&rq1->lock);
2976 __acquire(rq2->lock); /* Fake it out ;) */
2979 spin_lock(&rq1->lock);
2980 spin_lock(&rq2->lock);
2982 spin_lock(&rq2->lock);
2983 spin_lock(&rq1->lock);
2986 update_rq_clock(rq1);
2987 update_rq_clock(rq2);
2991 * double_rq_unlock - safely unlock two runqueues
2993 * Note this does not restore interrupts like task_rq_unlock,
2994 * you need to do so manually after calling.
2996 static void double_rq_unlock(struct rq *rq1, struct rq *rq2)
2997 __releases(rq1->lock)
2998 __releases(rq2->lock)
3000 spin_unlock(&rq1->lock);
3002 spin_unlock(&rq2->lock);
3004 __release(rq2->lock);
3008 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3010 static int double_lock_balance(struct rq *this_rq, struct rq *busiest)
3011 __releases(this_rq->lock)
3012 __acquires(busiest->lock)
3013 __acquires(this_rq->lock)
3017 if (unlikely(!irqs_disabled())) {
3018 /* printk() doesn't work good under rq->lock */
3019 spin_unlock(&this_rq->lock);
3022 if (unlikely(!spin_trylock(&busiest->lock))) {
3023 if (busiest < this_rq) {
3024 spin_unlock(&this_rq->lock);
3025 spin_lock(&busiest->lock);
3026 spin_lock(&this_rq->lock);
3029 spin_lock(&busiest->lock);
3035 * If dest_cpu is allowed for this process, migrate the task to it.
3036 * This is accomplished by forcing the cpu_allowed mask to only
3037 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3038 * the cpu_allowed mask is restored.
3040 static void sched_migrate_task(struct task_struct *p, int dest_cpu)
3042 struct migration_req req;
3043 unsigned long flags;
3046 rq = task_rq_lock(p, &flags);
3047 if (!cpu_isset(dest_cpu, p->cpus_allowed)
3048 || unlikely(cpu_is_offline(dest_cpu)))
3051 /* force the process onto the specified CPU */
3052 if (migrate_task(p, dest_cpu, &req)) {
3053 /* Need to wait for migration thread (might exit: take ref). */
3054 struct task_struct *mt = rq->migration_thread;
3056 get_task_struct(mt);
3057 task_rq_unlock(rq, &flags);
3058 wake_up_process(mt);
3059 put_task_struct(mt);
3060 wait_for_completion(&req.done);
3065 task_rq_unlock(rq, &flags);
3069 * sched_exec - execve() is a valuable balancing opportunity, because at
3070 * this point the task has the smallest effective memory and cache footprint.
3072 void sched_exec(void)
3074 int new_cpu, this_cpu = get_cpu();
3075 new_cpu = sched_balance_self(this_cpu, SD_BALANCE_EXEC);
3077 if (new_cpu != this_cpu)
3078 sched_migrate_task(current, new_cpu);
3082 * pull_task - move a task from a remote runqueue to the local runqueue.
3083 * Both runqueues must be locked.
3085 static void pull_task(struct rq *src_rq, struct task_struct *p,
3086 struct rq *this_rq, int this_cpu)
3088 deactivate_task(src_rq, p, 0);
3089 set_task_cpu(p, this_cpu);
3090 activate_task(this_rq, p, 0);
3092 * Note that idle threads have a prio of MAX_PRIO, for this test
3093 * to be always true for them.
3095 check_preempt_curr(this_rq, p);
3099 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3102 int can_migrate_task(struct task_struct *p, struct rq *rq, int this_cpu,
3103 struct sched_domain *sd, enum cpu_idle_type idle,
3107 * We do not migrate tasks that are:
3108 * 1) running (obviously), or
3109 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3110 * 3) are cache-hot on their current CPU.
3112 if (!cpu_isset(this_cpu, p->cpus_allowed)) {
3113 schedstat_inc(p, se.nr_failed_migrations_affine);
3118 if (task_running(rq, p)) {
3119 schedstat_inc(p, se.nr_failed_migrations_running);
3124 * Aggressive migration if:
3125 * 1) task is cache cold, or
3126 * 2) too many balance attempts have failed.
3129 if (!task_hot(p, rq->clock, sd) ||
3130 sd->nr_balance_failed > sd->cache_nice_tries) {
3131 #ifdef CONFIG_SCHEDSTATS
3132 if (task_hot(p, rq->clock, sd)) {
3133 schedstat_inc(sd, lb_hot_gained[idle]);
3134 schedstat_inc(p, se.nr_forced_migrations);
3140 if (task_hot(p, rq->clock, sd)) {
3141 schedstat_inc(p, se.nr_failed_migrations_hot);
3147 static unsigned long
3148 balance_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3149 unsigned long max_load_move, struct sched_domain *sd,
3150 enum cpu_idle_type idle, int *all_pinned,
3151 int *this_best_prio, struct rq_iterator *iterator)
3153 int loops = 0, pulled = 0, pinned = 0, skip_for_load;
3154 struct task_struct *p;
3155 long rem_load_move = max_load_move;
3157 if (max_load_move == 0)
3163 * Start the load-balancing iterator:
3165 p = iterator->start(iterator->arg);
3167 if (!p || loops++ > sysctl_sched_nr_migrate)
3170 * To help distribute high priority tasks across CPUs we don't
3171 * skip a task if it will be the highest priority task (i.e. smallest
3172 * prio value) on its new queue regardless of its load weight
3174 skip_for_load = (p->se.load.weight >> 1) > rem_load_move +
3175 SCHED_LOAD_SCALE_FUZZ;
3176 if ((skip_for_load && p->prio >= *this_best_prio) ||
3177 !can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3178 p = iterator->next(iterator->arg);
3182 pull_task(busiest, p, this_rq, this_cpu);
3184 rem_load_move -= p->se.load.weight;
3187 * We only want to steal up to the prescribed amount of weighted load.
3189 if (rem_load_move > 0) {
3190 if (p->prio < *this_best_prio)
3191 *this_best_prio = p->prio;
3192 p = iterator->next(iterator->arg);
3197 * Right now, this is one of only two places pull_task() is called,
3198 * so we can safely collect pull_task() stats here rather than
3199 * inside pull_task().
3201 schedstat_add(sd, lb_gained[idle], pulled);
3204 *all_pinned = pinned;
3206 return max_load_move - rem_load_move;
3210 * move_tasks tries to move up to max_load_move weighted load from busiest to
3211 * this_rq, as part of a balancing operation within domain "sd".
3212 * Returns 1 if successful and 0 otherwise.
3214 * Called with both runqueues locked.
3216 static int move_tasks(struct rq *this_rq, int this_cpu, struct rq *busiest,
3217 unsigned long max_load_move,
3218 struct sched_domain *sd, enum cpu_idle_type idle,
3221 const struct sched_class *class = sched_class_highest;
3222 unsigned long total_load_moved = 0;
3223 int this_best_prio = this_rq->curr->prio;
3227 class->load_balance(this_rq, this_cpu, busiest,
3228 max_load_move - total_load_moved,
3229 sd, idle, all_pinned, &this_best_prio);
3230 class = class->next;
3231 } while (class && max_load_move > total_load_moved);
3233 return total_load_moved > 0;
3237 iter_move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3238 struct sched_domain *sd, enum cpu_idle_type idle,
3239 struct rq_iterator *iterator)
3241 struct task_struct *p = iterator->start(iterator->arg);
3245 if (can_migrate_task(p, busiest, this_cpu, sd, idle, &pinned)) {
3246 pull_task(busiest, p, this_rq, this_cpu);
3248 * Right now, this is only the second place pull_task()
3249 * is called, so we can safely collect pull_task()
3250 * stats here rather than inside pull_task().
3252 schedstat_inc(sd, lb_gained[idle]);
3256 p = iterator->next(iterator->arg);
3263 * move_one_task tries to move exactly one task from busiest to this_rq, as
3264 * part of active balancing operations within "domain".
3265 * Returns 1 if successful and 0 otherwise.
3267 * Called with both runqueues locked.
3269 static int move_one_task(struct rq *this_rq, int this_cpu, struct rq *busiest,
3270 struct sched_domain *sd, enum cpu_idle_type idle)
3272 const struct sched_class *class;
3274 for (class = sched_class_highest; class; class = class->next)
3275 if (class->move_one_task(this_rq, this_cpu, busiest, sd, idle))
3282 * find_busiest_group finds and returns the busiest CPU group within the
3283 * domain. It calculates and returns the amount of weighted load which
3284 * should be moved to restore balance via the imbalance parameter.
3286 static struct sched_group *
3287 find_busiest_group(struct sched_domain *sd, int this_cpu,
3288 unsigned long *imbalance, enum cpu_idle_type idle,
3289 int *sd_idle, const cpumask_t *cpus, int *balance)
3291 struct sched_group *busiest = NULL, *this = NULL, *group = sd->groups;
3292 unsigned long max_load, avg_load, total_load, this_load, total_pwr;
3293 unsigned long max_pull;
3294 unsigned long busiest_load_per_task, busiest_nr_running;
3295 unsigned long this_load_per_task, this_nr_running;
3296 int load_idx, group_imb = 0;
3297 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3298 int power_savings_balance = 1;
3299 unsigned long leader_nr_running = 0, min_load_per_task = 0;
3300 unsigned long min_nr_running = ULONG_MAX;
3301 struct sched_group *group_min = NULL, *group_leader = NULL;
3304 max_load = this_load = total_load = total_pwr = 0;
3305 busiest_load_per_task = busiest_nr_running = 0;
3306 this_load_per_task = this_nr_running = 0;
3307 if (idle == CPU_NOT_IDLE)
3308 load_idx = sd->busy_idx;
3309 else if (idle == CPU_NEWLY_IDLE)
3310 load_idx = sd->newidle_idx;
3312 load_idx = sd->idle_idx;
3315 unsigned long load, group_capacity, max_cpu_load, min_cpu_load;
3318 int __group_imb = 0;
3319 unsigned int balance_cpu = -1, first_idle_cpu = 0;
3320 unsigned long sum_nr_running, sum_weighted_load;
3322 local_group = cpu_isset(this_cpu, group->cpumask);
3325 balance_cpu = first_cpu(group->cpumask);
3327 /* Tally up the load of all CPUs in the group */
3328 sum_weighted_load = sum_nr_running = avg_load = 0;
3330 min_cpu_load = ~0UL;
3332 for_each_cpu_mask(i, group->cpumask) {
3335 if (!cpu_isset(i, *cpus))
3340 if (*sd_idle && rq->nr_running)
3343 /* Bias balancing toward cpus of our domain */
3345 if (idle_cpu(i) && !first_idle_cpu) {
3350 load = target_load(i, load_idx);
3352 load = source_load(i, load_idx);
3353 if (load > max_cpu_load)
3354 max_cpu_load = load;
3355 if (min_cpu_load > load)
3356 min_cpu_load = load;
3360 sum_nr_running += rq->nr_running;
3361 sum_weighted_load += weighted_cpuload(i);
3365 * First idle cpu or the first cpu(busiest) in this sched group
3366 * is eligible for doing load balancing at this and above
3367 * domains. In the newly idle case, we will allow all the cpu's
3368 * to do the newly idle load balance.
3370 if (idle != CPU_NEWLY_IDLE && local_group &&
3371 balance_cpu != this_cpu && balance) {
3376 total_load += avg_load;
3377 total_pwr += group->__cpu_power;
3379 /* Adjust by relative CPU power of the group */
3380 avg_load = sg_div_cpu_power(group,
3381 avg_load * SCHED_LOAD_SCALE);
3383 if ((max_cpu_load - min_cpu_load) > SCHED_LOAD_SCALE)
3386 group_capacity = group->__cpu_power / SCHED_LOAD_SCALE;
3389 this_load = avg_load;
3391 this_nr_running = sum_nr_running;
3392 this_load_per_task = sum_weighted_load;
3393 } else if (avg_load > max_load &&
3394 (sum_nr_running > group_capacity || __group_imb)) {
3395 max_load = avg_load;
3397 busiest_nr_running = sum_nr_running;
3398 busiest_load_per_task = sum_weighted_load;
3399 group_imb = __group_imb;
3402 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3404 * Busy processors will not participate in power savings
3407 if (idle == CPU_NOT_IDLE ||
3408 !(sd->flags & SD_POWERSAVINGS_BALANCE))
3412 * If the local group is idle or completely loaded
3413 * no need to do power savings balance at this domain
3415 if (local_group && (this_nr_running >= group_capacity ||
3417 power_savings_balance = 0;
3420 * If a group is already running at full capacity or idle,
3421 * don't include that group in power savings calculations
3423 if (!power_savings_balance || sum_nr_running >= group_capacity
3428 * Calculate the group which has the least non-idle load.
3429 * This is the group from where we need to pick up the load
3432 if ((sum_nr_running < min_nr_running) ||
3433 (sum_nr_running == min_nr_running &&
3434 first_cpu(group->cpumask) <
3435 first_cpu(group_min->cpumask))) {
3437 min_nr_running = sum_nr_running;
3438 min_load_per_task = sum_weighted_load /
3443 * Calculate the group which is almost near its
3444 * capacity but still has some space to pick up some load
3445 * from other group and save more power
3447 if (sum_nr_running <= group_capacity - 1) {
3448 if (sum_nr_running > leader_nr_running ||
3449 (sum_nr_running == leader_nr_running &&
3450 first_cpu(group->cpumask) >
3451 first_cpu(group_leader->cpumask))) {
3452 group_leader = group;
3453 leader_nr_running = sum_nr_running;
3458 group = group->next;
3459 } while (group != sd->groups);
3461 if (!busiest || this_load >= max_load || busiest_nr_running == 0)
3464 avg_load = (SCHED_LOAD_SCALE * total_load) / total_pwr;
3466 if (this_load >= avg_load ||
3467 100*max_load <= sd->imbalance_pct*this_load)
3470 busiest_load_per_task /= busiest_nr_running;
3472 busiest_load_per_task = min(busiest_load_per_task, avg_load);
3475 * We're trying to get all the cpus to the average_load, so we don't
3476 * want to push ourselves above the average load, nor do we wish to
3477 * reduce the max loaded cpu below the average load, as either of these
3478 * actions would just result in more rebalancing later, and ping-pong
3479 * tasks around. Thus we look for the minimum possible imbalance.
3480 * Negative imbalances (*we* are more loaded than anyone else) will
3481 * be counted as no imbalance for these purposes -- we can't fix that
3482 * by pulling tasks to us. Be careful of negative numbers as they'll
3483 * appear as very large values with unsigned longs.
3485 if (max_load <= busiest_load_per_task)
3489 * In the presence of smp nice balancing, certain scenarios can have
3490 * max load less than avg load(as we skip the groups at or below
3491 * its cpu_power, while calculating max_load..)
3493 if (max_load < avg_load) {
3495 goto small_imbalance;
3498 /* Don't want to pull so many tasks that a group would go idle */
3499 max_pull = min(max_load - avg_load, max_load - busiest_load_per_task);
3501 /* How much load to actually move to equalise the imbalance */
3502 *imbalance = min(max_pull * busiest->__cpu_power,
3503 (avg_load - this_load) * this->__cpu_power)
3507 * if *imbalance is less than the average load per runnable task
3508 * there is no gaurantee that any tasks will be moved so we'll have
3509 * a think about bumping its value to force at least one task to be
3512 if (*imbalance < busiest_load_per_task) {
3513 unsigned long tmp, pwr_now, pwr_move;
3517 pwr_move = pwr_now = 0;
3519 if (this_nr_running) {
3520 this_load_per_task /= this_nr_running;
3521 if (busiest_load_per_task > this_load_per_task)
3524 this_load_per_task = SCHED_LOAD_SCALE;
3526 if (max_load - this_load + SCHED_LOAD_SCALE_FUZZ >=
3527 busiest_load_per_task * imbn) {
3528 *imbalance = busiest_load_per_task;
3533 * OK, we don't have enough imbalance to justify moving tasks,
3534 * however we may be able to increase total CPU power used by
3538 pwr_now += busiest->__cpu_power *
3539 min(busiest_load_per_task, max_load);
3540 pwr_now += this->__cpu_power *
3541 min(this_load_per_task, this_load);
3542 pwr_now /= SCHED_LOAD_SCALE;
3544 /* Amount of load we'd subtract */
3545 tmp = sg_div_cpu_power(busiest,
3546 busiest_load_per_task * SCHED_LOAD_SCALE);
3548 pwr_move += busiest->__cpu_power *
3549 min(busiest_load_per_task, max_load - tmp);
3551 /* Amount of load we'd add */
3552 if (max_load * busiest->__cpu_power <
3553 busiest_load_per_task * SCHED_LOAD_SCALE)
3554 tmp = sg_div_cpu_power(this,
3555 max_load * busiest->__cpu_power);
3557 tmp = sg_div_cpu_power(this,
3558 busiest_load_per_task * SCHED_LOAD_SCALE);
3559 pwr_move += this->__cpu_power *
3560 min(this_load_per_task, this_load + tmp);
3561 pwr_move /= SCHED_LOAD_SCALE;
3563 /* Move if we gain throughput */
3564 if (pwr_move > pwr_now)
3565 *imbalance = busiest_load_per_task;
3571 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3572 if (idle == CPU_NOT_IDLE || !(sd->flags & SD_POWERSAVINGS_BALANCE))
3575 if (this == group_leader && group_leader != group_min) {
3576 *imbalance = min_load_per_task;
3586 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3589 find_busiest_queue(struct sched_group *group, enum cpu_idle_type idle,
3590 unsigned long imbalance, const cpumask_t *cpus)
3592 struct rq *busiest = NULL, *rq;
3593 unsigned long max_load = 0;
3596 for_each_cpu_mask(i, group->cpumask) {
3599 if (!cpu_isset(i, *cpus))
3603 wl = weighted_cpuload(i);
3605 if (rq->nr_running == 1 && wl > imbalance)
3608 if (wl > max_load) {
3618 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3619 * so long as it is large enough.
3621 #define MAX_PINNED_INTERVAL 512
3624 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3625 * tasks if there is an imbalance.
3627 static int load_balance(int this_cpu, struct rq *this_rq,
3628 struct sched_domain *sd, enum cpu_idle_type idle,
3629 int *balance, cpumask_t *cpus)
3631 int ld_moved, all_pinned = 0, active_balance = 0, sd_idle = 0;
3632 struct sched_group *group;
3633 unsigned long imbalance;
3635 unsigned long flags;
3636 int unlock_aggregate;
3640 unlock_aggregate = get_aggregate(sd);
3643 * When power savings policy is enabled for the parent domain, idle
3644 * sibling can pick up load irrespective of busy siblings. In this case,
3645 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3646 * portraying it as CPU_NOT_IDLE.
3648 if (idle != CPU_NOT_IDLE && sd->flags & SD_SHARE_CPUPOWER &&
3649 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3652 schedstat_inc(sd, lb_count[idle]);
3655 group = find_busiest_group(sd, this_cpu, &imbalance, idle, &sd_idle,
3662 schedstat_inc(sd, lb_nobusyg[idle]);
3666 busiest = find_busiest_queue(group, idle, imbalance, cpus);
3668 schedstat_inc(sd, lb_nobusyq[idle]);
3672 BUG_ON(busiest == this_rq);
3674 schedstat_add(sd, lb_imbalance[idle], imbalance);
3677 if (busiest->nr_running > 1) {
3679 * Attempt to move tasks. If find_busiest_group has found
3680 * an imbalance but busiest->nr_running <= 1, the group is
3681 * still unbalanced. ld_moved simply stays zero, so it is
3682 * correctly treated as an imbalance.
3684 local_irq_save(flags);
3685 double_rq_lock(this_rq, busiest);
3686 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3687 imbalance, sd, idle, &all_pinned);
3688 double_rq_unlock(this_rq, busiest);
3689 local_irq_restore(flags);
3692 * some other cpu did the load balance for us.
3694 if (ld_moved && this_cpu != smp_processor_id())
3695 resched_cpu(this_cpu);
3697 /* All tasks on this runqueue were pinned by CPU affinity */
3698 if (unlikely(all_pinned)) {
3699 cpu_clear(cpu_of(busiest), *cpus);
3700 if (!cpus_empty(*cpus))
3707 schedstat_inc(sd, lb_failed[idle]);
3708 sd->nr_balance_failed++;
3710 if (unlikely(sd->nr_balance_failed > sd->cache_nice_tries+2)) {
3712 spin_lock_irqsave(&busiest->lock, flags);
3714 /* don't kick the migration_thread, if the curr
3715 * task on busiest cpu can't be moved to this_cpu
3717 if (!cpu_isset(this_cpu, busiest->curr->cpus_allowed)) {
3718 spin_unlock_irqrestore(&busiest->lock, flags);
3720 goto out_one_pinned;
3723 if (!busiest->active_balance) {
3724 busiest->active_balance = 1;
3725 busiest->push_cpu = this_cpu;
3728 spin_unlock_irqrestore(&busiest->lock, flags);
3730 wake_up_process(busiest->migration_thread);
3733 * We've kicked active balancing, reset the failure
3736 sd->nr_balance_failed = sd->cache_nice_tries+1;
3739 sd->nr_balance_failed = 0;
3741 if (likely(!active_balance)) {
3742 /* We were unbalanced, so reset the balancing interval */
3743 sd->balance_interval = sd->min_interval;
3746 * If we've begun active balancing, start to back off. This
3747 * case may not be covered by the all_pinned logic if there
3748 * is only 1 task on the busy runqueue (because we don't call
3751 if (sd->balance_interval < sd->max_interval)
3752 sd->balance_interval *= 2;
3755 if (!ld_moved && !sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3756 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3762 schedstat_inc(sd, lb_balanced[idle]);
3764 sd->nr_balance_failed = 0;
3767 /* tune up the balancing interval */
3768 if ((all_pinned && sd->balance_interval < MAX_PINNED_INTERVAL) ||
3769 (sd->balance_interval < sd->max_interval))
3770 sd->balance_interval *= 2;
3772 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3773 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3778 if (unlock_aggregate)
3784 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3785 * tasks if there is an imbalance.
3787 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3788 * this_rq is locked.
3791 load_balance_newidle(int this_cpu, struct rq *this_rq, struct sched_domain *sd,
3794 struct sched_group *group;
3795 struct rq *busiest = NULL;
3796 unsigned long imbalance;
3804 * When power savings policy is enabled for the parent domain, idle
3805 * sibling can pick up load irrespective of busy siblings. In this case,
3806 * let the state of idle sibling percolate up as IDLE, instead of
3807 * portraying it as CPU_NOT_IDLE.
3809 if (sd->flags & SD_SHARE_CPUPOWER &&
3810 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3813 schedstat_inc(sd, lb_count[CPU_NEWLY_IDLE]);
3815 group = find_busiest_group(sd, this_cpu, &imbalance, CPU_NEWLY_IDLE,
3816 &sd_idle, cpus, NULL);
3818 schedstat_inc(sd, lb_nobusyg[CPU_NEWLY_IDLE]);
3822 busiest = find_busiest_queue(group, CPU_NEWLY_IDLE, imbalance, cpus);
3824 schedstat_inc(sd, lb_nobusyq[CPU_NEWLY_IDLE]);
3828 BUG_ON(busiest == this_rq);
3830 schedstat_add(sd, lb_imbalance[CPU_NEWLY_IDLE], imbalance);
3833 if (busiest->nr_running > 1) {
3834 /* Attempt to move tasks */
3835 double_lock_balance(this_rq, busiest);
3836 /* this_rq->clock is already updated */
3837 update_rq_clock(busiest);
3838 ld_moved = move_tasks(this_rq, this_cpu, busiest,
3839 imbalance, sd, CPU_NEWLY_IDLE,
3841 spin_unlock(&busiest->lock);
3843 if (unlikely(all_pinned)) {
3844 cpu_clear(cpu_of(busiest), *cpus);
3845 if (!cpus_empty(*cpus))
3851 schedstat_inc(sd, lb_failed[CPU_NEWLY_IDLE]);
3852 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3853 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3856 sd->nr_balance_failed = 0;
3861 schedstat_inc(sd, lb_balanced[CPU_NEWLY_IDLE]);
3862 if (!sd_idle && sd->flags & SD_SHARE_CPUPOWER &&
3863 !test_sd_parent(sd, SD_POWERSAVINGS_BALANCE))
3865 sd->nr_balance_failed = 0;
3871 * idle_balance is called by schedule() if this_cpu is about to become
3872 * idle. Attempts to pull tasks from other CPUs.
3874 static void idle_balance(int this_cpu, struct rq *this_rq)
3876 struct sched_domain *sd;
3877 int pulled_task = -1;
3878 unsigned long next_balance = jiffies + HZ;
3881 for_each_domain(this_cpu, sd) {
3882 unsigned long interval;
3884 if (!(sd->flags & SD_LOAD_BALANCE))
3887 if (sd->flags & SD_BALANCE_NEWIDLE)
3888 /* If we've pulled tasks over stop searching: */
3889 pulled_task = load_balance_newidle(this_cpu, this_rq,
3892 interval = msecs_to_jiffies(sd->balance_interval);
3893 if (time_after(next_balance, sd->last_balance + interval))
3894 next_balance = sd->last_balance + interval;
3898 if (pulled_task || time_after(jiffies, this_rq->next_balance)) {
3900 * We are going idle. next_balance may be set based on
3901 * a busy processor. So reset next_balance.
3903 this_rq->next_balance = next_balance;
3908 * active_load_balance is run by migration threads. It pushes running tasks
3909 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3910 * running on each physical CPU where possible, and avoids physical /
3911 * logical imbalances.
3913 * Called with busiest_rq locked.
3915 static void active_load_balance(struct rq *busiest_rq, int busiest_cpu)
3917 int target_cpu = busiest_rq->push_cpu;
3918 struct sched_domain *sd;
3919 struct rq *target_rq;
3921 /* Is there any task to move? */
3922 if (busiest_rq->nr_running <= 1)
3925 target_rq = cpu_rq(target_cpu);
3928 * This condition is "impossible", if it occurs
3929 * we need to fix it. Originally reported by
3930 * Bjorn Helgaas on a 128-cpu setup.
3932 BUG_ON(busiest_rq == target_rq);
3934 /* move a task from busiest_rq to target_rq */
3935 double_lock_balance(busiest_rq, target_rq);
3936 update_rq_clock(busiest_rq);
3937 update_rq_clock(target_rq);
3939 /* Search for an sd spanning us and the target CPU. */
3940 for_each_domain(target_cpu, sd) {
3941 if ((sd->flags & SD_LOAD_BALANCE) &&
3942 cpu_isset(busiest_cpu, sd->span))
3947 schedstat_inc(sd, alb_count);
3949 if (move_one_task(target_rq, target_cpu, busiest_rq,
3951 schedstat_inc(sd, alb_pushed);
3953 schedstat_inc(sd, alb_failed);
3955 spin_unlock(&target_rq->lock);
3960 atomic_t load_balancer;
3962 } nohz ____cacheline_aligned = {
3963 .load_balancer = ATOMIC_INIT(-1),
3964 .cpu_mask = CPU_MASK_NONE,
3968 * This routine will try to nominate the ilb (idle load balancing)
3969 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3970 * load balancing on behalf of all those cpus. If all the cpus in the system
3971 * go into this tickless mode, then there will be no ilb owner (as there is
3972 * no need for one) and all the cpus will sleep till the next wakeup event
3975 * For the ilb owner, tick is not stopped. And this tick will be used
3976 * for idle load balancing. ilb owner will still be part of
3979 * While stopping the tick, this cpu will become the ilb owner if there
3980 * is no other owner. And will be the owner till that cpu becomes busy
3981 * or if all cpus in the system stop their ticks at which point
3982 * there is no need for ilb owner.
3984 * When the ilb owner becomes busy, it nominates another owner, during the
3985 * next busy scheduler_tick()
3987 int select_nohz_load_balancer(int stop_tick)
3989 int cpu = smp_processor_id();
3992 cpu_set(cpu, nohz.cpu_mask);
3993 cpu_rq(cpu)->in_nohz_recently = 1;
3996 * If we are going offline and still the leader, give up!
3998 if (cpu_is_offline(cpu) &&
3999 atomic_read(&nohz.load_balancer) == cpu) {
4000 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4005 /* time for ilb owner also to sleep */
4006 if (cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4007 if (atomic_read(&nohz.load_balancer) == cpu)
4008 atomic_set(&nohz.load_balancer, -1);
4012 if (atomic_read(&nohz.load_balancer) == -1) {
4013 /* make me the ilb owner */
4014 if (atomic_cmpxchg(&nohz.load_balancer, -1, cpu) == -1)
4016 } else if (atomic_read(&nohz.load_balancer) == cpu)
4019 if (!cpu_isset(cpu, nohz.cpu_mask))
4022 cpu_clear(cpu, nohz.cpu_mask);
4024 if (atomic_read(&nohz.load_balancer) == cpu)
4025 if (atomic_cmpxchg(&nohz.load_balancer, cpu, -1) != cpu)
4032 static DEFINE_SPINLOCK(balancing);
4035 * It checks each scheduling domain to see if it is due to be balanced,
4036 * and initiates a balancing operation if so.
4038 * Balancing parameters are set up in arch_init_sched_domains.
4040 static void rebalance_domains(int cpu, enum cpu_idle_type idle)
4043 struct rq *rq = cpu_rq(cpu);
4044 unsigned long interval;
4045 struct sched_domain *sd;
4046 /* Earliest time when we have to do rebalance again */
4047 unsigned long next_balance = jiffies + 60*HZ;
4048 int update_next_balance = 0;
4051 for_each_domain(cpu, sd) {
4052 if (!(sd->flags & SD_LOAD_BALANCE))
4055 interval = sd->balance_interval;
4056 if (idle != CPU_IDLE)
4057 interval *= sd->busy_factor;
4059 /* scale ms to jiffies */
4060 interval = msecs_to_jiffies(interval);
4061 if (unlikely(!interval))
4063 if (interval > HZ*NR_CPUS/10)
4064 interval = HZ*NR_CPUS/10;
4067 if (sd->flags & SD_SERIALIZE) {
4068 if (!spin_trylock(&balancing))
4072 if (time_after_eq(jiffies, sd->last_balance + interval)) {
4073 if (load_balance(cpu, rq, sd, idle, &balance, &tmp)) {
4075 * We've pulled tasks over so either we're no
4076 * longer idle, or one of our SMT siblings is
4079 idle = CPU_NOT_IDLE;
4081 sd->last_balance = jiffies;
4083 if (sd->flags & SD_SERIALIZE)
4084 spin_unlock(&balancing);
4086 if (time_after(next_balance, sd->last_balance + interval)) {
4087 next_balance = sd->last_balance + interval;
4088 update_next_balance = 1;
4092 * Stop the load balance at this level. There is another
4093 * CPU in our sched group which is doing load balancing more
4101 * next_balance will be updated only when there is a need.
4102 * When the cpu is attached to null domain for ex, it will not be
4105 if (likely(update_next_balance))
4106 rq->next_balance = next_balance;
4110 * run_rebalance_domains is triggered when needed from the scheduler tick.
4111 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4112 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4114 static void run_rebalance_domains(struct softirq_action *h)
4116 int this_cpu = smp_processor_id();
4117 struct rq *this_rq = cpu_rq(this_cpu);
4118 enum cpu_idle_type idle = this_rq->idle_at_tick ?
4119 CPU_IDLE : CPU_NOT_IDLE;
4121 rebalance_domains(this_cpu, idle);
4125 * If this cpu is the owner for idle load balancing, then do the
4126 * balancing on behalf of the other idle cpus whose ticks are
4129 if (this_rq->idle_at_tick &&
4130 atomic_read(&nohz.load_balancer) == this_cpu) {
4131 cpumask_t cpus = nohz.cpu_mask;
4135 cpu_clear(this_cpu, cpus);
4136 for_each_cpu_mask(balance_cpu, cpus) {
4138 * If this cpu gets work to do, stop the load balancing
4139 * work being done for other cpus. Next load
4140 * balancing owner will pick it up.
4145 rebalance_domains(balance_cpu, CPU_IDLE);
4147 rq = cpu_rq(balance_cpu);
4148 if (time_after(this_rq->next_balance, rq->next_balance))
4149 this_rq->next_balance = rq->next_balance;
4156 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4158 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4159 * idle load balancing owner or decide to stop the periodic load balancing,
4160 * if the whole system is idle.
4162 static inline void trigger_load_balance(struct rq *rq, int cpu)
4166 * If we were in the nohz mode recently and busy at the current
4167 * scheduler tick, then check if we need to nominate new idle
4170 if (rq->in_nohz_recently && !rq->idle_at_tick) {
4171 rq->in_nohz_recently = 0;
4173 if (atomic_read(&nohz.load_balancer) == cpu) {
4174 cpu_clear(cpu, nohz.cpu_mask);
4175 atomic_set(&nohz.load_balancer, -1);
4178 if (atomic_read(&nohz.load_balancer) == -1) {
4180 * simple selection for now: Nominate the
4181 * first cpu in the nohz list to be the next
4184 * TBD: Traverse the sched domains and nominate
4185 * the nearest cpu in the nohz.cpu_mask.
4187 int ilb = first_cpu(nohz.cpu_mask);
4189 if (ilb < nr_cpu_ids)
4195 * If this cpu is idle and doing idle load balancing for all the
4196 * cpus with ticks stopped, is it time for that to stop?
4198 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) == cpu &&
4199 cpus_weight(nohz.cpu_mask) == num_online_cpus()) {
4205 * If this cpu is idle and the idle load balancing is done by
4206 * someone else, then no need raise the SCHED_SOFTIRQ
4208 if (rq->idle_at_tick && atomic_read(&nohz.load_balancer) != cpu &&
4209 cpu_isset(cpu, nohz.cpu_mask))
4212 if (time_after_eq(jiffies, rq->next_balance))
4213 raise_softirq(SCHED_SOFTIRQ);
4216 #else /* CONFIG_SMP */
4219 * on UP we do not need to balance between CPUs:
4221 static inline void idle_balance(int cpu, struct rq *rq)
4227 DEFINE_PER_CPU(struct kernel_stat, kstat);
4229 EXPORT_PER_CPU_SYMBOL(kstat);
4232 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4233 * that have not yet been banked in case the task is currently running.
4235 unsigned long long task_sched_runtime(struct task_struct *p)
4237 unsigned long flags;
4241 rq = task_rq_lock(p, &flags);
4242 ns = p->se.sum_exec_runtime;
4243 if (task_current(rq, p)) {
4244 update_rq_clock(rq);
4245 delta_exec = rq->clock - p->se.exec_start;
4246 if ((s64)delta_exec > 0)
4249 task_rq_unlock(rq, &flags);
4255 * Account user cpu time to a process.
4256 * @p: the process that the cpu time gets accounted to
4257 * @cputime: the cpu time spent in user space since the last update
4259 void account_user_time(struct task_struct *p, cputime_t cputime)
4261 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4264 p->utime = cputime_add(p->utime, cputime);
4266 /* Add user time to cpustat. */
4267 tmp = cputime_to_cputime64(cputime);
4268 if (TASK_NICE(p) > 0)
4269 cpustat->nice = cputime64_add(cpustat->nice, tmp);
4271 cpustat->user = cputime64_add(cpustat->user, tmp);
4275 * Account guest cpu time to a process.
4276 * @p: the process that the cpu time gets accounted to
4277 * @cputime: the cpu time spent in virtual machine since the last update
4279 static void account_guest_time(struct task_struct *p, cputime_t cputime)
4282 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4284 tmp = cputime_to_cputime64(cputime);
4286 p->utime = cputime_add(p->utime, cputime);
4287 p->gtime = cputime_add(p->gtime, cputime);
4289 cpustat->user = cputime64_add(cpustat->user, tmp);
4290 cpustat->guest = cputime64_add(cpustat->guest, tmp);
4294 * Account scaled user cpu time to a process.
4295 * @p: the process that the cpu time gets accounted to
4296 * @cputime: the cpu time spent in user space since the last update
4298 void account_user_time_scaled(struct task_struct *p, cputime_t cputime)
4300 p->utimescaled = cputime_add(p->utimescaled, cputime);
4304 * Account system cpu time to a process.
4305 * @p: the process that the cpu time gets accounted to
4306 * @hardirq_offset: the offset to subtract from hardirq_count()
4307 * @cputime: the cpu time spent in kernel space since the last update
4309 void account_system_time(struct task_struct *p, int hardirq_offset,
4312 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4313 struct rq *rq = this_rq();
4316 if ((p->flags & PF_VCPU) && (irq_count() - hardirq_offset == 0)) {
4317 account_guest_time(p, cputime);
4321 p->stime = cputime_add(p->stime, cputime);
4323 /* Add system time to cpustat. */
4324 tmp = cputime_to_cputime64(cputime);
4325 if (hardirq_count() - hardirq_offset)
4326 cpustat->irq = cputime64_add(cpustat->irq, tmp);
4327 else if (softirq_count())
4328 cpustat->softirq = cputime64_add(cpustat->softirq, tmp);
4329 else if (p != rq->idle)
4330 cpustat->system = cputime64_add(cpustat->system, tmp);
4331 else if (atomic_read(&rq->nr_iowait) > 0)
4332 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4334 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4335 /* Account for system time used */
4336 acct_update_integrals(p);
4340 * Account scaled system cpu time to a process.
4341 * @p: the process that the cpu time gets accounted to
4342 * @hardirq_offset: the offset to subtract from hardirq_count()
4343 * @cputime: the cpu time spent in kernel space since the last update
4345 void account_system_time_scaled(struct task_struct *p, cputime_t cputime)
4347 p->stimescaled = cputime_add(p->stimescaled, cputime);
4351 * Account for involuntary wait time.
4352 * @p: the process from which the cpu time has been stolen
4353 * @steal: the cpu time spent in involuntary wait
4355 void account_steal_time(struct task_struct *p, cputime_t steal)
4357 struct cpu_usage_stat *cpustat = &kstat_this_cpu.cpustat;
4358 cputime64_t tmp = cputime_to_cputime64(steal);
4359 struct rq *rq = this_rq();
4361 if (p == rq->idle) {
4362 p->stime = cputime_add(p->stime, steal);
4363 if (atomic_read(&rq->nr_iowait) > 0)
4364 cpustat->iowait = cputime64_add(cpustat->iowait, tmp);
4366 cpustat->idle = cputime64_add(cpustat->idle, tmp);
4368 cpustat->steal = cputime64_add(cpustat->steal, tmp);
4372 * This function gets called by the timer code, with HZ frequency.
4373 * We call it with interrupts disabled.
4375 * It also gets called by the fork code, when changing the parent's
4378 void scheduler_tick(void)
4380 int cpu = smp_processor_id();
4381 struct rq *rq = cpu_rq(cpu);
4382 struct task_struct *curr = rq->curr;
4386 spin_lock(&rq->lock);
4387 update_rq_clock(rq);
4388 update_cpu_load(rq);
4389 curr->sched_class->task_tick(rq, curr, 0);
4390 spin_unlock(&rq->lock);
4393 rq->idle_at_tick = idle_cpu(cpu);
4394 trigger_load_balance(rq, cpu);
4398 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4399 defined(CONFIG_PREEMPT_TRACER))
4401 static inline unsigned long get_parent_ip(unsigned long addr)
4403 if (in_lock_functions(addr)) {
4404 addr = CALLER_ADDR2;
4405 if (in_lock_functions(addr))
4406 addr = CALLER_ADDR3;
4411 void __kprobes add_preempt_count(int val)
4413 #ifdef CONFIG_DEBUG_PREEMPT
4417 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4420 preempt_count() += val;
4421 #ifdef CONFIG_DEBUG_PREEMPT
4423 * Spinlock count overflowing soon?
4425 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK) >=
4428 if (preempt_count() == val)
4429 trace_preempt_off(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4431 EXPORT_SYMBOL(add_preempt_count);
4433 void __kprobes sub_preempt_count(int val)
4435 #ifdef CONFIG_DEBUG_PREEMPT
4439 if (DEBUG_LOCKS_WARN_ON(val > preempt_count()))
4442 * Is the spinlock portion underflowing?
4444 if (DEBUG_LOCKS_WARN_ON((val < PREEMPT_MASK) &&
4445 !(preempt_count() & PREEMPT_MASK)))
4449 if (preempt_count() == val)
4450 trace_preempt_on(CALLER_ADDR0, get_parent_ip(CALLER_ADDR1));
4451 preempt_count() -= val;
4453 EXPORT_SYMBOL(sub_preempt_count);
4458 * Print scheduling while atomic bug:
4460 static noinline void __schedule_bug(struct task_struct *prev)
4462 struct pt_regs *regs = get_irq_regs();
4464 printk(KERN_ERR "BUG: scheduling while atomic: %s/%d/0x%08x\n",
4465 prev->comm, prev->pid, preempt_count());
4467 debug_show_held_locks(prev);
4468 if (irqs_disabled())
4469 print_irqtrace_events(prev);
4478 * Various schedule()-time debugging checks and statistics:
4480 static inline void schedule_debug(struct task_struct *prev)
4483 * Test if we are atomic. Since do_exit() needs to call into
4484 * schedule() atomically, we ignore that path for now.
4485 * Otherwise, whine if we are scheduling when we should not be.
4487 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev->exit_state))
4488 __schedule_bug(prev);
4490 profile_hit(SCHED_PROFILING, __builtin_return_address(0));
4492 schedstat_inc(this_rq(), sched_count);
4493 #ifdef CONFIG_SCHEDSTATS
4494 if (unlikely(prev->lock_depth >= 0)) {
4495 schedstat_inc(this_rq(), bkl_count);
4496 schedstat_inc(prev, sched_info.bkl_count);
4502 * Pick up the highest-prio task:
4504 static inline struct task_struct *
4505 pick_next_task(struct rq *rq, struct task_struct *prev)
4507 const struct sched_class *class;
4508 struct task_struct *p;
4511 * Optimization: we know that if all tasks are in
4512 * the fair class we can call that function directly:
4514 if (likely(rq->nr_running == rq->cfs.nr_running)) {
4515 p = fair_sched_class.pick_next_task(rq);
4520 class = sched_class_highest;
4522 p = class->pick_next_task(rq);
4526 * Will never be NULL as the idle class always
4527 * returns a non-NULL p:
4529 class = class->next;
4534 * schedule() is the main scheduler function.
4536 asmlinkage void __sched schedule(void)
4538 struct task_struct *prev, *next;
4539 unsigned long *switch_count;
4545 cpu = smp_processor_id();
4549 switch_count = &prev->nivcsw;
4551 release_kernel_lock(prev);
4552 need_resched_nonpreemptible:
4554 schedule_debug(prev);
4559 * Do the rq-clock update outside the rq lock:
4561 local_irq_disable();
4562 update_rq_clock(rq);
4563 spin_lock(&rq->lock);
4564 clear_tsk_need_resched(prev);
4566 if (prev->state && !(preempt_count() & PREEMPT_ACTIVE)) {
4567 if (unlikely((prev->state & TASK_INTERRUPTIBLE) &&
4568 signal_pending(prev))) {
4569 prev->state = TASK_RUNNING;
4571 deactivate_task(rq, prev, 1);
4573 switch_count = &prev->nvcsw;
4577 if (prev->sched_class->pre_schedule)
4578 prev->sched_class->pre_schedule(rq, prev);
4581 if (unlikely(!rq->nr_running))
4582 idle_balance(cpu, rq);
4584 prev->sched_class->put_prev_task(rq, prev);
4585 next = pick_next_task(rq, prev);
4587 if (likely(prev != next)) {
4588 sched_info_switch(prev, next);
4594 context_switch(rq, prev, next); /* unlocks the rq */
4596 * the context switch might have flipped the stack from under
4597 * us, hence refresh the local variables.
4599 cpu = smp_processor_id();
4602 spin_unlock_irq(&rq->lock);
4606 if (unlikely(reacquire_kernel_lock(current) < 0))
4607 goto need_resched_nonpreemptible;
4609 preempt_enable_no_resched();
4610 if (unlikely(test_thread_flag(TIF_NEED_RESCHED)))
4613 EXPORT_SYMBOL(schedule);
4615 #ifdef CONFIG_PREEMPT
4617 * this is the entry point to schedule() from in-kernel preemption
4618 * off of preempt_enable. Kernel preemptions off return from interrupt
4619 * occur there and call schedule directly.
4621 asmlinkage void __sched preempt_schedule(void)
4623 struct thread_info *ti = current_thread_info();
4626 * If there is a non-zero preempt_count or interrupts are disabled,
4627 * we do not want to preempt the current task. Just return..
4629 if (likely(ti->preempt_count || irqs_disabled()))
4633 add_preempt_count(PREEMPT_ACTIVE);
4635 sub_preempt_count(PREEMPT_ACTIVE);
4638 * Check again in case we missed a preemption opportunity
4639 * between schedule and now.
4642 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4644 EXPORT_SYMBOL(preempt_schedule);
4647 * this is the entry point to schedule() from kernel preemption
4648 * off of irq context.
4649 * Note, that this is called and return with irqs disabled. This will
4650 * protect us against recursive calling from irq.
4652 asmlinkage void __sched preempt_schedule_irq(void)
4654 struct thread_info *ti = current_thread_info();
4656 /* Catch callers which need to be fixed */
4657 BUG_ON(ti->preempt_count || !irqs_disabled());
4660 add_preempt_count(PREEMPT_ACTIVE);
4663 local_irq_disable();
4664 sub_preempt_count(PREEMPT_ACTIVE);
4667 * Check again in case we missed a preemption opportunity
4668 * between schedule and now.
4671 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED)));
4674 #endif /* CONFIG_PREEMPT */
4676 int default_wake_function(wait_queue_t *curr, unsigned mode, int sync,
4679 return try_to_wake_up(curr->private, mode, sync);
4681 EXPORT_SYMBOL(default_wake_function);
4684 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4685 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4686 * number) then we wake all the non-exclusive tasks and one exclusive task.
4688 * There are circumstances in which we can try to wake a task which has already
4689 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4690 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4692 static void __wake_up_common(wait_queue_head_t *q, unsigned int mode,
4693 int nr_exclusive, int sync, void *key)
4695 wait_queue_t *curr, *next;
4697 list_for_each_entry_safe(curr, next, &q->task_list, task_list) {
4698 unsigned flags = curr->flags;
4700 if (curr->func(curr, mode, sync, key) &&
4701 (flags & WQ_FLAG_EXCLUSIVE) && !--nr_exclusive)
4707 * __wake_up - wake up threads blocked on a waitqueue.
4709 * @mode: which threads
4710 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4711 * @key: is directly passed to the wakeup function
4713 void __wake_up(wait_queue_head_t *q, unsigned int mode,
4714 int nr_exclusive, void *key)
4716 unsigned long flags;
4718 spin_lock_irqsave(&q->lock, flags);
4719 __wake_up_common(q, mode, nr_exclusive, 0, key);
4720 spin_unlock_irqrestore(&q->lock, flags);
4722 EXPORT_SYMBOL(__wake_up);
4725 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4727 void __wake_up_locked(wait_queue_head_t *q, unsigned int mode)
4729 __wake_up_common(q, mode, 1, 0, NULL);
4733 * __wake_up_sync - wake up threads blocked on a waitqueue.
4735 * @mode: which threads
4736 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4738 * The sync wakeup differs that the waker knows that it will schedule
4739 * away soon, so while the target thread will be woken up, it will not
4740 * be migrated to another CPU - ie. the two threads are 'synchronized'
4741 * with each other. This can prevent needless bouncing between CPUs.
4743 * On UP it can prevent extra preemption.
4746 __wake_up_sync(wait_queue_head_t *q, unsigned int mode, int nr_exclusive)
4748 unsigned long flags;
4754 if (unlikely(!nr_exclusive))
4757 spin_lock_irqsave(&q->lock, flags);
4758 __wake_up_common(q, mode, nr_exclusive, sync, NULL);
4759 spin_unlock_irqrestore(&q->lock, flags);
4761 EXPORT_SYMBOL_GPL(__wake_up_sync); /* For internal use only */
4763 void complete(struct completion *x)
4765 unsigned long flags;
4767 spin_lock_irqsave(&x->wait.lock, flags);
4769 __wake_up_common(&x->wait, TASK_NORMAL, 1, 0, NULL);
4770 spin_unlock_irqrestore(&x->wait.lock, flags);
4772 EXPORT_SYMBOL(complete);
4774 void complete_all(struct completion *x)
4776 unsigned long flags;
4778 spin_lock_irqsave(&x->wait.lock, flags);
4779 x->done += UINT_MAX/2;
4780 __wake_up_common(&x->wait, TASK_NORMAL, 0, 0, NULL);
4781 spin_unlock_irqrestore(&x->wait.lock, flags);
4783 EXPORT_SYMBOL(complete_all);
4785 static inline long __sched
4786 do_wait_for_common(struct completion *x, long timeout, int state)
4789 DECLARE_WAITQUEUE(wait, current);
4791 wait.flags |= WQ_FLAG_EXCLUSIVE;
4792 __add_wait_queue_tail(&x->wait, &wait);
4794 if ((state == TASK_INTERRUPTIBLE &&
4795 signal_pending(current)) ||
4796 (state == TASK_KILLABLE &&
4797 fatal_signal_pending(current))) {
4798 __remove_wait_queue(&x->wait, &wait);
4799 return -ERESTARTSYS;
4801 __set_current_state(state);
4802 spin_unlock_irq(&x->wait.lock);
4803 timeout = schedule_timeout(timeout);
4804 spin_lock_irq(&x->wait.lock);
4806 __remove_wait_queue(&x->wait, &wait);
4810 __remove_wait_queue(&x->wait, &wait);
4817 wait_for_common(struct completion *x, long timeout, int state)
4821 spin_lock_irq(&x->wait.lock);
4822 timeout = do_wait_for_common(x, timeout, state);
4823 spin_unlock_irq(&x->wait.lock);
4827 void __sched wait_for_completion(struct completion *x)
4829 wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_UNINTERRUPTIBLE);
4831 EXPORT_SYMBOL(wait_for_completion);
4833 unsigned long __sched
4834 wait_for_completion_timeout(struct completion *x, unsigned long timeout)
4836 return wait_for_common(x, timeout, TASK_UNINTERRUPTIBLE);
4838 EXPORT_SYMBOL(wait_for_completion_timeout);
4840 int __sched wait_for_completion_interruptible(struct completion *x)
4842 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_INTERRUPTIBLE);
4843 if (t == -ERESTARTSYS)
4847 EXPORT_SYMBOL(wait_for_completion_interruptible);
4849 unsigned long __sched
4850 wait_for_completion_interruptible_timeout(struct completion *x,
4851 unsigned long timeout)
4853 return wait_for_common(x, timeout, TASK_INTERRUPTIBLE);
4855 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout);
4857 int __sched wait_for_completion_killable(struct completion *x)
4859 long t = wait_for_common(x, MAX_SCHEDULE_TIMEOUT, TASK_KILLABLE);
4860 if (t == -ERESTARTSYS)
4864 EXPORT_SYMBOL(wait_for_completion_killable);
4867 sleep_on_common(wait_queue_head_t *q, int state, long timeout)
4869 unsigned long flags;
4872 init_waitqueue_entry(&wait, current);
4874 __set_current_state(state);
4876 spin_lock_irqsave(&q->lock, flags);
4877 __add_wait_queue(q, &wait);
4878 spin_unlock(&q->lock);
4879 timeout = schedule_timeout(timeout);
4880 spin_lock_irq(&q->lock);
4881 __remove_wait_queue(q, &wait);
4882 spin_unlock_irqrestore(&q->lock, flags);
4887 void __sched interruptible_sleep_on(wait_queue_head_t *q)
4889 sleep_on_common(q, TASK_INTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4891 EXPORT_SYMBOL(interruptible_sleep_on);
4894 interruptible_sleep_on_timeout(wait_queue_head_t *q, long timeout)
4896 return sleep_on_common(q, TASK_INTERRUPTIBLE, timeout);
4898 EXPORT_SYMBOL(interruptible_sleep_on_timeout);
4900 void __sched sleep_on(wait_queue_head_t *q)
4902 sleep_on_common(q, TASK_UNINTERRUPTIBLE, MAX_SCHEDULE_TIMEOUT);
4904 EXPORT_SYMBOL(sleep_on);
4906 long __sched sleep_on_timeout(wait_queue_head_t *q, long timeout)
4908 return sleep_on_common(q, TASK_UNINTERRUPTIBLE, timeout);
4910 EXPORT_SYMBOL(sleep_on_timeout);
4912 #ifdef CONFIG_RT_MUTEXES
4915 * rt_mutex_setprio - set the current priority of a task
4917 * @prio: prio value (kernel-internal form)
4919 * This function changes the 'effective' priority of a task. It does
4920 * not touch ->normal_prio like __setscheduler().
4922 * Used by the rt_mutex code to implement priority inheritance logic.
4924 void rt_mutex_setprio(struct task_struct *p, int prio)
4926 unsigned long flags;
4927 int oldprio, on_rq, running;
4929 const struct sched_class *prev_class = p->sched_class;
4931 BUG_ON(prio < 0 || prio > MAX_PRIO);
4933 rq = task_rq_lock(p, &flags);
4934 update_rq_clock(rq);
4937 on_rq = p->se.on_rq;
4938 running = task_current(rq, p);
4940 dequeue_task(rq, p, 0);
4942 p->sched_class->put_prev_task(rq, p);
4945 p->sched_class = &rt_sched_class;
4947 p->sched_class = &fair_sched_class;
4952 p->sched_class->set_curr_task(rq);
4954 enqueue_task(rq, p, 0);
4956 check_class_changed(rq, p, prev_class, oldprio, running);
4958 task_rq_unlock(rq, &flags);
4963 void set_user_nice(struct task_struct *p, long nice)
4965 int old_prio, delta, on_rq;
4966 unsigned long flags;
4969 if (TASK_NICE(p) == nice || nice < -20 || nice > 19)
4972 * We have to be careful, if called from sys_setpriority(),
4973 * the task might be in the middle of scheduling on another CPU.
4975 rq = task_rq_lock(p, &flags);
4976 update_rq_clock(rq);
4978 * The RT priorities are set via sched_setscheduler(), but we still
4979 * allow the 'normal' nice value to be set - but as expected
4980 * it wont have any effect on scheduling until the task is
4981 * SCHED_FIFO/SCHED_RR:
4983 if (task_has_rt_policy(p)) {
4984 p->static_prio = NICE_TO_PRIO(nice);
4987 on_rq = p->se.on_rq;
4989 dequeue_task(rq, p, 0);
4991 p->static_prio = NICE_TO_PRIO(nice);
4994 p->prio = effective_prio(p);
4995 delta = p->prio - old_prio;
4998 enqueue_task(rq, p, 0);
5000 * If the task increased its priority or is running and
5001 * lowered its priority, then reschedule its CPU:
5003 if (delta < 0 || (delta > 0 && task_running(rq, p)))
5004 resched_task(rq->curr);
5007 task_rq_unlock(rq, &flags);
5009 EXPORT_SYMBOL(set_user_nice);
5012 * can_nice - check if a task can reduce its nice value
5016 int can_nice(const struct task_struct *p, const int nice)
5018 /* convert nice value [19,-20] to rlimit style value [1,40] */
5019 int nice_rlim = 20 - nice;
5021 return (nice_rlim <= p->signal->rlim[RLIMIT_NICE].rlim_cur ||
5022 capable(CAP_SYS_NICE));
5025 #ifdef __ARCH_WANT_SYS_NICE
5028 * sys_nice - change the priority of the current process.
5029 * @increment: priority increment
5031 * sys_setpriority is a more generic, but much slower function that
5032 * does similar things.
5034 asmlinkage long sys_nice(int increment)
5039 * Setpriority might change our priority at the same moment.
5040 * We don't have to worry. Conceptually one call occurs first
5041 * and we have a single winner.
5043 if (increment < -40)
5048 nice = PRIO_TO_NICE(current->static_prio) + increment;
5054 if (increment < 0 && !can_nice(current, nice))
5057 retval = security_task_setnice(current, nice);
5061 set_user_nice(current, nice);
5068 * task_prio - return the priority value of a given task.
5069 * @p: the task in question.
5071 * This is the priority value as seen by users in /proc.
5072 * RT tasks are offset by -200. Normal tasks are centered
5073 * around 0, value goes from -16 to +15.
5075 int task_prio(const struct task_struct *p)
5077 return p->prio - MAX_RT_PRIO;
5081 * task_nice - return the nice value of a given task.
5082 * @p: the task in question.
5084 int task_nice(const struct task_struct *p)
5086 return TASK_NICE(p);
5088 EXPORT_SYMBOL(task_nice);
5091 * idle_cpu - is a given cpu idle currently?
5092 * @cpu: the processor in question.
5094 int idle_cpu(int cpu)
5096 return cpu_curr(cpu) == cpu_rq(cpu)->idle;
5100 * idle_task - return the idle task for a given cpu.
5101 * @cpu: the processor in question.
5103 struct task_struct *idle_task(int cpu)
5105 return cpu_rq(cpu)->idle;
5109 * find_process_by_pid - find a process with a matching PID value.
5110 * @pid: the pid in question.
5112 static struct task_struct *find_process_by_pid(pid_t pid)
5114 return pid ? find_task_by_vpid(pid) : current;
5117 /* Actually do priority change: must hold rq lock. */
5119 __setscheduler(struct rq *rq, struct task_struct *p, int policy, int prio)
5121 BUG_ON(p->se.on_rq);
5124 switch (p->policy) {
5128 p->sched_class = &fair_sched_class;
5132 p->sched_class = &rt_sched_class;
5136 p->rt_priority = prio;
5137 p->normal_prio = normal_prio(p);
5138 /* we are holding p->pi_lock already */
5139 p->prio = rt_mutex_getprio(p);
5144 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5145 * @p: the task in question.
5146 * @policy: new policy.
5147 * @param: structure containing the new RT priority.
5149 * NOTE that the task may be already dead.
5151 int sched_setscheduler(struct task_struct *p, int policy,
5152 struct sched_param *param)
5154 int retval, oldprio, oldpolicy = -1, on_rq, running;
5155 unsigned long flags;
5156 const struct sched_class *prev_class = p->sched_class;
5159 /* may grab non-irq protected spin_locks */
5160 BUG_ON(in_interrupt());
5162 /* double check policy once rq lock held */
5164 policy = oldpolicy = p->policy;
5165 else if (policy != SCHED_FIFO && policy != SCHED_RR &&
5166 policy != SCHED_NORMAL && policy != SCHED_BATCH &&
5167 policy != SCHED_IDLE)
5170 * Valid priorities for SCHED_FIFO and SCHED_RR are
5171 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5172 * SCHED_BATCH and SCHED_IDLE is 0.
5174 if (param->sched_priority < 0 ||
5175 (p->mm && param->sched_priority > MAX_USER_RT_PRIO-1) ||
5176 (!p->mm && param->sched_priority > MAX_RT_PRIO-1))
5178 if (rt_policy(policy) != (param->sched_priority != 0))
5182 * Allow unprivileged RT tasks to decrease priority:
5184 if (!capable(CAP_SYS_NICE)) {
5185 if (rt_policy(policy)) {
5186 unsigned long rlim_rtprio;
5188 if (!lock_task_sighand(p, &flags))
5190 rlim_rtprio = p->signal->rlim[RLIMIT_RTPRIO].rlim_cur;
5191 unlock_task_sighand(p, &flags);
5193 /* can't set/change the rt policy */
5194 if (policy != p->policy && !rlim_rtprio)
5197 /* can't increase priority */
5198 if (param->sched_priority > p->rt_priority &&
5199 param->sched_priority > rlim_rtprio)
5203 * Like positive nice levels, dont allow tasks to
5204 * move out of SCHED_IDLE either:
5206 if (p->policy == SCHED_IDLE && policy != SCHED_IDLE)
5209 /* can't change other user's priorities */
5210 if ((current->euid != p->euid) &&
5211 (current->euid != p->uid))
5215 #ifdef CONFIG_RT_GROUP_SCHED
5217 * Do not allow realtime tasks into groups that have no runtime
5220 if (rt_policy(policy) && task_group(p)->rt_bandwidth.rt_runtime == 0)
5224 retval = security_task_setscheduler(p, policy, param);
5228 * make sure no PI-waiters arrive (or leave) while we are
5229 * changing the priority of the task:
5231 spin_lock_irqsave(&p->pi_lock, flags);
5233 * To be able to change p->policy safely, the apropriate
5234 * runqueue lock must be held.
5236 rq = __task_rq_lock(p);
5237 /* recheck policy now with rq lock held */
5238 if (unlikely(oldpolicy != -1 && oldpolicy != p->policy)) {
5239 policy = oldpolicy = -1;
5240 __task_rq_unlock(rq);
5241 spin_unlock_irqrestore(&p->pi_lock, flags);
5244 update_rq_clock(rq);
5245 on_rq = p->se.on_rq;
5246 running = task_current(rq, p);
5248 deactivate_task(rq, p, 0);
5250 p->sched_class->put_prev_task(rq, p);
5253 __setscheduler(rq, p, policy, param->sched_priority);
5256 p->sched_class->set_curr_task(rq);
5258 activate_task(rq, p, 0);
5260 check_class_changed(rq, p, prev_class, oldprio, running);
5262 __task_rq_unlock(rq);
5263 spin_unlock_irqrestore(&p->pi_lock, flags);
5265 rt_mutex_adjust_pi(p);
5269 EXPORT_SYMBOL_GPL(sched_setscheduler);
5272 do_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5274 struct sched_param lparam;
5275 struct task_struct *p;
5278 if (!param || pid < 0)
5280 if (copy_from_user(&lparam, param, sizeof(struct sched_param)))
5285 p = find_process_by_pid(pid);
5287 retval = sched_setscheduler(p, policy, &lparam);
5294 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5295 * @pid: the pid in question.
5296 * @policy: new policy.
5297 * @param: structure containing the new RT priority.
5300 sys_sched_setscheduler(pid_t pid, int policy, struct sched_param __user *param)
5302 /* negative values for policy are not valid */
5306 return do_sched_setscheduler(pid, policy, param);
5310 * sys_sched_setparam - set/change the RT priority of a thread
5311 * @pid: the pid in question.
5312 * @param: structure containing the new RT priority.
5314 asmlinkage long sys_sched_setparam(pid_t pid, struct sched_param __user *param)
5316 return do_sched_setscheduler(pid, -1, param);
5320 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5321 * @pid: the pid in question.
5323 asmlinkage long sys_sched_getscheduler(pid_t pid)
5325 struct task_struct *p;
5332 read_lock(&tasklist_lock);
5333 p = find_process_by_pid(pid);
5335 retval = security_task_getscheduler(p);
5339 read_unlock(&tasklist_lock);
5344 * sys_sched_getscheduler - get the RT priority of a thread
5345 * @pid: the pid in question.
5346 * @param: structure containing the RT priority.
5348 asmlinkage long sys_sched_getparam(pid_t pid, struct sched_param __user *param)
5350 struct sched_param lp;
5351 struct task_struct *p;
5354 if (!param || pid < 0)
5357 read_lock(&tasklist_lock);
5358 p = find_process_by_pid(pid);
5363 retval = security_task_getscheduler(p);
5367 lp.sched_priority = p->rt_priority;
5368 read_unlock(&tasklist_lock);
5371 * This one might sleep, we cannot do it with a spinlock held ...
5373 retval = copy_to_user(param, &lp, sizeof(*param)) ? -EFAULT : 0;
5378 read_unlock(&tasklist_lock);
5382 long sched_setaffinity(pid_t pid, const cpumask_t *in_mask)
5384 cpumask_t cpus_allowed;
5385 cpumask_t new_mask = *in_mask;
5386 struct task_struct *p;
5390 read_lock(&tasklist_lock);
5392 p = find_process_by_pid(pid);
5394 read_unlock(&tasklist_lock);
5400 * It is not safe to call set_cpus_allowed with the
5401 * tasklist_lock held. We will bump the task_struct's
5402 * usage count and then drop tasklist_lock.
5405 read_unlock(&tasklist_lock);
5408 if ((current->euid != p->euid) && (current->euid != p->uid) &&
5409 !capable(CAP_SYS_NICE))
5412 retval = security_task_setscheduler(p, 0, NULL);
5416 cpuset_cpus_allowed(p, &cpus_allowed);
5417 cpus_and(new_mask, new_mask, cpus_allowed);
5419 retval = set_cpus_allowed_ptr(p, &new_mask);
5422 cpuset_cpus_allowed(p, &cpus_allowed);
5423 if (!cpus_subset(new_mask, cpus_allowed)) {
5425 * We must have raced with a concurrent cpuset
5426 * update. Just reset the cpus_allowed to the
5427 * cpuset's cpus_allowed
5429 new_mask = cpus_allowed;
5439 static int get_user_cpu_mask(unsigned long __user *user_mask_ptr, unsigned len,
5440 cpumask_t *new_mask)
5442 if (len < sizeof(cpumask_t)) {
5443 memset(new_mask, 0, sizeof(cpumask_t));
5444 } else if (len > sizeof(cpumask_t)) {
5445 len = sizeof(cpumask_t);
5447 return copy_from_user(new_mask, user_mask_ptr, len) ? -EFAULT : 0;
5451 * sys_sched_setaffinity - set the cpu affinity of a process
5452 * @pid: pid of the process
5453 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5454 * @user_mask_ptr: user-space pointer to the new cpu mask
5456 asmlinkage long sys_sched_setaffinity(pid_t pid, unsigned int len,
5457 unsigned long __user *user_mask_ptr)
5462 retval = get_user_cpu_mask(user_mask_ptr, len, &new_mask);
5466 return sched_setaffinity(pid, &new_mask);
5470 * Represents all cpu's present in the system
5471 * In systems capable of hotplug, this map could dynamically grow
5472 * as new cpu's are detected in the system via any platform specific
5473 * method, such as ACPI for e.g.
5476 cpumask_t cpu_present_map __read_mostly;
5477 EXPORT_SYMBOL(cpu_present_map);
5480 cpumask_t cpu_online_map __read_mostly = CPU_MASK_ALL;
5481 EXPORT_SYMBOL(cpu_online_map);
5483 cpumask_t cpu_possible_map __read_mostly = CPU_MASK_ALL;
5484 EXPORT_SYMBOL(cpu_possible_map);
5487 long sched_getaffinity(pid_t pid, cpumask_t *mask)
5489 struct task_struct *p;
5493 read_lock(&tasklist_lock);
5496 p = find_process_by_pid(pid);
5500 retval = security_task_getscheduler(p);
5504 cpus_and(*mask, p->cpus_allowed, cpu_online_map);
5507 read_unlock(&tasklist_lock);
5514 * sys_sched_getaffinity - get the cpu affinity of a process
5515 * @pid: pid of the process
5516 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5517 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5519 asmlinkage long sys_sched_getaffinity(pid_t pid, unsigned int len,
5520 unsigned long __user *user_mask_ptr)
5525 if (len < sizeof(cpumask_t))
5528 ret = sched_getaffinity(pid, &mask);
5532 if (copy_to_user(user_mask_ptr, &mask, sizeof(cpumask_t)))
5535 return sizeof(cpumask_t);
5539 * sys_sched_yield - yield the current processor to other threads.
5541 * This function yields the current CPU to other tasks. If there are no
5542 * other threads running on this CPU then this function will return.
5544 asmlinkage long sys_sched_yield(void)
5546 struct rq *rq = this_rq_lock();
5548 schedstat_inc(rq, yld_count);
5549 current->sched_class->yield_task(rq);
5552 * Since we are going to call schedule() anyway, there's
5553 * no need to preempt or enable interrupts:
5555 __release(rq->lock);
5556 spin_release(&rq->lock.dep_map, 1, _THIS_IP_);
5557 _raw_spin_unlock(&rq->lock);
5558 preempt_enable_no_resched();
5565 static void __cond_resched(void)
5567 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5568 __might_sleep(__FILE__, __LINE__);
5571 * The BKS might be reacquired before we have dropped
5572 * PREEMPT_ACTIVE, which could trigger a second
5573 * cond_resched() call.
5576 add_preempt_count(PREEMPT_ACTIVE);
5578 sub_preempt_count(PREEMPT_ACTIVE);
5579 } while (need_resched());
5582 int __sched _cond_resched(void)
5584 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE) &&
5585 system_state == SYSTEM_RUNNING) {
5591 EXPORT_SYMBOL(_cond_resched);
5594 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5595 * call schedule, and on return reacquire the lock.
5597 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5598 * operations here to prevent schedule() from being called twice (once via
5599 * spin_unlock(), once by hand).
5601 int cond_resched_lock(spinlock_t *lock)
5603 int resched = need_resched() && system_state == SYSTEM_RUNNING;
5606 if (spin_needbreak(lock) || resched) {
5608 if (resched && need_resched())
5617 EXPORT_SYMBOL(cond_resched_lock);
5619 int __sched cond_resched_softirq(void)
5621 BUG_ON(!in_softirq());
5623 if (need_resched() && system_state == SYSTEM_RUNNING) {
5631 EXPORT_SYMBOL(cond_resched_softirq);
5634 * yield - yield the current processor to other threads.
5636 * This is a shortcut for kernel-space yielding - it marks the
5637 * thread runnable and calls sys_sched_yield().
5639 void __sched yield(void)
5641 set_current_state(TASK_RUNNING);
5644 EXPORT_SYMBOL(yield);
5647 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5648 * that process accounting knows that this is a task in IO wait state.
5650 * But don't do that if it is a deliberate, throttling IO wait (this task
5651 * has set its backing_dev_info: the queue against which it should throttle)
5653 void __sched io_schedule(void)
5655 struct rq *rq = &__raw_get_cpu_var(runqueues);
5657 delayacct_blkio_start();
5658 atomic_inc(&rq->nr_iowait);
5660 atomic_dec(&rq->nr_iowait);
5661 delayacct_blkio_end();
5663 EXPORT_SYMBOL(io_schedule);
5665 long __sched io_schedule_timeout(long timeout)
5667 struct rq *rq = &__raw_get_cpu_var(runqueues);
5670 delayacct_blkio_start();
5671 atomic_inc(&rq->nr_iowait);
5672 ret = schedule_timeout(timeout);
5673 atomic_dec(&rq->nr_iowait);
5674 delayacct_blkio_end();
5679 * sys_sched_get_priority_max - return maximum RT priority.
5680 * @policy: scheduling class.
5682 * this syscall returns the maximum rt_priority that can be used
5683 * by a given scheduling class.
5685 asmlinkage long sys_sched_get_priority_max(int policy)
5692 ret = MAX_USER_RT_PRIO-1;
5704 * sys_sched_get_priority_min - return minimum RT priority.
5705 * @policy: scheduling class.
5707 * this syscall returns the minimum rt_priority that can be used
5708 * by a given scheduling class.
5710 asmlinkage long sys_sched_get_priority_min(int policy)
5728 * sys_sched_rr_get_interval - return the default timeslice of a process.
5729 * @pid: pid of the process.
5730 * @interval: userspace pointer to the timeslice value.
5732 * this syscall writes the default timeslice value of a given process
5733 * into the user-space timespec buffer. A value of '0' means infinity.
5736 long sys_sched_rr_get_interval(pid_t pid, struct timespec __user *interval)
5738 struct task_struct *p;
5739 unsigned int time_slice;
5747 read_lock(&tasklist_lock);
5748 p = find_process_by_pid(pid);
5752 retval = security_task_getscheduler(p);
5757 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5758 * tasks that are on an otherwise idle runqueue:
5761 if (p->policy == SCHED_RR) {
5762 time_slice = DEF_TIMESLICE;
5763 } else if (p->policy != SCHED_FIFO) {
5764 struct sched_entity *se = &p->se;
5765 unsigned long flags;
5768 rq = task_rq_lock(p, &flags);
5769 if (rq->cfs.load.weight)
5770 time_slice = NS_TO_JIFFIES(sched_slice(&rq->cfs, se));
5771 task_rq_unlock(rq, &flags);
5773 read_unlock(&tasklist_lock);
5774 jiffies_to_timespec(time_slice, &t);
5775 retval = copy_to_user(interval, &t, sizeof(t)) ? -EFAULT : 0;
5779 read_unlock(&tasklist_lock);
5783 static const char stat_nam[] = TASK_STATE_TO_CHAR_STR;
5785 void sched_show_task(struct task_struct *p)
5787 unsigned long free = 0;
5790 state = p->state ? __ffs(p->state) + 1 : 0;
5791 printk(KERN_INFO "%-13.13s %c", p->comm,
5792 state < sizeof(stat_nam) - 1 ? stat_nam[state] : '?');
5793 #if BITS_PER_LONG == 32
5794 if (state == TASK_RUNNING)
5795 printk(KERN_CONT " running ");
5797 printk(KERN_CONT " %08lx ", thread_saved_pc(p));
5799 if (state == TASK_RUNNING)
5800 printk(KERN_CONT " running task ");
5802 printk(KERN_CONT " %016lx ", thread_saved_pc(p));
5804 #ifdef CONFIG_DEBUG_STACK_USAGE
5806 unsigned long *n = end_of_stack(p);
5809 free = (unsigned long)n - (unsigned long)end_of_stack(p);
5812 printk(KERN_CONT "%5lu %5d %6d\n", free,
5813 task_pid_nr(p), task_pid_nr(p->real_parent));
5815 show_stack(p, NULL);
5818 void show_state_filter(unsigned long state_filter)
5820 struct task_struct *g, *p;
5822 #if BITS_PER_LONG == 32
5824 " task PC stack pid father\n");
5827 " task PC stack pid father\n");
5829 read_lock(&tasklist_lock);
5830 do_each_thread(g, p) {
5832 * reset the NMI-timeout, listing all files on a slow
5833 * console might take alot of time:
5835 touch_nmi_watchdog();
5836 if (!state_filter || (p->state & state_filter))
5838 } while_each_thread(g, p);
5840 touch_all_softlockup_watchdogs();
5842 #ifdef CONFIG_SCHED_DEBUG
5843 sysrq_sched_debug_show();
5845 read_unlock(&tasklist_lock);
5847 * Only show locks if all tasks are dumped:
5849 if (state_filter == -1)
5850 debug_show_all_locks();
5853 void __cpuinit init_idle_bootup_task(struct task_struct *idle)
5855 idle->sched_class = &idle_sched_class;
5859 * init_idle - set up an idle thread for a given CPU
5860 * @idle: task in question
5861 * @cpu: cpu the idle task belongs to
5863 * NOTE: this function does not set the idle thread's NEED_RESCHED
5864 * flag, to make booting more robust.
5866 void __cpuinit init_idle(struct task_struct *idle, int cpu)
5868 struct rq *rq = cpu_rq(cpu);
5869 unsigned long flags;
5872 idle->se.exec_start = sched_clock();
5874 idle->prio = idle->normal_prio = MAX_PRIO;
5875 idle->cpus_allowed = cpumask_of_cpu(cpu);
5876 __set_task_cpu(idle, cpu);
5878 spin_lock_irqsave(&rq->lock, flags);
5879 rq->curr = rq->idle = idle;
5880 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5883 spin_unlock_irqrestore(&rq->lock, flags);
5885 /* Set the preempt count _outside_ the spinlocks! */
5886 #if defined(CONFIG_PREEMPT)
5887 task_thread_info(idle)->preempt_count = (idle->lock_depth >= 0);
5889 task_thread_info(idle)->preempt_count = 0;
5892 * The idle tasks have their own, simple scheduling class:
5894 idle->sched_class = &idle_sched_class;
5898 * In a system that switches off the HZ timer nohz_cpu_mask
5899 * indicates which cpus entered this state. This is used
5900 * in the rcu update to wait only for active cpus. For system
5901 * which do not switch off the HZ timer nohz_cpu_mask should
5902 * always be CPU_MASK_NONE.
5904 cpumask_t nohz_cpu_mask = CPU_MASK_NONE;
5907 * Increase the granularity value when there are more CPUs,
5908 * because with more CPUs the 'effective latency' as visible
5909 * to users decreases. But the relationship is not linear,
5910 * so pick a second-best guess by going with the log2 of the
5913 * This idea comes from the SD scheduler of Con Kolivas:
5915 static inline void sched_init_granularity(void)
5917 unsigned int factor = 1 + ilog2(num_online_cpus());
5918 const unsigned long limit = 200000000;
5920 sysctl_sched_min_granularity *= factor;
5921 if (sysctl_sched_min_granularity > limit)
5922 sysctl_sched_min_granularity = limit;
5924 sysctl_sched_latency *= factor;
5925 if (sysctl_sched_latency > limit)
5926 sysctl_sched_latency = limit;
5928 sysctl_sched_wakeup_granularity *= factor;
5933 * This is how migration works:
5935 * 1) we queue a struct migration_req structure in the source CPU's
5936 * runqueue and wake up that CPU's migration thread.
5937 * 2) we down() the locked semaphore => thread blocks.
5938 * 3) migration thread wakes up (implicitly it forces the migrated
5939 * thread off the CPU)
5940 * 4) it gets the migration request and checks whether the migrated
5941 * task is still in the wrong runqueue.
5942 * 5) if it's in the wrong runqueue then the migration thread removes
5943 * it and puts it into the right queue.
5944 * 6) migration thread up()s the semaphore.
5945 * 7) we wake up and the migration is done.
5949 * Change a given task's CPU affinity. Migrate the thread to a
5950 * proper CPU and schedule it away if the CPU it's executing on
5951 * is removed from the allowed bitmask.
5953 * NOTE: the caller must have a valid reference to the task, the
5954 * task must not exit() & deallocate itself prematurely. The
5955 * call is not atomic; no spinlocks may be held.
5957 int set_cpus_allowed_ptr(struct task_struct *p, const cpumask_t *new_mask)
5959 struct migration_req req;
5960 unsigned long flags;
5964 rq = task_rq_lock(p, &flags);
5965 if (!cpus_intersects(*new_mask, cpu_online_map)) {
5970 if (p->sched_class->set_cpus_allowed)
5971 p->sched_class->set_cpus_allowed(p, new_mask);
5973 p->cpus_allowed = *new_mask;
5974 p->rt.nr_cpus_allowed = cpus_weight(*new_mask);
5977 /* Can the task run on the task's current CPU? If so, we're done */
5978 if (cpu_isset(task_cpu(p), *new_mask))
5981 if (migrate_task(p, any_online_cpu(*new_mask), &req)) {
5982 /* Need help from migration thread: drop lock and wait. */
5983 task_rq_unlock(rq, &flags);
5984 wake_up_process(rq->migration_thread);
5985 wait_for_completion(&req.done);
5986 tlb_migrate_finish(p->mm);
5990 task_rq_unlock(rq, &flags);
5994 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr);
5997 * Move (not current) task off this cpu, onto dest cpu. We're doing
5998 * this because either it can't run here any more (set_cpus_allowed()
5999 * away from this CPU, or CPU going down), or because we're
6000 * attempting to rebalance this task on exec (sched_exec).
6002 * So we race with normal scheduler movements, but that's OK, as long
6003 * as the task is no longer on this CPU.
6005 * Returns non-zero if task was successfully migrated.
6007 static int __migrate_task(struct task_struct *p, int src_cpu, int dest_cpu)
6009 struct rq *rq_dest, *rq_src;
6012 if (unlikely(cpu_is_offline(dest_cpu)))
6015 rq_src = cpu_rq(src_cpu);
6016 rq_dest = cpu_rq(dest_cpu);
6018 double_rq_lock(rq_src, rq_dest);
6019 /* Already moved. */
6020 if (task_cpu(p) != src_cpu)
6022 /* Affinity changed (again). */
6023 if (!cpu_isset(dest_cpu, p->cpus_allowed))
6026 on_rq = p->se.on_rq;
6028 deactivate_task(rq_src, p, 0);
6030 set_task_cpu(p, dest_cpu);
6032 activate_task(rq_dest, p, 0);
6033 check_preempt_curr(rq_dest, p);
6037 double_rq_unlock(rq_src, rq_dest);
6042 * migration_thread - this is a highprio system thread that performs
6043 * thread migration by bumping thread off CPU then 'pushing' onto
6046 static int migration_thread(void *data)
6048 int cpu = (long)data;
6052 BUG_ON(rq->migration_thread != current);
6054 set_current_state(TASK_INTERRUPTIBLE);
6055 while (!kthread_should_stop()) {
6056 struct migration_req *req;
6057 struct list_head *head;
6059 spin_lock_irq(&rq->lock);
6061 if (cpu_is_offline(cpu)) {
6062 spin_unlock_irq(&rq->lock);
6066 if (rq->active_balance) {
6067 active_load_balance(rq, cpu);
6068 rq->active_balance = 0;
6071 head = &rq->migration_queue;
6073 if (list_empty(head)) {
6074 spin_unlock_irq(&rq->lock);
6076 set_current_state(TASK_INTERRUPTIBLE);
6079 req = list_entry(head->next, struct migration_req, list);
6080 list_del_init(head->next);
6082 spin_unlock(&rq->lock);
6083 __migrate_task(req->task, cpu, req->dest_cpu);
6086 complete(&req->done);
6088 __set_current_state(TASK_RUNNING);
6092 /* Wait for kthread_stop */
6093 set_current_state(TASK_INTERRUPTIBLE);
6094 while (!kthread_should_stop()) {
6096 set_current_state(TASK_INTERRUPTIBLE);
6098 __set_current_state(TASK_RUNNING);
6102 #ifdef CONFIG_HOTPLUG_CPU
6104 static int __migrate_task_irq(struct task_struct *p, int src_cpu, int dest_cpu)
6108 local_irq_disable();
6109 ret = __migrate_task(p, src_cpu, dest_cpu);
6115 * Figure out where task on dead CPU should go, use force if necessary.
6116 * NOTE: interrupts should be disabled by the caller
6118 static void move_task_off_dead_cpu(int dead_cpu, struct task_struct *p)
6120 unsigned long flags;
6127 mask = node_to_cpumask(cpu_to_node(dead_cpu));
6128 cpus_and(mask, mask, p->cpus_allowed);
6129 dest_cpu = any_online_cpu(mask);
6131 /* On any allowed CPU? */
6132 if (dest_cpu >= nr_cpu_ids)
6133 dest_cpu = any_online_cpu(p->cpus_allowed);
6135 /* No more Mr. Nice Guy. */
6136 if (dest_cpu >= nr_cpu_ids) {
6137 cpumask_t cpus_allowed;
6139 cpuset_cpus_allowed_locked(p, &cpus_allowed);
6141 * Try to stay on the same cpuset, where the
6142 * current cpuset may be a subset of all cpus.
6143 * The cpuset_cpus_allowed_locked() variant of
6144 * cpuset_cpus_allowed() will not block. It must be
6145 * called within calls to cpuset_lock/cpuset_unlock.
6147 rq = task_rq_lock(p, &flags);
6148 p->cpus_allowed = cpus_allowed;
6149 dest_cpu = any_online_cpu(p->cpus_allowed);
6150 task_rq_unlock(rq, &flags);
6153 * Don't tell them about moving exiting tasks or
6154 * kernel threads (both mm NULL), since they never
6157 if (p->mm && printk_ratelimit()) {
6158 printk(KERN_INFO "process %d (%s) no "
6159 "longer affine to cpu%d\n",
6160 task_pid_nr(p), p->comm, dead_cpu);
6163 } while (!__migrate_task_irq(p, dead_cpu, dest_cpu));
6167 * While a dead CPU has no uninterruptible tasks queued at this point,
6168 * it might still have a nonzero ->nr_uninterruptible counter, because
6169 * for performance reasons the counter is not stricly tracking tasks to
6170 * their home CPUs. So we just add the counter to another CPU's counter,
6171 * to keep the global sum constant after CPU-down:
6173 static void migrate_nr_uninterruptible(struct rq *rq_src)
6175 struct rq *rq_dest = cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR));
6176 unsigned long flags;
6178 local_irq_save(flags);
6179 double_rq_lock(rq_src, rq_dest);
6180 rq_dest->nr_uninterruptible += rq_src->nr_uninterruptible;
6181 rq_src->nr_uninterruptible = 0;
6182 double_rq_unlock(rq_src, rq_dest);
6183 local_irq_restore(flags);
6186 /* Run through task list and migrate tasks from the dead cpu. */
6187 static void migrate_live_tasks(int src_cpu)
6189 struct task_struct *p, *t;
6191 read_lock(&tasklist_lock);
6193 do_each_thread(t, p) {
6197 if (task_cpu(p) == src_cpu)
6198 move_task_off_dead_cpu(src_cpu, p);
6199 } while_each_thread(t, p);
6201 read_unlock(&tasklist_lock);
6205 * Schedules idle task to be the next runnable task on current CPU.
6206 * It does so by boosting its priority to highest possible.
6207 * Used by CPU offline code.
6209 void sched_idle_next(void)
6211 int this_cpu = smp_processor_id();
6212 struct rq *rq = cpu_rq(this_cpu);
6213 struct task_struct *p = rq->idle;
6214 unsigned long flags;
6216 /* cpu has to be offline */
6217 BUG_ON(cpu_online(this_cpu));
6220 * Strictly not necessary since rest of the CPUs are stopped by now
6221 * and interrupts disabled on the current cpu.
6223 spin_lock_irqsave(&rq->lock, flags);
6225 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6227 update_rq_clock(rq);
6228 activate_task(rq, p, 0);
6230 spin_unlock_irqrestore(&rq->lock, flags);
6234 * Ensures that the idle task is using init_mm right before its cpu goes
6237 void idle_task_exit(void)
6239 struct mm_struct *mm = current->active_mm;
6241 BUG_ON(cpu_online(smp_processor_id()));
6244 switch_mm(mm, &init_mm, current);
6248 /* called under rq->lock with disabled interrupts */
6249 static void migrate_dead(unsigned int dead_cpu, struct task_struct *p)
6251 struct rq *rq = cpu_rq(dead_cpu);
6253 /* Must be exiting, otherwise would be on tasklist. */
6254 BUG_ON(!p->exit_state);
6256 /* Cannot have done final schedule yet: would have vanished. */
6257 BUG_ON(p->state == TASK_DEAD);
6262 * Drop lock around migration; if someone else moves it,
6263 * that's OK. No task can be added to this CPU, so iteration is
6266 spin_unlock_irq(&rq->lock);
6267 move_task_off_dead_cpu(dead_cpu, p);
6268 spin_lock_irq(&rq->lock);
6273 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6274 static void migrate_dead_tasks(unsigned int dead_cpu)
6276 struct rq *rq = cpu_rq(dead_cpu);
6277 struct task_struct *next;
6280 if (!rq->nr_running)
6282 update_rq_clock(rq);
6283 next = pick_next_task(rq, rq->curr);
6286 migrate_dead(dead_cpu, next);
6290 #endif /* CONFIG_HOTPLUG_CPU */
6292 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6294 static struct ctl_table sd_ctl_dir[] = {
6296 .procname = "sched_domain",
6302 static struct ctl_table sd_ctl_root[] = {
6304 .ctl_name = CTL_KERN,
6305 .procname = "kernel",
6307 .child = sd_ctl_dir,
6312 static struct ctl_table *sd_alloc_ctl_entry(int n)
6314 struct ctl_table *entry =
6315 kcalloc(n, sizeof(struct ctl_table), GFP_KERNEL);
6320 static void sd_free_ctl_entry(struct ctl_table **tablep)
6322 struct ctl_table *entry;
6325 * In the intermediate directories, both the child directory and
6326 * procname are dynamically allocated and could fail but the mode
6327 * will always be set. In the lowest directory the names are
6328 * static strings and all have proc handlers.
6330 for (entry = *tablep; entry->mode; entry++) {
6332 sd_free_ctl_entry(&entry->child);
6333 if (entry->proc_handler == NULL)
6334 kfree(entry->procname);
6342 set_table_entry(struct ctl_table *entry,
6343 const char *procname, void *data, int maxlen,
6344 mode_t mode, proc_handler *proc_handler)
6346 entry->procname = procname;
6348 entry->maxlen = maxlen;
6350 entry->proc_handler = proc_handler;
6353 static struct ctl_table *
6354 sd_alloc_ctl_domain_table(struct sched_domain *sd)
6356 struct ctl_table *table = sd_alloc_ctl_entry(12);
6361 set_table_entry(&table[0], "min_interval", &sd->min_interval,
6362 sizeof(long), 0644, proc_doulongvec_minmax);
6363 set_table_entry(&table[1], "max_interval", &sd->max_interval,
6364 sizeof(long), 0644, proc_doulongvec_minmax);
6365 set_table_entry(&table[2], "busy_idx", &sd->busy_idx,
6366 sizeof(int), 0644, proc_dointvec_minmax);
6367 set_table_entry(&table[3], "idle_idx", &sd->idle_idx,
6368 sizeof(int), 0644, proc_dointvec_minmax);
6369 set_table_entry(&table[4], "newidle_idx", &sd->newidle_idx,
6370 sizeof(int), 0644, proc_dointvec_minmax);
6371 set_table_entry(&table[5], "wake_idx", &sd->wake_idx,
6372 sizeof(int), 0644, proc_dointvec_minmax);
6373 set_table_entry(&table[6], "forkexec_idx", &sd->forkexec_idx,
6374 sizeof(int), 0644, proc_dointvec_minmax);
6375 set_table_entry(&table[7], "busy_factor", &sd->busy_factor,
6376 sizeof(int), 0644, proc_dointvec_minmax);
6377 set_table_entry(&table[8], "imbalance_pct", &sd->imbalance_pct,
6378 sizeof(int), 0644, proc_dointvec_minmax);
6379 set_table_entry(&table[9], "cache_nice_tries",
6380 &sd->cache_nice_tries,
6381 sizeof(int), 0644, proc_dointvec_minmax);
6382 set_table_entry(&table[10], "flags", &sd->flags,
6383 sizeof(int), 0644, proc_dointvec_minmax);
6384 /* &table[11] is terminator */
6389 static ctl_table *sd_alloc_ctl_cpu_table(int cpu)
6391 struct ctl_table *entry, *table;
6392 struct sched_domain *sd;
6393 int domain_num = 0, i;
6396 for_each_domain(cpu, sd)
6398 entry = table = sd_alloc_ctl_entry(domain_num + 1);
6403 for_each_domain(cpu, sd) {
6404 snprintf(buf, 32, "domain%d", i);
6405 entry->procname = kstrdup(buf, GFP_KERNEL);
6407 entry->child = sd_alloc_ctl_domain_table(sd);
6414 static struct ctl_table_header *sd_sysctl_header;
6415 static void register_sched_domain_sysctl(void)
6417 int i, cpu_num = num_online_cpus();
6418 struct ctl_table *entry = sd_alloc_ctl_entry(cpu_num + 1);
6421 WARN_ON(sd_ctl_dir[0].child);
6422 sd_ctl_dir[0].child = entry;
6427 for_each_online_cpu(i) {
6428 snprintf(buf, 32, "cpu%d", i);
6429 entry->procname = kstrdup(buf, GFP_KERNEL);
6431 entry->child = sd_alloc_ctl_cpu_table(i);
6435 WARN_ON(sd_sysctl_header);
6436 sd_sysctl_header = register_sysctl_table(sd_ctl_root);
6439 /* may be called multiple times per register */
6440 static void unregister_sched_domain_sysctl(void)
6442 if (sd_sysctl_header)
6443 unregister_sysctl_table(sd_sysctl_header);
6444 sd_sysctl_header = NULL;
6445 if (sd_ctl_dir[0].child)
6446 sd_free_ctl_entry(&sd_ctl_dir[0].child);
6449 static void register_sched_domain_sysctl(void)
6452 static void unregister_sched_domain_sysctl(void)
6458 * migration_call - callback that gets triggered when a CPU is added.
6459 * Here we can start up the necessary migration thread for the new CPU.
6461 static int __cpuinit
6462 migration_call(struct notifier_block *nfb, unsigned long action, void *hcpu)
6464 struct task_struct *p;
6465 int cpu = (long)hcpu;
6466 unsigned long flags;
6471 case CPU_UP_PREPARE:
6472 case CPU_UP_PREPARE_FROZEN:
6473 p = kthread_create(migration_thread, hcpu, "migration/%d", cpu);
6476 kthread_bind(p, cpu);
6477 /* Must be high prio: stop_machine expects to yield to it. */
6478 rq = task_rq_lock(p, &flags);
6479 __setscheduler(rq, p, SCHED_FIFO, MAX_RT_PRIO-1);
6480 task_rq_unlock(rq, &flags);
6481 cpu_rq(cpu)->migration_thread = p;
6485 case CPU_ONLINE_FROZEN:
6486 /* Strictly unnecessary, as first user will wake it. */
6487 wake_up_process(cpu_rq(cpu)->migration_thread);
6489 /* Update our root-domain */
6491 spin_lock_irqsave(&rq->lock, flags);
6493 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6494 cpu_set(cpu, rq->rd->online);
6496 spin_unlock_irqrestore(&rq->lock, flags);
6499 #ifdef CONFIG_HOTPLUG_CPU
6500 case CPU_UP_CANCELED:
6501 case CPU_UP_CANCELED_FROZEN:
6502 if (!cpu_rq(cpu)->migration_thread)
6504 /* Unbind it from offline cpu so it can run. Fall thru. */
6505 kthread_bind(cpu_rq(cpu)->migration_thread,
6506 any_online_cpu(cpu_online_map));
6507 kthread_stop(cpu_rq(cpu)->migration_thread);
6508 cpu_rq(cpu)->migration_thread = NULL;
6512 case CPU_DEAD_FROZEN:
6513 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6514 migrate_live_tasks(cpu);
6516 kthread_stop(rq->migration_thread);
6517 rq->migration_thread = NULL;
6518 /* Idle task back to normal (off runqueue, low prio) */
6519 spin_lock_irq(&rq->lock);
6520 update_rq_clock(rq);
6521 deactivate_task(rq, rq->idle, 0);
6522 rq->idle->static_prio = MAX_PRIO;
6523 __setscheduler(rq, rq->idle, SCHED_NORMAL, 0);
6524 rq->idle->sched_class = &idle_sched_class;
6525 migrate_dead_tasks(cpu);
6526 spin_unlock_irq(&rq->lock);
6528 migrate_nr_uninterruptible(rq);
6529 BUG_ON(rq->nr_running != 0);
6532 * No need to migrate the tasks: it was best-effort if
6533 * they didn't take sched_hotcpu_mutex. Just wake up
6536 spin_lock_irq(&rq->lock);
6537 while (!list_empty(&rq->migration_queue)) {
6538 struct migration_req *req;
6540 req = list_entry(rq->migration_queue.next,
6541 struct migration_req, list);
6542 list_del_init(&req->list);
6543 complete(&req->done);
6545 spin_unlock_irq(&rq->lock);
6549 case CPU_DYING_FROZEN:
6550 /* Update our root-domain */
6552 spin_lock_irqsave(&rq->lock, flags);
6554 BUG_ON(!cpu_isset(cpu, rq->rd->span));
6555 cpu_clear(cpu, rq->rd->online);
6557 spin_unlock_irqrestore(&rq->lock, flags);
6564 /* Register at highest priority so that task migration (migrate_all_tasks)
6565 * happens before everything else.
6567 static struct notifier_block __cpuinitdata migration_notifier = {
6568 .notifier_call = migration_call,
6572 void __init migration_init(void)
6574 void *cpu = (void *)(long)smp_processor_id();
6577 /* Start one for the boot CPU: */
6578 err = migration_call(&migration_notifier, CPU_UP_PREPARE, cpu);
6579 BUG_ON(err == NOTIFY_BAD);
6580 migration_call(&migration_notifier, CPU_ONLINE, cpu);
6581 register_cpu_notifier(&migration_notifier);
6587 #ifdef CONFIG_SCHED_DEBUG
6589 static int sched_domain_debug_one(struct sched_domain *sd, int cpu, int level,
6590 cpumask_t *groupmask)
6592 struct sched_group *group = sd->groups;
6595 cpulist_scnprintf(str, sizeof(str), sd->span);
6596 cpus_clear(*groupmask);
6598 printk(KERN_DEBUG "%*s domain %d: ", level, "", level);
6600 if (!(sd->flags & SD_LOAD_BALANCE)) {
6601 printk("does not load-balance\n");
6603 printk(KERN_ERR "ERROR: !SD_LOAD_BALANCE domain"
6608 printk(KERN_CONT "span %s\n", str);
6610 if (!cpu_isset(cpu, sd->span)) {
6611 printk(KERN_ERR "ERROR: domain->span does not contain "
6614 if (!cpu_isset(cpu, group->cpumask)) {
6615 printk(KERN_ERR "ERROR: domain->groups does not contain"
6619 printk(KERN_DEBUG "%*s groups:", level + 1, "");
6623 printk(KERN_ERR "ERROR: group is NULL\n");
6627 if (!group->__cpu_power) {
6628 printk(KERN_CONT "\n");
6629 printk(KERN_ERR "ERROR: domain->cpu_power not "
6634 if (!cpus_weight(group->cpumask)) {
6635 printk(KERN_CONT "\n");
6636 printk(KERN_ERR "ERROR: empty group\n");
6640 if (cpus_intersects(*groupmask, group->cpumask)) {
6641 printk(KERN_CONT "\n");
6642 printk(KERN_ERR "ERROR: repeated CPUs\n");
6646 cpus_or(*groupmask, *groupmask, group->cpumask);
6648 cpulist_scnprintf(str, sizeof(str), group->cpumask);
6649 printk(KERN_CONT " %s", str);
6651 group = group->next;
6652 } while (group != sd->groups);
6653 printk(KERN_CONT "\n");
6655 if (!cpus_equal(sd->span, *groupmask))
6656 printk(KERN_ERR "ERROR: groups don't span domain->span\n");
6658 if (sd->parent && !cpus_subset(*groupmask, sd->parent->span))
6659 printk(KERN_ERR "ERROR: parent span is not a superset "
6660 "of domain->span\n");
6664 static void sched_domain_debug(struct sched_domain *sd, int cpu)
6666 cpumask_t *groupmask;
6670 printk(KERN_DEBUG "CPU%d attaching NULL sched-domain.\n", cpu);
6674 printk(KERN_DEBUG "CPU%d attaching sched-domain:\n", cpu);
6676 groupmask = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
6678 printk(KERN_DEBUG "Cannot load-balance (out of memory)\n");
6683 if (sched_domain_debug_one(sd, cpu, level, groupmask))
6693 # define sched_domain_debug(sd, cpu) do { } while (0)
6696 static int sd_degenerate(struct sched_domain *sd)
6698 if (cpus_weight(sd->span) == 1)
6701 /* Following flags need at least 2 groups */
6702 if (sd->flags & (SD_LOAD_BALANCE |
6703 SD_BALANCE_NEWIDLE |
6707 SD_SHARE_PKG_RESOURCES)) {
6708 if (sd->groups != sd->groups->next)
6712 /* Following flags don't use groups */
6713 if (sd->flags & (SD_WAKE_IDLE |
6722 sd_parent_degenerate(struct sched_domain *sd, struct sched_domain *parent)
6724 unsigned long cflags = sd->flags, pflags = parent->flags;
6726 if (sd_degenerate(parent))
6729 if (!cpus_equal(sd->span, parent->span))
6732 /* Does parent contain flags not in child? */
6733 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6734 if (cflags & SD_WAKE_AFFINE)
6735 pflags &= ~SD_WAKE_BALANCE;
6736 /* Flags needing groups don't count if only 1 group in parent */
6737 if (parent->groups == parent->groups->next) {
6738 pflags &= ~(SD_LOAD_BALANCE |
6739 SD_BALANCE_NEWIDLE |
6743 SD_SHARE_PKG_RESOURCES);
6745 if (~cflags & pflags)
6751 static void rq_attach_root(struct rq *rq, struct root_domain *rd)
6753 unsigned long flags;
6754 const struct sched_class *class;
6756 spin_lock_irqsave(&rq->lock, flags);
6759 struct root_domain *old_rd = rq->rd;
6761 for (class = sched_class_highest; class; class = class->next) {
6762 if (class->leave_domain)
6763 class->leave_domain(rq);
6766 cpu_clear(rq->cpu, old_rd->span);
6767 cpu_clear(rq->cpu, old_rd->online);
6769 if (atomic_dec_and_test(&old_rd->refcount))
6773 atomic_inc(&rd->refcount);
6776 cpu_set(rq->cpu, rd->span);
6777 if (cpu_isset(rq->cpu, cpu_online_map))
6778 cpu_set(rq->cpu, rd->online);
6780 for (class = sched_class_highest; class; class = class->next) {
6781 if (class->join_domain)
6782 class->join_domain(rq);
6785 spin_unlock_irqrestore(&rq->lock, flags);
6788 static void init_rootdomain(struct root_domain *rd)
6790 memset(rd, 0, sizeof(*rd));
6792 cpus_clear(rd->span);
6793 cpus_clear(rd->online);
6796 static void init_defrootdomain(void)
6798 init_rootdomain(&def_root_domain);
6799 atomic_set(&def_root_domain.refcount, 1);
6802 static struct root_domain *alloc_rootdomain(void)
6804 struct root_domain *rd;
6806 rd = kmalloc(sizeof(*rd), GFP_KERNEL);
6810 init_rootdomain(rd);
6816 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6817 * hold the hotplug lock.
6820 cpu_attach_domain(struct sched_domain *sd, struct root_domain *rd, int cpu)
6822 struct rq *rq = cpu_rq(cpu);
6823 struct sched_domain *tmp;
6825 /* Remove the sched domains which do not contribute to scheduling. */
6826 for (tmp = sd; tmp; tmp = tmp->parent) {
6827 struct sched_domain *parent = tmp->parent;
6830 if (sd_parent_degenerate(tmp, parent)) {
6831 tmp->parent = parent->parent;
6833 parent->parent->child = tmp;
6837 if (sd && sd_degenerate(sd)) {
6843 sched_domain_debug(sd, cpu);
6845 rq_attach_root(rq, rd);
6846 rcu_assign_pointer(rq->sd, sd);
6849 /* cpus with isolated domains */
6850 static cpumask_t cpu_isolated_map = CPU_MASK_NONE;
6852 /* Setup the mask of cpus configured for isolated domains */
6853 static int __init isolated_cpu_setup(char *str)
6855 int ints[NR_CPUS], i;
6857 str = get_options(str, ARRAY_SIZE(ints), ints);
6858 cpus_clear(cpu_isolated_map);
6859 for (i = 1; i <= ints[0]; i++)
6860 if (ints[i] < NR_CPUS)
6861 cpu_set(ints[i], cpu_isolated_map);
6865 __setup("isolcpus=", isolated_cpu_setup);
6868 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6869 * to a function which identifies what group(along with sched group) a CPU
6870 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6871 * (due to the fact that we keep track of groups covered with a cpumask_t).
6873 * init_sched_build_groups will build a circular linked list of the groups
6874 * covered by the given span, and will set each group's ->cpumask correctly,
6875 * and ->cpu_power to 0.
6878 init_sched_build_groups(const cpumask_t *span, const cpumask_t *cpu_map,
6879 int (*group_fn)(int cpu, const cpumask_t *cpu_map,
6880 struct sched_group **sg,
6881 cpumask_t *tmpmask),
6882 cpumask_t *covered, cpumask_t *tmpmask)
6884 struct sched_group *first = NULL, *last = NULL;
6887 cpus_clear(*covered);
6889 for_each_cpu_mask(i, *span) {
6890 struct sched_group *sg;
6891 int group = group_fn(i, cpu_map, &sg, tmpmask);
6894 if (cpu_isset(i, *covered))
6897 cpus_clear(sg->cpumask);
6898 sg->__cpu_power = 0;
6900 for_each_cpu_mask(j, *span) {
6901 if (group_fn(j, cpu_map, NULL, tmpmask) != group)
6904 cpu_set(j, *covered);
6905 cpu_set(j, sg->cpumask);
6916 #define SD_NODES_PER_DOMAIN 16
6921 * find_next_best_node - find the next node to include in a sched_domain
6922 * @node: node whose sched_domain we're building
6923 * @used_nodes: nodes already in the sched_domain
6925 * Find the next node to include in a given scheduling domain. Simply
6926 * finds the closest node not already in the @used_nodes map.
6928 * Should use nodemask_t.
6930 static int find_next_best_node(int node, nodemask_t *used_nodes)
6932 int i, n, val, min_val, best_node = 0;
6936 for (i = 0; i < MAX_NUMNODES; i++) {
6937 /* Start at @node */
6938 n = (node + i) % MAX_NUMNODES;
6940 if (!nr_cpus_node(n))
6943 /* Skip already used nodes */
6944 if (node_isset(n, *used_nodes))
6947 /* Simple min distance search */
6948 val = node_distance(node, n);
6950 if (val < min_val) {
6956 node_set(best_node, *used_nodes);
6961 * sched_domain_node_span - get a cpumask for a node's sched_domain
6962 * @node: node whose cpumask we're constructing
6963 * @span: resulting cpumask
6965 * Given a node, construct a good cpumask for its sched_domain to span. It
6966 * should be one that prevents unnecessary balancing, but also spreads tasks
6969 static void sched_domain_node_span(int node, cpumask_t *span)
6971 nodemask_t used_nodes;
6972 node_to_cpumask_ptr(nodemask, node);
6976 nodes_clear(used_nodes);
6978 cpus_or(*span, *span, *nodemask);
6979 node_set(node, used_nodes);
6981 for (i = 1; i < SD_NODES_PER_DOMAIN; i++) {
6982 int next_node = find_next_best_node(node, &used_nodes);
6984 node_to_cpumask_ptr_next(nodemask, next_node);
6985 cpus_or(*span, *span, *nodemask);
6990 int sched_smt_power_savings = 0, sched_mc_power_savings = 0;
6993 * SMT sched-domains:
6995 #ifdef CONFIG_SCHED_SMT
6996 static DEFINE_PER_CPU(struct sched_domain, cpu_domains);
6997 static DEFINE_PER_CPU(struct sched_group, sched_group_cpus);
7000 cpu_to_cpu_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7004 *sg = &per_cpu(sched_group_cpus, cpu);
7010 * multi-core sched-domains:
7012 #ifdef CONFIG_SCHED_MC
7013 static DEFINE_PER_CPU(struct sched_domain, core_domains);
7014 static DEFINE_PER_CPU(struct sched_group, sched_group_core);
7017 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7019 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7024 *mask = per_cpu(cpu_sibling_map, cpu);
7025 cpus_and(*mask, *mask, *cpu_map);
7026 group = first_cpu(*mask);
7028 *sg = &per_cpu(sched_group_core, group);
7031 #elif defined(CONFIG_SCHED_MC)
7033 cpu_to_core_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7037 *sg = &per_cpu(sched_group_core, cpu);
7042 static DEFINE_PER_CPU(struct sched_domain, phys_domains);
7043 static DEFINE_PER_CPU(struct sched_group, sched_group_phys);
7046 cpu_to_phys_group(int cpu, const cpumask_t *cpu_map, struct sched_group **sg,
7050 #ifdef CONFIG_SCHED_MC
7051 *mask = cpu_coregroup_map(cpu);
7052 cpus_and(*mask, *mask, *cpu_map);
7053 group = first_cpu(*mask);
7054 #elif defined(CONFIG_SCHED_SMT)
7055 *mask = per_cpu(cpu_sibling_map, cpu);
7056 cpus_and(*mask, *mask, *cpu_map);
7057 group = first_cpu(*mask);
7062 *sg = &per_cpu(sched_group_phys, group);
7068 * The init_sched_build_groups can't handle what we want to do with node
7069 * groups, so roll our own. Now each node has its own list of groups which
7070 * gets dynamically allocated.
7072 static DEFINE_PER_CPU(struct sched_domain, node_domains);
7073 static struct sched_group ***sched_group_nodes_bycpu;
7075 static DEFINE_PER_CPU(struct sched_domain, allnodes_domains);
7076 static DEFINE_PER_CPU(struct sched_group, sched_group_allnodes);
7078 static int cpu_to_allnodes_group(int cpu, const cpumask_t *cpu_map,
7079 struct sched_group **sg, cpumask_t *nodemask)
7083 *nodemask = node_to_cpumask(cpu_to_node(cpu));
7084 cpus_and(*nodemask, *nodemask, *cpu_map);
7085 group = first_cpu(*nodemask);
7088 *sg = &per_cpu(sched_group_allnodes, group);
7092 static void init_numa_sched_groups_power(struct sched_group *group_head)
7094 struct sched_group *sg = group_head;
7100 for_each_cpu_mask(j, sg->cpumask) {
7101 struct sched_domain *sd;
7103 sd = &per_cpu(phys_domains, j);
7104 if (j != first_cpu(sd->groups->cpumask)) {
7106 * Only add "power" once for each
7112 sg_inc_cpu_power(sg, sd->groups->__cpu_power);
7115 } while (sg != group_head);
7120 /* Free memory allocated for various sched_group structures */
7121 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7125 for_each_cpu_mask(cpu, *cpu_map) {
7126 struct sched_group **sched_group_nodes
7127 = sched_group_nodes_bycpu[cpu];
7129 if (!sched_group_nodes)
7132 for (i = 0; i < MAX_NUMNODES; i++) {
7133 struct sched_group *oldsg, *sg = sched_group_nodes[i];
7135 *nodemask = node_to_cpumask(i);
7136 cpus_and(*nodemask, *nodemask, *cpu_map);
7137 if (cpus_empty(*nodemask))
7147 if (oldsg != sched_group_nodes[i])
7150 kfree(sched_group_nodes);
7151 sched_group_nodes_bycpu[cpu] = NULL;
7155 static void free_sched_groups(const cpumask_t *cpu_map, cpumask_t *nodemask)
7161 * Initialize sched groups cpu_power.
7163 * cpu_power indicates the capacity of sched group, which is used while
7164 * distributing the load between different sched groups in a sched domain.
7165 * Typically cpu_power for all the groups in a sched domain will be same unless
7166 * there are asymmetries in the topology. If there are asymmetries, group
7167 * having more cpu_power will pickup more load compared to the group having
7170 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7171 * the maximum number of tasks a group can handle in the presence of other idle
7172 * or lightly loaded groups in the same sched domain.
7174 static void init_sched_groups_power(int cpu, struct sched_domain *sd)
7176 struct sched_domain *child;
7177 struct sched_group *group;
7179 WARN_ON(!sd || !sd->groups);
7181 if (cpu != first_cpu(sd->groups->cpumask))
7186 sd->groups->__cpu_power = 0;
7189 * For perf policy, if the groups in child domain share resources
7190 * (for example cores sharing some portions of the cache hierarchy
7191 * or SMT), then set this domain groups cpu_power such that each group
7192 * can handle only one task, when there are other idle groups in the
7193 * same sched domain.
7195 if (!child || (!(sd->flags & SD_POWERSAVINGS_BALANCE) &&
7197 (SD_SHARE_CPUPOWER | SD_SHARE_PKG_RESOURCES)))) {
7198 sg_inc_cpu_power(sd->groups, SCHED_LOAD_SCALE);
7203 * add cpu_power of each child group to this groups cpu_power
7205 group = child->groups;
7207 sg_inc_cpu_power(sd->groups, group->__cpu_power);
7208 group = group->next;
7209 } while (group != child->groups);
7213 * Initializers for schedule domains
7214 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7217 #define SD_INIT(sd, type) sd_init_##type(sd)
7218 #define SD_INIT_FUNC(type) \
7219 static noinline void sd_init_##type(struct sched_domain *sd) \
7221 memset(sd, 0, sizeof(*sd)); \
7222 *sd = SD_##type##_INIT; \
7223 sd->level = SD_LV_##type; \
7228 SD_INIT_FUNC(ALLNODES)
7231 #ifdef CONFIG_SCHED_SMT
7232 SD_INIT_FUNC(SIBLING)
7234 #ifdef CONFIG_SCHED_MC
7239 * To minimize stack usage kmalloc room for cpumasks and share the
7240 * space as the usage in build_sched_domains() dictates. Used only
7241 * if the amount of space is significant.
7244 cpumask_t tmpmask; /* make this one first */
7247 cpumask_t this_sibling_map;
7248 cpumask_t this_core_map;
7250 cpumask_t send_covered;
7253 cpumask_t domainspan;
7255 cpumask_t notcovered;
7260 #define SCHED_CPUMASK_ALLOC 1
7261 #define SCHED_CPUMASK_FREE(v) kfree(v)
7262 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7264 #define SCHED_CPUMASK_ALLOC 0
7265 #define SCHED_CPUMASK_FREE(v)
7266 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7269 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7270 ((unsigned long)(a) + offsetof(struct allmasks, v))
7272 static int default_relax_domain_level = -1;
7274 static int __init setup_relax_domain_level(char *str)
7276 default_relax_domain_level = simple_strtoul(str, NULL, 0);
7279 __setup("relax_domain_level=", setup_relax_domain_level);
7281 static void set_domain_attribute(struct sched_domain *sd,
7282 struct sched_domain_attr *attr)
7286 if (!attr || attr->relax_domain_level < 0) {
7287 if (default_relax_domain_level < 0)
7290 request = default_relax_domain_level;
7292 request = attr->relax_domain_level;
7293 if (request < sd->level) {
7294 /* turn off idle balance on this domain */
7295 sd->flags &= ~(SD_WAKE_IDLE|SD_BALANCE_NEWIDLE);
7297 /* turn on idle balance on this domain */
7298 sd->flags |= (SD_WAKE_IDLE_FAR|SD_BALANCE_NEWIDLE);
7303 * Build sched domains for a given set of cpus and attach the sched domains
7304 * to the individual cpus
7306 static int __build_sched_domains(const cpumask_t *cpu_map,
7307 struct sched_domain_attr *attr)
7310 struct root_domain *rd;
7311 SCHED_CPUMASK_DECLARE(allmasks);
7314 struct sched_group **sched_group_nodes = NULL;
7315 int sd_allnodes = 0;
7318 * Allocate the per-node list of sched groups
7320 sched_group_nodes = kcalloc(MAX_NUMNODES, sizeof(struct sched_group *),
7322 if (!sched_group_nodes) {
7323 printk(KERN_WARNING "Can not alloc sched group node list\n");
7328 rd = alloc_rootdomain();
7330 printk(KERN_WARNING "Cannot alloc root domain\n");
7332 kfree(sched_group_nodes);
7337 #if SCHED_CPUMASK_ALLOC
7338 /* get space for all scratch cpumask variables */
7339 allmasks = kmalloc(sizeof(*allmasks), GFP_KERNEL);
7341 printk(KERN_WARNING "Cannot alloc cpumask array\n");
7344 kfree(sched_group_nodes);
7349 tmpmask = (cpumask_t *)allmasks;
7353 sched_group_nodes_bycpu[first_cpu(*cpu_map)] = sched_group_nodes;
7357 * Set up domains for cpus specified by the cpu_map.
7359 for_each_cpu_mask(i, *cpu_map) {
7360 struct sched_domain *sd = NULL, *p;
7361 SCHED_CPUMASK_VAR(nodemask, allmasks);
7363 *nodemask = node_to_cpumask(cpu_to_node(i));
7364 cpus_and(*nodemask, *nodemask, *cpu_map);
7367 if (cpus_weight(*cpu_map) >
7368 SD_NODES_PER_DOMAIN*cpus_weight(*nodemask)) {
7369 sd = &per_cpu(allnodes_domains, i);
7370 SD_INIT(sd, ALLNODES);
7371 set_domain_attribute(sd, attr);
7372 sd->span = *cpu_map;
7373 sd->first_cpu = first_cpu(sd->span);
7374 cpu_to_allnodes_group(i, cpu_map, &sd->groups, tmpmask);
7380 sd = &per_cpu(node_domains, i);
7382 set_domain_attribute(sd, attr);
7383 sched_domain_node_span(cpu_to_node(i), &sd->span);
7384 sd->first_cpu = first_cpu(sd->span);
7388 cpus_and(sd->span, sd->span, *cpu_map);
7392 sd = &per_cpu(phys_domains, i);
7394 set_domain_attribute(sd, attr);
7395 sd->span = *nodemask;
7396 sd->first_cpu = first_cpu(sd->span);
7400 cpu_to_phys_group(i, cpu_map, &sd->groups, tmpmask);
7402 #ifdef CONFIG_SCHED_MC
7404 sd = &per_cpu(core_domains, i);
7406 set_domain_attribute(sd, attr);
7407 sd->span = cpu_coregroup_map(i);
7408 sd->first_cpu = first_cpu(sd->span);
7409 cpus_and(sd->span, sd->span, *cpu_map);
7412 cpu_to_core_group(i, cpu_map, &sd->groups, tmpmask);
7415 #ifdef CONFIG_SCHED_SMT
7417 sd = &per_cpu(cpu_domains, i);
7418 SD_INIT(sd, SIBLING);
7419 set_domain_attribute(sd, attr);
7420 sd->span = per_cpu(cpu_sibling_map, i);
7421 sd->first_cpu = first_cpu(sd->span);
7422 cpus_and(sd->span, sd->span, *cpu_map);
7425 cpu_to_cpu_group(i, cpu_map, &sd->groups, tmpmask);
7429 #ifdef CONFIG_SCHED_SMT
7430 /* Set up CPU (sibling) groups */
7431 for_each_cpu_mask(i, *cpu_map) {
7432 SCHED_CPUMASK_VAR(this_sibling_map, allmasks);
7433 SCHED_CPUMASK_VAR(send_covered, allmasks);
7435 *this_sibling_map = per_cpu(cpu_sibling_map, i);
7436 cpus_and(*this_sibling_map, *this_sibling_map, *cpu_map);
7437 if (i != first_cpu(*this_sibling_map))
7440 init_sched_build_groups(this_sibling_map, cpu_map,
7442 send_covered, tmpmask);
7446 #ifdef CONFIG_SCHED_MC
7447 /* Set up multi-core groups */
7448 for_each_cpu_mask(i, *cpu_map) {
7449 SCHED_CPUMASK_VAR(this_core_map, allmasks);
7450 SCHED_CPUMASK_VAR(send_covered, allmasks);
7452 *this_core_map = cpu_coregroup_map(i);
7453 cpus_and(*this_core_map, *this_core_map, *cpu_map);
7454 if (i != first_cpu(*this_core_map))
7457 init_sched_build_groups(this_core_map, cpu_map,
7459 send_covered, tmpmask);
7463 /* Set up physical groups */
7464 for (i = 0; i < MAX_NUMNODES; i++) {
7465 SCHED_CPUMASK_VAR(nodemask, allmasks);
7466 SCHED_CPUMASK_VAR(send_covered, allmasks);
7468 *nodemask = node_to_cpumask(i);
7469 cpus_and(*nodemask, *nodemask, *cpu_map);
7470 if (cpus_empty(*nodemask))
7473 init_sched_build_groups(nodemask, cpu_map,
7475 send_covered, tmpmask);
7479 /* Set up node groups */
7481 SCHED_CPUMASK_VAR(send_covered, allmasks);
7483 init_sched_build_groups(cpu_map, cpu_map,
7484 &cpu_to_allnodes_group,
7485 send_covered, tmpmask);
7488 for (i = 0; i < MAX_NUMNODES; i++) {
7489 /* Set up node groups */
7490 struct sched_group *sg, *prev;
7491 SCHED_CPUMASK_VAR(nodemask, allmasks);
7492 SCHED_CPUMASK_VAR(domainspan, allmasks);
7493 SCHED_CPUMASK_VAR(covered, allmasks);
7496 *nodemask = node_to_cpumask(i);
7497 cpus_clear(*covered);
7499 cpus_and(*nodemask, *nodemask, *cpu_map);
7500 if (cpus_empty(*nodemask)) {
7501 sched_group_nodes[i] = NULL;
7505 sched_domain_node_span(i, domainspan);
7506 cpus_and(*domainspan, *domainspan, *cpu_map);
7508 sg = kmalloc_node(sizeof(struct sched_group), GFP_KERNEL, i);
7510 printk(KERN_WARNING "Can not alloc domain group for "
7514 sched_group_nodes[i] = sg;
7515 for_each_cpu_mask(j, *nodemask) {
7516 struct sched_domain *sd;
7518 sd = &per_cpu(node_domains, j);
7521 sg->__cpu_power = 0;
7522 sg->cpumask = *nodemask;
7524 cpus_or(*covered, *covered, *nodemask);
7527 for (j = 0; j < MAX_NUMNODES; j++) {
7528 SCHED_CPUMASK_VAR(notcovered, allmasks);
7529 int n = (i + j) % MAX_NUMNODES;
7530 node_to_cpumask_ptr(pnodemask, n);
7532 cpus_complement(*notcovered, *covered);
7533 cpus_and(*tmpmask, *notcovered, *cpu_map);
7534 cpus_and(*tmpmask, *tmpmask, *domainspan);
7535 if (cpus_empty(*tmpmask))
7538 cpus_and(*tmpmask, *tmpmask, *pnodemask);
7539 if (cpus_empty(*tmpmask))
7542 sg = kmalloc_node(sizeof(struct sched_group),
7546 "Can not alloc domain group for node %d\n", j);
7549 sg->__cpu_power = 0;
7550 sg->cpumask = *tmpmask;
7551 sg->next = prev->next;
7552 cpus_or(*covered, *covered, *tmpmask);
7559 /* Calculate CPU power for physical packages and nodes */
7560 #ifdef CONFIG_SCHED_SMT
7561 for_each_cpu_mask(i, *cpu_map) {
7562 struct sched_domain *sd = &per_cpu(cpu_domains, i);
7564 init_sched_groups_power(i, sd);
7567 #ifdef CONFIG_SCHED_MC
7568 for_each_cpu_mask(i, *cpu_map) {
7569 struct sched_domain *sd = &per_cpu(core_domains, i);
7571 init_sched_groups_power(i, sd);
7575 for_each_cpu_mask(i, *cpu_map) {
7576 struct sched_domain *sd = &per_cpu(phys_domains, i);
7578 init_sched_groups_power(i, sd);
7582 for (i = 0; i < MAX_NUMNODES; i++)
7583 init_numa_sched_groups_power(sched_group_nodes[i]);
7586 struct sched_group *sg;
7588 cpu_to_allnodes_group(first_cpu(*cpu_map), cpu_map, &sg,
7590 init_numa_sched_groups_power(sg);
7594 /* Attach the domains */
7595 for_each_cpu_mask(i, *cpu_map) {
7596 struct sched_domain *sd;
7597 #ifdef CONFIG_SCHED_SMT
7598 sd = &per_cpu(cpu_domains, i);
7599 #elif defined(CONFIG_SCHED_MC)
7600 sd = &per_cpu(core_domains, i);
7602 sd = &per_cpu(phys_domains, i);
7604 cpu_attach_domain(sd, rd, i);
7607 SCHED_CPUMASK_FREE((void *)allmasks);
7612 free_sched_groups(cpu_map, tmpmask);
7613 SCHED_CPUMASK_FREE((void *)allmasks);
7618 static int build_sched_domains(const cpumask_t *cpu_map)
7620 return __build_sched_domains(cpu_map, NULL);
7623 static cpumask_t *doms_cur; /* current sched domains */
7624 static int ndoms_cur; /* number of sched domains in 'doms_cur' */
7625 static struct sched_domain_attr *dattr_cur; /* attribues of custom domains
7629 * Special case: If a kmalloc of a doms_cur partition (array of
7630 * cpumask_t) fails, then fallback to a single sched domain,
7631 * as determined by the single cpumask_t fallback_doms.
7633 static cpumask_t fallback_doms;
7635 void __attribute__((weak)) arch_update_cpu_topology(void)
7640 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7641 * For now this just excludes isolated cpus, but could be used to
7642 * exclude other special cases in the future.
7644 static int arch_init_sched_domains(const cpumask_t *cpu_map)
7648 arch_update_cpu_topology();
7650 doms_cur = kmalloc(sizeof(cpumask_t), GFP_KERNEL);
7652 doms_cur = &fallback_doms;
7653 cpus_andnot(*doms_cur, *cpu_map, cpu_isolated_map);
7655 err = build_sched_domains(doms_cur);
7656 register_sched_domain_sysctl();
7661 static void arch_destroy_sched_domains(const cpumask_t *cpu_map,
7664 free_sched_groups(cpu_map, tmpmask);
7668 * Detach sched domains from a group of cpus specified in cpu_map
7669 * These cpus will now be attached to the NULL domain
7671 static void detach_destroy_domains(const cpumask_t *cpu_map)
7676 unregister_sched_domain_sysctl();
7678 for_each_cpu_mask(i, *cpu_map)
7679 cpu_attach_domain(NULL, &def_root_domain, i);
7680 synchronize_sched();
7681 arch_destroy_sched_domains(cpu_map, &tmpmask);
7684 /* handle null as "default" */
7685 static int dattrs_equal(struct sched_domain_attr *cur, int idx_cur,
7686 struct sched_domain_attr *new, int idx_new)
7688 struct sched_domain_attr tmp;
7695 return !memcmp(cur ? (cur + idx_cur) : &tmp,
7696 new ? (new + idx_new) : &tmp,
7697 sizeof(struct sched_domain_attr));
7701 * Partition sched domains as specified by the 'ndoms_new'
7702 * cpumasks in the array doms_new[] of cpumasks. This compares
7703 * doms_new[] to the current sched domain partitioning, doms_cur[].
7704 * It destroys each deleted domain and builds each new domain.
7706 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7707 * The masks don't intersect (don't overlap.) We should setup one
7708 * sched domain for each mask. CPUs not in any of the cpumasks will
7709 * not be load balanced. If the same cpumask appears both in the
7710 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7713 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7714 * ownership of it and will kfree it when done with it. If the caller
7715 * failed the kmalloc call, then it can pass in doms_new == NULL,
7716 * and partition_sched_domains() will fallback to the single partition
7719 * Call with hotplug lock held
7721 void partition_sched_domains(int ndoms_new, cpumask_t *doms_new,
7722 struct sched_domain_attr *dattr_new)
7726 mutex_lock(&sched_domains_mutex);
7728 /* always unregister in case we don't destroy any domains */
7729 unregister_sched_domain_sysctl();
7731 if (doms_new == NULL) {
7733 doms_new = &fallback_doms;
7734 cpus_andnot(doms_new[0], cpu_online_map, cpu_isolated_map);
7738 /* Destroy deleted domains */
7739 for (i = 0; i < ndoms_cur; i++) {
7740 for (j = 0; j < ndoms_new; j++) {
7741 if (cpus_equal(doms_cur[i], doms_new[j])
7742 && dattrs_equal(dattr_cur, i, dattr_new, j))
7745 /* no match - a current sched domain not in new doms_new[] */
7746 detach_destroy_domains(doms_cur + i);
7751 /* Build new domains */
7752 for (i = 0; i < ndoms_new; i++) {
7753 for (j = 0; j < ndoms_cur; j++) {
7754 if (cpus_equal(doms_new[i], doms_cur[j])
7755 && dattrs_equal(dattr_new, i, dattr_cur, j))
7758 /* no match - add a new doms_new */
7759 __build_sched_domains(doms_new + i,
7760 dattr_new ? dattr_new + i : NULL);
7765 /* Remember the new sched domains */
7766 if (doms_cur != &fallback_doms)
7768 kfree(dattr_cur); /* kfree(NULL) is safe */
7769 doms_cur = doms_new;
7770 dattr_cur = dattr_new;
7771 ndoms_cur = ndoms_new;
7773 register_sched_domain_sysctl();
7775 mutex_unlock(&sched_domains_mutex);
7778 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7779 int arch_reinit_sched_domains(void)
7784 mutex_lock(&sched_domains_mutex);
7785 detach_destroy_domains(&cpu_online_map);
7786 err = arch_init_sched_domains(&cpu_online_map);
7787 mutex_unlock(&sched_domains_mutex);
7793 static ssize_t sched_power_savings_store(const char *buf, size_t count, int smt)
7797 if (buf[0] != '0' && buf[0] != '1')
7801 sched_smt_power_savings = (buf[0] == '1');
7803 sched_mc_power_savings = (buf[0] == '1');
7805 ret = arch_reinit_sched_domains();
7807 return ret ? ret : count;
7810 #ifdef CONFIG_SCHED_MC
7811 static ssize_t sched_mc_power_savings_show(struct sys_device *dev, char *page)
7813 return sprintf(page, "%u\n", sched_mc_power_savings);
7815 static ssize_t sched_mc_power_savings_store(struct sys_device *dev,
7816 const char *buf, size_t count)
7818 return sched_power_savings_store(buf, count, 0);
7820 static SYSDEV_ATTR(sched_mc_power_savings, 0644, sched_mc_power_savings_show,
7821 sched_mc_power_savings_store);
7824 #ifdef CONFIG_SCHED_SMT
7825 static ssize_t sched_smt_power_savings_show(struct sys_device *dev, char *page)
7827 return sprintf(page, "%u\n", sched_smt_power_savings);
7829 static ssize_t sched_smt_power_savings_store(struct sys_device *dev,
7830 const char *buf, size_t count)
7832 return sched_power_savings_store(buf, count, 1);
7834 static SYSDEV_ATTR(sched_smt_power_savings, 0644, sched_smt_power_savings_show,
7835 sched_smt_power_savings_store);
7838 int sched_create_sysfs_power_savings_entries(struct sysdev_class *cls)
7842 #ifdef CONFIG_SCHED_SMT
7844 err = sysfs_create_file(&cls->kset.kobj,
7845 &attr_sched_smt_power_savings.attr);
7847 #ifdef CONFIG_SCHED_MC
7848 if (!err && mc_capable())
7849 err = sysfs_create_file(&cls->kset.kobj,
7850 &attr_sched_mc_power_savings.attr);
7857 * Force a reinitialization of the sched domains hierarchy. The domains
7858 * and groups cannot be updated in place without racing with the balancing
7859 * code, so we temporarily attach all running cpus to the NULL domain
7860 * which will prevent rebalancing while the sched domains are recalculated.
7862 static int update_sched_domains(struct notifier_block *nfb,
7863 unsigned long action, void *hcpu)
7866 case CPU_UP_PREPARE:
7867 case CPU_UP_PREPARE_FROZEN:
7868 case CPU_DOWN_PREPARE:
7869 case CPU_DOWN_PREPARE_FROZEN:
7870 detach_destroy_domains(&cpu_online_map);
7873 case CPU_UP_CANCELED:
7874 case CPU_UP_CANCELED_FROZEN:
7875 case CPU_DOWN_FAILED:
7876 case CPU_DOWN_FAILED_FROZEN:
7878 case CPU_ONLINE_FROZEN:
7880 case CPU_DEAD_FROZEN:
7882 * Fall through and re-initialise the domains.
7889 /* The hotplug lock is already held by cpu_up/cpu_down */
7890 arch_init_sched_domains(&cpu_online_map);
7895 void __init sched_init_smp(void)
7897 cpumask_t non_isolated_cpus;
7899 #if defined(CONFIG_NUMA)
7900 sched_group_nodes_bycpu = kzalloc(nr_cpu_ids * sizeof(void **),
7902 BUG_ON(sched_group_nodes_bycpu == NULL);
7905 mutex_lock(&sched_domains_mutex);
7906 arch_init_sched_domains(&cpu_online_map);
7907 cpus_andnot(non_isolated_cpus, cpu_possible_map, cpu_isolated_map);
7908 if (cpus_empty(non_isolated_cpus))
7909 cpu_set(smp_processor_id(), non_isolated_cpus);
7910 mutex_unlock(&sched_domains_mutex);
7912 /* XXX: Theoretical race here - CPU may be hotplugged now */
7913 hotcpu_notifier(update_sched_domains, 0);
7916 /* Move init over to a non-isolated CPU */
7917 if (set_cpus_allowed_ptr(current, &non_isolated_cpus) < 0)
7919 sched_init_granularity();
7922 void __init sched_init_smp(void)
7924 sched_init_granularity();
7926 #endif /* CONFIG_SMP */
7928 int in_sched_functions(unsigned long addr)
7930 return in_lock_functions(addr) ||
7931 (addr >= (unsigned long)__sched_text_start
7932 && addr < (unsigned long)__sched_text_end);
7935 static void init_cfs_rq(struct cfs_rq *cfs_rq, struct rq *rq)
7937 cfs_rq->tasks_timeline = RB_ROOT;
7938 INIT_LIST_HEAD(&cfs_rq->tasks);
7939 #ifdef CONFIG_FAIR_GROUP_SCHED
7942 cfs_rq->min_vruntime = (u64)(-(1LL << 20));
7945 static void init_rt_rq(struct rt_rq *rt_rq, struct rq *rq)
7947 struct rt_prio_array *array;
7950 array = &rt_rq->active;
7951 for (i = 0; i < MAX_RT_PRIO; i++) {
7952 INIT_LIST_HEAD(array->queue + i);
7953 __clear_bit(i, array->bitmap);
7955 /* delimiter for bitsearch: */
7956 __set_bit(MAX_RT_PRIO, array->bitmap);
7958 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7959 rt_rq->highest_prio = MAX_RT_PRIO;
7962 rt_rq->rt_nr_migratory = 0;
7963 rt_rq->overloaded = 0;
7967 rt_rq->rt_throttled = 0;
7968 rt_rq->rt_runtime = 0;
7969 spin_lock_init(&rt_rq->rt_runtime_lock);
7971 #ifdef CONFIG_RT_GROUP_SCHED
7972 rt_rq->rt_nr_boosted = 0;
7977 #ifdef CONFIG_FAIR_GROUP_SCHED
7978 static void init_tg_cfs_entry(struct task_group *tg, struct cfs_rq *cfs_rq,
7979 struct sched_entity *se, int cpu, int add,
7980 struct sched_entity *parent)
7982 struct rq *rq = cpu_rq(cpu);
7983 tg->cfs_rq[cpu] = cfs_rq;
7984 init_cfs_rq(cfs_rq, rq);
7987 list_add(&cfs_rq->leaf_cfs_rq_list, &rq->leaf_cfs_rq_list);
7990 /* se could be NULL for init_task_group */
7995 se->cfs_rq = &rq->cfs;
7997 se->cfs_rq = parent->my_q;
8000 se->load.weight = tg->shares;
8001 se->load.inv_weight = 0;
8002 se->parent = parent;
8006 #ifdef CONFIG_RT_GROUP_SCHED
8007 static void init_tg_rt_entry(struct task_group *tg, struct rt_rq *rt_rq,
8008 struct sched_rt_entity *rt_se, int cpu, int add,
8009 struct sched_rt_entity *parent)
8011 struct rq *rq = cpu_rq(cpu);
8013 tg->rt_rq[cpu] = rt_rq;
8014 init_rt_rq(rt_rq, rq);
8016 rt_rq->rt_se = rt_se;
8017 rt_rq->rt_runtime = tg->rt_bandwidth.rt_runtime;
8019 list_add(&rt_rq->leaf_rt_rq_list, &rq->leaf_rt_rq_list);
8021 tg->rt_se[cpu] = rt_se;
8026 rt_se->rt_rq = &rq->rt;
8028 rt_se->rt_rq = parent->my_q;
8030 rt_se->rt_rq = &rq->rt;
8031 rt_se->my_q = rt_rq;
8032 rt_se->parent = parent;
8033 INIT_LIST_HEAD(&rt_se->run_list);
8037 void __init sched_init(void)
8040 unsigned long alloc_size = 0, ptr;
8042 #ifdef CONFIG_FAIR_GROUP_SCHED
8043 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8045 #ifdef CONFIG_RT_GROUP_SCHED
8046 alloc_size += 2 * nr_cpu_ids * sizeof(void **);
8048 #ifdef CONFIG_USER_SCHED
8052 * As sched_init() is called before page_alloc is setup,
8053 * we use alloc_bootmem().
8056 ptr = (unsigned long)alloc_bootmem(alloc_size);
8058 #ifdef CONFIG_FAIR_GROUP_SCHED
8059 init_task_group.se = (struct sched_entity **)ptr;
8060 ptr += nr_cpu_ids * sizeof(void **);
8062 init_task_group.cfs_rq = (struct cfs_rq **)ptr;
8063 ptr += nr_cpu_ids * sizeof(void **);
8065 #ifdef CONFIG_USER_SCHED
8066 root_task_group.se = (struct sched_entity **)ptr;
8067 ptr += nr_cpu_ids * sizeof(void **);
8069 root_task_group.cfs_rq = (struct cfs_rq **)ptr;
8070 ptr += nr_cpu_ids * sizeof(void **);
8073 #ifdef CONFIG_RT_GROUP_SCHED
8074 init_task_group.rt_se = (struct sched_rt_entity **)ptr;
8075 ptr += nr_cpu_ids * sizeof(void **);
8077 init_task_group.rt_rq = (struct rt_rq **)ptr;
8078 ptr += nr_cpu_ids * sizeof(void **);
8080 #ifdef CONFIG_USER_SCHED
8081 root_task_group.rt_se = (struct sched_rt_entity **)ptr;
8082 ptr += nr_cpu_ids * sizeof(void **);
8084 root_task_group.rt_rq = (struct rt_rq **)ptr;
8085 ptr += nr_cpu_ids * sizeof(void **);
8092 init_defrootdomain();
8095 init_rt_bandwidth(&def_rt_bandwidth,
8096 global_rt_period(), global_rt_runtime());
8098 #ifdef CONFIG_RT_GROUP_SCHED
8099 init_rt_bandwidth(&init_task_group.rt_bandwidth,
8100 global_rt_period(), global_rt_runtime());
8101 #ifdef CONFIG_USER_SCHED
8102 init_rt_bandwidth(&root_task_group.rt_bandwidth,
8103 global_rt_period(), RUNTIME_INF);
8107 #ifdef CONFIG_GROUP_SCHED
8108 list_add(&init_task_group.list, &task_groups);
8109 INIT_LIST_HEAD(&init_task_group.children);
8111 #ifdef CONFIG_USER_SCHED
8112 INIT_LIST_HEAD(&root_task_group.children);
8113 init_task_group.parent = &root_task_group;
8114 list_add(&init_task_group.siblings, &root_task_group.children);
8118 for_each_possible_cpu(i) {
8122 spin_lock_init(&rq->lock);
8123 lockdep_set_class(&rq->lock, &rq->rq_lock_key);
8125 init_cfs_rq(&rq->cfs, rq);
8126 init_rt_rq(&rq->rt, rq);
8127 #ifdef CONFIG_FAIR_GROUP_SCHED
8128 init_task_group.shares = init_task_group_load;
8129 INIT_LIST_HEAD(&rq->leaf_cfs_rq_list);
8130 #ifdef CONFIG_CGROUP_SCHED
8132 * How much cpu bandwidth does init_task_group get?
8134 * In case of task-groups formed thr' the cgroup filesystem, it
8135 * gets 100% of the cpu resources in the system. This overall
8136 * system cpu resource is divided among the tasks of
8137 * init_task_group and its child task-groups in a fair manner,
8138 * based on each entity's (task or task-group's) weight
8139 * (se->load.weight).
8141 * In other words, if init_task_group has 10 tasks of weight
8142 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8143 * then A0's share of the cpu resource is:
8145 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8147 * We achieve this by letting init_task_group's tasks sit
8148 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8150 init_tg_cfs_entry(&init_task_group, &rq->cfs, NULL, i, 1, NULL);
8151 #elif defined CONFIG_USER_SCHED
8152 root_task_group.shares = NICE_0_LOAD;
8153 init_tg_cfs_entry(&root_task_group, &rq->cfs, NULL, i, 0, NULL);
8155 * In case of task-groups formed thr' the user id of tasks,
8156 * init_task_group represents tasks belonging to root user.
8157 * Hence it forms a sibling of all subsequent groups formed.
8158 * In this case, init_task_group gets only a fraction of overall
8159 * system cpu resource, based on the weight assigned to root
8160 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8161 * by letting tasks of init_task_group sit in a separate cfs_rq
8162 * (init_cfs_rq) and having one entity represent this group of
8163 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8165 init_tg_cfs_entry(&init_task_group,
8166 &per_cpu(init_cfs_rq, i),
8167 &per_cpu(init_sched_entity, i), i, 1,
8168 root_task_group.se[i]);
8171 #endif /* CONFIG_FAIR_GROUP_SCHED */
8173 rq->rt.rt_runtime = def_rt_bandwidth.rt_runtime;
8174 #ifdef CONFIG_RT_GROUP_SCHED
8175 INIT_LIST_HEAD(&rq->leaf_rt_rq_list);
8176 #ifdef CONFIG_CGROUP_SCHED
8177 init_tg_rt_entry(&init_task_group, &rq->rt, NULL, i, 1, NULL);
8178 #elif defined CONFIG_USER_SCHED
8179 init_tg_rt_entry(&root_task_group, &rq->rt, NULL, i, 0, NULL);
8180 init_tg_rt_entry(&init_task_group,
8181 &per_cpu(init_rt_rq, i),
8182 &per_cpu(init_sched_rt_entity, i), i, 1,
8183 root_task_group.rt_se[i]);
8187 for (j = 0; j < CPU_LOAD_IDX_MAX; j++)
8188 rq->cpu_load[j] = 0;
8192 rq->active_balance = 0;
8193 rq->next_balance = jiffies;
8196 rq->migration_thread = NULL;
8197 INIT_LIST_HEAD(&rq->migration_queue);
8198 rq_attach_root(rq, &def_root_domain);
8201 atomic_set(&rq->nr_iowait, 0);
8204 set_load_weight(&init_task);
8206 #ifdef CONFIG_PREEMPT_NOTIFIERS
8207 INIT_HLIST_HEAD(&init_task.preempt_notifiers);
8211 open_softirq(SCHED_SOFTIRQ, run_rebalance_domains, NULL);
8214 #ifdef CONFIG_RT_MUTEXES
8215 plist_head_init(&init_task.pi_waiters, &init_task.pi_lock);
8219 * The boot idle thread does lazy MMU switching as well:
8221 atomic_inc(&init_mm.mm_count);
8222 enter_lazy_tlb(&init_mm, current);
8225 * Make us the idle thread. Technically, schedule() should not be
8226 * called from this thread, however somewhere below it might be,
8227 * but because we are the idle thread, we just pick up running again
8228 * when this runqueue becomes "idle".
8230 init_idle(current, smp_processor_id());
8232 * During early bootup we pretend to be a normal task:
8234 current->sched_class = &fair_sched_class;
8236 scheduler_running = 1;
8239 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8240 void __might_sleep(char *file, int line)
8243 static unsigned long prev_jiffy; /* ratelimiting */
8245 if ((in_atomic() || irqs_disabled()) &&
8246 system_state == SYSTEM_RUNNING && !oops_in_progress) {
8247 if (time_before(jiffies, prev_jiffy + HZ) && prev_jiffy)
8249 prev_jiffy = jiffies;
8250 printk(KERN_ERR "BUG: sleeping function called from invalid"
8251 " context at %s:%d\n", file, line);
8252 printk("in_atomic():%d, irqs_disabled():%d\n",
8253 in_atomic(), irqs_disabled());
8254 debug_show_held_locks(current);
8255 if (irqs_disabled())
8256 print_irqtrace_events(current);
8261 EXPORT_SYMBOL(__might_sleep);
8264 #ifdef CONFIG_MAGIC_SYSRQ
8265 static void normalize_task(struct rq *rq, struct task_struct *p)
8269 update_rq_clock(rq);
8270 on_rq = p->se.on_rq;
8272 deactivate_task(rq, p, 0);
8273 __setscheduler(rq, p, SCHED_NORMAL, 0);
8275 activate_task(rq, p, 0);
8276 resched_task(rq->curr);
8280 void normalize_rt_tasks(void)
8282 struct task_struct *g, *p;
8283 unsigned long flags;
8286 read_lock_irqsave(&tasklist_lock, flags);
8287 do_each_thread(g, p) {
8289 * Only normalize user tasks:
8294 p->se.exec_start = 0;
8295 #ifdef CONFIG_SCHEDSTATS
8296 p->se.wait_start = 0;
8297 p->se.sleep_start = 0;
8298 p->se.block_start = 0;
8303 * Renice negative nice level userspace
8306 if (TASK_NICE(p) < 0 && p->mm)
8307 set_user_nice(p, 0);
8311 spin_lock(&p->pi_lock);
8312 rq = __task_rq_lock(p);
8314 normalize_task(rq, p);
8316 __task_rq_unlock(rq);
8317 spin_unlock(&p->pi_lock);
8318 } while_each_thread(g, p);
8320 read_unlock_irqrestore(&tasklist_lock, flags);
8323 #endif /* CONFIG_MAGIC_SYSRQ */
8327 * These functions are only useful for the IA64 MCA handling.
8329 * They can only be called when the whole system has been
8330 * stopped - every CPU needs to be quiescent, and no scheduling
8331 * activity can take place. Using them for anything else would
8332 * be a serious bug, and as a result, they aren't even visible
8333 * under any other configuration.
8337 * curr_task - return the current task for a given cpu.
8338 * @cpu: the processor in question.
8340 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8342 struct task_struct *curr_task(int cpu)
8344 return cpu_curr(cpu);
8348 * set_curr_task - set the current task for a given cpu.
8349 * @cpu: the processor in question.
8350 * @p: the task pointer to set.
8352 * Description: This function must only be used when non-maskable interrupts
8353 * are serviced on a separate stack. It allows the architecture to switch the
8354 * notion of the current task on a cpu in a non-blocking manner. This function
8355 * must be called with all CPU's synchronized, and interrupts disabled, the
8356 * and caller must save the original value of the current task (see
8357 * curr_task() above) and restore that value before reenabling interrupts and
8358 * re-starting the system.
8360 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8362 void set_curr_task(int cpu, struct task_struct *p)
8369 #ifdef CONFIG_FAIR_GROUP_SCHED
8370 static void free_fair_sched_group(struct task_group *tg)
8374 for_each_possible_cpu(i) {
8376 kfree(tg->cfs_rq[i]);
8386 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8388 struct cfs_rq *cfs_rq;
8389 struct sched_entity *se, *parent_se;
8393 tg->cfs_rq = kzalloc(sizeof(cfs_rq) * nr_cpu_ids, GFP_KERNEL);
8396 tg->se = kzalloc(sizeof(se) * nr_cpu_ids, GFP_KERNEL);
8400 tg->shares = NICE_0_LOAD;
8402 for_each_possible_cpu(i) {
8405 cfs_rq = kmalloc_node(sizeof(struct cfs_rq),
8406 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8410 se = kmalloc_node(sizeof(struct sched_entity),
8411 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8415 parent_se = parent ? parent->se[i] : NULL;
8416 init_tg_cfs_entry(tg, cfs_rq, se, i, 0, parent_se);
8425 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8427 list_add_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list,
8428 &cpu_rq(cpu)->leaf_cfs_rq_list);
8431 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8433 list_del_rcu(&tg->cfs_rq[cpu]->leaf_cfs_rq_list);
8436 static inline void free_fair_sched_group(struct task_group *tg)
8441 int alloc_fair_sched_group(struct task_group *tg, struct task_group *parent)
8446 static inline void register_fair_sched_group(struct task_group *tg, int cpu)
8450 static inline void unregister_fair_sched_group(struct task_group *tg, int cpu)
8455 #ifdef CONFIG_RT_GROUP_SCHED
8456 static void free_rt_sched_group(struct task_group *tg)
8460 destroy_rt_bandwidth(&tg->rt_bandwidth);
8462 for_each_possible_cpu(i) {
8464 kfree(tg->rt_rq[i]);
8466 kfree(tg->rt_se[i]);
8474 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8476 struct rt_rq *rt_rq;
8477 struct sched_rt_entity *rt_se, *parent_se;
8481 tg->rt_rq = kzalloc(sizeof(rt_rq) * nr_cpu_ids, GFP_KERNEL);
8484 tg->rt_se = kzalloc(sizeof(rt_se) * nr_cpu_ids, GFP_KERNEL);
8488 init_rt_bandwidth(&tg->rt_bandwidth,
8489 ktime_to_ns(def_rt_bandwidth.rt_period), 0);
8491 for_each_possible_cpu(i) {
8494 rt_rq = kmalloc_node(sizeof(struct rt_rq),
8495 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8499 rt_se = kmalloc_node(sizeof(struct sched_rt_entity),
8500 GFP_KERNEL|__GFP_ZERO, cpu_to_node(i));
8504 parent_se = parent ? parent->rt_se[i] : NULL;
8505 init_tg_rt_entry(tg, rt_rq, rt_se, i, 0, parent_se);
8514 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8516 list_add_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list,
8517 &cpu_rq(cpu)->leaf_rt_rq_list);
8520 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8522 list_del_rcu(&tg->rt_rq[cpu]->leaf_rt_rq_list);
8525 static inline void free_rt_sched_group(struct task_group *tg)
8530 int alloc_rt_sched_group(struct task_group *tg, struct task_group *parent)
8535 static inline void register_rt_sched_group(struct task_group *tg, int cpu)
8539 static inline void unregister_rt_sched_group(struct task_group *tg, int cpu)
8544 #ifdef CONFIG_GROUP_SCHED
8545 static void free_sched_group(struct task_group *tg)
8547 free_fair_sched_group(tg);
8548 free_rt_sched_group(tg);
8552 /* allocate runqueue etc for a new task group */
8553 struct task_group *sched_create_group(struct task_group *parent)
8555 struct task_group *tg;
8556 unsigned long flags;
8559 tg = kzalloc(sizeof(*tg), GFP_KERNEL);
8561 return ERR_PTR(-ENOMEM);
8563 if (!alloc_fair_sched_group(tg, parent))
8566 if (!alloc_rt_sched_group(tg, parent))
8569 spin_lock_irqsave(&task_group_lock, flags);
8570 for_each_possible_cpu(i) {
8571 register_fair_sched_group(tg, i);
8572 register_rt_sched_group(tg, i);
8574 list_add_rcu(&tg->list, &task_groups);
8576 WARN_ON(!parent); /* root should already exist */
8578 tg->parent = parent;
8579 list_add_rcu(&tg->siblings, &parent->children);
8580 INIT_LIST_HEAD(&tg->children);
8581 spin_unlock_irqrestore(&task_group_lock, flags);
8586 free_sched_group(tg);
8587 return ERR_PTR(-ENOMEM);
8590 /* rcu callback to free various structures associated with a task group */
8591 static void free_sched_group_rcu(struct rcu_head *rhp)
8593 /* now it should be safe to free those cfs_rqs */
8594 free_sched_group(container_of(rhp, struct task_group, rcu));
8597 /* Destroy runqueue etc associated with a task group */
8598 void sched_destroy_group(struct task_group *tg)
8600 unsigned long flags;
8603 spin_lock_irqsave(&task_group_lock, flags);
8604 for_each_possible_cpu(i) {
8605 unregister_fair_sched_group(tg, i);
8606 unregister_rt_sched_group(tg, i);
8608 list_del_rcu(&tg->list);
8609 list_del_rcu(&tg->siblings);
8610 spin_unlock_irqrestore(&task_group_lock, flags);
8612 /* wait for possible concurrent references to cfs_rqs complete */
8613 call_rcu(&tg->rcu, free_sched_group_rcu);
8616 /* change task's runqueue when it moves between groups.
8617 * The caller of this function should have put the task in its new group
8618 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8619 * reflect its new group.
8621 void sched_move_task(struct task_struct *tsk)
8624 unsigned long flags;
8627 rq = task_rq_lock(tsk, &flags);
8629 update_rq_clock(rq);
8631 running = task_current(rq, tsk);
8632 on_rq = tsk->se.on_rq;
8635 dequeue_task(rq, tsk, 0);
8636 if (unlikely(running))
8637 tsk->sched_class->put_prev_task(rq, tsk);
8639 set_task_rq(tsk, task_cpu(tsk));
8641 #ifdef CONFIG_FAIR_GROUP_SCHED
8642 if (tsk->sched_class->moved_group)
8643 tsk->sched_class->moved_group(tsk);
8646 if (unlikely(running))
8647 tsk->sched_class->set_curr_task(rq);
8649 enqueue_task(rq, tsk, 0);
8651 task_rq_unlock(rq, &flags);
8655 #ifdef CONFIG_FAIR_GROUP_SCHED
8656 static void __set_se_shares(struct sched_entity *se, unsigned long shares)
8658 struct cfs_rq *cfs_rq = se->cfs_rq;
8663 dequeue_entity(cfs_rq, se, 0);
8665 se->load.weight = shares;
8666 se->load.inv_weight = 0;
8669 enqueue_entity(cfs_rq, se, 0);
8672 static void set_se_shares(struct sched_entity *se, unsigned long shares)
8674 struct cfs_rq *cfs_rq = se->cfs_rq;
8675 struct rq *rq = cfs_rq->rq;
8676 unsigned long flags;
8678 spin_lock_irqsave(&rq->lock, flags);
8679 __set_se_shares(se, shares);
8680 spin_unlock_irqrestore(&rq->lock, flags);
8683 static DEFINE_MUTEX(shares_mutex);
8685 int sched_group_set_shares(struct task_group *tg, unsigned long shares)
8688 unsigned long flags;
8691 * We can't change the weight of the root cgroup.
8696 if (shares < MIN_SHARES)
8697 shares = MIN_SHARES;
8698 else if (shares > MAX_SHARES)
8699 shares = MAX_SHARES;
8701 mutex_lock(&shares_mutex);
8702 if (tg->shares == shares)
8705 spin_lock_irqsave(&task_group_lock, flags);
8706 for_each_possible_cpu(i)
8707 unregister_fair_sched_group(tg, i);
8708 list_del_rcu(&tg->siblings);
8709 spin_unlock_irqrestore(&task_group_lock, flags);
8711 /* wait for any ongoing reference to this group to finish */
8712 synchronize_sched();
8715 * Now we are free to modify the group's share on each cpu
8716 * w/o tripping rebalance_share or load_balance_fair.
8718 tg->shares = shares;
8719 for_each_possible_cpu(i) {
8723 cfs_rq_set_shares(tg->cfs_rq[i], 0);
8724 set_se_shares(tg->se[i], shares);
8728 * Enable load balance activity on this group, by inserting it back on
8729 * each cpu's rq->leaf_cfs_rq_list.
8731 spin_lock_irqsave(&task_group_lock, flags);
8732 for_each_possible_cpu(i)
8733 register_fair_sched_group(tg, i);
8734 list_add_rcu(&tg->siblings, &tg->parent->children);
8735 spin_unlock_irqrestore(&task_group_lock, flags);
8737 mutex_unlock(&shares_mutex);
8741 unsigned long sched_group_shares(struct task_group *tg)
8747 #ifdef CONFIG_RT_GROUP_SCHED
8749 * Ensure that the real time constraints are schedulable.
8751 static DEFINE_MUTEX(rt_constraints_mutex);
8753 static unsigned long to_ratio(u64 period, u64 runtime)
8755 if (runtime == RUNTIME_INF)
8758 return div64_u64(runtime << 16, period);
8761 #ifdef CONFIG_CGROUP_SCHED
8762 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8764 struct task_group *tgi, *parent = tg->parent;
8765 unsigned long total = 0;
8768 if (global_rt_period() < period)
8771 return to_ratio(period, runtime) <
8772 to_ratio(global_rt_period(), global_rt_runtime());
8775 if (ktime_to_ns(parent->rt_bandwidth.rt_period) < period)
8779 list_for_each_entry_rcu(tgi, &parent->children, siblings) {
8783 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8784 tgi->rt_bandwidth.rt_runtime);
8788 return total + to_ratio(period, runtime) <
8789 to_ratio(ktime_to_ns(parent->rt_bandwidth.rt_period),
8790 parent->rt_bandwidth.rt_runtime);
8792 #elif defined CONFIG_USER_SCHED
8793 static int __rt_schedulable(struct task_group *tg, u64 period, u64 runtime)
8795 struct task_group *tgi;
8796 unsigned long total = 0;
8797 unsigned long global_ratio =
8798 to_ratio(global_rt_period(), global_rt_runtime());
8801 list_for_each_entry_rcu(tgi, &task_groups, list) {
8805 total += to_ratio(ktime_to_ns(tgi->rt_bandwidth.rt_period),
8806 tgi->rt_bandwidth.rt_runtime);
8810 return total + to_ratio(period, runtime) < global_ratio;
8814 /* Must be called with tasklist_lock held */
8815 static inline int tg_has_rt_tasks(struct task_group *tg)
8817 struct task_struct *g, *p;
8818 do_each_thread(g, p) {
8819 if (rt_task(p) && rt_rq_of_se(&p->rt)->tg == tg)
8821 } while_each_thread(g, p);
8825 static int tg_set_bandwidth(struct task_group *tg,
8826 u64 rt_period, u64 rt_runtime)
8830 mutex_lock(&rt_constraints_mutex);
8831 read_lock(&tasklist_lock);
8832 if (rt_runtime == 0 && tg_has_rt_tasks(tg)) {
8836 if (!__rt_schedulable(tg, rt_period, rt_runtime)) {
8841 spin_lock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8842 tg->rt_bandwidth.rt_period = ns_to_ktime(rt_period);
8843 tg->rt_bandwidth.rt_runtime = rt_runtime;
8845 for_each_possible_cpu(i) {
8846 struct rt_rq *rt_rq = tg->rt_rq[i];
8848 spin_lock(&rt_rq->rt_runtime_lock);
8849 rt_rq->rt_runtime = rt_runtime;
8850 spin_unlock(&rt_rq->rt_runtime_lock);
8852 spin_unlock_irq(&tg->rt_bandwidth.rt_runtime_lock);
8854 read_unlock(&tasklist_lock);
8855 mutex_unlock(&rt_constraints_mutex);
8860 int sched_group_set_rt_runtime(struct task_group *tg, long rt_runtime_us)
8862 u64 rt_runtime, rt_period;
8864 rt_period = ktime_to_ns(tg->rt_bandwidth.rt_period);
8865 rt_runtime = (u64)rt_runtime_us * NSEC_PER_USEC;
8866 if (rt_runtime_us < 0)
8867 rt_runtime = RUNTIME_INF;
8869 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8872 long sched_group_rt_runtime(struct task_group *tg)
8876 if (tg->rt_bandwidth.rt_runtime == RUNTIME_INF)
8879 rt_runtime_us = tg->rt_bandwidth.rt_runtime;
8880 do_div(rt_runtime_us, NSEC_PER_USEC);
8881 return rt_runtime_us;
8884 int sched_group_set_rt_period(struct task_group *tg, long rt_period_us)
8886 u64 rt_runtime, rt_period;
8888 rt_period = (u64)rt_period_us * NSEC_PER_USEC;
8889 rt_runtime = tg->rt_bandwidth.rt_runtime;
8891 return tg_set_bandwidth(tg, rt_period, rt_runtime);
8894 long sched_group_rt_period(struct task_group *tg)
8898 rt_period_us = ktime_to_ns(tg->rt_bandwidth.rt_period);
8899 do_div(rt_period_us, NSEC_PER_USEC);
8900 return rt_period_us;
8903 static int sched_rt_global_constraints(void)
8907 mutex_lock(&rt_constraints_mutex);
8908 if (!__rt_schedulable(NULL, 1, 0))
8910 mutex_unlock(&rt_constraints_mutex);
8915 static int sched_rt_global_constraints(void)
8917 unsigned long flags;
8920 spin_lock_irqsave(&def_rt_bandwidth.rt_runtime_lock, flags);
8921 for_each_possible_cpu(i) {
8922 struct rt_rq *rt_rq = &cpu_rq(i)->rt;
8924 spin_lock(&rt_rq->rt_runtime_lock);
8925 rt_rq->rt_runtime = global_rt_runtime();
8926 spin_unlock(&rt_rq->rt_runtime_lock);
8928 spin_unlock_irqrestore(&def_rt_bandwidth.rt_runtime_lock, flags);
8934 int sched_rt_handler(struct ctl_table *table, int write,
8935 struct file *filp, void __user *buffer, size_t *lenp,
8939 int old_period, old_runtime;
8940 static DEFINE_MUTEX(mutex);
8943 old_period = sysctl_sched_rt_period;
8944 old_runtime = sysctl_sched_rt_runtime;
8946 ret = proc_dointvec(table, write, filp, buffer, lenp, ppos);
8948 if (!ret && write) {
8949 ret = sched_rt_global_constraints();
8951 sysctl_sched_rt_period = old_period;
8952 sysctl_sched_rt_runtime = old_runtime;
8954 def_rt_bandwidth.rt_runtime = global_rt_runtime();
8955 def_rt_bandwidth.rt_period =
8956 ns_to_ktime(global_rt_period());
8959 mutex_unlock(&mutex);
8964 #ifdef CONFIG_CGROUP_SCHED
8966 /* return corresponding task_group object of a cgroup */
8967 static inline struct task_group *cgroup_tg(struct cgroup *cgrp)
8969 return container_of(cgroup_subsys_state(cgrp, cpu_cgroup_subsys_id),
8970 struct task_group, css);
8973 static struct cgroup_subsys_state *
8974 cpu_cgroup_create(struct cgroup_subsys *ss, struct cgroup *cgrp)
8976 struct task_group *tg, *parent;
8978 if (!cgrp->parent) {
8979 /* This is early initialization for the top cgroup */
8980 init_task_group.css.cgroup = cgrp;
8981 return &init_task_group.css;
8984 parent = cgroup_tg(cgrp->parent);
8985 tg = sched_create_group(parent);
8987 return ERR_PTR(-ENOMEM);
8989 /* Bind the cgroup to task_group object we just created */
8990 tg->css.cgroup = cgrp;
8996 cpu_cgroup_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
8998 struct task_group *tg = cgroup_tg(cgrp);
9000 sched_destroy_group(tg);
9004 cpu_cgroup_can_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9005 struct task_struct *tsk)
9007 #ifdef CONFIG_RT_GROUP_SCHED
9008 /* Don't accept realtime tasks when there is no way for them to run */
9009 if (rt_task(tsk) && cgroup_tg(cgrp)->rt_bandwidth.rt_runtime == 0)
9012 /* We don't support RT-tasks being in separate groups */
9013 if (tsk->sched_class != &fair_sched_class)
9021 cpu_cgroup_attach(struct cgroup_subsys *ss, struct cgroup *cgrp,
9022 struct cgroup *old_cont, struct task_struct *tsk)
9024 sched_move_task(tsk);
9027 #ifdef CONFIG_FAIR_GROUP_SCHED
9028 static int cpu_shares_write_u64(struct cgroup *cgrp, struct cftype *cftype,
9031 return sched_group_set_shares(cgroup_tg(cgrp), shareval);
9034 static u64 cpu_shares_read_u64(struct cgroup *cgrp, struct cftype *cft)
9036 struct task_group *tg = cgroup_tg(cgrp);
9038 return (u64) tg->shares;
9042 #ifdef CONFIG_RT_GROUP_SCHED
9043 static int cpu_rt_runtime_write(struct cgroup *cgrp, struct cftype *cft,
9046 return sched_group_set_rt_runtime(cgroup_tg(cgrp), val);
9049 static s64 cpu_rt_runtime_read(struct cgroup *cgrp, struct cftype *cft)
9051 return sched_group_rt_runtime(cgroup_tg(cgrp));
9054 static int cpu_rt_period_write_uint(struct cgroup *cgrp, struct cftype *cftype,
9057 return sched_group_set_rt_period(cgroup_tg(cgrp), rt_period_us);
9060 static u64 cpu_rt_period_read_uint(struct cgroup *cgrp, struct cftype *cft)
9062 return sched_group_rt_period(cgroup_tg(cgrp));
9066 static struct cftype cpu_files[] = {
9067 #ifdef CONFIG_FAIR_GROUP_SCHED
9070 .read_u64 = cpu_shares_read_u64,
9071 .write_u64 = cpu_shares_write_u64,
9074 #ifdef CONFIG_RT_GROUP_SCHED
9076 .name = "rt_runtime_us",
9077 .read_s64 = cpu_rt_runtime_read,
9078 .write_s64 = cpu_rt_runtime_write,
9081 .name = "rt_period_us",
9082 .read_u64 = cpu_rt_period_read_uint,
9083 .write_u64 = cpu_rt_period_write_uint,
9088 static int cpu_cgroup_populate(struct cgroup_subsys *ss, struct cgroup *cont)
9090 return cgroup_add_files(cont, ss, cpu_files, ARRAY_SIZE(cpu_files));
9093 struct cgroup_subsys cpu_cgroup_subsys = {
9095 .create = cpu_cgroup_create,
9096 .destroy = cpu_cgroup_destroy,
9097 .can_attach = cpu_cgroup_can_attach,
9098 .attach = cpu_cgroup_attach,
9099 .populate = cpu_cgroup_populate,
9100 .subsys_id = cpu_cgroup_subsys_id,
9104 #endif /* CONFIG_CGROUP_SCHED */
9106 #ifdef CONFIG_CGROUP_CPUACCT
9109 * CPU accounting code for task groups.
9111 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9112 * (balbir@in.ibm.com).
9115 /* track cpu usage of a group of tasks */
9117 struct cgroup_subsys_state css;
9118 /* cpuusage holds pointer to a u64-type object on every cpu */
9122 struct cgroup_subsys cpuacct_subsys;
9124 /* return cpu accounting group corresponding to this container */
9125 static inline struct cpuacct *cgroup_ca(struct cgroup *cgrp)
9127 return container_of(cgroup_subsys_state(cgrp, cpuacct_subsys_id),
9128 struct cpuacct, css);
9131 /* return cpu accounting group to which this task belongs */
9132 static inline struct cpuacct *task_ca(struct task_struct *tsk)
9134 return container_of(task_subsys_state(tsk, cpuacct_subsys_id),
9135 struct cpuacct, css);
9138 /* create a new cpu accounting group */
9139 static struct cgroup_subsys_state *cpuacct_create(
9140 struct cgroup_subsys *ss, struct cgroup *cgrp)
9142 struct cpuacct *ca = kzalloc(sizeof(*ca), GFP_KERNEL);
9145 return ERR_PTR(-ENOMEM);
9147 ca->cpuusage = alloc_percpu(u64);
9148 if (!ca->cpuusage) {
9150 return ERR_PTR(-ENOMEM);
9156 /* destroy an existing cpu accounting group */
9158 cpuacct_destroy(struct cgroup_subsys *ss, struct cgroup *cgrp)
9160 struct cpuacct *ca = cgroup_ca(cgrp);
9162 free_percpu(ca->cpuusage);
9166 /* return total cpu usage (in nanoseconds) of a group */
9167 static u64 cpuusage_read(struct cgroup *cgrp, struct cftype *cft)
9169 struct cpuacct *ca = cgroup_ca(cgrp);
9170 u64 totalcpuusage = 0;
9173 for_each_possible_cpu(i) {
9174 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9177 * Take rq->lock to make 64-bit addition safe on 32-bit
9180 spin_lock_irq(&cpu_rq(i)->lock);
9181 totalcpuusage += *cpuusage;
9182 spin_unlock_irq(&cpu_rq(i)->lock);
9185 return totalcpuusage;
9188 static int cpuusage_write(struct cgroup *cgrp, struct cftype *cftype,
9191 struct cpuacct *ca = cgroup_ca(cgrp);
9200 for_each_possible_cpu(i) {
9201 u64 *cpuusage = percpu_ptr(ca->cpuusage, i);
9203 spin_lock_irq(&cpu_rq(i)->lock);
9205 spin_unlock_irq(&cpu_rq(i)->lock);
9211 static struct cftype files[] = {
9214 .read_u64 = cpuusage_read,
9215 .write_u64 = cpuusage_write,
9219 static int cpuacct_populate(struct cgroup_subsys *ss, struct cgroup *cgrp)
9221 return cgroup_add_files(cgrp, ss, files, ARRAY_SIZE(files));
9225 * charge this task's execution time to its accounting group.
9227 * called with rq->lock held.
9229 static void cpuacct_charge(struct task_struct *tsk, u64 cputime)
9233 if (!cpuacct_subsys.active)
9238 u64 *cpuusage = percpu_ptr(ca->cpuusage, task_cpu(tsk));
9240 *cpuusage += cputime;
9244 struct cgroup_subsys cpuacct_subsys = {
9246 .create = cpuacct_create,
9247 .destroy = cpuacct_destroy,
9248 .populate = cpuacct_populate,
9249 .subsys_id = cpuacct_subsys_id,
9251 #endif /* CONFIG_CGROUP_CPUACCT */