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/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy
)
125 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
130 static inline int task_has_rt_policy(struct task_struct
*p
)
132 return rt_policy(p
->policy
);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array
{
139 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
140 struct list_head queue
[MAX_RT_PRIO
];
143 struct rt_bandwidth
{
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock
;
148 struct hrtimer rt_period_timer
;
151 static struct rt_bandwidth def_rt_bandwidth
;
153 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
155 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
157 struct rt_bandwidth
*rt_b
=
158 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
164 now
= hrtimer_cb_get_time(timer
);
165 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
170 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
173 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
177 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
179 rt_b
->rt_period
= ns_to_ktime(period
);
180 rt_b
->rt_runtime
= runtime
;
182 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
184 hrtimer_init(&rt_b
->rt_period_timer
,
185 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
186 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime
>= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
198 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
201 if (hrtimer_active(&rt_b
->rt_period_timer
))
204 raw_spin_lock(&rt_b
->rt_runtime_lock
);
209 if (hrtimer_active(&rt_b
->rt_period_timer
))
212 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
213 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
215 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
216 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
217 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
218 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
219 HRTIMER_MODE_ABS_PINNED
, 0);
221 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
227 hrtimer_cancel(&rt_b
->rt_period_timer
);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex
);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups
);
245 /* task group related information */
247 struct cgroup_subsys_state css
;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity
**se
;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq
**cfs_rq
;
254 unsigned long shares
;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity
**rt_se
;
259 struct rt_rq
**rt_rq
;
261 struct rt_bandwidth rt_bandwidth
;
265 struct list_head list
;
267 struct task_group
*parent
;
268 struct list_head siblings
;
269 struct list_head children
;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock
);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group
.children
);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group
;
309 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load
;
314 unsigned long nr_running
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
322 struct list_head tasks
;
323 struct list_head
*balance_iterator
;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity
*curr
, *next
, *last
;
331 unsigned int nr_spread_over
;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list
;
345 struct task_group
*tg
; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight
;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load
;
362 * this cpu's part of tg->shares
364 unsigned long shares
;
367 * load.weight at the time we set shares
369 unsigned long rq_weight
;
374 /* Real-Time classes' related field in a runqueue: */
376 struct rt_prio_array active
;
377 unsigned long rt_nr_running
;
378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
380 int curr
; /* highest queued rt task prio */
382 int next
; /* next highest */
387 unsigned long rt_nr_migratory
;
388 unsigned long rt_nr_total
;
390 struct plist_head pushable_tasks
;
395 /* Nests inside the rq lock: */
396 raw_spinlock_t rt_runtime_lock
;
398 #ifdef CONFIG_RT_GROUP_SCHED
399 unsigned long rt_nr_boosted
;
402 struct list_head leaf_rt_rq_list
;
403 struct task_group
*tg
;
410 * We add the notion of a root-domain which will be used to define per-domain
411 * variables. Each exclusive cpuset essentially defines an island domain by
412 * fully partitioning the member cpus from any other cpuset. Whenever a new
413 * exclusive cpuset is created, we also create and attach a new root-domain
420 cpumask_var_t online
;
423 * The "RT overload" flag: it gets set if a CPU has more than
424 * one runnable RT task.
426 cpumask_var_t rto_mask
;
429 struct cpupri cpupri
;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain
;
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running
;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
461 unsigned char in_nohz_recently
;
463 unsigned int skip_clock_update
;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load
;
467 unsigned long nr_load_updates
;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list
;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list
;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible
;
489 struct task_struct
*curr
, *idle
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
498 struct root_domain
*rd
;
499 struct sched_domain
*sd
;
501 unsigned long cpu_power
;
503 unsigned char idle_at_tick
;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work
;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task
;
521 /* calc_load related fields */
522 unsigned long calc_load_update
;
523 long calc_load_active
;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending
;
528 struct call_single_data hrtick_csd
;
530 struct hrtimer hrtick_timer
;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info
;
536 unsigned long long rq_cpu_time
;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count
;
542 /* schedule() stats */
543 unsigned int sched_switch
;
544 unsigned int sched_count
;
545 unsigned int sched_goidle
;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count
;
549 unsigned int ttwu_local
;
552 unsigned int bkl_count
;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
559 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
561 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (rq
->curr
->se
.on_rq
&& test_tsk_need_resched(rq
->curr
))
568 rq
->skip_clock_update
= 1;
571 static inline int cpu_of(struct rq
*rq
)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group
*task_group(struct task_struct
*p
)
613 struct cgroup_subsys_state
*css
;
615 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
616 lockdep_is_held(&task_rq(p
)->lock
));
617 return container_of(css
, struct task_group
, css
);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
625 p
->se
.parent
= task_group(p
)->se
[cpu
];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
630 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
637 static inline struct task_group
*task_group(struct task_struct
*p
)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq
*rq
)
646 if (!rq
->skip_clock_update
)
647 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu
)
669 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug
unsigned int sysctl_sched_features
=
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly
char *sched_feat_names
[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file
*m
, void *v
)
709 for (i
= 0; sched_feat_names
[i
]; i
++) {
710 if (!(sysctl_sched_features
& (1UL << i
)))
712 seq_printf(m
, "%s ", sched_feat_names
[i
]);
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
))
737 if (strncmp(buf
, "NO_", 3) == 0) {
742 for (i
= 0; sched_feat_names
[i
]; i
++) {
743 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
745 sysctl_sched_features
&= ~(1UL << i
);
747 sysctl_sched_features
|= (1UL << i
);
752 if (!sched_feat_names
[i
])
760 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
762 return single_open(filp
, sched_feat_show
, NULL
);
765 static const struct file_operations sched_feat_fops
= {
766 .open
= sched_feat_open
,
767 .write
= sched_feat_write
,
770 .release
= single_release
,
773 static __init
int sched_init_debug(void)
775 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
780 late_initcall(sched_init_debug
);
784 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
787 * Number of tasks to iterate in a single balance run.
788 * Limited because this is done with IRQs disabled.
790 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
793 * ratelimit for updating the group shares.
796 unsigned int sysctl_sched_shares_ratelimit
= 250000;
797 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
800 * Inject some fuzzyness into changing the per-cpu group shares
801 * this avoids remote rq-locks at the expense of fairness.
804 unsigned int sysctl_sched_shares_thresh
= 4;
807 * period over which we average the RT time consumption, measured
812 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
815 * period over which we measure -rt task cpu usage in us.
818 unsigned int sysctl_sched_rt_period
= 1000000;
820 static __read_mostly
int scheduler_running
;
823 * part of the period that we allow rt tasks to run in us.
826 int sysctl_sched_rt_runtime
= 950000;
828 static inline u64
global_rt_period(void)
830 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
833 static inline u64
global_rt_runtime(void)
835 if (sysctl_sched_rt_runtime
< 0)
838 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
841 #ifndef prepare_arch_switch
842 # define prepare_arch_switch(next) do { } while (0)
844 #ifndef finish_arch_switch
845 # define finish_arch_switch(prev) do { } while (0)
848 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
850 return rq
->curr
== p
;
853 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
854 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
856 return task_current(rq
, p
);
859 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
863 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
865 #ifdef CONFIG_DEBUG_SPINLOCK
866 /* this is a valid case when another task releases the spinlock */
867 rq
->lock
.owner
= current
;
870 * If we are tracking spinlock dependencies then we have to
871 * fix up the runqueue lock - which gets 'carried over' from
874 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
876 raw_spin_unlock_irq(&rq
->lock
);
879 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
880 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
885 return task_current(rq
, p
);
889 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
893 * We can optimise this out completely for !SMP, because the
894 * SMP rebalancing from interrupt is the only thing that cares
899 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
900 raw_spin_unlock_irq(&rq
->lock
);
902 raw_spin_unlock(&rq
->lock
);
906 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
910 * After ->oncpu is cleared, the task can be moved to a different CPU.
911 * We must ensure this doesn't happen until the switch is completely
917 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
921 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
924 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
927 static inline int task_is_waking(struct task_struct
*p
)
929 return unlikely(p
->state
== TASK_WAKING
);
933 * __task_rq_lock - lock the runqueue a given task resides on.
934 * Must be called interrupts disabled.
936 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
943 raw_spin_lock(&rq
->lock
);
944 if (likely(rq
== task_rq(p
)))
946 raw_spin_unlock(&rq
->lock
);
951 * task_rq_lock - lock the runqueue a given task resides on and disable
952 * interrupts. Note the ordering: we can safely lookup the task_rq without
953 * explicitly disabling preemption.
955 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
961 local_irq_save(*flags
);
963 raw_spin_lock(&rq
->lock
);
964 if (likely(rq
== task_rq(p
)))
966 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
970 static void __task_rq_unlock(struct rq
*rq
)
973 raw_spin_unlock(&rq
->lock
);
976 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
979 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
983 * this_rq_lock - lock this runqueue and disable interrupts.
985 static struct rq
*this_rq_lock(void)
992 raw_spin_lock(&rq
->lock
);
997 #ifdef CONFIG_SCHED_HRTICK
999 * Use HR-timers to deliver accurate preemption points.
1001 * Its all a bit involved since we cannot program an hrt while holding the
1002 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1005 * When we get rescheduled we reprogram the hrtick_timer outside of the
1011 * - enabled by features
1012 * - hrtimer is actually high res
1014 static inline int hrtick_enabled(struct rq
*rq
)
1016 if (!sched_feat(HRTICK
))
1018 if (!cpu_active(cpu_of(rq
)))
1020 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1023 static void hrtick_clear(struct rq
*rq
)
1025 if (hrtimer_active(&rq
->hrtick_timer
))
1026 hrtimer_cancel(&rq
->hrtick_timer
);
1030 * High-resolution timer tick.
1031 * Runs from hardirq context with interrupts disabled.
1033 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1035 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1037 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1039 raw_spin_lock(&rq
->lock
);
1040 update_rq_clock(rq
);
1041 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1042 raw_spin_unlock(&rq
->lock
);
1044 return HRTIMER_NORESTART
;
1049 * called from hardirq (IPI) context
1051 static void __hrtick_start(void *arg
)
1053 struct rq
*rq
= arg
;
1055 raw_spin_lock(&rq
->lock
);
1056 hrtimer_restart(&rq
->hrtick_timer
);
1057 rq
->hrtick_csd_pending
= 0;
1058 raw_spin_unlock(&rq
->lock
);
1062 * Called to set the hrtick timer state.
1064 * called with rq->lock held and irqs disabled
1066 static void hrtick_start(struct rq
*rq
, u64 delay
)
1068 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1069 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1071 hrtimer_set_expires(timer
, time
);
1073 if (rq
== this_rq()) {
1074 hrtimer_restart(timer
);
1075 } else if (!rq
->hrtick_csd_pending
) {
1076 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1077 rq
->hrtick_csd_pending
= 1;
1082 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1084 int cpu
= (int)(long)hcpu
;
1087 case CPU_UP_CANCELED
:
1088 case CPU_UP_CANCELED_FROZEN
:
1089 case CPU_DOWN_PREPARE
:
1090 case CPU_DOWN_PREPARE_FROZEN
:
1092 case CPU_DEAD_FROZEN
:
1093 hrtick_clear(cpu_rq(cpu
));
1100 static __init
void init_hrtick(void)
1102 hotcpu_notifier(hotplug_hrtick
, 0);
1106 * Called to set the hrtick timer state.
1108 * called with rq->lock held and irqs disabled
1110 static void hrtick_start(struct rq
*rq
, u64 delay
)
1112 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1113 HRTIMER_MODE_REL_PINNED
, 0);
1116 static inline void init_hrtick(void)
1119 #endif /* CONFIG_SMP */
1121 static void init_rq_hrtick(struct rq
*rq
)
1124 rq
->hrtick_csd_pending
= 0;
1126 rq
->hrtick_csd
.flags
= 0;
1127 rq
->hrtick_csd
.func
= __hrtick_start
;
1128 rq
->hrtick_csd
.info
= rq
;
1131 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1132 rq
->hrtick_timer
.function
= hrtick
;
1134 #else /* CONFIG_SCHED_HRTICK */
1135 static inline void hrtick_clear(struct rq
*rq
)
1139 static inline void init_rq_hrtick(struct rq
*rq
)
1143 static inline void init_hrtick(void)
1146 #endif /* CONFIG_SCHED_HRTICK */
1149 * resched_task - mark a task 'to be rescheduled now'.
1151 * On UP this means the setting of the need_resched flag, on SMP it
1152 * might also involve a cross-CPU call to trigger the scheduler on
1157 #ifndef tsk_is_polling
1158 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1161 static void resched_task(struct task_struct
*p
)
1165 assert_raw_spin_locked(&task_rq(p
)->lock
);
1167 if (test_tsk_need_resched(p
))
1170 set_tsk_need_resched(p
);
1173 if (cpu
== smp_processor_id())
1176 /* NEED_RESCHED must be visible before we test polling */
1178 if (!tsk_is_polling(p
))
1179 smp_send_reschedule(cpu
);
1182 static void resched_cpu(int cpu
)
1184 struct rq
*rq
= cpu_rq(cpu
);
1185 unsigned long flags
;
1187 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1189 resched_task(cpu_curr(cpu
));
1190 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1195 * When add_timer_on() enqueues a timer into the timer wheel of an
1196 * idle CPU then this timer might expire before the next timer event
1197 * which is scheduled to wake up that CPU. In case of a completely
1198 * idle system the next event might even be infinite time into the
1199 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1200 * leaves the inner idle loop so the newly added timer is taken into
1201 * account when the CPU goes back to idle and evaluates the timer
1202 * wheel for the next timer event.
1204 void wake_up_idle_cpu(int cpu
)
1206 struct rq
*rq
= cpu_rq(cpu
);
1208 if (cpu
== smp_processor_id())
1212 * This is safe, as this function is called with the timer
1213 * wheel base lock of (cpu) held. When the CPU is on the way
1214 * to idle and has not yet set rq->curr to idle then it will
1215 * be serialized on the timer wheel base lock and take the new
1216 * timer into account automatically.
1218 if (rq
->curr
!= rq
->idle
)
1222 * We can set TIF_RESCHED on the idle task of the other CPU
1223 * lockless. The worst case is that the other CPU runs the
1224 * idle task through an additional NOOP schedule()
1226 set_tsk_need_resched(rq
->idle
);
1228 /* NEED_RESCHED must be visible before we test polling */
1230 if (!tsk_is_polling(rq
->idle
))
1231 smp_send_reschedule(cpu
);
1234 #endif /* CONFIG_NO_HZ */
1236 static u64
sched_avg_period(void)
1238 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1241 static void sched_avg_update(struct rq
*rq
)
1243 s64 period
= sched_avg_period();
1245 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1247 * Inline assembly required to prevent the compiler
1248 * optimising this loop into a divmod call.
1249 * See __iter_div_u64_rem() for another example of this.
1251 asm("" : "+rm" (rq
->age_stamp
));
1252 rq
->age_stamp
+= period
;
1257 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1259 rq
->rt_avg
+= rt_delta
;
1260 sched_avg_update(rq
);
1263 #else /* !CONFIG_SMP */
1264 static void resched_task(struct task_struct
*p
)
1266 assert_raw_spin_locked(&task_rq(p
)->lock
);
1267 set_tsk_need_resched(p
);
1270 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1273 #endif /* CONFIG_SMP */
1275 #if BITS_PER_LONG == 32
1276 # define WMULT_CONST (~0UL)
1278 # define WMULT_CONST (1UL << 32)
1281 #define WMULT_SHIFT 32
1284 * Shift right and round:
1286 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1289 * delta *= weight / lw
1291 static unsigned long
1292 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1293 struct load_weight
*lw
)
1297 if (!lw
->inv_weight
) {
1298 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1301 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1305 tmp
= (u64
)delta_exec
* weight
;
1307 * Check whether we'd overflow the 64-bit multiplication:
1309 if (unlikely(tmp
> WMULT_CONST
))
1310 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1313 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1315 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1318 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1324 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1331 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1332 * of tasks with abnormal "nice" values across CPUs the contribution that
1333 * each task makes to its run queue's load is weighted according to its
1334 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1335 * scaled version of the new time slice allocation that they receive on time
1339 #define WEIGHT_IDLEPRIO 3
1340 #define WMULT_IDLEPRIO 1431655765
1343 * Nice levels are multiplicative, with a gentle 10% change for every
1344 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1345 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1346 * that remained on nice 0.
1348 * The "10% effect" is relative and cumulative: from _any_ nice level,
1349 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1350 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1351 * If a task goes up by ~10% and another task goes down by ~10% then
1352 * the relative distance between them is ~25%.)
1354 static const int prio_to_weight
[40] = {
1355 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1356 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1357 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1358 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1359 /* 0 */ 1024, 820, 655, 526, 423,
1360 /* 5 */ 335, 272, 215, 172, 137,
1361 /* 10 */ 110, 87, 70, 56, 45,
1362 /* 15 */ 36, 29, 23, 18, 15,
1366 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1368 * In cases where the weight does not change often, we can use the
1369 * precalculated inverse to speed up arithmetics by turning divisions
1370 * into multiplications:
1372 static const u32 prio_to_wmult
[40] = {
1373 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1374 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1375 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1376 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1377 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1378 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1379 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1380 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1383 /* Time spent by the tasks of the cpu accounting group executing in ... */
1384 enum cpuacct_stat_index
{
1385 CPUACCT_STAT_USER
, /* ... user mode */
1386 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1388 CPUACCT_STAT_NSTATS
,
1391 #ifdef CONFIG_CGROUP_CPUACCT
1392 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1393 static void cpuacct_update_stats(struct task_struct
*tsk
,
1394 enum cpuacct_stat_index idx
, cputime_t val
);
1396 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1397 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1398 enum cpuacct_stat_index idx
, cputime_t val
) {}
1401 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1403 update_load_add(&rq
->load
, load
);
1406 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1408 update_load_sub(&rq
->load
, load
);
1411 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1412 typedef int (*tg_visitor
)(struct task_group
*, void *);
1415 * Iterate the full tree, calling @down when first entering a node and @up when
1416 * leaving it for the final time.
1418 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1420 struct task_group
*parent
, *child
;
1424 parent
= &root_task_group
;
1426 ret
= (*down
)(parent
, data
);
1429 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1436 ret
= (*up
)(parent
, data
);
1441 parent
= parent
->parent
;
1450 static int tg_nop(struct task_group
*tg
, void *data
)
1457 /* Used instead of source_load when we know the type == 0 */
1458 static unsigned long weighted_cpuload(const int cpu
)
1460 return cpu_rq(cpu
)->load
.weight
;
1464 * Return a low guess at the load of a migration-source cpu weighted
1465 * according to the scheduling class and "nice" value.
1467 * We want to under-estimate the load of migration sources, to
1468 * balance conservatively.
1470 static unsigned long source_load(int cpu
, int type
)
1472 struct rq
*rq
= cpu_rq(cpu
);
1473 unsigned long total
= weighted_cpuload(cpu
);
1475 if (type
== 0 || !sched_feat(LB_BIAS
))
1478 return min(rq
->cpu_load
[type
-1], total
);
1482 * Return a high guess at the load of a migration-target cpu weighted
1483 * according to the scheduling class and "nice" value.
1485 static unsigned long target_load(int cpu
, int type
)
1487 struct rq
*rq
= cpu_rq(cpu
);
1488 unsigned long total
= weighted_cpuload(cpu
);
1490 if (type
== 0 || !sched_feat(LB_BIAS
))
1493 return max(rq
->cpu_load
[type
-1], total
);
1496 static unsigned long power_of(int cpu
)
1498 return cpu_rq(cpu
)->cpu_power
;
1501 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1503 static unsigned long cpu_avg_load_per_task(int cpu
)
1505 struct rq
*rq
= cpu_rq(cpu
);
1506 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1509 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1511 rq
->avg_load_per_task
= 0;
1513 return rq
->avg_load_per_task
;
1516 #ifdef CONFIG_FAIR_GROUP_SCHED
1518 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1520 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1523 * Calculate and set the cpu's group shares.
1525 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1526 unsigned long sd_shares
,
1527 unsigned long sd_rq_weight
,
1528 unsigned long *usd_rq_weight
)
1530 unsigned long shares
, rq_weight
;
1533 rq_weight
= usd_rq_weight
[cpu
];
1536 rq_weight
= NICE_0_LOAD
;
1540 * \Sum_j shares_j * rq_weight_i
1541 * shares_i = -----------------------------
1542 * \Sum_j rq_weight_j
1544 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1545 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1547 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1548 sysctl_sched_shares_thresh
) {
1549 struct rq
*rq
= cpu_rq(cpu
);
1550 unsigned long flags
;
1552 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1553 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1554 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1555 __set_se_shares(tg
->se
[cpu
], shares
);
1556 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1561 * Re-compute the task group their per cpu shares over the given domain.
1562 * This needs to be done in a bottom-up fashion because the rq weight of a
1563 * parent group depends on the shares of its child groups.
1565 static int tg_shares_up(struct task_group
*tg
, void *data
)
1567 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1568 unsigned long *usd_rq_weight
;
1569 struct sched_domain
*sd
= data
;
1570 unsigned long flags
;
1576 local_irq_save(flags
);
1577 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1579 for_each_cpu(i
, sched_domain_span(sd
)) {
1580 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1581 usd_rq_weight
[i
] = weight
;
1583 rq_weight
+= weight
;
1585 * If there are currently no tasks on the cpu pretend there
1586 * is one of average load so that when a new task gets to
1587 * run here it will not get delayed by group starvation.
1590 weight
= NICE_0_LOAD
;
1592 sum_weight
+= weight
;
1593 shares
+= tg
->cfs_rq
[i
]->shares
;
1597 rq_weight
= sum_weight
;
1599 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1600 shares
= tg
->shares
;
1602 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1603 shares
= tg
->shares
;
1605 for_each_cpu(i
, sched_domain_span(sd
))
1606 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1608 local_irq_restore(flags
);
1614 * Compute the cpu's hierarchical load factor for each task group.
1615 * This needs to be done in a top-down fashion because the load of a child
1616 * group is a fraction of its parents load.
1618 static int tg_load_down(struct task_group
*tg
, void *data
)
1621 long cpu
= (long)data
;
1624 load
= cpu_rq(cpu
)->load
.weight
;
1626 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1627 load
*= tg
->cfs_rq
[cpu
]->shares
;
1628 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1631 tg
->cfs_rq
[cpu
]->h_load
= load
;
1636 static void update_shares(struct sched_domain
*sd
)
1641 if (root_task_group_empty())
1644 now
= cpu_clock(raw_smp_processor_id());
1645 elapsed
= now
- sd
->last_update
;
1647 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1648 sd
->last_update
= now
;
1649 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1653 static void update_h_load(long cpu
)
1655 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1660 static inline void update_shares(struct sched_domain
*sd
)
1666 #ifdef CONFIG_PREEMPT
1668 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1671 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1672 * way at the expense of forcing extra atomic operations in all
1673 * invocations. This assures that the double_lock is acquired using the
1674 * same underlying policy as the spinlock_t on this architecture, which
1675 * reduces latency compared to the unfair variant below. However, it
1676 * also adds more overhead and therefore may reduce throughput.
1678 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1679 __releases(this_rq
->lock
)
1680 __acquires(busiest
->lock
)
1681 __acquires(this_rq
->lock
)
1683 raw_spin_unlock(&this_rq
->lock
);
1684 double_rq_lock(this_rq
, busiest
);
1691 * Unfair double_lock_balance: Optimizes throughput at the expense of
1692 * latency by eliminating extra atomic operations when the locks are
1693 * already in proper order on entry. This favors lower cpu-ids and will
1694 * grant the double lock to lower cpus over higher ids under contention,
1695 * regardless of entry order into the function.
1697 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1698 __releases(this_rq
->lock
)
1699 __acquires(busiest
->lock
)
1700 __acquires(this_rq
->lock
)
1704 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1705 if (busiest
< this_rq
) {
1706 raw_spin_unlock(&this_rq
->lock
);
1707 raw_spin_lock(&busiest
->lock
);
1708 raw_spin_lock_nested(&this_rq
->lock
,
1709 SINGLE_DEPTH_NESTING
);
1712 raw_spin_lock_nested(&busiest
->lock
,
1713 SINGLE_DEPTH_NESTING
);
1718 #endif /* CONFIG_PREEMPT */
1721 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1723 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1725 if (unlikely(!irqs_disabled())) {
1726 /* printk() doesn't work good under rq->lock */
1727 raw_spin_unlock(&this_rq
->lock
);
1731 return _double_lock_balance(this_rq
, busiest
);
1734 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1735 __releases(busiest
->lock
)
1737 raw_spin_unlock(&busiest
->lock
);
1738 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1742 * double_rq_lock - safely lock two runqueues
1744 * Note this does not disable interrupts like task_rq_lock,
1745 * you need to do so manually before calling.
1747 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1748 __acquires(rq1
->lock
)
1749 __acquires(rq2
->lock
)
1751 BUG_ON(!irqs_disabled());
1753 raw_spin_lock(&rq1
->lock
);
1754 __acquire(rq2
->lock
); /* Fake it out ;) */
1757 raw_spin_lock(&rq1
->lock
);
1758 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1760 raw_spin_lock(&rq2
->lock
);
1761 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1767 * double_rq_unlock - safely unlock two runqueues
1769 * Note this does not restore interrupts like task_rq_unlock,
1770 * you need to do so manually after calling.
1772 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1773 __releases(rq1
->lock
)
1774 __releases(rq2
->lock
)
1776 raw_spin_unlock(&rq1
->lock
);
1778 raw_spin_unlock(&rq2
->lock
);
1780 __release(rq2
->lock
);
1785 #ifdef CONFIG_FAIR_GROUP_SCHED
1786 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1789 cfs_rq
->shares
= shares
;
1794 static void calc_load_account_idle(struct rq
*this_rq
);
1795 static void update_sysctl(void);
1796 static int get_update_sysctl_factor(void);
1798 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1800 set_task_rq(p
, cpu
);
1803 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1804 * successfuly executed on another CPU. We must ensure that updates of
1805 * per-task data have been completed by this moment.
1808 task_thread_info(p
)->cpu
= cpu
;
1812 static const struct sched_class rt_sched_class
;
1814 #define sched_class_highest (&rt_sched_class)
1815 #define for_each_class(class) \
1816 for (class = sched_class_highest; class; class = class->next)
1818 #include "sched_stats.h"
1820 static void inc_nr_running(struct rq
*rq
)
1825 static void dec_nr_running(struct rq
*rq
)
1830 static void set_load_weight(struct task_struct
*p
)
1833 * SCHED_IDLE tasks get minimal weight:
1835 if (p
->policy
== SCHED_IDLE
) {
1836 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1837 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1841 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1842 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1845 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1847 update_rq_clock(rq
);
1848 sched_info_queued(p
);
1849 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1853 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1855 update_rq_clock(rq
);
1856 sched_info_dequeued(p
);
1857 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1862 * activate_task - move a task to the runqueue.
1864 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1866 if (task_contributes_to_load(p
))
1867 rq
->nr_uninterruptible
--;
1869 enqueue_task(rq
, p
, flags
);
1874 * deactivate_task - remove a task from the runqueue.
1876 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1878 if (task_contributes_to_load(p
))
1879 rq
->nr_uninterruptible
++;
1881 dequeue_task(rq
, p
, flags
);
1885 #include "sched_idletask.c"
1886 #include "sched_fair.c"
1887 #include "sched_rt.c"
1888 #ifdef CONFIG_SCHED_DEBUG
1889 # include "sched_debug.c"
1893 * __normal_prio - return the priority that is based on the static prio
1895 static inline int __normal_prio(struct task_struct
*p
)
1897 return p
->static_prio
;
1901 * Calculate the expected normal priority: i.e. priority
1902 * without taking RT-inheritance into account. Might be
1903 * boosted by interactivity modifiers. Changes upon fork,
1904 * setprio syscalls, and whenever the interactivity
1905 * estimator recalculates.
1907 static inline int normal_prio(struct task_struct
*p
)
1911 if (task_has_rt_policy(p
))
1912 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1914 prio
= __normal_prio(p
);
1919 * Calculate the current priority, i.e. the priority
1920 * taken into account by the scheduler. This value might
1921 * be boosted by RT tasks, or might be boosted by
1922 * interactivity modifiers. Will be RT if the task got
1923 * RT-boosted. If not then it returns p->normal_prio.
1925 static int effective_prio(struct task_struct
*p
)
1927 p
->normal_prio
= normal_prio(p
);
1929 * If we are RT tasks or we were boosted to RT priority,
1930 * keep the priority unchanged. Otherwise, update priority
1931 * to the normal priority:
1933 if (!rt_prio(p
->prio
))
1934 return p
->normal_prio
;
1939 * task_curr - is this task currently executing on a CPU?
1940 * @p: the task in question.
1942 inline int task_curr(const struct task_struct
*p
)
1944 return cpu_curr(task_cpu(p
)) == p
;
1947 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1948 const struct sched_class
*prev_class
,
1949 int oldprio
, int running
)
1951 if (prev_class
!= p
->sched_class
) {
1952 if (prev_class
->switched_from
)
1953 prev_class
->switched_from(rq
, p
, running
);
1954 p
->sched_class
->switched_to(rq
, p
, running
);
1956 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1961 * Is this task likely cache-hot:
1964 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1968 if (p
->sched_class
!= &fair_sched_class
)
1972 * Buddy candidates are cache hot:
1974 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
1975 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1976 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1979 if (sysctl_sched_migration_cost
== -1)
1981 if (sysctl_sched_migration_cost
== 0)
1984 delta
= now
- p
->se
.exec_start
;
1986 return delta
< (s64
)sysctl_sched_migration_cost
;
1989 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1991 #ifdef CONFIG_SCHED_DEBUG
1993 * We should never call set_task_cpu() on a blocked task,
1994 * ttwu() will sort out the placement.
1996 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
1997 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2000 trace_sched_migrate_task(p
, new_cpu
);
2002 if (task_cpu(p
) != new_cpu
) {
2003 p
->se
.nr_migrations
++;
2004 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2007 __set_task_cpu(p
, new_cpu
);
2010 struct migration_arg
{
2011 struct task_struct
*task
;
2015 static int migration_cpu_stop(void *data
);
2018 * The task's runqueue lock must be held.
2019 * Returns true if you have to wait for migration thread.
2021 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2023 struct rq
*rq
= task_rq(p
);
2026 * If the task is not on a runqueue (and not running), then
2027 * the next wake-up will properly place the task.
2029 return p
->se
.on_rq
|| task_running(rq
, p
);
2033 * wait_task_inactive - wait for a thread to unschedule.
2035 * If @match_state is nonzero, it's the @p->state value just checked and
2036 * not expected to change. If it changes, i.e. @p might have woken up,
2037 * then return zero. When we succeed in waiting for @p to be off its CPU,
2038 * we return a positive number (its total switch count). If a second call
2039 * a short while later returns the same number, the caller can be sure that
2040 * @p has remained unscheduled the whole time.
2042 * The caller must ensure that the task *will* unschedule sometime soon,
2043 * else this function might spin for a *long* time. This function can't
2044 * be called with interrupts off, or it may introduce deadlock with
2045 * smp_call_function() if an IPI is sent by the same process we are
2046 * waiting to become inactive.
2048 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2050 unsigned long flags
;
2057 * We do the initial early heuristics without holding
2058 * any task-queue locks at all. We'll only try to get
2059 * the runqueue lock when things look like they will
2065 * If the task is actively running on another CPU
2066 * still, just relax and busy-wait without holding
2069 * NOTE! Since we don't hold any locks, it's not
2070 * even sure that "rq" stays as the right runqueue!
2071 * But we don't care, since "task_running()" will
2072 * return false if the runqueue has changed and p
2073 * is actually now running somewhere else!
2075 while (task_running(rq
, p
)) {
2076 if (match_state
&& unlikely(p
->state
!= match_state
))
2082 * Ok, time to look more closely! We need the rq
2083 * lock now, to be *sure*. If we're wrong, we'll
2084 * just go back and repeat.
2086 rq
= task_rq_lock(p
, &flags
);
2087 trace_sched_wait_task(p
);
2088 running
= task_running(rq
, p
);
2089 on_rq
= p
->se
.on_rq
;
2091 if (!match_state
|| p
->state
== match_state
)
2092 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2093 task_rq_unlock(rq
, &flags
);
2096 * If it changed from the expected state, bail out now.
2098 if (unlikely(!ncsw
))
2102 * Was it really running after all now that we
2103 * checked with the proper locks actually held?
2105 * Oops. Go back and try again..
2107 if (unlikely(running
)) {
2113 * It's not enough that it's not actively running,
2114 * it must be off the runqueue _entirely_, and not
2117 * So if it was still runnable (but just not actively
2118 * running right now), it's preempted, and we should
2119 * yield - it could be a while.
2121 if (unlikely(on_rq
)) {
2122 schedule_timeout_uninterruptible(1);
2127 * Ahh, all good. It wasn't running, and it wasn't
2128 * runnable, which means that it will never become
2129 * running in the future either. We're all done!
2138 * kick_process - kick a running thread to enter/exit the kernel
2139 * @p: the to-be-kicked thread
2141 * Cause a process which is running on another CPU to enter
2142 * kernel-mode, without any delay. (to get signals handled.)
2144 * NOTE: this function doesnt have to take the runqueue lock,
2145 * because all it wants to ensure is that the remote task enters
2146 * the kernel. If the IPI races and the task has been migrated
2147 * to another CPU then no harm is done and the purpose has been
2150 void kick_process(struct task_struct
*p
)
2156 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2157 smp_send_reschedule(cpu
);
2160 EXPORT_SYMBOL_GPL(kick_process
);
2161 #endif /* CONFIG_SMP */
2164 * task_oncpu_function_call - call a function on the cpu on which a task runs
2165 * @p: the task to evaluate
2166 * @func: the function to be called
2167 * @info: the function call argument
2169 * Calls the function @func when the task is currently running. This might
2170 * be on the current CPU, which just calls the function directly
2172 void task_oncpu_function_call(struct task_struct
*p
,
2173 void (*func
) (void *info
), void *info
)
2180 smp_call_function_single(cpu
, func
, info
, 1);
2186 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2188 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2191 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2193 /* Look for allowed, online CPU in same node. */
2194 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2195 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2198 /* Any allowed, online CPU? */
2199 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2200 if (dest_cpu
< nr_cpu_ids
)
2203 /* No more Mr. Nice Guy. */
2204 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2205 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2207 * Don't tell them about moving exiting tasks or
2208 * kernel threads (both mm NULL), since they never
2211 if (p
->mm
&& printk_ratelimit()) {
2212 printk(KERN_INFO
"process %d (%s) no "
2213 "longer affine to cpu%d\n",
2214 task_pid_nr(p
), p
->comm
, cpu
);
2222 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2225 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2227 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2230 * In order not to call set_task_cpu() on a blocking task we need
2231 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2234 * Since this is common to all placement strategies, this lives here.
2236 * [ this allows ->select_task() to simply return task_cpu(p) and
2237 * not worry about this generic constraint ]
2239 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2241 cpu
= select_fallback_rq(task_cpu(p
), p
);
2246 static void update_avg(u64
*avg
, u64 sample
)
2248 s64 diff
= sample
- *avg
;
2254 * try_to_wake_up - wake up a thread
2255 * @p: the to-be-woken-up thread
2256 * @state: the mask of task states that can be woken
2257 * @sync: do a synchronous wakeup?
2259 * Put it on the run-queue if it's not already there. The "current"
2260 * thread is always on the run-queue (except when the actual
2261 * re-schedule is in progress), and as such you're allowed to do
2262 * the simpler "current->state = TASK_RUNNING" to mark yourself
2263 * runnable without the overhead of this.
2265 * returns failure only if the task is already active.
2267 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2270 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2271 unsigned long flags
;
2272 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2275 this_cpu
= get_cpu();
2278 rq
= task_rq_lock(p
, &flags
);
2279 if (!(p
->state
& state
))
2289 if (unlikely(task_running(rq
, p
)))
2293 * In order to handle concurrent wakeups and release the rq->lock
2294 * we put the task in TASK_WAKING state.
2296 * First fix up the nr_uninterruptible count:
2298 if (task_contributes_to_load(p
)) {
2299 if (likely(cpu_online(orig_cpu
)))
2300 rq
->nr_uninterruptible
--;
2302 this_rq()->nr_uninterruptible
--;
2304 p
->state
= TASK_WAKING
;
2306 if (p
->sched_class
->task_waking
) {
2307 p
->sched_class
->task_waking(rq
, p
);
2308 en_flags
|= ENQUEUE_WAKING
;
2311 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2312 if (cpu
!= orig_cpu
)
2313 set_task_cpu(p
, cpu
);
2314 __task_rq_unlock(rq
);
2317 raw_spin_lock(&rq
->lock
);
2320 * We migrated the task without holding either rq->lock, however
2321 * since the task is not on the task list itself, nobody else
2322 * will try and migrate the task, hence the rq should match the
2323 * cpu we just moved it to.
2325 WARN_ON(task_cpu(p
) != cpu
);
2326 WARN_ON(p
->state
!= TASK_WAKING
);
2328 #ifdef CONFIG_SCHEDSTATS
2329 schedstat_inc(rq
, ttwu_count
);
2330 if (cpu
== this_cpu
)
2331 schedstat_inc(rq
, ttwu_local
);
2333 struct sched_domain
*sd
;
2334 for_each_domain(this_cpu
, sd
) {
2335 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2336 schedstat_inc(sd
, ttwu_wake_remote
);
2341 #endif /* CONFIG_SCHEDSTATS */
2344 #endif /* CONFIG_SMP */
2345 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2346 if (wake_flags
& WF_SYNC
)
2347 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2348 if (orig_cpu
!= cpu
)
2349 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2350 if (cpu
== this_cpu
)
2351 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2353 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2354 activate_task(rq
, p
, en_flags
);
2358 trace_sched_wakeup(p
, success
);
2359 check_preempt_curr(rq
, p
, wake_flags
);
2361 p
->state
= TASK_RUNNING
;
2363 if (p
->sched_class
->task_woken
)
2364 p
->sched_class
->task_woken(rq
, p
);
2366 if (unlikely(rq
->idle_stamp
)) {
2367 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2368 u64 max
= 2*sysctl_sched_migration_cost
;
2373 update_avg(&rq
->avg_idle
, delta
);
2378 task_rq_unlock(rq
, &flags
);
2385 * wake_up_process - Wake up a specific process
2386 * @p: The process to be woken up.
2388 * Attempt to wake up the nominated process and move it to the set of runnable
2389 * processes. Returns 1 if the process was woken up, 0 if it was already
2392 * It may be assumed that this function implies a write memory barrier before
2393 * changing the task state if and only if any tasks are woken up.
2395 int wake_up_process(struct task_struct
*p
)
2397 return try_to_wake_up(p
, TASK_ALL
, 0);
2399 EXPORT_SYMBOL(wake_up_process
);
2401 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2403 return try_to_wake_up(p
, state
, 0);
2407 * Perform scheduler related setup for a newly forked process p.
2408 * p is forked by current.
2410 * __sched_fork() is basic setup used by init_idle() too:
2412 static void __sched_fork(struct task_struct
*p
)
2414 p
->se
.exec_start
= 0;
2415 p
->se
.sum_exec_runtime
= 0;
2416 p
->se
.prev_sum_exec_runtime
= 0;
2417 p
->se
.nr_migrations
= 0;
2419 #ifdef CONFIG_SCHEDSTATS
2420 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2423 INIT_LIST_HEAD(&p
->rt
.run_list
);
2425 INIT_LIST_HEAD(&p
->se
.group_node
);
2427 #ifdef CONFIG_PREEMPT_NOTIFIERS
2428 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2433 * fork()/clone()-time setup:
2435 void sched_fork(struct task_struct
*p
, int clone_flags
)
2437 int cpu
= get_cpu();
2441 * We mark the process as running here. This guarantees that
2442 * nobody will actually run it, and a signal or other external
2443 * event cannot wake it up and insert it on the runqueue either.
2445 p
->state
= TASK_RUNNING
;
2448 * Revert to default priority/policy on fork if requested.
2450 if (unlikely(p
->sched_reset_on_fork
)) {
2451 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2452 p
->policy
= SCHED_NORMAL
;
2453 p
->normal_prio
= p
->static_prio
;
2456 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2457 p
->static_prio
= NICE_TO_PRIO(0);
2458 p
->normal_prio
= p
->static_prio
;
2463 * We don't need the reset flag anymore after the fork. It has
2464 * fulfilled its duty:
2466 p
->sched_reset_on_fork
= 0;
2470 * Make sure we do not leak PI boosting priority to the child.
2472 p
->prio
= current
->normal_prio
;
2474 if (!rt_prio(p
->prio
))
2475 p
->sched_class
= &fair_sched_class
;
2477 if (p
->sched_class
->task_fork
)
2478 p
->sched_class
->task_fork(p
);
2481 * The child is not yet in the pid-hash so no cgroup attach races,
2482 * and the cgroup is pinned to this child due to cgroup_fork()
2483 * is ran before sched_fork().
2485 * Silence PROVE_RCU.
2488 set_task_cpu(p
, cpu
);
2491 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2492 if (likely(sched_info_on()))
2493 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2495 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2498 #ifdef CONFIG_PREEMPT
2499 /* Want to start with kernel preemption disabled. */
2500 task_thread_info(p
)->preempt_count
= 1;
2502 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2508 * wake_up_new_task - wake up a newly created task for the first time.
2510 * This function will do some initial scheduler statistics housekeeping
2511 * that must be done for every newly created context, then puts the task
2512 * on the runqueue and wakes it.
2514 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2516 unsigned long flags
;
2518 int cpu __maybe_unused
= get_cpu();
2521 rq
= task_rq_lock(p
, &flags
);
2522 p
->state
= TASK_WAKING
;
2525 * Fork balancing, do it here and not earlier because:
2526 * - cpus_allowed can change in the fork path
2527 * - any previously selected cpu might disappear through hotplug
2529 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2530 * without people poking at ->cpus_allowed.
2532 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2533 set_task_cpu(p
, cpu
);
2535 p
->state
= TASK_RUNNING
;
2536 task_rq_unlock(rq
, &flags
);
2539 rq
= task_rq_lock(p
, &flags
);
2540 activate_task(rq
, p
, 0);
2541 trace_sched_wakeup_new(p
, 1);
2542 check_preempt_curr(rq
, p
, WF_FORK
);
2544 if (p
->sched_class
->task_woken
)
2545 p
->sched_class
->task_woken(rq
, p
);
2547 task_rq_unlock(rq
, &flags
);
2551 #ifdef CONFIG_PREEMPT_NOTIFIERS
2554 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2555 * @notifier: notifier struct to register
2557 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2559 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2561 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2564 * preempt_notifier_unregister - no longer interested in preemption notifications
2565 * @notifier: notifier struct to unregister
2567 * This is safe to call from within a preemption notifier.
2569 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2571 hlist_del(¬ifier
->link
);
2573 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2575 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2577 struct preempt_notifier
*notifier
;
2578 struct hlist_node
*node
;
2580 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2581 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2585 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2586 struct task_struct
*next
)
2588 struct preempt_notifier
*notifier
;
2589 struct hlist_node
*node
;
2591 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2592 notifier
->ops
->sched_out(notifier
, next
);
2595 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2597 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2602 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2603 struct task_struct
*next
)
2607 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2610 * prepare_task_switch - prepare to switch tasks
2611 * @rq: the runqueue preparing to switch
2612 * @prev: the current task that is being switched out
2613 * @next: the task we are going to switch to.
2615 * This is called with the rq lock held and interrupts off. It must
2616 * be paired with a subsequent finish_task_switch after the context
2619 * prepare_task_switch sets up locking and calls architecture specific
2623 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2624 struct task_struct
*next
)
2626 fire_sched_out_preempt_notifiers(prev
, next
);
2627 prepare_lock_switch(rq
, next
);
2628 prepare_arch_switch(next
);
2632 * finish_task_switch - clean up after a task-switch
2633 * @rq: runqueue associated with task-switch
2634 * @prev: the thread we just switched away from.
2636 * finish_task_switch must be called after the context switch, paired
2637 * with a prepare_task_switch call before the context switch.
2638 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2639 * and do any other architecture-specific cleanup actions.
2641 * Note that we may have delayed dropping an mm in context_switch(). If
2642 * so, we finish that here outside of the runqueue lock. (Doing it
2643 * with the lock held can cause deadlocks; see schedule() for
2646 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2647 __releases(rq
->lock
)
2649 struct mm_struct
*mm
= rq
->prev_mm
;
2655 * A task struct has one reference for the use as "current".
2656 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2657 * schedule one last time. The schedule call will never return, and
2658 * the scheduled task must drop that reference.
2659 * The test for TASK_DEAD must occur while the runqueue locks are
2660 * still held, otherwise prev could be scheduled on another cpu, die
2661 * there before we look at prev->state, and then the reference would
2663 * Manfred Spraul <manfred@colorfullife.com>
2665 prev_state
= prev
->state
;
2666 finish_arch_switch(prev
);
2667 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2668 local_irq_disable();
2669 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2670 perf_event_task_sched_in(current
);
2671 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2673 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2674 finish_lock_switch(rq
, prev
);
2676 fire_sched_in_preempt_notifiers(current
);
2679 if (unlikely(prev_state
== TASK_DEAD
)) {
2681 * Remove function-return probe instances associated with this
2682 * task and put them back on the free list.
2684 kprobe_flush_task(prev
);
2685 put_task_struct(prev
);
2691 /* assumes rq->lock is held */
2692 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2694 if (prev
->sched_class
->pre_schedule
)
2695 prev
->sched_class
->pre_schedule(rq
, prev
);
2698 /* rq->lock is NOT held, but preemption is disabled */
2699 static inline void post_schedule(struct rq
*rq
)
2701 if (rq
->post_schedule
) {
2702 unsigned long flags
;
2704 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2705 if (rq
->curr
->sched_class
->post_schedule
)
2706 rq
->curr
->sched_class
->post_schedule(rq
);
2707 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2709 rq
->post_schedule
= 0;
2715 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2719 static inline void post_schedule(struct rq
*rq
)
2726 * schedule_tail - first thing a freshly forked thread must call.
2727 * @prev: the thread we just switched away from.
2729 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2730 __releases(rq
->lock
)
2732 struct rq
*rq
= this_rq();
2734 finish_task_switch(rq
, prev
);
2737 * FIXME: do we need to worry about rq being invalidated by the
2742 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2743 /* In this case, finish_task_switch does not reenable preemption */
2746 if (current
->set_child_tid
)
2747 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2751 * context_switch - switch to the new MM and the new
2752 * thread's register state.
2755 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2756 struct task_struct
*next
)
2758 struct mm_struct
*mm
, *oldmm
;
2760 prepare_task_switch(rq
, prev
, next
);
2761 trace_sched_switch(prev
, next
);
2763 oldmm
= prev
->active_mm
;
2765 * For paravirt, this is coupled with an exit in switch_to to
2766 * combine the page table reload and the switch backend into
2769 arch_start_context_switch(prev
);
2772 next
->active_mm
= oldmm
;
2773 atomic_inc(&oldmm
->mm_count
);
2774 enter_lazy_tlb(oldmm
, next
);
2776 switch_mm(oldmm
, mm
, next
);
2778 if (likely(!prev
->mm
)) {
2779 prev
->active_mm
= NULL
;
2780 rq
->prev_mm
= oldmm
;
2783 * Since the runqueue lock will be released by the next
2784 * task (which is an invalid locking op but in the case
2785 * of the scheduler it's an obvious special-case), so we
2786 * do an early lockdep release here:
2788 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2789 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2792 /* Here we just switch the register state and the stack. */
2793 switch_to(prev
, next
, prev
);
2797 * this_rq must be evaluated again because prev may have moved
2798 * CPUs since it called schedule(), thus the 'rq' on its stack
2799 * frame will be invalid.
2801 finish_task_switch(this_rq(), prev
);
2805 * nr_running, nr_uninterruptible and nr_context_switches:
2807 * externally visible scheduler statistics: current number of runnable
2808 * threads, current number of uninterruptible-sleeping threads, total
2809 * number of context switches performed since bootup.
2811 unsigned long nr_running(void)
2813 unsigned long i
, sum
= 0;
2815 for_each_online_cpu(i
)
2816 sum
+= cpu_rq(i
)->nr_running
;
2821 unsigned long nr_uninterruptible(void)
2823 unsigned long i
, sum
= 0;
2825 for_each_possible_cpu(i
)
2826 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2829 * Since we read the counters lockless, it might be slightly
2830 * inaccurate. Do not allow it to go below zero though:
2832 if (unlikely((long)sum
< 0))
2838 unsigned long long nr_context_switches(void)
2841 unsigned long long sum
= 0;
2843 for_each_possible_cpu(i
)
2844 sum
+= cpu_rq(i
)->nr_switches
;
2849 unsigned long nr_iowait(void)
2851 unsigned long i
, sum
= 0;
2853 for_each_possible_cpu(i
)
2854 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2859 unsigned long nr_iowait_cpu(int cpu
)
2861 struct rq
*this = cpu_rq(cpu
);
2862 return atomic_read(&this->nr_iowait
);
2865 unsigned long this_cpu_load(void)
2867 struct rq
*this = this_rq();
2868 return this->cpu_load
[0];
2872 /* Variables and functions for calc_load */
2873 static atomic_long_t calc_load_tasks
;
2874 static unsigned long calc_load_update
;
2875 unsigned long avenrun
[3];
2876 EXPORT_SYMBOL(avenrun
);
2878 static long calc_load_fold_active(struct rq
*this_rq
)
2880 long nr_active
, delta
= 0;
2882 nr_active
= this_rq
->nr_running
;
2883 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2885 if (nr_active
!= this_rq
->calc_load_active
) {
2886 delta
= nr_active
- this_rq
->calc_load_active
;
2887 this_rq
->calc_load_active
= nr_active
;
2893 static unsigned long
2894 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2897 load
+= active
* (FIXED_1
- exp
);
2898 load
+= 1UL << (FSHIFT
- 1);
2899 return load
>> FSHIFT
;
2904 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2906 * When making the ILB scale, we should try to pull this in as well.
2908 static atomic_long_t calc_load_tasks_idle
;
2910 static void calc_load_account_idle(struct rq
*this_rq
)
2914 delta
= calc_load_fold_active(this_rq
);
2916 atomic_long_add(delta
, &calc_load_tasks_idle
);
2919 static long calc_load_fold_idle(void)
2924 * Its got a race, we don't care...
2926 if (atomic_long_read(&calc_load_tasks_idle
))
2927 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2933 * fixed_power_int - compute: x^n, in O(log n) time
2935 * @x: base of the power
2936 * @frac_bits: fractional bits of @x
2937 * @n: power to raise @x to.
2939 * By exploiting the relation between the definition of the natural power
2940 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
2941 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
2942 * (where: n_i \elem {0, 1}, the binary vector representing n),
2943 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
2944 * of course trivially computable in O(log_2 n), the length of our binary
2947 static unsigned long
2948 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
2950 unsigned long result
= 1UL << frac_bits
;
2955 result
+= 1UL << (frac_bits
- 1);
2956 result
>>= frac_bits
;
2962 x
+= 1UL << (frac_bits
- 1);
2970 * a1 = a0 * e + a * (1 - e)
2972 * a2 = a1 * e + a * (1 - e)
2973 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
2974 * = a0 * e^2 + a * (1 - e) * (1 + e)
2976 * a3 = a2 * e + a * (1 - e)
2977 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
2978 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
2982 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
2983 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
2984 * = a0 * e^n + a * (1 - e^n)
2986 * [1] application of the geometric series:
2989 * S_n := \Sum x^i = -------------
2992 static unsigned long
2993 calc_load_n(unsigned long load
, unsigned long exp
,
2994 unsigned long active
, unsigned int n
)
2997 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3001 * NO_HZ can leave us missing all per-cpu ticks calling
3002 * calc_load_account_active(), but since an idle CPU folds its delta into
3003 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3004 * in the pending idle delta if our idle period crossed a load cycle boundary.
3006 * Once we've updated the global active value, we need to apply the exponential
3007 * weights adjusted to the number of cycles missed.
3009 static void calc_global_nohz(unsigned long ticks
)
3011 long delta
, active
, n
;
3013 if (time_before(jiffies
, calc_load_update
))
3017 * If we crossed a calc_load_update boundary, make sure to fold
3018 * any pending idle changes, the respective CPUs might have
3019 * missed the tick driven calc_load_account_active() update
3022 delta
= calc_load_fold_idle();
3024 atomic_long_add(delta
, &calc_load_tasks
);
3027 * If we were idle for multiple load cycles, apply them.
3029 if (ticks
>= LOAD_FREQ
) {
3030 n
= ticks
/ LOAD_FREQ
;
3032 active
= atomic_long_read(&calc_load_tasks
);
3033 active
= active
> 0 ? active
* FIXED_1
: 0;
3035 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3036 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3037 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3039 calc_load_update
+= n
* LOAD_FREQ
;
3043 * Its possible the remainder of the above division also crosses
3044 * a LOAD_FREQ period, the regular check in calc_global_load()
3045 * which comes after this will take care of that.
3047 * Consider us being 11 ticks before a cycle completion, and us
3048 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3049 * age us 4 cycles, and the test in calc_global_load() will
3050 * pick up the final one.
3054 static void calc_load_account_idle(struct rq
*this_rq
)
3058 static inline long calc_load_fold_idle(void)
3063 static void calc_global_nohz(unsigned long ticks
)
3069 * get_avenrun - get the load average array
3070 * @loads: pointer to dest load array
3071 * @offset: offset to add
3072 * @shift: shift count to shift the result left
3074 * These values are estimates at best, so no need for locking.
3076 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3078 loads
[0] = (avenrun
[0] + offset
) << shift
;
3079 loads
[1] = (avenrun
[1] + offset
) << shift
;
3080 loads
[2] = (avenrun
[2] + offset
) << shift
;
3084 * calc_load - update the avenrun load estimates 10 ticks after the
3085 * CPUs have updated calc_load_tasks.
3087 void calc_global_load(unsigned long ticks
)
3091 calc_global_nohz(ticks
);
3093 if (time_before(jiffies
, calc_load_update
+ 10))
3096 active
= atomic_long_read(&calc_load_tasks
);
3097 active
= active
> 0 ? active
* FIXED_1
: 0;
3099 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3100 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3101 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3103 calc_load_update
+= LOAD_FREQ
;
3107 * Called from update_cpu_load() to periodically update this CPU's
3110 static void calc_load_account_active(struct rq
*this_rq
)
3114 if (time_before(jiffies
, this_rq
->calc_load_update
))
3117 delta
= calc_load_fold_active(this_rq
);
3118 delta
+= calc_load_fold_idle();
3120 atomic_long_add(delta
, &calc_load_tasks
);
3122 this_rq
->calc_load_update
+= LOAD_FREQ
;
3126 * Update rq->cpu_load[] statistics. This function is usually called every
3127 * scheduler tick (TICK_NSEC).
3129 static void update_cpu_load(struct rq
*this_rq
)
3131 unsigned long this_load
= this_rq
->load
.weight
;
3134 this_rq
->nr_load_updates
++;
3136 /* Update our load: */
3137 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3138 unsigned long old_load
, new_load
;
3140 /* scale is effectively 1 << i now, and >> i divides by scale */
3142 old_load
= this_rq
->cpu_load
[i
];
3143 new_load
= this_load
;
3145 * Round up the averaging division if load is increasing. This
3146 * prevents us from getting stuck on 9 if the load is 10, for
3149 if (new_load
> old_load
)
3150 new_load
+= scale
-1;
3151 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3154 calc_load_account_active(this_rq
);
3160 * sched_exec - execve() is a valuable balancing opportunity, because at
3161 * this point the task has the smallest effective memory and cache footprint.
3163 void sched_exec(void)
3165 struct task_struct
*p
= current
;
3166 unsigned long flags
;
3170 rq
= task_rq_lock(p
, &flags
);
3171 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3172 if (dest_cpu
== smp_processor_id())
3176 * select_task_rq() can race against ->cpus_allowed
3178 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3179 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3180 struct migration_arg arg
= { p
, dest_cpu
};
3182 task_rq_unlock(rq
, &flags
);
3183 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3187 task_rq_unlock(rq
, &flags
);
3192 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3194 EXPORT_PER_CPU_SYMBOL(kstat
);
3197 * Return any ns on the sched_clock that have not yet been accounted in
3198 * @p in case that task is currently running.
3200 * Called with task_rq_lock() held on @rq.
3202 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3206 if (task_current(rq
, p
)) {
3207 update_rq_clock(rq
);
3208 ns
= rq
->clock
- p
->se
.exec_start
;
3216 unsigned long long task_delta_exec(struct task_struct
*p
)
3218 unsigned long flags
;
3222 rq
= task_rq_lock(p
, &flags
);
3223 ns
= do_task_delta_exec(p
, rq
);
3224 task_rq_unlock(rq
, &flags
);
3230 * Return accounted runtime for the task.
3231 * In case the task is currently running, return the runtime plus current's
3232 * pending runtime that have not been accounted yet.
3234 unsigned long long task_sched_runtime(struct task_struct
*p
)
3236 unsigned long flags
;
3240 rq
= task_rq_lock(p
, &flags
);
3241 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3242 task_rq_unlock(rq
, &flags
);
3248 * Return sum_exec_runtime for the thread group.
3249 * In case the task is currently running, return the sum plus current's
3250 * pending runtime that have not been accounted yet.
3252 * Note that the thread group might have other running tasks as well,
3253 * so the return value not includes other pending runtime that other
3254 * running tasks might have.
3256 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3258 struct task_cputime totals
;
3259 unsigned long flags
;
3263 rq
= task_rq_lock(p
, &flags
);
3264 thread_group_cputime(p
, &totals
);
3265 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3266 task_rq_unlock(rq
, &flags
);
3272 * Account user cpu time to a process.
3273 * @p: the process that the cpu time gets accounted to
3274 * @cputime: the cpu time spent in user space since the last update
3275 * @cputime_scaled: cputime scaled by cpu frequency
3277 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3278 cputime_t cputime_scaled
)
3280 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3283 /* Add user time to process. */
3284 p
->utime
= cputime_add(p
->utime
, cputime
);
3285 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3286 account_group_user_time(p
, cputime
);
3288 /* Add user time to cpustat. */
3289 tmp
= cputime_to_cputime64(cputime
);
3290 if (TASK_NICE(p
) > 0)
3291 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3293 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3295 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3296 /* Account for user time used */
3297 acct_update_integrals(p
);
3301 * Account guest cpu time to a process.
3302 * @p: the process that the cpu time gets accounted to
3303 * @cputime: the cpu time spent in virtual machine since the last update
3304 * @cputime_scaled: cputime scaled by cpu frequency
3306 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3307 cputime_t cputime_scaled
)
3310 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3312 tmp
= cputime_to_cputime64(cputime
);
3314 /* Add guest time to process. */
3315 p
->utime
= cputime_add(p
->utime
, cputime
);
3316 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3317 account_group_user_time(p
, cputime
);
3318 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3320 /* Add guest time to cpustat. */
3321 if (TASK_NICE(p
) > 0) {
3322 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3323 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3325 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3326 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3331 * Account system cpu time to a process.
3332 * @p: the process that the cpu time gets accounted to
3333 * @hardirq_offset: the offset to subtract from hardirq_count()
3334 * @cputime: the cpu time spent in kernel space since the last update
3335 * @cputime_scaled: cputime scaled by cpu frequency
3337 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3338 cputime_t cputime
, cputime_t cputime_scaled
)
3340 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3343 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3344 account_guest_time(p
, cputime
, cputime_scaled
);
3348 /* Add system time to process. */
3349 p
->stime
= cputime_add(p
->stime
, cputime
);
3350 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3351 account_group_system_time(p
, cputime
);
3353 /* Add system time to cpustat. */
3354 tmp
= cputime_to_cputime64(cputime
);
3355 if (hardirq_count() - hardirq_offset
)
3356 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3357 else if (softirq_count())
3358 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3360 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3362 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3364 /* Account for system time used */
3365 acct_update_integrals(p
);
3369 * Account for involuntary wait time.
3370 * @steal: the cpu time spent in involuntary wait
3372 void account_steal_time(cputime_t cputime
)
3374 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3375 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3377 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3381 * Account for idle time.
3382 * @cputime: the cpu time spent in idle wait
3384 void account_idle_time(cputime_t cputime
)
3386 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3387 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3388 struct rq
*rq
= this_rq();
3390 if (atomic_read(&rq
->nr_iowait
) > 0)
3391 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3393 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3396 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3399 * Account a single tick of cpu time.
3400 * @p: the process that the cpu time gets accounted to
3401 * @user_tick: indicates if the tick is a user or a system tick
3403 void account_process_tick(struct task_struct
*p
, int user_tick
)
3405 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3406 struct rq
*rq
= this_rq();
3409 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3410 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3411 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3414 account_idle_time(cputime_one_jiffy
);
3418 * Account multiple ticks of steal time.
3419 * @p: the process from which the cpu time has been stolen
3420 * @ticks: number of stolen ticks
3422 void account_steal_ticks(unsigned long ticks
)
3424 account_steal_time(jiffies_to_cputime(ticks
));
3428 * Account multiple ticks of idle time.
3429 * @ticks: number of stolen ticks
3431 void account_idle_ticks(unsigned long ticks
)
3433 account_idle_time(jiffies_to_cputime(ticks
));
3439 * Use precise platform statistics if available:
3441 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3442 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3448 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3450 struct task_cputime cputime
;
3452 thread_group_cputime(p
, &cputime
);
3454 *ut
= cputime
.utime
;
3455 *st
= cputime
.stime
;
3459 #ifndef nsecs_to_cputime
3460 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3463 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3465 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3468 * Use CFS's precise accounting:
3470 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3476 do_div(temp
, total
);
3477 utime
= (cputime_t
)temp
;
3482 * Compare with previous values, to keep monotonicity:
3484 p
->prev_utime
= max(p
->prev_utime
, utime
);
3485 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3487 *ut
= p
->prev_utime
;
3488 *st
= p
->prev_stime
;
3492 * Must be called with siglock held.
3494 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3496 struct signal_struct
*sig
= p
->signal
;
3497 struct task_cputime cputime
;
3498 cputime_t rtime
, utime
, total
;
3500 thread_group_cputime(p
, &cputime
);
3502 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3503 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3508 temp
*= cputime
.utime
;
3509 do_div(temp
, total
);
3510 utime
= (cputime_t
)temp
;
3514 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3515 sig
->prev_stime
= max(sig
->prev_stime
,
3516 cputime_sub(rtime
, sig
->prev_utime
));
3518 *ut
= sig
->prev_utime
;
3519 *st
= sig
->prev_stime
;
3524 * This function gets called by the timer code, with HZ frequency.
3525 * We call it with interrupts disabled.
3527 * It also gets called by the fork code, when changing the parent's
3530 void scheduler_tick(void)
3532 int cpu
= smp_processor_id();
3533 struct rq
*rq
= cpu_rq(cpu
);
3534 struct task_struct
*curr
= rq
->curr
;
3538 raw_spin_lock(&rq
->lock
);
3539 update_rq_clock(rq
);
3540 update_cpu_load(rq
);
3541 curr
->sched_class
->task_tick(rq
, curr
, 0);
3542 raw_spin_unlock(&rq
->lock
);
3544 perf_event_task_tick(curr
);
3547 rq
->idle_at_tick
= idle_cpu(cpu
);
3548 trigger_load_balance(rq
, cpu
);
3552 notrace
unsigned long get_parent_ip(unsigned long addr
)
3554 if (in_lock_functions(addr
)) {
3555 addr
= CALLER_ADDR2
;
3556 if (in_lock_functions(addr
))
3557 addr
= CALLER_ADDR3
;
3562 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3563 defined(CONFIG_PREEMPT_TRACER))
3565 void __kprobes
add_preempt_count(int val
)
3567 #ifdef CONFIG_DEBUG_PREEMPT
3571 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3574 preempt_count() += val
;
3575 #ifdef CONFIG_DEBUG_PREEMPT
3577 * Spinlock count overflowing soon?
3579 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3582 if (preempt_count() == val
)
3583 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3585 EXPORT_SYMBOL(add_preempt_count
);
3587 void __kprobes
sub_preempt_count(int val
)
3589 #ifdef CONFIG_DEBUG_PREEMPT
3593 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3596 * Is the spinlock portion underflowing?
3598 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3599 !(preempt_count() & PREEMPT_MASK
)))
3603 if (preempt_count() == val
)
3604 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3605 preempt_count() -= val
;
3607 EXPORT_SYMBOL(sub_preempt_count
);
3612 * Print scheduling while atomic bug:
3614 static noinline
void __schedule_bug(struct task_struct
*prev
)
3616 struct pt_regs
*regs
= get_irq_regs();
3618 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3619 prev
->comm
, prev
->pid
, preempt_count());
3621 debug_show_held_locks(prev
);
3623 if (irqs_disabled())
3624 print_irqtrace_events(prev
);
3633 * Various schedule()-time debugging checks and statistics:
3635 static inline void schedule_debug(struct task_struct
*prev
)
3638 * Test if we are atomic. Since do_exit() needs to call into
3639 * schedule() atomically, we ignore that path for now.
3640 * Otherwise, whine if we are scheduling when we should not be.
3642 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3643 __schedule_bug(prev
);
3645 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3647 schedstat_inc(this_rq(), sched_count
);
3648 #ifdef CONFIG_SCHEDSTATS
3649 if (unlikely(prev
->lock_depth
>= 0)) {
3650 schedstat_inc(this_rq(), bkl_count
);
3651 schedstat_inc(prev
, sched_info
.bkl_count
);
3656 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3659 update_rq_clock(rq
);
3660 prev
->sched_class
->put_prev_task(rq
, prev
);
3664 * Pick up the highest-prio task:
3666 static inline struct task_struct
*
3667 pick_next_task(struct rq
*rq
)
3669 const struct sched_class
*class;
3670 struct task_struct
*p
;
3673 * Optimization: we know that if all tasks are in
3674 * the fair class we can call that function directly:
3676 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3677 p
= fair_sched_class
.pick_next_task(rq
);
3682 class = sched_class_highest
;
3684 p
= class->pick_next_task(rq
);
3688 * Will never be NULL as the idle class always
3689 * returns a non-NULL p:
3691 class = class->next
;
3696 * schedule() is the main scheduler function.
3698 asmlinkage
void __sched
schedule(void)
3700 struct task_struct
*prev
, *next
;
3701 unsigned long *switch_count
;
3707 cpu
= smp_processor_id();
3709 rcu_note_context_switch(cpu
);
3711 switch_count
= &prev
->nivcsw
;
3713 release_kernel_lock(prev
);
3714 need_resched_nonpreemptible
:
3716 schedule_debug(prev
);
3718 if (sched_feat(HRTICK
))
3721 raw_spin_lock_irq(&rq
->lock
);
3723 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3724 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3725 prev
->state
= TASK_RUNNING
;
3727 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3728 switch_count
= &prev
->nvcsw
;
3731 pre_schedule(rq
, prev
);
3733 if (unlikely(!rq
->nr_running
))
3734 idle_balance(cpu
, rq
);
3736 put_prev_task(rq
, prev
);
3737 next
= pick_next_task(rq
);
3738 clear_tsk_need_resched(prev
);
3739 rq
->skip_clock_update
= 0;
3741 if (likely(prev
!= next
)) {
3742 sched_info_switch(prev
, next
);
3743 perf_event_task_sched_out(prev
, next
);
3749 context_switch(rq
, prev
, next
); /* unlocks the rq */
3751 * the context switch might have flipped the stack from under
3752 * us, hence refresh the local variables.
3754 cpu
= smp_processor_id();
3757 raw_spin_unlock_irq(&rq
->lock
);
3761 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3763 switch_count
= &prev
->nivcsw
;
3764 goto need_resched_nonpreemptible
;
3767 preempt_enable_no_resched();
3771 EXPORT_SYMBOL(schedule
);
3773 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3775 * Look out! "owner" is an entirely speculative pointer
3776 * access and not reliable.
3778 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3783 if (!sched_feat(OWNER_SPIN
))
3786 #ifdef CONFIG_DEBUG_PAGEALLOC
3788 * Need to access the cpu field knowing that
3789 * DEBUG_PAGEALLOC could have unmapped it if
3790 * the mutex owner just released it and exited.
3792 if (probe_kernel_address(&owner
->cpu
, cpu
))
3799 * Even if the access succeeded (likely case),
3800 * the cpu field may no longer be valid.
3802 if (cpu
>= nr_cpumask_bits
)
3806 * We need to validate that we can do a
3807 * get_cpu() and that we have the percpu area.
3809 if (!cpu_online(cpu
))
3816 * Owner changed, break to re-assess state.
3818 if (lock
->owner
!= owner
) {
3820 * If the lock has switched to a different owner,
3821 * we likely have heavy contention. Return 0 to quit
3822 * optimistic spinning and not contend further:
3830 * Is that owner really running on that cpu?
3832 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3842 #ifdef CONFIG_PREEMPT
3844 * this is the entry point to schedule() from in-kernel preemption
3845 * off of preempt_enable. Kernel preemptions off return from interrupt
3846 * occur there and call schedule directly.
3848 asmlinkage
void __sched
preempt_schedule(void)
3850 struct thread_info
*ti
= current_thread_info();
3853 * If there is a non-zero preempt_count or interrupts are disabled,
3854 * we do not want to preempt the current task. Just return..
3856 if (likely(ti
->preempt_count
|| irqs_disabled()))
3860 add_preempt_count(PREEMPT_ACTIVE
);
3862 sub_preempt_count(PREEMPT_ACTIVE
);
3865 * Check again in case we missed a preemption opportunity
3866 * between schedule and now.
3869 } while (need_resched());
3871 EXPORT_SYMBOL(preempt_schedule
);
3874 * this is the entry point to schedule() from kernel preemption
3875 * off of irq context.
3876 * Note, that this is called and return with irqs disabled. This will
3877 * protect us against recursive calling from irq.
3879 asmlinkage
void __sched
preempt_schedule_irq(void)
3881 struct thread_info
*ti
= current_thread_info();
3883 /* Catch callers which need to be fixed */
3884 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3887 add_preempt_count(PREEMPT_ACTIVE
);
3890 local_irq_disable();
3891 sub_preempt_count(PREEMPT_ACTIVE
);
3894 * Check again in case we missed a preemption opportunity
3895 * between schedule and now.
3898 } while (need_resched());
3901 #endif /* CONFIG_PREEMPT */
3903 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3906 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3908 EXPORT_SYMBOL(default_wake_function
);
3911 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3912 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3913 * number) then we wake all the non-exclusive tasks and one exclusive task.
3915 * There are circumstances in which we can try to wake a task which has already
3916 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3917 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3919 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3920 int nr_exclusive
, int wake_flags
, void *key
)
3922 wait_queue_t
*curr
, *next
;
3924 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3925 unsigned flags
= curr
->flags
;
3927 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3928 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3934 * __wake_up - wake up threads blocked on a waitqueue.
3936 * @mode: which threads
3937 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3938 * @key: is directly passed to the wakeup function
3940 * It may be assumed that this function implies a write memory barrier before
3941 * changing the task state if and only if any tasks are woken up.
3943 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3944 int nr_exclusive
, void *key
)
3946 unsigned long flags
;
3948 spin_lock_irqsave(&q
->lock
, flags
);
3949 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3950 spin_unlock_irqrestore(&q
->lock
, flags
);
3952 EXPORT_SYMBOL(__wake_up
);
3955 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3957 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3959 __wake_up_common(q
, mode
, 1, 0, NULL
);
3961 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3963 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3965 __wake_up_common(q
, mode
, 1, 0, key
);
3969 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3971 * @mode: which threads
3972 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3973 * @key: opaque value to be passed to wakeup targets
3975 * The sync wakeup differs that the waker knows that it will schedule
3976 * away soon, so while the target thread will be woken up, it will not
3977 * be migrated to another CPU - ie. the two threads are 'synchronized'
3978 * with each other. This can prevent needless bouncing between CPUs.
3980 * On UP it can prevent extra preemption.
3982 * It may be assumed that this function implies a write memory barrier before
3983 * changing the task state if and only if any tasks are woken up.
3985 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3986 int nr_exclusive
, void *key
)
3988 unsigned long flags
;
3989 int wake_flags
= WF_SYNC
;
3994 if (unlikely(!nr_exclusive
))
3997 spin_lock_irqsave(&q
->lock
, flags
);
3998 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3999 spin_unlock_irqrestore(&q
->lock
, flags
);
4001 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4004 * __wake_up_sync - see __wake_up_sync_key()
4006 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4008 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4010 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4013 * complete: - signals a single thread waiting on this completion
4014 * @x: holds the state of this particular completion
4016 * This will wake up a single thread waiting on this completion. Threads will be
4017 * awakened in the same order in which they were queued.
4019 * See also complete_all(), wait_for_completion() and related routines.
4021 * It may be assumed that this function implies a write memory barrier before
4022 * changing the task state if and only if any tasks are woken up.
4024 void complete(struct completion
*x
)
4026 unsigned long flags
;
4028 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4030 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4031 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4033 EXPORT_SYMBOL(complete
);
4036 * complete_all: - signals all threads waiting on this completion
4037 * @x: holds the state of this particular completion
4039 * This will wake up all threads waiting on this particular completion event.
4041 * It may be assumed that this function implies a write memory barrier before
4042 * changing the task state if and only if any tasks are woken up.
4044 void complete_all(struct completion
*x
)
4046 unsigned long flags
;
4048 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4049 x
->done
+= UINT_MAX
/2;
4050 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4051 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4053 EXPORT_SYMBOL(complete_all
);
4055 static inline long __sched
4056 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4059 DECLARE_WAITQUEUE(wait
, current
);
4061 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4063 if (signal_pending_state(state
, current
)) {
4064 timeout
= -ERESTARTSYS
;
4067 __set_current_state(state
);
4068 spin_unlock_irq(&x
->wait
.lock
);
4069 timeout
= schedule_timeout(timeout
);
4070 spin_lock_irq(&x
->wait
.lock
);
4071 } while (!x
->done
&& timeout
);
4072 __remove_wait_queue(&x
->wait
, &wait
);
4077 return timeout
?: 1;
4081 wait_for_common(struct completion
*x
, long timeout
, int state
)
4085 spin_lock_irq(&x
->wait
.lock
);
4086 timeout
= do_wait_for_common(x
, timeout
, state
);
4087 spin_unlock_irq(&x
->wait
.lock
);
4092 * wait_for_completion: - waits for completion of a task
4093 * @x: holds the state of this particular completion
4095 * This waits to be signaled for completion of a specific task. It is NOT
4096 * interruptible and there is no timeout.
4098 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4099 * and interrupt capability. Also see complete().
4101 void __sched
wait_for_completion(struct completion
*x
)
4103 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4105 EXPORT_SYMBOL(wait_for_completion
);
4108 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4109 * @x: holds the state of this particular completion
4110 * @timeout: timeout value in jiffies
4112 * This waits for either a completion of a specific task to be signaled or for a
4113 * specified timeout to expire. The timeout is in jiffies. It is not
4116 unsigned long __sched
4117 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4119 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4121 EXPORT_SYMBOL(wait_for_completion_timeout
);
4124 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4125 * @x: holds the state of this particular completion
4127 * This waits for completion of a specific task to be signaled. It is
4130 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4132 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4133 if (t
== -ERESTARTSYS
)
4137 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4140 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4141 * @x: holds the state of this particular completion
4142 * @timeout: timeout value in jiffies
4144 * This waits for either a completion of a specific task to be signaled or for a
4145 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4147 unsigned long __sched
4148 wait_for_completion_interruptible_timeout(struct completion
*x
,
4149 unsigned long timeout
)
4151 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4153 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4156 * wait_for_completion_killable: - waits for completion of a task (killable)
4157 * @x: holds the state of this particular completion
4159 * This waits to be signaled for completion of a specific task. It can be
4160 * interrupted by a kill signal.
4162 int __sched
wait_for_completion_killable(struct completion
*x
)
4164 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4165 if (t
== -ERESTARTSYS
)
4169 EXPORT_SYMBOL(wait_for_completion_killable
);
4172 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4173 * @x: holds the state of this particular completion
4174 * @timeout: timeout value in jiffies
4176 * This waits for either a completion of a specific task to be
4177 * signaled or for a specified timeout to expire. It can be
4178 * interrupted by a kill signal. The timeout is in jiffies.
4180 unsigned long __sched
4181 wait_for_completion_killable_timeout(struct completion
*x
,
4182 unsigned long timeout
)
4184 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4186 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4189 * try_wait_for_completion - try to decrement a completion without blocking
4190 * @x: completion structure
4192 * Returns: 0 if a decrement cannot be done without blocking
4193 * 1 if a decrement succeeded.
4195 * If a completion is being used as a counting completion,
4196 * attempt to decrement the counter without blocking. This
4197 * enables us to avoid waiting if the resource the completion
4198 * is protecting is not available.
4200 bool try_wait_for_completion(struct completion
*x
)
4202 unsigned long flags
;
4205 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4210 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4213 EXPORT_SYMBOL(try_wait_for_completion
);
4216 * completion_done - Test to see if a completion has any waiters
4217 * @x: completion structure
4219 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4220 * 1 if there are no waiters.
4223 bool completion_done(struct completion
*x
)
4225 unsigned long flags
;
4228 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4231 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4234 EXPORT_SYMBOL(completion_done
);
4237 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4239 unsigned long flags
;
4242 init_waitqueue_entry(&wait
, current
);
4244 __set_current_state(state
);
4246 spin_lock_irqsave(&q
->lock
, flags
);
4247 __add_wait_queue(q
, &wait
);
4248 spin_unlock(&q
->lock
);
4249 timeout
= schedule_timeout(timeout
);
4250 spin_lock_irq(&q
->lock
);
4251 __remove_wait_queue(q
, &wait
);
4252 spin_unlock_irqrestore(&q
->lock
, flags
);
4257 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4259 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4261 EXPORT_SYMBOL(interruptible_sleep_on
);
4264 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4266 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4268 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4270 void __sched
sleep_on(wait_queue_head_t
*q
)
4272 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4274 EXPORT_SYMBOL(sleep_on
);
4276 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4278 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4280 EXPORT_SYMBOL(sleep_on_timeout
);
4282 #ifdef CONFIG_RT_MUTEXES
4285 * rt_mutex_setprio - set the current priority of a task
4287 * @prio: prio value (kernel-internal form)
4289 * This function changes the 'effective' priority of a task. It does
4290 * not touch ->normal_prio like __setscheduler().
4292 * Used by the rt_mutex code to implement priority inheritance logic.
4294 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4296 unsigned long flags
;
4297 int oldprio
, on_rq
, running
;
4299 const struct sched_class
*prev_class
;
4301 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4303 rq
= task_rq_lock(p
, &flags
);
4306 prev_class
= p
->sched_class
;
4307 on_rq
= p
->se
.on_rq
;
4308 running
= task_current(rq
, p
);
4310 dequeue_task(rq
, p
, 0);
4312 p
->sched_class
->put_prev_task(rq
, p
);
4315 p
->sched_class
= &rt_sched_class
;
4317 p
->sched_class
= &fair_sched_class
;
4322 p
->sched_class
->set_curr_task(rq
);
4324 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4326 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4328 task_rq_unlock(rq
, &flags
);
4333 void set_user_nice(struct task_struct
*p
, long nice
)
4335 int old_prio
, delta
, on_rq
;
4336 unsigned long flags
;
4339 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4342 * We have to be careful, if called from sys_setpriority(),
4343 * the task might be in the middle of scheduling on another CPU.
4345 rq
= task_rq_lock(p
, &flags
);
4347 * The RT priorities are set via sched_setscheduler(), but we still
4348 * allow the 'normal' nice value to be set - but as expected
4349 * it wont have any effect on scheduling until the task is
4350 * SCHED_FIFO/SCHED_RR:
4352 if (task_has_rt_policy(p
)) {
4353 p
->static_prio
= NICE_TO_PRIO(nice
);
4356 on_rq
= p
->se
.on_rq
;
4358 dequeue_task(rq
, p
, 0);
4360 p
->static_prio
= NICE_TO_PRIO(nice
);
4363 p
->prio
= effective_prio(p
);
4364 delta
= p
->prio
- old_prio
;
4367 enqueue_task(rq
, p
, 0);
4369 * If the task increased its priority or is running and
4370 * lowered its priority, then reschedule its CPU:
4372 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4373 resched_task(rq
->curr
);
4376 task_rq_unlock(rq
, &flags
);
4378 EXPORT_SYMBOL(set_user_nice
);
4381 * can_nice - check if a task can reduce its nice value
4385 int can_nice(const struct task_struct
*p
, const int nice
)
4387 /* convert nice value [19,-20] to rlimit style value [1,40] */
4388 int nice_rlim
= 20 - nice
;
4390 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4391 capable(CAP_SYS_NICE
));
4394 #ifdef __ARCH_WANT_SYS_NICE
4397 * sys_nice - change the priority of the current process.
4398 * @increment: priority increment
4400 * sys_setpriority is a more generic, but much slower function that
4401 * does similar things.
4403 SYSCALL_DEFINE1(nice
, int, increment
)
4408 * Setpriority might change our priority at the same moment.
4409 * We don't have to worry. Conceptually one call occurs first
4410 * and we have a single winner.
4412 if (increment
< -40)
4417 nice
= TASK_NICE(current
) + increment
;
4423 if (increment
< 0 && !can_nice(current
, nice
))
4426 retval
= security_task_setnice(current
, nice
);
4430 set_user_nice(current
, nice
);
4437 * task_prio - return the priority value of a given task.
4438 * @p: the task in question.
4440 * This is the priority value as seen by users in /proc.
4441 * RT tasks are offset by -200. Normal tasks are centered
4442 * around 0, value goes from -16 to +15.
4444 int task_prio(const struct task_struct
*p
)
4446 return p
->prio
- MAX_RT_PRIO
;
4450 * task_nice - return the nice value of a given task.
4451 * @p: the task in question.
4453 int task_nice(const struct task_struct
*p
)
4455 return TASK_NICE(p
);
4457 EXPORT_SYMBOL(task_nice
);
4460 * idle_cpu - is a given cpu idle currently?
4461 * @cpu: the processor in question.
4463 int idle_cpu(int cpu
)
4465 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4469 * idle_task - return the idle task for a given cpu.
4470 * @cpu: the processor in question.
4472 struct task_struct
*idle_task(int cpu
)
4474 return cpu_rq(cpu
)->idle
;
4478 * find_process_by_pid - find a process with a matching PID value.
4479 * @pid: the pid in question.
4481 static struct task_struct
*find_process_by_pid(pid_t pid
)
4483 return pid
? find_task_by_vpid(pid
) : current
;
4486 /* Actually do priority change: must hold rq lock. */
4488 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4490 BUG_ON(p
->se
.on_rq
);
4493 p
->rt_priority
= prio
;
4494 p
->normal_prio
= normal_prio(p
);
4495 /* we are holding p->pi_lock already */
4496 p
->prio
= rt_mutex_getprio(p
);
4497 if (rt_prio(p
->prio
))
4498 p
->sched_class
= &rt_sched_class
;
4500 p
->sched_class
= &fair_sched_class
;
4505 * check the target process has a UID that matches the current process's
4507 static bool check_same_owner(struct task_struct
*p
)
4509 const struct cred
*cred
= current_cred(), *pcred
;
4513 pcred
= __task_cred(p
);
4514 match
= (cred
->euid
== pcred
->euid
||
4515 cred
->euid
== pcred
->uid
);
4520 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4521 struct sched_param
*param
, bool user
)
4523 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4524 unsigned long flags
;
4525 const struct sched_class
*prev_class
;
4529 /* may grab non-irq protected spin_locks */
4530 BUG_ON(in_interrupt());
4532 /* double check policy once rq lock held */
4534 reset_on_fork
= p
->sched_reset_on_fork
;
4535 policy
= oldpolicy
= p
->policy
;
4537 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4538 policy
&= ~SCHED_RESET_ON_FORK
;
4540 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4541 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4542 policy
!= SCHED_IDLE
)
4547 * Valid priorities for SCHED_FIFO and SCHED_RR are
4548 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4549 * SCHED_BATCH and SCHED_IDLE is 0.
4551 if (param
->sched_priority
< 0 ||
4552 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4553 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4555 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4559 * Allow unprivileged RT tasks to decrease priority:
4561 if (user
&& !capable(CAP_SYS_NICE
)) {
4562 if (rt_policy(policy
)) {
4563 unsigned long rlim_rtprio
;
4565 if (!lock_task_sighand(p
, &flags
))
4567 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4568 unlock_task_sighand(p
, &flags
);
4570 /* can't set/change the rt policy */
4571 if (policy
!= p
->policy
&& !rlim_rtprio
)
4574 /* can't increase priority */
4575 if (param
->sched_priority
> p
->rt_priority
&&
4576 param
->sched_priority
> rlim_rtprio
)
4580 * Like positive nice levels, dont allow tasks to
4581 * move out of SCHED_IDLE either:
4583 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4586 /* can't change other user's priorities */
4587 if (!check_same_owner(p
))
4590 /* Normal users shall not reset the sched_reset_on_fork flag */
4591 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4596 retval
= security_task_setscheduler(p
, policy
, param
);
4602 * make sure no PI-waiters arrive (or leave) while we are
4603 * changing the priority of the task:
4605 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4607 * To be able to change p->policy safely, the apropriate
4608 * runqueue lock must be held.
4610 rq
= __task_rq_lock(p
);
4612 #ifdef CONFIG_RT_GROUP_SCHED
4615 * Do not allow realtime tasks into groups that have no runtime
4618 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4619 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4620 __task_rq_unlock(rq
);
4621 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4627 /* recheck policy now with rq lock held */
4628 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4629 policy
= oldpolicy
= -1;
4630 __task_rq_unlock(rq
);
4631 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4634 on_rq
= p
->se
.on_rq
;
4635 running
= task_current(rq
, p
);
4637 deactivate_task(rq
, p
, 0);
4639 p
->sched_class
->put_prev_task(rq
, p
);
4641 p
->sched_reset_on_fork
= reset_on_fork
;
4644 prev_class
= p
->sched_class
;
4645 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4648 p
->sched_class
->set_curr_task(rq
);
4650 activate_task(rq
, p
, 0);
4652 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4654 __task_rq_unlock(rq
);
4655 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4657 rt_mutex_adjust_pi(p
);
4663 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4664 * @p: the task in question.
4665 * @policy: new policy.
4666 * @param: structure containing the new RT priority.
4668 * NOTE that the task may be already dead.
4670 int sched_setscheduler(struct task_struct
*p
, int policy
,
4671 struct sched_param
*param
)
4673 return __sched_setscheduler(p
, policy
, param
, true);
4675 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4678 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4679 * @p: the task in question.
4680 * @policy: new policy.
4681 * @param: structure containing the new RT priority.
4683 * Just like sched_setscheduler, only don't bother checking if the
4684 * current context has permission. For example, this is needed in
4685 * stop_machine(): we create temporary high priority worker threads,
4686 * but our caller might not have that capability.
4688 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4689 struct sched_param
*param
)
4691 return __sched_setscheduler(p
, policy
, param
, false);
4695 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4697 struct sched_param lparam
;
4698 struct task_struct
*p
;
4701 if (!param
|| pid
< 0)
4703 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4708 p
= find_process_by_pid(pid
);
4710 retval
= sched_setscheduler(p
, policy
, &lparam
);
4717 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4718 * @pid: the pid in question.
4719 * @policy: new policy.
4720 * @param: structure containing the new RT priority.
4722 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4723 struct sched_param __user
*, param
)
4725 /* negative values for policy are not valid */
4729 return do_sched_setscheduler(pid
, policy
, param
);
4733 * sys_sched_setparam - set/change the RT priority of a thread
4734 * @pid: the pid in question.
4735 * @param: structure containing the new RT priority.
4737 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4739 return do_sched_setscheduler(pid
, -1, param
);
4743 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4744 * @pid: the pid in question.
4746 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4748 struct task_struct
*p
;
4756 p
= find_process_by_pid(pid
);
4758 retval
= security_task_getscheduler(p
);
4761 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4768 * sys_sched_getparam - get the RT priority of a thread
4769 * @pid: the pid in question.
4770 * @param: structure containing the RT priority.
4772 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4774 struct sched_param lp
;
4775 struct task_struct
*p
;
4778 if (!param
|| pid
< 0)
4782 p
= find_process_by_pid(pid
);
4787 retval
= security_task_getscheduler(p
);
4791 lp
.sched_priority
= p
->rt_priority
;
4795 * This one might sleep, we cannot do it with a spinlock held ...
4797 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4806 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4808 cpumask_var_t cpus_allowed
, new_mask
;
4809 struct task_struct
*p
;
4815 p
= find_process_by_pid(pid
);
4822 /* Prevent p going away */
4826 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4830 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4832 goto out_free_cpus_allowed
;
4835 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4838 retval
= security_task_setscheduler(p
, 0, NULL
);
4842 cpuset_cpus_allowed(p
, cpus_allowed
);
4843 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4845 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4848 cpuset_cpus_allowed(p
, cpus_allowed
);
4849 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4851 * We must have raced with a concurrent cpuset
4852 * update. Just reset the cpus_allowed to the
4853 * cpuset's cpus_allowed
4855 cpumask_copy(new_mask
, cpus_allowed
);
4860 free_cpumask_var(new_mask
);
4861 out_free_cpus_allowed
:
4862 free_cpumask_var(cpus_allowed
);
4869 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4870 struct cpumask
*new_mask
)
4872 if (len
< cpumask_size())
4873 cpumask_clear(new_mask
);
4874 else if (len
> cpumask_size())
4875 len
= cpumask_size();
4877 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4881 * sys_sched_setaffinity - set the cpu affinity of a process
4882 * @pid: pid of the process
4883 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4884 * @user_mask_ptr: user-space pointer to the new cpu mask
4886 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4887 unsigned long __user
*, user_mask_ptr
)
4889 cpumask_var_t new_mask
;
4892 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4895 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4897 retval
= sched_setaffinity(pid
, new_mask
);
4898 free_cpumask_var(new_mask
);
4902 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4904 struct task_struct
*p
;
4905 unsigned long flags
;
4913 p
= find_process_by_pid(pid
);
4917 retval
= security_task_getscheduler(p
);
4921 rq
= task_rq_lock(p
, &flags
);
4922 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4923 task_rq_unlock(rq
, &flags
);
4933 * sys_sched_getaffinity - get the cpu affinity of a process
4934 * @pid: pid of the process
4935 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4936 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4938 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4939 unsigned long __user
*, user_mask_ptr
)
4944 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4946 if (len
& (sizeof(unsigned long)-1))
4949 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4952 ret
= sched_getaffinity(pid
, mask
);
4954 size_t retlen
= min_t(size_t, len
, cpumask_size());
4956 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4961 free_cpumask_var(mask
);
4967 * sys_sched_yield - yield the current processor to other threads.
4969 * This function yields the current CPU to other tasks. If there are no
4970 * other threads running on this CPU then this function will return.
4972 SYSCALL_DEFINE0(sched_yield
)
4974 struct rq
*rq
= this_rq_lock();
4976 schedstat_inc(rq
, yld_count
);
4977 current
->sched_class
->yield_task(rq
);
4980 * Since we are going to call schedule() anyway, there's
4981 * no need to preempt or enable interrupts:
4983 __release(rq
->lock
);
4984 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4985 do_raw_spin_unlock(&rq
->lock
);
4986 preempt_enable_no_resched();
4993 static inline int should_resched(void)
4995 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4998 static void __cond_resched(void)
5000 add_preempt_count(PREEMPT_ACTIVE
);
5002 sub_preempt_count(PREEMPT_ACTIVE
);
5005 int __sched
_cond_resched(void)
5007 if (should_resched()) {
5013 EXPORT_SYMBOL(_cond_resched
);
5016 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5017 * call schedule, and on return reacquire the lock.
5019 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5020 * operations here to prevent schedule() from being called twice (once via
5021 * spin_unlock(), once by hand).
5023 int __cond_resched_lock(spinlock_t
*lock
)
5025 int resched
= should_resched();
5028 lockdep_assert_held(lock
);
5030 if (spin_needbreak(lock
) || resched
) {
5041 EXPORT_SYMBOL(__cond_resched_lock
);
5043 int __sched
__cond_resched_softirq(void)
5045 BUG_ON(!in_softirq());
5047 if (should_resched()) {
5055 EXPORT_SYMBOL(__cond_resched_softirq
);
5058 * yield - yield the current processor to other threads.
5060 * This is a shortcut for kernel-space yielding - it marks the
5061 * thread runnable and calls sys_sched_yield().
5063 void __sched
yield(void)
5065 set_current_state(TASK_RUNNING
);
5068 EXPORT_SYMBOL(yield
);
5071 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5072 * that process accounting knows that this is a task in IO wait state.
5074 void __sched
io_schedule(void)
5076 struct rq
*rq
= raw_rq();
5078 delayacct_blkio_start();
5079 atomic_inc(&rq
->nr_iowait
);
5080 current
->in_iowait
= 1;
5082 current
->in_iowait
= 0;
5083 atomic_dec(&rq
->nr_iowait
);
5084 delayacct_blkio_end();
5086 EXPORT_SYMBOL(io_schedule
);
5088 long __sched
io_schedule_timeout(long timeout
)
5090 struct rq
*rq
= raw_rq();
5093 delayacct_blkio_start();
5094 atomic_inc(&rq
->nr_iowait
);
5095 current
->in_iowait
= 1;
5096 ret
= schedule_timeout(timeout
);
5097 current
->in_iowait
= 0;
5098 atomic_dec(&rq
->nr_iowait
);
5099 delayacct_blkio_end();
5104 * sys_sched_get_priority_max - return maximum RT priority.
5105 * @policy: scheduling class.
5107 * this syscall returns the maximum rt_priority that can be used
5108 * by a given scheduling class.
5110 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5117 ret
= MAX_USER_RT_PRIO
-1;
5129 * sys_sched_get_priority_min - return minimum RT priority.
5130 * @policy: scheduling class.
5132 * this syscall returns the minimum rt_priority that can be used
5133 * by a given scheduling class.
5135 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5153 * sys_sched_rr_get_interval - return the default timeslice of a process.
5154 * @pid: pid of the process.
5155 * @interval: userspace pointer to the timeslice value.
5157 * this syscall writes the default timeslice value of a given process
5158 * into the user-space timespec buffer. A value of '0' means infinity.
5160 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5161 struct timespec __user
*, interval
)
5163 struct task_struct
*p
;
5164 unsigned int time_slice
;
5165 unsigned long flags
;
5175 p
= find_process_by_pid(pid
);
5179 retval
= security_task_getscheduler(p
);
5183 rq
= task_rq_lock(p
, &flags
);
5184 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5185 task_rq_unlock(rq
, &flags
);
5188 jiffies_to_timespec(time_slice
, &t
);
5189 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5197 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5199 void sched_show_task(struct task_struct
*p
)
5201 unsigned long free
= 0;
5204 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5205 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5206 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5207 #if BITS_PER_LONG == 32
5208 if (state
== TASK_RUNNING
)
5209 printk(KERN_CONT
" running ");
5211 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5213 if (state
== TASK_RUNNING
)
5214 printk(KERN_CONT
" running task ");
5216 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5218 #ifdef CONFIG_DEBUG_STACK_USAGE
5219 free
= stack_not_used(p
);
5221 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5222 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5223 (unsigned long)task_thread_info(p
)->flags
);
5225 show_stack(p
, NULL
);
5228 void show_state_filter(unsigned long state_filter
)
5230 struct task_struct
*g
, *p
;
5232 #if BITS_PER_LONG == 32
5234 " task PC stack pid father\n");
5237 " task PC stack pid father\n");
5239 read_lock(&tasklist_lock
);
5240 do_each_thread(g
, p
) {
5242 * reset the NMI-timeout, listing all files on a slow
5243 * console might take alot of time:
5245 touch_nmi_watchdog();
5246 if (!state_filter
|| (p
->state
& state_filter
))
5248 } while_each_thread(g
, p
);
5250 touch_all_softlockup_watchdogs();
5252 #ifdef CONFIG_SCHED_DEBUG
5253 sysrq_sched_debug_show();
5255 read_unlock(&tasklist_lock
);
5257 * Only show locks if all tasks are dumped:
5260 debug_show_all_locks();
5263 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5265 idle
->sched_class
= &idle_sched_class
;
5269 * init_idle - set up an idle thread for a given CPU
5270 * @idle: task in question
5271 * @cpu: cpu the idle task belongs to
5273 * NOTE: this function does not set the idle thread's NEED_RESCHED
5274 * flag, to make booting more robust.
5276 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5278 struct rq
*rq
= cpu_rq(cpu
);
5279 unsigned long flags
;
5281 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5284 idle
->state
= TASK_RUNNING
;
5285 idle
->se
.exec_start
= sched_clock();
5287 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5289 * We're having a chicken and egg problem, even though we are
5290 * holding rq->lock, the cpu isn't yet set to this cpu so the
5291 * lockdep check in task_group() will fail.
5293 * Similar case to sched_fork(). / Alternatively we could
5294 * use task_rq_lock() here and obtain the other rq->lock.
5299 __set_task_cpu(idle
, cpu
);
5302 rq
->curr
= rq
->idle
= idle
;
5303 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5306 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5308 /* Set the preempt count _outside_ the spinlocks! */
5309 #if defined(CONFIG_PREEMPT)
5310 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5312 task_thread_info(idle
)->preempt_count
= 0;
5315 * The idle tasks have their own, simple scheduling class:
5317 idle
->sched_class
= &idle_sched_class
;
5318 ftrace_graph_init_task(idle
);
5322 * In a system that switches off the HZ timer nohz_cpu_mask
5323 * indicates which cpus entered this state. This is used
5324 * in the rcu update to wait only for active cpus. For system
5325 * which do not switch off the HZ timer nohz_cpu_mask should
5326 * always be CPU_BITS_NONE.
5328 cpumask_var_t nohz_cpu_mask
;
5331 * Increase the granularity value when there are more CPUs,
5332 * because with more CPUs the 'effective latency' as visible
5333 * to users decreases. But the relationship is not linear,
5334 * so pick a second-best guess by going with the log2 of the
5337 * This idea comes from the SD scheduler of Con Kolivas:
5339 static int get_update_sysctl_factor(void)
5341 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5342 unsigned int factor
;
5344 switch (sysctl_sched_tunable_scaling
) {
5345 case SCHED_TUNABLESCALING_NONE
:
5348 case SCHED_TUNABLESCALING_LINEAR
:
5351 case SCHED_TUNABLESCALING_LOG
:
5353 factor
= 1 + ilog2(cpus
);
5360 static void update_sysctl(void)
5362 unsigned int factor
= get_update_sysctl_factor();
5364 #define SET_SYSCTL(name) \
5365 (sysctl_##name = (factor) * normalized_sysctl_##name)
5366 SET_SYSCTL(sched_min_granularity
);
5367 SET_SYSCTL(sched_latency
);
5368 SET_SYSCTL(sched_wakeup_granularity
);
5369 SET_SYSCTL(sched_shares_ratelimit
);
5373 static inline void sched_init_granularity(void)
5380 * This is how migration works:
5382 * 1) we invoke migration_cpu_stop() on the target CPU using
5384 * 2) stopper starts to run (implicitly forcing the migrated thread
5386 * 3) it checks whether the migrated task is still in the wrong runqueue.
5387 * 4) if it's in the wrong runqueue then the migration thread removes
5388 * it and puts it into the right queue.
5389 * 5) stopper completes and stop_one_cpu() returns and the migration
5394 * Change a given task's CPU affinity. Migrate the thread to a
5395 * proper CPU and schedule it away if the CPU it's executing on
5396 * is removed from the allowed bitmask.
5398 * NOTE: the caller must have a valid reference to the task, the
5399 * task must not exit() & deallocate itself prematurely. The
5400 * call is not atomic; no spinlocks may be held.
5402 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5404 unsigned long flags
;
5406 unsigned int dest_cpu
;
5410 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5411 * drop the rq->lock and still rely on ->cpus_allowed.
5414 while (task_is_waking(p
))
5416 rq
= task_rq_lock(p
, &flags
);
5417 if (task_is_waking(p
)) {
5418 task_rq_unlock(rq
, &flags
);
5422 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5427 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5428 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5433 if (p
->sched_class
->set_cpus_allowed
)
5434 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5436 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5437 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5440 /* Can the task run on the task's current CPU? If so, we're done */
5441 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5444 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5445 if (migrate_task(p
, dest_cpu
)) {
5446 struct migration_arg arg
= { p
, dest_cpu
};
5447 /* Need help from migration thread: drop lock and wait. */
5448 task_rq_unlock(rq
, &flags
);
5449 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5450 tlb_migrate_finish(p
->mm
);
5454 task_rq_unlock(rq
, &flags
);
5458 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5461 * Move (not current) task off this cpu, onto dest cpu. We're doing
5462 * this because either it can't run here any more (set_cpus_allowed()
5463 * away from this CPU, or CPU going down), or because we're
5464 * attempting to rebalance this task on exec (sched_exec).
5466 * So we race with normal scheduler movements, but that's OK, as long
5467 * as the task is no longer on this CPU.
5469 * Returns non-zero if task was successfully migrated.
5471 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5473 struct rq
*rq_dest
, *rq_src
;
5476 if (unlikely(!cpu_active(dest_cpu
)))
5479 rq_src
= cpu_rq(src_cpu
);
5480 rq_dest
= cpu_rq(dest_cpu
);
5482 double_rq_lock(rq_src
, rq_dest
);
5483 /* Already moved. */
5484 if (task_cpu(p
) != src_cpu
)
5486 /* Affinity changed (again). */
5487 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5491 * If we're not on a rq, the next wake-up will ensure we're
5495 deactivate_task(rq_src
, p
, 0);
5496 set_task_cpu(p
, dest_cpu
);
5497 activate_task(rq_dest
, p
, 0);
5498 check_preempt_curr(rq_dest
, p
, 0);
5503 double_rq_unlock(rq_src
, rq_dest
);
5508 * migration_cpu_stop - this will be executed by a highprio stopper thread
5509 * and performs thread migration by bumping thread off CPU then
5510 * 'pushing' onto another runqueue.
5512 static int migration_cpu_stop(void *data
)
5514 struct migration_arg
*arg
= data
;
5517 * The original target cpu might have gone down and we might
5518 * be on another cpu but it doesn't matter.
5520 local_irq_disable();
5521 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5526 #ifdef CONFIG_HOTPLUG_CPU
5528 * Figure out where task on dead CPU should go, use force if necessary.
5530 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5532 struct rq
*rq
= cpu_rq(dead_cpu
);
5533 int needs_cpu
, uninitialized_var(dest_cpu
);
5534 unsigned long flags
;
5536 local_irq_save(flags
);
5538 raw_spin_lock(&rq
->lock
);
5539 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5541 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5542 raw_spin_unlock(&rq
->lock
);
5544 * It can only fail if we race with set_cpus_allowed(),
5545 * in the racer should migrate the task anyway.
5548 __migrate_task(p
, dead_cpu
, dest_cpu
);
5549 local_irq_restore(flags
);
5553 * While a dead CPU has no uninterruptible tasks queued at this point,
5554 * it might still have a nonzero ->nr_uninterruptible counter, because
5555 * for performance reasons the counter is not stricly tracking tasks to
5556 * their home CPUs. So we just add the counter to another CPU's counter,
5557 * to keep the global sum constant after CPU-down:
5559 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5561 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5562 unsigned long flags
;
5564 local_irq_save(flags
);
5565 double_rq_lock(rq_src
, rq_dest
);
5566 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5567 rq_src
->nr_uninterruptible
= 0;
5568 double_rq_unlock(rq_src
, rq_dest
);
5569 local_irq_restore(flags
);
5572 /* Run through task list and migrate tasks from the dead cpu. */
5573 static void migrate_live_tasks(int src_cpu
)
5575 struct task_struct
*p
, *t
;
5577 read_lock(&tasklist_lock
);
5579 do_each_thread(t
, p
) {
5583 if (task_cpu(p
) == src_cpu
)
5584 move_task_off_dead_cpu(src_cpu
, p
);
5585 } while_each_thread(t
, p
);
5587 read_unlock(&tasklist_lock
);
5591 * Schedules idle task to be the next runnable task on current CPU.
5592 * It does so by boosting its priority to highest possible.
5593 * Used by CPU offline code.
5595 void sched_idle_next(void)
5597 int this_cpu
= smp_processor_id();
5598 struct rq
*rq
= cpu_rq(this_cpu
);
5599 struct task_struct
*p
= rq
->idle
;
5600 unsigned long flags
;
5602 /* cpu has to be offline */
5603 BUG_ON(cpu_online(this_cpu
));
5606 * Strictly not necessary since rest of the CPUs are stopped by now
5607 * and interrupts disabled on the current cpu.
5609 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5611 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5613 activate_task(rq
, p
, 0);
5615 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5619 * Ensures that the idle task is using init_mm right before its cpu goes
5622 void idle_task_exit(void)
5624 struct mm_struct
*mm
= current
->active_mm
;
5626 BUG_ON(cpu_online(smp_processor_id()));
5629 switch_mm(mm
, &init_mm
, current
);
5633 /* called under rq->lock with disabled interrupts */
5634 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5636 struct rq
*rq
= cpu_rq(dead_cpu
);
5638 /* Must be exiting, otherwise would be on tasklist. */
5639 BUG_ON(!p
->exit_state
);
5641 /* Cannot have done final schedule yet: would have vanished. */
5642 BUG_ON(p
->state
== TASK_DEAD
);
5647 * Drop lock around migration; if someone else moves it,
5648 * that's OK. No task can be added to this CPU, so iteration is
5651 raw_spin_unlock_irq(&rq
->lock
);
5652 move_task_off_dead_cpu(dead_cpu
, p
);
5653 raw_spin_lock_irq(&rq
->lock
);
5658 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5659 static void migrate_dead_tasks(unsigned int dead_cpu
)
5661 struct rq
*rq
= cpu_rq(dead_cpu
);
5662 struct task_struct
*next
;
5665 if (!rq
->nr_running
)
5667 next
= pick_next_task(rq
);
5670 next
->sched_class
->put_prev_task(rq
, next
);
5671 migrate_dead(dead_cpu
, next
);
5677 * remove the tasks which were accounted by rq from calc_load_tasks.
5679 static void calc_global_load_remove(struct rq
*rq
)
5681 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5682 rq
->calc_load_active
= 0;
5684 #endif /* CONFIG_HOTPLUG_CPU */
5686 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5688 static struct ctl_table sd_ctl_dir
[] = {
5690 .procname
= "sched_domain",
5696 static struct ctl_table sd_ctl_root
[] = {
5698 .procname
= "kernel",
5700 .child
= sd_ctl_dir
,
5705 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5707 struct ctl_table
*entry
=
5708 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5713 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5715 struct ctl_table
*entry
;
5718 * In the intermediate directories, both the child directory and
5719 * procname are dynamically allocated and could fail but the mode
5720 * will always be set. In the lowest directory the names are
5721 * static strings and all have proc handlers.
5723 for (entry
= *tablep
; entry
->mode
; entry
++) {
5725 sd_free_ctl_entry(&entry
->child
);
5726 if (entry
->proc_handler
== NULL
)
5727 kfree(entry
->procname
);
5735 set_table_entry(struct ctl_table
*entry
,
5736 const char *procname
, void *data
, int maxlen
,
5737 mode_t mode
, proc_handler
*proc_handler
)
5739 entry
->procname
= procname
;
5741 entry
->maxlen
= maxlen
;
5743 entry
->proc_handler
= proc_handler
;
5746 static struct ctl_table
*
5747 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5749 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5754 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5755 sizeof(long), 0644, proc_doulongvec_minmax
);
5756 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5757 sizeof(long), 0644, proc_doulongvec_minmax
);
5758 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5759 sizeof(int), 0644, proc_dointvec_minmax
);
5760 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5761 sizeof(int), 0644, proc_dointvec_minmax
);
5762 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5763 sizeof(int), 0644, proc_dointvec_minmax
);
5764 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5765 sizeof(int), 0644, proc_dointvec_minmax
);
5766 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5767 sizeof(int), 0644, proc_dointvec_minmax
);
5768 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5769 sizeof(int), 0644, proc_dointvec_minmax
);
5770 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5771 sizeof(int), 0644, proc_dointvec_minmax
);
5772 set_table_entry(&table
[9], "cache_nice_tries",
5773 &sd
->cache_nice_tries
,
5774 sizeof(int), 0644, proc_dointvec_minmax
);
5775 set_table_entry(&table
[10], "flags", &sd
->flags
,
5776 sizeof(int), 0644, proc_dointvec_minmax
);
5777 set_table_entry(&table
[11], "name", sd
->name
,
5778 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5779 /* &table[12] is terminator */
5784 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5786 struct ctl_table
*entry
, *table
;
5787 struct sched_domain
*sd
;
5788 int domain_num
= 0, i
;
5791 for_each_domain(cpu
, sd
)
5793 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5798 for_each_domain(cpu
, sd
) {
5799 snprintf(buf
, 32, "domain%d", i
);
5800 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5802 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5809 static struct ctl_table_header
*sd_sysctl_header
;
5810 static void register_sched_domain_sysctl(void)
5812 int i
, cpu_num
= num_possible_cpus();
5813 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5816 WARN_ON(sd_ctl_dir
[0].child
);
5817 sd_ctl_dir
[0].child
= entry
;
5822 for_each_possible_cpu(i
) {
5823 snprintf(buf
, 32, "cpu%d", i
);
5824 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5826 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5830 WARN_ON(sd_sysctl_header
);
5831 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5834 /* may be called multiple times per register */
5835 static void unregister_sched_domain_sysctl(void)
5837 if (sd_sysctl_header
)
5838 unregister_sysctl_table(sd_sysctl_header
);
5839 sd_sysctl_header
= NULL
;
5840 if (sd_ctl_dir
[0].child
)
5841 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5844 static void register_sched_domain_sysctl(void)
5847 static void unregister_sched_domain_sysctl(void)
5852 static void set_rq_online(struct rq
*rq
)
5855 const struct sched_class
*class;
5857 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5860 for_each_class(class) {
5861 if (class->rq_online
)
5862 class->rq_online(rq
);
5867 static void set_rq_offline(struct rq
*rq
)
5870 const struct sched_class
*class;
5872 for_each_class(class) {
5873 if (class->rq_offline
)
5874 class->rq_offline(rq
);
5877 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5883 * migration_call - callback that gets triggered when a CPU is added.
5884 * Here we can start up the necessary migration thread for the new CPU.
5886 static int __cpuinit
5887 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5889 int cpu
= (long)hcpu
;
5890 unsigned long flags
;
5891 struct rq
*rq
= cpu_rq(cpu
);
5895 case CPU_UP_PREPARE
:
5896 case CPU_UP_PREPARE_FROZEN
:
5897 rq
->calc_load_update
= calc_load_update
;
5901 case CPU_ONLINE_FROZEN
:
5902 /* Update our root-domain */
5903 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5905 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5909 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5912 #ifdef CONFIG_HOTPLUG_CPU
5914 case CPU_DEAD_FROZEN
:
5915 migrate_live_tasks(cpu
);
5916 /* Idle task back to normal (off runqueue, low prio) */
5917 raw_spin_lock_irq(&rq
->lock
);
5918 deactivate_task(rq
, rq
->idle
, 0);
5919 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5920 rq
->idle
->sched_class
= &idle_sched_class
;
5921 migrate_dead_tasks(cpu
);
5922 raw_spin_unlock_irq(&rq
->lock
);
5923 migrate_nr_uninterruptible(rq
);
5924 BUG_ON(rq
->nr_running
!= 0);
5925 calc_global_load_remove(rq
);
5929 case CPU_DYING_FROZEN
:
5930 /* Update our root-domain */
5931 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5933 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5936 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5944 * Register at high priority so that task migration (migrate_all_tasks)
5945 * happens before everything else. This has to be lower priority than
5946 * the notifier in the perf_event subsystem, though.
5948 static struct notifier_block __cpuinitdata migration_notifier
= {
5949 .notifier_call
= migration_call
,
5953 static int __init
migration_init(void)
5955 void *cpu
= (void *)(long)smp_processor_id();
5958 /* Start one for the boot CPU: */
5959 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5960 BUG_ON(err
== NOTIFY_BAD
);
5961 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5962 register_cpu_notifier(&migration_notifier
);
5966 early_initcall(migration_init
);
5971 #ifdef CONFIG_SCHED_DEBUG
5973 static __read_mostly
int sched_domain_debug_enabled
;
5975 static int __init
sched_domain_debug_setup(char *str
)
5977 sched_domain_debug_enabled
= 1;
5981 early_param("sched_debug", sched_domain_debug_setup
);
5983 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5984 struct cpumask
*groupmask
)
5986 struct sched_group
*group
= sd
->groups
;
5989 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5990 cpumask_clear(groupmask
);
5992 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5994 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5995 printk("does not load-balance\n");
5997 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6002 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6004 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6005 printk(KERN_ERR
"ERROR: domain->span does not contain "
6008 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6009 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6013 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6017 printk(KERN_ERR
"ERROR: group is NULL\n");
6021 if (!group
->cpu_power
) {
6022 printk(KERN_CONT
"\n");
6023 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6028 if (!cpumask_weight(sched_group_cpus(group
))) {
6029 printk(KERN_CONT
"\n");
6030 printk(KERN_ERR
"ERROR: empty group\n");
6034 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6035 printk(KERN_CONT
"\n");
6036 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6040 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6042 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6044 printk(KERN_CONT
" %s", str
);
6045 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6046 printk(KERN_CONT
" (cpu_power = %d)",
6050 group
= group
->next
;
6051 } while (group
!= sd
->groups
);
6052 printk(KERN_CONT
"\n");
6054 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6055 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6058 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6059 printk(KERN_ERR
"ERROR: parent span is not a superset "
6060 "of domain->span\n");
6064 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6066 cpumask_var_t groupmask
;
6069 if (!sched_domain_debug_enabled
)
6073 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6077 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6079 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6080 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6085 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6092 free_cpumask_var(groupmask
);
6094 #else /* !CONFIG_SCHED_DEBUG */
6095 # define sched_domain_debug(sd, cpu) do { } while (0)
6096 #endif /* CONFIG_SCHED_DEBUG */
6098 static int sd_degenerate(struct sched_domain
*sd
)
6100 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6103 /* Following flags need at least 2 groups */
6104 if (sd
->flags
& (SD_LOAD_BALANCE
|
6105 SD_BALANCE_NEWIDLE
|
6109 SD_SHARE_PKG_RESOURCES
)) {
6110 if (sd
->groups
!= sd
->groups
->next
)
6114 /* Following flags don't use groups */
6115 if (sd
->flags
& (SD_WAKE_AFFINE
))
6122 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6124 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6126 if (sd_degenerate(parent
))
6129 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6132 /* Flags needing groups don't count if only 1 group in parent */
6133 if (parent
->groups
== parent
->groups
->next
) {
6134 pflags
&= ~(SD_LOAD_BALANCE
|
6135 SD_BALANCE_NEWIDLE
|
6139 SD_SHARE_PKG_RESOURCES
);
6140 if (nr_node_ids
== 1)
6141 pflags
&= ~SD_SERIALIZE
;
6143 if (~cflags
& pflags
)
6149 static void free_rootdomain(struct root_domain
*rd
)
6151 synchronize_sched();
6153 cpupri_cleanup(&rd
->cpupri
);
6155 free_cpumask_var(rd
->rto_mask
);
6156 free_cpumask_var(rd
->online
);
6157 free_cpumask_var(rd
->span
);
6161 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6163 struct root_domain
*old_rd
= NULL
;
6164 unsigned long flags
;
6166 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6171 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6174 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6177 * If we dont want to free the old_rt yet then
6178 * set old_rd to NULL to skip the freeing later
6181 if (!atomic_dec_and_test(&old_rd
->refcount
))
6185 atomic_inc(&rd
->refcount
);
6188 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6189 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6192 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6195 free_rootdomain(old_rd
);
6198 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6200 gfp_t gfp
= GFP_KERNEL
;
6202 memset(rd
, 0, sizeof(*rd
));
6207 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6209 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6211 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6214 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6219 free_cpumask_var(rd
->rto_mask
);
6221 free_cpumask_var(rd
->online
);
6223 free_cpumask_var(rd
->span
);
6228 static void init_defrootdomain(void)
6230 init_rootdomain(&def_root_domain
, true);
6232 atomic_set(&def_root_domain
.refcount
, 1);
6235 static struct root_domain
*alloc_rootdomain(void)
6237 struct root_domain
*rd
;
6239 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6243 if (init_rootdomain(rd
, false) != 0) {
6252 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6253 * hold the hotplug lock.
6256 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6258 struct rq
*rq
= cpu_rq(cpu
);
6259 struct sched_domain
*tmp
;
6261 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6262 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6264 /* Remove the sched domains which do not contribute to scheduling. */
6265 for (tmp
= sd
; tmp
; ) {
6266 struct sched_domain
*parent
= tmp
->parent
;
6270 if (sd_parent_degenerate(tmp
, parent
)) {
6271 tmp
->parent
= parent
->parent
;
6273 parent
->parent
->child
= tmp
;
6278 if (sd
&& sd_degenerate(sd
)) {
6284 sched_domain_debug(sd
, cpu
);
6286 rq_attach_root(rq
, rd
);
6287 rcu_assign_pointer(rq
->sd
, sd
);
6290 /* cpus with isolated domains */
6291 static cpumask_var_t cpu_isolated_map
;
6293 /* Setup the mask of cpus configured for isolated domains */
6294 static int __init
isolated_cpu_setup(char *str
)
6296 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6297 cpulist_parse(str
, cpu_isolated_map
);
6301 __setup("isolcpus=", isolated_cpu_setup
);
6304 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6305 * to a function which identifies what group(along with sched group) a CPU
6306 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6307 * (due to the fact that we keep track of groups covered with a struct cpumask).
6309 * init_sched_build_groups will build a circular linked list of the groups
6310 * covered by the given span, and will set each group's ->cpumask correctly,
6311 * and ->cpu_power to 0.
6314 init_sched_build_groups(const struct cpumask
*span
,
6315 const struct cpumask
*cpu_map
,
6316 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6317 struct sched_group
**sg
,
6318 struct cpumask
*tmpmask
),
6319 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6321 struct sched_group
*first
= NULL
, *last
= NULL
;
6324 cpumask_clear(covered
);
6326 for_each_cpu(i
, span
) {
6327 struct sched_group
*sg
;
6328 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6331 if (cpumask_test_cpu(i
, covered
))
6334 cpumask_clear(sched_group_cpus(sg
));
6337 for_each_cpu(j
, span
) {
6338 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6341 cpumask_set_cpu(j
, covered
);
6342 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6353 #define SD_NODES_PER_DOMAIN 16
6358 * find_next_best_node - find the next node to include in a sched_domain
6359 * @node: node whose sched_domain we're building
6360 * @used_nodes: nodes already in the sched_domain
6362 * Find the next node to include in a given scheduling domain. Simply
6363 * finds the closest node not already in the @used_nodes map.
6365 * Should use nodemask_t.
6367 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6369 int i
, n
, val
, min_val
, best_node
= 0;
6373 for (i
= 0; i
< nr_node_ids
; i
++) {
6374 /* Start at @node */
6375 n
= (node
+ i
) % nr_node_ids
;
6377 if (!nr_cpus_node(n
))
6380 /* Skip already used nodes */
6381 if (node_isset(n
, *used_nodes
))
6384 /* Simple min distance search */
6385 val
= node_distance(node
, n
);
6387 if (val
< min_val
) {
6393 node_set(best_node
, *used_nodes
);
6398 * sched_domain_node_span - get a cpumask for a node's sched_domain
6399 * @node: node whose cpumask we're constructing
6400 * @span: resulting cpumask
6402 * Given a node, construct a good cpumask for its sched_domain to span. It
6403 * should be one that prevents unnecessary balancing, but also spreads tasks
6406 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6408 nodemask_t used_nodes
;
6411 cpumask_clear(span
);
6412 nodes_clear(used_nodes
);
6414 cpumask_or(span
, span
, cpumask_of_node(node
));
6415 node_set(node
, used_nodes
);
6417 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6418 int next_node
= find_next_best_node(node
, &used_nodes
);
6420 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6423 #endif /* CONFIG_NUMA */
6425 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6428 * The cpus mask in sched_group and sched_domain hangs off the end.
6430 * ( See the the comments in include/linux/sched.h:struct sched_group
6431 * and struct sched_domain. )
6433 struct static_sched_group
{
6434 struct sched_group sg
;
6435 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6438 struct static_sched_domain
{
6439 struct sched_domain sd
;
6440 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6446 cpumask_var_t domainspan
;
6447 cpumask_var_t covered
;
6448 cpumask_var_t notcovered
;
6450 cpumask_var_t nodemask
;
6451 cpumask_var_t this_sibling_map
;
6452 cpumask_var_t this_core_map
;
6453 cpumask_var_t send_covered
;
6454 cpumask_var_t tmpmask
;
6455 struct sched_group
**sched_group_nodes
;
6456 struct root_domain
*rd
;
6460 sa_sched_groups
= 0,
6465 sa_this_sibling_map
,
6467 sa_sched_group_nodes
,
6477 * SMT sched-domains:
6479 #ifdef CONFIG_SCHED_SMT
6480 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6481 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6484 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6485 struct sched_group
**sg
, struct cpumask
*unused
)
6488 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6491 #endif /* CONFIG_SCHED_SMT */
6494 * multi-core sched-domains:
6496 #ifdef CONFIG_SCHED_MC
6497 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6498 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6499 #endif /* CONFIG_SCHED_MC */
6501 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6503 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6504 struct sched_group
**sg
, struct cpumask
*mask
)
6508 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6509 group
= cpumask_first(mask
);
6511 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6514 #elif defined(CONFIG_SCHED_MC)
6516 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6517 struct sched_group
**sg
, struct cpumask
*unused
)
6520 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6525 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6526 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6529 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6530 struct sched_group
**sg
, struct cpumask
*mask
)
6533 #ifdef CONFIG_SCHED_MC
6534 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6535 group
= cpumask_first(mask
);
6536 #elif defined(CONFIG_SCHED_SMT)
6537 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6538 group
= cpumask_first(mask
);
6543 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6549 * The init_sched_build_groups can't handle what we want to do with node
6550 * groups, so roll our own. Now each node has its own list of groups which
6551 * gets dynamically allocated.
6553 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6554 static struct sched_group
***sched_group_nodes_bycpu
;
6556 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6557 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6559 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6560 struct sched_group
**sg
,
6561 struct cpumask
*nodemask
)
6565 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6566 group
= cpumask_first(nodemask
);
6569 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6573 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6575 struct sched_group
*sg
= group_head
;
6581 for_each_cpu(j
, sched_group_cpus(sg
)) {
6582 struct sched_domain
*sd
;
6584 sd
= &per_cpu(phys_domains
, j
).sd
;
6585 if (j
!= group_first_cpu(sd
->groups
)) {
6587 * Only add "power" once for each
6593 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6596 } while (sg
!= group_head
);
6599 static int build_numa_sched_groups(struct s_data
*d
,
6600 const struct cpumask
*cpu_map
, int num
)
6602 struct sched_domain
*sd
;
6603 struct sched_group
*sg
, *prev
;
6606 cpumask_clear(d
->covered
);
6607 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6608 if (cpumask_empty(d
->nodemask
)) {
6609 d
->sched_group_nodes
[num
] = NULL
;
6613 sched_domain_node_span(num
, d
->domainspan
);
6614 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6616 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6619 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6623 d
->sched_group_nodes
[num
] = sg
;
6625 for_each_cpu(j
, d
->nodemask
) {
6626 sd
= &per_cpu(node_domains
, j
).sd
;
6631 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6633 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6636 for (j
= 0; j
< nr_node_ids
; j
++) {
6637 n
= (num
+ j
) % nr_node_ids
;
6638 cpumask_complement(d
->notcovered
, d
->covered
);
6639 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6640 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6641 if (cpumask_empty(d
->tmpmask
))
6643 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6644 if (cpumask_empty(d
->tmpmask
))
6646 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6650 "Can not alloc domain group for node %d\n", j
);
6654 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6655 sg
->next
= prev
->next
;
6656 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6663 #endif /* CONFIG_NUMA */
6666 /* Free memory allocated for various sched_group structures */
6667 static void free_sched_groups(const struct cpumask
*cpu_map
,
6668 struct cpumask
*nodemask
)
6672 for_each_cpu(cpu
, cpu_map
) {
6673 struct sched_group
**sched_group_nodes
6674 = sched_group_nodes_bycpu
[cpu
];
6676 if (!sched_group_nodes
)
6679 for (i
= 0; i
< nr_node_ids
; i
++) {
6680 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6682 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6683 if (cpumask_empty(nodemask
))
6693 if (oldsg
!= sched_group_nodes
[i
])
6696 kfree(sched_group_nodes
);
6697 sched_group_nodes_bycpu
[cpu
] = NULL
;
6700 #else /* !CONFIG_NUMA */
6701 static void free_sched_groups(const struct cpumask
*cpu_map
,
6702 struct cpumask
*nodemask
)
6705 #endif /* CONFIG_NUMA */
6708 * Initialize sched groups cpu_power.
6710 * cpu_power indicates the capacity of sched group, which is used while
6711 * distributing the load between different sched groups in a sched domain.
6712 * Typically cpu_power for all the groups in a sched domain will be same unless
6713 * there are asymmetries in the topology. If there are asymmetries, group
6714 * having more cpu_power will pickup more load compared to the group having
6717 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6719 struct sched_domain
*child
;
6720 struct sched_group
*group
;
6724 WARN_ON(!sd
|| !sd
->groups
);
6726 if (cpu
!= group_first_cpu(sd
->groups
))
6731 sd
->groups
->cpu_power
= 0;
6734 power
= SCHED_LOAD_SCALE
;
6735 weight
= cpumask_weight(sched_domain_span(sd
));
6737 * SMT siblings share the power of a single core.
6738 * Usually multiple threads get a better yield out of
6739 * that one core than a single thread would have,
6740 * reflect that in sd->smt_gain.
6742 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6743 power
*= sd
->smt_gain
;
6745 power
>>= SCHED_LOAD_SHIFT
;
6747 sd
->groups
->cpu_power
+= power
;
6752 * Add cpu_power of each child group to this groups cpu_power.
6754 group
= child
->groups
;
6756 sd
->groups
->cpu_power
+= group
->cpu_power
;
6757 group
= group
->next
;
6758 } while (group
!= child
->groups
);
6762 * Initializers for schedule domains
6763 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6766 #ifdef CONFIG_SCHED_DEBUG
6767 # define SD_INIT_NAME(sd, type) sd->name = #type
6769 # define SD_INIT_NAME(sd, type) do { } while (0)
6772 #define SD_INIT(sd, type) sd_init_##type(sd)
6774 #define SD_INIT_FUNC(type) \
6775 static noinline void sd_init_##type(struct sched_domain *sd) \
6777 memset(sd, 0, sizeof(*sd)); \
6778 *sd = SD_##type##_INIT; \
6779 sd->level = SD_LV_##type; \
6780 SD_INIT_NAME(sd, type); \
6785 SD_INIT_FUNC(ALLNODES
)
6788 #ifdef CONFIG_SCHED_SMT
6789 SD_INIT_FUNC(SIBLING
)
6791 #ifdef CONFIG_SCHED_MC
6795 static int default_relax_domain_level
= -1;
6797 static int __init
setup_relax_domain_level(char *str
)
6801 val
= simple_strtoul(str
, NULL
, 0);
6802 if (val
< SD_LV_MAX
)
6803 default_relax_domain_level
= val
;
6807 __setup("relax_domain_level=", setup_relax_domain_level
);
6809 static void set_domain_attribute(struct sched_domain
*sd
,
6810 struct sched_domain_attr
*attr
)
6814 if (!attr
|| attr
->relax_domain_level
< 0) {
6815 if (default_relax_domain_level
< 0)
6818 request
= default_relax_domain_level
;
6820 request
= attr
->relax_domain_level
;
6821 if (request
< sd
->level
) {
6822 /* turn off idle balance on this domain */
6823 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6825 /* turn on idle balance on this domain */
6826 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6830 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6831 const struct cpumask
*cpu_map
)
6834 case sa_sched_groups
:
6835 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6836 d
->sched_group_nodes
= NULL
;
6838 free_rootdomain(d
->rd
); /* fall through */
6840 free_cpumask_var(d
->tmpmask
); /* fall through */
6841 case sa_send_covered
:
6842 free_cpumask_var(d
->send_covered
); /* fall through */
6843 case sa_this_core_map
:
6844 free_cpumask_var(d
->this_core_map
); /* fall through */
6845 case sa_this_sibling_map
:
6846 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6848 free_cpumask_var(d
->nodemask
); /* fall through */
6849 case sa_sched_group_nodes
:
6851 kfree(d
->sched_group_nodes
); /* fall through */
6853 free_cpumask_var(d
->notcovered
); /* fall through */
6855 free_cpumask_var(d
->covered
); /* fall through */
6857 free_cpumask_var(d
->domainspan
); /* fall through */
6864 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6865 const struct cpumask
*cpu_map
)
6868 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6870 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6871 return sa_domainspan
;
6872 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6874 /* Allocate the per-node list of sched groups */
6875 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6876 sizeof(struct sched_group
*), GFP_KERNEL
);
6877 if (!d
->sched_group_nodes
) {
6878 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6879 return sa_notcovered
;
6881 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6883 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6884 return sa_sched_group_nodes
;
6885 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6887 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6888 return sa_this_sibling_map
;
6889 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6890 return sa_this_core_map
;
6891 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6892 return sa_send_covered
;
6893 d
->rd
= alloc_rootdomain();
6895 printk(KERN_WARNING
"Cannot alloc root domain\n");
6898 return sa_rootdomain
;
6901 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6902 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6904 struct sched_domain
*sd
= NULL
;
6906 struct sched_domain
*parent
;
6909 if (cpumask_weight(cpu_map
) >
6910 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6911 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6912 SD_INIT(sd
, ALLNODES
);
6913 set_domain_attribute(sd
, attr
);
6914 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6915 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6920 sd
= &per_cpu(node_domains
, i
).sd
;
6922 set_domain_attribute(sd
, attr
);
6923 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6924 sd
->parent
= parent
;
6927 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6932 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6933 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6934 struct sched_domain
*parent
, int i
)
6936 struct sched_domain
*sd
;
6937 sd
= &per_cpu(phys_domains
, i
).sd
;
6939 set_domain_attribute(sd
, attr
);
6940 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6941 sd
->parent
= parent
;
6944 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6948 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6949 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6950 struct sched_domain
*parent
, int i
)
6952 struct sched_domain
*sd
= parent
;
6953 #ifdef CONFIG_SCHED_MC
6954 sd
= &per_cpu(core_domains
, i
).sd
;
6956 set_domain_attribute(sd
, attr
);
6957 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6958 sd
->parent
= parent
;
6960 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6965 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6966 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6967 struct sched_domain
*parent
, int i
)
6969 struct sched_domain
*sd
= parent
;
6970 #ifdef CONFIG_SCHED_SMT
6971 sd
= &per_cpu(cpu_domains
, i
).sd
;
6972 SD_INIT(sd
, SIBLING
);
6973 set_domain_attribute(sd
, attr
);
6974 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6975 sd
->parent
= parent
;
6977 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6982 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6983 const struct cpumask
*cpu_map
, int cpu
)
6986 #ifdef CONFIG_SCHED_SMT
6987 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6988 cpumask_and(d
->this_sibling_map
, cpu_map
,
6989 topology_thread_cpumask(cpu
));
6990 if (cpu
== cpumask_first(d
->this_sibling_map
))
6991 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
6993 d
->send_covered
, d
->tmpmask
);
6996 #ifdef CONFIG_SCHED_MC
6997 case SD_LV_MC
: /* set up multi-core groups */
6998 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
6999 if (cpu
== cpumask_first(d
->this_core_map
))
7000 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7002 d
->send_covered
, d
->tmpmask
);
7005 case SD_LV_CPU
: /* set up physical groups */
7006 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7007 if (!cpumask_empty(d
->nodemask
))
7008 init_sched_build_groups(d
->nodemask
, cpu_map
,
7010 d
->send_covered
, d
->tmpmask
);
7013 case SD_LV_ALLNODES
:
7014 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7015 d
->send_covered
, d
->tmpmask
);
7024 * Build sched domains for a given set of cpus and attach the sched domains
7025 * to the individual cpus
7027 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7028 struct sched_domain_attr
*attr
)
7030 enum s_alloc alloc_state
= sa_none
;
7032 struct sched_domain
*sd
;
7038 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7039 if (alloc_state
!= sa_rootdomain
)
7041 alloc_state
= sa_sched_groups
;
7044 * Set up domains for cpus specified by the cpu_map.
7046 for_each_cpu(i
, cpu_map
) {
7047 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7050 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7051 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7052 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7053 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7056 for_each_cpu(i
, cpu_map
) {
7057 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7058 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7061 /* Set up physical groups */
7062 for (i
= 0; i
< nr_node_ids
; i
++)
7063 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7066 /* Set up node groups */
7068 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7070 for (i
= 0; i
< nr_node_ids
; i
++)
7071 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7075 /* Calculate CPU power for physical packages and nodes */
7076 #ifdef CONFIG_SCHED_SMT
7077 for_each_cpu(i
, cpu_map
) {
7078 sd
= &per_cpu(cpu_domains
, i
).sd
;
7079 init_sched_groups_power(i
, sd
);
7082 #ifdef CONFIG_SCHED_MC
7083 for_each_cpu(i
, cpu_map
) {
7084 sd
= &per_cpu(core_domains
, i
).sd
;
7085 init_sched_groups_power(i
, sd
);
7089 for_each_cpu(i
, cpu_map
) {
7090 sd
= &per_cpu(phys_domains
, i
).sd
;
7091 init_sched_groups_power(i
, sd
);
7095 for (i
= 0; i
< nr_node_ids
; i
++)
7096 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7098 if (d
.sd_allnodes
) {
7099 struct sched_group
*sg
;
7101 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7103 init_numa_sched_groups_power(sg
);
7107 /* Attach the domains */
7108 for_each_cpu(i
, cpu_map
) {
7109 #ifdef CONFIG_SCHED_SMT
7110 sd
= &per_cpu(cpu_domains
, i
).sd
;
7111 #elif defined(CONFIG_SCHED_MC)
7112 sd
= &per_cpu(core_domains
, i
).sd
;
7114 sd
= &per_cpu(phys_domains
, i
).sd
;
7116 cpu_attach_domain(sd
, d
.rd
, i
);
7119 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7120 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7124 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7128 static int build_sched_domains(const struct cpumask
*cpu_map
)
7130 return __build_sched_domains(cpu_map
, NULL
);
7133 static cpumask_var_t
*doms_cur
; /* current sched domains */
7134 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7135 static struct sched_domain_attr
*dattr_cur
;
7136 /* attribues of custom domains in 'doms_cur' */
7139 * Special case: If a kmalloc of a doms_cur partition (array of
7140 * cpumask) fails, then fallback to a single sched domain,
7141 * as determined by the single cpumask fallback_doms.
7143 static cpumask_var_t fallback_doms
;
7146 * arch_update_cpu_topology lets virtualized architectures update the
7147 * cpu core maps. It is supposed to return 1 if the topology changed
7148 * or 0 if it stayed the same.
7150 int __attribute__((weak
)) arch_update_cpu_topology(void)
7155 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7158 cpumask_var_t
*doms
;
7160 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7163 for (i
= 0; i
< ndoms
; i
++) {
7164 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7165 free_sched_domains(doms
, i
);
7172 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7175 for (i
= 0; i
< ndoms
; i
++)
7176 free_cpumask_var(doms
[i
]);
7181 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7182 * For now this just excludes isolated cpus, but could be used to
7183 * exclude other special cases in the future.
7185 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7189 arch_update_cpu_topology();
7191 doms_cur
= alloc_sched_domains(ndoms_cur
);
7193 doms_cur
= &fallback_doms
;
7194 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7196 err
= build_sched_domains(doms_cur
[0]);
7197 register_sched_domain_sysctl();
7202 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7203 struct cpumask
*tmpmask
)
7205 free_sched_groups(cpu_map
, tmpmask
);
7209 * Detach sched domains from a group of cpus specified in cpu_map
7210 * These cpus will now be attached to the NULL domain
7212 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7214 /* Save because hotplug lock held. */
7215 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7218 for_each_cpu(i
, cpu_map
)
7219 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7220 synchronize_sched();
7221 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7224 /* handle null as "default" */
7225 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7226 struct sched_domain_attr
*new, int idx_new
)
7228 struct sched_domain_attr tmp
;
7235 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7236 new ? (new + idx_new
) : &tmp
,
7237 sizeof(struct sched_domain_attr
));
7241 * Partition sched domains as specified by the 'ndoms_new'
7242 * cpumasks in the array doms_new[] of cpumasks. This compares
7243 * doms_new[] to the current sched domain partitioning, doms_cur[].
7244 * It destroys each deleted domain and builds each new domain.
7246 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7247 * The masks don't intersect (don't overlap.) We should setup one
7248 * sched domain for each mask. CPUs not in any of the cpumasks will
7249 * not be load balanced. If the same cpumask appears both in the
7250 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7253 * The passed in 'doms_new' should be allocated using
7254 * alloc_sched_domains. This routine takes ownership of it and will
7255 * free_sched_domains it when done with it. If the caller failed the
7256 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7257 * and partition_sched_domains() will fallback to the single partition
7258 * 'fallback_doms', it also forces the domains to be rebuilt.
7260 * If doms_new == NULL it will be replaced with cpu_online_mask.
7261 * ndoms_new == 0 is a special case for destroying existing domains,
7262 * and it will not create the default domain.
7264 * Call with hotplug lock held
7266 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7267 struct sched_domain_attr
*dattr_new
)
7272 mutex_lock(&sched_domains_mutex
);
7274 /* always unregister in case we don't destroy any domains */
7275 unregister_sched_domain_sysctl();
7277 /* Let architecture update cpu core mappings. */
7278 new_topology
= arch_update_cpu_topology();
7280 n
= doms_new
? ndoms_new
: 0;
7282 /* Destroy deleted domains */
7283 for (i
= 0; i
< ndoms_cur
; i
++) {
7284 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7285 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7286 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7289 /* no match - a current sched domain not in new doms_new[] */
7290 detach_destroy_domains(doms_cur
[i
]);
7295 if (doms_new
== NULL
) {
7297 doms_new
= &fallback_doms
;
7298 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7299 WARN_ON_ONCE(dattr_new
);
7302 /* Build new domains */
7303 for (i
= 0; i
< ndoms_new
; i
++) {
7304 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7305 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7306 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7309 /* no match - add a new doms_new */
7310 __build_sched_domains(doms_new
[i
],
7311 dattr_new
? dattr_new
+ i
: NULL
);
7316 /* Remember the new sched domains */
7317 if (doms_cur
!= &fallback_doms
)
7318 free_sched_domains(doms_cur
, ndoms_cur
);
7319 kfree(dattr_cur
); /* kfree(NULL) is safe */
7320 doms_cur
= doms_new
;
7321 dattr_cur
= dattr_new
;
7322 ndoms_cur
= ndoms_new
;
7324 register_sched_domain_sysctl();
7326 mutex_unlock(&sched_domains_mutex
);
7329 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7330 static void arch_reinit_sched_domains(void)
7334 /* Destroy domains first to force the rebuild */
7335 partition_sched_domains(0, NULL
, NULL
);
7337 rebuild_sched_domains();
7341 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7343 unsigned int level
= 0;
7345 if (sscanf(buf
, "%u", &level
) != 1)
7349 * level is always be positive so don't check for
7350 * level < POWERSAVINGS_BALANCE_NONE which is 0
7351 * What happens on 0 or 1 byte write,
7352 * need to check for count as well?
7355 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7359 sched_smt_power_savings
= level
;
7361 sched_mc_power_savings
= level
;
7363 arch_reinit_sched_domains();
7368 #ifdef CONFIG_SCHED_MC
7369 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7370 struct sysdev_class_attribute
*attr
,
7373 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7375 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7376 struct sysdev_class_attribute
*attr
,
7377 const char *buf
, size_t count
)
7379 return sched_power_savings_store(buf
, count
, 0);
7381 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7382 sched_mc_power_savings_show
,
7383 sched_mc_power_savings_store
);
7386 #ifdef CONFIG_SCHED_SMT
7387 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7388 struct sysdev_class_attribute
*attr
,
7391 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7393 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7394 struct sysdev_class_attribute
*attr
,
7395 const char *buf
, size_t count
)
7397 return sched_power_savings_store(buf
, count
, 1);
7399 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7400 sched_smt_power_savings_show
,
7401 sched_smt_power_savings_store
);
7404 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7408 #ifdef CONFIG_SCHED_SMT
7410 err
= sysfs_create_file(&cls
->kset
.kobj
,
7411 &attr_sched_smt_power_savings
.attr
);
7413 #ifdef CONFIG_SCHED_MC
7414 if (!err
&& mc_capable())
7415 err
= sysfs_create_file(&cls
->kset
.kobj
,
7416 &attr_sched_mc_power_savings
.attr
);
7420 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7422 #ifndef CONFIG_CPUSETS
7424 * Add online and remove offline CPUs from the scheduler domains.
7425 * When cpusets are enabled they take over this function.
7427 static int update_sched_domains(struct notifier_block
*nfb
,
7428 unsigned long action
, void *hcpu
)
7432 case CPU_ONLINE_FROZEN
:
7433 case CPU_DOWN_PREPARE
:
7434 case CPU_DOWN_PREPARE_FROZEN
:
7435 case CPU_DOWN_FAILED
:
7436 case CPU_DOWN_FAILED_FROZEN
:
7437 partition_sched_domains(1, NULL
, NULL
);
7446 static int update_runtime(struct notifier_block
*nfb
,
7447 unsigned long action
, void *hcpu
)
7449 int cpu
= (int)(long)hcpu
;
7452 case CPU_DOWN_PREPARE
:
7453 case CPU_DOWN_PREPARE_FROZEN
:
7454 disable_runtime(cpu_rq(cpu
));
7457 case CPU_DOWN_FAILED
:
7458 case CPU_DOWN_FAILED_FROZEN
:
7460 case CPU_ONLINE_FROZEN
:
7461 enable_runtime(cpu_rq(cpu
));
7469 void __init
sched_init_smp(void)
7471 cpumask_var_t non_isolated_cpus
;
7473 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7474 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7476 #if defined(CONFIG_NUMA)
7477 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7479 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7482 mutex_lock(&sched_domains_mutex
);
7483 arch_init_sched_domains(cpu_active_mask
);
7484 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7485 if (cpumask_empty(non_isolated_cpus
))
7486 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7487 mutex_unlock(&sched_domains_mutex
);
7490 #ifndef CONFIG_CPUSETS
7491 /* XXX: Theoretical race here - CPU may be hotplugged now */
7492 hotcpu_notifier(update_sched_domains
, 0);
7495 /* RT runtime code needs to handle some hotplug events */
7496 hotcpu_notifier(update_runtime
, 0);
7500 /* Move init over to a non-isolated CPU */
7501 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7503 sched_init_granularity();
7504 free_cpumask_var(non_isolated_cpus
);
7506 init_sched_rt_class();
7509 void __init
sched_init_smp(void)
7511 sched_init_granularity();
7513 #endif /* CONFIG_SMP */
7515 const_debug
unsigned int sysctl_timer_migration
= 1;
7517 int in_sched_functions(unsigned long addr
)
7519 return in_lock_functions(addr
) ||
7520 (addr
>= (unsigned long)__sched_text_start
7521 && addr
< (unsigned long)__sched_text_end
);
7524 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7526 cfs_rq
->tasks_timeline
= RB_ROOT
;
7527 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7528 #ifdef CONFIG_FAIR_GROUP_SCHED
7531 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7534 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7536 struct rt_prio_array
*array
;
7539 array
= &rt_rq
->active
;
7540 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7541 INIT_LIST_HEAD(array
->queue
+ i
);
7542 __clear_bit(i
, array
->bitmap
);
7544 /* delimiter for bitsearch: */
7545 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7547 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7548 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7550 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7554 rt_rq
->rt_nr_migratory
= 0;
7555 rt_rq
->overloaded
= 0;
7556 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7560 rt_rq
->rt_throttled
= 0;
7561 rt_rq
->rt_runtime
= 0;
7562 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7564 #ifdef CONFIG_RT_GROUP_SCHED
7565 rt_rq
->rt_nr_boosted
= 0;
7570 #ifdef CONFIG_FAIR_GROUP_SCHED
7571 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7572 struct sched_entity
*se
, int cpu
, int add
,
7573 struct sched_entity
*parent
)
7575 struct rq
*rq
= cpu_rq(cpu
);
7576 tg
->cfs_rq
[cpu
] = cfs_rq
;
7577 init_cfs_rq(cfs_rq
, rq
);
7580 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7583 /* se could be NULL for init_task_group */
7588 se
->cfs_rq
= &rq
->cfs
;
7590 se
->cfs_rq
= parent
->my_q
;
7593 se
->load
.weight
= tg
->shares
;
7594 se
->load
.inv_weight
= 0;
7595 se
->parent
= parent
;
7599 #ifdef CONFIG_RT_GROUP_SCHED
7600 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7601 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7602 struct sched_rt_entity
*parent
)
7604 struct rq
*rq
= cpu_rq(cpu
);
7606 tg
->rt_rq
[cpu
] = rt_rq
;
7607 init_rt_rq(rt_rq
, rq
);
7609 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7611 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7613 tg
->rt_se
[cpu
] = rt_se
;
7618 rt_se
->rt_rq
= &rq
->rt
;
7620 rt_se
->rt_rq
= parent
->my_q
;
7622 rt_se
->my_q
= rt_rq
;
7623 rt_se
->parent
= parent
;
7624 INIT_LIST_HEAD(&rt_se
->run_list
);
7628 void __init
sched_init(void)
7631 unsigned long alloc_size
= 0, ptr
;
7633 #ifdef CONFIG_FAIR_GROUP_SCHED
7634 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7636 #ifdef CONFIG_RT_GROUP_SCHED
7637 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7639 #ifdef CONFIG_CPUMASK_OFFSTACK
7640 alloc_size
+= num_possible_cpus() * cpumask_size();
7643 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7645 #ifdef CONFIG_FAIR_GROUP_SCHED
7646 init_task_group
.se
= (struct sched_entity
**)ptr
;
7647 ptr
+= nr_cpu_ids
* sizeof(void **);
7649 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7650 ptr
+= nr_cpu_ids
* sizeof(void **);
7652 #endif /* CONFIG_FAIR_GROUP_SCHED */
7653 #ifdef CONFIG_RT_GROUP_SCHED
7654 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7655 ptr
+= nr_cpu_ids
* sizeof(void **);
7657 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7658 ptr
+= nr_cpu_ids
* sizeof(void **);
7660 #endif /* CONFIG_RT_GROUP_SCHED */
7661 #ifdef CONFIG_CPUMASK_OFFSTACK
7662 for_each_possible_cpu(i
) {
7663 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7664 ptr
+= cpumask_size();
7666 #endif /* CONFIG_CPUMASK_OFFSTACK */
7670 init_defrootdomain();
7673 init_rt_bandwidth(&def_rt_bandwidth
,
7674 global_rt_period(), global_rt_runtime());
7676 #ifdef CONFIG_RT_GROUP_SCHED
7677 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7678 global_rt_period(), global_rt_runtime());
7679 #endif /* CONFIG_RT_GROUP_SCHED */
7681 #ifdef CONFIG_CGROUP_SCHED
7682 list_add(&init_task_group
.list
, &task_groups
);
7683 INIT_LIST_HEAD(&init_task_group
.children
);
7685 #endif /* CONFIG_CGROUP_SCHED */
7687 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7688 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7689 __alignof__(unsigned long));
7691 for_each_possible_cpu(i
) {
7695 raw_spin_lock_init(&rq
->lock
);
7697 rq
->calc_load_active
= 0;
7698 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7699 init_cfs_rq(&rq
->cfs
, rq
);
7700 init_rt_rq(&rq
->rt
, rq
);
7701 #ifdef CONFIG_FAIR_GROUP_SCHED
7702 init_task_group
.shares
= init_task_group_load
;
7703 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7704 #ifdef CONFIG_CGROUP_SCHED
7706 * How much cpu bandwidth does init_task_group get?
7708 * In case of task-groups formed thr' the cgroup filesystem, it
7709 * gets 100% of the cpu resources in the system. This overall
7710 * system cpu resource is divided among the tasks of
7711 * init_task_group and its child task-groups in a fair manner,
7712 * based on each entity's (task or task-group's) weight
7713 * (se->load.weight).
7715 * In other words, if init_task_group has 10 tasks of weight
7716 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7717 * then A0's share of the cpu resource is:
7719 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7721 * We achieve this by letting init_task_group's tasks sit
7722 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7724 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7726 #endif /* CONFIG_FAIR_GROUP_SCHED */
7728 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7729 #ifdef CONFIG_RT_GROUP_SCHED
7730 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7731 #ifdef CONFIG_CGROUP_SCHED
7732 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7736 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7737 rq
->cpu_load
[j
] = 0;
7741 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7742 rq
->post_schedule
= 0;
7743 rq
->active_balance
= 0;
7744 rq
->next_balance
= jiffies
;
7749 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7750 rq_attach_root(rq
, &def_root_domain
);
7753 atomic_set(&rq
->nr_iowait
, 0);
7756 set_load_weight(&init_task
);
7758 #ifdef CONFIG_PREEMPT_NOTIFIERS
7759 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7763 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7766 #ifdef CONFIG_RT_MUTEXES
7767 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7771 * The boot idle thread does lazy MMU switching as well:
7773 atomic_inc(&init_mm
.mm_count
);
7774 enter_lazy_tlb(&init_mm
, current
);
7777 * Make us the idle thread. Technically, schedule() should not be
7778 * called from this thread, however somewhere below it might be,
7779 * but because we are the idle thread, we just pick up running again
7780 * when this runqueue becomes "idle".
7782 init_idle(current
, smp_processor_id());
7784 calc_load_update
= jiffies
+ LOAD_FREQ
;
7787 * During early bootup we pretend to be a normal task:
7789 current
->sched_class
= &fair_sched_class
;
7791 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7792 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7795 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7796 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7798 /* May be allocated at isolcpus cmdline parse time */
7799 if (cpu_isolated_map
== NULL
)
7800 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7805 scheduler_running
= 1;
7808 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7809 static inline int preempt_count_equals(int preempt_offset
)
7811 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7813 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7816 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7819 static unsigned long prev_jiffy
; /* ratelimiting */
7821 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7822 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7824 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7826 prev_jiffy
= jiffies
;
7829 "BUG: sleeping function called from invalid context at %s:%d\n",
7832 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7833 in_atomic(), irqs_disabled(),
7834 current
->pid
, current
->comm
);
7836 debug_show_held_locks(current
);
7837 if (irqs_disabled())
7838 print_irqtrace_events(current
);
7842 EXPORT_SYMBOL(__might_sleep
);
7845 #ifdef CONFIG_MAGIC_SYSRQ
7846 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7850 on_rq
= p
->se
.on_rq
;
7852 deactivate_task(rq
, p
, 0);
7853 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7855 activate_task(rq
, p
, 0);
7856 resched_task(rq
->curr
);
7860 void normalize_rt_tasks(void)
7862 struct task_struct
*g
, *p
;
7863 unsigned long flags
;
7866 read_lock_irqsave(&tasklist_lock
, flags
);
7867 do_each_thread(g
, p
) {
7869 * Only normalize user tasks:
7874 p
->se
.exec_start
= 0;
7875 #ifdef CONFIG_SCHEDSTATS
7876 p
->se
.statistics
.wait_start
= 0;
7877 p
->se
.statistics
.sleep_start
= 0;
7878 p
->se
.statistics
.block_start
= 0;
7883 * Renice negative nice level userspace
7886 if (TASK_NICE(p
) < 0 && p
->mm
)
7887 set_user_nice(p
, 0);
7891 raw_spin_lock(&p
->pi_lock
);
7892 rq
= __task_rq_lock(p
);
7894 normalize_task(rq
, p
);
7896 __task_rq_unlock(rq
);
7897 raw_spin_unlock(&p
->pi_lock
);
7898 } while_each_thread(g
, p
);
7900 read_unlock_irqrestore(&tasklist_lock
, flags
);
7903 #endif /* CONFIG_MAGIC_SYSRQ */
7905 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7907 * These functions are only useful for the IA64 MCA handling, or kdb.
7909 * They can only be called when the whole system has been
7910 * stopped - every CPU needs to be quiescent, and no scheduling
7911 * activity can take place. Using them for anything else would
7912 * be a serious bug, and as a result, they aren't even visible
7913 * under any other configuration.
7917 * curr_task - return the current task for a given cpu.
7918 * @cpu: the processor in question.
7920 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7922 struct task_struct
*curr_task(int cpu
)
7924 return cpu_curr(cpu
);
7927 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7931 * set_curr_task - set the current task for a given cpu.
7932 * @cpu: the processor in question.
7933 * @p: the task pointer to set.
7935 * Description: This function must only be used when non-maskable interrupts
7936 * are serviced on a separate stack. It allows the architecture to switch the
7937 * notion of the current task on a cpu in a non-blocking manner. This function
7938 * must be called with all CPU's synchronized, and interrupts disabled, the
7939 * and caller must save the original value of the current task (see
7940 * curr_task() above) and restore that value before reenabling interrupts and
7941 * re-starting the system.
7943 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7945 void set_curr_task(int cpu
, struct task_struct
*p
)
7952 #ifdef CONFIG_FAIR_GROUP_SCHED
7953 static void free_fair_sched_group(struct task_group
*tg
)
7957 for_each_possible_cpu(i
) {
7959 kfree(tg
->cfs_rq
[i
]);
7969 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7971 struct cfs_rq
*cfs_rq
;
7972 struct sched_entity
*se
;
7976 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7979 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7983 tg
->shares
= NICE_0_LOAD
;
7985 for_each_possible_cpu(i
) {
7988 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7989 GFP_KERNEL
, cpu_to_node(i
));
7993 se
= kzalloc_node(sizeof(struct sched_entity
),
7994 GFP_KERNEL
, cpu_to_node(i
));
7998 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8009 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8011 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8012 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8015 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8017 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8019 #else /* !CONFG_FAIR_GROUP_SCHED */
8020 static inline void free_fair_sched_group(struct task_group
*tg
)
8025 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8030 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8034 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8037 #endif /* CONFIG_FAIR_GROUP_SCHED */
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 static void free_rt_sched_group(struct task_group
*tg
)
8044 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8046 for_each_possible_cpu(i
) {
8048 kfree(tg
->rt_rq
[i
]);
8050 kfree(tg
->rt_se
[i
]);
8058 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8060 struct rt_rq
*rt_rq
;
8061 struct sched_rt_entity
*rt_se
;
8065 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8068 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8072 init_rt_bandwidth(&tg
->rt_bandwidth
,
8073 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8075 for_each_possible_cpu(i
) {
8078 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8079 GFP_KERNEL
, cpu_to_node(i
));
8083 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8084 GFP_KERNEL
, cpu_to_node(i
));
8088 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8099 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8101 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8102 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8105 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8107 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8109 #else /* !CONFIG_RT_GROUP_SCHED */
8110 static inline void free_rt_sched_group(struct task_group
*tg
)
8115 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8120 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8124 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8127 #endif /* CONFIG_RT_GROUP_SCHED */
8129 #ifdef CONFIG_CGROUP_SCHED
8130 static void free_sched_group(struct task_group
*tg
)
8132 free_fair_sched_group(tg
);
8133 free_rt_sched_group(tg
);
8137 /* allocate runqueue etc for a new task group */
8138 struct task_group
*sched_create_group(struct task_group
*parent
)
8140 struct task_group
*tg
;
8141 unsigned long flags
;
8144 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8146 return ERR_PTR(-ENOMEM
);
8148 if (!alloc_fair_sched_group(tg
, parent
))
8151 if (!alloc_rt_sched_group(tg
, parent
))
8154 spin_lock_irqsave(&task_group_lock
, flags
);
8155 for_each_possible_cpu(i
) {
8156 register_fair_sched_group(tg
, i
);
8157 register_rt_sched_group(tg
, i
);
8159 list_add_rcu(&tg
->list
, &task_groups
);
8161 WARN_ON(!parent
); /* root should already exist */
8163 tg
->parent
= parent
;
8164 INIT_LIST_HEAD(&tg
->children
);
8165 list_add_rcu(&tg
->siblings
, &parent
->children
);
8166 spin_unlock_irqrestore(&task_group_lock
, flags
);
8171 free_sched_group(tg
);
8172 return ERR_PTR(-ENOMEM
);
8175 /* rcu callback to free various structures associated with a task group */
8176 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8178 /* now it should be safe to free those cfs_rqs */
8179 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8182 /* Destroy runqueue etc associated with a task group */
8183 void sched_destroy_group(struct task_group
*tg
)
8185 unsigned long flags
;
8188 spin_lock_irqsave(&task_group_lock
, flags
);
8189 for_each_possible_cpu(i
) {
8190 unregister_fair_sched_group(tg
, i
);
8191 unregister_rt_sched_group(tg
, i
);
8193 list_del_rcu(&tg
->list
);
8194 list_del_rcu(&tg
->siblings
);
8195 spin_unlock_irqrestore(&task_group_lock
, flags
);
8197 /* wait for possible concurrent references to cfs_rqs complete */
8198 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8201 /* change task's runqueue when it moves between groups.
8202 * The caller of this function should have put the task in its new group
8203 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8204 * reflect its new group.
8206 void sched_move_task(struct task_struct
*tsk
)
8209 unsigned long flags
;
8212 rq
= task_rq_lock(tsk
, &flags
);
8214 running
= task_current(rq
, tsk
);
8215 on_rq
= tsk
->se
.on_rq
;
8218 dequeue_task(rq
, tsk
, 0);
8219 if (unlikely(running
))
8220 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8222 set_task_rq(tsk
, task_cpu(tsk
));
8224 #ifdef CONFIG_FAIR_GROUP_SCHED
8225 if (tsk
->sched_class
->moved_group
)
8226 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8229 if (unlikely(running
))
8230 tsk
->sched_class
->set_curr_task(rq
);
8232 enqueue_task(rq
, tsk
, 0);
8234 task_rq_unlock(rq
, &flags
);
8236 #endif /* CONFIG_CGROUP_SCHED */
8238 #ifdef CONFIG_FAIR_GROUP_SCHED
8239 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8241 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8246 dequeue_entity(cfs_rq
, se
, 0);
8248 se
->load
.weight
= shares
;
8249 se
->load
.inv_weight
= 0;
8252 enqueue_entity(cfs_rq
, se
, 0);
8255 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8257 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8258 struct rq
*rq
= cfs_rq
->rq
;
8259 unsigned long flags
;
8261 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8262 __set_se_shares(se
, shares
);
8263 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8266 static DEFINE_MUTEX(shares_mutex
);
8268 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8271 unsigned long flags
;
8274 * We can't change the weight of the root cgroup.
8279 if (shares
< MIN_SHARES
)
8280 shares
= MIN_SHARES
;
8281 else if (shares
> MAX_SHARES
)
8282 shares
= MAX_SHARES
;
8284 mutex_lock(&shares_mutex
);
8285 if (tg
->shares
== shares
)
8288 spin_lock_irqsave(&task_group_lock
, flags
);
8289 for_each_possible_cpu(i
)
8290 unregister_fair_sched_group(tg
, i
);
8291 list_del_rcu(&tg
->siblings
);
8292 spin_unlock_irqrestore(&task_group_lock
, flags
);
8294 /* wait for any ongoing reference to this group to finish */
8295 synchronize_sched();
8298 * Now we are free to modify the group's share on each cpu
8299 * w/o tripping rebalance_share or load_balance_fair.
8301 tg
->shares
= shares
;
8302 for_each_possible_cpu(i
) {
8306 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8307 set_se_shares(tg
->se
[i
], shares
);
8311 * Enable load balance activity on this group, by inserting it back on
8312 * each cpu's rq->leaf_cfs_rq_list.
8314 spin_lock_irqsave(&task_group_lock
, flags
);
8315 for_each_possible_cpu(i
)
8316 register_fair_sched_group(tg
, i
);
8317 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8318 spin_unlock_irqrestore(&task_group_lock
, flags
);
8320 mutex_unlock(&shares_mutex
);
8324 unsigned long sched_group_shares(struct task_group
*tg
)
8330 #ifdef CONFIG_RT_GROUP_SCHED
8332 * Ensure that the real time constraints are schedulable.
8334 static DEFINE_MUTEX(rt_constraints_mutex
);
8336 static unsigned long to_ratio(u64 period
, u64 runtime
)
8338 if (runtime
== RUNTIME_INF
)
8341 return div64_u64(runtime
<< 20, period
);
8344 /* Must be called with tasklist_lock held */
8345 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8347 struct task_struct
*g
, *p
;
8349 do_each_thread(g
, p
) {
8350 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8352 } while_each_thread(g
, p
);
8357 struct rt_schedulable_data
{
8358 struct task_group
*tg
;
8363 static int tg_schedulable(struct task_group
*tg
, void *data
)
8365 struct rt_schedulable_data
*d
= data
;
8366 struct task_group
*child
;
8367 unsigned long total
, sum
= 0;
8368 u64 period
, runtime
;
8370 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8371 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8374 period
= d
->rt_period
;
8375 runtime
= d
->rt_runtime
;
8379 * Cannot have more runtime than the period.
8381 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8385 * Ensure we don't starve existing RT tasks.
8387 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8390 total
= to_ratio(period
, runtime
);
8393 * Nobody can have more than the global setting allows.
8395 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8399 * The sum of our children's runtime should not exceed our own.
8401 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8402 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8403 runtime
= child
->rt_bandwidth
.rt_runtime
;
8405 if (child
== d
->tg
) {
8406 period
= d
->rt_period
;
8407 runtime
= d
->rt_runtime
;
8410 sum
+= to_ratio(period
, runtime
);
8419 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8421 struct rt_schedulable_data data
= {
8423 .rt_period
= period
,
8424 .rt_runtime
= runtime
,
8427 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8430 static int tg_set_bandwidth(struct task_group
*tg
,
8431 u64 rt_period
, u64 rt_runtime
)
8435 mutex_lock(&rt_constraints_mutex
);
8436 read_lock(&tasklist_lock
);
8437 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8441 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8442 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8443 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8445 for_each_possible_cpu(i
) {
8446 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8448 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8449 rt_rq
->rt_runtime
= rt_runtime
;
8450 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8452 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8454 read_unlock(&tasklist_lock
);
8455 mutex_unlock(&rt_constraints_mutex
);
8460 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8462 u64 rt_runtime
, rt_period
;
8464 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8465 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8466 if (rt_runtime_us
< 0)
8467 rt_runtime
= RUNTIME_INF
;
8469 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8472 long sched_group_rt_runtime(struct task_group
*tg
)
8476 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8479 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8480 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8481 return rt_runtime_us
;
8484 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8486 u64 rt_runtime
, rt_period
;
8488 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8489 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8494 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8497 long sched_group_rt_period(struct task_group
*tg
)
8501 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8502 do_div(rt_period_us
, NSEC_PER_USEC
);
8503 return rt_period_us
;
8506 static int sched_rt_global_constraints(void)
8508 u64 runtime
, period
;
8511 if (sysctl_sched_rt_period
<= 0)
8514 runtime
= global_rt_runtime();
8515 period
= global_rt_period();
8518 * Sanity check on the sysctl variables.
8520 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8523 mutex_lock(&rt_constraints_mutex
);
8524 read_lock(&tasklist_lock
);
8525 ret
= __rt_schedulable(NULL
, 0, 0);
8526 read_unlock(&tasklist_lock
);
8527 mutex_unlock(&rt_constraints_mutex
);
8532 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8534 /* Don't accept realtime tasks when there is no way for them to run */
8535 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8541 #else /* !CONFIG_RT_GROUP_SCHED */
8542 static int sched_rt_global_constraints(void)
8544 unsigned long flags
;
8547 if (sysctl_sched_rt_period
<= 0)
8551 * There's always some RT tasks in the root group
8552 * -- migration, kstopmachine etc..
8554 if (sysctl_sched_rt_runtime
== 0)
8557 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8558 for_each_possible_cpu(i
) {
8559 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8561 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8562 rt_rq
->rt_runtime
= global_rt_runtime();
8563 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8565 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8569 #endif /* CONFIG_RT_GROUP_SCHED */
8571 int sched_rt_handler(struct ctl_table
*table
, int write
,
8572 void __user
*buffer
, size_t *lenp
,
8576 int old_period
, old_runtime
;
8577 static DEFINE_MUTEX(mutex
);
8580 old_period
= sysctl_sched_rt_period
;
8581 old_runtime
= sysctl_sched_rt_runtime
;
8583 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8585 if (!ret
&& write
) {
8586 ret
= sched_rt_global_constraints();
8588 sysctl_sched_rt_period
= old_period
;
8589 sysctl_sched_rt_runtime
= old_runtime
;
8591 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8592 def_rt_bandwidth
.rt_period
=
8593 ns_to_ktime(global_rt_period());
8596 mutex_unlock(&mutex
);
8601 #ifdef CONFIG_CGROUP_SCHED
8603 /* return corresponding task_group object of a cgroup */
8604 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8606 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8607 struct task_group
, css
);
8610 static struct cgroup_subsys_state
*
8611 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8613 struct task_group
*tg
, *parent
;
8615 if (!cgrp
->parent
) {
8616 /* This is early initialization for the top cgroup */
8617 return &init_task_group
.css
;
8620 parent
= cgroup_tg(cgrp
->parent
);
8621 tg
= sched_create_group(parent
);
8623 return ERR_PTR(-ENOMEM
);
8629 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8631 struct task_group
*tg
= cgroup_tg(cgrp
);
8633 sched_destroy_group(tg
);
8637 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8639 #ifdef CONFIG_RT_GROUP_SCHED
8640 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8643 /* We don't support RT-tasks being in separate groups */
8644 if (tsk
->sched_class
!= &fair_sched_class
)
8651 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8652 struct task_struct
*tsk
, bool threadgroup
)
8654 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8658 struct task_struct
*c
;
8660 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8661 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8673 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8674 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8677 sched_move_task(tsk
);
8679 struct task_struct
*c
;
8681 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8688 #ifdef CONFIG_FAIR_GROUP_SCHED
8689 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8692 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8695 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8697 struct task_group
*tg
= cgroup_tg(cgrp
);
8699 return (u64
) tg
->shares
;
8701 #endif /* CONFIG_FAIR_GROUP_SCHED */
8703 #ifdef CONFIG_RT_GROUP_SCHED
8704 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8707 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8710 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8712 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8715 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8718 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8721 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8723 return sched_group_rt_period(cgroup_tg(cgrp
));
8725 #endif /* CONFIG_RT_GROUP_SCHED */
8727 static struct cftype cpu_files
[] = {
8728 #ifdef CONFIG_FAIR_GROUP_SCHED
8731 .read_u64
= cpu_shares_read_u64
,
8732 .write_u64
= cpu_shares_write_u64
,
8735 #ifdef CONFIG_RT_GROUP_SCHED
8737 .name
= "rt_runtime_us",
8738 .read_s64
= cpu_rt_runtime_read
,
8739 .write_s64
= cpu_rt_runtime_write
,
8742 .name
= "rt_period_us",
8743 .read_u64
= cpu_rt_period_read_uint
,
8744 .write_u64
= cpu_rt_period_write_uint
,
8749 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8751 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8754 struct cgroup_subsys cpu_cgroup_subsys
= {
8756 .create
= cpu_cgroup_create
,
8757 .destroy
= cpu_cgroup_destroy
,
8758 .can_attach
= cpu_cgroup_can_attach
,
8759 .attach
= cpu_cgroup_attach
,
8760 .populate
= cpu_cgroup_populate
,
8761 .subsys_id
= cpu_cgroup_subsys_id
,
8765 #endif /* CONFIG_CGROUP_SCHED */
8767 #ifdef CONFIG_CGROUP_CPUACCT
8770 * CPU accounting code for task groups.
8772 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8773 * (balbir@in.ibm.com).
8776 /* track cpu usage of a group of tasks and its child groups */
8778 struct cgroup_subsys_state css
;
8779 /* cpuusage holds pointer to a u64-type object on every cpu */
8780 u64 __percpu
*cpuusage
;
8781 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8782 struct cpuacct
*parent
;
8785 struct cgroup_subsys cpuacct_subsys
;
8787 /* return cpu accounting group corresponding to this container */
8788 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8790 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8791 struct cpuacct
, css
);
8794 /* return cpu accounting group to which this task belongs */
8795 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8797 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8798 struct cpuacct
, css
);
8801 /* create a new cpu accounting group */
8802 static struct cgroup_subsys_state
*cpuacct_create(
8803 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8805 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8811 ca
->cpuusage
= alloc_percpu(u64
);
8815 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8816 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8817 goto out_free_counters
;
8820 ca
->parent
= cgroup_ca(cgrp
->parent
);
8826 percpu_counter_destroy(&ca
->cpustat
[i
]);
8827 free_percpu(ca
->cpuusage
);
8831 return ERR_PTR(-ENOMEM
);
8834 /* destroy an existing cpu accounting group */
8836 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8838 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8841 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8842 percpu_counter_destroy(&ca
->cpustat
[i
]);
8843 free_percpu(ca
->cpuusage
);
8847 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8849 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8852 #ifndef CONFIG_64BIT
8854 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8856 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8858 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8866 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8868 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8870 #ifndef CONFIG_64BIT
8872 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8874 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8876 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8882 /* return total cpu usage (in nanoseconds) of a group */
8883 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8885 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8886 u64 totalcpuusage
= 0;
8889 for_each_present_cpu(i
)
8890 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8892 return totalcpuusage
;
8895 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8898 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8907 for_each_present_cpu(i
)
8908 cpuacct_cpuusage_write(ca
, i
, 0);
8914 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8917 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8921 for_each_present_cpu(i
) {
8922 percpu
= cpuacct_cpuusage_read(ca
, i
);
8923 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8925 seq_printf(m
, "\n");
8929 static const char *cpuacct_stat_desc
[] = {
8930 [CPUACCT_STAT_USER
] = "user",
8931 [CPUACCT_STAT_SYSTEM
] = "system",
8934 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8935 struct cgroup_map_cb
*cb
)
8937 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8940 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8941 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8942 val
= cputime64_to_clock_t(val
);
8943 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8948 static struct cftype files
[] = {
8951 .read_u64
= cpuusage_read
,
8952 .write_u64
= cpuusage_write
,
8955 .name
= "usage_percpu",
8956 .read_seq_string
= cpuacct_percpu_seq_read
,
8960 .read_map
= cpuacct_stats_show
,
8964 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8966 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8970 * charge this task's execution time to its accounting group.
8972 * called with rq->lock held.
8974 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8979 if (unlikely(!cpuacct_subsys
.active
))
8982 cpu
= task_cpu(tsk
);
8988 for (; ca
; ca
= ca
->parent
) {
8989 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8990 *cpuusage
+= cputime
;
8997 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8998 * in cputime_t units. As a result, cpuacct_update_stats calls
8999 * percpu_counter_add with values large enough to always overflow the
9000 * per cpu batch limit causing bad SMP scalability.
9002 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9003 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9004 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9007 #define CPUACCT_BATCH \
9008 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9010 #define CPUACCT_BATCH 0
9014 * Charge the system/user time to the task's accounting group.
9016 static void cpuacct_update_stats(struct task_struct
*tsk
,
9017 enum cpuacct_stat_index idx
, cputime_t val
)
9020 int batch
= CPUACCT_BATCH
;
9022 if (unlikely(!cpuacct_subsys
.active
))
9029 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9035 struct cgroup_subsys cpuacct_subsys
= {
9037 .create
= cpuacct_create
,
9038 .destroy
= cpuacct_destroy
,
9039 .populate
= cpuacct_populate
,
9040 .subsys_id
= cpuacct_subsys_id
,
9042 #endif /* CONFIG_CGROUP_CPUACCT */
9046 void synchronize_sched_expedited(void)
9050 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9052 #else /* #ifndef CONFIG_SMP */
9054 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
9056 static int synchronize_sched_expedited_cpu_stop(void *data
)
9059 * There must be a full memory barrier on each affected CPU
9060 * between the time that try_stop_cpus() is called and the
9061 * time that it returns.
9063 * In the current initial implementation of cpu_stop, the
9064 * above condition is already met when the control reaches
9065 * this point and the following smp_mb() is not strictly
9066 * necessary. Do smp_mb() anyway for documentation and
9067 * robustness against future implementation changes.
9069 smp_mb(); /* See above comment block. */
9074 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9075 * approach to force grace period to end quickly. This consumes
9076 * significant time on all CPUs, and is thus not recommended for
9077 * any sort of common-case code.
9079 * Note that it is illegal to call this function while holding any
9080 * lock that is acquired by a CPU-hotplug notifier. Failing to
9081 * observe this restriction will result in deadlock.
9083 void synchronize_sched_expedited(void)
9085 int snap
, trycount
= 0;
9087 smp_mb(); /* ensure prior mod happens before capturing snap. */
9088 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
9090 while (try_stop_cpus(cpu_online_mask
,
9091 synchronize_sched_expedited_cpu_stop
,
9094 if (trycount
++ < 10)
9095 udelay(trycount
* num_online_cpus());
9097 synchronize_sched();
9100 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
9101 smp_mb(); /* ensure test happens before caller kfree */
9106 atomic_inc(&synchronize_sched_expedited_count
);
9107 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
9110 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9112 #endif /* #else #ifndef CONFIG_SMP */