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 <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy
)
127 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
132 static inline int task_has_rt_policy(struct task_struct
*p
)
134 return rt_policy(p
->policy
);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array
{
141 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
142 struct list_head queue
[MAX_RT_PRIO
];
145 struct rt_bandwidth
{
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock
;
150 struct hrtimer rt_period_timer
;
153 static struct rt_bandwidth def_rt_bandwidth
;
155 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
157 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
159 struct rt_bandwidth
*rt_b
=
160 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
166 now
= hrtimer_cb_get_time(timer
);
167 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
172 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
175 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
179 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
181 rt_b
->rt_period
= ns_to_ktime(period
);
182 rt_b
->rt_runtime
= runtime
;
184 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
186 hrtimer_init(&rt_b
->rt_period_timer
,
187 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
188 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime
>= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
200 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
203 if (hrtimer_active(&rt_b
->rt_period_timer
))
206 raw_spin_lock(&rt_b
->rt_runtime_lock
);
211 if (hrtimer_active(&rt_b
->rt_period_timer
))
214 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
215 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
217 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
218 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
219 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
220 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
221 HRTIMER_MODE_ABS_PINNED
, 0);
223 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
229 hrtimer_cancel(&rt_b
->rt_period_timer
);
234 * sched_domains_mutex serializes calls to arch_init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex
);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups
);
247 /* task group related information */
249 struct cgroup_subsys_state css
;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity
**se
;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq
**cfs_rq
;
256 unsigned long shares
;
258 atomic_t load_weight
;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity
**rt_se
;
263 struct rt_rq
**rt_rq
;
265 struct rt_bandwidth rt_bandwidth
;
269 struct list_head list
;
271 struct task_group
*parent
;
272 struct list_head siblings
;
273 struct list_head children
;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup
*autogroup
;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock
);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group
;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load
;
311 unsigned long nr_running
;
316 struct rb_root tasks_timeline
;
317 struct rb_node
*rb_leftmost
;
319 struct list_head tasks
;
320 struct list_head
*balance_iterator
;
323 * 'curr' points to currently running entity on this cfs_rq.
324 * It is set to NULL otherwise (i.e when none are currently running).
326 struct sched_entity
*curr
, *next
, *last
;
328 unsigned int nr_spread_over
;
330 #ifdef CONFIG_FAIR_GROUP_SCHED
331 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
334 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
335 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
336 * (like users, containers etc.)
338 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
339 * list is used during load balance.
342 struct list_head leaf_cfs_rq_list
;
343 struct task_group
*tg
; /* group that "owns" this runqueue */
347 * the part of load.weight contributed by tasks
349 unsigned long task_weight
;
352 * h_load = weight * f(tg)
354 * Where f(tg) is the recursive weight fraction assigned to
357 unsigned long h_load
;
360 * Maintaining per-cpu shares distribution for group scheduling
362 * load_stamp is the last time we updated the load average
363 * load_last is the last time we updated the load average and saw load
364 * load_unacc_exec_time is currently unaccounted execution time
368 u64 load_stamp
, load_last
, load_unacc_exec_time
;
370 unsigned long load_contribution
;
375 /* Real-Time classes' related field in a runqueue: */
377 struct rt_prio_array active
;
378 unsigned long rt_nr_running
;
379 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
381 int curr
; /* highest queued rt task prio */
383 int next
; /* next highest */
388 unsigned long rt_nr_migratory
;
389 unsigned long rt_nr_total
;
391 struct plist_head pushable_tasks
;
396 /* Nests inside the rq lock: */
397 raw_spinlock_t rt_runtime_lock
;
399 #ifdef CONFIG_RT_GROUP_SCHED
400 unsigned long rt_nr_boosted
;
403 struct list_head leaf_rt_rq_list
;
404 struct task_group
*tg
;
411 * We add the notion of a root-domain which will be used to define per-domain
412 * variables. Each exclusive cpuset essentially defines an island domain by
413 * fully partitioning the member cpus from any other cpuset. Whenever a new
414 * exclusive cpuset is created, we also create and attach a new root-domain
421 cpumask_var_t online
;
424 * The "RT overload" flag: it gets set if a CPU has more than
425 * one runnable RT task.
427 cpumask_var_t rto_mask
;
429 struct cpupri cpupri
;
433 * By default the system creates a single root-domain with all cpus as
434 * members (mimicking the global state we have today).
436 static struct root_domain def_root_domain
;
438 #endif /* CONFIG_SMP */
441 * This is the main, per-CPU runqueue data structure.
443 * Locking rule: those places that want to lock multiple runqueues
444 * (such as the load balancing or the thread migration code), lock
445 * acquire operations must be ordered by ascending &runqueue.
452 * nr_running and cpu_load should be in the same cacheline because
453 * remote CPUs use both these fields when doing load calculation.
455 unsigned long nr_running
;
456 #define CPU_LOAD_IDX_MAX 5
457 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
458 unsigned long last_load_update_tick
;
461 unsigned char nohz_balance_kick
;
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
, *stop
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
499 struct root_domain
*rd
;
500 struct sched_domain
*sd
;
502 unsigned long cpu_power
;
504 unsigned char idle_at_tick
;
505 /* For active balancing */
509 struct cpu_stop_work active_balance_work
;
510 /* cpu of this runqueue: */
514 unsigned long avg_load_per_task
;
522 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
526 /* calc_load related fields */
527 unsigned long calc_load_update
;
528 long calc_load_active
;
530 #ifdef CONFIG_SCHED_HRTICK
532 int hrtick_csd_pending
;
533 struct call_single_data hrtick_csd
;
535 struct hrtimer hrtick_timer
;
538 #ifdef CONFIG_SCHEDSTATS
540 struct sched_info rq_sched_info
;
541 unsigned long long rq_cpu_time
;
542 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
544 /* sys_sched_yield() stats */
545 unsigned int yld_count
;
547 /* schedule() stats */
548 unsigned int sched_switch
;
549 unsigned int sched_count
;
550 unsigned int sched_goidle
;
552 /* try_to_wake_up() stats */
553 unsigned int ttwu_count
;
554 unsigned int ttwu_local
;
558 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
561 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
563 static inline int cpu_of(struct rq
*rq
)
572 #define rcu_dereference_check_sched_domain(p) \
573 rcu_dereference_check((p), \
574 rcu_read_lock_sched_held() || \
575 lockdep_is_held(&sched_domains_mutex))
578 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
579 * See detach_destroy_domains: synchronize_sched for details.
581 * The domain tree of any CPU may only be accessed from within
582 * preempt-disabled sections.
584 #define for_each_domain(cpu, __sd) \
585 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
587 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
588 #define this_rq() (&__get_cpu_var(runqueues))
589 #define task_rq(p) cpu_rq(task_cpu(p))
590 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
591 #define raw_rq() (&__raw_get_cpu_var(runqueues))
593 #ifdef CONFIG_CGROUP_SCHED
596 * Return the group to which this tasks belongs.
598 * We use task_subsys_state_check() and extend the RCU verification
599 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
600 * holds that lock for each task it moves into the cgroup. Therefore
601 * by holding that lock, we pin the task to the current cgroup.
603 static inline struct task_group
*task_group(struct task_struct
*p
)
605 struct task_group
*tg
;
606 struct cgroup_subsys_state
*css
;
608 if (p
->flags
& PF_EXITING
)
609 return &root_task_group
;
611 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
612 lockdep_is_held(&task_rq(p
)->lock
));
613 tg
= container_of(css
, struct task_group
, css
);
615 return autogroup_task_group(p
, tg
);
618 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
619 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
621 #ifdef CONFIG_FAIR_GROUP_SCHED
622 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
623 p
->se
.parent
= task_group(p
)->se
[cpu
];
626 #ifdef CONFIG_RT_GROUP_SCHED
627 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
628 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
632 #else /* CONFIG_CGROUP_SCHED */
634 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
635 static inline struct task_group
*task_group(struct task_struct
*p
)
640 #endif /* CONFIG_CGROUP_SCHED */
642 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
644 static void update_rq_clock(struct rq
*rq
)
648 if (rq
->skip_clock_update
)
651 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
653 update_rq_clock_task(rq
, delta
);
657 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
659 #ifdef CONFIG_SCHED_DEBUG
660 # define const_debug __read_mostly
662 # define const_debug static const
667 * @cpu: the processor in question.
669 * Returns true if the current cpu runqueue is locked.
670 * This interface allows printk to be called with the runqueue lock
671 * held and know whether or not it is OK to wake up the klogd.
673 int runqueue_is_locked(int cpu
)
675 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
679 * Debugging: various feature bits
682 #define SCHED_FEAT(name, enabled) \
683 __SCHED_FEAT_##name ,
686 #include "sched_features.h"
691 #define SCHED_FEAT(name, enabled) \
692 (1UL << __SCHED_FEAT_##name) * enabled |
694 const_debug
unsigned int sysctl_sched_features
=
695 #include "sched_features.h"
700 #ifdef CONFIG_SCHED_DEBUG
701 #define SCHED_FEAT(name, enabled) \
704 static __read_mostly
char *sched_feat_names
[] = {
705 #include "sched_features.h"
711 static int sched_feat_show(struct seq_file
*m
, void *v
)
715 for (i
= 0; sched_feat_names
[i
]; i
++) {
716 if (!(sysctl_sched_features
& (1UL << i
)))
718 seq_printf(m
, "%s ", sched_feat_names
[i
]);
726 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
727 size_t cnt
, loff_t
*ppos
)
737 if (copy_from_user(&buf
, ubuf
, cnt
))
743 if (strncmp(cmp
, "NO_", 3) == 0) {
748 for (i
= 0; sched_feat_names
[i
]; i
++) {
749 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
751 sysctl_sched_features
&= ~(1UL << i
);
753 sysctl_sched_features
|= (1UL << i
);
758 if (!sched_feat_names
[i
])
766 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
768 return single_open(filp
, sched_feat_show
, NULL
);
771 static const struct file_operations sched_feat_fops
= {
772 .open
= sched_feat_open
,
773 .write
= sched_feat_write
,
776 .release
= single_release
,
779 static __init
int sched_init_debug(void)
781 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
786 late_initcall(sched_init_debug
);
790 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
793 * Number of tasks to iterate in a single balance run.
794 * Limited because this is done with IRQs disabled.
796 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
799 * period over which we average the RT time consumption, measured
804 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
807 * period over which we measure -rt task cpu usage in us.
810 unsigned int sysctl_sched_rt_period
= 1000000;
812 static __read_mostly
int scheduler_running
;
815 * part of the period that we allow rt tasks to run in us.
818 int sysctl_sched_rt_runtime
= 950000;
820 static inline u64
global_rt_period(void)
822 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
825 static inline u64
global_rt_runtime(void)
827 if (sysctl_sched_rt_runtime
< 0)
830 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
833 #ifndef prepare_arch_switch
834 # define prepare_arch_switch(next) do { } while (0)
836 #ifndef finish_arch_switch
837 # define finish_arch_switch(prev) do { } while (0)
840 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
842 return rq
->curr
== p
;
845 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
846 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
848 return task_current(rq
, p
);
851 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
855 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
857 #ifdef CONFIG_DEBUG_SPINLOCK
858 /* this is a valid case when another task releases the spinlock */
859 rq
->lock
.owner
= current
;
862 * If we are tracking spinlock dependencies then we have to
863 * fix up the runqueue lock - which gets 'carried over' from
866 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
868 raw_spin_unlock_irq(&rq
->lock
);
871 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
872 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
877 return task_current(rq
, p
);
881 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
885 * We can optimise this out completely for !SMP, because the
886 * SMP rebalancing from interrupt is the only thing that cares
891 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
892 raw_spin_unlock_irq(&rq
->lock
);
894 raw_spin_unlock(&rq
->lock
);
898 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
902 * After ->oncpu is cleared, the task can be moved to a different CPU.
903 * We must ensure this doesn't happen until the switch is completely
909 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
913 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
916 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
919 static inline int task_is_waking(struct task_struct
*p
)
921 return unlikely(p
->state
== TASK_WAKING
);
925 * __task_rq_lock - lock the runqueue a given task resides on.
926 * Must be called interrupts disabled.
928 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
935 raw_spin_lock(&rq
->lock
);
936 if (likely(rq
== task_rq(p
)))
938 raw_spin_unlock(&rq
->lock
);
943 * task_rq_lock - lock the runqueue a given task resides on and disable
944 * interrupts. Note the ordering: we can safely lookup the task_rq without
945 * explicitly disabling preemption.
947 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
953 local_irq_save(*flags
);
955 raw_spin_lock(&rq
->lock
);
956 if (likely(rq
== task_rq(p
)))
958 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
962 static void __task_rq_unlock(struct rq
*rq
)
965 raw_spin_unlock(&rq
->lock
);
968 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
971 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
975 * this_rq_lock - lock this runqueue and disable interrupts.
977 static struct rq
*this_rq_lock(void)
984 raw_spin_lock(&rq
->lock
);
989 #ifdef CONFIG_SCHED_HRTICK
991 * Use HR-timers to deliver accurate preemption points.
993 * Its all a bit involved since we cannot program an hrt while holding the
994 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
997 * When we get rescheduled we reprogram the hrtick_timer outside of the
1003 * - enabled by features
1004 * - hrtimer is actually high res
1006 static inline int hrtick_enabled(struct rq
*rq
)
1008 if (!sched_feat(HRTICK
))
1010 if (!cpu_active(cpu_of(rq
)))
1012 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1015 static void hrtick_clear(struct rq
*rq
)
1017 if (hrtimer_active(&rq
->hrtick_timer
))
1018 hrtimer_cancel(&rq
->hrtick_timer
);
1022 * High-resolution timer tick.
1023 * Runs from hardirq context with interrupts disabled.
1025 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1027 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1029 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1031 raw_spin_lock(&rq
->lock
);
1032 update_rq_clock(rq
);
1033 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1034 raw_spin_unlock(&rq
->lock
);
1036 return HRTIMER_NORESTART
;
1041 * called from hardirq (IPI) context
1043 static void __hrtick_start(void *arg
)
1045 struct rq
*rq
= arg
;
1047 raw_spin_lock(&rq
->lock
);
1048 hrtimer_restart(&rq
->hrtick_timer
);
1049 rq
->hrtick_csd_pending
= 0;
1050 raw_spin_unlock(&rq
->lock
);
1054 * Called to set the hrtick timer state.
1056 * called with rq->lock held and irqs disabled
1058 static void hrtick_start(struct rq
*rq
, u64 delay
)
1060 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1061 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1063 hrtimer_set_expires(timer
, time
);
1065 if (rq
== this_rq()) {
1066 hrtimer_restart(timer
);
1067 } else if (!rq
->hrtick_csd_pending
) {
1068 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1069 rq
->hrtick_csd_pending
= 1;
1074 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1076 int cpu
= (int)(long)hcpu
;
1079 case CPU_UP_CANCELED
:
1080 case CPU_UP_CANCELED_FROZEN
:
1081 case CPU_DOWN_PREPARE
:
1082 case CPU_DOWN_PREPARE_FROZEN
:
1084 case CPU_DEAD_FROZEN
:
1085 hrtick_clear(cpu_rq(cpu
));
1092 static __init
void init_hrtick(void)
1094 hotcpu_notifier(hotplug_hrtick
, 0);
1098 * Called to set the hrtick timer state.
1100 * called with rq->lock held and irqs disabled
1102 static void hrtick_start(struct rq
*rq
, u64 delay
)
1104 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1105 HRTIMER_MODE_REL_PINNED
, 0);
1108 static inline void init_hrtick(void)
1111 #endif /* CONFIG_SMP */
1113 static void init_rq_hrtick(struct rq
*rq
)
1116 rq
->hrtick_csd_pending
= 0;
1118 rq
->hrtick_csd
.flags
= 0;
1119 rq
->hrtick_csd
.func
= __hrtick_start
;
1120 rq
->hrtick_csd
.info
= rq
;
1123 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1124 rq
->hrtick_timer
.function
= hrtick
;
1126 #else /* CONFIG_SCHED_HRTICK */
1127 static inline void hrtick_clear(struct rq
*rq
)
1131 static inline void init_rq_hrtick(struct rq
*rq
)
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SCHED_HRTICK */
1141 * resched_task - mark a task 'to be rescheduled now'.
1143 * On UP this means the setting of the need_resched flag, on SMP it
1144 * might also involve a cross-CPU call to trigger the scheduler on
1149 #ifndef tsk_is_polling
1150 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1153 static void resched_task(struct task_struct
*p
)
1157 assert_raw_spin_locked(&task_rq(p
)->lock
);
1159 if (test_tsk_need_resched(p
))
1162 set_tsk_need_resched(p
);
1165 if (cpu
== smp_processor_id())
1168 /* NEED_RESCHED must be visible before we test polling */
1170 if (!tsk_is_polling(p
))
1171 smp_send_reschedule(cpu
);
1174 static void resched_cpu(int cpu
)
1176 struct rq
*rq
= cpu_rq(cpu
);
1177 unsigned long flags
;
1179 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1181 resched_task(cpu_curr(cpu
));
1182 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1187 * In the semi idle case, use the nearest busy cpu for migrating timers
1188 * from an idle cpu. This is good for power-savings.
1190 * We don't do similar optimization for completely idle system, as
1191 * selecting an idle cpu will add more delays to the timers than intended
1192 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1194 int get_nohz_timer_target(void)
1196 int cpu
= smp_processor_id();
1198 struct sched_domain
*sd
;
1200 for_each_domain(cpu
, sd
) {
1201 for_each_cpu(i
, sched_domain_span(sd
))
1208 * When add_timer_on() enqueues a timer into the timer wheel of an
1209 * idle CPU then this timer might expire before the next timer event
1210 * which is scheduled to wake up that CPU. In case of a completely
1211 * idle system the next event might even be infinite time into the
1212 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1213 * leaves the inner idle loop so the newly added timer is taken into
1214 * account when the CPU goes back to idle and evaluates the timer
1215 * wheel for the next timer event.
1217 void wake_up_idle_cpu(int cpu
)
1219 struct rq
*rq
= cpu_rq(cpu
);
1221 if (cpu
== smp_processor_id())
1225 * This is safe, as this function is called with the timer
1226 * wheel base lock of (cpu) held. When the CPU is on the way
1227 * to idle and has not yet set rq->curr to idle then it will
1228 * be serialized on the timer wheel base lock and take the new
1229 * timer into account automatically.
1231 if (rq
->curr
!= rq
->idle
)
1235 * We can set TIF_RESCHED on the idle task of the other CPU
1236 * lockless. The worst case is that the other CPU runs the
1237 * idle task through an additional NOOP schedule()
1239 set_tsk_need_resched(rq
->idle
);
1241 /* NEED_RESCHED must be visible before we test polling */
1243 if (!tsk_is_polling(rq
->idle
))
1244 smp_send_reschedule(cpu
);
1247 #endif /* CONFIG_NO_HZ */
1249 static u64
sched_avg_period(void)
1251 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1254 static void sched_avg_update(struct rq
*rq
)
1256 s64 period
= sched_avg_period();
1258 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1260 * Inline assembly required to prevent the compiler
1261 * optimising this loop into a divmod call.
1262 * See __iter_div_u64_rem() for another example of this.
1264 asm("" : "+rm" (rq
->age_stamp
));
1265 rq
->age_stamp
+= period
;
1270 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1272 rq
->rt_avg
+= rt_delta
;
1273 sched_avg_update(rq
);
1276 #else /* !CONFIG_SMP */
1277 static void resched_task(struct task_struct
*p
)
1279 assert_raw_spin_locked(&task_rq(p
)->lock
);
1280 set_tsk_need_resched(p
);
1283 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1287 static void sched_avg_update(struct rq
*rq
)
1290 #endif /* CONFIG_SMP */
1292 #if BITS_PER_LONG == 32
1293 # define WMULT_CONST (~0UL)
1295 # define WMULT_CONST (1UL << 32)
1298 #define WMULT_SHIFT 32
1301 * Shift right and round:
1303 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1306 * delta *= weight / lw
1308 static unsigned long
1309 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1310 struct load_weight
*lw
)
1314 if (!lw
->inv_weight
) {
1315 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1318 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1322 tmp
= (u64
)delta_exec
* weight
;
1324 * Check whether we'd overflow the 64-bit multiplication:
1326 if (unlikely(tmp
> WMULT_CONST
))
1327 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1330 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1332 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1335 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1341 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1347 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
1354 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1355 * of tasks with abnormal "nice" values across CPUs the contribution that
1356 * each task makes to its run queue's load is weighted according to its
1357 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1358 * scaled version of the new time slice allocation that they receive on time
1362 #define WEIGHT_IDLEPRIO 3
1363 #define WMULT_IDLEPRIO 1431655765
1366 * Nice levels are multiplicative, with a gentle 10% change for every
1367 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1368 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1369 * that remained on nice 0.
1371 * The "10% effect" is relative and cumulative: from _any_ nice level,
1372 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1373 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1374 * If a task goes up by ~10% and another task goes down by ~10% then
1375 * the relative distance between them is ~25%.)
1377 static const int prio_to_weight
[40] = {
1378 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1379 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1380 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1381 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1382 /* 0 */ 1024, 820, 655, 526, 423,
1383 /* 5 */ 335, 272, 215, 172, 137,
1384 /* 10 */ 110, 87, 70, 56, 45,
1385 /* 15 */ 36, 29, 23, 18, 15,
1389 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1391 * In cases where the weight does not change often, we can use the
1392 * precalculated inverse to speed up arithmetics by turning divisions
1393 * into multiplications:
1395 static const u32 prio_to_wmult
[40] = {
1396 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1397 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1398 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1399 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1400 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1401 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1402 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1403 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1406 /* Time spent by the tasks of the cpu accounting group executing in ... */
1407 enum cpuacct_stat_index
{
1408 CPUACCT_STAT_USER
, /* ... user mode */
1409 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1411 CPUACCT_STAT_NSTATS
,
1414 #ifdef CONFIG_CGROUP_CPUACCT
1415 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1416 static void cpuacct_update_stats(struct task_struct
*tsk
,
1417 enum cpuacct_stat_index idx
, cputime_t val
);
1419 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1420 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1421 enum cpuacct_stat_index idx
, cputime_t val
) {}
1424 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1426 update_load_add(&rq
->load
, load
);
1429 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1431 update_load_sub(&rq
->load
, load
);
1434 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1435 typedef int (*tg_visitor
)(struct task_group
*, void *);
1438 * Iterate the full tree, calling @down when first entering a node and @up when
1439 * leaving it for the final time.
1441 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1443 struct task_group
*parent
, *child
;
1447 parent
= &root_task_group
;
1449 ret
= (*down
)(parent
, data
);
1452 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1459 ret
= (*up
)(parent
, data
);
1464 parent
= parent
->parent
;
1473 static int tg_nop(struct task_group
*tg
, void *data
)
1480 /* Used instead of source_load when we know the type == 0 */
1481 static unsigned long weighted_cpuload(const int cpu
)
1483 return cpu_rq(cpu
)->load
.weight
;
1487 * Return a low guess at the load of a migration-source cpu weighted
1488 * according to the scheduling class and "nice" value.
1490 * We want to under-estimate the load of migration sources, to
1491 * balance conservatively.
1493 static unsigned long source_load(int cpu
, int type
)
1495 struct rq
*rq
= cpu_rq(cpu
);
1496 unsigned long total
= weighted_cpuload(cpu
);
1498 if (type
== 0 || !sched_feat(LB_BIAS
))
1501 return min(rq
->cpu_load
[type
-1], total
);
1505 * Return a high guess at the load of a migration-target cpu weighted
1506 * according to the scheduling class and "nice" value.
1508 static unsigned long target_load(int cpu
, int type
)
1510 struct rq
*rq
= cpu_rq(cpu
);
1511 unsigned long total
= weighted_cpuload(cpu
);
1513 if (type
== 0 || !sched_feat(LB_BIAS
))
1516 return max(rq
->cpu_load
[type
-1], total
);
1519 static unsigned long power_of(int cpu
)
1521 return cpu_rq(cpu
)->cpu_power
;
1524 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1526 static unsigned long cpu_avg_load_per_task(int cpu
)
1528 struct rq
*rq
= cpu_rq(cpu
);
1529 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1532 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1534 rq
->avg_load_per_task
= 0;
1536 return rq
->avg_load_per_task
;
1539 #ifdef CONFIG_FAIR_GROUP_SCHED
1542 * Compute the cpu's hierarchical load factor for each task group.
1543 * This needs to be done in a top-down fashion because the load of a child
1544 * group is a fraction of its parents load.
1546 static int tg_load_down(struct task_group
*tg
, void *data
)
1549 long cpu
= (long)data
;
1552 load
= cpu_rq(cpu
)->load
.weight
;
1554 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1555 load
*= tg
->se
[cpu
]->load
.weight
;
1556 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1559 tg
->cfs_rq
[cpu
]->h_load
= load
;
1564 static void update_h_load(long cpu
)
1566 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1571 #ifdef CONFIG_PREEMPT
1573 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1576 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1577 * way at the expense of forcing extra atomic operations in all
1578 * invocations. This assures that the double_lock is acquired using the
1579 * same underlying policy as the spinlock_t on this architecture, which
1580 * reduces latency compared to the unfair variant below. However, it
1581 * also adds more overhead and therefore may reduce throughput.
1583 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1584 __releases(this_rq
->lock
)
1585 __acquires(busiest
->lock
)
1586 __acquires(this_rq
->lock
)
1588 raw_spin_unlock(&this_rq
->lock
);
1589 double_rq_lock(this_rq
, busiest
);
1596 * Unfair double_lock_balance: Optimizes throughput at the expense of
1597 * latency by eliminating extra atomic operations when the locks are
1598 * already in proper order on entry. This favors lower cpu-ids and will
1599 * grant the double lock to lower cpus over higher ids under contention,
1600 * regardless of entry order into the function.
1602 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1603 __releases(this_rq
->lock
)
1604 __acquires(busiest
->lock
)
1605 __acquires(this_rq
->lock
)
1609 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1610 if (busiest
< this_rq
) {
1611 raw_spin_unlock(&this_rq
->lock
);
1612 raw_spin_lock(&busiest
->lock
);
1613 raw_spin_lock_nested(&this_rq
->lock
,
1614 SINGLE_DEPTH_NESTING
);
1617 raw_spin_lock_nested(&busiest
->lock
,
1618 SINGLE_DEPTH_NESTING
);
1623 #endif /* CONFIG_PREEMPT */
1626 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1628 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1630 if (unlikely(!irqs_disabled())) {
1631 /* printk() doesn't work good under rq->lock */
1632 raw_spin_unlock(&this_rq
->lock
);
1636 return _double_lock_balance(this_rq
, busiest
);
1639 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1640 __releases(busiest
->lock
)
1642 raw_spin_unlock(&busiest
->lock
);
1643 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1647 * double_rq_lock - safely lock two runqueues
1649 * Note this does not disable interrupts like task_rq_lock,
1650 * you need to do so manually before calling.
1652 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1653 __acquires(rq1
->lock
)
1654 __acquires(rq2
->lock
)
1656 BUG_ON(!irqs_disabled());
1658 raw_spin_lock(&rq1
->lock
);
1659 __acquire(rq2
->lock
); /* Fake it out ;) */
1662 raw_spin_lock(&rq1
->lock
);
1663 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1665 raw_spin_lock(&rq2
->lock
);
1666 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1672 * double_rq_unlock - safely unlock two runqueues
1674 * Note this does not restore interrupts like task_rq_unlock,
1675 * you need to do so manually after calling.
1677 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1678 __releases(rq1
->lock
)
1679 __releases(rq2
->lock
)
1681 raw_spin_unlock(&rq1
->lock
);
1683 raw_spin_unlock(&rq2
->lock
);
1685 __release(rq2
->lock
);
1690 static void calc_load_account_idle(struct rq
*this_rq
);
1691 static void update_sysctl(void);
1692 static int get_update_sysctl_factor(void);
1693 static void update_cpu_load(struct rq
*this_rq
);
1695 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1697 set_task_rq(p
, cpu
);
1700 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1701 * successfuly executed on another CPU. We must ensure that updates of
1702 * per-task data have been completed by this moment.
1705 task_thread_info(p
)->cpu
= cpu
;
1709 static const struct sched_class rt_sched_class
;
1711 #define sched_class_highest (&stop_sched_class)
1712 #define for_each_class(class) \
1713 for (class = sched_class_highest; class; class = class->next)
1715 #include "sched_stats.h"
1717 static void inc_nr_running(struct rq
*rq
)
1722 static void dec_nr_running(struct rq
*rq
)
1727 static void set_load_weight(struct task_struct
*p
)
1730 * SCHED_IDLE tasks get minimal weight:
1732 if (p
->policy
== SCHED_IDLE
) {
1733 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1734 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1738 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1739 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1742 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1744 update_rq_clock(rq
);
1745 sched_info_queued(p
);
1746 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1750 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1752 update_rq_clock(rq
);
1753 sched_info_dequeued(p
);
1754 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1759 * activate_task - move a task to the runqueue.
1761 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1763 if (task_contributes_to_load(p
))
1764 rq
->nr_uninterruptible
--;
1766 enqueue_task(rq
, p
, flags
);
1771 * deactivate_task - remove a task from the runqueue.
1773 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1775 if (task_contributes_to_load(p
))
1776 rq
->nr_uninterruptible
++;
1778 dequeue_task(rq
, p
, flags
);
1782 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1785 * There are no locks covering percpu hardirq/softirq time.
1786 * They are only modified in account_system_vtime, on corresponding CPU
1787 * with interrupts disabled. So, writes are safe.
1788 * They are read and saved off onto struct rq in update_rq_clock().
1789 * This may result in other CPU reading this CPU's irq time and can
1790 * race with irq/account_system_vtime on this CPU. We would either get old
1791 * or new value with a side effect of accounting a slice of irq time to wrong
1792 * task when irq is in progress while we read rq->clock. That is a worthy
1793 * compromise in place of having locks on each irq in account_system_time.
1795 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1796 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1798 static DEFINE_PER_CPU(u64
, irq_start_time
);
1799 static int sched_clock_irqtime
;
1801 void enable_sched_clock_irqtime(void)
1803 sched_clock_irqtime
= 1;
1806 void disable_sched_clock_irqtime(void)
1808 sched_clock_irqtime
= 0;
1811 #ifndef CONFIG_64BIT
1812 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
1814 static inline void irq_time_write_begin(void)
1816 __this_cpu_inc(irq_time_seq
.sequence
);
1820 static inline void irq_time_write_end(void)
1823 __this_cpu_inc(irq_time_seq
.sequence
);
1826 static inline u64
irq_time_read(int cpu
)
1832 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1833 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1834 per_cpu(cpu_hardirq_time
, cpu
);
1835 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1839 #else /* CONFIG_64BIT */
1840 static inline void irq_time_write_begin(void)
1844 static inline void irq_time_write_end(void)
1848 static inline u64
irq_time_read(int cpu
)
1850 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1852 #endif /* CONFIG_64BIT */
1855 * Called before incrementing preempt_count on {soft,}irq_enter
1856 * and before decrementing preempt_count on {soft,}irq_exit.
1858 void account_system_vtime(struct task_struct
*curr
)
1860 unsigned long flags
;
1864 if (!sched_clock_irqtime
)
1867 local_irq_save(flags
);
1869 cpu
= smp_processor_id();
1870 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
1871 __this_cpu_add(irq_start_time
, delta
);
1873 irq_time_write_begin();
1875 * We do not account for softirq time from ksoftirqd here.
1876 * We want to continue accounting softirq time to ksoftirqd thread
1877 * in that case, so as not to confuse scheduler with a special task
1878 * that do not consume any time, but still wants to run.
1880 if (hardirq_count())
1881 __this_cpu_add(cpu_hardirq_time
, delta
);
1882 else if (in_serving_softirq() && !(curr
->flags
& PF_KSOFTIRQD
))
1883 __this_cpu_add(cpu_softirq_time
, delta
);
1885 irq_time_write_end();
1886 local_irq_restore(flags
);
1888 EXPORT_SYMBOL_GPL(account_system_vtime
);
1890 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1894 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
1897 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1898 * this case when a previous update_rq_clock() happened inside a
1899 * {soft,}irq region.
1901 * When this happens, we stop ->clock_task and only update the
1902 * prev_irq_time stamp to account for the part that fit, so that a next
1903 * update will consume the rest. This ensures ->clock_task is
1906 * It does however cause some slight miss-attribution of {soft,}irq
1907 * time, a more accurate solution would be to update the irq_time using
1908 * the current rq->clock timestamp, except that would require using
1911 if (irq_delta
> delta
)
1914 rq
->prev_irq_time
+= irq_delta
;
1916 rq
->clock_task
+= delta
;
1918 if (irq_delta
&& sched_feat(NONIRQ_POWER
))
1919 sched_rt_avg_update(rq
, irq_delta
);
1922 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1924 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1926 rq
->clock_task
+= delta
;
1929 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1931 #include "sched_idletask.c"
1932 #include "sched_fair.c"
1933 #include "sched_rt.c"
1934 #include "sched_autogroup.c"
1935 #include "sched_stoptask.c"
1936 #ifdef CONFIG_SCHED_DEBUG
1937 # include "sched_debug.c"
1940 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1942 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1943 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1947 * Make it appear like a SCHED_FIFO task, its something
1948 * userspace knows about and won't get confused about.
1950 * Also, it will make PI more or less work without too
1951 * much confusion -- but then, stop work should not
1952 * rely on PI working anyway.
1954 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1956 stop
->sched_class
= &stop_sched_class
;
1959 cpu_rq(cpu
)->stop
= stop
;
1963 * Reset it back to a normal scheduling class so that
1964 * it can die in pieces.
1966 old_stop
->sched_class
= &rt_sched_class
;
1971 * __normal_prio - return the priority that is based on the static prio
1973 static inline int __normal_prio(struct task_struct
*p
)
1975 return p
->static_prio
;
1979 * Calculate the expected normal priority: i.e. priority
1980 * without taking RT-inheritance into account. Might be
1981 * boosted by interactivity modifiers. Changes upon fork,
1982 * setprio syscalls, and whenever the interactivity
1983 * estimator recalculates.
1985 static inline int normal_prio(struct task_struct
*p
)
1989 if (task_has_rt_policy(p
))
1990 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1992 prio
= __normal_prio(p
);
1997 * Calculate the current priority, i.e. the priority
1998 * taken into account by the scheduler. This value might
1999 * be boosted by RT tasks, or might be boosted by
2000 * interactivity modifiers. Will be RT if the task got
2001 * RT-boosted. If not then it returns p->normal_prio.
2003 static int effective_prio(struct task_struct
*p
)
2005 p
->normal_prio
= normal_prio(p
);
2007 * If we are RT tasks or we were boosted to RT priority,
2008 * keep the priority unchanged. Otherwise, update priority
2009 * to the normal priority:
2011 if (!rt_prio(p
->prio
))
2012 return p
->normal_prio
;
2017 * task_curr - is this task currently executing on a CPU?
2018 * @p: the task in question.
2020 inline int task_curr(const struct task_struct
*p
)
2022 return cpu_curr(task_cpu(p
)) == p
;
2025 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2026 const struct sched_class
*prev_class
,
2027 int oldprio
, int running
)
2029 if (prev_class
!= p
->sched_class
) {
2030 if (prev_class
->switched_from
)
2031 prev_class
->switched_from(rq
, p
, running
);
2032 p
->sched_class
->switched_to(rq
, p
, running
);
2034 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2037 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2039 const struct sched_class
*class;
2041 if (p
->sched_class
== rq
->curr
->sched_class
) {
2042 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2044 for_each_class(class) {
2045 if (class == rq
->curr
->sched_class
)
2047 if (class == p
->sched_class
) {
2048 resched_task(rq
->curr
);
2055 * A queue event has occurred, and we're going to schedule. In
2056 * this case, we can save a useless back to back clock update.
2058 if (rq
->curr
->se
.on_rq
&& test_tsk_need_resched(rq
->curr
))
2059 rq
->skip_clock_update
= 1;
2064 * Is this task likely cache-hot:
2067 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2071 if (p
->sched_class
!= &fair_sched_class
)
2074 if (unlikely(p
->policy
== SCHED_IDLE
))
2078 * Buddy candidates are cache hot:
2080 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2081 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2082 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2085 if (sysctl_sched_migration_cost
== -1)
2087 if (sysctl_sched_migration_cost
== 0)
2090 delta
= now
- p
->se
.exec_start
;
2092 return delta
< (s64
)sysctl_sched_migration_cost
;
2095 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2097 #ifdef CONFIG_SCHED_DEBUG
2099 * We should never call set_task_cpu() on a blocked task,
2100 * ttwu() will sort out the placement.
2102 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2103 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2106 trace_sched_migrate_task(p
, new_cpu
);
2108 if (task_cpu(p
) != new_cpu
) {
2109 p
->se
.nr_migrations
++;
2110 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2113 __set_task_cpu(p
, new_cpu
);
2116 struct migration_arg
{
2117 struct task_struct
*task
;
2121 static int migration_cpu_stop(void *data
);
2124 * The task's runqueue lock must be held.
2125 * Returns true if you have to wait for migration thread.
2127 static bool migrate_task(struct task_struct
*p
, struct rq
*rq
)
2130 * If the task is not on a runqueue (and not running), then
2131 * the next wake-up will properly place the task.
2133 return p
->se
.on_rq
|| task_running(rq
, p
);
2137 * wait_task_inactive - wait for a thread to unschedule.
2139 * If @match_state is nonzero, it's the @p->state value just checked and
2140 * not expected to change. If it changes, i.e. @p might have woken up,
2141 * then return zero. When we succeed in waiting for @p to be off its CPU,
2142 * we return a positive number (its total switch count). If a second call
2143 * a short while later returns the same number, the caller can be sure that
2144 * @p has remained unscheduled the whole time.
2146 * The caller must ensure that the task *will* unschedule sometime soon,
2147 * else this function might spin for a *long* time. This function can't
2148 * be called with interrupts off, or it may introduce deadlock with
2149 * smp_call_function() if an IPI is sent by the same process we are
2150 * waiting to become inactive.
2152 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2154 unsigned long flags
;
2161 * We do the initial early heuristics without holding
2162 * any task-queue locks at all. We'll only try to get
2163 * the runqueue lock when things look like they will
2169 * If the task is actively running on another CPU
2170 * still, just relax and busy-wait without holding
2173 * NOTE! Since we don't hold any locks, it's not
2174 * even sure that "rq" stays as the right runqueue!
2175 * But we don't care, since "task_running()" will
2176 * return false if the runqueue has changed and p
2177 * is actually now running somewhere else!
2179 while (task_running(rq
, p
)) {
2180 if (match_state
&& unlikely(p
->state
!= match_state
))
2186 * Ok, time to look more closely! We need the rq
2187 * lock now, to be *sure*. If we're wrong, we'll
2188 * just go back and repeat.
2190 rq
= task_rq_lock(p
, &flags
);
2191 trace_sched_wait_task(p
);
2192 running
= task_running(rq
, p
);
2193 on_rq
= p
->se
.on_rq
;
2195 if (!match_state
|| p
->state
== match_state
)
2196 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2197 task_rq_unlock(rq
, &flags
);
2200 * If it changed from the expected state, bail out now.
2202 if (unlikely(!ncsw
))
2206 * Was it really running after all now that we
2207 * checked with the proper locks actually held?
2209 * Oops. Go back and try again..
2211 if (unlikely(running
)) {
2217 * It's not enough that it's not actively running,
2218 * it must be off the runqueue _entirely_, and not
2221 * So if it was still runnable (but just not actively
2222 * running right now), it's preempted, and we should
2223 * yield - it could be a while.
2225 if (unlikely(on_rq
)) {
2226 schedule_timeout_uninterruptible(1);
2231 * Ahh, all good. It wasn't running, and it wasn't
2232 * runnable, which means that it will never become
2233 * running in the future either. We're all done!
2242 * kick_process - kick a running thread to enter/exit the kernel
2243 * @p: the to-be-kicked thread
2245 * Cause a process which is running on another CPU to enter
2246 * kernel-mode, without any delay. (to get signals handled.)
2248 * NOTE: this function doesnt have to take the runqueue lock,
2249 * because all it wants to ensure is that the remote task enters
2250 * the kernel. If the IPI races and the task has been migrated
2251 * to another CPU then no harm is done and the purpose has been
2254 void kick_process(struct task_struct
*p
)
2260 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2261 smp_send_reschedule(cpu
);
2264 EXPORT_SYMBOL_GPL(kick_process
);
2265 #endif /* CONFIG_SMP */
2268 * task_oncpu_function_call - call a function on the cpu on which a task runs
2269 * @p: the task to evaluate
2270 * @func: the function to be called
2271 * @info: the function call argument
2273 * Calls the function @func when the task is currently running. This might
2274 * be on the current CPU, which just calls the function directly
2276 void task_oncpu_function_call(struct task_struct
*p
,
2277 void (*func
) (void *info
), void *info
)
2284 smp_call_function_single(cpu
, func
, info
, 1);
2290 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2292 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2295 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2297 /* Look for allowed, online CPU in same node. */
2298 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2299 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2302 /* Any allowed, online CPU? */
2303 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2304 if (dest_cpu
< nr_cpu_ids
)
2307 /* No more Mr. Nice Guy. */
2308 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2310 * Don't tell them about moving exiting tasks or
2311 * kernel threads (both mm NULL), since they never
2314 if (p
->mm
&& printk_ratelimit()) {
2315 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2316 task_pid_nr(p
), p
->comm
, cpu
);
2323 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2326 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2328 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2331 * In order not to call set_task_cpu() on a blocking task we need
2332 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2335 * Since this is common to all placement strategies, this lives here.
2337 * [ this allows ->select_task() to simply return task_cpu(p) and
2338 * not worry about this generic constraint ]
2340 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2342 cpu
= select_fallback_rq(task_cpu(p
), p
);
2347 static void update_avg(u64
*avg
, u64 sample
)
2349 s64 diff
= sample
- *avg
;
2354 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2355 bool is_sync
, bool is_migrate
, bool is_local
,
2356 unsigned long en_flags
)
2358 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2360 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2362 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2364 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2366 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2368 activate_task(rq
, p
, en_flags
);
2371 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2372 int wake_flags
, bool success
)
2374 trace_sched_wakeup(p
, success
);
2375 check_preempt_curr(rq
, p
, wake_flags
);
2377 p
->state
= TASK_RUNNING
;
2379 if (p
->sched_class
->task_woken
)
2380 p
->sched_class
->task_woken(rq
, p
);
2382 if (unlikely(rq
->idle_stamp
)) {
2383 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2384 u64 max
= 2*sysctl_sched_migration_cost
;
2389 update_avg(&rq
->avg_idle
, delta
);
2393 /* if a worker is waking up, notify workqueue */
2394 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2395 wq_worker_waking_up(p
, cpu_of(rq
));
2399 * try_to_wake_up - wake up a thread
2400 * @p: the thread to be awakened
2401 * @state: the mask of task states that can be woken
2402 * @wake_flags: wake modifier flags (WF_*)
2404 * Put it on the run-queue if it's not already there. The "current"
2405 * thread is always on the run-queue (except when the actual
2406 * re-schedule is in progress), and as such you're allowed to do
2407 * the simpler "current->state = TASK_RUNNING" to mark yourself
2408 * runnable without the overhead of this.
2410 * Returns %true if @p was woken up, %false if it was already running
2411 * or @state didn't match @p's state.
2413 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2416 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2417 unsigned long flags
;
2418 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2421 this_cpu
= get_cpu();
2424 rq
= task_rq_lock(p
, &flags
);
2425 if (!(p
->state
& state
))
2435 if (unlikely(task_running(rq
, p
)))
2439 * In order to handle concurrent wakeups and release the rq->lock
2440 * we put the task in TASK_WAKING state.
2442 * First fix up the nr_uninterruptible count:
2444 if (task_contributes_to_load(p
)) {
2445 if (likely(cpu_online(orig_cpu
)))
2446 rq
->nr_uninterruptible
--;
2448 this_rq()->nr_uninterruptible
--;
2450 p
->state
= TASK_WAKING
;
2452 if (p
->sched_class
->task_waking
) {
2453 p
->sched_class
->task_waking(rq
, p
);
2454 en_flags
|= ENQUEUE_WAKING
;
2457 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2458 if (cpu
!= orig_cpu
)
2459 set_task_cpu(p
, cpu
);
2460 __task_rq_unlock(rq
);
2463 raw_spin_lock(&rq
->lock
);
2466 * We migrated the task without holding either rq->lock, however
2467 * since the task is not on the task list itself, nobody else
2468 * will try and migrate the task, hence the rq should match the
2469 * cpu we just moved it to.
2471 WARN_ON(task_cpu(p
) != cpu
);
2472 WARN_ON(p
->state
!= TASK_WAKING
);
2474 #ifdef CONFIG_SCHEDSTATS
2475 schedstat_inc(rq
, ttwu_count
);
2476 if (cpu
== this_cpu
)
2477 schedstat_inc(rq
, ttwu_local
);
2479 struct sched_domain
*sd
;
2480 for_each_domain(this_cpu
, sd
) {
2481 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2482 schedstat_inc(sd
, ttwu_wake_remote
);
2487 #endif /* CONFIG_SCHEDSTATS */
2490 #endif /* CONFIG_SMP */
2491 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2492 cpu
== this_cpu
, en_flags
);
2495 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2497 task_rq_unlock(rq
, &flags
);
2504 * try_to_wake_up_local - try to wake up a local task with rq lock held
2505 * @p: the thread to be awakened
2507 * Put @p on the run-queue if it's not already there. The caller must
2508 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2509 * the current task. this_rq() stays locked over invocation.
2511 static void try_to_wake_up_local(struct task_struct
*p
)
2513 struct rq
*rq
= task_rq(p
);
2514 bool success
= false;
2516 BUG_ON(rq
!= this_rq());
2517 BUG_ON(p
== current
);
2518 lockdep_assert_held(&rq
->lock
);
2520 if (!(p
->state
& TASK_NORMAL
))
2524 if (likely(!task_running(rq
, p
))) {
2525 schedstat_inc(rq
, ttwu_count
);
2526 schedstat_inc(rq
, ttwu_local
);
2528 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2531 ttwu_post_activation(p
, rq
, 0, success
);
2535 * wake_up_process - Wake up a specific process
2536 * @p: The process to be woken up.
2538 * Attempt to wake up the nominated process and move it to the set of runnable
2539 * processes. Returns 1 if the process was woken up, 0 if it was already
2542 * It may be assumed that this function implies a write memory barrier before
2543 * changing the task state if and only if any tasks are woken up.
2545 int wake_up_process(struct task_struct
*p
)
2547 return try_to_wake_up(p
, TASK_ALL
, 0);
2549 EXPORT_SYMBOL(wake_up_process
);
2551 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2553 return try_to_wake_up(p
, state
, 0);
2557 * Perform scheduler related setup for a newly forked process p.
2558 * p is forked by current.
2560 * __sched_fork() is basic setup used by init_idle() too:
2562 static void __sched_fork(struct task_struct
*p
)
2564 p
->se
.exec_start
= 0;
2565 p
->se
.sum_exec_runtime
= 0;
2566 p
->se
.prev_sum_exec_runtime
= 0;
2567 p
->se
.nr_migrations
= 0;
2569 #ifdef CONFIG_SCHEDSTATS
2570 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2573 INIT_LIST_HEAD(&p
->rt
.run_list
);
2575 INIT_LIST_HEAD(&p
->se
.group_node
);
2577 #ifdef CONFIG_PREEMPT_NOTIFIERS
2578 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2583 * fork()/clone()-time setup:
2585 void sched_fork(struct task_struct
*p
, int clone_flags
)
2587 int cpu
= get_cpu();
2591 * We mark the process as running here. This guarantees that
2592 * nobody will actually run it, and a signal or other external
2593 * event cannot wake it up and insert it on the runqueue either.
2595 p
->state
= TASK_RUNNING
;
2598 * Revert to default priority/policy on fork if requested.
2600 if (unlikely(p
->sched_reset_on_fork
)) {
2601 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2602 p
->policy
= SCHED_NORMAL
;
2603 p
->normal_prio
= p
->static_prio
;
2606 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2607 p
->static_prio
= NICE_TO_PRIO(0);
2608 p
->normal_prio
= p
->static_prio
;
2613 * We don't need the reset flag anymore after the fork. It has
2614 * fulfilled its duty:
2616 p
->sched_reset_on_fork
= 0;
2620 * Make sure we do not leak PI boosting priority to the child.
2622 p
->prio
= current
->normal_prio
;
2624 if (!rt_prio(p
->prio
))
2625 p
->sched_class
= &fair_sched_class
;
2627 if (p
->sched_class
->task_fork
)
2628 p
->sched_class
->task_fork(p
);
2631 * The child is not yet in the pid-hash so no cgroup attach races,
2632 * and the cgroup is pinned to this child due to cgroup_fork()
2633 * is ran before sched_fork().
2635 * Silence PROVE_RCU.
2638 set_task_cpu(p
, cpu
);
2641 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2642 if (likely(sched_info_on()))
2643 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2645 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2648 #ifdef CONFIG_PREEMPT
2649 /* Want to start with kernel preemption disabled. */
2650 task_thread_info(p
)->preempt_count
= 1;
2653 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2660 * wake_up_new_task - wake up a newly created task for the first time.
2662 * This function will do some initial scheduler statistics housekeeping
2663 * that must be done for every newly created context, then puts the task
2664 * on the runqueue and wakes it.
2666 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2668 unsigned long flags
;
2670 int cpu __maybe_unused
= get_cpu();
2673 rq
= task_rq_lock(p
, &flags
);
2674 p
->state
= TASK_WAKING
;
2677 * Fork balancing, do it here and not earlier because:
2678 * - cpus_allowed can change in the fork path
2679 * - any previously selected cpu might disappear through hotplug
2681 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2682 * without people poking at ->cpus_allowed.
2684 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2685 set_task_cpu(p
, cpu
);
2687 p
->state
= TASK_RUNNING
;
2688 task_rq_unlock(rq
, &flags
);
2691 rq
= task_rq_lock(p
, &flags
);
2692 activate_task(rq
, p
, 0);
2693 trace_sched_wakeup_new(p
, 1);
2694 check_preempt_curr(rq
, p
, WF_FORK
);
2696 if (p
->sched_class
->task_woken
)
2697 p
->sched_class
->task_woken(rq
, p
);
2699 task_rq_unlock(rq
, &flags
);
2703 #ifdef CONFIG_PREEMPT_NOTIFIERS
2706 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2707 * @notifier: notifier struct to register
2709 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2711 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2713 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2716 * preempt_notifier_unregister - no longer interested in preemption notifications
2717 * @notifier: notifier struct to unregister
2719 * This is safe to call from within a preemption notifier.
2721 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2723 hlist_del(¬ifier
->link
);
2725 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2727 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2729 struct preempt_notifier
*notifier
;
2730 struct hlist_node
*node
;
2732 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2733 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2737 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2738 struct task_struct
*next
)
2740 struct preempt_notifier
*notifier
;
2741 struct hlist_node
*node
;
2743 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2744 notifier
->ops
->sched_out(notifier
, next
);
2747 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2749 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2754 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2755 struct task_struct
*next
)
2759 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2762 * prepare_task_switch - prepare to switch tasks
2763 * @rq: the runqueue preparing to switch
2764 * @prev: the current task that is being switched out
2765 * @next: the task we are going to switch to.
2767 * This is called with the rq lock held and interrupts off. It must
2768 * be paired with a subsequent finish_task_switch after the context
2771 * prepare_task_switch sets up locking and calls architecture specific
2775 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2776 struct task_struct
*next
)
2778 fire_sched_out_preempt_notifiers(prev
, next
);
2779 prepare_lock_switch(rq
, next
);
2780 prepare_arch_switch(next
);
2784 * finish_task_switch - clean up after a task-switch
2785 * @rq: runqueue associated with task-switch
2786 * @prev: the thread we just switched away from.
2788 * finish_task_switch must be called after the context switch, paired
2789 * with a prepare_task_switch call before the context switch.
2790 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2791 * and do any other architecture-specific cleanup actions.
2793 * Note that we may have delayed dropping an mm in context_switch(). If
2794 * so, we finish that here outside of the runqueue lock. (Doing it
2795 * with the lock held can cause deadlocks; see schedule() for
2798 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2799 __releases(rq
->lock
)
2801 struct mm_struct
*mm
= rq
->prev_mm
;
2807 * A task struct has one reference for the use as "current".
2808 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2809 * schedule one last time. The schedule call will never return, and
2810 * the scheduled task must drop that reference.
2811 * The test for TASK_DEAD must occur while the runqueue locks are
2812 * still held, otherwise prev could be scheduled on another cpu, die
2813 * there before we look at prev->state, and then the reference would
2815 * Manfred Spraul <manfred@colorfullife.com>
2817 prev_state
= prev
->state
;
2818 finish_arch_switch(prev
);
2819 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2820 local_irq_disable();
2821 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2822 perf_event_task_sched_in(current
);
2823 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2825 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2826 finish_lock_switch(rq
, prev
);
2828 fire_sched_in_preempt_notifiers(current
);
2831 if (unlikely(prev_state
== TASK_DEAD
)) {
2833 * Remove function-return probe instances associated with this
2834 * task and put them back on the free list.
2836 kprobe_flush_task(prev
);
2837 put_task_struct(prev
);
2843 /* assumes rq->lock is held */
2844 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2846 if (prev
->sched_class
->pre_schedule
)
2847 prev
->sched_class
->pre_schedule(rq
, prev
);
2850 /* rq->lock is NOT held, but preemption is disabled */
2851 static inline void post_schedule(struct rq
*rq
)
2853 if (rq
->post_schedule
) {
2854 unsigned long flags
;
2856 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2857 if (rq
->curr
->sched_class
->post_schedule
)
2858 rq
->curr
->sched_class
->post_schedule(rq
);
2859 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2861 rq
->post_schedule
= 0;
2867 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2871 static inline void post_schedule(struct rq
*rq
)
2878 * schedule_tail - first thing a freshly forked thread must call.
2879 * @prev: the thread we just switched away from.
2881 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2882 __releases(rq
->lock
)
2884 struct rq
*rq
= this_rq();
2886 finish_task_switch(rq
, prev
);
2889 * FIXME: do we need to worry about rq being invalidated by the
2894 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2895 /* In this case, finish_task_switch does not reenable preemption */
2898 if (current
->set_child_tid
)
2899 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2903 * context_switch - switch to the new MM and the new
2904 * thread's register state.
2907 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2908 struct task_struct
*next
)
2910 struct mm_struct
*mm
, *oldmm
;
2912 prepare_task_switch(rq
, prev
, next
);
2913 trace_sched_switch(prev
, next
);
2915 oldmm
= prev
->active_mm
;
2917 * For paravirt, this is coupled with an exit in switch_to to
2918 * combine the page table reload and the switch backend into
2921 arch_start_context_switch(prev
);
2924 next
->active_mm
= oldmm
;
2925 atomic_inc(&oldmm
->mm_count
);
2926 enter_lazy_tlb(oldmm
, next
);
2928 switch_mm(oldmm
, mm
, next
);
2931 prev
->active_mm
= NULL
;
2932 rq
->prev_mm
= oldmm
;
2935 * Since the runqueue lock will be released by the next
2936 * task (which is an invalid locking op but in the case
2937 * of the scheduler it's an obvious special-case), so we
2938 * do an early lockdep release here:
2940 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2941 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2944 /* Here we just switch the register state and the stack. */
2945 switch_to(prev
, next
, prev
);
2949 * this_rq must be evaluated again because prev may have moved
2950 * CPUs since it called schedule(), thus the 'rq' on its stack
2951 * frame will be invalid.
2953 finish_task_switch(this_rq(), prev
);
2957 * nr_running, nr_uninterruptible and nr_context_switches:
2959 * externally visible scheduler statistics: current number of runnable
2960 * threads, current number of uninterruptible-sleeping threads, total
2961 * number of context switches performed since bootup.
2963 unsigned long nr_running(void)
2965 unsigned long i
, sum
= 0;
2967 for_each_online_cpu(i
)
2968 sum
+= cpu_rq(i
)->nr_running
;
2973 unsigned long nr_uninterruptible(void)
2975 unsigned long i
, sum
= 0;
2977 for_each_possible_cpu(i
)
2978 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2981 * Since we read the counters lockless, it might be slightly
2982 * inaccurate. Do not allow it to go below zero though:
2984 if (unlikely((long)sum
< 0))
2990 unsigned long long nr_context_switches(void)
2993 unsigned long long sum
= 0;
2995 for_each_possible_cpu(i
)
2996 sum
+= cpu_rq(i
)->nr_switches
;
3001 unsigned long nr_iowait(void)
3003 unsigned long i
, sum
= 0;
3005 for_each_possible_cpu(i
)
3006 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3011 unsigned long nr_iowait_cpu(int cpu
)
3013 struct rq
*this = cpu_rq(cpu
);
3014 return atomic_read(&this->nr_iowait
);
3017 unsigned long this_cpu_load(void)
3019 struct rq
*this = this_rq();
3020 return this->cpu_load
[0];
3024 /* Variables and functions for calc_load */
3025 static atomic_long_t calc_load_tasks
;
3026 static unsigned long calc_load_update
;
3027 unsigned long avenrun
[3];
3028 EXPORT_SYMBOL(avenrun
);
3030 static long calc_load_fold_active(struct rq
*this_rq
)
3032 long nr_active
, delta
= 0;
3034 nr_active
= this_rq
->nr_running
;
3035 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3037 if (nr_active
!= this_rq
->calc_load_active
) {
3038 delta
= nr_active
- this_rq
->calc_load_active
;
3039 this_rq
->calc_load_active
= nr_active
;
3045 static unsigned long
3046 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3049 load
+= active
* (FIXED_1
- exp
);
3050 load
+= 1UL << (FSHIFT
- 1);
3051 return load
>> FSHIFT
;
3056 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3058 * When making the ILB scale, we should try to pull this in as well.
3060 static atomic_long_t calc_load_tasks_idle
;
3062 static void calc_load_account_idle(struct rq
*this_rq
)
3066 delta
= calc_load_fold_active(this_rq
);
3068 atomic_long_add(delta
, &calc_load_tasks_idle
);
3071 static long calc_load_fold_idle(void)
3076 * Its got a race, we don't care...
3078 if (atomic_long_read(&calc_load_tasks_idle
))
3079 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3085 * fixed_power_int - compute: x^n, in O(log n) time
3087 * @x: base of the power
3088 * @frac_bits: fractional bits of @x
3089 * @n: power to raise @x to.
3091 * By exploiting the relation between the definition of the natural power
3092 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3093 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3094 * (where: n_i \elem {0, 1}, the binary vector representing n),
3095 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3096 * of course trivially computable in O(log_2 n), the length of our binary
3099 static unsigned long
3100 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3102 unsigned long result
= 1UL << frac_bits
;
3107 result
+= 1UL << (frac_bits
- 1);
3108 result
>>= frac_bits
;
3114 x
+= 1UL << (frac_bits
- 1);
3122 * a1 = a0 * e + a * (1 - e)
3124 * a2 = a1 * e + a * (1 - e)
3125 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3126 * = a0 * e^2 + a * (1 - e) * (1 + e)
3128 * a3 = a2 * e + a * (1 - e)
3129 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3130 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3134 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3135 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3136 * = a0 * e^n + a * (1 - e^n)
3138 * [1] application of the geometric series:
3141 * S_n := \Sum x^i = -------------
3144 static unsigned long
3145 calc_load_n(unsigned long load
, unsigned long exp
,
3146 unsigned long active
, unsigned int n
)
3149 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3153 * NO_HZ can leave us missing all per-cpu ticks calling
3154 * calc_load_account_active(), but since an idle CPU folds its delta into
3155 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3156 * in the pending idle delta if our idle period crossed a load cycle boundary.
3158 * Once we've updated the global active value, we need to apply the exponential
3159 * weights adjusted to the number of cycles missed.
3161 static void calc_global_nohz(unsigned long ticks
)
3163 long delta
, active
, n
;
3165 if (time_before(jiffies
, calc_load_update
))
3169 * If we crossed a calc_load_update boundary, make sure to fold
3170 * any pending idle changes, the respective CPUs might have
3171 * missed the tick driven calc_load_account_active() update
3174 delta
= calc_load_fold_idle();
3176 atomic_long_add(delta
, &calc_load_tasks
);
3179 * If we were idle for multiple load cycles, apply them.
3181 if (ticks
>= LOAD_FREQ
) {
3182 n
= ticks
/ LOAD_FREQ
;
3184 active
= atomic_long_read(&calc_load_tasks
);
3185 active
= active
> 0 ? active
* FIXED_1
: 0;
3187 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3188 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3189 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3191 calc_load_update
+= n
* LOAD_FREQ
;
3195 * Its possible the remainder of the above division also crosses
3196 * a LOAD_FREQ period, the regular check in calc_global_load()
3197 * which comes after this will take care of that.
3199 * Consider us being 11 ticks before a cycle completion, and us
3200 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3201 * age us 4 cycles, and the test in calc_global_load() will
3202 * pick up the final one.
3206 static void calc_load_account_idle(struct rq
*this_rq
)
3210 static inline long calc_load_fold_idle(void)
3215 static void calc_global_nohz(unsigned long ticks
)
3221 * get_avenrun - get the load average array
3222 * @loads: pointer to dest load array
3223 * @offset: offset to add
3224 * @shift: shift count to shift the result left
3226 * These values are estimates at best, so no need for locking.
3228 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3230 loads
[0] = (avenrun
[0] + offset
) << shift
;
3231 loads
[1] = (avenrun
[1] + offset
) << shift
;
3232 loads
[2] = (avenrun
[2] + offset
) << shift
;
3236 * calc_load - update the avenrun load estimates 10 ticks after the
3237 * CPUs have updated calc_load_tasks.
3239 void calc_global_load(unsigned long ticks
)
3243 calc_global_nohz(ticks
);
3245 if (time_before(jiffies
, calc_load_update
+ 10))
3248 active
= atomic_long_read(&calc_load_tasks
);
3249 active
= active
> 0 ? active
* FIXED_1
: 0;
3251 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3252 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3253 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3255 calc_load_update
+= LOAD_FREQ
;
3259 * Called from update_cpu_load() to periodically update this CPU's
3262 static void calc_load_account_active(struct rq
*this_rq
)
3266 if (time_before(jiffies
, this_rq
->calc_load_update
))
3269 delta
= calc_load_fold_active(this_rq
);
3270 delta
+= calc_load_fold_idle();
3272 atomic_long_add(delta
, &calc_load_tasks
);
3274 this_rq
->calc_load_update
+= LOAD_FREQ
;
3278 * The exact cpuload at various idx values, calculated at every tick would be
3279 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3281 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3282 * on nth tick when cpu may be busy, then we have:
3283 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3284 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3286 * decay_load_missed() below does efficient calculation of
3287 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3288 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3290 * The calculation is approximated on a 128 point scale.
3291 * degrade_zero_ticks is the number of ticks after which load at any
3292 * particular idx is approximated to be zero.
3293 * degrade_factor is a precomputed table, a row for each load idx.
3294 * Each column corresponds to degradation factor for a power of two ticks,
3295 * based on 128 point scale.
3297 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3298 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3300 * With this power of 2 load factors, we can degrade the load n times
3301 * by looking at 1 bits in n and doing as many mult/shift instead of
3302 * n mult/shifts needed by the exact degradation.
3304 #define DEGRADE_SHIFT 7
3305 static const unsigned char
3306 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3307 static const unsigned char
3308 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3309 {0, 0, 0, 0, 0, 0, 0, 0},
3310 {64, 32, 8, 0, 0, 0, 0, 0},
3311 {96, 72, 40, 12, 1, 0, 0},
3312 {112, 98, 75, 43, 15, 1, 0},
3313 {120, 112, 98, 76, 45, 16, 2} };
3316 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3317 * would be when CPU is idle and so we just decay the old load without
3318 * adding any new load.
3320 static unsigned long
3321 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3325 if (!missed_updates
)
3328 if (missed_updates
>= degrade_zero_ticks
[idx
])
3332 return load
>> missed_updates
;
3334 while (missed_updates
) {
3335 if (missed_updates
% 2)
3336 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3338 missed_updates
>>= 1;
3345 * Update rq->cpu_load[] statistics. This function is usually called every
3346 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3347 * every tick. We fix it up based on jiffies.
3349 static void update_cpu_load(struct rq
*this_rq
)
3351 unsigned long this_load
= this_rq
->load
.weight
;
3352 unsigned long curr_jiffies
= jiffies
;
3353 unsigned long pending_updates
;
3356 this_rq
->nr_load_updates
++;
3358 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3359 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3362 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3363 this_rq
->last_load_update_tick
= curr_jiffies
;
3365 /* Update our load: */
3366 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3367 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3368 unsigned long old_load
, new_load
;
3370 /* scale is effectively 1 << i now, and >> i divides by scale */
3372 old_load
= this_rq
->cpu_load
[i
];
3373 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3374 new_load
= this_load
;
3376 * Round up the averaging division if load is increasing. This
3377 * prevents us from getting stuck on 9 if the load is 10, for
3380 if (new_load
> old_load
)
3381 new_load
+= scale
- 1;
3383 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3386 sched_avg_update(this_rq
);
3389 static void update_cpu_load_active(struct rq
*this_rq
)
3391 update_cpu_load(this_rq
);
3393 calc_load_account_active(this_rq
);
3399 * sched_exec - execve() is a valuable balancing opportunity, because at
3400 * this point the task has the smallest effective memory and cache footprint.
3402 void sched_exec(void)
3404 struct task_struct
*p
= current
;
3405 unsigned long flags
;
3409 rq
= task_rq_lock(p
, &flags
);
3410 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3411 if (dest_cpu
== smp_processor_id())
3415 * select_task_rq() can race against ->cpus_allowed
3417 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3418 likely(cpu_active(dest_cpu
)) && migrate_task(p
, rq
)) {
3419 struct migration_arg arg
= { p
, dest_cpu
};
3421 task_rq_unlock(rq
, &flags
);
3422 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3426 task_rq_unlock(rq
, &flags
);
3431 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3433 EXPORT_PER_CPU_SYMBOL(kstat
);
3436 * Return any ns on the sched_clock that have not yet been accounted in
3437 * @p in case that task is currently running.
3439 * Called with task_rq_lock() held on @rq.
3441 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3445 if (task_current(rq
, p
)) {
3446 update_rq_clock(rq
);
3447 ns
= rq
->clock_task
- p
->se
.exec_start
;
3455 unsigned long long task_delta_exec(struct task_struct
*p
)
3457 unsigned long flags
;
3461 rq
= task_rq_lock(p
, &flags
);
3462 ns
= do_task_delta_exec(p
, rq
);
3463 task_rq_unlock(rq
, &flags
);
3469 * Return accounted runtime for the task.
3470 * In case the task is currently running, return the runtime plus current's
3471 * pending runtime that have not been accounted yet.
3473 unsigned long long task_sched_runtime(struct task_struct
*p
)
3475 unsigned long flags
;
3479 rq
= task_rq_lock(p
, &flags
);
3480 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3481 task_rq_unlock(rq
, &flags
);
3487 * Return sum_exec_runtime for the thread group.
3488 * In case the task is currently running, return the sum plus current's
3489 * pending runtime that have not been accounted yet.
3491 * Note that the thread group might have other running tasks as well,
3492 * so the return value not includes other pending runtime that other
3493 * running tasks might have.
3495 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3497 struct task_cputime totals
;
3498 unsigned long flags
;
3502 rq
= task_rq_lock(p
, &flags
);
3503 thread_group_cputime(p
, &totals
);
3504 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3505 task_rq_unlock(rq
, &flags
);
3511 * Account user cpu time to a process.
3512 * @p: the process that the cpu time gets accounted to
3513 * @cputime: the cpu time spent in user space since the last update
3514 * @cputime_scaled: cputime scaled by cpu frequency
3516 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3517 cputime_t cputime_scaled
)
3519 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3522 /* Add user time to process. */
3523 p
->utime
= cputime_add(p
->utime
, cputime
);
3524 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3525 account_group_user_time(p
, cputime
);
3527 /* Add user time to cpustat. */
3528 tmp
= cputime_to_cputime64(cputime
);
3529 if (TASK_NICE(p
) > 0)
3530 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3532 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3534 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3535 /* Account for user time used */
3536 acct_update_integrals(p
);
3540 * Account guest cpu time to a process.
3541 * @p: the process that the cpu time gets accounted to
3542 * @cputime: the cpu time spent in virtual machine since the last update
3543 * @cputime_scaled: cputime scaled by cpu frequency
3545 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3546 cputime_t cputime_scaled
)
3549 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3551 tmp
= cputime_to_cputime64(cputime
);
3553 /* Add guest time to process. */
3554 p
->utime
= cputime_add(p
->utime
, cputime
);
3555 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3556 account_group_user_time(p
, cputime
);
3557 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3559 /* Add guest time to cpustat. */
3560 if (TASK_NICE(p
) > 0) {
3561 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3562 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3564 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3565 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3570 * Account system cpu time to a process.
3571 * @p: the process that the cpu time gets accounted to
3572 * @hardirq_offset: the offset to subtract from hardirq_count()
3573 * @cputime: the cpu time spent in kernel space since the last update
3574 * @cputime_scaled: cputime scaled by cpu frequency
3576 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3577 cputime_t cputime
, cputime_t cputime_scaled
)
3579 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3582 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3583 account_guest_time(p
, cputime
, cputime_scaled
);
3587 /* Add system time to process. */
3588 p
->stime
= cputime_add(p
->stime
, cputime
);
3589 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3590 account_group_system_time(p
, cputime
);
3592 /* Add system time to cpustat. */
3593 tmp
= cputime_to_cputime64(cputime
);
3594 if (hardirq_count() - hardirq_offset
)
3595 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3596 else if (in_serving_softirq())
3597 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3599 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3601 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3603 /* Account for system time used */
3604 acct_update_integrals(p
);
3608 * Account for involuntary wait time.
3609 * @steal: the cpu time spent in involuntary wait
3611 void account_steal_time(cputime_t cputime
)
3613 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3614 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3616 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3620 * Account for idle time.
3621 * @cputime: the cpu time spent in idle wait
3623 void account_idle_time(cputime_t cputime
)
3625 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3626 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3627 struct rq
*rq
= this_rq();
3629 if (atomic_read(&rq
->nr_iowait
) > 0)
3630 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3632 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3635 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3638 * Account a single tick of cpu time.
3639 * @p: the process that the cpu time gets accounted to
3640 * @user_tick: indicates if the tick is a user or a system tick
3642 void account_process_tick(struct task_struct
*p
, int user_tick
)
3644 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3645 struct rq
*rq
= this_rq();
3648 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3649 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3650 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3653 account_idle_time(cputime_one_jiffy
);
3657 * Account multiple ticks of steal time.
3658 * @p: the process from which the cpu time has been stolen
3659 * @ticks: number of stolen ticks
3661 void account_steal_ticks(unsigned long ticks
)
3663 account_steal_time(jiffies_to_cputime(ticks
));
3667 * Account multiple ticks of idle time.
3668 * @ticks: number of stolen ticks
3670 void account_idle_ticks(unsigned long ticks
)
3672 account_idle_time(jiffies_to_cputime(ticks
));
3678 * Use precise platform statistics if available:
3680 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3681 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3687 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3689 struct task_cputime cputime
;
3691 thread_group_cputime(p
, &cputime
);
3693 *ut
= cputime
.utime
;
3694 *st
= cputime
.stime
;
3698 #ifndef nsecs_to_cputime
3699 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3702 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3704 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3707 * Use CFS's precise accounting:
3709 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3715 do_div(temp
, total
);
3716 utime
= (cputime_t
)temp
;
3721 * Compare with previous values, to keep monotonicity:
3723 p
->prev_utime
= max(p
->prev_utime
, utime
);
3724 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3726 *ut
= p
->prev_utime
;
3727 *st
= p
->prev_stime
;
3731 * Must be called with siglock held.
3733 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3735 struct signal_struct
*sig
= p
->signal
;
3736 struct task_cputime cputime
;
3737 cputime_t rtime
, utime
, total
;
3739 thread_group_cputime(p
, &cputime
);
3741 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3742 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3747 temp
*= cputime
.utime
;
3748 do_div(temp
, total
);
3749 utime
= (cputime_t
)temp
;
3753 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3754 sig
->prev_stime
= max(sig
->prev_stime
,
3755 cputime_sub(rtime
, sig
->prev_utime
));
3757 *ut
= sig
->prev_utime
;
3758 *st
= sig
->prev_stime
;
3763 * This function gets called by the timer code, with HZ frequency.
3764 * We call it with interrupts disabled.
3766 * It also gets called by the fork code, when changing the parent's
3769 void scheduler_tick(void)
3771 int cpu
= smp_processor_id();
3772 struct rq
*rq
= cpu_rq(cpu
);
3773 struct task_struct
*curr
= rq
->curr
;
3777 raw_spin_lock(&rq
->lock
);
3778 update_rq_clock(rq
);
3779 update_cpu_load_active(rq
);
3780 curr
->sched_class
->task_tick(rq
, curr
, 0);
3781 raw_spin_unlock(&rq
->lock
);
3783 perf_event_task_tick();
3786 rq
->idle_at_tick
= idle_cpu(cpu
);
3787 trigger_load_balance(rq
, cpu
);
3791 notrace
unsigned long get_parent_ip(unsigned long addr
)
3793 if (in_lock_functions(addr
)) {
3794 addr
= CALLER_ADDR2
;
3795 if (in_lock_functions(addr
))
3796 addr
= CALLER_ADDR3
;
3801 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3802 defined(CONFIG_PREEMPT_TRACER))
3804 void __kprobes
add_preempt_count(int val
)
3806 #ifdef CONFIG_DEBUG_PREEMPT
3810 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3813 preempt_count() += val
;
3814 #ifdef CONFIG_DEBUG_PREEMPT
3816 * Spinlock count overflowing soon?
3818 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3821 if (preempt_count() == val
)
3822 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3824 EXPORT_SYMBOL(add_preempt_count
);
3826 void __kprobes
sub_preempt_count(int val
)
3828 #ifdef CONFIG_DEBUG_PREEMPT
3832 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3835 * Is the spinlock portion underflowing?
3837 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3838 !(preempt_count() & PREEMPT_MASK
)))
3842 if (preempt_count() == val
)
3843 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3844 preempt_count() -= val
;
3846 EXPORT_SYMBOL(sub_preempt_count
);
3851 * Print scheduling while atomic bug:
3853 static noinline
void __schedule_bug(struct task_struct
*prev
)
3855 struct pt_regs
*regs
= get_irq_regs();
3857 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3858 prev
->comm
, prev
->pid
, preempt_count());
3860 debug_show_held_locks(prev
);
3862 if (irqs_disabled())
3863 print_irqtrace_events(prev
);
3872 * Various schedule()-time debugging checks and statistics:
3874 static inline void schedule_debug(struct task_struct
*prev
)
3877 * Test if we are atomic. Since do_exit() needs to call into
3878 * schedule() atomically, we ignore that path for now.
3879 * Otherwise, whine if we are scheduling when we should not be.
3881 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3882 __schedule_bug(prev
);
3884 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3886 schedstat_inc(this_rq(), sched_count
);
3887 #ifdef CONFIG_SCHEDSTATS
3888 if (unlikely(prev
->lock_depth
>= 0)) {
3889 schedstat_inc(this_rq(), rq_sched_info
.bkl_count
);
3890 schedstat_inc(prev
, sched_info
.bkl_count
);
3895 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3898 update_rq_clock(rq
);
3899 prev
->sched_class
->put_prev_task(rq
, prev
);
3903 * Pick up the highest-prio task:
3905 static inline struct task_struct
*
3906 pick_next_task(struct rq
*rq
)
3908 const struct sched_class
*class;
3909 struct task_struct
*p
;
3912 * Optimization: we know that if all tasks are in
3913 * the fair class we can call that function directly:
3915 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3916 p
= fair_sched_class
.pick_next_task(rq
);
3921 for_each_class(class) {
3922 p
= class->pick_next_task(rq
);
3927 BUG(); /* the idle class will always have a runnable task */
3931 * schedule() is the main scheduler function.
3933 asmlinkage
void __sched
schedule(void)
3935 struct task_struct
*prev
, *next
;
3936 unsigned long *switch_count
;
3942 cpu
= smp_processor_id();
3944 rcu_note_context_switch(cpu
);
3947 schedule_debug(prev
);
3949 if (sched_feat(HRTICK
))
3952 raw_spin_lock_irq(&rq
->lock
);
3954 switch_count
= &prev
->nivcsw
;
3955 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3956 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
3957 prev
->state
= TASK_RUNNING
;
3960 * If a worker is going to sleep, notify and
3961 * ask workqueue whether it wants to wake up a
3962 * task to maintain concurrency. If so, wake
3965 if (prev
->flags
& PF_WQ_WORKER
) {
3966 struct task_struct
*to_wakeup
;
3968 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
3970 try_to_wake_up_local(to_wakeup
);
3972 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3974 switch_count
= &prev
->nvcsw
;
3977 pre_schedule(rq
, prev
);
3979 if (unlikely(!rq
->nr_running
))
3980 idle_balance(cpu
, rq
);
3982 put_prev_task(rq
, prev
);
3983 next
= pick_next_task(rq
);
3984 clear_tsk_need_resched(prev
);
3985 rq
->skip_clock_update
= 0;
3987 if (likely(prev
!= next
)) {
3988 sched_info_switch(prev
, next
);
3989 perf_event_task_sched_out(prev
, next
);
3995 context_switch(rq
, prev
, next
); /* unlocks the rq */
3997 * The context switch have flipped the stack from under us
3998 * and restored the local variables which were saved when
3999 * this task called schedule() in the past. prev == current
4000 * is still correct, but it can be moved to another cpu/rq.
4002 cpu
= smp_processor_id();
4005 raw_spin_unlock_irq(&rq
->lock
);
4009 preempt_enable_no_resched();
4013 EXPORT_SYMBOL(schedule
);
4015 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4017 * Look out! "owner" is an entirely speculative pointer
4018 * access and not reliable.
4020 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
4025 if (!sched_feat(OWNER_SPIN
))
4028 #ifdef CONFIG_DEBUG_PAGEALLOC
4030 * Need to access the cpu field knowing that
4031 * DEBUG_PAGEALLOC could have unmapped it if
4032 * the mutex owner just released it and exited.
4034 if (probe_kernel_address(&owner
->cpu
, cpu
))
4041 * Even if the access succeeded (likely case),
4042 * the cpu field may no longer be valid.
4044 if (cpu
>= nr_cpumask_bits
)
4048 * We need to validate that we can do a
4049 * get_cpu() and that we have the percpu area.
4051 if (!cpu_online(cpu
))
4058 * Owner changed, break to re-assess state.
4060 if (lock
->owner
!= owner
) {
4062 * If the lock has switched to a different owner,
4063 * we likely have heavy contention. Return 0 to quit
4064 * optimistic spinning and not contend further:
4072 * Is that owner really running on that cpu?
4074 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
4077 arch_mutex_cpu_relax();
4084 #ifdef CONFIG_PREEMPT
4086 * this is the entry point to schedule() from in-kernel preemption
4087 * off of preempt_enable. Kernel preemptions off return from interrupt
4088 * occur there and call schedule directly.
4090 asmlinkage
void __sched notrace
preempt_schedule(void)
4092 struct thread_info
*ti
= current_thread_info();
4095 * If there is a non-zero preempt_count or interrupts are disabled,
4096 * we do not want to preempt the current task. Just return..
4098 if (likely(ti
->preempt_count
|| irqs_disabled()))
4102 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4104 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4107 * Check again in case we missed a preemption opportunity
4108 * between schedule and now.
4111 } while (need_resched());
4113 EXPORT_SYMBOL(preempt_schedule
);
4116 * this is the entry point to schedule() from kernel preemption
4117 * off of irq context.
4118 * Note, that this is called and return with irqs disabled. This will
4119 * protect us against recursive calling from irq.
4121 asmlinkage
void __sched
preempt_schedule_irq(void)
4123 struct thread_info
*ti
= current_thread_info();
4125 /* Catch callers which need to be fixed */
4126 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4129 add_preempt_count(PREEMPT_ACTIVE
);
4132 local_irq_disable();
4133 sub_preempt_count(PREEMPT_ACTIVE
);
4136 * Check again in case we missed a preemption opportunity
4137 * between schedule and now.
4140 } while (need_resched());
4143 #endif /* CONFIG_PREEMPT */
4145 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4148 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4150 EXPORT_SYMBOL(default_wake_function
);
4153 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4154 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4155 * number) then we wake all the non-exclusive tasks and one exclusive task.
4157 * There are circumstances in which we can try to wake a task which has already
4158 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4159 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4161 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4162 int nr_exclusive
, int wake_flags
, void *key
)
4164 wait_queue_t
*curr
, *next
;
4166 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4167 unsigned flags
= curr
->flags
;
4169 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4170 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4176 * __wake_up - wake up threads blocked on a waitqueue.
4178 * @mode: which threads
4179 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4180 * @key: is directly passed to the wakeup function
4182 * It may be assumed that this function implies a write memory barrier before
4183 * changing the task state if and only if any tasks are woken up.
4185 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4186 int nr_exclusive
, void *key
)
4188 unsigned long flags
;
4190 spin_lock_irqsave(&q
->lock
, flags
);
4191 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4192 spin_unlock_irqrestore(&q
->lock
, flags
);
4194 EXPORT_SYMBOL(__wake_up
);
4197 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4199 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4201 __wake_up_common(q
, mode
, 1, 0, NULL
);
4203 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4205 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4207 __wake_up_common(q
, mode
, 1, 0, key
);
4211 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4213 * @mode: which threads
4214 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4215 * @key: opaque value to be passed to wakeup targets
4217 * The sync wakeup differs that the waker knows that it will schedule
4218 * away soon, so while the target thread will be woken up, it will not
4219 * be migrated to another CPU - ie. the two threads are 'synchronized'
4220 * with each other. This can prevent needless bouncing between CPUs.
4222 * On UP it can prevent extra preemption.
4224 * It may be assumed that this function implies a write memory barrier before
4225 * changing the task state if and only if any tasks are woken up.
4227 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4228 int nr_exclusive
, void *key
)
4230 unsigned long flags
;
4231 int wake_flags
= WF_SYNC
;
4236 if (unlikely(!nr_exclusive
))
4239 spin_lock_irqsave(&q
->lock
, flags
);
4240 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4241 spin_unlock_irqrestore(&q
->lock
, flags
);
4243 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4246 * __wake_up_sync - see __wake_up_sync_key()
4248 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4250 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4252 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4255 * complete: - signals a single thread waiting on this completion
4256 * @x: holds the state of this particular completion
4258 * This will wake up a single thread waiting on this completion. Threads will be
4259 * awakened in the same order in which they were queued.
4261 * See also complete_all(), wait_for_completion() and related routines.
4263 * It may be assumed that this function implies a write memory barrier before
4264 * changing the task state if and only if any tasks are woken up.
4266 void complete(struct completion
*x
)
4268 unsigned long flags
;
4270 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4272 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4273 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4275 EXPORT_SYMBOL(complete
);
4278 * complete_all: - signals all threads waiting on this completion
4279 * @x: holds the state of this particular completion
4281 * This will wake up all threads waiting on this particular completion event.
4283 * It may be assumed that this function implies a write memory barrier before
4284 * changing the task state if and only if any tasks are woken up.
4286 void complete_all(struct completion
*x
)
4288 unsigned long flags
;
4290 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4291 x
->done
+= UINT_MAX
/2;
4292 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4293 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4295 EXPORT_SYMBOL(complete_all
);
4297 static inline long __sched
4298 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4301 DECLARE_WAITQUEUE(wait
, current
);
4303 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4305 if (signal_pending_state(state
, current
)) {
4306 timeout
= -ERESTARTSYS
;
4309 __set_current_state(state
);
4310 spin_unlock_irq(&x
->wait
.lock
);
4311 timeout
= schedule_timeout(timeout
);
4312 spin_lock_irq(&x
->wait
.lock
);
4313 } while (!x
->done
&& timeout
);
4314 __remove_wait_queue(&x
->wait
, &wait
);
4319 return timeout
?: 1;
4323 wait_for_common(struct completion
*x
, long timeout
, int state
)
4327 spin_lock_irq(&x
->wait
.lock
);
4328 timeout
= do_wait_for_common(x
, timeout
, state
);
4329 spin_unlock_irq(&x
->wait
.lock
);
4334 * wait_for_completion: - waits for completion of a task
4335 * @x: holds the state of this particular completion
4337 * This waits to be signaled for completion of a specific task. It is NOT
4338 * interruptible and there is no timeout.
4340 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4341 * and interrupt capability. Also see complete().
4343 void __sched
wait_for_completion(struct completion
*x
)
4345 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4347 EXPORT_SYMBOL(wait_for_completion
);
4350 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4351 * @x: holds the state of this particular completion
4352 * @timeout: timeout value in jiffies
4354 * This waits for either a completion of a specific task to be signaled or for a
4355 * specified timeout to expire. The timeout is in jiffies. It is not
4358 unsigned long __sched
4359 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4361 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4363 EXPORT_SYMBOL(wait_for_completion_timeout
);
4366 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4367 * @x: holds the state of this particular completion
4369 * This waits for completion of a specific task to be signaled. It is
4372 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4374 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4375 if (t
== -ERESTARTSYS
)
4379 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4382 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4383 * @x: holds the state of this particular completion
4384 * @timeout: timeout value in jiffies
4386 * This waits for either a completion of a specific task to be signaled or for a
4387 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4390 wait_for_completion_interruptible_timeout(struct completion
*x
,
4391 unsigned long timeout
)
4393 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4395 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4398 * wait_for_completion_killable: - waits for completion of a task (killable)
4399 * @x: holds the state of this particular completion
4401 * This waits to be signaled for completion of a specific task. It can be
4402 * interrupted by a kill signal.
4404 int __sched
wait_for_completion_killable(struct completion
*x
)
4406 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4407 if (t
== -ERESTARTSYS
)
4411 EXPORT_SYMBOL(wait_for_completion_killable
);
4414 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4415 * @x: holds the state of this particular completion
4416 * @timeout: timeout value in jiffies
4418 * This waits for either a completion of a specific task to be
4419 * signaled or for a specified timeout to expire. It can be
4420 * interrupted by a kill signal. The timeout is in jiffies.
4423 wait_for_completion_killable_timeout(struct completion
*x
,
4424 unsigned long timeout
)
4426 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4428 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4431 * try_wait_for_completion - try to decrement a completion without blocking
4432 * @x: completion structure
4434 * Returns: 0 if a decrement cannot be done without blocking
4435 * 1 if a decrement succeeded.
4437 * If a completion is being used as a counting completion,
4438 * attempt to decrement the counter without blocking. This
4439 * enables us to avoid waiting if the resource the completion
4440 * is protecting is not available.
4442 bool try_wait_for_completion(struct completion
*x
)
4444 unsigned long flags
;
4447 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4452 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4455 EXPORT_SYMBOL(try_wait_for_completion
);
4458 * completion_done - Test to see if a completion has any waiters
4459 * @x: completion structure
4461 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4462 * 1 if there are no waiters.
4465 bool completion_done(struct completion
*x
)
4467 unsigned long flags
;
4470 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4473 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4476 EXPORT_SYMBOL(completion_done
);
4479 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4481 unsigned long flags
;
4484 init_waitqueue_entry(&wait
, current
);
4486 __set_current_state(state
);
4488 spin_lock_irqsave(&q
->lock
, flags
);
4489 __add_wait_queue(q
, &wait
);
4490 spin_unlock(&q
->lock
);
4491 timeout
= schedule_timeout(timeout
);
4492 spin_lock_irq(&q
->lock
);
4493 __remove_wait_queue(q
, &wait
);
4494 spin_unlock_irqrestore(&q
->lock
, flags
);
4499 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4501 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4503 EXPORT_SYMBOL(interruptible_sleep_on
);
4506 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4508 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4510 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4512 void __sched
sleep_on(wait_queue_head_t
*q
)
4514 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4516 EXPORT_SYMBOL(sleep_on
);
4518 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4520 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4522 EXPORT_SYMBOL(sleep_on_timeout
);
4524 #ifdef CONFIG_RT_MUTEXES
4527 * rt_mutex_setprio - set the current priority of a task
4529 * @prio: prio value (kernel-internal form)
4531 * This function changes the 'effective' priority of a task. It does
4532 * not touch ->normal_prio like __setscheduler().
4534 * Used by the rt_mutex code to implement priority inheritance logic.
4536 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4538 unsigned long flags
;
4539 int oldprio
, on_rq
, running
;
4541 const struct sched_class
*prev_class
;
4543 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4545 rq
= task_rq_lock(p
, &flags
);
4547 trace_sched_pi_setprio(p
, prio
);
4549 prev_class
= p
->sched_class
;
4550 on_rq
= p
->se
.on_rq
;
4551 running
= task_current(rq
, p
);
4553 dequeue_task(rq
, p
, 0);
4555 p
->sched_class
->put_prev_task(rq
, p
);
4558 p
->sched_class
= &rt_sched_class
;
4560 p
->sched_class
= &fair_sched_class
;
4565 p
->sched_class
->set_curr_task(rq
);
4567 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4569 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4571 task_rq_unlock(rq
, &flags
);
4576 void set_user_nice(struct task_struct
*p
, long nice
)
4578 int old_prio
, delta
, on_rq
;
4579 unsigned long flags
;
4582 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4585 * We have to be careful, if called from sys_setpriority(),
4586 * the task might be in the middle of scheduling on another CPU.
4588 rq
= task_rq_lock(p
, &flags
);
4590 * The RT priorities are set via sched_setscheduler(), but we still
4591 * allow the 'normal' nice value to be set - but as expected
4592 * it wont have any effect on scheduling until the task is
4593 * SCHED_FIFO/SCHED_RR:
4595 if (task_has_rt_policy(p
)) {
4596 p
->static_prio
= NICE_TO_PRIO(nice
);
4599 on_rq
= p
->se
.on_rq
;
4601 dequeue_task(rq
, p
, 0);
4603 p
->static_prio
= NICE_TO_PRIO(nice
);
4606 p
->prio
= effective_prio(p
);
4607 delta
= p
->prio
- old_prio
;
4610 enqueue_task(rq
, p
, 0);
4612 * If the task increased its priority or is running and
4613 * lowered its priority, then reschedule its CPU:
4615 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4616 resched_task(rq
->curr
);
4619 task_rq_unlock(rq
, &flags
);
4621 EXPORT_SYMBOL(set_user_nice
);
4624 * can_nice - check if a task can reduce its nice value
4628 int can_nice(const struct task_struct
*p
, const int nice
)
4630 /* convert nice value [19,-20] to rlimit style value [1,40] */
4631 int nice_rlim
= 20 - nice
;
4633 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4634 capable(CAP_SYS_NICE
));
4637 #ifdef __ARCH_WANT_SYS_NICE
4640 * sys_nice - change the priority of the current process.
4641 * @increment: priority increment
4643 * sys_setpriority is a more generic, but much slower function that
4644 * does similar things.
4646 SYSCALL_DEFINE1(nice
, int, increment
)
4651 * Setpriority might change our priority at the same moment.
4652 * We don't have to worry. Conceptually one call occurs first
4653 * and we have a single winner.
4655 if (increment
< -40)
4660 nice
= TASK_NICE(current
) + increment
;
4666 if (increment
< 0 && !can_nice(current
, nice
))
4669 retval
= security_task_setnice(current
, nice
);
4673 set_user_nice(current
, nice
);
4680 * task_prio - return the priority value of a given task.
4681 * @p: the task in question.
4683 * This is the priority value as seen by users in /proc.
4684 * RT tasks are offset by -200. Normal tasks are centered
4685 * around 0, value goes from -16 to +15.
4687 int task_prio(const struct task_struct
*p
)
4689 return p
->prio
- MAX_RT_PRIO
;
4693 * task_nice - return the nice value of a given task.
4694 * @p: the task in question.
4696 int task_nice(const struct task_struct
*p
)
4698 return TASK_NICE(p
);
4700 EXPORT_SYMBOL(task_nice
);
4703 * idle_cpu - is a given cpu idle currently?
4704 * @cpu: the processor in question.
4706 int idle_cpu(int cpu
)
4708 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4712 * idle_task - return the idle task for a given cpu.
4713 * @cpu: the processor in question.
4715 struct task_struct
*idle_task(int cpu
)
4717 return cpu_rq(cpu
)->idle
;
4721 * find_process_by_pid - find a process with a matching PID value.
4722 * @pid: the pid in question.
4724 static struct task_struct
*find_process_by_pid(pid_t pid
)
4726 return pid
? find_task_by_vpid(pid
) : current
;
4729 /* Actually do priority change: must hold rq lock. */
4731 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4733 BUG_ON(p
->se
.on_rq
);
4736 p
->rt_priority
= prio
;
4737 p
->normal_prio
= normal_prio(p
);
4738 /* we are holding p->pi_lock already */
4739 p
->prio
= rt_mutex_getprio(p
);
4740 if (rt_prio(p
->prio
))
4741 p
->sched_class
= &rt_sched_class
;
4743 p
->sched_class
= &fair_sched_class
;
4748 * check the target process has a UID that matches the current process's
4750 static bool check_same_owner(struct task_struct
*p
)
4752 const struct cred
*cred
= current_cred(), *pcred
;
4756 pcred
= __task_cred(p
);
4757 match
= (cred
->euid
== pcred
->euid
||
4758 cred
->euid
== pcred
->uid
);
4763 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4764 const struct sched_param
*param
, bool user
)
4766 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4767 unsigned long flags
;
4768 const struct sched_class
*prev_class
;
4772 /* may grab non-irq protected spin_locks */
4773 BUG_ON(in_interrupt());
4775 /* double check policy once rq lock held */
4777 reset_on_fork
= p
->sched_reset_on_fork
;
4778 policy
= oldpolicy
= p
->policy
;
4780 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4781 policy
&= ~SCHED_RESET_ON_FORK
;
4783 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4784 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4785 policy
!= SCHED_IDLE
)
4790 * Valid priorities for SCHED_FIFO and SCHED_RR are
4791 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4792 * SCHED_BATCH and SCHED_IDLE is 0.
4794 if (param
->sched_priority
< 0 ||
4795 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4796 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4798 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4802 * Allow unprivileged RT tasks to decrease priority:
4804 if (user
&& !capable(CAP_SYS_NICE
)) {
4805 if (rt_policy(policy
)) {
4806 unsigned long rlim_rtprio
=
4807 task_rlimit(p
, RLIMIT_RTPRIO
);
4809 /* can't set/change the rt policy */
4810 if (policy
!= p
->policy
&& !rlim_rtprio
)
4813 /* can't increase priority */
4814 if (param
->sched_priority
> p
->rt_priority
&&
4815 param
->sched_priority
> rlim_rtprio
)
4819 * Like positive nice levels, dont allow tasks to
4820 * move out of SCHED_IDLE either:
4822 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4825 /* can't change other user's priorities */
4826 if (!check_same_owner(p
))
4829 /* Normal users shall not reset the sched_reset_on_fork flag */
4830 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4835 retval
= security_task_setscheduler(p
);
4841 * make sure no PI-waiters arrive (or leave) while we are
4842 * changing the priority of the task:
4844 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4846 * To be able to change p->policy safely, the apropriate
4847 * runqueue lock must be held.
4849 rq
= __task_rq_lock(p
);
4852 * Changing the policy of the stop threads its a very bad idea
4854 if (p
== rq
->stop
) {
4855 __task_rq_unlock(rq
);
4856 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4860 #ifdef CONFIG_RT_GROUP_SCHED
4863 * Do not allow realtime tasks into groups that have no runtime
4866 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4867 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
4868 !task_group_is_autogroup(task_group(p
))) {
4869 __task_rq_unlock(rq
);
4870 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4876 /* recheck policy now with rq lock held */
4877 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4878 policy
= oldpolicy
= -1;
4879 __task_rq_unlock(rq
);
4880 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4883 on_rq
= p
->se
.on_rq
;
4884 running
= task_current(rq
, p
);
4886 deactivate_task(rq
, p
, 0);
4888 p
->sched_class
->put_prev_task(rq
, p
);
4890 p
->sched_reset_on_fork
= reset_on_fork
;
4893 prev_class
= p
->sched_class
;
4894 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4897 p
->sched_class
->set_curr_task(rq
);
4899 activate_task(rq
, p
, 0);
4901 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4903 __task_rq_unlock(rq
);
4904 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4906 rt_mutex_adjust_pi(p
);
4912 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4913 * @p: the task in question.
4914 * @policy: new policy.
4915 * @param: structure containing the new RT priority.
4917 * NOTE that the task may be already dead.
4919 int sched_setscheduler(struct task_struct
*p
, int policy
,
4920 const struct sched_param
*param
)
4922 return __sched_setscheduler(p
, policy
, param
, true);
4924 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4927 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4928 * @p: the task in question.
4929 * @policy: new policy.
4930 * @param: structure containing the new RT priority.
4932 * Just like sched_setscheduler, only don't bother checking if the
4933 * current context has permission. For example, this is needed in
4934 * stop_machine(): we create temporary high priority worker threads,
4935 * but our caller might not have that capability.
4937 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4938 const struct sched_param
*param
)
4940 return __sched_setscheduler(p
, policy
, param
, false);
4944 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4946 struct sched_param lparam
;
4947 struct task_struct
*p
;
4950 if (!param
|| pid
< 0)
4952 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4957 p
= find_process_by_pid(pid
);
4959 retval
= sched_setscheduler(p
, policy
, &lparam
);
4966 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4967 * @pid: the pid in question.
4968 * @policy: new policy.
4969 * @param: structure containing the new RT priority.
4971 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4972 struct sched_param __user
*, param
)
4974 /* negative values for policy are not valid */
4978 return do_sched_setscheduler(pid
, policy
, param
);
4982 * sys_sched_setparam - set/change the RT priority of a thread
4983 * @pid: the pid in question.
4984 * @param: structure containing the new RT priority.
4986 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4988 return do_sched_setscheduler(pid
, -1, param
);
4992 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4993 * @pid: the pid in question.
4995 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4997 struct task_struct
*p
;
5005 p
= find_process_by_pid(pid
);
5007 retval
= security_task_getscheduler(p
);
5010 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5017 * sys_sched_getparam - get the RT priority of a thread
5018 * @pid: the pid in question.
5019 * @param: structure containing the RT priority.
5021 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5023 struct sched_param lp
;
5024 struct task_struct
*p
;
5027 if (!param
|| pid
< 0)
5031 p
= find_process_by_pid(pid
);
5036 retval
= security_task_getscheduler(p
);
5040 lp
.sched_priority
= p
->rt_priority
;
5044 * This one might sleep, we cannot do it with a spinlock held ...
5046 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5055 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5057 cpumask_var_t cpus_allowed
, new_mask
;
5058 struct task_struct
*p
;
5064 p
= find_process_by_pid(pid
);
5071 /* Prevent p going away */
5075 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5079 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5081 goto out_free_cpus_allowed
;
5084 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5087 retval
= security_task_setscheduler(p
);
5091 cpuset_cpus_allowed(p
, cpus_allowed
);
5092 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5094 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5097 cpuset_cpus_allowed(p
, cpus_allowed
);
5098 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5100 * We must have raced with a concurrent cpuset
5101 * update. Just reset the cpus_allowed to the
5102 * cpuset's cpus_allowed
5104 cpumask_copy(new_mask
, cpus_allowed
);
5109 free_cpumask_var(new_mask
);
5110 out_free_cpus_allowed
:
5111 free_cpumask_var(cpus_allowed
);
5118 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5119 struct cpumask
*new_mask
)
5121 if (len
< cpumask_size())
5122 cpumask_clear(new_mask
);
5123 else if (len
> cpumask_size())
5124 len
= cpumask_size();
5126 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5130 * sys_sched_setaffinity - set the cpu affinity of a process
5131 * @pid: pid of the process
5132 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5133 * @user_mask_ptr: user-space pointer to the new cpu mask
5135 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5136 unsigned long __user
*, user_mask_ptr
)
5138 cpumask_var_t new_mask
;
5141 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5144 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5146 retval
= sched_setaffinity(pid
, new_mask
);
5147 free_cpumask_var(new_mask
);
5151 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5153 struct task_struct
*p
;
5154 unsigned long flags
;
5162 p
= find_process_by_pid(pid
);
5166 retval
= security_task_getscheduler(p
);
5170 rq
= task_rq_lock(p
, &flags
);
5171 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5172 task_rq_unlock(rq
, &flags
);
5182 * sys_sched_getaffinity - get the cpu affinity of a process
5183 * @pid: pid of the process
5184 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5185 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5187 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5188 unsigned long __user
*, user_mask_ptr
)
5193 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5195 if (len
& (sizeof(unsigned long)-1))
5198 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5201 ret
= sched_getaffinity(pid
, mask
);
5203 size_t retlen
= min_t(size_t, len
, cpumask_size());
5205 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5210 free_cpumask_var(mask
);
5216 * sys_sched_yield - yield the current processor to other threads.
5218 * This function yields the current CPU to other tasks. If there are no
5219 * other threads running on this CPU then this function will return.
5221 SYSCALL_DEFINE0(sched_yield
)
5223 struct rq
*rq
= this_rq_lock();
5225 schedstat_inc(rq
, yld_count
);
5226 current
->sched_class
->yield_task(rq
);
5229 * Since we are going to call schedule() anyway, there's
5230 * no need to preempt or enable interrupts:
5232 __release(rq
->lock
);
5233 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5234 do_raw_spin_unlock(&rq
->lock
);
5235 preempt_enable_no_resched();
5242 static inline int should_resched(void)
5244 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5247 static void __cond_resched(void)
5249 add_preempt_count(PREEMPT_ACTIVE
);
5251 sub_preempt_count(PREEMPT_ACTIVE
);
5254 int __sched
_cond_resched(void)
5256 if (should_resched()) {
5262 EXPORT_SYMBOL(_cond_resched
);
5265 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5266 * call schedule, and on return reacquire the lock.
5268 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5269 * operations here to prevent schedule() from being called twice (once via
5270 * spin_unlock(), once by hand).
5272 int __cond_resched_lock(spinlock_t
*lock
)
5274 int resched
= should_resched();
5277 lockdep_assert_held(lock
);
5279 if (spin_needbreak(lock
) || resched
) {
5290 EXPORT_SYMBOL(__cond_resched_lock
);
5292 int __sched
__cond_resched_softirq(void)
5294 BUG_ON(!in_softirq());
5296 if (should_resched()) {
5304 EXPORT_SYMBOL(__cond_resched_softirq
);
5307 * yield - yield the current processor to other threads.
5309 * This is a shortcut for kernel-space yielding - it marks the
5310 * thread runnable and calls sys_sched_yield().
5312 void __sched
yield(void)
5314 set_current_state(TASK_RUNNING
);
5317 EXPORT_SYMBOL(yield
);
5320 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5321 * that process accounting knows that this is a task in IO wait state.
5323 void __sched
io_schedule(void)
5325 struct rq
*rq
= raw_rq();
5327 delayacct_blkio_start();
5328 atomic_inc(&rq
->nr_iowait
);
5329 current
->in_iowait
= 1;
5331 current
->in_iowait
= 0;
5332 atomic_dec(&rq
->nr_iowait
);
5333 delayacct_blkio_end();
5335 EXPORT_SYMBOL(io_schedule
);
5337 long __sched
io_schedule_timeout(long timeout
)
5339 struct rq
*rq
= raw_rq();
5342 delayacct_blkio_start();
5343 atomic_inc(&rq
->nr_iowait
);
5344 current
->in_iowait
= 1;
5345 ret
= schedule_timeout(timeout
);
5346 current
->in_iowait
= 0;
5347 atomic_dec(&rq
->nr_iowait
);
5348 delayacct_blkio_end();
5353 * sys_sched_get_priority_max - return maximum RT priority.
5354 * @policy: scheduling class.
5356 * this syscall returns the maximum rt_priority that can be used
5357 * by a given scheduling class.
5359 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5366 ret
= MAX_USER_RT_PRIO
-1;
5378 * sys_sched_get_priority_min - return minimum RT priority.
5379 * @policy: scheduling class.
5381 * this syscall returns the minimum rt_priority that can be used
5382 * by a given scheduling class.
5384 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5402 * sys_sched_rr_get_interval - return the default timeslice of a process.
5403 * @pid: pid of the process.
5404 * @interval: userspace pointer to the timeslice value.
5406 * this syscall writes the default timeslice value of a given process
5407 * into the user-space timespec buffer. A value of '0' means infinity.
5409 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5410 struct timespec __user
*, interval
)
5412 struct task_struct
*p
;
5413 unsigned int time_slice
;
5414 unsigned long flags
;
5424 p
= find_process_by_pid(pid
);
5428 retval
= security_task_getscheduler(p
);
5432 rq
= task_rq_lock(p
, &flags
);
5433 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5434 task_rq_unlock(rq
, &flags
);
5437 jiffies_to_timespec(time_slice
, &t
);
5438 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5446 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5448 void sched_show_task(struct task_struct
*p
)
5450 unsigned long free
= 0;
5453 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5454 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5455 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5456 #if BITS_PER_LONG == 32
5457 if (state
== TASK_RUNNING
)
5458 printk(KERN_CONT
" running ");
5460 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5462 if (state
== TASK_RUNNING
)
5463 printk(KERN_CONT
" running task ");
5465 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5467 #ifdef CONFIG_DEBUG_STACK_USAGE
5468 free
= stack_not_used(p
);
5470 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5471 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5472 (unsigned long)task_thread_info(p
)->flags
);
5474 show_stack(p
, NULL
);
5477 void show_state_filter(unsigned long state_filter
)
5479 struct task_struct
*g
, *p
;
5481 #if BITS_PER_LONG == 32
5483 " task PC stack pid father\n");
5486 " task PC stack pid father\n");
5488 read_lock(&tasklist_lock
);
5489 do_each_thread(g
, p
) {
5491 * reset the NMI-timeout, listing all files on a slow
5492 * console might take alot of time:
5494 touch_nmi_watchdog();
5495 if (!state_filter
|| (p
->state
& state_filter
))
5497 } while_each_thread(g
, p
);
5499 touch_all_softlockup_watchdogs();
5501 #ifdef CONFIG_SCHED_DEBUG
5502 sysrq_sched_debug_show();
5504 read_unlock(&tasklist_lock
);
5506 * Only show locks if all tasks are dumped:
5509 debug_show_all_locks();
5512 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5514 idle
->sched_class
= &idle_sched_class
;
5518 * init_idle - set up an idle thread for a given CPU
5519 * @idle: task in question
5520 * @cpu: cpu the idle task belongs to
5522 * NOTE: this function does not set the idle thread's NEED_RESCHED
5523 * flag, to make booting more robust.
5525 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5527 struct rq
*rq
= cpu_rq(cpu
);
5528 unsigned long flags
;
5530 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5533 idle
->state
= TASK_RUNNING
;
5534 idle
->se
.exec_start
= sched_clock();
5536 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5538 * We're having a chicken and egg problem, even though we are
5539 * holding rq->lock, the cpu isn't yet set to this cpu so the
5540 * lockdep check in task_group() will fail.
5542 * Similar case to sched_fork(). / Alternatively we could
5543 * use task_rq_lock() here and obtain the other rq->lock.
5548 __set_task_cpu(idle
, cpu
);
5551 rq
->curr
= rq
->idle
= idle
;
5552 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5555 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5557 /* Set the preempt count _outside_ the spinlocks! */
5558 #if defined(CONFIG_PREEMPT)
5559 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5561 task_thread_info(idle
)->preempt_count
= 0;
5564 * The idle tasks have their own, simple scheduling class:
5566 idle
->sched_class
= &idle_sched_class
;
5567 ftrace_graph_init_task(idle
);
5571 * In a system that switches off the HZ timer nohz_cpu_mask
5572 * indicates which cpus entered this state. This is used
5573 * in the rcu update to wait only for active cpus. For system
5574 * which do not switch off the HZ timer nohz_cpu_mask should
5575 * always be CPU_BITS_NONE.
5577 cpumask_var_t nohz_cpu_mask
;
5580 * Increase the granularity value when there are more CPUs,
5581 * because with more CPUs the 'effective latency' as visible
5582 * to users decreases. But the relationship is not linear,
5583 * so pick a second-best guess by going with the log2 of the
5586 * This idea comes from the SD scheduler of Con Kolivas:
5588 static int get_update_sysctl_factor(void)
5590 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5591 unsigned int factor
;
5593 switch (sysctl_sched_tunable_scaling
) {
5594 case SCHED_TUNABLESCALING_NONE
:
5597 case SCHED_TUNABLESCALING_LINEAR
:
5600 case SCHED_TUNABLESCALING_LOG
:
5602 factor
= 1 + ilog2(cpus
);
5609 static void update_sysctl(void)
5611 unsigned int factor
= get_update_sysctl_factor();
5613 #define SET_SYSCTL(name) \
5614 (sysctl_##name = (factor) * normalized_sysctl_##name)
5615 SET_SYSCTL(sched_min_granularity
);
5616 SET_SYSCTL(sched_latency
);
5617 SET_SYSCTL(sched_wakeup_granularity
);
5621 static inline void sched_init_granularity(void)
5628 * This is how migration works:
5630 * 1) we invoke migration_cpu_stop() on the target CPU using
5632 * 2) stopper starts to run (implicitly forcing the migrated thread
5634 * 3) it checks whether the migrated task is still in the wrong runqueue.
5635 * 4) if it's in the wrong runqueue then the migration thread removes
5636 * it and puts it into the right queue.
5637 * 5) stopper completes and stop_one_cpu() returns and the migration
5642 * Change a given task's CPU affinity. Migrate the thread to a
5643 * proper CPU and schedule it away if the CPU it's executing on
5644 * is removed from the allowed bitmask.
5646 * NOTE: the caller must have a valid reference to the task, the
5647 * task must not exit() & deallocate itself prematurely. The
5648 * call is not atomic; no spinlocks may be held.
5650 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5652 unsigned long flags
;
5654 unsigned int dest_cpu
;
5658 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5659 * drop the rq->lock and still rely on ->cpus_allowed.
5662 while (task_is_waking(p
))
5664 rq
= task_rq_lock(p
, &flags
);
5665 if (task_is_waking(p
)) {
5666 task_rq_unlock(rq
, &flags
);
5670 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5675 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5676 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5681 if (p
->sched_class
->set_cpus_allowed
)
5682 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5684 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5685 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5688 /* Can the task run on the task's current CPU? If so, we're done */
5689 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5692 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5693 if (migrate_task(p
, rq
)) {
5694 struct migration_arg arg
= { p
, dest_cpu
};
5695 /* Need help from migration thread: drop lock and wait. */
5696 task_rq_unlock(rq
, &flags
);
5697 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5698 tlb_migrate_finish(p
->mm
);
5702 task_rq_unlock(rq
, &flags
);
5706 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5709 * Move (not current) task off this cpu, onto dest cpu. We're doing
5710 * this because either it can't run here any more (set_cpus_allowed()
5711 * away from this CPU, or CPU going down), or because we're
5712 * attempting to rebalance this task on exec (sched_exec).
5714 * So we race with normal scheduler movements, but that's OK, as long
5715 * as the task is no longer on this CPU.
5717 * Returns non-zero if task was successfully migrated.
5719 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5721 struct rq
*rq_dest
, *rq_src
;
5724 if (unlikely(!cpu_active(dest_cpu
)))
5727 rq_src
= cpu_rq(src_cpu
);
5728 rq_dest
= cpu_rq(dest_cpu
);
5730 double_rq_lock(rq_src
, rq_dest
);
5731 /* Already moved. */
5732 if (task_cpu(p
) != src_cpu
)
5734 /* Affinity changed (again). */
5735 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5739 * If we're not on a rq, the next wake-up will ensure we're
5743 deactivate_task(rq_src
, p
, 0);
5744 set_task_cpu(p
, dest_cpu
);
5745 activate_task(rq_dest
, p
, 0);
5746 check_preempt_curr(rq_dest
, p
, 0);
5751 double_rq_unlock(rq_src
, rq_dest
);
5756 * migration_cpu_stop - this will be executed by a highprio stopper thread
5757 * and performs thread migration by bumping thread off CPU then
5758 * 'pushing' onto another runqueue.
5760 static int migration_cpu_stop(void *data
)
5762 struct migration_arg
*arg
= data
;
5765 * The original target cpu might have gone down and we might
5766 * be on another cpu but it doesn't matter.
5768 local_irq_disable();
5769 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5774 #ifdef CONFIG_HOTPLUG_CPU
5777 * Ensures that the idle task is using init_mm right before its cpu goes
5780 void idle_task_exit(void)
5782 struct mm_struct
*mm
= current
->active_mm
;
5784 BUG_ON(cpu_online(smp_processor_id()));
5787 switch_mm(mm
, &init_mm
, current
);
5792 * While a dead CPU has no uninterruptible tasks queued at this point,
5793 * it might still have a nonzero ->nr_uninterruptible counter, because
5794 * for performance reasons the counter is not stricly tracking tasks to
5795 * their home CPUs. So we just add the counter to another CPU's counter,
5796 * to keep the global sum constant after CPU-down:
5798 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5800 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5802 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5803 rq_src
->nr_uninterruptible
= 0;
5807 * remove the tasks which were accounted by rq from calc_load_tasks.
5809 static void calc_global_load_remove(struct rq
*rq
)
5811 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5812 rq
->calc_load_active
= 0;
5816 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5817 * try_to_wake_up()->select_task_rq().
5819 * Called with rq->lock held even though we'er in stop_machine() and
5820 * there's no concurrency possible, we hold the required locks anyway
5821 * because of lock validation efforts.
5823 static void migrate_tasks(unsigned int dead_cpu
)
5825 struct rq
*rq
= cpu_rq(dead_cpu
);
5826 struct task_struct
*next
, *stop
= rq
->stop
;
5830 * Fudge the rq selection such that the below task selection loop
5831 * doesn't get stuck on the currently eligible stop task.
5833 * We're currently inside stop_machine() and the rq is either stuck
5834 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5835 * either way we should never end up calling schedule() until we're
5842 * There's this thread running, bail when that's the only
5845 if (rq
->nr_running
== 1)
5848 next
= pick_next_task(rq
);
5850 next
->sched_class
->put_prev_task(rq
, next
);
5852 /* Find suitable destination for @next, with force if needed. */
5853 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5854 raw_spin_unlock(&rq
->lock
);
5856 __migrate_task(next
, dead_cpu
, dest_cpu
);
5858 raw_spin_lock(&rq
->lock
);
5864 #endif /* CONFIG_HOTPLUG_CPU */
5866 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5868 static struct ctl_table sd_ctl_dir
[] = {
5870 .procname
= "sched_domain",
5876 static struct ctl_table sd_ctl_root
[] = {
5878 .procname
= "kernel",
5880 .child
= sd_ctl_dir
,
5885 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5887 struct ctl_table
*entry
=
5888 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5893 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5895 struct ctl_table
*entry
;
5898 * In the intermediate directories, both the child directory and
5899 * procname are dynamically allocated and could fail but the mode
5900 * will always be set. In the lowest directory the names are
5901 * static strings and all have proc handlers.
5903 for (entry
= *tablep
; entry
->mode
; entry
++) {
5905 sd_free_ctl_entry(&entry
->child
);
5906 if (entry
->proc_handler
== NULL
)
5907 kfree(entry
->procname
);
5915 set_table_entry(struct ctl_table
*entry
,
5916 const char *procname
, void *data
, int maxlen
,
5917 mode_t mode
, proc_handler
*proc_handler
)
5919 entry
->procname
= procname
;
5921 entry
->maxlen
= maxlen
;
5923 entry
->proc_handler
= proc_handler
;
5926 static struct ctl_table
*
5927 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5929 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5934 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5935 sizeof(long), 0644, proc_doulongvec_minmax
);
5936 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5937 sizeof(long), 0644, proc_doulongvec_minmax
);
5938 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5939 sizeof(int), 0644, proc_dointvec_minmax
);
5940 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5941 sizeof(int), 0644, proc_dointvec_minmax
);
5942 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5943 sizeof(int), 0644, proc_dointvec_minmax
);
5944 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5945 sizeof(int), 0644, proc_dointvec_minmax
);
5946 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5947 sizeof(int), 0644, proc_dointvec_minmax
);
5948 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5949 sizeof(int), 0644, proc_dointvec_minmax
);
5950 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5951 sizeof(int), 0644, proc_dointvec_minmax
);
5952 set_table_entry(&table
[9], "cache_nice_tries",
5953 &sd
->cache_nice_tries
,
5954 sizeof(int), 0644, proc_dointvec_minmax
);
5955 set_table_entry(&table
[10], "flags", &sd
->flags
,
5956 sizeof(int), 0644, proc_dointvec_minmax
);
5957 set_table_entry(&table
[11], "name", sd
->name
,
5958 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5959 /* &table[12] is terminator */
5964 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5966 struct ctl_table
*entry
, *table
;
5967 struct sched_domain
*sd
;
5968 int domain_num
= 0, i
;
5971 for_each_domain(cpu
, sd
)
5973 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5978 for_each_domain(cpu
, sd
) {
5979 snprintf(buf
, 32, "domain%d", i
);
5980 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5982 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5989 static struct ctl_table_header
*sd_sysctl_header
;
5990 static void register_sched_domain_sysctl(void)
5992 int i
, cpu_num
= num_possible_cpus();
5993 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5996 WARN_ON(sd_ctl_dir
[0].child
);
5997 sd_ctl_dir
[0].child
= entry
;
6002 for_each_possible_cpu(i
) {
6003 snprintf(buf
, 32, "cpu%d", i
);
6004 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6006 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6010 WARN_ON(sd_sysctl_header
);
6011 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6014 /* may be called multiple times per register */
6015 static void unregister_sched_domain_sysctl(void)
6017 if (sd_sysctl_header
)
6018 unregister_sysctl_table(sd_sysctl_header
);
6019 sd_sysctl_header
= NULL
;
6020 if (sd_ctl_dir
[0].child
)
6021 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6024 static void register_sched_domain_sysctl(void)
6027 static void unregister_sched_domain_sysctl(void)
6032 static void set_rq_online(struct rq
*rq
)
6035 const struct sched_class
*class;
6037 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6040 for_each_class(class) {
6041 if (class->rq_online
)
6042 class->rq_online(rq
);
6047 static void set_rq_offline(struct rq
*rq
)
6050 const struct sched_class
*class;
6052 for_each_class(class) {
6053 if (class->rq_offline
)
6054 class->rq_offline(rq
);
6057 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6063 * migration_call - callback that gets triggered when a CPU is added.
6064 * Here we can start up the necessary migration thread for the new CPU.
6066 static int __cpuinit
6067 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6069 int cpu
= (long)hcpu
;
6070 unsigned long flags
;
6071 struct rq
*rq
= cpu_rq(cpu
);
6073 switch (action
& ~CPU_TASKS_FROZEN
) {
6075 case CPU_UP_PREPARE
:
6076 rq
->calc_load_update
= calc_load_update
;
6080 /* Update our root-domain */
6081 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6083 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6087 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6090 #ifdef CONFIG_HOTPLUG_CPU
6092 /* Update our root-domain */
6093 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6095 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6099 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6100 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6102 migrate_nr_uninterruptible(rq
);
6103 calc_global_load_remove(rq
);
6111 * Register at high priority so that task migration (migrate_all_tasks)
6112 * happens before everything else. This has to be lower priority than
6113 * the notifier in the perf_event subsystem, though.
6115 static struct notifier_block __cpuinitdata migration_notifier
= {
6116 .notifier_call
= migration_call
,
6117 .priority
= CPU_PRI_MIGRATION
,
6120 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6121 unsigned long action
, void *hcpu
)
6123 switch (action
& ~CPU_TASKS_FROZEN
) {
6125 case CPU_DOWN_FAILED
:
6126 set_cpu_active((long)hcpu
, true);
6133 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6134 unsigned long action
, void *hcpu
)
6136 switch (action
& ~CPU_TASKS_FROZEN
) {
6137 case CPU_DOWN_PREPARE
:
6138 set_cpu_active((long)hcpu
, false);
6145 static int __init
migration_init(void)
6147 void *cpu
= (void *)(long)smp_processor_id();
6150 /* Initialize migration for the boot CPU */
6151 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6152 BUG_ON(err
== NOTIFY_BAD
);
6153 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6154 register_cpu_notifier(&migration_notifier
);
6156 /* Register cpu active notifiers */
6157 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6158 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6162 early_initcall(migration_init
);
6167 #ifdef CONFIG_SCHED_DEBUG
6169 static __read_mostly
int sched_domain_debug_enabled
;
6171 static int __init
sched_domain_debug_setup(char *str
)
6173 sched_domain_debug_enabled
= 1;
6177 early_param("sched_debug", sched_domain_debug_setup
);
6179 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6180 struct cpumask
*groupmask
)
6182 struct sched_group
*group
= sd
->groups
;
6185 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6186 cpumask_clear(groupmask
);
6188 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6190 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6191 printk("does not load-balance\n");
6193 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6198 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6200 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6201 printk(KERN_ERR
"ERROR: domain->span does not contain "
6204 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6205 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6209 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6213 printk(KERN_ERR
"ERROR: group is NULL\n");
6217 if (!group
->cpu_power
) {
6218 printk(KERN_CONT
"\n");
6219 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6224 if (!cpumask_weight(sched_group_cpus(group
))) {
6225 printk(KERN_CONT
"\n");
6226 printk(KERN_ERR
"ERROR: empty group\n");
6230 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6231 printk(KERN_CONT
"\n");
6232 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6236 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6238 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6240 printk(KERN_CONT
" %s", str
);
6241 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6242 printk(KERN_CONT
" (cpu_power = %d)",
6246 group
= group
->next
;
6247 } while (group
!= sd
->groups
);
6248 printk(KERN_CONT
"\n");
6250 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6251 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6254 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6255 printk(KERN_ERR
"ERROR: parent span is not a superset "
6256 "of domain->span\n");
6260 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6262 cpumask_var_t groupmask
;
6265 if (!sched_domain_debug_enabled
)
6269 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6273 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6275 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6276 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6281 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6288 free_cpumask_var(groupmask
);
6290 #else /* !CONFIG_SCHED_DEBUG */
6291 # define sched_domain_debug(sd, cpu) do { } while (0)
6292 #endif /* CONFIG_SCHED_DEBUG */
6294 static int sd_degenerate(struct sched_domain
*sd
)
6296 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6299 /* Following flags need at least 2 groups */
6300 if (sd
->flags
& (SD_LOAD_BALANCE
|
6301 SD_BALANCE_NEWIDLE
|
6305 SD_SHARE_PKG_RESOURCES
)) {
6306 if (sd
->groups
!= sd
->groups
->next
)
6310 /* Following flags don't use groups */
6311 if (sd
->flags
& (SD_WAKE_AFFINE
))
6318 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6320 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6322 if (sd_degenerate(parent
))
6325 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6328 /* Flags needing groups don't count if only 1 group in parent */
6329 if (parent
->groups
== parent
->groups
->next
) {
6330 pflags
&= ~(SD_LOAD_BALANCE
|
6331 SD_BALANCE_NEWIDLE
|
6335 SD_SHARE_PKG_RESOURCES
);
6336 if (nr_node_ids
== 1)
6337 pflags
&= ~SD_SERIALIZE
;
6339 if (~cflags
& pflags
)
6345 static void free_rootdomain(struct root_domain
*rd
)
6347 synchronize_sched();
6349 cpupri_cleanup(&rd
->cpupri
);
6351 free_cpumask_var(rd
->rto_mask
);
6352 free_cpumask_var(rd
->online
);
6353 free_cpumask_var(rd
->span
);
6357 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6359 struct root_domain
*old_rd
= NULL
;
6360 unsigned long flags
;
6362 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6367 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6370 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6373 * If we dont want to free the old_rt yet then
6374 * set old_rd to NULL to skip the freeing later
6377 if (!atomic_dec_and_test(&old_rd
->refcount
))
6381 atomic_inc(&rd
->refcount
);
6384 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6385 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6388 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6391 free_rootdomain(old_rd
);
6394 static int init_rootdomain(struct root_domain
*rd
)
6396 memset(rd
, 0, sizeof(*rd
));
6398 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6400 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6402 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6405 if (cpupri_init(&rd
->cpupri
) != 0)
6410 free_cpumask_var(rd
->rto_mask
);
6412 free_cpumask_var(rd
->online
);
6414 free_cpumask_var(rd
->span
);
6419 static void init_defrootdomain(void)
6421 init_rootdomain(&def_root_domain
);
6423 atomic_set(&def_root_domain
.refcount
, 1);
6426 static struct root_domain
*alloc_rootdomain(void)
6428 struct root_domain
*rd
;
6430 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6434 if (init_rootdomain(rd
) != 0) {
6443 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6444 * hold the hotplug lock.
6447 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6449 struct rq
*rq
= cpu_rq(cpu
);
6450 struct sched_domain
*tmp
;
6452 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6453 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6455 /* Remove the sched domains which do not contribute to scheduling. */
6456 for (tmp
= sd
; tmp
; ) {
6457 struct sched_domain
*parent
= tmp
->parent
;
6461 if (sd_parent_degenerate(tmp
, parent
)) {
6462 tmp
->parent
= parent
->parent
;
6464 parent
->parent
->child
= tmp
;
6469 if (sd
&& sd_degenerate(sd
)) {
6475 sched_domain_debug(sd
, cpu
);
6477 rq_attach_root(rq
, rd
);
6478 rcu_assign_pointer(rq
->sd
, sd
);
6481 /* cpus with isolated domains */
6482 static cpumask_var_t cpu_isolated_map
;
6484 /* Setup the mask of cpus configured for isolated domains */
6485 static int __init
isolated_cpu_setup(char *str
)
6487 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6488 cpulist_parse(str
, cpu_isolated_map
);
6492 __setup("isolcpus=", isolated_cpu_setup
);
6495 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6496 * to a function which identifies what group(along with sched group) a CPU
6497 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6498 * (due to the fact that we keep track of groups covered with a struct cpumask).
6500 * init_sched_build_groups will build a circular linked list of the groups
6501 * covered by the given span, and will set each group's ->cpumask correctly,
6502 * and ->cpu_power to 0.
6505 init_sched_build_groups(const struct cpumask
*span
,
6506 const struct cpumask
*cpu_map
,
6507 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6508 struct sched_group
**sg
,
6509 struct cpumask
*tmpmask
),
6510 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6512 struct sched_group
*first
= NULL
, *last
= NULL
;
6515 cpumask_clear(covered
);
6517 for_each_cpu(i
, span
) {
6518 struct sched_group
*sg
;
6519 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6522 if (cpumask_test_cpu(i
, covered
))
6525 cpumask_clear(sched_group_cpus(sg
));
6528 for_each_cpu(j
, span
) {
6529 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6532 cpumask_set_cpu(j
, covered
);
6533 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6544 #define SD_NODES_PER_DOMAIN 16
6549 * find_next_best_node - find the next node to include in a sched_domain
6550 * @node: node whose sched_domain we're building
6551 * @used_nodes: nodes already in the sched_domain
6553 * Find the next node to include in a given scheduling domain. Simply
6554 * finds the closest node not already in the @used_nodes map.
6556 * Should use nodemask_t.
6558 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6560 int i
, n
, val
, min_val
, best_node
= 0;
6564 for (i
= 0; i
< nr_node_ids
; i
++) {
6565 /* Start at @node */
6566 n
= (node
+ i
) % nr_node_ids
;
6568 if (!nr_cpus_node(n
))
6571 /* Skip already used nodes */
6572 if (node_isset(n
, *used_nodes
))
6575 /* Simple min distance search */
6576 val
= node_distance(node
, n
);
6578 if (val
< min_val
) {
6584 node_set(best_node
, *used_nodes
);
6589 * sched_domain_node_span - get a cpumask for a node's sched_domain
6590 * @node: node whose cpumask we're constructing
6591 * @span: resulting cpumask
6593 * Given a node, construct a good cpumask for its sched_domain to span. It
6594 * should be one that prevents unnecessary balancing, but also spreads tasks
6597 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6599 nodemask_t used_nodes
;
6602 cpumask_clear(span
);
6603 nodes_clear(used_nodes
);
6605 cpumask_or(span
, span
, cpumask_of_node(node
));
6606 node_set(node
, used_nodes
);
6608 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6609 int next_node
= find_next_best_node(node
, &used_nodes
);
6611 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6614 #endif /* CONFIG_NUMA */
6616 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6619 * The cpus mask in sched_group and sched_domain hangs off the end.
6621 * ( See the the comments in include/linux/sched.h:struct sched_group
6622 * and struct sched_domain. )
6624 struct static_sched_group
{
6625 struct sched_group sg
;
6626 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6629 struct static_sched_domain
{
6630 struct sched_domain sd
;
6631 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6637 cpumask_var_t domainspan
;
6638 cpumask_var_t covered
;
6639 cpumask_var_t notcovered
;
6641 cpumask_var_t nodemask
;
6642 cpumask_var_t this_sibling_map
;
6643 cpumask_var_t this_core_map
;
6644 cpumask_var_t this_book_map
;
6645 cpumask_var_t send_covered
;
6646 cpumask_var_t tmpmask
;
6647 struct sched_group
**sched_group_nodes
;
6648 struct root_domain
*rd
;
6652 sa_sched_groups
= 0,
6658 sa_this_sibling_map
,
6660 sa_sched_group_nodes
,
6670 * SMT sched-domains:
6672 #ifdef CONFIG_SCHED_SMT
6673 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6674 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6677 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6678 struct sched_group
**sg
, struct cpumask
*unused
)
6681 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6684 #endif /* CONFIG_SCHED_SMT */
6687 * multi-core sched-domains:
6689 #ifdef CONFIG_SCHED_MC
6690 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6691 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6694 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6695 struct sched_group
**sg
, struct cpumask
*mask
)
6698 #ifdef CONFIG_SCHED_SMT
6699 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6700 group
= cpumask_first(mask
);
6705 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6708 #endif /* CONFIG_SCHED_MC */
6711 * book sched-domains:
6713 #ifdef CONFIG_SCHED_BOOK
6714 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
6715 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
6718 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
6719 struct sched_group
**sg
, struct cpumask
*mask
)
6722 #ifdef CONFIG_SCHED_MC
6723 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6724 group
= cpumask_first(mask
);
6725 #elif defined(CONFIG_SCHED_SMT)
6726 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6727 group
= cpumask_first(mask
);
6730 *sg
= &per_cpu(sched_group_book
, group
).sg
;
6733 #endif /* CONFIG_SCHED_BOOK */
6735 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6736 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6739 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6740 struct sched_group
**sg
, struct cpumask
*mask
)
6743 #ifdef CONFIG_SCHED_BOOK
6744 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
6745 group
= cpumask_first(mask
);
6746 #elif defined(CONFIG_SCHED_MC)
6747 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6748 group
= cpumask_first(mask
);
6749 #elif defined(CONFIG_SCHED_SMT)
6750 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6751 group
= cpumask_first(mask
);
6756 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6762 * The init_sched_build_groups can't handle what we want to do with node
6763 * groups, so roll our own. Now each node has its own list of groups which
6764 * gets dynamically allocated.
6766 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6767 static struct sched_group
***sched_group_nodes_bycpu
;
6769 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6770 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6772 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6773 struct sched_group
**sg
,
6774 struct cpumask
*nodemask
)
6778 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6779 group
= cpumask_first(nodemask
);
6782 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6786 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6788 struct sched_group
*sg
= group_head
;
6794 for_each_cpu(j
, sched_group_cpus(sg
)) {
6795 struct sched_domain
*sd
;
6797 sd
= &per_cpu(phys_domains
, j
).sd
;
6798 if (j
!= group_first_cpu(sd
->groups
)) {
6800 * Only add "power" once for each
6806 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6809 } while (sg
!= group_head
);
6812 static int build_numa_sched_groups(struct s_data
*d
,
6813 const struct cpumask
*cpu_map
, int num
)
6815 struct sched_domain
*sd
;
6816 struct sched_group
*sg
, *prev
;
6819 cpumask_clear(d
->covered
);
6820 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6821 if (cpumask_empty(d
->nodemask
)) {
6822 d
->sched_group_nodes
[num
] = NULL
;
6826 sched_domain_node_span(num
, d
->domainspan
);
6827 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6829 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6832 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6836 d
->sched_group_nodes
[num
] = sg
;
6838 for_each_cpu(j
, d
->nodemask
) {
6839 sd
= &per_cpu(node_domains
, j
).sd
;
6844 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6846 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6849 for (j
= 0; j
< nr_node_ids
; j
++) {
6850 n
= (num
+ j
) % nr_node_ids
;
6851 cpumask_complement(d
->notcovered
, d
->covered
);
6852 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6853 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6854 if (cpumask_empty(d
->tmpmask
))
6856 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6857 if (cpumask_empty(d
->tmpmask
))
6859 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6863 "Can not alloc domain group for node %d\n", j
);
6867 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6868 sg
->next
= prev
->next
;
6869 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6876 #endif /* CONFIG_NUMA */
6879 /* Free memory allocated for various sched_group structures */
6880 static void free_sched_groups(const struct cpumask
*cpu_map
,
6881 struct cpumask
*nodemask
)
6885 for_each_cpu(cpu
, cpu_map
) {
6886 struct sched_group
**sched_group_nodes
6887 = sched_group_nodes_bycpu
[cpu
];
6889 if (!sched_group_nodes
)
6892 for (i
= 0; i
< nr_node_ids
; i
++) {
6893 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6895 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6896 if (cpumask_empty(nodemask
))
6906 if (oldsg
!= sched_group_nodes
[i
])
6909 kfree(sched_group_nodes
);
6910 sched_group_nodes_bycpu
[cpu
] = NULL
;
6913 #else /* !CONFIG_NUMA */
6914 static void free_sched_groups(const struct cpumask
*cpu_map
,
6915 struct cpumask
*nodemask
)
6918 #endif /* CONFIG_NUMA */
6921 * Initialize sched groups cpu_power.
6923 * cpu_power indicates the capacity of sched group, which is used while
6924 * distributing the load between different sched groups in a sched domain.
6925 * Typically cpu_power for all the groups in a sched domain will be same unless
6926 * there are asymmetries in the topology. If there are asymmetries, group
6927 * having more cpu_power will pickup more load compared to the group having
6930 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6932 struct sched_domain
*child
;
6933 struct sched_group
*group
;
6937 WARN_ON(!sd
|| !sd
->groups
);
6939 if (cpu
!= group_first_cpu(sd
->groups
))
6942 sd
->groups
->group_weight
= cpumask_weight(sched_group_cpus(sd
->groups
));
6946 sd
->groups
->cpu_power
= 0;
6949 power
= SCHED_LOAD_SCALE
;
6950 weight
= cpumask_weight(sched_domain_span(sd
));
6952 * SMT siblings share the power of a single core.
6953 * Usually multiple threads get a better yield out of
6954 * that one core than a single thread would have,
6955 * reflect that in sd->smt_gain.
6957 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6958 power
*= sd
->smt_gain
;
6960 power
>>= SCHED_LOAD_SHIFT
;
6962 sd
->groups
->cpu_power
+= power
;
6967 * Add cpu_power of each child group to this groups cpu_power.
6969 group
= child
->groups
;
6971 sd
->groups
->cpu_power
+= group
->cpu_power
;
6972 group
= group
->next
;
6973 } while (group
!= child
->groups
);
6977 * Initializers for schedule domains
6978 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6981 #ifdef CONFIG_SCHED_DEBUG
6982 # define SD_INIT_NAME(sd, type) sd->name = #type
6984 # define SD_INIT_NAME(sd, type) do { } while (0)
6987 #define SD_INIT(sd, type) sd_init_##type(sd)
6989 #define SD_INIT_FUNC(type) \
6990 static noinline void sd_init_##type(struct sched_domain *sd) \
6992 memset(sd, 0, sizeof(*sd)); \
6993 *sd = SD_##type##_INIT; \
6994 sd->level = SD_LV_##type; \
6995 SD_INIT_NAME(sd, type); \
7000 SD_INIT_FUNC(ALLNODES
)
7003 #ifdef CONFIG_SCHED_SMT
7004 SD_INIT_FUNC(SIBLING
)
7006 #ifdef CONFIG_SCHED_MC
7009 #ifdef CONFIG_SCHED_BOOK
7013 static int default_relax_domain_level
= -1;
7015 static int __init
setup_relax_domain_level(char *str
)
7019 val
= simple_strtoul(str
, NULL
, 0);
7020 if (val
< SD_LV_MAX
)
7021 default_relax_domain_level
= val
;
7025 __setup("relax_domain_level=", setup_relax_domain_level
);
7027 static void set_domain_attribute(struct sched_domain
*sd
,
7028 struct sched_domain_attr
*attr
)
7032 if (!attr
|| attr
->relax_domain_level
< 0) {
7033 if (default_relax_domain_level
< 0)
7036 request
= default_relax_domain_level
;
7038 request
= attr
->relax_domain_level
;
7039 if (request
< sd
->level
) {
7040 /* turn off idle balance on this domain */
7041 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7043 /* turn on idle balance on this domain */
7044 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7048 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7049 const struct cpumask
*cpu_map
)
7052 case sa_sched_groups
:
7053 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
7054 d
->sched_group_nodes
= NULL
;
7056 free_rootdomain(d
->rd
); /* fall through */
7058 free_cpumask_var(d
->tmpmask
); /* fall through */
7059 case sa_send_covered
:
7060 free_cpumask_var(d
->send_covered
); /* fall through */
7061 case sa_this_book_map
:
7062 free_cpumask_var(d
->this_book_map
); /* fall through */
7063 case sa_this_core_map
:
7064 free_cpumask_var(d
->this_core_map
); /* fall through */
7065 case sa_this_sibling_map
:
7066 free_cpumask_var(d
->this_sibling_map
); /* fall through */
7068 free_cpumask_var(d
->nodemask
); /* fall through */
7069 case sa_sched_group_nodes
:
7071 kfree(d
->sched_group_nodes
); /* fall through */
7073 free_cpumask_var(d
->notcovered
); /* fall through */
7075 free_cpumask_var(d
->covered
); /* fall through */
7077 free_cpumask_var(d
->domainspan
); /* fall through */
7084 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7085 const struct cpumask
*cpu_map
)
7088 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
7090 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
7091 return sa_domainspan
;
7092 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
7094 /* Allocate the per-node list of sched groups */
7095 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
7096 sizeof(struct sched_group
*), GFP_KERNEL
);
7097 if (!d
->sched_group_nodes
) {
7098 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7099 return sa_notcovered
;
7101 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
7103 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
7104 return sa_sched_group_nodes
;
7105 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
7107 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
7108 return sa_this_sibling_map
;
7109 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
7110 return sa_this_core_map
;
7111 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
7112 return sa_this_book_map
;
7113 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
7114 return sa_send_covered
;
7115 d
->rd
= alloc_rootdomain();
7117 printk(KERN_WARNING
"Cannot alloc root domain\n");
7120 return sa_rootdomain
;
7123 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
7124 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
7126 struct sched_domain
*sd
= NULL
;
7128 struct sched_domain
*parent
;
7131 if (cpumask_weight(cpu_map
) >
7132 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7133 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7134 SD_INIT(sd
, ALLNODES
);
7135 set_domain_attribute(sd
, attr
);
7136 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7137 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7142 sd
= &per_cpu(node_domains
, i
).sd
;
7144 set_domain_attribute(sd
, attr
);
7145 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7146 sd
->parent
= parent
;
7149 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7154 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7155 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7156 struct sched_domain
*parent
, int i
)
7158 struct sched_domain
*sd
;
7159 sd
= &per_cpu(phys_domains
, i
).sd
;
7161 set_domain_attribute(sd
, attr
);
7162 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7163 sd
->parent
= parent
;
7166 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7170 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7171 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7172 struct sched_domain
*parent
, int i
)
7174 struct sched_domain
*sd
= parent
;
7175 #ifdef CONFIG_SCHED_BOOK
7176 sd
= &per_cpu(book_domains
, i
).sd
;
7178 set_domain_attribute(sd
, attr
);
7179 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7180 sd
->parent
= parent
;
7182 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7187 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7188 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7189 struct sched_domain
*parent
, int i
)
7191 struct sched_domain
*sd
= parent
;
7192 #ifdef CONFIG_SCHED_MC
7193 sd
= &per_cpu(core_domains
, i
).sd
;
7195 set_domain_attribute(sd
, attr
);
7196 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7197 sd
->parent
= parent
;
7199 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7204 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7205 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7206 struct sched_domain
*parent
, int i
)
7208 struct sched_domain
*sd
= parent
;
7209 #ifdef CONFIG_SCHED_SMT
7210 sd
= &per_cpu(cpu_domains
, i
).sd
;
7211 SD_INIT(sd
, SIBLING
);
7212 set_domain_attribute(sd
, attr
);
7213 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7214 sd
->parent
= parent
;
7216 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7221 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7222 const struct cpumask
*cpu_map
, int cpu
)
7225 #ifdef CONFIG_SCHED_SMT
7226 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7227 cpumask_and(d
->this_sibling_map
, cpu_map
,
7228 topology_thread_cpumask(cpu
));
7229 if (cpu
== cpumask_first(d
->this_sibling_map
))
7230 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7232 d
->send_covered
, d
->tmpmask
);
7235 #ifdef CONFIG_SCHED_MC
7236 case SD_LV_MC
: /* set up multi-core groups */
7237 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7238 if (cpu
== cpumask_first(d
->this_core_map
))
7239 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7241 d
->send_covered
, d
->tmpmask
);
7244 #ifdef CONFIG_SCHED_BOOK
7245 case SD_LV_BOOK
: /* set up book groups */
7246 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7247 if (cpu
== cpumask_first(d
->this_book_map
))
7248 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7250 d
->send_covered
, d
->tmpmask
);
7253 case SD_LV_CPU
: /* set up physical groups */
7254 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7255 if (!cpumask_empty(d
->nodemask
))
7256 init_sched_build_groups(d
->nodemask
, cpu_map
,
7258 d
->send_covered
, d
->tmpmask
);
7261 case SD_LV_ALLNODES
:
7262 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7263 d
->send_covered
, d
->tmpmask
);
7272 * Build sched domains for a given set of cpus and attach the sched domains
7273 * to the individual cpus
7275 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7276 struct sched_domain_attr
*attr
)
7278 enum s_alloc alloc_state
= sa_none
;
7280 struct sched_domain
*sd
;
7286 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7287 if (alloc_state
!= sa_rootdomain
)
7289 alloc_state
= sa_sched_groups
;
7292 * Set up domains for cpus specified by the cpu_map.
7294 for_each_cpu(i
, cpu_map
) {
7295 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7298 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7299 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7300 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7301 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7302 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7305 for_each_cpu(i
, cpu_map
) {
7306 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7307 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7308 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7311 /* Set up physical groups */
7312 for (i
= 0; i
< nr_node_ids
; i
++)
7313 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7316 /* Set up node groups */
7318 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7320 for (i
= 0; i
< nr_node_ids
; i
++)
7321 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7325 /* Calculate CPU power for physical packages and nodes */
7326 #ifdef CONFIG_SCHED_SMT
7327 for_each_cpu(i
, cpu_map
) {
7328 sd
= &per_cpu(cpu_domains
, i
).sd
;
7329 init_sched_groups_power(i
, sd
);
7332 #ifdef CONFIG_SCHED_MC
7333 for_each_cpu(i
, cpu_map
) {
7334 sd
= &per_cpu(core_domains
, i
).sd
;
7335 init_sched_groups_power(i
, sd
);
7338 #ifdef CONFIG_SCHED_BOOK
7339 for_each_cpu(i
, cpu_map
) {
7340 sd
= &per_cpu(book_domains
, i
).sd
;
7341 init_sched_groups_power(i
, sd
);
7345 for_each_cpu(i
, cpu_map
) {
7346 sd
= &per_cpu(phys_domains
, i
).sd
;
7347 init_sched_groups_power(i
, sd
);
7351 for (i
= 0; i
< nr_node_ids
; i
++)
7352 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7354 if (d
.sd_allnodes
) {
7355 struct sched_group
*sg
;
7357 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7359 init_numa_sched_groups_power(sg
);
7363 /* Attach the domains */
7364 for_each_cpu(i
, cpu_map
) {
7365 #ifdef CONFIG_SCHED_SMT
7366 sd
= &per_cpu(cpu_domains
, i
).sd
;
7367 #elif defined(CONFIG_SCHED_MC)
7368 sd
= &per_cpu(core_domains
, i
).sd
;
7369 #elif defined(CONFIG_SCHED_BOOK)
7370 sd
= &per_cpu(book_domains
, i
).sd
;
7372 sd
= &per_cpu(phys_domains
, i
).sd
;
7374 cpu_attach_domain(sd
, d
.rd
, i
);
7377 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7378 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7382 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7386 static int build_sched_domains(const struct cpumask
*cpu_map
)
7388 return __build_sched_domains(cpu_map
, NULL
);
7391 static cpumask_var_t
*doms_cur
; /* current sched domains */
7392 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7393 static struct sched_domain_attr
*dattr_cur
;
7394 /* attribues of custom domains in 'doms_cur' */
7397 * Special case: If a kmalloc of a doms_cur partition (array of
7398 * cpumask) fails, then fallback to a single sched domain,
7399 * as determined by the single cpumask fallback_doms.
7401 static cpumask_var_t fallback_doms
;
7404 * arch_update_cpu_topology lets virtualized architectures update the
7405 * cpu core maps. It is supposed to return 1 if the topology changed
7406 * or 0 if it stayed the same.
7408 int __attribute__((weak
)) arch_update_cpu_topology(void)
7413 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7416 cpumask_var_t
*doms
;
7418 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7421 for (i
= 0; i
< ndoms
; i
++) {
7422 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7423 free_sched_domains(doms
, i
);
7430 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7433 for (i
= 0; i
< ndoms
; i
++)
7434 free_cpumask_var(doms
[i
]);
7439 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7440 * For now this just excludes isolated cpus, but could be used to
7441 * exclude other special cases in the future.
7443 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7447 arch_update_cpu_topology();
7449 doms_cur
= alloc_sched_domains(ndoms_cur
);
7451 doms_cur
= &fallback_doms
;
7452 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7454 err
= build_sched_domains(doms_cur
[0]);
7455 register_sched_domain_sysctl();
7460 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7461 struct cpumask
*tmpmask
)
7463 free_sched_groups(cpu_map
, tmpmask
);
7467 * Detach sched domains from a group of cpus specified in cpu_map
7468 * These cpus will now be attached to the NULL domain
7470 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7472 /* Save because hotplug lock held. */
7473 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7476 for_each_cpu(i
, cpu_map
)
7477 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7478 synchronize_sched();
7479 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7482 /* handle null as "default" */
7483 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7484 struct sched_domain_attr
*new, int idx_new
)
7486 struct sched_domain_attr tmp
;
7493 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7494 new ? (new + idx_new
) : &tmp
,
7495 sizeof(struct sched_domain_attr
));
7499 * Partition sched domains as specified by the 'ndoms_new'
7500 * cpumasks in the array doms_new[] of cpumasks. This compares
7501 * doms_new[] to the current sched domain partitioning, doms_cur[].
7502 * It destroys each deleted domain and builds each new domain.
7504 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7505 * The masks don't intersect (don't overlap.) We should setup one
7506 * sched domain for each mask. CPUs not in any of the cpumasks will
7507 * not be load balanced. If the same cpumask appears both in the
7508 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7511 * The passed in 'doms_new' should be allocated using
7512 * alloc_sched_domains. This routine takes ownership of it and will
7513 * free_sched_domains it when done with it. If the caller failed the
7514 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7515 * and partition_sched_domains() will fallback to the single partition
7516 * 'fallback_doms', it also forces the domains to be rebuilt.
7518 * If doms_new == NULL it will be replaced with cpu_online_mask.
7519 * ndoms_new == 0 is a special case for destroying existing domains,
7520 * and it will not create the default domain.
7522 * Call with hotplug lock held
7524 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7525 struct sched_domain_attr
*dattr_new
)
7530 mutex_lock(&sched_domains_mutex
);
7532 /* always unregister in case we don't destroy any domains */
7533 unregister_sched_domain_sysctl();
7535 /* Let architecture update cpu core mappings. */
7536 new_topology
= arch_update_cpu_topology();
7538 n
= doms_new
? ndoms_new
: 0;
7540 /* Destroy deleted domains */
7541 for (i
= 0; i
< ndoms_cur
; i
++) {
7542 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7543 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7544 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7547 /* no match - a current sched domain not in new doms_new[] */
7548 detach_destroy_domains(doms_cur
[i
]);
7553 if (doms_new
== NULL
) {
7555 doms_new
= &fallback_doms
;
7556 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7557 WARN_ON_ONCE(dattr_new
);
7560 /* Build new domains */
7561 for (i
= 0; i
< ndoms_new
; i
++) {
7562 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7563 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7564 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7567 /* no match - add a new doms_new */
7568 __build_sched_domains(doms_new
[i
],
7569 dattr_new
? dattr_new
+ i
: NULL
);
7574 /* Remember the new sched domains */
7575 if (doms_cur
!= &fallback_doms
)
7576 free_sched_domains(doms_cur
, ndoms_cur
);
7577 kfree(dattr_cur
); /* kfree(NULL) is safe */
7578 doms_cur
= doms_new
;
7579 dattr_cur
= dattr_new
;
7580 ndoms_cur
= ndoms_new
;
7582 register_sched_domain_sysctl();
7584 mutex_unlock(&sched_domains_mutex
);
7587 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7588 static void arch_reinit_sched_domains(void)
7592 /* Destroy domains first to force the rebuild */
7593 partition_sched_domains(0, NULL
, NULL
);
7595 rebuild_sched_domains();
7599 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7601 unsigned int level
= 0;
7603 if (sscanf(buf
, "%u", &level
) != 1)
7607 * level is always be positive so don't check for
7608 * level < POWERSAVINGS_BALANCE_NONE which is 0
7609 * What happens on 0 or 1 byte write,
7610 * need to check for count as well?
7613 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7617 sched_smt_power_savings
= level
;
7619 sched_mc_power_savings
= level
;
7621 arch_reinit_sched_domains();
7626 #ifdef CONFIG_SCHED_MC
7627 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7628 struct sysdev_class_attribute
*attr
,
7631 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7633 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7634 struct sysdev_class_attribute
*attr
,
7635 const char *buf
, size_t count
)
7637 return sched_power_savings_store(buf
, count
, 0);
7639 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7640 sched_mc_power_savings_show
,
7641 sched_mc_power_savings_store
);
7644 #ifdef CONFIG_SCHED_SMT
7645 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7646 struct sysdev_class_attribute
*attr
,
7649 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7651 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7652 struct sysdev_class_attribute
*attr
,
7653 const char *buf
, size_t count
)
7655 return sched_power_savings_store(buf
, count
, 1);
7657 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7658 sched_smt_power_savings_show
,
7659 sched_smt_power_savings_store
);
7662 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7666 #ifdef CONFIG_SCHED_SMT
7668 err
= sysfs_create_file(&cls
->kset
.kobj
,
7669 &attr_sched_smt_power_savings
.attr
);
7671 #ifdef CONFIG_SCHED_MC
7672 if (!err
&& mc_capable())
7673 err
= sysfs_create_file(&cls
->kset
.kobj
,
7674 &attr_sched_mc_power_savings
.attr
);
7678 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7681 * Update cpusets according to cpu_active mask. If cpusets are
7682 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7683 * around partition_sched_domains().
7685 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7688 switch (action
& ~CPU_TASKS_FROZEN
) {
7690 case CPU_DOWN_FAILED
:
7691 cpuset_update_active_cpus();
7698 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7701 switch (action
& ~CPU_TASKS_FROZEN
) {
7702 case CPU_DOWN_PREPARE
:
7703 cpuset_update_active_cpus();
7710 static int update_runtime(struct notifier_block
*nfb
,
7711 unsigned long action
, void *hcpu
)
7713 int cpu
= (int)(long)hcpu
;
7716 case CPU_DOWN_PREPARE
:
7717 case CPU_DOWN_PREPARE_FROZEN
:
7718 disable_runtime(cpu_rq(cpu
));
7721 case CPU_DOWN_FAILED
:
7722 case CPU_DOWN_FAILED_FROZEN
:
7724 case CPU_ONLINE_FROZEN
:
7725 enable_runtime(cpu_rq(cpu
));
7733 void __init
sched_init_smp(void)
7735 cpumask_var_t non_isolated_cpus
;
7737 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7738 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7740 #if defined(CONFIG_NUMA)
7741 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7743 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7746 mutex_lock(&sched_domains_mutex
);
7747 arch_init_sched_domains(cpu_active_mask
);
7748 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7749 if (cpumask_empty(non_isolated_cpus
))
7750 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7751 mutex_unlock(&sched_domains_mutex
);
7754 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7755 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7757 /* RT runtime code needs to handle some hotplug events */
7758 hotcpu_notifier(update_runtime
, 0);
7762 /* Move init over to a non-isolated CPU */
7763 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7765 sched_init_granularity();
7766 free_cpumask_var(non_isolated_cpus
);
7768 init_sched_rt_class();
7771 void __init
sched_init_smp(void)
7773 sched_init_granularity();
7775 #endif /* CONFIG_SMP */
7777 const_debug
unsigned int sysctl_timer_migration
= 1;
7779 int in_sched_functions(unsigned long addr
)
7781 return in_lock_functions(addr
) ||
7782 (addr
>= (unsigned long)__sched_text_start
7783 && addr
< (unsigned long)__sched_text_end
);
7786 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7788 cfs_rq
->tasks_timeline
= RB_ROOT
;
7789 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7790 #ifdef CONFIG_FAIR_GROUP_SCHED
7793 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7796 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7798 struct rt_prio_array
*array
;
7801 array
= &rt_rq
->active
;
7802 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7803 INIT_LIST_HEAD(array
->queue
+ i
);
7804 __clear_bit(i
, array
->bitmap
);
7806 /* delimiter for bitsearch: */
7807 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7809 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7810 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7812 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7816 rt_rq
->rt_nr_migratory
= 0;
7817 rt_rq
->overloaded
= 0;
7818 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7822 rt_rq
->rt_throttled
= 0;
7823 rt_rq
->rt_runtime
= 0;
7824 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7826 #ifdef CONFIG_RT_GROUP_SCHED
7827 rt_rq
->rt_nr_boosted
= 0;
7832 #ifdef CONFIG_FAIR_GROUP_SCHED
7833 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7834 struct sched_entity
*se
, int cpu
,
7835 struct sched_entity
*parent
)
7837 struct rq
*rq
= cpu_rq(cpu
);
7838 tg
->cfs_rq
[cpu
] = cfs_rq
;
7839 init_cfs_rq(cfs_rq
, rq
);
7843 /* se could be NULL for root_task_group */
7848 se
->cfs_rq
= &rq
->cfs
;
7850 se
->cfs_rq
= parent
->my_q
;
7853 update_load_set(&se
->load
, 0);
7854 se
->parent
= parent
;
7858 #ifdef CONFIG_RT_GROUP_SCHED
7859 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7860 struct sched_rt_entity
*rt_se
, int cpu
,
7861 struct sched_rt_entity
*parent
)
7863 struct rq
*rq
= cpu_rq(cpu
);
7865 tg
->rt_rq
[cpu
] = rt_rq
;
7866 init_rt_rq(rt_rq
, rq
);
7868 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7870 tg
->rt_se
[cpu
] = rt_se
;
7875 rt_se
->rt_rq
= &rq
->rt
;
7877 rt_se
->rt_rq
= parent
->my_q
;
7879 rt_se
->my_q
= rt_rq
;
7880 rt_se
->parent
= parent
;
7881 INIT_LIST_HEAD(&rt_se
->run_list
);
7885 void __init
sched_init(void)
7888 unsigned long alloc_size
= 0, ptr
;
7890 #ifdef CONFIG_FAIR_GROUP_SCHED
7891 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7893 #ifdef CONFIG_RT_GROUP_SCHED
7894 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7896 #ifdef CONFIG_CPUMASK_OFFSTACK
7897 alloc_size
+= num_possible_cpus() * cpumask_size();
7900 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7902 #ifdef CONFIG_FAIR_GROUP_SCHED
7903 root_task_group
.se
= (struct sched_entity
**)ptr
;
7904 ptr
+= nr_cpu_ids
* sizeof(void **);
7906 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7907 ptr
+= nr_cpu_ids
* sizeof(void **);
7909 #endif /* CONFIG_FAIR_GROUP_SCHED */
7910 #ifdef CONFIG_RT_GROUP_SCHED
7911 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7912 ptr
+= nr_cpu_ids
* sizeof(void **);
7914 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7915 ptr
+= nr_cpu_ids
* sizeof(void **);
7917 #endif /* CONFIG_RT_GROUP_SCHED */
7918 #ifdef CONFIG_CPUMASK_OFFSTACK
7919 for_each_possible_cpu(i
) {
7920 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7921 ptr
+= cpumask_size();
7923 #endif /* CONFIG_CPUMASK_OFFSTACK */
7927 init_defrootdomain();
7930 init_rt_bandwidth(&def_rt_bandwidth
,
7931 global_rt_period(), global_rt_runtime());
7933 #ifdef CONFIG_RT_GROUP_SCHED
7934 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7935 global_rt_period(), global_rt_runtime());
7936 #endif /* CONFIG_RT_GROUP_SCHED */
7938 #ifdef CONFIG_CGROUP_SCHED
7939 list_add(&root_task_group
.list
, &task_groups
);
7940 INIT_LIST_HEAD(&root_task_group
.children
);
7941 autogroup_init(&init_task
);
7942 #endif /* CONFIG_CGROUP_SCHED */
7944 for_each_possible_cpu(i
) {
7948 raw_spin_lock_init(&rq
->lock
);
7950 rq
->calc_load_active
= 0;
7951 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7952 init_cfs_rq(&rq
->cfs
, rq
);
7953 init_rt_rq(&rq
->rt
, rq
);
7954 #ifdef CONFIG_FAIR_GROUP_SCHED
7955 root_task_group
.shares
= root_task_group_load
;
7956 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7958 * How much cpu bandwidth does root_task_group get?
7960 * In case of task-groups formed thr' the cgroup filesystem, it
7961 * gets 100% of the cpu resources in the system. This overall
7962 * system cpu resource is divided among the tasks of
7963 * root_task_group and its child task-groups in a fair manner,
7964 * based on each entity's (task or task-group's) weight
7965 * (se->load.weight).
7967 * In other words, if root_task_group has 10 tasks of weight
7968 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7969 * then A0's share of the cpu resource is:
7971 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7973 * We achieve this by letting root_task_group's tasks sit
7974 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
7976 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
7977 #endif /* CONFIG_FAIR_GROUP_SCHED */
7979 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7980 #ifdef CONFIG_RT_GROUP_SCHED
7981 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7982 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
7985 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7986 rq
->cpu_load
[j
] = 0;
7988 rq
->last_load_update_tick
= jiffies
;
7993 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7994 rq
->post_schedule
= 0;
7995 rq
->active_balance
= 0;
7996 rq
->next_balance
= jiffies
;
8001 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8002 rq_attach_root(rq
, &def_root_domain
);
8004 rq
->nohz_balance_kick
= 0;
8005 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
8009 atomic_set(&rq
->nr_iowait
, 0);
8012 set_load_weight(&init_task
);
8014 #ifdef CONFIG_PREEMPT_NOTIFIERS
8015 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8019 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8022 #ifdef CONFIG_RT_MUTEXES
8023 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8027 * The boot idle thread does lazy MMU switching as well:
8029 atomic_inc(&init_mm
.mm_count
);
8030 enter_lazy_tlb(&init_mm
, current
);
8033 * Make us the idle thread. Technically, schedule() should not be
8034 * called from this thread, however somewhere below it might be,
8035 * but because we are the idle thread, we just pick up running again
8036 * when this runqueue becomes "idle".
8038 init_idle(current
, smp_processor_id());
8040 calc_load_update
= jiffies
+ LOAD_FREQ
;
8043 * During early bootup we pretend to be a normal task:
8045 current
->sched_class
= &fair_sched_class
;
8047 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8048 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
8051 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8052 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8053 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8054 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8055 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8057 /* May be allocated at isolcpus cmdline parse time */
8058 if (cpu_isolated_map
== NULL
)
8059 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8062 scheduler_running
= 1;
8065 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8066 static inline int preempt_count_equals(int preempt_offset
)
8068 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8070 return (nested
== preempt_offset
);
8073 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8076 static unsigned long prev_jiffy
; /* ratelimiting */
8078 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8079 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8081 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8083 prev_jiffy
= jiffies
;
8086 "BUG: sleeping function called from invalid context at %s:%d\n",
8089 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8090 in_atomic(), irqs_disabled(),
8091 current
->pid
, current
->comm
);
8093 debug_show_held_locks(current
);
8094 if (irqs_disabled())
8095 print_irqtrace_events(current
);
8099 EXPORT_SYMBOL(__might_sleep
);
8102 #ifdef CONFIG_MAGIC_SYSRQ
8103 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8107 on_rq
= p
->se
.on_rq
;
8109 deactivate_task(rq
, p
, 0);
8110 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8112 activate_task(rq
, p
, 0);
8113 resched_task(rq
->curr
);
8117 void normalize_rt_tasks(void)
8119 struct task_struct
*g
, *p
;
8120 unsigned long flags
;
8123 read_lock_irqsave(&tasklist_lock
, flags
);
8124 do_each_thread(g
, p
) {
8126 * Only normalize user tasks:
8131 p
->se
.exec_start
= 0;
8132 #ifdef CONFIG_SCHEDSTATS
8133 p
->se
.statistics
.wait_start
= 0;
8134 p
->se
.statistics
.sleep_start
= 0;
8135 p
->se
.statistics
.block_start
= 0;
8140 * Renice negative nice level userspace
8143 if (TASK_NICE(p
) < 0 && p
->mm
)
8144 set_user_nice(p
, 0);
8148 raw_spin_lock(&p
->pi_lock
);
8149 rq
= __task_rq_lock(p
);
8151 normalize_task(rq
, p
);
8153 __task_rq_unlock(rq
);
8154 raw_spin_unlock(&p
->pi_lock
);
8155 } while_each_thread(g
, p
);
8157 read_unlock_irqrestore(&tasklist_lock
, flags
);
8160 #endif /* CONFIG_MAGIC_SYSRQ */
8162 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8164 * These functions are only useful for the IA64 MCA handling, or kdb.
8166 * They can only be called when the whole system has been
8167 * stopped - every CPU needs to be quiescent, and no scheduling
8168 * activity can take place. Using them for anything else would
8169 * be a serious bug, and as a result, they aren't even visible
8170 * under any other configuration.
8174 * curr_task - return the current task for a given cpu.
8175 * @cpu: the processor in question.
8177 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8179 struct task_struct
*curr_task(int cpu
)
8181 return cpu_curr(cpu
);
8184 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8188 * set_curr_task - set the current task for a given cpu.
8189 * @cpu: the processor in question.
8190 * @p: the task pointer to set.
8192 * Description: This function must only be used when non-maskable interrupts
8193 * are serviced on a separate stack. It allows the architecture to switch the
8194 * notion of the current task on a cpu in a non-blocking manner. This function
8195 * must be called with all CPU's synchronized, and interrupts disabled, the
8196 * and caller must save the original value of the current task (see
8197 * curr_task() above) and restore that value before reenabling interrupts and
8198 * re-starting the system.
8200 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8202 void set_curr_task(int cpu
, struct task_struct
*p
)
8209 #ifdef CONFIG_FAIR_GROUP_SCHED
8210 static void free_fair_sched_group(struct task_group
*tg
)
8214 for_each_possible_cpu(i
) {
8216 kfree(tg
->cfs_rq
[i
]);
8226 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8228 struct cfs_rq
*cfs_rq
;
8229 struct sched_entity
*se
;
8233 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8236 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8240 tg
->shares
= NICE_0_LOAD
;
8242 for_each_possible_cpu(i
) {
8245 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8246 GFP_KERNEL
, cpu_to_node(i
));
8250 se
= kzalloc_node(sizeof(struct sched_entity
),
8251 GFP_KERNEL
, cpu_to_node(i
));
8255 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8266 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8268 struct rq
*rq
= cpu_rq(cpu
);
8269 unsigned long flags
;
8272 * Only empty task groups can be destroyed; so we can speculatively
8273 * check on_list without danger of it being re-added.
8275 if (!tg
->cfs_rq
[cpu
]->on_list
)
8278 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8279 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8280 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8282 #else /* !CONFG_FAIR_GROUP_SCHED */
8283 static inline void free_fair_sched_group(struct task_group
*tg
)
8288 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8293 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8296 #endif /* CONFIG_FAIR_GROUP_SCHED */
8298 #ifdef CONFIG_RT_GROUP_SCHED
8299 static void free_rt_sched_group(struct task_group
*tg
)
8303 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8305 for_each_possible_cpu(i
) {
8307 kfree(tg
->rt_rq
[i
]);
8309 kfree(tg
->rt_se
[i
]);
8317 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8319 struct rt_rq
*rt_rq
;
8320 struct sched_rt_entity
*rt_se
;
8324 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8327 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8331 init_rt_bandwidth(&tg
->rt_bandwidth
,
8332 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8334 for_each_possible_cpu(i
) {
8337 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8338 GFP_KERNEL
, cpu_to_node(i
));
8342 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8343 GFP_KERNEL
, cpu_to_node(i
));
8347 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8357 #else /* !CONFIG_RT_GROUP_SCHED */
8358 static inline void free_rt_sched_group(struct task_group
*tg
)
8363 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8367 #endif /* CONFIG_RT_GROUP_SCHED */
8369 #ifdef CONFIG_CGROUP_SCHED
8370 static void free_sched_group(struct task_group
*tg
)
8372 free_fair_sched_group(tg
);
8373 free_rt_sched_group(tg
);
8378 /* allocate runqueue etc for a new task group */
8379 struct task_group
*sched_create_group(struct task_group
*parent
)
8381 struct task_group
*tg
;
8382 unsigned long flags
;
8384 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8386 return ERR_PTR(-ENOMEM
);
8388 if (!alloc_fair_sched_group(tg
, parent
))
8391 if (!alloc_rt_sched_group(tg
, parent
))
8394 spin_lock_irqsave(&task_group_lock
, flags
);
8395 list_add_rcu(&tg
->list
, &task_groups
);
8397 WARN_ON(!parent
); /* root should already exist */
8399 tg
->parent
= parent
;
8400 INIT_LIST_HEAD(&tg
->children
);
8401 list_add_rcu(&tg
->siblings
, &parent
->children
);
8402 spin_unlock_irqrestore(&task_group_lock
, flags
);
8407 free_sched_group(tg
);
8408 return ERR_PTR(-ENOMEM
);
8411 /* rcu callback to free various structures associated with a task group */
8412 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8414 /* now it should be safe to free those cfs_rqs */
8415 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8418 /* Destroy runqueue etc associated with a task group */
8419 void sched_destroy_group(struct task_group
*tg
)
8421 unsigned long flags
;
8424 /* end participation in shares distribution */
8425 for_each_possible_cpu(i
)
8426 unregister_fair_sched_group(tg
, i
);
8428 spin_lock_irqsave(&task_group_lock
, flags
);
8429 list_del_rcu(&tg
->list
);
8430 list_del_rcu(&tg
->siblings
);
8431 spin_unlock_irqrestore(&task_group_lock
, flags
);
8433 /* wait for possible concurrent references to cfs_rqs complete */
8434 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8437 /* change task's runqueue when it moves between groups.
8438 * The caller of this function should have put the task in its new group
8439 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8440 * reflect its new group.
8442 void sched_move_task(struct task_struct
*tsk
)
8445 unsigned long flags
;
8448 rq
= task_rq_lock(tsk
, &flags
);
8450 running
= task_current(rq
, tsk
);
8451 on_rq
= tsk
->se
.on_rq
;
8454 dequeue_task(rq
, tsk
, 0);
8455 if (unlikely(running
))
8456 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8458 #ifdef CONFIG_FAIR_GROUP_SCHED
8459 if (tsk
->sched_class
->task_move_group
)
8460 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8463 set_task_rq(tsk
, task_cpu(tsk
));
8465 if (unlikely(running
))
8466 tsk
->sched_class
->set_curr_task(rq
);
8468 enqueue_task(rq
, tsk
, 0);
8470 task_rq_unlock(rq
, &flags
);
8472 #endif /* CONFIG_CGROUP_SCHED */
8474 #ifdef CONFIG_FAIR_GROUP_SCHED
8475 static DEFINE_MUTEX(shares_mutex
);
8477 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8480 unsigned long flags
;
8483 * We can't change the weight of the root cgroup.
8488 if (shares
< MIN_SHARES
)
8489 shares
= MIN_SHARES
;
8490 else if (shares
> MAX_SHARES
)
8491 shares
= MAX_SHARES
;
8493 mutex_lock(&shares_mutex
);
8494 if (tg
->shares
== shares
)
8497 tg
->shares
= shares
;
8498 for_each_possible_cpu(i
) {
8499 struct rq
*rq
= cpu_rq(i
);
8500 struct sched_entity
*se
;
8503 /* Propagate contribution to hierarchy */
8504 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8505 for_each_sched_entity(se
)
8506 update_cfs_shares(group_cfs_rq(se
), 0);
8507 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8511 mutex_unlock(&shares_mutex
);
8515 unsigned long sched_group_shares(struct task_group
*tg
)
8521 #ifdef CONFIG_RT_GROUP_SCHED
8523 * Ensure that the real time constraints are schedulable.
8525 static DEFINE_MUTEX(rt_constraints_mutex
);
8527 static unsigned long to_ratio(u64 period
, u64 runtime
)
8529 if (runtime
== RUNTIME_INF
)
8532 return div64_u64(runtime
<< 20, period
);
8535 /* Must be called with tasklist_lock held */
8536 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8538 struct task_struct
*g
, *p
;
8540 do_each_thread(g
, p
) {
8541 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8543 } while_each_thread(g
, p
);
8548 struct rt_schedulable_data
{
8549 struct task_group
*tg
;
8554 static int tg_schedulable(struct task_group
*tg
, void *data
)
8556 struct rt_schedulable_data
*d
= data
;
8557 struct task_group
*child
;
8558 unsigned long total
, sum
= 0;
8559 u64 period
, runtime
;
8561 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8562 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8565 period
= d
->rt_period
;
8566 runtime
= d
->rt_runtime
;
8570 * Cannot have more runtime than the period.
8572 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8576 * Ensure we don't starve existing RT tasks.
8578 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8581 total
= to_ratio(period
, runtime
);
8584 * Nobody can have more than the global setting allows.
8586 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8590 * The sum of our children's runtime should not exceed our own.
8592 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8593 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8594 runtime
= child
->rt_bandwidth
.rt_runtime
;
8596 if (child
== d
->tg
) {
8597 period
= d
->rt_period
;
8598 runtime
= d
->rt_runtime
;
8601 sum
+= to_ratio(period
, runtime
);
8610 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8612 struct rt_schedulable_data data
= {
8614 .rt_period
= period
,
8615 .rt_runtime
= runtime
,
8618 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8621 static int tg_set_bandwidth(struct task_group
*tg
,
8622 u64 rt_period
, u64 rt_runtime
)
8626 mutex_lock(&rt_constraints_mutex
);
8627 read_lock(&tasklist_lock
);
8628 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8632 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8633 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8634 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8636 for_each_possible_cpu(i
) {
8637 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8639 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8640 rt_rq
->rt_runtime
= rt_runtime
;
8641 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8643 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8645 read_unlock(&tasklist_lock
);
8646 mutex_unlock(&rt_constraints_mutex
);
8651 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8653 u64 rt_runtime
, rt_period
;
8655 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8656 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8657 if (rt_runtime_us
< 0)
8658 rt_runtime
= RUNTIME_INF
;
8660 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8663 long sched_group_rt_runtime(struct task_group
*tg
)
8667 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8670 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8671 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8672 return rt_runtime_us
;
8675 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8677 u64 rt_runtime
, rt_period
;
8679 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8680 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8685 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8688 long sched_group_rt_period(struct task_group
*tg
)
8692 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8693 do_div(rt_period_us
, NSEC_PER_USEC
);
8694 return rt_period_us
;
8697 static int sched_rt_global_constraints(void)
8699 u64 runtime
, period
;
8702 if (sysctl_sched_rt_period
<= 0)
8705 runtime
= global_rt_runtime();
8706 period
= global_rt_period();
8709 * Sanity check on the sysctl variables.
8711 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8714 mutex_lock(&rt_constraints_mutex
);
8715 read_lock(&tasklist_lock
);
8716 ret
= __rt_schedulable(NULL
, 0, 0);
8717 read_unlock(&tasklist_lock
);
8718 mutex_unlock(&rt_constraints_mutex
);
8723 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8725 /* Don't accept realtime tasks when there is no way for them to run */
8726 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8732 #else /* !CONFIG_RT_GROUP_SCHED */
8733 static int sched_rt_global_constraints(void)
8735 unsigned long flags
;
8738 if (sysctl_sched_rt_period
<= 0)
8742 * There's always some RT tasks in the root group
8743 * -- migration, kstopmachine etc..
8745 if (sysctl_sched_rt_runtime
== 0)
8748 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8749 for_each_possible_cpu(i
) {
8750 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8752 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8753 rt_rq
->rt_runtime
= global_rt_runtime();
8754 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8756 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8760 #endif /* CONFIG_RT_GROUP_SCHED */
8762 int sched_rt_handler(struct ctl_table
*table
, int write
,
8763 void __user
*buffer
, size_t *lenp
,
8767 int old_period
, old_runtime
;
8768 static DEFINE_MUTEX(mutex
);
8771 old_period
= sysctl_sched_rt_period
;
8772 old_runtime
= sysctl_sched_rt_runtime
;
8774 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8776 if (!ret
&& write
) {
8777 ret
= sched_rt_global_constraints();
8779 sysctl_sched_rt_period
= old_period
;
8780 sysctl_sched_rt_runtime
= old_runtime
;
8782 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8783 def_rt_bandwidth
.rt_period
=
8784 ns_to_ktime(global_rt_period());
8787 mutex_unlock(&mutex
);
8792 #ifdef CONFIG_CGROUP_SCHED
8794 /* return corresponding task_group object of a cgroup */
8795 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8797 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8798 struct task_group
, css
);
8801 static struct cgroup_subsys_state
*
8802 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8804 struct task_group
*tg
, *parent
;
8806 if (!cgrp
->parent
) {
8807 /* This is early initialization for the top cgroup */
8808 return &root_task_group
.css
;
8811 parent
= cgroup_tg(cgrp
->parent
);
8812 tg
= sched_create_group(parent
);
8814 return ERR_PTR(-ENOMEM
);
8820 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8822 struct task_group
*tg
= cgroup_tg(cgrp
);
8824 sched_destroy_group(tg
);
8828 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8830 #ifdef CONFIG_RT_GROUP_SCHED
8831 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8834 /* We don't support RT-tasks being in separate groups */
8835 if (tsk
->sched_class
!= &fair_sched_class
)
8842 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8843 struct task_struct
*tsk
, bool threadgroup
)
8845 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8849 struct task_struct
*c
;
8851 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8852 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8864 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8865 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8868 sched_move_task(tsk
);
8870 struct task_struct
*c
;
8872 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8880 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct task_struct
*task
)
8883 * cgroup_exit() is called in the copy_process() failure path.
8884 * Ignore this case since the task hasn't ran yet, this avoids
8885 * trying to poke a half freed task state from generic code.
8887 if (!(task
->flags
& PF_EXITING
))
8890 sched_move_task(task
);
8893 #ifdef CONFIG_FAIR_GROUP_SCHED
8894 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8897 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8900 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8902 struct task_group
*tg
= cgroup_tg(cgrp
);
8904 return (u64
) tg
->shares
;
8906 #endif /* CONFIG_FAIR_GROUP_SCHED */
8908 #ifdef CONFIG_RT_GROUP_SCHED
8909 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8912 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8915 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8917 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8920 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8923 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8926 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8928 return sched_group_rt_period(cgroup_tg(cgrp
));
8930 #endif /* CONFIG_RT_GROUP_SCHED */
8932 static struct cftype cpu_files
[] = {
8933 #ifdef CONFIG_FAIR_GROUP_SCHED
8936 .read_u64
= cpu_shares_read_u64
,
8937 .write_u64
= cpu_shares_write_u64
,
8940 #ifdef CONFIG_RT_GROUP_SCHED
8942 .name
= "rt_runtime_us",
8943 .read_s64
= cpu_rt_runtime_read
,
8944 .write_s64
= cpu_rt_runtime_write
,
8947 .name
= "rt_period_us",
8948 .read_u64
= cpu_rt_period_read_uint
,
8949 .write_u64
= cpu_rt_period_write_uint
,
8954 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8956 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8959 struct cgroup_subsys cpu_cgroup_subsys
= {
8961 .create
= cpu_cgroup_create
,
8962 .destroy
= cpu_cgroup_destroy
,
8963 .can_attach
= cpu_cgroup_can_attach
,
8964 .attach
= cpu_cgroup_attach
,
8965 .exit
= cpu_cgroup_exit
,
8966 .populate
= cpu_cgroup_populate
,
8967 .subsys_id
= cpu_cgroup_subsys_id
,
8971 #endif /* CONFIG_CGROUP_SCHED */
8973 #ifdef CONFIG_CGROUP_CPUACCT
8976 * CPU accounting code for task groups.
8978 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8979 * (balbir@in.ibm.com).
8982 /* track cpu usage of a group of tasks and its child groups */
8984 struct cgroup_subsys_state css
;
8985 /* cpuusage holds pointer to a u64-type object on every cpu */
8986 u64 __percpu
*cpuusage
;
8987 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8988 struct cpuacct
*parent
;
8991 struct cgroup_subsys cpuacct_subsys
;
8993 /* return cpu accounting group corresponding to this container */
8994 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8996 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8997 struct cpuacct
, css
);
9000 /* return cpu accounting group to which this task belongs */
9001 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9003 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9004 struct cpuacct
, css
);
9007 /* create a new cpu accounting group */
9008 static struct cgroup_subsys_state
*cpuacct_create(
9009 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9011 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9017 ca
->cpuusage
= alloc_percpu(u64
);
9021 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9022 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9023 goto out_free_counters
;
9026 ca
->parent
= cgroup_ca(cgrp
->parent
);
9032 percpu_counter_destroy(&ca
->cpustat
[i
]);
9033 free_percpu(ca
->cpuusage
);
9037 return ERR_PTR(-ENOMEM
);
9040 /* destroy an existing cpu accounting group */
9042 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9044 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9047 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9048 percpu_counter_destroy(&ca
->cpustat
[i
]);
9049 free_percpu(ca
->cpuusage
);
9053 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9055 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9058 #ifndef CONFIG_64BIT
9060 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9062 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9064 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9072 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9074 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9076 #ifndef CONFIG_64BIT
9078 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9080 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9082 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9088 /* return total cpu usage (in nanoseconds) of a group */
9089 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9091 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9092 u64 totalcpuusage
= 0;
9095 for_each_present_cpu(i
)
9096 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9098 return totalcpuusage
;
9101 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9104 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9113 for_each_present_cpu(i
)
9114 cpuacct_cpuusage_write(ca
, i
, 0);
9120 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9123 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9127 for_each_present_cpu(i
) {
9128 percpu
= cpuacct_cpuusage_read(ca
, i
);
9129 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9131 seq_printf(m
, "\n");
9135 static const char *cpuacct_stat_desc
[] = {
9136 [CPUACCT_STAT_USER
] = "user",
9137 [CPUACCT_STAT_SYSTEM
] = "system",
9140 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9141 struct cgroup_map_cb
*cb
)
9143 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9146 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9147 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9148 val
= cputime64_to_clock_t(val
);
9149 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9154 static struct cftype files
[] = {
9157 .read_u64
= cpuusage_read
,
9158 .write_u64
= cpuusage_write
,
9161 .name
= "usage_percpu",
9162 .read_seq_string
= cpuacct_percpu_seq_read
,
9166 .read_map
= cpuacct_stats_show
,
9170 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9172 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9176 * charge this task's execution time to its accounting group.
9178 * called with rq->lock held.
9180 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9185 if (unlikely(!cpuacct_subsys
.active
))
9188 cpu
= task_cpu(tsk
);
9194 for (; ca
; ca
= ca
->parent
) {
9195 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9196 *cpuusage
+= cputime
;
9203 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9204 * in cputime_t units. As a result, cpuacct_update_stats calls
9205 * percpu_counter_add with values large enough to always overflow the
9206 * per cpu batch limit causing bad SMP scalability.
9208 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9209 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9210 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9213 #define CPUACCT_BATCH \
9214 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9216 #define CPUACCT_BATCH 0
9220 * Charge the system/user time to the task's accounting group.
9222 static void cpuacct_update_stats(struct task_struct
*tsk
,
9223 enum cpuacct_stat_index idx
, cputime_t val
)
9226 int batch
= CPUACCT_BATCH
;
9228 if (unlikely(!cpuacct_subsys
.active
))
9235 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9241 struct cgroup_subsys cpuacct_subsys
= {
9243 .create
= cpuacct_create
,
9244 .destroy
= cpuacct_destroy
,
9245 .populate
= cpuacct_populate
,
9246 .subsys_id
= cpuacct_subsys_id
,
9248 #endif /* CONFIG_CGROUP_CPUACCT */