4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 raw_spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 struct cgroup_subsys_state css
;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity
**se
;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq
**cfs_rq
;
253 unsigned long shares
;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity
**rt_se
;
258 struct rt_rq
**rt_rq
;
260 struct rt_bandwidth rt_bandwidth
;
264 struct list_head list
;
266 struct task_group
*parent
;
267 struct list_head siblings
;
268 struct list_head children
;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock
);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group
.children
);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group
;
308 /* return group to which a task belongs */
309 static inline struct task_group
*task_group(struct task_struct
*p
)
311 struct task_group
*tg
;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
315 struct task_group
, css
);
317 tg
= &init_task_group
;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
327 p
->se
.parent
= task_group(p
)->se
[cpu
];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
332 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
338 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
339 static inline struct task_group
*task_group(struct task_struct
*p
)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load
;
349 unsigned long nr_running
;
354 struct rb_root tasks_timeline
;
355 struct rb_node
*rb_leftmost
;
357 struct list_head tasks
;
358 struct list_head
*balance_iterator
;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity
*curr
, *next
, *last
;
366 unsigned int nr_spread_over
;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list
;
380 struct task_group
*tg
; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight
;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load
;
397 * this cpu's part of tg->shares
399 unsigned long shares
;
402 * load.weight at the time we set shares
404 unsigned long rq_weight
;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active
;
412 unsigned long rt_nr_running
;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr
; /* highest queued rt task prio */
417 int next
; /* next highest */
422 unsigned long rt_nr_migratory
;
423 unsigned long rt_nr_total
;
425 struct plist_head pushable_tasks
;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock
;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted
;
437 struct list_head leaf_rt_rq_list
;
438 struct task_group
*tg
;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online
;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask
;
464 struct cpupri cpupri
;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain
;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running
;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
496 unsigned char in_nohz_recently
;
498 unsigned int skip_clock_update
;
500 /* capture load from *all* tasks on this cpu: */
501 struct load_weight load
;
502 unsigned long nr_load_updates
;
508 #ifdef CONFIG_FAIR_GROUP_SCHED
509 /* list of leaf cfs_rq on this cpu: */
510 struct list_head leaf_cfs_rq_list
;
512 #ifdef CONFIG_RT_GROUP_SCHED
513 struct list_head leaf_rt_rq_list
;
517 * This is part of a global counter where only the total sum
518 * over all CPUs matters. A task can increase this counter on
519 * one CPU and if it got migrated afterwards it may decrease
520 * it on another CPU. Always updated under the runqueue lock:
522 unsigned long nr_uninterruptible
;
524 struct task_struct
*curr
, *idle
;
525 unsigned long next_balance
;
526 struct mm_struct
*prev_mm
;
533 struct root_domain
*rd
;
534 struct sched_domain
*sd
;
536 unsigned char idle_at_tick
;
537 /* For active balancing */
541 /* cpu of this runqueue: */
545 unsigned long avg_load_per_task
;
547 struct task_struct
*migration_thread
;
548 struct list_head migration_queue
;
556 /* calc_load related fields */
557 unsigned long calc_load_update
;
558 long calc_load_active
;
560 #ifdef CONFIG_SCHED_HRTICK
562 int hrtick_csd_pending
;
563 struct call_single_data hrtick_csd
;
565 struct hrtimer hrtick_timer
;
568 #ifdef CONFIG_SCHEDSTATS
570 struct sched_info rq_sched_info
;
571 unsigned long long rq_cpu_time
;
572 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
574 /* sys_sched_yield() stats */
575 unsigned int yld_count
;
577 /* schedule() stats */
578 unsigned int sched_switch
;
579 unsigned int sched_count
;
580 unsigned int sched_goidle
;
582 /* try_to_wake_up() stats */
583 unsigned int ttwu_count
;
584 unsigned int ttwu_local
;
587 unsigned int bkl_count
;
591 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
594 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
596 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
599 * A queue event has occurred, and we're going to schedule. In
600 * this case, we can save a useless back to back clock update.
602 if (test_tsk_need_resched(p
))
603 rq
->skip_clock_update
= 1;
606 static inline int cpu_of(struct rq
*rq
)
615 #define rcu_dereference_check_sched_domain(p) \
616 rcu_dereference_check((p), \
617 rcu_read_lock_sched_held() || \
618 lockdep_is_held(&sched_domains_mutex))
621 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
622 * See detach_destroy_domains: synchronize_sched for details.
624 * The domain tree of any CPU may only be accessed from within
625 * preempt-disabled sections.
627 #define for_each_domain(cpu, __sd) \
628 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
630 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
631 #define this_rq() (&__get_cpu_var(runqueues))
632 #define task_rq(p) cpu_rq(task_cpu(p))
633 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
634 #define raw_rq() (&__raw_get_cpu_var(runqueues))
636 inline void update_rq_clock(struct rq
*rq
)
638 if (!rq
->skip_clock_update
)
639 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
643 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
645 #ifdef CONFIG_SCHED_DEBUG
646 # define const_debug __read_mostly
648 # define const_debug static const
653 * @cpu: the processor in question.
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(int cpu
)
661 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
665 * Debugging: various feature bits
668 #define SCHED_FEAT(name, enabled) \
669 __SCHED_FEAT_##name ,
672 #include "sched_features.h"
677 #define SCHED_FEAT(name, enabled) \
678 (1UL << __SCHED_FEAT_##name) * enabled |
680 const_debug
unsigned int sysctl_sched_features
=
681 #include "sched_features.h"
686 #ifdef CONFIG_SCHED_DEBUG
687 #define SCHED_FEAT(name, enabled) \
690 static __read_mostly
char *sched_feat_names
[] = {
691 #include "sched_features.h"
697 static int sched_feat_show(struct seq_file
*m
, void *v
)
701 for (i
= 0; sched_feat_names
[i
]; i
++) {
702 if (!(sysctl_sched_features
& (1UL << i
)))
704 seq_printf(m
, "%s ", sched_feat_names
[i
]);
712 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
713 size_t cnt
, loff_t
*ppos
)
723 if (copy_from_user(&buf
, ubuf
, cnt
))
728 if (strncmp(buf
, "NO_", 3) == 0) {
733 for (i
= 0; sched_feat_names
[i
]; i
++) {
734 int len
= strlen(sched_feat_names
[i
]);
736 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
738 sysctl_sched_features
&= ~(1UL << i
);
740 sysctl_sched_features
|= (1UL << i
);
745 if (!sched_feat_names
[i
])
753 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
755 return single_open(filp
, sched_feat_show
, NULL
);
758 static const struct file_operations sched_feat_fops
= {
759 .open
= sched_feat_open
,
760 .write
= sched_feat_write
,
763 .release
= single_release
,
766 static __init
int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
773 late_initcall(sched_init_debug
);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
786 * ratelimit for updating the group shares.
789 unsigned int sysctl_sched_shares_ratelimit
= 250000;
790 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
793 * Inject some fuzzyness into changing the per-cpu group shares
794 * this avoids remote rq-locks at the expense of fairness.
797 unsigned int sysctl_sched_shares_thresh
= 4;
800 * period over which we average the RT time consumption, measured
805 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period
= 1000000;
813 static __read_mostly
int scheduler_running
;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime
= 950000;
821 static inline u64
global_rt_period(void)
823 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
826 static inline u64
global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime
< 0)
831 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
843 return rq
->curr
== p
;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
849 return task_current(rq
, p
);
852 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
856 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq
->lock
.owner
= current
;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
869 raw_spin_unlock_irq(&rq
->lock
);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
878 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq
->lock
);
895 raw_spin_unlock(&rq
->lock
);
899 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize against
918 * ttwu() so that task_cpu() reports a stable number.
920 * We need to make an exception for PF_STARTING tasks because the fork
921 * path might require task_rq_lock() to work, eg. it can call
922 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
924 static inline int task_is_waking(struct task_struct
*p
)
926 return unlikely((p
->state
== TASK_WAKING
) && !(p
->flags
& PF_STARTING
));
930 * __task_rq_lock - lock the runqueue a given task resides on.
931 * Must be called interrupts disabled.
933 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
939 while (task_is_waking(p
))
942 raw_spin_lock(&rq
->lock
);
943 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
945 raw_spin_unlock(&rq
->lock
);
950 * task_rq_lock - lock the runqueue a given task resides on and disable
951 * interrupts. Note the ordering: we can safely lookup the task_rq without
952 * explicitly disabling preemption.
954 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
960 while (task_is_waking(p
))
962 local_irq_save(*flags
);
964 raw_spin_lock(&rq
->lock
);
965 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
967 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
971 void task_rq_unlock_wait(struct task_struct
*p
)
973 struct rq
*rq
= task_rq(p
);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 raw_spin_unlock_wait(&rq
->lock
);
979 static void __task_rq_unlock(struct rq
*rq
)
982 raw_spin_unlock(&rq
->lock
);
985 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
988 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq
*this_rq_lock(void)
1001 raw_spin_lock(&rq
->lock
);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq
*rq
)
1025 if (!sched_feat(HRTICK
))
1027 if (!cpu_active(cpu_of(rq
)))
1029 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1032 static void hrtick_clear(struct rq
*rq
)
1034 if (hrtimer_active(&rq
->hrtick_timer
))
1035 hrtimer_cancel(&rq
->hrtick_timer
);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1044 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1046 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1048 raw_spin_lock(&rq
->lock
);
1049 update_rq_clock(rq
);
1050 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1051 raw_spin_unlock(&rq
->lock
);
1053 return HRTIMER_NORESTART
;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg
)
1062 struct rq
*rq
= arg
;
1064 raw_spin_lock(&rq
->lock
);
1065 hrtimer_restart(&rq
->hrtick_timer
);
1066 rq
->hrtick_csd_pending
= 0;
1067 raw_spin_unlock(&rq
->lock
);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq
*rq
, u64 delay
)
1077 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1078 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1080 hrtimer_set_expires(timer
, time
);
1082 if (rq
== this_rq()) {
1083 hrtimer_restart(timer
);
1084 } else if (!rq
->hrtick_csd_pending
) {
1085 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1086 rq
->hrtick_csd_pending
= 1;
1091 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1093 int cpu
= (int)(long)hcpu
;
1096 case CPU_UP_CANCELED
:
1097 case CPU_UP_CANCELED_FROZEN
:
1098 case CPU_DOWN_PREPARE
:
1099 case CPU_DOWN_PREPARE_FROZEN
:
1101 case CPU_DEAD_FROZEN
:
1102 hrtick_clear(cpu_rq(cpu
));
1109 static __init
void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick
, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq
*rq
, u64 delay
)
1121 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1122 HRTIMER_MODE_REL_PINNED
, 0);
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SMP */
1130 static void init_rq_hrtick(struct rq
*rq
)
1133 rq
->hrtick_csd_pending
= 0;
1135 rq
->hrtick_csd
.flags
= 0;
1136 rq
->hrtick_csd
.func
= __hrtick_start
;
1137 rq
->hrtick_csd
.info
= rq
;
1140 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1141 rq
->hrtick_timer
.function
= hrtick
;
1143 #else /* CONFIG_SCHED_HRTICK */
1144 static inline void hrtick_clear(struct rq
*rq
)
1148 static inline void init_rq_hrtick(struct rq
*rq
)
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SCHED_HRTICK */
1158 * resched_task - mark a task 'to be rescheduled now'.
1160 * On UP this means the setting of the need_resched flag, on SMP it
1161 * might also involve a cross-CPU call to trigger the scheduler on
1166 #ifndef tsk_is_polling
1167 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1170 static void resched_task(struct task_struct
*p
)
1174 assert_raw_spin_locked(&task_rq(p
)->lock
);
1176 if (test_tsk_need_resched(p
))
1179 set_tsk_need_resched(p
);
1182 if (cpu
== smp_processor_id())
1185 /* NEED_RESCHED must be visible before we test polling */
1187 if (!tsk_is_polling(p
))
1188 smp_send_reschedule(cpu
);
1191 static void resched_cpu(int cpu
)
1193 struct rq
*rq
= cpu_rq(cpu
);
1194 unsigned long flags
;
1196 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1198 resched_task(cpu_curr(cpu
));
1199 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1204 * When add_timer_on() enqueues a timer into the timer wheel of an
1205 * idle CPU then this timer might expire before the next timer event
1206 * which is scheduled to wake up that CPU. In case of a completely
1207 * idle system the next event might even be infinite time into the
1208 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1209 * leaves the inner idle loop so the newly added timer is taken into
1210 * account when the CPU goes back to idle and evaluates the timer
1211 * wheel for the next timer event.
1213 void wake_up_idle_cpu(int cpu
)
1215 struct rq
*rq
= cpu_rq(cpu
);
1217 if (cpu
== smp_processor_id())
1221 * This is safe, as this function is called with the timer
1222 * wheel base lock of (cpu) held. When the CPU is on the way
1223 * to idle and has not yet set rq->curr to idle then it will
1224 * be serialized on the timer wheel base lock and take the new
1225 * timer into account automatically.
1227 if (rq
->curr
!= rq
->idle
)
1231 * We can set TIF_RESCHED on the idle task of the other CPU
1232 * lockless. The worst case is that the other CPU runs the
1233 * idle task through an additional NOOP schedule()
1235 set_tsk_need_resched(rq
->idle
);
1237 /* NEED_RESCHED must be visible before we test polling */
1239 if (!tsk_is_polling(rq
->idle
))
1240 smp_send_reschedule(cpu
);
1243 int nohz_ratelimit(int cpu
)
1245 struct rq
*rq
= cpu_rq(cpu
);
1246 u64 diff
= rq
->clock
- rq
->nohz_stamp
;
1248 rq
->nohz_stamp
= rq
->clock
;
1250 return diff
< (NSEC_PER_SEC
/ HZ
) >> 1;
1253 #endif /* CONFIG_NO_HZ */
1255 static u64
sched_avg_period(void)
1257 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1260 static void sched_avg_update(struct rq
*rq
)
1262 s64 period
= sched_avg_period();
1264 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
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
)
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1306 struct load_weight
*lw
)
1310 if (!lw
->inv_weight
) {
1311 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1314 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1318 tmp
= (u64
)delta_exec
* weight
;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp
> WMULT_CONST
))
1323 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1326 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1328 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1331 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1337 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight
[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult
[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 /* Time spent by the tasks of the cpu accounting group executing in ... */
1397 enum cpuacct_stat_index
{
1398 CPUACCT_STAT_USER
, /* ... user mode */
1399 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1401 CPUACCT_STAT_NSTATS
,
1404 #ifdef CONFIG_CGROUP_CPUACCT
1405 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1406 static void cpuacct_update_stats(struct task_struct
*tsk
,
1407 enum cpuacct_stat_index idx
, cputime_t val
);
1409 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1410 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1411 enum cpuacct_stat_index idx
, cputime_t val
) {}
1414 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1416 update_load_add(&rq
->load
, load
);
1419 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1421 update_load_sub(&rq
->load
, load
);
1424 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1425 typedef int (*tg_visitor
)(struct task_group
*, void *);
1428 * Iterate the full tree, calling @down when first entering a node and @up when
1429 * leaving it for the final time.
1431 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1433 struct task_group
*parent
, *child
;
1437 parent
= &root_task_group
;
1439 ret
= (*down
)(parent
, data
);
1442 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1449 ret
= (*up
)(parent
, data
);
1454 parent
= parent
->parent
;
1463 static int tg_nop(struct task_group
*tg
, void *data
)
1470 /* Used instead of source_load when we know the type == 0 */
1471 static unsigned long weighted_cpuload(const int cpu
)
1473 return cpu_rq(cpu
)->load
.weight
;
1477 * Return a low guess at the load of a migration-source cpu weighted
1478 * according to the scheduling class and "nice" value.
1480 * We want to under-estimate the load of migration sources, to
1481 * balance conservatively.
1483 static unsigned long source_load(int cpu
, int type
)
1485 struct rq
*rq
= cpu_rq(cpu
);
1486 unsigned long total
= weighted_cpuload(cpu
);
1488 if (type
== 0 || !sched_feat(LB_BIAS
))
1491 return min(rq
->cpu_load
[type
-1], total
);
1495 * Return a high guess at the load of a migration-target cpu weighted
1496 * according to the scheduling class and "nice" value.
1498 static unsigned long target_load(int cpu
, int type
)
1500 struct rq
*rq
= cpu_rq(cpu
);
1501 unsigned long total
= weighted_cpuload(cpu
);
1503 if (type
== 0 || !sched_feat(LB_BIAS
))
1506 return max(rq
->cpu_load
[type
-1], total
);
1509 static struct sched_group
*group_of(int cpu
)
1511 struct sched_domain
*sd
= rcu_dereference_sched(cpu_rq(cpu
)->sd
);
1519 static unsigned long power_of(int cpu
)
1521 struct sched_group
*group
= group_of(cpu
);
1524 return SCHED_LOAD_SCALE
;
1526 return group
->cpu_power
;
1529 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1531 static unsigned long cpu_avg_load_per_task(int cpu
)
1533 struct rq
*rq
= cpu_rq(cpu
);
1534 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1537 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1539 rq
->avg_load_per_task
= 0;
1541 return rq
->avg_load_per_task
;
1544 #ifdef CONFIG_FAIR_GROUP_SCHED
1546 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1548 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1551 * Calculate and set the cpu's group shares.
1553 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1554 unsigned long sd_shares
,
1555 unsigned long sd_rq_weight
,
1556 unsigned long *usd_rq_weight
)
1558 unsigned long shares
, rq_weight
;
1561 rq_weight
= usd_rq_weight
[cpu
];
1564 rq_weight
= NICE_0_LOAD
;
1568 * \Sum_j shares_j * rq_weight_i
1569 * shares_i = -----------------------------
1570 * \Sum_j rq_weight_j
1572 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1573 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1575 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1576 sysctl_sched_shares_thresh
) {
1577 struct rq
*rq
= cpu_rq(cpu
);
1578 unsigned long flags
;
1580 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1581 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1582 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1583 __set_se_shares(tg
->se
[cpu
], shares
);
1584 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1589 * Re-compute the task group their per cpu shares over the given domain.
1590 * This needs to be done in a bottom-up fashion because the rq weight of a
1591 * parent group depends on the shares of its child groups.
1593 static int tg_shares_up(struct task_group
*tg
, void *data
)
1595 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1596 unsigned long *usd_rq_weight
;
1597 struct sched_domain
*sd
= data
;
1598 unsigned long flags
;
1604 local_irq_save(flags
);
1605 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1607 for_each_cpu(i
, sched_domain_span(sd
)) {
1608 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1609 usd_rq_weight
[i
] = weight
;
1611 rq_weight
+= weight
;
1613 * If there are currently no tasks on the cpu pretend there
1614 * is one of average load so that when a new task gets to
1615 * run here it will not get delayed by group starvation.
1618 weight
= NICE_0_LOAD
;
1620 sum_weight
+= weight
;
1621 shares
+= tg
->cfs_rq
[i
]->shares
;
1625 rq_weight
= sum_weight
;
1627 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1628 shares
= tg
->shares
;
1630 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1631 shares
= tg
->shares
;
1633 for_each_cpu(i
, sched_domain_span(sd
))
1634 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1636 local_irq_restore(flags
);
1642 * Compute the cpu's hierarchical load factor for each task group.
1643 * This needs to be done in a top-down fashion because the load of a child
1644 * group is a fraction of its parents load.
1646 static int tg_load_down(struct task_group
*tg
, void *data
)
1649 long cpu
= (long)data
;
1652 load
= cpu_rq(cpu
)->load
.weight
;
1654 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1655 load
*= tg
->cfs_rq
[cpu
]->shares
;
1656 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1659 tg
->cfs_rq
[cpu
]->h_load
= load
;
1664 static void update_shares(struct sched_domain
*sd
)
1669 if (root_task_group_empty())
1672 now
= cpu_clock(raw_smp_processor_id());
1673 elapsed
= now
- sd
->last_update
;
1675 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1676 sd
->last_update
= now
;
1677 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1681 static void update_h_load(long cpu
)
1683 if (root_task_group_empty())
1686 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1691 static inline void update_shares(struct sched_domain
*sd
)
1697 #ifdef CONFIG_PREEMPT
1699 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1702 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1703 * way at the expense of forcing extra atomic operations in all
1704 * invocations. This assures that the double_lock is acquired using the
1705 * same underlying policy as the spinlock_t on this architecture, which
1706 * reduces latency compared to the unfair variant below. However, it
1707 * also adds more overhead and therefore may reduce throughput.
1709 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1710 __releases(this_rq
->lock
)
1711 __acquires(busiest
->lock
)
1712 __acquires(this_rq
->lock
)
1714 raw_spin_unlock(&this_rq
->lock
);
1715 double_rq_lock(this_rq
, busiest
);
1722 * Unfair double_lock_balance: Optimizes throughput at the expense of
1723 * latency by eliminating extra atomic operations when the locks are
1724 * already in proper order on entry. This favors lower cpu-ids and will
1725 * grant the double lock to lower cpus over higher ids under contention,
1726 * regardless of entry order into the function.
1728 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1729 __releases(this_rq
->lock
)
1730 __acquires(busiest
->lock
)
1731 __acquires(this_rq
->lock
)
1735 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1736 if (busiest
< this_rq
) {
1737 raw_spin_unlock(&this_rq
->lock
);
1738 raw_spin_lock(&busiest
->lock
);
1739 raw_spin_lock_nested(&this_rq
->lock
,
1740 SINGLE_DEPTH_NESTING
);
1743 raw_spin_lock_nested(&busiest
->lock
,
1744 SINGLE_DEPTH_NESTING
);
1749 #endif /* CONFIG_PREEMPT */
1752 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1754 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1756 if (unlikely(!irqs_disabled())) {
1757 /* printk() doesn't work good under rq->lock */
1758 raw_spin_unlock(&this_rq
->lock
);
1762 return _double_lock_balance(this_rq
, busiest
);
1765 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1766 __releases(busiest
->lock
)
1768 raw_spin_unlock(&busiest
->lock
);
1769 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1773 * double_rq_lock - safely lock two runqueues
1775 * Note this does not disable interrupts like task_rq_lock,
1776 * you need to do so manually before calling.
1778 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1779 __acquires(rq1
->lock
)
1780 __acquires(rq2
->lock
)
1782 BUG_ON(!irqs_disabled());
1784 raw_spin_lock(&rq1
->lock
);
1785 __acquire(rq2
->lock
); /* Fake it out ;) */
1788 raw_spin_lock(&rq1
->lock
);
1789 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1791 raw_spin_lock(&rq2
->lock
);
1792 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1798 * double_rq_unlock - safely unlock two runqueues
1800 * Note this does not restore interrupts like task_rq_unlock,
1801 * you need to do so manually after calling.
1803 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1804 __releases(rq1
->lock
)
1805 __releases(rq2
->lock
)
1807 raw_spin_unlock(&rq1
->lock
);
1809 raw_spin_unlock(&rq2
->lock
);
1811 __release(rq2
->lock
);
1816 #ifdef CONFIG_FAIR_GROUP_SCHED
1817 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1820 cfs_rq
->shares
= shares
;
1825 static void calc_load_account_active(struct rq
*this_rq
);
1826 static void update_sysctl(void);
1827 static int get_update_sysctl_factor(void);
1829 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1831 set_task_rq(p
, cpu
);
1834 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1835 * successfuly executed on another CPU. We must ensure that updates of
1836 * per-task data have been completed by this moment.
1839 task_thread_info(p
)->cpu
= cpu
;
1843 static const struct sched_class rt_sched_class
;
1845 #define sched_class_highest (&rt_sched_class)
1846 #define for_each_class(class) \
1847 for (class = sched_class_highest; class; class = class->next)
1849 #include "sched_stats.h"
1851 static void inc_nr_running(struct rq
*rq
)
1856 static void dec_nr_running(struct rq
*rq
)
1861 static void set_load_weight(struct task_struct
*p
)
1863 if (task_has_rt_policy(p
)) {
1864 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1865 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1870 * SCHED_IDLE tasks get minimal weight:
1872 if (p
->policy
== SCHED_IDLE
) {
1873 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1874 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1878 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1879 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1882 static void update_avg(u64
*avg
, u64 sample
)
1884 s64 diff
= sample
- *avg
;
1889 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1891 update_rq_clock(rq
);
1892 sched_info_queued(p
);
1893 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1897 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1899 update_rq_clock(rq
);
1900 sched_info_dequeued(p
);
1901 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1906 * activate_task - move a task to the runqueue.
1908 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1910 if (task_contributes_to_load(p
))
1911 rq
->nr_uninterruptible
--;
1913 enqueue_task(rq
, p
, wakeup
, false);
1918 * deactivate_task - remove a task from the runqueue.
1920 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1922 if (task_contributes_to_load(p
))
1923 rq
->nr_uninterruptible
++;
1925 dequeue_task(rq
, p
, sleep
);
1929 #include "sched_idletask.c"
1930 #include "sched_fair.c"
1931 #include "sched_rt.c"
1932 #ifdef CONFIG_SCHED_DEBUG
1933 # include "sched_debug.c"
1937 * __normal_prio - return the priority that is based on the static prio
1939 static inline int __normal_prio(struct task_struct
*p
)
1941 return p
->static_prio
;
1945 * Calculate the expected normal priority: i.e. priority
1946 * without taking RT-inheritance into account. Might be
1947 * boosted by interactivity modifiers. Changes upon fork,
1948 * setprio syscalls, and whenever the interactivity
1949 * estimator recalculates.
1951 static inline int normal_prio(struct task_struct
*p
)
1955 if (task_has_rt_policy(p
))
1956 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1958 prio
= __normal_prio(p
);
1963 * Calculate the current priority, i.e. the priority
1964 * taken into account by the scheduler. This value might
1965 * be boosted by RT tasks, or might be boosted by
1966 * interactivity modifiers. Will be RT if the task got
1967 * RT-boosted. If not then it returns p->normal_prio.
1969 static int effective_prio(struct task_struct
*p
)
1971 p
->normal_prio
= normal_prio(p
);
1973 * If we are RT tasks or we were boosted to RT priority,
1974 * keep the priority unchanged. Otherwise, update priority
1975 * to the normal priority:
1977 if (!rt_prio(p
->prio
))
1978 return p
->normal_prio
;
1983 * task_curr - is this task currently executing on a CPU?
1984 * @p: the task in question.
1986 inline int task_curr(const struct task_struct
*p
)
1988 return cpu_curr(task_cpu(p
)) == p
;
1991 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1992 const struct sched_class
*prev_class
,
1993 int oldprio
, int running
)
1995 if (prev_class
!= p
->sched_class
) {
1996 if (prev_class
->switched_from
)
1997 prev_class
->switched_from(rq
, p
, running
);
1998 p
->sched_class
->switched_to(rq
, p
, running
);
2000 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2005 * Is this task likely cache-hot:
2008 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2012 if (p
->sched_class
!= &fair_sched_class
)
2016 * Buddy candidates are cache hot:
2018 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2019 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2020 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2023 if (sysctl_sched_migration_cost
== -1)
2025 if (sysctl_sched_migration_cost
== 0)
2028 delta
= now
- p
->se
.exec_start
;
2030 return delta
< (s64
)sysctl_sched_migration_cost
;
2033 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2035 #ifdef CONFIG_SCHED_DEBUG
2037 * We should never call set_task_cpu() on a blocked task,
2038 * ttwu() will sort out the placement.
2040 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2041 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2044 trace_sched_migrate_task(p
, new_cpu
);
2046 if (task_cpu(p
) != new_cpu
) {
2047 p
->se
.nr_migrations
++;
2048 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2051 __set_task_cpu(p
, new_cpu
);
2054 struct migration_req
{
2055 struct list_head list
;
2057 struct task_struct
*task
;
2060 struct completion done
;
2064 * The task's runqueue lock must be held.
2065 * Returns true if you have to wait for migration thread.
2068 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2070 struct rq
*rq
= task_rq(p
);
2073 * If the task is not on a runqueue (and not running), then
2074 * the next wake-up will properly place the task.
2076 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2079 init_completion(&req
->done
);
2081 req
->dest_cpu
= dest_cpu
;
2082 list_add(&req
->list
, &rq
->migration_queue
);
2088 * wait_task_context_switch - wait for a thread to complete at least one
2091 * @p must not be current.
2093 void wait_task_context_switch(struct task_struct
*p
)
2095 unsigned long nvcsw
, nivcsw
, flags
;
2103 * The runqueue is assigned before the actual context
2104 * switch. We need to take the runqueue lock.
2106 * We could check initially without the lock but it is
2107 * very likely that we need to take the lock in every
2110 rq
= task_rq_lock(p
, &flags
);
2111 running
= task_running(rq
, p
);
2112 task_rq_unlock(rq
, &flags
);
2114 if (likely(!running
))
2117 * The switch count is incremented before the actual
2118 * context switch. We thus wait for two switches to be
2119 * sure at least one completed.
2121 if ((p
->nvcsw
- nvcsw
) > 1)
2123 if ((p
->nivcsw
- nivcsw
) > 1)
2131 * wait_task_inactive - wait for a thread to unschedule.
2133 * If @match_state is nonzero, it's the @p->state value just checked and
2134 * not expected to change. If it changes, i.e. @p might have woken up,
2135 * then return zero. When we succeed in waiting for @p to be off its CPU,
2136 * we return a positive number (its total switch count). If a second call
2137 * a short while later returns the same number, the caller can be sure that
2138 * @p has remained unscheduled the whole time.
2140 * The caller must ensure that the task *will* unschedule sometime soon,
2141 * else this function might spin for a *long* time. This function can't
2142 * be called with interrupts off, or it may introduce deadlock with
2143 * smp_call_function() if an IPI is sent by the same process we are
2144 * waiting to become inactive.
2146 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2148 unsigned long flags
;
2155 * We do the initial early heuristics without holding
2156 * any task-queue locks at all. We'll only try to get
2157 * the runqueue lock when things look like they will
2163 * If the task is actively running on another CPU
2164 * still, just relax and busy-wait without holding
2167 * NOTE! Since we don't hold any locks, it's not
2168 * even sure that "rq" stays as the right runqueue!
2169 * But we don't care, since "task_running()" will
2170 * return false if the runqueue has changed and p
2171 * is actually now running somewhere else!
2173 while (task_running(rq
, p
)) {
2174 if (match_state
&& unlikely(p
->state
!= match_state
))
2180 * Ok, time to look more closely! We need the rq
2181 * lock now, to be *sure*. If we're wrong, we'll
2182 * just go back and repeat.
2184 rq
= task_rq_lock(p
, &flags
);
2185 trace_sched_wait_task(rq
, p
);
2186 running
= task_running(rq
, p
);
2187 on_rq
= p
->se
.on_rq
;
2189 if (!match_state
|| p
->state
== match_state
)
2190 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2191 task_rq_unlock(rq
, &flags
);
2194 * If it changed from the expected state, bail out now.
2196 if (unlikely(!ncsw
))
2200 * Was it really running after all now that we
2201 * checked with the proper locks actually held?
2203 * Oops. Go back and try again..
2205 if (unlikely(running
)) {
2211 * It's not enough that it's not actively running,
2212 * it must be off the runqueue _entirely_, and not
2215 * So if it was still runnable (but just not actively
2216 * running right now), it's preempted, and we should
2217 * yield - it could be a while.
2219 if (unlikely(on_rq
)) {
2220 schedule_timeout_uninterruptible(1);
2225 * Ahh, all good. It wasn't running, and it wasn't
2226 * runnable, which means that it will never become
2227 * running in the future either. We're all done!
2236 * kick_process - kick a running thread to enter/exit the kernel
2237 * @p: the to-be-kicked thread
2239 * Cause a process which is running on another CPU to enter
2240 * kernel-mode, without any delay. (to get signals handled.)
2242 * NOTE: this function doesnt have to take the runqueue lock,
2243 * because all it wants to ensure is that the remote task enters
2244 * the kernel. If the IPI races and the task has been migrated
2245 * to another CPU then no harm is done and the purpose has been
2248 void kick_process(struct task_struct
*p
)
2254 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2255 smp_send_reschedule(cpu
);
2258 EXPORT_SYMBOL_GPL(kick_process
);
2259 #endif /* CONFIG_SMP */
2262 * task_oncpu_function_call - call a function on the cpu on which a task runs
2263 * @p: the task to evaluate
2264 * @func: the function to be called
2265 * @info: the function call argument
2267 * Calls the function @func when the task is currently running. This might
2268 * be on the current CPU, which just calls the function directly
2270 void task_oncpu_function_call(struct task_struct
*p
,
2271 void (*func
) (void *info
), void *info
)
2278 smp_call_function_single(cpu
, func
, info
, 1);
2284 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2286 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2289 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2291 /* Look for allowed, online CPU in same node. */
2292 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2293 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2296 /* Any allowed, online CPU? */
2297 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2298 if (dest_cpu
< nr_cpu_ids
)
2301 /* No more Mr. Nice Guy. */
2302 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2303 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2305 * Don't tell them about moving exiting tasks or
2306 * kernel threads (both mm NULL), since they never
2309 if (p
->mm
&& printk_ratelimit()) {
2310 printk(KERN_INFO
"process %d (%s) no "
2311 "longer affine to cpu%d\n",
2312 task_pid_nr(p
), p
->comm
, cpu
);
2320 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2323 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2325 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2328 * In order not to call set_task_cpu() on a blocking task we need
2329 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2332 * Since this is common to all placement strategies, this lives here.
2334 * [ this allows ->select_task() to simply return task_cpu(p) and
2335 * not worry about this generic constraint ]
2337 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2339 cpu
= select_fallback_rq(task_cpu(p
), p
);
2346 * try_to_wake_up - wake up a thread
2347 * @p: the to-be-woken-up thread
2348 * @state: the mask of task states that can be woken
2349 * @sync: do a synchronous wakeup?
2351 * Put it on the run-queue if it's not already there. The "current"
2352 * thread is always on the run-queue (except when the actual
2353 * re-schedule is in progress), and as such you're allowed to do
2354 * the simpler "current->state = TASK_RUNNING" to mark yourself
2355 * runnable without the overhead of this.
2357 * returns failure only if the task is already active.
2359 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2362 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2363 unsigned long flags
;
2366 this_cpu
= get_cpu();
2369 rq
= task_rq_lock(p
, &flags
);
2370 if (!(p
->state
& state
))
2380 if (unlikely(task_running(rq
, p
)))
2384 * In order to handle concurrent wakeups and release the rq->lock
2385 * we put the task in TASK_WAKING state.
2387 * First fix up the nr_uninterruptible count:
2389 if (task_contributes_to_load(p
))
2390 rq
->nr_uninterruptible
--;
2391 p
->state
= TASK_WAKING
;
2393 if (p
->sched_class
->task_waking
)
2394 p
->sched_class
->task_waking(rq
, p
);
2396 __task_rq_unlock(rq
);
2398 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2399 if (cpu
!= orig_cpu
) {
2401 * Since we migrate the task without holding any rq->lock,
2402 * we need to be careful with task_rq_lock(), since that
2403 * might end up locking an invalid rq.
2405 set_task_cpu(p
, cpu
);
2409 raw_spin_lock(&rq
->lock
);
2412 * We migrated the task without holding either rq->lock, however
2413 * since the task is not on the task list itself, nobody else
2414 * will try and migrate the task, hence the rq should match the
2415 * cpu we just moved it to.
2417 WARN_ON(task_cpu(p
) != cpu
);
2418 WARN_ON(p
->state
!= TASK_WAKING
);
2420 #ifdef CONFIG_SCHEDSTATS
2421 schedstat_inc(rq
, ttwu_count
);
2422 if (cpu
== this_cpu
)
2423 schedstat_inc(rq
, ttwu_local
);
2425 struct sched_domain
*sd
;
2426 for_each_domain(this_cpu
, sd
) {
2427 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2428 schedstat_inc(sd
, ttwu_wake_remote
);
2433 #endif /* CONFIG_SCHEDSTATS */
2436 #endif /* CONFIG_SMP */
2437 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2438 if (wake_flags
& WF_SYNC
)
2439 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2440 if (orig_cpu
!= cpu
)
2441 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2442 if (cpu
== this_cpu
)
2443 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2445 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2446 activate_task(rq
, p
, 1);
2450 trace_sched_wakeup(rq
, p
, success
);
2451 check_preempt_curr(rq
, p
, wake_flags
);
2453 p
->state
= TASK_RUNNING
;
2455 if (p
->sched_class
->task_woken
)
2456 p
->sched_class
->task_woken(rq
, p
);
2458 if (unlikely(rq
->idle_stamp
)) {
2459 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2460 u64 max
= 2*sysctl_sched_migration_cost
;
2465 update_avg(&rq
->avg_idle
, delta
);
2470 task_rq_unlock(rq
, &flags
);
2477 * wake_up_process - Wake up a specific process
2478 * @p: The process to be woken up.
2480 * Attempt to wake up the nominated process and move it to the set of runnable
2481 * processes. Returns 1 if the process was woken up, 0 if it was already
2484 * It may be assumed that this function implies a write memory barrier before
2485 * changing the task state if and only if any tasks are woken up.
2487 int wake_up_process(struct task_struct
*p
)
2489 return try_to_wake_up(p
, TASK_ALL
, 0);
2491 EXPORT_SYMBOL(wake_up_process
);
2493 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2495 return try_to_wake_up(p
, state
, 0);
2499 * Perform scheduler related setup for a newly forked process p.
2500 * p is forked by current.
2502 * __sched_fork() is basic setup used by init_idle() too:
2504 static void __sched_fork(struct task_struct
*p
)
2506 p
->se
.exec_start
= 0;
2507 p
->se
.sum_exec_runtime
= 0;
2508 p
->se
.prev_sum_exec_runtime
= 0;
2509 p
->se
.nr_migrations
= 0;
2511 #ifdef CONFIG_SCHEDSTATS
2512 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2515 INIT_LIST_HEAD(&p
->rt
.run_list
);
2517 INIT_LIST_HEAD(&p
->se
.group_node
);
2519 #ifdef CONFIG_PREEMPT_NOTIFIERS
2520 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2525 * fork()/clone()-time setup:
2527 void sched_fork(struct task_struct
*p
, int clone_flags
)
2529 int cpu
= get_cpu();
2533 * We mark the process as waking here. This guarantees that
2534 * nobody will actually run it, and a signal or other external
2535 * event cannot wake it up and insert it on the runqueue either.
2537 p
->state
= TASK_WAKING
;
2540 * Revert to default priority/policy on fork if requested.
2542 if (unlikely(p
->sched_reset_on_fork
)) {
2543 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2544 p
->policy
= SCHED_NORMAL
;
2545 p
->normal_prio
= p
->static_prio
;
2548 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2549 p
->static_prio
= NICE_TO_PRIO(0);
2550 p
->normal_prio
= p
->static_prio
;
2555 * We don't need the reset flag anymore after the fork. It has
2556 * fulfilled its duty:
2558 p
->sched_reset_on_fork
= 0;
2562 * Make sure we do not leak PI boosting priority to the child.
2564 p
->prio
= current
->normal_prio
;
2566 if (!rt_prio(p
->prio
))
2567 p
->sched_class
= &fair_sched_class
;
2569 if (p
->sched_class
->task_fork
)
2570 p
->sched_class
->task_fork(p
);
2572 set_task_cpu(p
, cpu
);
2574 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2575 if (likely(sched_info_on()))
2576 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2578 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2581 #ifdef CONFIG_PREEMPT
2582 /* Want to start with kernel preemption disabled. */
2583 task_thread_info(p
)->preempt_count
= 1;
2585 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2591 * wake_up_new_task - wake up a newly created task for the first time.
2593 * This function will do some initial scheduler statistics housekeeping
2594 * that must be done for every newly created context, then puts the task
2595 * on the runqueue and wakes it.
2597 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2599 unsigned long flags
;
2601 int cpu __maybe_unused
= get_cpu();
2605 * Fork balancing, do it here and not earlier because:
2606 * - cpus_allowed can change in the fork path
2607 * - any previously selected cpu might disappear through hotplug
2609 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2610 * ->cpus_allowed is stable, we have preemption disabled, meaning
2611 * cpu_online_mask is stable.
2613 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2614 set_task_cpu(p
, cpu
);
2618 * Since the task is not on the rq and we still have TASK_WAKING set
2619 * nobody else will migrate this task.
2622 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2624 BUG_ON(p
->state
!= TASK_WAKING
);
2625 p
->state
= TASK_RUNNING
;
2626 activate_task(rq
, p
, 0);
2627 trace_sched_wakeup_new(rq
, p
, 1);
2628 check_preempt_curr(rq
, p
, WF_FORK
);
2630 if (p
->sched_class
->task_woken
)
2631 p
->sched_class
->task_woken(rq
, p
);
2633 task_rq_unlock(rq
, &flags
);
2637 #ifdef CONFIG_PREEMPT_NOTIFIERS
2640 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2641 * @notifier: notifier struct to register
2643 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2645 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2647 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2650 * preempt_notifier_unregister - no longer interested in preemption notifications
2651 * @notifier: notifier struct to unregister
2653 * This is safe to call from within a preemption notifier.
2655 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2657 hlist_del(¬ifier
->link
);
2659 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2661 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2663 struct preempt_notifier
*notifier
;
2664 struct hlist_node
*node
;
2666 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2667 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2671 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2672 struct task_struct
*next
)
2674 struct preempt_notifier
*notifier
;
2675 struct hlist_node
*node
;
2677 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2678 notifier
->ops
->sched_out(notifier
, next
);
2681 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2683 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2688 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2689 struct task_struct
*next
)
2693 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2696 * prepare_task_switch - prepare to switch tasks
2697 * @rq: the runqueue preparing to switch
2698 * @prev: the current task that is being switched out
2699 * @next: the task we are going to switch to.
2701 * This is called with the rq lock held and interrupts off. It must
2702 * be paired with a subsequent finish_task_switch after the context
2705 * prepare_task_switch sets up locking and calls architecture specific
2709 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2710 struct task_struct
*next
)
2712 fire_sched_out_preempt_notifiers(prev
, next
);
2713 prepare_lock_switch(rq
, next
);
2714 prepare_arch_switch(next
);
2718 * finish_task_switch - clean up after a task-switch
2719 * @rq: runqueue associated with task-switch
2720 * @prev: the thread we just switched away from.
2722 * finish_task_switch must be called after the context switch, paired
2723 * with a prepare_task_switch call before the context switch.
2724 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2725 * and do any other architecture-specific cleanup actions.
2727 * Note that we may have delayed dropping an mm in context_switch(). If
2728 * so, we finish that here outside of the runqueue lock. (Doing it
2729 * with the lock held can cause deadlocks; see schedule() for
2732 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2733 __releases(rq
->lock
)
2735 struct mm_struct
*mm
= rq
->prev_mm
;
2741 * A task struct has one reference for the use as "current".
2742 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2743 * schedule one last time. The schedule call will never return, and
2744 * the scheduled task must drop that reference.
2745 * The test for TASK_DEAD must occur while the runqueue locks are
2746 * still held, otherwise prev could be scheduled on another cpu, die
2747 * there before we look at prev->state, and then the reference would
2749 * Manfred Spraul <manfred@colorfullife.com>
2751 prev_state
= prev
->state
;
2752 finish_arch_switch(prev
);
2753 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2754 local_irq_disable();
2755 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2756 perf_event_task_sched_in(current
);
2757 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2759 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2760 finish_lock_switch(rq
, prev
);
2762 fire_sched_in_preempt_notifiers(current
);
2765 if (unlikely(prev_state
== TASK_DEAD
)) {
2767 * Remove function-return probe instances associated with this
2768 * task and put them back on the free list.
2770 kprobe_flush_task(prev
);
2771 put_task_struct(prev
);
2777 /* assumes rq->lock is held */
2778 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2780 if (prev
->sched_class
->pre_schedule
)
2781 prev
->sched_class
->pre_schedule(rq
, prev
);
2784 /* rq->lock is NOT held, but preemption is disabled */
2785 static inline void post_schedule(struct rq
*rq
)
2787 if (rq
->post_schedule
) {
2788 unsigned long flags
;
2790 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2791 if (rq
->curr
->sched_class
->post_schedule
)
2792 rq
->curr
->sched_class
->post_schedule(rq
);
2793 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2795 rq
->post_schedule
= 0;
2801 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2805 static inline void post_schedule(struct rq
*rq
)
2812 * schedule_tail - first thing a freshly forked thread must call.
2813 * @prev: the thread we just switched away from.
2815 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2816 __releases(rq
->lock
)
2818 struct rq
*rq
= this_rq();
2820 finish_task_switch(rq
, prev
);
2823 * FIXME: do we need to worry about rq being invalidated by the
2828 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2829 /* In this case, finish_task_switch does not reenable preemption */
2832 if (current
->set_child_tid
)
2833 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2837 * context_switch - switch to the new MM and the new
2838 * thread's register state.
2841 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2842 struct task_struct
*next
)
2844 struct mm_struct
*mm
, *oldmm
;
2846 prepare_task_switch(rq
, prev
, next
);
2847 trace_sched_switch(rq
, prev
, next
);
2849 oldmm
= prev
->active_mm
;
2851 * For paravirt, this is coupled with an exit in switch_to to
2852 * combine the page table reload and the switch backend into
2855 arch_start_context_switch(prev
);
2858 next
->active_mm
= oldmm
;
2859 atomic_inc(&oldmm
->mm_count
);
2860 enter_lazy_tlb(oldmm
, next
);
2862 switch_mm(oldmm
, mm
, next
);
2864 if (likely(!prev
->mm
)) {
2865 prev
->active_mm
= NULL
;
2866 rq
->prev_mm
= oldmm
;
2869 * Since the runqueue lock will be released by the next
2870 * task (which is an invalid locking op but in the case
2871 * of the scheduler it's an obvious special-case), so we
2872 * do an early lockdep release here:
2874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2875 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2878 /* Here we just switch the register state and the stack. */
2879 switch_to(prev
, next
, prev
);
2883 * this_rq must be evaluated again because prev may have moved
2884 * CPUs since it called schedule(), thus the 'rq' on its stack
2885 * frame will be invalid.
2887 finish_task_switch(this_rq(), prev
);
2891 * nr_running, nr_uninterruptible and nr_context_switches:
2893 * externally visible scheduler statistics: current number of runnable
2894 * threads, current number of uninterruptible-sleeping threads, total
2895 * number of context switches performed since bootup.
2897 unsigned long nr_running(void)
2899 unsigned long i
, sum
= 0;
2901 for_each_online_cpu(i
)
2902 sum
+= cpu_rq(i
)->nr_running
;
2907 unsigned long nr_uninterruptible(void)
2909 unsigned long i
, sum
= 0;
2911 for_each_possible_cpu(i
)
2912 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2915 * Since we read the counters lockless, it might be slightly
2916 * inaccurate. Do not allow it to go below zero though:
2918 if (unlikely((long)sum
< 0))
2924 unsigned long long nr_context_switches(void)
2927 unsigned long long sum
= 0;
2929 for_each_possible_cpu(i
)
2930 sum
+= cpu_rq(i
)->nr_switches
;
2935 unsigned long nr_iowait(void)
2937 unsigned long i
, sum
= 0;
2939 for_each_possible_cpu(i
)
2940 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2945 unsigned long nr_iowait_cpu(void)
2947 struct rq
*this = this_rq();
2948 return atomic_read(&this->nr_iowait
);
2951 unsigned long this_cpu_load(void)
2953 struct rq
*this = this_rq();
2954 return this->cpu_load
[0];
2958 /* Variables and functions for calc_load */
2959 static atomic_long_t calc_load_tasks
;
2960 static unsigned long calc_load_update
;
2961 unsigned long avenrun
[3];
2962 EXPORT_SYMBOL(avenrun
);
2965 * get_avenrun - get the load average array
2966 * @loads: pointer to dest load array
2967 * @offset: offset to add
2968 * @shift: shift count to shift the result left
2970 * These values are estimates at best, so no need for locking.
2972 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2974 loads
[0] = (avenrun
[0] + offset
) << shift
;
2975 loads
[1] = (avenrun
[1] + offset
) << shift
;
2976 loads
[2] = (avenrun
[2] + offset
) << shift
;
2979 static unsigned long
2980 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2983 load
+= active
* (FIXED_1
- exp
);
2984 return load
>> FSHIFT
;
2988 * calc_load - update the avenrun load estimates 10 ticks after the
2989 * CPUs have updated calc_load_tasks.
2991 void calc_global_load(void)
2993 unsigned long upd
= calc_load_update
+ 10;
2996 if (time_before(jiffies
, upd
))
2999 active
= atomic_long_read(&calc_load_tasks
);
3000 active
= active
> 0 ? active
* FIXED_1
: 0;
3002 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3003 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3004 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3006 calc_load_update
+= LOAD_FREQ
;
3010 * Either called from update_cpu_load() or from a cpu going idle
3012 static void calc_load_account_active(struct rq
*this_rq
)
3014 long nr_active
, delta
;
3016 nr_active
= this_rq
->nr_running
;
3017 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3019 if (nr_active
!= this_rq
->calc_load_active
) {
3020 delta
= nr_active
- this_rq
->calc_load_active
;
3021 this_rq
->calc_load_active
= nr_active
;
3022 atomic_long_add(delta
, &calc_load_tasks
);
3027 * Update rq->cpu_load[] statistics. This function is usually called every
3028 * scheduler tick (TICK_NSEC).
3030 static void update_cpu_load(struct rq
*this_rq
)
3032 unsigned long this_load
= this_rq
->load
.weight
;
3035 this_rq
->nr_load_updates
++;
3037 /* Update our load: */
3038 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3039 unsigned long old_load
, new_load
;
3041 /* scale is effectively 1 << i now, and >> i divides by scale */
3043 old_load
= this_rq
->cpu_load
[i
];
3044 new_load
= this_load
;
3046 * Round up the averaging division if load is increasing. This
3047 * prevents us from getting stuck on 9 if the load is 10, for
3050 if (new_load
> old_load
)
3051 new_load
+= scale
-1;
3052 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3055 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3056 this_rq
->calc_load_update
+= LOAD_FREQ
;
3057 calc_load_account_active(this_rq
);
3064 * sched_exec - execve() is a valuable balancing opportunity, because at
3065 * this point the task has the smallest effective memory and cache footprint.
3067 void sched_exec(void)
3069 struct task_struct
*p
= current
;
3070 struct migration_req req
;
3071 int dest_cpu
, this_cpu
;
3072 unsigned long flags
;
3075 this_cpu
= get_cpu();
3076 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3077 if (dest_cpu
== this_cpu
) {
3082 rq
= task_rq_lock(p
, &flags
);
3085 * select_task_rq() can race against ->cpus_allowed
3087 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3088 likely(cpu_active(dest_cpu
)) &&
3089 migrate_task(p
, dest_cpu
, &req
)) {
3090 /* Need to wait for migration thread (might exit: take ref). */
3091 struct task_struct
*mt
= rq
->migration_thread
;
3093 get_task_struct(mt
);
3094 task_rq_unlock(rq
, &flags
);
3095 wake_up_process(mt
);
3096 put_task_struct(mt
);
3097 wait_for_completion(&req
.done
);
3101 task_rq_unlock(rq
, &flags
);
3106 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3108 EXPORT_PER_CPU_SYMBOL(kstat
);
3111 * Return any ns on the sched_clock that have not yet been accounted in
3112 * @p in case that task is currently running.
3114 * Called with task_rq_lock() held on @rq.
3116 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3120 if (task_current(rq
, p
)) {
3121 update_rq_clock(rq
);
3122 ns
= rq
->clock
- p
->se
.exec_start
;
3130 unsigned long long task_delta_exec(struct task_struct
*p
)
3132 unsigned long flags
;
3136 rq
= task_rq_lock(p
, &flags
);
3137 ns
= do_task_delta_exec(p
, rq
);
3138 task_rq_unlock(rq
, &flags
);
3144 * Return accounted runtime for the task.
3145 * In case the task is currently running, return the runtime plus current's
3146 * pending runtime that have not been accounted yet.
3148 unsigned long long task_sched_runtime(struct task_struct
*p
)
3150 unsigned long flags
;
3154 rq
= task_rq_lock(p
, &flags
);
3155 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3156 task_rq_unlock(rq
, &flags
);
3162 * Return sum_exec_runtime for the thread group.
3163 * In case the task is currently running, return the sum plus current's
3164 * pending runtime that have not been accounted yet.
3166 * Note that the thread group might have other running tasks as well,
3167 * so the return value not includes other pending runtime that other
3168 * running tasks might have.
3170 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3172 struct task_cputime totals
;
3173 unsigned long flags
;
3177 rq
= task_rq_lock(p
, &flags
);
3178 thread_group_cputime(p
, &totals
);
3179 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3180 task_rq_unlock(rq
, &flags
);
3186 * Account user cpu time to a process.
3187 * @p: the process that the cpu time gets accounted to
3188 * @cputime: the cpu time spent in user space since the last update
3189 * @cputime_scaled: cputime scaled by cpu frequency
3191 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3192 cputime_t cputime_scaled
)
3194 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3197 /* Add user time to process. */
3198 p
->utime
= cputime_add(p
->utime
, cputime
);
3199 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3200 account_group_user_time(p
, cputime
);
3202 /* Add user time to cpustat. */
3203 tmp
= cputime_to_cputime64(cputime
);
3204 if (TASK_NICE(p
) > 0)
3205 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3207 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3209 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3210 /* Account for user time used */
3211 acct_update_integrals(p
);
3215 * Account guest cpu time to a process.
3216 * @p: the process that the cpu time gets accounted to
3217 * @cputime: the cpu time spent in virtual machine since the last update
3218 * @cputime_scaled: cputime scaled by cpu frequency
3220 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3221 cputime_t cputime_scaled
)
3224 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3226 tmp
= cputime_to_cputime64(cputime
);
3228 /* Add guest time to process. */
3229 p
->utime
= cputime_add(p
->utime
, cputime
);
3230 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3231 account_group_user_time(p
, cputime
);
3232 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3234 /* Add guest time to cpustat. */
3235 if (TASK_NICE(p
) > 0) {
3236 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3237 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3239 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3240 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3245 * Account system cpu time to a process.
3246 * @p: the process that the cpu time gets accounted to
3247 * @hardirq_offset: the offset to subtract from hardirq_count()
3248 * @cputime: the cpu time spent in kernel space since the last update
3249 * @cputime_scaled: cputime scaled by cpu frequency
3251 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3252 cputime_t cputime
, cputime_t cputime_scaled
)
3254 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3257 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3258 account_guest_time(p
, cputime
, cputime_scaled
);
3262 /* Add system time to process. */
3263 p
->stime
= cputime_add(p
->stime
, cputime
);
3264 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3265 account_group_system_time(p
, cputime
);
3267 /* Add system time to cpustat. */
3268 tmp
= cputime_to_cputime64(cputime
);
3269 if (hardirq_count() - hardirq_offset
)
3270 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3271 else if (softirq_count())
3272 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3274 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3276 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3278 /* Account for system time used */
3279 acct_update_integrals(p
);
3283 * Account for involuntary wait time.
3284 * @steal: the cpu time spent in involuntary wait
3286 void account_steal_time(cputime_t cputime
)
3288 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3289 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3291 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3295 * Account for idle time.
3296 * @cputime: the cpu time spent in idle wait
3298 void account_idle_time(cputime_t cputime
)
3300 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3301 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3302 struct rq
*rq
= this_rq();
3304 if (atomic_read(&rq
->nr_iowait
) > 0)
3305 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3307 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3310 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3313 * Account a single tick of cpu time.
3314 * @p: the process that the cpu time gets accounted to
3315 * @user_tick: indicates if the tick is a user or a system tick
3317 void account_process_tick(struct task_struct
*p
, int user_tick
)
3319 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3320 struct rq
*rq
= this_rq();
3323 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3324 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3325 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3328 account_idle_time(cputime_one_jiffy
);
3332 * Account multiple ticks of steal time.
3333 * @p: the process from which the cpu time has been stolen
3334 * @ticks: number of stolen ticks
3336 void account_steal_ticks(unsigned long ticks
)
3338 account_steal_time(jiffies_to_cputime(ticks
));
3342 * Account multiple ticks of idle time.
3343 * @ticks: number of stolen ticks
3345 void account_idle_ticks(unsigned long ticks
)
3347 account_idle_time(jiffies_to_cputime(ticks
));
3353 * Use precise platform statistics if available:
3355 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3356 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3362 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3364 struct task_cputime cputime
;
3366 thread_group_cputime(p
, &cputime
);
3368 *ut
= cputime
.utime
;
3369 *st
= cputime
.stime
;
3373 #ifndef nsecs_to_cputime
3374 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3377 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3379 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3382 * Use CFS's precise accounting:
3384 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3389 temp
= (u64
)(rtime
* utime
);
3390 do_div(temp
, total
);
3391 utime
= (cputime_t
)temp
;
3396 * Compare with previous values, to keep monotonicity:
3398 p
->prev_utime
= max(p
->prev_utime
, utime
);
3399 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3401 *ut
= p
->prev_utime
;
3402 *st
= p
->prev_stime
;
3406 * Must be called with siglock held.
3408 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3410 struct signal_struct
*sig
= p
->signal
;
3411 struct task_cputime cputime
;
3412 cputime_t rtime
, utime
, total
;
3414 thread_group_cputime(p
, &cputime
);
3416 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3417 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3422 temp
= (u64
)(rtime
* cputime
.utime
);
3423 do_div(temp
, total
);
3424 utime
= (cputime_t
)temp
;
3428 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3429 sig
->prev_stime
= max(sig
->prev_stime
,
3430 cputime_sub(rtime
, sig
->prev_utime
));
3432 *ut
= sig
->prev_utime
;
3433 *st
= sig
->prev_stime
;
3438 * This function gets called by the timer code, with HZ frequency.
3439 * We call it with interrupts disabled.
3441 * It also gets called by the fork code, when changing the parent's
3444 void scheduler_tick(void)
3446 int cpu
= smp_processor_id();
3447 struct rq
*rq
= cpu_rq(cpu
);
3448 struct task_struct
*curr
= rq
->curr
;
3452 raw_spin_lock(&rq
->lock
);
3453 update_rq_clock(rq
);
3454 update_cpu_load(rq
);
3455 curr
->sched_class
->task_tick(rq
, curr
, 0);
3456 raw_spin_unlock(&rq
->lock
);
3458 perf_event_task_tick(curr
);
3461 rq
->idle_at_tick
= idle_cpu(cpu
);
3462 trigger_load_balance(rq
, cpu
);
3466 notrace
unsigned long get_parent_ip(unsigned long addr
)
3468 if (in_lock_functions(addr
)) {
3469 addr
= CALLER_ADDR2
;
3470 if (in_lock_functions(addr
))
3471 addr
= CALLER_ADDR3
;
3476 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3477 defined(CONFIG_PREEMPT_TRACER))
3479 void __kprobes
add_preempt_count(int val
)
3481 #ifdef CONFIG_DEBUG_PREEMPT
3485 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3488 preempt_count() += val
;
3489 #ifdef CONFIG_DEBUG_PREEMPT
3491 * Spinlock count overflowing soon?
3493 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3496 if (preempt_count() == val
)
3497 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3499 EXPORT_SYMBOL(add_preempt_count
);
3501 void __kprobes
sub_preempt_count(int val
)
3503 #ifdef CONFIG_DEBUG_PREEMPT
3507 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3510 * Is the spinlock portion underflowing?
3512 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3513 !(preempt_count() & PREEMPT_MASK
)))
3517 if (preempt_count() == val
)
3518 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3519 preempt_count() -= val
;
3521 EXPORT_SYMBOL(sub_preempt_count
);
3526 * Print scheduling while atomic bug:
3528 static noinline
void __schedule_bug(struct task_struct
*prev
)
3530 struct pt_regs
*regs
= get_irq_regs();
3532 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3533 prev
->comm
, prev
->pid
, preempt_count());
3535 debug_show_held_locks(prev
);
3537 if (irqs_disabled())
3538 print_irqtrace_events(prev
);
3547 * Various schedule()-time debugging checks and statistics:
3549 static inline void schedule_debug(struct task_struct
*prev
)
3552 * Test if we are atomic. Since do_exit() needs to call into
3553 * schedule() atomically, we ignore that path for now.
3554 * Otherwise, whine if we are scheduling when we should not be.
3556 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3557 __schedule_bug(prev
);
3559 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3561 schedstat_inc(this_rq(), sched_count
);
3562 #ifdef CONFIG_SCHEDSTATS
3563 if (unlikely(prev
->lock_depth
>= 0)) {
3564 schedstat_inc(this_rq(), bkl_count
);
3565 schedstat_inc(prev
, sched_info
.bkl_count
);
3570 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3573 update_rq_clock(rq
);
3574 rq
->skip_clock_update
= 0;
3575 prev
->sched_class
->put_prev_task(rq
, prev
);
3579 * Pick up the highest-prio task:
3581 static inline struct task_struct
*
3582 pick_next_task(struct rq
*rq
)
3584 const struct sched_class
*class;
3585 struct task_struct
*p
;
3588 * Optimization: we know that if all tasks are in
3589 * the fair class we can call that function directly:
3591 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3592 p
= fair_sched_class
.pick_next_task(rq
);
3597 class = sched_class_highest
;
3599 p
= class->pick_next_task(rq
);
3603 * Will never be NULL as the idle class always
3604 * returns a non-NULL p:
3606 class = class->next
;
3611 * schedule() is the main scheduler function.
3613 asmlinkage
void __sched
schedule(void)
3615 struct task_struct
*prev
, *next
;
3616 unsigned long *switch_count
;
3622 cpu
= smp_processor_id();
3626 switch_count
= &prev
->nivcsw
;
3628 release_kernel_lock(prev
);
3629 need_resched_nonpreemptible
:
3631 schedule_debug(prev
);
3633 if (sched_feat(HRTICK
))
3636 raw_spin_lock_irq(&rq
->lock
);
3637 clear_tsk_need_resched(prev
);
3639 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3640 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3641 prev
->state
= TASK_RUNNING
;
3643 deactivate_task(rq
, prev
, 1);
3644 switch_count
= &prev
->nvcsw
;
3647 pre_schedule(rq
, prev
);
3649 if (unlikely(!rq
->nr_running
))
3650 idle_balance(cpu
, rq
);
3652 put_prev_task(rq
, prev
);
3653 next
= pick_next_task(rq
);
3655 if (likely(prev
!= next
)) {
3656 sched_info_switch(prev
, next
);
3657 perf_event_task_sched_out(prev
, next
);
3663 context_switch(rq
, prev
, next
); /* unlocks the rq */
3665 * the context switch might have flipped the stack from under
3666 * us, hence refresh the local variables.
3668 cpu
= smp_processor_id();
3671 raw_spin_unlock_irq(&rq
->lock
);
3675 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3677 switch_count
= &prev
->nivcsw
;
3678 goto need_resched_nonpreemptible
;
3681 preempt_enable_no_resched();
3685 EXPORT_SYMBOL(schedule
);
3687 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3689 * Look out! "owner" is an entirely speculative pointer
3690 * access and not reliable.
3692 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3697 if (!sched_feat(OWNER_SPIN
))
3700 #ifdef CONFIG_DEBUG_PAGEALLOC
3702 * Need to access the cpu field knowing that
3703 * DEBUG_PAGEALLOC could have unmapped it if
3704 * the mutex owner just released it and exited.
3706 if (probe_kernel_address(&owner
->cpu
, cpu
))
3713 * Even if the access succeeded (likely case),
3714 * the cpu field may no longer be valid.
3716 if (cpu
>= nr_cpumask_bits
)
3720 * We need to validate that we can do a
3721 * get_cpu() and that we have the percpu area.
3723 if (!cpu_online(cpu
))
3730 * Owner changed, break to re-assess state.
3732 if (lock
->owner
!= owner
)
3736 * Is that owner really running on that cpu?
3738 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3748 #ifdef CONFIG_PREEMPT
3750 * this is the entry point to schedule() from in-kernel preemption
3751 * off of preempt_enable. Kernel preemptions off return from interrupt
3752 * occur there and call schedule directly.
3754 asmlinkage
void __sched
preempt_schedule(void)
3756 struct thread_info
*ti
= current_thread_info();
3759 * If there is a non-zero preempt_count or interrupts are disabled,
3760 * we do not want to preempt the current task. Just return..
3762 if (likely(ti
->preempt_count
|| irqs_disabled()))
3766 add_preempt_count(PREEMPT_ACTIVE
);
3768 sub_preempt_count(PREEMPT_ACTIVE
);
3771 * Check again in case we missed a preemption opportunity
3772 * between schedule and now.
3775 } while (need_resched());
3777 EXPORT_SYMBOL(preempt_schedule
);
3780 * this is the entry point to schedule() from kernel preemption
3781 * off of irq context.
3782 * Note, that this is called and return with irqs disabled. This will
3783 * protect us against recursive calling from irq.
3785 asmlinkage
void __sched
preempt_schedule_irq(void)
3787 struct thread_info
*ti
= current_thread_info();
3789 /* Catch callers which need to be fixed */
3790 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3793 add_preempt_count(PREEMPT_ACTIVE
);
3796 local_irq_disable();
3797 sub_preempt_count(PREEMPT_ACTIVE
);
3800 * Check again in case we missed a preemption opportunity
3801 * between schedule and now.
3804 } while (need_resched());
3807 #endif /* CONFIG_PREEMPT */
3809 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3812 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3814 EXPORT_SYMBOL(default_wake_function
);
3817 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3818 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3819 * number) then we wake all the non-exclusive tasks and one exclusive task.
3821 * There are circumstances in which we can try to wake a task which has already
3822 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3823 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3825 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3826 int nr_exclusive
, int wake_flags
, void *key
)
3828 wait_queue_t
*curr
, *next
;
3830 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3831 unsigned flags
= curr
->flags
;
3833 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3834 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3840 * __wake_up - wake up threads blocked on a waitqueue.
3842 * @mode: which threads
3843 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3844 * @key: is directly passed to the wakeup function
3846 * It may be assumed that this function implies a write memory barrier before
3847 * changing the task state if and only if any tasks are woken up.
3849 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3850 int nr_exclusive
, void *key
)
3852 unsigned long flags
;
3854 spin_lock_irqsave(&q
->lock
, flags
);
3855 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3856 spin_unlock_irqrestore(&q
->lock
, flags
);
3858 EXPORT_SYMBOL(__wake_up
);
3861 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3863 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3865 __wake_up_common(q
, mode
, 1, 0, NULL
);
3868 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3870 __wake_up_common(q
, mode
, 1, 0, key
);
3874 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3876 * @mode: which threads
3877 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3878 * @key: opaque value to be passed to wakeup targets
3880 * The sync wakeup differs that the waker knows that it will schedule
3881 * away soon, so while the target thread will be woken up, it will not
3882 * be migrated to another CPU - ie. the two threads are 'synchronized'
3883 * with each other. This can prevent needless bouncing between CPUs.
3885 * On UP it can prevent extra preemption.
3887 * It may be assumed that this function implies a write memory barrier before
3888 * changing the task state if and only if any tasks are woken up.
3890 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3891 int nr_exclusive
, void *key
)
3893 unsigned long flags
;
3894 int wake_flags
= WF_SYNC
;
3899 if (unlikely(!nr_exclusive
))
3902 spin_lock_irqsave(&q
->lock
, flags
);
3903 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3904 spin_unlock_irqrestore(&q
->lock
, flags
);
3906 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3909 * __wake_up_sync - see __wake_up_sync_key()
3911 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3913 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3915 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3918 * complete: - signals a single thread waiting on this completion
3919 * @x: holds the state of this particular completion
3921 * This will wake up a single thread waiting on this completion. Threads will be
3922 * awakened in the same order in which they were queued.
3924 * See also complete_all(), wait_for_completion() and related routines.
3926 * It may be assumed that this function implies a write memory barrier before
3927 * changing the task state if and only if any tasks are woken up.
3929 void complete(struct completion
*x
)
3931 unsigned long flags
;
3933 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3935 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3936 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3938 EXPORT_SYMBOL(complete
);
3941 * complete_all: - signals all threads waiting on this completion
3942 * @x: holds the state of this particular completion
3944 * This will wake up all threads waiting on this particular completion event.
3946 * It may be assumed that this function implies a write memory barrier before
3947 * changing the task state if and only if any tasks are woken up.
3949 void complete_all(struct completion
*x
)
3951 unsigned long flags
;
3953 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3954 x
->done
+= UINT_MAX
/2;
3955 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3956 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3958 EXPORT_SYMBOL(complete_all
);
3960 static inline long __sched
3961 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3964 DECLARE_WAITQUEUE(wait
, current
);
3966 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3967 __add_wait_queue_tail(&x
->wait
, &wait
);
3969 if (signal_pending_state(state
, current
)) {
3970 timeout
= -ERESTARTSYS
;
3973 __set_current_state(state
);
3974 spin_unlock_irq(&x
->wait
.lock
);
3975 timeout
= schedule_timeout(timeout
);
3976 spin_lock_irq(&x
->wait
.lock
);
3977 } while (!x
->done
&& timeout
);
3978 __remove_wait_queue(&x
->wait
, &wait
);
3983 return timeout
?: 1;
3987 wait_for_common(struct completion
*x
, long timeout
, int state
)
3991 spin_lock_irq(&x
->wait
.lock
);
3992 timeout
= do_wait_for_common(x
, timeout
, state
);
3993 spin_unlock_irq(&x
->wait
.lock
);
3998 * wait_for_completion: - waits for completion of a task
3999 * @x: holds the state of this particular completion
4001 * This waits to be signaled for completion of a specific task. It is NOT
4002 * interruptible and there is no timeout.
4004 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4005 * and interrupt capability. Also see complete().
4007 void __sched
wait_for_completion(struct completion
*x
)
4009 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4011 EXPORT_SYMBOL(wait_for_completion
);
4014 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4015 * @x: holds the state of this particular completion
4016 * @timeout: timeout value in jiffies
4018 * This waits for either a completion of a specific task to be signaled or for a
4019 * specified timeout to expire. The timeout is in jiffies. It is not
4022 unsigned long __sched
4023 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4025 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4027 EXPORT_SYMBOL(wait_for_completion_timeout
);
4030 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4031 * @x: holds the state of this particular completion
4033 * This waits for completion of a specific task to be signaled. It is
4036 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4038 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4039 if (t
== -ERESTARTSYS
)
4043 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4046 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4047 * @x: holds the state of this particular completion
4048 * @timeout: timeout value in jiffies
4050 * This waits for either a completion of a specific task to be signaled or for a
4051 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4053 unsigned long __sched
4054 wait_for_completion_interruptible_timeout(struct completion
*x
,
4055 unsigned long timeout
)
4057 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4059 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4062 * wait_for_completion_killable: - waits for completion of a task (killable)
4063 * @x: holds the state of this particular completion
4065 * This waits to be signaled for completion of a specific task. It can be
4066 * interrupted by a kill signal.
4068 int __sched
wait_for_completion_killable(struct completion
*x
)
4070 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4071 if (t
== -ERESTARTSYS
)
4075 EXPORT_SYMBOL(wait_for_completion_killable
);
4078 * try_wait_for_completion - try to decrement a completion without blocking
4079 * @x: completion structure
4081 * Returns: 0 if a decrement cannot be done without blocking
4082 * 1 if a decrement succeeded.
4084 * If a completion is being used as a counting completion,
4085 * attempt to decrement the counter without blocking. This
4086 * enables us to avoid waiting if the resource the completion
4087 * is protecting is not available.
4089 bool try_wait_for_completion(struct completion
*x
)
4091 unsigned long flags
;
4094 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4099 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4102 EXPORT_SYMBOL(try_wait_for_completion
);
4105 * completion_done - Test to see if a completion has any waiters
4106 * @x: completion structure
4108 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4109 * 1 if there are no waiters.
4112 bool completion_done(struct completion
*x
)
4114 unsigned long flags
;
4117 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4120 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4123 EXPORT_SYMBOL(completion_done
);
4126 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4128 unsigned long flags
;
4131 init_waitqueue_entry(&wait
, current
);
4133 __set_current_state(state
);
4135 spin_lock_irqsave(&q
->lock
, flags
);
4136 __add_wait_queue(q
, &wait
);
4137 spin_unlock(&q
->lock
);
4138 timeout
= schedule_timeout(timeout
);
4139 spin_lock_irq(&q
->lock
);
4140 __remove_wait_queue(q
, &wait
);
4141 spin_unlock_irqrestore(&q
->lock
, flags
);
4146 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4148 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4150 EXPORT_SYMBOL(interruptible_sleep_on
);
4153 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4155 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4157 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4159 void __sched
sleep_on(wait_queue_head_t
*q
)
4161 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4163 EXPORT_SYMBOL(sleep_on
);
4165 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4167 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4169 EXPORT_SYMBOL(sleep_on_timeout
);
4171 #ifdef CONFIG_RT_MUTEXES
4174 * rt_mutex_setprio - set the current priority of a task
4176 * @prio: prio value (kernel-internal form)
4178 * This function changes the 'effective' priority of a task. It does
4179 * not touch ->normal_prio like __setscheduler().
4181 * Used by the rt_mutex code to implement priority inheritance logic.
4183 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4185 unsigned long flags
;
4186 int oldprio
, on_rq
, running
;
4188 const struct sched_class
*prev_class
;
4190 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4192 rq
= task_rq_lock(p
, &flags
);
4195 prev_class
= p
->sched_class
;
4196 on_rq
= p
->se
.on_rq
;
4197 running
= task_current(rq
, p
);
4199 dequeue_task(rq
, p
, 0);
4201 p
->sched_class
->put_prev_task(rq
, p
);
4204 p
->sched_class
= &rt_sched_class
;
4206 p
->sched_class
= &fair_sched_class
;
4211 p
->sched_class
->set_curr_task(rq
);
4213 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4215 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4217 task_rq_unlock(rq
, &flags
);
4222 void set_user_nice(struct task_struct
*p
, long nice
)
4224 int old_prio
, delta
, on_rq
;
4225 unsigned long flags
;
4228 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4231 * We have to be careful, if called from sys_setpriority(),
4232 * the task might be in the middle of scheduling on another CPU.
4234 rq
= task_rq_lock(p
, &flags
);
4236 * The RT priorities are set via sched_setscheduler(), but we still
4237 * allow the 'normal' nice value to be set - but as expected
4238 * it wont have any effect on scheduling until the task is
4239 * SCHED_FIFO/SCHED_RR:
4241 if (task_has_rt_policy(p
)) {
4242 p
->static_prio
= NICE_TO_PRIO(nice
);
4245 on_rq
= p
->se
.on_rq
;
4247 dequeue_task(rq
, p
, 0);
4249 p
->static_prio
= NICE_TO_PRIO(nice
);
4252 p
->prio
= effective_prio(p
);
4253 delta
= p
->prio
- old_prio
;
4256 enqueue_task(rq
, p
, 0, false);
4258 * If the task increased its priority or is running and
4259 * lowered its priority, then reschedule its CPU:
4261 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4262 resched_task(rq
->curr
);
4265 task_rq_unlock(rq
, &flags
);
4267 EXPORT_SYMBOL(set_user_nice
);
4270 * can_nice - check if a task can reduce its nice value
4274 int can_nice(const struct task_struct
*p
, const int nice
)
4276 /* convert nice value [19,-20] to rlimit style value [1,40] */
4277 int nice_rlim
= 20 - nice
;
4279 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4280 capable(CAP_SYS_NICE
));
4283 #ifdef __ARCH_WANT_SYS_NICE
4286 * sys_nice - change the priority of the current process.
4287 * @increment: priority increment
4289 * sys_setpriority is a more generic, but much slower function that
4290 * does similar things.
4292 SYSCALL_DEFINE1(nice
, int, increment
)
4297 * Setpriority might change our priority at the same moment.
4298 * We don't have to worry. Conceptually one call occurs first
4299 * and we have a single winner.
4301 if (increment
< -40)
4306 nice
= TASK_NICE(current
) + increment
;
4312 if (increment
< 0 && !can_nice(current
, nice
))
4315 retval
= security_task_setnice(current
, nice
);
4319 set_user_nice(current
, nice
);
4326 * task_prio - return the priority value of a given task.
4327 * @p: the task in question.
4329 * This is the priority value as seen by users in /proc.
4330 * RT tasks are offset by -200. Normal tasks are centered
4331 * around 0, value goes from -16 to +15.
4333 int task_prio(const struct task_struct
*p
)
4335 return p
->prio
- MAX_RT_PRIO
;
4339 * task_nice - return the nice value of a given task.
4340 * @p: the task in question.
4342 int task_nice(const struct task_struct
*p
)
4344 return TASK_NICE(p
);
4346 EXPORT_SYMBOL(task_nice
);
4349 * idle_cpu - is a given cpu idle currently?
4350 * @cpu: the processor in question.
4352 int idle_cpu(int cpu
)
4354 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4358 * idle_task - return the idle task for a given cpu.
4359 * @cpu: the processor in question.
4361 struct task_struct
*idle_task(int cpu
)
4363 return cpu_rq(cpu
)->idle
;
4367 * find_process_by_pid - find a process with a matching PID value.
4368 * @pid: the pid in question.
4370 static struct task_struct
*find_process_by_pid(pid_t pid
)
4372 return pid
? find_task_by_vpid(pid
) : current
;
4375 /* Actually do priority change: must hold rq lock. */
4377 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4379 BUG_ON(p
->se
.on_rq
);
4382 p
->rt_priority
= prio
;
4383 p
->normal_prio
= normal_prio(p
);
4384 /* we are holding p->pi_lock already */
4385 p
->prio
= rt_mutex_getprio(p
);
4386 if (rt_prio(p
->prio
))
4387 p
->sched_class
= &rt_sched_class
;
4389 p
->sched_class
= &fair_sched_class
;
4394 * check the target process has a UID that matches the current process's
4396 static bool check_same_owner(struct task_struct
*p
)
4398 const struct cred
*cred
= current_cred(), *pcred
;
4402 pcred
= __task_cred(p
);
4403 match
= (cred
->euid
== pcred
->euid
||
4404 cred
->euid
== pcred
->uid
);
4409 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4410 struct sched_param
*param
, bool user
)
4412 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4413 unsigned long flags
;
4414 const struct sched_class
*prev_class
;
4418 /* may grab non-irq protected spin_locks */
4419 BUG_ON(in_interrupt());
4421 /* double check policy once rq lock held */
4423 reset_on_fork
= p
->sched_reset_on_fork
;
4424 policy
= oldpolicy
= p
->policy
;
4426 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4427 policy
&= ~SCHED_RESET_ON_FORK
;
4429 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4430 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4431 policy
!= SCHED_IDLE
)
4436 * Valid priorities for SCHED_FIFO and SCHED_RR are
4437 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4438 * SCHED_BATCH and SCHED_IDLE is 0.
4440 if (param
->sched_priority
< 0 ||
4441 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4442 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4444 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4448 * Allow unprivileged RT tasks to decrease priority:
4450 if (user
&& !capable(CAP_SYS_NICE
)) {
4451 if (rt_policy(policy
)) {
4452 unsigned long rlim_rtprio
;
4454 if (!lock_task_sighand(p
, &flags
))
4456 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4457 unlock_task_sighand(p
, &flags
);
4459 /* can't set/change the rt policy */
4460 if (policy
!= p
->policy
&& !rlim_rtprio
)
4463 /* can't increase priority */
4464 if (param
->sched_priority
> p
->rt_priority
&&
4465 param
->sched_priority
> rlim_rtprio
)
4469 * Like positive nice levels, dont allow tasks to
4470 * move out of SCHED_IDLE either:
4472 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4475 /* can't change other user's priorities */
4476 if (!check_same_owner(p
))
4479 /* Normal users shall not reset the sched_reset_on_fork flag */
4480 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4485 #ifdef CONFIG_RT_GROUP_SCHED
4487 * Do not allow realtime tasks into groups that have no runtime
4490 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4491 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4495 retval
= security_task_setscheduler(p
, policy
, param
);
4501 * make sure no PI-waiters arrive (or leave) while we are
4502 * changing the priority of the task:
4504 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4506 * To be able to change p->policy safely, the apropriate
4507 * runqueue lock must be held.
4509 rq
= __task_rq_lock(p
);
4510 /* recheck policy now with rq lock held */
4511 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4512 policy
= oldpolicy
= -1;
4513 __task_rq_unlock(rq
);
4514 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4517 on_rq
= p
->se
.on_rq
;
4518 running
= task_current(rq
, p
);
4520 deactivate_task(rq
, p
, 0);
4522 p
->sched_class
->put_prev_task(rq
, p
);
4524 p
->sched_reset_on_fork
= reset_on_fork
;
4527 prev_class
= p
->sched_class
;
4528 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4531 p
->sched_class
->set_curr_task(rq
);
4533 activate_task(rq
, p
, 0);
4535 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4537 __task_rq_unlock(rq
);
4538 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4540 rt_mutex_adjust_pi(p
);
4546 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4547 * @p: the task in question.
4548 * @policy: new policy.
4549 * @param: structure containing the new RT priority.
4551 * NOTE that the task may be already dead.
4553 int sched_setscheduler(struct task_struct
*p
, int policy
,
4554 struct sched_param
*param
)
4556 return __sched_setscheduler(p
, policy
, param
, true);
4558 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4561 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4562 * @p: the task in question.
4563 * @policy: new policy.
4564 * @param: structure containing the new RT priority.
4566 * Just like sched_setscheduler, only don't bother checking if the
4567 * current context has permission. For example, this is needed in
4568 * stop_machine(): we create temporary high priority worker threads,
4569 * but our caller might not have that capability.
4571 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4572 struct sched_param
*param
)
4574 return __sched_setscheduler(p
, policy
, param
, false);
4578 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4580 struct sched_param lparam
;
4581 struct task_struct
*p
;
4584 if (!param
|| pid
< 0)
4586 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4591 p
= find_process_by_pid(pid
);
4593 retval
= sched_setscheduler(p
, policy
, &lparam
);
4600 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4601 * @pid: the pid in question.
4602 * @policy: new policy.
4603 * @param: structure containing the new RT priority.
4605 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4606 struct sched_param __user
*, param
)
4608 /* negative values for policy are not valid */
4612 return do_sched_setscheduler(pid
, policy
, param
);
4616 * sys_sched_setparam - set/change the RT priority of a thread
4617 * @pid: the pid in question.
4618 * @param: structure containing the new RT priority.
4620 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4622 return do_sched_setscheduler(pid
, -1, param
);
4626 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4627 * @pid: the pid in question.
4629 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4631 struct task_struct
*p
;
4639 p
= find_process_by_pid(pid
);
4641 retval
= security_task_getscheduler(p
);
4644 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4651 * sys_sched_getparam - get the RT priority of a thread
4652 * @pid: the pid in question.
4653 * @param: structure containing the RT priority.
4655 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4657 struct sched_param lp
;
4658 struct task_struct
*p
;
4661 if (!param
|| pid
< 0)
4665 p
= find_process_by_pid(pid
);
4670 retval
= security_task_getscheduler(p
);
4674 lp
.sched_priority
= p
->rt_priority
;
4678 * This one might sleep, we cannot do it with a spinlock held ...
4680 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4689 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4691 cpumask_var_t cpus_allowed
, new_mask
;
4692 struct task_struct
*p
;
4698 p
= find_process_by_pid(pid
);
4705 /* Prevent p going away */
4709 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4713 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4715 goto out_free_cpus_allowed
;
4718 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4721 retval
= security_task_setscheduler(p
, 0, NULL
);
4725 cpuset_cpus_allowed(p
, cpus_allowed
);
4726 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4728 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4731 cpuset_cpus_allowed(p
, cpus_allowed
);
4732 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4734 * We must have raced with a concurrent cpuset
4735 * update. Just reset the cpus_allowed to the
4736 * cpuset's cpus_allowed
4738 cpumask_copy(new_mask
, cpus_allowed
);
4743 free_cpumask_var(new_mask
);
4744 out_free_cpus_allowed
:
4745 free_cpumask_var(cpus_allowed
);
4752 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4753 struct cpumask
*new_mask
)
4755 if (len
< cpumask_size())
4756 cpumask_clear(new_mask
);
4757 else if (len
> cpumask_size())
4758 len
= cpumask_size();
4760 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4764 * sys_sched_setaffinity - set the cpu affinity of a process
4765 * @pid: pid of the process
4766 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4767 * @user_mask_ptr: user-space pointer to the new cpu mask
4769 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4770 unsigned long __user
*, user_mask_ptr
)
4772 cpumask_var_t new_mask
;
4775 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4778 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4780 retval
= sched_setaffinity(pid
, new_mask
);
4781 free_cpumask_var(new_mask
);
4785 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4787 struct task_struct
*p
;
4788 unsigned long flags
;
4796 p
= find_process_by_pid(pid
);
4800 retval
= security_task_getscheduler(p
);
4804 rq
= task_rq_lock(p
, &flags
);
4805 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4806 task_rq_unlock(rq
, &flags
);
4816 * sys_sched_getaffinity - get the cpu affinity of a process
4817 * @pid: pid of the process
4818 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4819 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4821 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4822 unsigned long __user
*, user_mask_ptr
)
4827 if (len
< nr_cpu_ids
)
4829 if (len
& (sizeof(unsigned long)-1))
4832 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4835 ret
= sched_getaffinity(pid
, mask
);
4837 size_t retlen
= min_t(size_t, len
, cpumask_size());
4839 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4844 free_cpumask_var(mask
);
4850 * sys_sched_yield - yield the current processor to other threads.
4852 * This function yields the current CPU to other tasks. If there are no
4853 * other threads running on this CPU then this function will return.
4855 SYSCALL_DEFINE0(sched_yield
)
4857 struct rq
*rq
= this_rq_lock();
4859 schedstat_inc(rq
, yld_count
);
4860 current
->sched_class
->yield_task(rq
);
4863 * Since we are going to call schedule() anyway, there's
4864 * no need to preempt or enable interrupts:
4866 __release(rq
->lock
);
4867 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4868 do_raw_spin_unlock(&rq
->lock
);
4869 preempt_enable_no_resched();
4876 static inline int should_resched(void)
4878 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4881 static void __cond_resched(void)
4883 add_preempt_count(PREEMPT_ACTIVE
);
4885 sub_preempt_count(PREEMPT_ACTIVE
);
4888 int __sched
_cond_resched(void)
4890 if (should_resched()) {
4896 EXPORT_SYMBOL(_cond_resched
);
4899 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4900 * call schedule, and on return reacquire the lock.
4902 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4903 * operations here to prevent schedule() from being called twice (once via
4904 * spin_unlock(), once by hand).
4906 int __cond_resched_lock(spinlock_t
*lock
)
4908 int resched
= should_resched();
4911 lockdep_assert_held(lock
);
4913 if (spin_needbreak(lock
) || resched
) {
4924 EXPORT_SYMBOL(__cond_resched_lock
);
4926 int __sched
__cond_resched_softirq(void)
4928 BUG_ON(!in_softirq());
4930 if (should_resched()) {
4938 EXPORT_SYMBOL(__cond_resched_softirq
);
4941 * yield - yield the current processor to other threads.
4943 * This is a shortcut for kernel-space yielding - it marks the
4944 * thread runnable and calls sys_sched_yield().
4946 void __sched
yield(void)
4948 set_current_state(TASK_RUNNING
);
4951 EXPORT_SYMBOL(yield
);
4954 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4955 * that process accounting knows that this is a task in IO wait state.
4957 void __sched
io_schedule(void)
4959 struct rq
*rq
= raw_rq();
4961 delayacct_blkio_start();
4962 atomic_inc(&rq
->nr_iowait
);
4963 current
->in_iowait
= 1;
4965 current
->in_iowait
= 0;
4966 atomic_dec(&rq
->nr_iowait
);
4967 delayacct_blkio_end();
4969 EXPORT_SYMBOL(io_schedule
);
4971 long __sched
io_schedule_timeout(long timeout
)
4973 struct rq
*rq
= raw_rq();
4976 delayacct_blkio_start();
4977 atomic_inc(&rq
->nr_iowait
);
4978 current
->in_iowait
= 1;
4979 ret
= schedule_timeout(timeout
);
4980 current
->in_iowait
= 0;
4981 atomic_dec(&rq
->nr_iowait
);
4982 delayacct_blkio_end();
4987 * sys_sched_get_priority_max - return maximum RT priority.
4988 * @policy: scheduling class.
4990 * this syscall returns the maximum rt_priority that can be used
4991 * by a given scheduling class.
4993 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5000 ret
= MAX_USER_RT_PRIO
-1;
5012 * sys_sched_get_priority_min - return minimum RT priority.
5013 * @policy: scheduling class.
5015 * this syscall returns the minimum rt_priority that can be used
5016 * by a given scheduling class.
5018 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5036 * sys_sched_rr_get_interval - return the default timeslice of a process.
5037 * @pid: pid of the process.
5038 * @interval: userspace pointer to the timeslice value.
5040 * this syscall writes the default timeslice value of a given process
5041 * into the user-space timespec buffer. A value of '0' means infinity.
5043 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5044 struct timespec __user
*, interval
)
5046 struct task_struct
*p
;
5047 unsigned int time_slice
;
5048 unsigned long flags
;
5058 p
= find_process_by_pid(pid
);
5062 retval
= security_task_getscheduler(p
);
5066 rq
= task_rq_lock(p
, &flags
);
5067 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5068 task_rq_unlock(rq
, &flags
);
5071 jiffies_to_timespec(time_slice
, &t
);
5072 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5080 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5082 void sched_show_task(struct task_struct
*p
)
5084 unsigned long free
= 0;
5087 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5088 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5089 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5090 #if BITS_PER_LONG == 32
5091 if (state
== TASK_RUNNING
)
5092 printk(KERN_CONT
" running ");
5094 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5096 if (state
== TASK_RUNNING
)
5097 printk(KERN_CONT
" running task ");
5099 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5101 #ifdef CONFIG_DEBUG_STACK_USAGE
5102 free
= stack_not_used(p
);
5104 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5105 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5106 (unsigned long)task_thread_info(p
)->flags
);
5108 show_stack(p
, NULL
);
5111 void show_state_filter(unsigned long state_filter
)
5113 struct task_struct
*g
, *p
;
5115 #if BITS_PER_LONG == 32
5117 " task PC stack pid father\n");
5120 " task PC stack pid father\n");
5122 read_lock(&tasklist_lock
);
5123 do_each_thread(g
, p
) {
5125 * reset the NMI-timeout, listing all files on a slow
5126 * console might take alot of time:
5128 touch_nmi_watchdog();
5129 if (!state_filter
|| (p
->state
& state_filter
))
5131 } while_each_thread(g
, p
);
5133 touch_all_softlockup_watchdogs();
5135 #ifdef CONFIG_SCHED_DEBUG
5136 sysrq_sched_debug_show();
5138 read_unlock(&tasklist_lock
);
5140 * Only show locks if all tasks are dumped:
5143 debug_show_all_locks();
5146 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5148 idle
->sched_class
= &idle_sched_class
;
5152 * init_idle - set up an idle thread for a given CPU
5153 * @idle: task in question
5154 * @cpu: cpu the idle task belongs to
5156 * NOTE: this function does not set the idle thread's NEED_RESCHED
5157 * flag, to make booting more robust.
5159 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5161 struct rq
*rq
= cpu_rq(cpu
);
5162 unsigned long flags
;
5164 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5167 idle
->state
= TASK_RUNNING
;
5168 idle
->se
.exec_start
= sched_clock();
5170 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5171 __set_task_cpu(idle
, cpu
);
5173 rq
->curr
= rq
->idle
= idle
;
5174 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5177 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5179 /* Set the preempt count _outside_ the spinlocks! */
5180 #if defined(CONFIG_PREEMPT)
5181 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5183 task_thread_info(idle
)->preempt_count
= 0;
5186 * The idle tasks have their own, simple scheduling class:
5188 idle
->sched_class
= &idle_sched_class
;
5189 ftrace_graph_init_task(idle
);
5193 * In a system that switches off the HZ timer nohz_cpu_mask
5194 * indicates which cpus entered this state. This is used
5195 * in the rcu update to wait only for active cpus. For system
5196 * which do not switch off the HZ timer nohz_cpu_mask should
5197 * always be CPU_BITS_NONE.
5199 cpumask_var_t nohz_cpu_mask
;
5202 * Increase the granularity value when there are more CPUs,
5203 * because with more CPUs the 'effective latency' as visible
5204 * to users decreases. But the relationship is not linear,
5205 * so pick a second-best guess by going with the log2 of the
5208 * This idea comes from the SD scheduler of Con Kolivas:
5210 static int get_update_sysctl_factor(void)
5212 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5213 unsigned int factor
;
5215 switch (sysctl_sched_tunable_scaling
) {
5216 case SCHED_TUNABLESCALING_NONE
:
5219 case SCHED_TUNABLESCALING_LINEAR
:
5222 case SCHED_TUNABLESCALING_LOG
:
5224 factor
= 1 + ilog2(cpus
);
5231 static void update_sysctl(void)
5233 unsigned int factor
= get_update_sysctl_factor();
5235 #define SET_SYSCTL(name) \
5236 (sysctl_##name = (factor) * normalized_sysctl_##name)
5237 SET_SYSCTL(sched_min_granularity
);
5238 SET_SYSCTL(sched_latency
);
5239 SET_SYSCTL(sched_wakeup_granularity
);
5240 SET_SYSCTL(sched_shares_ratelimit
);
5244 static inline void sched_init_granularity(void)
5251 * This is how migration works:
5253 * 1) we queue a struct migration_req structure in the source CPU's
5254 * runqueue and wake up that CPU's migration thread.
5255 * 2) we down() the locked semaphore => thread blocks.
5256 * 3) migration thread wakes up (implicitly it forces the migrated
5257 * thread off the CPU)
5258 * 4) it gets the migration request and checks whether the migrated
5259 * task is still in the wrong runqueue.
5260 * 5) if it's in the wrong runqueue then the migration thread removes
5261 * it and puts it into the right queue.
5262 * 6) migration thread up()s the semaphore.
5263 * 7) we wake up and the migration is done.
5267 * Change a given task's CPU affinity. Migrate the thread to a
5268 * proper CPU and schedule it away if the CPU it's executing on
5269 * is removed from the allowed bitmask.
5271 * NOTE: the caller must have a valid reference to the task, the
5272 * task must not exit() & deallocate itself prematurely. The
5273 * call is not atomic; no spinlocks may be held.
5275 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5277 struct migration_req req
;
5278 unsigned long flags
;
5282 rq
= task_rq_lock(p
, &flags
);
5284 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5289 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5290 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5295 if (p
->sched_class
->set_cpus_allowed
)
5296 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5298 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5299 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5302 /* Can the task run on the task's current CPU? If so, we're done */
5303 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5306 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5307 /* Need help from migration thread: drop lock and wait. */
5308 struct task_struct
*mt
= rq
->migration_thread
;
5310 get_task_struct(mt
);
5311 task_rq_unlock(rq
, &flags
);
5312 wake_up_process(rq
->migration_thread
);
5313 put_task_struct(mt
);
5314 wait_for_completion(&req
.done
);
5315 tlb_migrate_finish(p
->mm
);
5319 task_rq_unlock(rq
, &flags
);
5323 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5326 * Move (not current) task off this cpu, onto dest cpu. We're doing
5327 * this because either it can't run here any more (set_cpus_allowed()
5328 * away from this CPU, or CPU going down), or because we're
5329 * attempting to rebalance this task on exec (sched_exec).
5331 * So we race with normal scheduler movements, but that's OK, as long
5332 * as the task is no longer on this CPU.
5334 * Returns non-zero if task was successfully migrated.
5336 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5338 struct rq
*rq_dest
, *rq_src
;
5341 if (unlikely(!cpu_active(dest_cpu
)))
5344 rq_src
= cpu_rq(src_cpu
);
5345 rq_dest
= cpu_rq(dest_cpu
);
5347 double_rq_lock(rq_src
, rq_dest
);
5348 /* Already moved. */
5349 if (task_cpu(p
) != src_cpu
)
5351 /* Affinity changed (again). */
5352 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5356 * If we're not on a rq, the next wake-up will ensure we're
5360 deactivate_task(rq_src
, p
, 0);
5361 set_task_cpu(p
, dest_cpu
);
5362 activate_task(rq_dest
, p
, 0);
5363 check_preempt_curr(rq_dest
, p
, 0);
5368 double_rq_unlock(rq_src
, rq_dest
);
5372 #define RCU_MIGRATION_IDLE 0
5373 #define RCU_MIGRATION_NEED_QS 1
5374 #define RCU_MIGRATION_GOT_QS 2
5375 #define RCU_MIGRATION_MUST_SYNC 3
5378 * migration_thread - this is a highprio system thread that performs
5379 * thread migration by bumping thread off CPU then 'pushing' onto
5382 static int migration_thread(void *data
)
5385 int cpu
= (long)data
;
5389 BUG_ON(rq
->migration_thread
!= current
);
5391 set_current_state(TASK_INTERRUPTIBLE
);
5392 while (!kthread_should_stop()) {
5393 struct migration_req
*req
;
5394 struct list_head
*head
;
5396 raw_spin_lock_irq(&rq
->lock
);
5398 if (cpu_is_offline(cpu
)) {
5399 raw_spin_unlock_irq(&rq
->lock
);
5403 if (rq
->active_balance
) {
5404 active_load_balance(rq
, cpu
);
5405 rq
->active_balance
= 0;
5408 head
= &rq
->migration_queue
;
5410 if (list_empty(head
)) {
5411 raw_spin_unlock_irq(&rq
->lock
);
5413 set_current_state(TASK_INTERRUPTIBLE
);
5416 req
= list_entry(head
->next
, struct migration_req
, list
);
5417 list_del_init(head
->next
);
5419 if (req
->task
!= NULL
) {
5420 raw_spin_unlock(&rq
->lock
);
5421 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5422 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5423 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5424 raw_spin_unlock(&rq
->lock
);
5426 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5427 raw_spin_unlock(&rq
->lock
);
5428 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5432 complete(&req
->done
);
5434 __set_current_state(TASK_RUNNING
);
5439 #ifdef CONFIG_HOTPLUG_CPU
5441 * Figure out where task on dead CPU should go, use force if necessary.
5443 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5445 struct rq
*rq
= cpu_rq(dead_cpu
);
5446 int needs_cpu
, uninitialized_var(dest_cpu
);
5447 unsigned long flags
;
5449 local_irq_save(flags
);
5451 raw_spin_lock(&rq
->lock
);
5452 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5454 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5455 raw_spin_unlock(&rq
->lock
);
5457 * It can only fail if we race with set_cpus_allowed(),
5458 * in the racer should migrate the task anyway.
5461 __migrate_task(p
, dead_cpu
, dest_cpu
);
5462 local_irq_restore(flags
);
5466 * While a dead CPU has no uninterruptible tasks queued at this point,
5467 * it might still have a nonzero ->nr_uninterruptible counter, because
5468 * for performance reasons the counter is not stricly tracking tasks to
5469 * their home CPUs. So we just add the counter to another CPU's counter,
5470 * to keep the global sum constant after CPU-down:
5472 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5474 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5475 unsigned long flags
;
5477 local_irq_save(flags
);
5478 double_rq_lock(rq_src
, rq_dest
);
5479 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5480 rq_src
->nr_uninterruptible
= 0;
5481 double_rq_unlock(rq_src
, rq_dest
);
5482 local_irq_restore(flags
);
5485 /* Run through task list and migrate tasks from the dead cpu. */
5486 static void migrate_live_tasks(int src_cpu
)
5488 struct task_struct
*p
, *t
;
5490 read_lock(&tasklist_lock
);
5492 do_each_thread(t
, p
) {
5496 if (task_cpu(p
) == src_cpu
)
5497 move_task_off_dead_cpu(src_cpu
, p
);
5498 } while_each_thread(t
, p
);
5500 read_unlock(&tasklist_lock
);
5504 * Schedules idle task to be the next runnable task on current CPU.
5505 * It does so by boosting its priority to highest possible.
5506 * Used by CPU offline code.
5508 void sched_idle_next(void)
5510 int this_cpu
= smp_processor_id();
5511 struct rq
*rq
= cpu_rq(this_cpu
);
5512 struct task_struct
*p
= rq
->idle
;
5513 unsigned long flags
;
5515 /* cpu has to be offline */
5516 BUG_ON(cpu_online(this_cpu
));
5519 * Strictly not necessary since rest of the CPUs are stopped by now
5520 * and interrupts disabled on the current cpu.
5522 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5524 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5526 activate_task(rq
, p
, 0);
5528 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5532 * Ensures that the idle task is using init_mm right before its cpu goes
5535 void idle_task_exit(void)
5537 struct mm_struct
*mm
= current
->active_mm
;
5539 BUG_ON(cpu_online(smp_processor_id()));
5542 switch_mm(mm
, &init_mm
, current
);
5546 /* called under rq->lock with disabled interrupts */
5547 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5549 struct rq
*rq
= cpu_rq(dead_cpu
);
5551 /* Must be exiting, otherwise would be on tasklist. */
5552 BUG_ON(!p
->exit_state
);
5554 /* Cannot have done final schedule yet: would have vanished. */
5555 BUG_ON(p
->state
== TASK_DEAD
);
5560 * Drop lock around migration; if someone else moves it,
5561 * that's OK. No task can be added to this CPU, so iteration is
5564 raw_spin_unlock_irq(&rq
->lock
);
5565 move_task_off_dead_cpu(dead_cpu
, p
);
5566 raw_spin_lock_irq(&rq
->lock
);
5571 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5572 static void migrate_dead_tasks(unsigned int dead_cpu
)
5574 struct rq
*rq
= cpu_rq(dead_cpu
);
5575 struct task_struct
*next
;
5578 if (!rq
->nr_running
)
5580 next
= pick_next_task(rq
);
5583 next
->sched_class
->put_prev_task(rq
, next
);
5584 migrate_dead(dead_cpu
, next
);
5590 * remove the tasks which were accounted by rq from calc_load_tasks.
5592 static void calc_global_load_remove(struct rq
*rq
)
5594 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5595 rq
->calc_load_active
= 0;
5597 #endif /* CONFIG_HOTPLUG_CPU */
5599 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5601 static struct ctl_table sd_ctl_dir
[] = {
5603 .procname
= "sched_domain",
5609 static struct ctl_table sd_ctl_root
[] = {
5611 .procname
= "kernel",
5613 .child
= sd_ctl_dir
,
5618 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5620 struct ctl_table
*entry
=
5621 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5626 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5628 struct ctl_table
*entry
;
5631 * In the intermediate directories, both the child directory and
5632 * procname are dynamically allocated and could fail but the mode
5633 * will always be set. In the lowest directory the names are
5634 * static strings and all have proc handlers.
5636 for (entry
= *tablep
; entry
->mode
; entry
++) {
5638 sd_free_ctl_entry(&entry
->child
);
5639 if (entry
->proc_handler
== NULL
)
5640 kfree(entry
->procname
);
5648 set_table_entry(struct ctl_table
*entry
,
5649 const char *procname
, void *data
, int maxlen
,
5650 mode_t mode
, proc_handler
*proc_handler
)
5652 entry
->procname
= procname
;
5654 entry
->maxlen
= maxlen
;
5656 entry
->proc_handler
= proc_handler
;
5659 static struct ctl_table
*
5660 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5662 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5667 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5668 sizeof(long), 0644, proc_doulongvec_minmax
);
5669 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5670 sizeof(long), 0644, proc_doulongvec_minmax
);
5671 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5672 sizeof(int), 0644, proc_dointvec_minmax
);
5673 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5674 sizeof(int), 0644, proc_dointvec_minmax
);
5675 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5676 sizeof(int), 0644, proc_dointvec_minmax
);
5677 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5678 sizeof(int), 0644, proc_dointvec_minmax
);
5679 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5680 sizeof(int), 0644, proc_dointvec_minmax
);
5681 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5682 sizeof(int), 0644, proc_dointvec_minmax
);
5683 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5684 sizeof(int), 0644, proc_dointvec_minmax
);
5685 set_table_entry(&table
[9], "cache_nice_tries",
5686 &sd
->cache_nice_tries
,
5687 sizeof(int), 0644, proc_dointvec_minmax
);
5688 set_table_entry(&table
[10], "flags", &sd
->flags
,
5689 sizeof(int), 0644, proc_dointvec_minmax
);
5690 set_table_entry(&table
[11], "name", sd
->name
,
5691 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5692 /* &table[12] is terminator */
5697 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5699 struct ctl_table
*entry
, *table
;
5700 struct sched_domain
*sd
;
5701 int domain_num
= 0, i
;
5704 for_each_domain(cpu
, sd
)
5706 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5711 for_each_domain(cpu
, sd
) {
5712 snprintf(buf
, 32, "domain%d", i
);
5713 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5715 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5722 static struct ctl_table_header
*sd_sysctl_header
;
5723 static void register_sched_domain_sysctl(void)
5725 int i
, cpu_num
= num_possible_cpus();
5726 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5729 WARN_ON(sd_ctl_dir
[0].child
);
5730 sd_ctl_dir
[0].child
= entry
;
5735 for_each_possible_cpu(i
) {
5736 snprintf(buf
, 32, "cpu%d", i
);
5737 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5739 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5743 WARN_ON(sd_sysctl_header
);
5744 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5747 /* may be called multiple times per register */
5748 static void unregister_sched_domain_sysctl(void)
5750 if (sd_sysctl_header
)
5751 unregister_sysctl_table(sd_sysctl_header
);
5752 sd_sysctl_header
= NULL
;
5753 if (sd_ctl_dir
[0].child
)
5754 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5757 static void register_sched_domain_sysctl(void)
5760 static void unregister_sched_domain_sysctl(void)
5765 static void set_rq_online(struct rq
*rq
)
5768 const struct sched_class
*class;
5770 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5773 for_each_class(class) {
5774 if (class->rq_online
)
5775 class->rq_online(rq
);
5780 static void set_rq_offline(struct rq
*rq
)
5783 const struct sched_class
*class;
5785 for_each_class(class) {
5786 if (class->rq_offline
)
5787 class->rq_offline(rq
);
5790 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5796 * migration_call - callback that gets triggered when a CPU is added.
5797 * Here we can start up the necessary migration thread for the new CPU.
5799 static int __cpuinit
5800 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5802 struct task_struct
*p
;
5803 int cpu
= (long)hcpu
;
5804 unsigned long flags
;
5809 case CPU_UP_PREPARE
:
5810 case CPU_UP_PREPARE_FROZEN
:
5811 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5814 kthread_bind(p
, cpu
);
5815 /* Must be high prio: stop_machine expects to yield to it. */
5816 rq
= task_rq_lock(p
, &flags
);
5817 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5818 task_rq_unlock(rq
, &flags
);
5820 cpu_rq(cpu
)->migration_thread
= p
;
5821 rq
->calc_load_update
= calc_load_update
;
5825 case CPU_ONLINE_FROZEN
:
5826 /* Strictly unnecessary, as first user will wake it. */
5827 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5829 /* Update our root-domain */
5831 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5833 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5837 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5840 #ifdef CONFIG_HOTPLUG_CPU
5841 case CPU_UP_CANCELED
:
5842 case CPU_UP_CANCELED_FROZEN
:
5843 if (!cpu_rq(cpu
)->migration_thread
)
5845 /* Unbind it from offline cpu so it can run. Fall thru. */
5846 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5847 cpumask_any(cpu_online_mask
));
5848 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5849 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5850 cpu_rq(cpu
)->migration_thread
= NULL
;
5854 case CPU_DEAD_FROZEN
:
5855 migrate_live_tasks(cpu
);
5857 kthread_stop(rq
->migration_thread
);
5858 put_task_struct(rq
->migration_thread
);
5859 rq
->migration_thread
= NULL
;
5860 /* Idle task back to normal (off runqueue, low prio) */
5861 raw_spin_lock_irq(&rq
->lock
);
5862 deactivate_task(rq
, rq
->idle
, 0);
5863 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5864 rq
->idle
->sched_class
= &idle_sched_class
;
5865 migrate_dead_tasks(cpu
);
5866 raw_spin_unlock_irq(&rq
->lock
);
5867 migrate_nr_uninterruptible(rq
);
5868 BUG_ON(rq
->nr_running
!= 0);
5869 calc_global_load_remove(rq
);
5871 * No need to migrate the tasks: it was best-effort if
5872 * they didn't take sched_hotcpu_mutex. Just wake up
5875 raw_spin_lock_irq(&rq
->lock
);
5876 while (!list_empty(&rq
->migration_queue
)) {
5877 struct migration_req
*req
;
5879 req
= list_entry(rq
->migration_queue
.next
,
5880 struct migration_req
, list
);
5881 list_del_init(&req
->list
);
5882 raw_spin_unlock_irq(&rq
->lock
);
5883 complete(&req
->done
);
5884 raw_spin_lock_irq(&rq
->lock
);
5886 raw_spin_unlock_irq(&rq
->lock
);
5890 case CPU_DYING_FROZEN
:
5891 /* Update our root-domain */
5893 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5895 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5898 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5906 * Register at high priority so that task migration (migrate_all_tasks)
5907 * happens before everything else. This has to be lower priority than
5908 * the notifier in the perf_event subsystem, though.
5910 static struct notifier_block __cpuinitdata migration_notifier
= {
5911 .notifier_call
= migration_call
,
5915 static int __init
migration_init(void)
5917 void *cpu
= (void *)(long)smp_processor_id();
5920 /* Start one for the boot CPU: */
5921 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5922 BUG_ON(err
== NOTIFY_BAD
);
5923 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5924 register_cpu_notifier(&migration_notifier
);
5928 early_initcall(migration_init
);
5933 #ifdef CONFIG_SCHED_DEBUG
5935 static __read_mostly
int sched_domain_debug_enabled
;
5937 static int __init
sched_domain_debug_setup(char *str
)
5939 sched_domain_debug_enabled
= 1;
5943 early_param("sched_debug", sched_domain_debug_setup
);
5945 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5946 struct cpumask
*groupmask
)
5948 struct sched_group
*group
= sd
->groups
;
5951 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5952 cpumask_clear(groupmask
);
5954 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5956 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5957 printk("does not load-balance\n");
5959 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5964 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5966 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5967 printk(KERN_ERR
"ERROR: domain->span does not contain "
5970 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5971 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5975 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5979 printk(KERN_ERR
"ERROR: group is NULL\n");
5983 if (!group
->cpu_power
) {
5984 printk(KERN_CONT
"\n");
5985 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5990 if (!cpumask_weight(sched_group_cpus(group
))) {
5991 printk(KERN_CONT
"\n");
5992 printk(KERN_ERR
"ERROR: empty group\n");
5996 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5997 printk(KERN_CONT
"\n");
5998 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6002 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6004 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6006 printk(KERN_CONT
" %s", str
);
6007 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6008 printk(KERN_CONT
" (cpu_power = %d)",
6012 group
= group
->next
;
6013 } while (group
!= sd
->groups
);
6014 printk(KERN_CONT
"\n");
6016 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6017 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6020 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6021 printk(KERN_ERR
"ERROR: parent span is not a superset "
6022 "of domain->span\n");
6026 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6028 cpumask_var_t groupmask
;
6031 if (!sched_domain_debug_enabled
)
6035 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6039 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6041 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6042 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6047 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6054 free_cpumask_var(groupmask
);
6056 #else /* !CONFIG_SCHED_DEBUG */
6057 # define sched_domain_debug(sd, cpu) do { } while (0)
6058 #endif /* CONFIG_SCHED_DEBUG */
6060 static int sd_degenerate(struct sched_domain
*sd
)
6062 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6065 /* Following flags need at least 2 groups */
6066 if (sd
->flags
& (SD_LOAD_BALANCE
|
6067 SD_BALANCE_NEWIDLE
|
6071 SD_SHARE_PKG_RESOURCES
)) {
6072 if (sd
->groups
!= sd
->groups
->next
)
6076 /* Following flags don't use groups */
6077 if (sd
->flags
& (SD_WAKE_AFFINE
))
6084 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6086 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6088 if (sd_degenerate(parent
))
6091 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6094 /* Flags needing groups don't count if only 1 group in parent */
6095 if (parent
->groups
== parent
->groups
->next
) {
6096 pflags
&= ~(SD_LOAD_BALANCE
|
6097 SD_BALANCE_NEWIDLE
|
6101 SD_SHARE_PKG_RESOURCES
);
6102 if (nr_node_ids
== 1)
6103 pflags
&= ~SD_SERIALIZE
;
6105 if (~cflags
& pflags
)
6111 static void free_rootdomain(struct root_domain
*rd
)
6113 synchronize_sched();
6115 cpupri_cleanup(&rd
->cpupri
);
6117 free_cpumask_var(rd
->rto_mask
);
6118 free_cpumask_var(rd
->online
);
6119 free_cpumask_var(rd
->span
);
6123 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6125 struct root_domain
*old_rd
= NULL
;
6126 unsigned long flags
;
6128 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6133 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6136 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6139 * If we dont want to free the old_rt yet then
6140 * set old_rd to NULL to skip the freeing later
6143 if (!atomic_dec_and_test(&old_rd
->refcount
))
6147 atomic_inc(&rd
->refcount
);
6150 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6151 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6154 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6157 free_rootdomain(old_rd
);
6160 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6162 gfp_t gfp
= GFP_KERNEL
;
6164 memset(rd
, 0, sizeof(*rd
));
6169 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6171 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6173 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6176 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6181 free_cpumask_var(rd
->rto_mask
);
6183 free_cpumask_var(rd
->online
);
6185 free_cpumask_var(rd
->span
);
6190 static void init_defrootdomain(void)
6192 init_rootdomain(&def_root_domain
, true);
6194 atomic_set(&def_root_domain
.refcount
, 1);
6197 static struct root_domain
*alloc_rootdomain(void)
6199 struct root_domain
*rd
;
6201 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6205 if (init_rootdomain(rd
, false) != 0) {
6214 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6215 * hold the hotplug lock.
6218 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6220 struct rq
*rq
= cpu_rq(cpu
);
6221 struct sched_domain
*tmp
;
6223 /* Remove the sched domains which do not contribute to scheduling. */
6224 for (tmp
= sd
; tmp
; ) {
6225 struct sched_domain
*parent
= tmp
->parent
;
6229 if (sd_parent_degenerate(tmp
, parent
)) {
6230 tmp
->parent
= parent
->parent
;
6232 parent
->parent
->child
= tmp
;
6237 if (sd
&& sd_degenerate(sd
)) {
6243 sched_domain_debug(sd
, cpu
);
6245 rq_attach_root(rq
, rd
);
6246 rcu_assign_pointer(rq
->sd
, sd
);
6249 /* cpus with isolated domains */
6250 static cpumask_var_t cpu_isolated_map
;
6252 /* Setup the mask of cpus configured for isolated domains */
6253 static int __init
isolated_cpu_setup(char *str
)
6255 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6256 cpulist_parse(str
, cpu_isolated_map
);
6260 __setup("isolcpus=", isolated_cpu_setup
);
6263 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6264 * to a function which identifies what group(along with sched group) a CPU
6265 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6266 * (due to the fact that we keep track of groups covered with a struct cpumask).
6268 * init_sched_build_groups will build a circular linked list of the groups
6269 * covered by the given span, and will set each group's ->cpumask correctly,
6270 * and ->cpu_power to 0.
6273 init_sched_build_groups(const struct cpumask
*span
,
6274 const struct cpumask
*cpu_map
,
6275 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6276 struct sched_group
**sg
,
6277 struct cpumask
*tmpmask
),
6278 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6280 struct sched_group
*first
= NULL
, *last
= NULL
;
6283 cpumask_clear(covered
);
6285 for_each_cpu(i
, span
) {
6286 struct sched_group
*sg
;
6287 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6290 if (cpumask_test_cpu(i
, covered
))
6293 cpumask_clear(sched_group_cpus(sg
));
6296 for_each_cpu(j
, span
) {
6297 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6300 cpumask_set_cpu(j
, covered
);
6301 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6312 #define SD_NODES_PER_DOMAIN 16
6317 * find_next_best_node - find the next node to include in a sched_domain
6318 * @node: node whose sched_domain we're building
6319 * @used_nodes: nodes already in the sched_domain
6321 * Find the next node to include in a given scheduling domain. Simply
6322 * finds the closest node not already in the @used_nodes map.
6324 * Should use nodemask_t.
6326 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6328 int i
, n
, val
, min_val
, best_node
= 0;
6332 for (i
= 0; i
< nr_node_ids
; i
++) {
6333 /* Start at @node */
6334 n
= (node
+ i
) % nr_node_ids
;
6336 if (!nr_cpus_node(n
))
6339 /* Skip already used nodes */
6340 if (node_isset(n
, *used_nodes
))
6343 /* Simple min distance search */
6344 val
= node_distance(node
, n
);
6346 if (val
< min_val
) {
6352 node_set(best_node
, *used_nodes
);
6357 * sched_domain_node_span - get a cpumask for a node's sched_domain
6358 * @node: node whose cpumask we're constructing
6359 * @span: resulting cpumask
6361 * Given a node, construct a good cpumask for its sched_domain to span. It
6362 * should be one that prevents unnecessary balancing, but also spreads tasks
6365 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6367 nodemask_t used_nodes
;
6370 cpumask_clear(span
);
6371 nodes_clear(used_nodes
);
6373 cpumask_or(span
, span
, cpumask_of_node(node
));
6374 node_set(node
, used_nodes
);
6376 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6377 int next_node
= find_next_best_node(node
, &used_nodes
);
6379 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6382 #endif /* CONFIG_NUMA */
6384 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6387 * The cpus mask in sched_group and sched_domain hangs off the end.
6389 * ( See the the comments in include/linux/sched.h:struct sched_group
6390 * and struct sched_domain. )
6392 struct static_sched_group
{
6393 struct sched_group sg
;
6394 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6397 struct static_sched_domain
{
6398 struct sched_domain sd
;
6399 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6405 cpumask_var_t domainspan
;
6406 cpumask_var_t covered
;
6407 cpumask_var_t notcovered
;
6409 cpumask_var_t nodemask
;
6410 cpumask_var_t this_sibling_map
;
6411 cpumask_var_t this_core_map
;
6412 cpumask_var_t send_covered
;
6413 cpumask_var_t tmpmask
;
6414 struct sched_group
**sched_group_nodes
;
6415 struct root_domain
*rd
;
6419 sa_sched_groups
= 0,
6424 sa_this_sibling_map
,
6426 sa_sched_group_nodes
,
6436 * SMT sched-domains:
6438 #ifdef CONFIG_SCHED_SMT
6439 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6440 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6443 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6444 struct sched_group
**sg
, struct cpumask
*unused
)
6447 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6450 #endif /* CONFIG_SCHED_SMT */
6453 * multi-core sched-domains:
6455 #ifdef CONFIG_SCHED_MC
6456 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6457 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6458 #endif /* CONFIG_SCHED_MC */
6460 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6462 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6463 struct sched_group
**sg
, struct cpumask
*mask
)
6467 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6468 group
= cpumask_first(mask
);
6470 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6473 #elif defined(CONFIG_SCHED_MC)
6475 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6476 struct sched_group
**sg
, struct cpumask
*unused
)
6479 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6484 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6485 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6488 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6489 struct sched_group
**sg
, struct cpumask
*mask
)
6492 #ifdef CONFIG_SCHED_MC
6493 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6494 group
= cpumask_first(mask
);
6495 #elif defined(CONFIG_SCHED_SMT)
6496 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6497 group
= cpumask_first(mask
);
6502 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6508 * The init_sched_build_groups can't handle what we want to do with node
6509 * groups, so roll our own. Now each node has its own list of groups which
6510 * gets dynamically allocated.
6512 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6513 static struct sched_group
***sched_group_nodes_bycpu
;
6515 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6516 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6518 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6519 struct sched_group
**sg
,
6520 struct cpumask
*nodemask
)
6524 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6525 group
= cpumask_first(nodemask
);
6528 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6532 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6534 struct sched_group
*sg
= group_head
;
6540 for_each_cpu(j
, sched_group_cpus(sg
)) {
6541 struct sched_domain
*sd
;
6543 sd
= &per_cpu(phys_domains
, j
).sd
;
6544 if (j
!= group_first_cpu(sd
->groups
)) {
6546 * Only add "power" once for each
6552 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6555 } while (sg
!= group_head
);
6558 static int build_numa_sched_groups(struct s_data
*d
,
6559 const struct cpumask
*cpu_map
, int num
)
6561 struct sched_domain
*sd
;
6562 struct sched_group
*sg
, *prev
;
6565 cpumask_clear(d
->covered
);
6566 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6567 if (cpumask_empty(d
->nodemask
)) {
6568 d
->sched_group_nodes
[num
] = NULL
;
6572 sched_domain_node_span(num
, d
->domainspan
);
6573 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6575 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6578 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6582 d
->sched_group_nodes
[num
] = sg
;
6584 for_each_cpu(j
, d
->nodemask
) {
6585 sd
= &per_cpu(node_domains
, j
).sd
;
6590 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6592 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6595 for (j
= 0; j
< nr_node_ids
; j
++) {
6596 n
= (num
+ j
) % nr_node_ids
;
6597 cpumask_complement(d
->notcovered
, d
->covered
);
6598 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6599 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6600 if (cpumask_empty(d
->tmpmask
))
6602 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6603 if (cpumask_empty(d
->tmpmask
))
6605 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6609 "Can not alloc domain group for node %d\n", j
);
6613 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6614 sg
->next
= prev
->next
;
6615 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6622 #endif /* CONFIG_NUMA */
6625 /* Free memory allocated for various sched_group structures */
6626 static void free_sched_groups(const struct cpumask
*cpu_map
,
6627 struct cpumask
*nodemask
)
6631 for_each_cpu(cpu
, cpu_map
) {
6632 struct sched_group
**sched_group_nodes
6633 = sched_group_nodes_bycpu
[cpu
];
6635 if (!sched_group_nodes
)
6638 for (i
= 0; i
< nr_node_ids
; i
++) {
6639 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6641 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6642 if (cpumask_empty(nodemask
))
6652 if (oldsg
!= sched_group_nodes
[i
])
6655 kfree(sched_group_nodes
);
6656 sched_group_nodes_bycpu
[cpu
] = NULL
;
6659 #else /* !CONFIG_NUMA */
6660 static void free_sched_groups(const struct cpumask
*cpu_map
,
6661 struct cpumask
*nodemask
)
6664 #endif /* CONFIG_NUMA */
6667 * Initialize sched groups cpu_power.
6669 * cpu_power indicates the capacity of sched group, which is used while
6670 * distributing the load between different sched groups in a sched domain.
6671 * Typically cpu_power for all the groups in a sched domain will be same unless
6672 * there are asymmetries in the topology. If there are asymmetries, group
6673 * having more cpu_power will pickup more load compared to the group having
6676 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6678 struct sched_domain
*child
;
6679 struct sched_group
*group
;
6683 WARN_ON(!sd
|| !sd
->groups
);
6685 if (cpu
!= group_first_cpu(sd
->groups
))
6690 sd
->groups
->cpu_power
= 0;
6693 power
= SCHED_LOAD_SCALE
;
6694 weight
= cpumask_weight(sched_domain_span(sd
));
6696 * SMT siblings share the power of a single core.
6697 * Usually multiple threads get a better yield out of
6698 * that one core than a single thread would have,
6699 * reflect that in sd->smt_gain.
6701 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6702 power
*= sd
->smt_gain
;
6704 power
>>= SCHED_LOAD_SHIFT
;
6706 sd
->groups
->cpu_power
+= power
;
6711 * Add cpu_power of each child group to this groups cpu_power.
6713 group
= child
->groups
;
6715 sd
->groups
->cpu_power
+= group
->cpu_power
;
6716 group
= group
->next
;
6717 } while (group
!= child
->groups
);
6721 * Initializers for schedule domains
6722 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6725 #ifdef CONFIG_SCHED_DEBUG
6726 # define SD_INIT_NAME(sd, type) sd->name = #type
6728 # define SD_INIT_NAME(sd, type) do { } while (0)
6731 #define SD_INIT(sd, type) sd_init_##type(sd)
6733 #define SD_INIT_FUNC(type) \
6734 static noinline void sd_init_##type(struct sched_domain *sd) \
6736 memset(sd, 0, sizeof(*sd)); \
6737 *sd = SD_##type##_INIT; \
6738 sd->level = SD_LV_##type; \
6739 SD_INIT_NAME(sd, type); \
6744 SD_INIT_FUNC(ALLNODES
)
6747 #ifdef CONFIG_SCHED_SMT
6748 SD_INIT_FUNC(SIBLING
)
6750 #ifdef CONFIG_SCHED_MC
6754 static int default_relax_domain_level
= -1;
6756 static int __init
setup_relax_domain_level(char *str
)
6760 val
= simple_strtoul(str
, NULL
, 0);
6761 if (val
< SD_LV_MAX
)
6762 default_relax_domain_level
= val
;
6766 __setup("relax_domain_level=", setup_relax_domain_level
);
6768 static void set_domain_attribute(struct sched_domain
*sd
,
6769 struct sched_domain_attr
*attr
)
6773 if (!attr
|| attr
->relax_domain_level
< 0) {
6774 if (default_relax_domain_level
< 0)
6777 request
= default_relax_domain_level
;
6779 request
= attr
->relax_domain_level
;
6780 if (request
< sd
->level
) {
6781 /* turn off idle balance on this domain */
6782 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6784 /* turn on idle balance on this domain */
6785 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6789 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6790 const struct cpumask
*cpu_map
)
6793 case sa_sched_groups
:
6794 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6795 d
->sched_group_nodes
= NULL
;
6797 free_rootdomain(d
->rd
); /* fall through */
6799 free_cpumask_var(d
->tmpmask
); /* fall through */
6800 case sa_send_covered
:
6801 free_cpumask_var(d
->send_covered
); /* fall through */
6802 case sa_this_core_map
:
6803 free_cpumask_var(d
->this_core_map
); /* fall through */
6804 case sa_this_sibling_map
:
6805 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6807 free_cpumask_var(d
->nodemask
); /* fall through */
6808 case sa_sched_group_nodes
:
6810 kfree(d
->sched_group_nodes
); /* fall through */
6812 free_cpumask_var(d
->notcovered
); /* fall through */
6814 free_cpumask_var(d
->covered
); /* fall through */
6816 free_cpumask_var(d
->domainspan
); /* fall through */
6823 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6824 const struct cpumask
*cpu_map
)
6827 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6829 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6830 return sa_domainspan
;
6831 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6833 /* Allocate the per-node list of sched groups */
6834 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6835 sizeof(struct sched_group
*), GFP_KERNEL
);
6836 if (!d
->sched_group_nodes
) {
6837 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6838 return sa_notcovered
;
6840 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6842 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6843 return sa_sched_group_nodes
;
6844 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6846 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6847 return sa_this_sibling_map
;
6848 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6849 return sa_this_core_map
;
6850 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6851 return sa_send_covered
;
6852 d
->rd
= alloc_rootdomain();
6854 printk(KERN_WARNING
"Cannot alloc root domain\n");
6857 return sa_rootdomain
;
6860 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6861 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6863 struct sched_domain
*sd
= NULL
;
6865 struct sched_domain
*parent
;
6868 if (cpumask_weight(cpu_map
) >
6869 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6870 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6871 SD_INIT(sd
, ALLNODES
);
6872 set_domain_attribute(sd
, attr
);
6873 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6874 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6879 sd
= &per_cpu(node_domains
, i
).sd
;
6881 set_domain_attribute(sd
, attr
);
6882 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6883 sd
->parent
= parent
;
6886 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6891 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6892 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6893 struct sched_domain
*parent
, int i
)
6895 struct sched_domain
*sd
;
6896 sd
= &per_cpu(phys_domains
, i
).sd
;
6898 set_domain_attribute(sd
, attr
);
6899 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6900 sd
->parent
= parent
;
6903 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6907 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6908 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6909 struct sched_domain
*parent
, int i
)
6911 struct sched_domain
*sd
= parent
;
6912 #ifdef CONFIG_SCHED_MC
6913 sd
= &per_cpu(core_domains
, i
).sd
;
6915 set_domain_attribute(sd
, attr
);
6916 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6917 sd
->parent
= parent
;
6919 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6924 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6925 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6926 struct sched_domain
*parent
, int i
)
6928 struct sched_domain
*sd
= parent
;
6929 #ifdef CONFIG_SCHED_SMT
6930 sd
= &per_cpu(cpu_domains
, i
).sd
;
6931 SD_INIT(sd
, SIBLING
);
6932 set_domain_attribute(sd
, attr
);
6933 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6934 sd
->parent
= parent
;
6936 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6941 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6942 const struct cpumask
*cpu_map
, int cpu
)
6945 #ifdef CONFIG_SCHED_SMT
6946 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6947 cpumask_and(d
->this_sibling_map
, cpu_map
,
6948 topology_thread_cpumask(cpu
));
6949 if (cpu
== cpumask_first(d
->this_sibling_map
))
6950 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
6952 d
->send_covered
, d
->tmpmask
);
6955 #ifdef CONFIG_SCHED_MC
6956 case SD_LV_MC
: /* set up multi-core groups */
6957 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
6958 if (cpu
== cpumask_first(d
->this_core_map
))
6959 init_sched_build_groups(d
->this_core_map
, cpu_map
,
6961 d
->send_covered
, d
->tmpmask
);
6964 case SD_LV_CPU
: /* set up physical groups */
6965 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
6966 if (!cpumask_empty(d
->nodemask
))
6967 init_sched_build_groups(d
->nodemask
, cpu_map
,
6969 d
->send_covered
, d
->tmpmask
);
6972 case SD_LV_ALLNODES
:
6973 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
6974 d
->send_covered
, d
->tmpmask
);
6983 * Build sched domains for a given set of cpus and attach the sched domains
6984 * to the individual cpus
6986 static int __build_sched_domains(const struct cpumask
*cpu_map
,
6987 struct sched_domain_attr
*attr
)
6989 enum s_alloc alloc_state
= sa_none
;
6991 struct sched_domain
*sd
;
6997 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6998 if (alloc_state
!= sa_rootdomain
)
7000 alloc_state
= sa_sched_groups
;
7003 * Set up domains for cpus specified by the cpu_map.
7005 for_each_cpu(i
, cpu_map
) {
7006 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7009 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7010 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7011 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7012 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7015 for_each_cpu(i
, cpu_map
) {
7016 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7017 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7020 /* Set up physical groups */
7021 for (i
= 0; i
< nr_node_ids
; i
++)
7022 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7025 /* Set up node groups */
7027 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7029 for (i
= 0; i
< nr_node_ids
; i
++)
7030 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7034 /* Calculate CPU power for physical packages and nodes */
7035 #ifdef CONFIG_SCHED_SMT
7036 for_each_cpu(i
, cpu_map
) {
7037 sd
= &per_cpu(cpu_domains
, i
).sd
;
7038 init_sched_groups_power(i
, sd
);
7041 #ifdef CONFIG_SCHED_MC
7042 for_each_cpu(i
, cpu_map
) {
7043 sd
= &per_cpu(core_domains
, i
).sd
;
7044 init_sched_groups_power(i
, sd
);
7048 for_each_cpu(i
, cpu_map
) {
7049 sd
= &per_cpu(phys_domains
, i
).sd
;
7050 init_sched_groups_power(i
, sd
);
7054 for (i
= 0; i
< nr_node_ids
; i
++)
7055 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7057 if (d
.sd_allnodes
) {
7058 struct sched_group
*sg
;
7060 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7062 init_numa_sched_groups_power(sg
);
7066 /* Attach the domains */
7067 for_each_cpu(i
, cpu_map
) {
7068 #ifdef CONFIG_SCHED_SMT
7069 sd
= &per_cpu(cpu_domains
, i
).sd
;
7070 #elif defined(CONFIG_SCHED_MC)
7071 sd
= &per_cpu(core_domains
, i
).sd
;
7073 sd
= &per_cpu(phys_domains
, i
).sd
;
7075 cpu_attach_domain(sd
, d
.rd
, i
);
7078 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7079 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7083 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7087 static int build_sched_domains(const struct cpumask
*cpu_map
)
7089 return __build_sched_domains(cpu_map
, NULL
);
7092 static cpumask_var_t
*doms_cur
; /* current sched domains */
7093 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7094 static struct sched_domain_attr
*dattr_cur
;
7095 /* attribues of custom domains in 'doms_cur' */
7098 * Special case: If a kmalloc of a doms_cur partition (array of
7099 * cpumask) fails, then fallback to a single sched domain,
7100 * as determined by the single cpumask fallback_doms.
7102 static cpumask_var_t fallback_doms
;
7105 * arch_update_cpu_topology lets virtualized architectures update the
7106 * cpu core maps. It is supposed to return 1 if the topology changed
7107 * or 0 if it stayed the same.
7109 int __attribute__((weak
)) arch_update_cpu_topology(void)
7114 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7117 cpumask_var_t
*doms
;
7119 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7122 for (i
= 0; i
< ndoms
; i
++) {
7123 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7124 free_sched_domains(doms
, i
);
7131 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7134 for (i
= 0; i
< ndoms
; i
++)
7135 free_cpumask_var(doms
[i
]);
7140 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7141 * For now this just excludes isolated cpus, but could be used to
7142 * exclude other special cases in the future.
7144 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7148 arch_update_cpu_topology();
7150 doms_cur
= alloc_sched_domains(ndoms_cur
);
7152 doms_cur
= &fallback_doms
;
7153 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7155 err
= build_sched_domains(doms_cur
[0]);
7156 register_sched_domain_sysctl();
7161 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7162 struct cpumask
*tmpmask
)
7164 free_sched_groups(cpu_map
, tmpmask
);
7168 * Detach sched domains from a group of cpus specified in cpu_map
7169 * These cpus will now be attached to the NULL domain
7171 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7173 /* Save because hotplug lock held. */
7174 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7177 for_each_cpu(i
, cpu_map
)
7178 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7179 synchronize_sched();
7180 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7183 /* handle null as "default" */
7184 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7185 struct sched_domain_attr
*new, int idx_new
)
7187 struct sched_domain_attr tmp
;
7194 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7195 new ? (new + idx_new
) : &tmp
,
7196 sizeof(struct sched_domain_attr
));
7200 * Partition sched domains as specified by the 'ndoms_new'
7201 * cpumasks in the array doms_new[] of cpumasks. This compares
7202 * doms_new[] to the current sched domain partitioning, doms_cur[].
7203 * It destroys each deleted domain and builds each new domain.
7205 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7206 * The masks don't intersect (don't overlap.) We should setup one
7207 * sched domain for each mask. CPUs not in any of the cpumasks will
7208 * not be load balanced. If the same cpumask appears both in the
7209 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7212 * The passed in 'doms_new' should be allocated using
7213 * alloc_sched_domains. This routine takes ownership of it and will
7214 * free_sched_domains it when done with it. If the caller failed the
7215 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7216 * and partition_sched_domains() will fallback to the single partition
7217 * 'fallback_doms', it also forces the domains to be rebuilt.
7219 * If doms_new == NULL it will be replaced with cpu_online_mask.
7220 * ndoms_new == 0 is a special case for destroying existing domains,
7221 * and it will not create the default domain.
7223 * Call with hotplug lock held
7225 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7226 struct sched_domain_attr
*dattr_new
)
7231 mutex_lock(&sched_domains_mutex
);
7233 /* always unregister in case we don't destroy any domains */
7234 unregister_sched_domain_sysctl();
7236 /* Let architecture update cpu core mappings. */
7237 new_topology
= arch_update_cpu_topology();
7239 n
= doms_new
? ndoms_new
: 0;
7241 /* Destroy deleted domains */
7242 for (i
= 0; i
< ndoms_cur
; i
++) {
7243 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7244 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7245 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7248 /* no match - a current sched domain not in new doms_new[] */
7249 detach_destroy_domains(doms_cur
[i
]);
7254 if (doms_new
== NULL
) {
7256 doms_new
= &fallback_doms
;
7257 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7258 WARN_ON_ONCE(dattr_new
);
7261 /* Build new domains */
7262 for (i
= 0; i
< ndoms_new
; i
++) {
7263 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7264 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7265 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7268 /* no match - add a new doms_new */
7269 __build_sched_domains(doms_new
[i
],
7270 dattr_new
? dattr_new
+ i
: NULL
);
7275 /* Remember the new sched domains */
7276 if (doms_cur
!= &fallback_doms
)
7277 free_sched_domains(doms_cur
, ndoms_cur
);
7278 kfree(dattr_cur
); /* kfree(NULL) is safe */
7279 doms_cur
= doms_new
;
7280 dattr_cur
= dattr_new
;
7281 ndoms_cur
= ndoms_new
;
7283 register_sched_domain_sysctl();
7285 mutex_unlock(&sched_domains_mutex
);
7288 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7289 static void arch_reinit_sched_domains(void)
7293 /* Destroy domains first to force the rebuild */
7294 partition_sched_domains(0, NULL
, NULL
);
7296 rebuild_sched_domains();
7300 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7302 unsigned int level
= 0;
7304 if (sscanf(buf
, "%u", &level
) != 1)
7308 * level is always be positive so don't check for
7309 * level < POWERSAVINGS_BALANCE_NONE which is 0
7310 * What happens on 0 or 1 byte write,
7311 * need to check for count as well?
7314 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7318 sched_smt_power_savings
= level
;
7320 sched_mc_power_savings
= level
;
7322 arch_reinit_sched_domains();
7327 #ifdef CONFIG_SCHED_MC
7328 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7329 struct sysdev_class_attribute
*attr
,
7332 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7334 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7335 struct sysdev_class_attribute
*attr
,
7336 const char *buf
, size_t count
)
7338 return sched_power_savings_store(buf
, count
, 0);
7340 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7341 sched_mc_power_savings_show
,
7342 sched_mc_power_savings_store
);
7345 #ifdef CONFIG_SCHED_SMT
7346 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7347 struct sysdev_class_attribute
*attr
,
7350 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7352 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7353 struct sysdev_class_attribute
*attr
,
7354 const char *buf
, size_t count
)
7356 return sched_power_savings_store(buf
, count
, 1);
7358 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7359 sched_smt_power_savings_show
,
7360 sched_smt_power_savings_store
);
7363 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7367 #ifdef CONFIG_SCHED_SMT
7369 err
= sysfs_create_file(&cls
->kset
.kobj
,
7370 &attr_sched_smt_power_savings
.attr
);
7372 #ifdef CONFIG_SCHED_MC
7373 if (!err
&& mc_capable())
7374 err
= sysfs_create_file(&cls
->kset
.kobj
,
7375 &attr_sched_mc_power_savings
.attr
);
7379 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7381 #ifndef CONFIG_CPUSETS
7383 * Add online and remove offline CPUs from the scheduler domains.
7384 * When cpusets are enabled they take over this function.
7386 static int update_sched_domains(struct notifier_block
*nfb
,
7387 unsigned long action
, void *hcpu
)
7391 case CPU_ONLINE_FROZEN
:
7392 case CPU_DOWN_PREPARE
:
7393 case CPU_DOWN_PREPARE_FROZEN
:
7394 case CPU_DOWN_FAILED
:
7395 case CPU_DOWN_FAILED_FROZEN
:
7396 partition_sched_domains(1, NULL
, NULL
);
7405 static int update_runtime(struct notifier_block
*nfb
,
7406 unsigned long action
, void *hcpu
)
7408 int cpu
= (int)(long)hcpu
;
7411 case CPU_DOWN_PREPARE
:
7412 case CPU_DOWN_PREPARE_FROZEN
:
7413 disable_runtime(cpu_rq(cpu
));
7416 case CPU_DOWN_FAILED
:
7417 case CPU_DOWN_FAILED_FROZEN
:
7419 case CPU_ONLINE_FROZEN
:
7420 enable_runtime(cpu_rq(cpu
));
7428 void __init
sched_init_smp(void)
7430 cpumask_var_t non_isolated_cpus
;
7432 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7433 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7435 #if defined(CONFIG_NUMA)
7436 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7438 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7441 mutex_lock(&sched_domains_mutex
);
7442 arch_init_sched_domains(cpu_active_mask
);
7443 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7444 if (cpumask_empty(non_isolated_cpus
))
7445 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7446 mutex_unlock(&sched_domains_mutex
);
7449 #ifndef CONFIG_CPUSETS
7450 /* XXX: Theoretical race here - CPU may be hotplugged now */
7451 hotcpu_notifier(update_sched_domains
, 0);
7454 /* RT runtime code needs to handle some hotplug events */
7455 hotcpu_notifier(update_runtime
, 0);
7459 /* Move init over to a non-isolated CPU */
7460 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7462 sched_init_granularity();
7463 free_cpumask_var(non_isolated_cpus
);
7465 init_sched_rt_class();
7468 void __init
sched_init_smp(void)
7470 sched_init_granularity();
7472 #endif /* CONFIG_SMP */
7474 const_debug
unsigned int sysctl_timer_migration
= 1;
7476 int in_sched_functions(unsigned long addr
)
7478 return in_lock_functions(addr
) ||
7479 (addr
>= (unsigned long)__sched_text_start
7480 && addr
< (unsigned long)__sched_text_end
);
7483 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7485 cfs_rq
->tasks_timeline
= RB_ROOT
;
7486 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7487 #ifdef CONFIG_FAIR_GROUP_SCHED
7490 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7493 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7495 struct rt_prio_array
*array
;
7498 array
= &rt_rq
->active
;
7499 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7500 INIT_LIST_HEAD(array
->queue
+ i
);
7501 __clear_bit(i
, array
->bitmap
);
7503 /* delimiter for bitsearch: */
7504 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7506 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7507 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7509 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7513 rt_rq
->rt_nr_migratory
= 0;
7514 rt_rq
->overloaded
= 0;
7515 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7519 rt_rq
->rt_throttled
= 0;
7520 rt_rq
->rt_runtime
= 0;
7521 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7523 #ifdef CONFIG_RT_GROUP_SCHED
7524 rt_rq
->rt_nr_boosted
= 0;
7529 #ifdef CONFIG_FAIR_GROUP_SCHED
7530 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7531 struct sched_entity
*se
, int cpu
, int add
,
7532 struct sched_entity
*parent
)
7534 struct rq
*rq
= cpu_rq(cpu
);
7535 tg
->cfs_rq
[cpu
] = cfs_rq
;
7536 init_cfs_rq(cfs_rq
, rq
);
7539 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7542 /* se could be NULL for init_task_group */
7547 se
->cfs_rq
= &rq
->cfs
;
7549 se
->cfs_rq
= parent
->my_q
;
7552 se
->load
.weight
= tg
->shares
;
7553 se
->load
.inv_weight
= 0;
7554 se
->parent
= parent
;
7558 #ifdef CONFIG_RT_GROUP_SCHED
7559 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7560 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7561 struct sched_rt_entity
*parent
)
7563 struct rq
*rq
= cpu_rq(cpu
);
7565 tg
->rt_rq
[cpu
] = rt_rq
;
7566 init_rt_rq(rt_rq
, rq
);
7568 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7570 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7572 tg
->rt_se
[cpu
] = rt_se
;
7577 rt_se
->rt_rq
= &rq
->rt
;
7579 rt_se
->rt_rq
= parent
->my_q
;
7581 rt_se
->my_q
= rt_rq
;
7582 rt_se
->parent
= parent
;
7583 INIT_LIST_HEAD(&rt_se
->run_list
);
7587 void __init
sched_init(void)
7590 unsigned long alloc_size
= 0, ptr
;
7592 #ifdef CONFIG_FAIR_GROUP_SCHED
7593 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7595 #ifdef CONFIG_RT_GROUP_SCHED
7596 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7598 #ifdef CONFIG_CPUMASK_OFFSTACK
7599 alloc_size
+= num_possible_cpus() * cpumask_size();
7602 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7604 #ifdef CONFIG_FAIR_GROUP_SCHED
7605 init_task_group
.se
= (struct sched_entity
**)ptr
;
7606 ptr
+= nr_cpu_ids
* sizeof(void **);
7608 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7609 ptr
+= nr_cpu_ids
* sizeof(void **);
7611 #endif /* CONFIG_FAIR_GROUP_SCHED */
7612 #ifdef CONFIG_RT_GROUP_SCHED
7613 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7614 ptr
+= nr_cpu_ids
* sizeof(void **);
7616 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7617 ptr
+= nr_cpu_ids
* sizeof(void **);
7619 #endif /* CONFIG_RT_GROUP_SCHED */
7620 #ifdef CONFIG_CPUMASK_OFFSTACK
7621 for_each_possible_cpu(i
) {
7622 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7623 ptr
+= cpumask_size();
7625 #endif /* CONFIG_CPUMASK_OFFSTACK */
7629 init_defrootdomain();
7632 init_rt_bandwidth(&def_rt_bandwidth
,
7633 global_rt_period(), global_rt_runtime());
7635 #ifdef CONFIG_RT_GROUP_SCHED
7636 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7637 global_rt_period(), global_rt_runtime());
7638 #endif /* CONFIG_RT_GROUP_SCHED */
7640 #ifdef CONFIG_CGROUP_SCHED
7641 list_add(&init_task_group
.list
, &task_groups
);
7642 INIT_LIST_HEAD(&init_task_group
.children
);
7644 #endif /* CONFIG_CGROUP_SCHED */
7646 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7647 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7648 __alignof__(unsigned long));
7650 for_each_possible_cpu(i
) {
7654 raw_spin_lock_init(&rq
->lock
);
7656 rq
->calc_load_active
= 0;
7657 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7658 init_cfs_rq(&rq
->cfs
, rq
);
7659 init_rt_rq(&rq
->rt
, rq
);
7660 #ifdef CONFIG_FAIR_GROUP_SCHED
7661 init_task_group
.shares
= init_task_group_load
;
7662 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7663 #ifdef CONFIG_CGROUP_SCHED
7665 * How much cpu bandwidth does init_task_group get?
7667 * In case of task-groups formed thr' the cgroup filesystem, it
7668 * gets 100% of the cpu resources in the system. This overall
7669 * system cpu resource is divided among the tasks of
7670 * init_task_group and its child task-groups in a fair manner,
7671 * based on each entity's (task or task-group's) weight
7672 * (se->load.weight).
7674 * In other words, if init_task_group has 10 tasks of weight
7675 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7676 * then A0's share of the cpu resource is:
7678 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7680 * We achieve this by letting init_task_group's tasks sit
7681 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7683 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7685 #endif /* CONFIG_FAIR_GROUP_SCHED */
7687 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7688 #ifdef CONFIG_RT_GROUP_SCHED
7689 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7690 #ifdef CONFIG_CGROUP_SCHED
7691 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7695 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7696 rq
->cpu_load
[j
] = 0;
7700 rq
->post_schedule
= 0;
7701 rq
->active_balance
= 0;
7702 rq
->next_balance
= jiffies
;
7706 rq
->migration_thread
= NULL
;
7708 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7709 INIT_LIST_HEAD(&rq
->migration_queue
);
7710 rq_attach_root(rq
, &def_root_domain
);
7713 atomic_set(&rq
->nr_iowait
, 0);
7716 set_load_weight(&init_task
);
7718 #ifdef CONFIG_PREEMPT_NOTIFIERS
7719 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7723 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7726 #ifdef CONFIG_RT_MUTEXES
7727 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7731 * The boot idle thread does lazy MMU switching as well:
7733 atomic_inc(&init_mm
.mm_count
);
7734 enter_lazy_tlb(&init_mm
, current
);
7737 * Make us the idle thread. Technically, schedule() should not be
7738 * called from this thread, however somewhere below it might be,
7739 * but because we are the idle thread, we just pick up running again
7740 * when this runqueue becomes "idle".
7742 init_idle(current
, smp_processor_id());
7744 calc_load_update
= jiffies
+ LOAD_FREQ
;
7747 * During early bootup we pretend to be a normal task:
7749 current
->sched_class
= &fair_sched_class
;
7751 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7752 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7755 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7756 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7758 /* May be allocated at isolcpus cmdline parse time */
7759 if (cpu_isolated_map
== NULL
)
7760 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7765 scheduler_running
= 1;
7768 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7769 static inline int preempt_count_equals(int preempt_offset
)
7771 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7773 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7776 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7779 static unsigned long prev_jiffy
; /* ratelimiting */
7781 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7782 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7784 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7786 prev_jiffy
= jiffies
;
7789 "BUG: sleeping function called from invalid context at %s:%d\n",
7792 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7793 in_atomic(), irqs_disabled(),
7794 current
->pid
, current
->comm
);
7796 debug_show_held_locks(current
);
7797 if (irqs_disabled())
7798 print_irqtrace_events(current
);
7802 EXPORT_SYMBOL(__might_sleep
);
7805 #ifdef CONFIG_MAGIC_SYSRQ
7806 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7810 on_rq
= p
->se
.on_rq
;
7812 deactivate_task(rq
, p
, 0);
7813 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7815 activate_task(rq
, p
, 0);
7816 resched_task(rq
->curr
);
7820 void normalize_rt_tasks(void)
7822 struct task_struct
*g
, *p
;
7823 unsigned long flags
;
7826 read_lock_irqsave(&tasklist_lock
, flags
);
7827 do_each_thread(g
, p
) {
7829 * Only normalize user tasks:
7834 p
->se
.exec_start
= 0;
7835 #ifdef CONFIG_SCHEDSTATS
7836 p
->se
.statistics
.wait_start
= 0;
7837 p
->se
.statistics
.sleep_start
= 0;
7838 p
->se
.statistics
.block_start
= 0;
7843 * Renice negative nice level userspace
7846 if (TASK_NICE(p
) < 0 && p
->mm
)
7847 set_user_nice(p
, 0);
7851 raw_spin_lock(&p
->pi_lock
);
7852 rq
= __task_rq_lock(p
);
7854 normalize_task(rq
, p
);
7856 __task_rq_unlock(rq
);
7857 raw_spin_unlock(&p
->pi_lock
);
7858 } while_each_thread(g
, p
);
7860 read_unlock_irqrestore(&tasklist_lock
, flags
);
7863 #endif /* CONFIG_MAGIC_SYSRQ */
7867 * These functions are only useful for the IA64 MCA handling.
7869 * They can only be called when the whole system has been
7870 * stopped - every CPU needs to be quiescent, and no scheduling
7871 * activity can take place. Using them for anything else would
7872 * be a serious bug, and as a result, they aren't even visible
7873 * under any other configuration.
7877 * curr_task - return the current task for a given cpu.
7878 * @cpu: the processor in question.
7880 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7882 struct task_struct
*curr_task(int cpu
)
7884 return cpu_curr(cpu
);
7888 * set_curr_task - set the current task for a given cpu.
7889 * @cpu: the processor in question.
7890 * @p: the task pointer to set.
7892 * Description: This function must only be used when non-maskable interrupts
7893 * are serviced on a separate stack. It allows the architecture to switch the
7894 * notion of the current task on a cpu in a non-blocking manner. This function
7895 * must be called with all CPU's synchronized, and interrupts disabled, the
7896 * and caller must save the original value of the current task (see
7897 * curr_task() above) and restore that value before reenabling interrupts and
7898 * re-starting the system.
7900 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7902 void set_curr_task(int cpu
, struct task_struct
*p
)
7909 #ifdef CONFIG_FAIR_GROUP_SCHED
7910 static void free_fair_sched_group(struct task_group
*tg
)
7914 for_each_possible_cpu(i
) {
7916 kfree(tg
->cfs_rq
[i
]);
7926 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7928 struct cfs_rq
*cfs_rq
;
7929 struct sched_entity
*se
;
7933 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7936 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7940 tg
->shares
= NICE_0_LOAD
;
7942 for_each_possible_cpu(i
) {
7945 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7946 GFP_KERNEL
, cpu_to_node(i
));
7950 se
= kzalloc_node(sizeof(struct sched_entity
),
7951 GFP_KERNEL
, cpu_to_node(i
));
7955 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
7966 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7968 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7969 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7972 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7974 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7976 #else /* !CONFG_FAIR_GROUP_SCHED */
7977 static inline void free_fair_sched_group(struct task_group
*tg
)
7982 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7987 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7991 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7994 #endif /* CONFIG_FAIR_GROUP_SCHED */
7996 #ifdef CONFIG_RT_GROUP_SCHED
7997 static void free_rt_sched_group(struct task_group
*tg
)
8001 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8003 for_each_possible_cpu(i
) {
8005 kfree(tg
->rt_rq
[i
]);
8007 kfree(tg
->rt_se
[i
]);
8015 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8017 struct rt_rq
*rt_rq
;
8018 struct sched_rt_entity
*rt_se
;
8022 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8025 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8029 init_rt_bandwidth(&tg
->rt_bandwidth
,
8030 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8032 for_each_possible_cpu(i
) {
8035 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8036 GFP_KERNEL
, cpu_to_node(i
));
8040 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8041 GFP_KERNEL
, cpu_to_node(i
));
8045 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8056 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8058 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8059 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8062 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8064 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8066 #else /* !CONFIG_RT_GROUP_SCHED */
8067 static inline void free_rt_sched_group(struct task_group
*tg
)
8072 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8077 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8081 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8084 #endif /* CONFIG_RT_GROUP_SCHED */
8086 #ifdef CONFIG_CGROUP_SCHED
8087 static void free_sched_group(struct task_group
*tg
)
8089 free_fair_sched_group(tg
);
8090 free_rt_sched_group(tg
);
8094 /* allocate runqueue etc for a new task group */
8095 struct task_group
*sched_create_group(struct task_group
*parent
)
8097 struct task_group
*tg
;
8098 unsigned long flags
;
8101 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8103 return ERR_PTR(-ENOMEM
);
8105 if (!alloc_fair_sched_group(tg
, parent
))
8108 if (!alloc_rt_sched_group(tg
, parent
))
8111 spin_lock_irqsave(&task_group_lock
, flags
);
8112 for_each_possible_cpu(i
) {
8113 register_fair_sched_group(tg
, i
);
8114 register_rt_sched_group(tg
, i
);
8116 list_add_rcu(&tg
->list
, &task_groups
);
8118 WARN_ON(!parent
); /* root should already exist */
8120 tg
->parent
= parent
;
8121 INIT_LIST_HEAD(&tg
->children
);
8122 list_add_rcu(&tg
->siblings
, &parent
->children
);
8123 spin_unlock_irqrestore(&task_group_lock
, flags
);
8128 free_sched_group(tg
);
8129 return ERR_PTR(-ENOMEM
);
8132 /* rcu callback to free various structures associated with a task group */
8133 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8135 /* now it should be safe to free those cfs_rqs */
8136 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8139 /* Destroy runqueue etc associated with a task group */
8140 void sched_destroy_group(struct task_group
*tg
)
8142 unsigned long flags
;
8145 spin_lock_irqsave(&task_group_lock
, flags
);
8146 for_each_possible_cpu(i
) {
8147 unregister_fair_sched_group(tg
, i
);
8148 unregister_rt_sched_group(tg
, i
);
8150 list_del_rcu(&tg
->list
);
8151 list_del_rcu(&tg
->siblings
);
8152 spin_unlock_irqrestore(&task_group_lock
, flags
);
8154 /* wait for possible concurrent references to cfs_rqs complete */
8155 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8158 /* change task's runqueue when it moves between groups.
8159 * The caller of this function should have put the task in its new group
8160 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8161 * reflect its new group.
8163 void sched_move_task(struct task_struct
*tsk
)
8166 unsigned long flags
;
8169 rq
= task_rq_lock(tsk
, &flags
);
8171 running
= task_current(rq
, tsk
);
8172 on_rq
= tsk
->se
.on_rq
;
8175 dequeue_task(rq
, tsk
, 0);
8176 if (unlikely(running
))
8177 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8179 set_task_rq(tsk
, task_cpu(tsk
));
8181 #ifdef CONFIG_FAIR_GROUP_SCHED
8182 if (tsk
->sched_class
->moved_group
)
8183 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8186 if (unlikely(running
))
8187 tsk
->sched_class
->set_curr_task(rq
);
8189 enqueue_task(rq
, tsk
, 0, false);
8191 task_rq_unlock(rq
, &flags
);
8193 #endif /* CONFIG_CGROUP_SCHED */
8195 #ifdef CONFIG_FAIR_GROUP_SCHED
8196 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8198 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8203 dequeue_entity(cfs_rq
, se
, 0);
8205 se
->load
.weight
= shares
;
8206 se
->load
.inv_weight
= 0;
8209 enqueue_entity(cfs_rq
, se
, 0);
8212 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8214 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8215 struct rq
*rq
= cfs_rq
->rq
;
8216 unsigned long flags
;
8218 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8219 __set_se_shares(se
, shares
);
8220 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8223 static DEFINE_MUTEX(shares_mutex
);
8225 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8228 unsigned long flags
;
8231 * We can't change the weight of the root cgroup.
8236 if (shares
< MIN_SHARES
)
8237 shares
= MIN_SHARES
;
8238 else if (shares
> MAX_SHARES
)
8239 shares
= MAX_SHARES
;
8241 mutex_lock(&shares_mutex
);
8242 if (tg
->shares
== shares
)
8245 spin_lock_irqsave(&task_group_lock
, flags
);
8246 for_each_possible_cpu(i
)
8247 unregister_fair_sched_group(tg
, i
);
8248 list_del_rcu(&tg
->siblings
);
8249 spin_unlock_irqrestore(&task_group_lock
, flags
);
8251 /* wait for any ongoing reference to this group to finish */
8252 synchronize_sched();
8255 * Now we are free to modify the group's share on each cpu
8256 * w/o tripping rebalance_share or load_balance_fair.
8258 tg
->shares
= shares
;
8259 for_each_possible_cpu(i
) {
8263 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8264 set_se_shares(tg
->se
[i
], shares
);
8268 * Enable load balance activity on this group, by inserting it back on
8269 * each cpu's rq->leaf_cfs_rq_list.
8271 spin_lock_irqsave(&task_group_lock
, flags
);
8272 for_each_possible_cpu(i
)
8273 register_fair_sched_group(tg
, i
);
8274 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8275 spin_unlock_irqrestore(&task_group_lock
, flags
);
8277 mutex_unlock(&shares_mutex
);
8281 unsigned long sched_group_shares(struct task_group
*tg
)
8287 #ifdef CONFIG_RT_GROUP_SCHED
8289 * Ensure that the real time constraints are schedulable.
8291 static DEFINE_MUTEX(rt_constraints_mutex
);
8293 static unsigned long to_ratio(u64 period
, u64 runtime
)
8295 if (runtime
== RUNTIME_INF
)
8298 return div64_u64(runtime
<< 20, period
);
8301 /* Must be called with tasklist_lock held */
8302 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8304 struct task_struct
*g
, *p
;
8306 do_each_thread(g
, p
) {
8307 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8309 } while_each_thread(g
, p
);
8314 struct rt_schedulable_data
{
8315 struct task_group
*tg
;
8320 static int tg_schedulable(struct task_group
*tg
, void *data
)
8322 struct rt_schedulable_data
*d
= data
;
8323 struct task_group
*child
;
8324 unsigned long total
, sum
= 0;
8325 u64 period
, runtime
;
8327 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8328 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8331 period
= d
->rt_period
;
8332 runtime
= d
->rt_runtime
;
8336 * Cannot have more runtime than the period.
8338 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8342 * Ensure we don't starve existing RT tasks.
8344 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8347 total
= to_ratio(period
, runtime
);
8350 * Nobody can have more than the global setting allows.
8352 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8356 * The sum of our children's runtime should not exceed our own.
8358 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8359 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8360 runtime
= child
->rt_bandwidth
.rt_runtime
;
8362 if (child
== d
->tg
) {
8363 period
= d
->rt_period
;
8364 runtime
= d
->rt_runtime
;
8367 sum
+= to_ratio(period
, runtime
);
8376 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8378 struct rt_schedulable_data data
= {
8380 .rt_period
= period
,
8381 .rt_runtime
= runtime
,
8384 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8387 static int tg_set_bandwidth(struct task_group
*tg
,
8388 u64 rt_period
, u64 rt_runtime
)
8392 mutex_lock(&rt_constraints_mutex
);
8393 read_lock(&tasklist_lock
);
8394 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8398 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8399 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8400 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8402 for_each_possible_cpu(i
) {
8403 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8405 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8406 rt_rq
->rt_runtime
= rt_runtime
;
8407 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8409 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8411 read_unlock(&tasklist_lock
);
8412 mutex_unlock(&rt_constraints_mutex
);
8417 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8419 u64 rt_runtime
, rt_period
;
8421 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8422 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8423 if (rt_runtime_us
< 0)
8424 rt_runtime
= RUNTIME_INF
;
8426 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8429 long sched_group_rt_runtime(struct task_group
*tg
)
8433 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8436 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8437 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8438 return rt_runtime_us
;
8441 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8443 u64 rt_runtime
, rt_period
;
8445 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8446 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8451 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8454 long sched_group_rt_period(struct task_group
*tg
)
8458 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8459 do_div(rt_period_us
, NSEC_PER_USEC
);
8460 return rt_period_us
;
8463 static int sched_rt_global_constraints(void)
8465 u64 runtime
, period
;
8468 if (sysctl_sched_rt_period
<= 0)
8471 runtime
= global_rt_runtime();
8472 period
= global_rt_period();
8475 * Sanity check on the sysctl variables.
8477 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8480 mutex_lock(&rt_constraints_mutex
);
8481 read_lock(&tasklist_lock
);
8482 ret
= __rt_schedulable(NULL
, 0, 0);
8483 read_unlock(&tasklist_lock
);
8484 mutex_unlock(&rt_constraints_mutex
);
8489 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8491 /* Don't accept realtime tasks when there is no way for them to run */
8492 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8498 #else /* !CONFIG_RT_GROUP_SCHED */
8499 static int sched_rt_global_constraints(void)
8501 unsigned long flags
;
8504 if (sysctl_sched_rt_period
<= 0)
8508 * There's always some RT tasks in the root group
8509 * -- migration, kstopmachine etc..
8511 if (sysctl_sched_rt_runtime
== 0)
8514 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8515 for_each_possible_cpu(i
) {
8516 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8518 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8519 rt_rq
->rt_runtime
= global_rt_runtime();
8520 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8522 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8526 #endif /* CONFIG_RT_GROUP_SCHED */
8528 int sched_rt_handler(struct ctl_table
*table
, int write
,
8529 void __user
*buffer
, size_t *lenp
,
8533 int old_period
, old_runtime
;
8534 static DEFINE_MUTEX(mutex
);
8537 old_period
= sysctl_sched_rt_period
;
8538 old_runtime
= sysctl_sched_rt_runtime
;
8540 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8542 if (!ret
&& write
) {
8543 ret
= sched_rt_global_constraints();
8545 sysctl_sched_rt_period
= old_period
;
8546 sysctl_sched_rt_runtime
= old_runtime
;
8548 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8549 def_rt_bandwidth
.rt_period
=
8550 ns_to_ktime(global_rt_period());
8553 mutex_unlock(&mutex
);
8558 #ifdef CONFIG_CGROUP_SCHED
8560 /* return corresponding task_group object of a cgroup */
8561 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8563 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8564 struct task_group
, css
);
8567 static struct cgroup_subsys_state
*
8568 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8570 struct task_group
*tg
, *parent
;
8572 if (!cgrp
->parent
) {
8573 /* This is early initialization for the top cgroup */
8574 return &init_task_group
.css
;
8577 parent
= cgroup_tg(cgrp
->parent
);
8578 tg
= sched_create_group(parent
);
8580 return ERR_PTR(-ENOMEM
);
8586 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8588 struct task_group
*tg
= cgroup_tg(cgrp
);
8590 sched_destroy_group(tg
);
8594 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8596 #ifdef CONFIG_RT_GROUP_SCHED
8597 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8600 /* We don't support RT-tasks being in separate groups */
8601 if (tsk
->sched_class
!= &fair_sched_class
)
8608 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8609 struct task_struct
*tsk
, bool threadgroup
)
8611 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8615 struct task_struct
*c
;
8617 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8618 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8630 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8631 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8634 sched_move_task(tsk
);
8636 struct task_struct
*c
;
8638 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8645 #ifdef CONFIG_FAIR_GROUP_SCHED
8646 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8649 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8652 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8654 struct task_group
*tg
= cgroup_tg(cgrp
);
8656 return (u64
) tg
->shares
;
8658 #endif /* CONFIG_FAIR_GROUP_SCHED */
8660 #ifdef CONFIG_RT_GROUP_SCHED
8661 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8664 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8667 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8669 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8672 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8675 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8678 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8680 return sched_group_rt_period(cgroup_tg(cgrp
));
8682 #endif /* CONFIG_RT_GROUP_SCHED */
8684 static struct cftype cpu_files
[] = {
8685 #ifdef CONFIG_FAIR_GROUP_SCHED
8688 .read_u64
= cpu_shares_read_u64
,
8689 .write_u64
= cpu_shares_write_u64
,
8692 #ifdef CONFIG_RT_GROUP_SCHED
8694 .name
= "rt_runtime_us",
8695 .read_s64
= cpu_rt_runtime_read
,
8696 .write_s64
= cpu_rt_runtime_write
,
8699 .name
= "rt_period_us",
8700 .read_u64
= cpu_rt_period_read_uint
,
8701 .write_u64
= cpu_rt_period_write_uint
,
8706 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8708 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8711 struct cgroup_subsys cpu_cgroup_subsys
= {
8713 .create
= cpu_cgroup_create
,
8714 .destroy
= cpu_cgroup_destroy
,
8715 .can_attach
= cpu_cgroup_can_attach
,
8716 .attach
= cpu_cgroup_attach
,
8717 .populate
= cpu_cgroup_populate
,
8718 .subsys_id
= cpu_cgroup_subsys_id
,
8722 #endif /* CONFIG_CGROUP_SCHED */
8724 #ifdef CONFIG_CGROUP_CPUACCT
8727 * CPU accounting code for task groups.
8729 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8730 * (balbir@in.ibm.com).
8733 /* track cpu usage of a group of tasks and its child groups */
8735 struct cgroup_subsys_state css
;
8736 /* cpuusage holds pointer to a u64-type object on every cpu */
8737 u64 __percpu
*cpuusage
;
8738 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8739 struct cpuacct
*parent
;
8742 struct cgroup_subsys cpuacct_subsys
;
8744 /* return cpu accounting group corresponding to this container */
8745 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8747 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8748 struct cpuacct
, css
);
8751 /* return cpu accounting group to which this task belongs */
8752 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8754 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8755 struct cpuacct
, css
);
8758 /* create a new cpu accounting group */
8759 static struct cgroup_subsys_state
*cpuacct_create(
8760 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8762 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8768 ca
->cpuusage
= alloc_percpu(u64
);
8772 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8773 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8774 goto out_free_counters
;
8777 ca
->parent
= cgroup_ca(cgrp
->parent
);
8783 percpu_counter_destroy(&ca
->cpustat
[i
]);
8784 free_percpu(ca
->cpuusage
);
8788 return ERR_PTR(-ENOMEM
);
8791 /* destroy an existing cpu accounting group */
8793 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8795 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8798 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8799 percpu_counter_destroy(&ca
->cpustat
[i
]);
8800 free_percpu(ca
->cpuusage
);
8804 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8806 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8809 #ifndef CONFIG_64BIT
8811 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8813 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8815 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8823 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8825 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8827 #ifndef CONFIG_64BIT
8829 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8831 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8833 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8839 /* return total cpu usage (in nanoseconds) of a group */
8840 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8842 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8843 u64 totalcpuusage
= 0;
8846 for_each_present_cpu(i
)
8847 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8849 return totalcpuusage
;
8852 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8855 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8864 for_each_present_cpu(i
)
8865 cpuacct_cpuusage_write(ca
, i
, 0);
8871 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8874 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8878 for_each_present_cpu(i
) {
8879 percpu
= cpuacct_cpuusage_read(ca
, i
);
8880 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8882 seq_printf(m
, "\n");
8886 static const char *cpuacct_stat_desc
[] = {
8887 [CPUACCT_STAT_USER
] = "user",
8888 [CPUACCT_STAT_SYSTEM
] = "system",
8891 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8892 struct cgroup_map_cb
*cb
)
8894 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8897 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8898 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8899 val
= cputime64_to_clock_t(val
);
8900 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8905 static struct cftype files
[] = {
8908 .read_u64
= cpuusage_read
,
8909 .write_u64
= cpuusage_write
,
8912 .name
= "usage_percpu",
8913 .read_seq_string
= cpuacct_percpu_seq_read
,
8917 .read_map
= cpuacct_stats_show
,
8921 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8923 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8927 * charge this task's execution time to its accounting group.
8929 * called with rq->lock held.
8931 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8936 if (unlikely(!cpuacct_subsys
.active
))
8939 cpu
= task_cpu(tsk
);
8945 for (; ca
; ca
= ca
->parent
) {
8946 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8947 *cpuusage
+= cputime
;
8954 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8955 * in cputime_t units. As a result, cpuacct_update_stats calls
8956 * percpu_counter_add with values large enough to always overflow the
8957 * per cpu batch limit causing bad SMP scalability.
8959 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8960 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8961 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8964 #define CPUACCT_BATCH \
8965 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8967 #define CPUACCT_BATCH 0
8971 * Charge the system/user time to the task's accounting group.
8973 static void cpuacct_update_stats(struct task_struct
*tsk
,
8974 enum cpuacct_stat_index idx
, cputime_t val
)
8977 int batch
= CPUACCT_BATCH
;
8979 if (unlikely(!cpuacct_subsys
.active
))
8986 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
8992 struct cgroup_subsys cpuacct_subsys
= {
8994 .create
= cpuacct_create
,
8995 .destroy
= cpuacct_destroy
,
8996 .populate
= cpuacct_populate
,
8997 .subsys_id
= cpuacct_subsys_id
,
8999 #endif /* CONFIG_CGROUP_CPUACCT */
9003 int rcu_expedited_torture_stats(char *page
)
9007 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9009 void synchronize_sched_expedited(void)
9012 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9014 #else /* #ifndef CONFIG_SMP */
9016 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9017 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9019 #define RCU_EXPEDITED_STATE_POST -2
9020 #define RCU_EXPEDITED_STATE_IDLE -1
9022 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9024 int rcu_expedited_torture_stats(char *page
)
9029 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9030 for_each_online_cpu(cpu
) {
9031 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9032 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9034 cnt
+= sprintf(&page
[cnt
], "\n");
9037 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9039 static long synchronize_sched_expedited_count
;
9042 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9043 * approach to force grace period to end quickly. This consumes
9044 * significant time on all CPUs, and is thus not recommended for
9045 * any sort of common-case code.
9047 * Note that it is illegal to call this function while holding any
9048 * lock that is acquired by a CPU-hotplug notifier. Failing to
9049 * observe this restriction will result in deadlock.
9051 void synchronize_sched_expedited(void)
9054 unsigned long flags
;
9055 bool need_full_sync
= 0;
9057 struct migration_req
*req
;
9061 smp_mb(); /* ensure prior mod happens before capturing snap. */
9062 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9064 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9066 if (trycount
++ < 10)
9067 udelay(trycount
* num_online_cpus());
9069 synchronize_sched();
9072 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9073 smp_mb(); /* ensure test happens before caller kfree */
9078 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9079 for_each_online_cpu(cpu
) {
9081 req
= &per_cpu(rcu_migration_req
, cpu
);
9082 init_completion(&req
->done
);
9084 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9085 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9086 list_add(&req
->list
, &rq
->migration_queue
);
9087 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9088 wake_up_process(rq
->migration_thread
);
9090 for_each_online_cpu(cpu
) {
9091 rcu_expedited_state
= cpu
;
9092 req
= &per_cpu(rcu_migration_req
, cpu
);
9094 wait_for_completion(&req
->done
);
9095 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9096 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9098 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9099 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9101 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9102 synchronize_sched_expedited_count
++;
9103 mutex_unlock(&rcu_sched_expedited_mutex
);
9106 synchronize_sched();
9108 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9110 #endif /* #else #ifndef CONFIG_SMP */