4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
78 #include <asm/mutex.h>
80 #include "sched_cpupri.h"
81 #include "workqueue_sched.h"
82 #include "sched_autogroup.h"
84 #define CREATE_TRACE_POINTS
85 #include <trace/events/sched.h>
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
126 static inline int rt_policy(int policy
)
128 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
133 static inline int task_has_rt_policy(struct task_struct
*p
)
135 return rt_policy(p
->policy
);
139 * This is the priority-queue data structure of the RT scheduling class:
141 struct rt_prio_array
{
142 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
143 struct list_head queue
[MAX_RT_PRIO
];
146 struct rt_bandwidth
{
147 /* nests inside the rq lock: */
148 raw_spinlock_t rt_runtime_lock
;
151 struct hrtimer rt_period_timer
;
154 static struct rt_bandwidth def_rt_bandwidth
;
156 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
158 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
160 struct rt_bandwidth
*rt_b
=
161 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
167 now
= hrtimer_cb_get_time(timer
);
168 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
173 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
176 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
180 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
182 rt_b
->rt_period
= ns_to_ktime(period
);
183 rt_b
->rt_runtime
= runtime
;
185 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
187 hrtimer_init(&rt_b
->rt_period_timer
,
188 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
189 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
192 static inline int rt_bandwidth_enabled(void)
194 return sysctl_sched_rt_runtime
>= 0;
197 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
201 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
204 if (hrtimer_active(&rt_b
->rt_period_timer
))
207 raw_spin_lock(&rt_b
->rt_runtime_lock
);
212 if (hrtimer_active(&rt_b
->rt_period_timer
))
215 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
216 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
218 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
219 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
220 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
221 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
222 HRTIMER_MODE_ABS_PINNED
, 0);
224 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
227 #ifdef CONFIG_RT_GROUP_SCHED
228 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
230 hrtimer_cancel(&rt_b
->rt_period_timer
);
235 * sched_domains_mutex serializes calls to arch_init_sched_domains,
236 * detach_destroy_domains and partition_sched_domains.
238 static DEFINE_MUTEX(sched_domains_mutex
);
240 #ifdef CONFIG_CGROUP_SCHED
242 #include <linux/cgroup.h>
246 static LIST_HEAD(task_groups
);
248 /* task group related information */
250 struct cgroup_subsys_state css
;
252 #ifdef CONFIG_FAIR_GROUP_SCHED
253 /* schedulable entities of this group on each cpu */
254 struct sched_entity
**se
;
255 /* runqueue "owned" by this group on each cpu */
256 struct cfs_rq
**cfs_rq
;
257 unsigned long shares
;
259 atomic_t load_weight
;
262 #ifdef CONFIG_RT_GROUP_SCHED
263 struct sched_rt_entity
**rt_se
;
264 struct rt_rq
**rt_rq
;
266 struct rt_bandwidth rt_bandwidth
;
270 struct list_head list
;
272 struct task_group
*parent
;
273 struct list_head siblings
;
274 struct list_head children
;
276 #ifdef CONFIG_SCHED_AUTOGROUP
277 struct autogroup
*autogroup
;
281 /* task_group_lock serializes the addition/removal of task groups */
282 static DEFINE_SPINLOCK(task_group_lock
);
284 #ifdef CONFIG_FAIR_GROUP_SCHED
286 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
289 * A weight of 0 or 1 can cause arithmetics problems.
290 * A weight of a cfs_rq is the sum of weights of which entities
291 * are queued on this cfs_rq, so a weight of a entity should not be
292 * too large, so as the shares value of a task group.
293 * (The default weight is 1024 - so there's no practical
294 * limitation from this.)
297 #define MAX_SHARES (1UL << 18)
299 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
302 /* Default task group.
303 * Every task in system belong to this group at bootup.
305 struct task_group root_task_group
;
307 #endif /* CONFIG_CGROUP_SCHED */
309 /* CFS-related fields in a runqueue */
311 struct load_weight load
;
312 unsigned long nr_running
;
317 struct rb_root tasks_timeline
;
318 struct rb_node
*rb_leftmost
;
320 struct list_head tasks
;
321 struct list_head
*balance_iterator
;
324 * 'curr' points to currently running entity on this cfs_rq.
325 * It is set to NULL otherwise (i.e when none are currently running).
327 struct sched_entity
*curr
, *next
, *last
, *skip
;
329 unsigned int nr_spread_over
;
331 #ifdef CONFIG_FAIR_GROUP_SCHED
332 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
335 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
336 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
337 * (like users, containers etc.)
339 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
340 * list is used during load balance.
343 struct list_head leaf_cfs_rq_list
;
344 struct task_group
*tg
; /* group that "owns" this runqueue */
348 * the part of load.weight contributed by tasks
350 unsigned long task_weight
;
353 * h_load = weight * f(tg)
355 * Where f(tg) is the recursive weight fraction assigned to
358 unsigned long h_load
;
361 * Maintaining per-cpu shares distribution for group scheduling
363 * load_stamp is the last time we updated the load average
364 * load_last is the last time we updated the load average and saw load
365 * load_unacc_exec_time is currently unaccounted execution time
369 u64 load_stamp
, load_last
, load_unacc_exec_time
;
371 unsigned long load_contribution
;
376 /* Real-Time classes' related field in a runqueue: */
378 struct rt_prio_array active
;
379 unsigned long rt_nr_running
;
380 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
382 int curr
; /* highest queued rt task prio */
384 int next
; /* next highest */
389 unsigned long rt_nr_migratory
;
390 unsigned long rt_nr_total
;
392 struct plist_head pushable_tasks
;
397 /* Nests inside the rq lock: */
398 raw_spinlock_t rt_runtime_lock
;
400 #ifdef CONFIG_RT_GROUP_SCHED
401 unsigned long rt_nr_boosted
;
404 struct list_head leaf_rt_rq_list
;
405 struct task_group
*tg
;
412 * We add the notion of a root-domain which will be used to define per-domain
413 * variables. Each exclusive cpuset essentially defines an island domain by
414 * fully partitioning the member cpus from any other cpuset. Whenever a new
415 * exclusive cpuset is created, we also create and attach a new root-domain
422 cpumask_var_t online
;
425 * The "RT overload" flag: it gets set if a CPU has more than
426 * one runnable RT task.
428 cpumask_var_t rto_mask
;
430 struct cpupri cpupri
;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain
;
439 #endif /* CONFIG_SMP */
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running
;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
459 unsigned long last_load_update_tick
;
462 unsigned char nohz_balance_kick
;
464 unsigned int skip_clock_update
;
466 /* capture load from *all* tasks on this cpu: */
467 struct load_weight load
;
468 unsigned long nr_load_updates
;
474 #ifdef CONFIG_FAIR_GROUP_SCHED
475 /* list of leaf cfs_rq on this cpu: */
476 struct list_head leaf_cfs_rq_list
;
478 #ifdef CONFIG_RT_GROUP_SCHED
479 struct list_head leaf_rt_rq_list
;
483 * This is part of a global counter where only the total sum
484 * over all CPUs matters. A task can increase this counter on
485 * one CPU and if it got migrated afterwards it may decrease
486 * it on another CPU. Always updated under the runqueue lock:
488 unsigned long nr_uninterruptible
;
490 struct task_struct
*curr
, *idle
, *stop
;
491 unsigned long next_balance
;
492 struct mm_struct
*prev_mm
;
500 struct root_domain
*rd
;
501 struct sched_domain
*sd
;
503 unsigned long cpu_power
;
505 unsigned char idle_at_tick
;
506 /* For active balancing */
510 struct cpu_stop_work active_balance_work
;
511 /* cpu of this runqueue: */
515 unsigned long avg_load_per_task
;
523 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
527 /* calc_load related fields */
528 unsigned long calc_load_update
;
529 long calc_load_active
;
531 #ifdef CONFIG_SCHED_HRTICK
533 int hrtick_csd_pending
;
534 struct call_single_data hrtick_csd
;
536 struct hrtimer hrtick_timer
;
539 #ifdef CONFIG_SCHEDSTATS
541 struct sched_info rq_sched_info
;
542 unsigned long long rq_cpu_time
;
543 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
545 /* sys_sched_yield() stats */
546 unsigned int yld_count
;
548 /* schedule() stats */
549 unsigned int sched_switch
;
550 unsigned int sched_count
;
551 unsigned int sched_goidle
;
553 /* try_to_wake_up() stats */
554 unsigned int ttwu_count
;
555 unsigned int ttwu_local
;
559 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
562 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
564 static inline int cpu_of(struct rq
*rq
)
573 #define rcu_dereference_check_sched_domain(p) \
574 rcu_dereference_check((p), \
575 rcu_read_lock_sched_held() || \
576 lockdep_is_held(&sched_domains_mutex))
579 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
580 * See detach_destroy_domains: synchronize_sched for details.
582 * The domain tree of any CPU may only be accessed from within
583 * preempt-disabled sections.
585 #define for_each_domain(cpu, __sd) \
586 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
588 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
589 #define this_rq() (&__get_cpu_var(runqueues))
590 #define task_rq(p) cpu_rq(task_cpu(p))
591 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
592 #define raw_rq() (&__raw_get_cpu_var(runqueues))
594 #ifdef CONFIG_CGROUP_SCHED
597 * Return the group to which this tasks belongs.
599 * We use task_subsys_state_check() and extend the RCU verification
600 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
601 * holds that lock for each task it moves into the cgroup. Therefore
602 * by holding that lock, we pin the task to the current cgroup.
604 static inline struct task_group
*task_group(struct task_struct
*p
)
606 struct task_group
*tg
;
607 struct cgroup_subsys_state
*css
;
609 if (p
->flags
& PF_EXITING
)
610 return &root_task_group
;
612 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
613 lockdep_is_held(&task_rq(p
)->lock
));
614 tg
= container_of(css
, struct task_group
, css
);
616 return autogroup_task_group(p
, tg
);
619 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
620 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
622 #ifdef CONFIG_FAIR_GROUP_SCHED
623 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
624 p
->se
.parent
= task_group(p
)->se
[cpu
];
627 #ifdef CONFIG_RT_GROUP_SCHED
628 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
629 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
633 #else /* CONFIG_CGROUP_SCHED */
635 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
636 static inline struct task_group
*task_group(struct task_struct
*p
)
641 #endif /* CONFIG_CGROUP_SCHED */
643 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
645 static void update_rq_clock(struct rq
*rq
)
649 if (rq
->skip_clock_update
)
652 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
654 update_rq_clock_task(rq
, delta
);
658 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
660 #ifdef CONFIG_SCHED_DEBUG
661 # define const_debug __read_mostly
663 # define const_debug static const
668 * @cpu: the processor in question.
670 * Returns true if the current cpu runqueue is locked.
671 * This interface allows printk to be called with the runqueue lock
672 * held and know whether or not it is OK to wake up the klogd.
674 int runqueue_is_locked(int cpu
)
676 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug
unsigned int sysctl_sched_features
=
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly
char *sched_feat_names
[] = {
706 #include "sched_features.h"
712 static int sched_feat_show(struct seq_file
*m
, void *v
)
716 for (i
= 0; sched_feat_names
[i
]; i
++) {
717 if (!(sysctl_sched_features
& (1UL << i
)))
719 seq_printf(m
, "%s ", sched_feat_names
[i
]);
727 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
728 size_t cnt
, loff_t
*ppos
)
738 if (copy_from_user(&buf
, ubuf
, cnt
))
744 if (strncmp(cmp
, "NO_", 3) == 0) {
749 for (i
= 0; sched_feat_names
[i
]; i
++) {
750 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
752 sysctl_sched_features
&= ~(1UL << i
);
754 sysctl_sched_features
|= (1UL << i
);
759 if (!sched_feat_names
[i
])
767 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
769 return single_open(filp
, sched_feat_show
, NULL
);
772 static const struct file_operations sched_feat_fops
= {
773 .open
= sched_feat_open
,
774 .write
= sched_feat_write
,
777 .release
= single_release
,
780 static __init
int sched_init_debug(void)
782 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
787 late_initcall(sched_init_debug
);
791 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
794 * Number of tasks to iterate in a single balance run.
795 * Limited because this is done with IRQs disabled.
797 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
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 ->cpus_allowed
920 static inline int task_is_waking(struct task_struct
*p
)
922 return unlikely(p
->state
== TASK_WAKING
);
926 * __task_rq_lock - lock the runqueue a given task resides on.
927 * Must be called interrupts disabled.
929 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
936 raw_spin_lock(&rq
->lock
);
937 if (likely(rq
== task_rq(p
)))
939 raw_spin_unlock(&rq
->lock
);
944 * task_rq_lock - lock the runqueue a given task resides on and disable
945 * interrupts. Note the ordering: we can safely lookup the task_rq without
946 * explicitly disabling preemption.
948 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
954 local_irq_save(*flags
);
956 raw_spin_lock(&rq
->lock
);
957 if (likely(rq
== task_rq(p
)))
959 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
963 static void __task_rq_unlock(struct rq
*rq
)
966 raw_spin_unlock(&rq
->lock
);
969 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
972 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
976 * this_rq_lock - lock this runqueue and disable interrupts.
978 static struct rq
*this_rq_lock(void)
985 raw_spin_lock(&rq
->lock
);
990 #ifdef CONFIG_SCHED_HRTICK
992 * Use HR-timers to deliver accurate preemption points.
994 * Its all a bit involved since we cannot program an hrt while holding the
995 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
998 * When we get rescheduled we reprogram the hrtick_timer outside of the
1004 * - enabled by features
1005 * - hrtimer is actually high res
1007 static inline int hrtick_enabled(struct rq
*rq
)
1009 if (!sched_feat(HRTICK
))
1011 if (!cpu_active(cpu_of(rq
)))
1013 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1016 static void hrtick_clear(struct rq
*rq
)
1018 if (hrtimer_active(&rq
->hrtick_timer
))
1019 hrtimer_cancel(&rq
->hrtick_timer
);
1023 * High-resolution timer tick.
1024 * Runs from hardirq context with interrupts disabled.
1026 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1028 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1030 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1032 raw_spin_lock(&rq
->lock
);
1033 update_rq_clock(rq
);
1034 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1035 raw_spin_unlock(&rq
->lock
);
1037 return HRTIMER_NORESTART
;
1042 * called from hardirq (IPI) context
1044 static void __hrtick_start(void *arg
)
1046 struct rq
*rq
= arg
;
1048 raw_spin_lock(&rq
->lock
);
1049 hrtimer_restart(&rq
->hrtick_timer
);
1050 rq
->hrtick_csd_pending
= 0;
1051 raw_spin_unlock(&rq
->lock
);
1055 * Called to set the hrtick timer state.
1057 * called with rq->lock held and irqs disabled
1059 static void hrtick_start(struct rq
*rq
, u64 delay
)
1061 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1062 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1064 hrtimer_set_expires(timer
, time
);
1066 if (rq
== this_rq()) {
1067 hrtimer_restart(timer
);
1068 } else if (!rq
->hrtick_csd_pending
) {
1069 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1070 rq
->hrtick_csd_pending
= 1;
1075 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1077 int cpu
= (int)(long)hcpu
;
1080 case CPU_UP_CANCELED
:
1081 case CPU_UP_CANCELED_FROZEN
:
1082 case CPU_DOWN_PREPARE
:
1083 case CPU_DOWN_PREPARE_FROZEN
:
1085 case CPU_DEAD_FROZEN
:
1086 hrtick_clear(cpu_rq(cpu
));
1093 static __init
void init_hrtick(void)
1095 hotcpu_notifier(hotplug_hrtick
, 0);
1099 * Called to set the hrtick timer state.
1101 * called with rq->lock held and irqs disabled
1103 static void hrtick_start(struct rq
*rq
, u64 delay
)
1105 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1106 HRTIMER_MODE_REL_PINNED
, 0);
1109 static inline void init_hrtick(void)
1112 #endif /* CONFIG_SMP */
1114 static void init_rq_hrtick(struct rq
*rq
)
1117 rq
->hrtick_csd_pending
= 0;
1119 rq
->hrtick_csd
.flags
= 0;
1120 rq
->hrtick_csd
.func
= __hrtick_start
;
1121 rq
->hrtick_csd
.info
= rq
;
1124 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1125 rq
->hrtick_timer
.function
= hrtick
;
1127 #else /* CONFIG_SCHED_HRTICK */
1128 static inline void hrtick_clear(struct rq
*rq
)
1132 static inline void init_rq_hrtick(struct rq
*rq
)
1136 static inline void init_hrtick(void)
1139 #endif /* CONFIG_SCHED_HRTICK */
1142 * resched_task - mark a task 'to be rescheduled now'.
1144 * On UP this means the setting of the need_resched flag, on SMP it
1145 * might also involve a cross-CPU call to trigger the scheduler on
1150 #ifndef tsk_is_polling
1151 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1154 static void resched_task(struct task_struct
*p
)
1158 assert_raw_spin_locked(&task_rq(p
)->lock
);
1160 if (test_tsk_need_resched(p
))
1163 set_tsk_need_resched(p
);
1166 if (cpu
== smp_processor_id())
1169 /* NEED_RESCHED must be visible before we test polling */
1171 if (!tsk_is_polling(p
))
1172 smp_send_reschedule(cpu
);
1175 static void resched_cpu(int cpu
)
1177 struct rq
*rq
= cpu_rq(cpu
);
1178 unsigned long flags
;
1180 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1182 resched_task(cpu_curr(cpu
));
1183 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1188 * In the semi idle case, use the nearest busy cpu for migrating timers
1189 * from an idle cpu. This is good for power-savings.
1191 * We don't do similar optimization for completely idle system, as
1192 * selecting an idle cpu will add more delays to the timers than intended
1193 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1195 int get_nohz_timer_target(void)
1197 int cpu
= smp_processor_id();
1199 struct sched_domain
*sd
;
1201 for_each_domain(cpu
, sd
) {
1202 for_each_cpu(i
, sched_domain_span(sd
))
1209 * When add_timer_on() enqueues a timer into the timer wheel of an
1210 * idle CPU then this timer might expire before the next timer event
1211 * which is scheduled to wake up that CPU. In case of a completely
1212 * idle system the next event might even be infinite time into the
1213 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1214 * leaves the inner idle loop so the newly added timer is taken into
1215 * account when the CPU goes back to idle and evaluates the timer
1216 * wheel for the next timer event.
1218 void wake_up_idle_cpu(int cpu
)
1220 struct rq
*rq
= cpu_rq(cpu
);
1222 if (cpu
== smp_processor_id())
1226 * This is safe, as this function is called with the timer
1227 * wheel base lock of (cpu) held. When the CPU is on the way
1228 * to idle and has not yet set rq->curr to idle then it will
1229 * be serialized on the timer wheel base lock and take the new
1230 * timer into account automatically.
1232 if (rq
->curr
!= rq
->idle
)
1236 * We can set TIF_RESCHED on the idle task of the other CPU
1237 * lockless. The worst case is that the other CPU runs the
1238 * idle task through an additional NOOP schedule()
1240 set_tsk_need_resched(rq
->idle
);
1242 /* NEED_RESCHED must be visible before we test polling */
1244 if (!tsk_is_polling(rq
->idle
))
1245 smp_send_reschedule(cpu
);
1248 #endif /* CONFIG_NO_HZ */
1250 static u64
sched_avg_period(void)
1252 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1255 static void sched_avg_update(struct rq
*rq
)
1257 s64 period
= sched_avg_period();
1259 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1261 * Inline assembly required to prevent the compiler
1262 * optimising this loop into a divmod call.
1263 * See __iter_div_u64_rem() for another example of this.
1265 asm("" : "+rm" (rq
->age_stamp
));
1266 rq
->age_stamp
+= period
;
1271 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1273 rq
->rt_avg
+= rt_delta
;
1274 sched_avg_update(rq
);
1277 #else /* !CONFIG_SMP */
1278 static void resched_task(struct task_struct
*p
)
1280 assert_raw_spin_locked(&task_rq(p
)->lock
);
1281 set_tsk_need_resched(p
);
1284 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1288 static void sched_avg_update(struct rq
*rq
)
1291 #endif /* CONFIG_SMP */
1293 #if BITS_PER_LONG == 32
1294 # define WMULT_CONST (~0UL)
1296 # define WMULT_CONST (1UL << 32)
1299 #define WMULT_SHIFT 32
1302 * Shift right and round:
1304 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1307 * delta *= weight / lw
1309 static unsigned long
1310 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1311 struct load_weight
*lw
)
1315 if (!lw
->inv_weight
) {
1316 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1319 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1323 tmp
= (u64
)delta_exec
* weight
;
1325 * Check whether we'd overflow the 64-bit multiplication:
1327 if (unlikely(tmp
> WMULT_CONST
))
1328 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1331 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1333 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1336 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1342 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1348 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
1355 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1356 * of tasks with abnormal "nice" values across CPUs the contribution that
1357 * each task makes to its run queue's load is weighted according to its
1358 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1359 * scaled version of the new time slice allocation that they receive on time
1363 #define WEIGHT_IDLEPRIO 3
1364 #define WMULT_IDLEPRIO 1431655765
1367 * Nice levels are multiplicative, with a gentle 10% change for every
1368 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1369 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1370 * that remained on nice 0.
1372 * The "10% effect" is relative and cumulative: from _any_ nice level,
1373 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1374 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1375 * If a task goes up by ~10% and another task goes down by ~10% then
1376 * the relative distance between them is ~25%.)
1378 static const int prio_to_weight
[40] = {
1379 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1380 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1381 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1382 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1383 /* 0 */ 1024, 820, 655, 526, 423,
1384 /* 5 */ 335, 272, 215, 172, 137,
1385 /* 10 */ 110, 87, 70, 56, 45,
1386 /* 15 */ 36, 29, 23, 18, 15,
1390 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1392 * In cases where the weight does not change often, we can use the
1393 * precalculated inverse to speed up arithmetics by turning divisions
1394 * into multiplications:
1396 static const u32 prio_to_wmult
[40] = {
1397 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1398 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1399 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1400 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1401 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1402 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1403 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1404 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1407 /* Time spent by the tasks of the cpu accounting group executing in ... */
1408 enum cpuacct_stat_index
{
1409 CPUACCT_STAT_USER
, /* ... user mode */
1410 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1412 CPUACCT_STAT_NSTATS
,
1415 #ifdef CONFIG_CGROUP_CPUACCT
1416 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1417 static void cpuacct_update_stats(struct task_struct
*tsk
,
1418 enum cpuacct_stat_index idx
, cputime_t val
);
1420 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1421 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1422 enum cpuacct_stat_index idx
, cputime_t val
) {}
1425 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1427 update_load_add(&rq
->load
, load
);
1430 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1432 update_load_sub(&rq
->load
, load
);
1435 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1436 typedef int (*tg_visitor
)(struct task_group
*, void *);
1439 * Iterate the full tree, calling @down when first entering a node and @up when
1440 * leaving it for the final time.
1442 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1444 struct task_group
*parent
, *child
;
1448 parent
= &root_task_group
;
1450 ret
= (*down
)(parent
, data
);
1453 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1460 ret
= (*up
)(parent
, data
);
1465 parent
= parent
->parent
;
1474 static int tg_nop(struct task_group
*tg
, void *data
)
1481 /* Used instead of source_load when we know the type == 0 */
1482 static unsigned long weighted_cpuload(const int cpu
)
1484 return cpu_rq(cpu
)->load
.weight
;
1488 * Return a low guess at the load of a migration-source cpu weighted
1489 * according to the scheduling class and "nice" value.
1491 * We want to under-estimate the load of migration sources, to
1492 * balance conservatively.
1494 static unsigned long source_load(int cpu
, int type
)
1496 struct rq
*rq
= cpu_rq(cpu
);
1497 unsigned long total
= weighted_cpuload(cpu
);
1499 if (type
== 0 || !sched_feat(LB_BIAS
))
1502 return min(rq
->cpu_load
[type
-1], total
);
1506 * Return a high guess at the load of a migration-target cpu weighted
1507 * according to the scheduling class and "nice" value.
1509 static unsigned long target_load(int cpu
, int type
)
1511 struct rq
*rq
= cpu_rq(cpu
);
1512 unsigned long total
= weighted_cpuload(cpu
);
1514 if (type
== 0 || !sched_feat(LB_BIAS
))
1517 return max(rq
->cpu_load
[type
-1], total
);
1520 static unsigned long power_of(int cpu
)
1522 return cpu_rq(cpu
)->cpu_power
;
1525 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1527 static unsigned long cpu_avg_load_per_task(int cpu
)
1529 struct rq
*rq
= cpu_rq(cpu
);
1530 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1533 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1535 rq
->avg_load_per_task
= 0;
1537 return rq
->avg_load_per_task
;
1540 #ifdef CONFIG_FAIR_GROUP_SCHED
1543 * Compute the cpu's hierarchical load factor for each task group.
1544 * This needs to be done in a top-down fashion because the load of a child
1545 * group is a fraction of its parents load.
1547 static int tg_load_down(struct task_group
*tg
, void *data
)
1550 long cpu
= (long)data
;
1553 load
= cpu_rq(cpu
)->load
.weight
;
1555 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1556 load
*= tg
->se
[cpu
]->load
.weight
;
1557 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1560 tg
->cfs_rq
[cpu
]->h_load
= load
;
1565 static void update_h_load(long cpu
)
1567 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1572 #ifdef CONFIG_PREEMPT
1574 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1577 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1578 * way at the expense of forcing extra atomic operations in all
1579 * invocations. This assures that the double_lock is acquired using the
1580 * same underlying policy as the spinlock_t on this architecture, which
1581 * reduces latency compared to the unfair variant below. However, it
1582 * also adds more overhead and therefore may reduce throughput.
1584 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1585 __releases(this_rq
->lock
)
1586 __acquires(busiest
->lock
)
1587 __acquires(this_rq
->lock
)
1589 raw_spin_unlock(&this_rq
->lock
);
1590 double_rq_lock(this_rq
, busiest
);
1597 * Unfair double_lock_balance: Optimizes throughput at the expense of
1598 * latency by eliminating extra atomic operations when the locks are
1599 * already in proper order on entry. This favors lower cpu-ids and will
1600 * grant the double lock to lower cpus over higher ids under contention,
1601 * regardless of entry order into the function.
1603 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1604 __releases(this_rq
->lock
)
1605 __acquires(busiest
->lock
)
1606 __acquires(this_rq
->lock
)
1610 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1611 if (busiest
< this_rq
) {
1612 raw_spin_unlock(&this_rq
->lock
);
1613 raw_spin_lock(&busiest
->lock
);
1614 raw_spin_lock_nested(&this_rq
->lock
,
1615 SINGLE_DEPTH_NESTING
);
1618 raw_spin_lock_nested(&busiest
->lock
,
1619 SINGLE_DEPTH_NESTING
);
1624 #endif /* CONFIG_PREEMPT */
1627 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1629 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1631 if (unlikely(!irqs_disabled())) {
1632 /* printk() doesn't work good under rq->lock */
1633 raw_spin_unlock(&this_rq
->lock
);
1637 return _double_lock_balance(this_rq
, busiest
);
1640 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1641 __releases(busiest
->lock
)
1643 raw_spin_unlock(&busiest
->lock
);
1644 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1648 * double_rq_lock - safely lock two runqueues
1650 * Note this does not disable interrupts like task_rq_lock,
1651 * you need to do so manually before calling.
1653 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1654 __acquires(rq1
->lock
)
1655 __acquires(rq2
->lock
)
1657 BUG_ON(!irqs_disabled());
1659 raw_spin_lock(&rq1
->lock
);
1660 __acquire(rq2
->lock
); /* Fake it out ;) */
1663 raw_spin_lock(&rq1
->lock
);
1664 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1666 raw_spin_lock(&rq2
->lock
);
1667 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1673 * double_rq_unlock - safely unlock two runqueues
1675 * Note this does not restore interrupts like task_rq_unlock,
1676 * you need to do so manually after calling.
1678 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1679 __releases(rq1
->lock
)
1680 __releases(rq2
->lock
)
1682 raw_spin_unlock(&rq1
->lock
);
1684 raw_spin_unlock(&rq2
->lock
);
1686 __release(rq2
->lock
);
1691 static void calc_load_account_idle(struct rq
*this_rq
);
1692 static void update_sysctl(void);
1693 static int get_update_sysctl_factor(void);
1694 static void update_cpu_load(struct rq
*this_rq
);
1696 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1698 set_task_rq(p
, cpu
);
1701 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1702 * successfuly executed on another CPU. We must ensure that updates of
1703 * per-task data have been completed by this moment.
1706 task_thread_info(p
)->cpu
= cpu
;
1710 static const struct sched_class rt_sched_class
;
1712 #define sched_class_highest (&stop_sched_class)
1713 #define for_each_class(class) \
1714 for (class = sched_class_highest; class; class = class->next)
1716 #include "sched_stats.h"
1718 static void inc_nr_running(struct rq
*rq
)
1723 static void dec_nr_running(struct rq
*rq
)
1728 static void set_load_weight(struct task_struct
*p
)
1731 * SCHED_IDLE tasks get minimal weight:
1733 if (p
->policy
== SCHED_IDLE
) {
1734 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1735 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1739 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1740 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1743 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1745 update_rq_clock(rq
);
1746 sched_info_queued(p
);
1747 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1751 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1753 update_rq_clock(rq
);
1754 sched_info_dequeued(p
);
1755 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1760 * activate_task - move a task to the runqueue.
1762 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1764 if (task_contributes_to_load(p
))
1765 rq
->nr_uninterruptible
--;
1767 enqueue_task(rq
, p
, flags
);
1772 * deactivate_task - remove a task from the runqueue.
1774 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1776 if (task_contributes_to_load(p
))
1777 rq
->nr_uninterruptible
++;
1779 dequeue_task(rq
, p
, flags
);
1783 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1786 * There are no locks covering percpu hardirq/softirq time.
1787 * They are only modified in account_system_vtime, on corresponding CPU
1788 * with interrupts disabled. So, writes are safe.
1789 * They are read and saved off onto struct rq in update_rq_clock().
1790 * This may result in other CPU reading this CPU's irq time and can
1791 * race with irq/account_system_vtime on this CPU. We would either get old
1792 * or new value with a side effect of accounting a slice of irq time to wrong
1793 * task when irq is in progress while we read rq->clock. That is a worthy
1794 * compromise in place of having locks on each irq in account_system_time.
1796 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1797 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1799 static DEFINE_PER_CPU(u64
, irq_start_time
);
1800 static int sched_clock_irqtime
;
1802 void enable_sched_clock_irqtime(void)
1804 sched_clock_irqtime
= 1;
1807 void disable_sched_clock_irqtime(void)
1809 sched_clock_irqtime
= 0;
1812 #ifndef CONFIG_64BIT
1813 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
1815 static inline void irq_time_write_begin(void)
1817 __this_cpu_inc(irq_time_seq
.sequence
);
1821 static inline void irq_time_write_end(void)
1824 __this_cpu_inc(irq_time_seq
.sequence
);
1827 static inline u64
irq_time_read(int cpu
)
1833 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1834 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1835 per_cpu(cpu_hardirq_time
, cpu
);
1836 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1840 #else /* CONFIG_64BIT */
1841 static inline void irq_time_write_begin(void)
1845 static inline void irq_time_write_end(void)
1849 static inline u64
irq_time_read(int cpu
)
1851 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1853 #endif /* CONFIG_64BIT */
1856 * Called before incrementing preempt_count on {soft,}irq_enter
1857 * and before decrementing preempt_count on {soft,}irq_exit.
1859 void account_system_vtime(struct task_struct
*curr
)
1861 unsigned long flags
;
1865 if (!sched_clock_irqtime
)
1868 local_irq_save(flags
);
1870 cpu
= smp_processor_id();
1871 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
1872 __this_cpu_add(irq_start_time
, delta
);
1874 irq_time_write_begin();
1876 * We do not account for softirq time from ksoftirqd here.
1877 * We want to continue accounting softirq time to ksoftirqd thread
1878 * in that case, so as not to confuse scheduler with a special task
1879 * that do not consume any time, but still wants to run.
1881 if (hardirq_count())
1882 __this_cpu_add(cpu_hardirq_time
, delta
);
1883 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
1884 __this_cpu_add(cpu_softirq_time
, delta
);
1886 irq_time_write_end();
1887 local_irq_restore(flags
);
1889 EXPORT_SYMBOL_GPL(account_system_vtime
);
1891 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1895 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
1898 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1899 * this case when a previous update_rq_clock() happened inside a
1900 * {soft,}irq region.
1902 * When this happens, we stop ->clock_task and only update the
1903 * prev_irq_time stamp to account for the part that fit, so that a next
1904 * update will consume the rest. This ensures ->clock_task is
1907 * It does however cause some slight miss-attribution of {soft,}irq
1908 * time, a more accurate solution would be to update the irq_time using
1909 * the current rq->clock timestamp, except that would require using
1912 if (irq_delta
> delta
)
1915 rq
->prev_irq_time
+= irq_delta
;
1917 rq
->clock_task
+= delta
;
1919 if (irq_delta
&& sched_feat(NONIRQ_POWER
))
1920 sched_rt_avg_update(rq
, irq_delta
);
1923 static int irqtime_account_hi_update(void)
1925 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
1926 unsigned long flags
;
1930 local_irq_save(flags
);
1931 latest_ns
= this_cpu_read(cpu_hardirq_time
);
1932 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->irq
))
1934 local_irq_restore(flags
);
1938 static int irqtime_account_si_update(void)
1940 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
1941 unsigned long flags
;
1945 local_irq_save(flags
);
1946 latest_ns
= this_cpu_read(cpu_softirq_time
);
1947 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->softirq
))
1949 local_irq_restore(flags
);
1953 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1955 #define sched_clock_irqtime (0)
1957 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1959 rq
->clock_task
+= delta
;
1962 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
1964 #include "sched_idletask.c"
1965 #include "sched_fair.c"
1966 #include "sched_rt.c"
1967 #include "sched_autogroup.c"
1968 #include "sched_stoptask.c"
1969 #ifdef CONFIG_SCHED_DEBUG
1970 # include "sched_debug.c"
1973 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
1975 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
1976 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
1980 * Make it appear like a SCHED_FIFO task, its something
1981 * userspace knows about and won't get confused about.
1983 * Also, it will make PI more or less work without too
1984 * much confusion -- but then, stop work should not
1985 * rely on PI working anyway.
1987 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
1989 stop
->sched_class
= &stop_sched_class
;
1992 cpu_rq(cpu
)->stop
= stop
;
1996 * Reset it back to a normal scheduling class so that
1997 * it can die in pieces.
1999 old_stop
->sched_class
= &rt_sched_class
;
2004 * __normal_prio - return the priority that is based on the static prio
2006 static inline int __normal_prio(struct task_struct
*p
)
2008 return p
->static_prio
;
2012 * Calculate the expected normal priority: i.e. priority
2013 * without taking RT-inheritance into account. Might be
2014 * boosted by interactivity modifiers. Changes upon fork,
2015 * setprio syscalls, and whenever the interactivity
2016 * estimator recalculates.
2018 static inline int normal_prio(struct task_struct
*p
)
2022 if (task_has_rt_policy(p
))
2023 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2025 prio
= __normal_prio(p
);
2030 * Calculate the current priority, i.e. the priority
2031 * taken into account by the scheduler. This value might
2032 * be boosted by RT tasks, or might be boosted by
2033 * interactivity modifiers. Will be RT if the task got
2034 * RT-boosted. If not then it returns p->normal_prio.
2036 static int effective_prio(struct task_struct
*p
)
2038 p
->normal_prio
= normal_prio(p
);
2040 * If we are RT tasks or we were boosted to RT priority,
2041 * keep the priority unchanged. Otherwise, update priority
2042 * to the normal priority:
2044 if (!rt_prio(p
->prio
))
2045 return p
->normal_prio
;
2050 * task_curr - is this task currently executing on a CPU?
2051 * @p: the task in question.
2053 inline int task_curr(const struct task_struct
*p
)
2055 return cpu_curr(task_cpu(p
)) == p
;
2058 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2059 const struct sched_class
*prev_class
,
2062 if (prev_class
!= p
->sched_class
) {
2063 if (prev_class
->switched_from
)
2064 prev_class
->switched_from(rq
, p
);
2065 p
->sched_class
->switched_to(rq
, p
);
2066 } else if (oldprio
!= p
->prio
)
2067 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2070 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2072 const struct sched_class
*class;
2074 if (p
->sched_class
== rq
->curr
->sched_class
) {
2075 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2077 for_each_class(class) {
2078 if (class == rq
->curr
->sched_class
)
2080 if (class == p
->sched_class
) {
2081 resched_task(rq
->curr
);
2088 * A queue event has occurred, and we're going to schedule. In
2089 * this case, we can save a useless back to back clock update.
2091 if (rq
->curr
->se
.on_rq
&& test_tsk_need_resched(rq
->curr
))
2092 rq
->skip_clock_update
= 1;
2097 * Is this task likely cache-hot:
2100 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2104 if (p
->sched_class
!= &fair_sched_class
)
2107 if (unlikely(p
->policy
== SCHED_IDLE
))
2111 * Buddy candidates are cache hot:
2113 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2114 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2115 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2118 if (sysctl_sched_migration_cost
== -1)
2120 if (sysctl_sched_migration_cost
== 0)
2123 delta
= now
- p
->se
.exec_start
;
2125 return delta
< (s64
)sysctl_sched_migration_cost
;
2128 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2130 #ifdef CONFIG_SCHED_DEBUG
2132 * We should never call set_task_cpu() on a blocked task,
2133 * ttwu() will sort out the placement.
2135 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2136 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2139 trace_sched_migrate_task(p
, new_cpu
);
2141 if (task_cpu(p
) != new_cpu
) {
2142 p
->se
.nr_migrations
++;
2143 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2146 __set_task_cpu(p
, new_cpu
);
2149 struct migration_arg
{
2150 struct task_struct
*task
;
2154 static int migration_cpu_stop(void *data
);
2157 * The task's runqueue lock must be held.
2158 * Returns true if you have to wait for migration thread.
2160 static bool migrate_task(struct task_struct
*p
, struct rq
*rq
)
2163 * If the task is not on a runqueue (and not running), then
2164 * the next wake-up will properly place the task.
2166 return p
->se
.on_rq
|| task_running(rq
, p
);
2170 * wait_task_inactive - wait for a thread to unschedule.
2172 * If @match_state is nonzero, it's the @p->state value just checked and
2173 * not expected to change. If it changes, i.e. @p might have woken up,
2174 * then return zero. When we succeed in waiting for @p to be off its CPU,
2175 * we return a positive number (its total switch count). If a second call
2176 * a short while later returns the same number, the caller can be sure that
2177 * @p has remained unscheduled the whole time.
2179 * The caller must ensure that the task *will* unschedule sometime soon,
2180 * else this function might spin for a *long* time. This function can't
2181 * be called with interrupts off, or it may introduce deadlock with
2182 * smp_call_function() if an IPI is sent by the same process we are
2183 * waiting to become inactive.
2185 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2187 unsigned long flags
;
2194 * We do the initial early heuristics without holding
2195 * any task-queue locks at all. We'll only try to get
2196 * the runqueue lock when things look like they will
2202 * If the task is actively running on another CPU
2203 * still, just relax and busy-wait without holding
2206 * NOTE! Since we don't hold any locks, it's not
2207 * even sure that "rq" stays as the right runqueue!
2208 * But we don't care, since "task_running()" will
2209 * return false if the runqueue has changed and p
2210 * is actually now running somewhere else!
2212 while (task_running(rq
, p
)) {
2213 if (match_state
&& unlikely(p
->state
!= match_state
))
2219 * Ok, time to look more closely! We need the rq
2220 * lock now, to be *sure*. If we're wrong, we'll
2221 * just go back and repeat.
2223 rq
= task_rq_lock(p
, &flags
);
2224 trace_sched_wait_task(p
);
2225 running
= task_running(rq
, p
);
2226 on_rq
= p
->se
.on_rq
;
2228 if (!match_state
|| p
->state
== match_state
)
2229 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2230 task_rq_unlock(rq
, &flags
);
2233 * If it changed from the expected state, bail out now.
2235 if (unlikely(!ncsw
))
2239 * Was it really running after all now that we
2240 * checked with the proper locks actually held?
2242 * Oops. Go back and try again..
2244 if (unlikely(running
)) {
2250 * It's not enough that it's not actively running,
2251 * it must be off the runqueue _entirely_, and not
2254 * So if it was still runnable (but just not actively
2255 * running right now), it's preempted, and we should
2256 * yield - it could be a while.
2258 if (unlikely(on_rq
)) {
2259 schedule_timeout_uninterruptible(1);
2264 * Ahh, all good. It wasn't running, and it wasn't
2265 * runnable, which means that it will never become
2266 * running in the future either. We're all done!
2275 * kick_process - kick a running thread to enter/exit the kernel
2276 * @p: the to-be-kicked thread
2278 * Cause a process which is running on another CPU to enter
2279 * kernel-mode, without any delay. (to get signals handled.)
2281 * NOTE: this function doesnt have to take the runqueue lock,
2282 * because all it wants to ensure is that the remote task enters
2283 * the kernel. If the IPI races and the task has been migrated
2284 * to another CPU then no harm is done and the purpose has been
2287 void kick_process(struct task_struct
*p
)
2293 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2294 smp_send_reschedule(cpu
);
2297 EXPORT_SYMBOL_GPL(kick_process
);
2298 #endif /* CONFIG_SMP */
2301 * task_oncpu_function_call - call a function on the cpu on which a task runs
2302 * @p: the task to evaluate
2303 * @func: the function to be called
2304 * @info: the function call argument
2306 * Calls the function @func when the task is currently running. This might
2307 * be on the current CPU, which just calls the function directly
2309 void task_oncpu_function_call(struct task_struct
*p
,
2310 void (*func
) (void *info
), void *info
)
2317 smp_call_function_single(cpu
, func
, info
, 1);
2323 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2325 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2328 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2330 /* Look for allowed, online CPU in same node. */
2331 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2332 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2335 /* Any allowed, online CPU? */
2336 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2337 if (dest_cpu
< nr_cpu_ids
)
2340 /* No more Mr. Nice Guy. */
2341 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2343 * Don't tell them about moving exiting tasks or
2344 * kernel threads (both mm NULL), since they never
2347 if (p
->mm
&& printk_ratelimit()) {
2348 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2349 task_pid_nr(p
), p
->comm
, cpu
);
2356 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2359 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2361 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2364 * In order not to call set_task_cpu() on a blocking task we need
2365 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2368 * Since this is common to all placement strategies, this lives here.
2370 * [ this allows ->select_task() to simply return task_cpu(p) and
2371 * not worry about this generic constraint ]
2373 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2375 cpu
= select_fallback_rq(task_cpu(p
), p
);
2380 static void update_avg(u64
*avg
, u64 sample
)
2382 s64 diff
= sample
- *avg
;
2387 static inline void ttwu_activate(struct task_struct
*p
, struct rq
*rq
,
2388 bool is_sync
, bool is_migrate
, bool is_local
,
2389 unsigned long en_flags
)
2391 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2393 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2395 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2397 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2399 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2401 activate_task(rq
, p
, en_flags
);
2404 static inline void ttwu_post_activation(struct task_struct
*p
, struct rq
*rq
,
2405 int wake_flags
, bool success
)
2407 trace_sched_wakeup(p
, success
);
2408 check_preempt_curr(rq
, p
, wake_flags
);
2410 p
->state
= TASK_RUNNING
;
2412 if (p
->sched_class
->task_woken
)
2413 p
->sched_class
->task_woken(rq
, p
);
2415 if (unlikely(rq
->idle_stamp
)) {
2416 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2417 u64 max
= 2*sysctl_sched_migration_cost
;
2422 update_avg(&rq
->avg_idle
, delta
);
2426 /* if a worker is waking up, notify workqueue */
2427 if ((p
->flags
& PF_WQ_WORKER
) && success
)
2428 wq_worker_waking_up(p
, cpu_of(rq
));
2432 * try_to_wake_up - wake up a thread
2433 * @p: the thread to be awakened
2434 * @state: the mask of task states that can be woken
2435 * @wake_flags: wake modifier flags (WF_*)
2437 * Put it on the run-queue if it's not already there. The "current"
2438 * thread is always on the run-queue (except when the actual
2439 * re-schedule is in progress), and as such you're allowed to do
2440 * the simpler "current->state = TASK_RUNNING" to mark yourself
2441 * runnable without the overhead of this.
2443 * Returns %true if @p was woken up, %false if it was already running
2444 * or @state didn't match @p's state.
2446 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2449 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2450 unsigned long flags
;
2451 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2454 this_cpu
= get_cpu();
2457 rq
= task_rq_lock(p
, &flags
);
2458 if (!(p
->state
& state
))
2468 if (unlikely(task_running(rq
, p
)))
2472 * In order to handle concurrent wakeups and release the rq->lock
2473 * we put the task in TASK_WAKING state.
2475 * First fix up the nr_uninterruptible count:
2477 if (task_contributes_to_load(p
)) {
2478 if (likely(cpu_online(orig_cpu
)))
2479 rq
->nr_uninterruptible
--;
2481 this_rq()->nr_uninterruptible
--;
2483 p
->state
= TASK_WAKING
;
2485 if (p
->sched_class
->task_waking
) {
2486 p
->sched_class
->task_waking(rq
, p
);
2487 en_flags
|= ENQUEUE_WAKING
;
2490 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2491 if (cpu
!= orig_cpu
)
2492 set_task_cpu(p
, cpu
);
2493 __task_rq_unlock(rq
);
2496 raw_spin_lock(&rq
->lock
);
2499 * We migrated the task without holding either rq->lock, however
2500 * since the task is not on the task list itself, nobody else
2501 * will try and migrate the task, hence the rq should match the
2502 * cpu we just moved it to.
2504 WARN_ON(task_cpu(p
) != cpu
);
2505 WARN_ON(p
->state
!= TASK_WAKING
);
2507 #ifdef CONFIG_SCHEDSTATS
2508 schedstat_inc(rq
, ttwu_count
);
2509 if (cpu
== this_cpu
)
2510 schedstat_inc(rq
, ttwu_local
);
2512 struct sched_domain
*sd
;
2513 for_each_domain(this_cpu
, sd
) {
2514 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2515 schedstat_inc(sd
, ttwu_wake_remote
);
2520 #endif /* CONFIG_SCHEDSTATS */
2523 #endif /* CONFIG_SMP */
2524 ttwu_activate(p
, rq
, wake_flags
& WF_SYNC
, orig_cpu
!= cpu
,
2525 cpu
== this_cpu
, en_flags
);
2528 ttwu_post_activation(p
, rq
, wake_flags
, success
);
2530 task_rq_unlock(rq
, &flags
);
2537 * try_to_wake_up_local - try to wake up a local task with rq lock held
2538 * @p: the thread to be awakened
2540 * Put @p on the run-queue if it's not already there. The caller must
2541 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2542 * the current task. this_rq() stays locked over invocation.
2544 static void try_to_wake_up_local(struct task_struct
*p
)
2546 struct rq
*rq
= task_rq(p
);
2547 bool success
= false;
2549 BUG_ON(rq
!= this_rq());
2550 BUG_ON(p
== current
);
2551 lockdep_assert_held(&rq
->lock
);
2553 if (!(p
->state
& TASK_NORMAL
))
2557 if (likely(!task_running(rq
, p
))) {
2558 schedstat_inc(rq
, ttwu_count
);
2559 schedstat_inc(rq
, ttwu_local
);
2561 ttwu_activate(p
, rq
, false, false, true, ENQUEUE_WAKEUP
);
2564 ttwu_post_activation(p
, rq
, 0, success
);
2568 * wake_up_process - Wake up a specific process
2569 * @p: The process to be woken up.
2571 * Attempt to wake up the nominated process and move it to the set of runnable
2572 * processes. Returns 1 if the process was woken up, 0 if it was already
2575 * It may be assumed that this function implies a write memory barrier before
2576 * changing the task state if and only if any tasks are woken up.
2578 int wake_up_process(struct task_struct
*p
)
2580 return try_to_wake_up(p
, TASK_ALL
, 0);
2582 EXPORT_SYMBOL(wake_up_process
);
2584 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2586 return try_to_wake_up(p
, state
, 0);
2590 * Perform scheduler related setup for a newly forked process p.
2591 * p is forked by current.
2593 * __sched_fork() is basic setup used by init_idle() too:
2595 static void __sched_fork(struct task_struct
*p
)
2597 p
->se
.exec_start
= 0;
2598 p
->se
.sum_exec_runtime
= 0;
2599 p
->se
.prev_sum_exec_runtime
= 0;
2600 p
->se
.nr_migrations
= 0;
2603 #ifdef CONFIG_SCHEDSTATS
2604 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2607 INIT_LIST_HEAD(&p
->rt
.run_list
);
2609 INIT_LIST_HEAD(&p
->se
.group_node
);
2611 #ifdef CONFIG_PREEMPT_NOTIFIERS
2612 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2617 * fork()/clone()-time setup:
2619 void sched_fork(struct task_struct
*p
, int clone_flags
)
2621 int cpu
= get_cpu();
2625 * We mark the process as running here. This guarantees that
2626 * nobody will actually run it, and a signal or other external
2627 * event cannot wake it up and insert it on the runqueue either.
2629 p
->state
= TASK_RUNNING
;
2632 * Revert to default priority/policy on fork if requested.
2634 if (unlikely(p
->sched_reset_on_fork
)) {
2635 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2636 p
->policy
= SCHED_NORMAL
;
2637 p
->normal_prio
= p
->static_prio
;
2640 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2641 p
->static_prio
= NICE_TO_PRIO(0);
2642 p
->normal_prio
= p
->static_prio
;
2647 * We don't need the reset flag anymore after the fork. It has
2648 * fulfilled its duty:
2650 p
->sched_reset_on_fork
= 0;
2654 * Make sure we do not leak PI boosting priority to the child.
2656 p
->prio
= current
->normal_prio
;
2658 if (!rt_prio(p
->prio
))
2659 p
->sched_class
= &fair_sched_class
;
2661 if (p
->sched_class
->task_fork
)
2662 p
->sched_class
->task_fork(p
);
2665 * The child is not yet in the pid-hash so no cgroup attach races,
2666 * and the cgroup is pinned to this child due to cgroup_fork()
2667 * is ran before sched_fork().
2669 * Silence PROVE_RCU.
2672 set_task_cpu(p
, cpu
);
2675 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2676 if (likely(sched_info_on()))
2677 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2679 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2682 #ifdef CONFIG_PREEMPT
2683 /* Want to start with kernel preemption disabled. */
2684 task_thread_info(p
)->preempt_count
= 1;
2687 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2694 * wake_up_new_task - wake up a newly created task for the first time.
2696 * This function will do some initial scheduler statistics housekeeping
2697 * that must be done for every newly created context, then puts the task
2698 * on the runqueue and wakes it.
2700 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2702 unsigned long flags
;
2704 int cpu __maybe_unused
= get_cpu();
2707 rq
= task_rq_lock(p
, &flags
);
2708 p
->state
= TASK_WAKING
;
2711 * Fork balancing, do it here and not earlier because:
2712 * - cpus_allowed can change in the fork path
2713 * - any previously selected cpu might disappear through hotplug
2715 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2716 * without people poking at ->cpus_allowed.
2718 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2719 set_task_cpu(p
, cpu
);
2721 p
->state
= TASK_RUNNING
;
2722 task_rq_unlock(rq
, &flags
);
2725 rq
= task_rq_lock(p
, &flags
);
2726 activate_task(rq
, p
, 0);
2727 trace_sched_wakeup_new(p
, 1);
2728 check_preempt_curr(rq
, p
, WF_FORK
);
2730 if (p
->sched_class
->task_woken
)
2731 p
->sched_class
->task_woken(rq
, p
);
2733 task_rq_unlock(rq
, &flags
);
2737 #ifdef CONFIG_PREEMPT_NOTIFIERS
2740 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2741 * @notifier: notifier struct to register
2743 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2745 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2747 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2750 * preempt_notifier_unregister - no longer interested in preemption notifications
2751 * @notifier: notifier struct to unregister
2753 * This is safe to call from within a preemption notifier.
2755 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2757 hlist_del(¬ifier
->link
);
2759 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2761 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2763 struct preempt_notifier
*notifier
;
2764 struct hlist_node
*node
;
2766 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2767 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2771 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2772 struct task_struct
*next
)
2774 struct preempt_notifier
*notifier
;
2775 struct hlist_node
*node
;
2777 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2778 notifier
->ops
->sched_out(notifier
, next
);
2781 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2783 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2788 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2789 struct task_struct
*next
)
2793 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2796 * prepare_task_switch - prepare to switch tasks
2797 * @rq: the runqueue preparing to switch
2798 * @prev: the current task that is being switched out
2799 * @next: the task we are going to switch to.
2801 * This is called with the rq lock held and interrupts off. It must
2802 * be paired with a subsequent finish_task_switch after the context
2805 * prepare_task_switch sets up locking and calls architecture specific
2809 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2810 struct task_struct
*next
)
2812 fire_sched_out_preempt_notifiers(prev
, next
);
2813 prepare_lock_switch(rq
, next
);
2814 prepare_arch_switch(next
);
2818 * finish_task_switch - clean up after a task-switch
2819 * @rq: runqueue associated with task-switch
2820 * @prev: the thread we just switched away from.
2822 * finish_task_switch must be called after the context switch, paired
2823 * with a prepare_task_switch call before the context switch.
2824 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2825 * and do any other architecture-specific cleanup actions.
2827 * Note that we may have delayed dropping an mm in context_switch(). If
2828 * so, we finish that here outside of the runqueue lock. (Doing it
2829 * with the lock held can cause deadlocks; see schedule() for
2832 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2833 __releases(rq
->lock
)
2835 struct mm_struct
*mm
= rq
->prev_mm
;
2841 * A task struct has one reference for the use as "current".
2842 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2843 * schedule one last time. The schedule call will never return, and
2844 * the scheduled task must drop that reference.
2845 * The test for TASK_DEAD must occur while the runqueue locks are
2846 * still held, otherwise prev could be scheduled on another cpu, die
2847 * there before we look at prev->state, and then the reference would
2849 * Manfred Spraul <manfred@colorfullife.com>
2851 prev_state
= prev
->state
;
2852 finish_arch_switch(prev
);
2853 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2854 local_irq_disable();
2855 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2856 perf_event_task_sched_in(current
);
2857 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2859 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2860 finish_lock_switch(rq
, prev
);
2862 fire_sched_in_preempt_notifiers(current
);
2865 if (unlikely(prev_state
== TASK_DEAD
)) {
2867 * Remove function-return probe instances associated with this
2868 * task and put them back on the free list.
2870 kprobe_flush_task(prev
);
2871 put_task_struct(prev
);
2877 /* assumes rq->lock is held */
2878 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2880 if (prev
->sched_class
->pre_schedule
)
2881 prev
->sched_class
->pre_schedule(rq
, prev
);
2884 /* rq->lock is NOT held, but preemption is disabled */
2885 static inline void post_schedule(struct rq
*rq
)
2887 if (rq
->post_schedule
) {
2888 unsigned long flags
;
2890 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2891 if (rq
->curr
->sched_class
->post_schedule
)
2892 rq
->curr
->sched_class
->post_schedule(rq
);
2893 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2895 rq
->post_schedule
= 0;
2901 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2905 static inline void post_schedule(struct rq
*rq
)
2912 * schedule_tail - first thing a freshly forked thread must call.
2913 * @prev: the thread we just switched away from.
2915 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2916 __releases(rq
->lock
)
2918 struct rq
*rq
= this_rq();
2920 finish_task_switch(rq
, prev
);
2923 * FIXME: do we need to worry about rq being invalidated by the
2928 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2929 /* In this case, finish_task_switch does not reenable preemption */
2932 if (current
->set_child_tid
)
2933 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2937 * context_switch - switch to the new MM and the new
2938 * thread's register state.
2941 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2942 struct task_struct
*next
)
2944 struct mm_struct
*mm
, *oldmm
;
2946 prepare_task_switch(rq
, prev
, next
);
2947 trace_sched_switch(prev
, next
);
2949 oldmm
= prev
->active_mm
;
2951 * For paravirt, this is coupled with an exit in switch_to to
2952 * combine the page table reload and the switch backend into
2955 arch_start_context_switch(prev
);
2958 next
->active_mm
= oldmm
;
2959 atomic_inc(&oldmm
->mm_count
);
2960 enter_lazy_tlb(oldmm
, next
);
2962 switch_mm(oldmm
, mm
, next
);
2965 prev
->active_mm
= NULL
;
2966 rq
->prev_mm
= oldmm
;
2969 * Since the runqueue lock will be released by the next
2970 * task (which is an invalid locking op but in the case
2971 * of the scheduler it's an obvious special-case), so we
2972 * do an early lockdep release here:
2974 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2975 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2978 /* Here we just switch the register state and the stack. */
2979 switch_to(prev
, next
, prev
);
2983 * this_rq must be evaluated again because prev may have moved
2984 * CPUs since it called schedule(), thus the 'rq' on its stack
2985 * frame will be invalid.
2987 finish_task_switch(this_rq(), prev
);
2991 * nr_running, nr_uninterruptible and nr_context_switches:
2993 * externally visible scheduler statistics: current number of runnable
2994 * threads, current number of uninterruptible-sleeping threads, total
2995 * number of context switches performed since bootup.
2997 unsigned long nr_running(void)
2999 unsigned long i
, sum
= 0;
3001 for_each_online_cpu(i
)
3002 sum
+= cpu_rq(i
)->nr_running
;
3007 unsigned long nr_uninterruptible(void)
3009 unsigned long i
, sum
= 0;
3011 for_each_possible_cpu(i
)
3012 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3015 * Since we read the counters lockless, it might be slightly
3016 * inaccurate. Do not allow it to go below zero though:
3018 if (unlikely((long)sum
< 0))
3024 unsigned long long nr_context_switches(void)
3027 unsigned long long sum
= 0;
3029 for_each_possible_cpu(i
)
3030 sum
+= cpu_rq(i
)->nr_switches
;
3035 unsigned long nr_iowait(void)
3037 unsigned long i
, sum
= 0;
3039 for_each_possible_cpu(i
)
3040 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3045 unsigned long nr_iowait_cpu(int cpu
)
3047 struct rq
*this = cpu_rq(cpu
);
3048 return atomic_read(&this->nr_iowait
);
3051 unsigned long this_cpu_load(void)
3053 struct rq
*this = this_rq();
3054 return this->cpu_load
[0];
3058 /* Variables and functions for calc_load */
3059 static atomic_long_t calc_load_tasks
;
3060 static unsigned long calc_load_update
;
3061 unsigned long avenrun
[3];
3062 EXPORT_SYMBOL(avenrun
);
3064 static long calc_load_fold_active(struct rq
*this_rq
)
3066 long nr_active
, delta
= 0;
3068 nr_active
= this_rq
->nr_running
;
3069 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3071 if (nr_active
!= this_rq
->calc_load_active
) {
3072 delta
= nr_active
- this_rq
->calc_load_active
;
3073 this_rq
->calc_load_active
= nr_active
;
3079 static unsigned long
3080 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3083 load
+= active
* (FIXED_1
- exp
);
3084 load
+= 1UL << (FSHIFT
- 1);
3085 return load
>> FSHIFT
;
3090 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3092 * When making the ILB scale, we should try to pull this in as well.
3094 static atomic_long_t calc_load_tasks_idle
;
3096 static void calc_load_account_idle(struct rq
*this_rq
)
3100 delta
= calc_load_fold_active(this_rq
);
3102 atomic_long_add(delta
, &calc_load_tasks_idle
);
3105 static long calc_load_fold_idle(void)
3110 * Its got a race, we don't care...
3112 if (atomic_long_read(&calc_load_tasks_idle
))
3113 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3119 * fixed_power_int - compute: x^n, in O(log n) time
3121 * @x: base of the power
3122 * @frac_bits: fractional bits of @x
3123 * @n: power to raise @x to.
3125 * By exploiting the relation between the definition of the natural power
3126 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3127 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3128 * (where: n_i \elem {0, 1}, the binary vector representing n),
3129 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3130 * of course trivially computable in O(log_2 n), the length of our binary
3133 static unsigned long
3134 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3136 unsigned long result
= 1UL << frac_bits
;
3141 result
+= 1UL << (frac_bits
- 1);
3142 result
>>= frac_bits
;
3148 x
+= 1UL << (frac_bits
- 1);
3156 * a1 = a0 * e + a * (1 - e)
3158 * a2 = a1 * e + a * (1 - e)
3159 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3160 * = a0 * e^2 + a * (1 - e) * (1 + e)
3162 * a3 = a2 * e + a * (1 - e)
3163 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3164 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3168 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3169 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3170 * = a0 * e^n + a * (1 - e^n)
3172 * [1] application of the geometric series:
3175 * S_n := \Sum x^i = -------------
3178 static unsigned long
3179 calc_load_n(unsigned long load
, unsigned long exp
,
3180 unsigned long active
, unsigned int n
)
3183 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3187 * NO_HZ can leave us missing all per-cpu ticks calling
3188 * calc_load_account_active(), but since an idle CPU folds its delta into
3189 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3190 * in the pending idle delta if our idle period crossed a load cycle boundary.
3192 * Once we've updated the global active value, we need to apply the exponential
3193 * weights adjusted to the number of cycles missed.
3195 static void calc_global_nohz(unsigned long ticks
)
3197 long delta
, active
, n
;
3199 if (time_before(jiffies
, calc_load_update
))
3203 * If we crossed a calc_load_update boundary, make sure to fold
3204 * any pending idle changes, the respective CPUs might have
3205 * missed the tick driven calc_load_account_active() update
3208 delta
= calc_load_fold_idle();
3210 atomic_long_add(delta
, &calc_load_tasks
);
3213 * If we were idle for multiple load cycles, apply them.
3215 if (ticks
>= LOAD_FREQ
) {
3216 n
= ticks
/ LOAD_FREQ
;
3218 active
= atomic_long_read(&calc_load_tasks
);
3219 active
= active
> 0 ? active
* FIXED_1
: 0;
3221 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3222 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3223 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3225 calc_load_update
+= n
* LOAD_FREQ
;
3229 * Its possible the remainder of the above division also crosses
3230 * a LOAD_FREQ period, the regular check in calc_global_load()
3231 * which comes after this will take care of that.
3233 * Consider us being 11 ticks before a cycle completion, and us
3234 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3235 * age us 4 cycles, and the test in calc_global_load() will
3236 * pick up the final one.
3240 static void calc_load_account_idle(struct rq
*this_rq
)
3244 static inline long calc_load_fold_idle(void)
3249 static void calc_global_nohz(unsigned long ticks
)
3255 * get_avenrun - get the load average array
3256 * @loads: pointer to dest load array
3257 * @offset: offset to add
3258 * @shift: shift count to shift the result left
3260 * These values are estimates at best, so no need for locking.
3262 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3264 loads
[0] = (avenrun
[0] + offset
) << shift
;
3265 loads
[1] = (avenrun
[1] + offset
) << shift
;
3266 loads
[2] = (avenrun
[2] + offset
) << shift
;
3270 * calc_load - update the avenrun load estimates 10 ticks after the
3271 * CPUs have updated calc_load_tasks.
3273 void calc_global_load(unsigned long ticks
)
3277 calc_global_nohz(ticks
);
3279 if (time_before(jiffies
, calc_load_update
+ 10))
3282 active
= atomic_long_read(&calc_load_tasks
);
3283 active
= active
> 0 ? active
* FIXED_1
: 0;
3285 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3286 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3287 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3289 calc_load_update
+= LOAD_FREQ
;
3293 * Called from update_cpu_load() to periodically update this CPU's
3296 static void calc_load_account_active(struct rq
*this_rq
)
3300 if (time_before(jiffies
, this_rq
->calc_load_update
))
3303 delta
= calc_load_fold_active(this_rq
);
3304 delta
+= calc_load_fold_idle();
3306 atomic_long_add(delta
, &calc_load_tasks
);
3308 this_rq
->calc_load_update
+= LOAD_FREQ
;
3312 * The exact cpuload at various idx values, calculated at every tick would be
3313 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3315 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3316 * on nth tick when cpu may be busy, then we have:
3317 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3318 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3320 * decay_load_missed() below does efficient calculation of
3321 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3322 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3324 * The calculation is approximated on a 128 point scale.
3325 * degrade_zero_ticks is the number of ticks after which load at any
3326 * particular idx is approximated to be zero.
3327 * degrade_factor is a precomputed table, a row for each load idx.
3328 * Each column corresponds to degradation factor for a power of two ticks,
3329 * based on 128 point scale.
3331 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3332 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3334 * With this power of 2 load factors, we can degrade the load n times
3335 * by looking at 1 bits in n and doing as many mult/shift instead of
3336 * n mult/shifts needed by the exact degradation.
3338 #define DEGRADE_SHIFT 7
3339 static const unsigned char
3340 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3341 static const unsigned char
3342 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3343 {0, 0, 0, 0, 0, 0, 0, 0},
3344 {64, 32, 8, 0, 0, 0, 0, 0},
3345 {96, 72, 40, 12, 1, 0, 0},
3346 {112, 98, 75, 43, 15, 1, 0},
3347 {120, 112, 98, 76, 45, 16, 2} };
3350 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3351 * would be when CPU is idle and so we just decay the old load without
3352 * adding any new load.
3354 static unsigned long
3355 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3359 if (!missed_updates
)
3362 if (missed_updates
>= degrade_zero_ticks
[idx
])
3366 return load
>> missed_updates
;
3368 while (missed_updates
) {
3369 if (missed_updates
% 2)
3370 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3372 missed_updates
>>= 1;
3379 * Update rq->cpu_load[] statistics. This function is usually called every
3380 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3381 * every tick. We fix it up based on jiffies.
3383 static void update_cpu_load(struct rq
*this_rq
)
3385 unsigned long this_load
= this_rq
->load
.weight
;
3386 unsigned long curr_jiffies
= jiffies
;
3387 unsigned long pending_updates
;
3390 this_rq
->nr_load_updates
++;
3392 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3393 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3396 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3397 this_rq
->last_load_update_tick
= curr_jiffies
;
3399 /* Update our load: */
3400 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3401 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3402 unsigned long old_load
, new_load
;
3404 /* scale is effectively 1 << i now, and >> i divides by scale */
3406 old_load
= this_rq
->cpu_load
[i
];
3407 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3408 new_load
= this_load
;
3410 * Round up the averaging division if load is increasing. This
3411 * prevents us from getting stuck on 9 if the load is 10, for
3414 if (new_load
> old_load
)
3415 new_load
+= scale
- 1;
3417 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3420 sched_avg_update(this_rq
);
3423 static void update_cpu_load_active(struct rq
*this_rq
)
3425 update_cpu_load(this_rq
);
3427 calc_load_account_active(this_rq
);
3433 * sched_exec - execve() is a valuable balancing opportunity, because at
3434 * this point the task has the smallest effective memory and cache footprint.
3436 void sched_exec(void)
3438 struct task_struct
*p
= current
;
3439 unsigned long flags
;
3443 rq
= task_rq_lock(p
, &flags
);
3444 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3445 if (dest_cpu
== smp_processor_id())
3449 * select_task_rq() can race against ->cpus_allowed
3451 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3452 likely(cpu_active(dest_cpu
)) && migrate_task(p
, rq
)) {
3453 struct migration_arg arg
= { p
, dest_cpu
};
3455 task_rq_unlock(rq
, &flags
);
3456 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3460 task_rq_unlock(rq
, &flags
);
3465 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3467 EXPORT_PER_CPU_SYMBOL(kstat
);
3470 * Return any ns on the sched_clock that have not yet been accounted in
3471 * @p in case that task is currently running.
3473 * Called with task_rq_lock() held on @rq.
3475 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3479 if (task_current(rq
, p
)) {
3480 update_rq_clock(rq
);
3481 ns
= rq
->clock_task
- p
->se
.exec_start
;
3489 unsigned long long task_delta_exec(struct task_struct
*p
)
3491 unsigned long flags
;
3495 rq
= task_rq_lock(p
, &flags
);
3496 ns
= do_task_delta_exec(p
, rq
);
3497 task_rq_unlock(rq
, &flags
);
3503 * Return accounted runtime for the task.
3504 * In case the task is currently running, return the runtime plus current's
3505 * pending runtime that have not been accounted yet.
3507 unsigned long long task_sched_runtime(struct task_struct
*p
)
3509 unsigned long flags
;
3513 rq
= task_rq_lock(p
, &flags
);
3514 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3515 task_rq_unlock(rq
, &flags
);
3521 * Return sum_exec_runtime for the thread group.
3522 * In case the task is currently running, return the sum plus current's
3523 * pending runtime that have not been accounted yet.
3525 * Note that the thread group might have other running tasks as well,
3526 * so the return value not includes other pending runtime that other
3527 * running tasks might have.
3529 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3531 struct task_cputime totals
;
3532 unsigned long flags
;
3536 rq
= task_rq_lock(p
, &flags
);
3537 thread_group_cputime(p
, &totals
);
3538 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3539 task_rq_unlock(rq
, &flags
);
3545 * Account user cpu time to a process.
3546 * @p: the process that the cpu time gets accounted to
3547 * @cputime: the cpu time spent in user space since the last update
3548 * @cputime_scaled: cputime scaled by cpu frequency
3550 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3551 cputime_t cputime_scaled
)
3553 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3556 /* Add user time to process. */
3557 p
->utime
= cputime_add(p
->utime
, cputime
);
3558 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3559 account_group_user_time(p
, cputime
);
3561 /* Add user time to cpustat. */
3562 tmp
= cputime_to_cputime64(cputime
);
3563 if (TASK_NICE(p
) > 0)
3564 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3566 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3568 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3569 /* Account for user time used */
3570 acct_update_integrals(p
);
3574 * Account guest cpu time to a process.
3575 * @p: the process that the cpu time gets accounted to
3576 * @cputime: the cpu time spent in virtual machine since the last update
3577 * @cputime_scaled: cputime scaled by cpu frequency
3579 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3580 cputime_t cputime_scaled
)
3583 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3585 tmp
= cputime_to_cputime64(cputime
);
3587 /* Add guest time to process. */
3588 p
->utime
= cputime_add(p
->utime
, cputime
);
3589 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3590 account_group_user_time(p
, cputime
);
3591 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3593 /* Add guest time to cpustat. */
3594 if (TASK_NICE(p
) > 0) {
3595 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3596 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3598 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3599 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3604 * Account system cpu time to a process and desired cpustat field
3605 * @p: the process that the cpu time gets accounted to
3606 * @cputime: the cpu time spent in kernel space since the last update
3607 * @cputime_scaled: cputime scaled by cpu frequency
3608 * @target_cputime64: pointer to cpustat field that has to be updated
3611 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
3612 cputime_t cputime_scaled
, cputime64_t
*target_cputime64
)
3614 cputime64_t tmp
= cputime_to_cputime64(cputime
);
3616 /* Add system time to process. */
3617 p
->stime
= cputime_add(p
->stime
, cputime
);
3618 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3619 account_group_system_time(p
, cputime
);
3621 /* Add system time to cpustat. */
3622 *target_cputime64
= cputime64_add(*target_cputime64
, tmp
);
3623 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3625 /* Account for system time used */
3626 acct_update_integrals(p
);
3630 * Account system cpu time to a process.
3631 * @p: the process that the cpu time gets accounted to
3632 * @hardirq_offset: the offset to subtract from hardirq_count()
3633 * @cputime: the cpu time spent in kernel space since the last update
3634 * @cputime_scaled: cputime scaled by cpu frequency
3636 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3637 cputime_t cputime
, cputime_t cputime_scaled
)
3639 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3640 cputime64_t
*target_cputime64
;
3642 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3643 account_guest_time(p
, cputime
, cputime_scaled
);
3647 if (hardirq_count() - hardirq_offset
)
3648 target_cputime64
= &cpustat
->irq
;
3649 else if (in_serving_softirq())
3650 target_cputime64
= &cpustat
->softirq
;
3652 target_cputime64
= &cpustat
->system
;
3654 __account_system_time(p
, cputime
, cputime_scaled
, target_cputime64
);
3657 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3659 * Account a tick to a process and cpustat
3660 * @p: the process that the cpu time gets accounted to
3661 * @user_tick: is the tick from userspace
3662 * @rq: the pointer to rq
3664 * Tick demultiplexing follows the order
3665 * - pending hardirq update
3666 * - pending softirq update
3670 * - check for guest_time
3671 * - else account as system_time
3673 * Check for hardirq is done both for system and user time as there is
3674 * no timer going off while we are on hardirq and hence we may never get an
3675 * opportunity to update it solely in system time.
3676 * p->stime and friends are only updated on system time and not on irq
3677 * softirq as those do not count in task exec_runtime any more.
3679 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3682 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3683 cputime64_t tmp
= cputime_to_cputime64(cputime_one_jiffy
);
3684 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3686 if (irqtime_account_hi_update()) {
3687 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3688 } else if (irqtime_account_si_update()) {
3689 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3690 } else if (this_cpu_ksoftirqd() == p
) {
3692 * ksoftirqd time do not get accounted in cpu_softirq_time.
3693 * So, we have to handle it separately here.
3694 * Also, p->stime needs to be updated for ksoftirqd.
3696 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3698 } else if (user_tick
) {
3699 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3700 } else if (p
== rq
->idle
) {
3701 account_idle_time(cputime_one_jiffy
);
3702 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
3703 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3705 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3710 static void irqtime_account_idle_ticks(int ticks
)
3713 struct rq
*rq
= this_rq();
3715 for (i
= 0; i
< ticks
; i
++)
3716 irqtime_account_process_tick(current
, 0, rq
);
3719 static void irqtime_account_idle_ticks(int ticks
) {}
3720 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3725 * Account for involuntary wait time.
3726 * @steal: the cpu time spent in involuntary wait
3728 void account_steal_time(cputime_t cputime
)
3730 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3731 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3733 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3737 * Account for idle time.
3738 * @cputime: the cpu time spent in idle wait
3740 void account_idle_time(cputime_t cputime
)
3742 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3743 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3744 struct rq
*rq
= this_rq();
3746 if (atomic_read(&rq
->nr_iowait
) > 0)
3747 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3749 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3752 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3755 * Account a single tick of cpu time.
3756 * @p: the process that the cpu time gets accounted to
3757 * @user_tick: indicates if the tick is a user or a system tick
3759 void account_process_tick(struct task_struct
*p
, int user_tick
)
3761 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3762 struct rq
*rq
= this_rq();
3764 if (sched_clock_irqtime
) {
3765 irqtime_account_process_tick(p
, user_tick
, rq
);
3770 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3771 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3772 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3775 account_idle_time(cputime_one_jiffy
);
3779 * Account multiple ticks of steal time.
3780 * @p: the process from which the cpu time has been stolen
3781 * @ticks: number of stolen ticks
3783 void account_steal_ticks(unsigned long ticks
)
3785 account_steal_time(jiffies_to_cputime(ticks
));
3789 * Account multiple ticks of idle time.
3790 * @ticks: number of stolen ticks
3792 void account_idle_ticks(unsigned long ticks
)
3795 if (sched_clock_irqtime
) {
3796 irqtime_account_idle_ticks(ticks
);
3800 account_idle_time(jiffies_to_cputime(ticks
));
3806 * Use precise platform statistics if available:
3808 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3809 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3815 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3817 struct task_cputime cputime
;
3819 thread_group_cputime(p
, &cputime
);
3821 *ut
= cputime
.utime
;
3822 *st
= cputime
.stime
;
3826 #ifndef nsecs_to_cputime
3827 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3830 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3832 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3835 * Use CFS's precise accounting:
3837 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3843 do_div(temp
, total
);
3844 utime
= (cputime_t
)temp
;
3849 * Compare with previous values, to keep monotonicity:
3851 p
->prev_utime
= max(p
->prev_utime
, utime
);
3852 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3854 *ut
= p
->prev_utime
;
3855 *st
= p
->prev_stime
;
3859 * Must be called with siglock held.
3861 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3863 struct signal_struct
*sig
= p
->signal
;
3864 struct task_cputime cputime
;
3865 cputime_t rtime
, utime
, total
;
3867 thread_group_cputime(p
, &cputime
);
3869 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3870 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3875 temp
*= cputime
.utime
;
3876 do_div(temp
, total
);
3877 utime
= (cputime_t
)temp
;
3881 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3882 sig
->prev_stime
= max(sig
->prev_stime
,
3883 cputime_sub(rtime
, sig
->prev_utime
));
3885 *ut
= sig
->prev_utime
;
3886 *st
= sig
->prev_stime
;
3891 * This function gets called by the timer code, with HZ frequency.
3892 * We call it with interrupts disabled.
3894 * It also gets called by the fork code, when changing the parent's
3897 void scheduler_tick(void)
3899 int cpu
= smp_processor_id();
3900 struct rq
*rq
= cpu_rq(cpu
);
3901 struct task_struct
*curr
= rq
->curr
;
3905 raw_spin_lock(&rq
->lock
);
3906 update_rq_clock(rq
);
3907 update_cpu_load_active(rq
);
3908 curr
->sched_class
->task_tick(rq
, curr
, 0);
3909 raw_spin_unlock(&rq
->lock
);
3911 perf_event_task_tick();
3914 rq
->idle_at_tick
= idle_cpu(cpu
);
3915 trigger_load_balance(rq
, cpu
);
3919 notrace
unsigned long get_parent_ip(unsigned long addr
)
3921 if (in_lock_functions(addr
)) {
3922 addr
= CALLER_ADDR2
;
3923 if (in_lock_functions(addr
))
3924 addr
= CALLER_ADDR3
;
3929 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3930 defined(CONFIG_PREEMPT_TRACER))
3932 void __kprobes
add_preempt_count(int val
)
3934 #ifdef CONFIG_DEBUG_PREEMPT
3938 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3941 preempt_count() += val
;
3942 #ifdef CONFIG_DEBUG_PREEMPT
3944 * Spinlock count overflowing soon?
3946 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3949 if (preempt_count() == val
)
3950 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3952 EXPORT_SYMBOL(add_preempt_count
);
3954 void __kprobes
sub_preempt_count(int val
)
3956 #ifdef CONFIG_DEBUG_PREEMPT
3960 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3963 * Is the spinlock portion underflowing?
3965 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3966 !(preempt_count() & PREEMPT_MASK
)))
3970 if (preempt_count() == val
)
3971 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3972 preempt_count() -= val
;
3974 EXPORT_SYMBOL(sub_preempt_count
);
3979 * Print scheduling while atomic bug:
3981 static noinline
void __schedule_bug(struct task_struct
*prev
)
3983 struct pt_regs
*regs
= get_irq_regs();
3985 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3986 prev
->comm
, prev
->pid
, preempt_count());
3988 debug_show_held_locks(prev
);
3990 if (irqs_disabled())
3991 print_irqtrace_events(prev
);
4000 * Various schedule()-time debugging checks and statistics:
4002 static inline void schedule_debug(struct task_struct
*prev
)
4005 * Test if we are atomic. Since do_exit() needs to call into
4006 * schedule() atomically, we ignore that path for now.
4007 * Otherwise, whine if we are scheduling when we should not be.
4009 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4010 __schedule_bug(prev
);
4012 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4014 schedstat_inc(this_rq(), sched_count
);
4015 #ifdef CONFIG_SCHEDSTATS
4016 if (unlikely(prev
->lock_depth
>= 0)) {
4017 schedstat_inc(this_rq(), rq_sched_info
.bkl_count
);
4018 schedstat_inc(prev
, sched_info
.bkl_count
);
4023 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4026 update_rq_clock(rq
);
4027 prev
->sched_class
->put_prev_task(rq
, prev
);
4031 * Pick up the highest-prio task:
4033 static inline struct task_struct
*
4034 pick_next_task(struct rq
*rq
)
4036 const struct sched_class
*class;
4037 struct task_struct
*p
;
4040 * Optimization: we know that if all tasks are in
4041 * the fair class we can call that function directly:
4043 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4044 p
= fair_sched_class
.pick_next_task(rq
);
4049 for_each_class(class) {
4050 p
= class->pick_next_task(rq
);
4055 BUG(); /* the idle class will always have a runnable task */
4059 * schedule() is the main scheduler function.
4061 asmlinkage
void __sched
schedule(void)
4063 struct task_struct
*prev
, *next
;
4064 unsigned long *switch_count
;
4070 cpu
= smp_processor_id();
4072 rcu_note_context_switch(cpu
);
4075 release_kernel_lock(prev
);
4076 need_resched_nonpreemptible
:
4078 schedule_debug(prev
);
4080 if (sched_feat(HRTICK
))
4083 raw_spin_lock_irq(&rq
->lock
);
4085 switch_count
= &prev
->nivcsw
;
4086 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4087 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
4088 prev
->state
= TASK_RUNNING
;
4091 * If a worker is going to sleep, notify and
4092 * ask workqueue whether it wants to wake up a
4093 * task to maintain concurrency. If so, wake
4096 if (prev
->flags
& PF_WQ_WORKER
) {
4097 struct task_struct
*to_wakeup
;
4099 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
4101 try_to_wake_up_local(to_wakeup
);
4103 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
4105 switch_count
= &prev
->nvcsw
;
4108 pre_schedule(rq
, prev
);
4110 if (unlikely(!rq
->nr_running
))
4111 idle_balance(cpu
, rq
);
4113 put_prev_task(rq
, prev
);
4114 next
= pick_next_task(rq
);
4115 clear_tsk_need_resched(prev
);
4116 rq
->skip_clock_update
= 0;
4118 if (likely(prev
!= next
)) {
4119 sched_info_switch(prev
, next
);
4120 perf_event_task_sched_out(prev
, next
);
4126 context_switch(rq
, prev
, next
); /* unlocks the rq */
4128 * The context switch have flipped the stack from under us
4129 * and restored the local variables which were saved when
4130 * this task called schedule() in the past. prev == current
4131 * is still correct, but it can be moved to another cpu/rq.
4133 cpu
= smp_processor_id();
4136 raw_spin_unlock_irq(&rq
->lock
);
4140 if (unlikely(reacquire_kernel_lock(prev
)))
4141 goto need_resched_nonpreemptible
;
4143 preempt_enable_no_resched();
4147 EXPORT_SYMBOL(schedule
);
4149 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4151 * Look out! "owner" is an entirely speculative pointer
4152 * access and not reliable.
4154 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
4159 if (!sched_feat(OWNER_SPIN
))
4162 #ifdef CONFIG_DEBUG_PAGEALLOC
4164 * Need to access the cpu field knowing that
4165 * DEBUG_PAGEALLOC could have unmapped it if
4166 * the mutex owner just released it and exited.
4168 if (probe_kernel_address(&owner
->cpu
, cpu
))
4175 * Even if the access succeeded (likely case),
4176 * the cpu field may no longer be valid.
4178 if (cpu
>= nr_cpumask_bits
)
4182 * We need to validate that we can do a
4183 * get_cpu() and that we have the percpu area.
4185 if (!cpu_online(cpu
))
4192 * Owner changed, break to re-assess state.
4194 if (lock
->owner
!= owner
) {
4196 * If the lock has switched to a different owner,
4197 * we likely have heavy contention. Return 0 to quit
4198 * optimistic spinning and not contend further:
4206 * Is that owner really running on that cpu?
4208 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
4211 arch_mutex_cpu_relax();
4218 #ifdef CONFIG_PREEMPT
4220 * this is the entry point to schedule() from in-kernel preemption
4221 * off of preempt_enable. Kernel preemptions off return from interrupt
4222 * occur there and call schedule directly.
4224 asmlinkage
void __sched notrace
preempt_schedule(void)
4226 struct thread_info
*ti
= current_thread_info();
4229 * If there is a non-zero preempt_count or interrupts are disabled,
4230 * we do not want to preempt the current task. Just return..
4232 if (likely(ti
->preempt_count
|| irqs_disabled()))
4236 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4238 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4241 * Check again in case we missed a preemption opportunity
4242 * between schedule and now.
4245 } while (need_resched());
4247 EXPORT_SYMBOL(preempt_schedule
);
4250 * this is the entry point to schedule() from kernel preemption
4251 * off of irq context.
4252 * Note, that this is called and return with irqs disabled. This will
4253 * protect us against recursive calling from irq.
4255 asmlinkage
void __sched
preempt_schedule_irq(void)
4257 struct thread_info
*ti
= current_thread_info();
4259 /* Catch callers which need to be fixed */
4260 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4263 add_preempt_count(PREEMPT_ACTIVE
);
4266 local_irq_disable();
4267 sub_preempt_count(PREEMPT_ACTIVE
);
4270 * Check again in case we missed a preemption opportunity
4271 * between schedule and now.
4274 } while (need_resched());
4277 #endif /* CONFIG_PREEMPT */
4279 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4282 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4284 EXPORT_SYMBOL(default_wake_function
);
4287 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4288 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4289 * number) then we wake all the non-exclusive tasks and one exclusive task.
4291 * There are circumstances in which we can try to wake a task which has already
4292 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4293 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4295 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4296 int nr_exclusive
, int wake_flags
, void *key
)
4298 wait_queue_t
*curr
, *next
;
4300 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4301 unsigned flags
= curr
->flags
;
4303 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4304 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4310 * __wake_up - wake up threads blocked on a waitqueue.
4312 * @mode: which threads
4313 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4314 * @key: is directly passed to the wakeup function
4316 * It may be assumed that this function implies a write memory barrier before
4317 * changing the task state if and only if any tasks are woken up.
4319 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4320 int nr_exclusive
, void *key
)
4322 unsigned long flags
;
4324 spin_lock_irqsave(&q
->lock
, flags
);
4325 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4326 spin_unlock_irqrestore(&q
->lock
, flags
);
4328 EXPORT_SYMBOL(__wake_up
);
4331 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4333 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4335 __wake_up_common(q
, mode
, 1, 0, NULL
);
4337 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4339 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4341 __wake_up_common(q
, mode
, 1, 0, key
);
4345 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4347 * @mode: which threads
4348 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4349 * @key: opaque value to be passed to wakeup targets
4351 * The sync wakeup differs that the waker knows that it will schedule
4352 * away soon, so while the target thread will be woken up, it will not
4353 * be migrated to another CPU - ie. the two threads are 'synchronized'
4354 * with each other. This can prevent needless bouncing between CPUs.
4356 * On UP it can prevent extra preemption.
4358 * It may be assumed that this function implies a write memory barrier before
4359 * changing the task state if and only if any tasks are woken up.
4361 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4362 int nr_exclusive
, void *key
)
4364 unsigned long flags
;
4365 int wake_flags
= WF_SYNC
;
4370 if (unlikely(!nr_exclusive
))
4373 spin_lock_irqsave(&q
->lock
, flags
);
4374 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4375 spin_unlock_irqrestore(&q
->lock
, flags
);
4377 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4380 * __wake_up_sync - see __wake_up_sync_key()
4382 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4384 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4386 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4389 * complete: - signals a single thread waiting on this completion
4390 * @x: holds the state of this particular completion
4392 * This will wake up a single thread waiting on this completion. Threads will be
4393 * awakened in the same order in which they were queued.
4395 * See also complete_all(), wait_for_completion() and related routines.
4397 * It may be assumed that this function implies a write memory barrier before
4398 * changing the task state if and only if any tasks are woken up.
4400 void complete(struct completion
*x
)
4402 unsigned long flags
;
4404 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4406 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4407 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4409 EXPORT_SYMBOL(complete
);
4412 * complete_all: - signals all threads waiting on this completion
4413 * @x: holds the state of this particular completion
4415 * This will wake up all threads waiting on this particular completion event.
4417 * It may be assumed that this function implies a write memory barrier before
4418 * changing the task state if and only if any tasks are woken up.
4420 void complete_all(struct completion
*x
)
4422 unsigned long flags
;
4424 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4425 x
->done
+= UINT_MAX
/2;
4426 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4427 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4429 EXPORT_SYMBOL(complete_all
);
4431 static inline long __sched
4432 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4435 DECLARE_WAITQUEUE(wait
, current
);
4437 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4439 if (signal_pending_state(state
, current
)) {
4440 timeout
= -ERESTARTSYS
;
4443 __set_current_state(state
);
4444 spin_unlock_irq(&x
->wait
.lock
);
4445 timeout
= schedule_timeout(timeout
);
4446 spin_lock_irq(&x
->wait
.lock
);
4447 } while (!x
->done
&& timeout
);
4448 __remove_wait_queue(&x
->wait
, &wait
);
4453 return timeout
?: 1;
4457 wait_for_common(struct completion
*x
, long timeout
, int state
)
4461 spin_lock_irq(&x
->wait
.lock
);
4462 timeout
= do_wait_for_common(x
, timeout
, state
);
4463 spin_unlock_irq(&x
->wait
.lock
);
4468 * wait_for_completion: - waits for completion of a task
4469 * @x: holds the state of this particular completion
4471 * This waits to be signaled for completion of a specific task. It is NOT
4472 * interruptible and there is no timeout.
4474 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4475 * and interrupt capability. Also see complete().
4477 void __sched
wait_for_completion(struct completion
*x
)
4479 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4481 EXPORT_SYMBOL(wait_for_completion
);
4484 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4485 * @x: holds the state of this particular completion
4486 * @timeout: timeout value in jiffies
4488 * This waits for either a completion of a specific task to be signaled or for a
4489 * specified timeout to expire. The timeout is in jiffies. It is not
4492 unsigned long __sched
4493 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4495 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4497 EXPORT_SYMBOL(wait_for_completion_timeout
);
4500 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4501 * @x: holds the state of this particular completion
4503 * This waits for completion of a specific task to be signaled. It is
4506 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4508 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4509 if (t
== -ERESTARTSYS
)
4513 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4516 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4517 * @x: holds the state of this particular completion
4518 * @timeout: timeout value in jiffies
4520 * This waits for either a completion of a specific task to be signaled or for a
4521 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4524 wait_for_completion_interruptible_timeout(struct completion
*x
,
4525 unsigned long timeout
)
4527 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4529 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4532 * wait_for_completion_killable: - waits for completion of a task (killable)
4533 * @x: holds the state of this particular completion
4535 * This waits to be signaled for completion of a specific task. It can be
4536 * interrupted by a kill signal.
4538 int __sched
wait_for_completion_killable(struct completion
*x
)
4540 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4541 if (t
== -ERESTARTSYS
)
4545 EXPORT_SYMBOL(wait_for_completion_killable
);
4548 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4549 * @x: holds the state of this particular completion
4550 * @timeout: timeout value in jiffies
4552 * This waits for either a completion of a specific task to be
4553 * signaled or for a specified timeout to expire. It can be
4554 * interrupted by a kill signal. The timeout is in jiffies.
4557 wait_for_completion_killable_timeout(struct completion
*x
,
4558 unsigned long timeout
)
4560 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4562 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4565 * try_wait_for_completion - try to decrement a completion without blocking
4566 * @x: completion structure
4568 * Returns: 0 if a decrement cannot be done without blocking
4569 * 1 if a decrement succeeded.
4571 * If a completion is being used as a counting completion,
4572 * attempt to decrement the counter without blocking. This
4573 * enables us to avoid waiting if the resource the completion
4574 * is protecting is not available.
4576 bool try_wait_for_completion(struct completion
*x
)
4578 unsigned long flags
;
4581 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4586 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4589 EXPORT_SYMBOL(try_wait_for_completion
);
4592 * completion_done - Test to see if a completion has any waiters
4593 * @x: completion structure
4595 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4596 * 1 if there are no waiters.
4599 bool completion_done(struct completion
*x
)
4601 unsigned long flags
;
4604 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4607 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4610 EXPORT_SYMBOL(completion_done
);
4613 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4615 unsigned long flags
;
4618 init_waitqueue_entry(&wait
, current
);
4620 __set_current_state(state
);
4622 spin_lock_irqsave(&q
->lock
, flags
);
4623 __add_wait_queue(q
, &wait
);
4624 spin_unlock(&q
->lock
);
4625 timeout
= schedule_timeout(timeout
);
4626 spin_lock_irq(&q
->lock
);
4627 __remove_wait_queue(q
, &wait
);
4628 spin_unlock_irqrestore(&q
->lock
, flags
);
4633 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4635 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4637 EXPORT_SYMBOL(interruptible_sleep_on
);
4640 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4642 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4644 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4646 void __sched
sleep_on(wait_queue_head_t
*q
)
4648 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4650 EXPORT_SYMBOL(sleep_on
);
4652 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4654 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4656 EXPORT_SYMBOL(sleep_on_timeout
);
4658 #ifdef CONFIG_RT_MUTEXES
4661 * rt_mutex_setprio - set the current priority of a task
4663 * @prio: prio value (kernel-internal form)
4665 * This function changes the 'effective' priority of a task. It does
4666 * not touch ->normal_prio like __setscheduler().
4668 * Used by the rt_mutex code to implement priority inheritance logic.
4670 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4672 unsigned long flags
;
4673 int oldprio
, on_rq
, running
;
4675 const struct sched_class
*prev_class
;
4677 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4679 rq
= task_rq_lock(p
, &flags
);
4681 trace_sched_pi_setprio(p
, prio
);
4683 prev_class
= p
->sched_class
;
4684 on_rq
= p
->se
.on_rq
;
4685 running
= task_current(rq
, p
);
4687 dequeue_task(rq
, p
, 0);
4689 p
->sched_class
->put_prev_task(rq
, p
);
4692 p
->sched_class
= &rt_sched_class
;
4694 p
->sched_class
= &fair_sched_class
;
4699 p
->sched_class
->set_curr_task(rq
);
4701 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4703 check_class_changed(rq
, p
, prev_class
, oldprio
);
4704 task_rq_unlock(rq
, &flags
);
4709 void set_user_nice(struct task_struct
*p
, long nice
)
4711 int old_prio
, delta
, on_rq
;
4712 unsigned long flags
;
4715 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4718 * We have to be careful, if called from sys_setpriority(),
4719 * the task might be in the middle of scheduling on another CPU.
4721 rq
= task_rq_lock(p
, &flags
);
4723 * The RT priorities are set via sched_setscheduler(), but we still
4724 * allow the 'normal' nice value to be set - but as expected
4725 * it wont have any effect on scheduling until the task is
4726 * SCHED_FIFO/SCHED_RR:
4728 if (task_has_rt_policy(p
)) {
4729 p
->static_prio
= NICE_TO_PRIO(nice
);
4732 on_rq
= p
->se
.on_rq
;
4734 dequeue_task(rq
, p
, 0);
4736 p
->static_prio
= NICE_TO_PRIO(nice
);
4739 p
->prio
= effective_prio(p
);
4740 delta
= p
->prio
- old_prio
;
4743 enqueue_task(rq
, p
, 0);
4745 * If the task increased its priority or is running and
4746 * lowered its priority, then reschedule its CPU:
4748 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4749 resched_task(rq
->curr
);
4752 task_rq_unlock(rq
, &flags
);
4754 EXPORT_SYMBOL(set_user_nice
);
4757 * can_nice - check if a task can reduce its nice value
4761 int can_nice(const struct task_struct
*p
, const int nice
)
4763 /* convert nice value [19,-20] to rlimit style value [1,40] */
4764 int nice_rlim
= 20 - nice
;
4766 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4767 capable(CAP_SYS_NICE
));
4770 #ifdef __ARCH_WANT_SYS_NICE
4773 * sys_nice - change the priority of the current process.
4774 * @increment: priority increment
4776 * sys_setpriority is a more generic, but much slower function that
4777 * does similar things.
4779 SYSCALL_DEFINE1(nice
, int, increment
)
4784 * Setpriority might change our priority at the same moment.
4785 * We don't have to worry. Conceptually one call occurs first
4786 * and we have a single winner.
4788 if (increment
< -40)
4793 nice
= TASK_NICE(current
) + increment
;
4799 if (increment
< 0 && !can_nice(current
, nice
))
4802 retval
= security_task_setnice(current
, nice
);
4806 set_user_nice(current
, nice
);
4813 * task_prio - return the priority value of a given task.
4814 * @p: the task in question.
4816 * This is the priority value as seen by users in /proc.
4817 * RT tasks are offset by -200. Normal tasks are centered
4818 * around 0, value goes from -16 to +15.
4820 int task_prio(const struct task_struct
*p
)
4822 return p
->prio
- MAX_RT_PRIO
;
4826 * task_nice - return the nice value of a given task.
4827 * @p: the task in question.
4829 int task_nice(const struct task_struct
*p
)
4831 return TASK_NICE(p
);
4833 EXPORT_SYMBOL(task_nice
);
4836 * idle_cpu - is a given cpu idle currently?
4837 * @cpu: the processor in question.
4839 int idle_cpu(int cpu
)
4841 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4845 * idle_task - return the idle task for a given cpu.
4846 * @cpu: the processor in question.
4848 struct task_struct
*idle_task(int cpu
)
4850 return cpu_rq(cpu
)->idle
;
4854 * find_process_by_pid - find a process with a matching PID value.
4855 * @pid: the pid in question.
4857 static struct task_struct
*find_process_by_pid(pid_t pid
)
4859 return pid
? find_task_by_vpid(pid
) : current
;
4862 /* Actually do priority change: must hold rq lock. */
4864 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4866 BUG_ON(p
->se
.on_rq
);
4869 p
->rt_priority
= prio
;
4870 p
->normal_prio
= normal_prio(p
);
4871 /* we are holding p->pi_lock already */
4872 p
->prio
= rt_mutex_getprio(p
);
4873 if (rt_prio(p
->prio
))
4874 p
->sched_class
= &rt_sched_class
;
4876 p
->sched_class
= &fair_sched_class
;
4881 * check the target process has a UID that matches the current process's
4883 static bool check_same_owner(struct task_struct
*p
)
4885 const struct cred
*cred
= current_cred(), *pcred
;
4889 pcred
= __task_cred(p
);
4890 match
= (cred
->euid
== pcred
->euid
||
4891 cred
->euid
== pcred
->uid
);
4896 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4897 const struct sched_param
*param
, bool user
)
4899 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4900 unsigned long flags
;
4901 const struct sched_class
*prev_class
;
4905 /* may grab non-irq protected spin_locks */
4906 BUG_ON(in_interrupt());
4908 /* double check policy once rq lock held */
4910 reset_on_fork
= p
->sched_reset_on_fork
;
4911 policy
= oldpolicy
= p
->policy
;
4913 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4914 policy
&= ~SCHED_RESET_ON_FORK
;
4916 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4917 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4918 policy
!= SCHED_IDLE
)
4923 * Valid priorities for SCHED_FIFO and SCHED_RR are
4924 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4925 * SCHED_BATCH and SCHED_IDLE is 0.
4927 if (param
->sched_priority
< 0 ||
4928 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4929 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4931 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4935 * Allow unprivileged RT tasks to decrease priority:
4937 if (user
&& !capable(CAP_SYS_NICE
)) {
4938 if (rt_policy(policy
)) {
4939 unsigned long rlim_rtprio
=
4940 task_rlimit(p
, RLIMIT_RTPRIO
);
4942 /* can't set/change the rt policy */
4943 if (policy
!= p
->policy
&& !rlim_rtprio
)
4946 /* can't increase priority */
4947 if (param
->sched_priority
> p
->rt_priority
&&
4948 param
->sched_priority
> rlim_rtprio
)
4952 * Like positive nice levels, dont allow tasks to
4953 * move out of SCHED_IDLE either:
4955 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4958 /* can't change other user's priorities */
4959 if (!check_same_owner(p
))
4962 /* Normal users shall not reset the sched_reset_on_fork flag */
4963 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4968 retval
= security_task_setscheduler(p
);
4974 * make sure no PI-waiters arrive (or leave) while we are
4975 * changing the priority of the task:
4977 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4979 * To be able to change p->policy safely, the apropriate
4980 * runqueue lock must be held.
4982 rq
= __task_rq_lock(p
);
4985 * Changing the policy of the stop threads its a very bad idea
4987 if (p
== rq
->stop
) {
4988 __task_rq_unlock(rq
);
4989 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4993 #ifdef CONFIG_RT_GROUP_SCHED
4996 * Do not allow realtime tasks into groups that have no runtime
4999 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5000 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5001 !task_group_is_autogroup(task_group(p
))) {
5002 __task_rq_unlock(rq
);
5003 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5009 /* recheck policy now with rq lock held */
5010 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5011 policy
= oldpolicy
= -1;
5012 __task_rq_unlock(rq
);
5013 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5016 on_rq
= p
->se
.on_rq
;
5017 running
= task_current(rq
, p
);
5019 deactivate_task(rq
, p
, 0);
5021 p
->sched_class
->put_prev_task(rq
, p
);
5023 p
->sched_reset_on_fork
= reset_on_fork
;
5026 prev_class
= p
->sched_class
;
5027 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5030 p
->sched_class
->set_curr_task(rq
);
5032 activate_task(rq
, p
, 0);
5034 check_class_changed(rq
, p
, prev_class
, oldprio
);
5035 __task_rq_unlock(rq
);
5036 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5038 rt_mutex_adjust_pi(p
);
5044 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5045 * @p: the task in question.
5046 * @policy: new policy.
5047 * @param: structure containing the new RT priority.
5049 * NOTE that the task may be already dead.
5051 int sched_setscheduler(struct task_struct
*p
, int policy
,
5052 const struct sched_param
*param
)
5054 return __sched_setscheduler(p
, policy
, param
, true);
5056 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5059 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5060 * @p: the task in question.
5061 * @policy: new policy.
5062 * @param: structure containing the new RT priority.
5064 * Just like sched_setscheduler, only don't bother checking if the
5065 * current context has permission. For example, this is needed in
5066 * stop_machine(): we create temporary high priority worker threads,
5067 * but our caller might not have that capability.
5069 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5070 const struct sched_param
*param
)
5072 return __sched_setscheduler(p
, policy
, param
, false);
5076 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5078 struct sched_param lparam
;
5079 struct task_struct
*p
;
5082 if (!param
|| pid
< 0)
5084 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5089 p
= find_process_by_pid(pid
);
5091 retval
= sched_setscheduler(p
, policy
, &lparam
);
5098 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5099 * @pid: the pid in question.
5100 * @policy: new policy.
5101 * @param: structure containing the new RT priority.
5103 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5104 struct sched_param __user
*, param
)
5106 /* negative values for policy are not valid */
5110 return do_sched_setscheduler(pid
, policy
, param
);
5114 * sys_sched_setparam - set/change the RT priority of a thread
5115 * @pid: the pid in question.
5116 * @param: structure containing the new RT priority.
5118 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5120 return do_sched_setscheduler(pid
, -1, param
);
5124 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5125 * @pid: the pid in question.
5127 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5129 struct task_struct
*p
;
5137 p
= find_process_by_pid(pid
);
5139 retval
= security_task_getscheduler(p
);
5142 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5149 * sys_sched_getparam - get the RT priority of a thread
5150 * @pid: the pid in question.
5151 * @param: structure containing the RT priority.
5153 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5155 struct sched_param lp
;
5156 struct task_struct
*p
;
5159 if (!param
|| pid
< 0)
5163 p
= find_process_by_pid(pid
);
5168 retval
= security_task_getscheduler(p
);
5172 lp
.sched_priority
= p
->rt_priority
;
5176 * This one might sleep, we cannot do it with a spinlock held ...
5178 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5187 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5189 cpumask_var_t cpus_allowed
, new_mask
;
5190 struct task_struct
*p
;
5196 p
= find_process_by_pid(pid
);
5203 /* Prevent p going away */
5207 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5211 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5213 goto out_free_cpus_allowed
;
5216 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5219 retval
= security_task_setscheduler(p
);
5223 cpuset_cpus_allowed(p
, cpus_allowed
);
5224 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5226 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5229 cpuset_cpus_allowed(p
, cpus_allowed
);
5230 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5232 * We must have raced with a concurrent cpuset
5233 * update. Just reset the cpus_allowed to the
5234 * cpuset's cpus_allowed
5236 cpumask_copy(new_mask
, cpus_allowed
);
5241 free_cpumask_var(new_mask
);
5242 out_free_cpus_allowed
:
5243 free_cpumask_var(cpus_allowed
);
5250 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5251 struct cpumask
*new_mask
)
5253 if (len
< cpumask_size())
5254 cpumask_clear(new_mask
);
5255 else if (len
> cpumask_size())
5256 len
= cpumask_size();
5258 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5262 * sys_sched_setaffinity - set the cpu affinity of a process
5263 * @pid: pid of the process
5264 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5265 * @user_mask_ptr: user-space pointer to the new cpu mask
5267 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5268 unsigned long __user
*, user_mask_ptr
)
5270 cpumask_var_t new_mask
;
5273 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5276 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5278 retval
= sched_setaffinity(pid
, new_mask
);
5279 free_cpumask_var(new_mask
);
5283 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5285 struct task_struct
*p
;
5286 unsigned long flags
;
5294 p
= find_process_by_pid(pid
);
5298 retval
= security_task_getscheduler(p
);
5302 rq
= task_rq_lock(p
, &flags
);
5303 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5304 task_rq_unlock(rq
, &flags
);
5314 * sys_sched_getaffinity - get the cpu affinity of a process
5315 * @pid: pid of the process
5316 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5317 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5319 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5320 unsigned long __user
*, user_mask_ptr
)
5325 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5327 if (len
& (sizeof(unsigned long)-1))
5330 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5333 ret
= sched_getaffinity(pid
, mask
);
5335 size_t retlen
= min_t(size_t, len
, cpumask_size());
5337 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5342 free_cpumask_var(mask
);
5348 * sys_sched_yield - yield the current processor to other threads.
5350 * This function yields the current CPU to other tasks. If there are no
5351 * other threads running on this CPU then this function will return.
5353 SYSCALL_DEFINE0(sched_yield
)
5355 struct rq
*rq
= this_rq_lock();
5357 schedstat_inc(rq
, yld_count
);
5358 current
->sched_class
->yield_task(rq
);
5361 * Since we are going to call schedule() anyway, there's
5362 * no need to preempt or enable interrupts:
5364 __release(rq
->lock
);
5365 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5366 do_raw_spin_unlock(&rq
->lock
);
5367 preempt_enable_no_resched();
5374 static inline int should_resched(void)
5376 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5379 static void __cond_resched(void)
5381 add_preempt_count(PREEMPT_ACTIVE
);
5383 sub_preempt_count(PREEMPT_ACTIVE
);
5386 int __sched
_cond_resched(void)
5388 if (should_resched()) {
5394 EXPORT_SYMBOL(_cond_resched
);
5397 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5398 * call schedule, and on return reacquire the lock.
5400 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5401 * operations here to prevent schedule() from being called twice (once via
5402 * spin_unlock(), once by hand).
5404 int __cond_resched_lock(spinlock_t
*lock
)
5406 int resched
= should_resched();
5409 lockdep_assert_held(lock
);
5411 if (spin_needbreak(lock
) || resched
) {
5422 EXPORT_SYMBOL(__cond_resched_lock
);
5424 int __sched
__cond_resched_softirq(void)
5426 BUG_ON(!in_softirq());
5428 if (should_resched()) {
5436 EXPORT_SYMBOL(__cond_resched_softirq
);
5439 * yield - yield the current processor to other threads.
5441 * This is a shortcut for kernel-space yielding - it marks the
5442 * thread runnable and calls sys_sched_yield().
5444 void __sched
yield(void)
5446 set_current_state(TASK_RUNNING
);
5449 EXPORT_SYMBOL(yield
);
5452 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5453 * that process accounting knows that this is a task in IO wait state.
5455 void __sched
io_schedule(void)
5457 struct rq
*rq
= raw_rq();
5459 delayacct_blkio_start();
5460 atomic_inc(&rq
->nr_iowait
);
5461 current
->in_iowait
= 1;
5463 current
->in_iowait
= 0;
5464 atomic_dec(&rq
->nr_iowait
);
5465 delayacct_blkio_end();
5467 EXPORT_SYMBOL(io_schedule
);
5469 long __sched
io_schedule_timeout(long timeout
)
5471 struct rq
*rq
= raw_rq();
5474 delayacct_blkio_start();
5475 atomic_inc(&rq
->nr_iowait
);
5476 current
->in_iowait
= 1;
5477 ret
= schedule_timeout(timeout
);
5478 current
->in_iowait
= 0;
5479 atomic_dec(&rq
->nr_iowait
);
5480 delayacct_blkio_end();
5485 * sys_sched_get_priority_max - return maximum RT priority.
5486 * @policy: scheduling class.
5488 * this syscall returns the maximum rt_priority that can be used
5489 * by a given scheduling class.
5491 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5498 ret
= MAX_USER_RT_PRIO
-1;
5510 * sys_sched_get_priority_min - return minimum RT priority.
5511 * @policy: scheduling class.
5513 * this syscall returns the minimum rt_priority that can be used
5514 * by a given scheduling class.
5516 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5534 * sys_sched_rr_get_interval - return the default timeslice of a process.
5535 * @pid: pid of the process.
5536 * @interval: userspace pointer to the timeslice value.
5538 * this syscall writes the default timeslice value of a given process
5539 * into the user-space timespec buffer. A value of '0' means infinity.
5541 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5542 struct timespec __user
*, interval
)
5544 struct task_struct
*p
;
5545 unsigned int time_slice
;
5546 unsigned long flags
;
5556 p
= find_process_by_pid(pid
);
5560 retval
= security_task_getscheduler(p
);
5564 rq
= task_rq_lock(p
, &flags
);
5565 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5566 task_rq_unlock(rq
, &flags
);
5569 jiffies_to_timespec(time_slice
, &t
);
5570 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5578 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5580 void sched_show_task(struct task_struct
*p
)
5582 unsigned long free
= 0;
5585 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5586 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5587 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5588 #if BITS_PER_LONG == 32
5589 if (state
== TASK_RUNNING
)
5590 printk(KERN_CONT
" running ");
5592 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5594 if (state
== TASK_RUNNING
)
5595 printk(KERN_CONT
" running task ");
5597 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5599 #ifdef CONFIG_DEBUG_STACK_USAGE
5600 free
= stack_not_used(p
);
5602 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5603 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5604 (unsigned long)task_thread_info(p
)->flags
);
5606 show_stack(p
, NULL
);
5609 void show_state_filter(unsigned long state_filter
)
5611 struct task_struct
*g
, *p
;
5613 #if BITS_PER_LONG == 32
5615 " task PC stack pid father\n");
5618 " task PC stack pid father\n");
5620 read_lock(&tasklist_lock
);
5621 do_each_thread(g
, p
) {
5623 * reset the NMI-timeout, listing all files on a slow
5624 * console might take alot of time:
5626 touch_nmi_watchdog();
5627 if (!state_filter
|| (p
->state
& state_filter
))
5629 } while_each_thread(g
, p
);
5631 touch_all_softlockup_watchdogs();
5633 #ifdef CONFIG_SCHED_DEBUG
5634 sysrq_sched_debug_show();
5636 read_unlock(&tasklist_lock
);
5638 * Only show locks if all tasks are dumped:
5641 debug_show_all_locks();
5644 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5646 idle
->sched_class
= &idle_sched_class
;
5650 * init_idle - set up an idle thread for a given CPU
5651 * @idle: task in question
5652 * @cpu: cpu the idle task belongs to
5654 * NOTE: this function does not set the idle thread's NEED_RESCHED
5655 * flag, to make booting more robust.
5657 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5659 struct rq
*rq
= cpu_rq(cpu
);
5660 unsigned long flags
;
5662 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5665 idle
->state
= TASK_RUNNING
;
5666 idle
->se
.exec_start
= sched_clock();
5668 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5670 * We're having a chicken and egg problem, even though we are
5671 * holding rq->lock, the cpu isn't yet set to this cpu so the
5672 * lockdep check in task_group() will fail.
5674 * Similar case to sched_fork(). / Alternatively we could
5675 * use task_rq_lock() here and obtain the other rq->lock.
5680 __set_task_cpu(idle
, cpu
);
5683 rq
->curr
= rq
->idle
= idle
;
5684 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5687 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5689 /* Set the preempt count _outside_ the spinlocks! */
5690 #if defined(CONFIG_PREEMPT)
5691 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5693 task_thread_info(idle
)->preempt_count
= 0;
5696 * The idle tasks have their own, simple scheduling class:
5698 idle
->sched_class
= &idle_sched_class
;
5699 ftrace_graph_init_task(idle
);
5703 * In a system that switches off the HZ timer nohz_cpu_mask
5704 * indicates which cpus entered this state. This is used
5705 * in the rcu update to wait only for active cpus. For system
5706 * which do not switch off the HZ timer nohz_cpu_mask should
5707 * always be CPU_BITS_NONE.
5709 cpumask_var_t nohz_cpu_mask
;
5712 * Increase the granularity value when there are more CPUs,
5713 * because with more CPUs the 'effective latency' as visible
5714 * to users decreases. But the relationship is not linear,
5715 * so pick a second-best guess by going with the log2 of the
5718 * This idea comes from the SD scheduler of Con Kolivas:
5720 static int get_update_sysctl_factor(void)
5722 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5723 unsigned int factor
;
5725 switch (sysctl_sched_tunable_scaling
) {
5726 case SCHED_TUNABLESCALING_NONE
:
5729 case SCHED_TUNABLESCALING_LINEAR
:
5732 case SCHED_TUNABLESCALING_LOG
:
5734 factor
= 1 + ilog2(cpus
);
5741 static void update_sysctl(void)
5743 unsigned int factor
= get_update_sysctl_factor();
5745 #define SET_SYSCTL(name) \
5746 (sysctl_##name = (factor) * normalized_sysctl_##name)
5747 SET_SYSCTL(sched_min_granularity
);
5748 SET_SYSCTL(sched_latency
);
5749 SET_SYSCTL(sched_wakeup_granularity
);
5753 static inline void sched_init_granularity(void)
5760 * This is how migration works:
5762 * 1) we invoke migration_cpu_stop() on the target CPU using
5764 * 2) stopper starts to run (implicitly forcing the migrated thread
5766 * 3) it checks whether the migrated task is still in the wrong runqueue.
5767 * 4) if it's in the wrong runqueue then the migration thread removes
5768 * it and puts it into the right queue.
5769 * 5) stopper completes and stop_one_cpu() returns and the migration
5774 * Change a given task's CPU affinity. Migrate the thread to a
5775 * proper CPU and schedule it away if the CPU it's executing on
5776 * is removed from the allowed bitmask.
5778 * NOTE: the caller must have a valid reference to the task, the
5779 * task must not exit() & deallocate itself prematurely. The
5780 * call is not atomic; no spinlocks may be held.
5782 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5784 unsigned long flags
;
5786 unsigned int dest_cpu
;
5790 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5791 * drop the rq->lock and still rely on ->cpus_allowed.
5794 while (task_is_waking(p
))
5796 rq
= task_rq_lock(p
, &flags
);
5797 if (task_is_waking(p
)) {
5798 task_rq_unlock(rq
, &flags
);
5802 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5807 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5808 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5813 if (p
->sched_class
->set_cpus_allowed
)
5814 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5816 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5817 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5820 /* Can the task run on the task's current CPU? If so, we're done */
5821 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5824 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5825 if (migrate_task(p
, rq
)) {
5826 struct migration_arg arg
= { p
, dest_cpu
};
5827 /* Need help from migration thread: drop lock and wait. */
5828 task_rq_unlock(rq
, &flags
);
5829 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5830 tlb_migrate_finish(p
->mm
);
5834 task_rq_unlock(rq
, &flags
);
5838 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5841 * Move (not current) task off this cpu, onto dest cpu. We're doing
5842 * this because either it can't run here any more (set_cpus_allowed()
5843 * away from this CPU, or CPU going down), or because we're
5844 * attempting to rebalance this task on exec (sched_exec).
5846 * So we race with normal scheduler movements, but that's OK, as long
5847 * as the task is no longer on this CPU.
5849 * Returns non-zero if task was successfully migrated.
5851 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5853 struct rq
*rq_dest
, *rq_src
;
5856 if (unlikely(!cpu_active(dest_cpu
)))
5859 rq_src
= cpu_rq(src_cpu
);
5860 rq_dest
= cpu_rq(dest_cpu
);
5862 double_rq_lock(rq_src
, rq_dest
);
5863 /* Already moved. */
5864 if (task_cpu(p
) != src_cpu
)
5866 /* Affinity changed (again). */
5867 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5871 * If we're not on a rq, the next wake-up will ensure we're
5875 deactivate_task(rq_src
, p
, 0);
5876 set_task_cpu(p
, dest_cpu
);
5877 activate_task(rq_dest
, p
, 0);
5878 check_preempt_curr(rq_dest
, p
, 0);
5883 double_rq_unlock(rq_src
, rq_dest
);
5888 * migration_cpu_stop - this will be executed by a highprio stopper thread
5889 * and performs thread migration by bumping thread off CPU then
5890 * 'pushing' onto another runqueue.
5892 static int migration_cpu_stop(void *data
)
5894 struct migration_arg
*arg
= data
;
5897 * The original target cpu might have gone down and we might
5898 * be on another cpu but it doesn't matter.
5900 local_irq_disable();
5901 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5906 #ifdef CONFIG_HOTPLUG_CPU
5909 * Ensures that the idle task is using init_mm right before its cpu goes
5912 void idle_task_exit(void)
5914 struct mm_struct
*mm
= current
->active_mm
;
5916 BUG_ON(cpu_online(smp_processor_id()));
5919 switch_mm(mm
, &init_mm
, current
);
5924 * While a dead CPU has no uninterruptible tasks queued at this point,
5925 * it might still have a nonzero ->nr_uninterruptible counter, because
5926 * for performance reasons the counter is not stricly tracking tasks to
5927 * their home CPUs. So we just add the counter to another CPU's counter,
5928 * to keep the global sum constant after CPU-down:
5930 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5932 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5934 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5935 rq_src
->nr_uninterruptible
= 0;
5939 * remove the tasks which were accounted by rq from calc_load_tasks.
5941 static void calc_global_load_remove(struct rq
*rq
)
5943 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5944 rq
->calc_load_active
= 0;
5948 * Migrate all tasks from the rq, sleeping tasks will be migrated by
5949 * try_to_wake_up()->select_task_rq().
5951 * Called with rq->lock held even though we'er in stop_machine() and
5952 * there's no concurrency possible, we hold the required locks anyway
5953 * because of lock validation efforts.
5955 static void migrate_tasks(unsigned int dead_cpu
)
5957 struct rq
*rq
= cpu_rq(dead_cpu
);
5958 struct task_struct
*next
, *stop
= rq
->stop
;
5962 * Fudge the rq selection such that the below task selection loop
5963 * doesn't get stuck on the currently eligible stop task.
5965 * We're currently inside stop_machine() and the rq is either stuck
5966 * in the stop_machine_cpu_stop() loop, or we're executing this code,
5967 * either way we should never end up calling schedule() until we're
5974 * There's this thread running, bail when that's the only
5977 if (rq
->nr_running
== 1)
5980 next
= pick_next_task(rq
);
5982 next
->sched_class
->put_prev_task(rq
, next
);
5984 /* Find suitable destination for @next, with force if needed. */
5985 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
5986 raw_spin_unlock(&rq
->lock
);
5988 __migrate_task(next
, dead_cpu
, dest_cpu
);
5990 raw_spin_lock(&rq
->lock
);
5996 #endif /* CONFIG_HOTPLUG_CPU */
5998 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6000 static struct ctl_table sd_ctl_dir
[] = {
6002 .procname
= "sched_domain",
6008 static struct ctl_table sd_ctl_root
[] = {
6010 .procname
= "kernel",
6012 .child
= sd_ctl_dir
,
6017 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6019 struct ctl_table
*entry
=
6020 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6025 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6027 struct ctl_table
*entry
;
6030 * In the intermediate directories, both the child directory and
6031 * procname are dynamically allocated and could fail but the mode
6032 * will always be set. In the lowest directory the names are
6033 * static strings and all have proc handlers.
6035 for (entry
= *tablep
; entry
->mode
; entry
++) {
6037 sd_free_ctl_entry(&entry
->child
);
6038 if (entry
->proc_handler
== NULL
)
6039 kfree(entry
->procname
);
6047 set_table_entry(struct ctl_table
*entry
,
6048 const char *procname
, void *data
, int maxlen
,
6049 mode_t mode
, proc_handler
*proc_handler
)
6051 entry
->procname
= procname
;
6053 entry
->maxlen
= maxlen
;
6055 entry
->proc_handler
= proc_handler
;
6058 static struct ctl_table
*
6059 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6061 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6066 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6067 sizeof(long), 0644, proc_doulongvec_minmax
);
6068 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6069 sizeof(long), 0644, proc_doulongvec_minmax
);
6070 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6071 sizeof(int), 0644, proc_dointvec_minmax
);
6072 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6073 sizeof(int), 0644, proc_dointvec_minmax
);
6074 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6075 sizeof(int), 0644, proc_dointvec_minmax
);
6076 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6077 sizeof(int), 0644, proc_dointvec_minmax
);
6078 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6079 sizeof(int), 0644, proc_dointvec_minmax
);
6080 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6081 sizeof(int), 0644, proc_dointvec_minmax
);
6082 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6083 sizeof(int), 0644, proc_dointvec_minmax
);
6084 set_table_entry(&table
[9], "cache_nice_tries",
6085 &sd
->cache_nice_tries
,
6086 sizeof(int), 0644, proc_dointvec_minmax
);
6087 set_table_entry(&table
[10], "flags", &sd
->flags
,
6088 sizeof(int), 0644, proc_dointvec_minmax
);
6089 set_table_entry(&table
[11], "name", sd
->name
,
6090 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6091 /* &table[12] is terminator */
6096 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6098 struct ctl_table
*entry
, *table
;
6099 struct sched_domain
*sd
;
6100 int domain_num
= 0, i
;
6103 for_each_domain(cpu
, sd
)
6105 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6110 for_each_domain(cpu
, sd
) {
6111 snprintf(buf
, 32, "domain%d", i
);
6112 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6114 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6121 static struct ctl_table_header
*sd_sysctl_header
;
6122 static void register_sched_domain_sysctl(void)
6124 int i
, cpu_num
= num_possible_cpus();
6125 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6128 WARN_ON(sd_ctl_dir
[0].child
);
6129 sd_ctl_dir
[0].child
= entry
;
6134 for_each_possible_cpu(i
) {
6135 snprintf(buf
, 32, "cpu%d", i
);
6136 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6138 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6142 WARN_ON(sd_sysctl_header
);
6143 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6146 /* may be called multiple times per register */
6147 static void unregister_sched_domain_sysctl(void)
6149 if (sd_sysctl_header
)
6150 unregister_sysctl_table(sd_sysctl_header
);
6151 sd_sysctl_header
= NULL
;
6152 if (sd_ctl_dir
[0].child
)
6153 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6156 static void register_sched_domain_sysctl(void)
6159 static void unregister_sched_domain_sysctl(void)
6164 static void set_rq_online(struct rq
*rq
)
6167 const struct sched_class
*class;
6169 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6172 for_each_class(class) {
6173 if (class->rq_online
)
6174 class->rq_online(rq
);
6179 static void set_rq_offline(struct rq
*rq
)
6182 const struct sched_class
*class;
6184 for_each_class(class) {
6185 if (class->rq_offline
)
6186 class->rq_offline(rq
);
6189 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6195 * migration_call - callback that gets triggered when a CPU is added.
6196 * Here we can start up the necessary migration thread for the new CPU.
6198 static int __cpuinit
6199 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6201 int cpu
= (long)hcpu
;
6202 unsigned long flags
;
6203 struct rq
*rq
= cpu_rq(cpu
);
6205 switch (action
& ~CPU_TASKS_FROZEN
) {
6207 case CPU_UP_PREPARE
:
6208 rq
->calc_load_update
= calc_load_update
;
6212 /* Update our root-domain */
6213 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6215 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6219 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6222 #ifdef CONFIG_HOTPLUG_CPU
6224 /* Update our root-domain */
6225 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6227 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6231 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6232 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6234 migrate_nr_uninterruptible(rq
);
6235 calc_global_load_remove(rq
);
6243 * Register at high priority so that task migration (migrate_all_tasks)
6244 * happens before everything else. This has to be lower priority than
6245 * the notifier in the perf_event subsystem, though.
6247 static struct notifier_block __cpuinitdata migration_notifier
= {
6248 .notifier_call
= migration_call
,
6249 .priority
= CPU_PRI_MIGRATION
,
6252 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6253 unsigned long action
, void *hcpu
)
6255 switch (action
& ~CPU_TASKS_FROZEN
) {
6257 case CPU_DOWN_FAILED
:
6258 set_cpu_active((long)hcpu
, true);
6265 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6266 unsigned long action
, void *hcpu
)
6268 switch (action
& ~CPU_TASKS_FROZEN
) {
6269 case CPU_DOWN_PREPARE
:
6270 set_cpu_active((long)hcpu
, false);
6277 static int __init
migration_init(void)
6279 void *cpu
= (void *)(long)smp_processor_id();
6282 /* Initialize migration for the boot CPU */
6283 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6284 BUG_ON(err
== NOTIFY_BAD
);
6285 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6286 register_cpu_notifier(&migration_notifier
);
6288 /* Register cpu active notifiers */
6289 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6290 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6294 early_initcall(migration_init
);
6299 #ifdef CONFIG_SCHED_DEBUG
6301 static __read_mostly
int sched_domain_debug_enabled
;
6303 static int __init
sched_domain_debug_setup(char *str
)
6305 sched_domain_debug_enabled
= 1;
6309 early_param("sched_debug", sched_domain_debug_setup
);
6311 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6312 struct cpumask
*groupmask
)
6314 struct sched_group
*group
= sd
->groups
;
6317 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6318 cpumask_clear(groupmask
);
6320 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6322 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6323 printk("does not load-balance\n");
6325 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6330 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6332 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6333 printk(KERN_ERR
"ERROR: domain->span does not contain "
6336 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6337 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6341 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6345 printk(KERN_ERR
"ERROR: group is NULL\n");
6349 if (!group
->cpu_power
) {
6350 printk(KERN_CONT
"\n");
6351 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6356 if (!cpumask_weight(sched_group_cpus(group
))) {
6357 printk(KERN_CONT
"\n");
6358 printk(KERN_ERR
"ERROR: empty group\n");
6362 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6363 printk(KERN_CONT
"\n");
6364 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6368 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6370 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6372 printk(KERN_CONT
" %s", str
);
6373 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6374 printk(KERN_CONT
" (cpu_power = %d)",
6378 group
= group
->next
;
6379 } while (group
!= sd
->groups
);
6380 printk(KERN_CONT
"\n");
6382 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6383 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6386 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6387 printk(KERN_ERR
"ERROR: parent span is not a superset "
6388 "of domain->span\n");
6392 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6394 cpumask_var_t groupmask
;
6397 if (!sched_domain_debug_enabled
)
6401 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6405 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6407 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6408 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6413 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6420 free_cpumask_var(groupmask
);
6422 #else /* !CONFIG_SCHED_DEBUG */
6423 # define sched_domain_debug(sd, cpu) do { } while (0)
6424 #endif /* CONFIG_SCHED_DEBUG */
6426 static int sd_degenerate(struct sched_domain
*sd
)
6428 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6431 /* Following flags need at least 2 groups */
6432 if (sd
->flags
& (SD_LOAD_BALANCE
|
6433 SD_BALANCE_NEWIDLE
|
6437 SD_SHARE_PKG_RESOURCES
)) {
6438 if (sd
->groups
!= sd
->groups
->next
)
6442 /* Following flags don't use groups */
6443 if (sd
->flags
& (SD_WAKE_AFFINE
))
6450 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6452 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6454 if (sd_degenerate(parent
))
6457 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6460 /* Flags needing groups don't count if only 1 group in parent */
6461 if (parent
->groups
== parent
->groups
->next
) {
6462 pflags
&= ~(SD_LOAD_BALANCE
|
6463 SD_BALANCE_NEWIDLE
|
6467 SD_SHARE_PKG_RESOURCES
);
6468 if (nr_node_ids
== 1)
6469 pflags
&= ~SD_SERIALIZE
;
6471 if (~cflags
& pflags
)
6477 static void free_rootdomain(struct root_domain
*rd
)
6479 synchronize_sched();
6481 cpupri_cleanup(&rd
->cpupri
);
6483 free_cpumask_var(rd
->rto_mask
);
6484 free_cpumask_var(rd
->online
);
6485 free_cpumask_var(rd
->span
);
6489 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6491 struct root_domain
*old_rd
= NULL
;
6492 unsigned long flags
;
6494 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6499 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6502 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6505 * If we dont want to free the old_rt yet then
6506 * set old_rd to NULL to skip the freeing later
6509 if (!atomic_dec_and_test(&old_rd
->refcount
))
6513 atomic_inc(&rd
->refcount
);
6516 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6517 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6520 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6523 free_rootdomain(old_rd
);
6526 static int init_rootdomain(struct root_domain
*rd
)
6528 memset(rd
, 0, sizeof(*rd
));
6530 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6532 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6534 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6537 if (cpupri_init(&rd
->cpupri
) != 0)
6542 free_cpumask_var(rd
->rto_mask
);
6544 free_cpumask_var(rd
->online
);
6546 free_cpumask_var(rd
->span
);
6551 static void init_defrootdomain(void)
6553 init_rootdomain(&def_root_domain
);
6555 atomic_set(&def_root_domain
.refcount
, 1);
6558 static struct root_domain
*alloc_rootdomain(void)
6560 struct root_domain
*rd
;
6562 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6566 if (init_rootdomain(rd
) != 0) {
6575 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6576 * hold the hotplug lock.
6579 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6581 struct rq
*rq
= cpu_rq(cpu
);
6582 struct sched_domain
*tmp
;
6584 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6585 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6587 /* Remove the sched domains which do not contribute to scheduling. */
6588 for (tmp
= sd
; tmp
; ) {
6589 struct sched_domain
*parent
= tmp
->parent
;
6593 if (sd_parent_degenerate(tmp
, parent
)) {
6594 tmp
->parent
= parent
->parent
;
6596 parent
->parent
->child
= tmp
;
6601 if (sd
&& sd_degenerate(sd
)) {
6607 sched_domain_debug(sd
, cpu
);
6609 rq_attach_root(rq
, rd
);
6610 rcu_assign_pointer(rq
->sd
, sd
);
6613 /* cpus with isolated domains */
6614 static cpumask_var_t cpu_isolated_map
;
6616 /* Setup the mask of cpus configured for isolated domains */
6617 static int __init
isolated_cpu_setup(char *str
)
6619 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6620 cpulist_parse(str
, cpu_isolated_map
);
6624 __setup("isolcpus=", isolated_cpu_setup
);
6627 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6628 * to a function which identifies what group(along with sched group) a CPU
6629 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6630 * (due to the fact that we keep track of groups covered with a struct cpumask).
6632 * init_sched_build_groups will build a circular linked list of the groups
6633 * covered by the given span, and will set each group's ->cpumask correctly,
6634 * and ->cpu_power to 0.
6637 init_sched_build_groups(const struct cpumask
*span
,
6638 const struct cpumask
*cpu_map
,
6639 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6640 struct sched_group
**sg
,
6641 struct cpumask
*tmpmask
),
6642 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6644 struct sched_group
*first
= NULL
, *last
= NULL
;
6647 cpumask_clear(covered
);
6649 for_each_cpu(i
, span
) {
6650 struct sched_group
*sg
;
6651 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6654 if (cpumask_test_cpu(i
, covered
))
6657 cpumask_clear(sched_group_cpus(sg
));
6660 for_each_cpu(j
, span
) {
6661 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6664 cpumask_set_cpu(j
, covered
);
6665 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6676 #define SD_NODES_PER_DOMAIN 16
6681 * find_next_best_node - find the next node to include in a sched_domain
6682 * @node: node whose sched_domain we're building
6683 * @used_nodes: nodes already in the sched_domain
6685 * Find the next node to include in a given scheduling domain. Simply
6686 * finds the closest node not already in the @used_nodes map.
6688 * Should use nodemask_t.
6690 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6692 int i
, n
, val
, min_val
, best_node
= 0;
6696 for (i
= 0; i
< nr_node_ids
; i
++) {
6697 /* Start at @node */
6698 n
= (node
+ i
) % nr_node_ids
;
6700 if (!nr_cpus_node(n
))
6703 /* Skip already used nodes */
6704 if (node_isset(n
, *used_nodes
))
6707 /* Simple min distance search */
6708 val
= node_distance(node
, n
);
6710 if (val
< min_val
) {
6716 node_set(best_node
, *used_nodes
);
6721 * sched_domain_node_span - get a cpumask for a node's sched_domain
6722 * @node: node whose cpumask we're constructing
6723 * @span: resulting cpumask
6725 * Given a node, construct a good cpumask for its sched_domain to span. It
6726 * should be one that prevents unnecessary balancing, but also spreads tasks
6729 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6731 nodemask_t used_nodes
;
6734 cpumask_clear(span
);
6735 nodes_clear(used_nodes
);
6737 cpumask_or(span
, span
, cpumask_of_node(node
));
6738 node_set(node
, used_nodes
);
6740 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6741 int next_node
= find_next_best_node(node
, &used_nodes
);
6743 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6746 #endif /* CONFIG_NUMA */
6748 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6751 * The cpus mask in sched_group and sched_domain hangs off the end.
6753 * ( See the the comments in include/linux/sched.h:struct sched_group
6754 * and struct sched_domain. )
6756 struct static_sched_group
{
6757 struct sched_group sg
;
6758 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6761 struct static_sched_domain
{
6762 struct sched_domain sd
;
6763 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6769 cpumask_var_t domainspan
;
6770 cpumask_var_t covered
;
6771 cpumask_var_t notcovered
;
6773 cpumask_var_t nodemask
;
6774 cpumask_var_t this_sibling_map
;
6775 cpumask_var_t this_core_map
;
6776 cpumask_var_t this_book_map
;
6777 cpumask_var_t send_covered
;
6778 cpumask_var_t tmpmask
;
6779 struct sched_group
**sched_group_nodes
;
6780 struct root_domain
*rd
;
6784 sa_sched_groups
= 0,
6790 sa_this_sibling_map
,
6792 sa_sched_group_nodes
,
6802 * SMT sched-domains:
6804 #ifdef CONFIG_SCHED_SMT
6805 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6806 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6809 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6810 struct sched_group
**sg
, struct cpumask
*unused
)
6813 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6816 #endif /* CONFIG_SCHED_SMT */
6819 * multi-core sched-domains:
6821 #ifdef CONFIG_SCHED_MC
6822 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6823 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6826 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6827 struct sched_group
**sg
, struct cpumask
*mask
)
6830 #ifdef CONFIG_SCHED_SMT
6831 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6832 group
= cpumask_first(mask
);
6837 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6840 #endif /* CONFIG_SCHED_MC */
6843 * book sched-domains:
6845 #ifdef CONFIG_SCHED_BOOK
6846 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
6847 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
6850 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
6851 struct sched_group
**sg
, struct cpumask
*mask
)
6854 #ifdef CONFIG_SCHED_MC
6855 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6856 group
= cpumask_first(mask
);
6857 #elif defined(CONFIG_SCHED_SMT)
6858 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6859 group
= cpumask_first(mask
);
6862 *sg
= &per_cpu(sched_group_book
, group
).sg
;
6865 #endif /* CONFIG_SCHED_BOOK */
6867 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6868 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6871 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6872 struct sched_group
**sg
, struct cpumask
*mask
)
6875 #ifdef CONFIG_SCHED_BOOK
6876 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
6877 group
= cpumask_first(mask
);
6878 #elif defined(CONFIG_SCHED_MC)
6879 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6880 group
= cpumask_first(mask
);
6881 #elif defined(CONFIG_SCHED_SMT)
6882 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6883 group
= cpumask_first(mask
);
6888 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6894 * The init_sched_build_groups can't handle what we want to do with node
6895 * groups, so roll our own. Now each node has its own list of groups which
6896 * gets dynamically allocated.
6898 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6899 static struct sched_group
***sched_group_nodes_bycpu
;
6901 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6902 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6904 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6905 struct sched_group
**sg
,
6906 struct cpumask
*nodemask
)
6910 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6911 group
= cpumask_first(nodemask
);
6914 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6918 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6920 struct sched_group
*sg
= group_head
;
6926 for_each_cpu(j
, sched_group_cpus(sg
)) {
6927 struct sched_domain
*sd
;
6929 sd
= &per_cpu(phys_domains
, j
).sd
;
6930 if (j
!= group_first_cpu(sd
->groups
)) {
6932 * Only add "power" once for each
6938 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6941 } while (sg
!= group_head
);
6944 static int build_numa_sched_groups(struct s_data
*d
,
6945 const struct cpumask
*cpu_map
, int num
)
6947 struct sched_domain
*sd
;
6948 struct sched_group
*sg
, *prev
;
6951 cpumask_clear(d
->covered
);
6952 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6953 if (cpumask_empty(d
->nodemask
)) {
6954 d
->sched_group_nodes
[num
] = NULL
;
6958 sched_domain_node_span(num
, d
->domainspan
);
6959 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6961 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6964 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6968 d
->sched_group_nodes
[num
] = sg
;
6970 for_each_cpu(j
, d
->nodemask
) {
6971 sd
= &per_cpu(node_domains
, j
).sd
;
6976 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6978 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6981 for (j
= 0; j
< nr_node_ids
; j
++) {
6982 n
= (num
+ j
) % nr_node_ids
;
6983 cpumask_complement(d
->notcovered
, d
->covered
);
6984 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6985 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6986 if (cpumask_empty(d
->tmpmask
))
6988 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6989 if (cpumask_empty(d
->tmpmask
))
6991 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6995 "Can not alloc domain group for node %d\n", j
);
6999 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
7000 sg
->next
= prev
->next
;
7001 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
7008 #endif /* CONFIG_NUMA */
7011 /* Free memory allocated for various sched_group structures */
7012 static void free_sched_groups(const struct cpumask
*cpu_map
,
7013 struct cpumask
*nodemask
)
7017 for_each_cpu(cpu
, cpu_map
) {
7018 struct sched_group
**sched_group_nodes
7019 = sched_group_nodes_bycpu
[cpu
];
7021 if (!sched_group_nodes
)
7024 for (i
= 0; i
< nr_node_ids
; i
++) {
7025 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7027 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7028 if (cpumask_empty(nodemask
))
7038 if (oldsg
!= sched_group_nodes
[i
])
7041 kfree(sched_group_nodes
);
7042 sched_group_nodes_bycpu
[cpu
] = NULL
;
7045 #else /* !CONFIG_NUMA */
7046 static void free_sched_groups(const struct cpumask
*cpu_map
,
7047 struct cpumask
*nodemask
)
7050 #endif /* CONFIG_NUMA */
7053 * Initialize sched groups cpu_power.
7055 * cpu_power indicates the capacity of sched group, which is used while
7056 * distributing the load between different sched groups in a sched domain.
7057 * Typically cpu_power for all the groups in a sched domain will be same unless
7058 * there are asymmetries in the topology. If there are asymmetries, group
7059 * having more cpu_power will pickup more load compared to the group having
7062 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7064 struct sched_domain
*child
;
7065 struct sched_group
*group
;
7069 WARN_ON(!sd
|| !sd
->groups
);
7071 if (cpu
!= group_first_cpu(sd
->groups
))
7074 sd
->groups
->group_weight
= cpumask_weight(sched_group_cpus(sd
->groups
));
7078 sd
->groups
->cpu_power
= 0;
7081 power
= SCHED_LOAD_SCALE
;
7082 weight
= cpumask_weight(sched_domain_span(sd
));
7084 * SMT siblings share the power of a single core.
7085 * Usually multiple threads get a better yield out of
7086 * that one core than a single thread would have,
7087 * reflect that in sd->smt_gain.
7089 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
7090 power
*= sd
->smt_gain
;
7092 power
>>= SCHED_LOAD_SHIFT
;
7094 sd
->groups
->cpu_power
+= power
;
7099 * Add cpu_power of each child group to this groups cpu_power.
7101 group
= child
->groups
;
7103 sd
->groups
->cpu_power
+= group
->cpu_power
;
7104 group
= group
->next
;
7105 } while (group
!= child
->groups
);
7109 * Initializers for schedule domains
7110 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7113 #ifdef CONFIG_SCHED_DEBUG
7114 # define SD_INIT_NAME(sd, type) sd->name = #type
7116 # define SD_INIT_NAME(sd, type) do { } while (0)
7119 #define SD_INIT(sd, type) sd_init_##type(sd)
7121 #define SD_INIT_FUNC(type) \
7122 static noinline void sd_init_##type(struct sched_domain *sd) \
7124 memset(sd, 0, sizeof(*sd)); \
7125 *sd = SD_##type##_INIT; \
7126 sd->level = SD_LV_##type; \
7127 SD_INIT_NAME(sd, type); \
7132 SD_INIT_FUNC(ALLNODES
)
7135 #ifdef CONFIG_SCHED_SMT
7136 SD_INIT_FUNC(SIBLING
)
7138 #ifdef CONFIG_SCHED_MC
7141 #ifdef CONFIG_SCHED_BOOK
7145 static int default_relax_domain_level
= -1;
7147 static int __init
setup_relax_domain_level(char *str
)
7151 val
= simple_strtoul(str
, NULL
, 0);
7152 if (val
< SD_LV_MAX
)
7153 default_relax_domain_level
= val
;
7157 __setup("relax_domain_level=", setup_relax_domain_level
);
7159 static void set_domain_attribute(struct sched_domain
*sd
,
7160 struct sched_domain_attr
*attr
)
7164 if (!attr
|| attr
->relax_domain_level
< 0) {
7165 if (default_relax_domain_level
< 0)
7168 request
= default_relax_domain_level
;
7170 request
= attr
->relax_domain_level
;
7171 if (request
< sd
->level
) {
7172 /* turn off idle balance on this domain */
7173 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7175 /* turn on idle balance on this domain */
7176 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7180 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7181 const struct cpumask
*cpu_map
)
7184 case sa_sched_groups
:
7185 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
7186 d
->sched_group_nodes
= NULL
;
7188 free_rootdomain(d
->rd
); /* fall through */
7190 free_cpumask_var(d
->tmpmask
); /* fall through */
7191 case sa_send_covered
:
7192 free_cpumask_var(d
->send_covered
); /* fall through */
7193 case sa_this_book_map
:
7194 free_cpumask_var(d
->this_book_map
); /* fall through */
7195 case sa_this_core_map
:
7196 free_cpumask_var(d
->this_core_map
); /* fall through */
7197 case sa_this_sibling_map
:
7198 free_cpumask_var(d
->this_sibling_map
); /* fall through */
7200 free_cpumask_var(d
->nodemask
); /* fall through */
7201 case sa_sched_group_nodes
:
7203 kfree(d
->sched_group_nodes
); /* fall through */
7205 free_cpumask_var(d
->notcovered
); /* fall through */
7207 free_cpumask_var(d
->covered
); /* fall through */
7209 free_cpumask_var(d
->domainspan
); /* fall through */
7216 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7217 const struct cpumask
*cpu_map
)
7220 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
7222 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
7223 return sa_domainspan
;
7224 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
7226 /* Allocate the per-node list of sched groups */
7227 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
7228 sizeof(struct sched_group
*), GFP_KERNEL
);
7229 if (!d
->sched_group_nodes
) {
7230 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7231 return sa_notcovered
;
7233 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
7235 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
7236 return sa_sched_group_nodes
;
7237 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
7239 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
7240 return sa_this_sibling_map
;
7241 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
7242 return sa_this_core_map
;
7243 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
7244 return sa_this_book_map
;
7245 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
7246 return sa_send_covered
;
7247 d
->rd
= alloc_rootdomain();
7249 printk(KERN_WARNING
"Cannot alloc root domain\n");
7252 return sa_rootdomain
;
7255 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
7256 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
7258 struct sched_domain
*sd
= NULL
;
7260 struct sched_domain
*parent
;
7263 if (cpumask_weight(cpu_map
) >
7264 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7265 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7266 SD_INIT(sd
, ALLNODES
);
7267 set_domain_attribute(sd
, attr
);
7268 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7269 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7274 sd
= &per_cpu(node_domains
, i
).sd
;
7276 set_domain_attribute(sd
, attr
);
7277 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7278 sd
->parent
= parent
;
7281 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7286 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7287 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7288 struct sched_domain
*parent
, int i
)
7290 struct sched_domain
*sd
;
7291 sd
= &per_cpu(phys_domains
, i
).sd
;
7293 set_domain_attribute(sd
, attr
);
7294 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7295 sd
->parent
= parent
;
7298 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7302 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7303 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7304 struct sched_domain
*parent
, int i
)
7306 struct sched_domain
*sd
= parent
;
7307 #ifdef CONFIG_SCHED_BOOK
7308 sd
= &per_cpu(book_domains
, i
).sd
;
7310 set_domain_attribute(sd
, attr
);
7311 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7312 sd
->parent
= parent
;
7314 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7319 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7320 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7321 struct sched_domain
*parent
, int i
)
7323 struct sched_domain
*sd
= parent
;
7324 #ifdef CONFIG_SCHED_MC
7325 sd
= &per_cpu(core_domains
, i
).sd
;
7327 set_domain_attribute(sd
, attr
);
7328 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7329 sd
->parent
= parent
;
7331 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7336 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7337 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7338 struct sched_domain
*parent
, int i
)
7340 struct sched_domain
*sd
= parent
;
7341 #ifdef CONFIG_SCHED_SMT
7342 sd
= &per_cpu(cpu_domains
, i
).sd
;
7343 SD_INIT(sd
, SIBLING
);
7344 set_domain_attribute(sd
, attr
);
7345 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7346 sd
->parent
= parent
;
7348 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7353 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7354 const struct cpumask
*cpu_map
, int cpu
)
7357 #ifdef CONFIG_SCHED_SMT
7358 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7359 cpumask_and(d
->this_sibling_map
, cpu_map
,
7360 topology_thread_cpumask(cpu
));
7361 if (cpu
== cpumask_first(d
->this_sibling_map
))
7362 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7364 d
->send_covered
, d
->tmpmask
);
7367 #ifdef CONFIG_SCHED_MC
7368 case SD_LV_MC
: /* set up multi-core groups */
7369 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7370 if (cpu
== cpumask_first(d
->this_core_map
))
7371 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7373 d
->send_covered
, d
->tmpmask
);
7376 #ifdef CONFIG_SCHED_BOOK
7377 case SD_LV_BOOK
: /* set up book groups */
7378 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7379 if (cpu
== cpumask_first(d
->this_book_map
))
7380 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7382 d
->send_covered
, d
->tmpmask
);
7385 case SD_LV_CPU
: /* set up physical groups */
7386 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7387 if (!cpumask_empty(d
->nodemask
))
7388 init_sched_build_groups(d
->nodemask
, cpu_map
,
7390 d
->send_covered
, d
->tmpmask
);
7393 case SD_LV_ALLNODES
:
7394 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7395 d
->send_covered
, d
->tmpmask
);
7404 * Build sched domains for a given set of cpus and attach the sched domains
7405 * to the individual cpus
7407 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7408 struct sched_domain_attr
*attr
)
7410 enum s_alloc alloc_state
= sa_none
;
7412 struct sched_domain
*sd
;
7418 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7419 if (alloc_state
!= sa_rootdomain
)
7421 alloc_state
= sa_sched_groups
;
7424 * Set up domains for cpus specified by the cpu_map.
7426 for_each_cpu(i
, cpu_map
) {
7427 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7430 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7431 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7432 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7433 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7434 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7437 for_each_cpu(i
, cpu_map
) {
7438 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7439 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7440 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7443 /* Set up physical groups */
7444 for (i
= 0; i
< nr_node_ids
; i
++)
7445 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7448 /* Set up node groups */
7450 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7452 for (i
= 0; i
< nr_node_ids
; i
++)
7453 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7457 /* Calculate CPU power for physical packages and nodes */
7458 #ifdef CONFIG_SCHED_SMT
7459 for_each_cpu(i
, cpu_map
) {
7460 sd
= &per_cpu(cpu_domains
, i
).sd
;
7461 init_sched_groups_power(i
, sd
);
7464 #ifdef CONFIG_SCHED_MC
7465 for_each_cpu(i
, cpu_map
) {
7466 sd
= &per_cpu(core_domains
, i
).sd
;
7467 init_sched_groups_power(i
, sd
);
7470 #ifdef CONFIG_SCHED_BOOK
7471 for_each_cpu(i
, cpu_map
) {
7472 sd
= &per_cpu(book_domains
, i
).sd
;
7473 init_sched_groups_power(i
, sd
);
7477 for_each_cpu(i
, cpu_map
) {
7478 sd
= &per_cpu(phys_domains
, i
).sd
;
7479 init_sched_groups_power(i
, sd
);
7483 for (i
= 0; i
< nr_node_ids
; i
++)
7484 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7486 if (d
.sd_allnodes
) {
7487 struct sched_group
*sg
;
7489 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7491 init_numa_sched_groups_power(sg
);
7495 /* Attach the domains */
7496 for_each_cpu(i
, cpu_map
) {
7497 #ifdef CONFIG_SCHED_SMT
7498 sd
= &per_cpu(cpu_domains
, i
).sd
;
7499 #elif defined(CONFIG_SCHED_MC)
7500 sd
= &per_cpu(core_domains
, i
).sd
;
7501 #elif defined(CONFIG_SCHED_BOOK)
7502 sd
= &per_cpu(book_domains
, i
).sd
;
7504 sd
= &per_cpu(phys_domains
, i
).sd
;
7506 cpu_attach_domain(sd
, d
.rd
, i
);
7509 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7510 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7514 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7518 static int build_sched_domains(const struct cpumask
*cpu_map
)
7520 return __build_sched_domains(cpu_map
, NULL
);
7523 static cpumask_var_t
*doms_cur
; /* current sched domains */
7524 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7525 static struct sched_domain_attr
*dattr_cur
;
7526 /* attribues of custom domains in 'doms_cur' */
7529 * Special case: If a kmalloc of a doms_cur partition (array of
7530 * cpumask) fails, then fallback to a single sched domain,
7531 * as determined by the single cpumask fallback_doms.
7533 static cpumask_var_t fallback_doms
;
7536 * arch_update_cpu_topology lets virtualized architectures update the
7537 * cpu core maps. It is supposed to return 1 if the topology changed
7538 * or 0 if it stayed the same.
7540 int __attribute__((weak
)) arch_update_cpu_topology(void)
7545 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7548 cpumask_var_t
*doms
;
7550 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7553 for (i
= 0; i
< ndoms
; i
++) {
7554 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7555 free_sched_domains(doms
, i
);
7562 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7565 for (i
= 0; i
< ndoms
; i
++)
7566 free_cpumask_var(doms
[i
]);
7571 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7572 * For now this just excludes isolated cpus, but could be used to
7573 * exclude other special cases in the future.
7575 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7579 arch_update_cpu_topology();
7581 doms_cur
= alloc_sched_domains(ndoms_cur
);
7583 doms_cur
= &fallback_doms
;
7584 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7586 err
= build_sched_domains(doms_cur
[0]);
7587 register_sched_domain_sysctl();
7592 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7593 struct cpumask
*tmpmask
)
7595 free_sched_groups(cpu_map
, tmpmask
);
7599 * Detach sched domains from a group of cpus specified in cpu_map
7600 * These cpus will now be attached to the NULL domain
7602 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7604 /* Save because hotplug lock held. */
7605 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7608 for_each_cpu(i
, cpu_map
)
7609 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7610 synchronize_sched();
7611 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7614 /* handle null as "default" */
7615 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7616 struct sched_domain_attr
*new, int idx_new
)
7618 struct sched_domain_attr tmp
;
7625 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7626 new ? (new + idx_new
) : &tmp
,
7627 sizeof(struct sched_domain_attr
));
7631 * Partition sched domains as specified by the 'ndoms_new'
7632 * cpumasks in the array doms_new[] of cpumasks. This compares
7633 * doms_new[] to the current sched domain partitioning, doms_cur[].
7634 * It destroys each deleted domain and builds each new domain.
7636 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7637 * The masks don't intersect (don't overlap.) We should setup one
7638 * sched domain for each mask. CPUs not in any of the cpumasks will
7639 * not be load balanced. If the same cpumask appears both in the
7640 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7643 * The passed in 'doms_new' should be allocated using
7644 * alloc_sched_domains. This routine takes ownership of it and will
7645 * free_sched_domains it when done with it. If the caller failed the
7646 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7647 * and partition_sched_domains() will fallback to the single partition
7648 * 'fallback_doms', it also forces the domains to be rebuilt.
7650 * If doms_new == NULL it will be replaced with cpu_online_mask.
7651 * ndoms_new == 0 is a special case for destroying existing domains,
7652 * and it will not create the default domain.
7654 * Call with hotplug lock held
7656 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7657 struct sched_domain_attr
*dattr_new
)
7662 mutex_lock(&sched_domains_mutex
);
7664 /* always unregister in case we don't destroy any domains */
7665 unregister_sched_domain_sysctl();
7667 /* Let architecture update cpu core mappings. */
7668 new_topology
= arch_update_cpu_topology();
7670 n
= doms_new
? ndoms_new
: 0;
7672 /* Destroy deleted domains */
7673 for (i
= 0; i
< ndoms_cur
; i
++) {
7674 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7675 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7676 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7679 /* no match - a current sched domain not in new doms_new[] */
7680 detach_destroy_domains(doms_cur
[i
]);
7685 if (doms_new
== NULL
) {
7687 doms_new
= &fallback_doms
;
7688 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7689 WARN_ON_ONCE(dattr_new
);
7692 /* Build new domains */
7693 for (i
= 0; i
< ndoms_new
; i
++) {
7694 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7695 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7696 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7699 /* no match - add a new doms_new */
7700 __build_sched_domains(doms_new
[i
],
7701 dattr_new
? dattr_new
+ i
: NULL
);
7706 /* Remember the new sched domains */
7707 if (doms_cur
!= &fallback_doms
)
7708 free_sched_domains(doms_cur
, ndoms_cur
);
7709 kfree(dattr_cur
); /* kfree(NULL) is safe */
7710 doms_cur
= doms_new
;
7711 dattr_cur
= dattr_new
;
7712 ndoms_cur
= ndoms_new
;
7714 register_sched_domain_sysctl();
7716 mutex_unlock(&sched_domains_mutex
);
7719 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7720 static void arch_reinit_sched_domains(void)
7724 /* Destroy domains first to force the rebuild */
7725 partition_sched_domains(0, NULL
, NULL
);
7727 rebuild_sched_domains();
7731 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7733 unsigned int level
= 0;
7735 if (sscanf(buf
, "%u", &level
) != 1)
7739 * level is always be positive so don't check for
7740 * level < POWERSAVINGS_BALANCE_NONE which is 0
7741 * What happens on 0 or 1 byte write,
7742 * need to check for count as well?
7745 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7749 sched_smt_power_savings
= level
;
7751 sched_mc_power_savings
= level
;
7753 arch_reinit_sched_domains();
7758 #ifdef CONFIG_SCHED_MC
7759 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7760 struct sysdev_class_attribute
*attr
,
7763 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7765 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7766 struct sysdev_class_attribute
*attr
,
7767 const char *buf
, size_t count
)
7769 return sched_power_savings_store(buf
, count
, 0);
7771 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7772 sched_mc_power_savings_show
,
7773 sched_mc_power_savings_store
);
7776 #ifdef CONFIG_SCHED_SMT
7777 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7778 struct sysdev_class_attribute
*attr
,
7781 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7783 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7784 struct sysdev_class_attribute
*attr
,
7785 const char *buf
, size_t count
)
7787 return sched_power_savings_store(buf
, count
, 1);
7789 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7790 sched_smt_power_savings_show
,
7791 sched_smt_power_savings_store
);
7794 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7798 #ifdef CONFIG_SCHED_SMT
7800 err
= sysfs_create_file(&cls
->kset
.kobj
,
7801 &attr_sched_smt_power_savings
.attr
);
7803 #ifdef CONFIG_SCHED_MC
7804 if (!err
&& mc_capable())
7805 err
= sysfs_create_file(&cls
->kset
.kobj
,
7806 &attr_sched_mc_power_savings
.attr
);
7810 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7813 * Update cpusets according to cpu_active mask. If cpusets are
7814 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7815 * around partition_sched_domains().
7817 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7820 switch (action
& ~CPU_TASKS_FROZEN
) {
7822 case CPU_DOWN_FAILED
:
7823 cpuset_update_active_cpus();
7830 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
7833 switch (action
& ~CPU_TASKS_FROZEN
) {
7834 case CPU_DOWN_PREPARE
:
7835 cpuset_update_active_cpus();
7842 static int update_runtime(struct notifier_block
*nfb
,
7843 unsigned long action
, void *hcpu
)
7845 int cpu
= (int)(long)hcpu
;
7848 case CPU_DOWN_PREPARE
:
7849 case CPU_DOWN_PREPARE_FROZEN
:
7850 disable_runtime(cpu_rq(cpu
));
7853 case CPU_DOWN_FAILED
:
7854 case CPU_DOWN_FAILED_FROZEN
:
7856 case CPU_ONLINE_FROZEN
:
7857 enable_runtime(cpu_rq(cpu
));
7865 void __init
sched_init_smp(void)
7867 cpumask_var_t non_isolated_cpus
;
7869 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7870 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7872 #if defined(CONFIG_NUMA)
7873 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7875 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7878 mutex_lock(&sched_domains_mutex
);
7879 arch_init_sched_domains(cpu_active_mask
);
7880 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7881 if (cpumask_empty(non_isolated_cpus
))
7882 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7883 mutex_unlock(&sched_domains_mutex
);
7886 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
7887 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
7889 /* RT runtime code needs to handle some hotplug events */
7890 hotcpu_notifier(update_runtime
, 0);
7894 /* Move init over to a non-isolated CPU */
7895 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7897 sched_init_granularity();
7898 free_cpumask_var(non_isolated_cpus
);
7900 init_sched_rt_class();
7903 void __init
sched_init_smp(void)
7905 sched_init_granularity();
7907 #endif /* CONFIG_SMP */
7909 const_debug
unsigned int sysctl_timer_migration
= 1;
7911 int in_sched_functions(unsigned long addr
)
7913 return in_lock_functions(addr
) ||
7914 (addr
>= (unsigned long)__sched_text_start
7915 && addr
< (unsigned long)__sched_text_end
);
7918 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7920 cfs_rq
->tasks_timeline
= RB_ROOT
;
7921 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7922 #ifdef CONFIG_FAIR_GROUP_SCHED
7924 /* allow initial update_cfs_load() to truncate */
7926 cfs_rq
->load_stamp
= 1;
7929 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7932 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7934 struct rt_prio_array
*array
;
7937 array
= &rt_rq
->active
;
7938 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7939 INIT_LIST_HEAD(array
->queue
+ i
);
7940 __clear_bit(i
, array
->bitmap
);
7942 /* delimiter for bitsearch: */
7943 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7945 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7946 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7948 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7952 rt_rq
->rt_nr_migratory
= 0;
7953 rt_rq
->overloaded
= 0;
7954 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7958 rt_rq
->rt_throttled
= 0;
7959 rt_rq
->rt_runtime
= 0;
7960 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7962 #ifdef CONFIG_RT_GROUP_SCHED
7963 rt_rq
->rt_nr_boosted
= 0;
7968 #ifdef CONFIG_FAIR_GROUP_SCHED
7969 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7970 struct sched_entity
*se
, int cpu
,
7971 struct sched_entity
*parent
)
7973 struct rq
*rq
= cpu_rq(cpu
);
7974 tg
->cfs_rq
[cpu
] = cfs_rq
;
7975 init_cfs_rq(cfs_rq
, rq
);
7979 /* se could be NULL for root_task_group */
7984 se
->cfs_rq
= &rq
->cfs
;
7986 se
->cfs_rq
= parent
->my_q
;
7989 update_load_set(&se
->load
, 0);
7990 se
->parent
= parent
;
7994 #ifdef CONFIG_RT_GROUP_SCHED
7995 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7996 struct sched_rt_entity
*rt_se
, int cpu
,
7997 struct sched_rt_entity
*parent
)
7999 struct rq
*rq
= cpu_rq(cpu
);
8001 tg
->rt_rq
[cpu
] = rt_rq
;
8002 init_rt_rq(rt_rq
, rq
);
8004 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8006 tg
->rt_se
[cpu
] = rt_se
;
8011 rt_se
->rt_rq
= &rq
->rt
;
8013 rt_se
->rt_rq
= parent
->my_q
;
8015 rt_se
->my_q
= rt_rq
;
8016 rt_se
->parent
= parent
;
8017 INIT_LIST_HEAD(&rt_se
->run_list
);
8021 void __init
sched_init(void)
8024 unsigned long alloc_size
= 0, ptr
;
8026 #ifdef CONFIG_FAIR_GROUP_SCHED
8027 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8029 #ifdef CONFIG_RT_GROUP_SCHED
8030 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8032 #ifdef CONFIG_CPUMASK_OFFSTACK
8033 alloc_size
+= num_possible_cpus() * cpumask_size();
8036 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
8038 #ifdef CONFIG_FAIR_GROUP_SCHED
8039 root_task_group
.se
= (struct sched_entity
**)ptr
;
8040 ptr
+= nr_cpu_ids
* sizeof(void **);
8042 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8043 ptr
+= nr_cpu_ids
* sizeof(void **);
8045 #endif /* CONFIG_FAIR_GROUP_SCHED */
8046 #ifdef CONFIG_RT_GROUP_SCHED
8047 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8048 ptr
+= nr_cpu_ids
* sizeof(void **);
8050 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8051 ptr
+= nr_cpu_ids
* sizeof(void **);
8053 #endif /* CONFIG_RT_GROUP_SCHED */
8054 #ifdef CONFIG_CPUMASK_OFFSTACK
8055 for_each_possible_cpu(i
) {
8056 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8057 ptr
+= cpumask_size();
8059 #endif /* CONFIG_CPUMASK_OFFSTACK */
8063 init_defrootdomain();
8066 init_rt_bandwidth(&def_rt_bandwidth
,
8067 global_rt_period(), global_rt_runtime());
8069 #ifdef CONFIG_RT_GROUP_SCHED
8070 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8071 global_rt_period(), global_rt_runtime());
8072 #endif /* CONFIG_RT_GROUP_SCHED */
8074 #ifdef CONFIG_CGROUP_SCHED
8075 list_add(&root_task_group
.list
, &task_groups
);
8076 INIT_LIST_HEAD(&root_task_group
.children
);
8077 autogroup_init(&init_task
);
8078 #endif /* CONFIG_CGROUP_SCHED */
8080 for_each_possible_cpu(i
) {
8084 raw_spin_lock_init(&rq
->lock
);
8086 rq
->calc_load_active
= 0;
8087 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8088 init_cfs_rq(&rq
->cfs
, rq
);
8089 init_rt_rq(&rq
->rt
, rq
);
8090 #ifdef CONFIG_FAIR_GROUP_SCHED
8091 root_task_group
.shares
= root_task_group_load
;
8092 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8094 * How much cpu bandwidth does root_task_group get?
8096 * In case of task-groups formed thr' the cgroup filesystem, it
8097 * gets 100% of the cpu resources in the system. This overall
8098 * system cpu resource is divided among the tasks of
8099 * root_task_group and its child task-groups in a fair manner,
8100 * based on each entity's (task or task-group's) weight
8101 * (se->load.weight).
8103 * In other words, if root_task_group has 10 tasks of weight
8104 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8105 * then A0's share of the cpu resource is:
8107 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8109 * We achieve this by letting root_task_group's tasks sit
8110 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8112 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8113 #endif /* CONFIG_FAIR_GROUP_SCHED */
8115 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8116 #ifdef CONFIG_RT_GROUP_SCHED
8117 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8118 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8121 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8122 rq
->cpu_load
[j
] = 0;
8124 rq
->last_load_update_tick
= jiffies
;
8129 rq
->cpu_power
= SCHED_LOAD_SCALE
;
8130 rq
->post_schedule
= 0;
8131 rq
->active_balance
= 0;
8132 rq
->next_balance
= jiffies
;
8137 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8138 rq_attach_root(rq
, &def_root_domain
);
8140 rq
->nohz_balance_kick
= 0;
8141 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
8145 atomic_set(&rq
->nr_iowait
, 0);
8148 set_load_weight(&init_task
);
8150 #ifdef CONFIG_PREEMPT_NOTIFIERS
8151 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8155 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8158 #ifdef CONFIG_RT_MUTEXES
8159 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8163 * The boot idle thread does lazy MMU switching as well:
8165 atomic_inc(&init_mm
.mm_count
);
8166 enter_lazy_tlb(&init_mm
, current
);
8169 * Make us the idle thread. Technically, schedule() should not be
8170 * called from this thread, however somewhere below it might be,
8171 * but because we are the idle thread, we just pick up running again
8172 * when this runqueue becomes "idle".
8174 init_idle(current
, smp_processor_id());
8176 calc_load_update
= jiffies
+ LOAD_FREQ
;
8179 * During early bootup we pretend to be a normal task:
8181 current
->sched_class
= &fair_sched_class
;
8183 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8184 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
8187 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8188 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8189 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8190 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8191 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8193 /* May be allocated at isolcpus cmdline parse time */
8194 if (cpu_isolated_map
== NULL
)
8195 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8198 scheduler_running
= 1;
8201 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8202 static inline int preempt_count_equals(int preempt_offset
)
8204 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8206 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
8209 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8212 static unsigned long prev_jiffy
; /* ratelimiting */
8214 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8215 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8217 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8219 prev_jiffy
= jiffies
;
8222 "BUG: sleeping function called from invalid context at %s:%d\n",
8225 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8226 in_atomic(), irqs_disabled(),
8227 current
->pid
, current
->comm
);
8229 debug_show_held_locks(current
);
8230 if (irqs_disabled())
8231 print_irqtrace_events(current
);
8235 EXPORT_SYMBOL(__might_sleep
);
8238 #ifdef CONFIG_MAGIC_SYSRQ
8239 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8241 const struct sched_class
*prev_class
= p
->sched_class
;
8242 int old_prio
= p
->prio
;
8245 on_rq
= p
->se
.on_rq
;
8247 deactivate_task(rq
, p
, 0);
8248 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8250 activate_task(rq
, p
, 0);
8251 resched_task(rq
->curr
);
8254 check_class_changed(rq
, p
, prev_class
, old_prio
);
8257 void normalize_rt_tasks(void)
8259 struct task_struct
*g
, *p
;
8260 unsigned long flags
;
8263 read_lock_irqsave(&tasklist_lock
, flags
);
8264 do_each_thread(g
, p
) {
8266 * Only normalize user tasks:
8271 p
->se
.exec_start
= 0;
8272 #ifdef CONFIG_SCHEDSTATS
8273 p
->se
.statistics
.wait_start
= 0;
8274 p
->se
.statistics
.sleep_start
= 0;
8275 p
->se
.statistics
.block_start
= 0;
8280 * Renice negative nice level userspace
8283 if (TASK_NICE(p
) < 0 && p
->mm
)
8284 set_user_nice(p
, 0);
8288 raw_spin_lock(&p
->pi_lock
);
8289 rq
= __task_rq_lock(p
);
8291 normalize_task(rq
, p
);
8293 __task_rq_unlock(rq
);
8294 raw_spin_unlock(&p
->pi_lock
);
8295 } while_each_thread(g
, p
);
8297 read_unlock_irqrestore(&tasklist_lock
, flags
);
8300 #endif /* CONFIG_MAGIC_SYSRQ */
8302 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8304 * These functions are only useful for the IA64 MCA handling, or kdb.
8306 * They can only be called when the whole system has been
8307 * stopped - every CPU needs to be quiescent, and no scheduling
8308 * activity can take place. Using them for anything else would
8309 * be a serious bug, and as a result, they aren't even visible
8310 * under any other configuration.
8314 * curr_task - return the current task for a given cpu.
8315 * @cpu: the processor in question.
8317 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8319 struct task_struct
*curr_task(int cpu
)
8321 return cpu_curr(cpu
);
8324 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8328 * set_curr_task - set the current task for a given cpu.
8329 * @cpu: the processor in question.
8330 * @p: the task pointer to set.
8332 * Description: This function must only be used when non-maskable interrupts
8333 * are serviced on a separate stack. It allows the architecture to switch the
8334 * notion of the current task on a cpu in a non-blocking manner. This function
8335 * must be called with all CPU's synchronized, and interrupts disabled, the
8336 * and caller must save the original value of the current task (see
8337 * curr_task() above) and restore that value before reenabling interrupts and
8338 * re-starting the system.
8340 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8342 void set_curr_task(int cpu
, struct task_struct
*p
)
8349 #ifdef CONFIG_FAIR_GROUP_SCHED
8350 static void free_fair_sched_group(struct task_group
*tg
)
8354 for_each_possible_cpu(i
) {
8356 kfree(tg
->cfs_rq
[i
]);
8366 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8368 struct cfs_rq
*cfs_rq
;
8369 struct sched_entity
*se
;
8373 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8376 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8380 tg
->shares
= NICE_0_LOAD
;
8382 for_each_possible_cpu(i
) {
8385 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8386 GFP_KERNEL
, cpu_to_node(i
));
8390 se
= kzalloc_node(sizeof(struct sched_entity
),
8391 GFP_KERNEL
, cpu_to_node(i
));
8395 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8406 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8408 struct rq
*rq
= cpu_rq(cpu
);
8409 unsigned long flags
;
8412 * Only empty task groups can be destroyed; so we can speculatively
8413 * check on_list without danger of it being re-added.
8415 if (!tg
->cfs_rq
[cpu
]->on_list
)
8418 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8419 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8420 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8422 #else /* !CONFG_FAIR_GROUP_SCHED */
8423 static inline void free_fair_sched_group(struct task_group
*tg
)
8428 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8433 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8436 #endif /* CONFIG_FAIR_GROUP_SCHED */
8438 #ifdef CONFIG_RT_GROUP_SCHED
8439 static void free_rt_sched_group(struct task_group
*tg
)
8443 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8445 for_each_possible_cpu(i
) {
8447 kfree(tg
->rt_rq
[i
]);
8449 kfree(tg
->rt_se
[i
]);
8457 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8459 struct rt_rq
*rt_rq
;
8460 struct sched_rt_entity
*rt_se
;
8464 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8467 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8471 init_rt_bandwidth(&tg
->rt_bandwidth
,
8472 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8474 for_each_possible_cpu(i
) {
8477 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8478 GFP_KERNEL
, cpu_to_node(i
));
8482 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8483 GFP_KERNEL
, cpu_to_node(i
));
8487 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8497 #else /* !CONFIG_RT_GROUP_SCHED */
8498 static inline void free_rt_sched_group(struct task_group
*tg
)
8503 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8507 #endif /* CONFIG_RT_GROUP_SCHED */
8509 #ifdef CONFIG_CGROUP_SCHED
8510 static void free_sched_group(struct task_group
*tg
)
8512 free_fair_sched_group(tg
);
8513 free_rt_sched_group(tg
);
8518 /* allocate runqueue etc for a new task group */
8519 struct task_group
*sched_create_group(struct task_group
*parent
)
8521 struct task_group
*tg
;
8522 unsigned long flags
;
8524 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8526 return ERR_PTR(-ENOMEM
);
8528 if (!alloc_fair_sched_group(tg
, parent
))
8531 if (!alloc_rt_sched_group(tg
, parent
))
8534 spin_lock_irqsave(&task_group_lock
, flags
);
8535 list_add_rcu(&tg
->list
, &task_groups
);
8537 WARN_ON(!parent
); /* root should already exist */
8539 tg
->parent
= parent
;
8540 INIT_LIST_HEAD(&tg
->children
);
8541 list_add_rcu(&tg
->siblings
, &parent
->children
);
8542 spin_unlock_irqrestore(&task_group_lock
, flags
);
8547 free_sched_group(tg
);
8548 return ERR_PTR(-ENOMEM
);
8551 /* rcu callback to free various structures associated with a task group */
8552 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8554 /* now it should be safe to free those cfs_rqs */
8555 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8558 /* Destroy runqueue etc associated with a task group */
8559 void sched_destroy_group(struct task_group
*tg
)
8561 unsigned long flags
;
8564 /* end participation in shares distribution */
8565 for_each_possible_cpu(i
)
8566 unregister_fair_sched_group(tg
, i
);
8568 spin_lock_irqsave(&task_group_lock
, flags
);
8569 list_del_rcu(&tg
->list
);
8570 list_del_rcu(&tg
->siblings
);
8571 spin_unlock_irqrestore(&task_group_lock
, flags
);
8573 /* wait for possible concurrent references to cfs_rqs complete */
8574 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8577 /* change task's runqueue when it moves between groups.
8578 * The caller of this function should have put the task in its new group
8579 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8580 * reflect its new group.
8582 void sched_move_task(struct task_struct
*tsk
)
8585 unsigned long flags
;
8588 rq
= task_rq_lock(tsk
, &flags
);
8590 running
= task_current(rq
, tsk
);
8591 on_rq
= tsk
->se
.on_rq
;
8594 dequeue_task(rq
, tsk
, 0);
8595 if (unlikely(running
))
8596 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8598 #ifdef CONFIG_FAIR_GROUP_SCHED
8599 if (tsk
->sched_class
->task_move_group
)
8600 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8603 set_task_rq(tsk
, task_cpu(tsk
));
8605 if (unlikely(running
))
8606 tsk
->sched_class
->set_curr_task(rq
);
8608 enqueue_task(rq
, tsk
, 0);
8610 task_rq_unlock(rq
, &flags
);
8612 #endif /* CONFIG_CGROUP_SCHED */
8614 #ifdef CONFIG_FAIR_GROUP_SCHED
8615 static DEFINE_MUTEX(shares_mutex
);
8617 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8620 unsigned long flags
;
8623 * We can't change the weight of the root cgroup.
8628 if (shares
< MIN_SHARES
)
8629 shares
= MIN_SHARES
;
8630 else if (shares
> MAX_SHARES
)
8631 shares
= MAX_SHARES
;
8633 mutex_lock(&shares_mutex
);
8634 if (tg
->shares
== shares
)
8637 tg
->shares
= shares
;
8638 for_each_possible_cpu(i
) {
8639 struct rq
*rq
= cpu_rq(i
);
8640 struct sched_entity
*se
;
8643 /* Propagate contribution to hierarchy */
8644 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8645 for_each_sched_entity(se
)
8646 update_cfs_shares(group_cfs_rq(se
));
8647 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8651 mutex_unlock(&shares_mutex
);
8655 unsigned long sched_group_shares(struct task_group
*tg
)
8661 #ifdef CONFIG_RT_GROUP_SCHED
8663 * Ensure that the real time constraints are schedulable.
8665 static DEFINE_MUTEX(rt_constraints_mutex
);
8667 static unsigned long to_ratio(u64 period
, u64 runtime
)
8669 if (runtime
== RUNTIME_INF
)
8672 return div64_u64(runtime
<< 20, period
);
8675 /* Must be called with tasklist_lock held */
8676 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8678 struct task_struct
*g
, *p
;
8680 do_each_thread(g
, p
) {
8681 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8683 } while_each_thread(g
, p
);
8688 struct rt_schedulable_data
{
8689 struct task_group
*tg
;
8694 static int tg_schedulable(struct task_group
*tg
, void *data
)
8696 struct rt_schedulable_data
*d
= data
;
8697 struct task_group
*child
;
8698 unsigned long total
, sum
= 0;
8699 u64 period
, runtime
;
8701 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8702 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8705 period
= d
->rt_period
;
8706 runtime
= d
->rt_runtime
;
8710 * Cannot have more runtime than the period.
8712 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8716 * Ensure we don't starve existing RT tasks.
8718 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8721 total
= to_ratio(period
, runtime
);
8724 * Nobody can have more than the global setting allows.
8726 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8730 * The sum of our children's runtime should not exceed our own.
8732 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8733 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8734 runtime
= child
->rt_bandwidth
.rt_runtime
;
8736 if (child
== d
->tg
) {
8737 period
= d
->rt_period
;
8738 runtime
= d
->rt_runtime
;
8741 sum
+= to_ratio(period
, runtime
);
8750 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8752 struct rt_schedulable_data data
= {
8754 .rt_period
= period
,
8755 .rt_runtime
= runtime
,
8758 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8761 static int tg_set_bandwidth(struct task_group
*tg
,
8762 u64 rt_period
, u64 rt_runtime
)
8766 mutex_lock(&rt_constraints_mutex
);
8767 read_lock(&tasklist_lock
);
8768 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8772 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8773 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8774 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8776 for_each_possible_cpu(i
) {
8777 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8779 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8780 rt_rq
->rt_runtime
= rt_runtime
;
8781 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8783 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8785 read_unlock(&tasklist_lock
);
8786 mutex_unlock(&rt_constraints_mutex
);
8791 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8793 u64 rt_runtime
, rt_period
;
8795 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8796 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8797 if (rt_runtime_us
< 0)
8798 rt_runtime
= RUNTIME_INF
;
8800 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8803 long sched_group_rt_runtime(struct task_group
*tg
)
8807 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8810 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8811 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8812 return rt_runtime_us
;
8815 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8817 u64 rt_runtime
, rt_period
;
8819 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8820 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8825 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8828 long sched_group_rt_period(struct task_group
*tg
)
8832 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8833 do_div(rt_period_us
, NSEC_PER_USEC
);
8834 return rt_period_us
;
8837 static int sched_rt_global_constraints(void)
8839 u64 runtime
, period
;
8842 if (sysctl_sched_rt_period
<= 0)
8845 runtime
= global_rt_runtime();
8846 period
= global_rt_period();
8849 * Sanity check on the sysctl variables.
8851 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8854 mutex_lock(&rt_constraints_mutex
);
8855 read_lock(&tasklist_lock
);
8856 ret
= __rt_schedulable(NULL
, 0, 0);
8857 read_unlock(&tasklist_lock
);
8858 mutex_unlock(&rt_constraints_mutex
);
8863 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8865 /* Don't accept realtime tasks when there is no way for them to run */
8866 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8872 #else /* !CONFIG_RT_GROUP_SCHED */
8873 static int sched_rt_global_constraints(void)
8875 unsigned long flags
;
8878 if (sysctl_sched_rt_period
<= 0)
8882 * There's always some RT tasks in the root group
8883 * -- migration, kstopmachine etc..
8885 if (sysctl_sched_rt_runtime
== 0)
8888 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8889 for_each_possible_cpu(i
) {
8890 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8892 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8893 rt_rq
->rt_runtime
= global_rt_runtime();
8894 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8896 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8900 #endif /* CONFIG_RT_GROUP_SCHED */
8902 int sched_rt_handler(struct ctl_table
*table
, int write
,
8903 void __user
*buffer
, size_t *lenp
,
8907 int old_period
, old_runtime
;
8908 static DEFINE_MUTEX(mutex
);
8911 old_period
= sysctl_sched_rt_period
;
8912 old_runtime
= sysctl_sched_rt_runtime
;
8914 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8916 if (!ret
&& write
) {
8917 ret
= sched_rt_global_constraints();
8919 sysctl_sched_rt_period
= old_period
;
8920 sysctl_sched_rt_runtime
= old_runtime
;
8922 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8923 def_rt_bandwidth
.rt_period
=
8924 ns_to_ktime(global_rt_period());
8927 mutex_unlock(&mutex
);
8932 #ifdef CONFIG_CGROUP_SCHED
8934 /* return corresponding task_group object of a cgroup */
8935 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8937 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8938 struct task_group
, css
);
8941 static struct cgroup_subsys_state
*
8942 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8944 struct task_group
*tg
, *parent
;
8946 if (!cgrp
->parent
) {
8947 /* This is early initialization for the top cgroup */
8948 return &root_task_group
.css
;
8951 parent
= cgroup_tg(cgrp
->parent
);
8952 tg
= sched_create_group(parent
);
8954 return ERR_PTR(-ENOMEM
);
8960 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8962 struct task_group
*tg
= cgroup_tg(cgrp
);
8964 sched_destroy_group(tg
);
8968 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8970 #ifdef CONFIG_RT_GROUP_SCHED
8971 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8974 /* We don't support RT-tasks being in separate groups */
8975 if (tsk
->sched_class
!= &fair_sched_class
)
8982 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8983 struct task_struct
*tsk
, bool threadgroup
)
8985 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8989 struct task_struct
*c
;
8991 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8992 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
9004 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9005 struct cgroup
*old_cont
, struct task_struct
*tsk
,
9008 sched_move_task(tsk
);
9010 struct task_struct
*c
;
9012 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
9020 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct task_struct
*task
)
9023 * cgroup_exit() is called in the copy_process() failure path.
9024 * Ignore this case since the task hasn't ran yet, this avoids
9025 * trying to poke a half freed task state from generic code.
9027 if (!(task
->flags
& PF_EXITING
))
9030 sched_move_task(task
);
9033 #ifdef CONFIG_FAIR_GROUP_SCHED
9034 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9037 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9040 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9042 struct task_group
*tg
= cgroup_tg(cgrp
);
9044 return (u64
) tg
->shares
;
9046 #endif /* CONFIG_FAIR_GROUP_SCHED */
9048 #ifdef CONFIG_RT_GROUP_SCHED
9049 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9052 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9055 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9057 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9060 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9063 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9066 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9068 return sched_group_rt_period(cgroup_tg(cgrp
));
9070 #endif /* CONFIG_RT_GROUP_SCHED */
9072 static struct cftype cpu_files
[] = {
9073 #ifdef CONFIG_FAIR_GROUP_SCHED
9076 .read_u64
= cpu_shares_read_u64
,
9077 .write_u64
= cpu_shares_write_u64
,
9080 #ifdef CONFIG_RT_GROUP_SCHED
9082 .name
= "rt_runtime_us",
9083 .read_s64
= cpu_rt_runtime_read
,
9084 .write_s64
= cpu_rt_runtime_write
,
9087 .name
= "rt_period_us",
9088 .read_u64
= cpu_rt_period_read_uint
,
9089 .write_u64
= cpu_rt_period_write_uint
,
9094 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9096 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9099 struct cgroup_subsys cpu_cgroup_subsys
= {
9101 .create
= cpu_cgroup_create
,
9102 .destroy
= cpu_cgroup_destroy
,
9103 .can_attach
= cpu_cgroup_can_attach
,
9104 .attach
= cpu_cgroup_attach
,
9105 .exit
= cpu_cgroup_exit
,
9106 .populate
= cpu_cgroup_populate
,
9107 .subsys_id
= cpu_cgroup_subsys_id
,
9111 #endif /* CONFIG_CGROUP_SCHED */
9113 #ifdef CONFIG_CGROUP_CPUACCT
9116 * CPU accounting code for task groups.
9118 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9119 * (balbir@in.ibm.com).
9122 /* track cpu usage of a group of tasks and its child groups */
9124 struct cgroup_subsys_state css
;
9125 /* cpuusage holds pointer to a u64-type object on every cpu */
9126 u64 __percpu
*cpuusage
;
9127 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9128 struct cpuacct
*parent
;
9131 struct cgroup_subsys cpuacct_subsys
;
9133 /* return cpu accounting group corresponding to this container */
9134 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9136 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9137 struct cpuacct
, css
);
9140 /* return cpu accounting group to which this task belongs */
9141 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9143 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9144 struct cpuacct
, css
);
9147 /* create a new cpu accounting group */
9148 static struct cgroup_subsys_state
*cpuacct_create(
9149 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9151 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9157 ca
->cpuusage
= alloc_percpu(u64
);
9161 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9162 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9163 goto out_free_counters
;
9166 ca
->parent
= cgroup_ca(cgrp
->parent
);
9172 percpu_counter_destroy(&ca
->cpustat
[i
]);
9173 free_percpu(ca
->cpuusage
);
9177 return ERR_PTR(-ENOMEM
);
9180 /* destroy an existing cpu accounting group */
9182 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9184 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9187 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9188 percpu_counter_destroy(&ca
->cpustat
[i
]);
9189 free_percpu(ca
->cpuusage
);
9193 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9195 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9198 #ifndef CONFIG_64BIT
9200 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9202 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9204 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9212 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9214 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9216 #ifndef CONFIG_64BIT
9218 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9220 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9222 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9228 /* return total cpu usage (in nanoseconds) of a group */
9229 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9231 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9232 u64 totalcpuusage
= 0;
9235 for_each_present_cpu(i
)
9236 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9238 return totalcpuusage
;
9241 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9244 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9253 for_each_present_cpu(i
)
9254 cpuacct_cpuusage_write(ca
, i
, 0);
9260 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9263 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9267 for_each_present_cpu(i
) {
9268 percpu
= cpuacct_cpuusage_read(ca
, i
);
9269 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9271 seq_printf(m
, "\n");
9275 static const char *cpuacct_stat_desc
[] = {
9276 [CPUACCT_STAT_USER
] = "user",
9277 [CPUACCT_STAT_SYSTEM
] = "system",
9280 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9281 struct cgroup_map_cb
*cb
)
9283 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9286 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9287 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9288 val
= cputime64_to_clock_t(val
);
9289 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9294 static struct cftype files
[] = {
9297 .read_u64
= cpuusage_read
,
9298 .write_u64
= cpuusage_write
,
9301 .name
= "usage_percpu",
9302 .read_seq_string
= cpuacct_percpu_seq_read
,
9306 .read_map
= cpuacct_stats_show
,
9310 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9312 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9316 * charge this task's execution time to its accounting group.
9318 * called with rq->lock held.
9320 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9325 if (unlikely(!cpuacct_subsys
.active
))
9328 cpu
= task_cpu(tsk
);
9334 for (; ca
; ca
= ca
->parent
) {
9335 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9336 *cpuusage
+= cputime
;
9343 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9344 * in cputime_t units. As a result, cpuacct_update_stats calls
9345 * percpu_counter_add with values large enough to always overflow the
9346 * per cpu batch limit causing bad SMP scalability.
9348 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9349 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9350 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9353 #define CPUACCT_BATCH \
9354 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9356 #define CPUACCT_BATCH 0
9360 * Charge the system/user time to the task's accounting group.
9362 static void cpuacct_update_stats(struct task_struct
*tsk
,
9363 enum cpuacct_stat_index idx
, cputime_t val
)
9366 int batch
= CPUACCT_BATCH
;
9368 if (unlikely(!cpuacct_subsys
.active
))
9375 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9381 struct cgroup_subsys cpuacct_subsys
= {
9383 .create
= cpuacct_create
,
9384 .destroy
= cpuacct_destroy
,
9385 .populate
= cpuacct_populate
,
9386 .subsys_id
= cpuacct_subsys_id
,
9388 #endif /* CONFIG_CGROUP_CPUACCT */