4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <asm/mmu_context.h>
36 #include <linux/interrupt.h>
37 #include <linux/capability.h>
38 #include <linux/completion.h>
39 #include <linux/kernel_stat.h>
40 #include <linux/debug_locks.h>
41 #include <linux/perf_event.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/proc_fs.h>
58 #include <linux/seq_file.h>
59 #include <linux/stop_machine.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/debugfs.h>
71 #include <linux/ctype.h>
72 #include <linux/ftrace.h>
73 #include <linux/slab.h>
76 #include <asm/irq_regs.h>
77 #include <asm/mutex.h>
79 #include "sched_cpupri.h"
80 #include "workqueue_sched.h"
81 #include "sched_autogroup.h"
83 #define CREATE_TRACE_POINTS
84 #include <trace/events/sched.h>
87 * Convert user-nice values [ -20 ... 0 ... 19 ]
88 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
91 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
92 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
93 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
96 * 'User priority' is the nice value converted to something we
97 * can work with better when scaling various scheduler parameters,
98 * it's a [ 0 ... 39 ] range.
100 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
101 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
102 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
105 * Helpers for converting nanosecond timing to jiffy resolution
107 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
109 #define NICE_0_LOAD SCHED_LOAD_SCALE
110 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
113 * These are the 'tuning knobs' of the scheduler:
115 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
116 * Timeslices get refilled after they expire.
118 #define DEF_TIMESLICE (100 * HZ / 1000)
121 * single value that denotes runtime == period, ie unlimited time.
123 #define RUNTIME_INF ((u64)~0ULL)
125 static inline int rt_policy(int policy
)
127 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
132 static inline int task_has_rt_policy(struct task_struct
*p
)
134 return rt_policy(p
->policy
);
138 * This is the priority-queue data structure of the RT scheduling class:
140 struct rt_prio_array
{
141 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
142 struct list_head queue
[MAX_RT_PRIO
];
145 struct rt_bandwidth
{
146 /* nests inside the rq lock: */
147 raw_spinlock_t rt_runtime_lock
;
150 struct hrtimer rt_period_timer
;
153 static struct rt_bandwidth def_rt_bandwidth
;
155 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
157 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
159 struct rt_bandwidth
*rt_b
=
160 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
166 now
= hrtimer_cb_get_time(timer
);
167 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
172 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
175 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
179 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
181 rt_b
->rt_period
= ns_to_ktime(period
);
182 rt_b
->rt_runtime
= runtime
;
184 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
186 hrtimer_init(&rt_b
->rt_period_timer
,
187 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
188 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
191 static inline int rt_bandwidth_enabled(void)
193 return sysctl_sched_rt_runtime
>= 0;
196 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
200 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
203 if (hrtimer_active(&rt_b
->rt_period_timer
))
206 raw_spin_lock(&rt_b
->rt_runtime_lock
);
211 if (hrtimer_active(&rt_b
->rt_period_timer
))
214 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
215 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
217 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
218 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
219 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
220 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
221 HRTIMER_MODE_ABS_PINNED
, 0);
223 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
226 #ifdef CONFIG_RT_GROUP_SCHED
227 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
229 hrtimer_cancel(&rt_b
->rt_period_timer
);
234 * sched_domains_mutex serializes calls to arch_init_sched_domains,
235 * detach_destroy_domains and partition_sched_domains.
237 static DEFINE_MUTEX(sched_domains_mutex
);
239 #ifdef CONFIG_CGROUP_SCHED
241 #include <linux/cgroup.h>
245 static LIST_HEAD(task_groups
);
247 /* task group related information */
249 struct cgroup_subsys_state css
;
251 #ifdef CONFIG_FAIR_GROUP_SCHED
252 /* schedulable entities of this group on each cpu */
253 struct sched_entity
**se
;
254 /* runqueue "owned" by this group on each cpu */
255 struct cfs_rq
**cfs_rq
;
256 unsigned long shares
;
258 atomic_t load_weight
;
261 #ifdef CONFIG_RT_GROUP_SCHED
262 struct sched_rt_entity
**rt_se
;
263 struct rt_rq
**rt_rq
;
265 struct rt_bandwidth rt_bandwidth
;
269 struct list_head list
;
271 struct task_group
*parent
;
272 struct list_head siblings
;
273 struct list_head children
;
275 #ifdef CONFIG_SCHED_AUTOGROUP
276 struct autogroup
*autogroup
;
280 /* task_group_lock serializes the addition/removal of task groups */
281 static DEFINE_SPINLOCK(task_group_lock
);
283 #ifdef CONFIG_FAIR_GROUP_SCHED
285 # define ROOT_TASK_GROUP_LOAD NICE_0_LOAD
288 * A weight of 0 or 1 can cause arithmetics problems.
289 * A weight of a cfs_rq is the sum of weights of which entities
290 * are queued on this cfs_rq, so a weight of a entity should not be
291 * too large, so as the shares value of a task group.
292 * (The default weight is 1024 - so there's no practical
293 * limitation from this.)
296 #define MAX_SHARES (1UL << 18)
298 static int root_task_group_load
= ROOT_TASK_GROUP_LOAD
;
301 /* Default task group.
302 * Every task in system belong to this group at bootup.
304 struct task_group root_task_group
;
306 #endif /* CONFIG_CGROUP_SCHED */
308 /* CFS-related fields in a runqueue */
310 struct load_weight load
;
311 unsigned long nr_running
;
316 u64 min_vruntime_copy
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
322 struct list_head tasks
;
323 struct list_head
*balance_iterator
;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity
*curr
, *next
, *last
, *skip
;
331 unsigned int nr_spread_over
;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
345 struct list_head leaf_cfs_rq_list
;
346 struct task_group
*tg
; /* group that "owns" this runqueue */
350 * the part of load.weight contributed by tasks
352 unsigned long task_weight
;
355 * h_load = weight * f(tg)
357 * Where f(tg) is the recursive weight fraction assigned to
360 unsigned long h_load
;
363 * Maintaining per-cpu shares distribution for group scheduling
365 * load_stamp is the last time we updated the load average
366 * load_last is the last time we updated the load average and saw load
367 * load_unacc_exec_time is currently unaccounted execution time
371 u64 load_stamp
, load_last
, load_unacc_exec_time
;
373 unsigned long load_contribution
;
378 /* Real-Time classes' related field in a runqueue: */
380 struct rt_prio_array active
;
381 unsigned long rt_nr_running
;
382 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
384 int curr
; /* highest queued rt task prio */
386 int next
; /* next highest */
391 unsigned long rt_nr_migratory
;
392 unsigned long rt_nr_total
;
394 struct plist_head pushable_tasks
;
399 /* Nests inside the rq lock: */
400 raw_spinlock_t rt_runtime_lock
;
402 #ifdef CONFIG_RT_GROUP_SCHED
403 unsigned long rt_nr_boosted
;
406 struct list_head leaf_rt_rq_list
;
407 struct task_group
*tg
;
414 * We add the notion of a root-domain which will be used to define per-domain
415 * variables. Each exclusive cpuset essentially defines an island domain by
416 * fully partitioning the member cpus from any other cpuset. Whenever a new
417 * exclusive cpuset is created, we also create and attach a new root-domain
424 cpumask_var_t online
;
427 * The "RT overload" flag: it gets set if a CPU has more than
428 * one runnable RT task.
430 cpumask_var_t rto_mask
;
432 struct cpupri cpupri
;
436 * By default the system creates a single root-domain with all cpus as
437 * members (mimicking the global state we have today).
439 static struct root_domain def_root_domain
;
441 #endif /* CONFIG_SMP */
444 * This is the main, per-CPU runqueue data structure.
446 * Locking rule: those places that want to lock multiple runqueues
447 * (such as the load balancing or the thread migration code), lock
448 * acquire operations must be ordered by ascending &runqueue.
455 * nr_running and cpu_load should be in the same cacheline because
456 * remote CPUs use both these fields when doing load calculation.
458 unsigned long nr_running
;
459 #define CPU_LOAD_IDX_MAX 5
460 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
461 unsigned long last_load_update_tick
;
464 unsigned char nohz_balance_kick
;
466 unsigned int skip_clock_update
;
468 /* capture load from *all* tasks on this cpu: */
469 struct load_weight load
;
470 unsigned long nr_load_updates
;
476 #ifdef CONFIG_FAIR_GROUP_SCHED
477 /* list of leaf cfs_rq on this cpu: */
478 struct list_head leaf_cfs_rq_list
;
480 #ifdef CONFIG_RT_GROUP_SCHED
481 struct list_head leaf_rt_rq_list
;
485 * This is part of a global counter where only the total sum
486 * over all CPUs matters. A task can increase this counter on
487 * one CPU and if it got migrated afterwards it may decrease
488 * it on another CPU. Always updated under the runqueue lock:
490 unsigned long nr_uninterruptible
;
492 struct task_struct
*curr
, *idle
, *stop
;
493 unsigned long next_balance
;
494 struct mm_struct
*prev_mm
;
502 struct root_domain
*rd
;
503 struct sched_domain
*sd
;
505 unsigned long cpu_power
;
507 unsigned char idle_at_tick
;
508 /* For active balancing */
512 struct cpu_stop_work active_balance_work
;
513 /* cpu of this runqueue: */
517 unsigned long avg_load_per_task
;
525 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
529 /* calc_load related fields */
530 unsigned long calc_load_update
;
531 long calc_load_active
;
533 #ifdef CONFIG_SCHED_HRTICK
535 int hrtick_csd_pending
;
536 struct call_single_data hrtick_csd
;
538 struct hrtimer hrtick_timer
;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info
;
544 unsigned long long rq_cpu_time
;
545 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
547 /* sys_sched_yield() stats */
548 unsigned int yld_count
;
550 /* schedule() stats */
551 unsigned int sched_switch
;
552 unsigned int sched_count
;
553 unsigned int sched_goidle
;
555 /* try_to_wake_up() stats */
556 unsigned int ttwu_count
;
557 unsigned int ttwu_local
;
561 struct task_struct
*wake_list
;
565 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
568 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
);
570 static inline int cpu_of(struct rq
*rq
)
579 #define rcu_dereference_check_sched_domain(p) \
580 rcu_dereference_check((p), \
581 rcu_read_lock_sched_held() || \
582 lockdep_is_held(&sched_domains_mutex))
585 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
586 * See detach_destroy_domains: synchronize_sched for details.
588 * The domain tree of any CPU may only be accessed from within
589 * preempt-disabled sections.
591 #define for_each_domain(cpu, __sd) \
592 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
594 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
595 #define this_rq() (&__get_cpu_var(runqueues))
596 #define task_rq(p) cpu_rq(task_cpu(p))
597 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
598 #define raw_rq() (&__raw_get_cpu_var(runqueues))
600 #ifdef CONFIG_CGROUP_SCHED
603 * Return the group to which this tasks belongs.
605 * We use task_subsys_state_check() and extend the RCU verification
606 * with lockdep_is_held(&p->pi_lock) because cpu_cgroup_attach()
607 * holds that lock for each task it moves into the cgroup. Therefore
608 * by holding that lock, we pin the task to the current cgroup.
610 static inline struct task_group
*task_group(struct task_struct
*p
)
612 struct task_group
*tg
;
613 struct cgroup_subsys_state
*css
;
615 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
616 lockdep_is_held(&p
->pi_lock
));
617 tg
= container_of(css
, struct task_group
, css
);
619 return autogroup_task_group(p
, tg
);
622 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
623 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
625 #ifdef CONFIG_FAIR_GROUP_SCHED
626 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
627 p
->se
.parent
= task_group(p
)->se
[cpu
];
630 #ifdef CONFIG_RT_GROUP_SCHED
631 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
632 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
636 #else /* CONFIG_CGROUP_SCHED */
638 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
639 static inline struct task_group
*task_group(struct task_struct
*p
)
644 #endif /* CONFIG_CGROUP_SCHED */
646 static void update_rq_clock_task(struct rq
*rq
, s64 delta
);
648 static void update_rq_clock(struct rq
*rq
)
652 if (rq
->skip_clock_update
)
655 delta
= sched_clock_cpu(cpu_of(rq
)) - rq
->clock
;
657 update_rq_clock_task(rq
, delta
);
661 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
663 #ifdef CONFIG_SCHED_DEBUG
664 # define const_debug __read_mostly
666 # define const_debug static const
670 * runqueue_is_locked - Returns true if the current cpu runqueue is locked
671 * @cpu: the processor in question.
673 * This interface allows printk to be called with the runqueue lock
674 * held and know whether or not it is OK to wake up the klogd.
676 int runqueue_is_locked(int cpu
)
678 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
682 * Debugging: various feature bits
685 #define SCHED_FEAT(name, enabled) \
686 __SCHED_FEAT_##name ,
689 #include "sched_features.h"
694 #define SCHED_FEAT(name, enabled) \
695 (1UL << __SCHED_FEAT_##name) * enabled |
697 const_debug
unsigned int sysctl_sched_features
=
698 #include "sched_features.h"
703 #ifdef CONFIG_SCHED_DEBUG
704 #define SCHED_FEAT(name, enabled) \
707 static __read_mostly
char *sched_feat_names
[] = {
708 #include "sched_features.h"
714 static int sched_feat_show(struct seq_file
*m
, void *v
)
718 for (i
= 0; sched_feat_names
[i
]; i
++) {
719 if (!(sysctl_sched_features
& (1UL << i
)))
721 seq_printf(m
, "%s ", sched_feat_names
[i
]);
729 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
730 size_t cnt
, loff_t
*ppos
)
740 if (copy_from_user(&buf
, ubuf
, cnt
))
746 if (strncmp(cmp
, "NO_", 3) == 0) {
751 for (i
= 0; sched_feat_names
[i
]; i
++) {
752 if (strcmp(cmp
, sched_feat_names
[i
]) == 0) {
754 sysctl_sched_features
&= ~(1UL << i
);
756 sysctl_sched_features
|= (1UL << i
);
761 if (!sched_feat_names
[i
])
769 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
771 return single_open(filp
, sched_feat_show
, NULL
);
774 static const struct file_operations sched_feat_fops
= {
775 .open
= sched_feat_open
,
776 .write
= sched_feat_write
,
779 .release
= single_release
,
782 static __init
int sched_init_debug(void)
784 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
789 late_initcall(sched_init_debug
);
793 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
796 * Number of tasks to iterate in a single balance run.
797 * Limited because this is done with IRQs disabled.
799 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
802 * period over which we average the RT time consumption, measured
807 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
810 * period over which we measure -rt task cpu usage in us.
813 unsigned int sysctl_sched_rt_period
= 1000000;
815 static __read_mostly
int scheduler_running
;
818 * part of the period that we allow rt tasks to run in us.
821 int sysctl_sched_rt_runtime
= 950000;
823 static inline u64
global_rt_period(void)
825 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
828 static inline u64
global_rt_runtime(void)
830 if (sysctl_sched_rt_runtime
< 0)
833 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
836 #ifndef prepare_arch_switch
837 # define prepare_arch_switch(next) do { } while (0)
839 #ifndef finish_arch_switch
840 # define finish_arch_switch(prev) do { } while (0)
843 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
845 return rq
->curr
== p
;
848 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
853 return task_current(rq
, p
);
857 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
858 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
862 * We can optimise this out completely for !SMP, because the
863 * SMP rebalancing from interrupt is the only thing that cares
870 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
874 * After ->on_cpu is cleared, the task can be moved to a different CPU.
875 * We must ensure this doesn't happen until the switch is completely
881 #ifdef CONFIG_DEBUG_SPINLOCK
882 /* this is a valid case when another task releases the spinlock */
883 rq
->lock
.owner
= current
;
886 * If we are tracking spinlock dependencies then we have to
887 * fix up the runqueue lock - which gets 'carried over' from
890 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
892 raw_spin_unlock_irq(&rq
->lock
);
895 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
896 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
900 * We can optimise this out completely for !SMP, because the
901 * SMP rebalancing from interrupt is the only thing that cares
906 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
907 raw_spin_unlock_irq(&rq
->lock
);
909 raw_spin_unlock(&rq
->lock
);
913 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
917 * After ->on_cpu is cleared, the task can be moved to a different CPU.
918 * We must ensure this doesn't happen until the switch is completely
924 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
928 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
931 * __task_rq_lock - lock the rq @p resides on.
933 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
938 lockdep_assert_held(&p
->pi_lock
);
942 raw_spin_lock(&rq
->lock
);
943 if (likely(rq
== task_rq(p
)))
945 raw_spin_unlock(&rq
->lock
);
950 * task_rq_lock - lock p->pi_lock and lock the rq @p resides on.
952 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
953 __acquires(p
->pi_lock
)
959 raw_spin_lock_irqsave(&p
->pi_lock
, *flags
);
961 raw_spin_lock(&rq
->lock
);
962 if (likely(rq
== task_rq(p
)))
964 raw_spin_unlock(&rq
->lock
);
965 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
969 static void __task_rq_unlock(struct rq
*rq
)
972 raw_spin_unlock(&rq
->lock
);
976 task_rq_unlock(struct rq
*rq
, struct task_struct
*p
, unsigned long *flags
)
978 __releases(p
->pi_lock
)
980 raw_spin_unlock(&rq
->lock
);
981 raw_spin_unlock_irqrestore(&p
->pi_lock
, *flags
);
985 * this_rq_lock - lock this runqueue and disable interrupts.
987 static struct rq
*this_rq_lock(void)
994 raw_spin_lock(&rq
->lock
);
999 #ifdef CONFIG_SCHED_HRTICK
1001 * Use HR-timers to deliver accurate preemption points.
1003 * Its all a bit involved since we cannot program an hrt while holding the
1004 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1007 * When we get rescheduled we reprogram the hrtick_timer outside of the
1013 * - enabled by features
1014 * - hrtimer is actually high res
1016 static inline int hrtick_enabled(struct rq
*rq
)
1018 if (!sched_feat(HRTICK
))
1020 if (!cpu_active(cpu_of(rq
)))
1022 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1025 static void hrtick_clear(struct rq
*rq
)
1027 if (hrtimer_active(&rq
->hrtick_timer
))
1028 hrtimer_cancel(&rq
->hrtick_timer
);
1032 * High-resolution timer tick.
1033 * Runs from hardirq context with interrupts disabled.
1035 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1037 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1039 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1041 raw_spin_lock(&rq
->lock
);
1042 update_rq_clock(rq
);
1043 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1044 raw_spin_unlock(&rq
->lock
);
1046 return HRTIMER_NORESTART
;
1051 * called from hardirq (IPI) context
1053 static void __hrtick_start(void *arg
)
1055 struct rq
*rq
= arg
;
1057 raw_spin_lock(&rq
->lock
);
1058 hrtimer_restart(&rq
->hrtick_timer
);
1059 rq
->hrtick_csd_pending
= 0;
1060 raw_spin_unlock(&rq
->lock
);
1064 * Called to set the hrtick timer state.
1066 * called with rq->lock held and irqs disabled
1068 static void hrtick_start(struct rq
*rq
, u64 delay
)
1070 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1071 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1073 hrtimer_set_expires(timer
, time
);
1075 if (rq
== this_rq()) {
1076 hrtimer_restart(timer
);
1077 } else if (!rq
->hrtick_csd_pending
) {
1078 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1079 rq
->hrtick_csd_pending
= 1;
1084 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1086 int cpu
= (int)(long)hcpu
;
1089 case CPU_UP_CANCELED
:
1090 case CPU_UP_CANCELED_FROZEN
:
1091 case CPU_DOWN_PREPARE
:
1092 case CPU_DOWN_PREPARE_FROZEN
:
1094 case CPU_DEAD_FROZEN
:
1095 hrtick_clear(cpu_rq(cpu
));
1102 static __init
void init_hrtick(void)
1104 hotcpu_notifier(hotplug_hrtick
, 0);
1108 * Called to set the hrtick timer state.
1110 * called with rq->lock held and irqs disabled
1112 static void hrtick_start(struct rq
*rq
, u64 delay
)
1114 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1115 HRTIMER_MODE_REL_PINNED
, 0);
1118 static inline void init_hrtick(void)
1121 #endif /* CONFIG_SMP */
1123 static void init_rq_hrtick(struct rq
*rq
)
1126 rq
->hrtick_csd_pending
= 0;
1128 rq
->hrtick_csd
.flags
= 0;
1129 rq
->hrtick_csd
.func
= __hrtick_start
;
1130 rq
->hrtick_csd
.info
= rq
;
1133 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1134 rq
->hrtick_timer
.function
= hrtick
;
1136 #else /* CONFIG_SCHED_HRTICK */
1137 static inline void hrtick_clear(struct rq
*rq
)
1141 static inline void init_rq_hrtick(struct rq
*rq
)
1145 static inline void init_hrtick(void)
1148 #endif /* CONFIG_SCHED_HRTICK */
1151 * resched_task - mark a task 'to be rescheduled now'.
1153 * On UP this means the setting of the need_resched flag, on SMP it
1154 * might also involve a cross-CPU call to trigger the scheduler on
1159 #ifndef tsk_is_polling
1160 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1163 static void resched_task(struct task_struct
*p
)
1167 assert_raw_spin_locked(&task_rq(p
)->lock
);
1169 if (test_tsk_need_resched(p
))
1172 set_tsk_need_resched(p
);
1175 if (cpu
== smp_processor_id())
1178 /* NEED_RESCHED must be visible before we test polling */
1180 if (!tsk_is_polling(p
))
1181 smp_send_reschedule(cpu
);
1184 static void resched_cpu(int cpu
)
1186 struct rq
*rq
= cpu_rq(cpu
);
1187 unsigned long flags
;
1189 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1191 resched_task(cpu_curr(cpu
));
1192 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1197 * In the semi idle case, use the nearest busy cpu for migrating timers
1198 * from an idle cpu. This is good for power-savings.
1200 * We don't do similar optimization for completely idle system, as
1201 * selecting an idle cpu will add more delays to the timers than intended
1202 * (as that cpu's timer base may not be uptodate wrt jiffies etc).
1204 int get_nohz_timer_target(void)
1206 int cpu
= smp_processor_id();
1208 struct sched_domain
*sd
;
1210 for_each_domain(cpu
, sd
) {
1211 for_each_cpu(i
, sched_domain_span(sd
))
1218 * When add_timer_on() enqueues a timer into the timer wheel of an
1219 * idle CPU then this timer might expire before the next timer event
1220 * which is scheduled to wake up that CPU. In case of a completely
1221 * idle system the next event might even be infinite time into the
1222 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1223 * leaves the inner idle loop so the newly added timer is taken into
1224 * account when the CPU goes back to idle and evaluates the timer
1225 * wheel for the next timer event.
1227 void wake_up_idle_cpu(int cpu
)
1229 struct rq
*rq
= cpu_rq(cpu
);
1231 if (cpu
== smp_processor_id())
1235 * This is safe, as this function is called with the timer
1236 * wheel base lock of (cpu) held. When the CPU is on the way
1237 * to idle and has not yet set rq->curr to idle then it will
1238 * be serialized on the timer wheel base lock and take the new
1239 * timer into account automatically.
1241 if (rq
->curr
!= rq
->idle
)
1245 * We can set TIF_RESCHED on the idle task of the other CPU
1246 * lockless. The worst case is that the other CPU runs the
1247 * idle task through an additional NOOP schedule()
1249 set_tsk_need_resched(rq
->idle
);
1251 /* NEED_RESCHED must be visible before we test polling */
1253 if (!tsk_is_polling(rq
->idle
))
1254 smp_send_reschedule(cpu
);
1257 #endif /* CONFIG_NO_HZ */
1259 static u64
sched_avg_period(void)
1261 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1264 static void sched_avg_update(struct rq
*rq
)
1266 s64 period
= sched_avg_period();
1268 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1270 * Inline assembly required to prevent the compiler
1271 * optimising this loop into a divmod call.
1272 * See __iter_div_u64_rem() for another example of this.
1274 asm("" : "+rm" (rq
->age_stamp
));
1275 rq
->age_stamp
+= period
;
1280 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1282 rq
->rt_avg
+= rt_delta
;
1283 sched_avg_update(rq
);
1286 #else /* !CONFIG_SMP */
1287 static void resched_task(struct task_struct
*p
)
1289 assert_raw_spin_locked(&task_rq(p
)->lock
);
1290 set_tsk_need_resched(p
);
1293 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1297 static void sched_avg_update(struct rq
*rq
)
1300 #endif /* CONFIG_SMP */
1302 #if BITS_PER_LONG == 32
1303 # define WMULT_CONST (~0UL)
1305 # define WMULT_CONST (1UL << 32)
1308 #define WMULT_SHIFT 32
1311 * Shift right and round:
1313 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1316 * delta *= weight / lw
1318 static unsigned long
1319 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1320 struct load_weight
*lw
)
1324 if (!lw
->inv_weight
) {
1325 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1328 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1332 tmp
= (u64
)delta_exec
* weight
;
1334 * Check whether we'd overflow the 64-bit multiplication:
1336 if (unlikely(tmp
> WMULT_CONST
))
1337 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1340 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1342 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1345 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1351 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1357 static inline void update_load_set(struct load_weight
*lw
, unsigned long w
)
1364 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1365 * of tasks with abnormal "nice" values across CPUs the contribution that
1366 * each task makes to its run queue's load is weighted according to its
1367 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1368 * scaled version of the new time slice allocation that they receive on time
1372 #define WEIGHT_IDLEPRIO 3
1373 #define WMULT_IDLEPRIO 1431655765
1376 * Nice levels are multiplicative, with a gentle 10% change for every
1377 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1378 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1379 * that remained on nice 0.
1381 * The "10% effect" is relative and cumulative: from _any_ nice level,
1382 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1383 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1384 * If a task goes up by ~10% and another task goes down by ~10% then
1385 * the relative distance between them is ~25%.)
1387 static const int prio_to_weight
[40] = {
1388 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1389 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1390 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1391 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1392 /* 0 */ 1024, 820, 655, 526, 423,
1393 /* 5 */ 335, 272, 215, 172, 137,
1394 /* 10 */ 110, 87, 70, 56, 45,
1395 /* 15 */ 36, 29, 23, 18, 15,
1399 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1401 * In cases where the weight does not change often, we can use the
1402 * precalculated inverse to speed up arithmetics by turning divisions
1403 * into multiplications:
1405 static const u32 prio_to_wmult
[40] = {
1406 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1407 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1408 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1409 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1410 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1411 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1412 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1413 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1416 /* Time spent by the tasks of the cpu accounting group executing in ... */
1417 enum cpuacct_stat_index
{
1418 CPUACCT_STAT_USER
, /* ... user mode */
1419 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1421 CPUACCT_STAT_NSTATS
,
1424 #ifdef CONFIG_CGROUP_CPUACCT
1425 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1426 static void cpuacct_update_stats(struct task_struct
*tsk
,
1427 enum cpuacct_stat_index idx
, cputime_t val
);
1429 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1430 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1431 enum cpuacct_stat_index idx
, cputime_t val
) {}
1434 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1436 update_load_add(&rq
->load
, load
);
1439 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1441 update_load_sub(&rq
->load
, load
);
1444 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1445 typedef int (*tg_visitor
)(struct task_group
*, void *);
1448 * Iterate the full tree, calling @down when first entering a node and @up when
1449 * leaving it for the final time.
1451 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1453 struct task_group
*parent
, *child
;
1457 parent
= &root_task_group
;
1459 ret
= (*down
)(parent
, data
);
1462 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1469 ret
= (*up
)(parent
, data
);
1474 parent
= parent
->parent
;
1483 static int tg_nop(struct task_group
*tg
, void *data
)
1490 /* Used instead of source_load when we know the type == 0 */
1491 static unsigned long weighted_cpuload(const int cpu
)
1493 return cpu_rq(cpu
)->load
.weight
;
1497 * Return a low guess at the load of a migration-source cpu weighted
1498 * according to the scheduling class and "nice" value.
1500 * We want to under-estimate the load of migration sources, to
1501 * balance conservatively.
1503 static unsigned long source_load(int cpu
, int type
)
1505 struct rq
*rq
= cpu_rq(cpu
);
1506 unsigned long total
= weighted_cpuload(cpu
);
1508 if (type
== 0 || !sched_feat(LB_BIAS
))
1511 return min(rq
->cpu_load
[type
-1], total
);
1515 * Return a high guess at the load of a migration-target cpu weighted
1516 * according to the scheduling class and "nice" value.
1518 static unsigned long target_load(int cpu
, int type
)
1520 struct rq
*rq
= cpu_rq(cpu
);
1521 unsigned long total
= weighted_cpuload(cpu
);
1523 if (type
== 0 || !sched_feat(LB_BIAS
))
1526 return max(rq
->cpu_load
[type
-1], total
);
1529 static unsigned long power_of(int cpu
)
1531 return cpu_rq(cpu
)->cpu_power
;
1534 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1536 static unsigned long cpu_avg_load_per_task(int cpu
)
1538 struct rq
*rq
= cpu_rq(cpu
);
1539 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1542 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1544 rq
->avg_load_per_task
= 0;
1546 return rq
->avg_load_per_task
;
1549 #ifdef CONFIG_FAIR_GROUP_SCHED
1552 * Compute the cpu's hierarchical load factor for each task group.
1553 * This needs to be done in a top-down fashion because the load of a child
1554 * group is a fraction of its parents load.
1556 static int tg_load_down(struct task_group
*tg
, void *data
)
1559 long cpu
= (long)data
;
1562 load
= cpu_rq(cpu
)->load
.weight
;
1564 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1565 load
*= tg
->se
[cpu
]->load
.weight
;
1566 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1569 tg
->cfs_rq
[cpu
]->h_load
= load
;
1574 static void update_h_load(long cpu
)
1576 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1581 #ifdef CONFIG_PREEMPT
1583 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1586 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1587 * way at the expense of forcing extra atomic operations in all
1588 * invocations. This assures that the double_lock is acquired using the
1589 * same underlying policy as the spinlock_t on this architecture, which
1590 * reduces latency compared to the unfair variant below. However, it
1591 * also adds more overhead and therefore may reduce throughput.
1593 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1594 __releases(this_rq
->lock
)
1595 __acquires(busiest
->lock
)
1596 __acquires(this_rq
->lock
)
1598 raw_spin_unlock(&this_rq
->lock
);
1599 double_rq_lock(this_rq
, busiest
);
1606 * Unfair double_lock_balance: Optimizes throughput at the expense of
1607 * latency by eliminating extra atomic operations when the locks are
1608 * already in proper order on entry. This favors lower cpu-ids and will
1609 * grant the double lock to lower cpus over higher ids under contention,
1610 * regardless of entry order into the function.
1612 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1613 __releases(this_rq
->lock
)
1614 __acquires(busiest
->lock
)
1615 __acquires(this_rq
->lock
)
1619 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1620 if (busiest
< this_rq
) {
1621 raw_spin_unlock(&this_rq
->lock
);
1622 raw_spin_lock(&busiest
->lock
);
1623 raw_spin_lock_nested(&this_rq
->lock
,
1624 SINGLE_DEPTH_NESTING
);
1627 raw_spin_lock_nested(&busiest
->lock
,
1628 SINGLE_DEPTH_NESTING
);
1633 #endif /* CONFIG_PREEMPT */
1636 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1638 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1640 if (unlikely(!irqs_disabled())) {
1641 /* printk() doesn't work good under rq->lock */
1642 raw_spin_unlock(&this_rq
->lock
);
1646 return _double_lock_balance(this_rq
, busiest
);
1649 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1650 __releases(busiest
->lock
)
1652 raw_spin_unlock(&busiest
->lock
);
1653 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1657 * double_rq_lock - safely lock two runqueues
1659 * Note this does not disable interrupts like task_rq_lock,
1660 * you need to do so manually before calling.
1662 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1663 __acquires(rq1
->lock
)
1664 __acquires(rq2
->lock
)
1666 BUG_ON(!irqs_disabled());
1668 raw_spin_lock(&rq1
->lock
);
1669 __acquire(rq2
->lock
); /* Fake it out ;) */
1672 raw_spin_lock(&rq1
->lock
);
1673 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1675 raw_spin_lock(&rq2
->lock
);
1676 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1682 * double_rq_unlock - safely unlock two runqueues
1684 * Note this does not restore interrupts like task_rq_unlock,
1685 * you need to do so manually after calling.
1687 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1688 __releases(rq1
->lock
)
1689 __releases(rq2
->lock
)
1691 raw_spin_unlock(&rq1
->lock
);
1693 raw_spin_unlock(&rq2
->lock
);
1695 __release(rq2
->lock
);
1698 #else /* CONFIG_SMP */
1701 * double_rq_lock - safely lock two runqueues
1703 * Note this does not disable interrupts like task_rq_lock,
1704 * you need to do so manually before calling.
1706 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1707 __acquires(rq1
->lock
)
1708 __acquires(rq2
->lock
)
1710 BUG_ON(!irqs_disabled());
1712 raw_spin_lock(&rq1
->lock
);
1713 __acquire(rq2
->lock
); /* Fake it out ;) */
1717 * double_rq_unlock - safely unlock two runqueues
1719 * Note this does not restore interrupts like task_rq_unlock,
1720 * you need to do so manually after calling.
1722 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1723 __releases(rq1
->lock
)
1724 __releases(rq2
->lock
)
1727 raw_spin_unlock(&rq1
->lock
);
1728 __release(rq2
->lock
);
1733 static void calc_load_account_idle(struct rq
*this_rq
);
1734 static void update_sysctl(void);
1735 static int get_update_sysctl_factor(void);
1736 static void update_cpu_load(struct rq
*this_rq
);
1738 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1740 set_task_rq(p
, cpu
);
1743 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1744 * successfuly executed on another CPU. We must ensure that updates of
1745 * per-task data have been completed by this moment.
1748 task_thread_info(p
)->cpu
= cpu
;
1752 static const struct sched_class rt_sched_class
;
1754 #define sched_class_highest (&stop_sched_class)
1755 #define for_each_class(class) \
1756 for (class = sched_class_highest; class; class = class->next)
1758 #include "sched_stats.h"
1760 static void inc_nr_running(struct rq
*rq
)
1765 static void dec_nr_running(struct rq
*rq
)
1770 static void set_load_weight(struct task_struct
*p
)
1773 * SCHED_IDLE tasks get minimal weight:
1775 if (p
->policy
== SCHED_IDLE
) {
1776 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1777 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1781 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1782 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1785 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1787 update_rq_clock(rq
);
1788 sched_info_queued(p
);
1789 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1792 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1794 update_rq_clock(rq
);
1795 sched_info_dequeued(p
);
1796 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1800 * activate_task - move a task to the runqueue.
1802 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1804 if (task_contributes_to_load(p
))
1805 rq
->nr_uninterruptible
--;
1807 enqueue_task(rq
, p
, flags
);
1812 * deactivate_task - remove a task from the runqueue.
1814 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1816 if (task_contributes_to_load(p
))
1817 rq
->nr_uninterruptible
++;
1819 dequeue_task(rq
, p
, flags
);
1823 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
1826 * There are no locks covering percpu hardirq/softirq time.
1827 * They are only modified in account_system_vtime, on corresponding CPU
1828 * with interrupts disabled. So, writes are safe.
1829 * They are read and saved off onto struct rq in update_rq_clock().
1830 * This may result in other CPU reading this CPU's irq time and can
1831 * race with irq/account_system_vtime on this CPU. We would either get old
1832 * or new value with a side effect of accounting a slice of irq time to wrong
1833 * task when irq is in progress while we read rq->clock. That is a worthy
1834 * compromise in place of having locks on each irq in account_system_time.
1836 static DEFINE_PER_CPU(u64
, cpu_hardirq_time
);
1837 static DEFINE_PER_CPU(u64
, cpu_softirq_time
);
1839 static DEFINE_PER_CPU(u64
, irq_start_time
);
1840 static int sched_clock_irqtime
;
1842 void enable_sched_clock_irqtime(void)
1844 sched_clock_irqtime
= 1;
1847 void disable_sched_clock_irqtime(void)
1849 sched_clock_irqtime
= 0;
1852 #ifndef CONFIG_64BIT
1853 static DEFINE_PER_CPU(seqcount_t
, irq_time_seq
);
1855 static inline void irq_time_write_begin(void)
1857 __this_cpu_inc(irq_time_seq
.sequence
);
1861 static inline void irq_time_write_end(void)
1864 __this_cpu_inc(irq_time_seq
.sequence
);
1867 static inline u64
irq_time_read(int cpu
)
1873 seq
= read_seqcount_begin(&per_cpu(irq_time_seq
, cpu
));
1874 irq_time
= per_cpu(cpu_softirq_time
, cpu
) +
1875 per_cpu(cpu_hardirq_time
, cpu
);
1876 } while (read_seqcount_retry(&per_cpu(irq_time_seq
, cpu
), seq
));
1880 #else /* CONFIG_64BIT */
1881 static inline void irq_time_write_begin(void)
1885 static inline void irq_time_write_end(void)
1889 static inline u64
irq_time_read(int cpu
)
1891 return per_cpu(cpu_softirq_time
, cpu
) + per_cpu(cpu_hardirq_time
, cpu
);
1893 #endif /* CONFIG_64BIT */
1896 * Called before incrementing preempt_count on {soft,}irq_enter
1897 * and before decrementing preempt_count on {soft,}irq_exit.
1899 void account_system_vtime(struct task_struct
*curr
)
1901 unsigned long flags
;
1905 if (!sched_clock_irqtime
)
1908 local_irq_save(flags
);
1910 cpu
= smp_processor_id();
1911 delta
= sched_clock_cpu(cpu
) - __this_cpu_read(irq_start_time
);
1912 __this_cpu_add(irq_start_time
, delta
);
1914 irq_time_write_begin();
1916 * We do not account for softirq time from ksoftirqd here.
1917 * We want to continue accounting softirq time to ksoftirqd thread
1918 * in that case, so as not to confuse scheduler with a special task
1919 * that do not consume any time, but still wants to run.
1921 if (hardirq_count())
1922 __this_cpu_add(cpu_hardirq_time
, delta
);
1923 else if (in_serving_softirq() && curr
!= this_cpu_ksoftirqd())
1924 __this_cpu_add(cpu_softirq_time
, delta
);
1926 irq_time_write_end();
1927 local_irq_restore(flags
);
1929 EXPORT_SYMBOL_GPL(account_system_vtime
);
1931 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1935 irq_delta
= irq_time_read(cpu_of(rq
)) - rq
->prev_irq_time
;
1938 * Since irq_time is only updated on {soft,}irq_exit, we might run into
1939 * this case when a previous update_rq_clock() happened inside a
1940 * {soft,}irq region.
1942 * When this happens, we stop ->clock_task and only update the
1943 * prev_irq_time stamp to account for the part that fit, so that a next
1944 * update will consume the rest. This ensures ->clock_task is
1947 * It does however cause some slight miss-attribution of {soft,}irq
1948 * time, a more accurate solution would be to update the irq_time using
1949 * the current rq->clock timestamp, except that would require using
1952 if (irq_delta
> delta
)
1955 rq
->prev_irq_time
+= irq_delta
;
1957 rq
->clock_task
+= delta
;
1959 if (irq_delta
&& sched_feat(NONIRQ_POWER
))
1960 sched_rt_avg_update(rq
, irq_delta
);
1963 static int irqtime_account_hi_update(void)
1965 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
1966 unsigned long flags
;
1970 local_irq_save(flags
);
1971 latest_ns
= this_cpu_read(cpu_hardirq_time
);
1972 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->irq
))
1974 local_irq_restore(flags
);
1978 static int irqtime_account_si_update(void)
1980 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
1981 unsigned long flags
;
1985 local_irq_save(flags
);
1986 latest_ns
= this_cpu_read(cpu_softirq_time
);
1987 if (cputime64_gt(nsecs_to_cputime64(latest_ns
), cpustat
->softirq
))
1989 local_irq_restore(flags
);
1993 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
1995 #define sched_clock_irqtime (0)
1997 static void update_rq_clock_task(struct rq
*rq
, s64 delta
)
1999 rq
->clock_task
+= delta
;
2002 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
2004 #include "sched_idletask.c"
2005 #include "sched_fair.c"
2006 #include "sched_rt.c"
2007 #include "sched_autogroup.c"
2008 #include "sched_stoptask.c"
2009 #ifdef CONFIG_SCHED_DEBUG
2010 # include "sched_debug.c"
2013 void sched_set_stop_task(int cpu
, struct task_struct
*stop
)
2015 struct sched_param param
= { .sched_priority
= MAX_RT_PRIO
- 1 };
2016 struct task_struct
*old_stop
= cpu_rq(cpu
)->stop
;
2020 * Make it appear like a SCHED_FIFO task, its something
2021 * userspace knows about and won't get confused about.
2023 * Also, it will make PI more or less work without too
2024 * much confusion -- but then, stop work should not
2025 * rely on PI working anyway.
2027 sched_setscheduler_nocheck(stop
, SCHED_FIFO
, ¶m
);
2029 stop
->sched_class
= &stop_sched_class
;
2032 cpu_rq(cpu
)->stop
= stop
;
2036 * Reset it back to a normal scheduling class so that
2037 * it can die in pieces.
2039 old_stop
->sched_class
= &rt_sched_class
;
2044 * __normal_prio - return the priority that is based on the static prio
2046 static inline int __normal_prio(struct task_struct
*p
)
2048 return p
->static_prio
;
2052 * Calculate the expected normal priority: i.e. priority
2053 * without taking RT-inheritance into account. Might be
2054 * boosted by interactivity modifiers. Changes upon fork,
2055 * setprio syscalls, and whenever the interactivity
2056 * estimator recalculates.
2058 static inline int normal_prio(struct task_struct
*p
)
2062 if (task_has_rt_policy(p
))
2063 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2065 prio
= __normal_prio(p
);
2070 * Calculate the current priority, i.e. the priority
2071 * taken into account by the scheduler. This value might
2072 * be boosted by RT tasks, or might be boosted by
2073 * interactivity modifiers. Will be RT if the task got
2074 * RT-boosted. If not then it returns p->normal_prio.
2076 static int effective_prio(struct task_struct
*p
)
2078 p
->normal_prio
= normal_prio(p
);
2080 * If we are RT tasks or we were boosted to RT priority,
2081 * keep the priority unchanged. Otherwise, update priority
2082 * to the normal priority:
2084 if (!rt_prio(p
->prio
))
2085 return p
->normal_prio
;
2090 * task_curr - is this task currently executing on a CPU?
2091 * @p: the task in question.
2093 inline int task_curr(const struct task_struct
*p
)
2095 return cpu_curr(task_cpu(p
)) == p
;
2098 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2099 const struct sched_class
*prev_class
,
2102 if (prev_class
!= p
->sched_class
) {
2103 if (prev_class
->switched_from
)
2104 prev_class
->switched_from(rq
, p
);
2105 p
->sched_class
->switched_to(rq
, p
);
2106 } else if (oldprio
!= p
->prio
)
2107 p
->sched_class
->prio_changed(rq
, p
, oldprio
);
2110 static void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
2112 const struct sched_class
*class;
2114 if (p
->sched_class
== rq
->curr
->sched_class
) {
2115 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
2117 for_each_class(class) {
2118 if (class == rq
->curr
->sched_class
)
2120 if (class == p
->sched_class
) {
2121 resched_task(rq
->curr
);
2128 * A queue event has occurred, and we're going to schedule. In
2129 * this case, we can save a useless back to back clock update.
2131 if (rq
->curr
->on_rq
&& test_tsk_need_resched(rq
->curr
))
2132 rq
->skip_clock_update
= 1;
2137 * Is this task likely cache-hot:
2140 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2144 if (p
->sched_class
!= &fair_sched_class
)
2147 if (unlikely(p
->policy
== SCHED_IDLE
))
2151 * Buddy candidates are cache hot:
2153 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2154 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2155 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2158 if (sysctl_sched_migration_cost
== -1)
2160 if (sysctl_sched_migration_cost
== 0)
2163 delta
= now
- p
->se
.exec_start
;
2165 return delta
< (s64
)sysctl_sched_migration_cost
;
2168 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2170 #ifdef CONFIG_SCHED_DEBUG
2172 * We should never call set_task_cpu() on a blocked task,
2173 * ttwu() will sort out the placement.
2175 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2176 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2178 #ifdef CONFIG_LOCKDEP
2179 WARN_ON_ONCE(debug_locks
&& !(lockdep_is_held(&p
->pi_lock
) ||
2180 lockdep_is_held(&task_rq(p
)->lock
)));
2184 trace_sched_migrate_task(p
, new_cpu
);
2186 if (task_cpu(p
) != new_cpu
) {
2187 p
->se
.nr_migrations
++;
2188 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2191 __set_task_cpu(p
, new_cpu
);
2194 struct migration_arg
{
2195 struct task_struct
*task
;
2199 static int migration_cpu_stop(void *data
);
2202 * The task's runqueue lock must be held.
2203 * Returns true if you have to wait for migration thread.
2205 static bool need_migrate_task(struct task_struct
*p
)
2208 * If the task is not on a runqueue (and not running), then
2209 * the next wake-up will properly place the task.
2211 bool running
= p
->on_rq
|| p
->on_cpu
;
2212 smp_rmb(); /* finish_lock_switch() */
2217 * wait_task_inactive - wait for a thread to unschedule.
2219 * If @match_state is nonzero, it's the @p->state value just checked and
2220 * not expected to change. If it changes, i.e. @p might have woken up,
2221 * then return zero. When we succeed in waiting for @p to be off its CPU,
2222 * we return a positive number (its total switch count). If a second call
2223 * a short while later returns the same number, the caller can be sure that
2224 * @p has remained unscheduled the whole time.
2226 * The caller must ensure that the task *will* unschedule sometime soon,
2227 * else this function might spin for a *long* time. This function can't
2228 * be called with interrupts off, or it may introduce deadlock with
2229 * smp_call_function() if an IPI is sent by the same process we are
2230 * waiting to become inactive.
2232 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2234 unsigned long flags
;
2241 * We do the initial early heuristics without holding
2242 * any task-queue locks at all. We'll only try to get
2243 * the runqueue lock when things look like they will
2249 * If the task is actively running on another CPU
2250 * still, just relax and busy-wait without holding
2253 * NOTE! Since we don't hold any locks, it's not
2254 * even sure that "rq" stays as the right runqueue!
2255 * But we don't care, since "task_running()" will
2256 * return false if the runqueue has changed and p
2257 * is actually now running somewhere else!
2259 while (task_running(rq
, p
)) {
2260 if (match_state
&& unlikely(p
->state
!= match_state
))
2266 * Ok, time to look more closely! We need the rq
2267 * lock now, to be *sure*. If we're wrong, we'll
2268 * just go back and repeat.
2270 rq
= task_rq_lock(p
, &flags
);
2271 trace_sched_wait_task(p
);
2272 running
= task_running(rq
, p
);
2275 if (!match_state
|| p
->state
== match_state
)
2276 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2277 task_rq_unlock(rq
, p
, &flags
);
2280 * If it changed from the expected state, bail out now.
2282 if (unlikely(!ncsw
))
2286 * Was it really running after all now that we
2287 * checked with the proper locks actually held?
2289 * Oops. Go back and try again..
2291 if (unlikely(running
)) {
2297 * It's not enough that it's not actively running,
2298 * it must be off the runqueue _entirely_, and not
2301 * So if it was still runnable (but just not actively
2302 * running right now), it's preempted, and we should
2303 * yield - it could be a while.
2305 if (unlikely(on_rq
)) {
2306 ktime_t to
= ktime_set(0, NSEC_PER_SEC
/HZ
);
2308 set_current_state(TASK_UNINTERRUPTIBLE
);
2309 schedule_hrtimeout(&to
, HRTIMER_MODE_REL
);
2314 * Ahh, all good. It wasn't running, and it wasn't
2315 * runnable, which means that it will never become
2316 * running in the future either. We're all done!
2325 * kick_process - kick a running thread to enter/exit the kernel
2326 * @p: the to-be-kicked thread
2328 * Cause a process which is running on another CPU to enter
2329 * kernel-mode, without any delay. (to get signals handled.)
2331 * NOTE: this function doesn't have to take the runqueue lock,
2332 * because all it wants to ensure is that the remote task enters
2333 * the kernel. If the IPI races and the task has been migrated
2334 * to another CPU then no harm is done and the purpose has been
2337 void kick_process(struct task_struct
*p
)
2343 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2344 smp_send_reschedule(cpu
);
2347 EXPORT_SYMBOL_GPL(kick_process
);
2348 #endif /* CONFIG_SMP */
2352 * ->cpus_allowed is protected by both rq->lock and p->pi_lock
2354 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2357 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2359 /* Look for allowed, online CPU in same node. */
2360 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2361 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2364 /* Any allowed, online CPU? */
2365 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2366 if (dest_cpu
< nr_cpu_ids
)
2369 /* No more Mr. Nice Guy. */
2370 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2372 * Don't tell them about moving exiting tasks or
2373 * kernel threads (both mm NULL), since they never
2376 if (p
->mm
&& printk_ratelimit()) {
2377 printk(KERN_INFO
"process %d (%s) no longer affine to cpu%d\n",
2378 task_pid_nr(p
), p
->comm
, cpu
);
2385 * The caller (fork, wakeup) owns p->pi_lock, ->cpus_allowed is stable.
2388 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2390 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2393 * In order not to call set_task_cpu() on a blocking task we need
2394 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2397 * Since this is common to all placement strategies, this lives here.
2399 * [ this allows ->select_task() to simply return task_cpu(p) and
2400 * not worry about this generic constraint ]
2402 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2404 cpu
= select_fallback_rq(task_cpu(p
), p
);
2409 static void update_avg(u64
*avg
, u64 sample
)
2411 s64 diff
= sample
- *avg
;
2417 ttwu_stat(struct task_struct
*p
, int cpu
, int wake_flags
)
2419 #ifdef CONFIG_SCHEDSTATS
2420 struct rq
*rq
= this_rq();
2423 int this_cpu
= smp_processor_id();
2425 if (cpu
== this_cpu
) {
2426 schedstat_inc(rq
, ttwu_local
);
2427 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2429 struct sched_domain
*sd
;
2431 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2432 for_each_domain(this_cpu
, sd
) {
2433 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2434 schedstat_inc(sd
, ttwu_wake_remote
);
2439 #endif /* CONFIG_SMP */
2441 schedstat_inc(rq
, ttwu_count
);
2442 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2444 if (wake_flags
& WF_SYNC
)
2445 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2447 if (cpu
!= task_cpu(p
))
2448 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2450 #endif /* CONFIG_SCHEDSTATS */
2453 static void ttwu_activate(struct rq
*rq
, struct task_struct
*p
, int en_flags
)
2455 activate_task(rq
, p
, en_flags
);
2458 /* if a worker is waking up, notify workqueue */
2459 if (p
->flags
& PF_WQ_WORKER
)
2460 wq_worker_waking_up(p
, cpu_of(rq
));
2464 * Mark the task runnable and perform wakeup-preemption.
2467 ttwu_do_wakeup(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2469 trace_sched_wakeup(p
, true);
2470 check_preempt_curr(rq
, p
, wake_flags
);
2472 p
->state
= TASK_RUNNING
;
2474 if (p
->sched_class
->task_woken
)
2475 p
->sched_class
->task_woken(rq
, p
);
2477 if (unlikely(rq
->idle_stamp
)) {
2478 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2479 u64 max
= 2*sysctl_sched_migration_cost
;
2484 update_avg(&rq
->avg_idle
, delta
);
2491 ttwu_do_activate(struct rq
*rq
, struct task_struct
*p
, int wake_flags
)
2494 if (p
->sched_contributes_to_load
)
2495 rq
->nr_uninterruptible
--;
2498 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
| ENQUEUE_WAKING
);
2499 ttwu_do_wakeup(rq
, p
, wake_flags
);
2503 * Called in case the task @p isn't fully descheduled from its runqueue,
2504 * in this case we must do a remote wakeup. Its a 'light' wakeup though,
2505 * since all we need to do is flip p->state to TASK_RUNNING, since
2506 * the task is still ->on_rq.
2508 static int ttwu_remote(struct task_struct
*p
, int wake_flags
)
2513 rq
= __task_rq_lock(p
);
2515 ttwu_do_wakeup(rq
, p
, wake_flags
);
2518 __task_rq_unlock(rq
);
2524 static void sched_ttwu_pending(void)
2526 struct rq
*rq
= this_rq();
2527 struct task_struct
*list
= xchg(&rq
->wake_list
, NULL
);
2532 raw_spin_lock(&rq
->lock
);
2535 struct task_struct
*p
= list
;
2536 list
= list
->wake_entry
;
2537 ttwu_do_activate(rq
, p
, 0);
2540 raw_spin_unlock(&rq
->lock
);
2543 void scheduler_ipi(void)
2545 sched_ttwu_pending();
2548 static void ttwu_queue_remote(struct task_struct
*p
, int cpu
)
2550 struct rq
*rq
= cpu_rq(cpu
);
2551 struct task_struct
*next
= rq
->wake_list
;
2554 struct task_struct
*old
= next
;
2556 p
->wake_entry
= next
;
2557 next
= cmpxchg(&rq
->wake_list
, old
, p
);
2563 smp_send_reschedule(cpu
);
2567 static void ttwu_queue(struct task_struct
*p
, int cpu
)
2569 struct rq
*rq
= cpu_rq(cpu
);
2571 #if defined(CONFIG_SMP) && defined(CONFIG_SCHED_TTWU_QUEUE)
2572 if (sched_feat(TTWU_QUEUE
) && cpu
!= smp_processor_id()) {
2573 ttwu_queue_remote(p
, cpu
);
2578 raw_spin_lock(&rq
->lock
);
2579 ttwu_do_activate(rq
, p
, 0);
2580 raw_spin_unlock(&rq
->lock
);
2584 * try_to_wake_up - wake up a thread
2585 * @p: the thread to be awakened
2586 * @state: the mask of task states that can be woken
2587 * @wake_flags: wake modifier flags (WF_*)
2589 * Put it on the run-queue if it's not already there. The "current"
2590 * thread is always on the run-queue (except when the actual
2591 * re-schedule is in progress), and as such you're allowed to do
2592 * the simpler "current->state = TASK_RUNNING" to mark yourself
2593 * runnable without the overhead of this.
2595 * Returns %true if @p was woken up, %false if it was already running
2596 * or @state didn't match @p's state.
2599 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int wake_flags
)
2601 unsigned long flags
;
2602 int cpu
, success
= 0;
2605 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2606 if (!(p
->state
& state
))
2609 success
= 1; /* we're going to change ->state */
2612 if (p
->on_rq
&& ttwu_remote(p
, wake_flags
))
2617 * If the owning (remote) cpu is still in the middle of schedule() with
2618 * this task as prev, wait until its done referencing the task.
2621 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2623 * If called from interrupt context we could have landed in the
2624 * middle of schedule(), in this case we should take care not
2625 * to spin on ->on_cpu if p is current, since that would
2636 * Pairs with the smp_wmb() in finish_lock_switch().
2640 p
->sched_contributes_to_load
= !!task_contributes_to_load(p
);
2641 p
->state
= TASK_WAKING
;
2643 if (p
->sched_class
->task_waking
)
2644 p
->sched_class
->task_waking(p
);
2646 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2647 if (task_cpu(p
) != cpu
)
2648 set_task_cpu(p
, cpu
);
2649 #endif /* CONFIG_SMP */
2653 ttwu_stat(p
, cpu
, wake_flags
);
2655 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2661 * try_to_wake_up_local - try to wake up a local task with rq lock held
2662 * @p: the thread to be awakened
2664 * Put @p on the run-queue if it's not already there. The caller must
2665 * ensure that this_rq() is locked, @p is bound to this_rq() and not
2668 static void try_to_wake_up_local(struct task_struct
*p
)
2670 struct rq
*rq
= task_rq(p
);
2672 BUG_ON(rq
!= this_rq());
2673 BUG_ON(p
== current
);
2674 lockdep_assert_held(&rq
->lock
);
2676 if (!raw_spin_trylock(&p
->pi_lock
)) {
2677 raw_spin_unlock(&rq
->lock
);
2678 raw_spin_lock(&p
->pi_lock
);
2679 raw_spin_lock(&rq
->lock
);
2682 if (!(p
->state
& TASK_NORMAL
))
2686 ttwu_activate(rq
, p
, ENQUEUE_WAKEUP
);
2688 ttwu_do_wakeup(rq
, p
, 0);
2689 ttwu_stat(p
, smp_processor_id(), 0);
2691 raw_spin_unlock(&p
->pi_lock
);
2695 * wake_up_process - Wake up a specific process
2696 * @p: The process to be woken up.
2698 * Attempt to wake up the nominated process and move it to the set of runnable
2699 * processes. Returns 1 if the process was woken up, 0 if it was already
2702 * It may be assumed that this function implies a write memory barrier before
2703 * changing the task state if and only if any tasks are woken up.
2705 int wake_up_process(struct task_struct
*p
)
2707 return try_to_wake_up(p
, TASK_ALL
, 0);
2709 EXPORT_SYMBOL(wake_up_process
);
2711 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2713 return try_to_wake_up(p
, state
, 0);
2717 * Perform scheduler related setup for a newly forked process p.
2718 * p is forked by current.
2720 * __sched_fork() is basic setup used by init_idle() too:
2722 static void __sched_fork(struct task_struct
*p
)
2727 p
->se
.exec_start
= 0;
2728 p
->se
.sum_exec_runtime
= 0;
2729 p
->se
.prev_sum_exec_runtime
= 0;
2730 p
->se
.nr_migrations
= 0;
2732 INIT_LIST_HEAD(&p
->se
.group_node
);
2734 #ifdef CONFIG_SCHEDSTATS
2735 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2738 INIT_LIST_HEAD(&p
->rt
.run_list
);
2740 #ifdef CONFIG_PREEMPT_NOTIFIERS
2741 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2746 * fork()/clone()-time setup:
2748 void sched_fork(struct task_struct
*p
, int clone_flags
)
2750 unsigned long flags
;
2751 int cpu
= get_cpu();
2755 * We mark the process as running here. This guarantees that
2756 * nobody will actually run it, and a signal or other external
2757 * event cannot wake it up and insert it on the runqueue either.
2759 p
->state
= TASK_RUNNING
;
2762 * Revert to default priority/policy on fork if requested.
2764 if (unlikely(p
->sched_reset_on_fork
)) {
2765 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2766 p
->policy
= SCHED_NORMAL
;
2767 p
->normal_prio
= p
->static_prio
;
2770 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2771 p
->static_prio
= NICE_TO_PRIO(0);
2772 p
->normal_prio
= p
->static_prio
;
2777 * We don't need the reset flag anymore after the fork. It has
2778 * fulfilled its duty:
2780 p
->sched_reset_on_fork
= 0;
2784 * Make sure we do not leak PI boosting priority to the child.
2786 p
->prio
= current
->normal_prio
;
2788 if (!rt_prio(p
->prio
))
2789 p
->sched_class
= &fair_sched_class
;
2791 if (p
->sched_class
->task_fork
)
2792 p
->sched_class
->task_fork(p
);
2795 * The child is not yet in the pid-hash so no cgroup attach races,
2796 * and the cgroup is pinned to this child due to cgroup_fork()
2797 * is ran before sched_fork().
2799 * Silence PROVE_RCU.
2801 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2802 set_task_cpu(p
, cpu
);
2803 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
2805 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2806 if (likely(sched_info_on()))
2807 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2809 #if defined(CONFIG_SMP)
2812 #ifdef CONFIG_PREEMPT
2813 /* Want to start with kernel preemption disabled. */
2814 task_thread_info(p
)->preempt_count
= 1;
2817 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2824 * wake_up_new_task - wake up a newly created task for the first time.
2826 * This function will do some initial scheduler statistics housekeeping
2827 * that must be done for every newly created context, then puts the task
2828 * on the runqueue and wakes it.
2830 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2832 unsigned long flags
;
2835 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
2838 * Fork balancing, do it here and not earlier because:
2839 * - cpus_allowed can change in the fork path
2840 * - any previously selected cpu might disappear through hotplug
2842 set_task_cpu(p
, select_task_rq(p
, SD_BALANCE_FORK
, 0));
2845 rq
= __task_rq_lock(p
);
2846 activate_task(rq
, p
, 0);
2848 trace_sched_wakeup_new(p
, true);
2849 check_preempt_curr(rq
, p
, WF_FORK
);
2851 if (p
->sched_class
->task_woken
)
2852 p
->sched_class
->task_woken(rq
, p
);
2854 task_rq_unlock(rq
, p
, &flags
);
2857 #ifdef CONFIG_PREEMPT_NOTIFIERS
2860 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2861 * @notifier: notifier struct to register
2863 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2865 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2867 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2870 * preempt_notifier_unregister - no longer interested in preemption notifications
2871 * @notifier: notifier struct to unregister
2873 * This is safe to call from within a preemption notifier.
2875 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2877 hlist_del(¬ifier
->link
);
2879 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2881 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2883 struct preempt_notifier
*notifier
;
2884 struct hlist_node
*node
;
2886 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2887 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2891 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2892 struct task_struct
*next
)
2894 struct preempt_notifier
*notifier
;
2895 struct hlist_node
*node
;
2897 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2898 notifier
->ops
->sched_out(notifier
, next
);
2901 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2903 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2908 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2909 struct task_struct
*next
)
2913 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2916 * prepare_task_switch - prepare to switch tasks
2917 * @rq: the runqueue preparing to switch
2918 * @prev: the current task that is being switched out
2919 * @next: the task we are going to switch to.
2921 * This is called with the rq lock held and interrupts off. It must
2922 * be paired with a subsequent finish_task_switch after the context
2925 * prepare_task_switch sets up locking and calls architecture specific
2929 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2930 struct task_struct
*next
)
2932 sched_info_switch(prev
, next
);
2933 perf_event_task_sched_out(prev
, next
);
2934 fire_sched_out_preempt_notifiers(prev
, next
);
2935 prepare_lock_switch(rq
, next
);
2936 prepare_arch_switch(next
);
2937 trace_sched_switch(prev
, next
);
2941 * finish_task_switch - clean up after a task-switch
2942 * @rq: runqueue associated with task-switch
2943 * @prev: the thread we just switched away from.
2945 * finish_task_switch must be called after the context switch, paired
2946 * with a prepare_task_switch call before the context switch.
2947 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2948 * and do any other architecture-specific cleanup actions.
2950 * Note that we may have delayed dropping an mm in context_switch(). If
2951 * so, we finish that here outside of the runqueue lock. (Doing it
2952 * with the lock held can cause deadlocks; see schedule() for
2955 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2956 __releases(rq
->lock
)
2958 struct mm_struct
*mm
= rq
->prev_mm
;
2964 * A task struct has one reference for the use as "current".
2965 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2966 * schedule one last time. The schedule call will never return, and
2967 * the scheduled task must drop that reference.
2968 * The test for TASK_DEAD must occur while the runqueue locks are
2969 * still held, otherwise prev could be scheduled on another cpu, die
2970 * there before we look at prev->state, and then the reference would
2972 * Manfred Spraul <manfred@colorfullife.com>
2974 prev_state
= prev
->state
;
2975 finish_arch_switch(prev
);
2976 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2977 local_irq_disable();
2978 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2979 perf_event_task_sched_in(current
);
2980 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2982 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2983 finish_lock_switch(rq
, prev
);
2985 fire_sched_in_preempt_notifiers(current
);
2988 if (unlikely(prev_state
== TASK_DEAD
)) {
2990 * Remove function-return probe instances associated with this
2991 * task and put them back on the free list.
2993 kprobe_flush_task(prev
);
2994 put_task_struct(prev
);
3000 /* assumes rq->lock is held */
3001 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
3003 if (prev
->sched_class
->pre_schedule
)
3004 prev
->sched_class
->pre_schedule(rq
, prev
);
3007 /* rq->lock is NOT held, but preemption is disabled */
3008 static inline void post_schedule(struct rq
*rq
)
3010 if (rq
->post_schedule
) {
3011 unsigned long flags
;
3013 raw_spin_lock_irqsave(&rq
->lock
, flags
);
3014 if (rq
->curr
->sched_class
->post_schedule
)
3015 rq
->curr
->sched_class
->post_schedule(rq
);
3016 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
3018 rq
->post_schedule
= 0;
3024 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
3028 static inline void post_schedule(struct rq
*rq
)
3035 * schedule_tail - first thing a freshly forked thread must call.
3036 * @prev: the thread we just switched away from.
3038 asmlinkage
void schedule_tail(struct task_struct
*prev
)
3039 __releases(rq
->lock
)
3041 struct rq
*rq
= this_rq();
3043 finish_task_switch(rq
, prev
);
3046 * FIXME: do we need to worry about rq being invalidated by the
3051 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
3052 /* In this case, finish_task_switch does not reenable preemption */
3055 if (current
->set_child_tid
)
3056 put_user(task_pid_vnr(current
), current
->set_child_tid
);
3060 * context_switch - switch to the new MM and the new
3061 * thread's register state.
3064 context_switch(struct rq
*rq
, struct task_struct
*prev
,
3065 struct task_struct
*next
)
3067 struct mm_struct
*mm
, *oldmm
;
3069 prepare_task_switch(rq
, prev
, next
);
3072 oldmm
= prev
->active_mm
;
3074 * For paravirt, this is coupled with an exit in switch_to to
3075 * combine the page table reload and the switch backend into
3078 arch_start_context_switch(prev
);
3081 next
->active_mm
= oldmm
;
3082 atomic_inc(&oldmm
->mm_count
);
3083 enter_lazy_tlb(oldmm
, next
);
3085 switch_mm(oldmm
, mm
, next
);
3088 prev
->active_mm
= NULL
;
3089 rq
->prev_mm
= oldmm
;
3092 * Since the runqueue lock will be released by the next
3093 * task (which is an invalid locking op but in the case
3094 * of the scheduler it's an obvious special-case), so we
3095 * do an early lockdep release here:
3097 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
3098 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
3101 /* Here we just switch the register state and the stack. */
3102 switch_to(prev
, next
, prev
);
3106 * this_rq must be evaluated again because prev may have moved
3107 * CPUs since it called schedule(), thus the 'rq' on its stack
3108 * frame will be invalid.
3110 finish_task_switch(this_rq(), prev
);
3114 * nr_running, nr_uninterruptible and nr_context_switches:
3116 * externally visible scheduler statistics: current number of runnable
3117 * threads, current number of uninterruptible-sleeping threads, total
3118 * number of context switches performed since bootup.
3120 unsigned long nr_running(void)
3122 unsigned long i
, sum
= 0;
3124 for_each_online_cpu(i
)
3125 sum
+= cpu_rq(i
)->nr_running
;
3130 unsigned long nr_uninterruptible(void)
3132 unsigned long i
, sum
= 0;
3134 for_each_possible_cpu(i
)
3135 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3138 * Since we read the counters lockless, it might be slightly
3139 * inaccurate. Do not allow it to go below zero though:
3141 if (unlikely((long)sum
< 0))
3147 unsigned long long nr_context_switches(void)
3150 unsigned long long sum
= 0;
3152 for_each_possible_cpu(i
)
3153 sum
+= cpu_rq(i
)->nr_switches
;
3158 unsigned long nr_iowait(void)
3160 unsigned long i
, sum
= 0;
3162 for_each_possible_cpu(i
)
3163 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3168 unsigned long nr_iowait_cpu(int cpu
)
3170 struct rq
*this = cpu_rq(cpu
);
3171 return atomic_read(&this->nr_iowait
);
3174 unsigned long this_cpu_load(void)
3176 struct rq
*this = this_rq();
3177 return this->cpu_load
[0];
3181 /* Variables and functions for calc_load */
3182 static atomic_long_t calc_load_tasks
;
3183 static unsigned long calc_load_update
;
3184 unsigned long avenrun
[3];
3185 EXPORT_SYMBOL(avenrun
);
3187 static long calc_load_fold_active(struct rq
*this_rq
)
3189 long nr_active
, delta
= 0;
3191 nr_active
= this_rq
->nr_running
;
3192 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3194 if (nr_active
!= this_rq
->calc_load_active
) {
3195 delta
= nr_active
- this_rq
->calc_load_active
;
3196 this_rq
->calc_load_active
= nr_active
;
3202 static unsigned long
3203 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3206 load
+= active
* (FIXED_1
- exp
);
3207 load
+= 1UL << (FSHIFT
- 1);
3208 return load
>> FSHIFT
;
3213 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
3215 * When making the ILB scale, we should try to pull this in as well.
3217 static atomic_long_t calc_load_tasks_idle
;
3219 static void calc_load_account_idle(struct rq
*this_rq
)
3223 delta
= calc_load_fold_active(this_rq
);
3225 atomic_long_add(delta
, &calc_load_tasks_idle
);
3228 static long calc_load_fold_idle(void)
3233 * Its got a race, we don't care...
3235 if (atomic_long_read(&calc_load_tasks_idle
))
3236 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
3242 * fixed_power_int - compute: x^n, in O(log n) time
3244 * @x: base of the power
3245 * @frac_bits: fractional bits of @x
3246 * @n: power to raise @x to.
3248 * By exploiting the relation between the definition of the natural power
3249 * function: x^n := x*x*...*x (x multiplied by itself for n times), and
3250 * the binary encoding of numbers used by computers: n := \Sum n_i * 2^i,
3251 * (where: n_i \elem {0, 1}, the binary vector representing n),
3252 * we find: x^n := x^(\Sum n_i * 2^i) := \Prod x^(n_i * 2^i), which is
3253 * of course trivially computable in O(log_2 n), the length of our binary
3256 static unsigned long
3257 fixed_power_int(unsigned long x
, unsigned int frac_bits
, unsigned int n
)
3259 unsigned long result
= 1UL << frac_bits
;
3264 result
+= 1UL << (frac_bits
- 1);
3265 result
>>= frac_bits
;
3271 x
+= 1UL << (frac_bits
- 1);
3279 * a1 = a0 * e + a * (1 - e)
3281 * a2 = a1 * e + a * (1 - e)
3282 * = (a0 * e + a * (1 - e)) * e + a * (1 - e)
3283 * = a0 * e^2 + a * (1 - e) * (1 + e)
3285 * a3 = a2 * e + a * (1 - e)
3286 * = (a0 * e^2 + a * (1 - e) * (1 + e)) * e + a * (1 - e)
3287 * = a0 * e^3 + a * (1 - e) * (1 + e + e^2)
3291 * an = a0 * e^n + a * (1 - e) * (1 + e + ... + e^n-1) [1]
3292 * = a0 * e^n + a * (1 - e) * (1 - e^n)/(1 - e)
3293 * = a0 * e^n + a * (1 - e^n)
3295 * [1] application of the geometric series:
3298 * S_n := \Sum x^i = -------------
3301 static unsigned long
3302 calc_load_n(unsigned long load
, unsigned long exp
,
3303 unsigned long active
, unsigned int n
)
3306 return calc_load(load
, fixed_power_int(exp
, FSHIFT
, n
), active
);
3310 * NO_HZ can leave us missing all per-cpu ticks calling
3311 * calc_load_account_active(), but since an idle CPU folds its delta into
3312 * calc_load_tasks_idle per calc_load_account_idle(), all we need to do is fold
3313 * in the pending idle delta if our idle period crossed a load cycle boundary.
3315 * Once we've updated the global active value, we need to apply the exponential
3316 * weights adjusted to the number of cycles missed.
3318 static void calc_global_nohz(unsigned long ticks
)
3320 long delta
, active
, n
;
3322 if (time_before(jiffies
, calc_load_update
))
3326 * If we crossed a calc_load_update boundary, make sure to fold
3327 * any pending idle changes, the respective CPUs might have
3328 * missed the tick driven calc_load_account_active() update
3331 delta
= calc_load_fold_idle();
3333 atomic_long_add(delta
, &calc_load_tasks
);
3336 * If we were idle for multiple load cycles, apply them.
3338 if (ticks
>= LOAD_FREQ
) {
3339 n
= ticks
/ LOAD_FREQ
;
3341 active
= atomic_long_read(&calc_load_tasks
);
3342 active
= active
> 0 ? active
* FIXED_1
: 0;
3344 avenrun
[0] = calc_load_n(avenrun
[0], EXP_1
, active
, n
);
3345 avenrun
[1] = calc_load_n(avenrun
[1], EXP_5
, active
, n
);
3346 avenrun
[2] = calc_load_n(avenrun
[2], EXP_15
, active
, n
);
3348 calc_load_update
+= n
* LOAD_FREQ
;
3352 * Its possible the remainder of the above division also crosses
3353 * a LOAD_FREQ period, the regular check in calc_global_load()
3354 * which comes after this will take care of that.
3356 * Consider us being 11 ticks before a cycle completion, and us
3357 * sleeping for 4*LOAD_FREQ + 22 ticks, then the above code will
3358 * age us 4 cycles, and the test in calc_global_load() will
3359 * pick up the final one.
3363 static void calc_load_account_idle(struct rq
*this_rq
)
3367 static inline long calc_load_fold_idle(void)
3372 static void calc_global_nohz(unsigned long ticks
)
3378 * get_avenrun - get the load average array
3379 * @loads: pointer to dest load array
3380 * @offset: offset to add
3381 * @shift: shift count to shift the result left
3383 * These values are estimates at best, so no need for locking.
3385 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3387 loads
[0] = (avenrun
[0] + offset
) << shift
;
3388 loads
[1] = (avenrun
[1] + offset
) << shift
;
3389 loads
[2] = (avenrun
[2] + offset
) << shift
;
3393 * calc_load - update the avenrun load estimates 10 ticks after the
3394 * CPUs have updated calc_load_tasks.
3396 void calc_global_load(unsigned long ticks
)
3400 calc_global_nohz(ticks
);
3402 if (time_before(jiffies
, calc_load_update
+ 10))
3405 active
= atomic_long_read(&calc_load_tasks
);
3406 active
= active
> 0 ? active
* FIXED_1
: 0;
3408 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3409 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3410 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3412 calc_load_update
+= LOAD_FREQ
;
3416 * Called from update_cpu_load() to periodically update this CPU's
3419 static void calc_load_account_active(struct rq
*this_rq
)
3423 if (time_before(jiffies
, this_rq
->calc_load_update
))
3426 delta
= calc_load_fold_active(this_rq
);
3427 delta
+= calc_load_fold_idle();
3429 atomic_long_add(delta
, &calc_load_tasks
);
3431 this_rq
->calc_load_update
+= LOAD_FREQ
;
3435 * The exact cpuload at various idx values, calculated at every tick would be
3436 * load = (2^idx - 1) / 2^idx * load + 1 / 2^idx * cur_load
3438 * If a cpu misses updates for n-1 ticks (as it was idle) and update gets called
3439 * on nth tick when cpu may be busy, then we have:
3440 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3441 * load = (2^idx - 1) / 2^idx) * load + 1 / 2^idx * cur_load
3443 * decay_load_missed() below does efficient calculation of
3444 * load = ((2^idx - 1) / 2^idx)^(n-1) * load
3445 * avoiding 0..n-1 loop doing load = ((2^idx - 1) / 2^idx) * load
3447 * The calculation is approximated on a 128 point scale.
3448 * degrade_zero_ticks is the number of ticks after which load at any
3449 * particular idx is approximated to be zero.
3450 * degrade_factor is a precomputed table, a row for each load idx.
3451 * Each column corresponds to degradation factor for a power of two ticks,
3452 * based on 128 point scale.
3454 * row 2, col 3 (=12) says that the degradation at load idx 2 after
3455 * 8 ticks is 12/128 (which is an approximation of exact factor 3^8/4^8).
3457 * With this power of 2 load factors, we can degrade the load n times
3458 * by looking at 1 bits in n and doing as many mult/shift instead of
3459 * n mult/shifts needed by the exact degradation.
3461 #define DEGRADE_SHIFT 7
3462 static const unsigned char
3463 degrade_zero_ticks
[CPU_LOAD_IDX_MAX
] = {0, 8, 32, 64, 128};
3464 static const unsigned char
3465 degrade_factor
[CPU_LOAD_IDX_MAX
][DEGRADE_SHIFT
+ 1] = {
3466 {0, 0, 0, 0, 0, 0, 0, 0},
3467 {64, 32, 8, 0, 0, 0, 0, 0},
3468 {96, 72, 40, 12, 1, 0, 0},
3469 {112, 98, 75, 43, 15, 1, 0},
3470 {120, 112, 98, 76, 45, 16, 2} };
3473 * Update cpu_load for any missed ticks, due to tickless idle. The backlog
3474 * would be when CPU is idle and so we just decay the old load without
3475 * adding any new load.
3477 static unsigned long
3478 decay_load_missed(unsigned long load
, unsigned long missed_updates
, int idx
)
3482 if (!missed_updates
)
3485 if (missed_updates
>= degrade_zero_ticks
[idx
])
3489 return load
>> missed_updates
;
3491 while (missed_updates
) {
3492 if (missed_updates
% 2)
3493 load
= (load
* degrade_factor
[idx
][j
]) >> DEGRADE_SHIFT
;
3495 missed_updates
>>= 1;
3502 * Update rq->cpu_load[] statistics. This function is usually called every
3503 * scheduler tick (TICK_NSEC). With tickless idle this will not be called
3504 * every tick. We fix it up based on jiffies.
3506 static void update_cpu_load(struct rq
*this_rq
)
3508 unsigned long this_load
= this_rq
->load
.weight
;
3509 unsigned long curr_jiffies
= jiffies
;
3510 unsigned long pending_updates
;
3513 this_rq
->nr_load_updates
++;
3515 /* Avoid repeated calls on same jiffy, when moving in and out of idle */
3516 if (curr_jiffies
== this_rq
->last_load_update_tick
)
3519 pending_updates
= curr_jiffies
- this_rq
->last_load_update_tick
;
3520 this_rq
->last_load_update_tick
= curr_jiffies
;
3522 /* Update our load: */
3523 this_rq
->cpu_load
[0] = this_load
; /* Fasttrack for idx 0 */
3524 for (i
= 1, scale
= 2; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3525 unsigned long old_load
, new_load
;
3527 /* scale is effectively 1 << i now, and >> i divides by scale */
3529 old_load
= this_rq
->cpu_load
[i
];
3530 old_load
= decay_load_missed(old_load
, pending_updates
- 1, i
);
3531 new_load
= this_load
;
3533 * Round up the averaging division if load is increasing. This
3534 * prevents us from getting stuck on 9 if the load is 10, for
3537 if (new_load
> old_load
)
3538 new_load
+= scale
- 1;
3540 this_rq
->cpu_load
[i
] = (old_load
* (scale
- 1) + new_load
) >> i
;
3543 sched_avg_update(this_rq
);
3546 static void update_cpu_load_active(struct rq
*this_rq
)
3548 update_cpu_load(this_rq
);
3550 calc_load_account_active(this_rq
);
3556 * sched_exec - execve() is a valuable balancing opportunity, because at
3557 * this point the task has the smallest effective memory and cache footprint.
3559 void sched_exec(void)
3561 struct task_struct
*p
= current
;
3562 unsigned long flags
;
3565 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
3566 dest_cpu
= p
->sched_class
->select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3567 if (dest_cpu
== smp_processor_id())
3570 if (likely(cpu_active(dest_cpu
))) {
3571 struct migration_arg arg
= { p
, dest_cpu
};
3573 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3574 stop_one_cpu(task_cpu(p
), migration_cpu_stop
, &arg
);
3578 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
3583 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3585 EXPORT_PER_CPU_SYMBOL(kstat
);
3588 * Return any ns on the sched_clock that have not yet been accounted in
3589 * @p in case that task is currently running.
3591 * Called with task_rq_lock() held on @rq.
3593 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3597 if (task_current(rq
, p
)) {
3598 update_rq_clock(rq
);
3599 ns
= rq
->clock_task
- p
->se
.exec_start
;
3607 unsigned long long task_delta_exec(struct task_struct
*p
)
3609 unsigned long flags
;
3613 rq
= task_rq_lock(p
, &flags
);
3614 ns
= do_task_delta_exec(p
, rq
);
3615 task_rq_unlock(rq
, p
, &flags
);
3621 * Return accounted runtime for the task.
3622 * In case the task is currently running, return the runtime plus current's
3623 * pending runtime that have not been accounted yet.
3625 unsigned long long task_sched_runtime(struct task_struct
*p
)
3627 unsigned long flags
;
3631 rq
= task_rq_lock(p
, &flags
);
3632 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3633 task_rq_unlock(rq
, p
, &flags
);
3639 * Return sum_exec_runtime for the thread group.
3640 * In case the task is currently running, return the sum plus current's
3641 * pending runtime that have not been accounted yet.
3643 * Note that the thread group might have other running tasks as well,
3644 * so the return value not includes other pending runtime that other
3645 * running tasks might have.
3647 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3649 struct task_cputime totals
;
3650 unsigned long flags
;
3654 rq
= task_rq_lock(p
, &flags
);
3655 thread_group_cputime(p
, &totals
);
3656 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3657 task_rq_unlock(rq
, p
, &flags
);
3663 * Account user cpu time to a process.
3664 * @p: the process that the cpu time gets accounted to
3665 * @cputime: the cpu time spent in user space since the last update
3666 * @cputime_scaled: cputime scaled by cpu frequency
3668 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3669 cputime_t cputime_scaled
)
3671 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3674 /* Add user time to process. */
3675 p
->utime
= cputime_add(p
->utime
, cputime
);
3676 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3677 account_group_user_time(p
, cputime
);
3679 /* Add user time to cpustat. */
3680 tmp
= cputime_to_cputime64(cputime
);
3681 if (TASK_NICE(p
) > 0)
3682 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3684 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3686 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3687 /* Account for user time used */
3688 acct_update_integrals(p
);
3692 * Account guest cpu time to a process.
3693 * @p: the process that the cpu time gets accounted to
3694 * @cputime: the cpu time spent in virtual machine since the last update
3695 * @cputime_scaled: cputime scaled by cpu frequency
3697 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3698 cputime_t cputime_scaled
)
3701 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3703 tmp
= cputime_to_cputime64(cputime
);
3705 /* Add guest time to process. */
3706 p
->utime
= cputime_add(p
->utime
, cputime
);
3707 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3708 account_group_user_time(p
, cputime
);
3709 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3711 /* Add guest time to cpustat. */
3712 if (TASK_NICE(p
) > 0) {
3713 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3714 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3716 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3717 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3722 * Account system cpu time to a process and desired cpustat field
3723 * @p: the process that the cpu time gets accounted to
3724 * @cputime: the cpu time spent in kernel space since the last update
3725 * @cputime_scaled: cputime scaled by cpu frequency
3726 * @target_cputime64: pointer to cpustat field that has to be updated
3729 void __account_system_time(struct task_struct
*p
, cputime_t cputime
,
3730 cputime_t cputime_scaled
, cputime64_t
*target_cputime64
)
3732 cputime64_t tmp
= cputime_to_cputime64(cputime
);
3734 /* Add system time to process. */
3735 p
->stime
= cputime_add(p
->stime
, cputime
);
3736 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3737 account_group_system_time(p
, cputime
);
3739 /* Add system time to cpustat. */
3740 *target_cputime64
= cputime64_add(*target_cputime64
, tmp
);
3741 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3743 /* Account for system time used */
3744 acct_update_integrals(p
);
3748 * Account system cpu time to a process.
3749 * @p: the process that the cpu time gets accounted to
3750 * @hardirq_offset: the offset to subtract from hardirq_count()
3751 * @cputime: the cpu time spent in kernel space since the last update
3752 * @cputime_scaled: cputime scaled by cpu frequency
3754 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3755 cputime_t cputime
, cputime_t cputime_scaled
)
3757 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3758 cputime64_t
*target_cputime64
;
3760 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3761 account_guest_time(p
, cputime
, cputime_scaled
);
3765 if (hardirq_count() - hardirq_offset
)
3766 target_cputime64
= &cpustat
->irq
;
3767 else if (in_serving_softirq())
3768 target_cputime64
= &cpustat
->softirq
;
3770 target_cputime64
= &cpustat
->system
;
3772 __account_system_time(p
, cputime
, cputime_scaled
, target_cputime64
);
3776 * Account for involuntary wait time.
3777 * @cputime: the cpu time spent in involuntary wait
3779 void account_steal_time(cputime_t cputime
)
3781 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3782 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3784 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3788 * Account for idle time.
3789 * @cputime: the cpu time spent in idle wait
3791 void account_idle_time(cputime_t cputime
)
3793 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3794 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3795 struct rq
*rq
= this_rq();
3797 if (atomic_read(&rq
->nr_iowait
) > 0)
3798 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3800 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3803 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3805 #ifdef CONFIG_IRQ_TIME_ACCOUNTING
3807 * Account a tick to a process and cpustat
3808 * @p: the process that the cpu time gets accounted to
3809 * @user_tick: is the tick from userspace
3810 * @rq: the pointer to rq
3812 * Tick demultiplexing follows the order
3813 * - pending hardirq update
3814 * - pending softirq update
3818 * - check for guest_time
3819 * - else account as system_time
3821 * Check for hardirq is done both for system and user time as there is
3822 * no timer going off while we are on hardirq and hence we may never get an
3823 * opportunity to update it solely in system time.
3824 * p->stime and friends are only updated on system time and not on irq
3825 * softirq as those do not count in task exec_runtime any more.
3827 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3830 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3831 cputime64_t tmp
= cputime_to_cputime64(cputime_one_jiffy
);
3832 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3834 if (irqtime_account_hi_update()) {
3835 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3836 } else if (irqtime_account_si_update()) {
3837 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3838 } else if (this_cpu_ksoftirqd() == p
) {
3840 * ksoftirqd time do not get accounted in cpu_softirq_time.
3841 * So, we have to handle it separately here.
3842 * Also, p->stime needs to be updated for ksoftirqd.
3844 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3846 } else if (user_tick
) {
3847 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3848 } else if (p
== rq
->idle
) {
3849 account_idle_time(cputime_one_jiffy
);
3850 } else if (p
->flags
& PF_VCPU
) { /* System time or guest time */
3851 account_guest_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3853 __account_system_time(p
, cputime_one_jiffy
, one_jiffy_scaled
,
3858 static void irqtime_account_idle_ticks(int ticks
)
3861 struct rq
*rq
= this_rq();
3863 for (i
= 0; i
< ticks
; i
++)
3864 irqtime_account_process_tick(current
, 0, rq
);
3866 #else /* CONFIG_IRQ_TIME_ACCOUNTING */
3867 static void irqtime_account_idle_ticks(int ticks
) {}
3868 static void irqtime_account_process_tick(struct task_struct
*p
, int user_tick
,
3870 #endif /* CONFIG_IRQ_TIME_ACCOUNTING */
3873 * Account a single tick of cpu time.
3874 * @p: the process that the cpu time gets accounted to
3875 * @user_tick: indicates if the tick is a user or a system tick
3877 void account_process_tick(struct task_struct
*p
, int user_tick
)
3879 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3880 struct rq
*rq
= this_rq();
3882 if (sched_clock_irqtime
) {
3883 irqtime_account_process_tick(p
, user_tick
, rq
);
3888 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3889 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3890 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3893 account_idle_time(cputime_one_jiffy
);
3897 * Account multiple ticks of steal time.
3898 * @p: the process from which the cpu time has been stolen
3899 * @ticks: number of stolen ticks
3901 void account_steal_ticks(unsigned long ticks
)
3903 account_steal_time(jiffies_to_cputime(ticks
));
3907 * Account multiple ticks of idle time.
3908 * @ticks: number of stolen ticks
3910 void account_idle_ticks(unsigned long ticks
)
3913 if (sched_clock_irqtime
) {
3914 irqtime_account_idle_ticks(ticks
);
3918 account_idle_time(jiffies_to_cputime(ticks
));
3924 * Use precise platform statistics if available:
3926 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3927 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3933 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3935 struct task_cputime cputime
;
3937 thread_group_cputime(p
, &cputime
);
3939 *ut
= cputime
.utime
;
3940 *st
= cputime
.stime
;
3944 #ifndef nsecs_to_cputime
3945 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3948 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3950 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3953 * Use CFS's precise accounting:
3955 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3961 do_div(temp
, total
);
3962 utime
= (cputime_t
)temp
;
3967 * Compare with previous values, to keep monotonicity:
3969 p
->prev_utime
= max(p
->prev_utime
, utime
);
3970 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3972 *ut
= p
->prev_utime
;
3973 *st
= p
->prev_stime
;
3977 * Must be called with siglock held.
3979 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3981 struct signal_struct
*sig
= p
->signal
;
3982 struct task_cputime cputime
;
3983 cputime_t rtime
, utime
, total
;
3985 thread_group_cputime(p
, &cputime
);
3987 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3988 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3993 temp
*= cputime
.utime
;
3994 do_div(temp
, total
);
3995 utime
= (cputime_t
)temp
;
3999 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
4000 sig
->prev_stime
= max(sig
->prev_stime
,
4001 cputime_sub(rtime
, sig
->prev_utime
));
4003 *ut
= sig
->prev_utime
;
4004 *st
= sig
->prev_stime
;
4009 * This function gets called by the timer code, with HZ frequency.
4010 * We call it with interrupts disabled.
4012 * It also gets called by the fork code, when changing the parent's
4015 void scheduler_tick(void)
4017 int cpu
= smp_processor_id();
4018 struct rq
*rq
= cpu_rq(cpu
);
4019 struct task_struct
*curr
= rq
->curr
;
4023 raw_spin_lock(&rq
->lock
);
4024 update_rq_clock(rq
);
4025 update_cpu_load_active(rq
);
4026 curr
->sched_class
->task_tick(rq
, curr
, 0);
4027 raw_spin_unlock(&rq
->lock
);
4029 perf_event_task_tick();
4032 rq
->idle_at_tick
= idle_cpu(cpu
);
4033 trigger_load_balance(rq
, cpu
);
4037 notrace
unsigned long get_parent_ip(unsigned long addr
)
4039 if (in_lock_functions(addr
)) {
4040 addr
= CALLER_ADDR2
;
4041 if (in_lock_functions(addr
))
4042 addr
= CALLER_ADDR3
;
4047 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4048 defined(CONFIG_PREEMPT_TRACER))
4050 void __kprobes
add_preempt_count(int val
)
4052 #ifdef CONFIG_DEBUG_PREEMPT
4056 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4059 preempt_count() += val
;
4060 #ifdef CONFIG_DEBUG_PREEMPT
4062 * Spinlock count overflowing soon?
4064 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4067 if (preempt_count() == val
)
4068 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4070 EXPORT_SYMBOL(add_preempt_count
);
4072 void __kprobes
sub_preempt_count(int val
)
4074 #ifdef CONFIG_DEBUG_PREEMPT
4078 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4081 * Is the spinlock portion underflowing?
4083 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4084 !(preempt_count() & PREEMPT_MASK
)))
4088 if (preempt_count() == val
)
4089 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4090 preempt_count() -= val
;
4092 EXPORT_SYMBOL(sub_preempt_count
);
4097 * Print scheduling while atomic bug:
4099 static noinline
void __schedule_bug(struct task_struct
*prev
)
4101 struct pt_regs
*regs
= get_irq_regs();
4103 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4104 prev
->comm
, prev
->pid
, preempt_count());
4106 debug_show_held_locks(prev
);
4108 if (irqs_disabled())
4109 print_irqtrace_events(prev
);
4118 * Various schedule()-time debugging checks and statistics:
4120 static inline void schedule_debug(struct task_struct
*prev
)
4123 * Test if we are atomic. Since do_exit() needs to call into
4124 * schedule() atomically, we ignore that path for now.
4125 * Otherwise, whine if we are scheduling when we should not be.
4127 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4128 __schedule_bug(prev
);
4130 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4132 schedstat_inc(this_rq(), sched_count
);
4133 #ifdef CONFIG_SCHEDSTATS
4134 if (unlikely(prev
->lock_depth
>= 0)) {
4135 schedstat_inc(this_rq(), rq_sched_info
.bkl_count
);
4136 schedstat_inc(prev
, sched_info
.bkl_count
);
4141 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4144 update_rq_clock(rq
);
4145 prev
->sched_class
->put_prev_task(rq
, prev
);
4149 * Pick up the highest-prio task:
4151 static inline struct task_struct
*
4152 pick_next_task(struct rq
*rq
)
4154 const struct sched_class
*class;
4155 struct task_struct
*p
;
4158 * Optimization: we know that if all tasks are in
4159 * the fair class we can call that function directly:
4161 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4162 p
= fair_sched_class
.pick_next_task(rq
);
4167 for_each_class(class) {
4168 p
= class->pick_next_task(rq
);
4173 BUG(); /* the idle class will always have a runnable task */
4177 * schedule() is the main scheduler function.
4179 asmlinkage
void __sched
schedule(void)
4181 struct task_struct
*prev
, *next
;
4182 unsigned long *switch_count
;
4188 cpu
= smp_processor_id();
4190 rcu_note_context_switch(cpu
);
4193 schedule_debug(prev
);
4195 if (sched_feat(HRTICK
))
4198 raw_spin_lock_irq(&rq
->lock
);
4200 switch_count
= &prev
->nivcsw
;
4201 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4202 if (unlikely(signal_pending_state(prev
->state
, prev
))) {
4203 prev
->state
= TASK_RUNNING
;
4205 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
4209 * If a worker went to sleep, notify and ask workqueue
4210 * whether it wants to wake up a task to maintain
4213 if (prev
->flags
& PF_WQ_WORKER
) {
4214 struct task_struct
*to_wakeup
;
4216 to_wakeup
= wq_worker_sleeping(prev
, cpu
);
4218 try_to_wake_up_local(to_wakeup
);
4222 * If we are going to sleep and we have plugged IO
4223 * queued, make sure to submit it to avoid deadlocks.
4225 if (blk_needs_flush_plug(prev
)) {
4226 raw_spin_unlock(&rq
->lock
);
4227 blk_schedule_flush_plug(prev
);
4228 raw_spin_lock(&rq
->lock
);
4231 switch_count
= &prev
->nvcsw
;
4234 pre_schedule(rq
, prev
);
4236 if (unlikely(!rq
->nr_running
))
4237 idle_balance(cpu
, rq
);
4239 put_prev_task(rq
, prev
);
4240 next
= pick_next_task(rq
);
4241 clear_tsk_need_resched(prev
);
4242 rq
->skip_clock_update
= 0;
4244 if (likely(prev
!= next
)) {
4249 context_switch(rq
, prev
, next
); /* unlocks the rq */
4251 * The context switch have flipped the stack from under us
4252 * and restored the local variables which were saved when
4253 * this task called schedule() in the past. prev == current
4254 * is still correct, but it can be moved to another cpu/rq.
4256 cpu
= smp_processor_id();
4259 raw_spin_unlock_irq(&rq
->lock
);
4263 preempt_enable_no_resched();
4267 EXPORT_SYMBOL(schedule
);
4269 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
4271 static inline bool owner_running(struct mutex
*lock
, struct task_struct
*owner
)
4276 if (lock
->owner
!= owner
)
4280 * Ensure we emit the owner->on_cpu, dereference _after_ checking
4281 * lock->owner still matches owner, if that fails, owner might
4282 * point to free()d memory, if it still matches, the rcu_read_lock()
4283 * ensures the memory stays valid.
4287 ret
= owner
->on_cpu
;
4295 * Look out! "owner" is an entirely speculative pointer
4296 * access and not reliable.
4298 int mutex_spin_on_owner(struct mutex
*lock
, struct task_struct
*owner
)
4300 if (!sched_feat(OWNER_SPIN
))
4303 while (owner_running(lock
, owner
)) {
4307 arch_mutex_cpu_relax();
4311 * If the owner changed to another task there is likely
4312 * heavy contention, stop spinning.
4321 #ifdef CONFIG_PREEMPT
4323 * this is the entry point to schedule() from in-kernel preemption
4324 * off of preempt_enable. Kernel preemptions off return from interrupt
4325 * occur there and call schedule directly.
4327 asmlinkage
void __sched notrace
preempt_schedule(void)
4329 struct thread_info
*ti
= current_thread_info();
4332 * If there is a non-zero preempt_count or interrupts are disabled,
4333 * we do not want to preempt the current task. Just return..
4335 if (likely(ti
->preempt_count
|| irqs_disabled()))
4339 add_preempt_count_notrace(PREEMPT_ACTIVE
);
4341 sub_preempt_count_notrace(PREEMPT_ACTIVE
);
4344 * Check again in case we missed a preemption opportunity
4345 * between schedule and now.
4348 } while (need_resched());
4350 EXPORT_SYMBOL(preempt_schedule
);
4353 * this is the entry point to schedule() from kernel preemption
4354 * off of irq context.
4355 * Note, that this is called and return with irqs disabled. This will
4356 * protect us against recursive calling from irq.
4358 asmlinkage
void __sched
preempt_schedule_irq(void)
4360 struct thread_info
*ti
= current_thread_info();
4362 /* Catch callers which need to be fixed */
4363 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4366 add_preempt_count(PREEMPT_ACTIVE
);
4369 local_irq_disable();
4370 sub_preempt_count(PREEMPT_ACTIVE
);
4373 * Check again in case we missed a preemption opportunity
4374 * between schedule and now.
4377 } while (need_resched());
4380 #endif /* CONFIG_PREEMPT */
4382 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
4385 return try_to_wake_up(curr
->private, mode
, wake_flags
);
4387 EXPORT_SYMBOL(default_wake_function
);
4390 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4391 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4392 * number) then we wake all the non-exclusive tasks and one exclusive task.
4394 * There are circumstances in which we can try to wake a task which has already
4395 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4396 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4398 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4399 int nr_exclusive
, int wake_flags
, void *key
)
4401 wait_queue_t
*curr
, *next
;
4403 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4404 unsigned flags
= curr
->flags
;
4406 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
4407 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4413 * __wake_up - wake up threads blocked on a waitqueue.
4415 * @mode: which threads
4416 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4417 * @key: is directly passed to the wakeup function
4419 * It may be assumed that this function implies a write memory barrier before
4420 * changing the task state if and only if any tasks are woken up.
4422 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4423 int nr_exclusive
, void *key
)
4425 unsigned long flags
;
4427 spin_lock_irqsave(&q
->lock
, flags
);
4428 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4429 spin_unlock_irqrestore(&q
->lock
, flags
);
4431 EXPORT_SYMBOL(__wake_up
);
4434 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4436 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4438 __wake_up_common(q
, mode
, 1, 0, NULL
);
4440 EXPORT_SYMBOL_GPL(__wake_up_locked
);
4442 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
4444 __wake_up_common(q
, mode
, 1, 0, key
);
4446 EXPORT_SYMBOL_GPL(__wake_up_locked_key
);
4449 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
4451 * @mode: which threads
4452 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4453 * @key: opaque value to be passed to wakeup targets
4455 * The sync wakeup differs that the waker knows that it will schedule
4456 * away soon, so while the target thread will be woken up, it will not
4457 * be migrated to another CPU - ie. the two threads are 'synchronized'
4458 * with each other. This can prevent needless bouncing between CPUs.
4460 * On UP it can prevent extra preemption.
4462 * It may be assumed that this function implies a write memory barrier before
4463 * changing the task state if and only if any tasks are woken up.
4465 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
4466 int nr_exclusive
, void *key
)
4468 unsigned long flags
;
4469 int wake_flags
= WF_SYNC
;
4474 if (unlikely(!nr_exclusive
))
4477 spin_lock_irqsave(&q
->lock
, flags
);
4478 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
4479 spin_unlock_irqrestore(&q
->lock
, flags
);
4481 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
4484 * __wake_up_sync - see __wake_up_sync_key()
4486 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4488 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
4490 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4493 * complete: - signals a single thread waiting on this completion
4494 * @x: holds the state of this particular completion
4496 * This will wake up a single thread waiting on this completion. Threads will be
4497 * awakened in the same order in which they were queued.
4499 * See also complete_all(), wait_for_completion() and related routines.
4501 * It may be assumed that this function implies a write memory barrier before
4502 * changing the task state if and only if any tasks are woken up.
4504 void complete(struct completion
*x
)
4506 unsigned long flags
;
4508 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4510 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4511 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4513 EXPORT_SYMBOL(complete
);
4516 * complete_all: - signals all threads waiting on this completion
4517 * @x: holds the state of this particular completion
4519 * This will wake up all threads waiting on this particular completion event.
4521 * It may be assumed that this function implies a write memory barrier before
4522 * changing the task state if and only if any tasks are woken up.
4524 void complete_all(struct completion
*x
)
4526 unsigned long flags
;
4528 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4529 x
->done
+= UINT_MAX
/2;
4530 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4531 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4533 EXPORT_SYMBOL(complete_all
);
4535 static inline long __sched
4536 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4539 DECLARE_WAITQUEUE(wait
, current
);
4541 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
4543 if (signal_pending_state(state
, current
)) {
4544 timeout
= -ERESTARTSYS
;
4547 __set_current_state(state
);
4548 spin_unlock_irq(&x
->wait
.lock
);
4549 timeout
= schedule_timeout(timeout
);
4550 spin_lock_irq(&x
->wait
.lock
);
4551 } while (!x
->done
&& timeout
);
4552 __remove_wait_queue(&x
->wait
, &wait
);
4557 return timeout
?: 1;
4561 wait_for_common(struct completion
*x
, long timeout
, int state
)
4565 spin_lock_irq(&x
->wait
.lock
);
4566 timeout
= do_wait_for_common(x
, timeout
, state
);
4567 spin_unlock_irq(&x
->wait
.lock
);
4572 * wait_for_completion: - waits for completion of a task
4573 * @x: holds the state of this particular completion
4575 * This waits to be signaled for completion of a specific task. It is NOT
4576 * interruptible and there is no timeout.
4578 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4579 * and interrupt capability. Also see complete().
4581 void __sched
wait_for_completion(struct completion
*x
)
4583 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4585 EXPORT_SYMBOL(wait_for_completion
);
4588 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4589 * @x: holds the state of this particular completion
4590 * @timeout: timeout value in jiffies
4592 * This waits for either a completion of a specific task to be signaled or for a
4593 * specified timeout to expire. The timeout is in jiffies. It is not
4596 unsigned long __sched
4597 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4599 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4601 EXPORT_SYMBOL(wait_for_completion_timeout
);
4604 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4605 * @x: holds the state of this particular completion
4607 * This waits for completion of a specific task to be signaled. It is
4610 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4612 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4613 if (t
== -ERESTARTSYS
)
4617 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4620 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4621 * @x: holds the state of this particular completion
4622 * @timeout: timeout value in jiffies
4624 * This waits for either a completion of a specific task to be signaled or for a
4625 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4628 wait_for_completion_interruptible_timeout(struct completion
*x
,
4629 unsigned long timeout
)
4631 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4633 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4636 * wait_for_completion_killable: - waits for completion of a task (killable)
4637 * @x: holds the state of this particular completion
4639 * This waits to be signaled for completion of a specific task. It can be
4640 * interrupted by a kill signal.
4642 int __sched
wait_for_completion_killable(struct completion
*x
)
4644 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4645 if (t
== -ERESTARTSYS
)
4649 EXPORT_SYMBOL(wait_for_completion_killable
);
4652 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4653 * @x: holds the state of this particular completion
4654 * @timeout: timeout value in jiffies
4656 * This waits for either a completion of a specific task to be
4657 * signaled or for a specified timeout to expire. It can be
4658 * interrupted by a kill signal. The timeout is in jiffies.
4661 wait_for_completion_killable_timeout(struct completion
*x
,
4662 unsigned long timeout
)
4664 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4666 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4669 * try_wait_for_completion - try to decrement a completion without blocking
4670 * @x: completion structure
4672 * Returns: 0 if a decrement cannot be done without blocking
4673 * 1 if a decrement succeeded.
4675 * If a completion is being used as a counting completion,
4676 * attempt to decrement the counter without blocking. This
4677 * enables us to avoid waiting if the resource the completion
4678 * is protecting is not available.
4680 bool try_wait_for_completion(struct completion
*x
)
4682 unsigned long flags
;
4685 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4690 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4693 EXPORT_SYMBOL(try_wait_for_completion
);
4696 * completion_done - Test to see if a completion has any waiters
4697 * @x: completion structure
4699 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4700 * 1 if there are no waiters.
4703 bool completion_done(struct completion
*x
)
4705 unsigned long flags
;
4708 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4711 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4714 EXPORT_SYMBOL(completion_done
);
4717 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4719 unsigned long flags
;
4722 init_waitqueue_entry(&wait
, current
);
4724 __set_current_state(state
);
4726 spin_lock_irqsave(&q
->lock
, flags
);
4727 __add_wait_queue(q
, &wait
);
4728 spin_unlock(&q
->lock
);
4729 timeout
= schedule_timeout(timeout
);
4730 spin_lock_irq(&q
->lock
);
4731 __remove_wait_queue(q
, &wait
);
4732 spin_unlock_irqrestore(&q
->lock
, flags
);
4737 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4739 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4741 EXPORT_SYMBOL(interruptible_sleep_on
);
4744 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4746 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4748 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4750 void __sched
sleep_on(wait_queue_head_t
*q
)
4752 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4754 EXPORT_SYMBOL(sleep_on
);
4756 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4758 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4760 EXPORT_SYMBOL(sleep_on_timeout
);
4762 #ifdef CONFIG_RT_MUTEXES
4765 * rt_mutex_setprio - set the current priority of a task
4767 * @prio: prio value (kernel-internal form)
4769 * This function changes the 'effective' priority of a task. It does
4770 * not touch ->normal_prio like __setscheduler().
4772 * Used by the rt_mutex code to implement priority inheritance logic.
4774 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4776 int oldprio
, on_rq
, running
;
4778 const struct sched_class
*prev_class
;
4780 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4782 rq
= __task_rq_lock(p
);
4784 trace_sched_pi_setprio(p
, prio
);
4786 prev_class
= p
->sched_class
;
4788 running
= task_current(rq
, p
);
4790 dequeue_task(rq
, p
, 0);
4792 p
->sched_class
->put_prev_task(rq
, p
);
4795 p
->sched_class
= &rt_sched_class
;
4797 p
->sched_class
= &fair_sched_class
;
4802 p
->sched_class
->set_curr_task(rq
);
4804 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4806 check_class_changed(rq
, p
, prev_class
, oldprio
);
4807 __task_rq_unlock(rq
);
4812 void set_user_nice(struct task_struct
*p
, long nice
)
4814 int old_prio
, delta
, on_rq
;
4815 unsigned long flags
;
4818 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4821 * We have to be careful, if called from sys_setpriority(),
4822 * the task might be in the middle of scheduling on another CPU.
4824 rq
= task_rq_lock(p
, &flags
);
4826 * The RT priorities are set via sched_setscheduler(), but we still
4827 * allow the 'normal' nice value to be set - but as expected
4828 * it wont have any effect on scheduling until the task is
4829 * SCHED_FIFO/SCHED_RR:
4831 if (task_has_rt_policy(p
)) {
4832 p
->static_prio
= NICE_TO_PRIO(nice
);
4837 dequeue_task(rq
, p
, 0);
4839 p
->static_prio
= NICE_TO_PRIO(nice
);
4842 p
->prio
= effective_prio(p
);
4843 delta
= p
->prio
- old_prio
;
4846 enqueue_task(rq
, p
, 0);
4848 * If the task increased its priority or is running and
4849 * lowered its priority, then reschedule its CPU:
4851 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4852 resched_task(rq
->curr
);
4855 task_rq_unlock(rq
, p
, &flags
);
4857 EXPORT_SYMBOL(set_user_nice
);
4860 * can_nice - check if a task can reduce its nice value
4864 int can_nice(const struct task_struct
*p
, const int nice
)
4866 /* convert nice value [19,-20] to rlimit style value [1,40] */
4867 int nice_rlim
= 20 - nice
;
4869 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4870 capable(CAP_SYS_NICE
));
4873 #ifdef __ARCH_WANT_SYS_NICE
4876 * sys_nice - change the priority of the current process.
4877 * @increment: priority increment
4879 * sys_setpriority is a more generic, but much slower function that
4880 * does similar things.
4882 SYSCALL_DEFINE1(nice
, int, increment
)
4887 * Setpriority might change our priority at the same moment.
4888 * We don't have to worry. Conceptually one call occurs first
4889 * and we have a single winner.
4891 if (increment
< -40)
4896 nice
= TASK_NICE(current
) + increment
;
4902 if (increment
< 0 && !can_nice(current
, nice
))
4905 retval
= security_task_setnice(current
, nice
);
4909 set_user_nice(current
, nice
);
4916 * task_prio - return the priority value of a given task.
4917 * @p: the task in question.
4919 * This is the priority value as seen by users in /proc.
4920 * RT tasks are offset by -200. Normal tasks are centered
4921 * around 0, value goes from -16 to +15.
4923 int task_prio(const struct task_struct
*p
)
4925 return p
->prio
- MAX_RT_PRIO
;
4929 * task_nice - return the nice value of a given task.
4930 * @p: the task in question.
4932 int task_nice(const struct task_struct
*p
)
4934 return TASK_NICE(p
);
4936 EXPORT_SYMBOL(task_nice
);
4939 * idle_cpu - is a given cpu idle currently?
4940 * @cpu: the processor in question.
4942 int idle_cpu(int cpu
)
4944 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4948 * idle_task - return the idle task for a given cpu.
4949 * @cpu: the processor in question.
4951 struct task_struct
*idle_task(int cpu
)
4953 return cpu_rq(cpu
)->idle
;
4957 * find_process_by_pid - find a process with a matching PID value.
4958 * @pid: the pid in question.
4960 static struct task_struct
*find_process_by_pid(pid_t pid
)
4962 return pid
? find_task_by_vpid(pid
) : current
;
4965 /* Actually do priority change: must hold rq lock. */
4967 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4970 p
->rt_priority
= prio
;
4971 p
->normal_prio
= normal_prio(p
);
4972 /* we are holding p->pi_lock already */
4973 p
->prio
= rt_mutex_getprio(p
);
4974 if (rt_prio(p
->prio
))
4975 p
->sched_class
= &rt_sched_class
;
4977 p
->sched_class
= &fair_sched_class
;
4982 * check the target process has a UID that matches the current process's
4984 static bool check_same_owner(struct task_struct
*p
)
4986 const struct cred
*cred
= current_cred(), *pcred
;
4990 pcred
= __task_cred(p
);
4991 if (cred
->user
->user_ns
== pcred
->user
->user_ns
)
4992 match
= (cred
->euid
== pcred
->euid
||
4993 cred
->euid
== pcred
->uid
);
5000 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5001 const struct sched_param
*param
, bool user
)
5003 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5004 unsigned long flags
;
5005 const struct sched_class
*prev_class
;
5009 /* may grab non-irq protected spin_locks */
5010 BUG_ON(in_interrupt());
5012 /* double check policy once rq lock held */
5014 reset_on_fork
= p
->sched_reset_on_fork
;
5015 policy
= oldpolicy
= p
->policy
;
5017 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
5018 policy
&= ~SCHED_RESET_ON_FORK
;
5020 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5021 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5022 policy
!= SCHED_IDLE
)
5027 * Valid priorities for SCHED_FIFO and SCHED_RR are
5028 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5029 * SCHED_BATCH and SCHED_IDLE is 0.
5031 if (param
->sched_priority
< 0 ||
5032 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5033 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5035 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5039 * Allow unprivileged RT tasks to decrease priority:
5041 if (user
&& !capable(CAP_SYS_NICE
)) {
5042 if (rt_policy(policy
)) {
5043 unsigned long rlim_rtprio
=
5044 task_rlimit(p
, RLIMIT_RTPRIO
);
5046 /* can't set/change the rt policy */
5047 if (policy
!= p
->policy
&& !rlim_rtprio
)
5050 /* can't increase priority */
5051 if (param
->sched_priority
> p
->rt_priority
&&
5052 param
->sched_priority
> rlim_rtprio
)
5057 * Treat SCHED_IDLE as nice 20. Only allow a switch to
5058 * SCHED_NORMAL if the RLIMIT_NICE would normally permit it.
5060 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
) {
5061 if (!can_nice(p
, TASK_NICE(p
)))
5065 /* can't change other user's priorities */
5066 if (!check_same_owner(p
))
5069 /* Normal users shall not reset the sched_reset_on_fork flag */
5070 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
5075 retval
= security_task_setscheduler(p
);
5081 * make sure no PI-waiters arrive (or leave) while we are
5082 * changing the priority of the task:
5084 * To be able to change p->policy safely, the appropriate
5085 * runqueue lock must be held.
5087 rq
= task_rq_lock(p
, &flags
);
5090 * Changing the policy of the stop threads its a very bad idea
5092 if (p
== rq
->stop
) {
5093 task_rq_unlock(rq
, p
, &flags
);
5098 * If not changing anything there's no need to proceed further:
5100 if (unlikely(policy
== p
->policy
&& (!rt_policy(policy
) ||
5101 param
->sched_priority
== p
->rt_priority
))) {
5103 __task_rq_unlock(rq
);
5104 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5108 #ifdef CONFIG_RT_GROUP_SCHED
5111 * Do not allow realtime tasks into groups that have no runtime
5114 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5115 task_group(p
)->rt_bandwidth
.rt_runtime
== 0 &&
5116 !task_group_is_autogroup(task_group(p
))) {
5117 task_rq_unlock(rq
, p
, &flags
);
5123 /* recheck policy now with rq lock held */
5124 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5125 policy
= oldpolicy
= -1;
5126 task_rq_unlock(rq
, p
, &flags
);
5130 running
= task_current(rq
, p
);
5132 deactivate_task(rq
, p
, 0);
5134 p
->sched_class
->put_prev_task(rq
, p
);
5136 p
->sched_reset_on_fork
= reset_on_fork
;
5139 prev_class
= p
->sched_class
;
5140 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5143 p
->sched_class
->set_curr_task(rq
);
5145 activate_task(rq
, p
, 0);
5147 check_class_changed(rq
, p
, prev_class
, oldprio
);
5148 task_rq_unlock(rq
, p
, &flags
);
5150 rt_mutex_adjust_pi(p
);
5156 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5157 * @p: the task in question.
5158 * @policy: new policy.
5159 * @param: structure containing the new RT priority.
5161 * NOTE that the task may be already dead.
5163 int sched_setscheduler(struct task_struct
*p
, int policy
,
5164 const struct sched_param
*param
)
5166 return __sched_setscheduler(p
, policy
, param
, true);
5168 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5171 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5172 * @p: the task in question.
5173 * @policy: new policy.
5174 * @param: structure containing the new RT priority.
5176 * Just like sched_setscheduler, only don't bother checking if the
5177 * current context has permission. For example, this is needed in
5178 * stop_machine(): we create temporary high priority worker threads,
5179 * but our caller might not have that capability.
5181 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5182 const struct sched_param
*param
)
5184 return __sched_setscheduler(p
, policy
, param
, false);
5188 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5190 struct sched_param lparam
;
5191 struct task_struct
*p
;
5194 if (!param
|| pid
< 0)
5196 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5201 p
= find_process_by_pid(pid
);
5203 retval
= sched_setscheduler(p
, policy
, &lparam
);
5210 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5211 * @pid: the pid in question.
5212 * @policy: new policy.
5213 * @param: structure containing the new RT priority.
5215 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5216 struct sched_param __user
*, param
)
5218 /* negative values for policy are not valid */
5222 return do_sched_setscheduler(pid
, policy
, param
);
5226 * sys_sched_setparam - set/change the RT priority of a thread
5227 * @pid: the pid in question.
5228 * @param: structure containing the new RT priority.
5230 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5232 return do_sched_setscheduler(pid
, -1, param
);
5236 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5237 * @pid: the pid in question.
5239 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
5241 struct task_struct
*p
;
5249 p
= find_process_by_pid(pid
);
5251 retval
= security_task_getscheduler(p
);
5254 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
5261 * sys_sched_getparam - get the RT priority of a thread
5262 * @pid: the pid in question.
5263 * @param: structure containing the RT priority.
5265 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
5267 struct sched_param lp
;
5268 struct task_struct
*p
;
5271 if (!param
|| pid
< 0)
5275 p
= find_process_by_pid(pid
);
5280 retval
= security_task_getscheduler(p
);
5284 lp
.sched_priority
= p
->rt_priority
;
5288 * This one might sleep, we cannot do it with a spinlock held ...
5290 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5299 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
5301 cpumask_var_t cpus_allowed
, new_mask
;
5302 struct task_struct
*p
;
5308 p
= find_process_by_pid(pid
);
5315 /* Prevent p going away */
5319 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
5323 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
5325 goto out_free_cpus_allowed
;
5328 if (!check_same_owner(p
) && !task_ns_capable(p
, CAP_SYS_NICE
))
5331 retval
= security_task_setscheduler(p
);
5335 cpuset_cpus_allowed(p
, cpus_allowed
);
5336 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
5338 retval
= set_cpus_allowed_ptr(p
, new_mask
);
5341 cpuset_cpus_allowed(p
, cpus_allowed
);
5342 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
5344 * We must have raced with a concurrent cpuset
5345 * update. Just reset the cpus_allowed to the
5346 * cpuset's cpus_allowed
5348 cpumask_copy(new_mask
, cpus_allowed
);
5353 free_cpumask_var(new_mask
);
5354 out_free_cpus_allowed
:
5355 free_cpumask_var(cpus_allowed
);
5362 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5363 struct cpumask
*new_mask
)
5365 if (len
< cpumask_size())
5366 cpumask_clear(new_mask
);
5367 else if (len
> cpumask_size())
5368 len
= cpumask_size();
5370 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5374 * sys_sched_setaffinity - set the cpu affinity of a process
5375 * @pid: pid of the process
5376 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5377 * @user_mask_ptr: user-space pointer to the new cpu mask
5379 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
5380 unsigned long __user
*, user_mask_ptr
)
5382 cpumask_var_t new_mask
;
5385 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
5388 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
5390 retval
= sched_setaffinity(pid
, new_mask
);
5391 free_cpumask_var(new_mask
);
5395 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
5397 struct task_struct
*p
;
5398 unsigned long flags
;
5405 p
= find_process_by_pid(pid
);
5409 retval
= security_task_getscheduler(p
);
5413 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
5414 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
5415 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5425 * sys_sched_getaffinity - get the cpu affinity of a process
5426 * @pid: pid of the process
5427 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5428 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5430 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
5431 unsigned long __user
*, user_mask_ptr
)
5436 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
5438 if (len
& (sizeof(unsigned long)-1))
5441 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
5444 ret
= sched_getaffinity(pid
, mask
);
5446 size_t retlen
= min_t(size_t, len
, cpumask_size());
5448 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
5453 free_cpumask_var(mask
);
5459 * sys_sched_yield - yield the current processor to other threads.
5461 * This function yields the current CPU to other tasks. If there are no
5462 * other threads running on this CPU then this function will return.
5464 SYSCALL_DEFINE0(sched_yield
)
5466 struct rq
*rq
= this_rq_lock();
5468 schedstat_inc(rq
, yld_count
);
5469 current
->sched_class
->yield_task(rq
);
5472 * Since we are going to call schedule() anyway, there's
5473 * no need to preempt or enable interrupts:
5475 __release(rq
->lock
);
5476 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5477 do_raw_spin_unlock(&rq
->lock
);
5478 preempt_enable_no_resched();
5485 static inline int should_resched(void)
5487 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
5490 static void __cond_resched(void)
5492 add_preempt_count(PREEMPT_ACTIVE
);
5494 sub_preempt_count(PREEMPT_ACTIVE
);
5497 int __sched
_cond_resched(void)
5499 if (should_resched()) {
5505 EXPORT_SYMBOL(_cond_resched
);
5508 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
5509 * call schedule, and on return reacquire the lock.
5511 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5512 * operations here to prevent schedule() from being called twice (once via
5513 * spin_unlock(), once by hand).
5515 int __cond_resched_lock(spinlock_t
*lock
)
5517 int resched
= should_resched();
5520 lockdep_assert_held(lock
);
5522 if (spin_needbreak(lock
) || resched
) {
5533 EXPORT_SYMBOL(__cond_resched_lock
);
5535 int __sched
__cond_resched_softirq(void)
5537 BUG_ON(!in_softirq());
5539 if (should_resched()) {
5547 EXPORT_SYMBOL(__cond_resched_softirq
);
5550 * yield - yield the current processor to other threads.
5552 * This is a shortcut for kernel-space yielding - it marks the
5553 * thread runnable and calls sys_sched_yield().
5555 void __sched
yield(void)
5557 set_current_state(TASK_RUNNING
);
5560 EXPORT_SYMBOL(yield
);
5563 * yield_to - yield the current processor to another thread in
5564 * your thread group, or accelerate that thread toward the
5565 * processor it's on.
5567 * @preempt: whether task preemption is allowed or not
5569 * It's the caller's job to ensure that the target task struct
5570 * can't go away on us before we can do any checks.
5572 * Returns true if we indeed boosted the target task.
5574 bool __sched
yield_to(struct task_struct
*p
, bool preempt
)
5576 struct task_struct
*curr
= current
;
5577 struct rq
*rq
, *p_rq
;
5578 unsigned long flags
;
5581 local_irq_save(flags
);
5586 double_rq_lock(rq
, p_rq
);
5587 while (task_rq(p
) != p_rq
) {
5588 double_rq_unlock(rq
, p_rq
);
5592 if (!curr
->sched_class
->yield_to_task
)
5595 if (curr
->sched_class
!= p
->sched_class
)
5598 if (task_running(p_rq
, p
) || p
->state
)
5601 yielded
= curr
->sched_class
->yield_to_task(rq
, p
, preempt
);
5603 schedstat_inc(rq
, yld_count
);
5605 * Make p's CPU reschedule; pick_next_entity takes care of
5608 if (preempt
&& rq
!= p_rq
)
5609 resched_task(p_rq
->curr
);
5613 double_rq_unlock(rq
, p_rq
);
5614 local_irq_restore(flags
);
5621 EXPORT_SYMBOL_GPL(yield_to
);
5624 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5625 * that process accounting knows that this is a task in IO wait state.
5627 void __sched
io_schedule(void)
5629 struct rq
*rq
= raw_rq();
5631 delayacct_blkio_start();
5632 atomic_inc(&rq
->nr_iowait
);
5633 blk_flush_plug(current
);
5634 current
->in_iowait
= 1;
5636 current
->in_iowait
= 0;
5637 atomic_dec(&rq
->nr_iowait
);
5638 delayacct_blkio_end();
5640 EXPORT_SYMBOL(io_schedule
);
5642 long __sched
io_schedule_timeout(long timeout
)
5644 struct rq
*rq
= raw_rq();
5647 delayacct_blkio_start();
5648 atomic_inc(&rq
->nr_iowait
);
5649 blk_flush_plug(current
);
5650 current
->in_iowait
= 1;
5651 ret
= schedule_timeout(timeout
);
5652 current
->in_iowait
= 0;
5653 atomic_dec(&rq
->nr_iowait
);
5654 delayacct_blkio_end();
5659 * sys_sched_get_priority_max - return maximum RT priority.
5660 * @policy: scheduling class.
5662 * this syscall returns the maximum rt_priority that can be used
5663 * by a given scheduling class.
5665 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5672 ret
= MAX_USER_RT_PRIO
-1;
5684 * sys_sched_get_priority_min - return minimum RT priority.
5685 * @policy: scheduling class.
5687 * this syscall returns the minimum rt_priority that can be used
5688 * by a given scheduling class.
5690 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5708 * sys_sched_rr_get_interval - return the default timeslice of a process.
5709 * @pid: pid of the process.
5710 * @interval: userspace pointer to the timeslice value.
5712 * this syscall writes the default timeslice value of a given process
5713 * into the user-space timespec buffer. A value of '0' means infinity.
5715 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5716 struct timespec __user
*, interval
)
5718 struct task_struct
*p
;
5719 unsigned int time_slice
;
5720 unsigned long flags
;
5730 p
= find_process_by_pid(pid
);
5734 retval
= security_task_getscheduler(p
);
5738 rq
= task_rq_lock(p
, &flags
);
5739 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5740 task_rq_unlock(rq
, p
, &flags
);
5743 jiffies_to_timespec(time_slice
, &t
);
5744 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5752 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5754 void sched_show_task(struct task_struct
*p
)
5756 unsigned long free
= 0;
5759 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5760 printk(KERN_INFO
"%-15.15s %c", p
->comm
,
5761 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5762 #if BITS_PER_LONG == 32
5763 if (state
== TASK_RUNNING
)
5764 printk(KERN_CONT
" running ");
5766 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5768 if (state
== TASK_RUNNING
)
5769 printk(KERN_CONT
" running task ");
5771 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5773 #ifdef CONFIG_DEBUG_STACK_USAGE
5774 free
= stack_not_used(p
);
5776 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5777 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5778 (unsigned long)task_thread_info(p
)->flags
);
5780 show_stack(p
, NULL
);
5783 void show_state_filter(unsigned long state_filter
)
5785 struct task_struct
*g
, *p
;
5787 #if BITS_PER_LONG == 32
5789 " task PC stack pid father\n");
5792 " task PC stack pid father\n");
5794 read_lock(&tasklist_lock
);
5795 do_each_thread(g
, p
) {
5797 * reset the NMI-timeout, listing all files on a slow
5798 * console might take a lot of time:
5800 touch_nmi_watchdog();
5801 if (!state_filter
|| (p
->state
& state_filter
))
5803 } while_each_thread(g
, p
);
5805 touch_all_softlockup_watchdogs();
5807 #ifdef CONFIG_SCHED_DEBUG
5808 sysrq_sched_debug_show();
5810 read_unlock(&tasklist_lock
);
5812 * Only show locks if all tasks are dumped:
5815 debug_show_all_locks();
5818 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5820 idle
->sched_class
= &idle_sched_class
;
5824 * init_idle - set up an idle thread for a given CPU
5825 * @idle: task in question
5826 * @cpu: cpu the idle task belongs to
5828 * NOTE: this function does not set the idle thread's NEED_RESCHED
5829 * flag, to make booting more robust.
5831 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5833 struct rq
*rq
= cpu_rq(cpu
);
5834 unsigned long flags
;
5836 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5839 idle
->state
= TASK_RUNNING
;
5840 idle
->se
.exec_start
= sched_clock();
5842 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5844 * We're having a chicken and egg problem, even though we are
5845 * holding rq->lock, the cpu isn't yet set to this cpu so the
5846 * lockdep check in task_group() will fail.
5848 * Similar case to sched_fork(). / Alternatively we could
5849 * use task_rq_lock() here and obtain the other rq->lock.
5854 __set_task_cpu(idle
, cpu
);
5857 rq
->curr
= rq
->idle
= idle
;
5858 #if defined(CONFIG_SMP)
5861 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5863 /* Set the preempt count _outside_ the spinlocks! */
5864 #if defined(CONFIG_PREEMPT)
5865 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5867 task_thread_info(idle
)->preempt_count
= 0;
5870 * The idle tasks have their own, simple scheduling class:
5872 idle
->sched_class
= &idle_sched_class
;
5873 ftrace_graph_init_idle_task(idle
, cpu
);
5877 * In a system that switches off the HZ timer nohz_cpu_mask
5878 * indicates which cpus entered this state. This is used
5879 * in the rcu update to wait only for active cpus. For system
5880 * which do not switch off the HZ timer nohz_cpu_mask should
5881 * always be CPU_BITS_NONE.
5883 cpumask_var_t nohz_cpu_mask
;
5886 * Increase the granularity value when there are more CPUs,
5887 * because with more CPUs the 'effective latency' as visible
5888 * to users decreases. But the relationship is not linear,
5889 * so pick a second-best guess by going with the log2 of the
5892 * This idea comes from the SD scheduler of Con Kolivas:
5894 static int get_update_sysctl_factor(void)
5896 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5897 unsigned int factor
;
5899 switch (sysctl_sched_tunable_scaling
) {
5900 case SCHED_TUNABLESCALING_NONE
:
5903 case SCHED_TUNABLESCALING_LINEAR
:
5906 case SCHED_TUNABLESCALING_LOG
:
5908 factor
= 1 + ilog2(cpus
);
5915 static void update_sysctl(void)
5917 unsigned int factor
= get_update_sysctl_factor();
5919 #define SET_SYSCTL(name) \
5920 (sysctl_##name = (factor) * normalized_sysctl_##name)
5921 SET_SYSCTL(sched_min_granularity
);
5922 SET_SYSCTL(sched_latency
);
5923 SET_SYSCTL(sched_wakeup_granularity
);
5927 static inline void sched_init_granularity(void)
5934 * This is how migration works:
5936 * 1) we invoke migration_cpu_stop() on the target CPU using
5938 * 2) stopper starts to run (implicitly forcing the migrated thread
5940 * 3) it checks whether the migrated task is still in the wrong runqueue.
5941 * 4) if it's in the wrong runqueue then the migration thread removes
5942 * it and puts it into the right queue.
5943 * 5) stopper completes and stop_one_cpu() returns and the migration
5948 * Change a given task's CPU affinity. Migrate the thread to a
5949 * proper CPU and schedule it away if the CPU it's executing on
5950 * is removed from the allowed bitmask.
5952 * NOTE: the caller must have a valid reference to the task, the
5953 * task must not exit() & deallocate itself prematurely. The
5954 * call is not atomic; no spinlocks may be held.
5956 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5958 unsigned long flags
;
5960 unsigned int dest_cpu
;
5963 rq
= task_rq_lock(p
, &flags
);
5965 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5970 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5971 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5976 if (p
->sched_class
->set_cpus_allowed
)
5977 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5979 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5980 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5983 /* Can the task run on the task's current CPU? If so, we're done */
5984 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5987 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5988 if (need_migrate_task(p
)) {
5989 struct migration_arg arg
= { p
, dest_cpu
};
5990 /* Need help from migration thread: drop lock and wait. */
5991 task_rq_unlock(rq
, p
, &flags
);
5992 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5993 tlb_migrate_finish(p
->mm
);
5997 task_rq_unlock(rq
, p
, &flags
);
6001 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6004 * Move (not current) task off this cpu, onto dest cpu. We're doing
6005 * this because either it can't run here any more (set_cpus_allowed()
6006 * away from this CPU, or CPU going down), or because we're
6007 * attempting to rebalance this task on exec (sched_exec).
6009 * So we race with normal scheduler movements, but that's OK, as long
6010 * as the task is no longer on this CPU.
6012 * Returns non-zero if task was successfully migrated.
6014 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6016 struct rq
*rq_dest
, *rq_src
;
6019 if (unlikely(!cpu_active(dest_cpu
)))
6022 rq_src
= cpu_rq(src_cpu
);
6023 rq_dest
= cpu_rq(dest_cpu
);
6025 raw_spin_lock(&p
->pi_lock
);
6026 double_rq_lock(rq_src
, rq_dest
);
6027 /* Already moved. */
6028 if (task_cpu(p
) != src_cpu
)
6030 /* Affinity changed (again). */
6031 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6035 * If we're not on a rq, the next wake-up will ensure we're
6039 deactivate_task(rq_src
, p
, 0);
6040 set_task_cpu(p
, dest_cpu
);
6041 activate_task(rq_dest
, p
, 0);
6042 check_preempt_curr(rq_dest
, p
, 0);
6047 double_rq_unlock(rq_src
, rq_dest
);
6048 raw_spin_unlock(&p
->pi_lock
);
6053 * migration_cpu_stop - this will be executed by a highprio stopper thread
6054 * and performs thread migration by bumping thread off CPU then
6055 * 'pushing' onto another runqueue.
6057 static int migration_cpu_stop(void *data
)
6059 struct migration_arg
*arg
= data
;
6062 * The original target cpu might have gone down and we might
6063 * be on another cpu but it doesn't matter.
6065 local_irq_disable();
6066 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
6071 #ifdef CONFIG_HOTPLUG_CPU
6074 * Ensures that the idle task is using init_mm right before its cpu goes
6077 void idle_task_exit(void)
6079 struct mm_struct
*mm
= current
->active_mm
;
6081 BUG_ON(cpu_online(smp_processor_id()));
6084 switch_mm(mm
, &init_mm
, current
);
6089 * While a dead CPU has no uninterruptible tasks queued at this point,
6090 * it might still have a nonzero ->nr_uninterruptible counter, because
6091 * for performance reasons the counter is not stricly tracking tasks to
6092 * their home CPUs. So we just add the counter to another CPU's counter,
6093 * to keep the global sum constant after CPU-down:
6095 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6097 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
6099 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6100 rq_src
->nr_uninterruptible
= 0;
6104 * remove the tasks which were accounted by rq from calc_load_tasks.
6106 static void calc_global_load_remove(struct rq
*rq
)
6108 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
6109 rq
->calc_load_active
= 0;
6113 * Migrate all tasks from the rq, sleeping tasks will be migrated by
6114 * try_to_wake_up()->select_task_rq().
6116 * Called with rq->lock held even though we'er in stop_machine() and
6117 * there's no concurrency possible, we hold the required locks anyway
6118 * because of lock validation efforts.
6120 static void migrate_tasks(unsigned int dead_cpu
)
6122 struct rq
*rq
= cpu_rq(dead_cpu
);
6123 struct task_struct
*next
, *stop
= rq
->stop
;
6127 * Fudge the rq selection such that the below task selection loop
6128 * doesn't get stuck on the currently eligible stop task.
6130 * We're currently inside stop_machine() and the rq is either stuck
6131 * in the stop_machine_cpu_stop() loop, or we're executing this code,
6132 * either way we should never end up calling schedule() until we're
6139 * There's this thread running, bail when that's the only
6142 if (rq
->nr_running
== 1)
6145 next
= pick_next_task(rq
);
6147 next
->sched_class
->put_prev_task(rq
, next
);
6149 /* Find suitable destination for @next, with force if needed. */
6150 dest_cpu
= select_fallback_rq(dead_cpu
, next
);
6151 raw_spin_unlock(&rq
->lock
);
6153 __migrate_task(next
, dead_cpu
, dest_cpu
);
6155 raw_spin_lock(&rq
->lock
);
6161 #endif /* CONFIG_HOTPLUG_CPU */
6163 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6165 static struct ctl_table sd_ctl_dir
[] = {
6167 .procname
= "sched_domain",
6173 static struct ctl_table sd_ctl_root
[] = {
6175 .procname
= "kernel",
6177 .child
= sd_ctl_dir
,
6182 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6184 struct ctl_table
*entry
=
6185 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6190 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6192 struct ctl_table
*entry
;
6195 * In the intermediate directories, both the child directory and
6196 * procname are dynamically allocated and could fail but the mode
6197 * will always be set. In the lowest directory the names are
6198 * static strings and all have proc handlers.
6200 for (entry
= *tablep
; entry
->mode
; entry
++) {
6202 sd_free_ctl_entry(&entry
->child
);
6203 if (entry
->proc_handler
== NULL
)
6204 kfree(entry
->procname
);
6212 set_table_entry(struct ctl_table
*entry
,
6213 const char *procname
, void *data
, int maxlen
,
6214 mode_t mode
, proc_handler
*proc_handler
)
6216 entry
->procname
= procname
;
6218 entry
->maxlen
= maxlen
;
6220 entry
->proc_handler
= proc_handler
;
6223 static struct ctl_table
*
6224 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6226 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6231 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6232 sizeof(long), 0644, proc_doulongvec_minmax
);
6233 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6234 sizeof(long), 0644, proc_doulongvec_minmax
);
6235 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6236 sizeof(int), 0644, proc_dointvec_minmax
);
6237 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6238 sizeof(int), 0644, proc_dointvec_minmax
);
6239 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6240 sizeof(int), 0644, proc_dointvec_minmax
);
6241 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6242 sizeof(int), 0644, proc_dointvec_minmax
);
6243 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6244 sizeof(int), 0644, proc_dointvec_minmax
);
6245 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6246 sizeof(int), 0644, proc_dointvec_minmax
);
6247 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6248 sizeof(int), 0644, proc_dointvec_minmax
);
6249 set_table_entry(&table
[9], "cache_nice_tries",
6250 &sd
->cache_nice_tries
,
6251 sizeof(int), 0644, proc_dointvec_minmax
);
6252 set_table_entry(&table
[10], "flags", &sd
->flags
,
6253 sizeof(int), 0644, proc_dointvec_minmax
);
6254 set_table_entry(&table
[11], "name", sd
->name
,
6255 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6256 /* &table[12] is terminator */
6261 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6263 struct ctl_table
*entry
, *table
;
6264 struct sched_domain
*sd
;
6265 int domain_num
= 0, i
;
6268 for_each_domain(cpu
, sd
)
6270 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6275 for_each_domain(cpu
, sd
) {
6276 snprintf(buf
, 32, "domain%d", i
);
6277 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6279 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6286 static struct ctl_table_header
*sd_sysctl_header
;
6287 static void register_sched_domain_sysctl(void)
6289 int i
, cpu_num
= num_possible_cpus();
6290 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6293 WARN_ON(sd_ctl_dir
[0].child
);
6294 sd_ctl_dir
[0].child
= entry
;
6299 for_each_possible_cpu(i
) {
6300 snprintf(buf
, 32, "cpu%d", i
);
6301 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6303 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6307 WARN_ON(sd_sysctl_header
);
6308 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6311 /* may be called multiple times per register */
6312 static void unregister_sched_domain_sysctl(void)
6314 if (sd_sysctl_header
)
6315 unregister_sysctl_table(sd_sysctl_header
);
6316 sd_sysctl_header
= NULL
;
6317 if (sd_ctl_dir
[0].child
)
6318 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6321 static void register_sched_domain_sysctl(void)
6324 static void unregister_sched_domain_sysctl(void)
6329 static void set_rq_online(struct rq
*rq
)
6332 const struct sched_class
*class;
6334 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
6337 for_each_class(class) {
6338 if (class->rq_online
)
6339 class->rq_online(rq
);
6344 static void set_rq_offline(struct rq
*rq
)
6347 const struct sched_class
*class;
6349 for_each_class(class) {
6350 if (class->rq_offline
)
6351 class->rq_offline(rq
);
6354 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
6360 * migration_call - callback that gets triggered when a CPU is added.
6361 * Here we can start up the necessary migration thread for the new CPU.
6363 static int __cpuinit
6364 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6366 int cpu
= (long)hcpu
;
6367 unsigned long flags
;
6368 struct rq
*rq
= cpu_rq(cpu
);
6370 switch (action
& ~CPU_TASKS_FROZEN
) {
6372 case CPU_UP_PREPARE
:
6373 rq
->calc_load_update
= calc_load_update
;
6377 /* Update our root-domain */
6378 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6380 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6384 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6387 #ifdef CONFIG_HOTPLUG_CPU
6389 sched_ttwu_pending();
6390 /* Update our root-domain */
6391 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6393 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
6397 BUG_ON(rq
->nr_running
!= 1); /* the migration thread */
6398 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6400 migrate_nr_uninterruptible(rq
);
6401 calc_global_load_remove(rq
);
6406 update_max_interval();
6412 * Register at high priority so that task migration (migrate_all_tasks)
6413 * happens before everything else. This has to be lower priority than
6414 * the notifier in the perf_event subsystem, though.
6416 static struct notifier_block __cpuinitdata migration_notifier
= {
6417 .notifier_call
= migration_call
,
6418 .priority
= CPU_PRI_MIGRATION
,
6421 static int __cpuinit
sched_cpu_active(struct notifier_block
*nfb
,
6422 unsigned long action
, void *hcpu
)
6424 switch (action
& ~CPU_TASKS_FROZEN
) {
6426 case CPU_DOWN_FAILED
:
6427 set_cpu_active((long)hcpu
, true);
6434 static int __cpuinit
sched_cpu_inactive(struct notifier_block
*nfb
,
6435 unsigned long action
, void *hcpu
)
6437 switch (action
& ~CPU_TASKS_FROZEN
) {
6438 case CPU_DOWN_PREPARE
:
6439 set_cpu_active((long)hcpu
, false);
6446 static int __init
migration_init(void)
6448 void *cpu
= (void *)(long)smp_processor_id();
6451 /* Initialize migration for the boot CPU */
6452 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6453 BUG_ON(err
== NOTIFY_BAD
);
6454 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6455 register_cpu_notifier(&migration_notifier
);
6457 /* Register cpu active notifiers */
6458 cpu_notifier(sched_cpu_active
, CPU_PRI_SCHED_ACTIVE
);
6459 cpu_notifier(sched_cpu_inactive
, CPU_PRI_SCHED_INACTIVE
);
6463 early_initcall(migration_init
);
6468 #ifdef CONFIG_SCHED_DEBUG
6470 static __read_mostly
int sched_domain_debug_enabled
;
6472 static int __init
sched_domain_debug_setup(char *str
)
6474 sched_domain_debug_enabled
= 1;
6478 early_param("sched_debug", sched_domain_debug_setup
);
6480 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6481 struct cpumask
*groupmask
)
6483 struct sched_group
*group
= sd
->groups
;
6486 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6487 cpumask_clear(groupmask
);
6489 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6491 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6492 printk("does not load-balance\n");
6494 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6499 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6501 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6502 printk(KERN_ERR
"ERROR: domain->span does not contain "
6505 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6506 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6510 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6514 printk(KERN_ERR
"ERROR: group is NULL\n");
6518 if (!group
->cpu_power
) {
6519 printk(KERN_CONT
"\n");
6520 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6525 if (!cpumask_weight(sched_group_cpus(group
))) {
6526 printk(KERN_CONT
"\n");
6527 printk(KERN_ERR
"ERROR: empty group\n");
6531 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6532 printk(KERN_CONT
"\n");
6533 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6537 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6539 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6541 printk(KERN_CONT
" %s", str
);
6542 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6543 printk(KERN_CONT
" (cpu_power = %d)",
6547 group
= group
->next
;
6548 } while (group
!= sd
->groups
);
6549 printk(KERN_CONT
"\n");
6551 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6552 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6555 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6556 printk(KERN_ERR
"ERROR: parent span is not a superset "
6557 "of domain->span\n");
6561 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6563 cpumask_var_t groupmask
;
6566 if (!sched_domain_debug_enabled
)
6570 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6574 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6576 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6577 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6582 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6589 free_cpumask_var(groupmask
);
6591 #else /* !CONFIG_SCHED_DEBUG */
6592 # define sched_domain_debug(sd, cpu) do { } while (0)
6593 #endif /* CONFIG_SCHED_DEBUG */
6595 static int sd_degenerate(struct sched_domain
*sd
)
6597 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6600 /* Following flags need at least 2 groups */
6601 if (sd
->flags
& (SD_LOAD_BALANCE
|
6602 SD_BALANCE_NEWIDLE
|
6606 SD_SHARE_PKG_RESOURCES
)) {
6607 if (sd
->groups
!= sd
->groups
->next
)
6611 /* Following flags don't use groups */
6612 if (sd
->flags
& (SD_WAKE_AFFINE
))
6619 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6621 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6623 if (sd_degenerate(parent
))
6626 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6629 /* Flags needing groups don't count if only 1 group in parent */
6630 if (parent
->groups
== parent
->groups
->next
) {
6631 pflags
&= ~(SD_LOAD_BALANCE
|
6632 SD_BALANCE_NEWIDLE
|
6636 SD_SHARE_PKG_RESOURCES
);
6637 if (nr_node_ids
== 1)
6638 pflags
&= ~SD_SERIALIZE
;
6640 if (~cflags
& pflags
)
6646 static void free_rootdomain(struct root_domain
*rd
)
6648 synchronize_sched();
6650 cpupri_cleanup(&rd
->cpupri
);
6652 free_cpumask_var(rd
->rto_mask
);
6653 free_cpumask_var(rd
->online
);
6654 free_cpumask_var(rd
->span
);
6658 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6660 struct root_domain
*old_rd
= NULL
;
6661 unsigned long flags
;
6663 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6668 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6671 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6674 * If we dont want to free the old_rt yet then
6675 * set old_rd to NULL to skip the freeing later
6678 if (!atomic_dec_and_test(&old_rd
->refcount
))
6682 atomic_inc(&rd
->refcount
);
6685 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6686 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6689 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6692 free_rootdomain(old_rd
);
6695 static int init_rootdomain(struct root_domain
*rd
)
6697 memset(rd
, 0, sizeof(*rd
));
6699 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
6701 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
6703 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
6706 if (cpupri_init(&rd
->cpupri
) != 0)
6711 free_cpumask_var(rd
->rto_mask
);
6713 free_cpumask_var(rd
->online
);
6715 free_cpumask_var(rd
->span
);
6720 static void init_defrootdomain(void)
6722 init_rootdomain(&def_root_domain
);
6724 atomic_set(&def_root_domain
.refcount
, 1);
6727 static struct root_domain
*alloc_rootdomain(void)
6729 struct root_domain
*rd
;
6731 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6735 if (init_rootdomain(rd
) != 0) {
6744 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6745 * hold the hotplug lock.
6748 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6750 struct rq
*rq
= cpu_rq(cpu
);
6751 struct sched_domain
*tmp
;
6753 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6754 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6756 /* Remove the sched domains which do not contribute to scheduling. */
6757 for (tmp
= sd
; tmp
; ) {
6758 struct sched_domain
*parent
= tmp
->parent
;
6762 if (sd_parent_degenerate(tmp
, parent
)) {
6763 tmp
->parent
= parent
->parent
;
6765 parent
->parent
->child
= tmp
;
6770 if (sd
&& sd_degenerate(sd
)) {
6776 sched_domain_debug(sd
, cpu
);
6778 rq_attach_root(rq
, rd
);
6779 rcu_assign_pointer(rq
->sd
, sd
);
6782 /* cpus with isolated domains */
6783 static cpumask_var_t cpu_isolated_map
;
6785 /* Setup the mask of cpus configured for isolated domains */
6786 static int __init
isolated_cpu_setup(char *str
)
6788 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6789 cpulist_parse(str
, cpu_isolated_map
);
6793 __setup("isolcpus=", isolated_cpu_setup
);
6796 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6797 * to a function which identifies what group(along with sched group) a CPU
6798 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6799 * (due to the fact that we keep track of groups covered with a struct cpumask).
6801 * init_sched_build_groups will build a circular linked list of the groups
6802 * covered by the given span, and will set each group's ->cpumask correctly,
6803 * and ->cpu_power to 0.
6806 init_sched_build_groups(const struct cpumask
*span
,
6807 const struct cpumask
*cpu_map
,
6808 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6809 struct sched_group
**sg
,
6810 struct cpumask
*tmpmask
),
6811 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6813 struct sched_group
*first
= NULL
, *last
= NULL
;
6816 cpumask_clear(covered
);
6818 for_each_cpu(i
, span
) {
6819 struct sched_group
*sg
;
6820 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6823 if (cpumask_test_cpu(i
, covered
))
6826 cpumask_clear(sched_group_cpus(sg
));
6829 for_each_cpu(j
, span
) {
6830 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6833 cpumask_set_cpu(j
, covered
);
6834 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6845 #define SD_NODES_PER_DOMAIN 16
6850 * find_next_best_node - find the next node to include in a sched_domain
6851 * @node: node whose sched_domain we're building
6852 * @used_nodes: nodes already in the sched_domain
6854 * Find the next node to include in a given scheduling domain. Simply
6855 * finds the closest node not already in the @used_nodes map.
6857 * Should use nodemask_t.
6859 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6861 int i
, n
, val
, min_val
, best_node
= 0;
6865 for (i
= 0; i
< nr_node_ids
; i
++) {
6866 /* Start at @node */
6867 n
= (node
+ i
) % nr_node_ids
;
6869 if (!nr_cpus_node(n
))
6872 /* Skip already used nodes */
6873 if (node_isset(n
, *used_nodes
))
6876 /* Simple min distance search */
6877 val
= node_distance(node
, n
);
6879 if (val
< min_val
) {
6885 node_set(best_node
, *used_nodes
);
6890 * sched_domain_node_span - get a cpumask for a node's sched_domain
6891 * @node: node whose cpumask we're constructing
6892 * @span: resulting cpumask
6894 * Given a node, construct a good cpumask for its sched_domain to span. It
6895 * should be one that prevents unnecessary balancing, but also spreads tasks
6898 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6900 nodemask_t used_nodes
;
6903 cpumask_clear(span
);
6904 nodes_clear(used_nodes
);
6906 cpumask_or(span
, span
, cpumask_of_node(node
));
6907 node_set(node
, used_nodes
);
6909 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6910 int next_node
= find_next_best_node(node
, &used_nodes
);
6912 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6915 #endif /* CONFIG_NUMA */
6917 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6920 * The cpus mask in sched_group and sched_domain hangs off the end.
6922 * ( See the the comments in include/linux/sched.h:struct sched_group
6923 * and struct sched_domain. )
6925 struct static_sched_group
{
6926 struct sched_group sg
;
6927 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6930 struct static_sched_domain
{
6931 struct sched_domain sd
;
6932 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6938 cpumask_var_t domainspan
;
6939 cpumask_var_t covered
;
6940 cpumask_var_t notcovered
;
6942 cpumask_var_t nodemask
;
6943 cpumask_var_t this_sibling_map
;
6944 cpumask_var_t this_core_map
;
6945 cpumask_var_t this_book_map
;
6946 cpumask_var_t send_covered
;
6947 cpumask_var_t tmpmask
;
6948 struct sched_group
**sched_group_nodes
;
6949 struct root_domain
*rd
;
6953 sa_sched_groups
= 0,
6959 sa_this_sibling_map
,
6961 sa_sched_group_nodes
,
6971 * SMT sched-domains:
6973 #ifdef CONFIG_SCHED_SMT
6974 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6975 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6978 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6979 struct sched_group
**sg
, struct cpumask
*unused
)
6982 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6985 #endif /* CONFIG_SCHED_SMT */
6988 * multi-core sched-domains:
6990 #ifdef CONFIG_SCHED_MC
6991 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6992 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6995 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6996 struct sched_group
**sg
, struct cpumask
*mask
)
6999 #ifdef CONFIG_SCHED_SMT
7000 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7001 group
= cpumask_first(mask
);
7006 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7009 #endif /* CONFIG_SCHED_MC */
7012 * book sched-domains:
7014 #ifdef CONFIG_SCHED_BOOK
7015 static DEFINE_PER_CPU(struct static_sched_domain
, book_domains
);
7016 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_book
);
7019 cpu_to_book_group(int cpu
, const struct cpumask
*cpu_map
,
7020 struct sched_group
**sg
, struct cpumask
*mask
)
7023 #ifdef CONFIG_SCHED_MC
7024 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7025 group
= cpumask_first(mask
);
7026 #elif defined(CONFIG_SCHED_SMT)
7027 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7028 group
= cpumask_first(mask
);
7031 *sg
= &per_cpu(sched_group_book
, group
).sg
;
7034 #endif /* CONFIG_SCHED_BOOK */
7036 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7037 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7040 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7041 struct sched_group
**sg
, struct cpumask
*mask
)
7044 #ifdef CONFIG_SCHED_BOOK
7045 cpumask_and(mask
, cpu_book_mask(cpu
), cpu_map
);
7046 group
= cpumask_first(mask
);
7047 #elif defined(CONFIG_SCHED_MC)
7048 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7049 group
= cpumask_first(mask
);
7050 #elif defined(CONFIG_SCHED_SMT)
7051 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7052 group
= cpumask_first(mask
);
7057 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7063 * The init_sched_build_groups can't handle what we want to do with node
7064 * groups, so roll our own. Now each node has its own list of groups which
7065 * gets dynamically allocated.
7067 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7068 static struct sched_group
***sched_group_nodes_bycpu
;
7070 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7071 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7073 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7074 struct sched_group
**sg
,
7075 struct cpumask
*nodemask
)
7079 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7080 group
= cpumask_first(nodemask
);
7083 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7087 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7089 struct sched_group
*sg
= group_head
;
7095 for_each_cpu(j
, sched_group_cpus(sg
)) {
7096 struct sched_domain
*sd
;
7098 sd
= &per_cpu(phys_domains
, j
).sd
;
7099 if (j
!= group_first_cpu(sd
->groups
)) {
7101 * Only add "power" once for each
7107 sg
->cpu_power
+= sd
->groups
->cpu_power
;
7110 } while (sg
!= group_head
);
7113 static int build_numa_sched_groups(struct s_data
*d
,
7114 const struct cpumask
*cpu_map
, int num
)
7116 struct sched_domain
*sd
;
7117 struct sched_group
*sg
, *prev
;
7120 cpumask_clear(d
->covered
);
7121 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
7122 if (cpumask_empty(d
->nodemask
)) {
7123 d
->sched_group_nodes
[num
] = NULL
;
7127 sched_domain_node_span(num
, d
->domainspan
);
7128 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
7130 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7133 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
7137 d
->sched_group_nodes
[num
] = sg
;
7139 for_each_cpu(j
, d
->nodemask
) {
7140 sd
= &per_cpu(node_domains
, j
).sd
;
7145 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
7147 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
7150 for (j
= 0; j
< nr_node_ids
; j
++) {
7151 n
= (num
+ j
) % nr_node_ids
;
7152 cpumask_complement(d
->notcovered
, d
->covered
);
7153 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
7154 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
7155 if (cpumask_empty(d
->tmpmask
))
7157 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
7158 if (cpumask_empty(d
->tmpmask
))
7160 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
7164 "Can not alloc domain group for node %d\n", j
);
7168 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
7169 sg
->next
= prev
->next
;
7170 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
7177 #endif /* CONFIG_NUMA */
7180 /* Free memory allocated for various sched_group structures */
7181 static void free_sched_groups(const struct cpumask
*cpu_map
,
7182 struct cpumask
*nodemask
)
7186 for_each_cpu(cpu
, cpu_map
) {
7187 struct sched_group
**sched_group_nodes
7188 = sched_group_nodes_bycpu
[cpu
];
7190 if (!sched_group_nodes
)
7193 for (i
= 0; i
< nr_node_ids
; i
++) {
7194 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7196 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7197 if (cpumask_empty(nodemask
))
7207 if (oldsg
!= sched_group_nodes
[i
])
7210 kfree(sched_group_nodes
);
7211 sched_group_nodes_bycpu
[cpu
] = NULL
;
7214 #else /* !CONFIG_NUMA */
7215 static void free_sched_groups(const struct cpumask
*cpu_map
,
7216 struct cpumask
*nodemask
)
7219 #endif /* CONFIG_NUMA */
7222 * Initialize sched groups cpu_power.
7224 * cpu_power indicates the capacity of sched group, which is used while
7225 * distributing the load between different sched groups in a sched domain.
7226 * Typically cpu_power for all the groups in a sched domain will be same unless
7227 * there are asymmetries in the topology. If there are asymmetries, group
7228 * having more cpu_power will pickup more load compared to the group having
7231 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7233 struct sched_domain
*child
;
7234 struct sched_group
*group
;
7238 WARN_ON(!sd
|| !sd
->groups
);
7240 if (cpu
!= group_first_cpu(sd
->groups
))
7243 sd
->groups
->group_weight
= cpumask_weight(sched_group_cpus(sd
->groups
));
7247 sd
->groups
->cpu_power
= 0;
7250 power
= SCHED_LOAD_SCALE
;
7251 weight
= cpumask_weight(sched_domain_span(sd
));
7253 * SMT siblings share the power of a single core.
7254 * Usually multiple threads get a better yield out of
7255 * that one core than a single thread would have,
7256 * reflect that in sd->smt_gain.
7258 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
7259 power
*= sd
->smt_gain
;
7261 power
>>= SCHED_LOAD_SHIFT
;
7263 sd
->groups
->cpu_power
+= power
;
7268 * Add cpu_power of each child group to this groups cpu_power.
7270 group
= child
->groups
;
7272 sd
->groups
->cpu_power
+= group
->cpu_power
;
7273 group
= group
->next
;
7274 } while (group
!= child
->groups
);
7278 * Initializers for schedule domains
7279 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7282 #ifdef CONFIG_SCHED_DEBUG
7283 # define SD_INIT_NAME(sd, type) sd->name = #type
7285 # define SD_INIT_NAME(sd, type) do { } while (0)
7288 #define SD_INIT(sd, type) sd_init_##type(sd)
7290 #define SD_INIT_FUNC(type) \
7291 static noinline void sd_init_##type(struct sched_domain *sd) \
7293 memset(sd, 0, sizeof(*sd)); \
7294 *sd = SD_##type##_INIT; \
7295 sd->level = SD_LV_##type; \
7296 SD_INIT_NAME(sd, type); \
7301 SD_INIT_FUNC(ALLNODES
)
7304 #ifdef CONFIG_SCHED_SMT
7305 SD_INIT_FUNC(SIBLING
)
7307 #ifdef CONFIG_SCHED_MC
7310 #ifdef CONFIG_SCHED_BOOK
7314 static int default_relax_domain_level
= -1;
7316 static int __init
setup_relax_domain_level(char *str
)
7320 val
= simple_strtoul(str
, NULL
, 0);
7321 if (val
< SD_LV_MAX
)
7322 default_relax_domain_level
= val
;
7326 __setup("relax_domain_level=", setup_relax_domain_level
);
7328 static void set_domain_attribute(struct sched_domain
*sd
,
7329 struct sched_domain_attr
*attr
)
7333 if (!attr
|| attr
->relax_domain_level
< 0) {
7334 if (default_relax_domain_level
< 0)
7337 request
= default_relax_domain_level
;
7339 request
= attr
->relax_domain_level
;
7340 if (request
< sd
->level
) {
7341 /* turn off idle balance on this domain */
7342 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7344 /* turn on idle balance on this domain */
7345 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
7349 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
7350 const struct cpumask
*cpu_map
)
7353 case sa_sched_groups
:
7354 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
7355 d
->sched_group_nodes
= NULL
;
7357 free_rootdomain(d
->rd
); /* fall through */
7359 free_cpumask_var(d
->tmpmask
); /* fall through */
7360 case sa_send_covered
:
7361 free_cpumask_var(d
->send_covered
); /* fall through */
7362 case sa_this_book_map
:
7363 free_cpumask_var(d
->this_book_map
); /* fall through */
7364 case sa_this_core_map
:
7365 free_cpumask_var(d
->this_core_map
); /* fall through */
7366 case sa_this_sibling_map
:
7367 free_cpumask_var(d
->this_sibling_map
); /* fall through */
7369 free_cpumask_var(d
->nodemask
); /* fall through */
7370 case sa_sched_group_nodes
:
7372 kfree(d
->sched_group_nodes
); /* fall through */
7374 free_cpumask_var(d
->notcovered
); /* fall through */
7376 free_cpumask_var(d
->covered
); /* fall through */
7378 free_cpumask_var(d
->domainspan
); /* fall through */
7385 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
7386 const struct cpumask
*cpu_map
)
7389 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
7391 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
7392 return sa_domainspan
;
7393 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
7395 /* Allocate the per-node list of sched groups */
7396 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
7397 sizeof(struct sched_group
*), GFP_KERNEL
);
7398 if (!d
->sched_group_nodes
) {
7399 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7400 return sa_notcovered
;
7402 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
7404 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
7405 return sa_sched_group_nodes
;
7406 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
7408 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
7409 return sa_this_sibling_map
;
7410 if (!alloc_cpumask_var(&d
->this_book_map
, GFP_KERNEL
))
7411 return sa_this_core_map
;
7412 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
7413 return sa_this_book_map
;
7414 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
7415 return sa_send_covered
;
7416 d
->rd
= alloc_rootdomain();
7418 printk(KERN_WARNING
"Cannot alloc root domain\n");
7421 return sa_rootdomain
;
7424 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
7425 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
7427 struct sched_domain
*sd
= NULL
;
7429 struct sched_domain
*parent
;
7432 if (cpumask_weight(cpu_map
) >
7433 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
7434 sd
= &per_cpu(allnodes_domains
, i
).sd
;
7435 SD_INIT(sd
, ALLNODES
);
7436 set_domain_attribute(sd
, attr
);
7437 cpumask_copy(sched_domain_span(sd
), cpu_map
);
7438 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7443 sd
= &per_cpu(node_domains
, i
).sd
;
7445 set_domain_attribute(sd
, attr
);
7446 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
7447 sd
->parent
= parent
;
7450 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
7455 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
7456 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7457 struct sched_domain
*parent
, int i
)
7459 struct sched_domain
*sd
;
7460 sd
= &per_cpu(phys_domains
, i
).sd
;
7462 set_domain_attribute(sd
, attr
);
7463 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
7464 sd
->parent
= parent
;
7467 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7471 static struct sched_domain
*__build_book_sched_domain(struct s_data
*d
,
7472 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7473 struct sched_domain
*parent
, int i
)
7475 struct sched_domain
*sd
= parent
;
7476 #ifdef CONFIG_SCHED_BOOK
7477 sd
= &per_cpu(book_domains
, i
).sd
;
7479 set_domain_attribute(sd
, attr
);
7480 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_book_mask(i
));
7481 sd
->parent
= parent
;
7483 cpu_to_book_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7488 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
7489 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7490 struct sched_domain
*parent
, int i
)
7492 struct sched_domain
*sd
= parent
;
7493 #ifdef CONFIG_SCHED_MC
7494 sd
= &per_cpu(core_domains
, i
).sd
;
7496 set_domain_attribute(sd
, attr
);
7497 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
7498 sd
->parent
= parent
;
7500 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7505 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
7506 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
7507 struct sched_domain
*parent
, int i
)
7509 struct sched_domain
*sd
= parent
;
7510 #ifdef CONFIG_SCHED_SMT
7511 sd
= &per_cpu(cpu_domains
, i
).sd
;
7512 SD_INIT(sd
, SIBLING
);
7513 set_domain_attribute(sd
, attr
);
7514 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
7515 sd
->parent
= parent
;
7517 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7522 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7523 const struct cpumask
*cpu_map
, int cpu
)
7526 #ifdef CONFIG_SCHED_SMT
7527 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7528 cpumask_and(d
->this_sibling_map
, cpu_map
,
7529 topology_thread_cpumask(cpu
));
7530 if (cpu
== cpumask_first(d
->this_sibling_map
))
7531 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7533 d
->send_covered
, d
->tmpmask
);
7536 #ifdef CONFIG_SCHED_MC
7537 case SD_LV_MC
: /* set up multi-core groups */
7538 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7539 if (cpu
== cpumask_first(d
->this_core_map
))
7540 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7542 d
->send_covered
, d
->tmpmask
);
7545 #ifdef CONFIG_SCHED_BOOK
7546 case SD_LV_BOOK
: /* set up book groups */
7547 cpumask_and(d
->this_book_map
, cpu_map
, cpu_book_mask(cpu
));
7548 if (cpu
== cpumask_first(d
->this_book_map
))
7549 init_sched_build_groups(d
->this_book_map
, cpu_map
,
7551 d
->send_covered
, d
->tmpmask
);
7554 case SD_LV_CPU
: /* set up physical groups */
7555 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7556 if (!cpumask_empty(d
->nodemask
))
7557 init_sched_build_groups(d
->nodemask
, cpu_map
,
7559 d
->send_covered
, d
->tmpmask
);
7562 case SD_LV_ALLNODES
:
7563 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7564 d
->send_covered
, d
->tmpmask
);
7573 * Build sched domains for a given set of cpus and attach the sched domains
7574 * to the individual cpus
7576 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7577 struct sched_domain_attr
*attr
)
7579 enum s_alloc alloc_state
= sa_none
;
7581 struct sched_domain
*sd
;
7587 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7588 if (alloc_state
!= sa_rootdomain
)
7590 alloc_state
= sa_sched_groups
;
7593 * Set up domains for cpus specified by the cpu_map.
7595 for_each_cpu(i
, cpu_map
) {
7596 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7599 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7600 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7601 sd
= __build_book_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7602 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7603 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7606 for_each_cpu(i
, cpu_map
) {
7607 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7608 build_sched_groups(&d
, SD_LV_BOOK
, cpu_map
, i
);
7609 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7612 /* Set up physical groups */
7613 for (i
= 0; i
< nr_node_ids
; i
++)
7614 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7617 /* Set up node groups */
7619 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7621 for (i
= 0; i
< nr_node_ids
; i
++)
7622 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7626 /* Calculate CPU power for physical packages and nodes */
7627 #ifdef CONFIG_SCHED_SMT
7628 for_each_cpu(i
, cpu_map
) {
7629 sd
= &per_cpu(cpu_domains
, i
).sd
;
7630 init_sched_groups_power(i
, sd
);
7633 #ifdef CONFIG_SCHED_MC
7634 for_each_cpu(i
, cpu_map
) {
7635 sd
= &per_cpu(core_domains
, i
).sd
;
7636 init_sched_groups_power(i
, sd
);
7639 #ifdef CONFIG_SCHED_BOOK
7640 for_each_cpu(i
, cpu_map
) {
7641 sd
= &per_cpu(book_domains
, i
).sd
;
7642 init_sched_groups_power(i
, sd
);
7646 for_each_cpu(i
, cpu_map
) {
7647 sd
= &per_cpu(phys_domains
, i
).sd
;
7648 init_sched_groups_power(i
, sd
);
7652 for (i
= 0; i
< nr_node_ids
; i
++)
7653 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7655 if (d
.sd_allnodes
) {
7656 struct sched_group
*sg
;
7658 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7660 init_numa_sched_groups_power(sg
);
7664 /* Attach the domains */
7665 for_each_cpu(i
, cpu_map
) {
7666 #ifdef CONFIG_SCHED_SMT
7667 sd
= &per_cpu(cpu_domains
, i
).sd
;
7668 #elif defined(CONFIG_SCHED_MC)
7669 sd
= &per_cpu(core_domains
, i
).sd
;
7670 #elif defined(CONFIG_SCHED_BOOK)
7671 sd
= &per_cpu(book_domains
, i
).sd
;
7673 sd
= &per_cpu(phys_domains
, i
).sd
;
7675 cpu_attach_domain(sd
, d
.rd
, i
);
7678 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7679 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7683 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7687 static int build_sched_domains(const struct cpumask
*cpu_map
)
7689 return __build_sched_domains(cpu_map
, NULL
);
7692 static cpumask_var_t
*doms_cur
; /* current sched domains */
7693 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7694 static struct sched_domain_attr
*dattr_cur
;
7695 /* attribues of custom domains in 'doms_cur' */
7698 * Special case: If a kmalloc of a doms_cur partition (array of
7699 * cpumask) fails, then fallback to a single sched domain,
7700 * as determined by the single cpumask fallback_doms.
7702 static cpumask_var_t fallback_doms
;
7705 * arch_update_cpu_topology lets virtualized architectures update the
7706 * cpu core maps. It is supposed to return 1 if the topology changed
7707 * or 0 if it stayed the same.
7709 int __attribute__((weak
)) arch_update_cpu_topology(void)
7714 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7717 cpumask_var_t
*doms
;
7719 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7722 for (i
= 0; i
< ndoms
; i
++) {
7723 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7724 free_sched_domains(doms
, i
);
7731 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7734 for (i
= 0; i
< ndoms
; i
++)
7735 free_cpumask_var(doms
[i
]);
7740 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7741 * For now this just excludes isolated cpus, but could be used to
7742 * exclude other special cases in the future.
7744 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7748 arch_update_cpu_topology();
7750 doms_cur
= alloc_sched_domains(ndoms_cur
);
7752 doms_cur
= &fallback_doms
;
7753 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7755 err
= build_sched_domains(doms_cur
[0]);
7756 register_sched_domain_sysctl();
7761 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7762 struct cpumask
*tmpmask
)
7764 free_sched_groups(cpu_map
, tmpmask
);
7768 * Detach sched domains from a group of cpus specified in cpu_map
7769 * These cpus will now be attached to the NULL domain
7771 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7773 /* Save because hotplug lock held. */
7774 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7777 for_each_cpu(i
, cpu_map
)
7778 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7779 synchronize_sched();
7780 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7783 /* handle null as "default" */
7784 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7785 struct sched_domain_attr
*new, int idx_new
)
7787 struct sched_domain_attr tmp
;
7794 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7795 new ? (new + idx_new
) : &tmp
,
7796 sizeof(struct sched_domain_attr
));
7800 * Partition sched domains as specified by the 'ndoms_new'
7801 * cpumasks in the array doms_new[] of cpumasks. This compares
7802 * doms_new[] to the current sched domain partitioning, doms_cur[].
7803 * It destroys each deleted domain and builds each new domain.
7805 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7806 * The masks don't intersect (don't overlap.) We should setup one
7807 * sched domain for each mask. CPUs not in any of the cpumasks will
7808 * not be load balanced. If the same cpumask appears both in the
7809 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7812 * The passed in 'doms_new' should be allocated using
7813 * alloc_sched_domains. This routine takes ownership of it and will
7814 * free_sched_domains it when done with it. If the caller failed the
7815 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7816 * and partition_sched_domains() will fallback to the single partition
7817 * 'fallback_doms', it also forces the domains to be rebuilt.
7819 * If doms_new == NULL it will be replaced with cpu_online_mask.
7820 * ndoms_new == 0 is a special case for destroying existing domains,
7821 * and it will not create the default domain.
7823 * Call with hotplug lock held
7825 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7826 struct sched_domain_attr
*dattr_new
)
7831 mutex_lock(&sched_domains_mutex
);
7833 /* always unregister in case we don't destroy any domains */
7834 unregister_sched_domain_sysctl();
7836 /* Let architecture update cpu core mappings. */
7837 new_topology
= arch_update_cpu_topology();
7839 n
= doms_new
? ndoms_new
: 0;
7841 /* Destroy deleted domains */
7842 for (i
= 0; i
< ndoms_cur
; i
++) {
7843 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7844 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7845 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7848 /* no match - a current sched domain not in new doms_new[] */
7849 detach_destroy_domains(doms_cur
[i
]);
7854 if (doms_new
== NULL
) {
7856 doms_new
= &fallback_doms
;
7857 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7858 WARN_ON_ONCE(dattr_new
);
7861 /* Build new domains */
7862 for (i
= 0; i
< ndoms_new
; i
++) {
7863 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7864 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7865 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7868 /* no match - add a new doms_new */
7869 __build_sched_domains(doms_new
[i
],
7870 dattr_new
? dattr_new
+ i
: NULL
);
7875 /* Remember the new sched domains */
7876 if (doms_cur
!= &fallback_doms
)
7877 free_sched_domains(doms_cur
, ndoms_cur
);
7878 kfree(dattr_cur
); /* kfree(NULL) is safe */
7879 doms_cur
= doms_new
;
7880 dattr_cur
= dattr_new
;
7881 ndoms_cur
= ndoms_new
;
7883 register_sched_domain_sysctl();
7885 mutex_unlock(&sched_domains_mutex
);
7888 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7889 static void arch_reinit_sched_domains(void)
7893 /* Destroy domains first to force the rebuild */
7894 partition_sched_domains(0, NULL
, NULL
);
7896 rebuild_sched_domains();
7900 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7902 unsigned int level
= 0;
7904 if (sscanf(buf
, "%u", &level
) != 1)
7908 * level is always be positive so don't check for
7909 * level < POWERSAVINGS_BALANCE_NONE which is 0
7910 * What happens on 0 or 1 byte write,
7911 * need to check for count as well?
7914 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7918 sched_smt_power_savings
= level
;
7920 sched_mc_power_savings
= level
;
7922 arch_reinit_sched_domains();
7927 #ifdef CONFIG_SCHED_MC
7928 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7929 struct sysdev_class_attribute
*attr
,
7932 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7934 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7935 struct sysdev_class_attribute
*attr
,
7936 const char *buf
, size_t count
)
7938 return sched_power_savings_store(buf
, count
, 0);
7940 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7941 sched_mc_power_savings_show
,
7942 sched_mc_power_savings_store
);
7945 #ifdef CONFIG_SCHED_SMT
7946 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7947 struct sysdev_class_attribute
*attr
,
7950 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7952 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7953 struct sysdev_class_attribute
*attr
,
7954 const char *buf
, size_t count
)
7956 return sched_power_savings_store(buf
, count
, 1);
7958 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7959 sched_smt_power_savings_show
,
7960 sched_smt_power_savings_store
);
7963 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7967 #ifdef CONFIG_SCHED_SMT
7969 err
= sysfs_create_file(&cls
->kset
.kobj
,
7970 &attr_sched_smt_power_savings
.attr
);
7972 #ifdef CONFIG_SCHED_MC
7973 if (!err
&& mc_capable())
7974 err
= sysfs_create_file(&cls
->kset
.kobj
,
7975 &attr_sched_mc_power_savings
.attr
);
7979 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7982 * Update cpusets according to cpu_active mask. If cpusets are
7983 * disabled, cpuset_update_active_cpus() becomes a simple wrapper
7984 * around partition_sched_domains().
7986 static int cpuset_cpu_active(struct notifier_block
*nfb
, unsigned long action
,
7989 switch (action
& ~CPU_TASKS_FROZEN
) {
7991 case CPU_DOWN_FAILED
:
7992 cpuset_update_active_cpus();
7999 static int cpuset_cpu_inactive(struct notifier_block
*nfb
, unsigned long action
,
8002 switch (action
& ~CPU_TASKS_FROZEN
) {
8003 case CPU_DOWN_PREPARE
:
8004 cpuset_update_active_cpus();
8011 static int update_runtime(struct notifier_block
*nfb
,
8012 unsigned long action
, void *hcpu
)
8014 int cpu
= (int)(long)hcpu
;
8017 case CPU_DOWN_PREPARE
:
8018 case CPU_DOWN_PREPARE_FROZEN
:
8019 disable_runtime(cpu_rq(cpu
));
8022 case CPU_DOWN_FAILED
:
8023 case CPU_DOWN_FAILED_FROZEN
:
8025 case CPU_ONLINE_FROZEN
:
8026 enable_runtime(cpu_rq(cpu
));
8034 void __init
sched_init_smp(void)
8036 cpumask_var_t non_isolated_cpus
;
8038 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8039 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8041 #if defined(CONFIG_NUMA)
8042 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8044 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8047 mutex_lock(&sched_domains_mutex
);
8048 arch_init_sched_domains(cpu_active_mask
);
8049 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8050 if (cpumask_empty(non_isolated_cpus
))
8051 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8052 mutex_unlock(&sched_domains_mutex
);
8055 hotcpu_notifier(cpuset_cpu_active
, CPU_PRI_CPUSET_ACTIVE
);
8056 hotcpu_notifier(cpuset_cpu_inactive
, CPU_PRI_CPUSET_INACTIVE
);
8058 /* RT runtime code needs to handle some hotplug events */
8059 hotcpu_notifier(update_runtime
, 0);
8063 /* Move init over to a non-isolated CPU */
8064 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8066 sched_init_granularity();
8067 free_cpumask_var(non_isolated_cpus
);
8069 init_sched_rt_class();
8072 void __init
sched_init_smp(void)
8074 sched_init_granularity();
8076 #endif /* CONFIG_SMP */
8078 const_debug
unsigned int sysctl_timer_migration
= 1;
8080 int in_sched_functions(unsigned long addr
)
8082 return in_lock_functions(addr
) ||
8083 (addr
>= (unsigned long)__sched_text_start
8084 && addr
< (unsigned long)__sched_text_end
);
8087 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8089 cfs_rq
->tasks_timeline
= RB_ROOT
;
8090 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8091 #ifdef CONFIG_FAIR_GROUP_SCHED
8093 /* allow initial update_cfs_load() to truncate */
8095 cfs_rq
->load_stamp
= 1;
8098 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8101 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8103 struct rt_prio_array
*array
;
8106 array
= &rt_rq
->active
;
8107 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8108 INIT_LIST_HEAD(array
->queue
+ i
);
8109 __clear_bit(i
, array
->bitmap
);
8111 /* delimiter for bitsearch: */
8112 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8114 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8115 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8117 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8121 rt_rq
->rt_nr_migratory
= 0;
8122 rt_rq
->overloaded
= 0;
8123 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
8127 rt_rq
->rt_throttled
= 0;
8128 rt_rq
->rt_runtime
= 0;
8129 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
8131 #ifdef CONFIG_RT_GROUP_SCHED
8132 rt_rq
->rt_nr_boosted
= 0;
8137 #ifdef CONFIG_FAIR_GROUP_SCHED
8138 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8139 struct sched_entity
*se
, int cpu
,
8140 struct sched_entity
*parent
)
8142 struct rq
*rq
= cpu_rq(cpu
);
8143 tg
->cfs_rq
[cpu
] = cfs_rq
;
8144 init_cfs_rq(cfs_rq
, rq
);
8148 /* se could be NULL for root_task_group */
8153 se
->cfs_rq
= &rq
->cfs
;
8155 se
->cfs_rq
= parent
->my_q
;
8158 update_load_set(&se
->load
, 0);
8159 se
->parent
= parent
;
8163 #ifdef CONFIG_RT_GROUP_SCHED
8164 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8165 struct sched_rt_entity
*rt_se
, int cpu
,
8166 struct sched_rt_entity
*parent
)
8168 struct rq
*rq
= cpu_rq(cpu
);
8170 tg
->rt_rq
[cpu
] = rt_rq
;
8171 init_rt_rq(rt_rq
, rq
);
8173 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8175 tg
->rt_se
[cpu
] = rt_se
;
8180 rt_se
->rt_rq
= &rq
->rt
;
8182 rt_se
->rt_rq
= parent
->my_q
;
8184 rt_se
->my_q
= rt_rq
;
8185 rt_se
->parent
= parent
;
8186 INIT_LIST_HEAD(&rt_se
->run_list
);
8190 void __init
sched_init(void)
8193 unsigned long alloc_size
= 0, ptr
;
8195 #ifdef CONFIG_FAIR_GROUP_SCHED
8196 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8198 #ifdef CONFIG_RT_GROUP_SCHED
8199 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8201 #ifdef CONFIG_CPUMASK_OFFSTACK
8202 alloc_size
+= num_possible_cpus() * cpumask_size();
8205 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
8207 #ifdef CONFIG_FAIR_GROUP_SCHED
8208 root_task_group
.se
= (struct sched_entity
**)ptr
;
8209 ptr
+= nr_cpu_ids
* sizeof(void **);
8211 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8212 ptr
+= nr_cpu_ids
* sizeof(void **);
8214 #endif /* CONFIG_FAIR_GROUP_SCHED */
8215 #ifdef CONFIG_RT_GROUP_SCHED
8216 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8217 ptr
+= nr_cpu_ids
* sizeof(void **);
8219 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8220 ptr
+= nr_cpu_ids
* sizeof(void **);
8222 #endif /* CONFIG_RT_GROUP_SCHED */
8223 #ifdef CONFIG_CPUMASK_OFFSTACK
8224 for_each_possible_cpu(i
) {
8225 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8226 ptr
+= cpumask_size();
8228 #endif /* CONFIG_CPUMASK_OFFSTACK */
8232 init_defrootdomain();
8235 init_rt_bandwidth(&def_rt_bandwidth
,
8236 global_rt_period(), global_rt_runtime());
8238 #ifdef CONFIG_RT_GROUP_SCHED
8239 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8240 global_rt_period(), global_rt_runtime());
8241 #endif /* CONFIG_RT_GROUP_SCHED */
8243 #ifdef CONFIG_CGROUP_SCHED
8244 list_add(&root_task_group
.list
, &task_groups
);
8245 INIT_LIST_HEAD(&root_task_group
.children
);
8246 autogroup_init(&init_task
);
8247 #endif /* CONFIG_CGROUP_SCHED */
8249 for_each_possible_cpu(i
) {
8253 raw_spin_lock_init(&rq
->lock
);
8255 rq
->calc_load_active
= 0;
8256 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
8257 init_cfs_rq(&rq
->cfs
, rq
);
8258 init_rt_rq(&rq
->rt
, rq
);
8259 #ifdef CONFIG_FAIR_GROUP_SCHED
8260 root_task_group
.shares
= root_task_group_load
;
8261 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8263 * How much cpu bandwidth does root_task_group get?
8265 * In case of task-groups formed thr' the cgroup filesystem, it
8266 * gets 100% of the cpu resources in the system. This overall
8267 * system cpu resource is divided among the tasks of
8268 * root_task_group and its child task-groups in a fair manner,
8269 * based on each entity's (task or task-group's) weight
8270 * (se->load.weight).
8272 * In other words, if root_task_group has 10 tasks of weight
8273 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8274 * then A0's share of the cpu resource is:
8276 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8278 * We achieve this by letting root_task_group's tasks sit
8279 * directly in rq->cfs (i.e root_task_group->se[] = NULL).
8281 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, NULL
);
8282 #endif /* CONFIG_FAIR_GROUP_SCHED */
8284 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8285 #ifdef CONFIG_RT_GROUP_SCHED
8286 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8287 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, NULL
);
8290 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8291 rq
->cpu_load
[j
] = 0;
8293 rq
->last_load_update_tick
= jiffies
;
8298 rq
->cpu_power
= SCHED_LOAD_SCALE
;
8299 rq
->post_schedule
= 0;
8300 rq
->active_balance
= 0;
8301 rq
->next_balance
= jiffies
;
8306 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
8307 rq_attach_root(rq
, &def_root_domain
);
8309 rq
->nohz_balance_kick
= 0;
8310 init_sched_softirq_csd(&per_cpu(remote_sched_softirq_cb
, i
));
8314 atomic_set(&rq
->nr_iowait
, 0);
8317 set_load_weight(&init_task
);
8319 #ifdef CONFIG_PREEMPT_NOTIFIERS
8320 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8324 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8327 #ifdef CONFIG_RT_MUTEXES
8328 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8332 * The boot idle thread does lazy MMU switching as well:
8334 atomic_inc(&init_mm
.mm_count
);
8335 enter_lazy_tlb(&init_mm
, current
);
8338 * Make us the idle thread. Technically, schedule() should not be
8339 * called from this thread, however somewhere below it might be,
8340 * but because we are the idle thread, we just pick up running again
8341 * when this runqueue becomes "idle".
8343 init_idle(current
, smp_processor_id());
8345 calc_load_update
= jiffies
+ LOAD_FREQ
;
8348 * During early bootup we pretend to be a normal task:
8350 current
->sched_class
= &fair_sched_class
;
8352 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
8353 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
8356 zalloc_cpumask_var(&nohz
.idle_cpus_mask
, GFP_NOWAIT
);
8357 alloc_cpumask_var(&nohz
.grp_idle_mask
, GFP_NOWAIT
);
8358 atomic_set(&nohz
.load_balancer
, nr_cpu_ids
);
8359 atomic_set(&nohz
.first_pick_cpu
, nr_cpu_ids
);
8360 atomic_set(&nohz
.second_pick_cpu
, nr_cpu_ids
);
8362 /* May be allocated at isolcpus cmdline parse time */
8363 if (cpu_isolated_map
== NULL
)
8364 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
8367 scheduler_running
= 1;
8370 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8371 static inline int preempt_count_equals(int preempt_offset
)
8373 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
8375 return (nested
== preempt_offset
);
8378 void __might_sleep(const char *file
, int line
, int preempt_offset
)
8381 static unsigned long prev_jiffy
; /* ratelimiting */
8383 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
8384 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8386 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8388 prev_jiffy
= jiffies
;
8391 "BUG: sleeping function called from invalid context at %s:%d\n",
8394 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8395 in_atomic(), irqs_disabled(),
8396 current
->pid
, current
->comm
);
8398 debug_show_held_locks(current
);
8399 if (irqs_disabled())
8400 print_irqtrace_events(current
);
8404 EXPORT_SYMBOL(__might_sleep
);
8407 #ifdef CONFIG_MAGIC_SYSRQ
8408 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8410 const struct sched_class
*prev_class
= p
->sched_class
;
8411 int old_prio
= p
->prio
;
8416 deactivate_task(rq
, p
, 0);
8417 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8419 activate_task(rq
, p
, 0);
8420 resched_task(rq
->curr
);
8423 check_class_changed(rq
, p
, prev_class
, old_prio
);
8426 void normalize_rt_tasks(void)
8428 struct task_struct
*g
, *p
;
8429 unsigned long flags
;
8432 read_lock_irqsave(&tasklist_lock
, flags
);
8433 do_each_thread(g
, p
) {
8435 * Only normalize user tasks:
8440 p
->se
.exec_start
= 0;
8441 #ifdef CONFIG_SCHEDSTATS
8442 p
->se
.statistics
.wait_start
= 0;
8443 p
->se
.statistics
.sleep_start
= 0;
8444 p
->se
.statistics
.block_start
= 0;
8449 * Renice negative nice level userspace
8452 if (TASK_NICE(p
) < 0 && p
->mm
)
8453 set_user_nice(p
, 0);
8457 raw_spin_lock(&p
->pi_lock
);
8458 rq
= __task_rq_lock(p
);
8460 normalize_task(rq
, p
);
8462 __task_rq_unlock(rq
);
8463 raw_spin_unlock(&p
->pi_lock
);
8464 } while_each_thread(g
, p
);
8466 read_unlock_irqrestore(&tasklist_lock
, flags
);
8469 #endif /* CONFIG_MAGIC_SYSRQ */
8471 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
8473 * These functions are only useful for the IA64 MCA handling, or kdb.
8475 * They can only be called when the whole system has been
8476 * stopped - every CPU needs to be quiescent, and no scheduling
8477 * activity can take place. Using them for anything else would
8478 * be a serious bug, and as a result, they aren't even visible
8479 * under any other configuration.
8483 * curr_task - return the current task for a given cpu.
8484 * @cpu: the processor in question.
8486 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8488 struct task_struct
*curr_task(int cpu
)
8490 return cpu_curr(cpu
);
8493 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
8497 * set_curr_task - set the current task for a given cpu.
8498 * @cpu: the processor in question.
8499 * @p: the task pointer to set.
8501 * Description: This function must only be used when non-maskable interrupts
8502 * are serviced on a separate stack. It allows the architecture to switch the
8503 * notion of the current task on a cpu in a non-blocking manner. This function
8504 * must be called with all CPU's synchronized, and interrupts disabled, the
8505 * and caller must save the original value of the current task (see
8506 * curr_task() above) and restore that value before reenabling interrupts and
8507 * re-starting the system.
8509 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8511 void set_curr_task(int cpu
, struct task_struct
*p
)
8518 #ifdef CONFIG_FAIR_GROUP_SCHED
8519 static void free_fair_sched_group(struct task_group
*tg
)
8523 for_each_possible_cpu(i
) {
8525 kfree(tg
->cfs_rq
[i
]);
8535 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8537 struct cfs_rq
*cfs_rq
;
8538 struct sched_entity
*se
;
8541 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8544 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8548 tg
->shares
= NICE_0_LOAD
;
8550 for_each_possible_cpu(i
) {
8551 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8552 GFP_KERNEL
, cpu_to_node(i
));
8556 se
= kzalloc_node(sizeof(struct sched_entity
),
8557 GFP_KERNEL
, cpu_to_node(i
));
8561 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, parent
->se
[i
]);
8572 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8574 struct rq
*rq
= cpu_rq(cpu
);
8575 unsigned long flags
;
8578 * Only empty task groups can be destroyed; so we can speculatively
8579 * check on_list without danger of it being re-added.
8581 if (!tg
->cfs_rq
[cpu
]->on_list
)
8584 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8585 list_del_leaf_cfs_rq(tg
->cfs_rq
[cpu
]);
8586 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8588 #else /* !CONFG_FAIR_GROUP_SCHED */
8589 static inline void free_fair_sched_group(struct task_group
*tg
)
8594 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8599 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8602 #endif /* CONFIG_FAIR_GROUP_SCHED */
8604 #ifdef CONFIG_RT_GROUP_SCHED
8605 static void free_rt_sched_group(struct task_group
*tg
)
8609 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8611 for_each_possible_cpu(i
) {
8613 kfree(tg
->rt_rq
[i
]);
8615 kfree(tg
->rt_se
[i
]);
8623 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8625 struct rt_rq
*rt_rq
;
8626 struct sched_rt_entity
*rt_se
;
8630 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8633 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8637 init_rt_bandwidth(&tg
->rt_bandwidth
,
8638 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8640 for_each_possible_cpu(i
) {
8643 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8644 GFP_KERNEL
, cpu_to_node(i
));
8648 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8649 GFP_KERNEL
, cpu_to_node(i
));
8653 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, parent
->rt_se
[i
]);
8663 #else /* !CONFIG_RT_GROUP_SCHED */
8664 static inline void free_rt_sched_group(struct task_group
*tg
)
8669 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8673 #endif /* CONFIG_RT_GROUP_SCHED */
8675 #ifdef CONFIG_CGROUP_SCHED
8676 static void free_sched_group(struct task_group
*tg
)
8678 free_fair_sched_group(tg
);
8679 free_rt_sched_group(tg
);
8684 /* allocate runqueue etc for a new task group */
8685 struct task_group
*sched_create_group(struct task_group
*parent
)
8687 struct task_group
*tg
;
8688 unsigned long flags
;
8690 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8692 return ERR_PTR(-ENOMEM
);
8694 if (!alloc_fair_sched_group(tg
, parent
))
8697 if (!alloc_rt_sched_group(tg
, parent
))
8700 spin_lock_irqsave(&task_group_lock
, flags
);
8701 list_add_rcu(&tg
->list
, &task_groups
);
8703 WARN_ON(!parent
); /* root should already exist */
8705 tg
->parent
= parent
;
8706 INIT_LIST_HEAD(&tg
->children
);
8707 list_add_rcu(&tg
->siblings
, &parent
->children
);
8708 spin_unlock_irqrestore(&task_group_lock
, flags
);
8713 free_sched_group(tg
);
8714 return ERR_PTR(-ENOMEM
);
8717 /* rcu callback to free various structures associated with a task group */
8718 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8720 /* now it should be safe to free those cfs_rqs */
8721 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8724 /* Destroy runqueue etc associated with a task group */
8725 void sched_destroy_group(struct task_group
*tg
)
8727 unsigned long flags
;
8730 /* end participation in shares distribution */
8731 for_each_possible_cpu(i
)
8732 unregister_fair_sched_group(tg
, i
);
8734 spin_lock_irqsave(&task_group_lock
, flags
);
8735 list_del_rcu(&tg
->list
);
8736 list_del_rcu(&tg
->siblings
);
8737 spin_unlock_irqrestore(&task_group_lock
, flags
);
8739 /* wait for possible concurrent references to cfs_rqs complete */
8740 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8743 /* change task's runqueue when it moves between groups.
8744 * The caller of this function should have put the task in its new group
8745 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8746 * reflect its new group.
8748 void sched_move_task(struct task_struct
*tsk
)
8751 unsigned long flags
;
8754 rq
= task_rq_lock(tsk
, &flags
);
8756 running
= task_current(rq
, tsk
);
8760 dequeue_task(rq
, tsk
, 0);
8761 if (unlikely(running
))
8762 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8764 #ifdef CONFIG_FAIR_GROUP_SCHED
8765 if (tsk
->sched_class
->task_move_group
)
8766 tsk
->sched_class
->task_move_group(tsk
, on_rq
);
8769 set_task_rq(tsk
, task_cpu(tsk
));
8771 if (unlikely(running
))
8772 tsk
->sched_class
->set_curr_task(rq
);
8774 enqueue_task(rq
, tsk
, 0);
8776 task_rq_unlock(rq
, tsk
, &flags
);
8778 #endif /* CONFIG_CGROUP_SCHED */
8780 #ifdef CONFIG_FAIR_GROUP_SCHED
8781 static DEFINE_MUTEX(shares_mutex
);
8783 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8786 unsigned long flags
;
8789 * We can't change the weight of the root cgroup.
8794 if (shares
< MIN_SHARES
)
8795 shares
= MIN_SHARES
;
8796 else if (shares
> MAX_SHARES
)
8797 shares
= MAX_SHARES
;
8799 mutex_lock(&shares_mutex
);
8800 if (tg
->shares
== shares
)
8803 tg
->shares
= shares
;
8804 for_each_possible_cpu(i
) {
8805 struct rq
*rq
= cpu_rq(i
);
8806 struct sched_entity
*se
;
8809 /* Propagate contribution to hierarchy */
8810 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8811 for_each_sched_entity(se
)
8812 update_cfs_shares(group_cfs_rq(se
));
8813 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8817 mutex_unlock(&shares_mutex
);
8821 unsigned long sched_group_shares(struct task_group
*tg
)
8827 #ifdef CONFIG_RT_GROUP_SCHED
8829 * Ensure that the real time constraints are schedulable.
8831 static DEFINE_MUTEX(rt_constraints_mutex
);
8833 static unsigned long to_ratio(u64 period
, u64 runtime
)
8835 if (runtime
== RUNTIME_INF
)
8838 return div64_u64(runtime
<< 20, period
);
8841 /* Must be called with tasklist_lock held */
8842 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8844 struct task_struct
*g
, *p
;
8846 do_each_thread(g
, p
) {
8847 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8849 } while_each_thread(g
, p
);
8854 struct rt_schedulable_data
{
8855 struct task_group
*tg
;
8860 static int tg_schedulable(struct task_group
*tg
, void *data
)
8862 struct rt_schedulable_data
*d
= data
;
8863 struct task_group
*child
;
8864 unsigned long total
, sum
= 0;
8865 u64 period
, runtime
;
8867 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8868 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8871 period
= d
->rt_period
;
8872 runtime
= d
->rt_runtime
;
8876 * Cannot have more runtime than the period.
8878 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8882 * Ensure we don't starve existing RT tasks.
8884 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8887 total
= to_ratio(period
, runtime
);
8890 * Nobody can have more than the global setting allows.
8892 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8896 * The sum of our children's runtime should not exceed our own.
8898 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8899 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8900 runtime
= child
->rt_bandwidth
.rt_runtime
;
8902 if (child
== d
->tg
) {
8903 period
= d
->rt_period
;
8904 runtime
= d
->rt_runtime
;
8907 sum
+= to_ratio(period
, runtime
);
8916 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8918 struct rt_schedulable_data data
= {
8920 .rt_period
= period
,
8921 .rt_runtime
= runtime
,
8924 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8927 static int tg_set_bandwidth(struct task_group
*tg
,
8928 u64 rt_period
, u64 rt_runtime
)
8932 mutex_lock(&rt_constraints_mutex
);
8933 read_lock(&tasklist_lock
);
8934 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8938 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8939 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8940 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8942 for_each_possible_cpu(i
) {
8943 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8945 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8946 rt_rq
->rt_runtime
= rt_runtime
;
8947 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8949 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8951 read_unlock(&tasklist_lock
);
8952 mutex_unlock(&rt_constraints_mutex
);
8957 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8959 u64 rt_runtime
, rt_period
;
8961 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8962 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8963 if (rt_runtime_us
< 0)
8964 rt_runtime
= RUNTIME_INF
;
8966 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8969 long sched_group_rt_runtime(struct task_group
*tg
)
8973 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8976 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8977 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8978 return rt_runtime_us
;
8981 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8983 u64 rt_runtime
, rt_period
;
8985 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8986 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8991 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8994 long sched_group_rt_period(struct task_group
*tg
)
8998 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8999 do_div(rt_period_us
, NSEC_PER_USEC
);
9000 return rt_period_us
;
9003 static int sched_rt_global_constraints(void)
9005 u64 runtime
, period
;
9008 if (sysctl_sched_rt_period
<= 0)
9011 runtime
= global_rt_runtime();
9012 period
= global_rt_period();
9015 * Sanity check on the sysctl variables.
9017 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9020 mutex_lock(&rt_constraints_mutex
);
9021 read_lock(&tasklist_lock
);
9022 ret
= __rt_schedulable(NULL
, 0, 0);
9023 read_unlock(&tasklist_lock
);
9024 mutex_unlock(&rt_constraints_mutex
);
9029 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9031 /* Don't accept realtime tasks when there is no way for them to run */
9032 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9038 #else /* !CONFIG_RT_GROUP_SCHED */
9039 static int sched_rt_global_constraints(void)
9041 unsigned long flags
;
9044 if (sysctl_sched_rt_period
<= 0)
9048 * There's always some RT tasks in the root group
9049 * -- migration, kstopmachine etc..
9051 if (sysctl_sched_rt_runtime
== 0)
9054 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9055 for_each_possible_cpu(i
) {
9056 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9058 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
9059 rt_rq
->rt_runtime
= global_rt_runtime();
9060 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
9062 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9066 #endif /* CONFIG_RT_GROUP_SCHED */
9068 int sched_rt_handler(struct ctl_table
*table
, int write
,
9069 void __user
*buffer
, size_t *lenp
,
9073 int old_period
, old_runtime
;
9074 static DEFINE_MUTEX(mutex
);
9077 old_period
= sysctl_sched_rt_period
;
9078 old_runtime
= sysctl_sched_rt_runtime
;
9080 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
9082 if (!ret
&& write
) {
9083 ret
= sched_rt_global_constraints();
9085 sysctl_sched_rt_period
= old_period
;
9086 sysctl_sched_rt_runtime
= old_runtime
;
9088 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9089 def_rt_bandwidth
.rt_period
=
9090 ns_to_ktime(global_rt_period());
9093 mutex_unlock(&mutex
);
9098 #ifdef CONFIG_CGROUP_SCHED
9100 /* return corresponding task_group object of a cgroup */
9101 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9103 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9104 struct task_group
, css
);
9107 static struct cgroup_subsys_state
*
9108 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9110 struct task_group
*tg
, *parent
;
9112 if (!cgrp
->parent
) {
9113 /* This is early initialization for the top cgroup */
9114 return &root_task_group
.css
;
9117 parent
= cgroup_tg(cgrp
->parent
);
9118 tg
= sched_create_group(parent
);
9120 return ERR_PTR(-ENOMEM
);
9126 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9128 struct task_group
*tg
= cgroup_tg(cgrp
);
9130 sched_destroy_group(tg
);
9134 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
9136 #ifdef CONFIG_RT_GROUP_SCHED
9137 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9140 /* We don't support RT-tasks being in separate groups */
9141 if (tsk
->sched_class
!= &fair_sched_class
)
9148 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9149 struct task_struct
*tsk
, bool threadgroup
)
9151 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
9155 struct task_struct
*c
;
9157 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
9158 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
9170 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9171 struct cgroup
*old_cont
, struct task_struct
*tsk
,
9174 sched_move_task(tsk
);
9176 struct task_struct
*c
;
9178 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
9186 cpu_cgroup_exit(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9187 struct cgroup
*old_cgrp
, struct task_struct
*task
)
9190 * cgroup_exit() is called in the copy_process() failure path.
9191 * Ignore this case since the task hasn't ran yet, this avoids
9192 * trying to poke a half freed task state from generic code.
9194 if (!(task
->flags
& PF_EXITING
))
9197 sched_move_task(task
);
9200 #ifdef CONFIG_FAIR_GROUP_SCHED
9201 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9204 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9207 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9209 struct task_group
*tg
= cgroup_tg(cgrp
);
9211 return (u64
) tg
->shares
;
9213 #endif /* CONFIG_FAIR_GROUP_SCHED */
9215 #ifdef CONFIG_RT_GROUP_SCHED
9216 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9219 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9222 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9224 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9227 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9230 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9233 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9235 return sched_group_rt_period(cgroup_tg(cgrp
));
9237 #endif /* CONFIG_RT_GROUP_SCHED */
9239 static struct cftype cpu_files
[] = {
9240 #ifdef CONFIG_FAIR_GROUP_SCHED
9243 .read_u64
= cpu_shares_read_u64
,
9244 .write_u64
= cpu_shares_write_u64
,
9247 #ifdef CONFIG_RT_GROUP_SCHED
9249 .name
= "rt_runtime_us",
9250 .read_s64
= cpu_rt_runtime_read
,
9251 .write_s64
= cpu_rt_runtime_write
,
9254 .name
= "rt_period_us",
9255 .read_u64
= cpu_rt_period_read_uint
,
9256 .write_u64
= cpu_rt_period_write_uint
,
9261 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9263 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9266 struct cgroup_subsys cpu_cgroup_subsys
= {
9268 .create
= cpu_cgroup_create
,
9269 .destroy
= cpu_cgroup_destroy
,
9270 .can_attach
= cpu_cgroup_can_attach
,
9271 .attach
= cpu_cgroup_attach
,
9272 .exit
= cpu_cgroup_exit
,
9273 .populate
= cpu_cgroup_populate
,
9274 .subsys_id
= cpu_cgroup_subsys_id
,
9278 #endif /* CONFIG_CGROUP_SCHED */
9280 #ifdef CONFIG_CGROUP_CPUACCT
9283 * CPU accounting code for task groups.
9285 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9286 * (balbir@in.ibm.com).
9289 /* track cpu usage of a group of tasks and its child groups */
9291 struct cgroup_subsys_state css
;
9292 /* cpuusage holds pointer to a u64-type object on every cpu */
9293 u64 __percpu
*cpuusage
;
9294 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
9295 struct cpuacct
*parent
;
9298 struct cgroup_subsys cpuacct_subsys
;
9300 /* return cpu accounting group corresponding to this container */
9301 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9303 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9304 struct cpuacct
, css
);
9307 /* return cpu accounting group to which this task belongs */
9308 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9310 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9311 struct cpuacct
, css
);
9314 /* create a new cpu accounting group */
9315 static struct cgroup_subsys_state
*cpuacct_create(
9316 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9318 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9324 ca
->cpuusage
= alloc_percpu(u64
);
9328 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9329 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
9330 goto out_free_counters
;
9333 ca
->parent
= cgroup_ca(cgrp
->parent
);
9339 percpu_counter_destroy(&ca
->cpustat
[i
]);
9340 free_percpu(ca
->cpuusage
);
9344 return ERR_PTR(-ENOMEM
);
9347 /* destroy an existing cpu accounting group */
9349 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9351 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9354 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
9355 percpu_counter_destroy(&ca
->cpustat
[i
]);
9356 free_percpu(ca
->cpuusage
);
9360 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
9362 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9365 #ifndef CONFIG_64BIT
9367 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
9369 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9371 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9379 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
9381 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9383 #ifndef CONFIG_64BIT
9385 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
9387 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
9389 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
9395 /* return total cpu usage (in nanoseconds) of a group */
9396 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9398 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9399 u64 totalcpuusage
= 0;
9402 for_each_present_cpu(i
)
9403 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
9405 return totalcpuusage
;
9408 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9411 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9420 for_each_present_cpu(i
)
9421 cpuacct_cpuusage_write(ca
, i
, 0);
9427 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
9430 struct cpuacct
*ca
= cgroup_ca(cgroup
);
9434 for_each_present_cpu(i
) {
9435 percpu
= cpuacct_cpuusage_read(ca
, i
);
9436 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
9438 seq_printf(m
, "\n");
9442 static const char *cpuacct_stat_desc
[] = {
9443 [CPUACCT_STAT_USER
] = "user",
9444 [CPUACCT_STAT_SYSTEM
] = "system",
9447 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
9448 struct cgroup_map_cb
*cb
)
9450 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9453 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
9454 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
9455 val
= cputime64_to_clock_t(val
);
9456 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
9461 static struct cftype files
[] = {
9464 .read_u64
= cpuusage_read
,
9465 .write_u64
= cpuusage_write
,
9468 .name
= "usage_percpu",
9469 .read_seq_string
= cpuacct_percpu_seq_read
,
9473 .read_map
= cpuacct_stats_show
,
9477 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9479 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9483 * charge this task's execution time to its accounting group.
9485 * called with rq->lock held.
9487 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9492 if (unlikely(!cpuacct_subsys
.active
))
9495 cpu
= task_cpu(tsk
);
9501 for (; ca
; ca
= ca
->parent
) {
9502 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9503 *cpuusage
+= cputime
;
9510 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9511 * in cputime_t units. As a result, cpuacct_update_stats calls
9512 * percpu_counter_add with values large enough to always overflow the
9513 * per cpu batch limit causing bad SMP scalability.
9515 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9516 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9517 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9520 #define CPUACCT_BATCH \
9521 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9523 #define CPUACCT_BATCH 0
9527 * Charge the system/user time to the task's accounting group.
9529 static void cpuacct_update_stats(struct task_struct
*tsk
,
9530 enum cpuacct_stat_index idx
, cputime_t val
)
9533 int batch
= CPUACCT_BATCH
;
9535 if (unlikely(!cpuacct_subsys
.active
))
9542 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9548 struct cgroup_subsys cpuacct_subsys
= {
9550 .create
= cpuacct_create
,
9551 .destroy
= cpuacct_destroy
,
9552 .populate
= cpuacct_populate
,
9553 .subsys_id
= cpuacct_subsys_id
,
9555 #endif /* CONFIG_CGROUP_CPUACCT */