4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy
)
125 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
130 static inline int task_has_rt_policy(struct task_struct
*p
)
132 return rt_policy(p
->policy
);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array
{
139 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
140 struct list_head queue
[MAX_RT_PRIO
];
143 struct rt_bandwidth
{
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock
;
148 struct hrtimer rt_period_timer
;
151 static struct rt_bandwidth def_rt_bandwidth
;
153 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
155 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
157 struct rt_bandwidth
*rt_b
=
158 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
164 now
= hrtimer_cb_get_time(timer
);
165 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
170 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
173 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
177 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
179 rt_b
->rt_period
= ns_to_ktime(period
);
180 rt_b
->rt_runtime
= runtime
;
182 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
184 hrtimer_init(&rt_b
->rt_period_timer
,
185 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
186 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime
>= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
198 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
201 if (hrtimer_active(&rt_b
->rt_period_timer
))
204 raw_spin_lock(&rt_b
->rt_runtime_lock
);
209 if (hrtimer_active(&rt_b
->rt_period_timer
))
212 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
213 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
215 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
216 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
217 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
218 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
219 HRTIMER_MODE_ABS_PINNED
, 0);
221 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
227 hrtimer_cancel(&rt_b
->rt_period_timer
);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex
);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups
);
245 /* task group related information */
247 struct cgroup_subsys_state css
;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity
**se
;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq
**cfs_rq
;
254 unsigned long shares
;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity
**rt_se
;
259 struct rt_rq
**rt_rq
;
261 struct rt_bandwidth rt_bandwidth
;
265 struct list_head list
;
267 struct task_group
*parent
;
268 struct list_head siblings
;
269 struct list_head children
;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock
);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group
.children
);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group
;
309 /* return group to which a task belongs */
310 static inline struct task_group
*task_group(struct task_struct
*p
)
312 struct task_group
*tg
;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
316 struct task_group
, css
);
318 tg
= &init_task_group
;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
327 * Strictly speaking this rcu_read_lock() is not needed since the
328 * task_group is tied to the cgroup, which in turn can never go away
329 * as long as there are tasks attached to it.
331 * However since task_group() uses task_subsys_state() which is an
332 * rcu_dereference() user, this quiets CONFIG_PROVE_RCU.
335 #ifdef CONFIG_FAIR_GROUP_SCHED
336 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
337 p
->se
.parent
= task_group(p
)->se
[cpu
];
340 #ifdef CONFIG_RT_GROUP_SCHED
341 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
342 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
349 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
350 static inline struct task_group
*task_group(struct task_struct
*p
)
355 #endif /* CONFIG_CGROUP_SCHED */
357 /* CFS-related fields in a runqueue */
359 struct load_weight load
;
360 unsigned long nr_running
;
365 struct rb_root tasks_timeline
;
366 struct rb_node
*rb_leftmost
;
368 struct list_head tasks
;
369 struct list_head
*balance_iterator
;
372 * 'curr' points to currently running entity on this cfs_rq.
373 * It is set to NULL otherwise (i.e when none are currently running).
375 struct sched_entity
*curr
, *next
, *last
;
377 unsigned int nr_spread_over
;
379 #ifdef CONFIG_FAIR_GROUP_SCHED
380 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
383 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
384 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
385 * (like users, containers etc.)
387 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
388 * list is used during load balance.
390 struct list_head leaf_cfs_rq_list
;
391 struct task_group
*tg
; /* group that "owns" this runqueue */
395 * the part of load.weight contributed by tasks
397 unsigned long task_weight
;
400 * h_load = weight * f(tg)
402 * Where f(tg) is the recursive weight fraction assigned to
405 unsigned long h_load
;
408 * this cpu's part of tg->shares
410 unsigned long shares
;
413 * load.weight at the time we set shares
415 unsigned long rq_weight
;
420 /* Real-Time classes' related field in a runqueue: */
422 struct rt_prio_array active
;
423 unsigned long rt_nr_running
;
424 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
426 int curr
; /* highest queued rt task prio */
428 int next
; /* next highest */
433 unsigned long rt_nr_migratory
;
434 unsigned long rt_nr_total
;
436 struct plist_head pushable_tasks
;
441 /* Nests inside the rq lock: */
442 raw_spinlock_t rt_runtime_lock
;
444 #ifdef CONFIG_RT_GROUP_SCHED
445 unsigned long rt_nr_boosted
;
448 struct list_head leaf_rt_rq_list
;
449 struct task_group
*tg
;
456 * We add the notion of a root-domain which will be used to define per-domain
457 * variables. Each exclusive cpuset essentially defines an island domain by
458 * fully partitioning the member cpus from any other cpuset. Whenever a new
459 * exclusive cpuset is created, we also create and attach a new root-domain
466 cpumask_var_t online
;
469 * The "RT overload" flag: it gets set if a CPU has more than
470 * one runnable RT task.
472 cpumask_var_t rto_mask
;
475 struct cpupri cpupri
;
480 * By default the system creates a single root-domain with all cpus as
481 * members (mimicking the global state we have today).
483 static struct root_domain def_root_domain
;
488 * This is the main, per-CPU runqueue data structure.
490 * Locking rule: those places that want to lock multiple runqueues
491 * (such as the load balancing or the thread migration code), lock
492 * acquire operations must be ordered by ascending &runqueue.
499 * nr_running and cpu_load should be in the same cacheline because
500 * remote CPUs use both these fields when doing load calculation.
502 unsigned long nr_running
;
503 #define CPU_LOAD_IDX_MAX 5
504 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
506 unsigned char in_nohz_recently
;
508 /* capture load from *all* tasks on this cpu: */
509 struct load_weight load
;
510 unsigned long nr_load_updates
;
516 #ifdef CONFIG_FAIR_GROUP_SCHED
517 /* list of leaf cfs_rq on this cpu: */
518 struct list_head leaf_cfs_rq_list
;
520 #ifdef CONFIG_RT_GROUP_SCHED
521 struct list_head leaf_rt_rq_list
;
525 * This is part of a global counter where only the total sum
526 * over all CPUs matters. A task can increase this counter on
527 * one CPU and if it got migrated afterwards it may decrease
528 * it on another CPU. Always updated under the runqueue lock:
530 unsigned long nr_uninterruptible
;
532 struct task_struct
*curr
, *idle
;
533 unsigned long next_balance
;
534 struct mm_struct
*prev_mm
;
541 struct root_domain
*rd
;
542 struct sched_domain
*sd
;
544 unsigned char idle_at_tick
;
545 /* For active balancing */
549 /* cpu of this runqueue: */
553 unsigned long avg_load_per_task
;
555 struct task_struct
*migration_thread
;
556 struct list_head migration_queue
;
564 /* calc_load related fields */
565 unsigned long calc_load_update
;
566 long calc_load_active
;
568 #ifdef CONFIG_SCHED_HRTICK
570 int hrtick_csd_pending
;
571 struct call_single_data hrtick_csd
;
573 struct hrtimer hrtick_timer
;
576 #ifdef CONFIG_SCHEDSTATS
578 struct sched_info rq_sched_info
;
579 unsigned long long rq_cpu_time
;
580 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
582 /* sys_sched_yield() stats */
583 unsigned int yld_count
;
585 /* schedule() stats */
586 unsigned int sched_switch
;
587 unsigned int sched_count
;
588 unsigned int sched_goidle
;
590 /* try_to_wake_up() stats */
591 unsigned int ttwu_count
;
592 unsigned int ttwu_local
;
595 unsigned int bkl_count
;
599 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
602 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
604 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
607 static inline int cpu_of(struct rq
*rq
)
616 #define rcu_dereference_check_sched_domain(p) \
617 rcu_dereference_check((p), \
618 rcu_read_lock_sched_held() || \
619 lockdep_is_held(&sched_domains_mutex))
622 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
623 * See detach_destroy_domains: synchronize_sched for details.
625 * The domain tree of any CPU may only be accessed from within
626 * preempt-disabled sections.
628 #define for_each_domain(cpu, __sd) \
629 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
631 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
632 #define this_rq() (&__get_cpu_var(runqueues))
633 #define task_rq(p) cpu_rq(task_cpu(p))
634 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
635 #define raw_rq() (&__raw_get_cpu_var(runqueues))
637 inline void update_rq_clock(struct rq
*rq
)
639 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
643 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
645 #ifdef CONFIG_SCHED_DEBUG
646 # define const_debug __read_mostly
648 # define const_debug static const
653 * @cpu: the processor in question.
655 * Returns true if the current cpu runqueue is locked.
656 * This interface allows printk to be called with the runqueue lock
657 * held and know whether or not it is OK to wake up the klogd.
659 int runqueue_is_locked(int cpu
)
661 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
665 * Debugging: various feature bits
668 #define SCHED_FEAT(name, enabled) \
669 __SCHED_FEAT_##name ,
672 #include "sched_features.h"
677 #define SCHED_FEAT(name, enabled) \
678 (1UL << __SCHED_FEAT_##name) * enabled |
680 const_debug
unsigned int sysctl_sched_features
=
681 #include "sched_features.h"
686 #ifdef CONFIG_SCHED_DEBUG
687 #define SCHED_FEAT(name, enabled) \
690 static __read_mostly
char *sched_feat_names
[] = {
691 #include "sched_features.h"
697 static int sched_feat_show(struct seq_file
*m
, void *v
)
701 for (i
= 0; sched_feat_names
[i
]; i
++) {
702 if (!(sysctl_sched_features
& (1UL << i
)))
704 seq_printf(m
, "%s ", sched_feat_names
[i
]);
712 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
713 size_t cnt
, loff_t
*ppos
)
723 if (copy_from_user(&buf
, ubuf
, cnt
))
728 if (strncmp(buf
, "NO_", 3) == 0) {
733 for (i
= 0; sched_feat_names
[i
]; i
++) {
734 int len
= strlen(sched_feat_names
[i
]);
736 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
738 sysctl_sched_features
&= ~(1UL << i
);
740 sysctl_sched_features
|= (1UL << i
);
745 if (!sched_feat_names
[i
])
753 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
755 return single_open(filp
, sched_feat_show
, NULL
);
758 static const struct file_operations sched_feat_fops
= {
759 .open
= sched_feat_open
,
760 .write
= sched_feat_write
,
763 .release
= single_release
,
766 static __init
int sched_init_debug(void)
768 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
773 late_initcall(sched_init_debug
);
777 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
780 * Number of tasks to iterate in a single balance run.
781 * Limited because this is done with IRQs disabled.
783 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
786 * ratelimit for updating the group shares.
789 unsigned int sysctl_sched_shares_ratelimit
= 250000;
790 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
793 * Inject some fuzzyness into changing the per-cpu group shares
794 * this avoids remote rq-locks at the expense of fairness.
797 unsigned int sysctl_sched_shares_thresh
= 4;
800 * period over which we average the RT time consumption, measured
805 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
808 * period over which we measure -rt task cpu usage in us.
811 unsigned int sysctl_sched_rt_period
= 1000000;
813 static __read_mostly
int scheduler_running
;
816 * part of the period that we allow rt tasks to run in us.
819 int sysctl_sched_rt_runtime
= 950000;
821 static inline u64
global_rt_period(void)
823 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
826 static inline u64
global_rt_runtime(void)
828 if (sysctl_sched_rt_runtime
< 0)
831 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
834 #ifndef prepare_arch_switch
835 # define prepare_arch_switch(next) do { } while (0)
837 #ifndef finish_arch_switch
838 # define finish_arch_switch(prev) do { } while (0)
841 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
843 return rq
->curr
== p
;
846 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
847 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
849 return task_current(rq
, p
);
852 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
856 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
858 #ifdef CONFIG_DEBUG_SPINLOCK
859 /* this is a valid case when another task releases the spinlock */
860 rq
->lock
.owner
= current
;
863 * If we are tracking spinlock dependencies then we have to
864 * fix up the runqueue lock - which gets 'carried over' from
867 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
869 raw_spin_unlock_irq(&rq
->lock
);
872 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
878 return task_current(rq
, p
);
882 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
886 * We can optimise this out completely for !SMP, because the
887 * SMP rebalancing from interrupt is the only thing that cares
892 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
893 raw_spin_unlock_irq(&rq
->lock
);
895 raw_spin_unlock(&rq
->lock
);
899 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
903 * After ->oncpu is cleared, the task can be moved to a different CPU.
904 * We must ensure this doesn't happen until the switch is completely
910 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
914 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
917 * Check whether the task is waking, we use this to synchronize against
918 * ttwu() so that task_cpu() reports a stable number.
920 * We need to make an exception for PF_STARTING tasks because the fork
921 * path might require task_rq_lock() to work, eg. it can call
922 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
924 static inline int task_is_waking(struct task_struct
*p
)
926 return unlikely((p
->state
== TASK_WAKING
) && !(p
->flags
& PF_STARTING
));
930 * __task_rq_lock - lock the runqueue a given task resides on.
931 * Must be called interrupts disabled.
933 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
939 while (task_is_waking(p
))
942 raw_spin_lock(&rq
->lock
);
943 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
945 raw_spin_unlock(&rq
->lock
);
950 * task_rq_lock - lock the runqueue a given task resides on and disable
951 * interrupts. Note the ordering: we can safely lookup the task_rq without
952 * explicitly disabling preemption.
954 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
960 while (task_is_waking(p
))
962 local_irq_save(*flags
);
964 raw_spin_lock(&rq
->lock
);
965 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
967 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
971 void task_rq_unlock_wait(struct task_struct
*p
)
973 struct rq
*rq
= task_rq(p
);
975 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
976 raw_spin_unlock_wait(&rq
->lock
);
979 static void __task_rq_unlock(struct rq
*rq
)
982 raw_spin_unlock(&rq
->lock
);
985 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
988 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
992 * this_rq_lock - lock this runqueue and disable interrupts.
994 static struct rq
*this_rq_lock(void)
1001 raw_spin_lock(&rq
->lock
);
1006 #ifdef CONFIG_SCHED_HRTICK
1008 * Use HR-timers to deliver accurate preemption points.
1010 * Its all a bit involved since we cannot program an hrt while holding the
1011 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1014 * When we get rescheduled we reprogram the hrtick_timer outside of the
1020 * - enabled by features
1021 * - hrtimer is actually high res
1023 static inline int hrtick_enabled(struct rq
*rq
)
1025 if (!sched_feat(HRTICK
))
1027 if (!cpu_active(cpu_of(rq
)))
1029 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1032 static void hrtick_clear(struct rq
*rq
)
1034 if (hrtimer_active(&rq
->hrtick_timer
))
1035 hrtimer_cancel(&rq
->hrtick_timer
);
1039 * High-resolution timer tick.
1040 * Runs from hardirq context with interrupts disabled.
1042 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1044 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1046 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1048 raw_spin_lock(&rq
->lock
);
1049 update_rq_clock(rq
);
1050 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1051 raw_spin_unlock(&rq
->lock
);
1053 return HRTIMER_NORESTART
;
1058 * called from hardirq (IPI) context
1060 static void __hrtick_start(void *arg
)
1062 struct rq
*rq
= arg
;
1064 raw_spin_lock(&rq
->lock
);
1065 hrtimer_restart(&rq
->hrtick_timer
);
1066 rq
->hrtick_csd_pending
= 0;
1067 raw_spin_unlock(&rq
->lock
);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq
*rq
, u64 delay
)
1077 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1078 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1080 hrtimer_set_expires(timer
, time
);
1082 if (rq
== this_rq()) {
1083 hrtimer_restart(timer
);
1084 } else if (!rq
->hrtick_csd_pending
) {
1085 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1086 rq
->hrtick_csd_pending
= 1;
1091 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1093 int cpu
= (int)(long)hcpu
;
1096 case CPU_UP_CANCELED
:
1097 case CPU_UP_CANCELED_FROZEN
:
1098 case CPU_DOWN_PREPARE
:
1099 case CPU_DOWN_PREPARE_FROZEN
:
1101 case CPU_DEAD_FROZEN
:
1102 hrtick_clear(cpu_rq(cpu
));
1109 static __init
void init_hrtick(void)
1111 hotcpu_notifier(hotplug_hrtick
, 0);
1115 * Called to set the hrtick timer state.
1117 * called with rq->lock held and irqs disabled
1119 static void hrtick_start(struct rq
*rq
, u64 delay
)
1121 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1122 HRTIMER_MODE_REL_PINNED
, 0);
1125 static inline void init_hrtick(void)
1128 #endif /* CONFIG_SMP */
1130 static void init_rq_hrtick(struct rq
*rq
)
1133 rq
->hrtick_csd_pending
= 0;
1135 rq
->hrtick_csd
.flags
= 0;
1136 rq
->hrtick_csd
.func
= __hrtick_start
;
1137 rq
->hrtick_csd
.info
= rq
;
1140 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1141 rq
->hrtick_timer
.function
= hrtick
;
1143 #else /* CONFIG_SCHED_HRTICK */
1144 static inline void hrtick_clear(struct rq
*rq
)
1148 static inline void init_rq_hrtick(struct rq
*rq
)
1152 static inline void init_hrtick(void)
1155 #endif /* CONFIG_SCHED_HRTICK */
1158 * resched_task - mark a task 'to be rescheduled now'.
1160 * On UP this means the setting of the need_resched flag, on SMP it
1161 * might also involve a cross-CPU call to trigger the scheduler on
1166 #ifndef tsk_is_polling
1167 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1170 static void resched_task(struct task_struct
*p
)
1174 assert_raw_spin_locked(&task_rq(p
)->lock
);
1176 if (test_tsk_need_resched(p
))
1179 set_tsk_need_resched(p
);
1182 if (cpu
== smp_processor_id())
1185 /* NEED_RESCHED must be visible before we test polling */
1187 if (!tsk_is_polling(p
))
1188 smp_send_reschedule(cpu
);
1191 static void resched_cpu(int cpu
)
1193 struct rq
*rq
= cpu_rq(cpu
);
1194 unsigned long flags
;
1196 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1198 resched_task(cpu_curr(cpu
));
1199 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1204 * When add_timer_on() enqueues a timer into the timer wheel of an
1205 * idle CPU then this timer might expire before the next timer event
1206 * which is scheduled to wake up that CPU. In case of a completely
1207 * idle system the next event might even be infinite time into the
1208 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1209 * leaves the inner idle loop so the newly added timer is taken into
1210 * account when the CPU goes back to idle and evaluates the timer
1211 * wheel for the next timer event.
1213 void wake_up_idle_cpu(int cpu
)
1215 struct rq
*rq
= cpu_rq(cpu
);
1217 if (cpu
== smp_processor_id())
1221 * This is safe, as this function is called with the timer
1222 * wheel base lock of (cpu) held. When the CPU is on the way
1223 * to idle and has not yet set rq->curr to idle then it will
1224 * be serialized on the timer wheel base lock and take the new
1225 * timer into account automatically.
1227 if (rq
->curr
!= rq
->idle
)
1231 * We can set TIF_RESCHED on the idle task of the other CPU
1232 * lockless. The worst case is that the other CPU runs the
1233 * idle task through an additional NOOP schedule()
1235 set_tsk_need_resched(rq
->idle
);
1237 /* NEED_RESCHED must be visible before we test polling */
1239 if (!tsk_is_polling(rq
->idle
))
1240 smp_send_reschedule(cpu
);
1242 #endif /* CONFIG_NO_HZ */
1244 static u64
sched_avg_period(void)
1246 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1249 static void sched_avg_update(struct rq
*rq
)
1251 s64 period
= sched_avg_period();
1253 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1254 rq
->age_stamp
+= period
;
1259 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1261 rq
->rt_avg
+= rt_delta
;
1262 sched_avg_update(rq
);
1265 #else /* !CONFIG_SMP */
1266 static void resched_task(struct task_struct
*p
)
1268 assert_raw_spin_locked(&task_rq(p
)->lock
);
1269 set_tsk_need_resched(p
);
1272 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1275 #endif /* CONFIG_SMP */
1277 #if BITS_PER_LONG == 32
1278 # define WMULT_CONST (~0UL)
1280 # define WMULT_CONST (1UL << 32)
1283 #define WMULT_SHIFT 32
1286 * Shift right and round:
1288 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1291 * delta *= weight / lw
1293 static unsigned long
1294 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1295 struct load_weight
*lw
)
1299 if (!lw
->inv_weight
) {
1300 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1303 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1307 tmp
= (u64
)delta_exec
* weight
;
1309 * Check whether we'd overflow the 64-bit multiplication:
1311 if (unlikely(tmp
> WMULT_CONST
))
1312 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1315 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1317 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1320 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1326 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1333 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1334 * of tasks with abnormal "nice" values across CPUs the contribution that
1335 * each task makes to its run queue's load is weighted according to its
1336 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1337 * scaled version of the new time slice allocation that they receive on time
1341 #define WEIGHT_IDLEPRIO 3
1342 #define WMULT_IDLEPRIO 1431655765
1345 * Nice levels are multiplicative, with a gentle 10% change for every
1346 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1347 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1348 * that remained on nice 0.
1350 * The "10% effect" is relative and cumulative: from _any_ nice level,
1351 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1352 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1353 * If a task goes up by ~10% and another task goes down by ~10% then
1354 * the relative distance between them is ~25%.)
1356 static const int prio_to_weight
[40] = {
1357 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1358 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1359 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1360 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1361 /* 0 */ 1024, 820, 655, 526, 423,
1362 /* 5 */ 335, 272, 215, 172, 137,
1363 /* 10 */ 110, 87, 70, 56, 45,
1364 /* 15 */ 36, 29, 23, 18, 15,
1368 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1370 * In cases where the weight does not change often, we can use the
1371 * precalculated inverse to speed up arithmetics by turning divisions
1372 * into multiplications:
1374 static const u32 prio_to_wmult
[40] = {
1375 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1376 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1377 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1378 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1379 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1380 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1381 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1382 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1385 /* Time spent by the tasks of the cpu accounting group executing in ... */
1386 enum cpuacct_stat_index
{
1387 CPUACCT_STAT_USER
, /* ... user mode */
1388 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1390 CPUACCT_STAT_NSTATS
,
1393 #ifdef CONFIG_CGROUP_CPUACCT
1394 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1395 static void cpuacct_update_stats(struct task_struct
*tsk
,
1396 enum cpuacct_stat_index idx
, cputime_t val
);
1398 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1399 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1400 enum cpuacct_stat_index idx
, cputime_t val
) {}
1403 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1405 update_load_add(&rq
->load
, load
);
1408 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1410 update_load_sub(&rq
->load
, load
);
1413 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1414 typedef int (*tg_visitor
)(struct task_group
*, void *);
1417 * Iterate the full tree, calling @down when first entering a node and @up when
1418 * leaving it for the final time.
1420 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1422 struct task_group
*parent
, *child
;
1426 parent
= &root_task_group
;
1428 ret
= (*down
)(parent
, data
);
1431 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1438 ret
= (*up
)(parent
, data
);
1443 parent
= parent
->parent
;
1452 static int tg_nop(struct task_group
*tg
, void *data
)
1459 /* Used instead of source_load when we know the type == 0 */
1460 static unsigned long weighted_cpuload(const int cpu
)
1462 return cpu_rq(cpu
)->load
.weight
;
1466 * Return a low guess at the load of a migration-source cpu weighted
1467 * according to the scheduling class and "nice" value.
1469 * We want to under-estimate the load of migration sources, to
1470 * balance conservatively.
1472 static unsigned long source_load(int cpu
, int type
)
1474 struct rq
*rq
= cpu_rq(cpu
);
1475 unsigned long total
= weighted_cpuload(cpu
);
1477 if (type
== 0 || !sched_feat(LB_BIAS
))
1480 return min(rq
->cpu_load
[type
-1], total
);
1484 * Return a high guess at the load of a migration-target cpu weighted
1485 * according to the scheduling class and "nice" value.
1487 static unsigned long target_load(int cpu
, int type
)
1489 struct rq
*rq
= cpu_rq(cpu
);
1490 unsigned long total
= weighted_cpuload(cpu
);
1492 if (type
== 0 || !sched_feat(LB_BIAS
))
1495 return max(rq
->cpu_load
[type
-1], total
);
1498 static struct sched_group
*group_of(int cpu
)
1500 struct sched_domain
*sd
= rcu_dereference_sched(cpu_rq(cpu
)->sd
);
1508 static unsigned long power_of(int cpu
)
1510 struct sched_group
*group
= group_of(cpu
);
1513 return SCHED_LOAD_SCALE
;
1515 return group
->cpu_power
;
1518 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1520 static unsigned long cpu_avg_load_per_task(int cpu
)
1522 struct rq
*rq
= cpu_rq(cpu
);
1523 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1526 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1528 rq
->avg_load_per_task
= 0;
1530 return rq
->avg_load_per_task
;
1533 #ifdef CONFIG_FAIR_GROUP_SCHED
1535 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1537 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1540 * Calculate and set the cpu's group shares.
1542 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1543 unsigned long sd_shares
,
1544 unsigned long sd_rq_weight
,
1545 unsigned long *usd_rq_weight
)
1547 unsigned long shares
, rq_weight
;
1550 rq_weight
= usd_rq_weight
[cpu
];
1553 rq_weight
= NICE_0_LOAD
;
1557 * \Sum_j shares_j * rq_weight_i
1558 * shares_i = -----------------------------
1559 * \Sum_j rq_weight_j
1561 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1562 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1564 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1565 sysctl_sched_shares_thresh
) {
1566 struct rq
*rq
= cpu_rq(cpu
);
1567 unsigned long flags
;
1569 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1570 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1571 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1572 __set_se_shares(tg
->se
[cpu
], shares
);
1573 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1578 * Re-compute the task group their per cpu shares over the given domain.
1579 * This needs to be done in a bottom-up fashion because the rq weight of a
1580 * parent group depends on the shares of its child groups.
1582 static int tg_shares_up(struct task_group
*tg
, void *data
)
1584 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1585 unsigned long *usd_rq_weight
;
1586 struct sched_domain
*sd
= data
;
1587 unsigned long flags
;
1593 local_irq_save(flags
);
1594 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1596 for_each_cpu(i
, sched_domain_span(sd
)) {
1597 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1598 usd_rq_weight
[i
] = weight
;
1600 rq_weight
+= weight
;
1602 * If there are currently no tasks on the cpu pretend there
1603 * is one of average load so that when a new task gets to
1604 * run here it will not get delayed by group starvation.
1607 weight
= NICE_0_LOAD
;
1609 sum_weight
+= weight
;
1610 shares
+= tg
->cfs_rq
[i
]->shares
;
1614 rq_weight
= sum_weight
;
1616 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1617 shares
= tg
->shares
;
1619 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1620 shares
= tg
->shares
;
1622 for_each_cpu(i
, sched_domain_span(sd
))
1623 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1625 local_irq_restore(flags
);
1631 * Compute the cpu's hierarchical load factor for each task group.
1632 * This needs to be done in a top-down fashion because the load of a child
1633 * group is a fraction of its parents load.
1635 static int tg_load_down(struct task_group
*tg
, void *data
)
1638 long cpu
= (long)data
;
1641 load
= cpu_rq(cpu
)->load
.weight
;
1643 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1644 load
*= tg
->cfs_rq
[cpu
]->shares
;
1645 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1648 tg
->cfs_rq
[cpu
]->h_load
= load
;
1653 static void update_shares(struct sched_domain
*sd
)
1658 if (root_task_group_empty())
1661 now
= cpu_clock(raw_smp_processor_id());
1662 elapsed
= now
- sd
->last_update
;
1664 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1665 sd
->last_update
= now
;
1666 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1670 static void update_h_load(long cpu
)
1672 if (root_task_group_empty())
1675 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1680 static inline void update_shares(struct sched_domain
*sd
)
1686 #ifdef CONFIG_PREEMPT
1688 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1691 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1692 * way at the expense of forcing extra atomic operations in all
1693 * invocations. This assures that the double_lock is acquired using the
1694 * same underlying policy as the spinlock_t on this architecture, which
1695 * reduces latency compared to the unfair variant below. However, it
1696 * also adds more overhead and therefore may reduce throughput.
1698 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1699 __releases(this_rq
->lock
)
1700 __acquires(busiest
->lock
)
1701 __acquires(this_rq
->lock
)
1703 raw_spin_unlock(&this_rq
->lock
);
1704 double_rq_lock(this_rq
, busiest
);
1711 * Unfair double_lock_balance: Optimizes throughput at the expense of
1712 * latency by eliminating extra atomic operations when the locks are
1713 * already in proper order on entry. This favors lower cpu-ids and will
1714 * grant the double lock to lower cpus over higher ids under contention,
1715 * regardless of entry order into the function.
1717 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1718 __releases(this_rq
->lock
)
1719 __acquires(busiest
->lock
)
1720 __acquires(this_rq
->lock
)
1724 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1725 if (busiest
< this_rq
) {
1726 raw_spin_unlock(&this_rq
->lock
);
1727 raw_spin_lock(&busiest
->lock
);
1728 raw_spin_lock_nested(&this_rq
->lock
,
1729 SINGLE_DEPTH_NESTING
);
1732 raw_spin_lock_nested(&busiest
->lock
,
1733 SINGLE_DEPTH_NESTING
);
1738 #endif /* CONFIG_PREEMPT */
1741 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1743 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1745 if (unlikely(!irqs_disabled())) {
1746 /* printk() doesn't work good under rq->lock */
1747 raw_spin_unlock(&this_rq
->lock
);
1751 return _double_lock_balance(this_rq
, busiest
);
1754 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1755 __releases(busiest
->lock
)
1757 raw_spin_unlock(&busiest
->lock
);
1758 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1762 * double_rq_lock - safely lock two runqueues
1764 * Note this does not disable interrupts like task_rq_lock,
1765 * you need to do so manually before calling.
1767 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1768 __acquires(rq1
->lock
)
1769 __acquires(rq2
->lock
)
1771 BUG_ON(!irqs_disabled());
1773 raw_spin_lock(&rq1
->lock
);
1774 __acquire(rq2
->lock
); /* Fake it out ;) */
1777 raw_spin_lock(&rq1
->lock
);
1778 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1780 raw_spin_lock(&rq2
->lock
);
1781 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1784 update_rq_clock(rq1
);
1785 update_rq_clock(rq2
);
1789 * double_rq_unlock - safely unlock two runqueues
1791 * Note this does not restore interrupts like task_rq_unlock,
1792 * you need to do so manually after calling.
1794 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1795 __releases(rq1
->lock
)
1796 __releases(rq2
->lock
)
1798 raw_spin_unlock(&rq1
->lock
);
1800 raw_spin_unlock(&rq2
->lock
);
1802 __release(rq2
->lock
);
1807 #ifdef CONFIG_FAIR_GROUP_SCHED
1808 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1811 cfs_rq
->shares
= shares
;
1816 static void calc_load_account_active(struct rq
*this_rq
);
1817 static void update_sysctl(void);
1818 static int get_update_sysctl_factor(void);
1820 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1822 set_task_rq(p
, cpu
);
1825 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1826 * successfuly executed on another CPU. We must ensure that updates of
1827 * per-task data have been completed by this moment.
1830 task_thread_info(p
)->cpu
= cpu
;
1834 static const struct sched_class rt_sched_class
;
1836 #define sched_class_highest (&rt_sched_class)
1837 #define for_each_class(class) \
1838 for (class = sched_class_highest; class; class = class->next)
1840 #include "sched_stats.h"
1842 static void inc_nr_running(struct rq
*rq
)
1847 static void dec_nr_running(struct rq
*rq
)
1852 static void set_load_weight(struct task_struct
*p
)
1854 if (task_has_rt_policy(p
)) {
1855 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1856 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1861 * SCHED_IDLE tasks get minimal weight:
1863 if (p
->policy
== SCHED_IDLE
) {
1864 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1865 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1869 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1870 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1873 static void update_avg(u64
*avg
, u64 sample
)
1875 s64 diff
= sample
- *avg
;
1880 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1883 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1885 sched_info_queued(p
);
1886 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1890 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1893 if (p
->se
.last_wakeup
) {
1894 update_avg(&p
->se
.avg_overlap
,
1895 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1896 p
->se
.last_wakeup
= 0;
1898 update_avg(&p
->se
.avg_wakeup
,
1899 sysctl_sched_wakeup_granularity
);
1903 sched_info_dequeued(p
);
1904 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1909 * activate_task - move a task to the runqueue.
1911 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1913 if (task_contributes_to_load(p
))
1914 rq
->nr_uninterruptible
--;
1916 enqueue_task(rq
, p
, wakeup
, false);
1921 * deactivate_task - remove a task from the runqueue.
1923 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1925 if (task_contributes_to_load(p
))
1926 rq
->nr_uninterruptible
++;
1928 dequeue_task(rq
, p
, sleep
);
1932 #include "sched_idletask.c"
1933 #include "sched_fair.c"
1934 #include "sched_rt.c"
1935 #ifdef CONFIG_SCHED_DEBUG
1936 # include "sched_debug.c"
1940 * __normal_prio - return the priority that is based on the static prio
1942 static inline int __normal_prio(struct task_struct
*p
)
1944 return p
->static_prio
;
1948 * Calculate the expected normal priority: i.e. priority
1949 * without taking RT-inheritance into account. Might be
1950 * boosted by interactivity modifiers. Changes upon fork,
1951 * setprio syscalls, and whenever the interactivity
1952 * estimator recalculates.
1954 static inline int normal_prio(struct task_struct
*p
)
1958 if (task_has_rt_policy(p
))
1959 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1961 prio
= __normal_prio(p
);
1966 * Calculate the current priority, i.e. the priority
1967 * taken into account by the scheduler. This value might
1968 * be boosted by RT tasks, or might be boosted by
1969 * interactivity modifiers. Will be RT if the task got
1970 * RT-boosted. If not then it returns p->normal_prio.
1972 static int effective_prio(struct task_struct
*p
)
1974 p
->normal_prio
= normal_prio(p
);
1976 * If we are RT tasks or we were boosted to RT priority,
1977 * keep the priority unchanged. Otherwise, update priority
1978 * to the normal priority:
1980 if (!rt_prio(p
->prio
))
1981 return p
->normal_prio
;
1986 * task_curr - is this task currently executing on a CPU?
1987 * @p: the task in question.
1989 inline int task_curr(const struct task_struct
*p
)
1991 return cpu_curr(task_cpu(p
)) == p
;
1994 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1995 const struct sched_class
*prev_class
,
1996 int oldprio
, int running
)
1998 if (prev_class
!= p
->sched_class
) {
1999 if (prev_class
->switched_from
)
2000 prev_class
->switched_from(rq
, p
, running
);
2001 p
->sched_class
->switched_to(rq
, p
, running
);
2003 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2008 * Is this task likely cache-hot:
2011 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2015 if (p
->sched_class
!= &fair_sched_class
)
2019 * Buddy candidates are cache hot:
2021 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2022 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2023 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2026 if (sysctl_sched_migration_cost
== -1)
2028 if (sysctl_sched_migration_cost
== 0)
2031 delta
= now
- p
->se
.exec_start
;
2033 return delta
< (s64
)sysctl_sched_migration_cost
;
2036 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2038 #ifdef CONFIG_SCHED_DEBUG
2040 * We should never call set_task_cpu() on a blocked task,
2041 * ttwu() will sort out the placement.
2043 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2044 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2047 trace_sched_migrate_task(p
, new_cpu
);
2049 if (task_cpu(p
) != new_cpu
) {
2050 p
->se
.nr_migrations
++;
2051 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2054 __set_task_cpu(p
, new_cpu
);
2057 struct migration_req
{
2058 struct list_head list
;
2060 struct task_struct
*task
;
2063 struct completion done
;
2067 * The task's runqueue lock must be held.
2068 * Returns true if you have to wait for migration thread.
2071 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2073 struct rq
*rq
= task_rq(p
);
2076 * If the task is not on a runqueue (and not running), then
2077 * the next wake-up will properly place the task.
2079 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2082 init_completion(&req
->done
);
2084 req
->dest_cpu
= dest_cpu
;
2085 list_add(&req
->list
, &rq
->migration_queue
);
2091 * wait_task_inactive - wait for a thread to unschedule.
2093 * If @match_state is nonzero, it's the @p->state value just checked and
2094 * not expected to change. If it changes, i.e. @p might have woken up,
2095 * then return zero. When we succeed in waiting for @p to be off its CPU,
2096 * we return a positive number (its total switch count). If a second call
2097 * a short while later returns the same number, the caller can be sure that
2098 * @p has remained unscheduled the whole time.
2100 * The caller must ensure that the task *will* unschedule sometime soon,
2101 * else this function might spin for a *long* time. This function can't
2102 * be called with interrupts off, or it may introduce deadlock with
2103 * smp_call_function() if an IPI is sent by the same process we are
2104 * waiting to become inactive.
2106 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2108 unsigned long flags
;
2115 * We do the initial early heuristics without holding
2116 * any task-queue locks at all. We'll only try to get
2117 * the runqueue lock when things look like they will
2123 * If the task is actively running on another CPU
2124 * still, just relax and busy-wait without holding
2127 * NOTE! Since we don't hold any locks, it's not
2128 * even sure that "rq" stays as the right runqueue!
2129 * But we don't care, since "task_running()" will
2130 * return false if the runqueue has changed and p
2131 * is actually now running somewhere else!
2133 while (task_running(rq
, p
)) {
2134 if (match_state
&& unlikely(p
->state
!= match_state
))
2140 * Ok, time to look more closely! We need the rq
2141 * lock now, to be *sure*. If we're wrong, we'll
2142 * just go back and repeat.
2144 rq
= task_rq_lock(p
, &flags
);
2145 trace_sched_wait_task(rq
, p
);
2146 running
= task_running(rq
, p
);
2147 on_rq
= p
->se
.on_rq
;
2149 if (!match_state
|| p
->state
== match_state
)
2150 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2151 task_rq_unlock(rq
, &flags
);
2154 * If it changed from the expected state, bail out now.
2156 if (unlikely(!ncsw
))
2160 * Was it really running after all now that we
2161 * checked with the proper locks actually held?
2163 * Oops. Go back and try again..
2165 if (unlikely(running
)) {
2171 * It's not enough that it's not actively running,
2172 * it must be off the runqueue _entirely_, and not
2175 * So if it was still runnable (but just not actively
2176 * running right now), it's preempted, and we should
2177 * yield - it could be a while.
2179 if (unlikely(on_rq
)) {
2180 schedule_timeout_uninterruptible(1);
2185 * Ahh, all good. It wasn't running, and it wasn't
2186 * runnable, which means that it will never become
2187 * running in the future either. We're all done!
2196 * kick_process - kick a running thread to enter/exit the kernel
2197 * @p: the to-be-kicked thread
2199 * Cause a process which is running on another CPU to enter
2200 * kernel-mode, without any delay. (to get signals handled.)
2202 * NOTE: this function doesnt have to take the runqueue lock,
2203 * because all it wants to ensure is that the remote task enters
2204 * the kernel. If the IPI races and the task has been migrated
2205 * to another CPU then no harm is done and the purpose has been
2208 void kick_process(struct task_struct
*p
)
2214 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2215 smp_send_reschedule(cpu
);
2218 EXPORT_SYMBOL_GPL(kick_process
);
2219 #endif /* CONFIG_SMP */
2222 * task_oncpu_function_call - call a function on the cpu on which a task runs
2223 * @p: the task to evaluate
2224 * @func: the function to be called
2225 * @info: the function call argument
2227 * Calls the function @func when the task is currently running. This might
2228 * be on the current CPU, which just calls the function directly
2230 void task_oncpu_function_call(struct task_struct
*p
,
2231 void (*func
) (void *info
), void *info
)
2238 smp_call_function_single(cpu
, func
, info
, 1);
2243 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2246 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2248 /* Look for allowed, online CPU in same node. */
2249 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2250 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2253 /* Any allowed, online CPU? */
2254 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2255 if (dest_cpu
< nr_cpu_ids
)
2258 /* No more Mr. Nice Guy. */
2259 if (dest_cpu
>= nr_cpu_ids
) {
2261 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
2263 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
2266 * Don't tell them about moving exiting tasks or
2267 * kernel threads (both mm NULL), since they never
2270 if (p
->mm
&& printk_ratelimit()) {
2271 printk(KERN_INFO
"process %d (%s) no "
2272 "longer affine to cpu%d\n",
2273 task_pid_nr(p
), p
->comm
, cpu
);
2281 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2282 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2285 * exec: is unstable, retry loop
2286 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2289 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2291 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2294 * In order not to call set_task_cpu() on a blocking task we need
2295 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2298 * Since this is common to all placement strategies, this lives here.
2300 * [ this allows ->select_task() to simply return task_cpu(p) and
2301 * not worry about this generic constraint ]
2303 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2305 cpu
= select_fallback_rq(task_cpu(p
), p
);
2312 * try_to_wake_up - wake up a thread
2313 * @p: the to-be-woken-up thread
2314 * @state: the mask of task states that can be woken
2315 * @sync: do a synchronous wakeup?
2317 * Put it on the run-queue if it's not already there. The "current"
2318 * thread is always on the run-queue (except when the actual
2319 * re-schedule is in progress), and as such you're allowed to do
2320 * the simpler "current->state = TASK_RUNNING" to mark yourself
2321 * runnable without the overhead of this.
2323 * returns failure only if the task is already active.
2325 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2328 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2329 unsigned long flags
;
2332 if (!sched_feat(SYNC_WAKEUPS
))
2333 wake_flags
&= ~WF_SYNC
;
2335 this_cpu
= get_cpu();
2338 rq
= task_rq_lock(p
, &flags
);
2339 update_rq_clock(rq
);
2340 if (!(p
->state
& state
))
2350 if (unlikely(task_running(rq
, p
)))
2354 * In order to handle concurrent wakeups and release the rq->lock
2355 * we put the task in TASK_WAKING state.
2357 * First fix up the nr_uninterruptible count:
2359 if (task_contributes_to_load(p
))
2360 rq
->nr_uninterruptible
--;
2361 p
->state
= TASK_WAKING
;
2363 if (p
->sched_class
->task_waking
)
2364 p
->sched_class
->task_waking(rq
, p
);
2366 __task_rq_unlock(rq
);
2368 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2369 if (cpu
!= orig_cpu
) {
2371 * Since we migrate the task without holding any rq->lock,
2372 * we need to be careful with task_rq_lock(), since that
2373 * might end up locking an invalid rq.
2375 set_task_cpu(p
, cpu
);
2379 raw_spin_lock(&rq
->lock
);
2380 update_rq_clock(rq
);
2383 * We migrated the task without holding either rq->lock, however
2384 * since the task is not on the task list itself, nobody else
2385 * will try and migrate the task, hence the rq should match the
2386 * cpu we just moved it to.
2388 WARN_ON(task_cpu(p
) != cpu
);
2389 WARN_ON(p
->state
!= TASK_WAKING
);
2391 #ifdef CONFIG_SCHEDSTATS
2392 schedstat_inc(rq
, ttwu_count
);
2393 if (cpu
== this_cpu
)
2394 schedstat_inc(rq
, ttwu_local
);
2396 struct sched_domain
*sd
;
2397 for_each_domain(this_cpu
, sd
) {
2398 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2399 schedstat_inc(sd
, ttwu_wake_remote
);
2404 #endif /* CONFIG_SCHEDSTATS */
2407 #endif /* CONFIG_SMP */
2408 schedstat_inc(p
, se
.nr_wakeups
);
2409 if (wake_flags
& WF_SYNC
)
2410 schedstat_inc(p
, se
.nr_wakeups_sync
);
2411 if (orig_cpu
!= cpu
)
2412 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2413 if (cpu
== this_cpu
)
2414 schedstat_inc(p
, se
.nr_wakeups_local
);
2416 schedstat_inc(p
, se
.nr_wakeups_remote
);
2417 activate_task(rq
, p
, 1);
2421 * Only attribute actual wakeups done by this task.
2423 if (!in_interrupt()) {
2424 struct sched_entity
*se
= ¤t
->se
;
2425 u64 sample
= se
->sum_exec_runtime
;
2427 if (se
->last_wakeup
)
2428 sample
-= se
->last_wakeup
;
2430 sample
-= se
->start_runtime
;
2431 update_avg(&se
->avg_wakeup
, sample
);
2433 se
->last_wakeup
= se
->sum_exec_runtime
;
2437 trace_sched_wakeup(rq
, p
, success
);
2438 check_preempt_curr(rq
, p
, wake_flags
);
2440 p
->state
= TASK_RUNNING
;
2442 if (p
->sched_class
->task_woken
)
2443 p
->sched_class
->task_woken(rq
, p
);
2445 if (unlikely(rq
->idle_stamp
)) {
2446 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2447 u64 max
= 2*sysctl_sched_migration_cost
;
2452 update_avg(&rq
->avg_idle
, delta
);
2457 task_rq_unlock(rq
, &flags
);
2464 * wake_up_process - Wake up a specific process
2465 * @p: The process to be woken up.
2467 * Attempt to wake up the nominated process and move it to the set of runnable
2468 * processes. Returns 1 if the process was woken up, 0 if it was already
2471 * It may be assumed that this function implies a write memory barrier before
2472 * changing the task state if and only if any tasks are woken up.
2474 int wake_up_process(struct task_struct
*p
)
2476 return try_to_wake_up(p
, TASK_ALL
, 0);
2478 EXPORT_SYMBOL(wake_up_process
);
2480 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2482 return try_to_wake_up(p
, state
, 0);
2486 * Perform scheduler related setup for a newly forked process p.
2487 * p is forked by current.
2489 * __sched_fork() is basic setup used by init_idle() too:
2491 static void __sched_fork(struct task_struct
*p
)
2493 p
->se
.exec_start
= 0;
2494 p
->se
.sum_exec_runtime
= 0;
2495 p
->se
.prev_sum_exec_runtime
= 0;
2496 p
->se
.nr_migrations
= 0;
2497 p
->se
.last_wakeup
= 0;
2498 p
->se
.avg_overlap
= 0;
2499 p
->se
.start_runtime
= 0;
2500 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2502 #ifdef CONFIG_SCHEDSTATS
2503 p
->se
.wait_start
= 0;
2505 p
->se
.wait_count
= 0;
2508 p
->se
.sleep_start
= 0;
2509 p
->se
.sleep_max
= 0;
2510 p
->se
.sum_sleep_runtime
= 0;
2512 p
->se
.block_start
= 0;
2513 p
->se
.block_max
= 0;
2515 p
->se
.slice_max
= 0;
2517 p
->se
.nr_migrations_cold
= 0;
2518 p
->se
.nr_failed_migrations_affine
= 0;
2519 p
->se
.nr_failed_migrations_running
= 0;
2520 p
->se
.nr_failed_migrations_hot
= 0;
2521 p
->se
.nr_forced_migrations
= 0;
2523 p
->se
.nr_wakeups
= 0;
2524 p
->se
.nr_wakeups_sync
= 0;
2525 p
->se
.nr_wakeups_migrate
= 0;
2526 p
->se
.nr_wakeups_local
= 0;
2527 p
->se
.nr_wakeups_remote
= 0;
2528 p
->se
.nr_wakeups_affine
= 0;
2529 p
->se
.nr_wakeups_affine_attempts
= 0;
2530 p
->se
.nr_wakeups_passive
= 0;
2531 p
->se
.nr_wakeups_idle
= 0;
2535 INIT_LIST_HEAD(&p
->rt
.run_list
);
2537 INIT_LIST_HEAD(&p
->se
.group_node
);
2539 #ifdef CONFIG_PREEMPT_NOTIFIERS
2540 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2545 * fork()/clone()-time setup:
2547 void sched_fork(struct task_struct
*p
, int clone_flags
)
2549 int cpu
= get_cpu();
2553 * We mark the process as waking here. This guarantees that
2554 * nobody will actually run it, and a signal or other external
2555 * event cannot wake it up and insert it on the runqueue either.
2557 p
->state
= TASK_WAKING
;
2560 * Revert to default priority/policy on fork if requested.
2562 if (unlikely(p
->sched_reset_on_fork
)) {
2563 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2564 p
->policy
= SCHED_NORMAL
;
2565 p
->normal_prio
= p
->static_prio
;
2568 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2569 p
->static_prio
= NICE_TO_PRIO(0);
2570 p
->normal_prio
= p
->static_prio
;
2575 * We don't need the reset flag anymore after the fork. It has
2576 * fulfilled its duty:
2578 p
->sched_reset_on_fork
= 0;
2582 * Make sure we do not leak PI boosting priority to the child.
2584 p
->prio
= current
->normal_prio
;
2586 if (!rt_prio(p
->prio
))
2587 p
->sched_class
= &fair_sched_class
;
2589 if (p
->sched_class
->task_fork
)
2590 p
->sched_class
->task_fork(p
);
2592 set_task_cpu(p
, cpu
);
2594 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2595 if (likely(sched_info_on()))
2596 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2598 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2601 #ifdef CONFIG_PREEMPT
2602 /* Want to start with kernel preemption disabled. */
2603 task_thread_info(p
)->preempt_count
= 1;
2605 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2611 * wake_up_new_task - wake up a newly created task for the first time.
2613 * This function will do some initial scheduler statistics housekeeping
2614 * that must be done for every newly created context, then puts the task
2615 * on the runqueue and wakes it.
2617 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2619 unsigned long flags
;
2621 int cpu __maybe_unused
= get_cpu();
2625 * Fork balancing, do it here and not earlier because:
2626 * - cpus_allowed can change in the fork path
2627 * - any previously selected cpu might disappear through hotplug
2629 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2630 * ->cpus_allowed is stable, we have preemption disabled, meaning
2631 * cpu_online_mask is stable.
2633 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2634 set_task_cpu(p
, cpu
);
2638 * Since the task is not on the rq and we still have TASK_WAKING set
2639 * nobody else will migrate this task.
2642 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2644 BUG_ON(p
->state
!= TASK_WAKING
);
2645 p
->state
= TASK_RUNNING
;
2646 update_rq_clock(rq
);
2647 activate_task(rq
, p
, 0);
2648 trace_sched_wakeup_new(rq
, p
, 1);
2649 check_preempt_curr(rq
, p
, WF_FORK
);
2651 if (p
->sched_class
->task_woken
)
2652 p
->sched_class
->task_woken(rq
, p
);
2654 task_rq_unlock(rq
, &flags
);
2658 #ifdef CONFIG_PREEMPT_NOTIFIERS
2661 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2662 * @notifier: notifier struct to register
2664 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2666 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2668 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2671 * preempt_notifier_unregister - no longer interested in preemption notifications
2672 * @notifier: notifier struct to unregister
2674 * This is safe to call from within a preemption notifier.
2676 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2678 hlist_del(¬ifier
->link
);
2680 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2682 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2684 struct preempt_notifier
*notifier
;
2685 struct hlist_node
*node
;
2687 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2688 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2692 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2693 struct task_struct
*next
)
2695 struct preempt_notifier
*notifier
;
2696 struct hlist_node
*node
;
2698 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2699 notifier
->ops
->sched_out(notifier
, next
);
2702 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2704 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2709 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2710 struct task_struct
*next
)
2714 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2717 * prepare_task_switch - prepare to switch tasks
2718 * @rq: the runqueue preparing to switch
2719 * @prev: the current task that is being switched out
2720 * @next: the task we are going to switch to.
2722 * This is called with the rq lock held and interrupts off. It must
2723 * be paired with a subsequent finish_task_switch after the context
2726 * prepare_task_switch sets up locking and calls architecture specific
2730 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2731 struct task_struct
*next
)
2733 fire_sched_out_preempt_notifiers(prev
, next
);
2734 prepare_lock_switch(rq
, next
);
2735 prepare_arch_switch(next
);
2739 * finish_task_switch - clean up after a task-switch
2740 * @rq: runqueue associated with task-switch
2741 * @prev: the thread we just switched away from.
2743 * finish_task_switch must be called after the context switch, paired
2744 * with a prepare_task_switch call before the context switch.
2745 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2746 * and do any other architecture-specific cleanup actions.
2748 * Note that we may have delayed dropping an mm in context_switch(). If
2749 * so, we finish that here outside of the runqueue lock. (Doing it
2750 * with the lock held can cause deadlocks; see schedule() for
2753 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2754 __releases(rq
->lock
)
2756 struct mm_struct
*mm
= rq
->prev_mm
;
2762 * A task struct has one reference for the use as "current".
2763 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2764 * schedule one last time. The schedule call will never return, and
2765 * the scheduled task must drop that reference.
2766 * The test for TASK_DEAD must occur while the runqueue locks are
2767 * still held, otherwise prev could be scheduled on another cpu, die
2768 * there before we look at prev->state, and then the reference would
2770 * Manfred Spraul <manfred@colorfullife.com>
2772 prev_state
= prev
->state
;
2773 finish_arch_switch(prev
);
2774 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2775 local_irq_disable();
2776 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2777 perf_event_task_sched_in(current
);
2778 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2780 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2781 finish_lock_switch(rq
, prev
);
2783 fire_sched_in_preempt_notifiers(current
);
2786 if (unlikely(prev_state
== TASK_DEAD
)) {
2788 * Remove function-return probe instances associated with this
2789 * task and put them back on the free list.
2791 kprobe_flush_task(prev
);
2792 put_task_struct(prev
);
2798 /* assumes rq->lock is held */
2799 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2801 if (prev
->sched_class
->pre_schedule
)
2802 prev
->sched_class
->pre_schedule(rq
, prev
);
2805 /* rq->lock is NOT held, but preemption is disabled */
2806 static inline void post_schedule(struct rq
*rq
)
2808 if (rq
->post_schedule
) {
2809 unsigned long flags
;
2811 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2812 if (rq
->curr
->sched_class
->post_schedule
)
2813 rq
->curr
->sched_class
->post_schedule(rq
);
2814 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2816 rq
->post_schedule
= 0;
2822 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2826 static inline void post_schedule(struct rq
*rq
)
2833 * schedule_tail - first thing a freshly forked thread must call.
2834 * @prev: the thread we just switched away from.
2836 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2837 __releases(rq
->lock
)
2839 struct rq
*rq
= this_rq();
2841 finish_task_switch(rq
, prev
);
2844 * FIXME: do we need to worry about rq being invalidated by the
2849 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2850 /* In this case, finish_task_switch does not reenable preemption */
2853 if (current
->set_child_tid
)
2854 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2858 * context_switch - switch to the new MM and the new
2859 * thread's register state.
2862 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2863 struct task_struct
*next
)
2865 struct mm_struct
*mm
, *oldmm
;
2867 prepare_task_switch(rq
, prev
, next
);
2868 trace_sched_switch(rq
, prev
, next
);
2870 oldmm
= prev
->active_mm
;
2872 * For paravirt, this is coupled with an exit in switch_to to
2873 * combine the page table reload and the switch backend into
2876 arch_start_context_switch(prev
);
2879 next
->active_mm
= oldmm
;
2880 atomic_inc(&oldmm
->mm_count
);
2881 enter_lazy_tlb(oldmm
, next
);
2883 switch_mm(oldmm
, mm
, next
);
2885 if (likely(!prev
->mm
)) {
2886 prev
->active_mm
= NULL
;
2887 rq
->prev_mm
= oldmm
;
2890 * Since the runqueue lock will be released by the next
2891 * task (which is an invalid locking op but in the case
2892 * of the scheduler it's an obvious special-case), so we
2893 * do an early lockdep release here:
2895 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2896 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2899 /* Here we just switch the register state and the stack. */
2900 switch_to(prev
, next
, prev
);
2904 * this_rq must be evaluated again because prev may have moved
2905 * CPUs since it called schedule(), thus the 'rq' on its stack
2906 * frame will be invalid.
2908 finish_task_switch(this_rq(), prev
);
2912 * nr_running, nr_uninterruptible and nr_context_switches:
2914 * externally visible scheduler statistics: current number of runnable
2915 * threads, current number of uninterruptible-sleeping threads, total
2916 * number of context switches performed since bootup.
2918 unsigned long nr_running(void)
2920 unsigned long i
, sum
= 0;
2922 for_each_online_cpu(i
)
2923 sum
+= cpu_rq(i
)->nr_running
;
2928 unsigned long nr_uninterruptible(void)
2930 unsigned long i
, sum
= 0;
2932 for_each_possible_cpu(i
)
2933 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2936 * Since we read the counters lockless, it might be slightly
2937 * inaccurate. Do not allow it to go below zero though:
2939 if (unlikely((long)sum
< 0))
2945 unsigned long long nr_context_switches(void)
2948 unsigned long long sum
= 0;
2950 for_each_possible_cpu(i
)
2951 sum
+= cpu_rq(i
)->nr_switches
;
2956 unsigned long nr_iowait(void)
2958 unsigned long i
, sum
= 0;
2960 for_each_possible_cpu(i
)
2961 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2966 unsigned long nr_iowait_cpu(void)
2968 struct rq
*this = this_rq();
2969 return atomic_read(&this->nr_iowait
);
2972 unsigned long this_cpu_load(void)
2974 struct rq
*this = this_rq();
2975 return this->cpu_load
[0];
2979 /* Variables and functions for calc_load */
2980 static atomic_long_t calc_load_tasks
;
2981 static unsigned long calc_load_update
;
2982 unsigned long avenrun
[3];
2983 EXPORT_SYMBOL(avenrun
);
2986 * get_avenrun - get the load average array
2987 * @loads: pointer to dest load array
2988 * @offset: offset to add
2989 * @shift: shift count to shift the result left
2991 * These values are estimates at best, so no need for locking.
2993 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2995 loads
[0] = (avenrun
[0] + offset
) << shift
;
2996 loads
[1] = (avenrun
[1] + offset
) << shift
;
2997 loads
[2] = (avenrun
[2] + offset
) << shift
;
3000 static unsigned long
3001 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3004 load
+= active
* (FIXED_1
- exp
);
3005 return load
>> FSHIFT
;
3009 * calc_load - update the avenrun load estimates 10 ticks after the
3010 * CPUs have updated calc_load_tasks.
3012 void calc_global_load(void)
3014 unsigned long upd
= calc_load_update
+ 10;
3017 if (time_before(jiffies
, upd
))
3020 active
= atomic_long_read(&calc_load_tasks
);
3021 active
= active
> 0 ? active
* FIXED_1
: 0;
3023 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3024 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3025 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3027 calc_load_update
+= LOAD_FREQ
;
3031 * Either called from update_cpu_load() or from a cpu going idle
3033 static void calc_load_account_active(struct rq
*this_rq
)
3035 long nr_active
, delta
;
3037 nr_active
= this_rq
->nr_running
;
3038 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3040 if (nr_active
!= this_rq
->calc_load_active
) {
3041 delta
= nr_active
- this_rq
->calc_load_active
;
3042 this_rq
->calc_load_active
= nr_active
;
3043 atomic_long_add(delta
, &calc_load_tasks
);
3048 * Update rq->cpu_load[] statistics. This function is usually called every
3049 * scheduler tick (TICK_NSEC).
3051 static void update_cpu_load(struct rq
*this_rq
)
3053 unsigned long this_load
= this_rq
->load
.weight
;
3056 this_rq
->nr_load_updates
++;
3058 /* Update our load: */
3059 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3060 unsigned long old_load
, new_load
;
3062 /* scale is effectively 1 << i now, and >> i divides by scale */
3064 old_load
= this_rq
->cpu_load
[i
];
3065 new_load
= this_load
;
3067 * Round up the averaging division if load is increasing. This
3068 * prevents us from getting stuck on 9 if the load is 10, for
3071 if (new_load
> old_load
)
3072 new_load
+= scale
-1;
3073 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3076 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3077 this_rq
->calc_load_update
+= LOAD_FREQ
;
3078 calc_load_account_active(this_rq
);
3085 * sched_exec - execve() is a valuable balancing opportunity, because at
3086 * this point the task has the smallest effective memory and cache footprint.
3088 void sched_exec(void)
3090 struct task_struct
*p
= current
;
3091 struct migration_req req
;
3092 int dest_cpu
, this_cpu
;
3093 unsigned long flags
;
3097 this_cpu
= get_cpu();
3098 dest_cpu
= select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3099 if (dest_cpu
== this_cpu
) {
3104 rq
= task_rq_lock(p
, &flags
);
3108 * select_task_rq() can race against ->cpus_allowed
3110 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3111 || unlikely(!cpu_active(dest_cpu
))) {
3112 task_rq_unlock(rq
, &flags
);
3116 /* force the process onto the specified CPU */
3117 if (migrate_task(p
, dest_cpu
, &req
)) {
3118 /* Need to wait for migration thread (might exit: take ref). */
3119 struct task_struct
*mt
= rq
->migration_thread
;
3121 get_task_struct(mt
);
3122 task_rq_unlock(rq
, &flags
);
3123 wake_up_process(mt
);
3124 put_task_struct(mt
);
3125 wait_for_completion(&req
.done
);
3129 task_rq_unlock(rq
, &flags
);
3134 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3136 EXPORT_PER_CPU_SYMBOL(kstat
);
3139 * Return any ns on the sched_clock that have not yet been accounted in
3140 * @p in case that task is currently running.
3142 * Called with task_rq_lock() held on @rq.
3144 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3148 if (task_current(rq
, p
)) {
3149 update_rq_clock(rq
);
3150 ns
= rq
->clock
- p
->se
.exec_start
;
3158 unsigned long long task_delta_exec(struct task_struct
*p
)
3160 unsigned long flags
;
3164 rq
= task_rq_lock(p
, &flags
);
3165 ns
= do_task_delta_exec(p
, rq
);
3166 task_rq_unlock(rq
, &flags
);
3172 * Return accounted runtime for the task.
3173 * In case the task is currently running, return the runtime plus current's
3174 * pending runtime that have not been accounted yet.
3176 unsigned long long task_sched_runtime(struct task_struct
*p
)
3178 unsigned long flags
;
3182 rq
= task_rq_lock(p
, &flags
);
3183 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3184 task_rq_unlock(rq
, &flags
);
3190 * Return sum_exec_runtime for the thread group.
3191 * In case the task is currently running, return the sum plus current's
3192 * pending runtime that have not been accounted yet.
3194 * Note that the thread group might have other running tasks as well,
3195 * so the return value not includes other pending runtime that other
3196 * running tasks might have.
3198 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3200 struct task_cputime totals
;
3201 unsigned long flags
;
3205 rq
= task_rq_lock(p
, &flags
);
3206 thread_group_cputime(p
, &totals
);
3207 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3208 task_rq_unlock(rq
, &flags
);
3214 * Account user cpu time to a process.
3215 * @p: the process that the cpu time gets accounted to
3216 * @cputime: the cpu time spent in user space since the last update
3217 * @cputime_scaled: cputime scaled by cpu frequency
3219 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3220 cputime_t cputime_scaled
)
3222 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3225 /* Add user time to process. */
3226 p
->utime
= cputime_add(p
->utime
, cputime
);
3227 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3228 account_group_user_time(p
, cputime
);
3230 /* Add user time to cpustat. */
3231 tmp
= cputime_to_cputime64(cputime
);
3232 if (TASK_NICE(p
) > 0)
3233 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3235 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3237 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3238 /* Account for user time used */
3239 acct_update_integrals(p
);
3243 * Account guest cpu time to a process.
3244 * @p: the process that the cpu time gets accounted to
3245 * @cputime: the cpu time spent in virtual machine since the last update
3246 * @cputime_scaled: cputime scaled by cpu frequency
3248 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3249 cputime_t cputime_scaled
)
3252 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3254 tmp
= cputime_to_cputime64(cputime
);
3256 /* Add guest time to process. */
3257 p
->utime
= cputime_add(p
->utime
, cputime
);
3258 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3259 account_group_user_time(p
, cputime
);
3260 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3262 /* Add guest time to cpustat. */
3263 if (TASK_NICE(p
) > 0) {
3264 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3265 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3267 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3268 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3273 * Account system cpu time to a process.
3274 * @p: the process that the cpu time gets accounted to
3275 * @hardirq_offset: the offset to subtract from hardirq_count()
3276 * @cputime: the cpu time spent in kernel space since the last update
3277 * @cputime_scaled: cputime scaled by cpu frequency
3279 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3280 cputime_t cputime
, cputime_t cputime_scaled
)
3282 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3285 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3286 account_guest_time(p
, cputime
, cputime_scaled
);
3290 /* Add system time to process. */
3291 p
->stime
= cputime_add(p
->stime
, cputime
);
3292 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3293 account_group_system_time(p
, cputime
);
3295 /* Add system time to cpustat. */
3296 tmp
= cputime_to_cputime64(cputime
);
3297 if (hardirq_count() - hardirq_offset
)
3298 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3299 else if (softirq_count())
3300 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3302 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3304 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3306 /* Account for system time used */
3307 acct_update_integrals(p
);
3311 * Account for involuntary wait time.
3312 * @steal: the cpu time spent in involuntary wait
3314 void account_steal_time(cputime_t cputime
)
3316 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3317 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3319 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3323 * Account for idle time.
3324 * @cputime: the cpu time spent in idle wait
3326 void account_idle_time(cputime_t cputime
)
3328 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3329 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3330 struct rq
*rq
= this_rq();
3332 if (atomic_read(&rq
->nr_iowait
) > 0)
3333 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3335 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3338 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3341 * Account a single tick of cpu time.
3342 * @p: the process that the cpu time gets accounted to
3343 * @user_tick: indicates if the tick is a user or a system tick
3345 void account_process_tick(struct task_struct
*p
, int user_tick
)
3347 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3348 struct rq
*rq
= this_rq();
3351 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3352 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3353 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3356 account_idle_time(cputime_one_jiffy
);
3360 * Account multiple ticks of steal time.
3361 * @p: the process from which the cpu time has been stolen
3362 * @ticks: number of stolen ticks
3364 void account_steal_ticks(unsigned long ticks
)
3366 account_steal_time(jiffies_to_cputime(ticks
));
3370 * Account multiple ticks of idle time.
3371 * @ticks: number of stolen ticks
3373 void account_idle_ticks(unsigned long ticks
)
3375 account_idle_time(jiffies_to_cputime(ticks
));
3381 * Use precise platform statistics if available:
3383 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3384 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3390 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3392 struct task_cputime cputime
;
3394 thread_group_cputime(p
, &cputime
);
3396 *ut
= cputime
.utime
;
3397 *st
= cputime
.stime
;
3401 #ifndef nsecs_to_cputime
3402 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3405 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3407 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3410 * Use CFS's precise accounting:
3412 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3417 temp
= (u64
)(rtime
* utime
);
3418 do_div(temp
, total
);
3419 utime
= (cputime_t
)temp
;
3424 * Compare with previous values, to keep monotonicity:
3426 p
->prev_utime
= max(p
->prev_utime
, utime
);
3427 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3429 *ut
= p
->prev_utime
;
3430 *st
= p
->prev_stime
;
3434 * Must be called with siglock held.
3436 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3438 struct signal_struct
*sig
= p
->signal
;
3439 struct task_cputime cputime
;
3440 cputime_t rtime
, utime
, total
;
3442 thread_group_cputime(p
, &cputime
);
3444 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3445 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3450 temp
= (u64
)(rtime
* cputime
.utime
);
3451 do_div(temp
, total
);
3452 utime
= (cputime_t
)temp
;
3456 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3457 sig
->prev_stime
= max(sig
->prev_stime
,
3458 cputime_sub(rtime
, sig
->prev_utime
));
3460 *ut
= sig
->prev_utime
;
3461 *st
= sig
->prev_stime
;
3466 * This function gets called by the timer code, with HZ frequency.
3467 * We call it with interrupts disabled.
3469 * It also gets called by the fork code, when changing the parent's
3472 void scheduler_tick(void)
3474 int cpu
= smp_processor_id();
3475 struct rq
*rq
= cpu_rq(cpu
);
3476 struct task_struct
*curr
= rq
->curr
;
3480 raw_spin_lock(&rq
->lock
);
3481 update_rq_clock(rq
);
3482 update_cpu_load(rq
);
3483 curr
->sched_class
->task_tick(rq
, curr
, 0);
3484 raw_spin_unlock(&rq
->lock
);
3486 perf_event_task_tick(curr
);
3489 rq
->idle_at_tick
= idle_cpu(cpu
);
3490 trigger_load_balance(rq
, cpu
);
3494 notrace
unsigned long get_parent_ip(unsigned long addr
)
3496 if (in_lock_functions(addr
)) {
3497 addr
= CALLER_ADDR2
;
3498 if (in_lock_functions(addr
))
3499 addr
= CALLER_ADDR3
;
3504 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3505 defined(CONFIG_PREEMPT_TRACER))
3507 void __kprobes
add_preempt_count(int val
)
3509 #ifdef CONFIG_DEBUG_PREEMPT
3513 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3516 preempt_count() += val
;
3517 #ifdef CONFIG_DEBUG_PREEMPT
3519 * Spinlock count overflowing soon?
3521 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3524 if (preempt_count() == val
)
3525 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3527 EXPORT_SYMBOL(add_preempt_count
);
3529 void __kprobes
sub_preempt_count(int val
)
3531 #ifdef CONFIG_DEBUG_PREEMPT
3535 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3538 * Is the spinlock portion underflowing?
3540 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3541 !(preempt_count() & PREEMPT_MASK
)))
3545 if (preempt_count() == val
)
3546 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3547 preempt_count() -= val
;
3549 EXPORT_SYMBOL(sub_preempt_count
);
3554 * Print scheduling while atomic bug:
3556 static noinline
void __schedule_bug(struct task_struct
*prev
)
3558 struct pt_regs
*regs
= get_irq_regs();
3560 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3561 prev
->comm
, prev
->pid
, preempt_count());
3563 debug_show_held_locks(prev
);
3565 if (irqs_disabled())
3566 print_irqtrace_events(prev
);
3575 * Various schedule()-time debugging checks and statistics:
3577 static inline void schedule_debug(struct task_struct
*prev
)
3580 * Test if we are atomic. Since do_exit() needs to call into
3581 * schedule() atomically, we ignore that path for now.
3582 * Otherwise, whine if we are scheduling when we should not be.
3584 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3585 __schedule_bug(prev
);
3587 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3589 schedstat_inc(this_rq(), sched_count
);
3590 #ifdef CONFIG_SCHEDSTATS
3591 if (unlikely(prev
->lock_depth
>= 0)) {
3592 schedstat_inc(this_rq(), bkl_count
);
3593 schedstat_inc(prev
, sched_info
.bkl_count
);
3598 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3600 if (prev
->state
== TASK_RUNNING
) {
3601 u64 runtime
= prev
->se
.sum_exec_runtime
;
3603 runtime
-= prev
->se
.prev_sum_exec_runtime
;
3604 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
3607 * In order to avoid avg_overlap growing stale when we are
3608 * indeed overlapping and hence not getting put to sleep, grow
3609 * the avg_overlap on preemption.
3611 * We use the average preemption runtime because that
3612 * correlates to the amount of cache footprint a task can
3615 update_avg(&prev
->se
.avg_overlap
, runtime
);
3617 prev
->sched_class
->put_prev_task(rq
, prev
);
3621 * Pick up the highest-prio task:
3623 static inline struct task_struct
*
3624 pick_next_task(struct rq
*rq
)
3626 const struct sched_class
*class;
3627 struct task_struct
*p
;
3630 * Optimization: we know that if all tasks are in
3631 * the fair class we can call that function directly:
3633 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3634 p
= fair_sched_class
.pick_next_task(rq
);
3639 class = sched_class_highest
;
3641 p
= class->pick_next_task(rq
);
3645 * Will never be NULL as the idle class always
3646 * returns a non-NULL p:
3648 class = class->next
;
3653 * schedule() is the main scheduler function.
3655 asmlinkage
void __sched
schedule(void)
3657 struct task_struct
*prev
, *next
;
3658 unsigned long *switch_count
;
3664 cpu
= smp_processor_id();
3666 rcu_note_context_switch(cpu
);
3668 switch_count
= &prev
->nivcsw
;
3670 release_kernel_lock(prev
);
3671 need_resched_nonpreemptible
:
3673 schedule_debug(prev
);
3675 if (sched_feat(HRTICK
))
3678 raw_spin_lock_irq(&rq
->lock
);
3679 update_rq_clock(rq
);
3680 clear_tsk_need_resched(prev
);
3682 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3683 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3684 prev
->state
= TASK_RUNNING
;
3686 deactivate_task(rq
, prev
, 1);
3687 switch_count
= &prev
->nvcsw
;
3690 pre_schedule(rq
, prev
);
3692 if (unlikely(!rq
->nr_running
))
3693 idle_balance(cpu
, rq
);
3695 put_prev_task(rq
, prev
);
3696 next
= pick_next_task(rq
);
3698 if (likely(prev
!= next
)) {
3699 sched_info_switch(prev
, next
);
3700 perf_event_task_sched_out(prev
, next
);
3706 context_switch(rq
, prev
, next
); /* unlocks the rq */
3708 * the context switch might have flipped the stack from under
3709 * us, hence refresh the local variables.
3711 cpu
= smp_processor_id();
3714 raw_spin_unlock_irq(&rq
->lock
);
3718 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3720 switch_count
= &prev
->nivcsw
;
3721 goto need_resched_nonpreemptible
;
3724 preempt_enable_no_resched();
3728 EXPORT_SYMBOL(schedule
);
3730 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3732 * Look out! "owner" is an entirely speculative pointer
3733 * access and not reliable.
3735 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3740 if (!sched_feat(OWNER_SPIN
))
3743 #ifdef CONFIG_DEBUG_PAGEALLOC
3745 * Need to access the cpu field knowing that
3746 * DEBUG_PAGEALLOC could have unmapped it if
3747 * the mutex owner just released it and exited.
3749 if (probe_kernel_address(&owner
->cpu
, cpu
))
3756 * Even if the access succeeded (likely case),
3757 * the cpu field may no longer be valid.
3759 if (cpu
>= nr_cpumask_bits
)
3763 * We need to validate that we can do a
3764 * get_cpu() and that we have the percpu area.
3766 if (!cpu_online(cpu
))
3773 * Owner changed, break to re-assess state.
3775 if (lock
->owner
!= owner
)
3779 * Is that owner really running on that cpu?
3781 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3791 #ifdef CONFIG_PREEMPT
3793 * this is the entry point to schedule() from in-kernel preemption
3794 * off of preempt_enable. Kernel preemptions off return from interrupt
3795 * occur there and call schedule directly.
3797 asmlinkage
void __sched
preempt_schedule(void)
3799 struct thread_info
*ti
= current_thread_info();
3802 * If there is a non-zero preempt_count or interrupts are disabled,
3803 * we do not want to preempt the current task. Just return..
3805 if (likely(ti
->preempt_count
|| irqs_disabled()))
3809 add_preempt_count(PREEMPT_ACTIVE
);
3811 sub_preempt_count(PREEMPT_ACTIVE
);
3814 * Check again in case we missed a preemption opportunity
3815 * between schedule and now.
3818 } while (need_resched());
3820 EXPORT_SYMBOL(preempt_schedule
);
3823 * this is the entry point to schedule() from kernel preemption
3824 * off of irq context.
3825 * Note, that this is called and return with irqs disabled. This will
3826 * protect us against recursive calling from irq.
3828 asmlinkage
void __sched
preempt_schedule_irq(void)
3830 struct thread_info
*ti
= current_thread_info();
3832 /* Catch callers which need to be fixed */
3833 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3836 add_preempt_count(PREEMPT_ACTIVE
);
3839 local_irq_disable();
3840 sub_preempt_count(PREEMPT_ACTIVE
);
3843 * Check again in case we missed a preemption opportunity
3844 * between schedule and now.
3847 } while (need_resched());
3850 #endif /* CONFIG_PREEMPT */
3852 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3855 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3857 EXPORT_SYMBOL(default_wake_function
);
3860 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3861 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3862 * number) then we wake all the non-exclusive tasks and one exclusive task.
3864 * There are circumstances in which we can try to wake a task which has already
3865 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3866 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3868 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3869 int nr_exclusive
, int wake_flags
, void *key
)
3871 wait_queue_t
*curr
, *next
;
3873 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3874 unsigned flags
= curr
->flags
;
3876 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3877 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3883 * __wake_up - wake up threads blocked on a waitqueue.
3885 * @mode: which threads
3886 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3887 * @key: is directly passed to the wakeup function
3889 * It may be assumed that this function implies a write memory barrier before
3890 * changing the task state if and only if any tasks are woken up.
3892 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3893 int nr_exclusive
, void *key
)
3895 unsigned long flags
;
3897 spin_lock_irqsave(&q
->lock
, flags
);
3898 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3899 spin_unlock_irqrestore(&q
->lock
, flags
);
3901 EXPORT_SYMBOL(__wake_up
);
3904 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3906 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3908 __wake_up_common(q
, mode
, 1, 0, NULL
);
3911 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3913 __wake_up_common(q
, mode
, 1, 0, key
);
3917 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3919 * @mode: which threads
3920 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3921 * @key: opaque value to be passed to wakeup targets
3923 * The sync wakeup differs that the waker knows that it will schedule
3924 * away soon, so while the target thread will be woken up, it will not
3925 * be migrated to another CPU - ie. the two threads are 'synchronized'
3926 * with each other. This can prevent needless bouncing between CPUs.
3928 * On UP it can prevent extra preemption.
3930 * It may be assumed that this function implies a write memory barrier before
3931 * changing the task state if and only if any tasks are woken up.
3933 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3934 int nr_exclusive
, void *key
)
3936 unsigned long flags
;
3937 int wake_flags
= WF_SYNC
;
3942 if (unlikely(!nr_exclusive
))
3945 spin_lock_irqsave(&q
->lock
, flags
);
3946 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3947 spin_unlock_irqrestore(&q
->lock
, flags
);
3949 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3952 * __wake_up_sync - see __wake_up_sync_key()
3954 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3956 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3958 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3961 * complete: - signals a single thread waiting on this completion
3962 * @x: holds the state of this particular completion
3964 * This will wake up a single thread waiting on this completion. Threads will be
3965 * awakened in the same order in which they were queued.
3967 * See also complete_all(), wait_for_completion() and related routines.
3969 * It may be assumed that this function implies a write memory barrier before
3970 * changing the task state if and only if any tasks are woken up.
3972 void complete(struct completion
*x
)
3974 unsigned long flags
;
3976 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3978 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3979 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3981 EXPORT_SYMBOL(complete
);
3984 * complete_all: - signals all threads waiting on this completion
3985 * @x: holds the state of this particular completion
3987 * This will wake up all threads waiting on this particular completion event.
3989 * It may be assumed that this function implies a write memory barrier before
3990 * changing the task state if and only if any tasks are woken up.
3992 void complete_all(struct completion
*x
)
3994 unsigned long flags
;
3996 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3997 x
->done
+= UINT_MAX
/2;
3998 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3999 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4001 EXPORT_SYMBOL(complete_all
);
4003 static inline long __sched
4004 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4007 DECLARE_WAITQUEUE(wait
, current
);
4009 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4010 __add_wait_queue_tail(&x
->wait
, &wait
);
4012 if (signal_pending_state(state
, current
)) {
4013 timeout
= -ERESTARTSYS
;
4016 __set_current_state(state
);
4017 spin_unlock_irq(&x
->wait
.lock
);
4018 timeout
= schedule_timeout(timeout
);
4019 spin_lock_irq(&x
->wait
.lock
);
4020 } while (!x
->done
&& timeout
);
4021 __remove_wait_queue(&x
->wait
, &wait
);
4026 return timeout
?: 1;
4030 wait_for_common(struct completion
*x
, long timeout
, int state
)
4034 spin_lock_irq(&x
->wait
.lock
);
4035 timeout
= do_wait_for_common(x
, timeout
, state
);
4036 spin_unlock_irq(&x
->wait
.lock
);
4041 * wait_for_completion: - waits for completion of a task
4042 * @x: holds the state of this particular completion
4044 * This waits to be signaled for completion of a specific task. It is NOT
4045 * interruptible and there is no timeout.
4047 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4048 * and interrupt capability. Also see complete().
4050 void __sched
wait_for_completion(struct completion
*x
)
4052 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4054 EXPORT_SYMBOL(wait_for_completion
);
4057 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4058 * @x: holds the state of this particular completion
4059 * @timeout: timeout value in jiffies
4061 * This waits for either a completion of a specific task to be signaled or for a
4062 * specified timeout to expire. The timeout is in jiffies. It is not
4065 unsigned long __sched
4066 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4068 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4070 EXPORT_SYMBOL(wait_for_completion_timeout
);
4073 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4074 * @x: holds the state of this particular completion
4076 * This waits for completion of a specific task to be signaled. It is
4079 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4081 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4082 if (t
== -ERESTARTSYS
)
4086 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4089 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4090 * @x: holds the state of this particular completion
4091 * @timeout: timeout value in jiffies
4093 * This waits for either a completion of a specific task to be signaled or for a
4094 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4096 unsigned long __sched
4097 wait_for_completion_interruptible_timeout(struct completion
*x
,
4098 unsigned long timeout
)
4100 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4102 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4105 * wait_for_completion_killable: - waits for completion of a task (killable)
4106 * @x: holds the state of this particular completion
4108 * This waits to be signaled for completion of a specific task. It can be
4109 * interrupted by a kill signal.
4111 int __sched
wait_for_completion_killable(struct completion
*x
)
4113 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4114 if (t
== -ERESTARTSYS
)
4118 EXPORT_SYMBOL(wait_for_completion_killable
);
4121 * try_wait_for_completion - try to decrement a completion without blocking
4122 * @x: completion structure
4124 * Returns: 0 if a decrement cannot be done without blocking
4125 * 1 if a decrement succeeded.
4127 * If a completion is being used as a counting completion,
4128 * attempt to decrement the counter without blocking. This
4129 * enables us to avoid waiting if the resource the completion
4130 * is protecting is not available.
4132 bool try_wait_for_completion(struct completion
*x
)
4134 unsigned long flags
;
4137 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4142 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4145 EXPORT_SYMBOL(try_wait_for_completion
);
4148 * completion_done - Test to see if a completion has any waiters
4149 * @x: completion structure
4151 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4152 * 1 if there are no waiters.
4155 bool completion_done(struct completion
*x
)
4157 unsigned long flags
;
4160 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4163 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4166 EXPORT_SYMBOL(completion_done
);
4169 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4171 unsigned long flags
;
4174 init_waitqueue_entry(&wait
, current
);
4176 __set_current_state(state
);
4178 spin_lock_irqsave(&q
->lock
, flags
);
4179 __add_wait_queue(q
, &wait
);
4180 spin_unlock(&q
->lock
);
4181 timeout
= schedule_timeout(timeout
);
4182 spin_lock_irq(&q
->lock
);
4183 __remove_wait_queue(q
, &wait
);
4184 spin_unlock_irqrestore(&q
->lock
, flags
);
4189 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4191 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4193 EXPORT_SYMBOL(interruptible_sleep_on
);
4196 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4198 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4200 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4202 void __sched
sleep_on(wait_queue_head_t
*q
)
4204 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4206 EXPORT_SYMBOL(sleep_on
);
4208 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4210 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4212 EXPORT_SYMBOL(sleep_on_timeout
);
4214 #ifdef CONFIG_RT_MUTEXES
4217 * rt_mutex_setprio - set the current priority of a task
4219 * @prio: prio value (kernel-internal form)
4221 * This function changes the 'effective' priority of a task. It does
4222 * not touch ->normal_prio like __setscheduler().
4224 * Used by the rt_mutex code to implement priority inheritance logic.
4226 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4228 unsigned long flags
;
4229 int oldprio
, on_rq
, running
;
4231 const struct sched_class
*prev_class
;
4233 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4235 rq
= task_rq_lock(p
, &flags
);
4236 update_rq_clock(rq
);
4239 prev_class
= p
->sched_class
;
4240 on_rq
= p
->se
.on_rq
;
4241 running
= task_current(rq
, p
);
4243 dequeue_task(rq
, p
, 0);
4245 p
->sched_class
->put_prev_task(rq
, p
);
4248 p
->sched_class
= &rt_sched_class
;
4250 p
->sched_class
= &fair_sched_class
;
4255 p
->sched_class
->set_curr_task(rq
);
4257 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4259 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4261 task_rq_unlock(rq
, &flags
);
4266 void set_user_nice(struct task_struct
*p
, long nice
)
4268 int old_prio
, delta
, on_rq
;
4269 unsigned long flags
;
4272 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4275 * We have to be careful, if called from sys_setpriority(),
4276 * the task might be in the middle of scheduling on another CPU.
4278 rq
= task_rq_lock(p
, &flags
);
4279 update_rq_clock(rq
);
4281 * The RT priorities are set via sched_setscheduler(), but we still
4282 * allow the 'normal' nice value to be set - but as expected
4283 * it wont have any effect on scheduling until the task is
4284 * SCHED_FIFO/SCHED_RR:
4286 if (task_has_rt_policy(p
)) {
4287 p
->static_prio
= NICE_TO_PRIO(nice
);
4290 on_rq
= p
->se
.on_rq
;
4292 dequeue_task(rq
, p
, 0);
4294 p
->static_prio
= NICE_TO_PRIO(nice
);
4297 p
->prio
= effective_prio(p
);
4298 delta
= p
->prio
- old_prio
;
4301 enqueue_task(rq
, p
, 0, false);
4303 * If the task increased its priority or is running and
4304 * lowered its priority, then reschedule its CPU:
4306 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4307 resched_task(rq
->curr
);
4310 task_rq_unlock(rq
, &flags
);
4312 EXPORT_SYMBOL(set_user_nice
);
4315 * can_nice - check if a task can reduce its nice value
4319 int can_nice(const struct task_struct
*p
, const int nice
)
4321 /* convert nice value [19,-20] to rlimit style value [1,40] */
4322 int nice_rlim
= 20 - nice
;
4324 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4325 capable(CAP_SYS_NICE
));
4328 #ifdef __ARCH_WANT_SYS_NICE
4331 * sys_nice - change the priority of the current process.
4332 * @increment: priority increment
4334 * sys_setpriority is a more generic, but much slower function that
4335 * does similar things.
4337 SYSCALL_DEFINE1(nice
, int, increment
)
4342 * Setpriority might change our priority at the same moment.
4343 * We don't have to worry. Conceptually one call occurs first
4344 * and we have a single winner.
4346 if (increment
< -40)
4351 nice
= TASK_NICE(current
) + increment
;
4357 if (increment
< 0 && !can_nice(current
, nice
))
4360 retval
= security_task_setnice(current
, nice
);
4364 set_user_nice(current
, nice
);
4371 * task_prio - return the priority value of a given task.
4372 * @p: the task in question.
4374 * This is the priority value as seen by users in /proc.
4375 * RT tasks are offset by -200. Normal tasks are centered
4376 * around 0, value goes from -16 to +15.
4378 int task_prio(const struct task_struct
*p
)
4380 return p
->prio
- MAX_RT_PRIO
;
4384 * task_nice - return the nice value of a given task.
4385 * @p: the task in question.
4387 int task_nice(const struct task_struct
*p
)
4389 return TASK_NICE(p
);
4391 EXPORT_SYMBOL(task_nice
);
4394 * idle_cpu - is a given cpu idle currently?
4395 * @cpu: the processor in question.
4397 int idle_cpu(int cpu
)
4399 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4403 * idle_task - return the idle task for a given cpu.
4404 * @cpu: the processor in question.
4406 struct task_struct
*idle_task(int cpu
)
4408 return cpu_rq(cpu
)->idle
;
4412 * find_process_by_pid - find a process with a matching PID value.
4413 * @pid: the pid in question.
4415 static struct task_struct
*find_process_by_pid(pid_t pid
)
4417 return pid
? find_task_by_vpid(pid
) : current
;
4420 /* Actually do priority change: must hold rq lock. */
4422 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4424 BUG_ON(p
->se
.on_rq
);
4427 p
->rt_priority
= prio
;
4428 p
->normal_prio
= normal_prio(p
);
4429 /* we are holding p->pi_lock already */
4430 p
->prio
= rt_mutex_getprio(p
);
4431 if (rt_prio(p
->prio
))
4432 p
->sched_class
= &rt_sched_class
;
4434 p
->sched_class
= &fair_sched_class
;
4439 * check the target process has a UID that matches the current process's
4441 static bool check_same_owner(struct task_struct
*p
)
4443 const struct cred
*cred
= current_cred(), *pcred
;
4447 pcred
= __task_cred(p
);
4448 match
= (cred
->euid
== pcred
->euid
||
4449 cred
->euid
== pcred
->uid
);
4454 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4455 struct sched_param
*param
, bool user
)
4457 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4458 unsigned long flags
;
4459 const struct sched_class
*prev_class
;
4463 /* may grab non-irq protected spin_locks */
4464 BUG_ON(in_interrupt());
4466 /* double check policy once rq lock held */
4468 reset_on_fork
= p
->sched_reset_on_fork
;
4469 policy
= oldpolicy
= p
->policy
;
4471 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4472 policy
&= ~SCHED_RESET_ON_FORK
;
4474 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4475 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4476 policy
!= SCHED_IDLE
)
4481 * Valid priorities for SCHED_FIFO and SCHED_RR are
4482 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4483 * SCHED_BATCH and SCHED_IDLE is 0.
4485 if (param
->sched_priority
< 0 ||
4486 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4487 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4489 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4493 * Allow unprivileged RT tasks to decrease priority:
4495 if (user
&& !capable(CAP_SYS_NICE
)) {
4496 if (rt_policy(policy
)) {
4497 unsigned long rlim_rtprio
;
4499 if (!lock_task_sighand(p
, &flags
))
4501 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4502 unlock_task_sighand(p
, &flags
);
4504 /* can't set/change the rt policy */
4505 if (policy
!= p
->policy
&& !rlim_rtprio
)
4508 /* can't increase priority */
4509 if (param
->sched_priority
> p
->rt_priority
&&
4510 param
->sched_priority
> rlim_rtprio
)
4514 * Like positive nice levels, dont allow tasks to
4515 * move out of SCHED_IDLE either:
4517 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4520 /* can't change other user's priorities */
4521 if (!check_same_owner(p
))
4524 /* Normal users shall not reset the sched_reset_on_fork flag */
4525 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4530 #ifdef CONFIG_RT_GROUP_SCHED
4532 * Do not allow realtime tasks into groups that have no runtime
4535 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4536 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4540 retval
= security_task_setscheduler(p
, policy
, param
);
4546 * make sure no PI-waiters arrive (or leave) while we are
4547 * changing the priority of the task:
4549 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4551 * To be able to change p->policy safely, the apropriate
4552 * runqueue lock must be held.
4554 rq
= __task_rq_lock(p
);
4555 /* recheck policy now with rq lock held */
4556 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4557 policy
= oldpolicy
= -1;
4558 __task_rq_unlock(rq
);
4559 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4562 update_rq_clock(rq
);
4563 on_rq
= p
->se
.on_rq
;
4564 running
= task_current(rq
, p
);
4566 deactivate_task(rq
, p
, 0);
4568 p
->sched_class
->put_prev_task(rq
, p
);
4570 p
->sched_reset_on_fork
= reset_on_fork
;
4573 prev_class
= p
->sched_class
;
4574 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4577 p
->sched_class
->set_curr_task(rq
);
4579 activate_task(rq
, p
, 0);
4581 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4583 __task_rq_unlock(rq
);
4584 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4586 rt_mutex_adjust_pi(p
);
4592 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4593 * @p: the task in question.
4594 * @policy: new policy.
4595 * @param: structure containing the new RT priority.
4597 * NOTE that the task may be already dead.
4599 int sched_setscheduler(struct task_struct
*p
, int policy
,
4600 struct sched_param
*param
)
4602 return __sched_setscheduler(p
, policy
, param
, true);
4604 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4607 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4608 * @p: the task in question.
4609 * @policy: new policy.
4610 * @param: structure containing the new RT priority.
4612 * Just like sched_setscheduler, only don't bother checking if the
4613 * current context has permission. For example, this is needed in
4614 * stop_machine(): we create temporary high priority worker threads,
4615 * but our caller might not have that capability.
4617 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4618 struct sched_param
*param
)
4620 return __sched_setscheduler(p
, policy
, param
, false);
4624 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4626 struct sched_param lparam
;
4627 struct task_struct
*p
;
4630 if (!param
|| pid
< 0)
4632 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4637 p
= find_process_by_pid(pid
);
4639 retval
= sched_setscheduler(p
, policy
, &lparam
);
4646 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4647 * @pid: the pid in question.
4648 * @policy: new policy.
4649 * @param: structure containing the new RT priority.
4651 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4652 struct sched_param __user
*, param
)
4654 /* negative values for policy are not valid */
4658 return do_sched_setscheduler(pid
, policy
, param
);
4662 * sys_sched_setparam - set/change the RT priority of a thread
4663 * @pid: the pid in question.
4664 * @param: structure containing the new RT priority.
4666 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4668 return do_sched_setscheduler(pid
, -1, param
);
4672 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4673 * @pid: the pid in question.
4675 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4677 struct task_struct
*p
;
4685 p
= find_process_by_pid(pid
);
4687 retval
= security_task_getscheduler(p
);
4690 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4697 * sys_sched_getparam - get the RT priority of a thread
4698 * @pid: the pid in question.
4699 * @param: structure containing the RT priority.
4701 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4703 struct sched_param lp
;
4704 struct task_struct
*p
;
4707 if (!param
|| pid
< 0)
4711 p
= find_process_by_pid(pid
);
4716 retval
= security_task_getscheduler(p
);
4720 lp
.sched_priority
= p
->rt_priority
;
4724 * This one might sleep, we cannot do it with a spinlock held ...
4726 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4735 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4737 cpumask_var_t cpus_allowed
, new_mask
;
4738 struct task_struct
*p
;
4744 p
= find_process_by_pid(pid
);
4751 /* Prevent p going away */
4755 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4759 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4761 goto out_free_cpus_allowed
;
4764 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4767 retval
= security_task_setscheduler(p
, 0, NULL
);
4771 cpuset_cpus_allowed(p
, cpus_allowed
);
4772 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4774 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4777 cpuset_cpus_allowed(p
, cpus_allowed
);
4778 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4780 * We must have raced with a concurrent cpuset
4781 * update. Just reset the cpus_allowed to the
4782 * cpuset's cpus_allowed
4784 cpumask_copy(new_mask
, cpus_allowed
);
4789 free_cpumask_var(new_mask
);
4790 out_free_cpus_allowed
:
4791 free_cpumask_var(cpus_allowed
);
4798 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4799 struct cpumask
*new_mask
)
4801 if (len
< cpumask_size())
4802 cpumask_clear(new_mask
);
4803 else if (len
> cpumask_size())
4804 len
= cpumask_size();
4806 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4810 * sys_sched_setaffinity - set the cpu affinity of a process
4811 * @pid: pid of the process
4812 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4813 * @user_mask_ptr: user-space pointer to the new cpu mask
4815 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4816 unsigned long __user
*, user_mask_ptr
)
4818 cpumask_var_t new_mask
;
4821 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4824 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4826 retval
= sched_setaffinity(pid
, new_mask
);
4827 free_cpumask_var(new_mask
);
4831 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4833 struct task_struct
*p
;
4834 unsigned long flags
;
4842 p
= find_process_by_pid(pid
);
4846 retval
= security_task_getscheduler(p
);
4850 rq
= task_rq_lock(p
, &flags
);
4851 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4852 task_rq_unlock(rq
, &flags
);
4862 * sys_sched_getaffinity - get the cpu affinity of a process
4863 * @pid: pid of the process
4864 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4865 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4867 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4868 unsigned long __user
*, user_mask_ptr
)
4873 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4875 if (len
& (sizeof(unsigned long)-1))
4878 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4881 ret
= sched_getaffinity(pid
, mask
);
4883 size_t retlen
= min_t(size_t, len
, cpumask_size());
4885 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4890 free_cpumask_var(mask
);
4896 * sys_sched_yield - yield the current processor to other threads.
4898 * This function yields the current CPU to other tasks. If there are no
4899 * other threads running on this CPU then this function will return.
4901 SYSCALL_DEFINE0(sched_yield
)
4903 struct rq
*rq
= this_rq_lock();
4905 schedstat_inc(rq
, yld_count
);
4906 current
->sched_class
->yield_task(rq
);
4909 * Since we are going to call schedule() anyway, there's
4910 * no need to preempt or enable interrupts:
4912 __release(rq
->lock
);
4913 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4914 do_raw_spin_unlock(&rq
->lock
);
4915 preempt_enable_no_resched();
4922 static inline int should_resched(void)
4924 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4927 static void __cond_resched(void)
4929 add_preempt_count(PREEMPT_ACTIVE
);
4931 sub_preempt_count(PREEMPT_ACTIVE
);
4934 int __sched
_cond_resched(void)
4936 if (should_resched()) {
4942 EXPORT_SYMBOL(_cond_resched
);
4945 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4946 * call schedule, and on return reacquire the lock.
4948 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4949 * operations here to prevent schedule() from being called twice (once via
4950 * spin_unlock(), once by hand).
4952 int __cond_resched_lock(spinlock_t
*lock
)
4954 int resched
= should_resched();
4957 lockdep_assert_held(lock
);
4959 if (spin_needbreak(lock
) || resched
) {
4970 EXPORT_SYMBOL(__cond_resched_lock
);
4972 int __sched
__cond_resched_softirq(void)
4974 BUG_ON(!in_softirq());
4976 if (should_resched()) {
4984 EXPORT_SYMBOL(__cond_resched_softirq
);
4987 * yield - yield the current processor to other threads.
4989 * This is a shortcut for kernel-space yielding - it marks the
4990 * thread runnable and calls sys_sched_yield().
4992 void __sched
yield(void)
4994 set_current_state(TASK_RUNNING
);
4997 EXPORT_SYMBOL(yield
);
5000 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5001 * that process accounting knows that this is a task in IO wait state.
5003 void __sched
io_schedule(void)
5005 struct rq
*rq
= raw_rq();
5007 delayacct_blkio_start();
5008 atomic_inc(&rq
->nr_iowait
);
5009 current
->in_iowait
= 1;
5011 current
->in_iowait
= 0;
5012 atomic_dec(&rq
->nr_iowait
);
5013 delayacct_blkio_end();
5015 EXPORT_SYMBOL(io_schedule
);
5017 long __sched
io_schedule_timeout(long timeout
)
5019 struct rq
*rq
= raw_rq();
5022 delayacct_blkio_start();
5023 atomic_inc(&rq
->nr_iowait
);
5024 current
->in_iowait
= 1;
5025 ret
= schedule_timeout(timeout
);
5026 current
->in_iowait
= 0;
5027 atomic_dec(&rq
->nr_iowait
);
5028 delayacct_blkio_end();
5033 * sys_sched_get_priority_max - return maximum RT priority.
5034 * @policy: scheduling class.
5036 * this syscall returns the maximum rt_priority that can be used
5037 * by a given scheduling class.
5039 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5046 ret
= MAX_USER_RT_PRIO
-1;
5058 * sys_sched_get_priority_min - return minimum RT priority.
5059 * @policy: scheduling class.
5061 * this syscall returns the minimum rt_priority that can be used
5062 * by a given scheduling class.
5064 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5082 * sys_sched_rr_get_interval - return the default timeslice of a process.
5083 * @pid: pid of the process.
5084 * @interval: userspace pointer to the timeslice value.
5086 * this syscall writes the default timeslice value of a given process
5087 * into the user-space timespec buffer. A value of '0' means infinity.
5089 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5090 struct timespec __user
*, interval
)
5092 struct task_struct
*p
;
5093 unsigned int time_slice
;
5094 unsigned long flags
;
5104 p
= find_process_by_pid(pid
);
5108 retval
= security_task_getscheduler(p
);
5112 rq
= task_rq_lock(p
, &flags
);
5113 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5114 task_rq_unlock(rq
, &flags
);
5117 jiffies_to_timespec(time_slice
, &t
);
5118 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5126 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5128 void sched_show_task(struct task_struct
*p
)
5130 unsigned long free
= 0;
5133 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5134 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5135 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5136 #if BITS_PER_LONG == 32
5137 if (state
== TASK_RUNNING
)
5138 printk(KERN_CONT
" running ");
5140 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5142 if (state
== TASK_RUNNING
)
5143 printk(KERN_CONT
" running task ");
5145 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5147 #ifdef CONFIG_DEBUG_STACK_USAGE
5148 free
= stack_not_used(p
);
5150 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5151 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5152 (unsigned long)task_thread_info(p
)->flags
);
5154 show_stack(p
, NULL
);
5157 void show_state_filter(unsigned long state_filter
)
5159 struct task_struct
*g
, *p
;
5161 #if BITS_PER_LONG == 32
5163 " task PC stack pid father\n");
5166 " task PC stack pid father\n");
5168 read_lock(&tasklist_lock
);
5169 do_each_thread(g
, p
) {
5171 * reset the NMI-timeout, listing all files on a slow
5172 * console might take alot of time:
5174 touch_nmi_watchdog();
5175 if (!state_filter
|| (p
->state
& state_filter
))
5177 } while_each_thread(g
, p
);
5179 touch_all_softlockup_watchdogs();
5181 #ifdef CONFIG_SCHED_DEBUG
5182 sysrq_sched_debug_show();
5184 read_unlock(&tasklist_lock
);
5186 * Only show locks if all tasks are dumped:
5189 debug_show_all_locks();
5192 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5194 idle
->sched_class
= &idle_sched_class
;
5198 * init_idle - set up an idle thread for a given CPU
5199 * @idle: task in question
5200 * @cpu: cpu the idle task belongs to
5202 * NOTE: this function does not set the idle thread's NEED_RESCHED
5203 * flag, to make booting more robust.
5205 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5207 struct rq
*rq
= cpu_rq(cpu
);
5208 unsigned long flags
;
5210 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5213 idle
->state
= TASK_RUNNING
;
5214 idle
->se
.exec_start
= sched_clock();
5216 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5217 __set_task_cpu(idle
, cpu
);
5219 rq
->curr
= rq
->idle
= idle
;
5220 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5223 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5225 /* Set the preempt count _outside_ the spinlocks! */
5226 #if defined(CONFIG_PREEMPT)
5227 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5229 task_thread_info(idle
)->preempt_count
= 0;
5232 * The idle tasks have their own, simple scheduling class:
5234 idle
->sched_class
= &idle_sched_class
;
5235 ftrace_graph_init_task(idle
);
5239 * In a system that switches off the HZ timer nohz_cpu_mask
5240 * indicates which cpus entered this state. This is used
5241 * in the rcu update to wait only for active cpus. For system
5242 * which do not switch off the HZ timer nohz_cpu_mask should
5243 * always be CPU_BITS_NONE.
5245 cpumask_var_t nohz_cpu_mask
;
5248 * Increase the granularity value when there are more CPUs,
5249 * because with more CPUs the 'effective latency' as visible
5250 * to users decreases. But the relationship is not linear,
5251 * so pick a second-best guess by going with the log2 of the
5254 * This idea comes from the SD scheduler of Con Kolivas:
5256 static int get_update_sysctl_factor(void)
5258 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5259 unsigned int factor
;
5261 switch (sysctl_sched_tunable_scaling
) {
5262 case SCHED_TUNABLESCALING_NONE
:
5265 case SCHED_TUNABLESCALING_LINEAR
:
5268 case SCHED_TUNABLESCALING_LOG
:
5270 factor
= 1 + ilog2(cpus
);
5277 static void update_sysctl(void)
5279 unsigned int factor
= get_update_sysctl_factor();
5281 #define SET_SYSCTL(name) \
5282 (sysctl_##name = (factor) * normalized_sysctl_##name)
5283 SET_SYSCTL(sched_min_granularity
);
5284 SET_SYSCTL(sched_latency
);
5285 SET_SYSCTL(sched_wakeup_granularity
);
5286 SET_SYSCTL(sched_shares_ratelimit
);
5290 static inline void sched_init_granularity(void)
5297 * This is how migration works:
5299 * 1) we queue a struct migration_req structure in the source CPU's
5300 * runqueue and wake up that CPU's migration thread.
5301 * 2) we down() the locked semaphore => thread blocks.
5302 * 3) migration thread wakes up (implicitly it forces the migrated
5303 * thread off the CPU)
5304 * 4) it gets the migration request and checks whether the migrated
5305 * task is still in the wrong runqueue.
5306 * 5) if it's in the wrong runqueue then the migration thread removes
5307 * it and puts it into the right queue.
5308 * 6) migration thread up()s the semaphore.
5309 * 7) we wake up and the migration is done.
5313 * Change a given task's CPU affinity. Migrate the thread to a
5314 * proper CPU and schedule it away if the CPU it's executing on
5315 * is removed from the allowed bitmask.
5317 * NOTE: the caller must have a valid reference to the task, the
5318 * task must not exit() & deallocate itself prematurely. The
5319 * call is not atomic; no spinlocks may be held.
5321 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5323 struct migration_req req
;
5324 unsigned long flags
;
5328 rq
= task_rq_lock(p
, &flags
);
5330 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5335 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5336 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5341 if (p
->sched_class
->set_cpus_allowed
)
5342 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5344 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5345 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5348 /* Can the task run on the task's current CPU? If so, we're done */
5349 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5352 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5353 /* Need help from migration thread: drop lock and wait. */
5354 struct task_struct
*mt
= rq
->migration_thread
;
5356 get_task_struct(mt
);
5357 task_rq_unlock(rq
, &flags
);
5358 wake_up_process(mt
);
5359 put_task_struct(mt
);
5360 wait_for_completion(&req
.done
);
5361 tlb_migrate_finish(p
->mm
);
5365 task_rq_unlock(rq
, &flags
);
5369 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5372 * Move (not current) task off this cpu, onto dest cpu. We're doing
5373 * this because either it can't run here any more (set_cpus_allowed()
5374 * away from this CPU, or CPU going down), or because we're
5375 * attempting to rebalance this task on exec (sched_exec).
5377 * So we race with normal scheduler movements, but that's OK, as long
5378 * as the task is no longer on this CPU.
5380 * Returns non-zero if task was successfully migrated.
5382 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5384 struct rq
*rq_dest
, *rq_src
;
5387 if (unlikely(!cpu_active(dest_cpu
)))
5390 rq_src
= cpu_rq(src_cpu
);
5391 rq_dest
= cpu_rq(dest_cpu
);
5393 double_rq_lock(rq_src
, rq_dest
);
5394 /* Already moved. */
5395 if (task_cpu(p
) != src_cpu
)
5397 /* Affinity changed (again). */
5398 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5402 * If we're not on a rq, the next wake-up will ensure we're
5406 deactivate_task(rq_src
, p
, 0);
5407 set_task_cpu(p
, dest_cpu
);
5408 activate_task(rq_dest
, p
, 0);
5409 check_preempt_curr(rq_dest
, p
, 0);
5414 double_rq_unlock(rq_src
, rq_dest
);
5418 #define RCU_MIGRATION_IDLE 0
5419 #define RCU_MIGRATION_NEED_QS 1
5420 #define RCU_MIGRATION_GOT_QS 2
5421 #define RCU_MIGRATION_MUST_SYNC 3
5424 * migration_thread - this is a highprio system thread that performs
5425 * thread migration by bumping thread off CPU then 'pushing' onto
5428 static int migration_thread(void *data
)
5431 int cpu
= (long)data
;
5435 BUG_ON(rq
->migration_thread
!= current
);
5437 set_current_state(TASK_INTERRUPTIBLE
);
5438 while (!kthread_should_stop()) {
5439 struct migration_req
*req
;
5440 struct list_head
*head
;
5442 raw_spin_lock_irq(&rq
->lock
);
5444 if (cpu_is_offline(cpu
)) {
5445 raw_spin_unlock_irq(&rq
->lock
);
5449 if (rq
->active_balance
) {
5450 active_load_balance(rq
, cpu
);
5451 rq
->active_balance
= 0;
5454 head
= &rq
->migration_queue
;
5456 if (list_empty(head
)) {
5457 raw_spin_unlock_irq(&rq
->lock
);
5459 set_current_state(TASK_INTERRUPTIBLE
);
5462 req
= list_entry(head
->next
, struct migration_req
, list
);
5463 list_del_init(head
->next
);
5465 if (req
->task
!= NULL
) {
5466 raw_spin_unlock(&rq
->lock
);
5467 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5468 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5469 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5470 raw_spin_unlock(&rq
->lock
);
5472 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5473 raw_spin_unlock(&rq
->lock
);
5474 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5478 complete(&req
->done
);
5480 __set_current_state(TASK_RUNNING
);
5485 #ifdef CONFIG_HOTPLUG_CPU
5487 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5491 local_irq_disable();
5492 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5498 * Figure out where task on dead CPU should go, use force if necessary.
5500 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5505 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5507 /* It can have affinity changed while we were choosing. */
5508 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
5513 * While a dead CPU has no uninterruptible tasks queued at this point,
5514 * it might still have a nonzero ->nr_uninterruptible counter, because
5515 * for performance reasons the counter is not stricly tracking tasks to
5516 * their home CPUs. So we just add the counter to another CPU's counter,
5517 * to keep the global sum constant after CPU-down:
5519 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5521 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5522 unsigned long flags
;
5524 local_irq_save(flags
);
5525 double_rq_lock(rq_src
, rq_dest
);
5526 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5527 rq_src
->nr_uninterruptible
= 0;
5528 double_rq_unlock(rq_src
, rq_dest
);
5529 local_irq_restore(flags
);
5532 /* Run through task list and migrate tasks from the dead cpu. */
5533 static void migrate_live_tasks(int src_cpu
)
5535 struct task_struct
*p
, *t
;
5537 read_lock(&tasklist_lock
);
5539 do_each_thread(t
, p
) {
5543 if (task_cpu(p
) == src_cpu
)
5544 move_task_off_dead_cpu(src_cpu
, p
);
5545 } while_each_thread(t
, p
);
5547 read_unlock(&tasklist_lock
);
5551 * Schedules idle task to be the next runnable task on current CPU.
5552 * It does so by boosting its priority to highest possible.
5553 * Used by CPU offline code.
5555 void sched_idle_next(void)
5557 int this_cpu
= smp_processor_id();
5558 struct rq
*rq
= cpu_rq(this_cpu
);
5559 struct task_struct
*p
= rq
->idle
;
5560 unsigned long flags
;
5562 /* cpu has to be offline */
5563 BUG_ON(cpu_online(this_cpu
));
5566 * Strictly not necessary since rest of the CPUs are stopped by now
5567 * and interrupts disabled on the current cpu.
5569 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5571 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5573 update_rq_clock(rq
);
5574 activate_task(rq
, p
, 0);
5576 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5580 * Ensures that the idle task is using init_mm right before its cpu goes
5583 void idle_task_exit(void)
5585 struct mm_struct
*mm
= current
->active_mm
;
5587 BUG_ON(cpu_online(smp_processor_id()));
5590 switch_mm(mm
, &init_mm
, current
);
5594 /* called under rq->lock with disabled interrupts */
5595 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5597 struct rq
*rq
= cpu_rq(dead_cpu
);
5599 /* Must be exiting, otherwise would be on tasklist. */
5600 BUG_ON(!p
->exit_state
);
5602 /* Cannot have done final schedule yet: would have vanished. */
5603 BUG_ON(p
->state
== TASK_DEAD
);
5608 * Drop lock around migration; if someone else moves it,
5609 * that's OK. No task can be added to this CPU, so iteration is
5612 raw_spin_unlock_irq(&rq
->lock
);
5613 move_task_off_dead_cpu(dead_cpu
, p
);
5614 raw_spin_lock_irq(&rq
->lock
);
5619 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5620 static void migrate_dead_tasks(unsigned int dead_cpu
)
5622 struct rq
*rq
= cpu_rq(dead_cpu
);
5623 struct task_struct
*next
;
5626 if (!rq
->nr_running
)
5628 update_rq_clock(rq
);
5629 next
= pick_next_task(rq
);
5632 next
->sched_class
->put_prev_task(rq
, next
);
5633 migrate_dead(dead_cpu
, next
);
5639 * remove the tasks which were accounted by rq from calc_load_tasks.
5641 static void calc_global_load_remove(struct rq
*rq
)
5643 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5644 rq
->calc_load_active
= 0;
5646 #endif /* CONFIG_HOTPLUG_CPU */
5648 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5650 static struct ctl_table sd_ctl_dir
[] = {
5652 .procname
= "sched_domain",
5658 static struct ctl_table sd_ctl_root
[] = {
5660 .procname
= "kernel",
5662 .child
= sd_ctl_dir
,
5667 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5669 struct ctl_table
*entry
=
5670 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5675 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5677 struct ctl_table
*entry
;
5680 * In the intermediate directories, both the child directory and
5681 * procname are dynamically allocated and could fail but the mode
5682 * will always be set. In the lowest directory the names are
5683 * static strings and all have proc handlers.
5685 for (entry
= *tablep
; entry
->mode
; entry
++) {
5687 sd_free_ctl_entry(&entry
->child
);
5688 if (entry
->proc_handler
== NULL
)
5689 kfree(entry
->procname
);
5697 set_table_entry(struct ctl_table
*entry
,
5698 const char *procname
, void *data
, int maxlen
,
5699 mode_t mode
, proc_handler
*proc_handler
)
5701 entry
->procname
= procname
;
5703 entry
->maxlen
= maxlen
;
5705 entry
->proc_handler
= proc_handler
;
5708 static struct ctl_table
*
5709 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5711 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5716 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5717 sizeof(long), 0644, proc_doulongvec_minmax
);
5718 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5719 sizeof(long), 0644, proc_doulongvec_minmax
);
5720 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5721 sizeof(int), 0644, proc_dointvec_minmax
);
5722 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5723 sizeof(int), 0644, proc_dointvec_minmax
);
5724 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5725 sizeof(int), 0644, proc_dointvec_minmax
);
5726 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5727 sizeof(int), 0644, proc_dointvec_minmax
);
5728 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5729 sizeof(int), 0644, proc_dointvec_minmax
);
5730 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5731 sizeof(int), 0644, proc_dointvec_minmax
);
5732 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5733 sizeof(int), 0644, proc_dointvec_minmax
);
5734 set_table_entry(&table
[9], "cache_nice_tries",
5735 &sd
->cache_nice_tries
,
5736 sizeof(int), 0644, proc_dointvec_minmax
);
5737 set_table_entry(&table
[10], "flags", &sd
->flags
,
5738 sizeof(int), 0644, proc_dointvec_minmax
);
5739 set_table_entry(&table
[11], "name", sd
->name
,
5740 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5741 /* &table[12] is terminator */
5746 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5748 struct ctl_table
*entry
, *table
;
5749 struct sched_domain
*sd
;
5750 int domain_num
= 0, i
;
5753 for_each_domain(cpu
, sd
)
5755 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5760 for_each_domain(cpu
, sd
) {
5761 snprintf(buf
, 32, "domain%d", i
);
5762 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5764 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5771 static struct ctl_table_header
*sd_sysctl_header
;
5772 static void register_sched_domain_sysctl(void)
5774 int i
, cpu_num
= num_possible_cpus();
5775 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5778 WARN_ON(sd_ctl_dir
[0].child
);
5779 sd_ctl_dir
[0].child
= entry
;
5784 for_each_possible_cpu(i
) {
5785 snprintf(buf
, 32, "cpu%d", i
);
5786 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5788 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5792 WARN_ON(sd_sysctl_header
);
5793 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5796 /* may be called multiple times per register */
5797 static void unregister_sched_domain_sysctl(void)
5799 if (sd_sysctl_header
)
5800 unregister_sysctl_table(sd_sysctl_header
);
5801 sd_sysctl_header
= NULL
;
5802 if (sd_ctl_dir
[0].child
)
5803 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5806 static void register_sched_domain_sysctl(void)
5809 static void unregister_sched_domain_sysctl(void)
5814 static void set_rq_online(struct rq
*rq
)
5817 const struct sched_class
*class;
5819 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5822 for_each_class(class) {
5823 if (class->rq_online
)
5824 class->rq_online(rq
);
5829 static void set_rq_offline(struct rq
*rq
)
5832 const struct sched_class
*class;
5834 for_each_class(class) {
5835 if (class->rq_offline
)
5836 class->rq_offline(rq
);
5839 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5845 * migration_call - callback that gets triggered when a CPU is added.
5846 * Here we can start up the necessary migration thread for the new CPU.
5848 static int __cpuinit
5849 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5851 struct task_struct
*p
;
5852 int cpu
= (long)hcpu
;
5853 unsigned long flags
;
5858 case CPU_UP_PREPARE
:
5859 case CPU_UP_PREPARE_FROZEN
:
5860 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5863 kthread_bind(p
, cpu
);
5864 /* Must be high prio: stop_machine expects to yield to it. */
5865 rq
= task_rq_lock(p
, &flags
);
5866 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5867 task_rq_unlock(rq
, &flags
);
5869 cpu_rq(cpu
)->migration_thread
= p
;
5870 rq
->calc_load_update
= calc_load_update
;
5874 case CPU_ONLINE_FROZEN
:
5875 /* Strictly unnecessary, as first user will wake it. */
5876 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5878 /* Update our root-domain */
5880 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5882 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5886 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5889 #ifdef CONFIG_HOTPLUG_CPU
5890 case CPU_UP_CANCELED
:
5891 case CPU_UP_CANCELED_FROZEN
:
5892 if (!cpu_rq(cpu
)->migration_thread
)
5894 /* Unbind it from offline cpu so it can run. Fall thru. */
5895 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5896 cpumask_any(cpu_online_mask
));
5897 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5898 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5899 cpu_rq(cpu
)->migration_thread
= NULL
;
5903 case CPU_DEAD_FROZEN
:
5904 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5905 migrate_live_tasks(cpu
);
5907 kthread_stop(rq
->migration_thread
);
5908 put_task_struct(rq
->migration_thread
);
5909 rq
->migration_thread
= NULL
;
5910 /* Idle task back to normal (off runqueue, low prio) */
5911 raw_spin_lock_irq(&rq
->lock
);
5912 update_rq_clock(rq
);
5913 deactivate_task(rq
, rq
->idle
, 0);
5914 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5915 rq
->idle
->sched_class
= &idle_sched_class
;
5916 migrate_dead_tasks(cpu
);
5917 raw_spin_unlock_irq(&rq
->lock
);
5919 migrate_nr_uninterruptible(rq
);
5920 BUG_ON(rq
->nr_running
!= 0);
5921 calc_global_load_remove(rq
);
5923 * No need to migrate the tasks: it was best-effort if
5924 * they didn't take sched_hotcpu_mutex. Just wake up
5927 raw_spin_lock_irq(&rq
->lock
);
5928 while (!list_empty(&rq
->migration_queue
)) {
5929 struct migration_req
*req
;
5931 req
= list_entry(rq
->migration_queue
.next
,
5932 struct migration_req
, list
);
5933 list_del_init(&req
->list
);
5934 raw_spin_unlock_irq(&rq
->lock
);
5935 complete(&req
->done
);
5936 raw_spin_lock_irq(&rq
->lock
);
5938 raw_spin_unlock_irq(&rq
->lock
);
5942 case CPU_DYING_FROZEN
:
5943 /* Update our root-domain */
5945 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5947 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5950 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5958 * Register at high priority so that task migration (migrate_all_tasks)
5959 * happens before everything else. This has to be lower priority than
5960 * the notifier in the perf_event subsystem, though.
5962 static struct notifier_block __cpuinitdata migration_notifier
= {
5963 .notifier_call
= migration_call
,
5967 static int __init
migration_init(void)
5969 void *cpu
= (void *)(long)smp_processor_id();
5972 /* Start one for the boot CPU: */
5973 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5974 BUG_ON(err
== NOTIFY_BAD
);
5975 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5976 register_cpu_notifier(&migration_notifier
);
5980 early_initcall(migration_init
);
5985 #ifdef CONFIG_SCHED_DEBUG
5987 static __read_mostly
int sched_domain_debug_enabled
;
5989 static int __init
sched_domain_debug_setup(char *str
)
5991 sched_domain_debug_enabled
= 1;
5995 early_param("sched_debug", sched_domain_debug_setup
);
5997 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5998 struct cpumask
*groupmask
)
6000 struct sched_group
*group
= sd
->groups
;
6003 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6004 cpumask_clear(groupmask
);
6006 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6008 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6009 printk("does not load-balance\n");
6011 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6016 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6018 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6019 printk(KERN_ERR
"ERROR: domain->span does not contain "
6022 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6023 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6027 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6031 printk(KERN_ERR
"ERROR: group is NULL\n");
6035 if (!group
->cpu_power
) {
6036 printk(KERN_CONT
"\n");
6037 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6042 if (!cpumask_weight(sched_group_cpus(group
))) {
6043 printk(KERN_CONT
"\n");
6044 printk(KERN_ERR
"ERROR: empty group\n");
6048 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6049 printk(KERN_CONT
"\n");
6050 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6054 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6056 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6058 printk(KERN_CONT
" %s", str
);
6059 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6060 printk(KERN_CONT
" (cpu_power = %d)",
6064 group
= group
->next
;
6065 } while (group
!= sd
->groups
);
6066 printk(KERN_CONT
"\n");
6068 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6069 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6072 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6073 printk(KERN_ERR
"ERROR: parent span is not a superset "
6074 "of domain->span\n");
6078 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6080 cpumask_var_t groupmask
;
6083 if (!sched_domain_debug_enabled
)
6087 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6091 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6093 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6094 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6099 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6106 free_cpumask_var(groupmask
);
6108 #else /* !CONFIG_SCHED_DEBUG */
6109 # define sched_domain_debug(sd, cpu) do { } while (0)
6110 #endif /* CONFIG_SCHED_DEBUG */
6112 static int sd_degenerate(struct sched_domain
*sd
)
6114 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6117 /* Following flags need at least 2 groups */
6118 if (sd
->flags
& (SD_LOAD_BALANCE
|
6119 SD_BALANCE_NEWIDLE
|
6123 SD_SHARE_PKG_RESOURCES
)) {
6124 if (sd
->groups
!= sd
->groups
->next
)
6128 /* Following flags don't use groups */
6129 if (sd
->flags
& (SD_WAKE_AFFINE
))
6136 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6138 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6140 if (sd_degenerate(parent
))
6143 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6146 /* Flags needing groups don't count if only 1 group in parent */
6147 if (parent
->groups
== parent
->groups
->next
) {
6148 pflags
&= ~(SD_LOAD_BALANCE
|
6149 SD_BALANCE_NEWIDLE
|
6153 SD_SHARE_PKG_RESOURCES
);
6154 if (nr_node_ids
== 1)
6155 pflags
&= ~SD_SERIALIZE
;
6157 if (~cflags
& pflags
)
6163 static void free_rootdomain(struct root_domain
*rd
)
6165 synchronize_sched();
6167 cpupri_cleanup(&rd
->cpupri
);
6169 free_cpumask_var(rd
->rto_mask
);
6170 free_cpumask_var(rd
->online
);
6171 free_cpumask_var(rd
->span
);
6175 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6177 struct root_domain
*old_rd
= NULL
;
6178 unsigned long flags
;
6180 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6185 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6188 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6191 * If we dont want to free the old_rt yet then
6192 * set old_rd to NULL to skip the freeing later
6195 if (!atomic_dec_and_test(&old_rd
->refcount
))
6199 atomic_inc(&rd
->refcount
);
6202 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6203 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6206 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6209 free_rootdomain(old_rd
);
6212 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6214 gfp_t gfp
= GFP_KERNEL
;
6216 memset(rd
, 0, sizeof(*rd
));
6221 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6223 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6225 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6228 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6233 free_cpumask_var(rd
->rto_mask
);
6235 free_cpumask_var(rd
->online
);
6237 free_cpumask_var(rd
->span
);
6242 static void init_defrootdomain(void)
6244 init_rootdomain(&def_root_domain
, true);
6246 atomic_set(&def_root_domain
.refcount
, 1);
6249 static struct root_domain
*alloc_rootdomain(void)
6251 struct root_domain
*rd
;
6253 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6257 if (init_rootdomain(rd
, false) != 0) {
6266 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6267 * hold the hotplug lock.
6270 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6272 struct rq
*rq
= cpu_rq(cpu
);
6273 struct sched_domain
*tmp
;
6275 /* Remove the sched domains which do not contribute to scheduling. */
6276 for (tmp
= sd
; tmp
; ) {
6277 struct sched_domain
*parent
= tmp
->parent
;
6281 if (sd_parent_degenerate(tmp
, parent
)) {
6282 tmp
->parent
= parent
->parent
;
6284 parent
->parent
->child
= tmp
;
6289 if (sd
&& sd_degenerate(sd
)) {
6295 sched_domain_debug(sd
, cpu
);
6297 rq_attach_root(rq
, rd
);
6298 rcu_assign_pointer(rq
->sd
, sd
);
6301 /* cpus with isolated domains */
6302 static cpumask_var_t cpu_isolated_map
;
6304 /* Setup the mask of cpus configured for isolated domains */
6305 static int __init
isolated_cpu_setup(char *str
)
6307 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6308 cpulist_parse(str
, cpu_isolated_map
);
6312 __setup("isolcpus=", isolated_cpu_setup
);
6315 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6316 * to a function which identifies what group(along with sched group) a CPU
6317 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6318 * (due to the fact that we keep track of groups covered with a struct cpumask).
6320 * init_sched_build_groups will build a circular linked list of the groups
6321 * covered by the given span, and will set each group's ->cpumask correctly,
6322 * and ->cpu_power to 0.
6325 init_sched_build_groups(const struct cpumask
*span
,
6326 const struct cpumask
*cpu_map
,
6327 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6328 struct sched_group
**sg
,
6329 struct cpumask
*tmpmask
),
6330 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6332 struct sched_group
*first
= NULL
, *last
= NULL
;
6335 cpumask_clear(covered
);
6337 for_each_cpu(i
, span
) {
6338 struct sched_group
*sg
;
6339 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6342 if (cpumask_test_cpu(i
, covered
))
6345 cpumask_clear(sched_group_cpus(sg
));
6348 for_each_cpu(j
, span
) {
6349 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6352 cpumask_set_cpu(j
, covered
);
6353 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6364 #define SD_NODES_PER_DOMAIN 16
6369 * find_next_best_node - find the next node to include in a sched_domain
6370 * @node: node whose sched_domain we're building
6371 * @used_nodes: nodes already in the sched_domain
6373 * Find the next node to include in a given scheduling domain. Simply
6374 * finds the closest node not already in the @used_nodes map.
6376 * Should use nodemask_t.
6378 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6380 int i
, n
, val
, min_val
, best_node
= 0;
6384 for (i
= 0; i
< nr_node_ids
; i
++) {
6385 /* Start at @node */
6386 n
= (node
+ i
) % nr_node_ids
;
6388 if (!nr_cpus_node(n
))
6391 /* Skip already used nodes */
6392 if (node_isset(n
, *used_nodes
))
6395 /* Simple min distance search */
6396 val
= node_distance(node
, n
);
6398 if (val
< min_val
) {
6404 node_set(best_node
, *used_nodes
);
6409 * sched_domain_node_span - get a cpumask for a node's sched_domain
6410 * @node: node whose cpumask we're constructing
6411 * @span: resulting cpumask
6413 * Given a node, construct a good cpumask for its sched_domain to span. It
6414 * should be one that prevents unnecessary balancing, but also spreads tasks
6417 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6419 nodemask_t used_nodes
;
6422 cpumask_clear(span
);
6423 nodes_clear(used_nodes
);
6425 cpumask_or(span
, span
, cpumask_of_node(node
));
6426 node_set(node
, used_nodes
);
6428 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6429 int next_node
= find_next_best_node(node
, &used_nodes
);
6431 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6434 #endif /* CONFIG_NUMA */
6436 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6439 * The cpus mask in sched_group and sched_domain hangs off the end.
6441 * ( See the the comments in include/linux/sched.h:struct sched_group
6442 * and struct sched_domain. )
6444 struct static_sched_group
{
6445 struct sched_group sg
;
6446 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6449 struct static_sched_domain
{
6450 struct sched_domain sd
;
6451 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6457 cpumask_var_t domainspan
;
6458 cpumask_var_t covered
;
6459 cpumask_var_t notcovered
;
6461 cpumask_var_t nodemask
;
6462 cpumask_var_t this_sibling_map
;
6463 cpumask_var_t this_core_map
;
6464 cpumask_var_t send_covered
;
6465 cpumask_var_t tmpmask
;
6466 struct sched_group
**sched_group_nodes
;
6467 struct root_domain
*rd
;
6471 sa_sched_groups
= 0,
6476 sa_this_sibling_map
,
6478 sa_sched_group_nodes
,
6488 * SMT sched-domains:
6490 #ifdef CONFIG_SCHED_SMT
6491 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6492 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6495 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6496 struct sched_group
**sg
, struct cpumask
*unused
)
6499 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6502 #endif /* CONFIG_SCHED_SMT */
6505 * multi-core sched-domains:
6507 #ifdef CONFIG_SCHED_MC
6508 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6509 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6510 #endif /* CONFIG_SCHED_MC */
6512 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6514 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6515 struct sched_group
**sg
, struct cpumask
*mask
)
6519 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6520 group
= cpumask_first(mask
);
6522 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6525 #elif defined(CONFIG_SCHED_MC)
6527 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6528 struct sched_group
**sg
, struct cpumask
*unused
)
6531 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6536 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6537 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6540 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6541 struct sched_group
**sg
, struct cpumask
*mask
)
6544 #ifdef CONFIG_SCHED_MC
6545 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6546 group
= cpumask_first(mask
);
6547 #elif defined(CONFIG_SCHED_SMT)
6548 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6549 group
= cpumask_first(mask
);
6554 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6560 * The init_sched_build_groups can't handle what we want to do with node
6561 * groups, so roll our own. Now each node has its own list of groups which
6562 * gets dynamically allocated.
6564 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6565 static struct sched_group
***sched_group_nodes_bycpu
;
6567 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6568 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6570 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6571 struct sched_group
**sg
,
6572 struct cpumask
*nodemask
)
6576 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6577 group
= cpumask_first(nodemask
);
6580 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6584 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6586 struct sched_group
*sg
= group_head
;
6592 for_each_cpu(j
, sched_group_cpus(sg
)) {
6593 struct sched_domain
*sd
;
6595 sd
= &per_cpu(phys_domains
, j
).sd
;
6596 if (j
!= group_first_cpu(sd
->groups
)) {
6598 * Only add "power" once for each
6604 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6607 } while (sg
!= group_head
);
6610 static int build_numa_sched_groups(struct s_data
*d
,
6611 const struct cpumask
*cpu_map
, int num
)
6613 struct sched_domain
*sd
;
6614 struct sched_group
*sg
, *prev
;
6617 cpumask_clear(d
->covered
);
6618 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6619 if (cpumask_empty(d
->nodemask
)) {
6620 d
->sched_group_nodes
[num
] = NULL
;
6624 sched_domain_node_span(num
, d
->domainspan
);
6625 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6627 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6630 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6634 d
->sched_group_nodes
[num
] = sg
;
6636 for_each_cpu(j
, d
->nodemask
) {
6637 sd
= &per_cpu(node_domains
, j
).sd
;
6642 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6644 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6647 for (j
= 0; j
< nr_node_ids
; j
++) {
6648 n
= (num
+ j
) % nr_node_ids
;
6649 cpumask_complement(d
->notcovered
, d
->covered
);
6650 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6651 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6652 if (cpumask_empty(d
->tmpmask
))
6654 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6655 if (cpumask_empty(d
->tmpmask
))
6657 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6661 "Can not alloc domain group for node %d\n", j
);
6665 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6666 sg
->next
= prev
->next
;
6667 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6674 #endif /* CONFIG_NUMA */
6677 /* Free memory allocated for various sched_group structures */
6678 static void free_sched_groups(const struct cpumask
*cpu_map
,
6679 struct cpumask
*nodemask
)
6683 for_each_cpu(cpu
, cpu_map
) {
6684 struct sched_group
**sched_group_nodes
6685 = sched_group_nodes_bycpu
[cpu
];
6687 if (!sched_group_nodes
)
6690 for (i
= 0; i
< nr_node_ids
; i
++) {
6691 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6693 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6694 if (cpumask_empty(nodemask
))
6704 if (oldsg
!= sched_group_nodes
[i
])
6707 kfree(sched_group_nodes
);
6708 sched_group_nodes_bycpu
[cpu
] = NULL
;
6711 #else /* !CONFIG_NUMA */
6712 static void free_sched_groups(const struct cpumask
*cpu_map
,
6713 struct cpumask
*nodemask
)
6716 #endif /* CONFIG_NUMA */
6719 * Initialize sched groups cpu_power.
6721 * cpu_power indicates the capacity of sched group, which is used while
6722 * distributing the load between different sched groups in a sched domain.
6723 * Typically cpu_power for all the groups in a sched domain will be same unless
6724 * there are asymmetries in the topology. If there are asymmetries, group
6725 * having more cpu_power will pickup more load compared to the group having
6728 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6730 struct sched_domain
*child
;
6731 struct sched_group
*group
;
6735 WARN_ON(!sd
|| !sd
->groups
);
6737 if (cpu
!= group_first_cpu(sd
->groups
))
6742 sd
->groups
->cpu_power
= 0;
6745 power
= SCHED_LOAD_SCALE
;
6746 weight
= cpumask_weight(sched_domain_span(sd
));
6748 * SMT siblings share the power of a single core.
6749 * Usually multiple threads get a better yield out of
6750 * that one core than a single thread would have,
6751 * reflect that in sd->smt_gain.
6753 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6754 power
*= sd
->smt_gain
;
6756 power
>>= SCHED_LOAD_SHIFT
;
6758 sd
->groups
->cpu_power
+= power
;
6763 * Add cpu_power of each child group to this groups cpu_power.
6765 group
= child
->groups
;
6767 sd
->groups
->cpu_power
+= group
->cpu_power
;
6768 group
= group
->next
;
6769 } while (group
!= child
->groups
);
6773 * Initializers for schedule domains
6774 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6777 #ifdef CONFIG_SCHED_DEBUG
6778 # define SD_INIT_NAME(sd, type) sd->name = #type
6780 # define SD_INIT_NAME(sd, type) do { } while (0)
6783 #define SD_INIT(sd, type) sd_init_##type(sd)
6785 #define SD_INIT_FUNC(type) \
6786 static noinline void sd_init_##type(struct sched_domain *sd) \
6788 memset(sd, 0, sizeof(*sd)); \
6789 *sd = SD_##type##_INIT; \
6790 sd->level = SD_LV_##type; \
6791 SD_INIT_NAME(sd, type); \
6796 SD_INIT_FUNC(ALLNODES
)
6799 #ifdef CONFIG_SCHED_SMT
6800 SD_INIT_FUNC(SIBLING
)
6802 #ifdef CONFIG_SCHED_MC
6806 static int default_relax_domain_level
= -1;
6808 static int __init
setup_relax_domain_level(char *str
)
6812 val
= simple_strtoul(str
, NULL
, 0);
6813 if (val
< SD_LV_MAX
)
6814 default_relax_domain_level
= val
;
6818 __setup("relax_domain_level=", setup_relax_domain_level
);
6820 static void set_domain_attribute(struct sched_domain
*sd
,
6821 struct sched_domain_attr
*attr
)
6825 if (!attr
|| attr
->relax_domain_level
< 0) {
6826 if (default_relax_domain_level
< 0)
6829 request
= default_relax_domain_level
;
6831 request
= attr
->relax_domain_level
;
6832 if (request
< sd
->level
) {
6833 /* turn off idle balance on this domain */
6834 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6836 /* turn on idle balance on this domain */
6837 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6841 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6842 const struct cpumask
*cpu_map
)
6845 case sa_sched_groups
:
6846 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6847 d
->sched_group_nodes
= NULL
;
6849 free_rootdomain(d
->rd
); /* fall through */
6851 free_cpumask_var(d
->tmpmask
); /* fall through */
6852 case sa_send_covered
:
6853 free_cpumask_var(d
->send_covered
); /* fall through */
6854 case sa_this_core_map
:
6855 free_cpumask_var(d
->this_core_map
); /* fall through */
6856 case sa_this_sibling_map
:
6857 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6859 free_cpumask_var(d
->nodemask
); /* fall through */
6860 case sa_sched_group_nodes
:
6862 kfree(d
->sched_group_nodes
); /* fall through */
6864 free_cpumask_var(d
->notcovered
); /* fall through */
6866 free_cpumask_var(d
->covered
); /* fall through */
6868 free_cpumask_var(d
->domainspan
); /* fall through */
6875 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6876 const struct cpumask
*cpu_map
)
6879 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6881 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6882 return sa_domainspan
;
6883 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6885 /* Allocate the per-node list of sched groups */
6886 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6887 sizeof(struct sched_group
*), GFP_KERNEL
);
6888 if (!d
->sched_group_nodes
) {
6889 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6890 return sa_notcovered
;
6892 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6894 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6895 return sa_sched_group_nodes
;
6896 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6898 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6899 return sa_this_sibling_map
;
6900 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6901 return sa_this_core_map
;
6902 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6903 return sa_send_covered
;
6904 d
->rd
= alloc_rootdomain();
6906 printk(KERN_WARNING
"Cannot alloc root domain\n");
6909 return sa_rootdomain
;
6912 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6913 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6915 struct sched_domain
*sd
= NULL
;
6917 struct sched_domain
*parent
;
6920 if (cpumask_weight(cpu_map
) >
6921 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6922 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6923 SD_INIT(sd
, ALLNODES
);
6924 set_domain_attribute(sd
, attr
);
6925 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6926 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6931 sd
= &per_cpu(node_domains
, i
).sd
;
6933 set_domain_attribute(sd
, attr
);
6934 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6935 sd
->parent
= parent
;
6938 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6943 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6944 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6945 struct sched_domain
*parent
, int i
)
6947 struct sched_domain
*sd
;
6948 sd
= &per_cpu(phys_domains
, i
).sd
;
6950 set_domain_attribute(sd
, attr
);
6951 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6952 sd
->parent
= parent
;
6955 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6959 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6960 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6961 struct sched_domain
*parent
, int i
)
6963 struct sched_domain
*sd
= parent
;
6964 #ifdef CONFIG_SCHED_MC
6965 sd
= &per_cpu(core_domains
, i
).sd
;
6967 set_domain_attribute(sd
, attr
);
6968 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6969 sd
->parent
= parent
;
6971 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6976 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6977 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6978 struct sched_domain
*parent
, int i
)
6980 struct sched_domain
*sd
= parent
;
6981 #ifdef CONFIG_SCHED_SMT
6982 sd
= &per_cpu(cpu_domains
, i
).sd
;
6983 SD_INIT(sd
, SIBLING
);
6984 set_domain_attribute(sd
, attr
);
6985 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6986 sd
->parent
= parent
;
6988 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6993 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6994 const struct cpumask
*cpu_map
, int cpu
)
6997 #ifdef CONFIG_SCHED_SMT
6998 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6999 cpumask_and(d
->this_sibling_map
, cpu_map
,
7000 topology_thread_cpumask(cpu
));
7001 if (cpu
== cpumask_first(d
->this_sibling_map
))
7002 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7004 d
->send_covered
, d
->tmpmask
);
7007 #ifdef CONFIG_SCHED_MC
7008 case SD_LV_MC
: /* set up multi-core groups */
7009 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7010 if (cpu
== cpumask_first(d
->this_core_map
))
7011 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7013 d
->send_covered
, d
->tmpmask
);
7016 case SD_LV_CPU
: /* set up physical groups */
7017 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7018 if (!cpumask_empty(d
->nodemask
))
7019 init_sched_build_groups(d
->nodemask
, cpu_map
,
7021 d
->send_covered
, d
->tmpmask
);
7024 case SD_LV_ALLNODES
:
7025 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7026 d
->send_covered
, d
->tmpmask
);
7035 * Build sched domains for a given set of cpus and attach the sched domains
7036 * to the individual cpus
7038 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7039 struct sched_domain_attr
*attr
)
7041 enum s_alloc alloc_state
= sa_none
;
7043 struct sched_domain
*sd
;
7049 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7050 if (alloc_state
!= sa_rootdomain
)
7052 alloc_state
= sa_sched_groups
;
7055 * Set up domains for cpus specified by the cpu_map.
7057 for_each_cpu(i
, cpu_map
) {
7058 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7061 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7062 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7063 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7064 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7067 for_each_cpu(i
, cpu_map
) {
7068 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7069 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7072 /* Set up physical groups */
7073 for (i
= 0; i
< nr_node_ids
; i
++)
7074 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7077 /* Set up node groups */
7079 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7081 for (i
= 0; i
< nr_node_ids
; i
++)
7082 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7086 /* Calculate CPU power for physical packages and nodes */
7087 #ifdef CONFIG_SCHED_SMT
7088 for_each_cpu(i
, cpu_map
) {
7089 sd
= &per_cpu(cpu_domains
, i
).sd
;
7090 init_sched_groups_power(i
, sd
);
7093 #ifdef CONFIG_SCHED_MC
7094 for_each_cpu(i
, cpu_map
) {
7095 sd
= &per_cpu(core_domains
, i
).sd
;
7096 init_sched_groups_power(i
, sd
);
7100 for_each_cpu(i
, cpu_map
) {
7101 sd
= &per_cpu(phys_domains
, i
).sd
;
7102 init_sched_groups_power(i
, sd
);
7106 for (i
= 0; i
< nr_node_ids
; i
++)
7107 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7109 if (d
.sd_allnodes
) {
7110 struct sched_group
*sg
;
7112 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7114 init_numa_sched_groups_power(sg
);
7118 /* Attach the domains */
7119 for_each_cpu(i
, cpu_map
) {
7120 #ifdef CONFIG_SCHED_SMT
7121 sd
= &per_cpu(cpu_domains
, i
).sd
;
7122 #elif defined(CONFIG_SCHED_MC)
7123 sd
= &per_cpu(core_domains
, i
).sd
;
7125 sd
= &per_cpu(phys_domains
, i
).sd
;
7127 cpu_attach_domain(sd
, d
.rd
, i
);
7130 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7131 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7135 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7139 static int build_sched_domains(const struct cpumask
*cpu_map
)
7141 return __build_sched_domains(cpu_map
, NULL
);
7144 static cpumask_var_t
*doms_cur
; /* current sched domains */
7145 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7146 static struct sched_domain_attr
*dattr_cur
;
7147 /* attribues of custom domains in 'doms_cur' */
7150 * Special case: If a kmalloc of a doms_cur partition (array of
7151 * cpumask) fails, then fallback to a single sched domain,
7152 * as determined by the single cpumask fallback_doms.
7154 static cpumask_var_t fallback_doms
;
7157 * arch_update_cpu_topology lets virtualized architectures update the
7158 * cpu core maps. It is supposed to return 1 if the topology changed
7159 * or 0 if it stayed the same.
7161 int __attribute__((weak
)) arch_update_cpu_topology(void)
7166 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7169 cpumask_var_t
*doms
;
7171 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7174 for (i
= 0; i
< ndoms
; i
++) {
7175 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7176 free_sched_domains(doms
, i
);
7183 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7186 for (i
= 0; i
< ndoms
; i
++)
7187 free_cpumask_var(doms
[i
]);
7192 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7193 * For now this just excludes isolated cpus, but could be used to
7194 * exclude other special cases in the future.
7196 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7200 arch_update_cpu_topology();
7202 doms_cur
= alloc_sched_domains(ndoms_cur
);
7204 doms_cur
= &fallback_doms
;
7205 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7207 err
= build_sched_domains(doms_cur
[0]);
7208 register_sched_domain_sysctl();
7213 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7214 struct cpumask
*tmpmask
)
7216 free_sched_groups(cpu_map
, tmpmask
);
7220 * Detach sched domains from a group of cpus specified in cpu_map
7221 * These cpus will now be attached to the NULL domain
7223 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7225 /* Save because hotplug lock held. */
7226 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7229 for_each_cpu(i
, cpu_map
)
7230 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7231 synchronize_sched();
7232 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7235 /* handle null as "default" */
7236 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7237 struct sched_domain_attr
*new, int idx_new
)
7239 struct sched_domain_attr tmp
;
7246 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7247 new ? (new + idx_new
) : &tmp
,
7248 sizeof(struct sched_domain_attr
));
7252 * Partition sched domains as specified by the 'ndoms_new'
7253 * cpumasks in the array doms_new[] of cpumasks. This compares
7254 * doms_new[] to the current sched domain partitioning, doms_cur[].
7255 * It destroys each deleted domain and builds each new domain.
7257 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7258 * The masks don't intersect (don't overlap.) We should setup one
7259 * sched domain for each mask. CPUs not in any of the cpumasks will
7260 * not be load balanced. If the same cpumask appears both in the
7261 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7264 * The passed in 'doms_new' should be allocated using
7265 * alloc_sched_domains. This routine takes ownership of it and will
7266 * free_sched_domains it when done with it. If the caller failed the
7267 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7268 * and partition_sched_domains() will fallback to the single partition
7269 * 'fallback_doms', it also forces the domains to be rebuilt.
7271 * If doms_new == NULL it will be replaced with cpu_online_mask.
7272 * ndoms_new == 0 is a special case for destroying existing domains,
7273 * and it will not create the default domain.
7275 * Call with hotplug lock held
7277 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7278 struct sched_domain_attr
*dattr_new
)
7283 mutex_lock(&sched_domains_mutex
);
7285 /* always unregister in case we don't destroy any domains */
7286 unregister_sched_domain_sysctl();
7288 /* Let architecture update cpu core mappings. */
7289 new_topology
= arch_update_cpu_topology();
7291 n
= doms_new
? ndoms_new
: 0;
7293 /* Destroy deleted domains */
7294 for (i
= 0; i
< ndoms_cur
; i
++) {
7295 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7296 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7297 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7300 /* no match - a current sched domain not in new doms_new[] */
7301 detach_destroy_domains(doms_cur
[i
]);
7306 if (doms_new
== NULL
) {
7308 doms_new
= &fallback_doms
;
7309 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7310 WARN_ON_ONCE(dattr_new
);
7313 /* Build new domains */
7314 for (i
= 0; i
< ndoms_new
; i
++) {
7315 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7316 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7317 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7320 /* no match - add a new doms_new */
7321 __build_sched_domains(doms_new
[i
],
7322 dattr_new
? dattr_new
+ i
: NULL
);
7327 /* Remember the new sched domains */
7328 if (doms_cur
!= &fallback_doms
)
7329 free_sched_domains(doms_cur
, ndoms_cur
);
7330 kfree(dattr_cur
); /* kfree(NULL) is safe */
7331 doms_cur
= doms_new
;
7332 dattr_cur
= dattr_new
;
7333 ndoms_cur
= ndoms_new
;
7335 register_sched_domain_sysctl();
7337 mutex_unlock(&sched_domains_mutex
);
7340 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7341 static void arch_reinit_sched_domains(void)
7345 /* Destroy domains first to force the rebuild */
7346 partition_sched_domains(0, NULL
, NULL
);
7348 rebuild_sched_domains();
7352 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7354 unsigned int level
= 0;
7356 if (sscanf(buf
, "%u", &level
) != 1)
7360 * level is always be positive so don't check for
7361 * level < POWERSAVINGS_BALANCE_NONE which is 0
7362 * What happens on 0 or 1 byte write,
7363 * need to check for count as well?
7366 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7370 sched_smt_power_savings
= level
;
7372 sched_mc_power_savings
= level
;
7374 arch_reinit_sched_domains();
7379 #ifdef CONFIG_SCHED_MC
7380 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7381 struct sysdev_class_attribute
*attr
,
7384 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7386 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7387 struct sysdev_class_attribute
*attr
,
7388 const char *buf
, size_t count
)
7390 return sched_power_savings_store(buf
, count
, 0);
7392 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7393 sched_mc_power_savings_show
,
7394 sched_mc_power_savings_store
);
7397 #ifdef CONFIG_SCHED_SMT
7398 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7399 struct sysdev_class_attribute
*attr
,
7402 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7404 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7405 struct sysdev_class_attribute
*attr
,
7406 const char *buf
, size_t count
)
7408 return sched_power_savings_store(buf
, count
, 1);
7410 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7411 sched_smt_power_savings_show
,
7412 sched_smt_power_savings_store
);
7415 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7419 #ifdef CONFIG_SCHED_SMT
7421 err
= sysfs_create_file(&cls
->kset
.kobj
,
7422 &attr_sched_smt_power_savings
.attr
);
7424 #ifdef CONFIG_SCHED_MC
7425 if (!err
&& mc_capable())
7426 err
= sysfs_create_file(&cls
->kset
.kobj
,
7427 &attr_sched_mc_power_savings
.attr
);
7431 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7433 #ifndef CONFIG_CPUSETS
7435 * Add online and remove offline CPUs from the scheduler domains.
7436 * When cpusets are enabled they take over this function.
7438 static int update_sched_domains(struct notifier_block
*nfb
,
7439 unsigned long action
, void *hcpu
)
7443 case CPU_ONLINE_FROZEN
:
7444 case CPU_DOWN_PREPARE
:
7445 case CPU_DOWN_PREPARE_FROZEN
:
7446 case CPU_DOWN_FAILED
:
7447 case CPU_DOWN_FAILED_FROZEN
:
7448 partition_sched_domains(1, NULL
, NULL
);
7457 static int update_runtime(struct notifier_block
*nfb
,
7458 unsigned long action
, void *hcpu
)
7460 int cpu
= (int)(long)hcpu
;
7463 case CPU_DOWN_PREPARE
:
7464 case CPU_DOWN_PREPARE_FROZEN
:
7465 disable_runtime(cpu_rq(cpu
));
7468 case CPU_DOWN_FAILED
:
7469 case CPU_DOWN_FAILED_FROZEN
:
7471 case CPU_ONLINE_FROZEN
:
7472 enable_runtime(cpu_rq(cpu
));
7480 void __init
sched_init_smp(void)
7482 cpumask_var_t non_isolated_cpus
;
7484 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7485 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7487 #if defined(CONFIG_NUMA)
7488 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7490 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7493 mutex_lock(&sched_domains_mutex
);
7494 arch_init_sched_domains(cpu_active_mask
);
7495 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7496 if (cpumask_empty(non_isolated_cpus
))
7497 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7498 mutex_unlock(&sched_domains_mutex
);
7501 #ifndef CONFIG_CPUSETS
7502 /* XXX: Theoretical race here - CPU may be hotplugged now */
7503 hotcpu_notifier(update_sched_domains
, 0);
7506 /* RT runtime code needs to handle some hotplug events */
7507 hotcpu_notifier(update_runtime
, 0);
7511 /* Move init over to a non-isolated CPU */
7512 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7514 sched_init_granularity();
7515 free_cpumask_var(non_isolated_cpus
);
7517 init_sched_rt_class();
7520 void __init
sched_init_smp(void)
7522 sched_init_granularity();
7524 #endif /* CONFIG_SMP */
7526 const_debug
unsigned int sysctl_timer_migration
= 1;
7528 int in_sched_functions(unsigned long addr
)
7530 return in_lock_functions(addr
) ||
7531 (addr
>= (unsigned long)__sched_text_start
7532 && addr
< (unsigned long)__sched_text_end
);
7535 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7537 cfs_rq
->tasks_timeline
= RB_ROOT
;
7538 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7539 #ifdef CONFIG_FAIR_GROUP_SCHED
7542 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7545 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7547 struct rt_prio_array
*array
;
7550 array
= &rt_rq
->active
;
7551 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7552 INIT_LIST_HEAD(array
->queue
+ i
);
7553 __clear_bit(i
, array
->bitmap
);
7555 /* delimiter for bitsearch: */
7556 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7558 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7559 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7561 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7565 rt_rq
->rt_nr_migratory
= 0;
7566 rt_rq
->overloaded
= 0;
7567 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7571 rt_rq
->rt_throttled
= 0;
7572 rt_rq
->rt_runtime
= 0;
7573 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7575 #ifdef CONFIG_RT_GROUP_SCHED
7576 rt_rq
->rt_nr_boosted
= 0;
7581 #ifdef CONFIG_FAIR_GROUP_SCHED
7582 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7583 struct sched_entity
*se
, int cpu
, int add
,
7584 struct sched_entity
*parent
)
7586 struct rq
*rq
= cpu_rq(cpu
);
7587 tg
->cfs_rq
[cpu
] = cfs_rq
;
7588 init_cfs_rq(cfs_rq
, rq
);
7591 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7594 /* se could be NULL for init_task_group */
7599 se
->cfs_rq
= &rq
->cfs
;
7601 se
->cfs_rq
= parent
->my_q
;
7604 se
->load
.weight
= tg
->shares
;
7605 se
->load
.inv_weight
= 0;
7606 se
->parent
= parent
;
7610 #ifdef CONFIG_RT_GROUP_SCHED
7611 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7612 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7613 struct sched_rt_entity
*parent
)
7615 struct rq
*rq
= cpu_rq(cpu
);
7617 tg
->rt_rq
[cpu
] = rt_rq
;
7618 init_rt_rq(rt_rq
, rq
);
7620 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7622 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7624 tg
->rt_se
[cpu
] = rt_se
;
7629 rt_se
->rt_rq
= &rq
->rt
;
7631 rt_se
->rt_rq
= parent
->my_q
;
7633 rt_se
->my_q
= rt_rq
;
7634 rt_se
->parent
= parent
;
7635 INIT_LIST_HEAD(&rt_se
->run_list
);
7639 void __init
sched_init(void)
7642 unsigned long alloc_size
= 0, ptr
;
7644 #ifdef CONFIG_FAIR_GROUP_SCHED
7645 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7647 #ifdef CONFIG_RT_GROUP_SCHED
7648 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7650 #ifdef CONFIG_CPUMASK_OFFSTACK
7651 alloc_size
+= num_possible_cpus() * cpumask_size();
7654 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7656 #ifdef CONFIG_FAIR_GROUP_SCHED
7657 init_task_group
.se
= (struct sched_entity
**)ptr
;
7658 ptr
+= nr_cpu_ids
* sizeof(void **);
7660 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7661 ptr
+= nr_cpu_ids
* sizeof(void **);
7663 #endif /* CONFIG_FAIR_GROUP_SCHED */
7664 #ifdef CONFIG_RT_GROUP_SCHED
7665 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7666 ptr
+= nr_cpu_ids
* sizeof(void **);
7668 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7669 ptr
+= nr_cpu_ids
* sizeof(void **);
7671 #endif /* CONFIG_RT_GROUP_SCHED */
7672 #ifdef CONFIG_CPUMASK_OFFSTACK
7673 for_each_possible_cpu(i
) {
7674 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7675 ptr
+= cpumask_size();
7677 #endif /* CONFIG_CPUMASK_OFFSTACK */
7681 init_defrootdomain();
7684 init_rt_bandwidth(&def_rt_bandwidth
,
7685 global_rt_period(), global_rt_runtime());
7687 #ifdef CONFIG_RT_GROUP_SCHED
7688 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7689 global_rt_period(), global_rt_runtime());
7690 #endif /* CONFIG_RT_GROUP_SCHED */
7692 #ifdef CONFIG_CGROUP_SCHED
7693 list_add(&init_task_group
.list
, &task_groups
);
7694 INIT_LIST_HEAD(&init_task_group
.children
);
7696 #endif /* CONFIG_CGROUP_SCHED */
7698 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7699 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7700 __alignof__(unsigned long));
7702 for_each_possible_cpu(i
) {
7706 raw_spin_lock_init(&rq
->lock
);
7708 rq
->calc_load_active
= 0;
7709 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7710 init_cfs_rq(&rq
->cfs
, rq
);
7711 init_rt_rq(&rq
->rt
, rq
);
7712 #ifdef CONFIG_FAIR_GROUP_SCHED
7713 init_task_group
.shares
= init_task_group_load
;
7714 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7715 #ifdef CONFIG_CGROUP_SCHED
7717 * How much cpu bandwidth does init_task_group get?
7719 * In case of task-groups formed thr' the cgroup filesystem, it
7720 * gets 100% of the cpu resources in the system. This overall
7721 * system cpu resource is divided among the tasks of
7722 * init_task_group and its child task-groups in a fair manner,
7723 * based on each entity's (task or task-group's) weight
7724 * (se->load.weight).
7726 * In other words, if init_task_group has 10 tasks of weight
7727 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7728 * then A0's share of the cpu resource is:
7730 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7732 * We achieve this by letting init_task_group's tasks sit
7733 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7735 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7737 #endif /* CONFIG_FAIR_GROUP_SCHED */
7739 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7740 #ifdef CONFIG_RT_GROUP_SCHED
7741 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7742 #ifdef CONFIG_CGROUP_SCHED
7743 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7747 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7748 rq
->cpu_load
[j
] = 0;
7752 rq
->post_schedule
= 0;
7753 rq
->active_balance
= 0;
7754 rq
->next_balance
= jiffies
;
7758 rq
->migration_thread
= NULL
;
7760 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7761 INIT_LIST_HEAD(&rq
->migration_queue
);
7762 rq_attach_root(rq
, &def_root_domain
);
7765 atomic_set(&rq
->nr_iowait
, 0);
7768 set_load_weight(&init_task
);
7770 #ifdef CONFIG_PREEMPT_NOTIFIERS
7771 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7775 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7778 #ifdef CONFIG_RT_MUTEXES
7779 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7783 * The boot idle thread does lazy MMU switching as well:
7785 atomic_inc(&init_mm
.mm_count
);
7786 enter_lazy_tlb(&init_mm
, current
);
7789 * Make us the idle thread. Technically, schedule() should not be
7790 * called from this thread, however somewhere below it might be,
7791 * but because we are the idle thread, we just pick up running again
7792 * when this runqueue becomes "idle".
7794 init_idle(current
, smp_processor_id());
7796 calc_load_update
= jiffies
+ LOAD_FREQ
;
7799 * During early bootup we pretend to be a normal task:
7801 current
->sched_class
= &fair_sched_class
;
7803 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7804 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7807 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7808 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7810 /* May be allocated at isolcpus cmdline parse time */
7811 if (cpu_isolated_map
== NULL
)
7812 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7817 scheduler_running
= 1;
7820 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7821 static inline int preempt_count_equals(int preempt_offset
)
7823 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7825 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7828 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7831 static unsigned long prev_jiffy
; /* ratelimiting */
7833 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7834 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7836 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7838 prev_jiffy
= jiffies
;
7841 "BUG: sleeping function called from invalid context at %s:%d\n",
7844 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7845 in_atomic(), irqs_disabled(),
7846 current
->pid
, current
->comm
);
7848 debug_show_held_locks(current
);
7849 if (irqs_disabled())
7850 print_irqtrace_events(current
);
7854 EXPORT_SYMBOL(__might_sleep
);
7857 #ifdef CONFIG_MAGIC_SYSRQ
7858 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7862 update_rq_clock(rq
);
7863 on_rq
= p
->se
.on_rq
;
7865 deactivate_task(rq
, p
, 0);
7866 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7868 activate_task(rq
, p
, 0);
7869 resched_task(rq
->curr
);
7873 void normalize_rt_tasks(void)
7875 struct task_struct
*g
, *p
;
7876 unsigned long flags
;
7879 read_lock_irqsave(&tasklist_lock
, flags
);
7880 do_each_thread(g
, p
) {
7882 * Only normalize user tasks:
7887 p
->se
.exec_start
= 0;
7888 #ifdef CONFIG_SCHEDSTATS
7889 p
->se
.wait_start
= 0;
7890 p
->se
.sleep_start
= 0;
7891 p
->se
.block_start
= 0;
7896 * Renice negative nice level userspace
7899 if (TASK_NICE(p
) < 0 && p
->mm
)
7900 set_user_nice(p
, 0);
7904 raw_spin_lock(&p
->pi_lock
);
7905 rq
= __task_rq_lock(p
);
7907 normalize_task(rq
, p
);
7909 __task_rq_unlock(rq
);
7910 raw_spin_unlock(&p
->pi_lock
);
7911 } while_each_thread(g
, p
);
7913 read_unlock_irqrestore(&tasklist_lock
, flags
);
7916 #endif /* CONFIG_MAGIC_SYSRQ */
7920 * These functions are only useful for the IA64 MCA handling.
7922 * They can only be called when the whole system has been
7923 * stopped - every CPU needs to be quiescent, and no scheduling
7924 * activity can take place. Using them for anything else would
7925 * be a serious bug, and as a result, they aren't even visible
7926 * under any other configuration.
7930 * curr_task - return the current task for a given cpu.
7931 * @cpu: the processor in question.
7933 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7935 struct task_struct
*curr_task(int cpu
)
7937 return cpu_curr(cpu
);
7941 * set_curr_task - set the current task for a given cpu.
7942 * @cpu: the processor in question.
7943 * @p: the task pointer to set.
7945 * Description: This function must only be used when non-maskable interrupts
7946 * are serviced on a separate stack. It allows the architecture to switch the
7947 * notion of the current task on a cpu in a non-blocking manner. This function
7948 * must be called with all CPU's synchronized, and interrupts disabled, the
7949 * and caller must save the original value of the current task (see
7950 * curr_task() above) and restore that value before reenabling interrupts and
7951 * re-starting the system.
7953 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7955 void set_curr_task(int cpu
, struct task_struct
*p
)
7962 #ifdef CONFIG_FAIR_GROUP_SCHED
7963 static void free_fair_sched_group(struct task_group
*tg
)
7967 for_each_possible_cpu(i
) {
7969 kfree(tg
->cfs_rq
[i
]);
7979 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7981 struct cfs_rq
*cfs_rq
;
7982 struct sched_entity
*se
;
7986 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7989 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7993 tg
->shares
= NICE_0_LOAD
;
7995 for_each_possible_cpu(i
) {
7998 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7999 GFP_KERNEL
, cpu_to_node(i
));
8003 se
= kzalloc_node(sizeof(struct sched_entity
),
8004 GFP_KERNEL
, cpu_to_node(i
));
8008 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8019 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8021 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8022 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8025 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8027 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8029 #else /* !CONFG_FAIR_GROUP_SCHED */
8030 static inline void free_fair_sched_group(struct task_group
*tg
)
8035 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8040 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8044 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8047 #endif /* CONFIG_FAIR_GROUP_SCHED */
8049 #ifdef CONFIG_RT_GROUP_SCHED
8050 static void free_rt_sched_group(struct task_group
*tg
)
8054 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8056 for_each_possible_cpu(i
) {
8058 kfree(tg
->rt_rq
[i
]);
8060 kfree(tg
->rt_se
[i
]);
8068 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8070 struct rt_rq
*rt_rq
;
8071 struct sched_rt_entity
*rt_se
;
8075 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8078 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8082 init_rt_bandwidth(&tg
->rt_bandwidth
,
8083 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8085 for_each_possible_cpu(i
) {
8088 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8089 GFP_KERNEL
, cpu_to_node(i
));
8093 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8094 GFP_KERNEL
, cpu_to_node(i
));
8098 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8109 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8111 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8112 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8115 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8117 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8119 #else /* !CONFIG_RT_GROUP_SCHED */
8120 static inline void free_rt_sched_group(struct task_group
*tg
)
8125 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8130 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8134 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8137 #endif /* CONFIG_RT_GROUP_SCHED */
8139 #ifdef CONFIG_CGROUP_SCHED
8140 static void free_sched_group(struct task_group
*tg
)
8142 free_fair_sched_group(tg
);
8143 free_rt_sched_group(tg
);
8147 /* allocate runqueue etc for a new task group */
8148 struct task_group
*sched_create_group(struct task_group
*parent
)
8150 struct task_group
*tg
;
8151 unsigned long flags
;
8154 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8156 return ERR_PTR(-ENOMEM
);
8158 if (!alloc_fair_sched_group(tg
, parent
))
8161 if (!alloc_rt_sched_group(tg
, parent
))
8164 spin_lock_irqsave(&task_group_lock
, flags
);
8165 for_each_possible_cpu(i
) {
8166 register_fair_sched_group(tg
, i
);
8167 register_rt_sched_group(tg
, i
);
8169 list_add_rcu(&tg
->list
, &task_groups
);
8171 WARN_ON(!parent
); /* root should already exist */
8173 tg
->parent
= parent
;
8174 INIT_LIST_HEAD(&tg
->children
);
8175 list_add_rcu(&tg
->siblings
, &parent
->children
);
8176 spin_unlock_irqrestore(&task_group_lock
, flags
);
8181 free_sched_group(tg
);
8182 return ERR_PTR(-ENOMEM
);
8185 /* rcu callback to free various structures associated with a task group */
8186 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8188 /* now it should be safe to free those cfs_rqs */
8189 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8192 /* Destroy runqueue etc associated with a task group */
8193 void sched_destroy_group(struct task_group
*tg
)
8195 unsigned long flags
;
8198 spin_lock_irqsave(&task_group_lock
, flags
);
8199 for_each_possible_cpu(i
) {
8200 unregister_fair_sched_group(tg
, i
);
8201 unregister_rt_sched_group(tg
, i
);
8203 list_del_rcu(&tg
->list
);
8204 list_del_rcu(&tg
->siblings
);
8205 spin_unlock_irqrestore(&task_group_lock
, flags
);
8207 /* wait for possible concurrent references to cfs_rqs complete */
8208 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8211 /* change task's runqueue when it moves between groups.
8212 * The caller of this function should have put the task in its new group
8213 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8214 * reflect its new group.
8216 void sched_move_task(struct task_struct
*tsk
)
8219 unsigned long flags
;
8222 rq
= task_rq_lock(tsk
, &flags
);
8224 update_rq_clock(rq
);
8226 running
= task_current(rq
, tsk
);
8227 on_rq
= tsk
->se
.on_rq
;
8230 dequeue_task(rq
, tsk
, 0);
8231 if (unlikely(running
))
8232 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8234 set_task_rq(tsk
, task_cpu(tsk
));
8236 #ifdef CONFIG_FAIR_GROUP_SCHED
8237 if (tsk
->sched_class
->moved_group
)
8238 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8241 if (unlikely(running
))
8242 tsk
->sched_class
->set_curr_task(rq
);
8244 enqueue_task(rq
, tsk
, 0, false);
8246 task_rq_unlock(rq
, &flags
);
8248 #endif /* CONFIG_CGROUP_SCHED */
8250 #ifdef CONFIG_FAIR_GROUP_SCHED
8251 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8253 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8258 dequeue_entity(cfs_rq
, se
, 0);
8260 se
->load
.weight
= shares
;
8261 se
->load
.inv_weight
= 0;
8264 enqueue_entity(cfs_rq
, se
, 0);
8267 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8269 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8270 struct rq
*rq
= cfs_rq
->rq
;
8271 unsigned long flags
;
8273 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8274 __set_se_shares(se
, shares
);
8275 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8278 static DEFINE_MUTEX(shares_mutex
);
8280 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8283 unsigned long flags
;
8286 * We can't change the weight of the root cgroup.
8291 if (shares
< MIN_SHARES
)
8292 shares
= MIN_SHARES
;
8293 else if (shares
> MAX_SHARES
)
8294 shares
= MAX_SHARES
;
8296 mutex_lock(&shares_mutex
);
8297 if (tg
->shares
== shares
)
8300 spin_lock_irqsave(&task_group_lock
, flags
);
8301 for_each_possible_cpu(i
)
8302 unregister_fair_sched_group(tg
, i
);
8303 list_del_rcu(&tg
->siblings
);
8304 spin_unlock_irqrestore(&task_group_lock
, flags
);
8306 /* wait for any ongoing reference to this group to finish */
8307 synchronize_sched();
8310 * Now we are free to modify the group's share on each cpu
8311 * w/o tripping rebalance_share or load_balance_fair.
8313 tg
->shares
= shares
;
8314 for_each_possible_cpu(i
) {
8318 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8319 set_se_shares(tg
->se
[i
], shares
);
8323 * Enable load balance activity on this group, by inserting it back on
8324 * each cpu's rq->leaf_cfs_rq_list.
8326 spin_lock_irqsave(&task_group_lock
, flags
);
8327 for_each_possible_cpu(i
)
8328 register_fair_sched_group(tg
, i
);
8329 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8330 spin_unlock_irqrestore(&task_group_lock
, flags
);
8332 mutex_unlock(&shares_mutex
);
8336 unsigned long sched_group_shares(struct task_group
*tg
)
8342 #ifdef CONFIG_RT_GROUP_SCHED
8344 * Ensure that the real time constraints are schedulable.
8346 static DEFINE_MUTEX(rt_constraints_mutex
);
8348 static unsigned long to_ratio(u64 period
, u64 runtime
)
8350 if (runtime
== RUNTIME_INF
)
8353 return div64_u64(runtime
<< 20, period
);
8356 /* Must be called with tasklist_lock held */
8357 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8359 struct task_struct
*g
, *p
;
8361 do_each_thread(g
, p
) {
8362 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8364 } while_each_thread(g
, p
);
8369 struct rt_schedulable_data
{
8370 struct task_group
*tg
;
8375 static int tg_schedulable(struct task_group
*tg
, void *data
)
8377 struct rt_schedulable_data
*d
= data
;
8378 struct task_group
*child
;
8379 unsigned long total
, sum
= 0;
8380 u64 period
, runtime
;
8382 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8383 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8386 period
= d
->rt_period
;
8387 runtime
= d
->rt_runtime
;
8391 * Cannot have more runtime than the period.
8393 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8397 * Ensure we don't starve existing RT tasks.
8399 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8402 total
= to_ratio(period
, runtime
);
8405 * Nobody can have more than the global setting allows.
8407 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8411 * The sum of our children's runtime should not exceed our own.
8413 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8414 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8415 runtime
= child
->rt_bandwidth
.rt_runtime
;
8417 if (child
== d
->tg
) {
8418 period
= d
->rt_period
;
8419 runtime
= d
->rt_runtime
;
8422 sum
+= to_ratio(period
, runtime
);
8431 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8433 struct rt_schedulable_data data
= {
8435 .rt_period
= period
,
8436 .rt_runtime
= runtime
,
8439 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8442 static int tg_set_bandwidth(struct task_group
*tg
,
8443 u64 rt_period
, u64 rt_runtime
)
8447 mutex_lock(&rt_constraints_mutex
);
8448 read_lock(&tasklist_lock
);
8449 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8453 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8454 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8455 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8457 for_each_possible_cpu(i
) {
8458 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8460 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8461 rt_rq
->rt_runtime
= rt_runtime
;
8462 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8464 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8466 read_unlock(&tasklist_lock
);
8467 mutex_unlock(&rt_constraints_mutex
);
8472 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8474 u64 rt_runtime
, rt_period
;
8476 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8477 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8478 if (rt_runtime_us
< 0)
8479 rt_runtime
= RUNTIME_INF
;
8481 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8484 long sched_group_rt_runtime(struct task_group
*tg
)
8488 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8491 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8492 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8493 return rt_runtime_us
;
8496 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8498 u64 rt_runtime
, rt_period
;
8500 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8501 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8506 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8509 long sched_group_rt_period(struct task_group
*tg
)
8513 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8514 do_div(rt_period_us
, NSEC_PER_USEC
);
8515 return rt_period_us
;
8518 static int sched_rt_global_constraints(void)
8520 u64 runtime
, period
;
8523 if (sysctl_sched_rt_period
<= 0)
8526 runtime
= global_rt_runtime();
8527 period
= global_rt_period();
8530 * Sanity check on the sysctl variables.
8532 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8535 mutex_lock(&rt_constraints_mutex
);
8536 read_lock(&tasklist_lock
);
8537 ret
= __rt_schedulable(NULL
, 0, 0);
8538 read_unlock(&tasklist_lock
);
8539 mutex_unlock(&rt_constraints_mutex
);
8544 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8546 /* Don't accept realtime tasks when there is no way for them to run */
8547 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8553 #else /* !CONFIG_RT_GROUP_SCHED */
8554 static int sched_rt_global_constraints(void)
8556 unsigned long flags
;
8559 if (sysctl_sched_rt_period
<= 0)
8563 * There's always some RT tasks in the root group
8564 * -- migration, kstopmachine etc..
8566 if (sysctl_sched_rt_runtime
== 0)
8569 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8570 for_each_possible_cpu(i
) {
8571 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8573 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8574 rt_rq
->rt_runtime
= global_rt_runtime();
8575 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8577 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8581 #endif /* CONFIG_RT_GROUP_SCHED */
8583 int sched_rt_handler(struct ctl_table
*table
, int write
,
8584 void __user
*buffer
, size_t *lenp
,
8588 int old_period
, old_runtime
;
8589 static DEFINE_MUTEX(mutex
);
8592 old_period
= sysctl_sched_rt_period
;
8593 old_runtime
= sysctl_sched_rt_runtime
;
8595 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8597 if (!ret
&& write
) {
8598 ret
= sched_rt_global_constraints();
8600 sysctl_sched_rt_period
= old_period
;
8601 sysctl_sched_rt_runtime
= old_runtime
;
8603 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8604 def_rt_bandwidth
.rt_period
=
8605 ns_to_ktime(global_rt_period());
8608 mutex_unlock(&mutex
);
8613 #ifdef CONFIG_CGROUP_SCHED
8615 /* return corresponding task_group object of a cgroup */
8616 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8618 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8619 struct task_group
, css
);
8622 static struct cgroup_subsys_state
*
8623 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8625 struct task_group
*tg
, *parent
;
8627 if (!cgrp
->parent
) {
8628 /* This is early initialization for the top cgroup */
8629 return &init_task_group
.css
;
8632 parent
= cgroup_tg(cgrp
->parent
);
8633 tg
= sched_create_group(parent
);
8635 return ERR_PTR(-ENOMEM
);
8641 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8643 struct task_group
*tg
= cgroup_tg(cgrp
);
8645 sched_destroy_group(tg
);
8649 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8651 #ifdef CONFIG_RT_GROUP_SCHED
8652 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8655 /* We don't support RT-tasks being in separate groups */
8656 if (tsk
->sched_class
!= &fair_sched_class
)
8663 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8664 struct task_struct
*tsk
, bool threadgroup
)
8666 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8670 struct task_struct
*c
;
8672 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8673 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8685 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8686 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8689 sched_move_task(tsk
);
8691 struct task_struct
*c
;
8693 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8700 #ifdef CONFIG_FAIR_GROUP_SCHED
8701 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8704 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8707 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8709 struct task_group
*tg
= cgroup_tg(cgrp
);
8711 return (u64
) tg
->shares
;
8713 #endif /* CONFIG_FAIR_GROUP_SCHED */
8715 #ifdef CONFIG_RT_GROUP_SCHED
8716 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8719 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8722 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8724 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8727 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8730 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8733 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8735 return sched_group_rt_period(cgroup_tg(cgrp
));
8737 #endif /* CONFIG_RT_GROUP_SCHED */
8739 static struct cftype cpu_files
[] = {
8740 #ifdef CONFIG_FAIR_GROUP_SCHED
8743 .read_u64
= cpu_shares_read_u64
,
8744 .write_u64
= cpu_shares_write_u64
,
8747 #ifdef CONFIG_RT_GROUP_SCHED
8749 .name
= "rt_runtime_us",
8750 .read_s64
= cpu_rt_runtime_read
,
8751 .write_s64
= cpu_rt_runtime_write
,
8754 .name
= "rt_period_us",
8755 .read_u64
= cpu_rt_period_read_uint
,
8756 .write_u64
= cpu_rt_period_write_uint
,
8761 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8763 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8766 struct cgroup_subsys cpu_cgroup_subsys
= {
8768 .create
= cpu_cgroup_create
,
8769 .destroy
= cpu_cgroup_destroy
,
8770 .can_attach
= cpu_cgroup_can_attach
,
8771 .attach
= cpu_cgroup_attach
,
8772 .populate
= cpu_cgroup_populate
,
8773 .subsys_id
= cpu_cgroup_subsys_id
,
8777 #endif /* CONFIG_CGROUP_SCHED */
8779 #ifdef CONFIG_CGROUP_CPUACCT
8782 * CPU accounting code for task groups.
8784 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8785 * (balbir@in.ibm.com).
8788 /* track cpu usage of a group of tasks and its child groups */
8790 struct cgroup_subsys_state css
;
8791 /* cpuusage holds pointer to a u64-type object on every cpu */
8792 u64 __percpu
*cpuusage
;
8793 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8794 struct cpuacct
*parent
;
8797 struct cgroup_subsys cpuacct_subsys
;
8799 /* return cpu accounting group corresponding to this container */
8800 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8802 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8803 struct cpuacct
, css
);
8806 /* return cpu accounting group to which this task belongs */
8807 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8809 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8810 struct cpuacct
, css
);
8813 /* create a new cpu accounting group */
8814 static struct cgroup_subsys_state
*cpuacct_create(
8815 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8817 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8823 ca
->cpuusage
= alloc_percpu(u64
);
8827 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8828 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8829 goto out_free_counters
;
8832 ca
->parent
= cgroup_ca(cgrp
->parent
);
8838 percpu_counter_destroy(&ca
->cpustat
[i
]);
8839 free_percpu(ca
->cpuusage
);
8843 return ERR_PTR(-ENOMEM
);
8846 /* destroy an existing cpu accounting group */
8848 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8850 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8853 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8854 percpu_counter_destroy(&ca
->cpustat
[i
]);
8855 free_percpu(ca
->cpuusage
);
8859 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8861 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8864 #ifndef CONFIG_64BIT
8866 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8868 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8870 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8878 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8880 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8882 #ifndef CONFIG_64BIT
8884 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8886 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8888 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8894 /* return total cpu usage (in nanoseconds) of a group */
8895 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8897 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8898 u64 totalcpuusage
= 0;
8901 for_each_present_cpu(i
)
8902 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8904 return totalcpuusage
;
8907 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8910 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8919 for_each_present_cpu(i
)
8920 cpuacct_cpuusage_write(ca
, i
, 0);
8926 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8929 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8933 for_each_present_cpu(i
) {
8934 percpu
= cpuacct_cpuusage_read(ca
, i
);
8935 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8937 seq_printf(m
, "\n");
8941 static const char *cpuacct_stat_desc
[] = {
8942 [CPUACCT_STAT_USER
] = "user",
8943 [CPUACCT_STAT_SYSTEM
] = "system",
8946 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8947 struct cgroup_map_cb
*cb
)
8949 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8952 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8953 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8954 val
= cputime64_to_clock_t(val
);
8955 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8960 static struct cftype files
[] = {
8963 .read_u64
= cpuusage_read
,
8964 .write_u64
= cpuusage_write
,
8967 .name
= "usage_percpu",
8968 .read_seq_string
= cpuacct_percpu_seq_read
,
8972 .read_map
= cpuacct_stats_show
,
8976 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8978 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8982 * charge this task's execution time to its accounting group.
8984 * called with rq->lock held.
8986 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8991 if (unlikely(!cpuacct_subsys
.active
))
8994 cpu
= task_cpu(tsk
);
9000 for (; ca
; ca
= ca
->parent
) {
9001 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9002 *cpuusage
+= cputime
;
9009 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9010 * in cputime_t units. As a result, cpuacct_update_stats calls
9011 * percpu_counter_add with values large enough to always overflow the
9012 * per cpu batch limit causing bad SMP scalability.
9014 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9015 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9016 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9019 #define CPUACCT_BATCH \
9020 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9022 #define CPUACCT_BATCH 0
9026 * Charge the system/user time to the task's accounting group.
9028 static void cpuacct_update_stats(struct task_struct
*tsk
,
9029 enum cpuacct_stat_index idx
, cputime_t val
)
9032 int batch
= CPUACCT_BATCH
;
9034 if (unlikely(!cpuacct_subsys
.active
))
9041 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9047 struct cgroup_subsys cpuacct_subsys
= {
9049 .create
= cpuacct_create
,
9050 .destroy
= cpuacct_destroy
,
9051 .populate
= cpuacct_populate
,
9052 .subsys_id
= cpuacct_subsys_id
,
9054 #endif /* CONFIG_CGROUP_CPUACCT */
9058 int rcu_expedited_torture_stats(char *page
)
9062 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9064 void synchronize_sched_expedited(void)
9067 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9069 #else /* #ifndef CONFIG_SMP */
9071 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9072 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9074 #define RCU_EXPEDITED_STATE_POST -2
9075 #define RCU_EXPEDITED_STATE_IDLE -1
9077 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9079 int rcu_expedited_torture_stats(char *page
)
9084 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9085 for_each_online_cpu(cpu
) {
9086 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9087 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9089 cnt
+= sprintf(&page
[cnt
], "\n");
9092 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9094 static long synchronize_sched_expedited_count
;
9097 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9098 * approach to force grace period to end quickly. This consumes
9099 * significant time on all CPUs, and is thus not recommended for
9100 * any sort of common-case code.
9102 * Note that it is illegal to call this function while holding any
9103 * lock that is acquired by a CPU-hotplug notifier. Failing to
9104 * observe this restriction will result in deadlock.
9106 void synchronize_sched_expedited(void)
9109 unsigned long flags
;
9110 bool need_full_sync
= 0;
9112 struct migration_req
*req
;
9116 smp_mb(); /* ensure prior mod happens before capturing snap. */
9117 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9119 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9121 if (trycount
++ < 10)
9122 udelay(trycount
* num_online_cpus());
9124 synchronize_sched();
9127 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9128 smp_mb(); /* ensure test happens before caller kfree */
9133 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9134 for_each_online_cpu(cpu
) {
9136 req
= &per_cpu(rcu_migration_req
, cpu
);
9137 init_completion(&req
->done
);
9139 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9140 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9141 list_add(&req
->list
, &rq
->migration_queue
);
9142 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9143 wake_up_process(rq
->migration_thread
);
9145 for_each_online_cpu(cpu
) {
9146 rcu_expedited_state
= cpu
;
9147 req
= &per_cpu(rcu_migration_req
, cpu
);
9149 wait_for_completion(&req
->done
);
9150 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9151 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9153 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9154 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9156 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9157 synchronize_sched_expedited_count
++;
9158 mutex_unlock(&rcu_sched_expedited_mutex
);
9161 synchronize_sched();
9163 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9165 #endif /* #else #ifndef CONFIG_SMP */