4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
78 #include "sched_cpupri.h"
80 #define CREATE_TRACE_POINTS
81 #include <trace/events/sched.h>
84 * Convert user-nice values [ -20 ... 0 ... 19 ]
85 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
88 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
89 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
90 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
93 * 'User priority' is the nice value converted to something we
94 * can work with better when scaling various scheduler parameters,
95 * it's a [ 0 ... 39 ] range.
97 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
98 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
99 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
102 * Helpers for converting nanosecond timing to jiffy resolution
104 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
106 #define NICE_0_LOAD SCHED_LOAD_SCALE
107 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
110 * These are the 'tuning knobs' of the scheduler:
112 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
113 * Timeslices get refilled after they expire.
115 #define DEF_TIMESLICE (100 * HZ / 1000)
118 * single value that denotes runtime == period, ie unlimited time.
120 #define RUNTIME_INF ((u64)~0ULL)
122 static inline int rt_policy(int policy
)
124 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
129 static inline int task_has_rt_policy(struct task_struct
*p
)
131 return rt_policy(p
->policy
);
135 * This is the priority-queue data structure of the RT scheduling class:
137 struct rt_prio_array
{
138 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
139 struct list_head queue
[MAX_RT_PRIO
];
142 struct rt_bandwidth
{
143 /* nests inside the rq lock: */
144 raw_spinlock_t rt_runtime_lock
;
147 struct hrtimer rt_period_timer
;
150 static struct rt_bandwidth def_rt_bandwidth
;
152 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
154 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
156 struct rt_bandwidth
*rt_b
=
157 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
163 now
= hrtimer_cb_get_time(timer
);
164 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
169 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
172 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
176 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
178 rt_b
->rt_period
= ns_to_ktime(period
);
179 rt_b
->rt_runtime
= runtime
;
181 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
183 hrtimer_init(&rt_b
->rt_period_timer
,
184 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
185 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
188 static inline int rt_bandwidth_enabled(void)
190 return sysctl_sched_rt_runtime
>= 0;
193 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
197 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
200 if (hrtimer_active(&rt_b
->rt_period_timer
))
203 raw_spin_lock(&rt_b
->rt_runtime_lock
);
208 if (hrtimer_active(&rt_b
->rt_period_timer
))
211 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
212 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
214 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
215 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
216 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
217 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
218 HRTIMER_MODE_ABS_PINNED
, 0);
220 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
223 #ifdef CONFIG_RT_GROUP_SCHED
224 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 hrtimer_cancel(&rt_b
->rt_period_timer
);
231 * sched_domains_mutex serializes calls to arch_init_sched_domains,
232 * detach_destroy_domains and partition_sched_domains.
234 static DEFINE_MUTEX(sched_domains_mutex
);
236 #ifdef CONFIG_CGROUP_SCHED
238 #include <linux/cgroup.h>
242 static LIST_HEAD(task_groups
);
244 /* task group related information */
246 struct cgroup_subsys_state css
;
248 #ifdef CONFIG_FAIR_GROUP_SCHED
249 /* schedulable entities of this group on each cpu */
250 struct sched_entity
**se
;
251 /* runqueue "owned" by this group on each cpu */
252 struct cfs_rq
**cfs_rq
;
253 unsigned long shares
;
256 #ifdef CONFIG_RT_GROUP_SCHED
257 struct sched_rt_entity
**rt_se
;
258 struct rt_rq
**rt_rq
;
260 struct rt_bandwidth rt_bandwidth
;
264 struct list_head list
;
266 struct task_group
*parent
;
267 struct list_head siblings
;
268 struct list_head children
;
271 #define root_task_group init_task_group
273 /* task_group_lock serializes add/remove of task groups and also changes to
274 * a task group's cpu shares.
276 static DEFINE_SPINLOCK(task_group_lock
);
278 #ifdef CONFIG_FAIR_GROUP_SCHED
281 static int root_task_group_empty(void)
283 return list_empty(&root_task_group
.children
);
287 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
290 * A weight of 0 or 1 can cause arithmetics problems.
291 * A weight of a cfs_rq is the sum of weights of which entities
292 * are queued on this cfs_rq, so a weight of a entity should not be
293 * too large, so as the shares value of a task group.
294 * (The default weight is 1024 - so there's no practical
295 * limitation from this.)
298 #define MAX_SHARES (1UL << 18)
300 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
303 /* Default task group.
304 * Every task in system belong to this group at bootup.
306 struct task_group init_task_group
;
308 /* return group to which a task belongs */
309 static inline struct task_group
*task_group(struct task_struct
*p
)
311 struct task_group
*tg
;
313 #ifdef CONFIG_CGROUP_SCHED
314 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
315 struct task_group
, css
);
317 tg
= &init_task_group
;
322 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
323 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
325 #ifdef CONFIG_FAIR_GROUP_SCHED
326 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
327 p
->se
.parent
= task_group(p
)->se
[cpu
];
330 #ifdef CONFIG_RT_GROUP_SCHED
331 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
332 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
338 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
339 static inline struct task_group
*task_group(struct task_struct
*p
)
344 #endif /* CONFIG_CGROUP_SCHED */
346 /* CFS-related fields in a runqueue */
348 struct load_weight load
;
349 unsigned long nr_running
;
354 struct rb_root tasks_timeline
;
355 struct rb_node
*rb_leftmost
;
357 struct list_head tasks
;
358 struct list_head
*balance_iterator
;
361 * 'curr' points to currently running entity on this cfs_rq.
362 * It is set to NULL otherwise (i.e when none are currently running).
364 struct sched_entity
*curr
, *next
, *last
;
366 unsigned int nr_spread_over
;
368 #ifdef CONFIG_FAIR_GROUP_SCHED
369 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
372 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
373 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
374 * (like users, containers etc.)
376 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
377 * list is used during load balance.
379 struct list_head leaf_cfs_rq_list
;
380 struct task_group
*tg
; /* group that "owns" this runqueue */
384 * the part of load.weight contributed by tasks
386 unsigned long task_weight
;
389 * h_load = weight * f(tg)
391 * Where f(tg) is the recursive weight fraction assigned to
394 unsigned long h_load
;
397 * this cpu's part of tg->shares
399 unsigned long shares
;
402 * load.weight at the time we set shares
404 unsigned long rq_weight
;
409 /* Real-Time classes' related field in a runqueue: */
411 struct rt_prio_array active
;
412 unsigned long rt_nr_running
;
413 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
415 int curr
; /* highest queued rt task prio */
417 int next
; /* next highest */
422 unsigned long rt_nr_migratory
;
423 unsigned long rt_nr_total
;
425 struct plist_head pushable_tasks
;
430 /* Nests inside the rq lock: */
431 raw_spinlock_t rt_runtime_lock
;
433 #ifdef CONFIG_RT_GROUP_SCHED
434 unsigned long rt_nr_boosted
;
437 struct list_head leaf_rt_rq_list
;
438 struct task_group
*tg
;
445 * We add the notion of a root-domain which will be used to define per-domain
446 * variables. Each exclusive cpuset essentially defines an island domain by
447 * fully partitioning the member cpus from any other cpuset. Whenever a new
448 * exclusive cpuset is created, we also create and attach a new root-domain
455 cpumask_var_t online
;
458 * The "RT overload" flag: it gets set if a CPU has more than
459 * one runnable RT task.
461 cpumask_var_t rto_mask
;
464 struct cpupri cpupri
;
469 * By default the system creates a single root-domain with all cpus as
470 * members (mimicking the global state we have today).
472 static struct root_domain def_root_domain
;
477 * This is the main, per-CPU runqueue data structure.
479 * Locking rule: those places that want to lock multiple runqueues
480 * (such as the load balancing or the thread migration code), lock
481 * acquire operations must be ordered by ascending &runqueue.
488 * nr_running and cpu_load should be in the same cacheline because
489 * remote CPUs use both these fields when doing load calculation.
491 unsigned long nr_running
;
492 #define CPU_LOAD_IDX_MAX 5
493 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
496 unsigned char in_nohz_recently
;
498 /* capture load from *all* tasks on this cpu: */
499 struct load_weight load
;
500 unsigned long nr_load_updates
;
506 #ifdef CONFIG_FAIR_GROUP_SCHED
507 /* list of leaf cfs_rq on this cpu: */
508 struct list_head leaf_cfs_rq_list
;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 struct list_head leaf_rt_rq_list
;
515 * This is part of a global counter where only the total sum
516 * over all CPUs matters. A task can increase this counter on
517 * one CPU and if it got migrated afterwards it may decrease
518 * it on another CPU. Always updated under the runqueue lock:
520 unsigned long nr_uninterruptible
;
522 struct task_struct
*curr
, *idle
;
523 unsigned long next_balance
;
524 struct mm_struct
*prev_mm
;
531 struct root_domain
*rd
;
532 struct sched_domain
*sd
;
534 unsigned char idle_at_tick
;
535 /* For active balancing */
539 /* cpu of this runqueue: */
543 unsigned long avg_load_per_task
;
545 struct task_struct
*migration_thread
;
546 struct list_head migration_queue
;
554 /* calc_load related fields */
555 unsigned long calc_load_update
;
556 long calc_load_active
;
558 #ifdef CONFIG_SCHED_HRTICK
560 int hrtick_csd_pending
;
561 struct call_single_data hrtick_csd
;
563 struct hrtimer hrtick_timer
;
566 #ifdef CONFIG_SCHEDSTATS
568 struct sched_info rq_sched_info
;
569 unsigned long long rq_cpu_time
;
570 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
572 /* sys_sched_yield() stats */
573 unsigned int yld_count
;
575 /* schedule() stats */
576 unsigned int sched_switch
;
577 unsigned int sched_count
;
578 unsigned int sched_goidle
;
580 /* try_to_wake_up() stats */
581 unsigned int ttwu_count
;
582 unsigned int ttwu_local
;
585 unsigned int bkl_count
;
589 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
592 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
594 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
597 static inline int cpu_of(struct rq
*rq
)
606 #define rcu_dereference_check_sched_domain(p) \
607 rcu_dereference_check((p), \
608 rcu_read_lock_sched_held() || \
609 lockdep_is_held(&sched_domains_mutex))
612 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
613 * See detach_destroy_domains: synchronize_sched for details.
615 * The domain tree of any CPU may only be accessed from within
616 * preempt-disabled sections.
618 #define for_each_domain(cpu, __sd) \
619 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
621 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
622 #define this_rq() (&__get_cpu_var(runqueues))
623 #define task_rq(p) cpu_rq(task_cpu(p))
624 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
625 #define raw_rq() (&__raw_get_cpu_var(runqueues))
627 inline void update_rq_clock(struct rq
*rq
)
629 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
633 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
635 #ifdef CONFIG_SCHED_DEBUG
636 # define const_debug __read_mostly
638 # define const_debug static const
643 * @cpu: the processor in question.
645 * Returns true if the current cpu runqueue is locked.
646 * This interface allows printk to be called with the runqueue lock
647 * held and know whether or not it is OK to wake up the klogd.
649 int runqueue_is_locked(int cpu
)
651 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
655 * Debugging: various feature bits
658 #define SCHED_FEAT(name, enabled) \
659 __SCHED_FEAT_##name ,
662 #include "sched_features.h"
667 #define SCHED_FEAT(name, enabled) \
668 (1UL << __SCHED_FEAT_##name) * enabled |
670 const_debug
unsigned int sysctl_sched_features
=
671 #include "sched_features.h"
676 #ifdef CONFIG_SCHED_DEBUG
677 #define SCHED_FEAT(name, enabled) \
680 static __read_mostly
char *sched_feat_names
[] = {
681 #include "sched_features.h"
687 static int sched_feat_show(struct seq_file
*m
, void *v
)
691 for (i
= 0; sched_feat_names
[i
]; i
++) {
692 if (!(sysctl_sched_features
& (1UL << i
)))
694 seq_printf(m
, "%s ", sched_feat_names
[i
]);
702 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
703 size_t cnt
, loff_t
*ppos
)
713 if (copy_from_user(&buf
, ubuf
, cnt
))
718 if (strncmp(buf
, "NO_", 3) == 0) {
723 for (i
= 0; sched_feat_names
[i
]; i
++) {
724 int len
= strlen(sched_feat_names
[i
]);
726 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
728 sysctl_sched_features
&= ~(1UL << i
);
730 sysctl_sched_features
|= (1UL << i
);
735 if (!sched_feat_names
[i
])
743 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
745 return single_open(filp
, sched_feat_show
, NULL
);
748 static const struct file_operations sched_feat_fops
= {
749 .open
= sched_feat_open
,
750 .write
= sched_feat_write
,
753 .release
= single_release
,
756 static __init
int sched_init_debug(void)
758 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
763 late_initcall(sched_init_debug
);
767 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
770 * Number of tasks to iterate in a single balance run.
771 * Limited because this is done with IRQs disabled.
773 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
776 * ratelimit for updating the group shares.
779 unsigned int sysctl_sched_shares_ratelimit
= 250000;
780 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
783 * Inject some fuzzyness into changing the per-cpu group shares
784 * this avoids remote rq-locks at the expense of fairness.
787 unsigned int sysctl_sched_shares_thresh
= 4;
790 * period over which we average the RT time consumption, measured
795 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
798 * period over which we measure -rt task cpu usage in us.
801 unsigned int sysctl_sched_rt_period
= 1000000;
803 static __read_mostly
int scheduler_running
;
806 * part of the period that we allow rt tasks to run in us.
809 int sysctl_sched_rt_runtime
= 950000;
811 static inline u64
global_rt_period(void)
813 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
816 static inline u64
global_rt_runtime(void)
818 if (sysctl_sched_rt_runtime
< 0)
821 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
824 #ifndef prepare_arch_switch
825 # define prepare_arch_switch(next) do { } while (0)
827 #ifndef finish_arch_switch
828 # define finish_arch_switch(prev) do { } while (0)
831 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
833 return rq
->curr
== p
;
836 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
837 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
839 return task_current(rq
, p
);
842 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
846 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
848 #ifdef CONFIG_DEBUG_SPINLOCK
849 /* this is a valid case when another task releases the spinlock */
850 rq
->lock
.owner
= current
;
853 * If we are tracking spinlock dependencies then we have to
854 * fix up the runqueue lock - which gets 'carried over' from
857 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
859 raw_spin_unlock_irq(&rq
->lock
);
862 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
863 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
868 return task_current(rq
, p
);
872 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
876 * We can optimise this out completely for !SMP, because the
877 * SMP rebalancing from interrupt is the only thing that cares
882 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
883 raw_spin_unlock_irq(&rq
->lock
);
885 raw_spin_unlock(&rq
->lock
);
889 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
893 * After ->oncpu is cleared, the task can be moved to a different CPU.
894 * We must ensure this doesn't happen until the switch is completely
900 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
904 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
907 * Check whether the task is waking, we use this to synchronize against
908 * ttwu() so that task_cpu() reports a stable number.
910 * We need to make an exception for PF_STARTING tasks because the fork
911 * path might require task_rq_lock() to work, eg. it can call
912 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
914 static inline int task_is_waking(struct task_struct
*p
)
916 return unlikely((p
->state
== TASK_WAKING
) && !(p
->flags
& PF_STARTING
));
920 * __task_rq_lock - lock the runqueue a given task resides on.
921 * Must be called interrupts disabled.
923 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
929 while (task_is_waking(p
))
932 raw_spin_lock(&rq
->lock
);
933 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
935 raw_spin_unlock(&rq
->lock
);
940 * task_rq_lock - lock the runqueue a given task resides on and disable
941 * interrupts. Note the ordering: we can safely lookup the task_rq without
942 * explicitly disabling preemption.
944 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
950 while (task_is_waking(p
))
952 local_irq_save(*flags
);
954 raw_spin_lock(&rq
->lock
);
955 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
957 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
961 void task_rq_unlock_wait(struct task_struct
*p
)
963 struct rq
*rq
= task_rq(p
);
965 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
966 raw_spin_unlock_wait(&rq
->lock
);
969 static void __task_rq_unlock(struct rq
*rq
)
972 raw_spin_unlock(&rq
->lock
);
975 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
978 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
982 * this_rq_lock - lock this runqueue and disable interrupts.
984 static struct rq
*this_rq_lock(void)
991 raw_spin_lock(&rq
->lock
);
996 #ifdef CONFIG_SCHED_HRTICK
998 * Use HR-timers to deliver accurate preemption points.
1000 * Its all a bit involved since we cannot program an hrt while holding the
1001 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1004 * When we get rescheduled we reprogram the hrtick_timer outside of the
1010 * - enabled by features
1011 * - hrtimer is actually high res
1013 static inline int hrtick_enabled(struct rq
*rq
)
1015 if (!sched_feat(HRTICK
))
1017 if (!cpu_active(cpu_of(rq
)))
1019 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1022 static void hrtick_clear(struct rq
*rq
)
1024 if (hrtimer_active(&rq
->hrtick_timer
))
1025 hrtimer_cancel(&rq
->hrtick_timer
);
1029 * High-resolution timer tick.
1030 * Runs from hardirq context with interrupts disabled.
1032 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1034 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1036 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1038 raw_spin_lock(&rq
->lock
);
1039 update_rq_clock(rq
);
1040 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1041 raw_spin_unlock(&rq
->lock
);
1043 return HRTIMER_NORESTART
;
1048 * called from hardirq (IPI) context
1050 static void __hrtick_start(void *arg
)
1052 struct rq
*rq
= arg
;
1054 raw_spin_lock(&rq
->lock
);
1055 hrtimer_restart(&rq
->hrtick_timer
);
1056 rq
->hrtick_csd_pending
= 0;
1057 raw_spin_unlock(&rq
->lock
);
1061 * Called to set the hrtick timer state.
1063 * called with rq->lock held and irqs disabled
1065 static void hrtick_start(struct rq
*rq
, u64 delay
)
1067 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1068 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1070 hrtimer_set_expires(timer
, time
);
1072 if (rq
== this_rq()) {
1073 hrtimer_restart(timer
);
1074 } else if (!rq
->hrtick_csd_pending
) {
1075 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1076 rq
->hrtick_csd_pending
= 1;
1081 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1083 int cpu
= (int)(long)hcpu
;
1086 case CPU_UP_CANCELED
:
1087 case CPU_UP_CANCELED_FROZEN
:
1088 case CPU_DOWN_PREPARE
:
1089 case CPU_DOWN_PREPARE_FROZEN
:
1091 case CPU_DEAD_FROZEN
:
1092 hrtick_clear(cpu_rq(cpu
));
1099 static __init
void init_hrtick(void)
1101 hotcpu_notifier(hotplug_hrtick
, 0);
1105 * Called to set the hrtick timer state.
1107 * called with rq->lock held and irqs disabled
1109 static void hrtick_start(struct rq
*rq
, u64 delay
)
1111 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1112 HRTIMER_MODE_REL_PINNED
, 0);
1115 static inline void init_hrtick(void)
1118 #endif /* CONFIG_SMP */
1120 static void init_rq_hrtick(struct rq
*rq
)
1123 rq
->hrtick_csd_pending
= 0;
1125 rq
->hrtick_csd
.flags
= 0;
1126 rq
->hrtick_csd
.func
= __hrtick_start
;
1127 rq
->hrtick_csd
.info
= rq
;
1130 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1131 rq
->hrtick_timer
.function
= hrtick
;
1133 #else /* CONFIG_SCHED_HRTICK */
1134 static inline void hrtick_clear(struct rq
*rq
)
1138 static inline void init_rq_hrtick(struct rq
*rq
)
1142 static inline void init_hrtick(void)
1145 #endif /* CONFIG_SCHED_HRTICK */
1148 * resched_task - mark a task 'to be rescheduled now'.
1150 * On UP this means the setting of the need_resched flag, on SMP it
1151 * might also involve a cross-CPU call to trigger the scheduler on
1156 #ifndef tsk_is_polling
1157 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1160 static void resched_task(struct task_struct
*p
)
1164 assert_raw_spin_locked(&task_rq(p
)->lock
);
1166 if (test_tsk_need_resched(p
))
1169 set_tsk_need_resched(p
);
1172 if (cpu
== smp_processor_id())
1175 /* NEED_RESCHED must be visible before we test polling */
1177 if (!tsk_is_polling(p
))
1178 smp_send_reschedule(cpu
);
1181 static void resched_cpu(int cpu
)
1183 struct rq
*rq
= cpu_rq(cpu
);
1184 unsigned long flags
;
1186 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1188 resched_task(cpu_curr(cpu
));
1189 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1194 * When add_timer_on() enqueues a timer into the timer wheel of an
1195 * idle CPU then this timer might expire before the next timer event
1196 * which is scheduled to wake up that CPU. In case of a completely
1197 * idle system the next event might even be infinite time into the
1198 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1199 * leaves the inner idle loop so the newly added timer is taken into
1200 * account when the CPU goes back to idle and evaluates the timer
1201 * wheel for the next timer event.
1203 void wake_up_idle_cpu(int cpu
)
1205 struct rq
*rq
= cpu_rq(cpu
);
1207 if (cpu
== smp_processor_id())
1211 * This is safe, as this function is called with the timer
1212 * wheel base lock of (cpu) held. When the CPU is on the way
1213 * to idle and has not yet set rq->curr to idle then it will
1214 * be serialized on the timer wheel base lock and take the new
1215 * timer into account automatically.
1217 if (rq
->curr
!= rq
->idle
)
1221 * We can set TIF_RESCHED on the idle task of the other CPU
1222 * lockless. The worst case is that the other CPU runs the
1223 * idle task through an additional NOOP schedule()
1225 set_tsk_need_resched(rq
->idle
);
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(rq
->idle
))
1230 smp_send_reschedule(cpu
);
1233 int nohz_ratelimit(int cpu
)
1235 struct rq
*rq
= cpu_rq(cpu
);
1236 u64 diff
= rq
->clock
- rq
->nohz_stamp
;
1238 rq
->nohz_stamp
= rq
->clock
;
1240 return diff
< (NSEC_PER_SEC
/ HZ
) >> 1;
1243 #endif /* CONFIG_NO_HZ */
1245 static u64
sched_avg_period(void)
1247 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1250 static void sched_avg_update(struct rq
*rq
)
1252 s64 period
= sched_avg_period();
1254 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1255 rq
->age_stamp
+= period
;
1260 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1262 rq
->rt_avg
+= rt_delta
;
1263 sched_avg_update(rq
);
1266 #else /* !CONFIG_SMP */
1267 static void resched_task(struct task_struct
*p
)
1269 assert_raw_spin_locked(&task_rq(p
)->lock
);
1270 set_tsk_need_resched(p
);
1273 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1276 #endif /* CONFIG_SMP */
1278 #if BITS_PER_LONG == 32
1279 # define WMULT_CONST (~0UL)
1281 # define WMULT_CONST (1UL << 32)
1284 #define WMULT_SHIFT 32
1287 * Shift right and round:
1289 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1292 * delta *= weight / lw
1294 static unsigned long
1295 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1296 struct load_weight
*lw
)
1300 if (!lw
->inv_weight
) {
1301 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1304 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1308 tmp
= (u64
)delta_exec
* weight
;
1310 * Check whether we'd overflow the 64-bit multiplication:
1312 if (unlikely(tmp
> WMULT_CONST
))
1313 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1316 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1318 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1321 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1327 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1334 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1335 * of tasks with abnormal "nice" values across CPUs the contribution that
1336 * each task makes to its run queue's load is weighted according to its
1337 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1338 * scaled version of the new time slice allocation that they receive on time
1342 #define WEIGHT_IDLEPRIO 3
1343 #define WMULT_IDLEPRIO 1431655765
1346 * Nice levels are multiplicative, with a gentle 10% change for every
1347 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1348 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1349 * that remained on nice 0.
1351 * The "10% effect" is relative and cumulative: from _any_ nice level,
1352 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1353 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1354 * If a task goes up by ~10% and another task goes down by ~10% then
1355 * the relative distance between them is ~25%.)
1357 static const int prio_to_weight
[40] = {
1358 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1359 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1360 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1361 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1362 /* 0 */ 1024, 820, 655, 526, 423,
1363 /* 5 */ 335, 272, 215, 172, 137,
1364 /* 10 */ 110, 87, 70, 56, 45,
1365 /* 15 */ 36, 29, 23, 18, 15,
1369 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1371 * In cases where the weight does not change often, we can use the
1372 * precalculated inverse to speed up arithmetics by turning divisions
1373 * into multiplications:
1375 static const u32 prio_to_wmult
[40] = {
1376 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1377 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1378 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1379 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1380 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1381 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1382 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1383 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1386 /* Time spent by the tasks of the cpu accounting group executing in ... */
1387 enum cpuacct_stat_index
{
1388 CPUACCT_STAT_USER
, /* ... user mode */
1389 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1391 CPUACCT_STAT_NSTATS
,
1394 #ifdef CONFIG_CGROUP_CPUACCT
1395 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1396 static void cpuacct_update_stats(struct task_struct
*tsk
,
1397 enum cpuacct_stat_index idx
, cputime_t val
);
1399 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1400 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1401 enum cpuacct_stat_index idx
, cputime_t val
) {}
1404 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1406 update_load_add(&rq
->load
, load
);
1409 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1411 update_load_sub(&rq
->load
, load
);
1414 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1415 typedef int (*tg_visitor
)(struct task_group
*, void *);
1418 * Iterate the full tree, calling @down when first entering a node and @up when
1419 * leaving it for the final time.
1421 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1423 struct task_group
*parent
, *child
;
1427 parent
= &root_task_group
;
1429 ret
= (*down
)(parent
, data
);
1432 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1439 ret
= (*up
)(parent
, data
);
1444 parent
= parent
->parent
;
1453 static int tg_nop(struct task_group
*tg
, void *data
)
1460 /* Used instead of source_load when we know the type == 0 */
1461 static unsigned long weighted_cpuload(const int cpu
)
1463 return cpu_rq(cpu
)->load
.weight
;
1467 * Return a low guess at the load of a migration-source cpu weighted
1468 * according to the scheduling class and "nice" value.
1470 * We want to under-estimate the load of migration sources, to
1471 * balance conservatively.
1473 static unsigned long source_load(int cpu
, int type
)
1475 struct rq
*rq
= cpu_rq(cpu
);
1476 unsigned long total
= weighted_cpuload(cpu
);
1478 if (type
== 0 || !sched_feat(LB_BIAS
))
1481 return min(rq
->cpu_load
[type
-1], total
);
1485 * Return a high guess at the load of a migration-target cpu weighted
1486 * according to the scheduling class and "nice" value.
1488 static unsigned long target_load(int cpu
, int type
)
1490 struct rq
*rq
= cpu_rq(cpu
);
1491 unsigned long total
= weighted_cpuload(cpu
);
1493 if (type
== 0 || !sched_feat(LB_BIAS
))
1496 return max(rq
->cpu_load
[type
-1], total
);
1499 static struct sched_group
*group_of(int cpu
)
1501 struct sched_domain
*sd
= rcu_dereference_sched(cpu_rq(cpu
)->sd
);
1509 static unsigned long power_of(int cpu
)
1511 struct sched_group
*group
= group_of(cpu
);
1514 return SCHED_LOAD_SCALE
;
1516 return group
->cpu_power
;
1519 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1521 static unsigned long cpu_avg_load_per_task(int cpu
)
1523 struct rq
*rq
= cpu_rq(cpu
);
1524 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1527 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1529 rq
->avg_load_per_task
= 0;
1531 return rq
->avg_load_per_task
;
1534 #ifdef CONFIG_FAIR_GROUP_SCHED
1536 static __read_mostly
unsigned long *update_shares_data
;
1538 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1541 * Calculate and set the cpu's group shares.
1543 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1544 unsigned long sd_shares
,
1545 unsigned long sd_rq_weight
,
1546 unsigned long *usd_rq_weight
)
1548 unsigned long shares
, rq_weight
;
1551 rq_weight
= usd_rq_weight
[cpu
];
1554 rq_weight
= NICE_0_LOAD
;
1558 * \Sum_j shares_j * rq_weight_i
1559 * shares_i = -----------------------------
1560 * \Sum_j rq_weight_j
1562 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1563 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1565 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1566 sysctl_sched_shares_thresh
) {
1567 struct rq
*rq
= cpu_rq(cpu
);
1568 unsigned long flags
;
1570 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1571 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1572 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1573 __set_se_shares(tg
->se
[cpu
], shares
);
1574 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1579 * Re-compute the task group their per cpu shares over the given domain.
1580 * This needs to be done in a bottom-up fashion because the rq weight of a
1581 * parent group depends on the shares of its child groups.
1583 static int tg_shares_up(struct task_group
*tg
, void *data
)
1585 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1586 unsigned long *usd_rq_weight
;
1587 struct sched_domain
*sd
= data
;
1588 unsigned long flags
;
1594 local_irq_save(flags
);
1595 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1597 for_each_cpu(i
, sched_domain_span(sd
)) {
1598 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1599 usd_rq_weight
[i
] = weight
;
1601 rq_weight
+= weight
;
1603 * If there are currently no tasks on the cpu pretend there
1604 * is one of average load so that when a new task gets to
1605 * run here it will not get delayed by group starvation.
1608 weight
= NICE_0_LOAD
;
1610 sum_weight
+= weight
;
1611 shares
+= tg
->cfs_rq
[i
]->shares
;
1615 rq_weight
= sum_weight
;
1617 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1618 shares
= tg
->shares
;
1620 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1621 shares
= tg
->shares
;
1623 for_each_cpu(i
, sched_domain_span(sd
))
1624 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1626 local_irq_restore(flags
);
1632 * Compute the cpu's hierarchical load factor for each task group.
1633 * This needs to be done in a top-down fashion because the load of a child
1634 * group is a fraction of its parents load.
1636 static int tg_load_down(struct task_group
*tg
, void *data
)
1639 long cpu
= (long)data
;
1642 load
= cpu_rq(cpu
)->load
.weight
;
1644 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1645 load
*= tg
->cfs_rq
[cpu
]->shares
;
1646 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1649 tg
->cfs_rq
[cpu
]->h_load
= load
;
1654 static void update_shares(struct sched_domain
*sd
)
1659 if (root_task_group_empty())
1662 now
= cpu_clock(raw_smp_processor_id());
1663 elapsed
= now
- sd
->last_update
;
1665 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1666 sd
->last_update
= now
;
1667 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1671 static void update_h_load(long cpu
)
1673 if (root_task_group_empty())
1676 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1681 static inline void update_shares(struct sched_domain
*sd
)
1687 #ifdef CONFIG_PREEMPT
1689 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1692 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1693 * way at the expense of forcing extra atomic operations in all
1694 * invocations. This assures that the double_lock is acquired using the
1695 * same underlying policy as the spinlock_t on this architecture, which
1696 * reduces latency compared to the unfair variant below. However, it
1697 * also adds more overhead and therefore may reduce throughput.
1699 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1700 __releases(this_rq
->lock
)
1701 __acquires(busiest
->lock
)
1702 __acquires(this_rq
->lock
)
1704 raw_spin_unlock(&this_rq
->lock
);
1705 double_rq_lock(this_rq
, busiest
);
1712 * Unfair double_lock_balance: Optimizes throughput at the expense of
1713 * latency by eliminating extra atomic operations when the locks are
1714 * already in proper order on entry. This favors lower cpu-ids and will
1715 * grant the double lock to lower cpus over higher ids under contention,
1716 * regardless of entry order into the function.
1718 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1719 __releases(this_rq
->lock
)
1720 __acquires(busiest
->lock
)
1721 __acquires(this_rq
->lock
)
1725 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1726 if (busiest
< this_rq
) {
1727 raw_spin_unlock(&this_rq
->lock
);
1728 raw_spin_lock(&busiest
->lock
);
1729 raw_spin_lock_nested(&this_rq
->lock
,
1730 SINGLE_DEPTH_NESTING
);
1733 raw_spin_lock_nested(&busiest
->lock
,
1734 SINGLE_DEPTH_NESTING
);
1739 #endif /* CONFIG_PREEMPT */
1742 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1744 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1746 if (unlikely(!irqs_disabled())) {
1747 /* printk() doesn't work good under rq->lock */
1748 raw_spin_unlock(&this_rq
->lock
);
1752 return _double_lock_balance(this_rq
, busiest
);
1755 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1756 __releases(busiest
->lock
)
1758 raw_spin_unlock(&busiest
->lock
);
1759 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1763 * double_rq_lock - safely lock two runqueues
1765 * Note this does not disable interrupts like task_rq_lock,
1766 * you need to do so manually before calling.
1768 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1769 __acquires(rq1
->lock
)
1770 __acquires(rq2
->lock
)
1772 BUG_ON(!irqs_disabled());
1774 raw_spin_lock(&rq1
->lock
);
1775 __acquire(rq2
->lock
); /* Fake it out ;) */
1778 raw_spin_lock(&rq1
->lock
);
1779 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1781 raw_spin_lock(&rq2
->lock
);
1782 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1785 update_rq_clock(rq1
);
1786 update_rq_clock(rq2
);
1790 * double_rq_unlock - safely unlock two runqueues
1792 * Note this does not restore interrupts like task_rq_unlock,
1793 * you need to do so manually after calling.
1795 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1796 __releases(rq1
->lock
)
1797 __releases(rq2
->lock
)
1799 raw_spin_unlock(&rq1
->lock
);
1801 raw_spin_unlock(&rq2
->lock
);
1803 __release(rq2
->lock
);
1808 #ifdef CONFIG_FAIR_GROUP_SCHED
1809 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1812 cfs_rq
->shares
= shares
;
1817 static void calc_load_account_active(struct rq
*this_rq
);
1818 static void update_sysctl(void);
1819 static int get_update_sysctl_factor(void);
1821 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1823 set_task_rq(p
, cpu
);
1826 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1827 * successfuly executed on another CPU. We must ensure that updates of
1828 * per-task data have been completed by this moment.
1831 task_thread_info(p
)->cpu
= cpu
;
1835 static const struct sched_class rt_sched_class
;
1837 #define sched_class_highest (&rt_sched_class)
1838 #define for_each_class(class) \
1839 for (class = sched_class_highest; class; class = class->next)
1841 #include "sched_stats.h"
1843 static void inc_nr_running(struct rq
*rq
)
1848 static void dec_nr_running(struct rq
*rq
)
1853 static void set_load_weight(struct task_struct
*p
)
1855 if (task_has_rt_policy(p
)) {
1856 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1857 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1862 * SCHED_IDLE tasks get minimal weight:
1864 if (p
->policy
== SCHED_IDLE
) {
1865 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1866 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1870 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1871 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1874 static void update_avg(u64
*avg
, u64 sample
)
1876 s64 diff
= sample
- *avg
;
1881 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1884 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1886 sched_info_queued(p
);
1887 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1891 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1894 if (p
->se
.last_wakeup
) {
1895 update_avg(&p
->se
.avg_overlap
,
1896 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1897 p
->se
.last_wakeup
= 0;
1899 update_avg(&p
->se
.avg_wakeup
,
1900 sysctl_sched_wakeup_granularity
);
1904 sched_info_dequeued(p
);
1905 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1910 * activate_task - move a task to the runqueue.
1912 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1914 if (task_contributes_to_load(p
))
1915 rq
->nr_uninterruptible
--;
1917 enqueue_task(rq
, p
, wakeup
, false);
1922 * deactivate_task - remove a task from the runqueue.
1924 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1926 if (task_contributes_to_load(p
))
1927 rq
->nr_uninterruptible
++;
1929 dequeue_task(rq
, p
, sleep
);
1933 #include "sched_idletask.c"
1934 #include "sched_fair.c"
1935 #include "sched_rt.c"
1936 #ifdef CONFIG_SCHED_DEBUG
1937 # include "sched_debug.c"
1941 * __normal_prio - return the priority that is based on the static prio
1943 static inline int __normal_prio(struct task_struct
*p
)
1945 return p
->static_prio
;
1949 * Calculate the expected normal priority: i.e. priority
1950 * without taking RT-inheritance into account. Might be
1951 * boosted by interactivity modifiers. Changes upon fork,
1952 * setprio syscalls, and whenever the interactivity
1953 * estimator recalculates.
1955 static inline int normal_prio(struct task_struct
*p
)
1959 if (task_has_rt_policy(p
))
1960 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1962 prio
= __normal_prio(p
);
1967 * Calculate the current priority, i.e. the priority
1968 * taken into account by the scheduler. This value might
1969 * be boosted by RT tasks, or might be boosted by
1970 * interactivity modifiers. Will be RT if the task got
1971 * RT-boosted. If not then it returns p->normal_prio.
1973 static int effective_prio(struct task_struct
*p
)
1975 p
->normal_prio
= normal_prio(p
);
1977 * If we are RT tasks or we were boosted to RT priority,
1978 * keep the priority unchanged. Otherwise, update priority
1979 * to the normal priority:
1981 if (!rt_prio(p
->prio
))
1982 return p
->normal_prio
;
1987 * task_curr - is this task currently executing on a CPU?
1988 * @p: the task in question.
1990 inline int task_curr(const struct task_struct
*p
)
1992 return cpu_curr(task_cpu(p
)) == p
;
1995 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1996 const struct sched_class
*prev_class
,
1997 int oldprio
, int running
)
1999 if (prev_class
!= p
->sched_class
) {
2000 if (prev_class
->switched_from
)
2001 prev_class
->switched_from(rq
, p
, running
);
2002 p
->sched_class
->switched_to(rq
, p
, running
);
2004 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2009 * Is this task likely cache-hot:
2012 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2016 if (p
->sched_class
!= &fair_sched_class
)
2020 * Buddy candidates are cache hot:
2022 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2023 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2024 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2027 if (sysctl_sched_migration_cost
== -1)
2029 if (sysctl_sched_migration_cost
== 0)
2032 delta
= now
- p
->se
.exec_start
;
2034 return delta
< (s64
)sysctl_sched_migration_cost
;
2037 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2039 #ifdef CONFIG_SCHED_DEBUG
2041 * We should never call set_task_cpu() on a blocked task,
2042 * ttwu() will sort out the placement.
2044 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2045 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2048 trace_sched_migrate_task(p
, new_cpu
);
2050 if (task_cpu(p
) != new_cpu
) {
2051 p
->se
.nr_migrations
++;
2052 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2055 __set_task_cpu(p
, new_cpu
);
2058 struct migration_req
{
2059 struct list_head list
;
2061 struct task_struct
*task
;
2064 struct completion done
;
2068 * The task's runqueue lock must be held.
2069 * Returns true if you have to wait for migration thread.
2072 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2074 struct rq
*rq
= task_rq(p
);
2077 * If the task is not on a runqueue (and not running), then
2078 * the next wake-up will properly place the task.
2080 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2083 init_completion(&req
->done
);
2085 req
->dest_cpu
= dest_cpu
;
2086 list_add(&req
->list
, &rq
->migration_queue
);
2092 * wait_task_context_switch - wait for a thread to complete at least one
2095 * @p must not be current.
2097 void wait_task_context_switch(struct task_struct
*p
)
2099 unsigned long nvcsw
, nivcsw
, flags
;
2107 * The runqueue is assigned before the actual context
2108 * switch. We need to take the runqueue lock.
2110 * We could check initially without the lock but it is
2111 * very likely that we need to take the lock in every
2114 rq
= task_rq_lock(p
, &flags
);
2115 running
= task_running(rq
, p
);
2116 task_rq_unlock(rq
, &flags
);
2118 if (likely(!running
))
2121 * The switch count is incremented before the actual
2122 * context switch. We thus wait for two switches to be
2123 * sure at least one completed.
2125 if ((p
->nvcsw
- nvcsw
) > 1)
2127 if ((p
->nivcsw
- nivcsw
) > 1)
2135 * wait_task_inactive - wait for a thread to unschedule.
2137 * If @match_state is nonzero, it's the @p->state value just checked and
2138 * not expected to change. If it changes, i.e. @p might have woken up,
2139 * then return zero. When we succeed in waiting for @p to be off its CPU,
2140 * we return a positive number (its total switch count). If a second call
2141 * a short while later returns the same number, the caller can be sure that
2142 * @p has remained unscheduled the whole time.
2144 * The caller must ensure that the task *will* unschedule sometime soon,
2145 * else this function might spin for a *long* time. This function can't
2146 * be called with interrupts off, or it may introduce deadlock with
2147 * smp_call_function() if an IPI is sent by the same process we are
2148 * waiting to become inactive.
2150 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2152 unsigned long flags
;
2159 * We do the initial early heuristics without holding
2160 * any task-queue locks at all. We'll only try to get
2161 * the runqueue lock when things look like they will
2167 * If the task is actively running on another CPU
2168 * still, just relax and busy-wait without holding
2171 * NOTE! Since we don't hold any locks, it's not
2172 * even sure that "rq" stays as the right runqueue!
2173 * But we don't care, since "task_running()" will
2174 * return false if the runqueue has changed and p
2175 * is actually now running somewhere else!
2177 while (task_running(rq
, p
)) {
2178 if (match_state
&& unlikely(p
->state
!= match_state
))
2184 * Ok, time to look more closely! We need the rq
2185 * lock now, to be *sure*. If we're wrong, we'll
2186 * just go back and repeat.
2188 rq
= task_rq_lock(p
, &flags
);
2189 trace_sched_wait_task(rq
, p
);
2190 running
= task_running(rq
, p
);
2191 on_rq
= p
->se
.on_rq
;
2193 if (!match_state
|| p
->state
== match_state
)
2194 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2195 task_rq_unlock(rq
, &flags
);
2198 * If it changed from the expected state, bail out now.
2200 if (unlikely(!ncsw
))
2204 * Was it really running after all now that we
2205 * checked with the proper locks actually held?
2207 * Oops. Go back and try again..
2209 if (unlikely(running
)) {
2215 * It's not enough that it's not actively running,
2216 * it must be off the runqueue _entirely_, and not
2219 * So if it was still runnable (but just not actively
2220 * running right now), it's preempted, and we should
2221 * yield - it could be a while.
2223 if (unlikely(on_rq
)) {
2224 schedule_timeout_uninterruptible(1);
2229 * Ahh, all good. It wasn't running, and it wasn't
2230 * runnable, which means that it will never become
2231 * running in the future either. We're all done!
2240 * kick_process - kick a running thread to enter/exit the kernel
2241 * @p: the to-be-kicked thread
2243 * Cause a process which is running on another CPU to enter
2244 * kernel-mode, without any delay. (to get signals handled.)
2246 * NOTE: this function doesnt have to take the runqueue lock,
2247 * because all it wants to ensure is that the remote task enters
2248 * the kernel. If the IPI races and the task has been migrated
2249 * to another CPU then no harm is done and the purpose has been
2252 void kick_process(struct task_struct
*p
)
2258 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2259 smp_send_reschedule(cpu
);
2262 EXPORT_SYMBOL_GPL(kick_process
);
2263 #endif /* CONFIG_SMP */
2266 * task_oncpu_function_call - call a function on the cpu on which a task runs
2267 * @p: the task to evaluate
2268 * @func: the function to be called
2269 * @info: the function call argument
2271 * Calls the function @func when the task is currently running. This might
2272 * be on the current CPU, which just calls the function directly
2274 void task_oncpu_function_call(struct task_struct
*p
,
2275 void (*func
) (void *info
), void *info
)
2282 smp_call_function_single(cpu
, func
, info
, 1);
2287 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2290 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2292 /* Look for allowed, online CPU in same node. */
2293 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2294 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2297 /* Any allowed, online CPU? */
2298 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2299 if (dest_cpu
< nr_cpu_ids
)
2302 /* No more Mr. Nice Guy. */
2303 if (dest_cpu
>= nr_cpu_ids
) {
2305 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
2307 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
2310 * Don't tell them about moving exiting tasks or
2311 * kernel threads (both mm NULL), since they never
2314 if (p
->mm
&& printk_ratelimit()) {
2315 printk(KERN_INFO
"process %d (%s) no "
2316 "longer affine to cpu%d\n",
2317 task_pid_nr(p
), p
->comm
, cpu
);
2325 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2326 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2329 * exec: is unstable, retry loop
2330 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2333 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2335 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2338 * In order not to call set_task_cpu() on a blocking task we need
2339 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2342 * Since this is common to all placement strategies, this lives here.
2344 * [ this allows ->select_task() to simply return task_cpu(p) and
2345 * not worry about this generic constraint ]
2347 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2349 cpu
= select_fallback_rq(task_cpu(p
), p
);
2356 * try_to_wake_up - wake up a thread
2357 * @p: the to-be-woken-up thread
2358 * @state: the mask of task states that can be woken
2359 * @sync: do a synchronous wakeup?
2361 * Put it on the run-queue if it's not already there. The "current"
2362 * thread is always on the run-queue (except when the actual
2363 * re-schedule is in progress), and as such you're allowed to do
2364 * the simpler "current->state = TASK_RUNNING" to mark yourself
2365 * runnable without the overhead of this.
2367 * returns failure only if the task is already active.
2369 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2372 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2373 unsigned long flags
;
2376 if (!sched_feat(SYNC_WAKEUPS
))
2377 wake_flags
&= ~WF_SYNC
;
2379 this_cpu
= get_cpu();
2382 rq
= task_rq_lock(p
, &flags
);
2383 update_rq_clock(rq
);
2384 if (!(p
->state
& state
))
2394 if (unlikely(task_running(rq
, p
)))
2398 * In order to handle concurrent wakeups and release the rq->lock
2399 * we put the task in TASK_WAKING state.
2401 * First fix up the nr_uninterruptible count:
2403 if (task_contributes_to_load(p
))
2404 rq
->nr_uninterruptible
--;
2405 p
->state
= TASK_WAKING
;
2407 if (p
->sched_class
->task_waking
)
2408 p
->sched_class
->task_waking(rq
, p
);
2410 __task_rq_unlock(rq
);
2412 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2413 if (cpu
!= orig_cpu
) {
2415 * Since we migrate the task without holding any rq->lock,
2416 * we need to be careful with task_rq_lock(), since that
2417 * might end up locking an invalid rq.
2419 set_task_cpu(p
, cpu
);
2423 raw_spin_lock(&rq
->lock
);
2424 update_rq_clock(rq
);
2427 * We migrated the task without holding either rq->lock, however
2428 * since the task is not on the task list itself, nobody else
2429 * will try and migrate the task, hence the rq should match the
2430 * cpu we just moved it to.
2432 WARN_ON(task_cpu(p
) != cpu
);
2433 WARN_ON(p
->state
!= TASK_WAKING
);
2435 #ifdef CONFIG_SCHEDSTATS
2436 schedstat_inc(rq
, ttwu_count
);
2437 if (cpu
== this_cpu
)
2438 schedstat_inc(rq
, ttwu_local
);
2440 struct sched_domain
*sd
;
2441 for_each_domain(this_cpu
, sd
) {
2442 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2443 schedstat_inc(sd
, ttwu_wake_remote
);
2448 #endif /* CONFIG_SCHEDSTATS */
2451 #endif /* CONFIG_SMP */
2452 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2453 if (wake_flags
& WF_SYNC
)
2454 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2455 if (orig_cpu
!= cpu
)
2456 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2457 if (cpu
== this_cpu
)
2458 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2460 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2461 activate_task(rq
, p
, 1);
2465 * Only attribute actual wakeups done by this task.
2467 if (!in_interrupt()) {
2468 struct sched_entity
*se
= ¤t
->se
;
2469 u64 sample
= se
->sum_exec_runtime
;
2471 if (se
->last_wakeup
)
2472 sample
-= se
->last_wakeup
;
2474 sample
-= se
->start_runtime
;
2475 update_avg(&se
->avg_wakeup
, sample
);
2477 se
->last_wakeup
= se
->sum_exec_runtime
;
2481 trace_sched_wakeup(rq
, p
, success
);
2482 check_preempt_curr(rq
, p
, wake_flags
);
2484 p
->state
= TASK_RUNNING
;
2486 if (p
->sched_class
->task_woken
)
2487 p
->sched_class
->task_woken(rq
, p
);
2489 if (unlikely(rq
->idle_stamp
)) {
2490 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2491 u64 max
= 2*sysctl_sched_migration_cost
;
2496 update_avg(&rq
->avg_idle
, delta
);
2501 task_rq_unlock(rq
, &flags
);
2508 * wake_up_process - Wake up a specific process
2509 * @p: The process to be woken up.
2511 * Attempt to wake up the nominated process and move it to the set of runnable
2512 * processes. Returns 1 if the process was woken up, 0 if it was already
2515 * It may be assumed that this function implies a write memory barrier before
2516 * changing the task state if and only if any tasks are woken up.
2518 int wake_up_process(struct task_struct
*p
)
2520 return try_to_wake_up(p
, TASK_ALL
, 0);
2522 EXPORT_SYMBOL(wake_up_process
);
2524 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2526 return try_to_wake_up(p
, state
, 0);
2530 * Perform scheduler related setup for a newly forked process p.
2531 * p is forked by current.
2533 * __sched_fork() is basic setup used by init_idle() too:
2535 static void __sched_fork(struct task_struct
*p
)
2537 p
->se
.exec_start
= 0;
2538 p
->se
.sum_exec_runtime
= 0;
2539 p
->se
.prev_sum_exec_runtime
= 0;
2540 p
->se
.nr_migrations
= 0;
2541 p
->se
.last_wakeup
= 0;
2542 p
->se
.avg_overlap
= 0;
2543 p
->se
.start_runtime
= 0;
2544 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2546 #ifdef CONFIG_SCHEDSTATS
2547 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2550 INIT_LIST_HEAD(&p
->rt
.run_list
);
2552 INIT_LIST_HEAD(&p
->se
.group_node
);
2554 #ifdef CONFIG_PREEMPT_NOTIFIERS
2555 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2560 * fork()/clone()-time setup:
2562 void sched_fork(struct task_struct
*p
, int clone_flags
)
2564 int cpu
= get_cpu();
2568 * We mark the process as waking here. This guarantees that
2569 * nobody will actually run it, and a signal or other external
2570 * event cannot wake it up and insert it on the runqueue either.
2572 p
->state
= TASK_WAKING
;
2575 * Revert to default priority/policy on fork if requested.
2577 if (unlikely(p
->sched_reset_on_fork
)) {
2578 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2579 p
->policy
= SCHED_NORMAL
;
2580 p
->normal_prio
= p
->static_prio
;
2583 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2584 p
->static_prio
= NICE_TO_PRIO(0);
2585 p
->normal_prio
= p
->static_prio
;
2590 * We don't need the reset flag anymore after the fork. It has
2591 * fulfilled its duty:
2593 p
->sched_reset_on_fork
= 0;
2597 * Make sure we do not leak PI boosting priority to the child.
2599 p
->prio
= current
->normal_prio
;
2601 if (!rt_prio(p
->prio
))
2602 p
->sched_class
= &fair_sched_class
;
2604 if (p
->sched_class
->task_fork
)
2605 p
->sched_class
->task_fork(p
);
2607 set_task_cpu(p
, cpu
);
2609 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2610 if (likely(sched_info_on()))
2611 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2613 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2616 #ifdef CONFIG_PREEMPT
2617 /* Want to start with kernel preemption disabled. */
2618 task_thread_info(p
)->preempt_count
= 1;
2620 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2626 * wake_up_new_task - wake up a newly created task for the first time.
2628 * This function will do some initial scheduler statistics housekeeping
2629 * that must be done for every newly created context, then puts the task
2630 * on the runqueue and wakes it.
2632 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2634 unsigned long flags
;
2636 int cpu
= get_cpu();
2640 * Fork balancing, do it here and not earlier because:
2641 * - cpus_allowed can change in the fork path
2642 * - any previously selected cpu might disappear through hotplug
2644 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2645 * ->cpus_allowed is stable, we have preemption disabled, meaning
2646 * cpu_online_mask is stable.
2648 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2649 set_task_cpu(p
, cpu
);
2653 * Since the task is not on the rq and we still have TASK_WAKING set
2654 * nobody else will migrate this task.
2657 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2659 BUG_ON(p
->state
!= TASK_WAKING
);
2660 p
->state
= TASK_RUNNING
;
2661 update_rq_clock(rq
);
2662 activate_task(rq
, p
, 0);
2663 trace_sched_wakeup_new(rq
, p
, 1);
2664 check_preempt_curr(rq
, p
, WF_FORK
);
2666 if (p
->sched_class
->task_woken
)
2667 p
->sched_class
->task_woken(rq
, p
);
2669 task_rq_unlock(rq
, &flags
);
2673 #ifdef CONFIG_PREEMPT_NOTIFIERS
2676 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2677 * @notifier: notifier struct to register
2679 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2681 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2683 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2686 * preempt_notifier_unregister - no longer interested in preemption notifications
2687 * @notifier: notifier struct to unregister
2689 * This is safe to call from within a preemption notifier.
2691 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2693 hlist_del(¬ifier
->link
);
2695 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2697 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2699 struct preempt_notifier
*notifier
;
2700 struct hlist_node
*node
;
2702 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2703 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2707 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2708 struct task_struct
*next
)
2710 struct preempt_notifier
*notifier
;
2711 struct hlist_node
*node
;
2713 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2714 notifier
->ops
->sched_out(notifier
, next
);
2717 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2719 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2724 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2725 struct task_struct
*next
)
2729 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2732 * prepare_task_switch - prepare to switch tasks
2733 * @rq: the runqueue preparing to switch
2734 * @prev: the current task that is being switched out
2735 * @next: the task we are going to switch to.
2737 * This is called with the rq lock held and interrupts off. It must
2738 * be paired with a subsequent finish_task_switch after the context
2741 * prepare_task_switch sets up locking and calls architecture specific
2745 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2746 struct task_struct
*next
)
2748 fire_sched_out_preempt_notifiers(prev
, next
);
2749 prepare_lock_switch(rq
, next
);
2750 prepare_arch_switch(next
);
2754 * finish_task_switch - clean up after a task-switch
2755 * @rq: runqueue associated with task-switch
2756 * @prev: the thread we just switched away from.
2758 * finish_task_switch must be called after the context switch, paired
2759 * with a prepare_task_switch call before the context switch.
2760 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2761 * and do any other architecture-specific cleanup actions.
2763 * Note that we may have delayed dropping an mm in context_switch(). If
2764 * so, we finish that here outside of the runqueue lock. (Doing it
2765 * with the lock held can cause deadlocks; see schedule() for
2768 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2769 __releases(rq
->lock
)
2771 struct mm_struct
*mm
= rq
->prev_mm
;
2777 * A task struct has one reference for the use as "current".
2778 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2779 * schedule one last time. The schedule call will never return, and
2780 * the scheduled task must drop that reference.
2781 * The test for TASK_DEAD must occur while the runqueue locks are
2782 * still held, otherwise prev could be scheduled on another cpu, die
2783 * there before we look at prev->state, and then the reference would
2785 * Manfred Spraul <manfred@colorfullife.com>
2787 prev_state
= prev
->state
;
2788 finish_arch_switch(prev
);
2789 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2790 local_irq_disable();
2791 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2792 perf_event_task_sched_in(current
);
2793 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2795 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2796 finish_lock_switch(rq
, prev
);
2798 fire_sched_in_preempt_notifiers(current
);
2801 if (unlikely(prev_state
== TASK_DEAD
)) {
2803 * Remove function-return probe instances associated with this
2804 * task and put them back on the free list.
2806 kprobe_flush_task(prev
);
2807 put_task_struct(prev
);
2813 /* assumes rq->lock is held */
2814 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2816 if (prev
->sched_class
->pre_schedule
)
2817 prev
->sched_class
->pre_schedule(rq
, prev
);
2820 /* rq->lock is NOT held, but preemption is disabled */
2821 static inline void post_schedule(struct rq
*rq
)
2823 if (rq
->post_schedule
) {
2824 unsigned long flags
;
2826 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2827 if (rq
->curr
->sched_class
->post_schedule
)
2828 rq
->curr
->sched_class
->post_schedule(rq
);
2829 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2831 rq
->post_schedule
= 0;
2837 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2841 static inline void post_schedule(struct rq
*rq
)
2848 * schedule_tail - first thing a freshly forked thread must call.
2849 * @prev: the thread we just switched away from.
2851 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2852 __releases(rq
->lock
)
2854 struct rq
*rq
= this_rq();
2856 finish_task_switch(rq
, prev
);
2859 * FIXME: do we need to worry about rq being invalidated by the
2864 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2865 /* In this case, finish_task_switch does not reenable preemption */
2868 if (current
->set_child_tid
)
2869 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2873 * context_switch - switch to the new MM and the new
2874 * thread's register state.
2877 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2878 struct task_struct
*next
)
2880 struct mm_struct
*mm
, *oldmm
;
2882 prepare_task_switch(rq
, prev
, next
);
2883 trace_sched_switch(rq
, prev
, next
);
2885 oldmm
= prev
->active_mm
;
2887 * For paravirt, this is coupled with an exit in switch_to to
2888 * combine the page table reload and the switch backend into
2891 arch_start_context_switch(prev
);
2894 next
->active_mm
= oldmm
;
2895 atomic_inc(&oldmm
->mm_count
);
2896 enter_lazy_tlb(oldmm
, next
);
2898 switch_mm(oldmm
, mm
, next
);
2900 if (likely(!prev
->mm
)) {
2901 prev
->active_mm
= NULL
;
2902 rq
->prev_mm
= oldmm
;
2905 * Since the runqueue lock will be released by the next
2906 * task (which is an invalid locking op but in the case
2907 * of the scheduler it's an obvious special-case), so we
2908 * do an early lockdep release here:
2910 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2911 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2914 /* Here we just switch the register state and the stack. */
2915 switch_to(prev
, next
, prev
);
2919 * this_rq must be evaluated again because prev may have moved
2920 * CPUs since it called schedule(), thus the 'rq' on its stack
2921 * frame will be invalid.
2923 finish_task_switch(this_rq(), prev
);
2927 * nr_running, nr_uninterruptible and nr_context_switches:
2929 * externally visible scheduler statistics: current number of runnable
2930 * threads, current number of uninterruptible-sleeping threads, total
2931 * number of context switches performed since bootup.
2933 unsigned long nr_running(void)
2935 unsigned long i
, sum
= 0;
2937 for_each_online_cpu(i
)
2938 sum
+= cpu_rq(i
)->nr_running
;
2943 unsigned long nr_uninterruptible(void)
2945 unsigned long i
, sum
= 0;
2947 for_each_possible_cpu(i
)
2948 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2951 * Since we read the counters lockless, it might be slightly
2952 * inaccurate. Do not allow it to go below zero though:
2954 if (unlikely((long)sum
< 0))
2960 unsigned long long nr_context_switches(void)
2963 unsigned long long sum
= 0;
2965 for_each_possible_cpu(i
)
2966 sum
+= cpu_rq(i
)->nr_switches
;
2971 unsigned long nr_iowait(void)
2973 unsigned long i
, sum
= 0;
2975 for_each_possible_cpu(i
)
2976 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2981 unsigned long nr_iowait_cpu(void)
2983 struct rq
*this = this_rq();
2984 return atomic_read(&this->nr_iowait
);
2987 unsigned long this_cpu_load(void)
2989 struct rq
*this = this_rq();
2990 return this->cpu_load
[0];
2994 /* Variables and functions for calc_load */
2995 static atomic_long_t calc_load_tasks
;
2996 static unsigned long calc_load_update
;
2997 unsigned long avenrun
[3];
2998 EXPORT_SYMBOL(avenrun
);
3001 * get_avenrun - get the load average array
3002 * @loads: pointer to dest load array
3003 * @offset: offset to add
3004 * @shift: shift count to shift the result left
3006 * These values are estimates at best, so no need for locking.
3008 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3010 loads
[0] = (avenrun
[0] + offset
) << shift
;
3011 loads
[1] = (avenrun
[1] + offset
) << shift
;
3012 loads
[2] = (avenrun
[2] + offset
) << shift
;
3015 static unsigned long
3016 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3019 load
+= active
* (FIXED_1
- exp
);
3020 return load
>> FSHIFT
;
3024 * calc_load - update the avenrun load estimates 10 ticks after the
3025 * CPUs have updated calc_load_tasks.
3027 void calc_global_load(void)
3029 unsigned long upd
= calc_load_update
+ 10;
3032 if (time_before(jiffies
, upd
))
3035 active
= atomic_long_read(&calc_load_tasks
);
3036 active
= active
> 0 ? active
* FIXED_1
: 0;
3038 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3039 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3040 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3042 calc_load_update
+= LOAD_FREQ
;
3046 * Either called from update_cpu_load() or from a cpu going idle
3048 static void calc_load_account_active(struct rq
*this_rq
)
3050 long nr_active
, delta
;
3052 nr_active
= this_rq
->nr_running
;
3053 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3055 if (nr_active
!= this_rq
->calc_load_active
) {
3056 delta
= nr_active
- this_rq
->calc_load_active
;
3057 this_rq
->calc_load_active
= nr_active
;
3058 atomic_long_add(delta
, &calc_load_tasks
);
3063 * Update rq->cpu_load[] statistics. This function is usually called every
3064 * scheduler tick (TICK_NSEC).
3066 static void update_cpu_load(struct rq
*this_rq
)
3068 unsigned long this_load
= this_rq
->load
.weight
;
3071 this_rq
->nr_load_updates
++;
3073 /* Update our load: */
3074 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3075 unsigned long old_load
, new_load
;
3077 /* scale is effectively 1 << i now, and >> i divides by scale */
3079 old_load
= this_rq
->cpu_load
[i
];
3080 new_load
= this_load
;
3082 * Round up the averaging division if load is increasing. This
3083 * prevents us from getting stuck on 9 if the load is 10, for
3086 if (new_load
> old_load
)
3087 new_load
+= scale
-1;
3088 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3091 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3092 this_rq
->calc_load_update
+= LOAD_FREQ
;
3093 calc_load_account_active(this_rq
);
3100 * sched_exec - execve() is a valuable balancing opportunity, because at
3101 * this point the task has the smallest effective memory and cache footprint.
3103 void sched_exec(void)
3105 struct task_struct
*p
= current
;
3106 struct migration_req req
;
3107 int dest_cpu
, this_cpu
;
3108 unsigned long flags
;
3112 this_cpu
= get_cpu();
3113 dest_cpu
= select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3114 if (dest_cpu
== this_cpu
) {
3119 rq
= task_rq_lock(p
, &flags
);
3123 * select_task_rq() can race against ->cpus_allowed
3125 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3126 || unlikely(!cpu_active(dest_cpu
))) {
3127 task_rq_unlock(rq
, &flags
);
3131 /* force the process onto the specified CPU */
3132 if (migrate_task(p
, dest_cpu
, &req
)) {
3133 /* Need to wait for migration thread (might exit: take ref). */
3134 struct task_struct
*mt
= rq
->migration_thread
;
3136 get_task_struct(mt
);
3137 task_rq_unlock(rq
, &flags
);
3138 wake_up_process(mt
);
3139 put_task_struct(mt
);
3140 wait_for_completion(&req
.done
);
3144 task_rq_unlock(rq
, &flags
);
3149 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3151 EXPORT_PER_CPU_SYMBOL(kstat
);
3154 * Return any ns on the sched_clock that have not yet been accounted in
3155 * @p in case that task is currently running.
3157 * Called with task_rq_lock() held on @rq.
3159 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3163 if (task_current(rq
, p
)) {
3164 update_rq_clock(rq
);
3165 ns
= rq
->clock
- p
->se
.exec_start
;
3173 unsigned long long task_delta_exec(struct task_struct
*p
)
3175 unsigned long flags
;
3179 rq
= task_rq_lock(p
, &flags
);
3180 ns
= do_task_delta_exec(p
, rq
);
3181 task_rq_unlock(rq
, &flags
);
3187 * Return accounted runtime for the task.
3188 * In case the task is currently running, return the runtime plus current's
3189 * pending runtime that have not been accounted yet.
3191 unsigned long long task_sched_runtime(struct task_struct
*p
)
3193 unsigned long flags
;
3197 rq
= task_rq_lock(p
, &flags
);
3198 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3199 task_rq_unlock(rq
, &flags
);
3205 * Return sum_exec_runtime for the thread group.
3206 * In case the task is currently running, return the sum plus current's
3207 * pending runtime that have not been accounted yet.
3209 * Note that the thread group might have other running tasks as well,
3210 * so the return value not includes other pending runtime that other
3211 * running tasks might have.
3213 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3215 struct task_cputime totals
;
3216 unsigned long flags
;
3220 rq
= task_rq_lock(p
, &flags
);
3221 thread_group_cputime(p
, &totals
);
3222 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3223 task_rq_unlock(rq
, &flags
);
3229 * Account user cpu time to a process.
3230 * @p: the process that the cpu time gets accounted to
3231 * @cputime: the cpu time spent in user space since the last update
3232 * @cputime_scaled: cputime scaled by cpu frequency
3234 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3235 cputime_t cputime_scaled
)
3237 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3240 /* Add user time to process. */
3241 p
->utime
= cputime_add(p
->utime
, cputime
);
3242 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3243 account_group_user_time(p
, cputime
);
3245 /* Add user time to cpustat. */
3246 tmp
= cputime_to_cputime64(cputime
);
3247 if (TASK_NICE(p
) > 0)
3248 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3250 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3252 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3253 /* Account for user time used */
3254 acct_update_integrals(p
);
3258 * Account guest cpu time to a process.
3259 * @p: the process that the cpu time gets accounted to
3260 * @cputime: the cpu time spent in virtual machine since the last update
3261 * @cputime_scaled: cputime scaled by cpu frequency
3263 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3264 cputime_t cputime_scaled
)
3267 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3269 tmp
= cputime_to_cputime64(cputime
);
3271 /* Add guest time to process. */
3272 p
->utime
= cputime_add(p
->utime
, cputime
);
3273 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3274 account_group_user_time(p
, cputime
);
3275 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3277 /* Add guest time to cpustat. */
3278 if (TASK_NICE(p
) > 0) {
3279 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3280 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3282 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3283 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3288 * Account system cpu time to a process.
3289 * @p: the process that the cpu time gets accounted to
3290 * @hardirq_offset: the offset to subtract from hardirq_count()
3291 * @cputime: the cpu time spent in kernel space since the last update
3292 * @cputime_scaled: cputime scaled by cpu frequency
3294 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3295 cputime_t cputime
, cputime_t cputime_scaled
)
3297 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3300 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3301 account_guest_time(p
, cputime
, cputime_scaled
);
3305 /* Add system time to process. */
3306 p
->stime
= cputime_add(p
->stime
, cputime
);
3307 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3308 account_group_system_time(p
, cputime
);
3310 /* Add system time to cpustat. */
3311 tmp
= cputime_to_cputime64(cputime
);
3312 if (hardirq_count() - hardirq_offset
)
3313 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3314 else if (softirq_count())
3315 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3317 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3319 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3321 /* Account for system time used */
3322 acct_update_integrals(p
);
3326 * Account for involuntary wait time.
3327 * @steal: the cpu time spent in involuntary wait
3329 void account_steal_time(cputime_t cputime
)
3331 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3332 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3334 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3338 * Account for idle time.
3339 * @cputime: the cpu time spent in idle wait
3341 void account_idle_time(cputime_t cputime
)
3343 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3344 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3345 struct rq
*rq
= this_rq();
3347 if (atomic_read(&rq
->nr_iowait
) > 0)
3348 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3350 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3353 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3356 * Account a single tick of cpu time.
3357 * @p: the process that the cpu time gets accounted to
3358 * @user_tick: indicates if the tick is a user or a system tick
3360 void account_process_tick(struct task_struct
*p
, int user_tick
)
3362 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3363 struct rq
*rq
= this_rq();
3366 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3367 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3368 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3371 account_idle_time(cputime_one_jiffy
);
3375 * Account multiple ticks of steal time.
3376 * @p: the process from which the cpu time has been stolen
3377 * @ticks: number of stolen ticks
3379 void account_steal_ticks(unsigned long ticks
)
3381 account_steal_time(jiffies_to_cputime(ticks
));
3385 * Account multiple ticks of idle time.
3386 * @ticks: number of stolen ticks
3388 void account_idle_ticks(unsigned long ticks
)
3390 account_idle_time(jiffies_to_cputime(ticks
));
3396 * Use precise platform statistics if available:
3398 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3399 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3405 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3407 struct task_cputime cputime
;
3409 thread_group_cputime(p
, &cputime
);
3411 *ut
= cputime
.utime
;
3412 *st
= cputime
.stime
;
3416 #ifndef nsecs_to_cputime
3417 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3420 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3422 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3425 * Use CFS's precise accounting:
3427 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3432 temp
= (u64
)(rtime
* utime
);
3433 do_div(temp
, total
);
3434 utime
= (cputime_t
)temp
;
3439 * Compare with previous values, to keep monotonicity:
3441 p
->prev_utime
= max(p
->prev_utime
, utime
);
3442 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3444 *ut
= p
->prev_utime
;
3445 *st
= p
->prev_stime
;
3449 * Must be called with siglock held.
3451 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3453 struct signal_struct
*sig
= p
->signal
;
3454 struct task_cputime cputime
;
3455 cputime_t rtime
, utime
, total
;
3457 thread_group_cputime(p
, &cputime
);
3459 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3460 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3465 temp
= (u64
)(rtime
* cputime
.utime
);
3466 do_div(temp
, total
);
3467 utime
= (cputime_t
)temp
;
3471 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3472 sig
->prev_stime
= max(sig
->prev_stime
,
3473 cputime_sub(rtime
, sig
->prev_utime
));
3475 *ut
= sig
->prev_utime
;
3476 *st
= sig
->prev_stime
;
3481 * This function gets called by the timer code, with HZ frequency.
3482 * We call it with interrupts disabled.
3484 * It also gets called by the fork code, when changing the parent's
3487 void scheduler_tick(void)
3489 int cpu
= smp_processor_id();
3490 struct rq
*rq
= cpu_rq(cpu
);
3491 struct task_struct
*curr
= rq
->curr
;
3495 raw_spin_lock(&rq
->lock
);
3496 update_rq_clock(rq
);
3497 update_cpu_load(rq
);
3498 curr
->sched_class
->task_tick(rq
, curr
, 0);
3499 raw_spin_unlock(&rq
->lock
);
3501 perf_event_task_tick(curr
);
3504 rq
->idle_at_tick
= idle_cpu(cpu
);
3505 trigger_load_balance(rq
, cpu
);
3509 notrace
unsigned long get_parent_ip(unsigned long addr
)
3511 if (in_lock_functions(addr
)) {
3512 addr
= CALLER_ADDR2
;
3513 if (in_lock_functions(addr
))
3514 addr
= CALLER_ADDR3
;
3519 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3520 defined(CONFIG_PREEMPT_TRACER))
3522 void __kprobes
add_preempt_count(int val
)
3524 #ifdef CONFIG_DEBUG_PREEMPT
3528 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3531 preempt_count() += val
;
3532 #ifdef CONFIG_DEBUG_PREEMPT
3534 * Spinlock count overflowing soon?
3536 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3539 if (preempt_count() == val
)
3540 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3542 EXPORT_SYMBOL(add_preempt_count
);
3544 void __kprobes
sub_preempt_count(int val
)
3546 #ifdef CONFIG_DEBUG_PREEMPT
3550 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3553 * Is the spinlock portion underflowing?
3555 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3556 !(preempt_count() & PREEMPT_MASK
)))
3560 if (preempt_count() == val
)
3561 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3562 preempt_count() -= val
;
3564 EXPORT_SYMBOL(sub_preempt_count
);
3569 * Print scheduling while atomic bug:
3571 static noinline
void __schedule_bug(struct task_struct
*prev
)
3573 struct pt_regs
*regs
= get_irq_regs();
3575 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3576 prev
->comm
, prev
->pid
, preempt_count());
3578 debug_show_held_locks(prev
);
3580 if (irqs_disabled())
3581 print_irqtrace_events(prev
);
3590 * Various schedule()-time debugging checks and statistics:
3592 static inline void schedule_debug(struct task_struct
*prev
)
3595 * Test if we are atomic. Since do_exit() needs to call into
3596 * schedule() atomically, we ignore that path for now.
3597 * Otherwise, whine if we are scheduling when we should not be.
3599 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3600 __schedule_bug(prev
);
3602 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3604 schedstat_inc(this_rq(), sched_count
);
3605 #ifdef CONFIG_SCHEDSTATS
3606 if (unlikely(prev
->lock_depth
>= 0)) {
3607 schedstat_inc(this_rq(), bkl_count
);
3608 schedstat_inc(prev
, sched_info
.bkl_count
);
3613 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3615 if (prev
->state
== TASK_RUNNING
) {
3616 u64 runtime
= prev
->se
.sum_exec_runtime
;
3618 runtime
-= prev
->se
.prev_sum_exec_runtime
;
3619 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
3622 * In order to avoid avg_overlap growing stale when we are
3623 * indeed overlapping and hence not getting put to sleep, grow
3624 * the avg_overlap on preemption.
3626 * We use the average preemption runtime because that
3627 * correlates to the amount of cache footprint a task can
3630 update_avg(&prev
->se
.avg_overlap
, runtime
);
3632 prev
->sched_class
->put_prev_task(rq
, prev
);
3636 * Pick up the highest-prio task:
3638 static inline struct task_struct
*
3639 pick_next_task(struct rq
*rq
)
3641 const struct sched_class
*class;
3642 struct task_struct
*p
;
3645 * Optimization: we know that if all tasks are in
3646 * the fair class we can call that function directly:
3648 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3649 p
= fair_sched_class
.pick_next_task(rq
);
3654 class = sched_class_highest
;
3656 p
= class->pick_next_task(rq
);
3660 * Will never be NULL as the idle class always
3661 * returns a non-NULL p:
3663 class = class->next
;
3668 * schedule() is the main scheduler function.
3670 asmlinkage
void __sched
schedule(void)
3672 struct task_struct
*prev
, *next
;
3673 unsigned long *switch_count
;
3679 cpu
= smp_processor_id();
3683 switch_count
= &prev
->nivcsw
;
3685 release_kernel_lock(prev
);
3686 need_resched_nonpreemptible
:
3688 schedule_debug(prev
);
3690 if (sched_feat(HRTICK
))
3693 raw_spin_lock_irq(&rq
->lock
);
3694 update_rq_clock(rq
);
3695 clear_tsk_need_resched(prev
);
3697 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3698 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3699 prev
->state
= TASK_RUNNING
;
3701 deactivate_task(rq
, prev
, 1);
3702 switch_count
= &prev
->nvcsw
;
3705 pre_schedule(rq
, prev
);
3707 if (unlikely(!rq
->nr_running
))
3708 idle_balance(cpu
, rq
);
3710 put_prev_task(rq
, prev
);
3711 next
= pick_next_task(rq
);
3713 if (likely(prev
!= next
)) {
3714 sched_info_switch(prev
, next
);
3715 perf_event_task_sched_out(prev
, next
);
3721 context_switch(rq
, prev
, next
); /* unlocks the rq */
3723 * the context switch might have flipped the stack from under
3724 * us, hence refresh the local variables.
3726 cpu
= smp_processor_id();
3729 raw_spin_unlock_irq(&rq
->lock
);
3733 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3735 switch_count
= &prev
->nivcsw
;
3736 goto need_resched_nonpreemptible
;
3739 preempt_enable_no_resched();
3743 EXPORT_SYMBOL(schedule
);
3745 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3747 * Look out! "owner" is an entirely speculative pointer
3748 * access and not reliable.
3750 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3755 if (!sched_feat(OWNER_SPIN
))
3758 #ifdef CONFIG_DEBUG_PAGEALLOC
3760 * Need to access the cpu field knowing that
3761 * DEBUG_PAGEALLOC could have unmapped it if
3762 * the mutex owner just released it and exited.
3764 if (probe_kernel_address(&owner
->cpu
, cpu
))
3771 * Even if the access succeeded (likely case),
3772 * the cpu field may no longer be valid.
3774 if (cpu
>= nr_cpumask_bits
)
3778 * We need to validate that we can do a
3779 * get_cpu() and that we have the percpu area.
3781 if (!cpu_online(cpu
))
3788 * Owner changed, break to re-assess state.
3790 if (lock
->owner
!= owner
)
3794 * Is that owner really running on that cpu?
3796 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3806 #ifdef CONFIG_PREEMPT
3808 * this is the entry point to schedule() from in-kernel preemption
3809 * off of preempt_enable. Kernel preemptions off return from interrupt
3810 * occur there and call schedule directly.
3812 asmlinkage
void __sched
preempt_schedule(void)
3814 struct thread_info
*ti
= current_thread_info();
3817 * If there is a non-zero preempt_count or interrupts are disabled,
3818 * we do not want to preempt the current task. Just return..
3820 if (likely(ti
->preempt_count
|| irqs_disabled()))
3824 add_preempt_count(PREEMPT_ACTIVE
);
3826 sub_preempt_count(PREEMPT_ACTIVE
);
3829 * Check again in case we missed a preemption opportunity
3830 * between schedule and now.
3833 } while (need_resched());
3835 EXPORT_SYMBOL(preempt_schedule
);
3838 * this is the entry point to schedule() from kernel preemption
3839 * off of irq context.
3840 * Note, that this is called and return with irqs disabled. This will
3841 * protect us against recursive calling from irq.
3843 asmlinkage
void __sched
preempt_schedule_irq(void)
3845 struct thread_info
*ti
= current_thread_info();
3847 /* Catch callers which need to be fixed */
3848 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3851 add_preempt_count(PREEMPT_ACTIVE
);
3854 local_irq_disable();
3855 sub_preempt_count(PREEMPT_ACTIVE
);
3858 * Check again in case we missed a preemption opportunity
3859 * between schedule and now.
3862 } while (need_resched());
3865 #endif /* CONFIG_PREEMPT */
3867 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3870 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3872 EXPORT_SYMBOL(default_wake_function
);
3875 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3876 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3877 * number) then we wake all the non-exclusive tasks and one exclusive task.
3879 * There are circumstances in which we can try to wake a task which has already
3880 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3881 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3883 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3884 int nr_exclusive
, int wake_flags
, void *key
)
3886 wait_queue_t
*curr
, *next
;
3888 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3889 unsigned flags
= curr
->flags
;
3891 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3892 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3898 * __wake_up - wake up threads blocked on a waitqueue.
3900 * @mode: which threads
3901 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3902 * @key: is directly passed to the wakeup function
3904 * It may be assumed that this function implies a write memory barrier before
3905 * changing the task state if and only if any tasks are woken up.
3907 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3908 int nr_exclusive
, void *key
)
3910 unsigned long flags
;
3912 spin_lock_irqsave(&q
->lock
, flags
);
3913 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3914 spin_unlock_irqrestore(&q
->lock
, flags
);
3916 EXPORT_SYMBOL(__wake_up
);
3919 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3921 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3923 __wake_up_common(q
, mode
, 1, 0, NULL
);
3926 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3928 __wake_up_common(q
, mode
, 1, 0, key
);
3932 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3934 * @mode: which threads
3935 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3936 * @key: opaque value to be passed to wakeup targets
3938 * The sync wakeup differs that the waker knows that it will schedule
3939 * away soon, so while the target thread will be woken up, it will not
3940 * be migrated to another CPU - ie. the two threads are 'synchronized'
3941 * with each other. This can prevent needless bouncing between CPUs.
3943 * On UP it can prevent extra preemption.
3945 * It may be assumed that this function implies a write memory barrier before
3946 * changing the task state if and only if any tasks are woken up.
3948 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3949 int nr_exclusive
, void *key
)
3951 unsigned long flags
;
3952 int wake_flags
= WF_SYNC
;
3957 if (unlikely(!nr_exclusive
))
3960 spin_lock_irqsave(&q
->lock
, flags
);
3961 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3962 spin_unlock_irqrestore(&q
->lock
, flags
);
3964 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3967 * __wake_up_sync - see __wake_up_sync_key()
3969 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3971 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3973 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3976 * complete: - signals a single thread waiting on this completion
3977 * @x: holds the state of this particular completion
3979 * This will wake up a single thread waiting on this completion. Threads will be
3980 * awakened in the same order in which they were queued.
3982 * See also complete_all(), wait_for_completion() and related routines.
3984 * It may be assumed that this function implies a write memory barrier before
3985 * changing the task state if and only if any tasks are woken up.
3987 void complete(struct completion
*x
)
3989 unsigned long flags
;
3991 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3993 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3994 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3996 EXPORT_SYMBOL(complete
);
3999 * complete_all: - signals all threads waiting on this completion
4000 * @x: holds the state of this particular completion
4002 * This will wake up all threads waiting on this particular completion event.
4004 * It may be assumed that this function implies a write memory barrier before
4005 * changing the task state if and only if any tasks are woken up.
4007 void complete_all(struct completion
*x
)
4009 unsigned long flags
;
4011 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4012 x
->done
+= UINT_MAX
/2;
4013 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4014 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4016 EXPORT_SYMBOL(complete_all
);
4018 static inline long __sched
4019 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4022 DECLARE_WAITQUEUE(wait
, current
);
4024 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4025 __add_wait_queue_tail(&x
->wait
, &wait
);
4027 if (signal_pending_state(state
, current
)) {
4028 timeout
= -ERESTARTSYS
;
4031 __set_current_state(state
);
4032 spin_unlock_irq(&x
->wait
.lock
);
4033 timeout
= schedule_timeout(timeout
);
4034 spin_lock_irq(&x
->wait
.lock
);
4035 } while (!x
->done
&& timeout
);
4036 __remove_wait_queue(&x
->wait
, &wait
);
4041 return timeout
?: 1;
4045 wait_for_common(struct completion
*x
, long timeout
, int state
)
4049 spin_lock_irq(&x
->wait
.lock
);
4050 timeout
= do_wait_for_common(x
, timeout
, state
);
4051 spin_unlock_irq(&x
->wait
.lock
);
4056 * wait_for_completion: - waits for completion of a task
4057 * @x: holds the state of this particular completion
4059 * This waits to be signaled for completion of a specific task. It is NOT
4060 * interruptible and there is no timeout.
4062 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4063 * and interrupt capability. Also see complete().
4065 void __sched
wait_for_completion(struct completion
*x
)
4067 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4069 EXPORT_SYMBOL(wait_for_completion
);
4072 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4073 * @x: holds the state of this particular completion
4074 * @timeout: timeout value in jiffies
4076 * This waits for either a completion of a specific task to be signaled or for a
4077 * specified timeout to expire. The timeout is in jiffies. It is not
4080 unsigned long __sched
4081 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4083 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4085 EXPORT_SYMBOL(wait_for_completion_timeout
);
4088 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4089 * @x: holds the state of this particular completion
4091 * This waits for completion of a specific task to be signaled. It is
4094 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4096 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4097 if (t
== -ERESTARTSYS
)
4101 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4104 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4105 * @x: holds the state of this particular completion
4106 * @timeout: timeout value in jiffies
4108 * This waits for either a completion of a specific task to be signaled or for a
4109 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4111 unsigned long __sched
4112 wait_for_completion_interruptible_timeout(struct completion
*x
,
4113 unsigned long timeout
)
4115 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4117 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4120 * wait_for_completion_killable: - waits for completion of a task (killable)
4121 * @x: holds the state of this particular completion
4123 * This waits to be signaled for completion of a specific task. It can be
4124 * interrupted by a kill signal.
4126 int __sched
wait_for_completion_killable(struct completion
*x
)
4128 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4129 if (t
== -ERESTARTSYS
)
4133 EXPORT_SYMBOL(wait_for_completion_killable
);
4136 * try_wait_for_completion - try to decrement a completion without blocking
4137 * @x: completion structure
4139 * Returns: 0 if a decrement cannot be done without blocking
4140 * 1 if a decrement succeeded.
4142 * If a completion is being used as a counting completion,
4143 * attempt to decrement the counter without blocking. This
4144 * enables us to avoid waiting if the resource the completion
4145 * is protecting is not available.
4147 bool try_wait_for_completion(struct completion
*x
)
4149 unsigned long flags
;
4152 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4157 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4160 EXPORT_SYMBOL(try_wait_for_completion
);
4163 * completion_done - Test to see if a completion has any waiters
4164 * @x: completion structure
4166 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4167 * 1 if there are no waiters.
4170 bool completion_done(struct completion
*x
)
4172 unsigned long flags
;
4175 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4178 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4181 EXPORT_SYMBOL(completion_done
);
4184 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4186 unsigned long flags
;
4189 init_waitqueue_entry(&wait
, current
);
4191 __set_current_state(state
);
4193 spin_lock_irqsave(&q
->lock
, flags
);
4194 __add_wait_queue(q
, &wait
);
4195 spin_unlock(&q
->lock
);
4196 timeout
= schedule_timeout(timeout
);
4197 spin_lock_irq(&q
->lock
);
4198 __remove_wait_queue(q
, &wait
);
4199 spin_unlock_irqrestore(&q
->lock
, flags
);
4204 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4206 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4208 EXPORT_SYMBOL(interruptible_sleep_on
);
4211 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4213 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4215 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4217 void __sched
sleep_on(wait_queue_head_t
*q
)
4219 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4221 EXPORT_SYMBOL(sleep_on
);
4223 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4225 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4227 EXPORT_SYMBOL(sleep_on_timeout
);
4229 #ifdef CONFIG_RT_MUTEXES
4232 * rt_mutex_setprio - set the current priority of a task
4234 * @prio: prio value (kernel-internal form)
4236 * This function changes the 'effective' priority of a task. It does
4237 * not touch ->normal_prio like __setscheduler().
4239 * Used by the rt_mutex code to implement priority inheritance logic.
4241 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4243 unsigned long flags
;
4244 int oldprio
, on_rq
, running
;
4246 const struct sched_class
*prev_class
;
4248 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4250 rq
= task_rq_lock(p
, &flags
);
4251 update_rq_clock(rq
);
4254 prev_class
= p
->sched_class
;
4255 on_rq
= p
->se
.on_rq
;
4256 running
= task_current(rq
, p
);
4258 dequeue_task(rq
, p
, 0);
4260 p
->sched_class
->put_prev_task(rq
, p
);
4263 p
->sched_class
= &rt_sched_class
;
4265 p
->sched_class
= &fair_sched_class
;
4270 p
->sched_class
->set_curr_task(rq
);
4272 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4274 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4276 task_rq_unlock(rq
, &flags
);
4281 void set_user_nice(struct task_struct
*p
, long nice
)
4283 int old_prio
, delta
, on_rq
;
4284 unsigned long flags
;
4287 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4290 * We have to be careful, if called from sys_setpriority(),
4291 * the task might be in the middle of scheduling on another CPU.
4293 rq
= task_rq_lock(p
, &flags
);
4294 update_rq_clock(rq
);
4296 * The RT priorities are set via sched_setscheduler(), but we still
4297 * allow the 'normal' nice value to be set - but as expected
4298 * it wont have any effect on scheduling until the task is
4299 * SCHED_FIFO/SCHED_RR:
4301 if (task_has_rt_policy(p
)) {
4302 p
->static_prio
= NICE_TO_PRIO(nice
);
4305 on_rq
= p
->se
.on_rq
;
4307 dequeue_task(rq
, p
, 0);
4309 p
->static_prio
= NICE_TO_PRIO(nice
);
4312 p
->prio
= effective_prio(p
);
4313 delta
= p
->prio
- old_prio
;
4316 enqueue_task(rq
, p
, 0, false);
4318 * If the task increased its priority or is running and
4319 * lowered its priority, then reschedule its CPU:
4321 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4322 resched_task(rq
->curr
);
4325 task_rq_unlock(rq
, &flags
);
4327 EXPORT_SYMBOL(set_user_nice
);
4330 * can_nice - check if a task can reduce its nice value
4334 int can_nice(const struct task_struct
*p
, const int nice
)
4336 /* convert nice value [19,-20] to rlimit style value [1,40] */
4337 int nice_rlim
= 20 - nice
;
4339 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4340 capable(CAP_SYS_NICE
));
4343 #ifdef __ARCH_WANT_SYS_NICE
4346 * sys_nice - change the priority of the current process.
4347 * @increment: priority increment
4349 * sys_setpriority is a more generic, but much slower function that
4350 * does similar things.
4352 SYSCALL_DEFINE1(nice
, int, increment
)
4357 * Setpriority might change our priority at the same moment.
4358 * We don't have to worry. Conceptually one call occurs first
4359 * and we have a single winner.
4361 if (increment
< -40)
4366 nice
= TASK_NICE(current
) + increment
;
4372 if (increment
< 0 && !can_nice(current
, nice
))
4375 retval
= security_task_setnice(current
, nice
);
4379 set_user_nice(current
, nice
);
4386 * task_prio - return the priority value of a given task.
4387 * @p: the task in question.
4389 * This is the priority value as seen by users in /proc.
4390 * RT tasks are offset by -200. Normal tasks are centered
4391 * around 0, value goes from -16 to +15.
4393 int task_prio(const struct task_struct
*p
)
4395 return p
->prio
- MAX_RT_PRIO
;
4399 * task_nice - return the nice value of a given task.
4400 * @p: the task in question.
4402 int task_nice(const struct task_struct
*p
)
4404 return TASK_NICE(p
);
4406 EXPORT_SYMBOL(task_nice
);
4409 * idle_cpu - is a given cpu idle currently?
4410 * @cpu: the processor in question.
4412 int idle_cpu(int cpu
)
4414 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4418 * idle_task - return the idle task for a given cpu.
4419 * @cpu: the processor in question.
4421 struct task_struct
*idle_task(int cpu
)
4423 return cpu_rq(cpu
)->idle
;
4427 * find_process_by_pid - find a process with a matching PID value.
4428 * @pid: the pid in question.
4430 static struct task_struct
*find_process_by_pid(pid_t pid
)
4432 return pid
? find_task_by_vpid(pid
) : current
;
4435 /* Actually do priority change: must hold rq lock. */
4437 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4439 BUG_ON(p
->se
.on_rq
);
4442 p
->rt_priority
= prio
;
4443 p
->normal_prio
= normal_prio(p
);
4444 /* we are holding p->pi_lock already */
4445 p
->prio
= rt_mutex_getprio(p
);
4446 if (rt_prio(p
->prio
))
4447 p
->sched_class
= &rt_sched_class
;
4449 p
->sched_class
= &fair_sched_class
;
4454 * check the target process has a UID that matches the current process's
4456 static bool check_same_owner(struct task_struct
*p
)
4458 const struct cred
*cred
= current_cred(), *pcred
;
4462 pcred
= __task_cred(p
);
4463 match
= (cred
->euid
== pcred
->euid
||
4464 cred
->euid
== pcred
->uid
);
4469 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4470 struct sched_param
*param
, bool user
)
4472 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4473 unsigned long flags
;
4474 const struct sched_class
*prev_class
;
4478 /* may grab non-irq protected spin_locks */
4479 BUG_ON(in_interrupt());
4481 /* double check policy once rq lock held */
4483 reset_on_fork
= p
->sched_reset_on_fork
;
4484 policy
= oldpolicy
= p
->policy
;
4486 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4487 policy
&= ~SCHED_RESET_ON_FORK
;
4489 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4490 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4491 policy
!= SCHED_IDLE
)
4496 * Valid priorities for SCHED_FIFO and SCHED_RR are
4497 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4498 * SCHED_BATCH and SCHED_IDLE is 0.
4500 if (param
->sched_priority
< 0 ||
4501 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4502 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4504 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4508 * Allow unprivileged RT tasks to decrease priority:
4510 if (user
&& !capable(CAP_SYS_NICE
)) {
4511 if (rt_policy(policy
)) {
4512 unsigned long rlim_rtprio
;
4514 if (!lock_task_sighand(p
, &flags
))
4516 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4517 unlock_task_sighand(p
, &flags
);
4519 /* can't set/change the rt policy */
4520 if (policy
!= p
->policy
&& !rlim_rtprio
)
4523 /* can't increase priority */
4524 if (param
->sched_priority
> p
->rt_priority
&&
4525 param
->sched_priority
> rlim_rtprio
)
4529 * Like positive nice levels, dont allow tasks to
4530 * move out of SCHED_IDLE either:
4532 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4535 /* can't change other user's priorities */
4536 if (!check_same_owner(p
))
4539 /* Normal users shall not reset the sched_reset_on_fork flag */
4540 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4545 #ifdef CONFIG_RT_GROUP_SCHED
4547 * Do not allow realtime tasks into groups that have no runtime
4550 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4551 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4555 retval
= security_task_setscheduler(p
, policy
, param
);
4561 * make sure no PI-waiters arrive (or leave) while we are
4562 * changing the priority of the task:
4564 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4566 * To be able to change p->policy safely, the apropriate
4567 * runqueue lock must be held.
4569 rq
= __task_rq_lock(p
);
4570 /* recheck policy now with rq lock held */
4571 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4572 policy
= oldpolicy
= -1;
4573 __task_rq_unlock(rq
);
4574 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4577 update_rq_clock(rq
);
4578 on_rq
= p
->se
.on_rq
;
4579 running
= task_current(rq
, p
);
4581 deactivate_task(rq
, p
, 0);
4583 p
->sched_class
->put_prev_task(rq
, p
);
4585 p
->sched_reset_on_fork
= reset_on_fork
;
4588 prev_class
= p
->sched_class
;
4589 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4592 p
->sched_class
->set_curr_task(rq
);
4594 activate_task(rq
, p
, 0);
4596 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4598 __task_rq_unlock(rq
);
4599 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4601 rt_mutex_adjust_pi(p
);
4607 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4608 * @p: the task in question.
4609 * @policy: new policy.
4610 * @param: structure containing the new RT priority.
4612 * NOTE that the task may be already dead.
4614 int sched_setscheduler(struct task_struct
*p
, int policy
,
4615 struct sched_param
*param
)
4617 return __sched_setscheduler(p
, policy
, param
, true);
4619 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4622 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4623 * @p: the task in question.
4624 * @policy: new policy.
4625 * @param: structure containing the new RT priority.
4627 * Just like sched_setscheduler, only don't bother checking if the
4628 * current context has permission. For example, this is needed in
4629 * stop_machine(): we create temporary high priority worker threads,
4630 * but our caller might not have that capability.
4632 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4633 struct sched_param
*param
)
4635 return __sched_setscheduler(p
, policy
, param
, false);
4639 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4641 struct sched_param lparam
;
4642 struct task_struct
*p
;
4645 if (!param
|| pid
< 0)
4647 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4652 p
= find_process_by_pid(pid
);
4654 retval
= sched_setscheduler(p
, policy
, &lparam
);
4661 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4662 * @pid: the pid in question.
4663 * @policy: new policy.
4664 * @param: structure containing the new RT priority.
4666 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4667 struct sched_param __user
*, param
)
4669 /* negative values for policy are not valid */
4673 return do_sched_setscheduler(pid
, policy
, param
);
4677 * sys_sched_setparam - set/change the RT priority of a thread
4678 * @pid: the pid in question.
4679 * @param: structure containing the new RT priority.
4681 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4683 return do_sched_setscheduler(pid
, -1, param
);
4687 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4688 * @pid: the pid in question.
4690 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4692 struct task_struct
*p
;
4700 p
= find_process_by_pid(pid
);
4702 retval
= security_task_getscheduler(p
);
4705 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4712 * sys_sched_getparam - get the RT priority of a thread
4713 * @pid: the pid in question.
4714 * @param: structure containing the RT priority.
4716 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4718 struct sched_param lp
;
4719 struct task_struct
*p
;
4722 if (!param
|| pid
< 0)
4726 p
= find_process_by_pid(pid
);
4731 retval
= security_task_getscheduler(p
);
4735 lp
.sched_priority
= p
->rt_priority
;
4739 * This one might sleep, we cannot do it with a spinlock held ...
4741 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4750 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4752 cpumask_var_t cpus_allowed
, new_mask
;
4753 struct task_struct
*p
;
4759 p
= find_process_by_pid(pid
);
4766 /* Prevent p going away */
4770 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4774 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4776 goto out_free_cpus_allowed
;
4779 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4782 retval
= security_task_setscheduler(p
, 0, NULL
);
4786 cpuset_cpus_allowed(p
, cpus_allowed
);
4787 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4789 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4792 cpuset_cpus_allowed(p
, cpus_allowed
);
4793 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4795 * We must have raced with a concurrent cpuset
4796 * update. Just reset the cpus_allowed to the
4797 * cpuset's cpus_allowed
4799 cpumask_copy(new_mask
, cpus_allowed
);
4804 free_cpumask_var(new_mask
);
4805 out_free_cpus_allowed
:
4806 free_cpumask_var(cpus_allowed
);
4813 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4814 struct cpumask
*new_mask
)
4816 if (len
< cpumask_size())
4817 cpumask_clear(new_mask
);
4818 else if (len
> cpumask_size())
4819 len
= cpumask_size();
4821 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4825 * sys_sched_setaffinity - set the cpu affinity of a process
4826 * @pid: pid of the process
4827 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4828 * @user_mask_ptr: user-space pointer to the new cpu mask
4830 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4831 unsigned long __user
*, user_mask_ptr
)
4833 cpumask_var_t new_mask
;
4836 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4839 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4841 retval
= sched_setaffinity(pid
, new_mask
);
4842 free_cpumask_var(new_mask
);
4846 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4848 struct task_struct
*p
;
4849 unsigned long flags
;
4857 p
= find_process_by_pid(pid
);
4861 retval
= security_task_getscheduler(p
);
4865 rq
= task_rq_lock(p
, &flags
);
4866 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4867 task_rq_unlock(rq
, &flags
);
4877 * sys_sched_getaffinity - get the cpu affinity of a process
4878 * @pid: pid of the process
4879 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4880 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4882 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4883 unsigned long __user
*, user_mask_ptr
)
4888 if (len
< cpumask_size())
4891 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4894 ret
= sched_getaffinity(pid
, mask
);
4896 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
4899 ret
= cpumask_size();
4901 free_cpumask_var(mask
);
4907 * sys_sched_yield - yield the current processor to other threads.
4909 * This function yields the current CPU to other tasks. If there are no
4910 * other threads running on this CPU then this function will return.
4912 SYSCALL_DEFINE0(sched_yield
)
4914 struct rq
*rq
= this_rq_lock();
4916 schedstat_inc(rq
, yld_count
);
4917 current
->sched_class
->yield_task(rq
);
4920 * Since we are going to call schedule() anyway, there's
4921 * no need to preempt or enable interrupts:
4923 __release(rq
->lock
);
4924 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4925 do_raw_spin_unlock(&rq
->lock
);
4926 preempt_enable_no_resched();
4933 static inline int should_resched(void)
4935 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4938 static void __cond_resched(void)
4940 add_preempt_count(PREEMPT_ACTIVE
);
4942 sub_preempt_count(PREEMPT_ACTIVE
);
4945 int __sched
_cond_resched(void)
4947 if (should_resched()) {
4953 EXPORT_SYMBOL(_cond_resched
);
4956 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4957 * call schedule, and on return reacquire the lock.
4959 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4960 * operations here to prevent schedule() from being called twice (once via
4961 * spin_unlock(), once by hand).
4963 int __cond_resched_lock(spinlock_t
*lock
)
4965 int resched
= should_resched();
4968 lockdep_assert_held(lock
);
4970 if (spin_needbreak(lock
) || resched
) {
4981 EXPORT_SYMBOL(__cond_resched_lock
);
4983 int __sched
__cond_resched_softirq(void)
4985 BUG_ON(!in_softirq());
4987 if (should_resched()) {
4995 EXPORT_SYMBOL(__cond_resched_softirq
);
4998 * yield - yield the current processor to other threads.
5000 * This is a shortcut for kernel-space yielding - it marks the
5001 * thread runnable and calls sys_sched_yield().
5003 void __sched
yield(void)
5005 set_current_state(TASK_RUNNING
);
5008 EXPORT_SYMBOL(yield
);
5011 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5012 * that process accounting knows that this is a task in IO wait state.
5014 void __sched
io_schedule(void)
5016 struct rq
*rq
= raw_rq();
5018 delayacct_blkio_start();
5019 atomic_inc(&rq
->nr_iowait
);
5020 current
->in_iowait
= 1;
5022 current
->in_iowait
= 0;
5023 atomic_dec(&rq
->nr_iowait
);
5024 delayacct_blkio_end();
5026 EXPORT_SYMBOL(io_schedule
);
5028 long __sched
io_schedule_timeout(long timeout
)
5030 struct rq
*rq
= raw_rq();
5033 delayacct_blkio_start();
5034 atomic_inc(&rq
->nr_iowait
);
5035 current
->in_iowait
= 1;
5036 ret
= schedule_timeout(timeout
);
5037 current
->in_iowait
= 0;
5038 atomic_dec(&rq
->nr_iowait
);
5039 delayacct_blkio_end();
5044 * sys_sched_get_priority_max - return maximum RT priority.
5045 * @policy: scheduling class.
5047 * this syscall returns the maximum rt_priority that can be used
5048 * by a given scheduling class.
5050 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5057 ret
= MAX_USER_RT_PRIO
-1;
5069 * sys_sched_get_priority_min - return minimum RT priority.
5070 * @policy: scheduling class.
5072 * this syscall returns the minimum rt_priority that can be used
5073 * by a given scheduling class.
5075 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5093 * sys_sched_rr_get_interval - return the default timeslice of a process.
5094 * @pid: pid of the process.
5095 * @interval: userspace pointer to the timeslice value.
5097 * this syscall writes the default timeslice value of a given process
5098 * into the user-space timespec buffer. A value of '0' means infinity.
5100 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5101 struct timespec __user
*, interval
)
5103 struct task_struct
*p
;
5104 unsigned int time_slice
;
5105 unsigned long flags
;
5115 p
= find_process_by_pid(pid
);
5119 retval
= security_task_getscheduler(p
);
5123 rq
= task_rq_lock(p
, &flags
);
5124 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5125 task_rq_unlock(rq
, &flags
);
5128 jiffies_to_timespec(time_slice
, &t
);
5129 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5137 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5139 void sched_show_task(struct task_struct
*p
)
5141 unsigned long free
= 0;
5144 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5145 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5146 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5147 #if BITS_PER_LONG == 32
5148 if (state
== TASK_RUNNING
)
5149 printk(KERN_CONT
" running ");
5151 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5153 if (state
== TASK_RUNNING
)
5154 printk(KERN_CONT
" running task ");
5156 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5158 #ifdef CONFIG_DEBUG_STACK_USAGE
5159 free
= stack_not_used(p
);
5161 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5162 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5163 (unsigned long)task_thread_info(p
)->flags
);
5165 show_stack(p
, NULL
);
5168 void show_state_filter(unsigned long state_filter
)
5170 struct task_struct
*g
, *p
;
5172 #if BITS_PER_LONG == 32
5174 " task PC stack pid father\n");
5177 " task PC stack pid father\n");
5179 read_lock(&tasklist_lock
);
5180 do_each_thread(g
, p
) {
5182 * reset the NMI-timeout, listing all files on a slow
5183 * console might take alot of time:
5185 touch_nmi_watchdog();
5186 if (!state_filter
|| (p
->state
& state_filter
))
5188 } while_each_thread(g
, p
);
5190 touch_all_softlockup_watchdogs();
5192 #ifdef CONFIG_SCHED_DEBUG
5193 sysrq_sched_debug_show();
5195 read_unlock(&tasklist_lock
);
5197 * Only show locks if all tasks are dumped:
5200 debug_show_all_locks();
5203 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5205 idle
->sched_class
= &idle_sched_class
;
5209 * init_idle - set up an idle thread for a given CPU
5210 * @idle: task in question
5211 * @cpu: cpu the idle task belongs to
5213 * NOTE: this function does not set the idle thread's NEED_RESCHED
5214 * flag, to make booting more robust.
5216 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5218 struct rq
*rq
= cpu_rq(cpu
);
5219 unsigned long flags
;
5221 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5224 idle
->state
= TASK_RUNNING
;
5225 idle
->se
.exec_start
= sched_clock();
5227 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5228 __set_task_cpu(idle
, cpu
);
5230 rq
->curr
= rq
->idle
= idle
;
5231 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5234 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5236 /* Set the preempt count _outside_ the spinlocks! */
5237 #if defined(CONFIG_PREEMPT)
5238 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5240 task_thread_info(idle
)->preempt_count
= 0;
5243 * The idle tasks have their own, simple scheduling class:
5245 idle
->sched_class
= &idle_sched_class
;
5246 ftrace_graph_init_task(idle
);
5250 * In a system that switches off the HZ timer nohz_cpu_mask
5251 * indicates which cpus entered this state. This is used
5252 * in the rcu update to wait only for active cpus. For system
5253 * which do not switch off the HZ timer nohz_cpu_mask should
5254 * always be CPU_BITS_NONE.
5256 cpumask_var_t nohz_cpu_mask
;
5259 * Increase the granularity value when there are more CPUs,
5260 * because with more CPUs the 'effective latency' as visible
5261 * to users decreases. But the relationship is not linear,
5262 * so pick a second-best guess by going with the log2 of the
5265 * This idea comes from the SD scheduler of Con Kolivas:
5267 static int get_update_sysctl_factor(void)
5269 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5270 unsigned int factor
;
5272 switch (sysctl_sched_tunable_scaling
) {
5273 case SCHED_TUNABLESCALING_NONE
:
5276 case SCHED_TUNABLESCALING_LINEAR
:
5279 case SCHED_TUNABLESCALING_LOG
:
5281 factor
= 1 + ilog2(cpus
);
5288 static void update_sysctl(void)
5290 unsigned int factor
= get_update_sysctl_factor();
5292 #define SET_SYSCTL(name) \
5293 (sysctl_##name = (factor) * normalized_sysctl_##name)
5294 SET_SYSCTL(sched_min_granularity
);
5295 SET_SYSCTL(sched_latency
);
5296 SET_SYSCTL(sched_wakeup_granularity
);
5297 SET_SYSCTL(sched_shares_ratelimit
);
5301 static inline void sched_init_granularity(void)
5308 * This is how migration works:
5310 * 1) we queue a struct migration_req structure in the source CPU's
5311 * runqueue and wake up that CPU's migration thread.
5312 * 2) we down() the locked semaphore => thread blocks.
5313 * 3) migration thread wakes up (implicitly it forces the migrated
5314 * thread off the CPU)
5315 * 4) it gets the migration request and checks whether the migrated
5316 * task is still in the wrong runqueue.
5317 * 5) if it's in the wrong runqueue then the migration thread removes
5318 * it and puts it into the right queue.
5319 * 6) migration thread up()s the semaphore.
5320 * 7) we wake up and the migration is done.
5324 * Change a given task's CPU affinity. Migrate the thread to a
5325 * proper CPU and schedule it away if the CPU it's executing on
5326 * is removed from the allowed bitmask.
5328 * NOTE: the caller must have a valid reference to the task, the
5329 * task must not exit() & deallocate itself prematurely. The
5330 * call is not atomic; no spinlocks may be held.
5332 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5334 struct migration_req req
;
5335 unsigned long flags
;
5339 rq
= task_rq_lock(p
, &flags
);
5341 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5346 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5347 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5352 if (p
->sched_class
->set_cpus_allowed
)
5353 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5355 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5356 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5359 /* Can the task run on the task's current CPU? If so, we're done */
5360 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5363 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5364 /* Need help from migration thread: drop lock and wait. */
5365 struct task_struct
*mt
= rq
->migration_thread
;
5367 get_task_struct(mt
);
5368 task_rq_unlock(rq
, &flags
);
5369 wake_up_process(rq
->migration_thread
);
5370 put_task_struct(mt
);
5371 wait_for_completion(&req
.done
);
5372 tlb_migrate_finish(p
->mm
);
5376 task_rq_unlock(rq
, &flags
);
5380 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5383 * Move (not current) task off this cpu, onto dest cpu. We're doing
5384 * this because either it can't run here any more (set_cpus_allowed()
5385 * away from this CPU, or CPU going down), or because we're
5386 * attempting to rebalance this task on exec (sched_exec).
5388 * So we race with normal scheduler movements, but that's OK, as long
5389 * as the task is no longer on this CPU.
5391 * Returns non-zero if task was successfully migrated.
5393 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5395 struct rq
*rq_dest
, *rq_src
;
5398 if (unlikely(!cpu_active(dest_cpu
)))
5401 rq_src
= cpu_rq(src_cpu
);
5402 rq_dest
= cpu_rq(dest_cpu
);
5404 double_rq_lock(rq_src
, rq_dest
);
5405 /* Already moved. */
5406 if (task_cpu(p
) != src_cpu
)
5408 /* Affinity changed (again). */
5409 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5413 * If we're not on a rq, the next wake-up will ensure we're
5417 deactivate_task(rq_src
, p
, 0);
5418 set_task_cpu(p
, dest_cpu
);
5419 activate_task(rq_dest
, p
, 0);
5420 check_preempt_curr(rq_dest
, p
, 0);
5425 double_rq_unlock(rq_src
, rq_dest
);
5429 #define RCU_MIGRATION_IDLE 0
5430 #define RCU_MIGRATION_NEED_QS 1
5431 #define RCU_MIGRATION_GOT_QS 2
5432 #define RCU_MIGRATION_MUST_SYNC 3
5435 * migration_thread - this is a highprio system thread that performs
5436 * thread migration by bumping thread off CPU then 'pushing' onto
5439 static int migration_thread(void *data
)
5442 int cpu
= (long)data
;
5446 BUG_ON(rq
->migration_thread
!= current
);
5448 set_current_state(TASK_INTERRUPTIBLE
);
5449 while (!kthread_should_stop()) {
5450 struct migration_req
*req
;
5451 struct list_head
*head
;
5453 raw_spin_lock_irq(&rq
->lock
);
5455 if (cpu_is_offline(cpu
)) {
5456 raw_spin_unlock_irq(&rq
->lock
);
5460 if (rq
->active_balance
) {
5461 active_load_balance(rq
, cpu
);
5462 rq
->active_balance
= 0;
5465 head
= &rq
->migration_queue
;
5467 if (list_empty(head
)) {
5468 raw_spin_unlock_irq(&rq
->lock
);
5470 set_current_state(TASK_INTERRUPTIBLE
);
5473 req
= list_entry(head
->next
, struct migration_req
, list
);
5474 list_del_init(head
->next
);
5476 if (req
->task
!= NULL
) {
5477 raw_spin_unlock(&rq
->lock
);
5478 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5479 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5480 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5481 raw_spin_unlock(&rq
->lock
);
5483 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5484 raw_spin_unlock(&rq
->lock
);
5485 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5489 complete(&req
->done
);
5491 __set_current_state(TASK_RUNNING
);
5496 #ifdef CONFIG_HOTPLUG_CPU
5498 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5502 local_irq_disable();
5503 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5509 * Figure out where task on dead CPU should go, use force if necessary.
5511 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5516 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5518 /* It can have affinity changed while we were choosing. */
5519 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
5524 * While a dead CPU has no uninterruptible tasks queued at this point,
5525 * it might still have a nonzero ->nr_uninterruptible counter, because
5526 * for performance reasons the counter is not stricly tracking tasks to
5527 * their home CPUs. So we just add the counter to another CPU's counter,
5528 * to keep the global sum constant after CPU-down:
5530 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5532 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5533 unsigned long flags
;
5535 local_irq_save(flags
);
5536 double_rq_lock(rq_src
, rq_dest
);
5537 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5538 rq_src
->nr_uninterruptible
= 0;
5539 double_rq_unlock(rq_src
, rq_dest
);
5540 local_irq_restore(flags
);
5543 /* Run through task list and migrate tasks from the dead cpu. */
5544 static void migrate_live_tasks(int src_cpu
)
5546 struct task_struct
*p
, *t
;
5548 read_lock(&tasklist_lock
);
5550 do_each_thread(t
, p
) {
5554 if (task_cpu(p
) == src_cpu
)
5555 move_task_off_dead_cpu(src_cpu
, p
);
5556 } while_each_thread(t
, p
);
5558 read_unlock(&tasklist_lock
);
5562 * Schedules idle task to be the next runnable task on current CPU.
5563 * It does so by boosting its priority to highest possible.
5564 * Used by CPU offline code.
5566 void sched_idle_next(void)
5568 int this_cpu
= smp_processor_id();
5569 struct rq
*rq
= cpu_rq(this_cpu
);
5570 struct task_struct
*p
= rq
->idle
;
5571 unsigned long flags
;
5573 /* cpu has to be offline */
5574 BUG_ON(cpu_online(this_cpu
));
5577 * Strictly not necessary since rest of the CPUs are stopped by now
5578 * and interrupts disabled on the current cpu.
5580 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5582 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5584 update_rq_clock(rq
);
5585 activate_task(rq
, p
, 0);
5587 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5591 * Ensures that the idle task is using init_mm right before its cpu goes
5594 void idle_task_exit(void)
5596 struct mm_struct
*mm
= current
->active_mm
;
5598 BUG_ON(cpu_online(smp_processor_id()));
5601 switch_mm(mm
, &init_mm
, current
);
5605 /* called under rq->lock with disabled interrupts */
5606 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5608 struct rq
*rq
= cpu_rq(dead_cpu
);
5610 /* Must be exiting, otherwise would be on tasklist. */
5611 BUG_ON(!p
->exit_state
);
5613 /* Cannot have done final schedule yet: would have vanished. */
5614 BUG_ON(p
->state
== TASK_DEAD
);
5619 * Drop lock around migration; if someone else moves it,
5620 * that's OK. No task can be added to this CPU, so iteration is
5623 raw_spin_unlock_irq(&rq
->lock
);
5624 move_task_off_dead_cpu(dead_cpu
, p
);
5625 raw_spin_lock_irq(&rq
->lock
);
5630 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5631 static void migrate_dead_tasks(unsigned int dead_cpu
)
5633 struct rq
*rq
= cpu_rq(dead_cpu
);
5634 struct task_struct
*next
;
5637 if (!rq
->nr_running
)
5639 update_rq_clock(rq
);
5640 next
= pick_next_task(rq
);
5643 next
->sched_class
->put_prev_task(rq
, next
);
5644 migrate_dead(dead_cpu
, next
);
5650 * remove the tasks which were accounted by rq from calc_load_tasks.
5652 static void calc_global_load_remove(struct rq
*rq
)
5654 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5655 rq
->calc_load_active
= 0;
5657 #endif /* CONFIG_HOTPLUG_CPU */
5659 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5661 static struct ctl_table sd_ctl_dir
[] = {
5663 .procname
= "sched_domain",
5669 static struct ctl_table sd_ctl_root
[] = {
5671 .procname
= "kernel",
5673 .child
= sd_ctl_dir
,
5678 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5680 struct ctl_table
*entry
=
5681 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5686 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5688 struct ctl_table
*entry
;
5691 * In the intermediate directories, both the child directory and
5692 * procname are dynamically allocated and could fail but the mode
5693 * will always be set. In the lowest directory the names are
5694 * static strings and all have proc handlers.
5696 for (entry
= *tablep
; entry
->mode
; entry
++) {
5698 sd_free_ctl_entry(&entry
->child
);
5699 if (entry
->proc_handler
== NULL
)
5700 kfree(entry
->procname
);
5708 set_table_entry(struct ctl_table
*entry
,
5709 const char *procname
, void *data
, int maxlen
,
5710 mode_t mode
, proc_handler
*proc_handler
)
5712 entry
->procname
= procname
;
5714 entry
->maxlen
= maxlen
;
5716 entry
->proc_handler
= proc_handler
;
5719 static struct ctl_table
*
5720 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5722 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5727 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5728 sizeof(long), 0644, proc_doulongvec_minmax
);
5729 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5730 sizeof(long), 0644, proc_doulongvec_minmax
);
5731 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5732 sizeof(int), 0644, proc_dointvec_minmax
);
5733 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5734 sizeof(int), 0644, proc_dointvec_minmax
);
5735 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5736 sizeof(int), 0644, proc_dointvec_minmax
);
5737 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5738 sizeof(int), 0644, proc_dointvec_minmax
);
5739 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5740 sizeof(int), 0644, proc_dointvec_minmax
);
5741 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5742 sizeof(int), 0644, proc_dointvec_minmax
);
5743 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5744 sizeof(int), 0644, proc_dointvec_minmax
);
5745 set_table_entry(&table
[9], "cache_nice_tries",
5746 &sd
->cache_nice_tries
,
5747 sizeof(int), 0644, proc_dointvec_minmax
);
5748 set_table_entry(&table
[10], "flags", &sd
->flags
,
5749 sizeof(int), 0644, proc_dointvec_minmax
);
5750 set_table_entry(&table
[11], "name", sd
->name
,
5751 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5752 /* &table[12] is terminator */
5757 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5759 struct ctl_table
*entry
, *table
;
5760 struct sched_domain
*sd
;
5761 int domain_num
= 0, i
;
5764 for_each_domain(cpu
, sd
)
5766 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5771 for_each_domain(cpu
, sd
) {
5772 snprintf(buf
, 32, "domain%d", i
);
5773 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5775 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5782 static struct ctl_table_header
*sd_sysctl_header
;
5783 static void register_sched_domain_sysctl(void)
5785 int i
, cpu_num
= num_possible_cpus();
5786 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5789 WARN_ON(sd_ctl_dir
[0].child
);
5790 sd_ctl_dir
[0].child
= entry
;
5795 for_each_possible_cpu(i
) {
5796 snprintf(buf
, 32, "cpu%d", i
);
5797 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5799 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5803 WARN_ON(sd_sysctl_header
);
5804 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5807 /* may be called multiple times per register */
5808 static void unregister_sched_domain_sysctl(void)
5810 if (sd_sysctl_header
)
5811 unregister_sysctl_table(sd_sysctl_header
);
5812 sd_sysctl_header
= NULL
;
5813 if (sd_ctl_dir
[0].child
)
5814 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5817 static void register_sched_domain_sysctl(void)
5820 static void unregister_sched_domain_sysctl(void)
5825 static void set_rq_online(struct rq
*rq
)
5828 const struct sched_class
*class;
5830 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5833 for_each_class(class) {
5834 if (class->rq_online
)
5835 class->rq_online(rq
);
5840 static void set_rq_offline(struct rq
*rq
)
5843 const struct sched_class
*class;
5845 for_each_class(class) {
5846 if (class->rq_offline
)
5847 class->rq_offline(rq
);
5850 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5856 * migration_call - callback that gets triggered when a CPU is added.
5857 * Here we can start up the necessary migration thread for the new CPU.
5859 static int __cpuinit
5860 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5862 struct task_struct
*p
;
5863 int cpu
= (long)hcpu
;
5864 unsigned long flags
;
5869 case CPU_UP_PREPARE
:
5870 case CPU_UP_PREPARE_FROZEN
:
5871 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5874 kthread_bind(p
, cpu
);
5875 /* Must be high prio: stop_machine expects to yield to it. */
5876 rq
= task_rq_lock(p
, &flags
);
5877 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5878 task_rq_unlock(rq
, &flags
);
5880 cpu_rq(cpu
)->migration_thread
= p
;
5881 rq
->calc_load_update
= calc_load_update
;
5885 case CPU_ONLINE_FROZEN
:
5886 /* Strictly unnecessary, as first user will wake it. */
5887 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5889 /* Update our root-domain */
5891 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5893 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5897 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5900 #ifdef CONFIG_HOTPLUG_CPU
5901 case CPU_UP_CANCELED
:
5902 case CPU_UP_CANCELED_FROZEN
:
5903 if (!cpu_rq(cpu
)->migration_thread
)
5905 /* Unbind it from offline cpu so it can run. Fall thru. */
5906 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5907 cpumask_any(cpu_online_mask
));
5908 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5909 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5910 cpu_rq(cpu
)->migration_thread
= NULL
;
5914 case CPU_DEAD_FROZEN
:
5915 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5916 migrate_live_tasks(cpu
);
5918 kthread_stop(rq
->migration_thread
);
5919 put_task_struct(rq
->migration_thread
);
5920 rq
->migration_thread
= NULL
;
5921 /* Idle task back to normal (off runqueue, low prio) */
5922 raw_spin_lock_irq(&rq
->lock
);
5923 update_rq_clock(rq
);
5924 deactivate_task(rq
, rq
->idle
, 0);
5925 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5926 rq
->idle
->sched_class
= &idle_sched_class
;
5927 migrate_dead_tasks(cpu
);
5928 raw_spin_unlock_irq(&rq
->lock
);
5930 migrate_nr_uninterruptible(rq
);
5931 BUG_ON(rq
->nr_running
!= 0);
5932 calc_global_load_remove(rq
);
5934 * No need to migrate the tasks: it was best-effort if
5935 * they didn't take sched_hotcpu_mutex. Just wake up
5938 raw_spin_lock_irq(&rq
->lock
);
5939 while (!list_empty(&rq
->migration_queue
)) {
5940 struct migration_req
*req
;
5942 req
= list_entry(rq
->migration_queue
.next
,
5943 struct migration_req
, list
);
5944 list_del_init(&req
->list
);
5945 raw_spin_unlock_irq(&rq
->lock
);
5946 complete(&req
->done
);
5947 raw_spin_lock_irq(&rq
->lock
);
5949 raw_spin_unlock_irq(&rq
->lock
);
5953 case CPU_DYING_FROZEN
:
5954 /* Update our root-domain */
5956 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5958 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5961 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5969 * Register at high priority so that task migration (migrate_all_tasks)
5970 * happens before everything else. This has to be lower priority than
5971 * the notifier in the perf_event subsystem, though.
5973 static struct notifier_block __cpuinitdata migration_notifier
= {
5974 .notifier_call
= migration_call
,
5978 static int __init
migration_init(void)
5980 void *cpu
= (void *)(long)smp_processor_id();
5983 /* Start one for the boot CPU: */
5984 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5985 BUG_ON(err
== NOTIFY_BAD
);
5986 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5987 register_cpu_notifier(&migration_notifier
);
5991 early_initcall(migration_init
);
5996 #ifdef CONFIG_SCHED_DEBUG
5998 static __read_mostly
int sched_domain_debug_enabled
;
6000 static int __init
sched_domain_debug_setup(char *str
)
6002 sched_domain_debug_enabled
= 1;
6006 early_param("sched_debug", sched_domain_debug_setup
);
6008 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6009 struct cpumask
*groupmask
)
6011 struct sched_group
*group
= sd
->groups
;
6014 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6015 cpumask_clear(groupmask
);
6017 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6019 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6020 printk("does not load-balance\n");
6022 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6027 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6029 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6030 printk(KERN_ERR
"ERROR: domain->span does not contain "
6033 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6034 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6038 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6042 printk(KERN_ERR
"ERROR: group is NULL\n");
6046 if (!group
->cpu_power
) {
6047 printk(KERN_CONT
"\n");
6048 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6053 if (!cpumask_weight(sched_group_cpus(group
))) {
6054 printk(KERN_CONT
"\n");
6055 printk(KERN_ERR
"ERROR: empty group\n");
6059 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6060 printk(KERN_CONT
"\n");
6061 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6065 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6067 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6069 printk(KERN_CONT
" %s", str
);
6070 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6071 printk(KERN_CONT
" (cpu_power = %d)",
6075 group
= group
->next
;
6076 } while (group
!= sd
->groups
);
6077 printk(KERN_CONT
"\n");
6079 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6080 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6083 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6084 printk(KERN_ERR
"ERROR: parent span is not a superset "
6085 "of domain->span\n");
6089 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6091 cpumask_var_t groupmask
;
6094 if (!sched_domain_debug_enabled
)
6098 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6102 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6104 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6105 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6110 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6117 free_cpumask_var(groupmask
);
6119 #else /* !CONFIG_SCHED_DEBUG */
6120 # define sched_domain_debug(sd, cpu) do { } while (0)
6121 #endif /* CONFIG_SCHED_DEBUG */
6123 static int sd_degenerate(struct sched_domain
*sd
)
6125 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6128 /* Following flags need at least 2 groups */
6129 if (sd
->flags
& (SD_LOAD_BALANCE
|
6130 SD_BALANCE_NEWIDLE
|
6134 SD_SHARE_PKG_RESOURCES
)) {
6135 if (sd
->groups
!= sd
->groups
->next
)
6139 /* Following flags don't use groups */
6140 if (sd
->flags
& (SD_WAKE_AFFINE
))
6147 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6149 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6151 if (sd_degenerate(parent
))
6154 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6157 /* Flags needing groups don't count if only 1 group in parent */
6158 if (parent
->groups
== parent
->groups
->next
) {
6159 pflags
&= ~(SD_LOAD_BALANCE
|
6160 SD_BALANCE_NEWIDLE
|
6164 SD_SHARE_PKG_RESOURCES
);
6165 if (nr_node_ids
== 1)
6166 pflags
&= ~SD_SERIALIZE
;
6168 if (~cflags
& pflags
)
6174 static void free_rootdomain(struct root_domain
*rd
)
6176 synchronize_sched();
6178 cpupri_cleanup(&rd
->cpupri
);
6180 free_cpumask_var(rd
->rto_mask
);
6181 free_cpumask_var(rd
->online
);
6182 free_cpumask_var(rd
->span
);
6186 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6188 struct root_domain
*old_rd
= NULL
;
6189 unsigned long flags
;
6191 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6196 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6199 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6202 * If we dont want to free the old_rt yet then
6203 * set old_rd to NULL to skip the freeing later
6206 if (!atomic_dec_and_test(&old_rd
->refcount
))
6210 atomic_inc(&rd
->refcount
);
6213 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6214 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6217 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6220 free_rootdomain(old_rd
);
6223 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6225 gfp_t gfp
= GFP_KERNEL
;
6227 memset(rd
, 0, sizeof(*rd
));
6232 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6234 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6236 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6239 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6244 free_cpumask_var(rd
->rto_mask
);
6246 free_cpumask_var(rd
->online
);
6248 free_cpumask_var(rd
->span
);
6253 static void init_defrootdomain(void)
6255 init_rootdomain(&def_root_domain
, true);
6257 atomic_set(&def_root_domain
.refcount
, 1);
6260 static struct root_domain
*alloc_rootdomain(void)
6262 struct root_domain
*rd
;
6264 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6268 if (init_rootdomain(rd
, false) != 0) {
6277 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6278 * hold the hotplug lock.
6281 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6283 struct rq
*rq
= cpu_rq(cpu
);
6284 struct sched_domain
*tmp
;
6286 /* Remove the sched domains which do not contribute to scheduling. */
6287 for (tmp
= sd
; tmp
; ) {
6288 struct sched_domain
*parent
= tmp
->parent
;
6292 if (sd_parent_degenerate(tmp
, parent
)) {
6293 tmp
->parent
= parent
->parent
;
6295 parent
->parent
->child
= tmp
;
6300 if (sd
&& sd_degenerate(sd
)) {
6306 sched_domain_debug(sd
, cpu
);
6308 rq_attach_root(rq
, rd
);
6309 rcu_assign_pointer(rq
->sd
, sd
);
6312 /* cpus with isolated domains */
6313 static cpumask_var_t cpu_isolated_map
;
6315 /* Setup the mask of cpus configured for isolated domains */
6316 static int __init
isolated_cpu_setup(char *str
)
6318 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6319 cpulist_parse(str
, cpu_isolated_map
);
6323 __setup("isolcpus=", isolated_cpu_setup
);
6326 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6327 * to a function which identifies what group(along with sched group) a CPU
6328 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6329 * (due to the fact that we keep track of groups covered with a struct cpumask).
6331 * init_sched_build_groups will build a circular linked list of the groups
6332 * covered by the given span, and will set each group's ->cpumask correctly,
6333 * and ->cpu_power to 0.
6336 init_sched_build_groups(const struct cpumask
*span
,
6337 const struct cpumask
*cpu_map
,
6338 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6339 struct sched_group
**sg
,
6340 struct cpumask
*tmpmask
),
6341 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6343 struct sched_group
*first
= NULL
, *last
= NULL
;
6346 cpumask_clear(covered
);
6348 for_each_cpu(i
, span
) {
6349 struct sched_group
*sg
;
6350 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6353 if (cpumask_test_cpu(i
, covered
))
6356 cpumask_clear(sched_group_cpus(sg
));
6359 for_each_cpu(j
, span
) {
6360 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6363 cpumask_set_cpu(j
, covered
);
6364 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6375 #define SD_NODES_PER_DOMAIN 16
6380 * find_next_best_node - find the next node to include in a sched_domain
6381 * @node: node whose sched_domain we're building
6382 * @used_nodes: nodes already in the sched_domain
6384 * Find the next node to include in a given scheduling domain. Simply
6385 * finds the closest node not already in the @used_nodes map.
6387 * Should use nodemask_t.
6389 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6391 int i
, n
, val
, min_val
, best_node
= 0;
6395 for (i
= 0; i
< nr_node_ids
; i
++) {
6396 /* Start at @node */
6397 n
= (node
+ i
) % nr_node_ids
;
6399 if (!nr_cpus_node(n
))
6402 /* Skip already used nodes */
6403 if (node_isset(n
, *used_nodes
))
6406 /* Simple min distance search */
6407 val
= node_distance(node
, n
);
6409 if (val
< min_val
) {
6415 node_set(best_node
, *used_nodes
);
6420 * sched_domain_node_span - get a cpumask for a node's sched_domain
6421 * @node: node whose cpumask we're constructing
6422 * @span: resulting cpumask
6424 * Given a node, construct a good cpumask for its sched_domain to span. It
6425 * should be one that prevents unnecessary balancing, but also spreads tasks
6428 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6430 nodemask_t used_nodes
;
6433 cpumask_clear(span
);
6434 nodes_clear(used_nodes
);
6436 cpumask_or(span
, span
, cpumask_of_node(node
));
6437 node_set(node
, used_nodes
);
6439 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6440 int next_node
= find_next_best_node(node
, &used_nodes
);
6442 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6445 #endif /* CONFIG_NUMA */
6447 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6450 * The cpus mask in sched_group and sched_domain hangs off the end.
6452 * ( See the the comments in include/linux/sched.h:struct sched_group
6453 * and struct sched_domain. )
6455 struct static_sched_group
{
6456 struct sched_group sg
;
6457 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6460 struct static_sched_domain
{
6461 struct sched_domain sd
;
6462 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6468 cpumask_var_t domainspan
;
6469 cpumask_var_t covered
;
6470 cpumask_var_t notcovered
;
6472 cpumask_var_t nodemask
;
6473 cpumask_var_t this_sibling_map
;
6474 cpumask_var_t this_core_map
;
6475 cpumask_var_t send_covered
;
6476 cpumask_var_t tmpmask
;
6477 struct sched_group
**sched_group_nodes
;
6478 struct root_domain
*rd
;
6482 sa_sched_groups
= 0,
6487 sa_this_sibling_map
,
6489 sa_sched_group_nodes
,
6499 * SMT sched-domains:
6501 #ifdef CONFIG_SCHED_SMT
6502 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6503 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6506 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6507 struct sched_group
**sg
, struct cpumask
*unused
)
6510 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6513 #endif /* CONFIG_SCHED_SMT */
6516 * multi-core sched-domains:
6518 #ifdef CONFIG_SCHED_MC
6519 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6520 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6521 #endif /* CONFIG_SCHED_MC */
6523 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6525 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6526 struct sched_group
**sg
, struct cpumask
*mask
)
6530 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6531 group
= cpumask_first(mask
);
6533 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6536 #elif defined(CONFIG_SCHED_MC)
6538 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6539 struct sched_group
**sg
, struct cpumask
*unused
)
6542 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6547 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6548 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6551 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6552 struct sched_group
**sg
, struct cpumask
*mask
)
6555 #ifdef CONFIG_SCHED_MC
6556 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6557 group
= cpumask_first(mask
);
6558 #elif defined(CONFIG_SCHED_SMT)
6559 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6560 group
= cpumask_first(mask
);
6565 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6571 * The init_sched_build_groups can't handle what we want to do with node
6572 * groups, so roll our own. Now each node has its own list of groups which
6573 * gets dynamically allocated.
6575 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6576 static struct sched_group
***sched_group_nodes_bycpu
;
6578 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6579 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6581 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6582 struct sched_group
**sg
,
6583 struct cpumask
*nodemask
)
6587 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6588 group
= cpumask_first(nodemask
);
6591 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6595 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6597 struct sched_group
*sg
= group_head
;
6603 for_each_cpu(j
, sched_group_cpus(sg
)) {
6604 struct sched_domain
*sd
;
6606 sd
= &per_cpu(phys_domains
, j
).sd
;
6607 if (j
!= group_first_cpu(sd
->groups
)) {
6609 * Only add "power" once for each
6615 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6618 } while (sg
!= group_head
);
6621 static int build_numa_sched_groups(struct s_data
*d
,
6622 const struct cpumask
*cpu_map
, int num
)
6624 struct sched_domain
*sd
;
6625 struct sched_group
*sg
, *prev
;
6628 cpumask_clear(d
->covered
);
6629 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6630 if (cpumask_empty(d
->nodemask
)) {
6631 d
->sched_group_nodes
[num
] = NULL
;
6635 sched_domain_node_span(num
, d
->domainspan
);
6636 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6638 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6641 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6645 d
->sched_group_nodes
[num
] = sg
;
6647 for_each_cpu(j
, d
->nodemask
) {
6648 sd
= &per_cpu(node_domains
, j
).sd
;
6653 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6655 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6658 for (j
= 0; j
< nr_node_ids
; j
++) {
6659 n
= (num
+ j
) % nr_node_ids
;
6660 cpumask_complement(d
->notcovered
, d
->covered
);
6661 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6662 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6663 if (cpumask_empty(d
->tmpmask
))
6665 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6666 if (cpumask_empty(d
->tmpmask
))
6668 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6672 "Can not alloc domain group for node %d\n", j
);
6676 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6677 sg
->next
= prev
->next
;
6678 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6685 #endif /* CONFIG_NUMA */
6688 /* Free memory allocated for various sched_group structures */
6689 static void free_sched_groups(const struct cpumask
*cpu_map
,
6690 struct cpumask
*nodemask
)
6694 for_each_cpu(cpu
, cpu_map
) {
6695 struct sched_group
**sched_group_nodes
6696 = sched_group_nodes_bycpu
[cpu
];
6698 if (!sched_group_nodes
)
6701 for (i
= 0; i
< nr_node_ids
; i
++) {
6702 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6704 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6705 if (cpumask_empty(nodemask
))
6715 if (oldsg
!= sched_group_nodes
[i
])
6718 kfree(sched_group_nodes
);
6719 sched_group_nodes_bycpu
[cpu
] = NULL
;
6722 #else /* !CONFIG_NUMA */
6723 static void free_sched_groups(const struct cpumask
*cpu_map
,
6724 struct cpumask
*nodemask
)
6727 #endif /* CONFIG_NUMA */
6730 * Initialize sched groups cpu_power.
6732 * cpu_power indicates the capacity of sched group, which is used while
6733 * distributing the load between different sched groups in a sched domain.
6734 * Typically cpu_power for all the groups in a sched domain will be same unless
6735 * there are asymmetries in the topology. If there are asymmetries, group
6736 * having more cpu_power will pickup more load compared to the group having
6739 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6741 struct sched_domain
*child
;
6742 struct sched_group
*group
;
6746 WARN_ON(!sd
|| !sd
->groups
);
6748 if (cpu
!= group_first_cpu(sd
->groups
))
6753 sd
->groups
->cpu_power
= 0;
6756 power
= SCHED_LOAD_SCALE
;
6757 weight
= cpumask_weight(sched_domain_span(sd
));
6759 * SMT siblings share the power of a single core.
6760 * Usually multiple threads get a better yield out of
6761 * that one core than a single thread would have,
6762 * reflect that in sd->smt_gain.
6764 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6765 power
*= sd
->smt_gain
;
6767 power
>>= SCHED_LOAD_SHIFT
;
6769 sd
->groups
->cpu_power
+= power
;
6774 * Add cpu_power of each child group to this groups cpu_power.
6776 group
= child
->groups
;
6778 sd
->groups
->cpu_power
+= group
->cpu_power
;
6779 group
= group
->next
;
6780 } while (group
!= child
->groups
);
6784 * Initializers for schedule domains
6785 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6788 #ifdef CONFIG_SCHED_DEBUG
6789 # define SD_INIT_NAME(sd, type) sd->name = #type
6791 # define SD_INIT_NAME(sd, type) do { } while (0)
6794 #define SD_INIT(sd, type) sd_init_##type(sd)
6796 #define SD_INIT_FUNC(type) \
6797 static noinline void sd_init_##type(struct sched_domain *sd) \
6799 memset(sd, 0, sizeof(*sd)); \
6800 *sd = SD_##type##_INIT; \
6801 sd->level = SD_LV_##type; \
6802 SD_INIT_NAME(sd, type); \
6807 SD_INIT_FUNC(ALLNODES
)
6810 #ifdef CONFIG_SCHED_SMT
6811 SD_INIT_FUNC(SIBLING
)
6813 #ifdef CONFIG_SCHED_MC
6817 static int default_relax_domain_level
= -1;
6819 static int __init
setup_relax_domain_level(char *str
)
6823 val
= simple_strtoul(str
, NULL
, 0);
6824 if (val
< SD_LV_MAX
)
6825 default_relax_domain_level
= val
;
6829 __setup("relax_domain_level=", setup_relax_domain_level
);
6831 static void set_domain_attribute(struct sched_domain
*sd
,
6832 struct sched_domain_attr
*attr
)
6836 if (!attr
|| attr
->relax_domain_level
< 0) {
6837 if (default_relax_domain_level
< 0)
6840 request
= default_relax_domain_level
;
6842 request
= attr
->relax_domain_level
;
6843 if (request
< sd
->level
) {
6844 /* turn off idle balance on this domain */
6845 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6847 /* turn on idle balance on this domain */
6848 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6852 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6853 const struct cpumask
*cpu_map
)
6856 case sa_sched_groups
:
6857 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6858 d
->sched_group_nodes
= NULL
;
6860 free_rootdomain(d
->rd
); /* fall through */
6862 free_cpumask_var(d
->tmpmask
); /* fall through */
6863 case sa_send_covered
:
6864 free_cpumask_var(d
->send_covered
); /* fall through */
6865 case sa_this_core_map
:
6866 free_cpumask_var(d
->this_core_map
); /* fall through */
6867 case sa_this_sibling_map
:
6868 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6870 free_cpumask_var(d
->nodemask
); /* fall through */
6871 case sa_sched_group_nodes
:
6873 kfree(d
->sched_group_nodes
); /* fall through */
6875 free_cpumask_var(d
->notcovered
); /* fall through */
6877 free_cpumask_var(d
->covered
); /* fall through */
6879 free_cpumask_var(d
->domainspan
); /* fall through */
6886 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6887 const struct cpumask
*cpu_map
)
6890 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6892 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6893 return sa_domainspan
;
6894 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6896 /* Allocate the per-node list of sched groups */
6897 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6898 sizeof(struct sched_group
*), GFP_KERNEL
);
6899 if (!d
->sched_group_nodes
) {
6900 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6901 return sa_notcovered
;
6903 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6905 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6906 return sa_sched_group_nodes
;
6907 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6909 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6910 return sa_this_sibling_map
;
6911 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6912 return sa_this_core_map
;
6913 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6914 return sa_send_covered
;
6915 d
->rd
= alloc_rootdomain();
6917 printk(KERN_WARNING
"Cannot alloc root domain\n");
6920 return sa_rootdomain
;
6923 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6924 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6926 struct sched_domain
*sd
= NULL
;
6928 struct sched_domain
*parent
;
6931 if (cpumask_weight(cpu_map
) >
6932 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6933 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6934 SD_INIT(sd
, ALLNODES
);
6935 set_domain_attribute(sd
, attr
);
6936 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6937 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6942 sd
= &per_cpu(node_domains
, i
).sd
;
6944 set_domain_attribute(sd
, attr
);
6945 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6946 sd
->parent
= parent
;
6949 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6954 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6955 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6956 struct sched_domain
*parent
, int i
)
6958 struct sched_domain
*sd
;
6959 sd
= &per_cpu(phys_domains
, i
).sd
;
6961 set_domain_attribute(sd
, attr
);
6962 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6963 sd
->parent
= parent
;
6966 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6970 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6971 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6972 struct sched_domain
*parent
, int i
)
6974 struct sched_domain
*sd
= parent
;
6975 #ifdef CONFIG_SCHED_MC
6976 sd
= &per_cpu(core_domains
, i
).sd
;
6978 set_domain_attribute(sd
, attr
);
6979 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6980 sd
->parent
= parent
;
6982 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6987 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6988 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6989 struct sched_domain
*parent
, int i
)
6991 struct sched_domain
*sd
= parent
;
6992 #ifdef CONFIG_SCHED_SMT
6993 sd
= &per_cpu(cpu_domains
, i
).sd
;
6994 SD_INIT(sd
, SIBLING
);
6995 set_domain_attribute(sd
, attr
);
6996 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6997 sd
->parent
= parent
;
6999 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
7004 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
7005 const struct cpumask
*cpu_map
, int cpu
)
7008 #ifdef CONFIG_SCHED_SMT
7009 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
7010 cpumask_and(d
->this_sibling_map
, cpu_map
,
7011 topology_thread_cpumask(cpu
));
7012 if (cpu
== cpumask_first(d
->this_sibling_map
))
7013 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7015 d
->send_covered
, d
->tmpmask
);
7018 #ifdef CONFIG_SCHED_MC
7019 case SD_LV_MC
: /* set up multi-core groups */
7020 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7021 if (cpu
== cpumask_first(d
->this_core_map
))
7022 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7024 d
->send_covered
, d
->tmpmask
);
7027 case SD_LV_CPU
: /* set up physical groups */
7028 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7029 if (!cpumask_empty(d
->nodemask
))
7030 init_sched_build_groups(d
->nodemask
, cpu_map
,
7032 d
->send_covered
, d
->tmpmask
);
7035 case SD_LV_ALLNODES
:
7036 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7037 d
->send_covered
, d
->tmpmask
);
7046 * Build sched domains for a given set of cpus and attach the sched domains
7047 * to the individual cpus
7049 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7050 struct sched_domain_attr
*attr
)
7052 enum s_alloc alloc_state
= sa_none
;
7054 struct sched_domain
*sd
;
7060 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7061 if (alloc_state
!= sa_rootdomain
)
7063 alloc_state
= sa_sched_groups
;
7066 * Set up domains for cpus specified by the cpu_map.
7068 for_each_cpu(i
, cpu_map
) {
7069 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7072 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7073 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7074 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7075 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7078 for_each_cpu(i
, cpu_map
) {
7079 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7080 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7083 /* Set up physical groups */
7084 for (i
= 0; i
< nr_node_ids
; i
++)
7085 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7088 /* Set up node groups */
7090 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7092 for (i
= 0; i
< nr_node_ids
; i
++)
7093 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7097 /* Calculate CPU power for physical packages and nodes */
7098 #ifdef CONFIG_SCHED_SMT
7099 for_each_cpu(i
, cpu_map
) {
7100 sd
= &per_cpu(cpu_domains
, i
).sd
;
7101 init_sched_groups_power(i
, sd
);
7104 #ifdef CONFIG_SCHED_MC
7105 for_each_cpu(i
, cpu_map
) {
7106 sd
= &per_cpu(core_domains
, i
).sd
;
7107 init_sched_groups_power(i
, sd
);
7111 for_each_cpu(i
, cpu_map
) {
7112 sd
= &per_cpu(phys_domains
, i
).sd
;
7113 init_sched_groups_power(i
, sd
);
7117 for (i
= 0; i
< nr_node_ids
; i
++)
7118 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7120 if (d
.sd_allnodes
) {
7121 struct sched_group
*sg
;
7123 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7125 init_numa_sched_groups_power(sg
);
7129 /* Attach the domains */
7130 for_each_cpu(i
, cpu_map
) {
7131 #ifdef CONFIG_SCHED_SMT
7132 sd
= &per_cpu(cpu_domains
, i
).sd
;
7133 #elif defined(CONFIG_SCHED_MC)
7134 sd
= &per_cpu(core_domains
, i
).sd
;
7136 sd
= &per_cpu(phys_domains
, i
).sd
;
7138 cpu_attach_domain(sd
, d
.rd
, i
);
7141 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7142 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7146 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7150 static int build_sched_domains(const struct cpumask
*cpu_map
)
7152 return __build_sched_domains(cpu_map
, NULL
);
7155 static cpumask_var_t
*doms_cur
; /* current sched domains */
7156 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7157 static struct sched_domain_attr
*dattr_cur
;
7158 /* attribues of custom domains in 'doms_cur' */
7161 * Special case: If a kmalloc of a doms_cur partition (array of
7162 * cpumask) fails, then fallback to a single sched domain,
7163 * as determined by the single cpumask fallback_doms.
7165 static cpumask_var_t fallback_doms
;
7168 * arch_update_cpu_topology lets virtualized architectures update the
7169 * cpu core maps. It is supposed to return 1 if the topology changed
7170 * or 0 if it stayed the same.
7172 int __attribute__((weak
)) arch_update_cpu_topology(void)
7177 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7180 cpumask_var_t
*doms
;
7182 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7185 for (i
= 0; i
< ndoms
; i
++) {
7186 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7187 free_sched_domains(doms
, i
);
7194 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7197 for (i
= 0; i
< ndoms
; i
++)
7198 free_cpumask_var(doms
[i
]);
7203 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7204 * For now this just excludes isolated cpus, but could be used to
7205 * exclude other special cases in the future.
7207 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7211 arch_update_cpu_topology();
7213 doms_cur
= alloc_sched_domains(ndoms_cur
);
7215 doms_cur
= &fallback_doms
;
7216 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7218 err
= build_sched_domains(doms_cur
[0]);
7219 register_sched_domain_sysctl();
7224 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7225 struct cpumask
*tmpmask
)
7227 free_sched_groups(cpu_map
, tmpmask
);
7231 * Detach sched domains from a group of cpus specified in cpu_map
7232 * These cpus will now be attached to the NULL domain
7234 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7236 /* Save because hotplug lock held. */
7237 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7240 for_each_cpu(i
, cpu_map
)
7241 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7242 synchronize_sched();
7243 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7246 /* handle null as "default" */
7247 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7248 struct sched_domain_attr
*new, int idx_new
)
7250 struct sched_domain_attr tmp
;
7257 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7258 new ? (new + idx_new
) : &tmp
,
7259 sizeof(struct sched_domain_attr
));
7263 * Partition sched domains as specified by the 'ndoms_new'
7264 * cpumasks in the array doms_new[] of cpumasks. This compares
7265 * doms_new[] to the current sched domain partitioning, doms_cur[].
7266 * It destroys each deleted domain and builds each new domain.
7268 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7269 * The masks don't intersect (don't overlap.) We should setup one
7270 * sched domain for each mask. CPUs not in any of the cpumasks will
7271 * not be load balanced. If the same cpumask appears both in the
7272 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7275 * The passed in 'doms_new' should be allocated using
7276 * alloc_sched_domains. This routine takes ownership of it and will
7277 * free_sched_domains it when done with it. If the caller failed the
7278 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7279 * and partition_sched_domains() will fallback to the single partition
7280 * 'fallback_doms', it also forces the domains to be rebuilt.
7282 * If doms_new == NULL it will be replaced with cpu_online_mask.
7283 * ndoms_new == 0 is a special case for destroying existing domains,
7284 * and it will not create the default domain.
7286 * Call with hotplug lock held
7288 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7289 struct sched_domain_attr
*dattr_new
)
7294 mutex_lock(&sched_domains_mutex
);
7296 /* always unregister in case we don't destroy any domains */
7297 unregister_sched_domain_sysctl();
7299 /* Let architecture update cpu core mappings. */
7300 new_topology
= arch_update_cpu_topology();
7302 n
= doms_new
? ndoms_new
: 0;
7304 /* Destroy deleted domains */
7305 for (i
= 0; i
< ndoms_cur
; i
++) {
7306 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7307 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7308 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7311 /* no match - a current sched domain not in new doms_new[] */
7312 detach_destroy_domains(doms_cur
[i
]);
7317 if (doms_new
== NULL
) {
7319 doms_new
= &fallback_doms
;
7320 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7321 WARN_ON_ONCE(dattr_new
);
7324 /* Build new domains */
7325 for (i
= 0; i
< ndoms_new
; i
++) {
7326 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7327 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7328 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7331 /* no match - add a new doms_new */
7332 __build_sched_domains(doms_new
[i
],
7333 dattr_new
? dattr_new
+ i
: NULL
);
7338 /* Remember the new sched domains */
7339 if (doms_cur
!= &fallback_doms
)
7340 free_sched_domains(doms_cur
, ndoms_cur
);
7341 kfree(dattr_cur
); /* kfree(NULL) is safe */
7342 doms_cur
= doms_new
;
7343 dattr_cur
= dattr_new
;
7344 ndoms_cur
= ndoms_new
;
7346 register_sched_domain_sysctl();
7348 mutex_unlock(&sched_domains_mutex
);
7351 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7352 static void arch_reinit_sched_domains(void)
7356 /* Destroy domains first to force the rebuild */
7357 partition_sched_domains(0, NULL
, NULL
);
7359 rebuild_sched_domains();
7363 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7365 unsigned int level
= 0;
7367 if (sscanf(buf
, "%u", &level
) != 1)
7371 * level is always be positive so don't check for
7372 * level < POWERSAVINGS_BALANCE_NONE which is 0
7373 * What happens on 0 or 1 byte write,
7374 * need to check for count as well?
7377 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7381 sched_smt_power_savings
= level
;
7383 sched_mc_power_savings
= level
;
7385 arch_reinit_sched_domains();
7390 #ifdef CONFIG_SCHED_MC
7391 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7394 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7396 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7397 const char *buf
, size_t count
)
7399 return sched_power_savings_store(buf
, count
, 0);
7401 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7402 sched_mc_power_savings_show
,
7403 sched_mc_power_savings_store
);
7406 #ifdef CONFIG_SCHED_SMT
7407 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7410 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7412 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7413 const char *buf
, size_t count
)
7415 return sched_power_savings_store(buf
, count
, 1);
7417 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7418 sched_smt_power_savings_show
,
7419 sched_smt_power_savings_store
);
7422 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7426 #ifdef CONFIG_SCHED_SMT
7428 err
= sysfs_create_file(&cls
->kset
.kobj
,
7429 &attr_sched_smt_power_savings
.attr
);
7431 #ifdef CONFIG_SCHED_MC
7432 if (!err
&& mc_capable())
7433 err
= sysfs_create_file(&cls
->kset
.kobj
,
7434 &attr_sched_mc_power_savings
.attr
);
7438 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7440 #ifndef CONFIG_CPUSETS
7442 * Add online and remove offline CPUs from the scheduler domains.
7443 * When cpusets are enabled they take over this function.
7445 static int update_sched_domains(struct notifier_block
*nfb
,
7446 unsigned long action
, void *hcpu
)
7450 case CPU_ONLINE_FROZEN
:
7451 case CPU_DOWN_PREPARE
:
7452 case CPU_DOWN_PREPARE_FROZEN
:
7453 case CPU_DOWN_FAILED
:
7454 case CPU_DOWN_FAILED_FROZEN
:
7455 partition_sched_domains(1, NULL
, NULL
);
7464 static int update_runtime(struct notifier_block
*nfb
,
7465 unsigned long action
, void *hcpu
)
7467 int cpu
= (int)(long)hcpu
;
7470 case CPU_DOWN_PREPARE
:
7471 case CPU_DOWN_PREPARE_FROZEN
:
7472 disable_runtime(cpu_rq(cpu
));
7475 case CPU_DOWN_FAILED
:
7476 case CPU_DOWN_FAILED_FROZEN
:
7478 case CPU_ONLINE_FROZEN
:
7479 enable_runtime(cpu_rq(cpu
));
7487 void __init
sched_init_smp(void)
7489 cpumask_var_t non_isolated_cpus
;
7491 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7492 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7494 #if defined(CONFIG_NUMA)
7495 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7497 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7500 mutex_lock(&sched_domains_mutex
);
7501 arch_init_sched_domains(cpu_active_mask
);
7502 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7503 if (cpumask_empty(non_isolated_cpus
))
7504 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7505 mutex_unlock(&sched_domains_mutex
);
7508 #ifndef CONFIG_CPUSETS
7509 /* XXX: Theoretical race here - CPU may be hotplugged now */
7510 hotcpu_notifier(update_sched_domains
, 0);
7513 /* RT runtime code needs to handle some hotplug events */
7514 hotcpu_notifier(update_runtime
, 0);
7518 /* Move init over to a non-isolated CPU */
7519 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7521 sched_init_granularity();
7522 free_cpumask_var(non_isolated_cpus
);
7524 init_sched_rt_class();
7527 void __init
sched_init_smp(void)
7529 sched_init_granularity();
7531 #endif /* CONFIG_SMP */
7533 const_debug
unsigned int sysctl_timer_migration
= 1;
7535 int in_sched_functions(unsigned long addr
)
7537 return in_lock_functions(addr
) ||
7538 (addr
>= (unsigned long)__sched_text_start
7539 && addr
< (unsigned long)__sched_text_end
);
7542 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7544 cfs_rq
->tasks_timeline
= RB_ROOT
;
7545 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7546 #ifdef CONFIG_FAIR_GROUP_SCHED
7549 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7552 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7554 struct rt_prio_array
*array
;
7557 array
= &rt_rq
->active
;
7558 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7559 INIT_LIST_HEAD(array
->queue
+ i
);
7560 __clear_bit(i
, array
->bitmap
);
7562 /* delimiter for bitsearch: */
7563 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7565 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7566 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7568 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7572 rt_rq
->rt_nr_migratory
= 0;
7573 rt_rq
->overloaded
= 0;
7574 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7578 rt_rq
->rt_throttled
= 0;
7579 rt_rq
->rt_runtime
= 0;
7580 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7582 #ifdef CONFIG_RT_GROUP_SCHED
7583 rt_rq
->rt_nr_boosted
= 0;
7588 #ifdef CONFIG_FAIR_GROUP_SCHED
7589 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7590 struct sched_entity
*se
, int cpu
, int add
,
7591 struct sched_entity
*parent
)
7593 struct rq
*rq
= cpu_rq(cpu
);
7594 tg
->cfs_rq
[cpu
] = cfs_rq
;
7595 init_cfs_rq(cfs_rq
, rq
);
7598 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7601 /* se could be NULL for init_task_group */
7606 se
->cfs_rq
= &rq
->cfs
;
7608 se
->cfs_rq
= parent
->my_q
;
7611 se
->load
.weight
= tg
->shares
;
7612 se
->load
.inv_weight
= 0;
7613 se
->parent
= parent
;
7617 #ifdef CONFIG_RT_GROUP_SCHED
7618 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7619 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7620 struct sched_rt_entity
*parent
)
7622 struct rq
*rq
= cpu_rq(cpu
);
7624 tg
->rt_rq
[cpu
] = rt_rq
;
7625 init_rt_rq(rt_rq
, rq
);
7627 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7629 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7631 tg
->rt_se
[cpu
] = rt_se
;
7636 rt_se
->rt_rq
= &rq
->rt
;
7638 rt_se
->rt_rq
= parent
->my_q
;
7640 rt_se
->my_q
= rt_rq
;
7641 rt_se
->parent
= parent
;
7642 INIT_LIST_HEAD(&rt_se
->run_list
);
7646 void __init
sched_init(void)
7649 unsigned long alloc_size
= 0, ptr
;
7651 #ifdef CONFIG_FAIR_GROUP_SCHED
7652 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7654 #ifdef CONFIG_RT_GROUP_SCHED
7655 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7657 #ifdef CONFIG_CPUMASK_OFFSTACK
7658 alloc_size
+= num_possible_cpus() * cpumask_size();
7661 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7663 #ifdef CONFIG_FAIR_GROUP_SCHED
7664 init_task_group
.se
= (struct sched_entity
**)ptr
;
7665 ptr
+= nr_cpu_ids
* sizeof(void **);
7667 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7668 ptr
+= nr_cpu_ids
* sizeof(void **);
7670 #endif /* CONFIG_FAIR_GROUP_SCHED */
7671 #ifdef CONFIG_RT_GROUP_SCHED
7672 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7673 ptr
+= nr_cpu_ids
* sizeof(void **);
7675 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7676 ptr
+= nr_cpu_ids
* sizeof(void **);
7678 #endif /* CONFIG_RT_GROUP_SCHED */
7679 #ifdef CONFIG_CPUMASK_OFFSTACK
7680 for_each_possible_cpu(i
) {
7681 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7682 ptr
+= cpumask_size();
7684 #endif /* CONFIG_CPUMASK_OFFSTACK */
7688 init_defrootdomain();
7691 init_rt_bandwidth(&def_rt_bandwidth
,
7692 global_rt_period(), global_rt_runtime());
7694 #ifdef CONFIG_RT_GROUP_SCHED
7695 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7696 global_rt_period(), global_rt_runtime());
7697 #endif /* CONFIG_RT_GROUP_SCHED */
7699 #ifdef CONFIG_CGROUP_SCHED
7700 list_add(&init_task_group
.list
, &task_groups
);
7701 INIT_LIST_HEAD(&init_task_group
.children
);
7703 #endif /* CONFIG_CGROUP_SCHED */
7705 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7706 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7707 __alignof__(unsigned long));
7709 for_each_possible_cpu(i
) {
7713 raw_spin_lock_init(&rq
->lock
);
7715 rq
->calc_load_active
= 0;
7716 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7717 init_cfs_rq(&rq
->cfs
, rq
);
7718 init_rt_rq(&rq
->rt
, rq
);
7719 #ifdef CONFIG_FAIR_GROUP_SCHED
7720 init_task_group
.shares
= init_task_group_load
;
7721 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7722 #ifdef CONFIG_CGROUP_SCHED
7724 * How much cpu bandwidth does init_task_group get?
7726 * In case of task-groups formed thr' the cgroup filesystem, it
7727 * gets 100% of the cpu resources in the system. This overall
7728 * system cpu resource is divided among the tasks of
7729 * init_task_group and its child task-groups in a fair manner,
7730 * based on each entity's (task or task-group's) weight
7731 * (se->load.weight).
7733 * In other words, if init_task_group has 10 tasks of weight
7734 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7735 * then A0's share of the cpu resource is:
7737 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7739 * We achieve this by letting init_task_group's tasks sit
7740 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7742 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7744 #endif /* CONFIG_FAIR_GROUP_SCHED */
7746 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7747 #ifdef CONFIG_RT_GROUP_SCHED
7748 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7749 #ifdef CONFIG_CGROUP_SCHED
7750 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7754 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7755 rq
->cpu_load
[j
] = 0;
7759 rq
->post_schedule
= 0;
7760 rq
->active_balance
= 0;
7761 rq
->next_balance
= jiffies
;
7765 rq
->migration_thread
= NULL
;
7767 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7768 INIT_LIST_HEAD(&rq
->migration_queue
);
7769 rq_attach_root(rq
, &def_root_domain
);
7772 atomic_set(&rq
->nr_iowait
, 0);
7775 set_load_weight(&init_task
);
7777 #ifdef CONFIG_PREEMPT_NOTIFIERS
7778 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7782 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7785 #ifdef CONFIG_RT_MUTEXES
7786 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7790 * The boot idle thread does lazy MMU switching as well:
7792 atomic_inc(&init_mm
.mm_count
);
7793 enter_lazy_tlb(&init_mm
, current
);
7796 * Make us the idle thread. Technically, schedule() should not be
7797 * called from this thread, however somewhere below it might be,
7798 * but because we are the idle thread, we just pick up running again
7799 * when this runqueue becomes "idle".
7801 init_idle(current
, smp_processor_id());
7803 calc_load_update
= jiffies
+ LOAD_FREQ
;
7806 * During early bootup we pretend to be a normal task:
7808 current
->sched_class
= &fair_sched_class
;
7810 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7811 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7814 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7815 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7817 /* May be allocated at isolcpus cmdline parse time */
7818 if (cpu_isolated_map
== NULL
)
7819 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7824 scheduler_running
= 1;
7827 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7828 static inline int preempt_count_equals(int preempt_offset
)
7830 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7832 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7835 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7838 static unsigned long prev_jiffy
; /* ratelimiting */
7840 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7841 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7843 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7845 prev_jiffy
= jiffies
;
7848 "BUG: sleeping function called from invalid context at %s:%d\n",
7851 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7852 in_atomic(), irqs_disabled(),
7853 current
->pid
, current
->comm
);
7855 debug_show_held_locks(current
);
7856 if (irqs_disabled())
7857 print_irqtrace_events(current
);
7861 EXPORT_SYMBOL(__might_sleep
);
7864 #ifdef CONFIG_MAGIC_SYSRQ
7865 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7869 update_rq_clock(rq
);
7870 on_rq
= p
->se
.on_rq
;
7872 deactivate_task(rq
, p
, 0);
7873 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7875 activate_task(rq
, p
, 0);
7876 resched_task(rq
->curr
);
7880 void normalize_rt_tasks(void)
7882 struct task_struct
*g
, *p
;
7883 unsigned long flags
;
7886 read_lock_irqsave(&tasklist_lock
, flags
);
7887 do_each_thread(g
, p
) {
7889 * Only normalize user tasks:
7894 p
->se
.exec_start
= 0;
7895 #ifdef CONFIG_SCHEDSTATS
7896 p
->se
.statistics
.wait_start
= 0;
7897 p
->se
.statistics
.sleep_start
= 0;
7898 p
->se
.statistics
.block_start
= 0;
7903 * Renice negative nice level userspace
7906 if (TASK_NICE(p
) < 0 && p
->mm
)
7907 set_user_nice(p
, 0);
7911 raw_spin_lock(&p
->pi_lock
);
7912 rq
= __task_rq_lock(p
);
7914 normalize_task(rq
, p
);
7916 __task_rq_unlock(rq
);
7917 raw_spin_unlock(&p
->pi_lock
);
7918 } while_each_thread(g
, p
);
7920 read_unlock_irqrestore(&tasklist_lock
, flags
);
7923 #endif /* CONFIG_MAGIC_SYSRQ */
7927 * These functions are only useful for the IA64 MCA handling.
7929 * They can only be called when the whole system has been
7930 * stopped - every CPU needs to be quiescent, and no scheduling
7931 * activity can take place. Using them for anything else would
7932 * be a serious bug, and as a result, they aren't even visible
7933 * under any other configuration.
7937 * curr_task - return the current task for a given cpu.
7938 * @cpu: the processor in question.
7940 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7942 struct task_struct
*curr_task(int cpu
)
7944 return cpu_curr(cpu
);
7948 * set_curr_task - set the current task for a given cpu.
7949 * @cpu: the processor in question.
7950 * @p: the task pointer to set.
7952 * Description: This function must only be used when non-maskable interrupts
7953 * are serviced on a separate stack. It allows the architecture to switch the
7954 * notion of the current task on a cpu in a non-blocking manner. This function
7955 * must be called with all CPU's synchronized, and interrupts disabled, the
7956 * and caller must save the original value of the current task (see
7957 * curr_task() above) and restore that value before reenabling interrupts and
7958 * re-starting the system.
7960 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7962 void set_curr_task(int cpu
, struct task_struct
*p
)
7969 #ifdef CONFIG_FAIR_GROUP_SCHED
7970 static void free_fair_sched_group(struct task_group
*tg
)
7974 for_each_possible_cpu(i
) {
7976 kfree(tg
->cfs_rq
[i
]);
7986 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7988 struct cfs_rq
*cfs_rq
;
7989 struct sched_entity
*se
;
7993 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7996 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8000 tg
->shares
= NICE_0_LOAD
;
8002 for_each_possible_cpu(i
) {
8005 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
8006 GFP_KERNEL
, cpu_to_node(i
));
8010 se
= kzalloc_node(sizeof(struct sched_entity
),
8011 GFP_KERNEL
, cpu_to_node(i
));
8015 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8026 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8028 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8029 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8032 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8034 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8036 #else /* !CONFG_FAIR_GROUP_SCHED */
8037 static inline void free_fair_sched_group(struct task_group
*tg
)
8042 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8047 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8051 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8054 #endif /* CONFIG_FAIR_GROUP_SCHED */
8056 #ifdef CONFIG_RT_GROUP_SCHED
8057 static void free_rt_sched_group(struct task_group
*tg
)
8061 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8063 for_each_possible_cpu(i
) {
8065 kfree(tg
->rt_rq
[i
]);
8067 kfree(tg
->rt_se
[i
]);
8075 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8077 struct rt_rq
*rt_rq
;
8078 struct sched_rt_entity
*rt_se
;
8082 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8085 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8089 init_rt_bandwidth(&tg
->rt_bandwidth
,
8090 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8092 for_each_possible_cpu(i
) {
8095 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8096 GFP_KERNEL
, cpu_to_node(i
));
8100 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8101 GFP_KERNEL
, cpu_to_node(i
));
8105 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8116 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8118 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8119 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8122 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8124 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8126 #else /* !CONFIG_RT_GROUP_SCHED */
8127 static inline void free_rt_sched_group(struct task_group
*tg
)
8132 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8137 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8141 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8144 #endif /* CONFIG_RT_GROUP_SCHED */
8146 #ifdef CONFIG_CGROUP_SCHED
8147 static void free_sched_group(struct task_group
*tg
)
8149 free_fair_sched_group(tg
);
8150 free_rt_sched_group(tg
);
8154 /* allocate runqueue etc for a new task group */
8155 struct task_group
*sched_create_group(struct task_group
*parent
)
8157 struct task_group
*tg
;
8158 unsigned long flags
;
8161 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8163 return ERR_PTR(-ENOMEM
);
8165 if (!alloc_fair_sched_group(tg
, parent
))
8168 if (!alloc_rt_sched_group(tg
, parent
))
8171 spin_lock_irqsave(&task_group_lock
, flags
);
8172 for_each_possible_cpu(i
) {
8173 register_fair_sched_group(tg
, i
);
8174 register_rt_sched_group(tg
, i
);
8176 list_add_rcu(&tg
->list
, &task_groups
);
8178 WARN_ON(!parent
); /* root should already exist */
8180 tg
->parent
= parent
;
8181 INIT_LIST_HEAD(&tg
->children
);
8182 list_add_rcu(&tg
->siblings
, &parent
->children
);
8183 spin_unlock_irqrestore(&task_group_lock
, flags
);
8188 free_sched_group(tg
);
8189 return ERR_PTR(-ENOMEM
);
8192 /* rcu callback to free various structures associated with a task group */
8193 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8195 /* now it should be safe to free those cfs_rqs */
8196 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8199 /* Destroy runqueue etc associated with a task group */
8200 void sched_destroy_group(struct task_group
*tg
)
8202 unsigned long flags
;
8205 spin_lock_irqsave(&task_group_lock
, flags
);
8206 for_each_possible_cpu(i
) {
8207 unregister_fair_sched_group(tg
, i
);
8208 unregister_rt_sched_group(tg
, i
);
8210 list_del_rcu(&tg
->list
);
8211 list_del_rcu(&tg
->siblings
);
8212 spin_unlock_irqrestore(&task_group_lock
, flags
);
8214 /* wait for possible concurrent references to cfs_rqs complete */
8215 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8218 /* change task's runqueue when it moves between groups.
8219 * The caller of this function should have put the task in its new group
8220 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8221 * reflect its new group.
8223 void sched_move_task(struct task_struct
*tsk
)
8226 unsigned long flags
;
8229 rq
= task_rq_lock(tsk
, &flags
);
8231 update_rq_clock(rq
);
8233 running
= task_current(rq
, tsk
);
8234 on_rq
= tsk
->se
.on_rq
;
8237 dequeue_task(rq
, tsk
, 0);
8238 if (unlikely(running
))
8239 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8241 set_task_rq(tsk
, task_cpu(tsk
));
8243 #ifdef CONFIG_FAIR_GROUP_SCHED
8244 if (tsk
->sched_class
->moved_group
)
8245 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8248 if (unlikely(running
))
8249 tsk
->sched_class
->set_curr_task(rq
);
8251 enqueue_task(rq
, tsk
, 0, false);
8253 task_rq_unlock(rq
, &flags
);
8255 #endif /* CONFIG_CGROUP_SCHED */
8257 #ifdef CONFIG_FAIR_GROUP_SCHED
8258 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8260 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8265 dequeue_entity(cfs_rq
, se
, 0);
8267 se
->load
.weight
= shares
;
8268 se
->load
.inv_weight
= 0;
8271 enqueue_entity(cfs_rq
, se
, 0);
8274 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8276 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8277 struct rq
*rq
= cfs_rq
->rq
;
8278 unsigned long flags
;
8280 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8281 __set_se_shares(se
, shares
);
8282 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8285 static DEFINE_MUTEX(shares_mutex
);
8287 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8290 unsigned long flags
;
8293 * We can't change the weight of the root cgroup.
8298 if (shares
< MIN_SHARES
)
8299 shares
= MIN_SHARES
;
8300 else if (shares
> MAX_SHARES
)
8301 shares
= MAX_SHARES
;
8303 mutex_lock(&shares_mutex
);
8304 if (tg
->shares
== shares
)
8307 spin_lock_irqsave(&task_group_lock
, flags
);
8308 for_each_possible_cpu(i
)
8309 unregister_fair_sched_group(tg
, i
);
8310 list_del_rcu(&tg
->siblings
);
8311 spin_unlock_irqrestore(&task_group_lock
, flags
);
8313 /* wait for any ongoing reference to this group to finish */
8314 synchronize_sched();
8317 * Now we are free to modify the group's share on each cpu
8318 * w/o tripping rebalance_share or load_balance_fair.
8320 tg
->shares
= shares
;
8321 for_each_possible_cpu(i
) {
8325 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8326 set_se_shares(tg
->se
[i
], shares
);
8330 * Enable load balance activity on this group, by inserting it back on
8331 * each cpu's rq->leaf_cfs_rq_list.
8333 spin_lock_irqsave(&task_group_lock
, flags
);
8334 for_each_possible_cpu(i
)
8335 register_fair_sched_group(tg
, i
);
8336 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8337 spin_unlock_irqrestore(&task_group_lock
, flags
);
8339 mutex_unlock(&shares_mutex
);
8343 unsigned long sched_group_shares(struct task_group
*tg
)
8349 #ifdef CONFIG_RT_GROUP_SCHED
8351 * Ensure that the real time constraints are schedulable.
8353 static DEFINE_MUTEX(rt_constraints_mutex
);
8355 static unsigned long to_ratio(u64 period
, u64 runtime
)
8357 if (runtime
== RUNTIME_INF
)
8360 return div64_u64(runtime
<< 20, period
);
8363 /* Must be called with tasklist_lock held */
8364 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8366 struct task_struct
*g
, *p
;
8368 do_each_thread(g
, p
) {
8369 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8371 } while_each_thread(g
, p
);
8376 struct rt_schedulable_data
{
8377 struct task_group
*tg
;
8382 static int tg_schedulable(struct task_group
*tg
, void *data
)
8384 struct rt_schedulable_data
*d
= data
;
8385 struct task_group
*child
;
8386 unsigned long total
, sum
= 0;
8387 u64 period
, runtime
;
8389 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8390 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8393 period
= d
->rt_period
;
8394 runtime
= d
->rt_runtime
;
8398 * Cannot have more runtime than the period.
8400 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8404 * Ensure we don't starve existing RT tasks.
8406 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8409 total
= to_ratio(period
, runtime
);
8412 * Nobody can have more than the global setting allows.
8414 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8418 * The sum of our children's runtime should not exceed our own.
8420 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8421 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8422 runtime
= child
->rt_bandwidth
.rt_runtime
;
8424 if (child
== d
->tg
) {
8425 period
= d
->rt_period
;
8426 runtime
= d
->rt_runtime
;
8429 sum
+= to_ratio(period
, runtime
);
8438 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8440 struct rt_schedulable_data data
= {
8442 .rt_period
= period
,
8443 .rt_runtime
= runtime
,
8446 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8449 static int tg_set_bandwidth(struct task_group
*tg
,
8450 u64 rt_period
, u64 rt_runtime
)
8454 mutex_lock(&rt_constraints_mutex
);
8455 read_lock(&tasklist_lock
);
8456 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8460 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8461 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8462 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8464 for_each_possible_cpu(i
) {
8465 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8467 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8468 rt_rq
->rt_runtime
= rt_runtime
;
8469 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8471 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8473 read_unlock(&tasklist_lock
);
8474 mutex_unlock(&rt_constraints_mutex
);
8479 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8481 u64 rt_runtime
, rt_period
;
8483 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8484 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8485 if (rt_runtime_us
< 0)
8486 rt_runtime
= RUNTIME_INF
;
8488 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8491 long sched_group_rt_runtime(struct task_group
*tg
)
8495 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8498 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8499 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8500 return rt_runtime_us
;
8503 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8505 u64 rt_runtime
, rt_period
;
8507 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8508 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8513 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8516 long sched_group_rt_period(struct task_group
*tg
)
8520 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8521 do_div(rt_period_us
, NSEC_PER_USEC
);
8522 return rt_period_us
;
8525 static int sched_rt_global_constraints(void)
8527 u64 runtime
, period
;
8530 if (sysctl_sched_rt_period
<= 0)
8533 runtime
= global_rt_runtime();
8534 period
= global_rt_period();
8537 * Sanity check on the sysctl variables.
8539 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8542 mutex_lock(&rt_constraints_mutex
);
8543 read_lock(&tasklist_lock
);
8544 ret
= __rt_schedulable(NULL
, 0, 0);
8545 read_unlock(&tasklist_lock
);
8546 mutex_unlock(&rt_constraints_mutex
);
8551 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8553 /* Don't accept realtime tasks when there is no way for them to run */
8554 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8560 #else /* !CONFIG_RT_GROUP_SCHED */
8561 static int sched_rt_global_constraints(void)
8563 unsigned long flags
;
8566 if (sysctl_sched_rt_period
<= 0)
8570 * There's always some RT tasks in the root group
8571 * -- migration, kstopmachine etc..
8573 if (sysctl_sched_rt_runtime
== 0)
8576 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8577 for_each_possible_cpu(i
) {
8578 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8580 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8581 rt_rq
->rt_runtime
= global_rt_runtime();
8582 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8584 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8588 #endif /* CONFIG_RT_GROUP_SCHED */
8590 int sched_rt_handler(struct ctl_table
*table
, int write
,
8591 void __user
*buffer
, size_t *lenp
,
8595 int old_period
, old_runtime
;
8596 static DEFINE_MUTEX(mutex
);
8599 old_period
= sysctl_sched_rt_period
;
8600 old_runtime
= sysctl_sched_rt_runtime
;
8602 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8604 if (!ret
&& write
) {
8605 ret
= sched_rt_global_constraints();
8607 sysctl_sched_rt_period
= old_period
;
8608 sysctl_sched_rt_runtime
= old_runtime
;
8610 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8611 def_rt_bandwidth
.rt_period
=
8612 ns_to_ktime(global_rt_period());
8615 mutex_unlock(&mutex
);
8620 #ifdef CONFIG_CGROUP_SCHED
8622 /* return corresponding task_group object of a cgroup */
8623 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8625 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8626 struct task_group
, css
);
8629 static struct cgroup_subsys_state
*
8630 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8632 struct task_group
*tg
, *parent
;
8634 if (!cgrp
->parent
) {
8635 /* This is early initialization for the top cgroup */
8636 return &init_task_group
.css
;
8639 parent
= cgroup_tg(cgrp
->parent
);
8640 tg
= sched_create_group(parent
);
8642 return ERR_PTR(-ENOMEM
);
8648 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8650 struct task_group
*tg
= cgroup_tg(cgrp
);
8652 sched_destroy_group(tg
);
8656 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8658 #ifdef CONFIG_RT_GROUP_SCHED
8659 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8662 /* We don't support RT-tasks being in separate groups */
8663 if (tsk
->sched_class
!= &fair_sched_class
)
8670 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8671 struct task_struct
*tsk
, bool threadgroup
)
8673 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8677 struct task_struct
*c
;
8679 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8680 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8692 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8693 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8696 sched_move_task(tsk
);
8698 struct task_struct
*c
;
8700 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8707 #ifdef CONFIG_FAIR_GROUP_SCHED
8708 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8711 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8714 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8716 struct task_group
*tg
= cgroup_tg(cgrp
);
8718 return (u64
) tg
->shares
;
8720 #endif /* CONFIG_FAIR_GROUP_SCHED */
8722 #ifdef CONFIG_RT_GROUP_SCHED
8723 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8726 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8729 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8731 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8734 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8737 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8740 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8742 return sched_group_rt_period(cgroup_tg(cgrp
));
8744 #endif /* CONFIG_RT_GROUP_SCHED */
8746 static struct cftype cpu_files
[] = {
8747 #ifdef CONFIG_FAIR_GROUP_SCHED
8750 .read_u64
= cpu_shares_read_u64
,
8751 .write_u64
= cpu_shares_write_u64
,
8754 #ifdef CONFIG_RT_GROUP_SCHED
8756 .name
= "rt_runtime_us",
8757 .read_s64
= cpu_rt_runtime_read
,
8758 .write_s64
= cpu_rt_runtime_write
,
8761 .name
= "rt_period_us",
8762 .read_u64
= cpu_rt_period_read_uint
,
8763 .write_u64
= cpu_rt_period_write_uint
,
8768 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8770 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8773 struct cgroup_subsys cpu_cgroup_subsys
= {
8775 .create
= cpu_cgroup_create
,
8776 .destroy
= cpu_cgroup_destroy
,
8777 .can_attach
= cpu_cgroup_can_attach
,
8778 .attach
= cpu_cgroup_attach
,
8779 .populate
= cpu_cgroup_populate
,
8780 .subsys_id
= cpu_cgroup_subsys_id
,
8784 #endif /* CONFIG_CGROUP_SCHED */
8786 #ifdef CONFIG_CGROUP_CPUACCT
8789 * CPU accounting code for task groups.
8791 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8792 * (balbir@in.ibm.com).
8795 /* track cpu usage of a group of tasks and its child groups */
8797 struct cgroup_subsys_state css
;
8798 /* cpuusage holds pointer to a u64-type object on every cpu */
8800 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8801 struct cpuacct
*parent
;
8804 struct cgroup_subsys cpuacct_subsys
;
8806 /* return cpu accounting group corresponding to this container */
8807 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8809 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8810 struct cpuacct
, css
);
8813 /* return cpu accounting group to which this task belongs */
8814 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8816 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8817 struct cpuacct
, css
);
8820 /* create a new cpu accounting group */
8821 static struct cgroup_subsys_state
*cpuacct_create(
8822 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8824 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8830 ca
->cpuusage
= alloc_percpu(u64
);
8834 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8835 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8836 goto out_free_counters
;
8839 ca
->parent
= cgroup_ca(cgrp
->parent
);
8845 percpu_counter_destroy(&ca
->cpustat
[i
]);
8846 free_percpu(ca
->cpuusage
);
8850 return ERR_PTR(-ENOMEM
);
8853 /* destroy an existing cpu accounting group */
8855 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8857 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8860 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8861 percpu_counter_destroy(&ca
->cpustat
[i
]);
8862 free_percpu(ca
->cpuusage
);
8866 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8868 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8871 #ifndef CONFIG_64BIT
8873 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8875 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8877 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8885 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8887 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8889 #ifndef CONFIG_64BIT
8891 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8893 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8895 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8901 /* return total cpu usage (in nanoseconds) of a group */
8902 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8904 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8905 u64 totalcpuusage
= 0;
8908 for_each_present_cpu(i
)
8909 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8911 return totalcpuusage
;
8914 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8917 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8926 for_each_present_cpu(i
)
8927 cpuacct_cpuusage_write(ca
, i
, 0);
8933 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8936 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8940 for_each_present_cpu(i
) {
8941 percpu
= cpuacct_cpuusage_read(ca
, i
);
8942 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8944 seq_printf(m
, "\n");
8948 static const char *cpuacct_stat_desc
[] = {
8949 [CPUACCT_STAT_USER
] = "user",
8950 [CPUACCT_STAT_SYSTEM
] = "system",
8953 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8954 struct cgroup_map_cb
*cb
)
8956 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8959 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8960 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8961 val
= cputime64_to_clock_t(val
);
8962 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8967 static struct cftype files
[] = {
8970 .read_u64
= cpuusage_read
,
8971 .write_u64
= cpuusage_write
,
8974 .name
= "usage_percpu",
8975 .read_seq_string
= cpuacct_percpu_seq_read
,
8979 .read_map
= cpuacct_stats_show
,
8983 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8985 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8989 * charge this task's execution time to its accounting group.
8991 * called with rq->lock held.
8993 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8998 if (unlikely(!cpuacct_subsys
.active
))
9001 cpu
= task_cpu(tsk
);
9007 for (; ca
; ca
= ca
->parent
) {
9008 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
9009 *cpuusage
+= cputime
;
9016 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9017 * in cputime_t units. As a result, cpuacct_update_stats calls
9018 * percpu_counter_add with values large enough to always overflow the
9019 * per cpu batch limit causing bad SMP scalability.
9021 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9022 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9023 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9026 #define CPUACCT_BATCH \
9027 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9029 #define CPUACCT_BATCH 0
9033 * Charge the system/user time to the task's accounting group.
9035 static void cpuacct_update_stats(struct task_struct
*tsk
,
9036 enum cpuacct_stat_index idx
, cputime_t val
)
9039 int batch
= CPUACCT_BATCH
;
9041 if (unlikely(!cpuacct_subsys
.active
))
9048 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9054 struct cgroup_subsys cpuacct_subsys
= {
9056 .create
= cpuacct_create
,
9057 .destroy
= cpuacct_destroy
,
9058 .populate
= cpuacct_populate
,
9059 .subsys_id
= cpuacct_subsys_id
,
9061 #endif /* CONFIG_CGROUP_CPUACCT */
9065 int rcu_expedited_torture_stats(char *page
)
9069 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9071 void synchronize_sched_expedited(void)
9074 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9076 #else /* #ifndef CONFIG_SMP */
9078 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9079 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9081 #define RCU_EXPEDITED_STATE_POST -2
9082 #define RCU_EXPEDITED_STATE_IDLE -1
9084 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9086 int rcu_expedited_torture_stats(char *page
)
9091 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9092 for_each_online_cpu(cpu
) {
9093 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9094 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9096 cnt
+= sprintf(&page
[cnt
], "\n");
9099 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9101 static long synchronize_sched_expedited_count
;
9104 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9105 * approach to force grace period to end quickly. This consumes
9106 * significant time on all CPUs, and is thus not recommended for
9107 * any sort of common-case code.
9109 * Note that it is illegal to call this function while holding any
9110 * lock that is acquired by a CPU-hotplug notifier. Failing to
9111 * observe this restriction will result in deadlock.
9113 void synchronize_sched_expedited(void)
9116 unsigned long flags
;
9117 bool need_full_sync
= 0;
9119 struct migration_req
*req
;
9123 smp_mb(); /* ensure prior mod happens before capturing snap. */
9124 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9126 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9128 if (trycount
++ < 10)
9129 udelay(trycount
* num_online_cpus());
9131 synchronize_sched();
9134 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9135 smp_mb(); /* ensure test happens before caller kfree */
9140 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9141 for_each_online_cpu(cpu
) {
9143 req
= &per_cpu(rcu_migration_req
, cpu
);
9144 init_completion(&req
->done
);
9146 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9147 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9148 list_add(&req
->list
, &rq
->migration_queue
);
9149 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9150 wake_up_process(rq
->migration_thread
);
9152 for_each_online_cpu(cpu
) {
9153 rcu_expedited_state
= cpu
;
9154 req
= &per_cpu(rcu_migration_req
, cpu
);
9156 wait_for_completion(&req
->done
);
9157 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9158 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9160 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9161 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9163 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9164 synchronize_sched_expedited_count
++;
9165 mutex_unlock(&rcu_sched_expedited_mutex
);
9168 synchronize_sched();
9170 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9172 #endif /* #else #ifndef CONFIG_SMP */