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
];
495 unsigned char in_nohz_recently
;
497 /* capture load from *all* tasks on this cpu: */
498 struct load_weight load
;
499 unsigned long nr_load_updates
;
505 #ifdef CONFIG_FAIR_GROUP_SCHED
506 /* list of leaf cfs_rq on this cpu: */
507 struct list_head leaf_cfs_rq_list
;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 struct list_head leaf_rt_rq_list
;
514 * This is part of a global counter where only the total sum
515 * over all CPUs matters. A task can increase this counter on
516 * one CPU and if it got migrated afterwards it may decrease
517 * it on another CPU. Always updated under the runqueue lock:
519 unsigned long nr_uninterruptible
;
521 struct task_struct
*curr
, *idle
;
522 unsigned long next_balance
;
523 struct mm_struct
*prev_mm
;
530 struct root_domain
*rd
;
531 struct sched_domain
*sd
;
533 unsigned char idle_at_tick
;
534 /* For active balancing */
538 /* cpu of this runqueue: */
542 unsigned long avg_load_per_task
;
544 struct task_struct
*migration_thread
;
545 struct list_head migration_queue
;
553 /* calc_load related fields */
554 unsigned long calc_load_update
;
555 long calc_load_active
;
557 #ifdef CONFIG_SCHED_HRTICK
559 int hrtick_csd_pending
;
560 struct call_single_data hrtick_csd
;
562 struct hrtimer hrtick_timer
;
565 #ifdef CONFIG_SCHEDSTATS
567 struct sched_info rq_sched_info
;
568 unsigned long long rq_cpu_time
;
569 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
571 /* sys_sched_yield() stats */
572 unsigned int yld_count
;
574 /* schedule() stats */
575 unsigned int sched_switch
;
576 unsigned int sched_count
;
577 unsigned int sched_goidle
;
579 /* try_to_wake_up() stats */
580 unsigned int ttwu_count
;
581 unsigned int ttwu_local
;
584 unsigned int bkl_count
;
588 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
591 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
593 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
596 static inline int cpu_of(struct rq
*rq
)
605 #define rcu_dereference_check_sched_domain(p) \
606 rcu_dereference_check((p), \
607 rcu_read_lock_sched_held() || \
608 lockdep_is_held(&sched_domains_mutex))
611 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
612 * See detach_destroy_domains: synchronize_sched for details.
614 * The domain tree of any CPU may only be accessed from within
615 * preempt-disabled sections.
617 #define for_each_domain(cpu, __sd) \
618 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
620 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
621 #define this_rq() (&__get_cpu_var(runqueues))
622 #define task_rq(p) cpu_rq(task_cpu(p))
623 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
624 #define raw_rq() (&__raw_get_cpu_var(runqueues))
626 inline void update_rq_clock(struct rq
*rq
)
628 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
632 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
634 #ifdef CONFIG_SCHED_DEBUG
635 # define const_debug __read_mostly
637 # define const_debug static const
642 * @cpu: the processor in question.
644 * Returns true if the current cpu runqueue is locked.
645 * This interface allows printk to be called with the runqueue lock
646 * held and know whether or not it is OK to wake up the klogd.
648 int runqueue_is_locked(int cpu
)
650 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
654 * Debugging: various feature bits
657 #define SCHED_FEAT(name, enabled) \
658 __SCHED_FEAT_##name ,
661 #include "sched_features.h"
666 #define SCHED_FEAT(name, enabled) \
667 (1UL << __SCHED_FEAT_##name) * enabled |
669 const_debug
unsigned int sysctl_sched_features
=
670 #include "sched_features.h"
675 #ifdef CONFIG_SCHED_DEBUG
676 #define SCHED_FEAT(name, enabled) \
679 static __read_mostly
char *sched_feat_names
[] = {
680 #include "sched_features.h"
686 static int sched_feat_show(struct seq_file
*m
, void *v
)
690 for (i
= 0; sched_feat_names
[i
]; i
++) {
691 if (!(sysctl_sched_features
& (1UL << i
)))
693 seq_printf(m
, "%s ", sched_feat_names
[i
]);
701 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
702 size_t cnt
, loff_t
*ppos
)
712 if (copy_from_user(&buf
, ubuf
, cnt
))
717 if (strncmp(buf
, "NO_", 3) == 0) {
722 for (i
= 0; sched_feat_names
[i
]; i
++) {
723 int len
= strlen(sched_feat_names
[i
]);
725 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
727 sysctl_sched_features
&= ~(1UL << i
);
729 sysctl_sched_features
|= (1UL << i
);
734 if (!sched_feat_names
[i
])
742 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
744 return single_open(filp
, sched_feat_show
, NULL
);
747 static const struct file_operations sched_feat_fops
= {
748 .open
= sched_feat_open
,
749 .write
= sched_feat_write
,
752 .release
= single_release
,
755 static __init
int sched_init_debug(void)
757 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
762 late_initcall(sched_init_debug
);
766 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
769 * Number of tasks to iterate in a single balance run.
770 * Limited because this is done with IRQs disabled.
772 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
775 * ratelimit for updating the group shares.
778 unsigned int sysctl_sched_shares_ratelimit
= 250000;
779 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
782 * Inject some fuzzyness into changing the per-cpu group shares
783 * this avoids remote rq-locks at the expense of fairness.
786 unsigned int sysctl_sched_shares_thresh
= 4;
789 * period over which we average the RT time consumption, measured
794 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
797 * period over which we measure -rt task cpu usage in us.
800 unsigned int sysctl_sched_rt_period
= 1000000;
802 static __read_mostly
int scheduler_running
;
805 * part of the period that we allow rt tasks to run in us.
808 int sysctl_sched_rt_runtime
= 950000;
810 static inline u64
global_rt_period(void)
812 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
815 static inline u64
global_rt_runtime(void)
817 if (sysctl_sched_rt_runtime
< 0)
820 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
823 #ifndef prepare_arch_switch
824 # define prepare_arch_switch(next) do { } while (0)
826 #ifndef finish_arch_switch
827 # define finish_arch_switch(prev) do { } while (0)
830 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
832 return rq
->curr
== p
;
835 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
836 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
838 return task_current(rq
, p
);
841 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
845 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
847 #ifdef CONFIG_DEBUG_SPINLOCK
848 /* this is a valid case when another task releases the spinlock */
849 rq
->lock
.owner
= current
;
852 * If we are tracking spinlock dependencies then we have to
853 * fix up the runqueue lock - which gets 'carried over' from
856 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
858 raw_spin_unlock_irq(&rq
->lock
);
861 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
862 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
867 return task_current(rq
, p
);
871 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
875 * We can optimise this out completely for !SMP, because the
876 * SMP rebalancing from interrupt is the only thing that cares
881 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
882 raw_spin_unlock_irq(&rq
->lock
);
884 raw_spin_unlock(&rq
->lock
);
888 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
892 * After ->oncpu is cleared, the task can be moved to a different CPU.
893 * We must ensure this doesn't happen until the switch is completely
899 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
903 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
906 * Check whether the task is waking, we use this to synchronize against
907 * ttwu() so that task_cpu() reports a stable number.
909 * We need to make an exception for PF_STARTING tasks because the fork
910 * path might require task_rq_lock() to work, eg. it can call
911 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
913 static inline int task_is_waking(struct task_struct
*p
)
915 return unlikely((p
->state
== TASK_WAKING
) && !(p
->flags
& PF_STARTING
));
919 * __task_rq_lock - lock the runqueue a given task resides on.
920 * Must be called interrupts disabled.
922 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
928 while (task_is_waking(p
))
931 raw_spin_lock(&rq
->lock
);
932 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
934 raw_spin_unlock(&rq
->lock
);
939 * task_rq_lock - lock the runqueue a given task resides on and disable
940 * interrupts. Note the ordering: we can safely lookup the task_rq without
941 * explicitly disabling preemption.
943 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
949 while (task_is_waking(p
))
951 local_irq_save(*flags
);
953 raw_spin_lock(&rq
->lock
);
954 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
956 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
960 void task_rq_unlock_wait(struct task_struct
*p
)
962 struct rq
*rq
= task_rq(p
);
964 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
965 raw_spin_unlock_wait(&rq
->lock
);
968 static void __task_rq_unlock(struct rq
*rq
)
971 raw_spin_unlock(&rq
->lock
);
974 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
977 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
981 * this_rq_lock - lock this runqueue and disable interrupts.
983 static struct rq
*this_rq_lock(void)
990 raw_spin_lock(&rq
->lock
);
995 #ifdef CONFIG_SCHED_HRTICK
997 * Use HR-timers to deliver accurate preemption points.
999 * Its all a bit involved since we cannot program an hrt while holding the
1000 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1003 * When we get rescheduled we reprogram the hrtick_timer outside of the
1009 * - enabled by features
1010 * - hrtimer is actually high res
1012 static inline int hrtick_enabled(struct rq
*rq
)
1014 if (!sched_feat(HRTICK
))
1016 if (!cpu_active(cpu_of(rq
)))
1018 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1021 static void hrtick_clear(struct rq
*rq
)
1023 if (hrtimer_active(&rq
->hrtick_timer
))
1024 hrtimer_cancel(&rq
->hrtick_timer
);
1028 * High-resolution timer tick.
1029 * Runs from hardirq context with interrupts disabled.
1031 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1033 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1035 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1037 raw_spin_lock(&rq
->lock
);
1038 update_rq_clock(rq
);
1039 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1040 raw_spin_unlock(&rq
->lock
);
1042 return HRTIMER_NORESTART
;
1047 * called from hardirq (IPI) context
1049 static void __hrtick_start(void *arg
)
1051 struct rq
*rq
= arg
;
1053 raw_spin_lock(&rq
->lock
);
1054 hrtimer_restart(&rq
->hrtick_timer
);
1055 rq
->hrtick_csd_pending
= 0;
1056 raw_spin_unlock(&rq
->lock
);
1060 * Called to set the hrtick timer state.
1062 * called with rq->lock held and irqs disabled
1064 static void hrtick_start(struct rq
*rq
, u64 delay
)
1066 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1067 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1069 hrtimer_set_expires(timer
, time
);
1071 if (rq
== this_rq()) {
1072 hrtimer_restart(timer
);
1073 } else if (!rq
->hrtick_csd_pending
) {
1074 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1075 rq
->hrtick_csd_pending
= 1;
1080 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1082 int cpu
= (int)(long)hcpu
;
1085 case CPU_UP_CANCELED
:
1086 case CPU_UP_CANCELED_FROZEN
:
1087 case CPU_DOWN_PREPARE
:
1088 case CPU_DOWN_PREPARE_FROZEN
:
1090 case CPU_DEAD_FROZEN
:
1091 hrtick_clear(cpu_rq(cpu
));
1098 static __init
void init_hrtick(void)
1100 hotcpu_notifier(hotplug_hrtick
, 0);
1104 * Called to set the hrtick timer state.
1106 * called with rq->lock held and irqs disabled
1108 static void hrtick_start(struct rq
*rq
, u64 delay
)
1110 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1111 HRTIMER_MODE_REL_PINNED
, 0);
1114 static inline void init_hrtick(void)
1117 #endif /* CONFIG_SMP */
1119 static void init_rq_hrtick(struct rq
*rq
)
1122 rq
->hrtick_csd_pending
= 0;
1124 rq
->hrtick_csd
.flags
= 0;
1125 rq
->hrtick_csd
.func
= __hrtick_start
;
1126 rq
->hrtick_csd
.info
= rq
;
1129 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1130 rq
->hrtick_timer
.function
= hrtick
;
1132 #else /* CONFIG_SCHED_HRTICK */
1133 static inline void hrtick_clear(struct rq
*rq
)
1137 static inline void init_rq_hrtick(struct rq
*rq
)
1141 static inline void init_hrtick(void)
1144 #endif /* CONFIG_SCHED_HRTICK */
1147 * resched_task - mark a task 'to be rescheduled now'.
1149 * On UP this means the setting of the need_resched flag, on SMP it
1150 * might also involve a cross-CPU call to trigger the scheduler on
1155 #ifndef tsk_is_polling
1156 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1159 static void resched_task(struct task_struct
*p
)
1163 assert_raw_spin_locked(&task_rq(p
)->lock
);
1165 if (test_tsk_need_resched(p
))
1168 set_tsk_need_resched(p
);
1171 if (cpu
== smp_processor_id())
1174 /* NEED_RESCHED must be visible before we test polling */
1176 if (!tsk_is_polling(p
))
1177 smp_send_reschedule(cpu
);
1180 static void resched_cpu(int cpu
)
1182 struct rq
*rq
= cpu_rq(cpu
);
1183 unsigned long flags
;
1185 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1187 resched_task(cpu_curr(cpu
));
1188 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1193 * When add_timer_on() enqueues a timer into the timer wheel of an
1194 * idle CPU then this timer might expire before the next timer event
1195 * which is scheduled to wake up that CPU. In case of a completely
1196 * idle system the next event might even be infinite time into the
1197 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1198 * leaves the inner idle loop so the newly added timer is taken into
1199 * account when the CPU goes back to idle and evaluates the timer
1200 * wheel for the next timer event.
1202 void wake_up_idle_cpu(int cpu
)
1204 struct rq
*rq
= cpu_rq(cpu
);
1206 if (cpu
== smp_processor_id())
1210 * This is safe, as this function is called with the timer
1211 * wheel base lock of (cpu) held. When the CPU is on the way
1212 * to idle and has not yet set rq->curr to idle then it will
1213 * be serialized on the timer wheel base lock and take the new
1214 * timer into account automatically.
1216 if (rq
->curr
!= rq
->idle
)
1220 * We can set TIF_RESCHED on the idle task of the other CPU
1221 * lockless. The worst case is that the other CPU runs the
1222 * idle task through an additional NOOP schedule()
1224 set_tsk_need_resched(rq
->idle
);
1226 /* NEED_RESCHED must be visible before we test polling */
1228 if (!tsk_is_polling(rq
->idle
))
1229 smp_send_reschedule(cpu
);
1231 #endif /* CONFIG_NO_HZ */
1233 static u64
sched_avg_period(void)
1235 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1238 static void sched_avg_update(struct rq
*rq
)
1240 s64 period
= sched_avg_period();
1242 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1243 rq
->age_stamp
+= period
;
1248 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1250 rq
->rt_avg
+= rt_delta
;
1251 sched_avg_update(rq
);
1254 #else /* !CONFIG_SMP */
1255 static void resched_task(struct task_struct
*p
)
1257 assert_raw_spin_locked(&task_rq(p
)->lock
);
1258 set_tsk_need_resched(p
);
1261 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1264 #endif /* CONFIG_SMP */
1266 #if BITS_PER_LONG == 32
1267 # define WMULT_CONST (~0UL)
1269 # define WMULT_CONST (1UL << 32)
1272 #define WMULT_SHIFT 32
1275 * Shift right and round:
1277 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1280 * delta *= weight / lw
1282 static unsigned long
1283 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1284 struct load_weight
*lw
)
1288 if (!lw
->inv_weight
) {
1289 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1292 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1296 tmp
= (u64
)delta_exec
* weight
;
1298 * Check whether we'd overflow the 64-bit multiplication:
1300 if (unlikely(tmp
> WMULT_CONST
))
1301 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1304 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1306 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1309 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1315 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1322 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1323 * of tasks with abnormal "nice" values across CPUs the contribution that
1324 * each task makes to its run queue's load is weighted according to its
1325 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1326 * scaled version of the new time slice allocation that they receive on time
1330 #define WEIGHT_IDLEPRIO 3
1331 #define WMULT_IDLEPRIO 1431655765
1334 * Nice levels are multiplicative, with a gentle 10% change for every
1335 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1336 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1337 * that remained on nice 0.
1339 * The "10% effect" is relative and cumulative: from _any_ nice level,
1340 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1341 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1342 * If a task goes up by ~10% and another task goes down by ~10% then
1343 * the relative distance between them is ~25%.)
1345 static const int prio_to_weight
[40] = {
1346 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1347 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1348 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1349 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1350 /* 0 */ 1024, 820, 655, 526, 423,
1351 /* 5 */ 335, 272, 215, 172, 137,
1352 /* 10 */ 110, 87, 70, 56, 45,
1353 /* 15 */ 36, 29, 23, 18, 15,
1357 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1359 * In cases where the weight does not change often, we can use the
1360 * precalculated inverse to speed up arithmetics by turning divisions
1361 * into multiplications:
1363 static const u32 prio_to_wmult
[40] = {
1364 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1365 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1366 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1367 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1368 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1369 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1370 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1371 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1374 /* Time spent by the tasks of the cpu accounting group executing in ... */
1375 enum cpuacct_stat_index
{
1376 CPUACCT_STAT_USER
, /* ... user mode */
1377 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1379 CPUACCT_STAT_NSTATS
,
1382 #ifdef CONFIG_CGROUP_CPUACCT
1383 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1384 static void cpuacct_update_stats(struct task_struct
*tsk
,
1385 enum cpuacct_stat_index idx
, cputime_t val
);
1387 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1388 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1389 enum cpuacct_stat_index idx
, cputime_t val
) {}
1392 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1394 update_load_add(&rq
->load
, load
);
1397 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1399 update_load_sub(&rq
->load
, load
);
1402 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1403 typedef int (*tg_visitor
)(struct task_group
*, void *);
1406 * Iterate the full tree, calling @down when first entering a node and @up when
1407 * leaving it for the final time.
1409 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1411 struct task_group
*parent
, *child
;
1415 parent
= &root_task_group
;
1417 ret
= (*down
)(parent
, data
);
1420 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1427 ret
= (*up
)(parent
, data
);
1432 parent
= parent
->parent
;
1441 static int tg_nop(struct task_group
*tg
, void *data
)
1448 /* Used instead of source_load when we know the type == 0 */
1449 static unsigned long weighted_cpuload(const int cpu
)
1451 return cpu_rq(cpu
)->load
.weight
;
1455 * Return a low guess at the load of a migration-source cpu weighted
1456 * according to the scheduling class and "nice" value.
1458 * We want to under-estimate the load of migration sources, to
1459 * balance conservatively.
1461 static unsigned long source_load(int cpu
, int type
)
1463 struct rq
*rq
= cpu_rq(cpu
);
1464 unsigned long total
= weighted_cpuload(cpu
);
1466 if (type
== 0 || !sched_feat(LB_BIAS
))
1469 return min(rq
->cpu_load
[type
-1], total
);
1473 * Return a high guess at the load of a migration-target cpu weighted
1474 * according to the scheduling class and "nice" value.
1476 static unsigned long target_load(int cpu
, int type
)
1478 struct rq
*rq
= cpu_rq(cpu
);
1479 unsigned long total
= weighted_cpuload(cpu
);
1481 if (type
== 0 || !sched_feat(LB_BIAS
))
1484 return max(rq
->cpu_load
[type
-1], total
);
1487 static struct sched_group
*group_of(int cpu
)
1489 struct sched_domain
*sd
= rcu_dereference_sched(cpu_rq(cpu
)->sd
);
1497 static unsigned long power_of(int cpu
)
1499 struct sched_group
*group
= group_of(cpu
);
1502 return SCHED_LOAD_SCALE
;
1504 return group
->cpu_power
;
1507 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1509 static unsigned long cpu_avg_load_per_task(int cpu
)
1511 struct rq
*rq
= cpu_rq(cpu
);
1512 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1515 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1517 rq
->avg_load_per_task
= 0;
1519 return rq
->avg_load_per_task
;
1522 #ifdef CONFIG_FAIR_GROUP_SCHED
1524 static __read_mostly
unsigned long *update_shares_data
;
1526 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1529 * Calculate and set the cpu's group shares.
1531 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1532 unsigned long sd_shares
,
1533 unsigned long sd_rq_weight
,
1534 unsigned long *usd_rq_weight
)
1536 unsigned long shares
, rq_weight
;
1539 rq_weight
= usd_rq_weight
[cpu
];
1542 rq_weight
= NICE_0_LOAD
;
1546 * \Sum_j shares_j * rq_weight_i
1547 * shares_i = -----------------------------
1548 * \Sum_j rq_weight_j
1550 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1551 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1553 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1554 sysctl_sched_shares_thresh
) {
1555 struct rq
*rq
= cpu_rq(cpu
);
1556 unsigned long flags
;
1558 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1559 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1560 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1561 __set_se_shares(tg
->se
[cpu
], shares
);
1562 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1567 * Re-compute the task group their per cpu shares over the given domain.
1568 * This needs to be done in a bottom-up fashion because the rq weight of a
1569 * parent group depends on the shares of its child groups.
1571 static int tg_shares_up(struct task_group
*tg
, void *data
)
1573 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1574 unsigned long *usd_rq_weight
;
1575 struct sched_domain
*sd
= data
;
1576 unsigned long flags
;
1582 local_irq_save(flags
);
1583 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1585 for_each_cpu(i
, sched_domain_span(sd
)) {
1586 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1587 usd_rq_weight
[i
] = weight
;
1589 rq_weight
+= weight
;
1591 * If there are currently no tasks on the cpu pretend there
1592 * is one of average load so that when a new task gets to
1593 * run here it will not get delayed by group starvation.
1596 weight
= NICE_0_LOAD
;
1598 sum_weight
+= weight
;
1599 shares
+= tg
->cfs_rq
[i
]->shares
;
1603 rq_weight
= sum_weight
;
1605 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1606 shares
= tg
->shares
;
1608 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1609 shares
= tg
->shares
;
1611 for_each_cpu(i
, sched_domain_span(sd
))
1612 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1614 local_irq_restore(flags
);
1620 * Compute the cpu's hierarchical load factor for each task group.
1621 * This needs to be done in a top-down fashion because the load of a child
1622 * group is a fraction of its parents load.
1624 static int tg_load_down(struct task_group
*tg
, void *data
)
1627 long cpu
= (long)data
;
1630 load
= cpu_rq(cpu
)->load
.weight
;
1632 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1633 load
*= tg
->cfs_rq
[cpu
]->shares
;
1634 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1637 tg
->cfs_rq
[cpu
]->h_load
= load
;
1642 static void update_shares(struct sched_domain
*sd
)
1647 if (root_task_group_empty())
1650 now
= cpu_clock(raw_smp_processor_id());
1651 elapsed
= now
- sd
->last_update
;
1653 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1654 sd
->last_update
= now
;
1655 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1659 static void update_h_load(long cpu
)
1661 if (root_task_group_empty())
1664 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1669 static inline void update_shares(struct sched_domain
*sd
)
1675 #ifdef CONFIG_PREEMPT
1677 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1680 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1681 * way at the expense of forcing extra atomic operations in all
1682 * invocations. This assures that the double_lock is acquired using the
1683 * same underlying policy as the spinlock_t on this architecture, which
1684 * reduces latency compared to the unfair variant below. However, it
1685 * also adds more overhead and therefore may reduce throughput.
1687 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1688 __releases(this_rq
->lock
)
1689 __acquires(busiest
->lock
)
1690 __acquires(this_rq
->lock
)
1692 raw_spin_unlock(&this_rq
->lock
);
1693 double_rq_lock(this_rq
, busiest
);
1700 * Unfair double_lock_balance: Optimizes throughput at the expense of
1701 * latency by eliminating extra atomic operations when the locks are
1702 * already in proper order on entry. This favors lower cpu-ids and will
1703 * grant the double lock to lower cpus over higher ids under contention,
1704 * regardless of entry order into the function.
1706 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1707 __releases(this_rq
->lock
)
1708 __acquires(busiest
->lock
)
1709 __acquires(this_rq
->lock
)
1713 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1714 if (busiest
< this_rq
) {
1715 raw_spin_unlock(&this_rq
->lock
);
1716 raw_spin_lock(&busiest
->lock
);
1717 raw_spin_lock_nested(&this_rq
->lock
,
1718 SINGLE_DEPTH_NESTING
);
1721 raw_spin_lock_nested(&busiest
->lock
,
1722 SINGLE_DEPTH_NESTING
);
1727 #endif /* CONFIG_PREEMPT */
1730 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1732 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1734 if (unlikely(!irqs_disabled())) {
1735 /* printk() doesn't work good under rq->lock */
1736 raw_spin_unlock(&this_rq
->lock
);
1740 return _double_lock_balance(this_rq
, busiest
);
1743 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1744 __releases(busiest
->lock
)
1746 raw_spin_unlock(&busiest
->lock
);
1747 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1751 * double_rq_lock - safely lock two runqueues
1753 * Note this does not disable interrupts like task_rq_lock,
1754 * you need to do so manually before calling.
1756 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1757 __acquires(rq1
->lock
)
1758 __acquires(rq2
->lock
)
1760 BUG_ON(!irqs_disabled());
1762 raw_spin_lock(&rq1
->lock
);
1763 __acquire(rq2
->lock
); /* Fake it out ;) */
1766 raw_spin_lock(&rq1
->lock
);
1767 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1769 raw_spin_lock(&rq2
->lock
);
1770 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1773 update_rq_clock(rq1
);
1774 update_rq_clock(rq2
);
1778 * double_rq_unlock - safely unlock two runqueues
1780 * Note this does not restore interrupts like task_rq_unlock,
1781 * you need to do so manually after calling.
1783 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1784 __releases(rq1
->lock
)
1785 __releases(rq2
->lock
)
1787 raw_spin_unlock(&rq1
->lock
);
1789 raw_spin_unlock(&rq2
->lock
);
1791 __release(rq2
->lock
);
1796 #ifdef CONFIG_FAIR_GROUP_SCHED
1797 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1800 cfs_rq
->shares
= shares
;
1805 static void calc_load_account_active(struct rq
*this_rq
);
1806 static void update_sysctl(void);
1807 static int get_update_sysctl_factor(void);
1809 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1811 set_task_rq(p
, cpu
);
1814 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1815 * successfuly executed on another CPU. We must ensure that updates of
1816 * per-task data have been completed by this moment.
1819 task_thread_info(p
)->cpu
= cpu
;
1823 static const struct sched_class rt_sched_class
;
1825 #define sched_class_highest (&rt_sched_class)
1826 #define for_each_class(class) \
1827 for (class = sched_class_highest; class; class = class->next)
1829 #include "sched_stats.h"
1831 static void inc_nr_running(struct rq
*rq
)
1836 static void dec_nr_running(struct rq
*rq
)
1841 static void set_load_weight(struct task_struct
*p
)
1843 if (task_has_rt_policy(p
)) {
1844 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1845 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1850 * SCHED_IDLE tasks get minimal weight:
1852 if (p
->policy
== SCHED_IDLE
) {
1853 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1854 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1858 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1859 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1862 static void update_avg(u64
*avg
, u64 sample
)
1864 s64 diff
= sample
- *avg
;
1869 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1872 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1874 sched_info_queued(p
);
1875 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1879 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1882 if (p
->se
.last_wakeup
) {
1883 update_avg(&p
->se
.avg_overlap
,
1884 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1885 p
->se
.last_wakeup
= 0;
1887 update_avg(&p
->se
.avg_wakeup
,
1888 sysctl_sched_wakeup_granularity
);
1892 sched_info_dequeued(p
);
1893 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1898 * activate_task - move a task to the runqueue.
1900 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1902 if (task_contributes_to_load(p
))
1903 rq
->nr_uninterruptible
--;
1905 enqueue_task(rq
, p
, wakeup
, false);
1910 * deactivate_task - remove a task from the runqueue.
1912 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1914 if (task_contributes_to_load(p
))
1915 rq
->nr_uninterruptible
++;
1917 dequeue_task(rq
, p
, sleep
);
1921 #include "sched_idletask.c"
1922 #include "sched_fair.c"
1923 #include "sched_rt.c"
1924 #ifdef CONFIG_SCHED_DEBUG
1925 # include "sched_debug.c"
1929 * __normal_prio - return the priority that is based on the static prio
1931 static inline int __normal_prio(struct task_struct
*p
)
1933 return p
->static_prio
;
1937 * Calculate the expected normal priority: i.e. priority
1938 * without taking RT-inheritance into account. Might be
1939 * boosted by interactivity modifiers. Changes upon fork,
1940 * setprio syscalls, and whenever the interactivity
1941 * estimator recalculates.
1943 static inline int normal_prio(struct task_struct
*p
)
1947 if (task_has_rt_policy(p
))
1948 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1950 prio
= __normal_prio(p
);
1955 * Calculate the current priority, i.e. the priority
1956 * taken into account by the scheduler. This value might
1957 * be boosted by RT tasks, or might be boosted by
1958 * interactivity modifiers. Will be RT if the task got
1959 * RT-boosted. If not then it returns p->normal_prio.
1961 static int effective_prio(struct task_struct
*p
)
1963 p
->normal_prio
= normal_prio(p
);
1965 * If we are RT tasks or we were boosted to RT priority,
1966 * keep the priority unchanged. Otherwise, update priority
1967 * to the normal priority:
1969 if (!rt_prio(p
->prio
))
1970 return p
->normal_prio
;
1975 * task_curr - is this task currently executing on a CPU?
1976 * @p: the task in question.
1978 inline int task_curr(const struct task_struct
*p
)
1980 return cpu_curr(task_cpu(p
)) == p
;
1983 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1984 const struct sched_class
*prev_class
,
1985 int oldprio
, int running
)
1987 if (prev_class
!= p
->sched_class
) {
1988 if (prev_class
->switched_from
)
1989 prev_class
->switched_from(rq
, p
, running
);
1990 p
->sched_class
->switched_to(rq
, p
, running
);
1992 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1997 * Is this task likely cache-hot:
2000 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2004 if (p
->sched_class
!= &fair_sched_class
)
2008 * Buddy candidates are cache hot:
2010 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2011 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2012 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2015 if (sysctl_sched_migration_cost
== -1)
2017 if (sysctl_sched_migration_cost
== 0)
2020 delta
= now
- p
->se
.exec_start
;
2022 return delta
< (s64
)sysctl_sched_migration_cost
;
2025 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2027 #ifdef CONFIG_SCHED_DEBUG
2029 * We should never call set_task_cpu() on a blocked task,
2030 * ttwu() will sort out the placement.
2032 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2033 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2036 trace_sched_migrate_task(p
, new_cpu
);
2038 if (task_cpu(p
) != new_cpu
) {
2039 p
->se
.nr_migrations
++;
2040 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2043 __set_task_cpu(p
, new_cpu
);
2046 struct migration_req
{
2047 struct list_head list
;
2049 struct task_struct
*task
;
2052 struct completion done
;
2056 * The task's runqueue lock must be held.
2057 * Returns true if you have to wait for migration thread.
2060 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2062 struct rq
*rq
= task_rq(p
);
2065 * If the task is not on a runqueue (and not running), then
2066 * the next wake-up will properly place the task.
2068 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2071 init_completion(&req
->done
);
2073 req
->dest_cpu
= dest_cpu
;
2074 list_add(&req
->list
, &rq
->migration_queue
);
2080 * wait_task_context_switch - wait for a thread to complete at least one
2083 * @p must not be current.
2085 void wait_task_context_switch(struct task_struct
*p
)
2087 unsigned long nvcsw
, nivcsw
, flags
;
2095 * The runqueue is assigned before the actual context
2096 * switch. We need to take the runqueue lock.
2098 * We could check initially without the lock but it is
2099 * very likely that we need to take the lock in every
2102 rq
= task_rq_lock(p
, &flags
);
2103 running
= task_running(rq
, p
);
2104 task_rq_unlock(rq
, &flags
);
2106 if (likely(!running
))
2109 * The switch count is incremented before the actual
2110 * context switch. We thus wait for two switches to be
2111 * sure at least one completed.
2113 if ((p
->nvcsw
- nvcsw
) > 1)
2115 if ((p
->nivcsw
- nivcsw
) > 1)
2123 * wait_task_inactive - wait for a thread to unschedule.
2125 * If @match_state is nonzero, it's the @p->state value just checked and
2126 * not expected to change. If it changes, i.e. @p might have woken up,
2127 * then return zero. When we succeed in waiting for @p to be off its CPU,
2128 * we return a positive number (its total switch count). If a second call
2129 * a short while later returns the same number, the caller can be sure that
2130 * @p has remained unscheduled the whole time.
2132 * The caller must ensure that the task *will* unschedule sometime soon,
2133 * else this function might spin for a *long* time. This function can't
2134 * be called with interrupts off, or it may introduce deadlock with
2135 * smp_call_function() if an IPI is sent by the same process we are
2136 * waiting to become inactive.
2138 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2140 unsigned long flags
;
2147 * We do the initial early heuristics without holding
2148 * any task-queue locks at all. We'll only try to get
2149 * the runqueue lock when things look like they will
2155 * If the task is actively running on another CPU
2156 * still, just relax and busy-wait without holding
2159 * NOTE! Since we don't hold any locks, it's not
2160 * even sure that "rq" stays as the right runqueue!
2161 * But we don't care, since "task_running()" will
2162 * return false if the runqueue has changed and p
2163 * is actually now running somewhere else!
2165 while (task_running(rq
, p
)) {
2166 if (match_state
&& unlikely(p
->state
!= match_state
))
2172 * Ok, time to look more closely! We need the rq
2173 * lock now, to be *sure*. If we're wrong, we'll
2174 * just go back and repeat.
2176 rq
= task_rq_lock(p
, &flags
);
2177 trace_sched_wait_task(rq
, p
);
2178 running
= task_running(rq
, p
);
2179 on_rq
= p
->se
.on_rq
;
2181 if (!match_state
|| p
->state
== match_state
)
2182 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2183 task_rq_unlock(rq
, &flags
);
2186 * If it changed from the expected state, bail out now.
2188 if (unlikely(!ncsw
))
2192 * Was it really running after all now that we
2193 * checked with the proper locks actually held?
2195 * Oops. Go back and try again..
2197 if (unlikely(running
)) {
2203 * It's not enough that it's not actively running,
2204 * it must be off the runqueue _entirely_, and not
2207 * So if it was still runnable (but just not actively
2208 * running right now), it's preempted, and we should
2209 * yield - it could be a while.
2211 if (unlikely(on_rq
)) {
2212 schedule_timeout_uninterruptible(1);
2217 * Ahh, all good. It wasn't running, and it wasn't
2218 * runnable, which means that it will never become
2219 * running in the future either. We're all done!
2228 * kick_process - kick a running thread to enter/exit the kernel
2229 * @p: the to-be-kicked thread
2231 * Cause a process which is running on another CPU to enter
2232 * kernel-mode, without any delay. (to get signals handled.)
2234 * NOTE: this function doesnt have to take the runqueue lock,
2235 * because all it wants to ensure is that the remote task enters
2236 * the kernel. If the IPI races and the task has been migrated
2237 * to another CPU then no harm is done and the purpose has been
2240 void kick_process(struct task_struct
*p
)
2246 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2247 smp_send_reschedule(cpu
);
2250 EXPORT_SYMBOL_GPL(kick_process
);
2251 #endif /* CONFIG_SMP */
2254 * task_oncpu_function_call - call a function on the cpu on which a task runs
2255 * @p: the task to evaluate
2256 * @func: the function to be called
2257 * @info: the function call argument
2259 * Calls the function @func when the task is currently running. This might
2260 * be on the current CPU, which just calls the function directly
2262 void task_oncpu_function_call(struct task_struct
*p
,
2263 void (*func
) (void *info
), void *info
)
2270 smp_call_function_single(cpu
, func
, info
, 1);
2275 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2278 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2280 /* Look for allowed, online CPU in same node. */
2281 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2282 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2285 /* Any allowed, online CPU? */
2286 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2287 if (dest_cpu
< nr_cpu_ids
)
2290 /* No more Mr. Nice Guy. */
2291 if (dest_cpu
>= nr_cpu_ids
) {
2293 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
2295 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
2298 * Don't tell them about moving exiting tasks or
2299 * kernel threads (both mm NULL), since they never
2302 if (p
->mm
&& printk_ratelimit()) {
2303 printk(KERN_INFO
"process %d (%s) no "
2304 "longer affine to cpu%d\n",
2305 task_pid_nr(p
), p
->comm
, cpu
);
2313 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2314 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2317 * exec: is unstable, retry loop
2318 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2321 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2323 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2326 * In order not to call set_task_cpu() on a blocking task we need
2327 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2330 * Since this is common to all placement strategies, this lives here.
2332 * [ this allows ->select_task() to simply return task_cpu(p) and
2333 * not worry about this generic constraint ]
2335 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2337 cpu
= select_fallback_rq(task_cpu(p
), p
);
2344 * try_to_wake_up - wake up a thread
2345 * @p: the to-be-woken-up thread
2346 * @state: the mask of task states that can be woken
2347 * @sync: do a synchronous wakeup?
2349 * Put it on the run-queue if it's not already there. The "current"
2350 * thread is always on the run-queue (except when the actual
2351 * re-schedule is in progress), and as such you're allowed to do
2352 * the simpler "current->state = TASK_RUNNING" to mark yourself
2353 * runnable without the overhead of this.
2355 * returns failure only if the task is already active.
2357 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2360 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2361 unsigned long flags
;
2364 if (!sched_feat(SYNC_WAKEUPS
))
2365 wake_flags
&= ~WF_SYNC
;
2367 this_cpu
= get_cpu();
2370 rq
= task_rq_lock(p
, &flags
);
2371 update_rq_clock(rq
);
2372 if (!(p
->state
& state
))
2382 if (unlikely(task_running(rq
, p
)))
2386 * In order to handle concurrent wakeups and release the rq->lock
2387 * we put the task in TASK_WAKING state.
2389 * First fix up the nr_uninterruptible count:
2391 if (task_contributes_to_load(p
))
2392 rq
->nr_uninterruptible
--;
2393 p
->state
= TASK_WAKING
;
2395 if (p
->sched_class
->task_waking
)
2396 p
->sched_class
->task_waking(rq
, p
);
2398 __task_rq_unlock(rq
);
2400 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2401 if (cpu
!= orig_cpu
) {
2403 * Since we migrate the task without holding any rq->lock,
2404 * we need to be careful with task_rq_lock(), since that
2405 * might end up locking an invalid rq.
2407 set_task_cpu(p
, cpu
);
2411 raw_spin_lock(&rq
->lock
);
2412 update_rq_clock(rq
);
2415 * We migrated the task without holding either rq->lock, however
2416 * since the task is not on the task list itself, nobody else
2417 * will try and migrate the task, hence the rq should match the
2418 * cpu we just moved it to.
2420 WARN_ON(task_cpu(p
) != cpu
);
2421 WARN_ON(p
->state
!= TASK_WAKING
);
2423 #ifdef CONFIG_SCHEDSTATS
2424 schedstat_inc(rq
, ttwu_count
);
2425 if (cpu
== this_cpu
)
2426 schedstat_inc(rq
, ttwu_local
);
2428 struct sched_domain
*sd
;
2429 for_each_domain(this_cpu
, sd
) {
2430 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2431 schedstat_inc(sd
, ttwu_wake_remote
);
2436 #endif /* CONFIG_SCHEDSTATS */
2439 #endif /* CONFIG_SMP */
2440 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2441 if (wake_flags
& WF_SYNC
)
2442 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2443 if (orig_cpu
!= cpu
)
2444 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2445 if (cpu
== this_cpu
)
2446 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2448 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2449 activate_task(rq
, p
, 1);
2453 * Only attribute actual wakeups done by this task.
2455 if (!in_interrupt()) {
2456 struct sched_entity
*se
= ¤t
->se
;
2457 u64 sample
= se
->sum_exec_runtime
;
2459 if (se
->last_wakeup
)
2460 sample
-= se
->last_wakeup
;
2462 sample
-= se
->start_runtime
;
2463 update_avg(&se
->avg_wakeup
, sample
);
2465 se
->last_wakeup
= se
->sum_exec_runtime
;
2469 trace_sched_wakeup(rq
, p
, success
);
2470 check_preempt_curr(rq
, p
, wake_flags
);
2472 p
->state
= TASK_RUNNING
;
2474 if (p
->sched_class
->task_woken
)
2475 p
->sched_class
->task_woken(rq
, p
);
2477 if (unlikely(rq
->idle_stamp
)) {
2478 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2479 u64 max
= 2*sysctl_sched_migration_cost
;
2484 update_avg(&rq
->avg_idle
, delta
);
2489 task_rq_unlock(rq
, &flags
);
2496 * wake_up_process - Wake up a specific process
2497 * @p: The process to be woken up.
2499 * Attempt to wake up the nominated process and move it to the set of runnable
2500 * processes. Returns 1 if the process was woken up, 0 if it was already
2503 * It may be assumed that this function implies a write memory barrier before
2504 * changing the task state if and only if any tasks are woken up.
2506 int wake_up_process(struct task_struct
*p
)
2508 return try_to_wake_up(p
, TASK_ALL
, 0);
2510 EXPORT_SYMBOL(wake_up_process
);
2512 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2514 return try_to_wake_up(p
, state
, 0);
2518 * Perform scheduler related setup for a newly forked process p.
2519 * p is forked by current.
2521 * __sched_fork() is basic setup used by init_idle() too:
2523 static void __sched_fork(struct task_struct
*p
)
2525 p
->se
.exec_start
= 0;
2526 p
->se
.sum_exec_runtime
= 0;
2527 p
->se
.prev_sum_exec_runtime
= 0;
2528 p
->se
.nr_migrations
= 0;
2529 p
->se
.last_wakeup
= 0;
2530 p
->se
.avg_overlap
= 0;
2531 p
->se
.start_runtime
= 0;
2532 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2534 #ifdef CONFIG_SCHEDSTATS
2535 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2538 INIT_LIST_HEAD(&p
->rt
.run_list
);
2540 INIT_LIST_HEAD(&p
->se
.group_node
);
2542 #ifdef CONFIG_PREEMPT_NOTIFIERS
2543 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2548 * fork()/clone()-time setup:
2550 void sched_fork(struct task_struct
*p
, int clone_flags
)
2552 int cpu
= get_cpu();
2556 * We mark the process as waking here. This guarantees that
2557 * nobody will actually run it, and a signal or other external
2558 * event cannot wake it up and insert it on the runqueue either.
2560 p
->state
= TASK_WAKING
;
2563 * Revert to default priority/policy on fork if requested.
2565 if (unlikely(p
->sched_reset_on_fork
)) {
2566 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2567 p
->policy
= SCHED_NORMAL
;
2568 p
->normal_prio
= p
->static_prio
;
2571 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2572 p
->static_prio
= NICE_TO_PRIO(0);
2573 p
->normal_prio
= p
->static_prio
;
2578 * We don't need the reset flag anymore after the fork. It has
2579 * fulfilled its duty:
2581 p
->sched_reset_on_fork
= 0;
2585 * Make sure we do not leak PI boosting priority to the child.
2587 p
->prio
= current
->normal_prio
;
2589 if (!rt_prio(p
->prio
))
2590 p
->sched_class
= &fair_sched_class
;
2592 if (p
->sched_class
->task_fork
)
2593 p
->sched_class
->task_fork(p
);
2595 set_task_cpu(p
, cpu
);
2597 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2598 if (likely(sched_info_on()))
2599 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2601 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2604 #ifdef CONFIG_PREEMPT
2605 /* Want to start with kernel preemption disabled. */
2606 task_thread_info(p
)->preempt_count
= 1;
2608 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2614 * wake_up_new_task - wake up a newly created task for the first time.
2616 * This function will do some initial scheduler statistics housekeeping
2617 * that must be done for every newly created context, then puts the task
2618 * on the runqueue and wakes it.
2620 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2622 unsigned long flags
;
2624 int cpu
= get_cpu();
2628 * Fork balancing, do it here and not earlier because:
2629 * - cpus_allowed can change in the fork path
2630 * - any previously selected cpu might disappear through hotplug
2632 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2633 * ->cpus_allowed is stable, we have preemption disabled, meaning
2634 * cpu_online_mask is stable.
2636 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2637 set_task_cpu(p
, cpu
);
2641 * Since the task is not on the rq and we still have TASK_WAKING set
2642 * nobody else will migrate this task.
2645 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2647 BUG_ON(p
->state
!= TASK_WAKING
);
2648 p
->state
= TASK_RUNNING
;
2649 update_rq_clock(rq
);
2650 activate_task(rq
, p
, 0);
2651 trace_sched_wakeup_new(rq
, p
, 1);
2652 check_preempt_curr(rq
, p
, WF_FORK
);
2654 if (p
->sched_class
->task_woken
)
2655 p
->sched_class
->task_woken(rq
, p
);
2657 task_rq_unlock(rq
, &flags
);
2661 #ifdef CONFIG_PREEMPT_NOTIFIERS
2664 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2665 * @notifier: notifier struct to register
2667 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2669 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2671 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2674 * preempt_notifier_unregister - no longer interested in preemption notifications
2675 * @notifier: notifier struct to unregister
2677 * This is safe to call from within a preemption notifier.
2679 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2681 hlist_del(¬ifier
->link
);
2683 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2685 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2687 struct preempt_notifier
*notifier
;
2688 struct hlist_node
*node
;
2690 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2691 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2695 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2696 struct task_struct
*next
)
2698 struct preempt_notifier
*notifier
;
2699 struct hlist_node
*node
;
2701 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2702 notifier
->ops
->sched_out(notifier
, next
);
2705 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2707 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2712 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2713 struct task_struct
*next
)
2717 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2720 * prepare_task_switch - prepare to switch tasks
2721 * @rq: the runqueue preparing to switch
2722 * @prev: the current task that is being switched out
2723 * @next: the task we are going to switch to.
2725 * This is called with the rq lock held and interrupts off. It must
2726 * be paired with a subsequent finish_task_switch after the context
2729 * prepare_task_switch sets up locking and calls architecture specific
2733 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2734 struct task_struct
*next
)
2736 fire_sched_out_preempt_notifiers(prev
, next
);
2737 prepare_lock_switch(rq
, next
);
2738 prepare_arch_switch(next
);
2742 * finish_task_switch - clean up after a task-switch
2743 * @rq: runqueue associated with task-switch
2744 * @prev: the thread we just switched away from.
2746 * finish_task_switch must be called after the context switch, paired
2747 * with a prepare_task_switch call before the context switch.
2748 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2749 * and do any other architecture-specific cleanup actions.
2751 * Note that we may have delayed dropping an mm in context_switch(). If
2752 * so, we finish that here outside of the runqueue lock. (Doing it
2753 * with the lock held can cause deadlocks; see schedule() for
2756 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2757 __releases(rq
->lock
)
2759 struct mm_struct
*mm
= rq
->prev_mm
;
2765 * A task struct has one reference for the use as "current".
2766 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2767 * schedule one last time. The schedule call will never return, and
2768 * the scheduled task must drop that reference.
2769 * The test for TASK_DEAD must occur while the runqueue locks are
2770 * still held, otherwise prev could be scheduled on another cpu, die
2771 * there before we look at prev->state, and then the reference would
2773 * Manfred Spraul <manfred@colorfullife.com>
2775 prev_state
= prev
->state
;
2776 finish_arch_switch(prev
);
2777 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2778 local_irq_disable();
2779 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2780 perf_event_task_sched_in(current
);
2781 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2783 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2784 finish_lock_switch(rq
, prev
);
2786 fire_sched_in_preempt_notifiers(current
);
2789 if (unlikely(prev_state
== TASK_DEAD
)) {
2791 * Remove function-return probe instances associated with this
2792 * task and put them back on the free list.
2794 kprobe_flush_task(prev
);
2795 put_task_struct(prev
);
2801 /* assumes rq->lock is held */
2802 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2804 if (prev
->sched_class
->pre_schedule
)
2805 prev
->sched_class
->pre_schedule(rq
, prev
);
2808 /* rq->lock is NOT held, but preemption is disabled */
2809 static inline void post_schedule(struct rq
*rq
)
2811 if (rq
->post_schedule
) {
2812 unsigned long flags
;
2814 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2815 if (rq
->curr
->sched_class
->post_schedule
)
2816 rq
->curr
->sched_class
->post_schedule(rq
);
2817 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2819 rq
->post_schedule
= 0;
2825 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2829 static inline void post_schedule(struct rq
*rq
)
2836 * schedule_tail - first thing a freshly forked thread must call.
2837 * @prev: the thread we just switched away from.
2839 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2840 __releases(rq
->lock
)
2842 struct rq
*rq
= this_rq();
2844 finish_task_switch(rq
, prev
);
2847 * FIXME: do we need to worry about rq being invalidated by the
2852 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2853 /* In this case, finish_task_switch does not reenable preemption */
2856 if (current
->set_child_tid
)
2857 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2861 * context_switch - switch to the new MM and the new
2862 * thread's register state.
2865 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2866 struct task_struct
*next
)
2868 struct mm_struct
*mm
, *oldmm
;
2870 prepare_task_switch(rq
, prev
, next
);
2871 trace_sched_switch(rq
, prev
, next
);
2873 oldmm
= prev
->active_mm
;
2875 * For paravirt, this is coupled with an exit in switch_to to
2876 * combine the page table reload and the switch backend into
2879 arch_start_context_switch(prev
);
2882 next
->active_mm
= oldmm
;
2883 atomic_inc(&oldmm
->mm_count
);
2884 enter_lazy_tlb(oldmm
, next
);
2886 switch_mm(oldmm
, mm
, next
);
2888 if (likely(!prev
->mm
)) {
2889 prev
->active_mm
= NULL
;
2890 rq
->prev_mm
= oldmm
;
2893 * Since the runqueue lock will be released by the next
2894 * task (which is an invalid locking op but in the case
2895 * of the scheduler it's an obvious special-case), so we
2896 * do an early lockdep release here:
2898 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2899 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2902 /* Here we just switch the register state and the stack. */
2903 switch_to(prev
, next
, prev
);
2907 * this_rq must be evaluated again because prev may have moved
2908 * CPUs since it called schedule(), thus the 'rq' on its stack
2909 * frame will be invalid.
2911 finish_task_switch(this_rq(), prev
);
2915 * nr_running, nr_uninterruptible and nr_context_switches:
2917 * externally visible scheduler statistics: current number of runnable
2918 * threads, current number of uninterruptible-sleeping threads, total
2919 * number of context switches performed since bootup.
2921 unsigned long nr_running(void)
2923 unsigned long i
, sum
= 0;
2925 for_each_online_cpu(i
)
2926 sum
+= cpu_rq(i
)->nr_running
;
2931 unsigned long nr_uninterruptible(void)
2933 unsigned long i
, sum
= 0;
2935 for_each_possible_cpu(i
)
2936 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2939 * Since we read the counters lockless, it might be slightly
2940 * inaccurate. Do not allow it to go below zero though:
2942 if (unlikely((long)sum
< 0))
2948 unsigned long long nr_context_switches(void)
2951 unsigned long long sum
= 0;
2953 for_each_possible_cpu(i
)
2954 sum
+= cpu_rq(i
)->nr_switches
;
2959 unsigned long nr_iowait(void)
2961 unsigned long i
, sum
= 0;
2963 for_each_possible_cpu(i
)
2964 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2969 unsigned long nr_iowait_cpu(void)
2971 struct rq
*this = this_rq();
2972 return atomic_read(&this->nr_iowait
);
2975 unsigned long this_cpu_load(void)
2977 struct rq
*this = this_rq();
2978 return this->cpu_load
[0];
2982 /* Variables and functions for calc_load */
2983 static atomic_long_t calc_load_tasks
;
2984 static unsigned long calc_load_update
;
2985 unsigned long avenrun
[3];
2986 EXPORT_SYMBOL(avenrun
);
2989 * get_avenrun - get the load average array
2990 * @loads: pointer to dest load array
2991 * @offset: offset to add
2992 * @shift: shift count to shift the result left
2994 * These values are estimates at best, so no need for locking.
2996 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2998 loads
[0] = (avenrun
[0] + offset
) << shift
;
2999 loads
[1] = (avenrun
[1] + offset
) << shift
;
3000 loads
[2] = (avenrun
[2] + offset
) << shift
;
3003 static unsigned long
3004 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3007 load
+= active
* (FIXED_1
- exp
);
3008 return load
>> FSHIFT
;
3012 * calc_load - update the avenrun load estimates 10 ticks after the
3013 * CPUs have updated calc_load_tasks.
3015 void calc_global_load(void)
3017 unsigned long upd
= calc_load_update
+ 10;
3020 if (time_before(jiffies
, upd
))
3023 active
= atomic_long_read(&calc_load_tasks
);
3024 active
= active
> 0 ? active
* FIXED_1
: 0;
3026 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3027 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3028 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3030 calc_load_update
+= LOAD_FREQ
;
3034 * Either called from update_cpu_load() or from a cpu going idle
3036 static void calc_load_account_active(struct rq
*this_rq
)
3038 long nr_active
, delta
;
3040 nr_active
= this_rq
->nr_running
;
3041 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3043 if (nr_active
!= this_rq
->calc_load_active
) {
3044 delta
= nr_active
- this_rq
->calc_load_active
;
3045 this_rq
->calc_load_active
= nr_active
;
3046 atomic_long_add(delta
, &calc_load_tasks
);
3051 * Update rq->cpu_load[] statistics. This function is usually called every
3052 * scheduler tick (TICK_NSEC).
3054 static void update_cpu_load(struct rq
*this_rq
)
3056 unsigned long this_load
= this_rq
->load
.weight
;
3059 this_rq
->nr_load_updates
++;
3061 /* Update our load: */
3062 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3063 unsigned long old_load
, new_load
;
3065 /* scale is effectively 1 << i now, and >> i divides by scale */
3067 old_load
= this_rq
->cpu_load
[i
];
3068 new_load
= this_load
;
3070 * Round up the averaging division if load is increasing. This
3071 * prevents us from getting stuck on 9 if the load is 10, for
3074 if (new_load
> old_load
)
3075 new_load
+= scale
-1;
3076 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3079 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3080 this_rq
->calc_load_update
+= LOAD_FREQ
;
3081 calc_load_account_active(this_rq
);
3088 * sched_exec - execve() is a valuable balancing opportunity, because at
3089 * this point the task has the smallest effective memory and cache footprint.
3091 void sched_exec(void)
3093 struct task_struct
*p
= current
;
3094 struct migration_req req
;
3095 int dest_cpu
, this_cpu
;
3096 unsigned long flags
;
3100 this_cpu
= get_cpu();
3101 dest_cpu
= select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3102 if (dest_cpu
== this_cpu
) {
3107 rq
= task_rq_lock(p
, &flags
);
3111 * select_task_rq() can race against ->cpus_allowed
3113 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3114 || unlikely(!cpu_active(dest_cpu
))) {
3115 task_rq_unlock(rq
, &flags
);
3119 /* force the process onto the specified CPU */
3120 if (migrate_task(p
, dest_cpu
, &req
)) {
3121 /* Need to wait for migration thread (might exit: take ref). */
3122 struct task_struct
*mt
= rq
->migration_thread
;
3124 get_task_struct(mt
);
3125 task_rq_unlock(rq
, &flags
);
3126 wake_up_process(mt
);
3127 put_task_struct(mt
);
3128 wait_for_completion(&req
.done
);
3132 task_rq_unlock(rq
, &flags
);
3137 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3139 EXPORT_PER_CPU_SYMBOL(kstat
);
3142 * Return any ns on the sched_clock that have not yet been accounted in
3143 * @p in case that task is currently running.
3145 * Called with task_rq_lock() held on @rq.
3147 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3151 if (task_current(rq
, p
)) {
3152 update_rq_clock(rq
);
3153 ns
= rq
->clock
- p
->se
.exec_start
;
3161 unsigned long long task_delta_exec(struct task_struct
*p
)
3163 unsigned long flags
;
3167 rq
= task_rq_lock(p
, &flags
);
3168 ns
= do_task_delta_exec(p
, rq
);
3169 task_rq_unlock(rq
, &flags
);
3175 * Return accounted runtime for the task.
3176 * In case the task is currently running, return the runtime plus current's
3177 * pending runtime that have not been accounted yet.
3179 unsigned long long task_sched_runtime(struct task_struct
*p
)
3181 unsigned long flags
;
3185 rq
= task_rq_lock(p
, &flags
);
3186 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3187 task_rq_unlock(rq
, &flags
);
3193 * Return sum_exec_runtime for the thread group.
3194 * In case the task is currently running, return the sum plus current's
3195 * pending runtime that have not been accounted yet.
3197 * Note that the thread group might have other running tasks as well,
3198 * so the return value not includes other pending runtime that other
3199 * running tasks might have.
3201 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3203 struct task_cputime totals
;
3204 unsigned long flags
;
3208 rq
= task_rq_lock(p
, &flags
);
3209 thread_group_cputime(p
, &totals
);
3210 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3211 task_rq_unlock(rq
, &flags
);
3217 * Account user cpu time to a process.
3218 * @p: the process that the cpu time gets accounted to
3219 * @cputime: the cpu time spent in user space since the last update
3220 * @cputime_scaled: cputime scaled by cpu frequency
3222 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3223 cputime_t cputime_scaled
)
3225 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3228 /* Add user time to process. */
3229 p
->utime
= cputime_add(p
->utime
, cputime
);
3230 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3231 account_group_user_time(p
, cputime
);
3233 /* Add user time to cpustat. */
3234 tmp
= cputime_to_cputime64(cputime
);
3235 if (TASK_NICE(p
) > 0)
3236 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3238 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3240 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3241 /* Account for user time used */
3242 acct_update_integrals(p
);
3246 * Account guest cpu time to a process.
3247 * @p: the process that the cpu time gets accounted to
3248 * @cputime: the cpu time spent in virtual machine since the last update
3249 * @cputime_scaled: cputime scaled by cpu frequency
3251 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3252 cputime_t cputime_scaled
)
3255 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3257 tmp
= cputime_to_cputime64(cputime
);
3259 /* Add guest time to process. */
3260 p
->utime
= cputime_add(p
->utime
, cputime
);
3261 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3262 account_group_user_time(p
, cputime
);
3263 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3265 /* Add guest time to cpustat. */
3266 if (TASK_NICE(p
) > 0) {
3267 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3268 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3270 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3271 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3276 * Account system cpu time to a process.
3277 * @p: the process that the cpu time gets accounted to
3278 * @hardirq_offset: the offset to subtract from hardirq_count()
3279 * @cputime: the cpu time spent in kernel space since the last update
3280 * @cputime_scaled: cputime scaled by cpu frequency
3282 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3283 cputime_t cputime
, cputime_t cputime_scaled
)
3285 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3288 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3289 account_guest_time(p
, cputime
, cputime_scaled
);
3293 /* Add system time to process. */
3294 p
->stime
= cputime_add(p
->stime
, cputime
);
3295 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3296 account_group_system_time(p
, cputime
);
3298 /* Add system time to cpustat. */
3299 tmp
= cputime_to_cputime64(cputime
);
3300 if (hardirq_count() - hardirq_offset
)
3301 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3302 else if (softirq_count())
3303 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3305 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3307 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3309 /* Account for system time used */
3310 acct_update_integrals(p
);
3314 * Account for involuntary wait time.
3315 * @steal: the cpu time spent in involuntary wait
3317 void account_steal_time(cputime_t cputime
)
3319 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3320 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3322 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3326 * Account for idle time.
3327 * @cputime: the cpu time spent in idle wait
3329 void account_idle_time(cputime_t cputime
)
3331 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3332 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3333 struct rq
*rq
= this_rq();
3335 if (atomic_read(&rq
->nr_iowait
) > 0)
3336 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3338 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3341 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3344 * Account a single tick of cpu time.
3345 * @p: the process that the cpu time gets accounted to
3346 * @user_tick: indicates if the tick is a user or a system tick
3348 void account_process_tick(struct task_struct
*p
, int user_tick
)
3350 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3351 struct rq
*rq
= this_rq();
3354 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3355 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3356 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3359 account_idle_time(cputime_one_jiffy
);
3363 * Account multiple ticks of steal time.
3364 * @p: the process from which the cpu time has been stolen
3365 * @ticks: number of stolen ticks
3367 void account_steal_ticks(unsigned long ticks
)
3369 account_steal_time(jiffies_to_cputime(ticks
));
3373 * Account multiple ticks of idle time.
3374 * @ticks: number of stolen ticks
3376 void account_idle_ticks(unsigned long ticks
)
3378 account_idle_time(jiffies_to_cputime(ticks
));
3384 * Use precise platform statistics if available:
3386 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3387 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3393 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3395 struct task_cputime cputime
;
3397 thread_group_cputime(p
, &cputime
);
3399 *ut
= cputime
.utime
;
3400 *st
= cputime
.stime
;
3404 #ifndef nsecs_to_cputime
3405 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3408 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3410 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3413 * Use CFS's precise accounting:
3415 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3420 temp
= (u64
)(rtime
* utime
);
3421 do_div(temp
, total
);
3422 utime
= (cputime_t
)temp
;
3427 * Compare with previous values, to keep monotonicity:
3429 p
->prev_utime
= max(p
->prev_utime
, utime
);
3430 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3432 *ut
= p
->prev_utime
;
3433 *st
= p
->prev_stime
;
3437 * Must be called with siglock held.
3439 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3441 struct signal_struct
*sig
= p
->signal
;
3442 struct task_cputime cputime
;
3443 cputime_t rtime
, utime
, total
;
3445 thread_group_cputime(p
, &cputime
);
3447 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3448 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3453 temp
= (u64
)(rtime
* cputime
.utime
);
3454 do_div(temp
, total
);
3455 utime
= (cputime_t
)temp
;
3459 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3460 sig
->prev_stime
= max(sig
->prev_stime
,
3461 cputime_sub(rtime
, sig
->prev_utime
));
3463 *ut
= sig
->prev_utime
;
3464 *st
= sig
->prev_stime
;
3469 * This function gets called by the timer code, with HZ frequency.
3470 * We call it with interrupts disabled.
3472 * It also gets called by the fork code, when changing the parent's
3475 void scheduler_tick(void)
3477 int cpu
= smp_processor_id();
3478 struct rq
*rq
= cpu_rq(cpu
);
3479 struct task_struct
*curr
= rq
->curr
;
3483 raw_spin_lock(&rq
->lock
);
3484 update_rq_clock(rq
);
3485 update_cpu_load(rq
);
3486 curr
->sched_class
->task_tick(rq
, curr
, 0);
3487 raw_spin_unlock(&rq
->lock
);
3489 perf_event_task_tick(curr
);
3492 rq
->idle_at_tick
= idle_cpu(cpu
);
3493 trigger_load_balance(rq
, cpu
);
3497 notrace
unsigned long get_parent_ip(unsigned long addr
)
3499 if (in_lock_functions(addr
)) {
3500 addr
= CALLER_ADDR2
;
3501 if (in_lock_functions(addr
))
3502 addr
= CALLER_ADDR3
;
3507 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3508 defined(CONFIG_PREEMPT_TRACER))
3510 void __kprobes
add_preempt_count(int val
)
3512 #ifdef CONFIG_DEBUG_PREEMPT
3516 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3519 preempt_count() += val
;
3520 #ifdef CONFIG_DEBUG_PREEMPT
3522 * Spinlock count overflowing soon?
3524 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3527 if (preempt_count() == val
)
3528 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3530 EXPORT_SYMBOL(add_preempt_count
);
3532 void __kprobes
sub_preempt_count(int val
)
3534 #ifdef CONFIG_DEBUG_PREEMPT
3538 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3541 * Is the spinlock portion underflowing?
3543 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3544 !(preempt_count() & PREEMPT_MASK
)))
3548 if (preempt_count() == val
)
3549 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3550 preempt_count() -= val
;
3552 EXPORT_SYMBOL(sub_preempt_count
);
3557 * Print scheduling while atomic bug:
3559 static noinline
void __schedule_bug(struct task_struct
*prev
)
3561 struct pt_regs
*regs
= get_irq_regs();
3563 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3564 prev
->comm
, prev
->pid
, preempt_count());
3566 debug_show_held_locks(prev
);
3568 if (irqs_disabled())
3569 print_irqtrace_events(prev
);
3578 * Various schedule()-time debugging checks and statistics:
3580 static inline void schedule_debug(struct task_struct
*prev
)
3583 * Test if we are atomic. Since do_exit() needs to call into
3584 * schedule() atomically, we ignore that path for now.
3585 * Otherwise, whine if we are scheduling when we should not be.
3587 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3588 __schedule_bug(prev
);
3590 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3592 schedstat_inc(this_rq(), sched_count
);
3593 #ifdef CONFIG_SCHEDSTATS
3594 if (unlikely(prev
->lock_depth
>= 0)) {
3595 schedstat_inc(this_rq(), bkl_count
);
3596 schedstat_inc(prev
, sched_info
.bkl_count
);
3601 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3603 if (prev
->state
== TASK_RUNNING
) {
3604 u64 runtime
= prev
->se
.sum_exec_runtime
;
3606 runtime
-= prev
->se
.prev_sum_exec_runtime
;
3607 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
3610 * In order to avoid avg_overlap growing stale when we are
3611 * indeed overlapping and hence not getting put to sleep, grow
3612 * the avg_overlap on preemption.
3614 * We use the average preemption runtime because that
3615 * correlates to the amount of cache footprint a task can
3618 update_avg(&prev
->se
.avg_overlap
, runtime
);
3620 prev
->sched_class
->put_prev_task(rq
, prev
);
3624 * Pick up the highest-prio task:
3626 static inline struct task_struct
*
3627 pick_next_task(struct rq
*rq
)
3629 const struct sched_class
*class;
3630 struct task_struct
*p
;
3633 * Optimization: we know that if all tasks are in
3634 * the fair class we can call that function directly:
3636 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3637 p
= fair_sched_class
.pick_next_task(rq
);
3642 class = sched_class_highest
;
3644 p
= class->pick_next_task(rq
);
3648 * Will never be NULL as the idle class always
3649 * returns a non-NULL p:
3651 class = class->next
;
3656 * schedule() is the main scheduler function.
3658 asmlinkage
void __sched
schedule(void)
3660 struct task_struct
*prev
, *next
;
3661 unsigned long *switch_count
;
3667 cpu
= smp_processor_id();
3671 switch_count
= &prev
->nivcsw
;
3673 release_kernel_lock(prev
);
3674 need_resched_nonpreemptible
:
3676 schedule_debug(prev
);
3678 if (sched_feat(HRTICK
))
3681 raw_spin_lock_irq(&rq
->lock
);
3682 update_rq_clock(rq
);
3683 clear_tsk_need_resched(prev
);
3685 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3686 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3687 prev
->state
= TASK_RUNNING
;
3689 deactivate_task(rq
, prev
, 1);
3690 switch_count
= &prev
->nvcsw
;
3693 pre_schedule(rq
, prev
);
3695 if (unlikely(!rq
->nr_running
))
3696 idle_balance(cpu
, rq
);
3698 put_prev_task(rq
, prev
);
3699 next
= pick_next_task(rq
);
3701 if (likely(prev
!= next
)) {
3702 sched_info_switch(prev
, next
);
3703 perf_event_task_sched_out(prev
, next
);
3709 context_switch(rq
, prev
, next
); /* unlocks the rq */
3711 * the context switch might have flipped the stack from under
3712 * us, hence refresh the local variables.
3714 cpu
= smp_processor_id();
3717 raw_spin_unlock_irq(&rq
->lock
);
3721 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3723 switch_count
= &prev
->nivcsw
;
3724 goto need_resched_nonpreemptible
;
3727 preempt_enable_no_resched();
3731 EXPORT_SYMBOL(schedule
);
3733 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3735 * Look out! "owner" is an entirely speculative pointer
3736 * access and not reliable.
3738 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3743 if (!sched_feat(OWNER_SPIN
))
3746 #ifdef CONFIG_DEBUG_PAGEALLOC
3748 * Need to access the cpu field knowing that
3749 * DEBUG_PAGEALLOC could have unmapped it if
3750 * the mutex owner just released it and exited.
3752 if (probe_kernel_address(&owner
->cpu
, cpu
))
3759 * Even if the access succeeded (likely case),
3760 * the cpu field may no longer be valid.
3762 if (cpu
>= nr_cpumask_bits
)
3766 * We need to validate that we can do a
3767 * get_cpu() and that we have the percpu area.
3769 if (!cpu_online(cpu
))
3776 * Owner changed, break to re-assess state.
3778 if (lock
->owner
!= owner
)
3782 * Is that owner really running on that cpu?
3784 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3794 #ifdef CONFIG_PREEMPT
3796 * this is the entry point to schedule() from in-kernel preemption
3797 * off of preempt_enable. Kernel preemptions off return from interrupt
3798 * occur there and call schedule directly.
3800 asmlinkage
void __sched
preempt_schedule(void)
3802 struct thread_info
*ti
= current_thread_info();
3805 * If there is a non-zero preempt_count or interrupts are disabled,
3806 * we do not want to preempt the current task. Just return..
3808 if (likely(ti
->preempt_count
|| irqs_disabled()))
3812 add_preempt_count(PREEMPT_ACTIVE
);
3814 sub_preempt_count(PREEMPT_ACTIVE
);
3817 * Check again in case we missed a preemption opportunity
3818 * between schedule and now.
3821 } while (need_resched());
3823 EXPORT_SYMBOL(preempt_schedule
);
3826 * this is the entry point to schedule() from kernel preemption
3827 * off of irq context.
3828 * Note, that this is called and return with irqs disabled. This will
3829 * protect us against recursive calling from irq.
3831 asmlinkage
void __sched
preempt_schedule_irq(void)
3833 struct thread_info
*ti
= current_thread_info();
3835 /* Catch callers which need to be fixed */
3836 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3839 add_preempt_count(PREEMPT_ACTIVE
);
3842 local_irq_disable();
3843 sub_preempt_count(PREEMPT_ACTIVE
);
3846 * Check again in case we missed a preemption opportunity
3847 * between schedule and now.
3850 } while (need_resched());
3853 #endif /* CONFIG_PREEMPT */
3855 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3858 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3860 EXPORT_SYMBOL(default_wake_function
);
3863 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3864 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3865 * number) then we wake all the non-exclusive tasks and one exclusive task.
3867 * There are circumstances in which we can try to wake a task which has already
3868 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3869 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3871 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3872 int nr_exclusive
, int wake_flags
, void *key
)
3874 wait_queue_t
*curr
, *next
;
3876 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3877 unsigned flags
= curr
->flags
;
3879 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3880 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3886 * __wake_up - wake up threads blocked on a waitqueue.
3888 * @mode: which threads
3889 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3890 * @key: is directly passed to the wakeup function
3892 * It may be assumed that this function implies a write memory barrier before
3893 * changing the task state if and only if any tasks are woken up.
3895 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3896 int nr_exclusive
, void *key
)
3898 unsigned long flags
;
3900 spin_lock_irqsave(&q
->lock
, flags
);
3901 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3902 spin_unlock_irqrestore(&q
->lock
, flags
);
3904 EXPORT_SYMBOL(__wake_up
);
3907 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3909 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3911 __wake_up_common(q
, mode
, 1, 0, NULL
);
3914 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3916 __wake_up_common(q
, mode
, 1, 0, key
);
3920 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3922 * @mode: which threads
3923 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3924 * @key: opaque value to be passed to wakeup targets
3926 * The sync wakeup differs that the waker knows that it will schedule
3927 * away soon, so while the target thread will be woken up, it will not
3928 * be migrated to another CPU - ie. the two threads are 'synchronized'
3929 * with each other. This can prevent needless bouncing between CPUs.
3931 * On UP it can prevent extra preemption.
3933 * It may be assumed that this function implies a write memory barrier before
3934 * changing the task state if and only if any tasks are woken up.
3936 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3937 int nr_exclusive
, void *key
)
3939 unsigned long flags
;
3940 int wake_flags
= WF_SYNC
;
3945 if (unlikely(!nr_exclusive
))
3948 spin_lock_irqsave(&q
->lock
, flags
);
3949 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3950 spin_unlock_irqrestore(&q
->lock
, flags
);
3952 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3955 * __wake_up_sync - see __wake_up_sync_key()
3957 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3959 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3961 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3964 * complete: - signals a single thread waiting on this completion
3965 * @x: holds the state of this particular completion
3967 * This will wake up a single thread waiting on this completion. Threads will be
3968 * awakened in the same order in which they were queued.
3970 * See also complete_all(), wait_for_completion() and related routines.
3972 * It may be assumed that this function implies a write memory barrier before
3973 * changing the task state if and only if any tasks are woken up.
3975 void complete(struct completion
*x
)
3977 unsigned long flags
;
3979 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3981 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3982 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3984 EXPORT_SYMBOL(complete
);
3987 * complete_all: - signals all threads waiting on this completion
3988 * @x: holds the state of this particular completion
3990 * This will wake up all threads waiting on this particular completion event.
3992 * It may be assumed that this function implies a write memory barrier before
3993 * changing the task state if and only if any tasks are woken up.
3995 void complete_all(struct completion
*x
)
3997 unsigned long flags
;
3999 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4000 x
->done
+= UINT_MAX
/2;
4001 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4002 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4004 EXPORT_SYMBOL(complete_all
);
4006 static inline long __sched
4007 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4010 DECLARE_WAITQUEUE(wait
, current
);
4012 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4013 __add_wait_queue_tail(&x
->wait
, &wait
);
4015 if (signal_pending_state(state
, current
)) {
4016 timeout
= -ERESTARTSYS
;
4019 __set_current_state(state
);
4020 spin_unlock_irq(&x
->wait
.lock
);
4021 timeout
= schedule_timeout(timeout
);
4022 spin_lock_irq(&x
->wait
.lock
);
4023 } while (!x
->done
&& timeout
);
4024 __remove_wait_queue(&x
->wait
, &wait
);
4029 return timeout
?: 1;
4033 wait_for_common(struct completion
*x
, long timeout
, int state
)
4037 spin_lock_irq(&x
->wait
.lock
);
4038 timeout
= do_wait_for_common(x
, timeout
, state
);
4039 spin_unlock_irq(&x
->wait
.lock
);
4044 * wait_for_completion: - waits for completion of a task
4045 * @x: holds the state of this particular completion
4047 * This waits to be signaled for completion of a specific task. It is NOT
4048 * interruptible and there is no timeout.
4050 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4051 * and interrupt capability. Also see complete().
4053 void __sched
wait_for_completion(struct completion
*x
)
4055 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4057 EXPORT_SYMBOL(wait_for_completion
);
4060 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4061 * @x: holds the state of this particular completion
4062 * @timeout: timeout value in jiffies
4064 * This waits for either a completion of a specific task to be signaled or for a
4065 * specified timeout to expire. The timeout is in jiffies. It is not
4068 unsigned long __sched
4069 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4071 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4073 EXPORT_SYMBOL(wait_for_completion_timeout
);
4076 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4077 * @x: holds the state of this particular completion
4079 * This waits for completion of a specific task to be signaled. It is
4082 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4084 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4085 if (t
== -ERESTARTSYS
)
4089 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4092 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4093 * @x: holds the state of this particular completion
4094 * @timeout: timeout value in jiffies
4096 * This waits for either a completion of a specific task to be signaled or for a
4097 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4099 unsigned long __sched
4100 wait_for_completion_interruptible_timeout(struct completion
*x
,
4101 unsigned long timeout
)
4103 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4105 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4108 * wait_for_completion_killable: - waits for completion of a task (killable)
4109 * @x: holds the state of this particular completion
4111 * This waits to be signaled for completion of a specific task. It can be
4112 * interrupted by a kill signal.
4114 int __sched
wait_for_completion_killable(struct completion
*x
)
4116 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4117 if (t
== -ERESTARTSYS
)
4121 EXPORT_SYMBOL(wait_for_completion_killable
);
4124 * try_wait_for_completion - try to decrement a completion without blocking
4125 * @x: completion structure
4127 * Returns: 0 if a decrement cannot be done without blocking
4128 * 1 if a decrement succeeded.
4130 * If a completion is being used as a counting completion,
4131 * attempt to decrement the counter without blocking. This
4132 * enables us to avoid waiting if the resource the completion
4133 * is protecting is not available.
4135 bool try_wait_for_completion(struct completion
*x
)
4137 unsigned long flags
;
4140 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4145 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4148 EXPORT_SYMBOL(try_wait_for_completion
);
4151 * completion_done - Test to see if a completion has any waiters
4152 * @x: completion structure
4154 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4155 * 1 if there are no waiters.
4158 bool completion_done(struct completion
*x
)
4160 unsigned long flags
;
4163 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4166 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4169 EXPORT_SYMBOL(completion_done
);
4172 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4174 unsigned long flags
;
4177 init_waitqueue_entry(&wait
, current
);
4179 __set_current_state(state
);
4181 spin_lock_irqsave(&q
->lock
, flags
);
4182 __add_wait_queue(q
, &wait
);
4183 spin_unlock(&q
->lock
);
4184 timeout
= schedule_timeout(timeout
);
4185 spin_lock_irq(&q
->lock
);
4186 __remove_wait_queue(q
, &wait
);
4187 spin_unlock_irqrestore(&q
->lock
, flags
);
4192 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4194 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4196 EXPORT_SYMBOL(interruptible_sleep_on
);
4199 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4201 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4203 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4205 void __sched
sleep_on(wait_queue_head_t
*q
)
4207 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4209 EXPORT_SYMBOL(sleep_on
);
4211 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4213 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4215 EXPORT_SYMBOL(sleep_on_timeout
);
4217 #ifdef CONFIG_RT_MUTEXES
4220 * rt_mutex_setprio - set the current priority of a task
4222 * @prio: prio value (kernel-internal form)
4224 * This function changes the 'effective' priority of a task. It does
4225 * not touch ->normal_prio like __setscheduler().
4227 * Used by the rt_mutex code to implement priority inheritance logic.
4229 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4231 unsigned long flags
;
4232 int oldprio
, on_rq
, running
;
4234 const struct sched_class
*prev_class
;
4236 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4238 rq
= task_rq_lock(p
, &flags
);
4239 update_rq_clock(rq
);
4242 prev_class
= p
->sched_class
;
4243 on_rq
= p
->se
.on_rq
;
4244 running
= task_current(rq
, p
);
4246 dequeue_task(rq
, p
, 0);
4248 p
->sched_class
->put_prev_task(rq
, p
);
4251 p
->sched_class
= &rt_sched_class
;
4253 p
->sched_class
= &fair_sched_class
;
4258 p
->sched_class
->set_curr_task(rq
);
4260 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4262 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4264 task_rq_unlock(rq
, &flags
);
4269 void set_user_nice(struct task_struct
*p
, long nice
)
4271 int old_prio
, delta
, on_rq
;
4272 unsigned long flags
;
4275 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4278 * We have to be careful, if called from sys_setpriority(),
4279 * the task might be in the middle of scheduling on another CPU.
4281 rq
= task_rq_lock(p
, &flags
);
4282 update_rq_clock(rq
);
4284 * The RT priorities are set via sched_setscheduler(), but we still
4285 * allow the 'normal' nice value to be set - but as expected
4286 * it wont have any effect on scheduling until the task is
4287 * SCHED_FIFO/SCHED_RR:
4289 if (task_has_rt_policy(p
)) {
4290 p
->static_prio
= NICE_TO_PRIO(nice
);
4293 on_rq
= p
->se
.on_rq
;
4295 dequeue_task(rq
, p
, 0);
4297 p
->static_prio
= NICE_TO_PRIO(nice
);
4300 p
->prio
= effective_prio(p
);
4301 delta
= p
->prio
- old_prio
;
4304 enqueue_task(rq
, p
, 0, false);
4306 * If the task increased its priority or is running and
4307 * lowered its priority, then reschedule its CPU:
4309 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4310 resched_task(rq
->curr
);
4313 task_rq_unlock(rq
, &flags
);
4315 EXPORT_SYMBOL(set_user_nice
);
4318 * can_nice - check if a task can reduce its nice value
4322 int can_nice(const struct task_struct
*p
, const int nice
)
4324 /* convert nice value [19,-20] to rlimit style value [1,40] */
4325 int nice_rlim
= 20 - nice
;
4327 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4328 capable(CAP_SYS_NICE
));
4331 #ifdef __ARCH_WANT_SYS_NICE
4334 * sys_nice - change the priority of the current process.
4335 * @increment: priority increment
4337 * sys_setpriority is a more generic, but much slower function that
4338 * does similar things.
4340 SYSCALL_DEFINE1(nice
, int, increment
)
4345 * Setpriority might change our priority at the same moment.
4346 * We don't have to worry. Conceptually one call occurs first
4347 * and we have a single winner.
4349 if (increment
< -40)
4354 nice
= TASK_NICE(current
) + increment
;
4360 if (increment
< 0 && !can_nice(current
, nice
))
4363 retval
= security_task_setnice(current
, nice
);
4367 set_user_nice(current
, nice
);
4374 * task_prio - return the priority value of a given task.
4375 * @p: the task in question.
4377 * This is the priority value as seen by users in /proc.
4378 * RT tasks are offset by -200. Normal tasks are centered
4379 * around 0, value goes from -16 to +15.
4381 int task_prio(const struct task_struct
*p
)
4383 return p
->prio
- MAX_RT_PRIO
;
4387 * task_nice - return the nice value of a given task.
4388 * @p: the task in question.
4390 int task_nice(const struct task_struct
*p
)
4392 return TASK_NICE(p
);
4394 EXPORT_SYMBOL(task_nice
);
4397 * idle_cpu - is a given cpu idle currently?
4398 * @cpu: the processor in question.
4400 int idle_cpu(int cpu
)
4402 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4406 * idle_task - return the idle task for a given cpu.
4407 * @cpu: the processor in question.
4409 struct task_struct
*idle_task(int cpu
)
4411 return cpu_rq(cpu
)->idle
;
4415 * find_process_by_pid - find a process with a matching PID value.
4416 * @pid: the pid in question.
4418 static struct task_struct
*find_process_by_pid(pid_t pid
)
4420 return pid
? find_task_by_vpid(pid
) : current
;
4423 /* Actually do priority change: must hold rq lock. */
4425 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4427 BUG_ON(p
->se
.on_rq
);
4430 p
->rt_priority
= prio
;
4431 p
->normal_prio
= normal_prio(p
);
4432 /* we are holding p->pi_lock already */
4433 p
->prio
= rt_mutex_getprio(p
);
4434 if (rt_prio(p
->prio
))
4435 p
->sched_class
= &rt_sched_class
;
4437 p
->sched_class
= &fair_sched_class
;
4442 * check the target process has a UID that matches the current process's
4444 static bool check_same_owner(struct task_struct
*p
)
4446 const struct cred
*cred
= current_cred(), *pcred
;
4450 pcred
= __task_cred(p
);
4451 match
= (cred
->euid
== pcred
->euid
||
4452 cred
->euid
== pcred
->uid
);
4457 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4458 struct sched_param
*param
, bool user
)
4460 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4461 unsigned long flags
;
4462 const struct sched_class
*prev_class
;
4466 /* may grab non-irq protected spin_locks */
4467 BUG_ON(in_interrupt());
4469 /* double check policy once rq lock held */
4471 reset_on_fork
= p
->sched_reset_on_fork
;
4472 policy
= oldpolicy
= p
->policy
;
4474 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4475 policy
&= ~SCHED_RESET_ON_FORK
;
4477 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4478 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4479 policy
!= SCHED_IDLE
)
4484 * Valid priorities for SCHED_FIFO and SCHED_RR are
4485 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4486 * SCHED_BATCH and SCHED_IDLE is 0.
4488 if (param
->sched_priority
< 0 ||
4489 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4490 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4492 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4496 * Allow unprivileged RT tasks to decrease priority:
4498 if (user
&& !capable(CAP_SYS_NICE
)) {
4499 if (rt_policy(policy
)) {
4500 unsigned long rlim_rtprio
;
4502 if (!lock_task_sighand(p
, &flags
))
4504 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4505 unlock_task_sighand(p
, &flags
);
4507 /* can't set/change the rt policy */
4508 if (policy
!= p
->policy
&& !rlim_rtprio
)
4511 /* can't increase priority */
4512 if (param
->sched_priority
> p
->rt_priority
&&
4513 param
->sched_priority
> rlim_rtprio
)
4517 * Like positive nice levels, dont allow tasks to
4518 * move out of SCHED_IDLE either:
4520 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4523 /* can't change other user's priorities */
4524 if (!check_same_owner(p
))
4527 /* Normal users shall not reset the sched_reset_on_fork flag */
4528 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4533 #ifdef CONFIG_RT_GROUP_SCHED
4535 * Do not allow realtime tasks into groups that have no runtime
4538 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4539 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4543 retval
= security_task_setscheduler(p
, policy
, param
);
4549 * make sure no PI-waiters arrive (or leave) while we are
4550 * changing the priority of the task:
4552 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4554 * To be able to change p->policy safely, the apropriate
4555 * runqueue lock must be held.
4557 rq
= __task_rq_lock(p
);
4558 /* recheck policy now with rq lock held */
4559 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4560 policy
= oldpolicy
= -1;
4561 __task_rq_unlock(rq
);
4562 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4565 update_rq_clock(rq
);
4566 on_rq
= p
->se
.on_rq
;
4567 running
= task_current(rq
, p
);
4569 deactivate_task(rq
, p
, 0);
4571 p
->sched_class
->put_prev_task(rq
, p
);
4573 p
->sched_reset_on_fork
= reset_on_fork
;
4576 prev_class
= p
->sched_class
;
4577 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4580 p
->sched_class
->set_curr_task(rq
);
4582 activate_task(rq
, p
, 0);
4584 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4586 __task_rq_unlock(rq
);
4587 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4589 rt_mutex_adjust_pi(p
);
4595 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4596 * @p: the task in question.
4597 * @policy: new policy.
4598 * @param: structure containing the new RT priority.
4600 * NOTE that the task may be already dead.
4602 int sched_setscheduler(struct task_struct
*p
, int policy
,
4603 struct sched_param
*param
)
4605 return __sched_setscheduler(p
, policy
, param
, true);
4607 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4610 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4611 * @p: the task in question.
4612 * @policy: new policy.
4613 * @param: structure containing the new RT priority.
4615 * Just like sched_setscheduler, only don't bother checking if the
4616 * current context has permission. For example, this is needed in
4617 * stop_machine(): we create temporary high priority worker threads,
4618 * but our caller might not have that capability.
4620 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4621 struct sched_param
*param
)
4623 return __sched_setscheduler(p
, policy
, param
, false);
4627 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4629 struct sched_param lparam
;
4630 struct task_struct
*p
;
4633 if (!param
|| pid
< 0)
4635 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4640 p
= find_process_by_pid(pid
);
4642 retval
= sched_setscheduler(p
, policy
, &lparam
);
4649 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4650 * @pid: the pid in question.
4651 * @policy: new policy.
4652 * @param: structure containing the new RT priority.
4654 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4655 struct sched_param __user
*, param
)
4657 /* negative values for policy are not valid */
4661 return do_sched_setscheduler(pid
, policy
, param
);
4665 * sys_sched_setparam - set/change the RT priority of a thread
4666 * @pid: the pid in question.
4667 * @param: structure containing the new RT priority.
4669 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4671 return do_sched_setscheduler(pid
, -1, param
);
4675 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4676 * @pid: the pid in question.
4678 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4680 struct task_struct
*p
;
4688 p
= find_process_by_pid(pid
);
4690 retval
= security_task_getscheduler(p
);
4693 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4700 * sys_sched_getparam - get the RT priority of a thread
4701 * @pid: the pid in question.
4702 * @param: structure containing the RT priority.
4704 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4706 struct sched_param lp
;
4707 struct task_struct
*p
;
4710 if (!param
|| pid
< 0)
4714 p
= find_process_by_pid(pid
);
4719 retval
= security_task_getscheduler(p
);
4723 lp
.sched_priority
= p
->rt_priority
;
4727 * This one might sleep, we cannot do it with a spinlock held ...
4729 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4738 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4740 cpumask_var_t cpus_allowed
, new_mask
;
4741 struct task_struct
*p
;
4747 p
= find_process_by_pid(pid
);
4754 /* Prevent p going away */
4758 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4762 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4764 goto out_free_cpus_allowed
;
4767 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4770 retval
= security_task_setscheduler(p
, 0, NULL
);
4774 cpuset_cpus_allowed(p
, cpus_allowed
);
4775 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4777 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4780 cpuset_cpus_allowed(p
, cpus_allowed
);
4781 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4783 * We must have raced with a concurrent cpuset
4784 * update. Just reset the cpus_allowed to the
4785 * cpuset's cpus_allowed
4787 cpumask_copy(new_mask
, cpus_allowed
);
4792 free_cpumask_var(new_mask
);
4793 out_free_cpus_allowed
:
4794 free_cpumask_var(cpus_allowed
);
4801 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4802 struct cpumask
*new_mask
)
4804 if (len
< cpumask_size())
4805 cpumask_clear(new_mask
);
4806 else if (len
> cpumask_size())
4807 len
= cpumask_size();
4809 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4813 * sys_sched_setaffinity - set the cpu affinity of a process
4814 * @pid: pid of the process
4815 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4816 * @user_mask_ptr: user-space pointer to the new cpu mask
4818 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4819 unsigned long __user
*, user_mask_ptr
)
4821 cpumask_var_t new_mask
;
4824 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4827 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4829 retval
= sched_setaffinity(pid
, new_mask
);
4830 free_cpumask_var(new_mask
);
4834 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4836 struct task_struct
*p
;
4837 unsigned long flags
;
4845 p
= find_process_by_pid(pid
);
4849 retval
= security_task_getscheduler(p
);
4853 rq
= task_rq_lock(p
, &flags
);
4854 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4855 task_rq_unlock(rq
, &flags
);
4865 * sys_sched_getaffinity - get the cpu affinity of a process
4866 * @pid: pid of the process
4867 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4868 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4870 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4871 unsigned long __user
*, user_mask_ptr
)
4876 if (len
< cpumask_size())
4879 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4882 ret
= sched_getaffinity(pid
, mask
);
4884 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
4887 ret
= cpumask_size();
4889 free_cpumask_var(mask
);
4895 * sys_sched_yield - yield the current processor to other threads.
4897 * This function yields the current CPU to other tasks. If there are no
4898 * other threads running on this CPU then this function will return.
4900 SYSCALL_DEFINE0(sched_yield
)
4902 struct rq
*rq
= this_rq_lock();
4904 schedstat_inc(rq
, yld_count
);
4905 current
->sched_class
->yield_task(rq
);
4908 * Since we are going to call schedule() anyway, there's
4909 * no need to preempt or enable interrupts:
4911 __release(rq
->lock
);
4912 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4913 do_raw_spin_unlock(&rq
->lock
);
4914 preempt_enable_no_resched();
4921 static inline int should_resched(void)
4923 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4926 static void __cond_resched(void)
4928 add_preempt_count(PREEMPT_ACTIVE
);
4930 sub_preempt_count(PREEMPT_ACTIVE
);
4933 int __sched
_cond_resched(void)
4935 if (should_resched()) {
4941 EXPORT_SYMBOL(_cond_resched
);
4944 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4945 * call schedule, and on return reacquire the lock.
4947 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4948 * operations here to prevent schedule() from being called twice (once via
4949 * spin_unlock(), once by hand).
4951 int __cond_resched_lock(spinlock_t
*lock
)
4953 int resched
= should_resched();
4956 lockdep_assert_held(lock
);
4958 if (spin_needbreak(lock
) || resched
) {
4969 EXPORT_SYMBOL(__cond_resched_lock
);
4971 int __sched
__cond_resched_softirq(void)
4973 BUG_ON(!in_softirq());
4975 if (should_resched()) {
4983 EXPORT_SYMBOL(__cond_resched_softirq
);
4986 * yield - yield the current processor to other threads.
4988 * This is a shortcut for kernel-space yielding - it marks the
4989 * thread runnable and calls sys_sched_yield().
4991 void __sched
yield(void)
4993 set_current_state(TASK_RUNNING
);
4996 EXPORT_SYMBOL(yield
);
4999 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5000 * that process accounting knows that this is a task in IO wait state.
5002 void __sched
io_schedule(void)
5004 struct rq
*rq
= raw_rq();
5006 delayacct_blkio_start();
5007 atomic_inc(&rq
->nr_iowait
);
5008 current
->in_iowait
= 1;
5010 current
->in_iowait
= 0;
5011 atomic_dec(&rq
->nr_iowait
);
5012 delayacct_blkio_end();
5014 EXPORT_SYMBOL(io_schedule
);
5016 long __sched
io_schedule_timeout(long timeout
)
5018 struct rq
*rq
= raw_rq();
5021 delayacct_blkio_start();
5022 atomic_inc(&rq
->nr_iowait
);
5023 current
->in_iowait
= 1;
5024 ret
= schedule_timeout(timeout
);
5025 current
->in_iowait
= 0;
5026 atomic_dec(&rq
->nr_iowait
);
5027 delayacct_blkio_end();
5032 * sys_sched_get_priority_max - return maximum RT priority.
5033 * @policy: scheduling class.
5035 * this syscall returns the maximum rt_priority that can be used
5036 * by a given scheduling class.
5038 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5045 ret
= MAX_USER_RT_PRIO
-1;
5057 * sys_sched_get_priority_min - return minimum RT priority.
5058 * @policy: scheduling class.
5060 * this syscall returns the minimum rt_priority that can be used
5061 * by a given scheduling class.
5063 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5081 * sys_sched_rr_get_interval - return the default timeslice of a process.
5082 * @pid: pid of the process.
5083 * @interval: userspace pointer to the timeslice value.
5085 * this syscall writes the default timeslice value of a given process
5086 * into the user-space timespec buffer. A value of '0' means infinity.
5088 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5089 struct timespec __user
*, interval
)
5091 struct task_struct
*p
;
5092 unsigned int time_slice
;
5093 unsigned long flags
;
5103 p
= find_process_by_pid(pid
);
5107 retval
= security_task_getscheduler(p
);
5111 rq
= task_rq_lock(p
, &flags
);
5112 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5113 task_rq_unlock(rq
, &flags
);
5116 jiffies_to_timespec(time_slice
, &t
);
5117 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5125 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5127 void sched_show_task(struct task_struct
*p
)
5129 unsigned long free
= 0;
5132 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5133 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5134 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5135 #if BITS_PER_LONG == 32
5136 if (state
== TASK_RUNNING
)
5137 printk(KERN_CONT
" running ");
5139 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5141 if (state
== TASK_RUNNING
)
5142 printk(KERN_CONT
" running task ");
5144 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5146 #ifdef CONFIG_DEBUG_STACK_USAGE
5147 free
= stack_not_used(p
);
5149 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5150 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5151 (unsigned long)task_thread_info(p
)->flags
);
5153 show_stack(p
, NULL
);
5156 void show_state_filter(unsigned long state_filter
)
5158 struct task_struct
*g
, *p
;
5160 #if BITS_PER_LONG == 32
5162 " task PC stack pid father\n");
5165 " task PC stack pid father\n");
5167 read_lock(&tasklist_lock
);
5168 do_each_thread(g
, p
) {
5170 * reset the NMI-timeout, listing all files on a slow
5171 * console might take alot of time:
5173 touch_nmi_watchdog();
5174 if (!state_filter
|| (p
->state
& state_filter
))
5176 } while_each_thread(g
, p
);
5178 touch_all_softlockup_watchdogs();
5180 #ifdef CONFIG_SCHED_DEBUG
5181 sysrq_sched_debug_show();
5183 read_unlock(&tasklist_lock
);
5185 * Only show locks if all tasks are dumped:
5188 debug_show_all_locks();
5191 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5193 idle
->sched_class
= &idle_sched_class
;
5197 * init_idle - set up an idle thread for a given CPU
5198 * @idle: task in question
5199 * @cpu: cpu the idle task belongs to
5201 * NOTE: this function does not set the idle thread's NEED_RESCHED
5202 * flag, to make booting more robust.
5204 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5206 struct rq
*rq
= cpu_rq(cpu
);
5207 unsigned long flags
;
5209 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5212 idle
->state
= TASK_RUNNING
;
5213 idle
->se
.exec_start
= sched_clock();
5215 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5216 __set_task_cpu(idle
, cpu
);
5218 rq
->curr
= rq
->idle
= idle
;
5219 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5222 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5224 /* Set the preempt count _outside_ the spinlocks! */
5225 #if defined(CONFIG_PREEMPT)
5226 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5228 task_thread_info(idle
)->preempt_count
= 0;
5231 * The idle tasks have their own, simple scheduling class:
5233 idle
->sched_class
= &idle_sched_class
;
5234 ftrace_graph_init_task(idle
);
5238 * In a system that switches off the HZ timer nohz_cpu_mask
5239 * indicates which cpus entered this state. This is used
5240 * in the rcu update to wait only for active cpus. For system
5241 * which do not switch off the HZ timer nohz_cpu_mask should
5242 * always be CPU_BITS_NONE.
5244 cpumask_var_t nohz_cpu_mask
;
5247 * Increase the granularity value when there are more CPUs,
5248 * because with more CPUs the 'effective latency' as visible
5249 * to users decreases. But the relationship is not linear,
5250 * so pick a second-best guess by going with the log2 of the
5253 * This idea comes from the SD scheduler of Con Kolivas:
5255 static int get_update_sysctl_factor(void)
5257 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5258 unsigned int factor
;
5260 switch (sysctl_sched_tunable_scaling
) {
5261 case SCHED_TUNABLESCALING_NONE
:
5264 case SCHED_TUNABLESCALING_LINEAR
:
5267 case SCHED_TUNABLESCALING_LOG
:
5269 factor
= 1 + ilog2(cpus
);
5276 static void update_sysctl(void)
5278 unsigned int factor
= get_update_sysctl_factor();
5280 #define SET_SYSCTL(name) \
5281 (sysctl_##name = (factor) * normalized_sysctl_##name)
5282 SET_SYSCTL(sched_min_granularity
);
5283 SET_SYSCTL(sched_latency
);
5284 SET_SYSCTL(sched_wakeup_granularity
);
5285 SET_SYSCTL(sched_shares_ratelimit
);
5289 static inline void sched_init_granularity(void)
5296 * This is how migration works:
5298 * 1) we queue a struct migration_req structure in the source CPU's
5299 * runqueue and wake up that CPU's migration thread.
5300 * 2) we down() the locked semaphore => thread blocks.
5301 * 3) migration thread wakes up (implicitly it forces the migrated
5302 * thread off the CPU)
5303 * 4) it gets the migration request and checks whether the migrated
5304 * task is still in the wrong runqueue.
5305 * 5) if it's in the wrong runqueue then the migration thread removes
5306 * it and puts it into the right queue.
5307 * 6) migration thread up()s the semaphore.
5308 * 7) we wake up and the migration is done.
5312 * Change a given task's CPU affinity. Migrate the thread to a
5313 * proper CPU and schedule it away if the CPU it's executing on
5314 * is removed from the allowed bitmask.
5316 * NOTE: the caller must have a valid reference to the task, the
5317 * task must not exit() & deallocate itself prematurely. The
5318 * call is not atomic; no spinlocks may be held.
5320 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5322 struct migration_req req
;
5323 unsigned long flags
;
5327 rq
= task_rq_lock(p
, &flags
);
5329 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5334 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5335 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5340 if (p
->sched_class
->set_cpus_allowed
)
5341 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5343 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5344 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5347 /* Can the task run on the task's current CPU? If so, we're done */
5348 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5351 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5352 /* Need help from migration thread: drop lock and wait. */
5353 struct task_struct
*mt
= rq
->migration_thread
;
5355 get_task_struct(mt
);
5356 task_rq_unlock(rq
, &flags
);
5357 wake_up_process(rq
->migration_thread
);
5358 put_task_struct(mt
);
5359 wait_for_completion(&req
.done
);
5360 tlb_migrate_finish(p
->mm
);
5364 task_rq_unlock(rq
, &flags
);
5368 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5371 * Move (not current) task off this cpu, onto dest cpu. We're doing
5372 * this because either it can't run here any more (set_cpus_allowed()
5373 * away from this CPU, or CPU going down), or because we're
5374 * attempting to rebalance this task on exec (sched_exec).
5376 * So we race with normal scheduler movements, but that's OK, as long
5377 * as the task is no longer on this CPU.
5379 * Returns non-zero if task was successfully migrated.
5381 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5383 struct rq
*rq_dest
, *rq_src
;
5386 if (unlikely(!cpu_active(dest_cpu
)))
5389 rq_src
= cpu_rq(src_cpu
);
5390 rq_dest
= cpu_rq(dest_cpu
);
5392 double_rq_lock(rq_src
, rq_dest
);
5393 /* Already moved. */
5394 if (task_cpu(p
) != src_cpu
)
5396 /* Affinity changed (again). */
5397 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5401 * If we're not on a rq, the next wake-up will ensure we're
5405 deactivate_task(rq_src
, p
, 0);
5406 set_task_cpu(p
, dest_cpu
);
5407 activate_task(rq_dest
, p
, 0);
5408 check_preempt_curr(rq_dest
, p
, 0);
5413 double_rq_unlock(rq_src
, rq_dest
);
5417 #define RCU_MIGRATION_IDLE 0
5418 #define RCU_MIGRATION_NEED_QS 1
5419 #define RCU_MIGRATION_GOT_QS 2
5420 #define RCU_MIGRATION_MUST_SYNC 3
5423 * migration_thread - this is a highprio system thread that performs
5424 * thread migration by bumping thread off CPU then 'pushing' onto
5427 static int migration_thread(void *data
)
5430 int cpu
= (long)data
;
5434 BUG_ON(rq
->migration_thread
!= current
);
5436 set_current_state(TASK_INTERRUPTIBLE
);
5437 while (!kthread_should_stop()) {
5438 struct migration_req
*req
;
5439 struct list_head
*head
;
5441 raw_spin_lock_irq(&rq
->lock
);
5443 if (cpu_is_offline(cpu
)) {
5444 raw_spin_unlock_irq(&rq
->lock
);
5448 if (rq
->active_balance
) {
5449 active_load_balance(rq
, cpu
);
5450 rq
->active_balance
= 0;
5453 head
= &rq
->migration_queue
;
5455 if (list_empty(head
)) {
5456 raw_spin_unlock_irq(&rq
->lock
);
5458 set_current_state(TASK_INTERRUPTIBLE
);
5461 req
= list_entry(head
->next
, struct migration_req
, list
);
5462 list_del_init(head
->next
);
5464 if (req
->task
!= NULL
) {
5465 raw_spin_unlock(&rq
->lock
);
5466 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5467 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5468 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5469 raw_spin_unlock(&rq
->lock
);
5471 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5472 raw_spin_unlock(&rq
->lock
);
5473 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5477 complete(&req
->done
);
5479 __set_current_state(TASK_RUNNING
);
5484 #ifdef CONFIG_HOTPLUG_CPU
5486 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5490 local_irq_disable();
5491 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5497 * Figure out where task on dead CPU should go, use force if necessary.
5499 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5504 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5506 /* It can have affinity changed while we were choosing. */
5507 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
5512 * While a dead CPU has no uninterruptible tasks queued at this point,
5513 * it might still have a nonzero ->nr_uninterruptible counter, because
5514 * for performance reasons the counter is not stricly tracking tasks to
5515 * their home CPUs. So we just add the counter to another CPU's counter,
5516 * to keep the global sum constant after CPU-down:
5518 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5520 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5521 unsigned long flags
;
5523 local_irq_save(flags
);
5524 double_rq_lock(rq_src
, rq_dest
);
5525 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5526 rq_src
->nr_uninterruptible
= 0;
5527 double_rq_unlock(rq_src
, rq_dest
);
5528 local_irq_restore(flags
);
5531 /* Run through task list and migrate tasks from the dead cpu. */
5532 static void migrate_live_tasks(int src_cpu
)
5534 struct task_struct
*p
, *t
;
5536 read_lock(&tasklist_lock
);
5538 do_each_thread(t
, p
) {
5542 if (task_cpu(p
) == src_cpu
)
5543 move_task_off_dead_cpu(src_cpu
, p
);
5544 } while_each_thread(t
, p
);
5546 read_unlock(&tasklist_lock
);
5550 * Schedules idle task to be the next runnable task on current CPU.
5551 * It does so by boosting its priority to highest possible.
5552 * Used by CPU offline code.
5554 void sched_idle_next(void)
5556 int this_cpu
= smp_processor_id();
5557 struct rq
*rq
= cpu_rq(this_cpu
);
5558 struct task_struct
*p
= rq
->idle
;
5559 unsigned long flags
;
5561 /* cpu has to be offline */
5562 BUG_ON(cpu_online(this_cpu
));
5565 * Strictly not necessary since rest of the CPUs are stopped by now
5566 * and interrupts disabled on the current cpu.
5568 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5570 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5572 update_rq_clock(rq
);
5573 activate_task(rq
, p
, 0);
5575 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5579 * Ensures that the idle task is using init_mm right before its cpu goes
5582 void idle_task_exit(void)
5584 struct mm_struct
*mm
= current
->active_mm
;
5586 BUG_ON(cpu_online(smp_processor_id()));
5589 switch_mm(mm
, &init_mm
, current
);
5593 /* called under rq->lock with disabled interrupts */
5594 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5596 struct rq
*rq
= cpu_rq(dead_cpu
);
5598 /* Must be exiting, otherwise would be on tasklist. */
5599 BUG_ON(!p
->exit_state
);
5601 /* Cannot have done final schedule yet: would have vanished. */
5602 BUG_ON(p
->state
== TASK_DEAD
);
5607 * Drop lock around migration; if someone else moves it,
5608 * that's OK. No task can be added to this CPU, so iteration is
5611 raw_spin_unlock_irq(&rq
->lock
);
5612 move_task_off_dead_cpu(dead_cpu
, p
);
5613 raw_spin_lock_irq(&rq
->lock
);
5618 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5619 static void migrate_dead_tasks(unsigned int dead_cpu
)
5621 struct rq
*rq
= cpu_rq(dead_cpu
);
5622 struct task_struct
*next
;
5625 if (!rq
->nr_running
)
5627 update_rq_clock(rq
);
5628 next
= pick_next_task(rq
);
5631 next
->sched_class
->put_prev_task(rq
, next
);
5632 migrate_dead(dead_cpu
, next
);
5638 * remove the tasks which were accounted by rq from calc_load_tasks.
5640 static void calc_global_load_remove(struct rq
*rq
)
5642 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5643 rq
->calc_load_active
= 0;
5645 #endif /* CONFIG_HOTPLUG_CPU */
5647 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5649 static struct ctl_table sd_ctl_dir
[] = {
5651 .procname
= "sched_domain",
5657 static struct ctl_table sd_ctl_root
[] = {
5659 .procname
= "kernel",
5661 .child
= sd_ctl_dir
,
5666 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5668 struct ctl_table
*entry
=
5669 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5674 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5676 struct ctl_table
*entry
;
5679 * In the intermediate directories, both the child directory and
5680 * procname are dynamically allocated and could fail but the mode
5681 * will always be set. In the lowest directory the names are
5682 * static strings and all have proc handlers.
5684 for (entry
= *tablep
; entry
->mode
; entry
++) {
5686 sd_free_ctl_entry(&entry
->child
);
5687 if (entry
->proc_handler
== NULL
)
5688 kfree(entry
->procname
);
5696 set_table_entry(struct ctl_table
*entry
,
5697 const char *procname
, void *data
, int maxlen
,
5698 mode_t mode
, proc_handler
*proc_handler
)
5700 entry
->procname
= procname
;
5702 entry
->maxlen
= maxlen
;
5704 entry
->proc_handler
= proc_handler
;
5707 static struct ctl_table
*
5708 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5710 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5715 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5716 sizeof(long), 0644, proc_doulongvec_minmax
);
5717 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5718 sizeof(long), 0644, proc_doulongvec_minmax
);
5719 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5720 sizeof(int), 0644, proc_dointvec_minmax
);
5721 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5722 sizeof(int), 0644, proc_dointvec_minmax
);
5723 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5724 sizeof(int), 0644, proc_dointvec_minmax
);
5725 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5726 sizeof(int), 0644, proc_dointvec_minmax
);
5727 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5728 sizeof(int), 0644, proc_dointvec_minmax
);
5729 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5730 sizeof(int), 0644, proc_dointvec_minmax
);
5731 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5732 sizeof(int), 0644, proc_dointvec_minmax
);
5733 set_table_entry(&table
[9], "cache_nice_tries",
5734 &sd
->cache_nice_tries
,
5735 sizeof(int), 0644, proc_dointvec_minmax
);
5736 set_table_entry(&table
[10], "flags", &sd
->flags
,
5737 sizeof(int), 0644, proc_dointvec_minmax
);
5738 set_table_entry(&table
[11], "name", sd
->name
,
5739 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5740 /* &table[12] is terminator */
5745 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5747 struct ctl_table
*entry
, *table
;
5748 struct sched_domain
*sd
;
5749 int domain_num
= 0, i
;
5752 for_each_domain(cpu
, sd
)
5754 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5759 for_each_domain(cpu
, sd
) {
5760 snprintf(buf
, 32, "domain%d", i
);
5761 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5763 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5770 static struct ctl_table_header
*sd_sysctl_header
;
5771 static void register_sched_domain_sysctl(void)
5773 int i
, cpu_num
= num_possible_cpus();
5774 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5777 WARN_ON(sd_ctl_dir
[0].child
);
5778 sd_ctl_dir
[0].child
= entry
;
5783 for_each_possible_cpu(i
) {
5784 snprintf(buf
, 32, "cpu%d", i
);
5785 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5787 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5791 WARN_ON(sd_sysctl_header
);
5792 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5795 /* may be called multiple times per register */
5796 static void unregister_sched_domain_sysctl(void)
5798 if (sd_sysctl_header
)
5799 unregister_sysctl_table(sd_sysctl_header
);
5800 sd_sysctl_header
= NULL
;
5801 if (sd_ctl_dir
[0].child
)
5802 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5805 static void register_sched_domain_sysctl(void)
5808 static void unregister_sched_domain_sysctl(void)
5813 static void set_rq_online(struct rq
*rq
)
5816 const struct sched_class
*class;
5818 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5821 for_each_class(class) {
5822 if (class->rq_online
)
5823 class->rq_online(rq
);
5828 static void set_rq_offline(struct rq
*rq
)
5831 const struct sched_class
*class;
5833 for_each_class(class) {
5834 if (class->rq_offline
)
5835 class->rq_offline(rq
);
5838 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5844 * migration_call - callback that gets triggered when a CPU is added.
5845 * Here we can start up the necessary migration thread for the new CPU.
5847 static int __cpuinit
5848 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5850 struct task_struct
*p
;
5851 int cpu
= (long)hcpu
;
5852 unsigned long flags
;
5857 case CPU_UP_PREPARE
:
5858 case CPU_UP_PREPARE_FROZEN
:
5859 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5862 kthread_bind(p
, cpu
);
5863 /* Must be high prio: stop_machine expects to yield to it. */
5864 rq
= task_rq_lock(p
, &flags
);
5865 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5866 task_rq_unlock(rq
, &flags
);
5868 cpu_rq(cpu
)->migration_thread
= p
;
5869 rq
->calc_load_update
= calc_load_update
;
5873 case CPU_ONLINE_FROZEN
:
5874 /* Strictly unnecessary, as first user will wake it. */
5875 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5877 /* Update our root-domain */
5879 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5881 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5885 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5888 #ifdef CONFIG_HOTPLUG_CPU
5889 case CPU_UP_CANCELED
:
5890 case CPU_UP_CANCELED_FROZEN
:
5891 if (!cpu_rq(cpu
)->migration_thread
)
5893 /* Unbind it from offline cpu so it can run. Fall thru. */
5894 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5895 cpumask_any(cpu_online_mask
));
5896 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5897 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5898 cpu_rq(cpu
)->migration_thread
= NULL
;
5902 case CPU_DEAD_FROZEN
:
5903 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5904 migrate_live_tasks(cpu
);
5906 kthread_stop(rq
->migration_thread
);
5907 put_task_struct(rq
->migration_thread
);
5908 rq
->migration_thread
= NULL
;
5909 /* Idle task back to normal (off runqueue, low prio) */
5910 raw_spin_lock_irq(&rq
->lock
);
5911 update_rq_clock(rq
);
5912 deactivate_task(rq
, rq
->idle
, 0);
5913 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5914 rq
->idle
->sched_class
= &idle_sched_class
;
5915 migrate_dead_tasks(cpu
);
5916 raw_spin_unlock_irq(&rq
->lock
);
5918 migrate_nr_uninterruptible(rq
);
5919 BUG_ON(rq
->nr_running
!= 0);
5920 calc_global_load_remove(rq
);
5922 * No need to migrate the tasks: it was best-effort if
5923 * they didn't take sched_hotcpu_mutex. Just wake up
5926 raw_spin_lock_irq(&rq
->lock
);
5927 while (!list_empty(&rq
->migration_queue
)) {
5928 struct migration_req
*req
;
5930 req
= list_entry(rq
->migration_queue
.next
,
5931 struct migration_req
, list
);
5932 list_del_init(&req
->list
);
5933 raw_spin_unlock_irq(&rq
->lock
);
5934 complete(&req
->done
);
5935 raw_spin_lock_irq(&rq
->lock
);
5937 raw_spin_unlock_irq(&rq
->lock
);
5941 case CPU_DYING_FROZEN
:
5942 /* Update our root-domain */
5944 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5946 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5949 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5957 * Register at high priority so that task migration (migrate_all_tasks)
5958 * happens before everything else. This has to be lower priority than
5959 * the notifier in the perf_event subsystem, though.
5961 static struct notifier_block __cpuinitdata migration_notifier
= {
5962 .notifier_call
= migration_call
,
5966 static int __init
migration_init(void)
5968 void *cpu
= (void *)(long)smp_processor_id();
5971 /* Start one for the boot CPU: */
5972 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5973 BUG_ON(err
== NOTIFY_BAD
);
5974 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5975 register_cpu_notifier(&migration_notifier
);
5979 early_initcall(migration_init
);
5984 #ifdef CONFIG_SCHED_DEBUG
5986 static __read_mostly
int sched_domain_debug_enabled
;
5988 static int __init
sched_domain_debug_setup(char *str
)
5990 sched_domain_debug_enabled
= 1;
5994 early_param("sched_debug", sched_domain_debug_setup
);
5996 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5997 struct cpumask
*groupmask
)
5999 struct sched_group
*group
= sd
->groups
;
6002 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
6003 cpumask_clear(groupmask
);
6005 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6007 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6008 printk("does not load-balance\n");
6010 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6015 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6017 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6018 printk(KERN_ERR
"ERROR: domain->span does not contain "
6021 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6022 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6026 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6030 printk(KERN_ERR
"ERROR: group is NULL\n");
6034 if (!group
->cpu_power
) {
6035 printk(KERN_CONT
"\n");
6036 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6041 if (!cpumask_weight(sched_group_cpus(group
))) {
6042 printk(KERN_CONT
"\n");
6043 printk(KERN_ERR
"ERROR: empty group\n");
6047 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6048 printk(KERN_CONT
"\n");
6049 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6053 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6055 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6057 printk(KERN_CONT
" %s", str
);
6058 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6059 printk(KERN_CONT
" (cpu_power = %d)",
6063 group
= group
->next
;
6064 } while (group
!= sd
->groups
);
6065 printk(KERN_CONT
"\n");
6067 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6068 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6071 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6072 printk(KERN_ERR
"ERROR: parent span is not a superset "
6073 "of domain->span\n");
6077 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6079 cpumask_var_t groupmask
;
6082 if (!sched_domain_debug_enabled
)
6086 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6090 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6092 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6093 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6098 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6105 free_cpumask_var(groupmask
);
6107 #else /* !CONFIG_SCHED_DEBUG */
6108 # define sched_domain_debug(sd, cpu) do { } while (0)
6109 #endif /* CONFIG_SCHED_DEBUG */
6111 static int sd_degenerate(struct sched_domain
*sd
)
6113 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6116 /* Following flags need at least 2 groups */
6117 if (sd
->flags
& (SD_LOAD_BALANCE
|
6118 SD_BALANCE_NEWIDLE
|
6122 SD_SHARE_PKG_RESOURCES
)) {
6123 if (sd
->groups
!= sd
->groups
->next
)
6127 /* Following flags don't use groups */
6128 if (sd
->flags
& (SD_WAKE_AFFINE
))
6135 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6137 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6139 if (sd_degenerate(parent
))
6142 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6145 /* Flags needing groups don't count if only 1 group in parent */
6146 if (parent
->groups
== parent
->groups
->next
) {
6147 pflags
&= ~(SD_LOAD_BALANCE
|
6148 SD_BALANCE_NEWIDLE
|
6152 SD_SHARE_PKG_RESOURCES
);
6153 if (nr_node_ids
== 1)
6154 pflags
&= ~SD_SERIALIZE
;
6156 if (~cflags
& pflags
)
6162 static void free_rootdomain(struct root_domain
*rd
)
6164 synchronize_sched();
6166 cpupri_cleanup(&rd
->cpupri
);
6168 free_cpumask_var(rd
->rto_mask
);
6169 free_cpumask_var(rd
->online
);
6170 free_cpumask_var(rd
->span
);
6174 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6176 struct root_domain
*old_rd
= NULL
;
6177 unsigned long flags
;
6179 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6184 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6187 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6190 * If we dont want to free the old_rt yet then
6191 * set old_rd to NULL to skip the freeing later
6194 if (!atomic_dec_and_test(&old_rd
->refcount
))
6198 atomic_inc(&rd
->refcount
);
6201 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6202 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6205 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6208 free_rootdomain(old_rd
);
6211 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6213 gfp_t gfp
= GFP_KERNEL
;
6215 memset(rd
, 0, sizeof(*rd
));
6220 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6222 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6224 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6227 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6232 free_cpumask_var(rd
->rto_mask
);
6234 free_cpumask_var(rd
->online
);
6236 free_cpumask_var(rd
->span
);
6241 static void init_defrootdomain(void)
6243 init_rootdomain(&def_root_domain
, true);
6245 atomic_set(&def_root_domain
.refcount
, 1);
6248 static struct root_domain
*alloc_rootdomain(void)
6250 struct root_domain
*rd
;
6252 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6256 if (init_rootdomain(rd
, false) != 0) {
6265 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6266 * hold the hotplug lock.
6269 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6271 struct rq
*rq
= cpu_rq(cpu
);
6272 struct sched_domain
*tmp
;
6274 /* Remove the sched domains which do not contribute to scheduling. */
6275 for (tmp
= sd
; tmp
; ) {
6276 struct sched_domain
*parent
= tmp
->parent
;
6280 if (sd_parent_degenerate(tmp
, parent
)) {
6281 tmp
->parent
= parent
->parent
;
6283 parent
->parent
->child
= tmp
;
6288 if (sd
&& sd_degenerate(sd
)) {
6294 sched_domain_debug(sd
, cpu
);
6296 rq_attach_root(rq
, rd
);
6297 rcu_assign_pointer(rq
->sd
, sd
);
6300 /* cpus with isolated domains */
6301 static cpumask_var_t cpu_isolated_map
;
6303 /* Setup the mask of cpus configured for isolated domains */
6304 static int __init
isolated_cpu_setup(char *str
)
6306 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6307 cpulist_parse(str
, cpu_isolated_map
);
6311 __setup("isolcpus=", isolated_cpu_setup
);
6314 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6315 * to a function which identifies what group(along with sched group) a CPU
6316 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6317 * (due to the fact that we keep track of groups covered with a struct cpumask).
6319 * init_sched_build_groups will build a circular linked list of the groups
6320 * covered by the given span, and will set each group's ->cpumask correctly,
6321 * and ->cpu_power to 0.
6324 init_sched_build_groups(const struct cpumask
*span
,
6325 const struct cpumask
*cpu_map
,
6326 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6327 struct sched_group
**sg
,
6328 struct cpumask
*tmpmask
),
6329 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6331 struct sched_group
*first
= NULL
, *last
= NULL
;
6334 cpumask_clear(covered
);
6336 for_each_cpu(i
, span
) {
6337 struct sched_group
*sg
;
6338 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6341 if (cpumask_test_cpu(i
, covered
))
6344 cpumask_clear(sched_group_cpus(sg
));
6347 for_each_cpu(j
, span
) {
6348 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6351 cpumask_set_cpu(j
, covered
);
6352 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6363 #define SD_NODES_PER_DOMAIN 16
6368 * find_next_best_node - find the next node to include in a sched_domain
6369 * @node: node whose sched_domain we're building
6370 * @used_nodes: nodes already in the sched_domain
6372 * Find the next node to include in a given scheduling domain. Simply
6373 * finds the closest node not already in the @used_nodes map.
6375 * Should use nodemask_t.
6377 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6379 int i
, n
, val
, min_val
, best_node
= 0;
6383 for (i
= 0; i
< nr_node_ids
; i
++) {
6384 /* Start at @node */
6385 n
= (node
+ i
) % nr_node_ids
;
6387 if (!nr_cpus_node(n
))
6390 /* Skip already used nodes */
6391 if (node_isset(n
, *used_nodes
))
6394 /* Simple min distance search */
6395 val
= node_distance(node
, n
);
6397 if (val
< min_val
) {
6403 node_set(best_node
, *used_nodes
);
6408 * sched_domain_node_span - get a cpumask for a node's sched_domain
6409 * @node: node whose cpumask we're constructing
6410 * @span: resulting cpumask
6412 * Given a node, construct a good cpumask for its sched_domain to span. It
6413 * should be one that prevents unnecessary balancing, but also spreads tasks
6416 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6418 nodemask_t used_nodes
;
6421 cpumask_clear(span
);
6422 nodes_clear(used_nodes
);
6424 cpumask_or(span
, span
, cpumask_of_node(node
));
6425 node_set(node
, used_nodes
);
6427 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6428 int next_node
= find_next_best_node(node
, &used_nodes
);
6430 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6433 #endif /* CONFIG_NUMA */
6435 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6438 * The cpus mask in sched_group and sched_domain hangs off the end.
6440 * ( See the the comments in include/linux/sched.h:struct sched_group
6441 * and struct sched_domain. )
6443 struct static_sched_group
{
6444 struct sched_group sg
;
6445 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6448 struct static_sched_domain
{
6449 struct sched_domain sd
;
6450 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6456 cpumask_var_t domainspan
;
6457 cpumask_var_t covered
;
6458 cpumask_var_t notcovered
;
6460 cpumask_var_t nodemask
;
6461 cpumask_var_t this_sibling_map
;
6462 cpumask_var_t this_core_map
;
6463 cpumask_var_t send_covered
;
6464 cpumask_var_t tmpmask
;
6465 struct sched_group
**sched_group_nodes
;
6466 struct root_domain
*rd
;
6470 sa_sched_groups
= 0,
6475 sa_this_sibling_map
,
6477 sa_sched_group_nodes
,
6487 * SMT sched-domains:
6489 #ifdef CONFIG_SCHED_SMT
6490 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6491 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6494 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6495 struct sched_group
**sg
, struct cpumask
*unused
)
6498 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6501 #endif /* CONFIG_SCHED_SMT */
6504 * multi-core sched-domains:
6506 #ifdef CONFIG_SCHED_MC
6507 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6508 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6509 #endif /* CONFIG_SCHED_MC */
6511 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6513 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6514 struct sched_group
**sg
, struct cpumask
*mask
)
6518 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6519 group
= cpumask_first(mask
);
6521 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6524 #elif defined(CONFIG_SCHED_MC)
6526 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6527 struct sched_group
**sg
, struct cpumask
*unused
)
6530 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6535 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6536 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6539 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6540 struct sched_group
**sg
, struct cpumask
*mask
)
6543 #ifdef CONFIG_SCHED_MC
6544 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6545 group
= cpumask_first(mask
);
6546 #elif defined(CONFIG_SCHED_SMT)
6547 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6548 group
= cpumask_first(mask
);
6553 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6559 * The init_sched_build_groups can't handle what we want to do with node
6560 * groups, so roll our own. Now each node has its own list of groups which
6561 * gets dynamically allocated.
6563 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6564 static struct sched_group
***sched_group_nodes_bycpu
;
6566 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6567 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6569 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6570 struct sched_group
**sg
,
6571 struct cpumask
*nodemask
)
6575 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6576 group
= cpumask_first(nodemask
);
6579 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6583 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6585 struct sched_group
*sg
= group_head
;
6591 for_each_cpu(j
, sched_group_cpus(sg
)) {
6592 struct sched_domain
*sd
;
6594 sd
= &per_cpu(phys_domains
, j
).sd
;
6595 if (j
!= group_first_cpu(sd
->groups
)) {
6597 * Only add "power" once for each
6603 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6606 } while (sg
!= group_head
);
6609 static int build_numa_sched_groups(struct s_data
*d
,
6610 const struct cpumask
*cpu_map
, int num
)
6612 struct sched_domain
*sd
;
6613 struct sched_group
*sg
, *prev
;
6616 cpumask_clear(d
->covered
);
6617 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6618 if (cpumask_empty(d
->nodemask
)) {
6619 d
->sched_group_nodes
[num
] = NULL
;
6623 sched_domain_node_span(num
, d
->domainspan
);
6624 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6626 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6629 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6633 d
->sched_group_nodes
[num
] = sg
;
6635 for_each_cpu(j
, d
->nodemask
) {
6636 sd
= &per_cpu(node_domains
, j
).sd
;
6641 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6643 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6646 for (j
= 0; j
< nr_node_ids
; j
++) {
6647 n
= (num
+ j
) % nr_node_ids
;
6648 cpumask_complement(d
->notcovered
, d
->covered
);
6649 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6650 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6651 if (cpumask_empty(d
->tmpmask
))
6653 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6654 if (cpumask_empty(d
->tmpmask
))
6656 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6660 "Can not alloc domain group for node %d\n", j
);
6664 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6665 sg
->next
= prev
->next
;
6666 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6673 #endif /* CONFIG_NUMA */
6676 /* Free memory allocated for various sched_group structures */
6677 static void free_sched_groups(const struct cpumask
*cpu_map
,
6678 struct cpumask
*nodemask
)
6682 for_each_cpu(cpu
, cpu_map
) {
6683 struct sched_group
**sched_group_nodes
6684 = sched_group_nodes_bycpu
[cpu
];
6686 if (!sched_group_nodes
)
6689 for (i
= 0; i
< nr_node_ids
; i
++) {
6690 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6692 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6693 if (cpumask_empty(nodemask
))
6703 if (oldsg
!= sched_group_nodes
[i
])
6706 kfree(sched_group_nodes
);
6707 sched_group_nodes_bycpu
[cpu
] = NULL
;
6710 #else /* !CONFIG_NUMA */
6711 static void free_sched_groups(const struct cpumask
*cpu_map
,
6712 struct cpumask
*nodemask
)
6715 #endif /* CONFIG_NUMA */
6718 * Initialize sched groups cpu_power.
6720 * cpu_power indicates the capacity of sched group, which is used while
6721 * distributing the load between different sched groups in a sched domain.
6722 * Typically cpu_power for all the groups in a sched domain will be same unless
6723 * there are asymmetries in the topology. If there are asymmetries, group
6724 * having more cpu_power will pickup more load compared to the group having
6727 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6729 struct sched_domain
*child
;
6730 struct sched_group
*group
;
6734 WARN_ON(!sd
|| !sd
->groups
);
6736 if (cpu
!= group_first_cpu(sd
->groups
))
6741 sd
->groups
->cpu_power
= 0;
6744 power
= SCHED_LOAD_SCALE
;
6745 weight
= cpumask_weight(sched_domain_span(sd
));
6747 * SMT siblings share the power of a single core.
6748 * Usually multiple threads get a better yield out of
6749 * that one core than a single thread would have,
6750 * reflect that in sd->smt_gain.
6752 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6753 power
*= sd
->smt_gain
;
6755 power
>>= SCHED_LOAD_SHIFT
;
6757 sd
->groups
->cpu_power
+= power
;
6762 * Add cpu_power of each child group to this groups cpu_power.
6764 group
= child
->groups
;
6766 sd
->groups
->cpu_power
+= group
->cpu_power
;
6767 group
= group
->next
;
6768 } while (group
!= child
->groups
);
6772 * Initializers for schedule domains
6773 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6776 #ifdef CONFIG_SCHED_DEBUG
6777 # define SD_INIT_NAME(sd, type) sd->name = #type
6779 # define SD_INIT_NAME(sd, type) do { } while (0)
6782 #define SD_INIT(sd, type) sd_init_##type(sd)
6784 #define SD_INIT_FUNC(type) \
6785 static noinline void sd_init_##type(struct sched_domain *sd) \
6787 memset(sd, 0, sizeof(*sd)); \
6788 *sd = SD_##type##_INIT; \
6789 sd->level = SD_LV_##type; \
6790 SD_INIT_NAME(sd, type); \
6795 SD_INIT_FUNC(ALLNODES
)
6798 #ifdef CONFIG_SCHED_SMT
6799 SD_INIT_FUNC(SIBLING
)
6801 #ifdef CONFIG_SCHED_MC
6805 static int default_relax_domain_level
= -1;
6807 static int __init
setup_relax_domain_level(char *str
)
6811 val
= simple_strtoul(str
, NULL
, 0);
6812 if (val
< SD_LV_MAX
)
6813 default_relax_domain_level
= val
;
6817 __setup("relax_domain_level=", setup_relax_domain_level
);
6819 static void set_domain_attribute(struct sched_domain
*sd
,
6820 struct sched_domain_attr
*attr
)
6824 if (!attr
|| attr
->relax_domain_level
< 0) {
6825 if (default_relax_domain_level
< 0)
6828 request
= default_relax_domain_level
;
6830 request
= attr
->relax_domain_level
;
6831 if (request
< sd
->level
) {
6832 /* turn off idle balance on this domain */
6833 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6835 /* turn on idle balance on this domain */
6836 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6840 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6841 const struct cpumask
*cpu_map
)
6844 case sa_sched_groups
:
6845 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6846 d
->sched_group_nodes
= NULL
;
6848 free_rootdomain(d
->rd
); /* fall through */
6850 free_cpumask_var(d
->tmpmask
); /* fall through */
6851 case sa_send_covered
:
6852 free_cpumask_var(d
->send_covered
); /* fall through */
6853 case sa_this_core_map
:
6854 free_cpumask_var(d
->this_core_map
); /* fall through */
6855 case sa_this_sibling_map
:
6856 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6858 free_cpumask_var(d
->nodemask
); /* fall through */
6859 case sa_sched_group_nodes
:
6861 kfree(d
->sched_group_nodes
); /* fall through */
6863 free_cpumask_var(d
->notcovered
); /* fall through */
6865 free_cpumask_var(d
->covered
); /* fall through */
6867 free_cpumask_var(d
->domainspan
); /* fall through */
6874 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6875 const struct cpumask
*cpu_map
)
6878 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6880 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6881 return sa_domainspan
;
6882 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6884 /* Allocate the per-node list of sched groups */
6885 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6886 sizeof(struct sched_group
*), GFP_KERNEL
);
6887 if (!d
->sched_group_nodes
) {
6888 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6889 return sa_notcovered
;
6891 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6893 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6894 return sa_sched_group_nodes
;
6895 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6897 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6898 return sa_this_sibling_map
;
6899 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6900 return sa_this_core_map
;
6901 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6902 return sa_send_covered
;
6903 d
->rd
= alloc_rootdomain();
6905 printk(KERN_WARNING
"Cannot alloc root domain\n");
6908 return sa_rootdomain
;
6911 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6912 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6914 struct sched_domain
*sd
= NULL
;
6916 struct sched_domain
*parent
;
6919 if (cpumask_weight(cpu_map
) >
6920 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6921 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6922 SD_INIT(sd
, ALLNODES
);
6923 set_domain_attribute(sd
, attr
);
6924 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6925 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6930 sd
= &per_cpu(node_domains
, i
).sd
;
6932 set_domain_attribute(sd
, attr
);
6933 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6934 sd
->parent
= parent
;
6937 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6942 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6943 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6944 struct sched_domain
*parent
, int i
)
6946 struct sched_domain
*sd
;
6947 sd
= &per_cpu(phys_domains
, i
).sd
;
6949 set_domain_attribute(sd
, attr
);
6950 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6951 sd
->parent
= parent
;
6954 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6958 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6959 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6960 struct sched_domain
*parent
, int i
)
6962 struct sched_domain
*sd
= parent
;
6963 #ifdef CONFIG_SCHED_MC
6964 sd
= &per_cpu(core_domains
, i
).sd
;
6966 set_domain_attribute(sd
, attr
);
6967 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6968 sd
->parent
= parent
;
6970 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6975 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6976 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6977 struct sched_domain
*parent
, int i
)
6979 struct sched_domain
*sd
= parent
;
6980 #ifdef CONFIG_SCHED_SMT
6981 sd
= &per_cpu(cpu_domains
, i
).sd
;
6982 SD_INIT(sd
, SIBLING
);
6983 set_domain_attribute(sd
, attr
);
6984 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6985 sd
->parent
= parent
;
6987 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6992 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6993 const struct cpumask
*cpu_map
, int cpu
)
6996 #ifdef CONFIG_SCHED_SMT
6997 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6998 cpumask_and(d
->this_sibling_map
, cpu_map
,
6999 topology_thread_cpumask(cpu
));
7000 if (cpu
== cpumask_first(d
->this_sibling_map
))
7001 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
7003 d
->send_covered
, d
->tmpmask
);
7006 #ifdef CONFIG_SCHED_MC
7007 case SD_LV_MC
: /* set up multi-core groups */
7008 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7009 if (cpu
== cpumask_first(d
->this_core_map
))
7010 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7012 d
->send_covered
, d
->tmpmask
);
7015 case SD_LV_CPU
: /* set up physical groups */
7016 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7017 if (!cpumask_empty(d
->nodemask
))
7018 init_sched_build_groups(d
->nodemask
, cpu_map
,
7020 d
->send_covered
, d
->tmpmask
);
7023 case SD_LV_ALLNODES
:
7024 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7025 d
->send_covered
, d
->tmpmask
);
7034 * Build sched domains for a given set of cpus and attach the sched domains
7035 * to the individual cpus
7037 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7038 struct sched_domain_attr
*attr
)
7040 enum s_alloc alloc_state
= sa_none
;
7042 struct sched_domain
*sd
;
7048 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7049 if (alloc_state
!= sa_rootdomain
)
7051 alloc_state
= sa_sched_groups
;
7054 * Set up domains for cpus specified by the cpu_map.
7056 for_each_cpu(i
, cpu_map
) {
7057 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7060 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7061 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7062 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7063 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7066 for_each_cpu(i
, cpu_map
) {
7067 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7068 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7071 /* Set up physical groups */
7072 for (i
= 0; i
< nr_node_ids
; i
++)
7073 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7076 /* Set up node groups */
7078 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7080 for (i
= 0; i
< nr_node_ids
; i
++)
7081 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7085 /* Calculate CPU power for physical packages and nodes */
7086 #ifdef CONFIG_SCHED_SMT
7087 for_each_cpu(i
, cpu_map
) {
7088 sd
= &per_cpu(cpu_domains
, i
).sd
;
7089 init_sched_groups_power(i
, sd
);
7092 #ifdef CONFIG_SCHED_MC
7093 for_each_cpu(i
, cpu_map
) {
7094 sd
= &per_cpu(core_domains
, i
).sd
;
7095 init_sched_groups_power(i
, sd
);
7099 for_each_cpu(i
, cpu_map
) {
7100 sd
= &per_cpu(phys_domains
, i
).sd
;
7101 init_sched_groups_power(i
, sd
);
7105 for (i
= 0; i
< nr_node_ids
; i
++)
7106 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7108 if (d
.sd_allnodes
) {
7109 struct sched_group
*sg
;
7111 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7113 init_numa_sched_groups_power(sg
);
7117 /* Attach the domains */
7118 for_each_cpu(i
, cpu_map
) {
7119 #ifdef CONFIG_SCHED_SMT
7120 sd
= &per_cpu(cpu_domains
, i
).sd
;
7121 #elif defined(CONFIG_SCHED_MC)
7122 sd
= &per_cpu(core_domains
, i
).sd
;
7124 sd
= &per_cpu(phys_domains
, i
).sd
;
7126 cpu_attach_domain(sd
, d
.rd
, i
);
7129 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7130 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7134 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7138 static int build_sched_domains(const struct cpumask
*cpu_map
)
7140 return __build_sched_domains(cpu_map
, NULL
);
7143 static cpumask_var_t
*doms_cur
; /* current sched domains */
7144 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7145 static struct sched_domain_attr
*dattr_cur
;
7146 /* attribues of custom domains in 'doms_cur' */
7149 * Special case: If a kmalloc of a doms_cur partition (array of
7150 * cpumask) fails, then fallback to a single sched domain,
7151 * as determined by the single cpumask fallback_doms.
7153 static cpumask_var_t fallback_doms
;
7156 * arch_update_cpu_topology lets virtualized architectures update the
7157 * cpu core maps. It is supposed to return 1 if the topology changed
7158 * or 0 if it stayed the same.
7160 int __attribute__((weak
)) arch_update_cpu_topology(void)
7165 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7168 cpumask_var_t
*doms
;
7170 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7173 for (i
= 0; i
< ndoms
; i
++) {
7174 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7175 free_sched_domains(doms
, i
);
7182 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7185 for (i
= 0; i
< ndoms
; i
++)
7186 free_cpumask_var(doms
[i
]);
7191 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7192 * For now this just excludes isolated cpus, but could be used to
7193 * exclude other special cases in the future.
7195 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7199 arch_update_cpu_topology();
7201 doms_cur
= alloc_sched_domains(ndoms_cur
);
7203 doms_cur
= &fallback_doms
;
7204 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7206 err
= build_sched_domains(doms_cur
[0]);
7207 register_sched_domain_sysctl();
7212 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7213 struct cpumask
*tmpmask
)
7215 free_sched_groups(cpu_map
, tmpmask
);
7219 * Detach sched domains from a group of cpus specified in cpu_map
7220 * These cpus will now be attached to the NULL domain
7222 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7224 /* Save because hotplug lock held. */
7225 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7228 for_each_cpu(i
, cpu_map
)
7229 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7230 synchronize_sched();
7231 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7234 /* handle null as "default" */
7235 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7236 struct sched_domain_attr
*new, int idx_new
)
7238 struct sched_domain_attr tmp
;
7245 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7246 new ? (new + idx_new
) : &tmp
,
7247 sizeof(struct sched_domain_attr
));
7251 * Partition sched domains as specified by the 'ndoms_new'
7252 * cpumasks in the array doms_new[] of cpumasks. This compares
7253 * doms_new[] to the current sched domain partitioning, doms_cur[].
7254 * It destroys each deleted domain and builds each new domain.
7256 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7257 * The masks don't intersect (don't overlap.) We should setup one
7258 * sched domain for each mask. CPUs not in any of the cpumasks will
7259 * not be load balanced. If the same cpumask appears both in the
7260 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7263 * The passed in 'doms_new' should be allocated using
7264 * alloc_sched_domains. This routine takes ownership of it and will
7265 * free_sched_domains it when done with it. If the caller failed the
7266 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7267 * and partition_sched_domains() will fallback to the single partition
7268 * 'fallback_doms', it also forces the domains to be rebuilt.
7270 * If doms_new == NULL it will be replaced with cpu_online_mask.
7271 * ndoms_new == 0 is a special case for destroying existing domains,
7272 * and it will not create the default domain.
7274 * Call with hotplug lock held
7276 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7277 struct sched_domain_attr
*dattr_new
)
7282 mutex_lock(&sched_domains_mutex
);
7284 /* always unregister in case we don't destroy any domains */
7285 unregister_sched_domain_sysctl();
7287 /* Let architecture update cpu core mappings. */
7288 new_topology
= arch_update_cpu_topology();
7290 n
= doms_new
? ndoms_new
: 0;
7292 /* Destroy deleted domains */
7293 for (i
= 0; i
< ndoms_cur
; i
++) {
7294 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7295 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7296 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7299 /* no match - a current sched domain not in new doms_new[] */
7300 detach_destroy_domains(doms_cur
[i
]);
7305 if (doms_new
== NULL
) {
7307 doms_new
= &fallback_doms
;
7308 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7309 WARN_ON_ONCE(dattr_new
);
7312 /* Build new domains */
7313 for (i
= 0; i
< ndoms_new
; i
++) {
7314 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7315 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7316 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7319 /* no match - add a new doms_new */
7320 __build_sched_domains(doms_new
[i
],
7321 dattr_new
? dattr_new
+ i
: NULL
);
7326 /* Remember the new sched domains */
7327 if (doms_cur
!= &fallback_doms
)
7328 free_sched_domains(doms_cur
, ndoms_cur
);
7329 kfree(dattr_cur
); /* kfree(NULL) is safe */
7330 doms_cur
= doms_new
;
7331 dattr_cur
= dattr_new
;
7332 ndoms_cur
= ndoms_new
;
7334 register_sched_domain_sysctl();
7336 mutex_unlock(&sched_domains_mutex
);
7339 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7340 static void arch_reinit_sched_domains(void)
7344 /* Destroy domains first to force the rebuild */
7345 partition_sched_domains(0, NULL
, NULL
);
7347 rebuild_sched_domains();
7351 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7353 unsigned int level
= 0;
7355 if (sscanf(buf
, "%u", &level
) != 1)
7359 * level is always be positive so don't check for
7360 * level < POWERSAVINGS_BALANCE_NONE which is 0
7361 * What happens on 0 or 1 byte write,
7362 * need to check for count as well?
7365 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7369 sched_smt_power_savings
= level
;
7371 sched_mc_power_savings
= level
;
7373 arch_reinit_sched_domains();
7378 #ifdef CONFIG_SCHED_MC
7379 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7382 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7384 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7385 const char *buf
, size_t count
)
7387 return sched_power_savings_store(buf
, count
, 0);
7389 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7390 sched_mc_power_savings_show
,
7391 sched_mc_power_savings_store
);
7394 #ifdef CONFIG_SCHED_SMT
7395 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7398 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7400 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7401 const char *buf
, size_t count
)
7403 return sched_power_savings_store(buf
, count
, 1);
7405 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7406 sched_smt_power_savings_show
,
7407 sched_smt_power_savings_store
);
7410 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7414 #ifdef CONFIG_SCHED_SMT
7416 err
= sysfs_create_file(&cls
->kset
.kobj
,
7417 &attr_sched_smt_power_savings
.attr
);
7419 #ifdef CONFIG_SCHED_MC
7420 if (!err
&& mc_capable())
7421 err
= sysfs_create_file(&cls
->kset
.kobj
,
7422 &attr_sched_mc_power_savings
.attr
);
7426 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7428 #ifndef CONFIG_CPUSETS
7430 * Add online and remove offline CPUs from the scheduler domains.
7431 * When cpusets are enabled they take over this function.
7433 static int update_sched_domains(struct notifier_block
*nfb
,
7434 unsigned long action
, void *hcpu
)
7438 case CPU_ONLINE_FROZEN
:
7439 case CPU_DOWN_PREPARE
:
7440 case CPU_DOWN_PREPARE_FROZEN
:
7441 case CPU_DOWN_FAILED
:
7442 case CPU_DOWN_FAILED_FROZEN
:
7443 partition_sched_domains(1, NULL
, NULL
);
7452 static int update_runtime(struct notifier_block
*nfb
,
7453 unsigned long action
, void *hcpu
)
7455 int cpu
= (int)(long)hcpu
;
7458 case CPU_DOWN_PREPARE
:
7459 case CPU_DOWN_PREPARE_FROZEN
:
7460 disable_runtime(cpu_rq(cpu
));
7463 case CPU_DOWN_FAILED
:
7464 case CPU_DOWN_FAILED_FROZEN
:
7466 case CPU_ONLINE_FROZEN
:
7467 enable_runtime(cpu_rq(cpu
));
7475 void __init
sched_init_smp(void)
7477 cpumask_var_t non_isolated_cpus
;
7479 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7480 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7482 #if defined(CONFIG_NUMA)
7483 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7485 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7488 mutex_lock(&sched_domains_mutex
);
7489 arch_init_sched_domains(cpu_active_mask
);
7490 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7491 if (cpumask_empty(non_isolated_cpus
))
7492 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7493 mutex_unlock(&sched_domains_mutex
);
7496 #ifndef CONFIG_CPUSETS
7497 /* XXX: Theoretical race here - CPU may be hotplugged now */
7498 hotcpu_notifier(update_sched_domains
, 0);
7501 /* RT runtime code needs to handle some hotplug events */
7502 hotcpu_notifier(update_runtime
, 0);
7506 /* Move init over to a non-isolated CPU */
7507 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7509 sched_init_granularity();
7510 free_cpumask_var(non_isolated_cpus
);
7512 init_sched_rt_class();
7515 void __init
sched_init_smp(void)
7517 sched_init_granularity();
7519 #endif /* CONFIG_SMP */
7521 const_debug
unsigned int sysctl_timer_migration
= 1;
7523 int in_sched_functions(unsigned long addr
)
7525 return in_lock_functions(addr
) ||
7526 (addr
>= (unsigned long)__sched_text_start
7527 && addr
< (unsigned long)__sched_text_end
);
7530 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7532 cfs_rq
->tasks_timeline
= RB_ROOT
;
7533 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7534 #ifdef CONFIG_FAIR_GROUP_SCHED
7537 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7540 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7542 struct rt_prio_array
*array
;
7545 array
= &rt_rq
->active
;
7546 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7547 INIT_LIST_HEAD(array
->queue
+ i
);
7548 __clear_bit(i
, array
->bitmap
);
7550 /* delimiter for bitsearch: */
7551 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7553 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7554 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7556 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7560 rt_rq
->rt_nr_migratory
= 0;
7561 rt_rq
->overloaded
= 0;
7562 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7566 rt_rq
->rt_throttled
= 0;
7567 rt_rq
->rt_runtime
= 0;
7568 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7570 #ifdef CONFIG_RT_GROUP_SCHED
7571 rt_rq
->rt_nr_boosted
= 0;
7576 #ifdef CONFIG_FAIR_GROUP_SCHED
7577 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7578 struct sched_entity
*se
, int cpu
, int add
,
7579 struct sched_entity
*parent
)
7581 struct rq
*rq
= cpu_rq(cpu
);
7582 tg
->cfs_rq
[cpu
] = cfs_rq
;
7583 init_cfs_rq(cfs_rq
, rq
);
7586 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7589 /* se could be NULL for init_task_group */
7594 se
->cfs_rq
= &rq
->cfs
;
7596 se
->cfs_rq
= parent
->my_q
;
7599 se
->load
.weight
= tg
->shares
;
7600 se
->load
.inv_weight
= 0;
7601 se
->parent
= parent
;
7605 #ifdef CONFIG_RT_GROUP_SCHED
7606 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7607 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7608 struct sched_rt_entity
*parent
)
7610 struct rq
*rq
= cpu_rq(cpu
);
7612 tg
->rt_rq
[cpu
] = rt_rq
;
7613 init_rt_rq(rt_rq
, rq
);
7615 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7617 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7619 tg
->rt_se
[cpu
] = rt_se
;
7624 rt_se
->rt_rq
= &rq
->rt
;
7626 rt_se
->rt_rq
= parent
->my_q
;
7628 rt_se
->my_q
= rt_rq
;
7629 rt_se
->parent
= parent
;
7630 INIT_LIST_HEAD(&rt_se
->run_list
);
7634 void __init
sched_init(void)
7637 unsigned long alloc_size
= 0, ptr
;
7639 #ifdef CONFIG_FAIR_GROUP_SCHED
7640 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7642 #ifdef CONFIG_RT_GROUP_SCHED
7643 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7645 #ifdef CONFIG_CPUMASK_OFFSTACK
7646 alloc_size
+= num_possible_cpus() * cpumask_size();
7649 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7651 #ifdef CONFIG_FAIR_GROUP_SCHED
7652 init_task_group
.se
= (struct sched_entity
**)ptr
;
7653 ptr
+= nr_cpu_ids
* sizeof(void **);
7655 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7656 ptr
+= nr_cpu_ids
* sizeof(void **);
7658 #endif /* CONFIG_FAIR_GROUP_SCHED */
7659 #ifdef CONFIG_RT_GROUP_SCHED
7660 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7661 ptr
+= nr_cpu_ids
* sizeof(void **);
7663 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7664 ptr
+= nr_cpu_ids
* sizeof(void **);
7666 #endif /* CONFIG_RT_GROUP_SCHED */
7667 #ifdef CONFIG_CPUMASK_OFFSTACK
7668 for_each_possible_cpu(i
) {
7669 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7670 ptr
+= cpumask_size();
7672 #endif /* CONFIG_CPUMASK_OFFSTACK */
7676 init_defrootdomain();
7679 init_rt_bandwidth(&def_rt_bandwidth
,
7680 global_rt_period(), global_rt_runtime());
7682 #ifdef CONFIG_RT_GROUP_SCHED
7683 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7684 global_rt_period(), global_rt_runtime());
7685 #endif /* CONFIG_RT_GROUP_SCHED */
7687 #ifdef CONFIG_CGROUP_SCHED
7688 list_add(&init_task_group
.list
, &task_groups
);
7689 INIT_LIST_HEAD(&init_task_group
.children
);
7691 #endif /* CONFIG_CGROUP_SCHED */
7693 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7694 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7695 __alignof__(unsigned long));
7697 for_each_possible_cpu(i
) {
7701 raw_spin_lock_init(&rq
->lock
);
7703 rq
->calc_load_active
= 0;
7704 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7705 init_cfs_rq(&rq
->cfs
, rq
);
7706 init_rt_rq(&rq
->rt
, rq
);
7707 #ifdef CONFIG_FAIR_GROUP_SCHED
7708 init_task_group
.shares
= init_task_group_load
;
7709 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7710 #ifdef CONFIG_CGROUP_SCHED
7712 * How much cpu bandwidth does init_task_group get?
7714 * In case of task-groups formed thr' the cgroup filesystem, it
7715 * gets 100% of the cpu resources in the system. This overall
7716 * system cpu resource is divided among the tasks of
7717 * init_task_group and its child task-groups in a fair manner,
7718 * based on each entity's (task or task-group's) weight
7719 * (se->load.weight).
7721 * In other words, if init_task_group has 10 tasks of weight
7722 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7723 * then A0's share of the cpu resource is:
7725 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7727 * We achieve this by letting init_task_group's tasks sit
7728 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7730 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7732 #endif /* CONFIG_FAIR_GROUP_SCHED */
7734 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7735 #ifdef CONFIG_RT_GROUP_SCHED
7736 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7737 #ifdef CONFIG_CGROUP_SCHED
7738 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7742 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7743 rq
->cpu_load
[j
] = 0;
7747 rq
->post_schedule
= 0;
7748 rq
->active_balance
= 0;
7749 rq
->next_balance
= jiffies
;
7753 rq
->migration_thread
= NULL
;
7755 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7756 INIT_LIST_HEAD(&rq
->migration_queue
);
7757 rq_attach_root(rq
, &def_root_domain
);
7760 atomic_set(&rq
->nr_iowait
, 0);
7763 set_load_weight(&init_task
);
7765 #ifdef CONFIG_PREEMPT_NOTIFIERS
7766 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7770 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7773 #ifdef CONFIG_RT_MUTEXES
7774 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7778 * The boot idle thread does lazy MMU switching as well:
7780 atomic_inc(&init_mm
.mm_count
);
7781 enter_lazy_tlb(&init_mm
, current
);
7784 * Make us the idle thread. Technically, schedule() should not be
7785 * called from this thread, however somewhere below it might be,
7786 * but because we are the idle thread, we just pick up running again
7787 * when this runqueue becomes "idle".
7789 init_idle(current
, smp_processor_id());
7791 calc_load_update
= jiffies
+ LOAD_FREQ
;
7794 * During early bootup we pretend to be a normal task:
7796 current
->sched_class
= &fair_sched_class
;
7798 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7799 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7802 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7803 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7805 /* May be allocated at isolcpus cmdline parse time */
7806 if (cpu_isolated_map
== NULL
)
7807 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7812 scheduler_running
= 1;
7815 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7816 static inline int preempt_count_equals(int preempt_offset
)
7818 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7820 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7823 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7826 static unsigned long prev_jiffy
; /* ratelimiting */
7828 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7829 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7831 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7833 prev_jiffy
= jiffies
;
7836 "BUG: sleeping function called from invalid context at %s:%d\n",
7839 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7840 in_atomic(), irqs_disabled(),
7841 current
->pid
, current
->comm
);
7843 debug_show_held_locks(current
);
7844 if (irqs_disabled())
7845 print_irqtrace_events(current
);
7849 EXPORT_SYMBOL(__might_sleep
);
7852 #ifdef CONFIG_MAGIC_SYSRQ
7853 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7857 update_rq_clock(rq
);
7858 on_rq
= p
->se
.on_rq
;
7860 deactivate_task(rq
, p
, 0);
7861 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7863 activate_task(rq
, p
, 0);
7864 resched_task(rq
->curr
);
7868 void normalize_rt_tasks(void)
7870 struct task_struct
*g
, *p
;
7871 unsigned long flags
;
7874 read_lock_irqsave(&tasklist_lock
, flags
);
7875 do_each_thread(g
, p
) {
7877 * Only normalize user tasks:
7882 p
->se
.exec_start
= 0;
7883 #ifdef CONFIG_SCHEDSTATS
7884 p
->se
.statistics
.wait_start
= 0;
7885 p
->se
.statistics
.sleep_start
= 0;
7886 p
->se
.statistics
.block_start
= 0;
7891 * Renice negative nice level userspace
7894 if (TASK_NICE(p
) < 0 && p
->mm
)
7895 set_user_nice(p
, 0);
7899 raw_spin_lock(&p
->pi_lock
);
7900 rq
= __task_rq_lock(p
);
7902 normalize_task(rq
, p
);
7904 __task_rq_unlock(rq
);
7905 raw_spin_unlock(&p
->pi_lock
);
7906 } while_each_thread(g
, p
);
7908 read_unlock_irqrestore(&tasklist_lock
, flags
);
7911 #endif /* CONFIG_MAGIC_SYSRQ */
7915 * These functions are only useful for the IA64 MCA handling.
7917 * They can only be called when the whole system has been
7918 * stopped - every CPU needs to be quiescent, and no scheduling
7919 * activity can take place. Using them for anything else would
7920 * be a serious bug, and as a result, they aren't even visible
7921 * under any other configuration.
7925 * curr_task - return the current task for a given cpu.
7926 * @cpu: the processor in question.
7928 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7930 struct task_struct
*curr_task(int cpu
)
7932 return cpu_curr(cpu
);
7936 * set_curr_task - set the current task for a given cpu.
7937 * @cpu: the processor in question.
7938 * @p: the task pointer to set.
7940 * Description: This function must only be used when non-maskable interrupts
7941 * are serviced on a separate stack. It allows the architecture to switch the
7942 * notion of the current task on a cpu in a non-blocking manner. This function
7943 * must be called with all CPU's synchronized, and interrupts disabled, the
7944 * and caller must save the original value of the current task (see
7945 * curr_task() above) and restore that value before reenabling interrupts and
7946 * re-starting the system.
7948 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7950 void set_curr_task(int cpu
, struct task_struct
*p
)
7957 #ifdef CONFIG_FAIR_GROUP_SCHED
7958 static void free_fair_sched_group(struct task_group
*tg
)
7962 for_each_possible_cpu(i
) {
7964 kfree(tg
->cfs_rq
[i
]);
7974 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7976 struct cfs_rq
*cfs_rq
;
7977 struct sched_entity
*se
;
7981 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7984 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7988 tg
->shares
= NICE_0_LOAD
;
7990 for_each_possible_cpu(i
) {
7993 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7994 GFP_KERNEL
, cpu_to_node(i
));
7998 se
= kzalloc_node(sizeof(struct sched_entity
),
7999 GFP_KERNEL
, cpu_to_node(i
));
8003 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8014 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8016 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8017 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8020 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8022 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8024 #else /* !CONFG_FAIR_GROUP_SCHED */
8025 static inline void free_fair_sched_group(struct task_group
*tg
)
8030 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8035 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8039 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8042 #endif /* CONFIG_FAIR_GROUP_SCHED */
8044 #ifdef CONFIG_RT_GROUP_SCHED
8045 static void free_rt_sched_group(struct task_group
*tg
)
8049 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8051 for_each_possible_cpu(i
) {
8053 kfree(tg
->rt_rq
[i
]);
8055 kfree(tg
->rt_se
[i
]);
8063 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8065 struct rt_rq
*rt_rq
;
8066 struct sched_rt_entity
*rt_se
;
8070 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8073 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8077 init_rt_bandwidth(&tg
->rt_bandwidth
,
8078 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8080 for_each_possible_cpu(i
) {
8083 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8084 GFP_KERNEL
, cpu_to_node(i
));
8088 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8089 GFP_KERNEL
, cpu_to_node(i
));
8093 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8104 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8106 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8107 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8110 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8112 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8114 #else /* !CONFIG_RT_GROUP_SCHED */
8115 static inline void free_rt_sched_group(struct task_group
*tg
)
8120 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8125 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8129 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8132 #endif /* CONFIG_RT_GROUP_SCHED */
8134 #ifdef CONFIG_CGROUP_SCHED
8135 static void free_sched_group(struct task_group
*tg
)
8137 free_fair_sched_group(tg
);
8138 free_rt_sched_group(tg
);
8142 /* allocate runqueue etc for a new task group */
8143 struct task_group
*sched_create_group(struct task_group
*parent
)
8145 struct task_group
*tg
;
8146 unsigned long flags
;
8149 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8151 return ERR_PTR(-ENOMEM
);
8153 if (!alloc_fair_sched_group(tg
, parent
))
8156 if (!alloc_rt_sched_group(tg
, parent
))
8159 spin_lock_irqsave(&task_group_lock
, flags
);
8160 for_each_possible_cpu(i
) {
8161 register_fair_sched_group(tg
, i
);
8162 register_rt_sched_group(tg
, i
);
8164 list_add_rcu(&tg
->list
, &task_groups
);
8166 WARN_ON(!parent
); /* root should already exist */
8168 tg
->parent
= parent
;
8169 INIT_LIST_HEAD(&tg
->children
);
8170 list_add_rcu(&tg
->siblings
, &parent
->children
);
8171 spin_unlock_irqrestore(&task_group_lock
, flags
);
8176 free_sched_group(tg
);
8177 return ERR_PTR(-ENOMEM
);
8180 /* rcu callback to free various structures associated with a task group */
8181 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8183 /* now it should be safe to free those cfs_rqs */
8184 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8187 /* Destroy runqueue etc associated with a task group */
8188 void sched_destroy_group(struct task_group
*tg
)
8190 unsigned long flags
;
8193 spin_lock_irqsave(&task_group_lock
, flags
);
8194 for_each_possible_cpu(i
) {
8195 unregister_fair_sched_group(tg
, i
);
8196 unregister_rt_sched_group(tg
, i
);
8198 list_del_rcu(&tg
->list
);
8199 list_del_rcu(&tg
->siblings
);
8200 spin_unlock_irqrestore(&task_group_lock
, flags
);
8202 /* wait for possible concurrent references to cfs_rqs complete */
8203 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8206 /* change task's runqueue when it moves between groups.
8207 * The caller of this function should have put the task in its new group
8208 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8209 * reflect its new group.
8211 void sched_move_task(struct task_struct
*tsk
)
8214 unsigned long flags
;
8217 rq
= task_rq_lock(tsk
, &flags
);
8219 update_rq_clock(rq
);
8221 running
= task_current(rq
, tsk
);
8222 on_rq
= tsk
->se
.on_rq
;
8225 dequeue_task(rq
, tsk
, 0);
8226 if (unlikely(running
))
8227 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8229 set_task_rq(tsk
, task_cpu(tsk
));
8231 #ifdef CONFIG_FAIR_GROUP_SCHED
8232 if (tsk
->sched_class
->moved_group
)
8233 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8236 if (unlikely(running
))
8237 tsk
->sched_class
->set_curr_task(rq
);
8239 enqueue_task(rq
, tsk
, 0, false);
8241 task_rq_unlock(rq
, &flags
);
8243 #endif /* CONFIG_CGROUP_SCHED */
8245 #ifdef CONFIG_FAIR_GROUP_SCHED
8246 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8248 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8253 dequeue_entity(cfs_rq
, se
, 0);
8255 se
->load
.weight
= shares
;
8256 se
->load
.inv_weight
= 0;
8259 enqueue_entity(cfs_rq
, se
, 0);
8262 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8264 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8265 struct rq
*rq
= cfs_rq
->rq
;
8266 unsigned long flags
;
8268 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8269 __set_se_shares(se
, shares
);
8270 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8273 static DEFINE_MUTEX(shares_mutex
);
8275 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8278 unsigned long flags
;
8281 * We can't change the weight of the root cgroup.
8286 if (shares
< MIN_SHARES
)
8287 shares
= MIN_SHARES
;
8288 else if (shares
> MAX_SHARES
)
8289 shares
= MAX_SHARES
;
8291 mutex_lock(&shares_mutex
);
8292 if (tg
->shares
== shares
)
8295 spin_lock_irqsave(&task_group_lock
, flags
);
8296 for_each_possible_cpu(i
)
8297 unregister_fair_sched_group(tg
, i
);
8298 list_del_rcu(&tg
->siblings
);
8299 spin_unlock_irqrestore(&task_group_lock
, flags
);
8301 /* wait for any ongoing reference to this group to finish */
8302 synchronize_sched();
8305 * Now we are free to modify the group's share on each cpu
8306 * w/o tripping rebalance_share or load_balance_fair.
8308 tg
->shares
= shares
;
8309 for_each_possible_cpu(i
) {
8313 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8314 set_se_shares(tg
->se
[i
], shares
);
8318 * Enable load balance activity on this group, by inserting it back on
8319 * each cpu's rq->leaf_cfs_rq_list.
8321 spin_lock_irqsave(&task_group_lock
, flags
);
8322 for_each_possible_cpu(i
)
8323 register_fair_sched_group(tg
, i
);
8324 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8325 spin_unlock_irqrestore(&task_group_lock
, flags
);
8327 mutex_unlock(&shares_mutex
);
8331 unsigned long sched_group_shares(struct task_group
*tg
)
8337 #ifdef CONFIG_RT_GROUP_SCHED
8339 * Ensure that the real time constraints are schedulable.
8341 static DEFINE_MUTEX(rt_constraints_mutex
);
8343 static unsigned long to_ratio(u64 period
, u64 runtime
)
8345 if (runtime
== RUNTIME_INF
)
8348 return div64_u64(runtime
<< 20, period
);
8351 /* Must be called with tasklist_lock held */
8352 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8354 struct task_struct
*g
, *p
;
8356 do_each_thread(g
, p
) {
8357 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8359 } while_each_thread(g
, p
);
8364 struct rt_schedulable_data
{
8365 struct task_group
*tg
;
8370 static int tg_schedulable(struct task_group
*tg
, void *data
)
8372 struct rt_schedulable_data
*d
= data
;
8373 struct task_group
*child
;
8374 unsigned long total
, sum
= 0;
8375 u64 period
, runtime
;
8377 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8378 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8381 period
= d
->rt_period
;
8382 runtime
= d
->rt_runtime
;
8386 * Cannot have more runtime than the period.
8388 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8392 * Ensure we don't starve existing RT tasks.
8394 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8397 total
= to_ratio(period
, runtime
);
8400 * Nobody can have more than the global setting allows.
8402 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8406 * The sum of our children's runtime should not exceed our own.
8408 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8409 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8410 runtime
= child
->rt_bandwidth
.rt_runtime
;
8412 if (child
== d
->tg
) {
8413 period
= d
->rt_period
;
8414 runtime
= d
->rt_runtime
;
8417 sum
+= to_ratio(period
, runtime
);
8426 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8428 struct rt_schedulable_data data
= {
8430 .rt_period
= period
,
8431 .rt_runtime
= runtime
,
8434 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8437 static int tg_set_bandwidth(struct task_group
*tg
,
8438 u64 rt_period
, u64 rt_runtime
)
8442 mutex_lock(&rt_constraints_mutex
);
8443 read_lock(&tasklist_lock
);
8444 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8448 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8449 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8450 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8452 for_each_possible_cpu(i
) {
8453 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8455 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8456 rt_rq
->rt_runtime
= rt_runtime
;
8457 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8459 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8461 read_unlock(&tasklist_lock
);
8462 mutex_unlock(&rt_constraints_mutex
);
8467 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8469 u64 rt_runtime
, rt_period
;
8471 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8472 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8473 if (rt_runtime_us
< 0)
8474 rt_runtime
= RUNTIME_INF
;
8476 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8479 long sched_group_rt_runtime(struct task_group
*tg
)
8483 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8486 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8487 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8488 return rt_runtime_us
;
8491 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8493 u64 rt_runtime
, rt_period
;
8495 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8496 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8501 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8504 long sched_group_rt_period(struct task_group
*tg
)
8508 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8509 do_div(rt_period_us
, NSEC_PER_USEC
);
8510 return rt_period_us
;
8513 static int sched_rt_global_constraints(void)
8515 u64 runtime
, period
;
8518 if (sysctl_sched_rt_period
<= 0)
8521 runtime
= global_rt_runtime();
8522 period
= global_rt_period();
8525 * Sanity check on the sysctl variables.
8527 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8530 mutex_lock(&rt_constraints_mutex
);
8531 read_lock(&tasklist_lock
);
8532 ret
= __rt_schedulable(NULL
, 0, 0);
8533 read_unlock(&tasklist_lock
);
8534 mutex_unlock(&rt_constraints_mutex
);
8539 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8541 /* Don't accept realtime tasks when there is no way for them to run */
8542 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8548 #else /* !CONFIG_RT_GROUP_SCHED */
8549 static int sched_rt_global_constraints(void)
8551 unsigned long flags
;
8554 if (sysctl_sched_rt_period
<= 0)
8558 * There's always some RT tasks in the root group
8559 * -- migration, kstopmachine etc..
8561 if (sysctl_sched_rt_runtime
== 0)
8564 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8565 for_each_possible_cpu(i
) {
8566 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8568 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8569 rt_rq
->rt_runtime
= global_rt_runtime();
8570 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8572 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8576 #endif /* CONFIG_RT_GROUP_SCHED */
8578 int sched_rt_handler(struct ctl_table
*table
, int write
,
8579 void __user
*buffer
, size_t *lenp
,
8583 int old_period
, old_runtime
;
8584 static DEFINE_MUTEX(mutex
);
8587 old_period
= sysctl_sched_rt_period
;
8588 old_runtime
= sysctl_sched_rt_runtime
;
8590 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8592 if (!ret
&& write
) {
8593 ret
= sched_rt_global_constraints();
8595 sysctl_sched_rt_period
= old_period
;
8596 sysctl_sched_rt_runtime
= old_runtime
;
8598 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8599 def_rt_bandwidth
.rt_period
=
8600 ns_to_ktime(global_rt_period());
8603 mutex_unlock(&mutex
);
8608 #ifdef CONFIG_CGROUP_SCHED
8610 /* return corresponding task_group object of a cgroup */
8611 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8613 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8614 struct task_group
, css
);
8617 static struct cgroup_subsys_state
*
8618 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8620 struct task_group
*tg
, *parent
;
8622 if (!cgrp
->parent
) {
8623 /* This is early initialization for the top cgroup */
8624 return &init_task_group
.css
;
8627 parent
= cgroup_tg(cgrp
->parent
);
8628 tg
= sched_create_group(parent
);
8630 return ERR_PTR(-ENOMEM
);
8636 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8638 struct task_group
*tg
= cgroup_tg(cgrp
);
8640 sched_destroy_group(tg
);
8644 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8646 #ifdef CONFIG_RT_GROUP_SCHED
8647 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8650 /* We don't support RT-tasks being in separate groups */
8651 if (tsk
->sched_class
!= &fair_sched_class
)
8658 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8659 struct task_struct
*tsk
, bool threadgroup
)
8661 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8665 struct task_struct
*c
;
8667 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8668 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8680 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8681 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8684 sched_move_task(tsk
);
8686 struct task_struct
*c
;
8688 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8695 #ifdef CONFIG_FAIR_GROUP_SCHED
8696 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8699 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8702 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8704 struct task_group
*tg
= cgroup_tg(cgrp
);
8706 return (u64
) tg
->shares
;
8708 #endif /* CONFIG_FAIR_GROUP_SCHED */
8710 #ifdef CONFIG_RT_GROUP_SCHED
8711 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8714 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8717 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8719 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8722 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8725 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8728 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8730 return sched_group_rt_period(cgroup_tg(cgrp
));
8732 #endif /* CONFIG_RT_GROUP_SCHED */
8734 static struct cftype cpu_files
[] = {
8735 #ifdef CONFIG_FAIR_GROUP_SCHED
8738 .read_u64
= cpu_shares_read_u64
,
8739 .write_u64
= cpu_shares_write_u64
,
8742 #ifdef CONFIG_RT_GROUP_SCHED
8744 .name
= "rt_runtime_us",
8745 .read_s64
= cpu_rt_runtime_read
,
8746 .write_s64
= cpu_rt_runtime_write
,
8749 .name
= "rt_period_us",
8750 .read_u64
= cpu_rt_period_read_uint
,
8751 .write_u64
= cpu_rt_period_write_uint
,
8756 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8758 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8761 struct cgroup_subsys cpu_cgroup_subsys
= {
8763 .create
= cpu_cgroup_create
,
8764 .destroy
= cpu_cgroup_destroy
,
8765 .can_attach
= cpu_cgroup_can_attach
,
8766 .attach
= cpu_cgroup_attach
,
8767 .populate
= cpu_cgroup_populate
,
8768 .subsys_id
= cpu_cgroup_subsys_id
,
8772 #endif /* CONFIG_CGROUP_SCHED */
8774 #ifdef CONFIG_CGROUP_CPUACCT
8777 * CPU accounting code for task groups.
8779 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8780 * (balbir@in.ibm.com).
8783 /* track cpu usage of a group of tasks and its child groups */
8785 struct cgroup_subsys_state css
;
8786 /* cpuusage holds pointer to a u64-type object on every cpu */
8788 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8789 struct cpuacct
*parent
;
8792 struct cgroup_subsys cpuacct_subsys
;
8794 /* return cpu accounting group corresponding to this container */
8795 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8797 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8798 struct cpuacct
, css
);
8801 /* return cpu accounting group to which this task belongs */
8802 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8804 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8805 struct cpuacct
, css
);
8808 /* create a new cpu accounting group */
8809 static struct cgroup_subsys_state
*cpuacct_create(
8810 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8812 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8818 ca
->cpuusage
= alloc_percpu(u64
);
8822 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8823 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8824 goto out_free_counters
;
8827 ca
->parent
= cgroup_ca(cgrp
->parent
);
8833 percpu_counter_destroy(&ca
->cpustat
[i
]);
8834 free_percpu(ca
->cpuusage
);
8838 return ERR_PTR(-ENOMEM
);
8841 /* destroy an existing cpu accounting group */
8843 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8845 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8848 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8849 percpu_counter_destroy(&ca
->cpustat
[i
]);
8850 free_percpu(ca
->cpuusage
);
8854 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8856 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8859 #ifndef CONFIG_64BIT
8861 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8863 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8865 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8873 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8875 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8877 #ifndef CONFIG_64BIT
8879 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8881 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8883 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8889 /* return total cpu usage (in nanoseconds) of a group */
8890 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8892 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8893 u64 totalcpuusage
= 0;
8896 for_each_present_cpu(i
)
8897 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8899 return totalcpuusage
;
8902 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8905 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8914 for_each_present_cpu(i
)
8915 cpuacct_cpuusage_write(ca
, i
, 0);
8921 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8924 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8928 for_each_present_cpu(i
) {
8929 percpu
= cpuacct_cpuusage_read(ca
, i
);
8930 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8932 seq_printf(m
, "\n");
8936 static const char *cpuacct_stat_desc
[] = {
8937 [CPUACCT_STAT_USER
] = "user",
8938 [CPUACCT_STAT_SYSTEM
] = "system",
8941 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8942 struct cgroup_map_cb
*cb
)
8944 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8947 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8948 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8949 val
= cputime64_to_clock_t(val
);
8950 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8955 static struct cftype files
[] = {
8958 .read_u64
= cpuusage_read
,
8959 .write_u64
= cpuusage_write
,
8962 .name
= "usage_percpu",
8963 .read_seq_string
= cpuacct_percpu_seq_read
,
8967 .read_map
= cpuacct_stats_show
,
8971 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8973 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8977 * charge this task's execution time to its accounting group.
8979 * called with rq->lock held.
8981 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8986 if (unlikely(!cpuacct_subsys
.active
))
8989 cpu
= task_cpu(tsk
);
8995 for (; ca
; ca
= ca
->parent
) {
8996 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8997 *cpuusage
+= cputime
;
9004 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9005 * in cputime_t units. As a result, cpuacct_update_stats calls
9006 * percpu_counter_add with values large enough to always overflow the
9007 * per cpu batch limit causing bad SMP scalability.
9009 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9010 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9011 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9014 #define CPUACCT_BATCH \
9015 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9017 #define CPUACCT_BATCH 0
9021 * Charge the system/user time to the task's accounting group.
9023 static void cpuacct_update_stats(struct task_struct
*tsk
,
9024 enum cpuacct_stat_index idx
, cputime_t val
)
9027 int batch
= CPUACCT_BATCH
;
9029 if (unlikely(!cpuacct_subsys
.active
))
9036 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9042 struct cgroup_subsys cpuacct_subsys
= {
9044 .create
= cpuacct_create
,
9045 .destroy
= cpuacct_destroy
,
9046 .populate
= cpuacct_populate
,
9047 .subsys_id
= cpuacct_subsys_id
,
9049 #endif /* CONFIG_CGROUP_CPUACCT */
9053 int rcu_expedited_torture_stats(char *page
)
9057 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9059 void synchronize_sched_expedited(void)
9062 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9064 #else /* #ifndef CONFIG_SMP */
9066 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9067 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9069 #define RCU_EXPEDITED_STATE_POST -2
9070 #define RCU_EXPEDITED_STATE_IDLE -1
9072 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9074 int rcu_expedited_torture_stats(char *page
)
9079 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9080 for_each_online_cpu(cpu
) {
9081 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9082 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9084 cnt
+= sprintf(&page
[cnt
], "\n");
9087 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9089 static long synchronize_sched_expedited_count
;
9092 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9093 * approach to force grace period to end quickly. This consumes
9094 * significant time on all CPUs, and is thus not recommended for
9095 * any sort of common-case code.
9097 * Note that it is illegal to call this function while holding any
9098 * lock that is acquired by a CPU-hotplug notifier. Failing to
9099 * observe this restriction will result in deadlock.
9101 void synchronize_sched_expedited(void)
9104 unsigned long flags
;
9105 bool need_full_sync
= 0;
9107 struct migration_req
*req
;
9111 smp_mb(); /* ensure prior mod happens before capturing snap. */
9112 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9114 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9116 if (trycount
++ < 10)
9117 udelay(trycount
* num_online_cpus());
9119 synchronize_sched();
9122 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9123 smp_mb(); /* ensure test happens before caller kfree */
9128 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9129 for_each_online_cpu(cpu
) {
9131 req
= &per_cpu(rcu_migration_req
, cpu
);
9132 init_completion(&req
->done
);
9134 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9135 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9136 list_add(&req
->list
, &rq
->migration_queue
);
9137 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9138 wake_up_process(rq
->migration_thread
);
9140 for_each_online_cpu(cpu
) {
9141 rcu_expedited_state
= cpu
;
9142 req
= &per_cpu(rcu_migration_req
, cpu
);
9144 wait_for_completion(&req
->done
);
9145 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9146 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9148 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9149 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9151 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9152 synchronize_sched_expedited_count
++;
9153 mutex_unlock(&rcu_sched_expedited_mutex
);
9156 synchronize_sched();
9158 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9160 #endif /* #else #ifndef CONFIG_SMP */