4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/perf_event.h>
43 #include <linux/security.h>
44 #include <linux/notifier.h>
45 #include <linux/profile.h>
46 #include <linux/freezer.h>
47 #include <linux/vmalloc.h>
48 #include <linux/blkdev.h>
49 #include <linux/delay.h>
50 #include <linux/pid_namespace.h>
51 #include <linux/smp.h>
52 #include <linux/threads.h>
53 #include <linux/timer.h>
54 #include <linux/rcupdate.h>
55 #include <linux/cpu.h>
56 #include <linux/cpuset.h>
57 #include <linux/percpu.h>
58 #include <linux/kthread.h>
59 #include <linux/proc_fs.h>
60 #include <linux/seq_file.h>
61 #include <linux/sysctl.h>
62 #include <linux/syscalls.h>
63 #include <linux/times.h>
64 #include <linux/tsacct_kern.h>
65 #include <linux/kprobes.h>
66 #include <linux/delayacct.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
73 #include <linux/ftrace.h>
74 #include <linux/slab.h>
77 #include <asm/irq_regs.h>
79 #include "sched_cpupri.h"
81 #define CREATE_TRACE_POINTS
82 #include <trace/events/sched.h>
85 * Convert user-nice values [ -20 ... 0 ... 19 ]
86 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
89 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
90 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
91 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
94 * 'User priority' is the nice value converted to something we
95 * can work with better when scaling various scheduler parameters,
96 * it's a [ 0 ... 39 ] range.
98 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
99 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
100 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
103 * Helpers for converting nanosecond timing to jiffy resolution
105 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
107 #define NICE_0_LOAD SCHED_LOAD_SCALE
108 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 * These are the 'tuning knobs' of the scheduler:
113 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
114 * Timeslices get refilled after they expire.
116 #define DEF_TIMESLICE (100 * HZ / 1000)
119 * single value that denotes runtime == period, ie unlimited time.
121 #define RUNTIME_INF ((u64)~0ULL)
123 static inline int rt_policy(int policy
)
125 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
130 static inline int task_has_rt_policy(struct task_struct
*p
)
132 return rt_policy(p
->policy
);
136 * This is the priority-queue data structure of the RT scheduling class:
138 struct rt_prio_array
{
139 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
140 struct list_head queue
[MAX_RT_PRIO
];
143 struct rt_bandwidth
{
144 /* nests inside the rq lock: */
145 raw_spinlock_t rt_runtime_lock
;
148 struct hrtimer rt_period_timer
;
151 static struct rt_bandwidth def_rt_bandwidth
;
153 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
155 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
157 struct rt_bandwidth
*rt_b
=
158 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
164 now
= hrtimer_cb_get_time(timer
);
165 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
170 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
173 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
177 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
179 rt_b
->rt_period
= ns_to_ktime(period
);
180 rt_b
->rt_runtime
= runtime
;
182 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
184 hrtimer_init(&rt_b
->rt_period_timer
,
185 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
186 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
189 static inline int rt_bandwidth_enabled(void)
191 return sysctl_sched_rt_runtime
>= 0;
194 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
198 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
201 if (hrtimer_active(&rt_b
->rt_period_timer
))
204 raw_spin_lock(&rt_b
->rt_runtime_lock
);
209 if (hrtimer_active(&rt_b
->rt_period_timer
))
212 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
213 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
215 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
216 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
217 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
218 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
219 HRTIMER_MODE_ABS_PINNED
, 0);
221 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
224 #ifdef CONFIG_RT_GROUP_SCHED
225 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
227 hrtimer_cancel(&rt_b
->rt_period_timer
);
232 * sched_domains_mutex serializes calls to arch_init_sched_domains,
233 * detach_destroy_domains and partition_sched_domains.
235 static DEFINE_MUTEX(sched_domains_mutex
);
237 #ifdef CONFIG_CGROUP_SCHED
239 #include <linux/cgroup.h>
243 static LIST_HEAD(task_groups
);
245 /* task group related information */
247 struct cgroup_subsys_state css
;
249 #ifdef CONFIG_FAIR_GROUP_SCHED
250 /* schedulable entities of this group on each cpu */
251 struct sched_entity
**se
;
252 /* runqueue "owned" by this group on each cpu */
253 struct cfs_rq
**cfs_rq
;
254 unsigned long shares
;
257 #ifdef CONFIG_RT_GROUP_SCHED
258 struct sched_rt_entity
**rt_se
;
259 struct rt_rq
**rt_rq
;
261 struct rt_bandwidth rt_bandwidth
;
265 struct list_head list
;
267 struct task_group
*parent
;
268 struct list_head siblings
;
269 struct list_head children
;
272 #define root_task_group init_task_group
274 /* task_group_lock serializes add/remove of task groups and also changes to
275 * a task group's cpu shares.
277 static DEFINE_SPINLOCK(task_group_lock
);
279 #ifdef CONFIG_FAIR_GROUP_SCHED
282 static int root_task_group_empty(void)
284 return list_empty(&root_task_group
.children
);
288 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
291 * A weight of 0 or 1 can cause arithmetics problems.
292 * A weight of a cfs_rq is the sum of weights of which entities
293 * are queued on this cfs_rq, so a weight of a entity should not be
294 * too large, so as the shares value of a task group.
295 * (The default weight is 1024 - so there's no practical
296 * limitation from this.)
299 #define MAX_SHARES (1UL << 18)
301 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
304 /* Default task group.
305 * Every task in system belong to this group at bootup.
307 struct task_group init_task_group
;
309 /* return group to which a task belongs */
310 static inline struct task_group
*task_group(struct task_struct
*p
)
312 struct task_group
*tg
;
314 #ifdef CONFIG_CGROUP_SCHED
315 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
316 struct task_group
, css
);
318 tg
= &init_task_group
;
323 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
324 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
326 #ifdef CONFIG_FAIR_GROUP_SCHED
327 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
328 p
->se
.parent
= task_group(p
)->se
[cpu
];
331 #ifdef CONFIG_RT_GROUP_SCHED
332 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
333 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
339 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
340 static inline struct task_group
*task_group(struct task_struct
*p
)
345 #endif /* CONFIG_CGROUP_SCHED */
347 /* CFS-related fields in a runqueue */
349 struct load_weight load
;
350 unsigned long nr_running
;
355 struct rb_root tasks_timeline
;
356 struct rb_node
*rb_leftmost
;
358 struct list_head tasks
;
359 struct list_head
*balance_iterator
;
362 * 'curr' points to currently running entity on this cfs_rq.
363 * It is set to NULL otherwise (i.e when none are currently running).
365 struct sched_entity
*curr
, *next
, *last
;
367 unsigned int nr_spread_over
;
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
373 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
374 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
375 * (like users, containers etc.)
377 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
378 * list is used during load balance.
380 struct list_head leaf_cfs_rq_list
;
381 struct task_group
*tg
; /* group that "owns" this runqueue */
385 * the part of load.weight contributed by tasks
387 unsigned long task_weight
;
390 * h_load = weight * f(tg)
392 * Where f(tg) is the recursive weight fraction assigned to
395 unsigned long h_load
;
398 * this cpu's part of tg->shares
400 unsigned long shares
;
403 * load.weight at the time we set shares
405 unsigned long rq_weight
;
410 /* Real-Time classes' related field in a runqueue: */
412 struct rt_prio_array active
;
413 unsigned long rt_nr_running
;
414 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
416 int curr
; /* highest queued rt task prio */
418 int next
; /* next highest */
423 unsigned long rt_nr_migratory
;
424 unsigned long rt_nr_total
;
426 struct plist_head pushable_tasks
;
431 /* Nests inside the rq lock: */
432 raw_spinlock_t rt_runtime_lock
;
434 #ifdef CONFIG_RT_GROUP_SCHED
435 unsigned long rt_nr_boosted
;
438 struct list_head leaf_rt_rq_list
;
439 struct task_group
*tg
;
446 * We add the notion of a root-domain which will be used to define per-domain
447 * variables. Each exclusive cpuset essentially defines an island domain by
448 * fully partitioning the member cpus from any other cpuset. Whenever a new
449 * exclusive cpuset is created, we also create and attach a new root-domain
456 cpumask_var_t online
;
459 * The "RT overload" flag: it gets set if a CPU has more than
460 * one runnable RT task.
462 cpumask_var_t rto_mask
;
465 struct cpupri cpupri
;
470 * By default the system creates a single root-domain with all cpus as
471 * members (mimicking the global state we have today).
473 static struct root_domain def_root_domain
;
478 * This is the main, per-CPU runqueue data structure.
480 * Locking rule: those places that want to lock multiple runqueues
481 * (such as the load balancing or the thread migration code), lock
482 * acquire operations must be ordered by ascending &runqueue.
489 * nr_running and cpu_load should be in the same cacheline because
490 * remote CPUs use both these fields when doing load calculation.
492 unsigned long nr_running
;
493 #define CPU_LOAD_IDX_MAX 5
494 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
496 unsigned char in_nohz_recently
;
498 /* capture load from *all* tasks on this cpu: */
499 struct load_weight load
;
500 unsigned long nr_load_updates
;
506 #ifdef CONFIG_FAIR_GROUP_SCHED
507 /* list of leaf cfs_rq on this cpu: */
508 struct list_head leaf_cfs_rq_list
;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 struct list_head leaf_rt_rq_list
;
515 * This is part of a global counter where only the total sum
516 * over all CPUs matters. A task can increase this counter on
517 * one CPU and if it got migrated afterwards it may decrease
518 * it on another CPU. Always updated under the runqueue lock:
520 unsigned long nr_uninterruptible
;
522 struct task_struct
*curr
, *idle
;
523 unsigned long next_balance
;
524 struct mm_struct
*prev_mm
;
531 struct root_domain
*rd
;
532 struct sched_domain
*sd
;
534 unsigned char idle_at_tick
;
535 /* For active balancing */
539 /* cpu of this runqueue: */
543 unsigned long avg_load_per_task
;
545 struct task_struct
*migration_thread
;
546 struct list_head migration_queue
;
554 /* calc_load related fields */
555 unsigned long calc_load_update
;
556 long calc_load_active
;
558 #ifdef CONFIG_SCHED_HRTICK
560 int hrtick_csd_pending
;
561 struct call_single_data hrtick_csd
;
563 struct hrtimer hrtick_timer
;
566 #ifdef CONFIG_SCHEDSTATS
568 struct sched_info rq_sched_info
;
569 unsigned long long rq_cpu_time
;
570 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
572 /* sys_sched_yield() stats */
573 unsigned int yld_count
;
575 /* schedule() stats */
576 unsigned int sched_switch
;
577 unsigned int sched_count
;
578 unsigned int sched_goidle
;
580 /* try_to_wake_up() stats */
581 unsigned int ttwu_count
;
582 unsigned int ttwu_local
;
585 unsigned int bkl_count
;
589 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
592 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
594 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
597 static inline int cpu_of(struct rq
*rq
)
606 #define rcu_dereference_check_sched_domain(p) \
607 rcu_dereference_check((p), \
608 rcu_read_lock_sched_held() || \
609 lockdep_is_held(&sched_domains_mutex))
612 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
613 * See detach_destroy_domains: synchronize_sched for details.
615 * The domain tree of any CPU may only be accessed from within
616 * preempt-disabled sections.
618 #define for_each_domain(cpu, __sd) \
619 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
621 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
622 #define this_rq() (&__get_cpu_var(runqueues))
623 #define task_rq(p) cpu_rq(task_cpu(p))
624 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
625 #define raw_rq() (&__raw_get_cpu_var(runqueues))
627 inline void update_rq_clock(struct rq
*rq
)
629 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
633 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
635 #ifdef CONFIG_SCHED_DEBUG
636 # define const_debug __read_mostly
638 # define const_debug static const
643 * @cpu: the processor in question.
645 * Returns true if the current cpu runqueue is locked.
646 * This interface allows printk to be called with the runqueue lock
647 * held and know whether or not it is OK to wake up the klogd.
649 int runqueue_is_locked(int cpu
)
651 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
655 * Debugging: various feature bits
658 #define SCHED_FEAT(name, enabled) \
659 __SCHED_FEAT_##name ,
662 #include "sched_features.h"
667 #define SCHED_FEAT(name, enabled) \
668 (1UL << __SCHED_FEAT_##name) * enabled |
670 const_debug
unsigned int sysctl_sched_features
=
671 #include "sched_features.h"
676 #ifdef CONFIG_SCHED_DEBUG
677 #define SCHED_FEAT(name, enabled) \
680 static __read_mostly
char *sched_feat_names
[] = {
681 #include "sched_features.h"
687 static int sched_feat_show(struct seq_file
*m
, void *v
)
691 for (i
= 0; sched_feat_names
[i
]; i
++) {
692 if (!(sysctl_sched_features
& (1UL << i
)))
694 seq_printf(m
, "%s ", sched_feat_names
[i
]);
702 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
703 size_t cnt
, loff_t
*ppos
)
713 if (copy_from_user(&buf
, ubuf
, cnt
))
718 if (strncmp(buf
, "NO_", 3) == 0) {
723 for (i
= 0; sched_feat_names
[i
]; i
++) {
724 int len
= strlen(sched_feat_names
[i
]);
726 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
728 sysctl_sched_features
&= ~(1UL << i
);
730 sysctl_sched_features
|= (1UL << i
);
735 if (!sched_feat_names
[i
])
743 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
745 return single_open(filp
, sched_feat_show
, NULL
);
748 static const struct file_operations sched_feat_fops
= {
749 .open
= sched_feat_open
,
750 .write
= sched_feat_write
,
753 .release
= single_release
,
756 static __init
int sched_init_debug(void)
758 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
763 late_initcall(sched_init_debug
);
767 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
770 * Number of tasks to iterate in a single balance run.
771 * Limited because this is done with IRQs disabled.
773 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
776 * ratelimit for updating the group shares.
779 unsigned int sysctl_sched_shares_ratelimit
= 250000;
780 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
783 * Inject some fuzzyness into changing the per-cpu group shares
784 * this avoids remote rq-locks at the expense of fairness.
787 unsigned int sysctl_sched_shares_thresh
= 4;
790 * period over which we average the RT time consumption, measured
795 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
798 * period over which we measure -rt task cpu usage in us.
801 unsigned int sysctl_sched_rt_period
= 1000000;
803 static __read_mostly
int scheduler_running
;
806 * part of the period that we allow rt tasks to run in us.
809 int sysctl_sched_rt_runtime
= 950000;
811 static inline u64
global_rt_period(void)
813 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
816 static inline u64
global_rt_runtime(void)
818 if (sysctl_sched_rt_runtime
< 0)
821 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
824 #ifndef prepare_arch_switch
825 # define prepare_arch_switch(next) do { } while (0)
827 #ifndef finish_arch_switch
828 # define finish_arch_switch(prev) do { } while (0)
831 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
833 return rq
->curr
== p
;
836 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
837 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
839 return task_current(rq
, p
);
842 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
846 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
848 #ifdef CONFIG_DEBUG_SPINLOCK
849 /* this is a valid case when another task releases the spinlock */
850 rq
->lock
.owner
= current
;
853 * If we are tracking spinlock dependencies then we have to
854 * fix up the runqueue lock - which gets 'carried over' from
857 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
859 raw_spin_unlock_irq(&rq
->lock
);
862 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
863 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
868 return task_current(rq
, p
);
872 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
876 * We can optimise this out completely for !SMP, because the
877 * SMP rebalancing from interrupt is the only thing that cares
882 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
883 raw_spin_unlock_irq(&rq
->lock
);
885 raw_spin_unlock(&rq
->lock
);
889 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
893 * After ->oncpu is cleared, the task can be moved to a different CPU.
894 * We must ensure this doesn't happen until the switch is completely
900 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
904 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
907 * Check whether the task is waking, we use this to synchronize against
908 * ttwu() so that task_cpu() reports a stable number.
910 * We need to make an exception for PF_STARTING tasks because the fork
911 * path might require task_rq_lock() to work, eg. it can call
912 * set_cpus_allowed_ptr() from the cpuset clone_ns code.
914 static inline int task_is_waking(struct task_struct
*p
)
916 return unlikely((p
->state
== TASK_WAKING
) && !(p
->flags
& PF_STARTING
));
920 * __task_rq_lock - lock the runqueue a given task resides on.
921 * Must be called interrupts disabled.
923 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
929 while (task_is_waking(p
))
932 raw_spin_lock(&rq
->lock
);
933 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
935 raw_spin_unlock(&rq
->lock
);
940 * task_rq_lock - lock the runqueue a given task resides on and disable
941 * interrupts. Note the ordering: we can safely lookup the task_rq without
942 * explicitly disabling preemption.
944 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
950 while (task_is_waking(p
))
952 local_irq_save(*flags
);
954 raw_spin_lock(&rq
->lock
);
955 if (likely(rq
== task_rq(p
) && !task_is_waking(p
)))
957 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
961 void task_rq_unlock_wait(struct task_struct
*p
)
963 struct rq
*rq
= task_rq(p
);
965 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
966 raw_spin_unlock_wait(&rq
->lock
);
969 static void __task_rq_unlock(struct rq
*rq
)
972 raw_spin_unlock(&rq
->lock
);
975 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
978 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
982 * this_rq_lock - lock this runqueue and disable interrupts.
984 static struct rq
*this_rq_lock(void)
991 raw_spin_lock(&rq
->lock
);
996 #ifdef CONFIG_SCHED_HRTICK
998 * Use HR-timers to deliver accurate preemption points.
1000 * Its all a bit involved since we cannot program an hrt while holding the
1001 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1004 * When we get rescheduled we reprogram the hrtick_timer outside of the
1010 * - enabled by features
1011 * - hrtimer is actually high res
1013 static inline int hrtick_enabled(struct rq
*rq
)
1015 if (!sched_feat(HRTICK
))
1017 if (!cpu_active(cpu_of(rq
)))
1019 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1022 static void hrtick_clear(struct rq
*rq
)
1024 if (hrtimer_active(&rq
->hrtick_timer
))
1025 hrtimer_cancel(&rq
->hrtick_timer
);
1029 * High-resolution timer tick.
1030 * Runs from hardirq context with interrupts disabled.
1032 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1034 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1036 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1038 raw_spin_lock(&rq
->lock
);
1039 update_rq_clock(rq
);
1040 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1041 raw_spin_unlock(&rq
->lock
);
1043 return HRTIMER_NORESTART
;
1048 * called from hardirq (IPI) context
1050 static void __hrtick_start(void *arg
)
1052 struct rq
*rq
= arg
;
1054 raw_spin_lock(&rq
->lock
);
1055 hrtimer_restart(&rq
->hrtick_timer
);
1056 rq
->hrtick_csd_pending
= 0;
1057 raw_spin_unlock(&rq
->lock
);
1061 * Called to set the hrtick timer state.
1063 * called with rq->lock held and irqs disabled
1065 static void hrtick_start(struct rq
*rq
, u64 delay
)
1067 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1068 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1070 hrtimer_set_expires(timer
, time
);
1072 if (rq
== this_rq()) {
1073 hrtimer_restart(timer
);
1074 } else if (!rq
->hrtick_csd_pending
) {
1075 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1076 rq
->hrtick_csd_pending
= 1;
1081 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1083 int cpu
= (int)(long)hcpu
;
1086 case CPU_UP_CANCELED
:
1087 case CPU_UP_CANCELED_FROZEN
:
1088 case CPU_DOWN_PREPARE
:
1089 case CPU_DOWN_PREPARE_FROZEN
:
1091 case CPU_DEAD_FROZEN
:
1092 hrtick_clear(cpu_rq(cpu
));
1099 static __init
void init_hrtick(void)
1101 hotcpu_notifier(hotplug_hrtick
, 0);
1105 * Called to set the hrtick timer state.
1107 * called with rq->lock held and irqs disabled
1109 static void hrtick_start(struct rq
*rq
, u64 delay
)
1111 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1112 HRTIMER_MODE_REL_PINNED
, 0);
1115 static inline void init_hrtick(void)
1118 #endif /* CONFIG_SMP */
1120 static void init_rq_hrtick(struct rq
*rq
)
1123 rq
->hrtick_csd_pending
= 0;
1125 rq
->hrtick_csd
.flags
= 0;
1126 rq
->hrtick_csd
.func
= __hrtick_start
;
1127 rq
->hrtick_csd
.info
= rq
;
1130 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1131 rq
->hrtick_timer
.function
= hrtick
;
1133 #else /* CONFIG_SCHED_HRTICK */
1134 static inline void hrtick_clear(struct rq
*rq
)
1138 static inline void init_rq_hrtick(struct rq
*rq
)
1142 static inline void init_hrtick(void)
1145 #endif /* CONFIG_SCHED_HRTICK */
1148 * resched_task - mark a task 'to be rescheduled now'.
1150 * On UP this means the setting of the need_resched flag, on SMP it
1151 * might also involve a cross-CPU call to trigger the scheduler on
1156 #ifndef tsk_is_polling
1157 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1160 static void resched_task(struct task_struct
*p
)
1164 assert_raw_spin_locked(&task_rq(p
)->lock
);
1166 if (test_tsk_need_resched(p
))
1169 set_tsk_need_resched(p
);
1172 if (cpu
== smp_processor_id())
1175 /* NEED_RESCHED must be visible before we test polling */
1177 if (!tsk_is_polling(p
))
1178 smp_send_reschedule(cpu
);
1181 static void resched_cpu(int cpu
)
1183 struct rq
*rq
= cpu_rq(cpu
);
1184 unsigned long flags
;
1186 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1188 resched_task(cpu_curr(cpu
));
1189 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1194 * When add_timer_on() enqueues a timer into the timer wheel of an
1195 * idle CPU then this timer might expire before the next timer event
1196 * which is scheduled to wake up that CPU. In case of a completely
1197 * idle system the next event might even be infinite time into the
1198 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1199 * leaves the inner idle loop so the newly added timer is taken into
1200 * account when the CPU goes back to idle and evaluates the timer
1201 * wheel for the next timer event.
1203 void wake_up_idle_cpu(int cpu
)
1205 struct rq
*rq
= cpu_rq(cpu
);
1207 if (cpu
== smp_processor_id())
1211 * This is safe, as this function is called with the timer
1212 * wheel base lock of (cpu) held. When the CPU is on the way
1213 * to idle and has not yet set rq->curr to idle then it will
1214 * be serialized on the timer wheel base lock and take the new
1215 * timer into account automatically.
1217 if (rq
->curr
!= rq
->idle
)
1221 * We can set TIF_RESCHED on the idle task of the other CPU
1222 * lockless. The worst case is that the other CPU runs the
1223 * idle task through an additional NOOP schedule()
1225 set_tsk_need_resched(rq
->idle
);
1227 /* NEED_RESCHED must be visible before we test polling */
1229 if (!tsk_is_polling(rq
->idle
))
1230 smp_send_reschedule(cpu
);
1232 #endif /* CONFIG_NO_HZ */
1234 static u64
sched_avg_period(void)
1236 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1239 static void sched_avg_update(struct rq
*rq
)
1241 s64 period
= sched_avg_period();
1243 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1244 rq
->age_stamp
+= period
;
1249 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1251 rq
->rt_avg
+= rt_delta
;
1252 sched_avg_update(rq
);
1255 #else /* !CONFIG_SMP */
1256 static void resched_task(struct task_struct
*p
)
1258 assert_raw_spin_locked(&task_rq(p
)->lock
);
1259 set_tsk_need_resched(p
);
1262 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1265 #endif /* CONFIG_SMP */
1267 #if BITS_PER_LONG == 32
1268 # define WMULT_CONST (~0UL)
1270 # define WMULT_CONST (1UL << 32)
1273 #define WMULT_SHIFT 32
1276 * Shift right and round:
1278 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1281 * delta *= weight / lw
1283 static unsigned long
1284 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1285 struct load_weight
*lw
)
1289 if (!lw
->inv_weight
) {
1290 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1293 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1297 tmp
= (u64
)delta_exec
* weight
;
1299 * Check whether we'd overflow the 64-bit multiplication:
1301 if (unlikely(tmp
> WMULT_CONST
))
1302 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1305 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1307 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1310 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1316 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1323 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1324 * of tasks with abnormal "nice" values across CPUs the contribution that
1325 * each task makes to its run queue's load is weighted according to its
1326 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1327 * scaled version of the new time slice allocation that they receive on time
1331 #define WEIGHT_IDLEPRIO 3
1332 #define WMULT_IDLEPRIO 1431655765
1335 * Nice levels are multiplicative, with a gentle 10% change for every
1336 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1337 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1338 * that remained on nice 0.
1340 * The "10% effect" is relative and cumulative: from _any_ nice level,
1341 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1342 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1343 * If a task goes up by ~10% and another task goes down by ~10% then
1344 * the relative distance between them is ~25%.)
1346 static const int prio_to_weight
[40] = {
1347 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1348 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1349 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1350 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1351 /* 0 */ 1024, 820, 655, 526, 423,
1352 /* 5 */ 335, 272, 215, 172, 137,
1353 /* 10 */ 110, 87, 70, 56, 45,
1354 /* 15 */ 36, 29, 23, 18, 15,
1358 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1360 * In cases where the weight does not change often, we can use the
1361 * precalculated inverse to speed up arithmetics by turning divisions
1362 * into multiplications:
1364 static const u32 prio_to_wmult
[40] = {
1365 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1366 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1367 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1368 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1369 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1370 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1371 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1372 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1375 /* Time spent by the tasks of the cpu accounting group executing in ... */
1376 enum cpuacct_stat_index
{
1377 CPUACCT_STAT_USER
, /* ... user mode */
1378 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1380 CPUACCT_STAT_NSTATS
,
1383 #ifdef CONFIG_CGROUP_CPUACCT
1384 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1385 static void cpuacct_update_stats(struct task_struct
*tsk
,
1386 enum cpuacct_stat_index idx
, cputime_t val
);
1388 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1389 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1390 enum cpuacct_stat_index idx
, cputime_t val
) {}
1393 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1395 update_load_add(&rq
->load
, load
);
1398 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1400 update_load_sub(&rq
->load
, load
);
1403 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1404 typedef int (*tg_visitor
)(struct task_group
*, void *);
1407 * Iterate the full tree, calling @down when first entering a node and @up when
1408 * leaving it for the final time.
1410 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1412 struct task_group
*parent
, *child
;
1416 parent
= &root_task_group
;
1418 ret
= (*down
)(parent
, data
);
1421 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1428 ret
= (*up
)(parent
, data
);
1433 parent
= parent
->parent
;
1442 static int tg_nop(struct task_group
*tg
, void *data
)
1449 /* Used instead of source_load when we know the type == 0 */
1450 static unsigned long weighted_cpuload(const int cpu
)
1452 return cpu_rq(cpu
)->load
.weight
;
1456 * Return a low guess at the load of a migration-source cpu weighted
1457 * according to the scheduling class and "nice" value.
1459 * We want to under-estimate the load of migration sources, to
1460 * balance conservatively.
1462 static unsigned long source_load(int cpu
, int type
)
1464 struct rq
*rq
= cpu_rq(cpu
);
1465 unsigned long total
= weighted_cpuload(cpu
);
1467 if (type
== 0 || !sched_feat(LB_BIAS
))
1470 return min(rq
->cpu_load
[type
-1], total
);
1474 * Return a high guess at the load of a migration-target cpu weighted
1475 * according to the scheduling class and "nice" value.
1477 static unsigned long target_load(int cpu
, int type
)
1479 struct rq
*rq
= cpu_rq(cpu
);
1480 unsigned long total
= weighted_cpuload(cpu
);
1482 if (type
== 0 || !sched_feat(LB_BIAS
))
1485 return max(rq
->cpu_load
[type
-1], total
);
1488 static struct sched_group
*group_of(int cpu
)
1490 struct sched_domain
*sd
= rcu_dereference_sched(cpu_rq(cpu
)->sd
);
1498 static unsigned long power_of(int cpu
)
1500 struct sched_group
*group
= group_of(cpu
);
1503 return SCHED_LOAD_SCALE
;
1505 return group
->cpu_power
;
1508 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1510 static unsigned long cpu_avg_load_per_task(int cpu
)
1512 struct rq
*rq
= cpu_rq(cpu
);
1513 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1516 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1518 rq
->avg_load_per_task
= 0;
1520 return rq
->avg_load_per_task
;
1523 #ifdef CONFIG_FAIR_GROUP_SCHED
1525 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1527 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1530 * Calculate and set the cpu's group shares.
1532 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1533 unsigned long sd_shares
,
1534 unsigned long sd_rq_weight
,
1535 unsigned long *usd_rq_weight
)
1537 unsigned long shares
, rq_weight
;
1540 rq_weight
= usd_rq_weight
[cpu
];
1543 rq_weight
= NICE_0_LOAD
;
1547 * \Sum_j shares_j * rq_weight_i
1548 * shares_i = -----------------------------
1549 * \Sum_j rq_weight_j
1551 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1552 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1554 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1555 sysctl_sched_shares_thresh
) {
1556 struct rq
*rq
= cpu_rq(cpu
);
1557 unsigned long flags
;
1559 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1560 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1561 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1562 __set_se_shares(tg
->se
[cpu
], shares
);
1563 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1568 * Re-compute the task group their per cpu shares over the given domain.
1569 * This needs to be done in a bottom-up fashion because the rq weight of a
1570 * parent group depends on the shares of its child groups.
1572 static int tg_shares_up(struct task_group
*tg
, void *data
)
1574 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1575 unsigned long *usd_rq_weight
;
1576 struct sched_domain
*sd
= data
;
1577 unsigned long flags
;
1583 local_irq_save(flags
);
1584 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1586 for_each_cpu(i
, sched_domain_span(sd
)) {
1587 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1588 usd_rq_weight
[i
] = weight
;
1590 rq_weight
+= weight
;
1592 * If there are currently no tasks on the cpu pretend there
1593 * is one of average load so that when a new task gets to
1594 * run here it will not get delayed by group starvation.
1597 weight
= NICE_0_LOAD
;
1599 sum_weight
+= weight
;
1600 shares
+= tg
->cfs_rq
[i
]->shares
;
1604 rq_weight
= sum_weight
;
1606 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1607 shares
= tg
->shares
;
1609 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1610 shares
= tg
->shares
;
1612 for_each_cpu(i
, sched_domain_span(sd
))
1613 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1615 local_irq_restore(flags
);
1621 * Compute the cpu's hierarchical load factor for each task group.
1622 * This needs to be done in a top-down fashion because the load of a child
1623 * group is a fraction of its parents load.
1625 static int tg_load_down(struct task_group
*tg
, void *data
)
1628 long cpu
= (long)data
;
1631 load
= cpu_rq(cpu
)->load
.weight
;
1633 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1634 load
*= tg
->cfs_rq
[cpu
]->shares
;
1635 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1638 tg
->cfs_rq
[cpu
]->h_load
= load
;
1643 static void update_shares(struct sched_domain
*sd
)
1648 if (root_task_group_empty())
1651 now
= cpu_clock(raw_smp_processor_id());
1652 elapsed
= now
- sd
->last_update
;
1654 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1655 sd
->last_update
= now
;
1656 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1660 static void update_h_load(long cpu
)
1662 if (root_task_group_empty())
1665 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1670 static inline void update_shares(struct sched_domain
*sd
)
1676 #ifdef CONFIG_PREEMPT
1678 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1681 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1682 * way at the expense of forcing extra atomic operations in all
1683 * invocations. This assures that the double_lock is acquired using the
1684 * same underlying policy as the spinlock_t on this architecture, which
1685 * reduces latency compared to the unfair variant below. However, it
1686 * also adds more overhead and therefore may reduce throughput.
1688 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1689 __releases(this_rq
->lock
)
1690 __acquires(busiest
->lock
)
1691 __acquires(this_rq
->lock
)
1693 raw_spin_unlock(&this_rq
->lock
);
1694 double_rq_lock(this_rq
, busiest
);
1701 * Unfair double_lock_balance: Optimizes throughput at the expense of
1702 * latency by eliminating extra atomic operations when the locks are
1703 * already in proper order on entry. This favors lower cpu-ids and will
1704 * grant the double lock to lower cpus over higher ids under contention,
1705 * regardless of entry order into the function.
1707 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1708 __releases(this_rq
->lock
)
1709 __acquires(busiest
->lock
)
1710 __acquires(this_rq
->lock
)
1714 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1715 if (busiest
< this_rq
) {
1716 raw_spin_unlock(&this_rq
->lock
);
1717 raw_spin_lock(&busiest
->lock
);
1718 raw_spin_lock_nested(&this_rq
->lock
,
1719 SINGLE_DEPTH_NESTING
);
1722 raw_spin_lock_nested(&busiest
->lock
,
1723 SINGLE_DEPTH_NESTING
);
1728 #endif /* CONFIG_PREEMPT */
1731 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1733 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1735 if (unlikely(!irqs_disabled())) {
1736 /* printk() doesn't work good under rq->lock */
1737 raw_spin_unlock(&this_rq
->lock
);
1741 return _double_lock_balance(this_rq
, busiest
);
1744 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1745 __releases(busiest
->lock
)
1747 raw_spin_unlock(&busiest
->lock
);
1748 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1752 * double_rq_lock - safely lock two runqueues
1754 * Note this does not disable interrupts like task_rq_lock,
1755 * you need to do so manually before calling.
1757 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1758 __acquires(rq1
->lock
)
1759 __acquires(rq2
->lock
)
1761 BUG_ON(!irqs_disabled());
1763 raw_spin_lock(&rq1
->lock
);
1764 __acquire(rq2
->lock
); /* Fake it out ;) */
1767 raw_spin_lock(&rq1
->lock
);
1768 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1770 raw_spin_lock(&rq2
->lock
);
1771 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1774 update_rq_clock(rq1
);
1775 update_rq_clock(rq2
);
1779 * double_rq_unlock - safely unlock two runqueues
1781 * Note this does not restore interrupts like task_rq_unlock,
1782 * you need to do so manually after calling.
1784 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1785 __releases(rq1
->lock
)
1786 __releases(rq2
->lock
)
1788 raw_spin_unlock(&rq1
->lock
);
1790 raw_spin_unlock(&rq2
->lock
);
1792 __release(rq2
->lock
);
1797 #ifdef CONFIG_FAIR_GROUP_SCHED
1798 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1801 cfs_rq
->shares
= shares
;
1806 static void calc_load_account_active(struct rq
*this_rq
);
1807 static void update_sysctl(void);
1808 static int get_update_sysctl_factor(void);
1810 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1812 set_task_rq(p
, cpu
);
1815 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1816 * successfuly executed on another CPU. We must ensure that updates of
1817 * per-task data have been completed by this moment.
1820 task_thread_info(p
)->cpu
= cpu
;
1824 static const struct sched_class rt_sched_class
;
1826 #define sched_class_highest (&rt_sched_class)
1827 #define for_each_class(class) \
1828 for (class = sched_class_highest; class; class = class->next)
1830 #include "sched_stats.h"
1832 static void inc_nr_running(struct rq
*rq
)
1837 static void dec_nr_running(struct rq
*rq
)
1842 static void set_load_weight(struct task_struct
*p
)
1844 if (task_has_rt_policy(p
)) {
1845 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1846 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1851 * SCHED_IDLE tasks get minimal weight:
1853 if (p
->policy
== SCHED_IDLE
) {
1854 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1855 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1859 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1860 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1863 static void update_avg(u64
*avg
, u64 sample
)
1865 s64 diff
= sample
- *avg
;
1870 enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
, bool head
)
1873 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1875 sched_info_queued(p
);
1876 p
->sched_class
->enqueue_task(rq
, p
, wakeup
, head
);
1880 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1883 if (p
->se
.last_wakeup
) {
1884 update_avg(&p
->se
.avg_overlap
,
1885 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1886 p
->se
.last_wakeup
= 0;
1888 update_avg(&p
->se
.avg_wakeup
,
1889 sysctl_sched_wakeup_granularity
);
1893 sched_info_dequeued(p
);
1894 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1899 * activate_task - move a task to the runqueue.
1901 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1903 if (task_contributes_to_load(p
))
1904 rq
->nr_uninterruptible
--;
1906 enqueue_task(rq
, p
, wakeup
, false);
1911 * deactivate_task - remove a task from the runqueue.
1913 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1915 if (task_contributes_to_load(p
))
1916 rq
->nr_uninterruptible
++;
1918 dequeue_task(rq
, p
, sleep
);
1922 #include "sched_idletask.c"
1923 #include "sched_fair.c"
1924 #include "sched_rt.c"
1925 #ifdef CONFIG_SCHED_DEBUG
1926 # include "sched_debug.c"
1930 * __normal_prio - return the priority that is based on the static prio
1932 static inline int __normal_prio(struct task_struct
*p
)
1934 return p
->static_prio
;
1938 * Calculate the expected normal priority: i.e. priority
1939 * without taking RT-inheritance into account. Might be
1940 * boosted by interactivity modifiers. Changes upon fork,
1941 * setprio syscalls, and whenever the interactivity
1942 * estimator recalculates.
1944 static inline int normal_prio(struct task_struct
*p
)
1948 if (task_has_rt_policy(p
))
1949 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1951 prio
= __normal_prio(p
);
1956 * Calculate the current priority, i.e. the priority
1957 * taken into account by the scheduler. This value might
1958 * be boosted by RT tasks, or might be boosted by
1959 * interactivity modifiers. Will be RT if the task got
1960 * RT-boosted. If not then it returns p->normal_prio.
1962 static int effective_prio(struct task_struct
*p
)
1964 p
->normal_prio
= normal_prio(p
);
1966 * If we are RT tasks or we were boosted to RT priority,
1967 * keep the priority unchanged. Otherwise, update priority
1968 * to the normal priority:
1970 if (!rt_prio(p
->prio
))
1971 return p
->normal_prio
;
1976 * task_curr - is this task currently executing on a CPU?
1977 * @p: the task in question.
1979 inline int task_curr(const struct task_struct
*p
)
1981 return cpu_curr(task_cpu(p
)) == p
;
1984 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1985 const struct sched_class
*prev_class
,
1986 int oldprio
, int running
)
1988 if (prev_class
!= p
->sched_class
) {
1989 if (prev_class
->switched_from
)
1990 prev_class
->switched_from(rq
, p
, running
);
1991 p
->sched_class
->switched_to(rq
, p
, running
);
1993 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1998 * Is this task likely cache-hot:
2001 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2005 if (p
->sched_class
!= &fair_sched_class
)
2009 * Buddy candidates are cache hot:
2011 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2012 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2013 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2016 if (sysctl_sched_migration_cost
== -1)
2018 if (sysctl_sched_migration_cost
== 0)
2021 delta
= now
- p
->se
.exec_start
;
2023 return delta
< (s64
)sysctl_sched_migration_cost
;
2026 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2028 #ifdef CONFIG_SCHED_DEBUG
2030 * We should never call set_task_cpu() on a blocked task,
2031 * ttwu() will sort out the placement.
2033 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2034 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2037 trace_sched_migrate_task(p
, new_cpu
);
2039 if (task_cpu(p
) != new_cpu
) {
2040 p
->se
.nr_migrations
++;
2041 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2044 __set_task_cpu(p
, new_cpu
);
2047 struct migration_req
{
2048 struct list_head list
;
2050 struct task_struct
*task
;
2053 struct completion done
;
2057 * The task's runqueue lock must be held.
2058 * Returns true if you have to wait for migration thread.
2061 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2063 struct rq
*rq
= task_rq(p
);
2066 * If the task is not on a runqueue (and not running), then
2067 * the next wake-up will properly place the task.
2069 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2072 init_completion(&req
->done
);
2074 req
->dest_cpu
= dest_cpu
;
2075 list_add(&req
->list
, &rq
->migration_queue
);
2081 * wait_task_inactive - wait for a thread to unschedule.
2083 * If @match_state is nonzero, it's the @p->state value just checked and
2084 * not expected to change. If it changes, i.e. @p might have woken up,
2085 * then return zero. When we succeed in waiting for @p to be off its CPU,
2086 * we return a positive number (its total switch count). If a second call
2087 * a short while later returns the same number, the caller can be sure that
2088 * @p has remained unscheduled the whole time.
2090 * The caller must ensure that the task *will* unschedule sometime soon,
2091 * else this function might spin for a *long* time. This function can't
2092 * be called with interrupts off, or it may introduce deadlock with
2093 * smp_call_function() if an IPI is sent by the same process we are
2094 * waiting to become inactive.
2096 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2098 unsigned long flags
;
2105 * We do the initial early heuristics without holding
2106 * any task-queue locks at all. We'll only try to get
2107 * the runqueue lock when things look like they will
2113 * If the task is actively running on another CPU
2114 * still, just relax and busy-wait without holding
2117 * NOTE! Since we don't hold any locks, it's not
2118 * even sure that "rq" stays as the right runqueue!
2119 * But we don't care, since "task_running()" will
2120 * return false if the runqueue has changed and p
2121 * is actually now running somewhere else!
2123 while (task_running(rq
, p
)) {
2124 if (match_state
&& unlikely(p
->state
!= match_state
))
2130 * Ok, time to look more closely! We need the rq
2131 * lock now, to be *sure*. If we're wrong, we'll
2132 * just go back and repeat.
2134 rq
= task_rq_lock(p
, &flags
);
2135 trace_sched_wait_task(rq
, p
);
2136 running
= task_running(rq
, p
);
2137 on_rq
= p
->se
.on_rq
;
2139 if (!match_state
|| p
->state
== match_state
)
2140 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2141 task_rq_unlock(rq
, &flags
);
2144 * If it changed from the expected state, bail out now.
2146 if (unlikely(!ncsw
))
2150 * Was it really running after all now that we
2151 * checked with the proper locks actually held?
2153 * Oops. Go back and try again..
2155 if (unlikely(running
)) {
2161 * It's not enough that it's not actively running,
2162 * it must be off the runqueue _entirely_, and not
2165 * So if it was still runnable (but just not actively
2166 * running right now), it's preempted, and we should
2167 * yield - it could be a while.
2169 if (unlikely(on_rq
)) {
2170 schedule_timeout_uninterruptible(1);
2175 * Ahh, all good. It wasn't running, and it wasn't
2176 * runnable, which means that it will never become
2177 * running in the future either. We're all done!
2186 * kick_process - kick a running thread to enter/exit the kernel
2187 * @p: the to-be-kicked thread
2189 * Cause a process which is running on another CPU to enter
2190 * kernel-mode, without any delay. (to get signals handled.)
2192 * NOTE: this function doesnt have to take the runqueue lock,
2193 * because all it wants to ensure is that the remote task enters
2194 * the kernel. If the IPI races and the task has been migrated
2195 * to another CPU then no harm is done and the purpose has been
2198 void kick_process(struct task_struct
*p
)
2204 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2205 smp_send_reschedule(cpu
);
2208 EXPORT_SYMBOL_GPL(kick_process
);
2209 #endif /* CONFIG_SMP */
2212 * task_oncpu_function_call - call a function on the cpu on which a task runs
2213 * @p: the task to evaluate
2214 * @func: the function to be called
2215 * @info: the function call argument
2217 * Calls the function @func when the task is currently running. This might
2218 * be on the current CPU, which just calls the function directly
2220 void task_oncpu_function_call(struct task_struct
*p
,
2221 void (*func
) (void *info
), void *info
)
2228 smp_call_function_single(cpu
, func
, info
, 1);
2233 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2236 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2238 /* Look for allowed, online CPU in same node. */
2239 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2240 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2243 /* Any allowed, online CPU? */
2244 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2245 if (dest_cpu
< nr_cpu_ids
)
2248 /* No more Mr. Nice Guy. */
2249 if (dest_cpu
>= nr_cpu_ids
) {
2251 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
2253 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
2256 * Don't tell them about moving exiting tasks or
2257 * kernel threads (both mm NULL), since they never
2260 if (p
->mm
&& printk_ratelimit()) {
2261 printk(KERN_INFO
"process %d (%s) no "
2262 "longer affine to cpu%d\n",
2263 task_pid_nr(p
), p
->comm
, cpu
);
2271 * Gets called from 3 sites (exec, fork, wakeup), since it is called without
2272 * holding rq->lock we need to ensure ->cpus_allowed is stable, this is done
2275 * exec: is unstable, retry loop
2276 * fork & wake-up: serialize ->cpus_allowed against TASK_WAKING
2279 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2281 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2284 * In order not to call set_task_cpu() on a blocking task we need
2285 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2288 * Since this is common to all placement strategies, this lives here.
2290 * [ this allows ->select_task() to simply return task_cpu(p) and
2291 * not worry about this generic constraint ]
2293 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2295 cpu
= select_fallback_rq(task_cpu(p
), p
);
2302 * try_to_wake_up - wake up a thread
2303 * @p: the to-be-woken-up thread
2304 * @state: the mask of task states that can be woken
2305 * @sync: do a synchronous wakeup?
2307 * Put it on the run-queue if it's not already there. The "current"
2308 * thread is always on the run-queue (except when the actual
2309 * re-schedule is in progress), and as such you're allowed to do
2310 * the simpler "current->state = TASK_RUNNING" to mark yourself
2311 * runnable without the overhead of this.
2313 * returns failure only if the task is already active.
2315 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2318 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2319 unsigned long flags
;
2322 if (!sched_feat(SYNC_WAKEUPS
))
2323 wake_flags
&= ~WF_SYNC
;
2325 this_cpu
= get_cpu();
2328 rq
= task_rq_lock(p
, &flags
);
2329 update_rq_clock(rq
);
2330 if (!(p
->state
& state
))
2340 if (unlikely(task_running(rq
, p
)))
2344 * In order to handle concurrent wakeups and release the rq->lock
2345 * we put the task in TASK_WAKING state.
2347 * First fix up the nr_uninterruptible count:
2349 if (task_contributes_to_load(p
))
2350 rq
->nr_uninterruptible
--;
2351 p
->state
= TASK_WAKING
;
2353 if (p
->sched_class
->task_waking
)
2354 p
->sched_class
->task_waking(rq
, p
);
2356 __task_rq_unlock(rq
);
2358 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2359 if (cpu
!= orig_cpu
) {
2361 * Since we migrate the task without holding any rq->lock,
2362 * we need to be careful with task_rq_lock(), since that
2363 * might end up locking an invalid rq.
2365 set_task_cpu(p
, cpu
);
2369 raw_spin_lock(&rq
->lock
);
2370 update_rq_clock(rq
);
2373 * We migrated the task without holding either rq->lock, however
2374 * since the task is not on the task list itself, nobody else
2375 * will try and migrate the task, hence the rq should match the
2376 * cpu we just moved it to.
2378 WARN_ON(task_cpu(p
) != cpu
);
2379 WARN_ON(p
->state
!= TASK_WAKING
);
2381 #ifdef CONFIG_SCHEDSTATS
2382 schedstat_inc(rq
, ttwu_count
);
2383 if (cpu
== this_cpu
)
2384 schedstat_inc(rq
, ttwu_local
);
2386 struct sched_domain
*sd
;
2387 for_each_domain(this_cpu
, sd
) {
2388 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2389 schedstat_inc(sd
, ttwu_wake_remote
);
2394 #endif /* CONFIG_SCHEDSTATS */
2397 #endif /* CONFIG_SMP */
2398 schedstat_inc(p
, se
.nr_wakeups
);
2399 if (wake_flags
& WF_SYNC
)
2400 schedstat_inc(p
, se
.nr_wakeups_sync
);
2401 if (orig_cpu
!= cpu
)
2402 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2403 if (cpu
== this_cpu
)
2404 schedstat_inc(p
, se
.nr_wakeups_local
);
2406 schedstat_inc(p
, se
.nr_wakeups_remote
);
2407 activate_task(rq
, p
, 1);
2411 * Only attribute actual wakeups done by this task.
2413 if (!in_interrupt()) {
2414 struct sched_entity
*se
= ¤t
->se
;
2415 u64 sample
= se
->sum_exec_runtime
;
2417 if (se
->last_wakeup
)
2418 sample
-= se
->last_wakeup
;
2420 sample
-= se
->start_runtime
;
2421 update_avg(&se
->avg_wakeup
, sample
);
2423 se
->last_wakeup
= se
->sum_exec_runtime
;
2427 trace_sched_wakeup(rq
, p
, success
);
2428 check_preempt_curr(rq
, p
, wake_flags
);
2430 p
->state
= TASK_RUNNING
;
2432 if (p
->sched_class
->task_woken
)
2433 p
->sched_class
->task_woken(rq
, p
);
2435 if (unlikely(rq
->idle_stamp
)) {
2436 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2437 u64 max
= 2*sysctl_sched_migration_cost
;
2442 update_avg(&rq
->avg_idle
, delta
);
2447 task_rq_unlock(rq
, &flags
);
2454 * wake_up_process - Wake up a specific process
2455 * @p: The process to be woken up.
2457 * Attempt to wake up the nominated process and move it to the set of runnable
2458 * processes. Returns 1 if the process was woken up, 0 if it was already
2461 * It may be assumed that this function implies a write memory barrier before
2462 * changing the task state if and only if any tasks are woken up.
2464 int wake_up_process(struct task_struct
*p
)
2466 return try_to_wake_up(p
, TASK_ALL
, 0);
2468 EXPORT_SYMBOL(wake_up_process
);
2470 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2472 return try_to_wake_up(p
, state
, 0);
2476 * Perform scheduler related setup for a newly forked process p.
2477 * p is forked by current.
2479 * __sched_fork() is basic setup used by init_idle() too:
2481 static void __sched_fork(struct task_struct
*p
)
2483 p
->se
.exec_start
= 0;
2484 p
->se
.sum_exec_runtime
= 0;
2485 p
->se
.prev_sum_exec_runtime
= 0;
2486 p
->se
.nr_migrations
= 0;
2487 p
->se
.last_wakeup
= 0;
2488 p
->se
.avg_overlap
= 0;
2489 p
->se
.start_runtime
= 0;
2490 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2492 #ifdef CONFIG_SCHEDSTATS
2493 p
->se
.wait_start
= 0;
2495 p
->se
.wait_count
= 0;
2498 p
->se
.sleep_start
= 0;
2499 p
->se
.sleep_max
= 0;
2500 p
->se
.sum_sleep_runtime
= 0;
2502 p
->se
.block_start
= 0;
2503 p
->se
.block_max
= 0;
2505 p
->se
.slice_max
= 0;
2507 p
->se
.nr_migrations_cold
= 0;
2508 p
->se
.nr_failed_migrations_affine
= 0;
2509 p
->se
.nr_failed_migrations_running
= 0;
2510 p
->se
.nr_failed_migrations_hot
= 0;
2511 p
->se
.nr_forced_migrations
= 0;
2513 p
->se
.nr_wakeups
= 0;
2514 p
->se
.nr_wakeups_sync
= 0;
2515 p
->se
.nr_wakeups_migrate
= 0;
2516 p
->se
.nr_wakeups_local
= 0;
2517 p
->se
.nr_wakeups_remote
= 0;
2518 p
->se
.nr_wakeups_affine
= 0;
2519 p
->se
.nr_wakeups_affine_attempts
= 0;
2520 p
->se
.nr_wakeups_passive
= 0;
2521 p
->se
.nr_wakeups_idle
= 0;
2525 INIT_LIST_HEAD(&p
->rt
.run_list
);
2527 INIT_LIST_HEAD(&p
->se
.group_node
);
2529 #ifdef CONFIG_PREEMPT_NOTIFIERS
2530 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2535 * fork()/clone()-time setup:
2537 void sched_fork(struct task_struct
*p
, int clone_flags
)
2539 int cpu
= get_cpu();
2543 * We mark the process as waking here. This guarantees that
2544 * nobody will actually run it, and a signal or other external
2545 * event cannot wake it up and insert it on the runqueue either.
2547 p
->state
= TASK_WAKING
;
2550 * Revert to default priority/policy on fork if requested.
2552 if (unlikely(p
->sched_reset_on_fork
)) {
2553 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2554 p
->policy
= SCHED_NORMAL
;
2555 p
->normal_prio
= p
->static_prio
;
2558 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2559 p
->static_prio
= NICE_TO_PRIO(0);
2560 p
->normal_prio
= p
->static_prio
;
2565 * We don't need the reset flag anymore after the fork. It has
2566 * fulfilled its duty:
2568 p
->sched_reset_on_fork
= 0;
2572 * Make sure we do not leak PI boosting priority to the child.
2574 p
->prio
= current
->normal_prio
;
2576 if (!rt_prio(p
->prio
))
2577 p
->sched_class
= &fair_sched_class
;
2579 if (p
->sched_class
->task_fork
)
2580 p
->sched_class
->task_fork(p
);
2582 set_task_cpu(p
, cpu
);
2584 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2585 if (likely(sched_info_on()))
2586 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2588 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2591 #ifdef CONFIG_PREEMPT
2592 /* Want to start with kernel preemption disabled. */
2593 task_thread_info(p
)->preempt_count
= 1;
2595 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2601 * wake_up_new_task - wake up a newly created task for the first time.
2603 * This function will do some initial scheduler statistics housekeeping
2604 * that must be done for every newly created context, then puts the task
2605 * on the runqueue and wakes it.
2607 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2609 unsigned long flags
;
2611 int cpu __maybe_unused
= get_cpu();
2615 * Fork balancing, do it here and not earlier because:
2616 * - cpus_allowed can change in the fork path
2617 * - any previously selected cpu might disappear through hotplug
2619 * We still have TASK_WAKING but PF_STARTING is gone now, meaning
2620 * ->cpus_allowed is stable, we have preemption disabled, meaning
2621 * cpu_online_mask is stable.
2623 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2624 set_task_cpu(p
, cpu
);
2628 * Since the task is not on the rq and we still have TASK_WAKING set
2629 * nobody else will migrate this task.
2632 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2634 BUG_ON(p
->state
!= TASK_WAKING
);
2635 p
->state
= TASK_RUNNING
;
2636 update_rq_clock(rq
);
2637 activate_task(rq
, p
, 0);
2638 trace_sched_wakeup_new(rq
, p
, 1);
2639 check_preempt_curr(rq
, p
, WF_FORK
);
2641 if (p
->sched_class
->task_woken
)
2642 p
->sched_class
->task_woken(rq
, p
);
2644 task_rq_unlock(rq
, &flags
);
2648 #ifdef CONFIG_PREEMPT_NOTIFIERS
2651 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2652 * @notifier: notifier struct to register
2654 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2656 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2658 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2661 * preempt_notifier_unregister - no longer interested in preemption notifications
2662 * @notifier: notifier struct to unregister
2664 * This is safe to call from within a preemption notifier.
2666 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2668 hlist_del(¬ifier
->link
);
2670 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2672 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2674 struct preempt_notifier
*notifier
;
2675 struct hlist_node
*node
;
2677 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2678 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2682 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2683 struct task_struct
*next
)
2685 struct preempt_notifier
*notifier
;
2686 struct hlist_node
*node
;
2688 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2689 notifier
->ops
->sched_out(notifier
, next
);
2692 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2694 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2699 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2700 struct task_struct
*next
)
2704 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2707 * prepare_task_switch - prepare to switch tasks
2708 * @rq: the runqueue preparing to switch
2709 * @prev: the current task that is being switched out
2710 * @next: the task we are going to switch to.
2712 * This is called with the rq lock held and interrupts off. It must
2713 * be paired with a subsequent finish_task_switch after the context
2716 * prepare_task_switch sets up locking and calls architecture specific
2720 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2721 struct task_struct
*next
)
2723 fire_sched_out_preempt_notifiers(prev
, next
);
2724 prepare_lock_switch(rq
, next
);
2725 prepare_arch_switch(next
);
2729 * finish_task_switch - clean up after a task-switch
2730 * @rq: runqueue associated with task-switch
2731 * @prev: the thread we just switched away from.
2733 * finish_task_switch must be called after the context switch, paired
2734 * with a prepare_task_switch call before the context switch.
2735 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2736 * and do any other architecture-specific cleanup actions.
2738 * Note that we may have delayed dropping an mm in context_switch(). If
2739 * so, we finish that here outside of the runqueue lock. (Doing it
2740 * with the lock held can cause deadlocks; see schedule() for
2743 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2744 __releases(rq
->lock
)
2746 struct mm_struct
*mm
= rq
->prev_mm
;
2752 * A task struct has one reference for the use as "current".
2753 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2754 * schedule one last time. The schedule call will never return, and
2755 * the scheduled task must drop that reference.
2756 * The test for TASK_DEAD must occur while the runqueue locks are
2757 * still held, otherwise prev could be scheduled on another cpu, die
2758 * there before we look at prev->state, and then the reference would
2760 * Manfred Spraul <manfred@colorfullife.com>
2762 prev_state
= prev
->state
;
2763 finish_arch_switch(prev
);
2764 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2765 local_irq_disable();
2766 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2767 perf_event_task_sched_in(current
);
2768 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2770 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2771 finish_lock_switch(rq
, prev
);
2773 fire_sched_in_preempt_notifiers(current
);
2776 if (unlikely(prev_state
== TASK_DEAD
)) {
2778 * Remove function-return probe instances associated with this
2779 * task and put them back on the free list.
2781 kprobe_flush_task(prev
);
2782 put_task_struct(prev
);
2788 /* assumes rq->lock is held */
2789 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2791 if (prev
->sched_class
->pre_schedule
)
2792 prev
->sched_class
->pre_schedule(rq
, prev
);
2795 /* rq->lock is NOT held, but preemption is disabled */
2796 static inline void post_schedule(struct rq
*rq
)
2798 if (rq
->post_schedule
) {
2799 unsigned long flags
;
2801 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2802 if (rq
->curr
->sched_class
->post_schedule
)
2803 rq
->curr
->sched_class
->post_schedule(rq
);
2804 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2806 rq
->post_schedule
= 0;
2812 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2816 static inline void post_schedule(struct rq
*rq
)
2823 * schedule_tail - first thing a freshly forked thread must call.
2824 * @prev: the thread we just switched away from.
2826 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2827 __releases(rq
->lock
)
2829 struct rq
*rq
= this_rq();
2831 finish_task_switch(rq
, prev
);
2834 * FIXME: do we need to worry about rq being invalidated by the
2839 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2840 /* In this case, finish_task_switch does not reenable preemption */
2843 if (current
->set_child_tid
)
2844 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2848 * context_switch - switch to the new MM and the new
2849 * thread's register state.
2852 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2853 struct task_struct
*next
)
2855 struct mm_struct
*mm
, *oldmm
;
2857 prepare_task_switch(rq
, prev
, next
);
2858 trace_sched_switch(rq
, prev
, next
);
2860 oldmm
= prev
->active_mm
;
2862 * For paravirt, this is coupled with an exit in switch_to to
2863 * combine the page table reload and the switch backend into
2866 arch_start_context_switch(prev
);
2869 next
->active_mm
= oldmm
;
2870 atomic_inc(&oldmm
->mm_count
);
2871 enter_lazy_tlb(oldmm
, next
);
2873 switch_mm(oldmm
, mm
, next
);
2875 if (likely(!prev
->mm
)) {
2876 prev
->active_mm
= NULL
;
2877 rq
->prev_mm
= oldmm
;
2880 * Since the runqueue lock will be released by the next
2881 * task (which is an invalid locking op but in the case
2882 * of the scheduler it's an obvious special-case), so we
2883 * do an early lockdep release here:
2885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2886 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2889 /* Here we just switch the register state and the stack. */
2890 switch_to(prev
, next
, prev
);
2894 * this_rq must be evaluated again because prev may have moved
2895 * CPUs since it called schedule(), thus the 'rq' on its stack
2896 * frame will be invalid.
2898 finish_task_switch(this_rq(), prev
);
2902 * nr_running, nr_uninterruptible and nr_context_switches:
2904 * externally visible scheduler statistics: current number of runnable
2905 * threads, current number of uninterruptible-sleeping threads, total
2906 * number of context switches performed since bootup.
2908 unsigned long nr_running(void)
2910 unsigned long i
, sum
= 0;
2912 for_each_online_cpu(i
)
2913 sum
+= cpu_rq(i
)->nr_running
;
2918 unsigned long nr_uninterruptible(void)
2920 unsigned long i
, sum
= 0;
2922 for_each_possible_cpu(i
)
2923 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2926 * Since we read the counters lockless, it might be slightly
2927 * inaccurate. Do not allow it to go below zero though:
2929 if (unlikely((long)sum
< 0))
2935 unsigned long long nr_context_switches(void)
2938 unsigned long long sum
= 0;
2940 for_each_possible_cpu(i
)
2941 sum
+= cpu_rq(i
)->nr_switches
;
2946 unsigned long nr_iowait(void)
2948 unsigned long i
, sum
= 0;
2950 for_each_possible_cpu(i
)
2951 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2956 unsigned long nr_iowait_cpu(void)
2958 struct rq
*this = this_rq();
2959 return atomic_read(&this->nr_iowait
);
2962 unsigned long this_cpu_load(void)
2964 struct rq
*this = this_rq();
2965 return this->cpu_load
[0];
2969 /* Variables and functions for calc_load */
2970 static atomic_long_t calc_load_tasks
;
2971 static unsigned long calc_load_update
;
2972 unsigned long avenrun
[3];
2973 EXPORT_SYMBOL(avenrun
);
2976 * get_avenrun - get the load average array
2977 * @loads: pointer to dest load array
2978 * @offset: offset to add
2979 * @shift: shift count to shift the result left
2981 * These values are estimates at best, so no need for locking.
2983 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2985 loads
[0] = (avenrun
[0] + offset
) << shift
;
2986 loads
[1] = (avenrun
[1] + offset
) << shift
;
2987 loads
[2] = (avenrun
[2] + offset
) << shift
;
2990 static unsigned long
2991 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2994 load
+= active
* (FIXED_1
- exp
);
2995 return load
>> FSHIFT
;
2999 * calc_load - update the avenrun load estimates 10 ticks after the
3000 * CPUs have updated calc_load_tasks.
3002 void calc_global_load(void)
3004 unsigned long upd
= calc_load_update
+ 10;
3007 if (time_before(jiffies
, upd
))
3010 active
= atomic_long_read(&calc_load_tasks
);
3011 active
= active
> 0 ? active
* FIXED_1
: 0;
3013 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3014 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3015 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3017 calc_load_update
+= LOAD_FREQ
;
3021 * Either called from update_cpu_load() or from a cpu going idle
3023 static void calc_load_account_active(struct rq
*this_rq
)
3025 long nr_active
, delta
;
3027 nr_active
= this_rq
->nr_running
;
3028 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3030 if (nr_active
!= this_rq
->calc_load_active
) {
3031 delta
= nr_active
- this_rq
->calc_load_active
;
3032 this_rq
->calc_load_active
= nr_active
;
3033 atomic_long_add(delta
, &calc_load_tasks
);
3038 * Update rq->cpu_load[] statistics. This function is usually called every
3039 * scheduler tick (TICK_NSEC).
3041 static void update_cpu_load(struct rq
*this_rq
)
3043 unsigned long this_load
= this_rq
->load
.weight
;
3046 this_rq
->nr_load_updates
++;
3048 /* Update our load: */
3049 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3050 unsigned long old_load
, new_load
;
3052 /* scale is effectively 1 << i now, and >> i divides by scale */
3054 old_load
= this_rq
->cpu_load
[i
];
3055 new_load
= this_load
;
3057 * Round up the averaging division if load is increasing. This
3058 * prevents us from getting stuck on 9 if the load is 10, for
3061 if (new_load
> old_load
)
3062 new_load
+= scale
-1;
3063 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3066 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3067 this_rq
->calc_load_update
+= LOAD_FREQ
;
3068 calc_load_account_active(this_rq
);
3075 * sched_exec - execve() is a valuable balancing opportunity, because at
3076 * this point the task has the smallest effective memory and cache footprint.
3078 void sched_exec(void)
3080 struct task_struct
*p
= current
;
3081 struct migration_req req
;
3082 int dest_cpu
, this_cpu
;
3083 unsigned long flags
;
3087 this_cpu
= get_cpu();
3088 dest_cpu
= select_task_rq(p
, SD_BALANCE_EXEC
, 0);
3089 if (dest_cpu
== this_cpu
) {
3094 rq
= task_rq_lock(p
, &flags
);
3098 * select_task_rq() can race against ->cpus_allowed
3100 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3101 || unlikely(!cpu_active(dest_cpu
))) {
3102 task_rq_unlock(rq
, &flags
);
3106 /* force the process onto the specified CPU */
3107 if (migrate_task(p
, dest_cpu
, &req
)) {
3108 /* Need to wait for migration thread (might exit: take ref). */
3109 struct task_struct
*mt
= rq
->migration_thread
;
3111 get_task_struct(mt
);
3112 task_rq_unlock(rq
, &flags
);
3113 wake_up_process(mt
);
3114 put_task_struct(mt
);
3115 wait_for_completion(&req
.done
);
3119 task_rq_unlock(rq
, &flags
);
3124 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3126 EXPORT_PER_CPU_SYMBOL(kstat
);
3129 * Return any ns on the sched_clock that have not yet been accounted in
3130 * @p in case that task is currently running.
3132 * Called with task_rq_lock() held on @rq.
3134 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3138 if (task_current(rq
, p
)) {
3139 update_rq_clock(rq
);
3140 ns
= rq
->clock
- p
->se
.exec_start
;
3148 unsigned long long task_delta_exec(struct task_struct
*p
)
3150 unsigned long flags
;
3154 rq
= task_rq_lock(p
, &flags
);
3155 ns
= do_task_delta_exec(p
, rq
);
3156 task_rq_unlock(rq
, &flags
);
3162 * Return accounted runtime for the task.
3163 * In case the task is currently running, return the runtime plus current's
3164 * pending runtime that have not been accounted yet.
3166 unsigned long long task_sched_runtime(struct task_struct
*p
)
3168 unsigned long flags
;
3172 rq
= task_rq_lock(p
, &flags
);
3173 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3174 task_rq_unlock(rq
, &flags
);
3180 * Return sum_exec_runtime for the thread group.
3181 * In case the task is currently running, return the sum plus current's
3182 * pending runtime that have not been accounted yet.
3184 * Note that the thread group might have other running tasks as well,
3185 * so the return value not includes other pending runtime that other
3186 * running tasks might have.
3188 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3190 struct task_cputime totals
;
3191 unsigned long flags
;
3195 rq
= task_rq_lock(p
, &flags
);
3196 thread_group_cputime(p
, &totals
);
3197 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3198 task_rq_unlock(rq
, &flags
);
3204 * Account user cpu time to a process.
3205 * @p: the process that the cpu time gets accounted to
3206 * @cputime: the cpu time spent in user space since the last update
3207 * @cputime_scaled: cputime scaled by cpu frequency
3209 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3210 cputime_t cputime_scaled
)
3212 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3215 /* Add user time to process. */
3216 p
->utime
= cputime_add(p
->utime
, cputime
);
3217 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3218 account_group_user_time(p
, cputime
);
3220 /* Add user time to cpustat. */
3221 tmp
= cputime_to_cputime64(cputime
);
3222 if (TASK_NICE(p
) > 0)
3223 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3225 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3227 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3228 /* Account for user time used */
3229 acct_update_integrals(p
);
3233 * Account guest cpu time to a process.
3234 * @p: the process that the cpu time gets accounted to
3235 * @cputime: the cpu time spent in virtual machine since the last update
3236 * @cputime_scaled: cputime scaled by cpu frequency
3238 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3239 cputime_t cputime_scaled
)
3242 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3244 tmp
= cputime_to_cputime64(cputime
);
3246 /* Add guest time to process. */
3247 p
->utime
= cputime_add(p
->utime
, cputime
);
3248 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3249 account_group_user_time(p
, cputime
);
3250 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3252 /* Add guest time to cpustat. */
3253 if (TASK_NICE(p
) > 0) {
3254 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3255 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3257 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3258 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3263 * Account system cpu time to a process.
3264 * @p: the process that the cpu time gets accounted to
3265 * @hardirq_offset: the offset to subtract from hardirq_count()
3266 * @cputime: the cpu time spent in kernel space since the last update
3267 * @cputime_scaled: cputime scaled by cpu frequency
3269 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3270 cputime_t cputime
, cputime_t cputime_scaled
)
3272 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3275 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3276 account_guest_time(p
, cputime
, cputime_scaled
);
3280 /* Add system time to process. */
3281 p
->stime
= cputime_add(p
->stime
, cputime
);
3282 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3283 account_group_system_time(p
, cputime
);
3285 /* Add system time to cpustat. */
3286 tmp
= cputime_to_cputime64(cputime
);
3287 if (hardirq_count() - hardirq_offset
)
3288 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3289 else if (softirq_count())
3290 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3292 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3294 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3296 /* Account for system time used */
3297 acct_update_integrals(p
);
3301 * Account for involuntary wait time.
3302 * @steal: the cpu time spent in involuntary wait
3304 void account_steal_time(cputime_t cputime
)
3306 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3307 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3309 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3313 * Account for idle time.
3314 * @cputime: the cpu time spent in idle wait
3316 void account_idle_time(cputime_t cputime
)
3318 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3319 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3320 struct rq
*rq
= this_rq();
3322 if (atomic_read(&rq
->nr_iowait
) > 0)
3323 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3325 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3328 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3331 * Account a single tick of cpu time.
3332 * @p: the process that the cpu time gets accounted to
3333 * @user_tick: indicates if the tick is a user or a system tick
3335 void account_process_tick(struct task_struct
*p
, int user_tick
)
3337 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3338 struct rq
*rq
= this_rq();
3341 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3342 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3343 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3346 account_idle_time(cputime_one_jiffy
);
3350 * Account multiple ticks of steal time.
3351 * @p: the process from which the cpu time has been stolen
3352 * @ticks: number of stolen ticks
3354 void account_steal_ticks(unsigned long ticks
)
3356 account_steal_time(jiffies_to_cputime(ticks
));
3360 * Account multiple ticks of idle time.
3361 * @ticks: number of stolen ticks
3363 void account_idle_ticks(unsigned long ticks
)
3365 account_idle_time(jiffies_to_cputime(ticks
));
3371 * Use precise platform statistics if available:
3373 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3374 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3380 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3382 struct task_cputime cputime
;
3384 thread_group_cputime(p
, &cputime
);
3386 *ut
= cputime
.utime
;
3387 *st
= cputime
.stime
;
3391 #ifndef nsecs_to_cputime
3392 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3395 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3397 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3400 * Use CFS's precise accounting:
3402 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3407 temp
= (u64
)(rtime
* utime
);
3408 do_div(temp
, total
);
3409 utime
= (cputime_t
)temp
;
3414 * Compare with previous values, to keep monotonicity:
3416 p
->prev_utime
= max(p
->prev_utime
, utime
);
3417 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3419 *ut
= p
->prev_utime
;
3420 *st
= p
->prev_stime
;
3424 * Must be called with siglock held.
3426 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3428 struct signal_struct
*sig
= p
->signal
;
3429 struct task_cputime cputime
;
3430 cputime_t rtime
, utime
, total
;
3432 thread_group_cputime(p
, &cputime
);
3434 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3435 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3440 temp
= (u64
)(rtime
* cputime
.utime
);
3441 do_div(temp
, total
);
3442 utime
= (cputime_t
)temp
;
3446 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3447 sig
->prev_stime
= max(sig
->prev_stime
,
3448 cputime_sub(rtime
, sig
->prev_utime
));
3450 *ut
= sig
->prev_utime
;
3451 *st
= sig
->prev_stime
;
3456 * This function gets called by the timer code, with HZ frequency.
3457 * We call it with interrupts disabled.
3459 * It also gets called by the fork code, when changing the parent's
3462 void scheduler_tick(void)
3464 int cpu
= smp_processor_id();
3465 struct rq
*rq
= cpu_rq(cpu
);
3466 struct task_struct
*curr
= rq
->curr
;
3470 raw_spin_lock(&rq
->lock
);
3471 update_rq_clock(rq
);
3472 update_cpu_load(rq
);
3473 curr
->sched_class
->task_tick(rq
, curr
, 0);
3474 raw_spin_unlock(&rq
->lock
);
3476 perf_event_task_tick(curr
);
3479 rq
->idle_at_tick
= idle_cpu(cpu
);
3480 trigger_load_balance(rq
, cpu
);
3484 notrace
unsigned long get_parent_ip(unsigned long addr
)
3486 if (in_lock_functions(addr
)) {
3487 addr
= CALLER_ADDR2
;
3488 if (in_lock_functions(addr
))
3489 addr
= CALLER_ADDR3
;
3494 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3495 defined(CONFIG_PREEMPT_TRACER))
3497 void __kprobes
add_preempt_count(int val
)
3499 #ifdef CONFIG_DEBUG_PREEMPT
3503 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3506 preempt_count() += val
;
3507 #ifdef CONFIG_DEBUG_PREEMPT
3509 * Spinlock count overflowing soon?
3511 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3514 if (preempt_count() == val
)
3515 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3517 EXPORT_SYMBOL(add_preempt_count
);
3519 void __kprobes
sub_preempt_count(int val
)
3521 #ifdef CONFIG_DEBUG_PREEMPT
3525 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3528 * Is the spinlock portion underflowing?
3530 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3531 !(preempt_count() & PREEMPT_MASK
)))
3535 if (preempt_count() == val
)
3536 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3537 preempt_count() -= val
;
3539 EXPORT_SYMBOL(sub_preempt_count
);
3544 * Print scheduling while atomic bug:
3546 static noinline
void __schedule_bug(struct task_struct
*prev
)
3548 struct pt_regs
*regs
= get_irq_regs();
3550 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3551 prev
->comm
, prev
->pid
, preempt_count());
3553 debug_show_held_locks(prev
);
3555 if (irqs_disabled())
3556 print_irqtrace_events(prev
);
3565 * Various schedule()-time debugging checks and statistics:
3567 static inline void schedule_debug(struct task_struct
*prev
)
3570 * Test if we are atomic. Since do_exit() needs to call into
3571 * schedule() atomically, we ignore that path for now.
3572 * Otherwise, whine if we are scheduling when we should not be.
3574 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3575 __schedule_bug(prev
);
3577 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3579 schedstat_inc(this_rq(), sched_count
);
3580 #ifdef CONFIG_SCHEDSTATS
3581 if (unlikely(prev
->lock_depth
>= 0)) {
3582 schedstat_inc(this_rq(), bkl_count
);
3583 schedstat_inc(prev
, sched_info
.bkl_count
);
3588 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3590 if (prev
->state
== TASK_RUNNING
) {
3591 u64 runtime
= prev
->se
.sum_exec_runtime
;
3593 runtime
-= prev
->se
.prev_sum_exec_runtime
;
3594 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
3597 * In order to avoid avg_overlap growing stale when we are
3598 * indeed overlapping and hence not getting put to sleep, grow
3599 * the avg_overlap on preemption.
3601 * We use the average preemption runtime because that
3602 * correlates to the amount of cache footprint a task can
3605 update_avg(&prev
->se
.avg_overlap
, runtime
);
3607 prev
->sched_class
->put_prev_task(rq
, prev
);
3611 * Pick up the highest-prio task:
3613 static inline struct task_struct
*
3614 pick_next_task(struct rq
*rq
)
3616 const struct sched_class
*class;
3617 struct task_struct
*p
;
3620 * Optimization: we know that if all tasks are in
3621 * the fair class we can call that function directly:
3623 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3624 p
= fair_sched_class
.pick_next_task(rq
);
3629 class = sched_class_highest
;
3631 p
= class->pick_next_task(rq
);
3635 * Will never be NULL as the idle class always
3636 * returns a non-NULL p:
3638 class = class->next
;
3643 * schedule() is the main scheduler function.
3645 asmlinkage
void __sched
schedule(void)
3647 struct task_struct
*prev
, *next
;
3648 unsigned long *switch_count
;
3654 cpu
= smp_processor_id();
3658 switch_count
= &prev
->nivcsw
;
3660 release_kernel_lock(prev
);
3661 need_resched_nonpreemptible
:
3663 schedule_debug(prev
);
3665 if (sched_feat(HRTICK
))
3668 raw_spin_lock_irq(&rq
->lock
);
3669 update_rq_clock(rq
);
3670 clear_tsk_need_resched(prev
);
3672 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3673 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3674 prev
->state
= TASK_RUNNING
;
3676 deactivate_task(rq
, prev
, 1);
3677 switch_count
= &prev
->nvcsw
;
3680 pre_schedule(rq
, prev
);
3682 if (unlikely(!rq
->nr_running
))
3683 idle_balance(cpu
, rq
);
3685 put_prev_task(rq
, prev
);
3686 next
= pick_next_task(rq
);
3688 if (likely(prev
!= next
)) {
3689 sched_info_switch(prev
, next
);
3690 perf_event_task_sched_out(prev
, next
);
3696 context_switch(rq
, prev
, next
); /* unlocks the rq */
3698 * the context switch might have flipped the stack from under
3699 * us, hence refresh the local variables.
3701 cpu
= smp_processor_id();
3704 raw_spin_unlock_irq(&rq
->lock
);
3708 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3710 switch_count
= &prev
->nivcsw
;
3711 goto need_resched_nonpreemptible
;
3714 preempt_enable_no_resched();
3718 EXPORT_SYMBOL(schedule
);
3720 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3722 * Look out! "owner" is an entirely speculative pointer
3723 * access and not reliable.
3725 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3730 if (!sched_feat(OWNER_SPIN
))
3733 #ifdef CONFIG_DEBUG_PAGEALLOC
3735 * Need to access the cpu field knowing that
3736 * DEBUG_PAGEALLOC could have unmapped it if
3737 * the mutex owner just released it and exited.
3739 if (probe_kernel_address(&owner
->cpu
, cpu
))
3746 * Even if the access succeeded (likely case),
3747 * the cpu field may no longer be valid.
3749 if (cpu
>= nr_cpumask_bits
)
3753 * We need to validate that we can do a
3754 * get_cpu() and that we have the percpu area.
3756 if (!cpu_online(cpu
))
3763 * Owner changed, break to re-assess state.
3765 if (lock
->owner
!= owner
)
3769 * Is that owner really running on that cpu?
3771 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3781 #ifdef CONFIG_PREEMPT
3783 * this is the entry point to schedule() from in-kernel preemption
3784 * off of preempt_enable. Kernel preemptions off return from interrupt
3785 * occur there and call schedule directly.
3787 asmlinkage
void __sched
preempt_schedule(void)
3789 struct thread_info
*ti
= current_thread_info();
3792 * If there is a non-zero preempt_count or interrupts are disabled,
3793 * we do not want to preempt the current task. Just return..
3795 if (likely(ti
->preempt_count
|| irqs_disabled()))
3799 add_preempt_count(PREEMPT_ACTIVE
);
3801 sub_preempt_count(PREEMPT_ACTIVE
);
3804 * Check again in case we missed a preemption opportunity
3805 * between schedule and now.
3808 } while (need_resched());
3810 EXPORT_SYMBOL(preempt_schedule
);
3813 * this is the entry point to schedule() from kernel preemption
3814 * off of irq context.
3815 * Note, that this is called and return with irqs disabled. This will
3816 * protect us against recursive calling from irq.
3818 asmlinkage
void __sched
preempt_schedule_irq(void)
3820 struct thread_info
*ti
= current_thread_info();
3822 /* Catch callers which need to be fixed */
3823 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3826 add_preempt_count(PREEMPT_ACTIVE
);
3829 local_irq_disable();
3830 sub_preempt_count(PREEMPT_ACTIVE
);
3833 * Check again in case we missed a preemption opportunity
3834 * between schedule and now.
3837 } while (need_resched());
3840 #endif /* CONFIG_PREEMPT */
3842 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3845 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3847 EXPORT_SYMBOL(default_wake_function
);
3850 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3851 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3852 * number) then we wake all the non-exclusive tasks and one exclusive task.
3854 * There are circumstances in which we can try to wake a task which has already
3855 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3856 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3858 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3859 int nr_exclusive
, int wake_flags
, void *key
)
3861 wait_queue_t
*curr
, *next
;
3863 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3864 unsigned flags
= curr
->flags
;
3866 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3867 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3873 * __wake_up - wake up threads blocked on a waitqueue.
3875 * @mode: which threads
3876 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3877 * @key: is directly passed to the wakeup function
3879 * It may be assumed that this function implies a write memory barrier before
3880 * changing the task state if and only if any tasks are woken up.
3882 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3883 int nr_exclusive
, void *key
)
3885 unsigned long flags
;
3887 spin_lock_irqsave(&q
->lock
, flags
);
3888 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3889 spin_unlock_irqrestore(&q
->lock
, flags
);
3891 EXPORT_SYMBOL(__wake_up
);
3894 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3896 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3898 __wake_up_common(q
, mode
, 1, 0, NULL
);
3901 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3903 __wake_up_common(q
, mode
, 1, 0, key
);
3907 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3909 * @mode: which threads
3910 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3911 * @key: opaque value to be passed to wakeup targets
3913 * The sync wakeup differs that the waker knows that it will schedule
3914 * away soon, so while the target thread will be woken up, it will not
3915 * be migrated to another CPU - ie. the two threads are 'synchronized'
3916 * with each other. This can prevent needless bouncing between CPUs.
3918 * On UP it can prevent extra preemption.
3920 * It may be assumed that this function implies a write memory barrier before
3921 * changing the task state if and only if any tasks are woken up.
3923 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3924 int nr_exclusive
, void *key
)
3926 unsigned long flags
;
3927 int wake_flags
= WF_SYNC
;
3932 if (unlikely(!nr_exclusive
))
3935 spin_lock_irqsave(&q
->lock
, flags
);
3936 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3937 spin_unlock_irqrestore(&q
->lock
, flags
);
3939 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3942 * __wake_up_sync - see __wake_up_sync_key()
3944 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3946 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3948 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3951 * complete: - signals a single thread waiting on this completion
3952 * @x: holds the state of this particular completion
3954 * This will wake up a single thread waiting on this completion. Threads will be
3955 * awakened in the same order in which they were queued.
3957 * See also complete_all(), wait_for_completion() and related routines.
3959 * It may be assumed that this function implies a write memory barrier before
3960 * changing the task state if and only if any tasks are woken up.
3962 void complete(struct completion
*x
)
3964 unsigned long flags
;
3966 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3968 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3969 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3971 EXPORT_SYMBOL(complete
);
3974 * complete_all: - signals all threads waiting on this completion
3975 * @x: holds the state of this particular completion
3977 * This will wake up all threads waiting on this particular completion event.
3979 * It may be assumed that this function implies a write memory barrier before
3980 * changing the task state if and only if any tasks are woken up.
3982 void complete_all(struct completion
*x
)
3984 unsigned long flags
;
3986 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3987 x
->done
+= UINT_MAX
/2;
3988 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3989 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3991 EXPORT_SYMBOL(complete_all
);
3993 static inline long __sched
3994 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3997 DECLARE_WAITQUEUE(wait
, current
);
3999 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4000 __add_wait_queue_tail(&x
->wait
, &wait
);
4002 if (signal_pending_state(state
, current
)) {
4003 timeout
= -ERESTARTSYS
;
4006 __set_current_state(state
);
4007 spin_unlock_irq(&x
->wait
.lock
);
4008 timeout
= schedule_timeout(timeout
);
4009 spin_lock_irq(&x
->wait
.lock
);
4010 } while (!x
->done
&& timeout
);
4011 __remove_wait_queue(&x
->wait
, &wait
);
4016 return timeout
?: 1;
4020 wait_for_common(struct completion
*x
, long timeout
, int state
)
4024 spin_lock_irq(&x
->wait
.lock
);
4025 timeout
= do_wait_for_common(x
, timeout
, state
);
4026 spin_unlock_irq(&x
->wait
.lock
);
4031 * wait_for_completion: - waits for completion of a task
4032 * @x: holds the state of this particular completion
4034 * This waits to be signaled for completion of a specific task. It is NOT
4035 * interruptible and there is no timeout.
4037 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4038 * and interrupt capability. Also see complete().
4040 void __sched
wait_for_completion(struct completion
*x
)
4042 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4044 EXPORT_SYMBOL(wait_for_completion
);
4047 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4048 * @x: holds the state of this particular completion
4049 * @timeout: timeout value in jiffies
4051 * This waits for either a completion of a specific task to be signaled or for a
4052 * specified timeout to expire. The timeout is in jiffies. It is not
4055 unsigned long __sched
4056 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4058 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4060 EXPORT_SYMBOL(wait_for_completion_timeout
);
4063 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4064 * @x: holds the state of this particular completion
4066 * This waits for completion of a specific task to be signaled. It is
4069 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4071 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4072 if (t
== -ERESTARTSYS
)
4076 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4079 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4080 * @x: holds the state of this particular completion
4081 * @timeout: timeout value in jiffies
4083 * This waits for either a completion of a specific task to be signaled or for a
4084 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4086 unsigned long __sched
4087 wait_for_completion_interruptible_timeout(struct completion
*x
,
4088 unsigned long timeout
)
4090 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4092 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4095 * wait_for_completion_killable: - waits for completion of a task (killable)
4096 * @x: holds the state of this particular completion
4098 * This waits to be signaled for completion of a specific task. It can be
4099 * interrupted by a kill signal.
4101 int __sched
wait_for_completion_killable(struct completion
*x
)
4103 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4104 if (t
== -ERESTARTSYS
)
4108 EXPORT_SYMBOL(wait_for_completion_killable
);
4111 * try_wait_for_completion - try to decrement a completion without blocking
4112 * @x: completion structure
4114 * Returns: 0 if a decrement cannot be done without blocking
4115 * 1 if a decrement succeeded.
4117 * If a completion is being used as a counting completion,
4118 * attempt to decrement the counter without blocking. This
4119 * enables us to avoid waiting if the resource the completion
4120 * is protecting is not available.
4122 bool try_wait_for_completion(struct completion
*x
)
4124 unsigned long flags
;
4127 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4132 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4135 EXPORT_SYMBOL(try_wait_for_completion
);
4138 * completion_done - Test to see if a completion has any waiters
4139 * @x: completion structure
4141 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4142 * 1 if there are no waiters.
4145 bool completion_done(struct completion
*x
)
4147 unsigned long flags
;
4150 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4153 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4156 EXPORT_SYMBOL(completion_done
);
4159 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4161 unsigned long flags
;
4164 init_waitqueue_entry(&wait
, current
);
4166 __set_current_state(state
);
4168 spin_lock_irqsave(&q
->lock
, flags
);
4169 __add_wait_queue(q
, &wait
);
4170 spin_unlock(&q
->lock
);
4171 timeout
= schedule_timeout(timeout
);
4172 spin_lock_irq(&q
->lock
);
4173 __remove_wait_queue(q
, &wait
);
4174 spin_unlock_irqrestore(&q
->lock
, flags
);
4179 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4181 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4183 EXPORT_SYMBOL(interruptible_sleep_on
);
4186 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4188 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4190 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4192 void __sched
sleep_on(wait_queue_head_t
*q
)
4194 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4196 EXPORT_SYMBOL(sleep_on
);
4198 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4200 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4202 EXPORT_SYMBOL(sleep_on_timeout
);
4204 #ifdef CONFIG_RT_MUTEXES
4207 * rt_mutex_setprio - set the current priority of a task
4209 * @prio: prio value (kernel-internal form)
4211 * This function changes the 'effective' priority of a task. It does
4212 * not touch ->normal_prio like __setscheduler().
4214 * Used by the rt_mutex code to implement priority inheritance logic.
4216 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4218 unsigned long flags
;
4219 int oldprio
, on_rq
, running
;
4221 const struct sched_class
*prev_class
;
4223 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4225 rq
= task_rq_lock(p
, &flags
);
4226 update_rq_clock(rq
);
4229 prev_class
= p
->sched_class
;
4230 on_rq
= p
->se
.on_rq
;
4231 running
= task_current(rq
, p
);
4233 dequeue_task(rq
, p
, 0);
4235 p
->sched_class
->put_prev_task(rq
, p
);
4238 p
->sched_class
= &rt_sched_class
;
4240 p
->sched_class
= &fair_sched_class
;
4245 p
->sched_class
->set_curr_task(rq
);
4247 enqueue_task(rq
, p
, 0, oldprio
< prio
);
4249 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4251 task_rq_unlock(rq
, &flags
);
4256 void set_user_nice(struct task_struct
*p
, long nice
)
4258 int old_prio
, delta
, on_rq
;
4259 unsigned long flags
;
4262 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4265 * We have to be careful, if called from sys_setpriority(),
4266 * the task might be in the middle of scheduling on another CPU.
4268 rq
= task_rq_lock(p
, &flags
);
4269 update_rq_clock(rq
);
4271 * The RT priorities are set via sched_setscheduler(), but we still
4272 * allow the 'normal' nice value to be set - but as expected
4273 * it wont have any effect on scheduling until the task is
4274 * SCHED_FIFO/SCHED_RR:
4276 if (task_has_rt_policy(p
)) {
4277 p
->static_prio
= NICE_TO_PRIO(nice
);
4280 on_rq
= p
->se
.on_rq
;
4282 dequeue_task(rq
, p
, 0);
4284 p
->static_prio
= NICE_TO_PRIO(nice
);
4287 p
->prio
= effective_prio(p
);
4288 delta
= p
->prio
- old_prio
;
4291 enqueue_task(rq
, p
, 0, false);
4293 * If the task increased its priority or is running and
4294 * lowered its priority, then reschedule its CPU:
4296 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4297 resched_task(rq
->curr
);
4300 task_rq_unlock(rq
, &flags
);
4302 EXPORT_SYMBOL(set_user_nice
);
4305 * can_nice - check if a task can reduce its nice value
4309 int can_nice(const struct task_struct
*p
, const int nice
)
4311 /* convert nice value [19,-20] to rlimit style value [1,40] */
4312 int nice_rlim
= 20 - nice
;
4314 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4315 capable(CAP_SYS_NICE
));
4318 #ifdef __ARCH_WANT_SYS_NICE
4321 * sys_nice - change the priority of the current process.
4322 * @increment: priority increment
4324 * sys_setpriority is a more generic, but much slower function that
4325 * does similar things.
4327 SYSCALL_DEFINE1(nice
, int, increment
)
4332 * Setpriority might change our priority at the same moment.
4333 * We don't have to worry. Conceptually one call occurs first
4334 * and we have a single winner.
4336 if (increment
< -40)
4341 nice
= TASK_NICE(current
) + increment
;
4347 if (increment
< 0 && !can_nice(current
, nice
))
4350 retval
= security_task_setnice(current
, nice
);
4354 set_user_nice(current
, nice
);
4361 * task_prio - return the priority value of a given task.
4362 * @p: the task in question.
4364 * This is the priority value as seen by users in /proc.
4365 * RT tasks are offset by -200. Normal tasks are centered
4366 * around 0, value goes from -16 to +15.
4368 int task_prio(const struct task_struct
*p
)
4370 return p
->prio
- MAX_RT_PRIO
;
4374 * task_nice - return the nice value of a given task.
4375 * @p: the task in question.
4377 int task_nice(const struct task_struct
*p
)
4379 return TASK_NICE(p
);
4381 EXPORT_SYMBOL(task_nice
);
4384 * idle_cpu - is a given cpu idle currently?
4385 * @cpu: the processor in question.
4387 int idle_cpu(int cpu
)
4389 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4393 * idle_task - return the idle task for a given cpu.
4394 * @cpu: the processor in question.
4396 struct task_struct
*idle_task(int cpu
)
4398 return cpu_rq(cpu
)->idle
;
4402 * find_process_by_pid - find a process with a matching PID value.
4403 * @pid: the pid in question.
4405 static struct task_struct
*find_process_by_pid(pid_t pid
)
4407 return pid
? find_task_by_vpid(pid
) : current
;
4410 /* Actually do priority change: must hold rq lock. */
4412 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4414 BUG_ON(p
->se
.on_rq
);
4417 p
->rt_priority
= prio
;
4418 p
->normal_prio
= normal_prio(p
);
4419 /* we are holding p->pi_lock already */
4420 p
->prio
= rt_mutex_getprio(p
);
4421 if (rt_prio(p
->prio
))
4422 p
->sched_class
= &rt_sched_class
;
4424 p
->sched_class
= &fair_sched_class
;
4429 * check the target process has a UID that matches the current process's
4431 static bool check_same_owner(struct task_struct
*p
)
4433 const struct cred
*cred
= current_cred(), *pcred
;
4437 pcred
= __task_cred(p
);
4438 match
= (cred
->euid
== pcred
->euid
||
4439 cred
->euid
== pcred
->uid
);
4444 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4445 struct sched_param
*param
, bool user
)
4447 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4448 unsigned long flags
;
4449 const struct sched_class
*prev_class
;
4453 /* may grab non-irq protected spin_locks */
4454 BUG_ON(in_interrupt());
4456 /* double check policy once rq lock held */
4458 reset_on_fork
= p
->sched_reset_on_fork
;
4459 policy
= oldpolicy
= p
->policy
;
4461 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4462 policy
&= ~SCHED_RESET_ON_FORK
;
4464 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4465 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4466 policy
!= SCHED_IDLE
)
4471 * Valid priorities for SCHED_FIFO and SCHED_RR are
4472 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4473 * SCHED_BATCH and SCHED_IDLE is 0.
4475 if (param
->sched_priority
< 0 ||
4476 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4477 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4479 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4483 * Allow unprivileged RT tasks to decrease priority:
4485 if (user
&& !capable(CAP_SYS_NICE
)) {
4486 if (rt_policy(policy
)) {
4487 unsigned long rlim_rtprio
;
4489 if (!lock_task_sighand(p
, &flags
))
4491 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4492 unlock_task_sighand(p
, &flags
);
4494 /* can't set/change the rt policy */
4495 if (policy
!= p
->policy
&& !rlim_rtprio
)
4498 /* can't increase priority */
4499 if (param
->sched_priority
> p
->rt_priority
&&
4500 param
->sched_priority
> rlim_rtprio
)
4504 * Like positive nice levels, dont allow tasks to
4505 * move out of SCHED_IDLE either:
4507 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4510 /* can't change other user's priorities */
4511 if (!check_same_owner(p
))
4514 /* Normal users shall not reset the sched_reset_on_fork flag */
4515 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4520 #ifdef CONFIG_RT_GROUP_SCHED
4522 * Do not allow realtime tasks into groups that have no runtime
4525 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4526 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4530 retval
= security_task_setscheduler(p
, policy
, param
);
4536 * make sure no PI-waiters arrive (or leave) while we are
4537 * changing the priority of the task:
4539 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4541 * To be able to change p->policy safely, the apropriate
4542 * runqueue lock must be held.
4544 rq
= __task_rq_lock(p
);
4545 /* recheck policy now with rq lock held */
4546 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4547 policy
= oldpolicy
= -1;
4548 __task_rq_unlock(rq
);
4549 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4552 update_rq_clock(rq
);
4553 on_rq
= p
->se
.on_rq
;
4554 running
= task_current(rq
, p
);
4556 deactivate_task(rq
, p
, 0);
4558 p
->sched_class
->put_prev_task(rq
, p
);
4560 p
->sched_reset_on_fork
= reset_on_fork
;
4563 prev_class
= p
->sched_class
;
4564 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4567 p
->sched_class
->set_curr_task(rq
);
4569 activate_task(rq
, p
, 0);
4571 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4573 __task_rq_unlock(rq
);
4574 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4576 rt_mutex_adjust_pi(p
);
4582 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4583 * @p: the task in question.
4584 * @policy: new policy.
4585 * @param: structure containing the new RT priority.
4587 * NOTE that the task may be already dead.
4589 int sched_setscheduler(struct task_struct
*p
, int policy
,
4590 struct sched_param
*param
)
4592 return __sched_setscheduler(p
, policy
, param
, true);
4594 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4597 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4598 * @p: the task in question.
4599 * @policy: new policy.
4600 * @param: structure containing the new RT priority.
4602 * Just like sched_setscheduler, only don't bother checking if the
4603 * current context has permission. For example, this is needed in
4604 * stop_machine(): we create temporary high priority worker threads,
4605 * but our caller might not have that capability.
4607 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4608 struct sched_param
*param
)
4610 return __sched_setscheduler(p
, policy
, param
, false);
4614 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4616 struct sched_param lparam
;
4617 struct task_struct
*p
;
4620 if (!param
|| pid
< 0)
4622 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4627 p
= find_process_by_pid(pid
);
4629 retval
= sched_setscheduler(p
, policy
, &lparam
);
4636 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4637 * @pid: the pid in question.
4638 * @policy: new policy.
4639 * @param: structure containing the new RT priority.
4641 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4642 struct sched_param __user
*, param
)
4644 /* negative values for policy are not valid */
4648 return do_sched_setscheduler(pid
, policy
, param
);
4652 * sys_sched_setparam - set/change the RT priority of a thread
4653 * @pid: the pid in question.
4654 * @param: structure containing the new RT priority.
4656 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4658 return do_sched_setscheduler(pid
, -1, param
);
4662 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4663 * @pid: the pid in question.
4665 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4667 struct task_struct
*p
;
4675 p
= find_process_by_pid(pid
);
4677 retval
= security_task_getscheduler(p
);
4680 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4687 * sys_sched_getparam - get the RT priority of a thread
4688 * @pid: the pid in question.
4689 * @param: structure containing the RT priority.
4691 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4693 struct sched_param lp
;
4694 struct task_struct
*p
;
4697 if (!param
|| pid
< 0)
4701 p
= find_process_by_pid(pid
);
4706 retval
= security_task_getscheduler(p
);
4710 lp
.sched_priority
= p
->rt_priority
;
4714 * This one might sleep, we cannot do it with a spinlock held ...
4716 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4725 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4727 cpumask_var_t cpus_allowed
, new_mask
;
4728 struct task_struct
*p
;
4734 p
= find_process_by_pid(pid
);
4741 /* Prevent p going away */
4745 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4749 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4751 goto out_free_cpus_allowed
;
4754 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4757 retval
= security_task_setscheduler(p
, 0, NULL
);
4761 cpuset_cpus_allowed(p
, cpus_allowed
);
4762 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4764 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4767 cpuset_cpus_allowed(p
, cpus_allowed
);
4768 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4770 * We must have raced with a concurrent cpuset
4771 * update. Just reset the cpus_allowed to the
4772 * cpuset's cpus_allowed
4774 cpumask_copy(new_mask
, cpus_allowed
);
4779 free_cpumask_var(new_mask
);
4780 out_free_cpus_allowed
:
4781 free_cpumask_var(cpus_allowed
);
4788 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4789 struct cpumask
*new_mask
)
4791 if (len
< cpumask_size())
4792 cpumask_clear(new_mask
);
4793 else if (len
> cpumask_size())
4794 len
= cpumask_size();
4796 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4800 * sys_sched_setaffinity - set the cpu affinity of a process
4801 * @pid: pid of the process
4802 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4803 * @user_mask_ptr: user-space pointer to the new cpu mask
4805 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4806 unsigned long __user
*, user_mask_ptr
)
4808 cpumask_var_t new_mask
;
4811 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4814 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4816 retval
= sched_setaffinity(pid
, new_mask
);
4817 free_cpumask_var(new_mask
);
4821 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4823 struct task_struct
*p
;
4824 unsigned long flags
;
4832 p
= find_process_by_pid(pid
);
4836 retval
= security_task_getscheduler(p
);
4840 rq
= task_rq_lock(p
, &flags
);
4841 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4842 task_rq_unlock(rq
, &flags
);
4852 * sys_sched_getaffinity - get the cpu affinity of a process
4853 * @pid: pid of the process
4854 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4855 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4857 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4858 unsigned long __user
*, user_mask_ptr
)
4863 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4865 if (len
& (sizeof(unsigned long)-1))
4868 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4871 ret
= sched_getaffinity(pid
, mask
);
4873 size_t retlen
= min_t(size_t, len
, cpumask_size());
4875 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4880 free_cpumask_var(mask
);
4886 * sys_sched_yield - yield the current processor to other threads.
4888 * This function yields the current CPU to other tasks. If there are no
4889 * other threads running on this CPU then this function will return.
4891 SYSCALL_DEFINE0(sched_yield
)
4893 struct rq
*rq
= this_rq_lock();
4895 schedstat_inc(rq
, yld_count
);
4896 current
->sched_class
->yield_task(rq
);
4899 * Since we are going to call schedule() anyway, there's
4900 * no need to preempt or enable interrupts:
4902 __release(rq
->lock
);
4903 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4904 do_raw_spin_unlock(&rq
->lock
);
4905 preempt_enable_no_resched();
4912 static inline int should_resched(void)
4914 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4917 static void __cond_resched(void)
4919 add_preempt_count(PREEMPT_ACTIVE
);
4921 sub_preempt_count(PREEMPT_ACTIVE
);
4924 int __sched
_cond_resched(void)
4926 if (should_resched()) {
4932 EXPORT_SYMBOL(_cond_resched
);
4935 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4936 * call schedule, and on return reacquire the lock.
4938 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4939 * operations here to prevent schedule() from being called twice (once via
4940 * spin_unlock(), once by hand).
4942 int __cond_resched_lock(spinlock_t
*lock
)
4944 int resched
= should_resched();
4947 lockdep_assert_held(lock
);
4949 if (spin_needbreak(lock
) || resched
) {
4960 EXPORT_SYMBOL(__cond_resched_lock
);
4962 int __sched
__cond_resched_softirq(void)
4964 BUG_ON(!in_softirq());
4966 if (should_resched()) {
4974 EXPORT_SYMBOL(__cond_resched_softirq
);
4977 * yield - yield the current processor to other threads.
4979 * This is a shortcut for kernel-space yielding - it marks the
4980 * thread runnable and calls sys_sched_yield().
4982 void __sched
yield(void)
4984 set_current_state(TASK_RUNNING
);
4987 EXPORT_SYMBOL(yield
);
4990 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4991 * that process accounting knows that this is a task in IO wait state.
4993 void __sched
io_schedule(void)
4995 struct rq
*rq
= raw_rq();
4997 delayacct_blkio_start();
4998 atomic_inc(&rq
->nr_iowait
);
4999 current
->in_iowait
= 1;
5001 current
->in_iowait
= 0;
5002 atomic_dec(&rq
->nr_iowait
);
5003 delayacct_blkio_end();
5005 EXPORT_SYMBOL(io_schedule
);
5007 long __sched
io_schedule_timeout(long timeout
)
5009 struct rq
*rq
= raw_rq();
5012 delayacct_blkio_start();
5013 atomic_inc(&rq
->nr_iowait
);
5014 current
->in_iowait
= 1;
5015 ret
= schedule_timeout(timeout
);
5016 current
->in_iowait
= 0;
5017 atomic_dec(&rq
->nr_iowait
);
5018 delayacct_blkio_end();
5023 * sys_sched_get_priority_max - return maximum RT priority.
5024 * @policy: scheduling class.
5026 * this syscall returns the maximum rt_priority that can be used
5027 * by a given scheduling class.
5029 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
5036 ret
= MAX_USER_RT_PRIO
-1;
5048 * sys_sched_get_priority_min - return minimum RT priority.
5049 * @policy: scheduling class.
5051 * this syscall returns the minimum rt_priority that can be used
5052 * by a given scheduling class.
5054 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5072 * sys_sched_rr_get_interval - return the default timeslice of a process.
5073 * @pid: pid of the process.
5074 * @interval: userspace pointer to the timeslice value.
5076 * this syscall writes the default timeslice value of a given process
5077 * into the user-space timespec buffer. A value of '0' means infinity.
5079 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5080 struct timespec __user
*, interval
)
5082 struct task_struct
*p
;
5083 unsigned int time_slice
;
5084 unsigned long flags
;
5094 p
= find_process_by_pid(pid
);
5098 retval
= security_task_getscheduler(p
);
5102 rq
= task_rq_lock(p
, &flags
);
5103 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5104 task_rq_unlock(rq
, &flags
);
5107 jiffies_to_timespec(time_slice
, &t
);
5108 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5116 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5118 void sched_show_task(struct task_struct
*p
)
5120 unsigned long free
= 0;
5123 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5124 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5125 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5126 #if BITS_PER_LONG == 32
5127 if (state
== TASK_RUNNING
)
5128 printk(KERN_CONT
" running ");
5130 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5132 if (state
== TASK_RUNNING
)
5133 printk(KERN_CONT
" running task ");
5135 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5137 #ifdef CONFIG_DEBUG_STACK_USAGE
5138 free
= stack_not_used(p
);
5140 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5141 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5142 (unsigned long)task_thread_info(p
)->flags
);
5144 show_stack(p
, NULL
);
5147 void show_state_filter(unsigned long state_filter
)
5149 struct task_struct
*g
, *p
;
5151 #if BITS_PER_LONG == 32
5153 " task PC stack pid father\n");
5156 " task PC stack pid father\n");
5158 read_lock(&tasklist_lock
);
5159 do_each_thread(g
, p
) {
5161 * reset the NMI-timeout, listing all files on a slow
5162 * console might take alot of time:
5164 touch_nmi_watchdog();
5165 if (!state_filter
|| (p
->state
& state_filter
))
5167 } while_each_thread(g
, p
);
5169 touch_all_softlockup_watchdogs();
5171 #ifdef CONFIG_SCHED_DEBUG
5172 sysrq_sched_debug_show();
5174 read_unlock(&tasklist_lock
);
5176 * Only show locks if all tasks are dumped:
5179 debug_show_all_locks();
5182 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5184 idle
->sched_class
= &idle_sched_class
;
5188 * init_idle - set up an idle thread for a given CPU
5189 * @idle: task in question
5190 * @cpu: cpu the idle task belongs to
5192 * NOTE: this function does not set the idle thread's NEED_RESCHED
5193 * flag, to make booting more robust.
5195 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5197 struct rq
*rq
= cpu_rq(cpu
);
5198 unsigned long flags
;
5200 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5203 idle
->state
= TASK_RUNNING
;
5204 idle
->se
.exec_start
= sched_clock();
5206 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5207 __set_task_cpu(idle
, cpu
);
5209 rq
->curr
= rq
->idle
= idle
;
5210 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5213 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5215 /* Set the preempt count _outside_ the spinlocks! */
5216 #if defined(CONFIG_PREEMPT)
5217 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5219 task_thread_info(idle
)->preempt_count
= 0;
5222 * The idle tasks have their own, simple scheduling class:
5224 idle
->sched_class
= &idle_sched_class
;
5225 ftrace_graph_init_task(idle
);
5229 * In a system that switches off the HZ timer nohz_cpu_mask
5230 * indicates which cpus entered this state. This is used
5231 * in the rcu update to wait only for active cpus. For system
5232 * which do not switch off the HZ timer nohz_cpu_mask should
5233 * always be CPU_BITS_NONE.
5235 cpumask_var_t nohz_cpu_mask
;
5238 * Increase the granularity value when there are more CPUs,
5239 * because with more CPUs the 'effective latency' as visible
5240 * to users decreases. But the relationship is not linear,
5241 * so pick a second-best guess by going with the log2 of the
5244 * This idea comes from the SD scheduler of Con Kolivas:
5246 static int get_update_sysctl_factor(void)
5248 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5249 unsigned int factor
;
5251 switch (sysctl_sched_tunable_scaling
) {
5252 case SCHED_TUNABLESCALING_NONE
:
5255 case SCHED_TUNABLESCALING_LINEAR
:
5258 case SCHED_TUNABLESCALING_LOG
:
5260 factor
= 1 + ilog2(cpus
);
5267 static void update_sysctl(void)
5269 unsigned int factor
= get_update_sysctl_factor();
5271 #define SET_SYSCTL(name) \
5272 (sysctl_##name = (factor) * normalized_sysctl_##name)
5273 SET_SYSCTL(sched_min_granularity
);
5274 SET_SYSCTL(sched_latency
);
5275 SET_SYSCTL(sched_wakeup_granularity
);
5276 SET_SYSCTL(sched_shares_ratelimit
);
5280 static inline void sched_init_granularity(void)
5287 * This is how migration works:
5289 * 1) we queue a struct migration_req structure in the source CPU's
5290 * runqueue and wake up that CPU's migration thread.
5291 * 2) we down() the locked semaphore => thread blocks.
5292 * 3) migration thread wakes up (implicitly it forces the migrated
5293 * thread off the CPU)
5294 * 4) it gets the migration request and checks whether the migrated
5295 * task is still in the wrong runqueue.
5296 * 5) if it's in the wrong runqueue then the migration thread removes
5297 * it and puts it into the right queue.
5298 * 6) migration thread up()s the semaphore.
5299 * 7) we wake up and the migration is done.
5303 * Change a given task's CPU affinity. Migrate the thread to a
5304 * proper CPU and schedule it away if the CPU it's executing on
5305 * is removed from the allowed bitmask.
5307 * NOTE: the caller must have a valid reference to the task, the
5308 * task must not exit() & deallocate itself prematurely. The
5309 * call is not atomic; no spinlocks may be held.
5311 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5313 struct migration_req req
;
5314 unsigned long flags
;
5318 rq
= task_rq_lock(p
, &flags
);
5320 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5325 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5326 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5331 if (p
->sched_class
->set_cpus_allowed
)
5332 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5334 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5335 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5338 /* Can the task run on the task's current CPU? If so, we're done */
5339 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5342 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
5343 /* Need help from migration thread: drop lock and wait. */
5344 struct task_struct
*mt
= rq
->migration_thread
;
5346 get_task_struct(mt
);
5347 task_rq_unlock(rq
, &flags
);
5348 wake_up_process(mt
);
5349 put_task_struct(mt
);
5350 wait_for_completion(&req
.done
);
5351 tlb_migrate_finish(p
->mm
);
5355 task_rq_unlock(rq
, &flags
);
5359 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5362 * Move (not current) task off this cpu, onto dest cpu. We're doing
5363 * this because either it can't run here any more (set_cpus_allowed()
5364 * away from this CPU, or CPU going down), or because we're
5365 * attempting to rebalance this task on exec (sched_exec).
5367 * So we race with normal scheduler movements, but that's OK, as long
5368 * as the task is no longer on this CPU.
5370 * Returns non-zero if task was successfully migrated.
5372 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5374 struct rq
*rq_dest
, *rq_src
;
5377 if (unlikely(!cpu_active(dest_cpu
)))
5380 rq_src
= cpu_rq(src_cpu
);
5381 rq_dest
= cpu_rq(dest_cpu
);
5383 double_rq_lock(rq_src
, rq_dest
);
5384 /* Already moved. */
5385 if (task_cpu(p
) != src_cpu
)
5387 /* Affinity changed (again). */
5388 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5392 * If we're not on a rq, the next wake-up will ensure we're
5396 deactivate_task(rq_src
, p
, 0);
5397 set_task_cpu(p
, dest_cpu
);
5398 activate_task(rq_dest
, p
, 0);
5399 check_preempt_curr(rq_dest
, p
, 0);
5404 double_rq_unlock(rq_src
, rq_dest
);
5408 #define RCU_MIGRATION_IDLE 0
5409 #define RCU_MIGRATION_NEED_QS 1
5410 #define RCU_MIGRATION_GOT_QS 2
5411 #define RCU_MIGRATION_MUST_SYNC 3
5414 * migration_thread - this is a highprio system thread that performs
5415 * thread migration by bumping thread off CPU then 'pushing' onto
5418 static int migration_thread(void *data
)
5421 int cpu
= (long)data
;
5425 BUG_ON(rq
->migration_thread
!= current
);
5427 set_current_state(TASK_INTERRUPTIBLE
);
5428 while (!kthread_should_stop()) {
5429 struct migration_req
*req
;
5430 struct list_head
*head
;
5432 raw_spin_lock_irq(&rq
->lock
);
5434 if (cpu_is_offline(cpu
)) {
5435 raw_spin_unlock_irq(&rq
->lock
);
5439 if (rq
->active_balance
) {
5440 active_load_balance(rq
, cpu
);
5441 rq
->active_balance
= 0;
5444 head
= &rq
->migration_queue
;
5446 if (list_empty(head
)) {
5447 raw_spin_unlock_irq(&rq
->lock
);
5449 set_current_state(TASK_INTERRUPTIBLE
);
5452 req
= list_entry(head
->next
, struct migration_req
, list
);
5453 list_del_init(head
->next
);
5455 if (req
->task
!= NULL
) {
5456 raw_spin_unlock(&rq
->lock
);
5457 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5458 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
5459 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
5460 raw_spin_unlock(&rq
->lock
);
5462 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
5463 raw_spin_unlock(&rq
->lock
);
5464 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
5468 complete(&req
->done
);
5470 __set_current_state(TASK_RUNNING
);
5475 #ifdef CONFIG_HOTPLUG_CPU
5477 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5481 local_irq_disable();
5482 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5488 * Figure out where task on dead CPU should go, use force if necessary.
5490 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5495 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5497 /* It can have affinity changed while we were choosing. */
5498 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
5503 * While a dead CPU has no uninterruptible tasks queued at this point,
5504 * it might still have a nonzero ->nr_uninterruptible counter, because
5505 * for performance reasons the counter is not stricly tracking tasks to
5506 * their home CPUs. So we just add the counter to another CPU's counter,
5507 * to keep the global sum constant after CPU-down:
5509 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5511 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5512 unsigned long flags
;
5514 local_irq_save(flags
);
5515 double_rq_lock(rq_src
, rq_dest
);
5516 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5517 rq_src
->nr_uninterruptible
= 0;
5518 double_rq_unlock(rq_src
, rq_dest
);
5519 local_irq_restore(flags
);
5522 /* Run through task list and migrate tasks from the dead cpu. */
5523 static void migrate_live_tasks(int src_cpu
)
5525 struct task_struct
*p
, *t
;
5527 read_lock(&tasklist_lock
);
5529 do_each_thread(t
, p
) {
5533 if (task_cpu(p
) == src_cpu
)
5534 move_task_off_dead_cpu(src_cpu
, p
);
5535 } while_each_thread(t
, p
);
5537 read_unlock(&tasklist_lock
);
5541 * Schedules idle task to be the next runnable task on current CPU.
5542 * It does so by boosting its priority to highest possible.
5543 * Used by CPU offline code.
5545 void sched_idle_next(void)
5547 int this_cpu
= smp_processor_id();
5548 struct rq
*rq
= cpu_rq(this_cpu
);
5549 struct task_struct
*p
= rq
->idle
;
5550 unsigned long flags
;
5552 /* cpu has to be offline */
5553 BUG_ON(cpu_online(this_cpu
));
5556 * Strictly not necessary since rest of the CPUs are stopped by now
5557 * and interrupts disabled on the current cpu.
5559 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5561 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5563 update_rq_clock(rq
);
5564 activate_task(rq
, p
, 0);
5566 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5570 * Ensures that the idle task is using init_mm right before its cpu goes
5573 void idle_task_exit(void)
5575 struct mm_struct
*mm
= current
->active_mm
;
5577 BUG_ON(cpu_online(smp_processor_id()));
5580 switch_mm(mm
, &init_mm
, current
);
5584 /* called under rq->lock with disabled interrupts */
5585 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5587 struct rq
*rq
= cpu_rq(dead_cpu
);
5589 /* Must be exiting, otherwise would be on tasklist. */
5590 BUG_ON(!p
->exit_state
);
5592 /* Cannot have done final schedule yet: would have vanished. */
5593 BUG_ON(p
->state
== TASK_DEAD
);
5598 * Drop lock around migration; if someone else moves it,
5599 * that's OK. No task can be added to this CPU, so iteration is
5602 raw_spin_unlock_irq(&rq
->lock
);
5603 move_task_off_dead_cpu(dead_cpu
, p
);
5604 raw_spin_lock_irq(&rq
->lock
);
5609 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5610 static void migrate_dead_tasks(unsigned int dead_cpu
)
5612 struct rq
*rq
= cpu_rq(dead_cpu
);
5613 struct task_struct
*next
;
5616 if (!rq
->nr_running
)
5618 update_rq_clock(rq
);
5619 next
= pick_next_task(rq
);
5622 next
->sched_class
->put_prev_task(rq
, next
);
5623 migrate_dead(dead_cpu
, next
);
5629 * remove the tasks which were accounted by rq from calc_load_tasks.
5631 static void calc_global_load_remove(struct rq
*rq
)
5633 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5634 rq
->calc_load_active
= 0;
5636 #endif /* CONFIG_HOTPLUG_CPU */
5638 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5640 static struct ctl_table sd_ctl_dir
[] = {
5642 .procname
= "sched_domain",
5648 static struct ctl_table sd_ctl_root
[] = {
5650 .procname
= "kernel",
5652 .child
= sd_ctl_dir
,
5657 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5659 struct ctl_table
*entry
=
5660 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5665 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5667 struct ctl_table
*entry
;
5670 * In the intermediate directories, both the child directory and
5671 * procname are dynamically allocated and could fail but the mode
5672 * will always be set. In the lowest directory the names are
5673 * static strings and all have proc handlers.
5675 for (entry
= *tablep
; entry
->mode
; entry
++) {
5677 sd_free_ctl_entry(&entry
->child
);
5678 if (entry
->proc_handler
== NULL
)
5679 kfree(entry
->procname
);
5687 set_table_entry(struct ctl_table
*entry
,
5688 const char *procname
, void *data
, int maxlen
,
5689 mode_t mode
, proc_handler
*proc_handler
)
5691 entry
->procname
= procname
;
5693 entry
->maxlen
= maxlen
;
5695 entry
->proc_handler
= proc_handler
;
5698 static struct ctl_table
*
5699 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5701 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5706 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5707 sizeof(long), 0644, proc_doulongvec_minmax
);
5708 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5709 sizeof(long), 0644, proc_doulongvec_minmax
);
5710 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5711 sizeof(int), 0644, proc_dointvec_minmax
);
5712 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5713 sizeof(int), 0644, proc_dointvec_minmax
);
5714 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5715 sizeof(int), 0644, proc_dointvec_minmax
);
5716 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5717 sizeof(int), 0644, proc_dointvec_minmax
);
5718 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5719 sizeof(int), 0644, proc_dointvec_minmax
);
5720 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5721 sizeof(int), 0644, proc_dointvec_minmax
);
5722 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5723 sizeof(int), 0644, proc_dointvec_minmax
);
5724 set_table_entry(&table
[9], "cache_nice_tries",
5725 &sd
->cache_nice_tries
,
5726 sizeof(int), 0644, proc_dointvec_minmax
);
5727 set_table_entry(&table
[10], "flags", &sd
->flags
,
5728 sizeof(int), 0644, proc_dointvec_minmax
);
5729 set_table_entry(&table
[11], "name", sd
->name
,
5730 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5731 /* &table[12] is terminator */
5736 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5738 struct ctl_table
*entry
, *table
;
5739 struct sched_domain
*sd
;
5740 int domain_num
= 0, i
;
5743 for_each_domain(cpu
, sd
)
5745 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5750 for_each_domain(cpu
, sd
) {
5751 snprintf(buf
, 32, "domain%d", i
);
5752 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5754 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5761 static struct ctl_table_header
*sd_sysctl_header
;
5762 static void register_sched_domain_sysctl(void)
5764 int i
, cpu_num
= num_possible_cpus();
5765 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5768 WARN_ON(sd_ctl_dir
[0].child
);
5769 sd_ctl_dir
[0].child
= entry
;
5774 for_each_possible_cpu(i
) {
5775 snprintf(buf
, 32, "cpu%d", i
);
5776 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5778 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5782 WARN_ON(sd_sysctl_header
);
5783 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5786 /* may be called multiple times per register */
5787 static void unregister_sched_domain_sysctl(void)
5789 if (sd_sysctl_header
)
5790 unregister_sysctl_table(sd_sysctl_header
);
5791 sd_sysctl_header
= NULL
;
5792 if (sd_ctl_dir
[0].child
)
5793 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5796 static void register_sched_domain_sysctl(void)
5799 static void unregister_sched_domain_sysctl(void)
5804 static void set_rq_online(struct rq
*rq
)
5807 const struct sched_class
*class;
5809 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5812 for_each_class(class) {
5813 if (class->rq_online
)
5814 class->rq_online(rq
);
5819 static void set_rq_offline(struct rq
*rq
)
5822 const struct sched_class
*class;
5824 for_each_class(class) {
5825 if (class->rq_offline
)
5826 class->rq_offline(rq
);
5829 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5835 * migration_call - callback that gets triggered when a CPU is added.
5836 * Here we can start up the necessary migration thread for the new CPU.
5838 static int __cpuinit
5839 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5841 struct task_struct
*p
;
5842 int cpu
= (long)hcpu
;
5843 unsigned long flags
;
5848 case CPU_UP_PREPARE
:
5849 case CPU_UP_PREPARE_FROZEN
:
5850 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
5853 kthread_bind(p
, cpu
);
5854 /* Must be high prio: stop_machine expects to yield to it. */
5855 rq
= task_rq_lock(p
, &flags
);
5856 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5857 task_rq_unlock(rq
, &flags
);
5859 cpu_rq(cpu
)->migration_thread
= p
;
5860 rq
->calc_load_update
= calc_load_update
;
5864 case CPU_ONLINE_FROZEN
:
5865 /* Strictly unnecessary, as first user will wake it. */
5866 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5868 /* Update our root-domain */
5870 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5872 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5876 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5879 #ifdef CONFIG_HOTPLUG_CPU
5880 case CPU_UP_CANCELED
:
5881 case CPU_UP_CANCELED_FROZEN
:
5882 if (!cpu_rq(cpu
)->migration_thread
)
5884 /* Unbind it from offline cpu so it can run. Fall thru. */
5885 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5886 cpumask_any(cpu_online_mask
));
5887 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5888 put_task_struct(cpu_rq(cpu
)->migration_thread
);
5889 cpu_rq(cpu
)->migration_thread
= NULL
;
5893 case CPU_DEAD_FROZEN
:
5894 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
5895 migrate_live_tasks(cpu
);
5897 kthread_stop(rq
->migration_thread
);
5898 put_task_struct(rq
->migration_thread
);
5899 rq
->migration_thread
= NULL
;
5900 /* Idle task back to normal (off runqueue, low prio) */
5901 raw_spin_lock_irq(&rq
->lock
);
5902 update_rq_clock(rq
);
5903 deactivate_task(rq
, rq
->idle
, 0);
5904 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5905 rq
->idle
->sched_class
= &idle_sched_class
;
5906 migrate_dead_tasks(cpu
);
5907 raw_spin_unlock_irq(&rq
->lock
);
5909 migrate_nr_uninterruptible(rq
);
5910 BUG_ON(rq
->nr_running
!= 0);
5911 calc_global_load_remove(rq
);
5913 * No need to migrate the tasks: it was best-effort if
5914 * they didn't take sched_hotcpu_mutex. Just wake up
5917 raw_spin_lock_irq(&rq
->lock
);
5918 while (!list_empty(&rq
->migration_queue
)) {
5919 struct migration_req
*req
;
5921 req
= list_entry(rq
->migration_queue
.next
,
5922 struct migration_req
, list
);
5923 list_del_init(&req
->list
);
5924 raw_spin_unlock_irq(&rq
->lock
);
5925 complete(&req
->done
);
5926 raw_spin_lock_irq(&rq
->lock
);
5928 raw_spin_unlock_irq(&rq
->lock
);
5932 case CPU_DYING_FROZEN
:
5933 /* Update our root-domain */
5935 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5937 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5940 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5948 * Register at high priority so that task migration (migrate_all_tasks)
5949 * happens before everything else. This has to be lower priority than
5950 * the notifier in the perf_event subsystem, though.
5952 static struct notifier_block __cpuinitdata migration_notifier
= {
5953 .notifier_call
= migration_call
,
5957 static int __init
migration_init(void)
5959 void *cpu
= (void *)(long)smp_processor_id();
5962 /* Start one for the boot CPU: */
5963 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5964 BUG_ON(err
== NOTIFY_BAD
);
5965 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5966 register_cpu_notifier(&migration_notifier
);
5970 early_initcall(migration_init
);
5975 #ifdef CONFIG_SCHED_DEBUG
5977 static __read_mostly
int sched_domain_debug_enabled
;
5979 static int __init
sched_domain_debug_setup(char *str
)
5981 sched_domain_debug_enabled
= 1;
5985 early_param("sched_debug", sched_domain_debug_setup
);
5987 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5988 struct cpumask
*groupmask
)
5990 struct sched_group
*group
= sd
->groups
;
5993 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5994 cpumask_clear(groupmask
);
5996 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5998 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5999 printk("does not load-balance\n");
6001 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6006 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
6008 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
6009 printk(KERN_ERR
"ERROR: domain->span does not contain "
6012 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
6013 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6017 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6021 printk(KERN_ERR
"ERROR: group is NULL\n");
6025 if (!group
->cpu_power
) {
6026 printk(KERN_CONT
"\n");
6027 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6032 if (!cpumask_weight(sched_group_cpus(group
))) {
6033 printk(KERN_CONT
"\n");
6034 printk(KERN_ERR
"ERROR: empty group\n");
6038 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
6039 printk(KERN_CONT
"\n");
6040 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6044 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
6046 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
6048 printk(KERN_CONT
" %s", str
);
6049 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
6050 printk(KERN_CONT
" (cpu_power = %d)",
6054 group
= group
->next
;
6055 } while (group
!= sd
->groups
);
6056 printk(KERN_CONT
"\n");
6058 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
6059 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6062 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
6063 printk(KERN_ERR
"ERROR: parent span is not a superset "
6064 "of domain->span\n");
6068 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6070 cpumask_var_t groupmask
;
6073 if (!sched_domain_debug_enabled
)
6077 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6081 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6083 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
6084 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6089 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6096 free_cpumask_var(groupmask
);
6098 #else /* !CONFIG_SCHED_DEBUG */
6099 # define sched_domain_debug(sd, cpu) do { } while (0)
6100 #endif /* CONFIG_SCHED_DEBUG */
6102 static int sd_degenerate(struct sched_domain
*sd
)
6104 if (cpumask_weight(sched_domain_span(sd
)) == 1)
6107 /* Following flags need at least 2 groups */
6108 if (sd
->flags
& (SD_LOAD_BALANCE
|
6109 SD_BALANCE_NEWIDLE
|
6113 SD_SHARE_PKG_RESOURCES
)) {
6114 if (sd
->groups
!= sd
->groups
->next
)
6118 /* Following flags don't use groups */
6119 if (sd
->flags
& (SD_WAKE_AFFINE
))
6126 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6128 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6130 if (sd_degenerate(parent
))
6133 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
6136 /* Flags needing groups don't count if only 1 group in parent */
6137 if (parent
->groups
== parent
->groups
->next
) {
6138 pflags
&= ~(SD_LOAD_BALANCE
|
6139 SD_BALANCE_NEWIDLE
|
6143 SD_SHARE_PKG_RESOURCES
);
6144 if (nr_node_ids
== 1)
6145 pflags
&= ~SD_SERIALIZE
;
6147 if (~cflags
& pflags
)
6153 static void free_rootdomain(struct root_domain
*rd
)
6155 synchronize_sched();
6157 cpupri_cleanup(&rd
->cpupri
);
6159 free_cpumask_var(rd
->rto_mask
);
6160 free_cpumask_var(rd
->online
);
6161 free_cpumask_var(rd
->span
);
6165 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6167 struct root_domain
*old_rd
= NULL
;
6168 unsigned long flags
;
6170 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6175 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6178 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6181 * If we dont want to free the old_rt yet then
6182 * set old_rd to NULL to skip the freeing later
6185 if (!atomic_dec_and_test(&old_rd
->refcount
))
6189 atomic_inc(&rd
->refcount
);
6192 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6193 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6196 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6199 free_rootdomain(old_rd
);
6202 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6204 gfp_t gfp
= GFP_KERNEL
;
6206 memset(rd
, 0, sizeof(*rd
));
6211 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6213 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6215 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6218 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6223 free_cpumask_var(rd
->rto_mask
);
6225 free_cpumask_var(rd
->online
);
6227 free_cpumask_var(rd
->span
);
6232 static void init_defrootdomain(void)
6234 init_rootdomain(&def_root_domain
, true);
6236 atomic_set(&def_root_domain
.refcount
, 1);
6239 static struct root_domain
*alloc_rootdomain(void)
6241 struct root_domain
*rd
;
6243 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6247 if (init_rootdomain(rd
, false) != 0) {
6256 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6257 * hold the hotplug lock.
6260 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6262 struct rq
*rq
= cpu_rq(cpu
);
6263 struct sched_domain
*tmp
;
6265 /* Remove the sched domains which do not contribute to scheduling. */
6266 for (tmp
= sd
; tmp
; ) {
6267 struct sched_domain
*parent
= tmp
->parent
;
6271 if (sd_parent_degenerate(tmp
, parent
)) {
6272 tmp
->parent
= parent
->parent
;
6274 parent
->parent
->child
= tmp
;
6279 if (sd
&& sd_degenerate(sd
)) {
6285 sched_domain_debug(sd
, cpu
);
6287 rq_attach_root(rq
, rd
);
6288 rcu_assign_pointer(rq
->sd
, sd
);
6291 /* cpus with isolated domains */
6292 static cpumask_var_t cpu_isolated_map
;
6294 /* Setup the mask of cpus configured for isolated domains */
6295 static int __init
isolated_cpu_setup(char *str
)
6297 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6298 cpulist_parse(str
, cpu_isolated_map
);
6302 __setup("isolcpus=", isolated_cpu_setup
);
6305 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6306 * to a function which identifies what group(along with sched group) a CPU
6307 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6308 * (due to the fact that we keep track of groups covered with a struct cpumask).
6310 * init_sched_build_groups will build a circular linked list of the groups
6311 * covered by the given span, and will set each group's ->cpumask correctly,
6312 * and ->cpu_power to 0.
6315 init_sched_build_groups(const struct cpumask
*span
,
6316 const struct cpumask
*cpu_map
,
6317 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6318 struct sched_group
**sg
,
6319 struct cpumask
*tmpmask
),
6320 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6322 struct sched_group
*first
= NULL
, *last
= NULL
;
6325 cpumask_clear(covered
);
6327 for_each_cpu(i
, span
) {
6328 struct sched_group
*sg
;
6329 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6332 if (cpumask_test_cpu(i
, covered
))
6335 cpumask_clear(sched_group_cpus(sg
));
6338 for_each_cpu(j
, span
) {
6339 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6342 cpumask_set_cpu(j
, covered
);
6343 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6354 #define SD_NODES_PER_DOMAIN 16
6359 * find_next_best_node - find the next node to include in a sched_domain
6360 * @node: node whose sched_domain we're building
6361 * @used_nodes: nodes already in the sched_domain
6363 * Find the next node to include in a given scheduling domain. Simply
6364 * finds the closest node not already in the @used_nodes map.
6366 * Should use nodemask_t.
6368 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6370 int i
, n
, val
, min_val
, best_node
= 0;
6374 for (i
= 0; i
< nr_node_ids
; i
++) {
6375 /* Start at @node */
6376 n
= (node
+ i
) % nr_node_ids
;
6378 if (!nr_cpus_node(n
))
6381 /* Skip already used nodes */
6382 if (node_isset(n
, *used_nodes
))
6385 /* Simple min distance search */
6386 val
= node_distance(node
, n
);
6388 if (val
< min_val
) {
6394 node_set(best_node
, *used_nodes
);
6399 * sched_domain_node_span - get a cpumask for a node's sched_domain
6400 * @node: node whose cpumask we're constructing
6401 * @span: resulting cpumask
6403 * Given a node, construct a good cpumask for its sched_domain to span. It
6404 * should be one that prevents unnecessary balancing, but also spreads tasks
6407 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6409 nodemask_t used_nodes
;
6412 cpumask_clear(span
);
6413 nodes_clear(used_nodes
);
6415 cpumask_or(span
, span
, cpumask_of_node(node
));
6416 node_set(node
, used_nodes
);
6418 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6419 int next_node
= find_next_best_node(node
, &used_nodes
);
6421 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6424 #endif /* CONFIG_NUMA */
6426 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6429 * The cpus mask in sched_group and sched_domain hangs off the end.
6431 * ( See the the comments in include/linux/sched.h:struct sched_group
6432 * and struct sched_domain. )
6434 struct static_sched_group
{
6435 struct sched_group sg
;
6436 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6439 struct static_sched_domain
{
6440 struct sched_domain sd
;
6441 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6447 cpumask_var_t domainspan
;
6448 cpumask_var_t covered
;
6449 cpumask_var_t notcovered
;
6451 cpumask_var_t nodemask
;
6452 cpumask_var_t this_sibling_map
;
6453 cpumask_var_t this_core_map
;
6454 cpumask_var_t send_covered
;
6455 cpumask_var_t tmpmask
;
6456 struct sched_group
**sched_group_nodes
;
6457 struct root_domain
*rd
;
6461 sa_sched_groups
= 0,
6466 sa_this_sibling_map
,
6468 sa_sched_group_nodes
,
6478 * SMT sched-domains:
6480 #ifdef CONFIG_SCHED_SMT
6481 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6482 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6485 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6486 struct sched_group
**sg
, struct cpumask
*unused
)
6489 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6492 #endif /* CONFIG_SCHED_SMT */
6495 * multi-core sched-domains:
6497 #ifdef CONFIG_SCHED_MC
6498 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6499 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6500 #endif /* CONFIG_SCHED_MC */
6502 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6504 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6505 struct sched_group
**sg
, struct cpumask
*mask
)
6509 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6510 group
= cpumask_first(mask
);
6512 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6515 #elif defined(CONFIG_SCHED_MC)
6517 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6518 struct sched_group
**sg
, struct cpumask
*unused
)
6521 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6526 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6527 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6530 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6531 struct sched_group
**sg
, struct cpumask
*mask
)
6534 #ifdef CONFIG_SCHED_MC
6535 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6536 group
= cpumask_first(mask
);
6537 #elif defined(CONFIG_SCHED_SMT)
6538 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6539 group
= cpumask_first(mask
);
6544 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6550 * The init_sched_build_groups can't handle what we want to do with node
6551 * groups, so roll our own. Now each node has its own list of groups which
6552 * gets dynamically allocated.
6554 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6555 static struct sched_group
***sched_group_nodes_bycpu
;
6557 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6558 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6560 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6561 struct sched_group
**sg
,
6562 struct cpumask
*nodemask
)
6566 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6567 group
= cpumask_first(nodemask
);
6570 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6574 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6576 struct sched_group
*sg
= group_head
;
6582 for_each_cpu(j
, sched_group_cpus(sg
)) {
6583 struct sched_domain
*sd
;
6585 sd
= &per_cpu(phys_domains
, j
).sd
;
6586 if (j
!= group_first_cpu(sd
->groups
)) {
6588 * Only add "power" once for each
6594 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6597 } while (sg
!= group_head
);
6600 static int build_numa_sched_groups(struct s_data
*d
,
6601 const struct cpumask
*cpu_map
, int num
)
6603 struct sched_domain
*sd
;
6604 struct sched_group
*sg
, *prev
;
6607 cpumask_clear(d
->covered
);
6608 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6609 if (cpumask_empty(d
->nodemask
)) {
6610 d
->sched_group_nodes
[num
] = NULL
;
6614 sched_domain_node_span(num
, d
->domainspan
);
6615 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6617 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6620 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6624 d
->sched_group_nodes
[num
] = sg
;
6626 for_each_cpu(j
, d
->nodemask
) {
6627 sd
= &per_cpu(node_domains
, j
).sd
;
6632 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6634 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6637 for (j
= 0; j
< nr_node_ids
; j
++) {
6638 n
= (num
+ j
) % nr_node_ids
;
6639 cpumask_complement(d
->notcovered
, d
->covered
);
6640 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6641 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6642 if (cpumask_empty(d
->tmpmask
))
6644 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6645 if (cpumask_empty(d
->tmpmask
))
6647 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6651 "Can not alloc domain group for node %d\n", j
);
6655 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6656 sg
->next
= prev
->next
;
6657 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6664 #endif /* CONFIG_NUMA */
6667 /* Free memory allocated for various sched_group structures */
6668 static void free_sched_groups(const struct cpumask
*cpu_map
,
6669 struct cpumask
*nodemask
)
6673 for_each_cpu(cpu
, cpu_map
) {
6674 struct sched_group
**sched_group_nodes
6675 = sched_group_nodes_bycpu
[cpu
];
6677 if (!sched_group_nodes
)
6680 for (i
= 0; i
< nr_node_ids
; i
++) {
6681 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6683 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6684 if (cpumask_empty(nodemask
))
6694 if (oldsg
!= sched_group_nodes
[i
])
6697 kfree(sched_group_nodes
);
6698 sched_group_nodes_bycpu
[cpu
] = NULL
;
6701 #else /* !CONFIG_NUMA */
6702 static void free_sched_groups(const struct cpumask
*cpu_map
,
6703 struct cpumask
*nodemask
)
6706 #endif /* CONFIG_NUMA */
6709 * Initialize sched groups cpu_power.
6711 * cpu_power indicates the capacity of sched group, which is used while
6712 * distributing the load between different sched groups in a sched domain.
6713 * Typically cpu_power for all the groups in a sched domain will be same unless
6714 * there are asymmetries in the topology. If there are asymmetries, group
6715 * having more cpu_power will pickup more load compared to the group having
6718 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6720 struct sched_domain
*child
;
6721 struct sched_group
*group
;
6725 WARN_ON(!sd
|| !sd
->groups
);
6727 if (cpu
!= group_first_cpu(sd
->groups
))
6732 sd
->groups
->cpu_power
= 0;
6735 power
= SCHED_LOAD_SCALE
;
6736 weight
= cpumask_weight(sched_domain_span(sd
));
6738 * SMT siblings share the power of a single core.
6739 * Usually multiple threads get a better yield out of
6740 * that one core than a single thread would have,
6741 * reflect that in sd->smt_gain.
6743 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6744 power
*= sd
->smt_gain
;
6746 power
>>= SCHED_LOAD_SHIFT
;
6748 sd
->groups
->cpu_power
+= power
;
6753 * Add cpu_power of each child group to this groups cpu_power.
6755 group
= child
->groups
;
6757 sd
->groups
->cpu_power
+= group
->cpu_power
;
6758 group
= group
->next
;
6759 } while (group
!= child
->groups
);
6763 * Initializers for schedule domains
6764 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6767 #ifdef CONFIG_SCHED_DEBUG
6768 # define SD_INIT_NAME(sd, type) sd->name = #type
6770 # define SD_INIT_NAME(sd, type) do { } while (0)
6773 #define SD_INIT(sd, type) sd_init_##type(sd)
6775 #define SD_INIT_FUNC(type) \
6776 static noinline void sd_init_##type(struct sched_domain *sd) \
6778 memset(sd, 0, sizeof(*sd)); \
6779 *sd = SD_##type##_INIT; \
6780 sd->level = SD_LV_##type; \
6781 SD_INIT_NAME(sd, type); \
6786 SD_INIT_FUNC(ALLNODES
)
6789 #ifdef CONFIG_SCHED_SMT
6790 SD_INIT_FUNC(SIBLING
)
6792 #ifdef CONFIG_SCHED_MC
6796 static int default_relax_domain_level
= -1;
6798 static int __init
setup_relax_domain_level(char *str
)
6802 val
= simple_strtoul(str
, NULL
, 0);
6803 if (val
< SD_LV_MAX
)
6804 default_relax_domain_level
= val
;
6808 __setup("relax_domain_level=", setup_relax_domain_level
);
6810 static void set_domain_attribute(struct sched_domain
*sd
,
6811 struct sched_domain_attr
*attr
)
6815 if (!attr
|| attr
->relax_domain_level
< 0) {
6816 if (default_relax_domain_level
< 0)
6819 request
= default_relax_domain_level
;
6821 request
= attr
->relax_domain_level
;
6822 if (request
< sd
->level
) {
6823 /* turn off idle balance on this domain */
6824 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6826 /* turn on idle balance on this domain */
6827 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6831 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6832 const struct cpumask
*cpu_map
)
6835 case sa_sched_groups
:
6836 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6837 d
->sched_group_nodes
= NULL
;
6839 free_rootdomain(d
->rd
); /* fall through */
6841 free_cpumask_var(d
->tmpmask
); /* fall through */
6842 case sa_send_covered
:
6843 free_cpumask_var(d
->send_covered
); /* fall through */
6844 case sa_this_core_map
:
6845 free_cpumask_var(d
->this_core_map
); /* fall through */
6846 case sa_this_sibling_map
:
6847 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6849 free_cpumask_var(d
->nodemask
); /* fall through */
6850 case sa_sched_group_nodes
:
6852 kfree(d
->sched_group_nodes
); /* fall through */
6854 free_cpumask_var(d
->notcovered
); /* fall through */
6856 free_cpumask_var(d
->covered
); /* fall through */
6858 free_cpumask_var(d
->domainspan
); /* fall through */
6865 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6866 const struct cpumask
*cpu_map
)
6869 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6871 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6872 return sa_domainspan
;
6873 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6875 /* Allocate the per-node list of sched groups */
6876 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6877 sizeof(struct sched_group
*), GFP_KERNEL
);
6878 if (!d
->sched_group_nodes
) {
6879 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6880 return sa_notcovered
;
6882 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6884 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6885 return sa_sched_group_nodes
;
6886 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6888 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6889 return sa_this_sibling_map
;
6890 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6891 return sa_this_core_map
;
6892 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6893 return sa_send_covered
;
6894 d
->rd
= alloc_rootdomain();
6896 printk(KERN_WARNING
"Cannot alloc root domain\n");
6899 return sa_rootdomain
;
6902 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6903 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6905 struct sched_domain
*sd
= NULL
;
6907 struct sched_domain
*parent
;
6910 if (cpumask_weight(cpu_map
) >
6911 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6912 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6913 SD_INIT(sd
, ALLNODES
);
6914 set_domain_attribute(sd
, attr
);
6915 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6916 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6921 sd
= &per_cpu(node_domains
, i
).sd
;
6923 set_domain_attribute(sd
, attr
);
6924 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6925 sd
->parent
= parent
;
6928 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6933 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6934 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6935 struct sched_domain
*parent
, int i
)
6937 struct sched_domain
*sd
;
6938 sd
= &per_cpu(phys_domains
, i
).sd
;
6940 set_domain_attribute(sd
, attr
);
6941 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6942 sd
->parent
= parent
;
6945 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6949 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6950 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6951 struct sched_domain
*parent
, int i
)
6953 struct sched_domain
*sd
= parent
;
6954 #ifdef CONFIG_SCHED_MC
6955 sd
= &per_cpu(core_domains
, i
).sd
;
6957 set_domain_attribute(sd
, attr
);
6958 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6959 sd
->parent
= parent
;
6961 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6966 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6967 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6968 struct sched_domain
*parent
, int i
)
6970 struct sched_domain
*sd
= parent
;
6971 #ifdef CONFIG_SCHED_SMT
6972 sd
= &per_cpu(cpu_domains
, i
).sd
;
6973 SD_INIT(sd
, SIBLING
);
6974 set_domain_attribute(sd
, attr
);
6975 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6976 sd
->parent
= parent
;
6978 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6983 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6984 const struct cpumask
*cpu_map
, int cpu
)
6987 #ifdef CONFIG_SCHED_SMT
6988 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6989 cpumask_and(d
->this_sibling_map
, cpu_map
,
6990 topology_thread_cpumask(cpu
));
6991 if (cpu
== cpumask_first(d
->this_sibling_map
))
6992 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
6994 d
->send_covered
, d
->tmpmask
);
6997 #ifdef CONFIG_SCHED_MC
6998 case SD_LV_MC
: /* set up multi-core groups */
6999 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
7000 if (cpu
== cpumask_first(d
->this_core_map
))
7001 init_sched_build_groups(d
->this_core_map
, cpu_map
,
7003 d
->send_covered
, d
->tmpmask
);
7006 case SD_LV_CPU
: /* set up physical groups */
7007 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
7008 if (!cpumask_empty(d
->nodemask
))
7009 init_sched_build_groups(d
->nodemask
, cpu_map
,
7011 d
->send_covered
, d
->tmpmask
);
7014 case SD_LV_ALLNODES
:
7015 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
7016 d
->send_covered
, d
->tmpmask
);
7025 * Build sched domains for a given set of cpus and attach the sched domains
7026 * to the individual cpus
7028 static int __build_sched_domains(const struct cpumask
*cpu_map
,
7029 struct sched_domain_attr
*attr
)
7031 enum s_alloc alloc_state
= sa_none
;
7033 struct sched_domain
*sd
;
7039 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
7040 if (alloc_state
!= sa_rootdomain
)
7042 alloc_state
= sa_sched_groups
;
7045 * Set up domains for cpus specified by the cpu_map.
7047 for_each_cpu(i
, cpu_map
) {
7048 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
7051 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
7052 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7053 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7054 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
7057 for_each_cpu(i
, cpu_map
) {
7058 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
7059 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
7062 /* Set up physical groups */
7063 for (i
= 0; i
< nr_node_ids
; i
++)
7064 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
7067 /* Set up node groups */
7069 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
7071 for (i
= 0; i
< nr_node_ids
; i
++)
7072 if (build_numa_sched_groups(&d
, cpu_map
, i
))
7076 /* Calculate CPU power for physical packages and nodes */
7077 #ifdef CONFIG_SCHED_SMT
7078 for_each_cpu(i
, cpu_map
) {
7079 sd
= &per_cpu(cpu_domains
, i
).sd
;
7080 init_sched_groups_power(i
, sd
);
7083 #ifdef CONFIG_SCHED_MC
7084 for_each_cpu(i
, cpu_map
) {
7085 sd
= &per_cpu(core_domains
, i
).sd
;
7086 init_sched_groups_power(i
, sd
);
7090 for_each_cpu(i
, cpu_map
) {
7091 sd
= &per_cpu(phys_domains
, i
).sd
;
7092 init_sched_groups_power(i
, sd
);
7096 for (i
= 0; i
< nr_node_ids
; i
++)
7097 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
7099 if (d
.sd_allnodes
) {
7100 struct sched_group
*sg
;
7102 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
7104 init_numa_sched_groups_power(sg
);
7108 /* Attach the domains */
7109 for_each_cpu(i
, cpu_map
) {
7110 #ifdef CONFIG_SCHED_SMT
7111 sd
= &per_cpu(cpu_domains
, i
).sd
;
7112 #elif defined(CONFIG_SCHED_MC)
7113 sd
= &per_cpu(core_domains
, i
).sd
;
7115 sd
= &per_cpu(phys_domains
, i
).sd
;
7117 cpu_attach_domain(sd
, d
.rd
, i
);
7120 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
7121 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
7125 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
7129 static int build_sched_domains(const struct cpumask
*cpu_map
)
7131 return __build_sched_domains(cpu_map
, NULL
);
7134 static cpumask_var_t
*doms_cur
; /* current sched domains */
7135 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7136 static struct sched_domain_attr
*dattr_cur
;
7137 /* attribues of custom domains in 'doms_cur' */
7140 * Special case: If a kmalloc of a doms_cur partition (array of
7141 * cpumask) fails, then fallback to a single sched domain,
7142 * as determined by the single cpumask fallback_doms.
7144 static cpumask_var_t fallback_doms
;
7147 * arch_update_cpu_topology lets virtualized architectures update the
7148 * cpu core maps. It is supposed to return 1 if the topology changed
7149 * or 0 if it stayed the same.
7151 int __attribute__((weak
)) arch_update_cpu_topology(void)
7156 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7159 cpumask_var_t
*doms
;
7161 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7164 for (i
= 0; i
< ndoms
; i
++) {
7165 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7166 free_sched_domains(doms
, i
);
7173 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7176 for (i
= 0; i
< ndoms
; i
++)
7177 free_cpumask_var(doms
[i
]);
7182 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7183 * For now this just excludes isolated cpus, but could be used to
7184 * exclude other special cases in the future.
7186 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7190 arch_update_cpu_topology();
7192 doms_cur
= alloc_sched_domains(ndoms_cur
);
7194 doms_cur
= &fallback_doms
;
7195 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7197 err
= build_sched_domains(doms_cur
[0]);
7198 register_sched_domain_sysctl();
7203 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7204 struct cpumask
*tmpmask
)
7206 free_sched_groups(cpu_map
, tmpmask
);
7210 * Detach sched domains from a group of cpus specified in cpu_map
7211 * These cpus will now be attached to the NULL domain
7213 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7215 /* Save because hotplug lock held. */
7216 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7219 for_each_cpu(i
, cpu_map
)
7220 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7221 synchronize_sched();
7222 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7225 /* handle null as "default" */
7226 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7227 struct sched_domain_attr
*new, int idx_new
)
7229 struct sched_domain_attr tmp
;
7236 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7237 new ? (new + idx_new
) : &tmp
,
7238 sizeof(struct sched_domain_attr
));
7242 * Partition sched domains as specified by the 'ndoms_new'
7243 * cpumasks in the array doms_new[] of cpumasks. This compares
7244 * doms_new[] to the current sched domain partitioning, doms_cur[].
7245 * It destroys each deleted domain and builds each new domain.
7247 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7248 * The masks don't intersect (don't overlap.) We should setup one
7249 * sched domain for each mask. CPUs not in any of the cpumasks will
7250 * not be load balanced. If the same cpumask appears both in the
7251 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7254 * The passed in 'doms_new' should be allocated using
7255 * alloc_sched_domains. This routine takes ownership of it and will
7256 * free_sched_domains it when done with it. If the caller failed the
7257 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7258 * and partition_sched_domains() will fallback to the single partition
7259 * 'fallback_doms', it also forces the domains to be rebuilt.
7261 * If doms_new == NULL it will be replaced with cpu_online_mask.
7262 * ndoms_new == 0 is a special case for destroying existing domains,
7263 * and it will not create the default domain.
7265 * Call with hotplug lock held
7267 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7268 struct sched_domain_attr
*dattr_new
)
7273 mutex_lock(&sched_domains_mutex
);
7275 /* always unregister in case we don't destroy any domains */
7276 unregister_sched_domain_sysctl();
7278 /* Let architecture update cpu core mappings. */
7279 new_topology
= arch_update_cpu_topology();
7281 n
= doms_new
? ndoms_new
: 0;
7283 /* Destroy deleted domains */
7284 for (i
= 0; i
< ndoms_cur
; i
++) {
7285 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7286 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7287 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7290 /* no match - a current sched domain not in new doms_new[] */
7291 detach_destroy_domains(doms_cur
[i
]);
7296 if (doms_new
== NULL
) {
7298 doms_new
= &fallback_doms
;
7299 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7300 WARN_ON_ONCE(dattr_new
);
7303 /* Build new domains */
7304 for (i
= 0; i
< ndoms_new
; i
++) {
7305 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7306 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7307 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7310 /* no match - add a new doms_new */
7311 __build_sched_domains(doms_new
[i
],
7312 dattr_new
? dattr_new
+ i
: NULL
);
7317 /* Remember the new sched domains */
7318 if (doms_cur
!= &fallback_doms
)
7319 free_sched_domains(doms_cur
, ndoms_cur
);
7320 kfree(dattr_cur
); /* kfree(NULL) is safe */
7321 doms_cur
= doms_new
;
7322 dattr_cur
= dattr_new
;
7323 ndoms_cur
= ndoms_new
;
7325 register_sched_domain_sysctl();
7327 mutex_unlock(&sched_domains_mutex
);
7330 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7331 static void arch_reinit_sched_domains(void)
7335 /* Destroy domains first to force the rebuild */
7336 partition_sched_domains(0, NULL
, NULL
);
7338 rebuild_sched_domains();
7342 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7344 unsigned int level
= 0;
7346 if (sscanf(buf
, "%u", &level
) != 1)
7350 * level is always be positive so don't check for
7351 * level < POWERSAVINGS_BALANCE_NONE which is 0
7352 * What happens on 0 or 1 byte write,
7353 * need to check for count as well?
7356 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7360 sched_smt_power_savings
= level
;
7362 sched_mc_power_savings
= level
;
7364 arch_reinit_sched_domains();
7369 #ifdef CONFIG_SCHED_MC
7370 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7371 struct sysdev_class_attribute
*attr
,
7374 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7376 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7377 struct sysdev_class_attribute
*attr
,
7378 const char *buf
, size_t count
)
7380 return sched_power_savings_store(buf
, count
, 0);
7382 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7383 sched_mc_power_savings_show
,
7384 sched_mc_power_savings_store
);
7387 #ifdef CONFIG_SCHED_SMT
7388 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7389 struct sysdev_class_attribute
*attr
,
7392 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7394 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7395 struct sysdev_class_attribute
*attr
,
7396 const char *buf
, size_t count
)
7398 return sched_power_savings_store(buf
, count
, 1);
7400 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7401 sched_smt_power_savings_show
,
7402 sched_smt_power_savings_store
);
7405 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7409 #ifdef CONFIG_SCHED_SMT
7411 err
= sysfs_create_file(&cls
->kset
.kobj
,
7412 &attr_sched_smt_power_savings
.attr
);
7414 #ifdef CONFIG_SCHED_MC
7415 if (!err
&& mc_capable())
7416 err
= sysfs_create_file(&cls
->kset
.kobj
,
7417 &attr_sched_mc_power_savings
.attr
);
7421 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7423 #ifndef CONFIG_CPUSETS
7425 * Add online and remove offline CPUs from the scheduler domains.
7426 * When cpusets are enabled they take over this function.
7428 static int update_sched_domains(struct notifier_block
*nfb
,
7429 unsigned long action
, void *hcpu
)
7433 case CPU_ONLINE_FROZEN
:
7434 case CPU_DOWN_PREPARE
:
7435 case CPU_DOWN_PREPARE_FROZEN
:
7436 case CPU_DOWN_FAILED
:
7437 case CPU_DOWN_FAILED_FROZEN
:
7438 partition_sched_domains(1, NULL
, NULL
);
7447 static int update_runtime(struct notifier_block
*nfb
,
7448 unsigned long action
, void *hcpu
)
7450 int cpu
= (int)(long)hcpu
;
7453 case CPU_DOWN_PREPARE
:
7454 case CPU_DOWN_PREPARE_FROZEN
:
7455 disable_runtime(cpu_rq(cpu
));
7458 case CPU_DOWN_FAILED
:
7459 case CPU_DOWN_FAILED_FROZEN
:
7461 case CPU_ONLINE_FROZEN
:
7462 enable_runtime(cpu_rq(cpu
));
7470 void __init
sched_init_smp(void)
7472 cpumask_var_t non_isolated_cpus
;
7474 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7475 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7477 #if defined(CONFIG_NUMA)
7478 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7480 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7483 mutex_lock(&sched_domains_mutex
);
7484 arch_init_sched_domains(cpu_active_mask
);
7485 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7486 if (cpumask_empty(non_isolated_cpus
))
7487 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7488 mutex_unlock(&sched_domains_mutex
);
7491 #ifndef CONFIG_CPUSETS
7492 /* XXX: Theoretical race here - CPU may be hotplugged now */
7493 hotcpu_notifier(update_sched_domains
, 0);
7496 /* RT runtime code needs to handle some hotplug events */
7497 hotcpu_notifier(update_runtime
, 0);
7501 /* Move init over to a non-isolated CPU */
7502 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7504 sched_init_granularity();
7505 free_cpumask_var(non_isolated_cpus
);
7507 init_sched_rt_class();
7510 void __init
sched_init_smp(void)
7512 sched_init_granularity();
7514 #endif /* CONFIG_SMP */
7516 const_debug
unsigned int sysctl_timer_migration
= 1;
7518 int in_sched_functions(unsigned long addr
)
7520 return in_lock_functions(addr
) ||
7521 (addr
>= (unsigned long)__sched_text_start
7522 && addr
< (unsigned long)__sched_text_end
);
7525 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7527 cfs_rq
->tasks_timeline
= RB_ROOT
;
7528 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7529 #ifdef CONFIG_FAIR_GROUP_SCHED
7532 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7535 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7537 struct rt_prio_array
*array
;
7540 array
= &rt_rq
->active
;
7541 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7542 INIT_LIST_HEAD(array
->queue
+ i
);
7543 __clear_bit(i
, array
->bitmap
);
7545 /* delimiter for bitsearch: */
7546 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7548 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7549 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7551 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7555 rt_rq
->rt_nr_migratory
= 0;
7556 rt_rq
->overloaded
= 0;
7557 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7561 rt_rq
->rt_throttled
= 0;
7562 rt_rq
->rt_runtime
= 0;
7563 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7565 #ifdef CONFIG_RT_GROUP_SCHED
7566 rt_rq
->rt_nr_boosted
= 0;
7571 #ifdef CONFIG_FAIR_GROUP_SCHED
7572 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7573 struct sched_entity
*se
, int cpu
, int add
,
7574 struct sched_entity
*parent
)
7576 struct rq
*rq
= cpu_rq(cpu
);
7577 tg
->cfs_rq
[cpu
] = cfs_rq
;
7578 init_cfs_rq(cfs_rq
, rq
);
7581 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7584 /* se could be NULL for init_task_group */
7589 se
->cfs_rq
= &rq
->cfs
;
7591 se
->cfs_rq
= parent
->my_q
;
7594 se
->load
.weight
= tg
->shares
;
7595 se
->load
.inv_weight
= 0;
7596 se
->parent
= parent
;
7600 #ifdef CONFIG_RT_GROUP_SCHED
7601 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7602 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7603 struct sched_rt_entity
*parent
)
7605 struct rq
*rq
= cpu_rq(cpu
);
7607 tg
->rt_rq
[cpu
] = rt_rq
;
7608 init_rt_rq(rt_rq
, rq
);
7610 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7612 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7614 tg
->rt_se
[cpu
] = rt_se
;
7619 rt_se
->rt_rq
= &rq
->rt
;
7621 rt_se
->rt_rq
= parent
->my_q
;
7623 rt_se
->my_q
= rt_rq
;
7624 rt_se
->parent
= parent
;
7625 INIT_LIST_HEAD(&rt_se
->run_list
);
7629 void __init
sched_init(void)
7632 unsigned long alloc_size
= 0, ptr
;
7634 #ifdef CONFIG_FAIR_GROUP_SCHED
7635 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7637 #ifdef CONFIG_RT_GROUP_SCHED
7638 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7640 #ifdef CONFIG_CPUMASK_OFFSTACK
7641 alloc_size
+= num_possible_cpus() * cpumask_size();
7644 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7646 #ifdef CONFIG_FAIR_GROUP_SCHED
7647 init_task_group
.se
= (struct sched_entity
**)ptr
;
7648 ptr
+= nr_cpu_ids
* sizeof(void **);
7650 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7651 ptr
+= nr_cpu_ids
* sizeof(void **);
7653 #endif /* CONFIG_FAIR_GROUP_SCHED */
7654 #ifdef CONFIG_RT_GROUP_SCHED
7655 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7656 ptr
+= nr_cpu_ids
* sizeof(void **);
7658 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7659 ptr
+= nr_cpu_ids
* sizeof(void **);
7661 #endif /* CONFIG_RT_GROUP_SCHED */
7662 #ifdef CONFIG_CPUMASK_OFFSTACK
7663 for_each_possible_cpu(i
) {
7664 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7665 ptr
+= cpumask_size();
7667 #endif /* CONFIG_CPUMASK_OFFSTACK */
7671 init_defrootdomain();
7674 init_rt_bandwidth(&def_rt_bandwidth
,
7675 global_rt_period(), global_rt_runtime());
7677 #ifdef CONFIG_RT_GROUP_SCHED
7678 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7679 global_rt_period(), global_rt_runtime());
7680 #endif /* CONFIG_RT_GROUP_SCHED */
7682 #ifdef CONFIG_CGROUP_SCHED
7683 list_add(&init_task_group
.list
, &task_groups
);
7684 INIT_LIST_HEAD(&init_task_group
.children
);
7686 #endif /* CONFIG_CGROUP_SCHED */
7688 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7689 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7690 __alignof__(unsigned long));
7692 for_each_possible_cpu(i
) {
7696 raw_spin_lock_init(&rq
->lock
);
7698 rq
->calc_load_active
= 0;
7699 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7700 init_cfs_rq(&rq
->cfs
, rq
);
7701 init_rt_rq(&rq
->rt
, rq
);
7702 #ifdef CONFIG_FAIR_GROUP_SCHED
7703 init_task_group
.shares
= init_task_group_load
;
7704 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7705 #ifdef CONFIG_CGROUP_SCHED
7707 * How much cpu bandwidth does init_task_group get?
7709 * In case of task-groups formed thr' the cgroup filesystem, it
7710 * gets 100% of the cpu resources in the system. This overall
7711 * system cpu resource is divided among the tasks of
7712 * init_task_group and its child task-groups in a fair manner,
7713 * based on each entity's (task or task-group's) weight
7714 * (se->load.weight).
7716 * In other words, if init_task_group has 10 tasks of weight
7717 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7718 * then A0's share of the cpu resource is:
7720 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7722 * We achieve this by letting init_task_group's tasks sit
7723 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7725 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7727 #endif /* CONFIG_FAIR_GROUP_SCHED */
7729 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7730 #ifdef CONFIG_RT_GROUP_SCHED
7731 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7732 #ifdef CONFIG_CGROUP_SCHED
7733 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7737 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7738 rq
->cpu_load
[j
] = 0;
7742 rq
->post_schedule
= 0;
7743 rq
->active_balance
= 0;
7744 rq
->next_balance
= jiffies
;
7748 rq
->migration_thread
= NULL
;
7750 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7751 INIT_LIST_HEAD(&rq
->migration_queue
);
7752 rq_attach_root(rq
, &def_root_domain
);
7755 atomic_set(&rq
->nr_iowait
, 0);
7758 set_load_weight(&init_task
);
7760 #ifdef CONFIG_PREEMPT_NOTIFIERS
7761 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7765 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7768 #ifdef CONFIG_RT_MUTEXES
7769 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7773 * The boot idle thread does lazy MMU switching as well:
7775 atomic_inc(&init_mm
.mm_count
);
7776 enter_lazy_tlb(&init_mm
, current
);
7779 * Make us the idle thread. Technically, schedule() should not be
7780 * called from this thread, however somewhere below it might be,
7781 * but because we are the idle thread, we just pick up running again
7782 * when this runqueue becomes "idle".
7784 init_idle(current
, smp_processor_id());
7786 calc_load_update
= jiffies
+ LOAD_FREQ
;
7789 * During early bootup we pretend to be a normal task:
7791 current
->sched_class
= &fair_sched_class
;
7793 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7794 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7797 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7798 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7800 /* May be allocated at isolcpus cmdline parse time */
7801 if (cpu_isolated_map
== NULL
)
7802 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7807 scheduler_running
= 1;
7810 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7811 static inline int preempt_count_equals(int preempt_offset
)
7813 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7815 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7818 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7821 static unsigned long prev_jiffy
; /* ratelimiting */
7823 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7824 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7826 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7828 prev_jiffy
= jiffies
;
7831 "BUG: sleeping function called from invalid context at %s:%d\n",
7834 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7835 in_atomic(), irqs_disabled(),
7836 current
->pid
, current
->comm
);
7838 debug_show_held_locks(current
);
7839 if (irqs_disabled())
7840 print_irqtrace_events(current
);
7844 EXPORT_SYMBOL(__might_sleep
);
7847 #ifdef CONFIG_MAGIC_SYSRQ
7848 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7852 update_rq_clock(rq
);
7853 on_rq
= p
->se
.on_rq
;
7855 deactivate_task(rq
, p
, 0);
7856 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7858 activate_task(rq
, p
, 0);
7859 resched_task(rq
->curr
);
7863 void normalize_rt_tasks(void)
7865 struct task_struct
*g
, *p
;
7866 unsigned long flags
;
7869 read_lock_irqsave(&tasklist_lock
, flags
);
7870 do_each_thread(g
, p
) {
7872 * Only normalize user tasks:
7877 p
->se
.exec_start
= 0;
7878 #ifdef CONFIG_SCHEDSTATS
7879 p
->se
.wait_start
= 0;
7880 p
->se
.sleep_start
= 0;
7881 p
->se
.block_start
= 0;
7886 * Renice negative nice level userspace
7889 if (TASK_NICE(p
) < 0 && p
->mm
)
7890 set_user_nice(p
, 0);
7894 raw_spin_lock(&p
->pi_lock
);
7895 rq
= __task_rq_lock(p
);
7897 normalize_task(rq
, p
);
7899 __task_rq_unlock(rq
);
7900 raw_spin_unlock(&p
->pi_lock
);
7901 } while_each_thread(g
, p
);
7903 read_unlock_irqrestore(&tasklist_lock
, flags
);
7906 #endif /* CONFIG_MAGIC_SYSRQ */
7910 * These functions are only useful for the IA64 MCA handling.
7912 * They can only be called when the whole system has been
7913 * stopped - every CPU needs to be quiescent, and no scheduling
7914 * activity can take place. Using them for anything else would
7915 * be a serious bug, and as a result, they aren't even visible
7916 * under any other configuration.
7920 * curr_task - return the current task for a given cpu.
7921 * @cpu: the processor in question.
7923 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7925 struct task_struct
*curr_task(int cpu
)
7927 return cpu_curr(cpu
);
7931 * set_curr_task - set the current task for a given cpu.
7932 * @cpu: the processor in question.
7933 * @p: the task pointer to set.
7935 * Description: This function must only be used when non-maskable interrupts
7936 * are serviced on a separate stack. It allows the architecture to switch the
7937 * notion of the current task on a cpu in a non-blocking manner. This function
7938 * must be called with all CPU's synchronized, and interrupts disabled, the
7939 * and caller must save the original value of the current task (see
7940 * curr_task() above) and restore that value before reenabling interrupts and
7941 * re-starting the system.
7943 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7945 void set_curr_task(int cpu
, struct task_struct
*p
)
7952 #ifdef CONFIG_FAIR_GROUP_SCHED
7953 static void free_fair_sched_group(struct task_group
*tg
)
7957 for_each_possible_cpu(i
) {
7959 kfree(tg
->cfs_rq
[i
]);
7969 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7971 struct cfs_rq
*cfs_rq
;
7972 struct sched_entity
*se
;
7976 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7979 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7983 tg
->shares
= NICE_0_LOAD
;
7985 for_each_possible_cpu(i
) {
7988 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7989 GFP_KERNEL
, cpu_to_node(i
));
7993 se
= kzalloc_node(sizeof(struct sched_entity
),
7994 GFP_KERNEL
, cpu_to_node(i
));
7998 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
8009 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8011 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8012 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8015 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8017 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8019 #else /* !CONFG_FAIR_GROUP_SCHED */
8020 static inline void free_fair_sched_group(struct task_group
*tg
)
8025 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8030 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8034 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8037 #endif /* CONFIG_FAIR_GROUP_SCHED */
8039 #ifdef CONFIG_RT_GROUP_SCHED
8040 static void free_rt_sched_group(struct task_group
*tg
)
8044 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8046 for_each_possible_cpu(i
) {
8048 kfree(tg
->rt_rq
[i
]);
8050 kfree(tg
->rt_se
[i
]);
8058 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8060 struct rt_rq
*rt_rq
;
8061 struct sched_rt_entity
*rt_se
;
8065 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8068 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8072 init_rt_bandwidth(&tg
->rt_bandwidth
,
8073 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8075 for_each_possible_cpu(i
) {
8078 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
8079 GFP_KERNEL
, cpu_to_node(i
));
8083 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
8084 GFP_KERNEL
, cpu_to_node(i
));
8088 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
8099 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8101 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8102 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8105 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8107 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8109 #else /* !CONFIG_RT_GROUP_SCHED */
8110 static inline void free_rt_sched_group(struct task_group
*tg
)
8115 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8120 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8124 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8127 #endif /* CONFIG_RT_GROUP_SCHED */
8129 #ifdef CONFIG_CGROUP_SCHED
8130 static void free_sched_group(struct task_group
*tg
)
8132 free_fair_sched_group(tg
);
8133 free_rt_sched_group(tg
);
8137 /* allocate runqueue etc for a new task group */
8138 struct task_group
*sched_create_group(struct task_group
*parent
)
8140 struct task_group
*tg
;
8141 unsigned long flags
;
8144 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8146 return ERR_PTR(-ENOMEM
);
8148 if (!alloc_fair_sched_group(tg
, parent
))
8151 if (!alloc_rt_sched_group(tg
, parent
))
8154 spin_lock_irqsave(&task_group_lock
, flags
);
8155 for_each_possible_cpu(i
) {
8156 register_fair_sched_group(tg
, i
);
8157 register_rt_sched_group(tg
, i
);
8159 list_add_rcu(&tg
->list
, &task_groups
);
8161 WARN_ON(!parent
); /* root should already exist */
8163 tg
->parent
= parent
;
8164 INIT_LIST_HEAD(&tg
->children
);
8165 list_add_rcu(&tg
->siblings
, &parent
->children
);
8166 spin_unlock_irqrestore(&task_group_lock
, flags
);
8171 free_sched_group(tg
);
8172 return ERR_PTR(-ENOMEM
);
8175 /* rcu callback to free various structures associated with a task group */
8176 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8178 /* now it should be safe to free those cfs_rqs */
8179 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8182 /* Destroy runqueue etc associated with a task group */
8183 void sched_destroy_group(struct task_group
*tg
)
8185 unsigned long flags
;
8188 spin_lock_irqsave(&task_group_lock
, flags
);
8189 for_each_possible_cpu(i
) {
8190 unregister_fair_sched_group(tg
, i
);
8191 unregister_rt_sched_group(tg
, i
);
8193 list_del_rcu(&tg
->list
);
8194 list_del_rcu(&tg
->siblings
);
8195 spin_unlock_irqrestore(&task_group_lock
, flags
);
8197 /* wait for possible concurrent references to cfs_rqs complete */
8198 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8201 /* change task's runqueue when it moves between groups.
8202 * The caller of this function should have put the task in its new group
8203 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8204 * reflect its new group.
8206 void sched_move_task(struct task_struct
*tsk
)
8209 unsigned long flags
;
8212 rq
= task_rq_lock(tsk
, &flags
);
8214 update_rq_clock(rq
);
8216 running
= task_current(rq
, tsk
);
8217 on_rq
= tsk
->se
.on_rq
;
8220 dequeue_task(rq
, tsk
, 0);
8221 if (unlikely(running
))
8222 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8224 set_task_rq(tsk
, task_cpu(tsk
));
8226 #ifdef CONFIG_FAIR_GROUP_SCHED
8227 if (tsk
->sched_class
->moved_group
)
8228 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8231 if (unlikely(running
))
8232 tsk
->sched_class
->set_curr_task(rq
);
8234 enqueue_task(rq
, tsk
, 0, false);
8236 task_rq_unlock(rq
, &flags
);
8238 #endif /* CONFIG_CGROUP_SCHED */
8240 #ifdef CONFIG_FAIR_GROUP_SCHED
8241 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8243 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8248 dequeue_entity(cfs_rq
, se
, 0);
8250 se
->load
.weight
= shares
;
8251 se
->load
.inv_weight
= 0;
8254 enqueue_entity(cfs_rq
, se
, 0);
8257 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8259 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8260 struct rq
*rq
= cfs_rq
->rq
;
8261 unsigned long flags
;
8263 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8264 __set_se_shares(se
, shares
);
8265 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8268 static DEFINE_MUTEX(shares_mutex
);
8270 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8273 unsigned long flags
;
8276 * We can't change the weight of the root cgroup.
8281 if (shares
< MIN_SHARES
)
8282 shares
= MIN_SHARES
;
8283 else if (shares
> MAX_SHARES
)
8284 shares
= MAX_SHARES
;
8286 mutex_lock(&shares_mutex
);
8287 if (tg
->shares
== shares
)
8290 spin_lock_irqsave(&task_group_lock
, flags
);
8291 for_each_possible_cpu(i
)
8292 unregister_fair_sched_group(tg
, i
);
8293 list_del_rcu(&tg
->siblings
);
8294 spin_unlock_irqrestore(&task_group_lock
, flags
);
8296 /* wait for any ongoing reference to this group to finish */
8297 synchronize_sched();
8300 * Now we are free to modify the group's share on each cpu
8301 * w/o tripping rebalance_share or load_balance_fair.
8303 tg
->shares
= shares
;
8304 for_each_possible_cpu(i
) {
8308 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8309 set_se_shares(tg
->se
[i
], shares
);
8313 * Enable load balance activity on this group, by inserting it back on
8314 * each cpu's rq->leaf_cfs_rq_list.
8316 spin_lock_irqsave(&task_group_lock
, flags
);
8317 for_each_possible_cpu(i
)
8318 register_fair_sched_group(tg
, i
);
8319 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8320 spin_unlock_irqrestore(&task_group_lock
, flags
);
8322 mutex_unlock(&shares_mutex
);
8326 unsigned long sched_group_shares(struct task_group
*tg
)
8332 #ifdef CONFIG_RT_GROUP_SCHED
8334 * Ensure that the real time constraints are schedulable.
8336 static DEFINE_MUTEX(rt_constraints_mutex
);
8338 static unsigned long to_ratio(u64 period
, u64 runtime
)
8340 if (runtime
== RUNTIME_INF
)
8343 return div64_u64(runtime
<< 20, period
);
8346 /* Must be called with tasklist_lock held */
8347 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8349 struct task_struct
*g
, *p
;
8351 do_each_thread(g
, p
) {
8352 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8354 } while_each_thread(g
, p
);
8359 struct rt_schedulable_data
{
8360 struct task_group
*tg
;
8365 static int tg_schedulable(struct task_group
*tg
, void *data
)
8367 struct rt_schedulable_data
*d
= data
;
8368 struct task_group
*child
;
8369 unsigned long total
, sum
= 0;
8370 u64 period
, runtime
;
8372 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8373 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8376 period
= d
->rt_period
;
8377 runtime
= d
->rt_runtime
;
8381 * Cannot have more runtime than the period.
8383 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8387 * Ensure we don't starve existing RT tasks.
8389 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8392 total
= to_ratio(period
, runtime
);
8395 * Nobody can have more than the global setting allows.
8397 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8401 * The sum of our children's runtime should not exceed our own.
8403 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8404 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8405 runtime
= child
->rt_bandwidth
.rt_runtime
;
8407 if (child
== d
->tg
) {
8408 period
= d
->rt_period
;
8409 runtime
= d
->rt_runtime
;
8412 sum
+= to_ratio(period
, runtime
);
8421 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8423 struct rt_schedulable_data data
= {
8425 .rt_period
= period
,
8426 .rt_runtime
= runtime
,
8429 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8432 static int tg_set_bandwidth(struct task_group
*tg
,
8433 u64 rt_period
, u64 rt_runtime
)
8437 mutex_lock(&rt_constraints_mutex
);
8438 read_lock(&tasklist_lock
);
8439 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8443 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8444 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8445 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8447 for_each_possible_cpu(i
) {
8448 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8450 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8451 rt_rq
->rt_runtime
= rt_runtime
;
8452 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8454 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8456 read_unlock(&tasklist_lock
);
8457 mutex_unlock(&rt_constraints_mutex
);
8462 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8464 u64 rt_runtime
, rt_period
;
8466 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8467 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8468 if (rt_runtime_us
< 0)
8469 rt_runtime
= RUNTIME_INF
;
8471 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8474 long sched_group_rt_runtime(struct task_group
*tg
)
8478 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8481 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8482 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8483 return rt_runtime_us
;
8486 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8488 u64 rt_runtime
, rt_period
;
8490 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8491 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8496 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8499 long sched_group_rt_period(struct task_group
*tg
)
8503 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8504 do_div(rt_period_us
, NSEC_PER_USEC
);
8505 return rt_period_us
;
8508 static int sched_rt_global_constraints(void)
8510 u64 runtime
, period
;
8513 if (sysctl_sched_rt_period
<= 0)
8516 runtime
= global_rt_runtime();
8517 period
= global_rt_period();
8520 * Sanity check on the sysctl variables.
8522 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8525 mutex_lock(&rt_constraints_mutex
);
8526 read_lock(&tasklist_lock
);
8527 ret
= __rt_schedulable(NULL
, 0, 0);
8528 read_unlock(&tasklist_lock
);
8529 mutex_unlock(&rt_constraints_mutex
);
8534 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8536 /* Don't accept realtime tasks when there is no way for them to run */
8537 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8543 #else /* !CONFIG_RT_GROUP_SCHED */
8544 static int sched_rt_global_constraints(void)
8546 unsigned long flags
;
8549 if (sysctl_sched_rt_period
<= 0)
8553 * There's always some RT tasks in the root group
8554 * -- migration, kstopmachine etc..
8556 if (sysctl_sched_rt_runtime
== 0)
8559 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8560 for_each_possible_cpu(i
) {
8561 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8563 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8564 rt_rq
->rt_runtime
= global_rt_runtime();
8565 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8567 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8571 #endif /* CONFIG_RT_GROUP_SCHED */
8573 int sched_rt_handler(struct ctl_table
*table
, int write
,
8574 void __user
*buffer
, size_t *lenp
,
8578 int old_period
, old_runtime
;
8579 static DEFINE_MUTEX(mutex
);
8582 old_period
= sysctl_sched_rt_period
;
8583 old_runtime
= sysctl_sched_rt_runtime
;
8585 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8587 if (!ret
&& write
) {
8588 ret
= sched_rt_global_constraints();
8590 sysctl_sched_rt_period
= old_period
;
8591 sysctl_sched_rt_runtime
= old_runtime
;
8593 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8594 def_rt_bandwidth
.rt_period
=
8595 ns_to_ktime(global_rt_period());
8598 mutex_unlock(&mutex
);
8603 #ifdef CONFIG_CGROUP_SCHED
8605 /* return corresponding task_group object of a cgroup */
8606 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8608 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8609 struct task_group
, css
);
8612 static struct cgroup_subsys_state
*
8613 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8615 struct task_group
*tg
, *parent
;
8617 if (!cgrp
->parent
) {
8618 /* This is early initialization for the top cgroup */
8619 return &init_task_group
.css
;
8622 parent
= cgroup_tg(cgrp
->parent
);
8623 tg
= sched_create_group(parent
);
8625 return ERR_PTR(-ENOMEM
);
8631 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8633 struct task_group
*tg
= cgroup_tg(cgrp
);
8635 sched_destroy_group(tg
);
8639 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8641 #ifdef CONFIG_RT_GROUP_SCHED
8642 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8645 /* We don't support RT-tasks being in separate groups */
8646 if (tsk
->sched_class
!= &fair_sched_class
)
8653 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8654 struct task_struct
*tsk
, bool threadgroup
)
8656 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8660 struct task_struct
*c
;
8662 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8663 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8675 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8676 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8679 sched_move_task(tsk
);
8681 struct task_struct
*c
;
8683 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8690 #ifdef CONFIG_FAIR_GROUP_SCHED
8691 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8694 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8697 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8699 struct task_group
*tg
= cgroup_tg(cgrp
);
8701 return (u64
) tg
->shares
;
8703 #endif /* CONFIG_FAIR_GROUP_SCHED */
8705 #ifdef CONFIG_RT_GROUP_SCHED
8706 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8709 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8712 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8714 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8717 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8720 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8723 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8725 return sched_group_rt_period(cgroup_tg(cgrp
));
8727 #endif /* CONFIG_RT_GROUP_SCHED */
8729 static struct cftype cpu_files
[] = {
8730 #ifdef CONFIG_FAIR_GROUP_SCHED
8733 .read_u64
= cpu_shares_read_u64
,
8734 .write_u64
= cpu_shares_write_u64
,
8737 #ifdef CONFIG_RT_GROUP_SCHED
8739 .name
= "rt_runtime_us",
8740 .read_s64
= cpu_rt_runtime_read
,
8741 .write_s64
= cpu_rt_runtime_write
,
8744 .name
= "rt_period_us",
8745 .read_u64
= cpu_rt_period_read_uint
,
8746 .write_u64
= cpu_rt_period_write_uint
,
8751 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8753 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8756 struct cgroup_subsys cpu_cgroup_subsys
= {
8758 .create
= cpu_cgroup_create
,
8759 .destroy
= cpu_cgroup_destroy
,
8760 .can_attach
= cpu_cgroup_can_attach
,
8761 .attach
= cpu_cgroup_attach
,
8762 .populate
= cpu_cgroup_populate
,
8763 .subsys_id
= cpu_cgroup_subsys_id
,
8767 #endif /* CONFIG_CGROUP_SCHED */
8769 #ifdef CONFIG_CGROUP_CPUACCT
8772 * CPU accounting code for task groups.
8774 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8775 * (balbir@in.ibm.com).
8778 /* track cpu usage of a group of tasks and its child groups */
8780 struct cgroup_subsys_state css
;
8781 /* cpuusage holds pointer to a u64-type object on every cpu */
8782 u64 __percpu
*cpuusage
;
8783 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8784 struct cpuacct
*parent
;
8787 struct cgroup_subsys cpuacct_subsys
;
8789 /* return cpu accounting group corresponding to this container */
8790 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8792 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8793 struct cpuacct
, css
);
8796 /* return cpu accounting group to which this task belongs */
8797 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8799 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8800 struct cpuacct
, css
);
8803 /* create a new cpu accounting group */
8804 static struct cgroup_subsys_state
*cpuacct_create(
8805 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8807 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8813 ca
->cpuusage
= alloc_percpu(u64
);
8817 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8818 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8819 goto out_free_counters
;
8822 ca
->parent
= cgroup_ca(cgrp
->parent
);
8828 percpu_counter_destroy(&ca
->cpustat
[i
]);
8829 free_percpu(ca
->cpuusage
);
8833 return ERR_PTR(-ENOMEM
);
8836 /* destroy an existing cpu accounting group */
8838 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8840 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8843 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8844 percpu_counter_destroy(&ca
->cpustat
[i
]);
8845 free_percpu(ca
->cpuusage
);
8849 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8851 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8854 #ifndef CONFIG_64BIT
8856 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8858 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8860 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8868 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8870 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8872 #ifndef CONFIG_64BIT
8874 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8876 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8878 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8884 /* return total cpu usage (in nanoseconds) of a group */
8885 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8887 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8888 u64 totalcpuusage
= 0;
8891 for_each_present_cpu(i
)
8892 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8894 return totalcpuusage
;
8897 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8900 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8909 for_each_present_cpu(i
)
8910 cpuacct_cpuusage_write(ca
, i
, 0);
8916 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8919 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8923 for_each_present_cpu(i
) {
8924 percpu
= cpuacct_cpuusage_read(ca
, i
);
8925 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8927 seq_printf(m
, "\n");
8931 static const char *cpuacct_stat_desc
[] = {
8932 [CPUACCT_STAT_USER
] = "user",
8933 [CPUACCT_STAT_SYSTEM
] = "system",
8936 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8937 struct cgroup_map_cb
*cb
)
8939 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8942 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8943 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8944 val
= cputime64_to_clock_t(val
);
8945 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8950 static struct cftype files
[] = {
8953 .read_u64
= cpuusage_read
,
8954 .write_u64
= cpuusage_write
,
8957 .name
= "usage_percpu",
8958 .read_seq_string
= cpuacct_percpu_seq_read
,
8962 .read_map
= cpuacct_stats_show
,
8966 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8968 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8972 * charge this task's execution time to its accounting group.
8974 * called with rq->lock held.
8976 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8981 if (unlikely(!cpuacct_subsys
.active
))
8984 cpu
= task_cpu(tsk
);
8990 for (; ca
; ca
= ca
->parent
) {
8991 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8992 *cpuusage
+= cputime
;
8999 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
9000 * in cputime_t units. As a result, cpuacct_update_stats calls
9001 * percpu_counter_add with values large enough to always overflow the
9002 * per cpu batch limit causing bad SMP scalability.
9004 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
9005 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
9006 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
9009 #define CPUACCT_BATCH \
9010 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
9012 #define CPUACCT_BATCH 0
9016 * Charge the system/user time to the task's accounting group.
9018 static void cpuacct_update_stats(struct task_struct
*tsk
,
9019 enum cpuacct_stat_index idx
, cputime_t val
)
9022 int batch
= CPUACCT_BATCH
;
9024 if (unlikely(!cpuacct_subsys
.active
))
9031 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
9037 struct cgroup_subsys cpuacct_subsys
= {
9039 .create
= cpuacct_create
,
9040 .destroy
= cpuacct_destroy
,
9041 .populate
= cpuacct_populate
,
9042 .subsys_id
= cpuacct_subsys_id
,
9044 #endif /* CONFIG_CGROUP_CPUACCT */
9048 int rcu_expedited_torture_stats(char *page
)
9052 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9054 void synchronize_sched_expedited(void)
9057 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9059 #else /* #ifndef CONFIG_SMP */
9061 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
9062 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
9064 #define RCU_EXPEDITED_STATE_POST -2
9065 #define RCU_EXPEDITED_STATE_IDLE -1
9067 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9069 int rcu_expedited_torture_stats(char *page
)
9074 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
9075 for_each_online_cpu(cpu
) {
9076 cnt
+= sprintf(&page
[cnt
], " %d:%d",
9077 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
9079 cnt
+= sprintf(&page
[cnt
], "\n");
9082 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
9084 static long synchronize_sched_expedited_count
;
9087 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
9088 * approach to force grace period to end quickly. This consumes
9089 * significant time on all CPUs, and is thus not recommended for
9090 * any sort of common-case code.
9092 * Note that it is illegal to call this function while holding any
9093 * lock that is acquired by a CPU-hotplug notifier. Failing to
9094 * observe this restriction will result in deadlock.
9096 void synchronize_sched_expedited(void)
9099 unsigned long flags
;
9100 bool need_full_sync
= 0;
9102 struct migration_req
*req
;
9106 smp_mb(); /* ensure prior mod happens before capturing snap. */
9107 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
9109 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
9111 if (trycount
++ < 10)
9112 udelay(trycount
* num_online_cpus());
9114 synchronize_sched();
9117 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
9118 smp_mb(); /* ensure test happens before caller kfree */
9123 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
9124 for_each_online_cpu(cpu
) {
9126 req
= &per_cpu(rcu_migration_req
, cpu
);
9127 init_completion(&req
->done
);
9129 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
9130 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9131 list_add(&req
->list
, &rq
->migration_queue
);
9132 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9133 wake_up_process(rq
->migration_thread
);
9135 for_each_online_cpu(cpu
) {
9136 rcu_expedited_state
= cpu
;
9137 req
= &per_cpu(rcu_migration_req
, cpu
);
9139 wait_for_completion(&req
->done
);
9140 raw_spin_lock_irqsave(&rq
->lock
, flags
);
9141 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
9143 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
9144 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
9146 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
9147 synchronize_sched_expedited_count
++;
9148 mutex_unlock(&rcu_sched_expedited_mutex
);
9151 synchronize_sched();
9153 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
9155 #endif /* #else #ifndef CONFIG_SMP */