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/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/stop_machine.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 #endif /* CONFIG_CGROUP_SCHED */
311 /* CFS-related fields in a runqueue */
313 struct load_weight load
;
314 unsigned long nr_running
;
319 struct rb_root tasks_timeline
;
320 struct rb_node
*rb_leftmost
;
322 struct list_head tasks
;
323 struct list_head
*balance_iterator
;
326 * 'curr' points to currently running entity on this cfs_rq.
327 * It is set to NULL otherwise (i.e when none are currently running).
329 struct sched_entity
*curr
, *next
, *last
;
331 unsigned int nr_spread_over
;
333 #ifdef CONFIG_FAIR_GROUP_SCHED
334 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
337 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
338 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
339 * (like users, containers etc.)
341 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
342 * list is used during load balance.
344 struct list_head leaf_cfs_rq_list
;
345 struct task_group
*tg
; /* group that "owns" this runqueue */
349 * the part of load.weight contributed by tasks
351 unsigned long task_weight
;
354 * h_load = weight * f(tg)
356 * Where f(tg) is the recursive weight fraction assigned to
359 unsigned long h_load
;
362 * this cpu's part of tg->shares
364 unsigned long shares
;
367 * load.weight at the time we set shares
369 unsigned long rq_weight
;
374 /* Real-Time classes' related field in a runqueue: */
376 struct rt_prio_array active
;
377 unsigned long rt_nr_running
;
378 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
380 int curr
; /* highest queued rt task prio */
382 int next
; /* next highest */
387 unsigned long rt_nr_migratory
;
388 unsigned long rt_nr_total
;
390 struct plist_head pushable_tasks
;
395 /* Nests inside the rq lock: */
396 raw_spinlock_t rt_runtime_lock
;
398 #ifdef CONFIG_RT_GROUP_SCHED
399 unsigned long rt_nr_boosted
;
402 struct list_head leaf_rt_rq_list
;
403 struct task_group
*tg
;
410 * We add the notion of a root-domain which will be used to define per-domain
411 * variables. Each exclusive cpuset essentially defines an island domain by
412 * fully partitioning the member cpus from any other cpuset. Whenever a new
413 * exclusive cpuset is created, we also create and attach a new root-domain
420 cpumask_var_t online
;
423 * The "RT overload" flag: it gets set if a CPU has more than
424 * one runnable RT task.
426 cpumask_var_t rto_mask
;
429 struct cpupri cpupri
;
434 * By default the system creates a single root-domain with all cpus as
435 * members (mimicking the global state we have today).
437 static struct root_domain def_root_domain
;
442 * This is the main, per-CPU runqueue data structure.
444 * Locking rule: those places that want to lock multiple runqueues
445 * (such as the load balancing or the thread migration code), lock
446 * acquire operations must be ordered by ascending &runqueue.
453 * nr_running and cpu_load should be in the same cacheline because
454 * remote CPUs use both these fields when doing load calculation.
456 unsigned long nr_running
;
457 #define CPU_LOAD_IDX_MAX 5
458 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
461 unsigned char in_nohz_recently
;
463 unsigned int skip_clock_update
;
465 /* capture load from *all* tasks on this cpu: */
466 struct load_weight load
;
467 unsigned long nr_load_updates
;
473 #ifdef CONFIG_FAIR_GROUP_SCHED
474 /* list of leaf cfs_rq on this cpu: */
475 struct list_head leaf_cfs_rq_list
;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 struct list_head leaf_rt_rq_list
;
482 * This is part of a global counter where only the total sum
483 * over all CPUs matters. A task can increase this counter on
484 * one CPU and if it got migrated afterwards it may decrease
485 * it on another CPU. Always updated under the runqueue lock:
487 unsigned long nr_uninterruptible
;
489 struct task_struct
*curr
, *idle
;
490 unsigned long next_balance
;
491 struct mm_struct
*prev_mm
;
498 struct root_domain
*rd
;
499 struct sched_domain
*sd
;
501 unsigned long cpu_power
;
503 unsigned char idle_at_tick
;
504 /* For active balancing */
508 struct cpu_stop_work active_balance_work
;
509 /* cpu of this runqueue: */
513 unsigned long avg_load_per_task
;
521 /* calc_load related fields */
522 unsigned long calc_load_update
;
523 long calc_load_active
;
525 #ifdef CONFIG_SCHED_HRTICK
527 int hrtick_csd_pending
;
528 struct call_single_data hrtick_csd
;
530 struct hrtimer hrtick_timer
;
533 #ifdef CONFIG_SCHEDSTATS
535 struct sched_info rq_sched_info
;
536 unsigned long long rq_cpu_time
;
537 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
539 /* sys_sched_yield() stats */
540 unsigned int yld_count
;
542 /* schedule() stats */
543 unsigned int sched_switch
;
544 unsigned int sched_count
;
545 unsigned int sched_goidle
;
547 /* try_to_wake_up() stats */
548 unsigned int ttwu_count
;
549 unsigned int ttwu_local
;
552 unsigned int bkl_count
;
556 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
559 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
561 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
564 * A queue event has occurred, and we're going to schedule. In
565 * this case, we can save a useless back to back clock update.
567 if (test_tsk_need_resched(p
))
568 rq
->skip_clock_update
= 1;
571 static inline int cpu_of(struct rq
*rq
)
580 #define rcu_dereference_check_sched_domain(p) \
581 rcu_dereference_check((p), \
582 rcu_read_lock_sched_held() || \
583 lockdep_is_held(&sched_domains_mutex))
586 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
587 * See detach_destroy_domains: synchronize_sched for details.
589 * The domain tree of any CPU may only be accessed from within
590 * preempt-disabled sections.
592 #define for_each_domain(cpu, __sd) \
593 for (__sd = rcu_dereference_check_sched_domain(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
595 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
596 #define this_rq() (&__get_cpu_var(runqueues))
597 #define task_rq(p) cpu_rq(task_cpu(p))
598 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
599 #define raw_rq() (&__raw_get_cpu_var(runqueues))
601 #ifdef CONFIG_CGROUP_SCHED
604 * Return the group to which this tasks belongs.
606 * We use task_subsys_state_check() and extend the RCU verification
607 * with lockdep_is_held(&task_rq(p)->lock) because cpu_cgroup_attach()
608 * holds that lock for each task it moves into the cgroup. Therefore
609 * by holding that lock, we pin the task to the current cgroup.
611 static inline struct task_group
*task_group(struct task_struct
*p
)
613 struct cgroup_subsys_state
*css
;
615 css
= task_subsys_state_check(p
, cpu_cgroup_subsys_id
,
616 lockdep_is_held(&task_rq(p
)->lock
));
617 return container_of(css
, struct task_group
, css
);
620 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
621 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
623 #ifdef CONFIG_FAIR_GROUP_SCHED
624 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
625 p
->se
.parent
= task_group(p
)->se
[cpu
];
628 #ifdef CONFIG_RT_GROUP_SCHED
629 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
630 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
634 #else /* CONFIG_CGROUP_SCHED */
636 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
637 static inline struct task_group
*task_group(struct task_struct
*p
)
642 #endif /* CONFIG_CGROUP_SCHED */
644 inline void update_rq_clock(struct rq
*rq
)
646 if (!rq
->skip_clock_update
)
647 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
651 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
653 #ifdef CONFIG_SCHED_DEBUG
654 # define const_debug __read_mostly
656 # define const_debug static const
661 * @cpu: the processor in question.
663 * Returns true if the current cpu runqueue is locked.
664 * This interface allows printk to be called with the runqueue lock
665 * held and know whether or not it is OK to wake up the klogd.
667 int runqueue_is_locked(int cpu
)
669 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
673 * Debugging: various feature bits
676 #define SCHED_FEAT(name, enabled) \
677 __SCHED_FEAT_##name ,
680 #include "sched_features.h"
685 #define SCHED_FEAT(name, enabled) \
686 (1UL << __SCHED_FEAT_##name) * enabled |
688 const_debug
unsigned int sysctl_sched_features
=
689 #include "sched_features.h"
694 #ifdef CONFIG_SCHED_DEBUG
695 #define SCHED_FEAT(name, enabled) \
698 static __read_mostly
char *sched_feat_names
[] = {
699 #include "sched_features.h"
705 static int sched_feat_show(struct seq_file
*m
, void *v
)
709 for (i
= 0; sched_feat_names
[i
]; i
++) {
710 if (!(sysctl_sched_features
& (1UL << i
)))
712 seq_printf(m
, "%s ", sched_feat_names
[i
]);
720 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
721 size_t cnt
, loff_t
*ppos
)
731 if (copy_from_user(&buf
, ubuf
, cnt
))
736 if (strncmp(buf
, "NO_", 3) == 0) {
741 for (i
= 0; sched_feat_names
[i
]; i
++) {
742 int len
= strlen(sched_feat_names
[i
]);
744 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
746 sysctl_sched_features
&= ~(1UL << i
);
748 sysctl_sched_features
|= (1UL << i
);
753 if (!sched_feat_names
[i
])
761 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
763 return single_open(filp
, sched_feat_show
, NULL
);
766 static const struct file_operations sched_feat_fops
= {
767 .open
= sched_feat_open
,
768 .write
= sched_feat_write
,
771 .release
= single_release
,
774 static __init
int sched_init_debug(void)
776 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
781 late_initcall(sched_init_debug
);
785 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
788 * Number of tasks to iterate in a single balance run.
789 * Limited because this is done with IRQs disabled.
791 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
794 * ratelimit for updating the group shares.
797 unsigned int sysctl_sched_shares_ratelimit
= 250000;
798 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
801 * Inject some fuzzyness into changing the per-cpu group shares
802 * this avoids remote rq-locks at the expense of fairness.
805 unsigned int sysctl_sched_shares_thresh
= 4;
808 * period over which we average the RT time consumption, measured
813 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
816 * period over which we measure -rt task cpu usage in us.
819 unsigned int sysctl_sched_rt_period
= 1000000;
821 static __read_mostly
int scheduler_running
;
824 * part of the period that we allow rt tasks to run in us.
827 int sysctl_sched_rt_runtime
= 950000;
829 static inline u64
global_rt_period(void)
831 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
834 static inline u64
global_rt_runtime(void)
836 if (sysctl_sched_rt_runtime
< 0)
839 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
842 #ifndef prepare_arch_switch
843 # define prepare_arch_switch(next) do { } while (0)
845 #ifndef finish_arch_switch
846 # define finish_arch_switch(prev) do { } while (0)
849 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
851 return rq
->curr
== p
;
854 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
855 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
857 return task_current(rq
, p
);
860 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
864 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
866 #ifdef CONFIG_DEBUG_SPINLOCK
867 /* this is a valid case when another task releases the spinlock */
868 rq
->lock
.owner
= current
;
871 * If we are tracking spinlock dependencies then we have to
872 * fix up the runqueue lock - which gets 'carried over' from
875 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
877 raw_spin_unlock_irq(&rq
->lock
);
880 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
881 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
886 return task_current(rq
, p
);
890 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
894 * We can optimise this out completely for !SMP, because the
895 * SMP rebalancing from interrupt is the only thing that cares
900 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
901 raw_spin_unlock_irq(&rq
->lock
);
903 raw_spin_unlock(&rq
->lock
);
907 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
911 * After ->oncpu is cleared, the task can be moved to a different CPU.
912 * We must ensure this doesn't happen until the switch is completely
918 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
925 * Check whether the task is waking, we use this to synchronize ->cpus_allowed
928 static inline int task_is_waking(struct task_struct
*p
)
930 return unlikely(p
->state
== TASK_WAKING
);
934 * __task_rq_lock - lock the runqueue a given task resides on.
935 * Must be called interrupts disabled.
937 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
944 raw_spin_lock(&rq
->lock
);
945 if (likely(rq
== task_rq(p
)))
947 raw_spin_unlock(&rq
->lock
);
952 * task_rq_lock - lock the runqueue a given task resides on and disable
953 * interrupts. Note the ordering: we can safely lookup the task_rq without
954 * explicitly disabling preemption.
956 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
962 local_irq_save(*flags
);
964 raw_spin_lock(&rq
->lock
);
965 if (likely(rq
== task_rq(p
)))
967 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
971 static void __task_rq_unlock(struct rq
*rq
)
974 raw_spin_unlock(&rq
->lock
);
977 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
980 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
984 * this_rq_lock - lock this runqueue and disable interrupts.
986 static struct rq
*this_rq_lock(void)
993 raw_spin_lock(&rq
->lock
);
998 #ifdef CONFIG_SCHED_HRTICK
1000 * Use HR-timers to deliver accurate preemption points.
1002 * Its all a bit involved since we cannot program an hrt while holding the
1003 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1006 * When we get rescheduled we reprogram the hrtick_timer outside of the
1012 * - enabled by features
1013 * - hrtimer is actually high res
1015 static inline int hrtick_enabled(struct rq
*rq
)
1017 if (!sched_feat(HRTICK
))
1019 if (!cpu_active(cpu_of(rq
)))
1021 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1024 static void hrtick_clear(struct rq
*rq
)
1026 if (hrtimer_active(&rq
->hrtick_timer
))
1027 hrtimer_cancel(&rq
->hrtick_timer
);
1031 * High-resolution timer tick.
1032 * Runs from hardirq context with interrupts disabled.
1034 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1036 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1038 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1040 raw_spin_lock(&rq
->lock
);
1041 update_rq_clock(rq
);
1042 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1043 raw_spin_unlock(&rq
->lock
);
1045 return HRTIMER_NORESTART
;
1050 * called from hardirq (IPI) context
1052 static void __hrtick_start(void *arg
)
1054 struct rq
*rq
= arg
;
1056 raw_spin_lock(&rq
->lock
);
1057 hrtimer_restart(&rq
->hrtick_timer
);
1058 rq
->hrtick_csd_pending
= 0;
1059 raw_spin_unlock(&rq
->lock
);
1063 * Called to set the hrtick timer state.
1065 * called with rq->lock held and irqs disabled
1067 static void hrtick_start(struct rq
*rq
, u64 delay
)
1069 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1070 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1072 hrtimer_set_expires(timer
, time
);
1074 if (rq
== this_rq()) {
1075 hrtimer_restart(timer
);
1076 } else if (!rq
->hrtick_csd_pending
) {
1077 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1078 rq
->hrtick_csd_pending
= 1;
1083 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1085 int cpu
= (int)(long)hcpu
;
1088 case CPU_UP_CANCELED
:
1089 case CPU_UP_CANCELED_FROZEN
:
1090 case CPU_DOWN_PREPARE
:
1091 case CPU_DOWN_PREPARE_FROZEN
:
1093 case CPU_DEAD_FROZEN
:
1094 hrtick_clear(cpu_rq(cpu
));
1101 static __init
void init_hrtick(void)
1103 hotcpu_notifier(hotplug_hrtick
, 0);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq
*rq
, u64 delay
)
1113 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1114 HRTIMER_MODE_REL_PINNED
, 0);
1117 static inline void init_hrtick(void)
1120 #endif /* CONFIG_SMP */
1122 static void init_rq_hrtick(struct rq
*rq
)
1125 rq
->hrtick_csd_pending
= 0;
1127 rq
->hrtick_csd
.flags
= 0;
1128 rq
->hrtick_csd
.func
= __hrtick_start
;
1129 rq
->hrtick_csd
.info
= rq
;
1132 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1133 rq
->hrtick_timer
.function
= hrtick
;
1135 #else /* CONFIG_SCHED_HRTICK */
1136 static inline void hrtick_clear(struct rq
*rq
)
1140 static inline void init_rq_hrtick(struct rq
*rq
)
1144 static inline void init_hrtick(void)
1147 #endif /* CONFIG_SCHED_HRTICK */
1150 * resched_task - mark a task 'to be rescheduled now'.
1152 * On UP this means the setting of the need_resched flag, on SMP it
1153 * might also involve a cross-CPU call to trigger the scheduler on
1158 #ifndef tsk_is_polling
1159 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1162 static void resched_task(struct task_struct
*p
)
1166 assert_raw_spin_locked(&task_rq(p
)->lock
);
1168 if (test_tsk_need_resched(p
))
1171 set_tsk_need_resched(p
);
1174 if (cpu
== smp_processor_id())
1177 /* NEED_RESCHED must be visible before we test polling */
1179 if (!tsk_is_polling(p
))
1180 smp_send_reschedule(cpu
);
1183 static void resched_cpu(int cpu
)
1185 struct rq
*rq
= cpu_rq(cpu
);
1186 unsigned long flags
;
1188 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1190 resched_task(cpu_curr(cpu
));
1191 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1196 * When add_timer_on() enqueues a timer into the timer wheel of an
1197 * idle CPU then this timer might expire before the next timer event
1198 * which is scheduled to wake up that CPU. In case of a completely
1199 * idle system the next event might even be infinite time into the
1200 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1201 * leaves the inner idle loop so the newly added timer is taken into
1202 * account when the CPU goes back to idle and evaluates the timer
1203 * wheel for the next timer event.
1205 void wake_up_idle_cpu(int cpu
)
1207 struct rq
*rq
= cpu_rq(cpu
);
1209 if (cpu
== smp_processor_id())
1213 * This is safe, as this function is called with the timer
1214 * wheel base lock of (cpu) held. When the CPU is on the way
1215 * to idle and has not yet set rq->curr to idle then it will
1216 * be serialized on the timer wheel base lock and take the new
1217 * timer into account automatically.
1219 if (rq
->curr
!= rq
->idle
)
1223 * We can set TIF_RESCHED on the idle task of the other CPU
1224 * lockless. The worst case is that the other CPU runs the
1225 * idle task through an additional NOOP schedule()
1227 set_tsk_need_resched(rq
->idle
);
1229 /* NEED_RESCHED must be visible before we test polling */
1231 if (!tsk_is_polling(rq
->idle
))
1232 smp_send_reschedule(cpu
);
1235 #endif /* CONFIG_NO_HZ */
1237 static u64
sched_avg_period(void)
1239 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1242 static void sched_avg_update(struct rq
*rq
)
1244 s64 period
= sched_avg_period();
1246 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1248 * Inline assembly required to prevent the compiler
1249 * optimising this loop into a divmod call.
1250 * See __iter_div_u64_rem() for another example of this.
1252 asm("" : "+rm" (rq
->age_stamp
));
1253 rq
->age_stamp
+= period
;
1258 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1260 rq
->rt_avg
+= rt_delta
;
1261 sched_avg_update(rq
);
1264 #else /* !CONFIG_SMP */
1265 static void resched_task(struct task_struct
*p
)
1267 assert_raw_spin_locked(&task_rq(p
)->lock
);
1268 set_tsk_need_resched(p
);
1271 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1274 #endif /* CONFIG_SMP */
1276 #if BITS_PER_LONG == 32
1277 # define WMULT_CONST (~0UL)
1279 # define WMULT_CONST (1UL << 32)
1282 #define WMULT_SHIFT 32
1285 * Shift right and round:
1287 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1290 * delta *= weight / lw
1292 static unsigned long
1293 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1294 struct load_weight
*lw
)
1298 if (!lw
->inv_weight
) {
1299 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1302 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1306 tmp
= (u64
)delta_exec
* weight
;
1308 * Check whether we'd overflow the 64-bit multiplication:
1310 if (unlikely(tmp
> WMULT_CONST
))
1311 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1314 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1316 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1319 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1325 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1332 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1333 * of tasks with abnormal "nice" values across CPUs the contribution that
1334 * each task makes to its run queue's load is weighted according to its
1335 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1336 * scaled version of the new time slice allocation that they receive on time
1340 #define WEIGHT_IDLEPRIO 3
1341 #define WMULT_IDLEPRIO 1431655765
1344 * Nice levels are multiplicative, with a gentle 10% change for every
1345 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1346 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1347 * that remained on nice 0.
1349 * The "10% effect" is relative and cumulative: from _any_ nice level,
1350 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1351 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1352 * If a task goes up by ~10% and another task goes down by ~10% then
1353 * the relative distance between them is ~25%.)
1355 static const int prio_to_weight
[40] = {
1356 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1357 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1358 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1359 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1360 /* 0 */ 1024, 820, 655, 526, 423,
1361 /* 5 */ 335, 272, 215, 172, 137,
1362 /* 10 */ 110, 87, 70, 56, 45,
1363 /* 15 */ 36, 29, 23, 18, 15,
1367 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1369 * In cases where the weight does not change often, we can use the
1370 * precalculated inverse to speed up arithmetics by turning divisions
1371 * into multiplications:
1373 static const u32 prio_to_wmult
[40] = {
1374 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1375 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1376 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1377 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1378 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1379 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1380 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1381 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1384 /* Time spent by the tasks of the cpu accounting group executing in ... */
1385 enum cpuacct_stat_index
{
1386 CPUACCT_STAT_USER
, /* ... user mode */
1387 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1389 CPUACCT_STAT_NSTATS
,
1392 #ifdef CONFIG_CGROUP_CPUACCT
1393 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1394 static void cpuacct_update_stats(struct task_struct
*tsk
,
1395 enum cpuacct_stat_index idx
, cputime_t val
);
1397 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1398 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1399 enum cpuacct_stat_index idx
, cputime_t val
) {}
1402 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1404 update_load_add(&rq
->load
, load
);
1407 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1409 update_load_sub(&rq
->load
, load
);
1412 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1413 typedef int (*tg_visitor
)(struct task_group
*, void *);
1416 * Iterate the full tree, calling @down when first entering a node and @up when
1417 * leaving it for the final time.
1419 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1421 struct task_group
*parent
, *child
;
1425 parent
= &root_task_group
;
1427 ret
= (*down
)(parent
, data
);
1430 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1437 ret
= (*up
)(parent
, data
);
1442 parent
= parent
->parent
;
1451 static int tg_nop(struct task_group
*tg
, void *data
)
1458 /* Used instead of source_load when we know the type == 0 */
1459 static unsigned long weighted_cpuload(const int cpu
)
1461 return cpu_rq(cpu
)->load
.weight
;
1465 * Return a low guess at the load of a migration-source cpu weighted
1466 * according to the scheduling class and "nice" value.
1468 * We want to under-estimate the load of migration sources, to
1469 * balance conservatively.
1471 static unsigned long source_load(int cpu
, int type
)
1473 struct rq
*rq
= cpu_rq(cpu
);
1474 unsigned long total
= weighted_cpuload(cpu
);
1476 if (type
== 0 || !sched_feat(LB_BIAS
))
1479 return min(rq
->cpu_load
[type
-1], total
);
1483 * Return a high guess at the load of a migration-target cpu weighted
1484 * according to the scheduling class and "nice" value.
1486 static unsigned long target_load(int cpu
, int type
)
1488 struct rq
*rq
= cpu_rq(cpu
);
1489 unsigned long total
= weighted_cpuload(cpu
);
1491 if (type
== 0 || !sched_feat(LB_BIAS
))
1494 return max(rq
->cpu_load
[type
-1], total
);
1497 static unsigned long power_of(int cpu
)
1499 return cpu_rq(cpu
)->cpu_power
;
1502 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1504 static unsigned long cpu_avg_load_per_task(int cpu
)
1506 struct rq
*rq
= cpu_rq(cpu
);
1507 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1510 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1512 rq
->avg_load_per_task
= 0;
1514 return rq
->avg_load_per_task
;
1517 #ifdef CONFIG_FAIR_GROUP_SCHED
1519 static __read_mostly
unsigned long __percpu
*update_shares_data
;
1521 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1524 * Calculate and set the cpu's group shares.
1526 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1527 unsigned long sd_shares
,
1528 unsigned long sd_rq_weight
,
1529 unsigned long *usd_rq_weight
)
1531 unsigned long shares
, rq_weight
;
1534 rq_weight
= usd_rq_weight
[cpu
];
1537 rq_weight
= NICE_0_LOAD
;
1541 * \Sum_j shares_j * rq_weight_i
1542 * shares_i = -----------------------------
1543 * \Sum_j rq_weight_j
1545 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1546 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1548 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1549 sysctl_sched_shares_thresh
) {
1550 struct rq
*rq
= cpu_rq(cpu
);
1551 unsigned long flags
;
1553 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1554 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1555 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1556 __set_se_shares(tg
->se
[cpu
], shares
);
1557 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1562 * Re-compute the task group their per cpu shares over the given domain.
1563 * This needs to be done in a bottom-up fashion because the rq weight of a
1564 * parent group depends on the shares of its child groups.
1566 static int tg_shares_up(struct task_group
*tg
, void *data
)
1568 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1569 unsigned long *usd_rq_weight
;
1570 struct sched_domain
*sd
= data
;
1571 unsigned long flags
;
1577 local_irq_save(flags
);
1578 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1580 for_each_cpu(i
, sched_domain_span(sd
)) {
1581 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1582 usd_rq_weight
[i
] = weight
;
1584 rq_weight
+= weight
;
1586 * If there are currently no tasks on the cpu pretend there
1587 * is one of average load so that when a new task gets to
1588 * run here it will not get delayed by group starvation.
1591 weight
= NICE_0_LOAD
;
1593 sum_weight
+= weight
;
1594 shares
+= tg
->cfs_rq
[i
]->shares
;
1598 rq_weight
= sum_weight
;
1600 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1601 shares
= tg
->shares
;
1603 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1604 shares
= tg
->shares
;
1606 for_each_cpu(i
, sched_domain_span(sd
))
1607 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1609 local_irq_restore(flags
);
1615 * Compute the cpu's hierarchical load factor for each task group.
1616 * This needs to be done in a top-down fashion because the load of a child
1617 * group is a fraction of its parents load.
1619 static int tg_load_down(struct task_group
*tg
, void *data
)
1622 long cpu
= (long)data
;
1625 load
= cpu_rq(cpu
)->load
.weight
;
1627 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1628 load
*= tg
->cfs_rq
[cpu
]->shares
;
1629 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1632 tg
->cfs_rq
[cpu
]->h_load
= load
;
1637 static void update_shares(struct sched_domain
*sd
)
1642 if (root_task_group_empty())
1645 now
= cpu_clock(raw_smp_processor_id());
1646 elapsed
= now
- sd
->last_update
;
1648 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1649 sd
->last_update
= now
;
1650 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1654 static void update_h_load(long cpu
)
1656 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1661 static inline void update_shares(struct sched_domain
*sd
)
1667 #ifdef CONFIG_PREEMPT
1669 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1672 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1673 * way at the expense of forcing extra atomic operations in all
1674 * invocations. This assures that the double_lock is acquired using the
1675 * same underlying policy as the spinlock_t on this architecture, which
1676 * reduces latency compared to the unfair variant below. However, it
1677 * also adds more overhead and therefore may reduce throughput.
1679 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1680 __releases(this_rq
->lock
)
1681 __acquires(busiest
->lock
)
1682 __acquires(this_rq
->lock
)
1684 raw_spin_unlock(&this_rq
->lock
);
1685 double_rq_lock(this_rq
, busiest
);
1692 * Unfair double_lock_balance: Optimizes throughput at the expense of
1693 * latency by eliminating extra atomic operations when the locks are
1694 * already in proper order on entry. This favors lower cpu-ids and will
1695 * grant the double lock to lower cpus over higher ids under contention,
1696 * regardless of entry order into the function.
1698 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1699 __releases(this_rq
->lock
)
1700 __acquires(busiest
->lock
)
1701 __acquires(this_rq
->lock
)
1705 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1706 if (busiest
< this_rq
) {
1707 raw_spin_unlock(&this_rq
->lock
);
1708 raw_spin_lock(&busiest
->lock
);
1709 raw_spin_lock_nested(&this_rq
->lock
,
1710 SINGLE_DEPTH_NESTING
);
1713 raw_spin_lock_nested(&busiest
->lock
,
1714 SINGLE_DEPTH_NESTING
);
1719 #endif /* CONFIG_PREEMPT */
1722 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1724 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1726 if (unlikely(!irqs_disabled())) {
1727 /* printk() doesn't work good under rq->lock */
1728 raw_spin_unlock(&this_rq
->lock
);
1732 return _double_lock_balance(this_rq
, busiest
);
1735 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1736 __releases(busiest
->lock
)
1738 raw_spin_unlock(&busiest
->lock
);
1739 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1743 * double_rq_lock - safely lock two runqueues
1745 * Note this does not disable interrupts like task_rq_lock,
1746 * you need to do so manually before calling.
1748 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1749 __acquires(rq1
->lock
)
1750 __acquires(rq2
->lock
)
1752 BUG_ON(!irqs_disabled());
1754 raw_spin_lock(&rq1
->lock
);
1755 __acquire(rq2
->lock
); /* Fake it out ;) */
1758 raw_spin_lock(&rq1
->lock
);
1759 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
1761 raw_spin_lock(&rq2
->lock
);
1762 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
1768 * double_rq_unlock - safely unlock two runqueues
1770 * Note this does not restore interrupts like task_rq_unlock,
1771 * you need to do so manually after calling.
1773 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1774 __releases(rq1
->lock
)
1775 __releases(rq2
->lock
)
1777 raw_spin_unlock(&rq1
->lock
);
1779 raw_spin_unlock(&rq2
->lock
);
1781 __release(rq2
->lock
);
1786 #ifdef CONFIG_FAIR_GROUP_SCHED
1787 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1790 cfs_rq
->shares
= shares
;
1795 static void calc_load_account_idle(struct rq
*this_rq
);
1796 static void update_sysctl(void);
1797 static int get_update_sysctl_factor(void);
1799 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1801 set_task_rq(p
, cpu
);
1804 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1805 * successfuly executed on another CPU. We must ensure that updates of
1806 * per-task data have been completed by this moment.
1809 task_thread_info(p
)->cpu
= cpu
;
1813 static const struct sched_class rt_sched_class
;
1815 #define sched_class_highest (&rt_sched_class)
1816 #define for_each_class(class) \
1817 for (class = sched_class_highest; class; class = class->next)
1819 #include "sched_stats.h"
1821 static void inc_nr_running(struct rq
*rq
)
1826 static void dec_nr_running(struct rq
*rq
)
1831 static void set_load_weight(struct task_struct
*p
)
1833 if (task_has_rt_policy(p
)) {
1834 p
->se
.load
.weight
= 0;
1835 p
->se
.load
.inv_weight
= WMULT_CONST
;
1840 * SCHED_IDLE tasks get minimal weight:
1842 if (p
->policy
== SCHED_IDLE
) {
1843 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1844 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1848 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1849 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1852 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1854 update_rq_clock(rq
);
1855 sched_info_queued(p
);
1856 p
->sched_class
->enqueue_task(rq
, p
, flags
);
1860 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1862 update_rq_clock(rq
);
1863 sched_info_dequeued(p
);
1864 p
->sched_class
->dequeue_task(rq
, p
, flags
);
1869 * activate_task - move a task to the runqueue.
1871 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1873 if (task_contributes_to_load(p
))
1874 rq
->nr_uninterruptible
--;
1876 enqueue_task(rq
, p
, flags
);
1881 * deactivate_task - remove a task from the runqueue.
1883 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int flags
)
1885 if (task_contributes_to_load(p
))
1886 rq
->nr_uninterruptible
++;
1888 dequeue_task(rq
, p
, flags
);
1892 #include "sched_idletask.c"
1893 #include "sched_fair.c"
1894 #include "sched_rt.c"
1895 #ifdef CONFIG_SCHED_DEBUG
1896 # include "sched_debug.c"
1900 * __normal_prio - return the priority that is based on the static prio
1902 static inline int __normal_prio(struct task_struct
*p
)
1904 return p
->static_prio
;
1908 * Calculate the expected normal priority: i.e. priority
1909 * without taking RT-inheritance into account. Might be
1910 * boosted by interactivity modifiers. Changes upon fork,
1911 * setprio syscalls, and whenever the interactivity
1912 * estimator recalculates.
1914 static inline int normal_prio(struct task_struct
*p
)
1918 if (task_has_rt_policy(p
))
1919 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1921 prio
= __normal_prio(p
);
1926 * Calculate the current priority, i.e. the priority
1927 * taken into account by the scheduler. This value might
1928 * be boosted by RT tasks, or might be boosted by
1929 * interactivity modifiers. Will be RT if the task got
1930 * RT-boosted. If not then it returns p->normal_prio.
1932 static int effective_prio(struct task_struct
*p
)
1934 p
->normal_prio
= normal_prio(p
);
1936 * If we are RT tasks or we were boosted to RT priority,
1937 * keep the priority unchanged. Otherwise, update priority
1938 * to the normal priority:
1940 if (!rt_prio(p
->prio
))
1941 return p
->normal_prio
;
1946 * task_curr - is this task currently executing on a CPU?
1947 * @p: the task in question.
1949 inline int task_curr(const struct task_struct
*p
)
1951 return cpu_curr(task_cpu(p
)) == p
;
1954 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1955 const struct sched_class
*prev_class
,
1956 int oldprio
, int running
)
1958 if (prev_class
!= p
->sched_class
) {
1959 if (prev_class
->switched_from
)
1960 prev_class
->switched_from(rq
, p
, running
);
1961 p
->sched_class
->switched_to(rq
, p
, running
);
1963 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1968 * Is this task likely cache-hot:
1971 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1975 if (p
->sched_class
!= &fair_sched_class
)
1979 * Buddy candidates are cache hot:
1981 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
1982 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1983 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1986 if (sysctl_sched_migration_cost
== -1)
1988 if (sysctl_sched_migration_cost
== 0)
1991 delta
= now
- p
->se
.exec_start
;
1993 return delta
< (s64
)sysctl_sched_migration_cost
;
1996 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1998 #ifdef CONFIG_SCHED_DEBUG
2000 * We should never call set_task_cpu() on a blocked task,
2001 * ttwu() will sort out the placement.
2003 WARN_ON_ONCE(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
&&
2004 !(task_thread_info(p
)->preempt_count
& PREEMPT_ACTIVE
));
2007 trace_sched_migrate_task(p
, new_cpu
);
2009 if (task_cpu(p
) != new_cpu
) {
2010 p
->se
.nr_migrations
++;
2011 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
, 1, 1, NULL
, 0);
2014 __set_task_cpu(p
, new_cpu
);
2017 struct migration_arg
{
2018 struct task_struct
*task
;
2022 static int migration_cpu_stop(void *data
);
2025 * The task's runqueue lock must be held.
2026 * Returns true if you have to wait for migration thread.
2028 static bool migrate_task(struct task_struct
*p
, int dest_cpu
)
2030 struct rq
*rq
= task_rq(p
);
2033 * If the task is not on a runqueue (and not running), then
2034 * the next wake-up will properly place the task.
2036 return p
->se
.on_rq
|| task_running(rq
, p
);
2040 * wait_task_inactive - wait for a thread to unschedule.
2042 * If @match_state is nonzero, it's the @p->state value just checked and
2043 * not expected to change. If it changes, i.e. @p might have woken up,
2044 * then return zero. When we succeed in waiting for @p to be off its CPU,
2045 * we return a positive number (its total switch count). If a second call
2046 * a short while later returns the same number, the caller can be sure that
2047 * @p has remained unscheduled the whole time.
2049 * The caller must ensure that the task *will* unschedule sometime soon,
2050 * else this function might spin for a *long* time. This function can't
2051 * be called with interrupts off, or it may introduce deadlock with
2052 * smp_call_function() if an IPI is sent by the same process we are
2053 * waiting to become inactive.
2055 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2057 unsigned long flags
;
2064 * We do the initial early heuristics without holding
2065 * any task-queue locks at all. We'll only try to get
2066 * the runqueue lock when things look like they will
2072 * If the task is actively running on another CPU
2073 * still, just relax and busy-wait without holding
2076 * NOTE! Since we don't hold any locks, it's not
2077 * even sure that "rq" stays as the right runqueue!
2078 * But we don't care, since "task_running()" will
2079 * return false if the runqueue has changed and p
2080 * is actually now running somewhere else!
2082 while (task_running(rq
, p
)) {
2083 if (match_state
&& unlikely(p
->state
!= match_state
))
2089 * Ok, time to look more closely! We need the rq
2090 * lock now, to be *sure*. If we're wrong, we'll
2091 * just go back and repeat.
2093 rq
= task_rq_lock(p
, &flags
);
2094 trace_sched_wait_task(p
);
2095 running
= task_running(rq
, p
);
2096 on_rq
= p
->se
.on_rq
;
2098 if (!match_state
|| p
->state
== match_state
)
2099 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2100 task_rq_unlock(rq
, &flags
);
2103 * If it changed from the expected state, bail out now.
2105 if (unlikely(!ncsw
))
2109 * Was it really running after all now that we
2110 * checked with the proper locks actually held?
2112 * Oops. Go back and try again..
2114 if (unlikely(running
)) {
2120 * It's not enough that it's not actively running,
2121 * it must be off the runqueue _entirely_, and not
2124 * So if it was still runnable (but just not actively
2125 * running right now), it's preempted, and we should
2126 * yield - it could be a while.
2128 if (unlikely(on_rq
)) {
2129 schedule_timeout_uninterruptible(1);
2134 * Ahh, all good. It wasn't running, and it wasn't
2135 * runnable, which means that it will never become
2136 * running in the future either. We're all done!
2145 * kick_process - kick a running thread to enter/exit the kernel
2146 * @p: the to-be-kicked thread
2148 * Cause a process which is running on another CPU to enter
2149 * kernel-mode, without any delay. (to get signals handled.)
2151 * NOTE: this function doesnt have to take the runqueue lock,
2152 * because all it wants to ensure is that the remote task enters
2153 * the kernel. If the IPI races and the task has been migrated
2154 * to another CPU then no harm is done and the purpose has been
2157 void kick_process(struct task_struct
*p
)
2163 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2164 smp_send_reschedule(cpu
);
2167 EXPORT_SYMBOL_GPL(kick_process
);
2168 #endif /* CONFIG_SMP */
2171 * task_oncpu_function_call - call a function on the cpu on which a task runs
2172 * @p: the task to evaluate
2173 * @func: the function to be called
2174 * @info: the function call argument
2176 * Calls the function @func when the task is currently running. This might
2177 * be on the current CPU, which just calls the function directly
2179 void task_oncpu_function_call(struct task_struct
*p
,
2180 void (*func
) (void *info
), void *info
)
2187 smp_call_function_single(cpu
, func
, info
, 1);
2193 * ->cpus_allowed is protected by either TASK_WAKING or rq->lock held.
2195 static int select_fallback_rq(int cpu
, struct task_struct
*p
)
2198 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(cpu
));
2200 /* Look for allowed, online CPU in same node. */
2201 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
2202 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
2205 /* Any allowed, online CPU? */
2206 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2207 if (dest_cpu
< nr_cpu_ids
)
2210 /* No more Mr. Nice Guy. */
2211 if (unlikely(dest_cpu
>= nr_cpu_ids
)) {
2212 dest_cpu
= cpuset_cpus_allowed_fallback(p
);
2214 * Don't tell them about moving exiting tasks or
2215 * kernel threads (both mm NULL), since they never
2218 if (p
->mm
&& printk_ratelimit()) {
2219 printk(KERN_INFO
"process %d (%s) no "
2220 "longer affine to cpu%d\n",
2221 task_pid_nr(p
), p
->comm
, cpu
);
2229 * The caller (fork, wakeup) owns TASK_WAKING, ->cpus_allowed is stable.
2232 int select_task_rq(struct rq
*rq
, struct task_struct
*p
, int sd_flags
, int wake_flags
)
2234 int cpu
= p
->sched_class
->select_task_rq(rq
, p
, sd_flags
, wake_flags
);
2237 * In order not to call set_task_cpu() on a blocking task we need
2238 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2241 * Since this is common to all placement strategies, this lives here.
2243 * [ this allows ->select_task() to simply return task_cpu(p) and
2244 * not worry about this generic constraint ]
2246 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2248 cpu
= select_fallback_rq(task_cpu(p
), p
);
2253 static void update_avg(u64
*avg
, u64 sample
)
2255 s64 diff
= sample
- *avg
;
2261 * try_to_wake_up - wake up a thread
2262 * @p: the to-be-woken-up thread
2263 * @state: the mask of task states that can be woken
2264 * @sync: do a synchronous wakeup?
2266 * Put it on the run-queue if it's not already there. The "current"
2267 * thread is always on the run-queue (except when the actual
2268 * re-schedule is in progress), and as such you're allowed to do
2269 * the simpler "current->state = TASK_RUNNING" to mark yourself
2270 * runnable without the overhead of this.
2272 * returns failure only if the task is already active.
2274 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2277 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2278 unsigned long flags
;
2279 unsigned long en_flags
= ENQUEUE_WAKEUP
;
2282 this_cpu
= get_cpu();
2285 rq
= task_rq_lock(p
, &flags
);
2286 if (!(p
->state
& state
))
2296 if (unlikely(task_running(rq
, p
)))
2300 * In order to handle concurrent wakeups and release the rq->lock
2301 * we put the task in TASK_WAKING state.
2303 * First fix up the nr_uninterruptible count:
2305 if (task_contributes_to_load(p
)) {
2306 if (likely(cpu_online(orig_cpu
)))
2307 rq
->nr_uninterruptible
--;
2309 this_rq()->nr_uninterruptible
--;
2311 p
->state
= TASK_WAKING
;
2313 if (p
->sched_class
->task_waking
) {
2314 p
->sched_class
->task_waking(rq
, p
);
2315 en_flags
|= ENQUEUE_WAKING
;
2318 cpu
= select_task_rq(rq
, p
, SD_BALANCE_WAKE
, wake_flags
);
2319 if (cpu
!= orig_cpu
)
2320 set_task_cpu(p
, cpu
);
2321 __task_rq_unlock(rq
);
2324 raw_spin_lock(&rq
->lock
);
2327 * We migrated the task without holding either rq->lock, however
2328 * since the task is not on the task list itself, nobody else
2329 * will try and migrate the task, hence the rq should match the
2330 * cpu we just moved it to.
2332 WARN_ON(task_cpu(p
) != cpu
);
2333 WARN_ON(p
->state
!= TASK_WAKING
);
2335 #ifdef CONFIG_SCHEDSTATS
2336 schedstat_inc(rq
, ttwu_count
);
2337 if (cpu
== this_cpu
)
2338 schedstat_inc(rq
, ttwu_local
);
2340 struct sched_domain
*sd
;
2341 for_each_domain(this_cpu
, sd
) {
2342 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2343 schedstat_inc(sd
, ttwu_wake_remote
);
2348 #endif /* CONFIG_SCHEDSTATS */
2351 #endif /* CONFIG_SMP */
2352 schedstat_inc(p
, se
.statistics
.nr_wakeups
);
2353 if (wake_flags
& WF_SYNC
)
2354 schedstat_inc(p
, se
.statistics
.nr_wakeups_sync
);
2355 if (orig_cpu
!= cpu
)
2356 schedstat_inc(p
, se
.statistics
.nr_wakeups_migrate
);
2357 if (cpu
== this_cpu
)
2358 schedstat_inc(p
, se
.statistics
.nr_wakeups_local
);
2360 schedstat_inc(p
, se
.statistics
.nr_wakeups_remote
);
2361 activate_task(rq
, p
, en_flags
);
2365 trace_sched_wakeup(p
, success
);
2366 check_preempt_curr(rq
, p
, wake_flags
);
2368 p
->state
= TASK_RUNNING
;
2370 if (p
->sched_class
->task_woken
)
2371 p
->sched_class
->task_woken(rq
, p
);
2373 if (unlikely(rq
->idle_stamp
)) {
2374 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2375 u64 max
= 2*sysctl_sched_migration_cost
;
2380 update_avg(&rq
->avg_idle
, delta
);
2385 task_rq_unlock(rq
, &flags
);
2392 * wake_up_process - Wake up a specific process
2393 * @p: The process to be woken up.
2395 * Attempt to wake up the nominated process and move it to the set of runnable
2396 * processes. Returns 1 if the process was woken up, 0 if it was already
2399 * It may be assumed that this function implies a write memory barrier before
2400 * changing the task state if and only if any tasks are woken up.
2402 int wake_up_process(struct task_struct
*p
)
2404 return try_to_wake_up(p
, TASK_ALL
, 0);
2406 EXPORT_SYMBOL(wake_up_process
);
2408 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2410 return try_to_wake_up(p
, state
, 0);
2414 * Perform scheduler related setup for a newly forked process p.
2415 * p is forked by current.
2417 * __sched_fork() is basic setup used by init_idle() too:
2419 static void __sched_fork(struct task_struct
*p
)
2421 p
->se
.exec_start
= 0;
2422 p
->se
.sum_exec_runtime
= 0;
2423 p
->se
.prev_sum_exec_runtime
= 0;
2424 p
->se
.nr_migrations
= 0;
2426 #ifdef CONFIG_SCHEDSTATS
2427 memset(&p
->se
.statistics
, 0, sizeof(p
->se
.statistics
));
2430 INIT_LIST_HEAD(&p
->rt
.run_list
);
2432 INIT_LIST_HEAD(&p
->se
.group_node
);
2434 #ifdef CONFIG_PREEMPT_NOTIFIERS
2435 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2440 * fork()/clone()-time setup:
2442 void sched_fork(struct task_struct
*p
, int clone_flags
)
2444 int cpu
= get_cpu();
2448 * We mark the process as running here. This guarantees that
2449 * nobody will actually run it, and a signal or other external
2450 * event cannot wake it up and insert it on the runqueue either.
2452 p
->state
= TASK_RUNNING
;
2455 * Revert to default priority/policy on fork if requested.
2457 if (unlikely(p
->sched_reset_on_fork
)) {
2458 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2459 p
->policy
= SCHED_NORMAL
;
2460 p
->normal_prio
= p
->static_prio
;
2463 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2464 p
->static_prio
= NICE_TO_PRIO(0);
2465 p
->normal_prio
= p
->static_prio
;
2470 * We don't need the reset flag anymore after the fork. It has
2471 * fulfilled its duty:
2473 p
->sched_reset_on_fork
= 0;
2477 * Make sure we do not leak PI boosting priority to the child.
2479 p
->prio
= current
->normal_prio
;
2481 if (!rt_prio(p
->prio
))
2482 p
->sched_class
= &fair_sched_class
;
2484 if (p
->sched_class
->task_fork
)
2485 p
->sched_class
->task_fork(p
);
2488 * The child is not yet in the pid-hash so no cgroup attach races,
2489 * and the cgroup is pinned to this child due to cgroup_fork()
2490 * is ran before sched_fork().
2492 * Silence PROVE_RCU.
2495 set_task_cpu(p
, cpu
);
2498 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2499 if (likely(sched_info_on()))
2500 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2502 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2505 #ifdef CONFIG_PREEMPT
2506 /* Want to start with kernel preemption disabled. */
2507 task_thread_info(p
)->preempt_count
= 1;
2509 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2515 * wake_up_new_task - wake up a newly created task for the first time.
2517 * This function will do some initial scheduler statistics housekeeping
2518 * that must be done for every newly created context, then puts the task
2519 * on the runqueue and wakes it.
2521 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2523 unsigned long flags
;
2525 int cpu __maybe_unused
= get_cpu();
2528 rq
= task_rq_lock(p
, &flags
);
2529 p
->state
= TASK_WAKING
;
2532 * Fork balancing, do it here and not earlier because:
2533 * - cpus_allowed can change in the fork path
2534 * - any previously selected cpu might disappear through hotplug
2536 * We set TASK_WAKING so that select_task_rq() can drop rq->lock
2537 * without people poking at ->cpus_allowed.
2539 cpu
= select_task_rq(rq
, p
, SD_BALANCE_FORK
, 0);
2540 set_task_cpu(p
, cpu
);
2542 p
->state
= TASK_RUNNING
;
2543 task_rq_unlock(rq
, &flags
);
2546 rq
= task_rq_lock(p
, &flags
);
2547 activate_task(rq
, p
, 0);
2548 trace_sched_wakeup_new(p
, 1);
2549 check_preempt_curr(rq
, p
, WF_FORK
);
2551 if (p
->sched_class
->task_woken
)
2552 p
->sched_class
->task_woken(rq
, p
);
2554 task_rq_unlock(rq
, &flags
);
2558 #ifdef CONFIG_PREEMPT_NOTIFIERS
2561 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2562 * @notifier: notifier struct to register
2564 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2566 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2568 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2571 * preempt_notifier_unregister - no longer interested in preemption notifications
2572 * @notifier: notifier struct to unregister
2574 * This is safe to call from within a preemption notifier.
2576 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2578 hlist_del(¬ifier
->link
);
2580 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2582 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2584 struct preempt_notifier
*notifier
;
2585 struct hlist_node
*node
;
2587 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2588 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2592 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2593 struct task_struct
*next
)
2595 struct preempt_notifier
*notifier
;
2596 struct hlist_node
*node
;
2598 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2599 notifier
->ops
->sched_out(notifier
, next
);
2602 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2604 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2609 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2610 struct task_struct
*next
)
2614 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2617 * prepare_task_switch - prepare to switch tasks
2618 * @rq: the runqueue preparing to switch
2619 * @prev: the current task that is being switched out
2620 * @next: the task we are going to switch to.
2622 * This is called with the rq lock held and interrupts off. It must
2623 * be paired with a subsequent finish_task_switch after the context
2626 * prepare_task_switch sets up locking and calls architecture specific
2630 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2631 struct task_struct
*next
)
2633 fire_sched_out_preempt_notifiers(prev
, next
);
2634 prepare_lock_switch(rq
, next
);
2635 prepare_arch_switch(next
);
2639 * finish_task_switch - clean up after a task-switch
2640 * @rq: runqueue associated with task-switch
2641 * @prev: the thread we just switched away from.
2643 * finish_task_switch must be called after the context switch, paired
2644 * with a prepare_task_switch call before the context switch.
2645 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2646 * and do any other architecture-specific cleanup actions.
2648 * Note that we may have delayed dropping an mm in context_switch(). If
2649 * so, we finish that here outside of the runqueue lock. (Doing it
2650 * with the lock held can cause deadlocks; see schedule() for
2653 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2654 __releases(rq
->lock
)
2656 struct mm_struct
*mm
= rq
->prev_mm
;
2662 * A task struct has one reference for the use as "current".
2663 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2664 * schedule one last time. The schedule call will never return, and
2665 * the scheduled task must drop that reference.
2666 * The test for TASK_DEAD must occur while the runqueue locks are
2667 * still held, otherwise prev could be scheduled on another cpu, die
2668 * there before we look at prev->state, and then the reference would
2670 * Manfred Spraul <manfred@colorfullife.com>
2672 prev_state
= prev
->state
;
2673 finish_arch_switch(prev
);
2674 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2675 local_irq_disable();
2676 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2677 perf_event_task_sched_in(current
);
2678 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
2680 #endif /* __ARCH_WANT_INTERRUPTS_ON_CTXSW */
2681 finish_lock_switch(rq
, prev
);
2683 fire_sched_in_preempt_notifiers(current
);
2686 if (unlikely(prev_state
== TASK_DEAD
)) {
2688 * Remove function-return probe instances associated with this
2689 * task and put them back on the free list.
2691 kprobe_flush_task(prev
);
2692 put_task_struct(prev
);
2698 /* assumes rq->lock is held */
2699 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2701 if (prev
->sched_class
->pre_schedule
)
2702 prev
->sched_class
->pre_schedule(rq
, prev
);
2705 /* rq->lock is NOT held, but preemption is disabled */
2706 static inline void post_schedule(struct rq
*rq
)
2708 if (rq
->post_schedule
) {
2709 unsigned long flags
;
2711 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2712 if (rq
->curr
->sched_class
->post_schedule
)
2713 rq
->curr
->sched_class
->post_schedule(rq
);
2714 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2716 rq
->post_schedule
= 0;
2722 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2726 static inline void post_schedule(struct rq
*rq
)
2733 * schedule_tail - first thing a freshly forked thread must call.
2734 * @prev: the thread we just switched away from.
2736 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2737 __releases(rq
->lock
)
2739 struct rq
*rq
= this_rq();
2741 finish_task_switch(rq
, prev
);
2744 * FIXME: do we need to worry about rq being invalidated by the
2749 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2750 /* In this case, finish_task_switch does not reenable preemption */
2753 if (current
->set_child_tid
)
2754 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2758 * context_switch - switch to the new MM and the new
2759 * thread's register state.
2762 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2763 struct task_struct
*next
)
2765 struct mm_struct
*mm
, *oldmm
;
2767 prepare_task_switch(rq
, prev
, next
);
2768 trace_sched_switch(prev
, next
);
2770 oldmm
= prev
->active_mm
;
2772 * For paravirt, this is coupled with an exit in switch_to to
2773 * combine the page table reload and the switch backend into
2776 arch_start_context_switch(prev
);
2779 next
->active_mm
= oldmm
;
2780 atomic_inc(&oldmm
->mm_count
);
2781 enter_lazy_tlb(oldmm
, next
);
2783 switch_mm(oldmm
, mm
, next
);
2785 if (likely(!prev
->mm
)) {
2786 prev
->active_mm
= NULL
;
2787 rq
->prev_mm
= oldmm
;
2790 * Since the runqueue lock will be released by the next
2791 * task (which is an invalid locking op but in the case
2792 * of the scheduler it's an obvious special-case), so we
2793 * do an early lockdep release here:
2795 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2796 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2799 /* Here we just switch the register state and the stack. */
2800 switch_to(prev
, next
, prev
);
2804 * this_rq must be evaluated again because prev may have moved
2805 * CPUs since it called schedule(), thus the 'rq' on its stack
2806 * frame will be invalid.
2808 finish_task_switch(this_rq(), prev
);
2812 * nr_running, nr_uninterruptible and nr_context_switches:
2814 * externally visible scheduler statistics: current number of runnable
2815 * threads, current number of uninterruptible-sleeping threads, total
2816 * number of context switches performed since bootup.
2818 unsigned long nr_running(void)
2820 unsigned long i
, sum
= 0;
2822 for_each_online_cpu(i
)
2823 sum
+= cpu_rq(i
)->nr_running
;
2828 unsigned long nr_uninterruptible(void)
2830 unsigned long i
, sum
= 0;
2832 for_each_possible_cpu(i
)
2833 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2836 * Since we read the counters lockless, it might be slightly
2837 * inaccurate. Do not allow it to go below zero though:
2839 if (unlikely((long)sum
< 0))
2845 unsigned long long nr_context_switches(void)
2848 unsigned long long sum
= 0;
2850 for_each_possible_cpu(i
)
2851 sum
+= cpu_rq(i
)->nr_switches
;
2856 unsigned long nr_iowait(void)
2858 unsigned long i
, sum
= 0;
2860 for_each_possible_cpu(i
)
2861 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2866 unsigned long nr_iowait_cpu(int cpu
)
2868 struct rq
*this = cpu_rq(cpu
);
2869 return atomic_read(&this->nr_iowait
);
2872 unsigned long this_cpu_load(void)
2874 struct rq
*this = this_rq();
2875 return this->cpu_load
[0];
2879 /* Variables and functions for calc_load */
2880 static atomic_long_t calc_load_tasks
;
2881 static unsigned long calc_load_update
;
2882 unsigned long avenrun
[3];
2883 EXPORT_SYMBOL(avenrun
);
2885 static long calc_load_fold_active(struct rq
*this_rq
)
2887 long nr_active
, delta
= 0;
2889 nr_active
= this_rq
->nr_running
;
2890 nr_active
+= (long) this_rq
->nr_uninterruptible
;
2892 if (nr_active
!= this_rq
->calc_load_active
) {
2893 delta
= nr_active
- this_rq
->calc_load_active
;
2894 this_rq
->calc_load_active
= nr_active
;
2902 * For NO_HZ we delay the active fold to the next LOAD_FREQ update.
2904 * When making the ILB scale, we should try to pull this in as well.
2906 static atomic_long_t calc_load_tasks_idle
;
2908 static void calc_load_account_idle(struct rq
*this_rq
)
2912 delta
= calc_load_fold_active(this_rq
);
2914 atomic_long_add(delta
, &calc_load_tasks_idle
);
2917 static long calc_load_fold_idle(void)
2922 * Its got a race, we don't care...
2924 if (atomic_long_read(&calc_load_tasks_idle
))
2925 delta
= atomic_long_xchg(&calc_load_tasks_idle
, 0);
2930 static void calc_load_account_idle(struct rq
*this_rq
)
2934 static inline long calc_load_fold_idle(void)
2941 * get_avenrun - get the load average array
2942 * @loads: pointer to dest load array
2943 * @offset: offset to add
2944 * @shift: shift count to shift the result left
2946 * These values are estimates at best, so no need for locking.
2948 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
2950 loads
[0] = (avenrun
[0] + offset
) << shift
;
2951 loads
[1] = (avenrun
[1] + offset
) << shift
;
2952 loads
[2] = (avenrun
[2] + offset
) << shift
;
2955 static unsigned long
2956 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
2959 load
+= active
* (FIXED_1
- exp
);
2960 return load
>> FSHIFT
;
2964 * calc_load - update the avenrun load estimates 10 ticks after the
2965 * CPUs have updated calc_load_tasks.
2967 void calc_global_load(void)
2969 unsigned long upd
= calc_load_update
+ 10;
2972 if (time_before(jiffies
, upd
))
2975 active
= atomic_long_read(&calc_load_tasks
);
2976 active
= active
> 0 ? active
* FIXED_1
: 0;
2978 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
2979 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
2980 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
2982 calc_load_update
+= LOAD_FREQ
;
2986 * Called from update_cpu_load() to periodically update this CPU's
2989 static void calc_load_account_active(struct rq
*this_rq
)
2993 if (time_before(jiffies
, this_rq
->calc_load_update
))
2996 delta
= calc_load_fold_active(this_rq
);
2997 delta
+= calc_load_fold_idle();
2999 atomic_long_add(delta
, &calc_load_tasks
);
3001 this_rq
->calc_load_update
+= LOAD_FREQ
;
3005 * Update rq->cpu_load[] statistics. This function is usually called every
3006 * scheduler tick (TICK_NSEC).
3008 static void update_cpu_load(struct rq
*this_rq
)
3010 unsigned long this_load
= this_rq
->load
.weight
;
3013 this_rq
->nr_load_updates
++;
3015 /* Update our load: */
3016 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3017 unsigned long old_load
, new_load
;
3019 /* scale is effectively 1 << i now, and >> i divides by scale */
3021 old_load
= this_rq
->cpu_load
[i
];
3022 new_load
= this_load
;
3024 * Round up the averaging division if load is increasing. This
3025 * prevents us from getting stuck on 9 if the load is 10, for
3028 if (new_load
> old_load
)
3029 new_load
+= scale
-1;
3030 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3033 calc_load_account_active(this_rq
);
3039 * sched_exec - execve() is a valuable balancing opportunity, because at
3040 * this point the task has the smallest effective memory and cache footprint.
3042 void sched_exec(void)
3044 struct task_struct
*p
= current
;
3045 unsigned long flags
;
3049 rq
= task_rq_lock(p
, &flags
);
3050 dest_cpu
= p
->sched_class
->select_task_rq(rq
, p
, SD_BALANCE_EXEC
, 0);
3051 if (dest_cpu
== smp_processor_id())
3055 * select_task_rq() can race against ->cpus_allowed
3057 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
) &&
3058 likely(cpu_active(dest_cpu
)) && migrate_task(p
, dest_cpu
)) {
3059 struct migration_arg arg
= { p
, dest_cpu
};
3061 task_rq_unlock(rq
, &flags
);
3062 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
3066 task_rq_unlock(rq
, &flags
);
3071 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3073 EXPORT_PER_CPU_SYMBOL(kstat
);
3076 * Return any ns on the sched_clock that have not yet been accounted in
3077 * @p in case that task is currently running.
3079 * Called with task_rq_lock() held on @rq.
3081 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
3085 if (task_current(rq
, p
)) {
3086 update_rq_clock(rq
);
3087 ns
= rq
->clock
- p
->se
.exec_start
;
3095 unsigned long long task_delta_exec(struct task_struct
*p
)
3097 unsigned long flags
;
3101 rq
= task_rq_lock(p
, &flags
);
3102 ns
= do_task_delta_exec(p
, rq
);
3103 task_rq_unlock(rq
, &flags
);
3109 * Return accounted runtime for the task.
3110 * In case the task is currently running, return the runtime plus current's
3111 * pending runtime that have not been accounted yet.
3113 unsigned long long task_sched_runtime(struct task_struct
*p
)
3115 unsigned long flags
;
3119 rq
= task_rq_lock(p
, &flags
);
3120 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3121 task_rq_unlock(rq
, &flags
);
3127 * Return sum_exec_runtime for the thread group.
3128 * In case the task is currently running, return the sum plus current's
3129 * pending runtime that have not been accounted yet.
3131 * Note that the thread group might have other running tasks as well,
3132 * so the return value not includes other pending runtime that other
3133 * running tasks might have.
3135 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
3137 struct task_cputime totals
;
3138 unsigned long flags
;
3142 rq
= task_rq_lock(p
, &flags
);
3143 thread_group_cputime(p
, &totals
);
3144 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
3145 task_rq_unlock(rq
, &flags
);
3151 * Account user cpu time to a process.
3152 * @p: the process that the cpu time gets accounted to
3153 * @cputime: the cpu time spent in user space since the last update
3154 * @cputime_scaled: cputime scaled by cpu frequency
3156 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
3157 cputime_t cputime_scaled
)
3159 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3162 /* Add user time to process. */
3163 p
->utime
= cputime_add(p
->utime
, cputime
);
3164 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3165 account_group_user_time(p
, cputime
);
3167 /* Add user time to cpustat. */
3168 tmp
= cputime_to_cputime64(cputime
);
3169 if (TASK_NICE(p
) > 0)
3170 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3172 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3174 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
3175 /* Account for user time used */
3176 acct_update_integrals(p
);
3180 * Account guest cpu time to a process.
3181 * @p: the process that the cpu time gets accounted to
3182 * @cputime: the cpu time spent in virtual machine since the last update
3183 * @cputime_scaled: cputime scaled by cpu frequency
3185 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
3186 cputime_t cputime_scaled
)
3189 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3191 tmp
= cputime_to_cputime64(cputime
);
3193 /* Add guest time to process. */
3194 p
->utime
= cputime_add(p
->utime
, cputime
);
3195 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
3196 account_group_user_time(p
, cputime
);
3197 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3199 /* Add guest time to cpustat. */
3200 if (TASK_NICE(p
) > 0) {
3201 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3202 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
3204 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3205 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3210 * Account system cpu time to a process.
3211 * @p: the process that the cpu time gets accounted to
3212 * @hardirq_offset: the offset to subtract from hardirq_count()
3213 * @cputime: the cpu time spent in kernel space since the last update
3214 * @cputime_scaled: cputime scaled by cpu frequency
3216 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3217 cputime_t cputime
, cputime_t cputime_scaled
)
3219 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3222 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3223 account_guest_time(p
, cputime
, cputime_scaled
);
3227 /* Add system time to process. */
3228 p
->stime
= cputime_add(p
->stime
, cputime
);
3229 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
3230 account_group_system_time(p
, cputime
);
3232 /* Add system time to cpustat. */
3233 tmp
= cputime_to_cputime64(cputime
);
3234 if (hardirq_count() - hardirq_offset
)
3235 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3236 else if (softirq_count())
3237 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3239 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3241 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
3243 /* Account for system time used */
3244 acct_update_integrals(p
);
3248 * Account for involuntary wait time.
3249 * @steal: the cpu time spent in involuntary wait
3251 void account_steal_time(cputime_t cputime
)
3253 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3254 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3256 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
3260 * Account for idle time.
3261 * @cputime: the cpu time spent in idle wait
3263 void account_idle_time(cputime_t cputime
)
3265 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3266 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
3267 struct rq
*rq
= this_rq();
3269 if (atomic_read(&rq
->nr_iowait
) > 0)
3270 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
3272 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
3275 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
3278 * Account a single tick of cpu time.
3279 * @p: the process that the cpu time gets accounted to
3280 * @user_tick: indicates if the tick is a user or a system tick
3282 void account_process_tick(struct task_struct
*p
, int user_tick
)
3284 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
3285 struct rq
*rq
= this_rq();
3288 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
3289 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
3290 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
3293 account_idle_time(cputime_one_jiffy
);
3297 * Account multiple ticks of steal time.
3298 * @p: the process from which the cpu time has been stolen
3299 * @ticks: number of stolen ticks
3301 void account_steal_ticks(unsigned long ticks
)
3303 account_steal_time(jiffies_to_cputime(ticks
));
3307 * Account multiple ticks of idle time.
3308 * @ticks: number of stolen ticks
3310 void account_idle_ticks(unsigned long ticks
)
3312 account_idle_time(jiffies_to_cputime(ticks
));
3318 * Use precise platform statistics if available:
3320 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
3321 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3327 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3329 struct task_cputime cputime
;
3331 thread_group_cputime(p
, &cputime
);
3333 *ut
= cputime
.utime
;
3334 *st
= cputime
.stime
;
3338 #ifndef nsecs_to_cputime
3339 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
3342 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3344 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
3347 * Use CFS's precise accounting:
3349 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
3354 temp
= (u64
)(rtime
* utime
);
3355 do_div(temp
, total
);
3356 utime
= (cputime_t
)temp
;
3361 * Compare with previous values, to keep monotonicity:
3363 p
->prev_utime
= max(p
->prev_utime
, utime
);
3364 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
3366 *ut
= p
->prev_utime
;
3367 *st
= p
->prev_stime
;
3371 * Must be called with siglock held.
3373 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
3375 struct signal_struct
*sig
= p
->signal
;
3376 struct task_cputime cputime
;
3377 cputime_t rtime
, utime
, total
;
3379 thread_group_cputime(p
, &cputime
);
3381 total
= cputime_add(cputime
.utime
, cputime
.stime
);
3382 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
3387 temp
= (u64
)(rtime
* cputime
.utime
);
3388 do_div(temp
, total
);
3389 utime
= (cputime_t
)temp
;
3393 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
3394 sig
->prev_stime
= max(sig
->prev_stime
,
3395 cputime_sub(rtime
, sig
->prev_utime
));
3397 *ut
= sig
->prev_utime
;
3398 *st
= sig
->prev_stime
;
3403 * This function gets called by the timer code, with HZ frequency.
3404 * We call it with interrupts disabled.
3406 * It also gets called by the fork code, when changing the parent's
3409 void scheduler_tick(void)
3411 int cpu
= smp_processor_id();
3412 struct rq
*rq
= cpu_rq(cpu
);
3413 struct task_struct
*curr
= rq
->curr
;
3417 raw_spin_lock(&rq
->lock
);
3418 update_rq_clock(rq
);
3419 update_cpu_load(rq
);
3420 curr
->sched_class
->task_tick(rq
, curr
, 0);
3421 raw_spin_unlock(&rq
->lock
);
3423 perf_event_task_tick(curr
);
3426 rq
->idle_at_tick
= idle_cpu(cpu
);
3427 trigger_load_balance(rq
, cpu
);
3431 notrace
unsigned long get_parent_ip(unsigned long addr
)
3433 if (in_lock_functions(addr
)) {
3434 addr
= CALLER_ADDR2
;
3435 if (in_lock_functions(addr
))
3436 addr
= CALLER_ADDR3
;
3441 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
3442 defined(CONFIG_PREEMPT_TRACER))
3444 void __kprobes
add_preempt_count(int val
)
3446 #ifdef CONFIG_DEBUG_PREEMPT
3450 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3453 preempt_count() += val
;
3454 #ifdef CONFIG_DEBUG_PREEMPT
3456 * Spinlock count overflowing soon?
3458 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3461 if (preempt_count() == val
)
3462 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3464 EXPORT_SYMBOL(add_preempt_count
);
3466 void __kprobes
sub_preempt_count(int val
)
3468 #ifdef CONFIG_DEBUG_PREEMPT
3472 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3475 * Is the spinlock portion underflowing?
3477 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3478 !(preempt_count() & PREEMPT_MASK
)))
3482 if (preempt_count() == val
)
3483 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
3484 preempt_count() -= val
;
3486 EXPORT_SYMBOL(sub_preempt_count
);
3491 * Print scheduling while atomic bug:
3493 static noinline
void __schedule_bug(struct task_struct
*prev
)
3495 struct pt_regs
*regs
= get_irq_regs();
3497 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
3498 prev
->comm
, prev
->pid
, preempt_count());
3500 debug_show_held_locks(prev
);
3502 if (irqs_disabled())
3503 print_irqtrace_events(prev
);
3512 * Various schedule()-time debugging checks and statistics:
3514 static inline void schedule_debug(struct task_struct
*prev
)
3517 * Test if we are atomic. Since do_exit() needs to call into
3518 * schedule() atomically, we ignore that path for now.
3519 * Otherwise, whine if we are scheduling when we should not be.
3521 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
3522 __schedule_bug(prev
);
3524 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3526 schedstat_inc(this_rq(), sched_count
);
3527 #ifdef CONFIG_SCHEDSTATS
3528 if (unlikely(prev
->lock_depth
>= 0)) {
3529 schedstat_inc(this_rq(), bkl_count
);
3530 schedstat_inc(prev
, sched_info
.bkl_count
);
3535 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
3538 update_rq_clock(rq
);
3539 rq
->skip_clock_update
= 0;
3540 prev
->sched_class
->put_prev_task(rq
, prev
);
3544 * Pick up the highest-prio task:
3546 static inline struct task_struct
*
3547 pick_next_task(struct rq
*rq
)
3549 const struct sched_class
*class;
3550 struct task_struct
*p
;
3553 * Optimization: we know that if all tasks are in
3554 * the fair class we can call that function directly:
3556 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
3557 p
= fair_sched_class
.pick_next_task(rq
);
3562 class = sched_class_highest
;
3564 p
= class->pick_next_task(rq
);
3568 * Will never be NULL as the idle class always
3569 * returns a non-NULL p:
3571 class = class->next
;
3576 * schedule() is the main scheduler function.
3578 asmlinkage
void __sched
schedule(void)
3580 struct task_struct
*prev
, *next
;
3581 unsigned long *switch_count
;
3587 cpu
= smp_processor_id();
3589 rcu_note_context_switch(cpu
);
3591 switch_count
= &prev
->nivcsw
;
3593 release_kernel_lock(prev
);
3594 need_resched_nonpreemptible
:
3596 schedule_debug(prev
);
3598 if (sched_feat(HRTICK
))
3601 raw_spin_lock_irq(&rq
->lock
);
3602 clear_tsk_need_resched(prev
);
3604 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3605 if (unlikely(signal_pending_state(prev
->state
, prev
)))
3606 prev
->state
= TASK_RUNNING
;
3608 deactivate_task(rq
, prev
, DEQUEUE_SLEEP
);
3609 switch_count
= &prev
->nvcsw
;
3612 pre_schedule(rq
, prev
);
3614 if (unlikely(!rq
->nr_running
))
3615 idle_balance(cpu
, rq
);
3617 put_prev_task(rq
, prev
);
3618 next
= pick_next_task(rq
);
3620 if (likely(prev
!= next
)) {
3621 sched_info_switch(prev
, next
);
3622 perf_event_task_sched_out(prev
, next
);
3628 context_switch(rq
, prev
, next
); /* unlocks the rq */
3630 * the context switch might have flipped the stack from under
3631 * us, hence refresh the local variables.
3633 cpu
= smp_processor_id();
3636 raw_spin_unlock_irq(&rq
->lock
);
3640 if (unlikely(reacquire_kernel_lock(current
) < 0)) {
3642 switch_count
= &prev
->nivcsw
;
3643 goto need_resched_nonpreemptible
;
3646 preempt_enable_no_resched();
3650 EXPORT_SYMBOL(schedule
);
3652 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
3654 * Look out! "owner" is an entirely speculative pointer
3655 * access and not reliable.
3657 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
3662 if (!sched_feat(OWNER_SPIN
))
3665 #ifdef CONFIG_DEBUG_PAGEALLOC
3667 * Need to access the cpu field knowing that
3668 * DEBUG_PAGEALLOC could have unmapped it if
3669 * the mutex owner just released it and exited.
3671 if (probe_kernel_address(&owner
->cpu
, cpu
))
3678 * Even if the access succeeded (likely case),
3679 * the cpu field may no longer be valid.
3681 if (cpu
>= nr_cpumask_bits
)
3685 * We need to validate that we can do a
3686 * get_cpu() and that we have the percpu area.
3688 if (!cpu_online(cpu
))
3695 * Owner changed, break to re-assess state.
3697 if (lock
->owner
!= owner
)
3701 * Is that owner really running on that cpu?
3703 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
3713 #ifdef CONFIG_PREEMPT
3715 * this is the entry point to schedule() from in-kernel preemption
3716 * off of preempt_enable. Kernel preemptions off return from interrupt
3717 * occur there and call schedule directly.
3719 asmlinkage
void __sched
preempt_schedule(void)
3721 struct thread_info
*ti
= current_thread_info();
3724 * If there is a non-zero preempt_count or interrupts are disabled,
3725 * we do not want to preempt the current task. Just return..
3727 if (likely(ti
->preempt_count
|| irqs_disabled()))
3731 add_preempt_count(PREEMPT_ACTIVE
);
3733 sub_preempt_count(PREEMPT_ACTIVE
);
3736 * Check again in case we missed a preemption opportunity
3737 * between schedule and now.
3740 } while (need_resched());
3742 EXPORT_SYMBOL(preempt_schedule
);
3745 * this is the entry point to schedule() from kernel preemption
3746 * off of irq context.
3747 * Note, that this is called and return with irqs disabled. This will
3748 * protect us against recursive calling from irq.
3750 asmlinkage
void __sched
preempt_schedule_irq(void)
3752 struct thread_info
*ti
= current_thread_info();
3754 /* Catch callers which need to be fixed */
3755 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3758 add_preempt_count(PREEMPT_ACTIVE
);
3761 local_irq_disable();
3762 sub_preempt_count(PREEMPT_ACTIVE
);
3765 * Check again in case we missed a preemption opportunity
3766 * between schedule and now.
3769 } while (need_resched());
3772 #endif /* CONFIG_PREEMPT */
3774 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
3777 return try_to_wake_up(curr
->private, mode
, wake_flags
);
3779 EXPORT_SYMBOL(default_wake_function
);
3782 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3783 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3784 * number) then we wake all the non-exclusive tasks and one exclusive task.
3786 * There are circumstances in which we can try to wake a task which has already
3787 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3788 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3790 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3791 int nr_exclusive
, int wake_flags
, void *key
)
3793 wait_queue_t
*curr
, *next
;
3795 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
3796 unsigned flags
= curr
->flags
;
3798 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
3799 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3805 * __wake_up - wake up threads blocked on a waitqueue.
3807 * @mode: which threads
3808 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3809 * @key: is directly passed to the wakeup function
3811 * It may be assumed that this function implies a write memory barrier before
3812 * changing the task state if and only if any tasks are woken up.
3814 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3815 int nr_exclusive
, void *key
)
3817 unsigned long flags
;
3819 spin_lock_irqsave(&q
->lock
, flags
);
3820 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3821 spin_unlock_irqrestore(&q
->lock
, flags
);
3823 EXPORT_SYMBOL(__wake_up
);
3826 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3828 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3830 __wake_up_common(q
, mode
, 1, 0, NULL
);
3832 EXPORT_SYMBOL_GPL(__wake_up_locked
);
3834 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
3836 __wake_up_common(q
, mode
, 1, 0, key
);
3840 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
3842 * @mode: which threads
3843 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3844 * @key: opaque value to be passed to wakeup targets
3846 * The sync wakeup differs that the waker knows that it will schedule
3847 * away soon, so while the target thread will be woken up, it will not
3848 * be migrated to another CPU - ie. the two threads are 'synchronized'
3849 * with each other. This can prevent needless bouncing between CPUs.
3851 * On UP it can prevent extra preemption.
3853 * It may be assumed that this function implies a write memory barrier before
3854 * changing the task state if and only if any tasks are woken up.
3856 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
3857 int nr_exclusive
, void *key
)
3859 unsigned long flags
;
3860 int wake_flags
= WF_SYNC
;
3865 if (unlikely(!nr_exclusive
))
3868 spin_lock_irqsave(&q
->lock
, flags
);
3869 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
3870 spin_unlock_irqrestore(&q
->lock
, flags
);
3872 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
3875 * __wake_up_sync - see __wake_up_sync_key()
3877 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3879 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
3881 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3884 * complete: - signals a single thread waiting on this completion
3885 * @x: holds the state of this particular completion
3887 * This will wake up a single thread waiting on this completion. Threads will be
3888 * awakened in the same order in which they were queued.
3890 * See also complete_all(), wait_for_completion() and related routines.
3892 * It may be assumed that this function implies a write memory barrier before
3893 * changing the task state if and only if any tasks are woken up.
3895 void complete(struct completion
*x
)
3897 unsigned long flags
;
3899 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3901 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
3902 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3904 EXPORT_SYMBOL(complete
);
3907 * complete_all: - signals all threads waiting on this completion
3908 * @x: holds the state of this particular completion
3910 * This will wake up all threads waiting on this particular completion event.
3912 * It may be assumed that this function implies a write memory barrier before
3913 * changing the task state if and only if any tasks are woken up.
3915 void complete_all(struct completion
*x
)
3917 unsigned long flags
;
3919 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3920 x
->done
+= UINT_MAX
/2;
3921 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
3922 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3924 EXPORT_SYMBOL(complete_all
);
3926 static inline long __sched
3927 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
3930 DECLARE_WAITQUEUE(wait
, current
);
3932 __add_wait_queue_tail_exclusive(&x
->wait
, &wait
);
3934 if (signal_pending_state(state
, current
)) {
3935 timeout
= -ERESTARTSYS
;
3938 __set_current_state(state
);
3939 spin_unlock_irq(&x
->wait
.lock
);
3940 timeout
= schedule_timeout(timeout
);
3941 spin_lock_irq(&x
->wait
.lock
);
3942 } while (!x
->done
&& timeout
);
3943 __remove_wait_queue(&x
->wait
, &wait
);
3948 return timeout
?: 1;
3952 wait_for_common(struct completion
*x
, long timeout
, int state
)
3956 spin_lock_irq(&x
->wait
.lock
);
3957 timeout
= do_wait_for_common(x
, timeout
, state
);
3958 spin_unlock_irq(&x
->wait
.lock
);
3963 * wait_for_completion: - waits for completion of a task
3964 * @x: holds the state of this particular completion
3966 * This waits to be signaled for completion of a specific task. It is NOT
3967 * interruptible and there is no timeout.
3969 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
3970 * and interrupt capability. Also see complete().
3972 void __sched
wait_for_completion(struct completion
*x
)
3974 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
3976 EXPORT_SYMBOL(wait_for_completion
);
3979 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
3980 * @x: holds the state of this particular completion
3981 * @timeout: timeout value in jiffies
3983 * This waits for either a completion of a specific task to be signaled or for a
3984 * specified timeout to expire. The timeout is in jiffies. It is not
3987 unsigned long __sched
3988 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3990 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
3992 EXPORT_SYMBOL(wait_for_completion_timeout
);
3995 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
3996 * @x: holds the state of this particular completion
3998 * This waits for completion of a specific task to be signaled. It is
4001 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4003 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4004 if (t
== -ERESTARTSYS
)
4008 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4011 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4012 * @x: holds the state of this particular completion
4013 * @timeout: timeout value in jiffies
4015 * This waits for either a completion of a specific task to be signaled or for a
4016 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4018 unsigned long __sched
4019 wait_for_completion_interruptible_timeout(struct completion
*x
,
4020 unsigned long timeout
)
4022 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4024 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4027 * wait_for_completion_killable: - waits for completion of a task (killable)
4028 * @x: holds the state of this particular completion
4030 * This waits to be signaled for completion of a specific task. It can be
4031 * interrupted by a kill signal.
4033 int __sched
wait_for_completion_killable(struct completion
*x
)
4035 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4036 if (t
== -ERESTARTSYS
)
4040 EXPORT_SYMBOL(wait_for_completion_killable
);
4043 * wait_for_completion_killable_timeout: - waits for completion of a task (w/(to,killable))
4044 * @x: holds the state of this particular completion
4045 * @timeout: timeout value in jiffies
4047 * This waits for either a completion of a specific task to be
4048 * signaled or for a specified timeout to expire. It can be
4049 * interrupted by a kill signal. The timeout is in jiffies.
4051 unsigned long __sched
4052 wait_for_completion_killable_timeout(struct completion
*x
,
4053 unsigned long timeout
)
4055 return wait_for_common(x
, timeout
, TASK_KILLABLE
);
4057 EXPORT_SYMBOL(wait_for_completion_killable_timeout
);
4060 * try_wait_for_completion - try to decrement a completion without blocking
4061 * @x: completion structure
4063 * Returns: 0 if a decrement cannot be done without blocking
4064 * 1 if a decrement succeeded.
4066 * If a completion is being used as a counting completion,
4067 * attempt to decrement the counter without blocking. This
4068 * enables us to avoid waiting if the resource the completion
4069 * is protecting is not available.
4071 bool try_wait_for_completion(struct completion
*x
)
4073 unsigned long flags
;
4076 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4081 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4084 EXPORT_SYMBOL(try_wait_for_completion
);
4087 * completion_done - Test to see if a completion has any waiters
4088 * @x: completion structure
4090 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4091 * 1 if there are no waiters.
4094 bool completion_done(struct completion
*x
)
4096 unsigned long flags
;
4099 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4102 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4105 EXPORT_SYMBOL(completion_done
);
4108 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4110 unsigned long flags
;
4113 init_waitqueue_entry(&wait
, current
);
4115 __set_current_state(state
);
4117 spin_lock_irqsave(&q
->lock
, flags
);
4118 __add_wait_queue(q
, &wait
);
4119 spin_unlock(&q
->lock
);
4120 timeout
= schedule_timeout(timeout
);
4121 spin_lock_irq(&q
->lock
);
4122 __remove_wait_queue(q
, &wait
);
4123 spin_unlock_irqrestore(&q
->lock
, flags
);
4128 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4130 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4132 EXPORT_SYMBOL(interruptible_sleep_on
);
4135 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4137 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4139 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4141 void __sched
sleep_on(wait_queue_head_t
*q
)
4143 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4145 EXPORT_SYMBOL(sleep_on
);
4147 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4149 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4151 EXPORT_SYMBOL(sleep_on_timeout
);
4153 #ifdef CONFIG_RT_MUTEXES
4156 * rt_mutex_setprio - set the current priority of a task
4158 * @prio: prio value (kernel-internal form)
4160 * This function changes the 'effective' priority of a task. It does
4161 * not touch ->normal_prio like __setscheduler().
4163 * Used by the rt_mutex code to implement priority inheritance logic.
4165 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4167 unsigned long flags
;
4168 int oldprio
, on_rq
, running
;
4170 const struct sched_class
*prev_class
;
4172 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4174 rq
= task_rq_lock(p
, &flags
);
4177 prev_class
= p
->sched_class
;
4178 on_rq
= p
->se
.on_rq
;
4179 running
= task_current(rq
, p
);
4181 dequeue_task(rq
, p
, 0);
4183 p
->sched_class
->put_prev_task(rq
, p
);
4186 p
->sched_class
= &rt_sched_class
;
4188 p
->sched_class
= &fair_sched_class
;
4193 p
->sched_class
->set_curr_task(rq
);
4195 enqueue_task(rq
, p
, oldprio
< prio
? ENQUEUE_HEAD
: 0);
4197 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4199 task_rq_unlock(rq
, &flags
);
4204 void set_user_nice(struct task_struct
*p
, long nice
)
4206 int old_prio
, delta
, on_rq
;
4207 unsigned long flags
;
4210 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4213 * We have to be careful, if called from sys_setpriority(),
4214 * the task might be in the middle of scheduling on another CPU.
4216 rq
= task_rq_lock(p
, &flags
);
4218 * The RT priorities are set via sched_setscheduler(), but we still
4219 * allow the 'normal' nice value to be set - but as expected
4220 * it wont have any effect on scheduling until the task is
4221 * SCHED_FIFO/SCHED_RR:
4223 if (task_has_rt_policy(p
)) {
4224 p
->static_prio
= NICE_TO_PRIO(nice
);
4227 on_rq
= p
->se
.on_rq
;
4229 dequeue_task(rq
, p
, 0);
4231 p
->static_prio
= NICE_TO_PRIO(nice
);
4234 p
->prio
= effective_prio(p
);
4235 delta
= p
->prio
- old_prio
;
4238 enqueue_task(rq
, p
, 0);
4240 * If the task increased its priority or is running and
4241 * lowered its priority, then reschedule its CPU:
4243 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4244 resched_task(rq
->curr
);
4247 task_rq_unlock(rq
, &flags
);
4249 EXPORT_SYMBOL(set_user_nice
);
4252 * can_nice - check if a task can reduce its nice value
4256 int can_nice(const struct task_struct
*p
, const int nice
)
4258 /* convert nice value [19,-20] to rlimit style value [1,40] */
4259 int nice_rlim
= 20 - nice
;
4261 return (nice_rlim
<= task_rlimit(p
, RLIMIT_NICE
) ||
4262 capable(CAP_SYS_NICE
));
4265 #ifdef __ARCH_WANT_SYS_NICE
4268 * sys_nice - change the priority of the current process.
4269 * @increment: priority increment
4271 * sys_setpriority is a more generic, but much slower function that
4272 * does similar things.
4274 SYSCALL_DEFINE1(nice
, int, increment
)
4279 * Setpriority might change our priority at the same moment.
4280 * We don't have to worry. Conceptually one call occurs first
4281 * and we have a single winner.
4283 if (increment
< -40)
4288 nice
= TASK_NICE(current
) + increment
;
4294 if (increment
< 0 && !can_nice(current
, nice
))
4297 retval
= security_task_setnice(current
, nice
);
4301 set_user_nice(current
, nice
);
4308 * task_prio - return the priority value of a given task.
4309 * @p: the task in question.
4311 * This is the priority value as seen by users in /proc.
4312 * RT tasks are offset by -200. Normal tasks are centered
4313 * around 0, value goes from -16 to +15.
4315 int task_prio(const struct task_struct
*p
)
4317 return p
->prio
- MAX_RT_PRIO
;
4321 * task_nice - return the nice value of a given task.
4322 * @p: the task in question.
4324 int task_nice(const struct task_struct
*p
)
4326 return TASK_NICE(p
);
4328 EXPORT_SYMBOL(task_nice
);
4331 * idle_cpu - is a given cpu idle currently?
4332 * @cpu: the processor in question.
4334 int idle_cpu(int cpu
)
4336 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4340 * idle_task - return the idle task for a given cpu.
4341 * @cpu: the processor in question.
4343 struct task_struct
*idle_task(int cpu
)
4345 return cpu_rq(cpu
)->idle
;
4349 * find_process_by_pid - find a process with a matching PID value.
4350 * @pid: the pid in question.
4352 static struct task_struct
*find_process_by_pid(pid_t pid
)
4354 return pid
? find_task_by_vpid(pid
) : current
;
4357 /* Actually do priority change: must hold rq lock. */
4359 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4361 BUG_ON(p
->se
.on_rq
);
4364 p
->rt_priority
= prio
;
4365 p
->normal_prio
= normal_prio(p
);
4366 /* we are holding p->pi_lock already */
4367 p
->prio
= rt_mutex_getprio(p
);
4368 if (rt_prio(p
->prio
))
4369 p
->sched_class
= &rt_sched_class
;
4371 p
->sched_class
= &fair_sched_class
;
4376 * check the target process has a UID that matches the current process's
4378 static bool check_same_owner(struct task_struct
*p
)
4380 const struct cred
*cred
= current_cred(), *pcred
;
4384 pcred
= __task_cred(p
);
4385 match
= (cred
->euid
== pcred
->euid
||
4386 cred
->euid
== pcred
->uid
);
4391 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
4392 struct sched_param
*param
, bool user
)
4394 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4395 unsigned long flags
;
4396 const struct sched_class
*prev_class
;
4400 /* may grab non-irq protected spin_locks */
4401 BUG_ON(in_interrupt());
4403 /* double check policy once rq lock held */
4405 reset_on_fork
= p
->sched_reset_on_fork
;
4406 policy
= oldpolicy
= p
->policy
;
4408 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
4409 policy
&= ~SCHED_RESET_ON_FORK
;
4411 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4412 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4413 policy
!= SCHED_IDLE
)
4418 * Valid priorities for SCHED_FIFO and SCHED_RR are
4419 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4420 * SCHED_BATCH and SCHED_IDLE is 0.
4422 if (param
->sched_priority
< 0 ||
4423 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4424 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4426 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4430 * Allow unprivileged RT tasks to decrease priority:
4432 if (user
&& !capable(CAP_SYS_NICE
)) {
4433 if (rt_policy(policy
)) {
4434 unsigned long rlim_rtprio
;
4436 if (!lock_task_sighand(p
, &flags
))
4438 rlim_rtprio
= task_rlimit(p
, RLIMIT_RTPRIO
);
4439 unlock_task_sighand(p
, &flags
);
4441 /* can't set/change the rt policy */
4442 if (policy
!= p
->policy
&& !rlim_rtprio
)
4445 /* can't increase priority */
4446 if (param
->sched_priority
> p
->rt_priority
&&
4447 param
->sched_priority
> rlim_rtprio
)
4451 * Like positive nice levels, dont allow tasks to
4452 * move out of SCHED_IDLE either:
4454 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4457 /* can't change other user's priorities */
4458 if (!check_same_owner(p
))
4461 /* Normal users shall not reset the sched_reset_on_fork flag */
4462 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
4467 retval
= security_task_setscheduler(p
, policy
, param
);
4473 * make sure no PI-waiters arrive (or leave) while we are
4474 * changing the priority of the task:
4476 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
4478 * To be able to change p->policy safely, the apropriate
4479 * runqueue lock must be held.
4481 rq
= __task_rq_lock(p
);
4483 #ifdef CONFIG_RT_GROUP_SCHED
4486 * Do not allow realtime tasks into groups that have no runtime
4489 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
4490 task_group(p
)->rt_bandwidth
.rt_runtime
== 0) {
4491 __task_rq_unlock(rq
);
4492 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4498 /* recheck policy now with rq lock held */
4499 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4500 policy
= oldpolicy
= -1;
4501 __task_rq_unlock(rq
);
4502 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4505 on_rq
= p
->se
.on_rq
;
4506 running
= task_current(rq
, p
);
4508 deactivate_task(rq
, p
, 0);
4510 p
->sched_class
->put_prev_task(rq
, p
);
4512 p
->sched_reset_on_fork
= reset_on_fork
;
4515 prev_class
= p
->sched_class
;
4516 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4519 p
->sched_class
->set_curr_task(rq
);
4521 activate_task(rq
, p
, 0);
4523 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4525 __task_rq_unlock(rq
);
4526 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4528 rt_mutex_adjust_pi(p
);
4534 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4535 * @p: the task in question.
4536 * @policy: new policy.
4537 * @param: structure containing the new RT priority.
4539 * NOTE that the task may be already dead.
4541 int sched_setscheduler(struct task_struct
*p
, int policy
,
4542 struct sched_param
*param
)
4544 return __sched_setscheduler(p
, policy
, param
, true);
4546 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4549 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
4550 * @p: the task in question.
4551 * @policy: new policy.
4552 * @param: structure containing the new RT priority.
4554 * Just like sched_setscheduler, only don't bother checking if the
4555 * current context has permission. For example, this is needed in
4556 * stop_machine(): we create temporary high priority worker threads,
4557 * but our caller might not have that capability.
4559 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
4560 struct sched_param
*param
)
4562 return __sched_setscheduler(p
, policy
, param
, false);
4566 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4568 struct sched_param lparam
;
4569 struct task_struct
*p
;
4572 if (!param
|| pid
< 0)
4574 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4579 p
= find_process_by_pid(pid
);
4581 retval
= sched_setscheduler(p
, policy
, &lparam
);
4588 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4589 * @pid: the pid in question.
4590 * @policy: new policy.
4591 * @param: structure containing the new RT priority.
4593 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
4594 struct sched_param __user
*, param
)
4596 /* negative values for policy are not valid */
4600 return do_sched_setscheduler(pid
, policy
, param
);
4604 * sys_sched_setparam - set/change the RT priority of a thread
4605 * @pid: the pid in question.
4606 * @param: structure containing the new RT priority.
4608 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4610 return do_sched_setscheduler(pid
, -1, param
);
4614 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4615 * @pid: the pid in question.
4617 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
4619 struct task_struct
*p
;
4627 p
= find_process_by_pid(pid
);
4629 retval
= security_task_getscheduler(p
);
4632 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
4639 * sys_sched_getparam - get the RT priority of a thread
4640 * @pid: the pid in question.
4641 * @param: structure containing the RT priority.
4643 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
4645 struct sched_param lp
;
4646 struct task_struct
*p
;
4649 if (!param
|| pid
< 0)
4653 p
= find_process_by_pid(pid
);
4658 retval
= security_task_getscheduler(p
);
4662 lp
.sched_priority
= p
->rt_priority
;
4666 * This one might sleep, we cannot do it with a spinlock held ...
4668 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4677 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
4679 cpumask_var_t cpus_allowed
, new_mask
;
4680 struct task_struct
*p
;
4686 p
= find_process_by_pid(pid
);
4693 /* Prevent p going away */
4697 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
4701 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
4703 goto out_free_cpus_allowed
;
4706 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
4709 retval
= security_task_setscheduler(p
, 0, NULL
);
4713 cpuset_cpus_allowed(p
, cpus_allowed
);
4714 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
4716 retval
= set_cpus_allowed_ptr(p
, new_mask
);
4719 cpuset_cpus_allowed(p
, cpus_allowed
);
4720 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
4722 * We must have raced with a concurrent cpuset
4723 * update. Just reset the cpus_allowed to the
4724 * cpuset's cpus_allowed
4726 cpumask_copy(new_mask
, cpus_allowed
);
4731 free_cpumask_var(new_mask
);
4732 out_free_cpus_allowed
:
4733 free_cpumask_var(cpus_allowed
);
4740 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4741 struct cpumask
*new_mask
)
4743 if (len
< cpumask_size())
4744 cpumask_clear(new_mask
);
4745 else if (len
> cpumask_size())
4746 len
= cpumask_size();
4748 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4752 * sys_sched_setaffinity - set the cpu affinity of a process
4753 * @pid: pid of the process
4754 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4755 * @user_mask_ptr: user-space pointer to the new cpu mask
4757 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
4758 unsigned long __user
*, user_mask_ptr
)
4760 cpumask_var_t new_mask
;
4763 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
4766 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
4768 retval
= sched_setaffinity(pid
, new_mask
);
4769 free_cpumask_var(new_mask
);
4773 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
4775 struct task_struct
*p
;
4776 unsigned long flags
;
4784 p
= find_process_by_pid(pid
);
4788 retval
= security_task_getscheduler(p
);
4792 rq
= task_rq_lock(p
, &flags
);
4793 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
4794 task_rq_unlock(rq
, &flags
);
4804 * sys_sched_getaffinity - get the cpu affinity of a process
4805 * @pid: pid of the process
4806 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4807 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4809 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
4810 unsigned long __user
*, user_mask_ptr
)
4815 if ((len
* BITS_PER_BYTE
) < nr_cpu_ids
)
4817 if (len
& (sizeof(unsigned long)-1))
4820 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
4823 ret
= sched_getaffinity(pid
, mask
);
4825 size_t retlen
= min_t(size_t, len
, cpumask_size());
4827 if (copy_to_user(user_mask_ptr
, mask
, retlen
))
4832 free_cpumask_var(mask
);
4838 * sys_sched_yield - yield the current processor to other threads.
4840 * This function yields the current CPU to other tasks. If there are no
4841 * other threads running on this CPU then this function will return.
4843 SYSCALL_DEFINE0(sched_yield
)
4845 struct rq
*rq
= this_rq_lock();
4847 schedstat_inc(rq
, yld_count
);
4848 current
->sched_class
->yield_task(rq
);
4851 * Since we are going to call schedule() anyway, there's
4852 * no need to preempt or enable interrupts:
4854 __release(rq
->lock
);
4855 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4856 do_raw_spin_unlock(&rq
->lock
);
4857 preempt_enable_no_resched();
4864 static inline int should_resched(void)
4866 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
4869 static void __cond_resched(void)
4871 add_preempt_count(PREEMPT_ACTIVE
);
4873 sub_preempt_count(PREEMPT_ACTIVE
);
4876 int __sched
_cond_resched(void)
4878 if (should_resched()) {
4884 EXPORT_SYMBOL(_cond_resched
);
4887 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
4888 * call schedule, and on return reacquire the lock.
4890 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4891 * operations here to prevent schedule() from being called twice (once via
4892 * spin_unlock(), once by hand).
4894 int __cond_resched_lock(spinlock_t
*lock
)
4896 int resched
= should_resched();
4899 lockdep_assert_held(lock
);
4901 if (spin_needbreak(lock
) || resched
) {
4912 EXPORT_SYMBOL(__cond_resched_lock
);
4914 int __sched
__cond_resched_softirq(void)
4916 BUG_ON(!in_softirq());
4918 if (should_resched()) {
4926 EXPORT_SYMBOL(__cond_resched_softirq
);
4929 * yield - yield the current processor to other threads.
4931 * This is a shortcut for kernel-space yielding - it marks the
4932 * thread runnable and calls sys_sched_yield().
4934 void __sched
yield(void)
4936 set_current_state(TASK_RUNNING
);
4939 EXPORT_SYMBOL(yield
);
4942 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4943 * that process accounting knows that this is a task in IO wait state.
4945 void __sched
io_schedule(void)
4947 struct rq
*rq
= raw_rq();
4949 delayacct_blkio_start();
4950 atomic_inc(&rq
->nr_iowait
);
4951 current
->in_iowait
= 1;
4953 current
->in_iowait
= 0;
4954 atomic_dec(&rq
->nr_iowait
);
4955 delayacct_blkio_end();
4957 EXPORT_SYMBOL(io_schedule
);
4959 long __sched
io_schedule_timeout(long timeout
)
4961 struct rq
*rq
= raw_rq();
4964 delayacct_blkio_start();
4965 atomic_inc(&rq
->nr_iowait
);
4966 current
->in_iowait
= 1;
4967 ret
= schedule_timeout(timeout
);
4968 current
->in_iowait
= 0;
4969 atomic_dec(&rq
->nr_iowait
);
4970 delayacct_blkio_end();
4975 * sys_sched_get_priority_max - return maximum RT priority.
4976 * @policy: scheduling class.
4978 * this syscall returns the maximum rt_priority that can be used
4979 * by a given scheduling class.
4981 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
4988 ret
= MAX_USER_RT_PRIO
-1;
5000 * sys_sched_get_priority_min - return minimum RT priority.
5001 * @policy: scheduling class.
5003 * this syscall returns the minimum rt_priority that can be used
5004 * by a given scheduling class.
5006 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
5024 * sys_sched_rr_get_interval - return the default timeslice of a process.
5025 * @pid: pid of the process.
5026 * @interval: userspace pointer to the timeslice value.
5028 * this syscall writes the default timeslice value of a given process
5029 * into the user-space timespec buffer. A value of '0' means infinity.
5031 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
5032 struct timespec __user
*, interval
)
5034 struct task_struct
*p
;
5035 unsigned int time_slice
;
5036 unsigned long flags
;
5046 p
= find_process_by_pid(pid
);
5050 retval
= security_task_getscheduler(p
);
5054 rq
= task_rq_lock(p
, &flags
);
5055 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
5056 task_rq_unlock(rq
, &flags
);
5059 jiffies_to_timespec(time_slice
, &t
);
5060 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5068 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5070 void sched_show_task(struct task_struct
*p
)
5072 unsigned long free
= 0;
5075 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5076 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5077 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5078 #if BITS_PER_LONG == 32
5079 if (state
== TASK_RUNNING
)
5080 printk(KERN_CONT
" running ");
5082 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5084 if (state
== TASK_RUNNING
)
5085 printk(KERN_CONT
" running task ");
5087 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5089 #ifdef CONFIG_DEBUG_STACK_USAGE
5090 free
= stack_not_used(p
);
5092 printk(KERN_CONT
"%5lu %5d %6d 0x%08lx\n", free
,
5093 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
5094 (unsigned long)task_thread_info(p
)->flags
);
5096 show_stack(p
, NULL
);
5099 void show_state_filter(unsigned long state_filter
)
5101 struct task_struct
*g
, *p
;
5103 #if BITS_PER_LONG == 32
5105 " task PC stack pid father\n");
5108 " task PC stack pid father\n");
5110 read_lock(&tasklist_lock
);
5111 do_each_thread(g
, p
) {
5113 * reset the NMI-timeout, listing all files on a slow
5114 * console might take alot of time:
5116 touch_nmi_watchdog();
5117 if (!state_filter
|| (p
->state
& state_filter
))
5119 } while_each_thread(g
, p
);
5121 touch_all_softlockup_watchdogs();
5123 #ifdef CONFIG_SCHED_DEBUG
5124 sysrq_sched_debug_show();
5126 read_unlock(&tasklist_lock
);
5128 * Only show locks if all tasks are dumped:
5131 debug_show_all_locks();
5134 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5136 idle
->sched_class
= &idle_sched_class
;
5140 * init_idle - set up an idle thread for a given CPU
5141 * @idle: task in question
5142 * @cpu: cpu the idle task belongs to
5144 * NOTE: this function does not set the idle thread's NEED_RESCHED
5145 * flag, to make booting more robust.
5147 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5149 struct rq
*rq
= cpu_rq(cpu
);
5150 unsigned long flags
;
5152 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5155 idle
->state
= TASK_RUNNING
;
5156 idle
->se
.exec_start
= sched_clock();
5158 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
5159 __set_task_cpu(idle
, cpu
);
5161 rq
->curr
= rq
->idle
= idle
;
5162 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5165 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5167 /* Set the preempt count _outside_ the spinlocks! */
5168 #if defined(CONFIG_PREEMPT)
5169 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5171 task_thread_info(idle
)->preempt_count
= 0;
5174 * The idle tasks have their own, simple scheduling class:
5176 idle
->sched_class
= &idle_sched_class
;
5177 ftrace_graph_init_task(idle
);
5181 * In a system that switches off the HZ timer nohz_cpu_mask
5182 * indicates which cpus entered this state. This is used
5183 * in the rcu update to wait only for active cpus. For system
5184 * which do not switch off the HZ timer nohz_cpu_mask should
5185 * always be CPU_BITS_NONE.
5187 cpumask_var_t nohz_cpu_mask
;
5190 * Increase the granularity value when there are more CPUs,
5191 * because with more CPUs the 'effective latency' as visible
5192 * to users decreases. But the relationship is not linear,
5193 * so pick a second-best guess by going with the log2 of the
5196 * This idea comes from the SD scheduler of Con Kolivas:
5198 static int get_update_sysctl_factor(void)
5200 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
5201 unsigned int factor
;
5203 switch (sysctl_sched_tunable_scaling
) {
5204 case SCHED_TUNABLESCALING_NONE
:
5207 case SCHED_TUNABLESCALING_LINEAR
:
5210 case SCHED_TUNABLESCALING_LOG
:
5212 factor
= 1 + ilog2(cpus
);
5219 static void update_sysctl(void)
5221 unsigned int factor
= get_update_sysctl_factor();
5223 #define SET_SYSCTL(name) \
5224 (sysctl_##name = (factor) * normalized_sysctl_##name)
5225 SET_SYSCTL(sched_min_granularity
);
5226 SET_SYSCTL(sched_latency
);
5227 SET_SYSCTL(sched_wakeup_granularity
);
5228 SET_SYSCTL(sched_shares_ratelimit
);
5232 static inline void sched_init_granularity(void)
5239 * This is how migration works:
5241 * 1) we invoke migration_cpu_stop() on the target CPU using
5243 * 2) stopper starts to run (implicitly forcing the migrated thread
5245 * 3) it checks whether the migrated task is still in the wrong runqueue.
5246 * 4) if it's in the wrong runqueue then the migration thread removes
5247 * it and puts it into the right queue.
5248 * 5) stopper completes and stop_one_cpu() returns and the migration
5253 * Change a given task's CPU affinity. Migrate the thread to a
5254 * proper CPU and schedule it away if the CPU it's executing on
5255 * is removed from the allowed bitmask.
5257 * NOTE: the caller must have a valid reference to the task, the
5258 * task must not exit() & deallocate itself prematurely. The
5259 * call is not atomic; no spinlocks may be held.
5261 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
5263 unsigned long flags
;
5265 unsigned int dest_cpu
;
5269 * Serialize against TASK_WAKING so that ttwu() and wunt() can
5270 * drop the rq->lock and still rely on ->cpus_allowed.
5273 while (task_is_waking(p
))
5275 rq
= task_rq_lock(p
, &flags
);
5276 if (task_is_waking(p
)) {
5277 task_rq_unlock(rq
, &flags
);
5281 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
5286 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5287 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
5292 if (p
->sched_class
->set_cpus_allowed
)
5293 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5295 cpumask_copy(&p
->cpus_allowed
, new_mask
);
5296 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
5299 /* Can the task run on the task's current CPU? If so, we're done */
5300 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
5303 dest_cpu
= cpumask_any_and(cpu_active_mask
, new_mask
);
5304 if (migrate_task(p
, dest_cpu
)) {
5305 struct migration_arg arg
= { p
, dest_cpu
};
5306 /* Need help from migration thread: drop lock and wait. */
5307 task_rq_unlock(rq
, &flags
);
5308 stop_one_cpu(cpu_of(rq
), migration_cpu_stop
, &arg
);
5309 tlb_migrate_finish(p
->mm
);
5313 task_rq_unlock(rq
, &flags
);
5317 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5320 * Move (not current) task off this cpu, onto dest cpu. We're doing
5321 * this because either it can't run here any more (set_cpus_allowed()
5322 * away from this CPU, or CPU going down), or because we're
5323 * attempting to rebalance this task on exec (sched_exec).
5325 * So we race with normal scheduler movements, but that's OK, as long
5326 * as the task is no longer on this CPU.
5328 * Returns non-zero if task was successfully migrated.
5330 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5332 struct rq
*rq_dest
, *rq_src
;
5335 if (unlikely(!cpu_active(dest_cpu
)))
5338 rq_src
= cpu_rq(src_cpu
);
5339 rq_dest
= cpu_rq(dest_cpu
);
5341 double_rq_lock(rq_src
, rq_dest
);
5342 /* Already moved. */
5343 if (task_cpu(p
) != src_cpu
)
5345 /* Affinity changed (again). */
5346 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
5350 * If we're not on a rq, the next wake-up will ensure we're
5354 deactivate_task(rq_src
, p
, 0);
5355 set_task_cpu(p
, dest_cpu
);
5356 activate_task(rq_dest
, p
, 0);
5357 check_preempt_curr(rq_dest
, p
, 0);
5362 double_rq_unlock(rq_src
, rq_dest
);
5367 * migration_cpu_stop - this will be executed by a highprio stopper thread
5368 * and performs thread migration by bumping thread off CPU then
5369 * 'pushing' onto another runqueue.
5371 static int migration_cpu_stop(void *data
)
5373 struct migration_arg
*arg
= data
;
5376 * The original target cpu might have gone down and we might
5377 * be on another cpu but it doesn't matter.
5379 local_irq_disable();
5380 __migrate_task(arg
->task
, raw_smp_processor_id(), arg
->dest_cpu
);
5385 #ifdef CONFIG_HOTPLUG_CPU
5387 * Figure out where task on dead CPU should go, use force if necessary.
5389 void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5391 struct rq
*rq
= cpu_rq(dead_cpu
);
5392 int needs_cpu
, uninitialized_var(dest_cpu
);
5393 unsigned long flags
;
5395 local_irq_save(flags
);
5397 raw_spin_lock(&rq
->lock
);
5398 needs_cpu
= (task_cpu(p
) == dead_cpu
) && (p
->state
!= TASK_WAKING
);
5400 dest_cpu
= select_fallback_rq(dead_cpu
, p
);
5401 raw_spin_unlock(&rq
->lock
);
5403 * It can only fail if we race with set_cpus_allowed(),
5404 * in the racer should migrate the task anyway.
5407 __migrate_task(p
, dead_cpu
, dest_cpu
);
5408 local_irq_restore(flags
);
5412 * While a dead CPU has no uninterruptible tasks queued at this point,
5413 * it might still have a nonzero ->nr_uninterruptible counter, because
5414 * for performance reasons the counter is not stricly tracking tasks to
5415 * their home CPUs. So we just add the counter to another CPU's counter,
5416 * to keep the global sum constant after CPU-down:
5418 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5420 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
5421 unsigned long flags
;
5423 local_irq_save(flags
);
5424 double_rq_lock(rq_src
, rq_dest
);
5425 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5426 rq_src
->nr_uninterruptible
= 0;
5427 double_rq_unlock(rq_src
, rq_dest
);
5428 local_irq_restore(flags
);
5431 /* Run through task list and migrate tasks from the dead cpu. */
5432 static void migrate_live_tasks(int src_cpu
)
5434 struct task_struct
*p
, *t
;
5436 read_lock(&tasklist_lock
);
5438 do_each_thread(t
, p
) {
5442 if (task_cpu(p
) == src_cpu
)
5443 move_task_off_dead_cpu(src_cpu
, p
);
5444 } while_each_thread(t
, p
);
5446 read_unlock(&tasklist_lock
);
5450 * Schedules idle task to be the next runnable task on current CPU.
5451 * It does so by boosting its priority to highest possible.
5452 * Used by CPU offline code.
5454 void sched_idle_next(void)
5456 int this_cpu
= smp_processor_id();
5457 struct rq
*rq
= cpu_rq(this_cpu
);
5458 struct task_struct
*p
= rq
->idle
;
5459 unsigned long flags
;
5461 /* cpu has to be offline */
5462 BUG_ON(cpu_online(this_cpu
));
5465 * Strictly not necessary since rest of the CPUs are stopped by now
5466 * and interrupts disabled on the current cpu.
5468 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5470 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5472 activate_task(rq
, p
, 0);
5474 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5478 * Ensures that the idle task is using init_mm right before its cpu goes
5481 void idle_task_exit(void)
5483 struct mm_struct
*mm
= current
->active_mm
;
5485 BUG_ON(cpu_online(smp_processor_id()));
5488 switch_mm(mm
, &init_mm
, current
);
5492 /* called under rq->lock with disabled interrupts */
5493 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5495 struct rq
*rq
= cpu_rq(dead_cpu
);
5497 /* Must be exiting, otherwise would be on tasklist. */
5498 BUG_ON(!p
->exit_state
);
5500 /* Cannot have done final schedule yet: would have vanished. */
5501 BUG_ON(p
->state
== TASK_DEAD
);
5506 * Drop lock around migration; if someone else moves it,
5507 * that's OK. No task can be added to this CPU, so iteration is
5510 raw_spin_unlock_irq(&rq
->lock
);
5511 move_task_off_dead_cpu(dead_cpu
, p
);
5512 raw_spin_lock_irq(&rq
->lock
);
5517 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5518 static void migrate_dead_tasks(unsigned int dead_cpu
)
5520 struct rq
*rq
= cpu_rq(dead_cpu
);
5521 struct task_struct
*next
;
5524 if (!rq
->nr_running
)
5526 next
= pick_next_task(rq
);
5529 next
->sched_class
->put_prev_task(rq
, next
);
5530 migrate_dead(dead_cpu
, next
);
5536 * remove the tasks which were accounted by rq from calc_load_tasks.
5538 static void calc_global_load_remove(struct rq
*rq
)
5540 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
5541 rq
->calc_load_active
= 0;
5543 #endif /* CONFIG_HOTPLUG_CPU */
5545 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5547 static struct ctl_table sd_ctl_dir
[] = {
5549 .procname
= "sched_domain",
5555 static struct ctl_table sd_ctl_root
[] = {
5557 .procname
= "kernel",
5559 .child
= sd_ctl_dir
,
5564 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5566 struct ctl_table
*entry
=
5567 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5572 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5574 struct ctl_table
*entry
;
5577 * In the intermediate directories, both the child directory and
5578 * procname are dynamically allocated and could fail but the mode
5579 * will always be set. In the lowest directory the names are
5580 * static strings and all have proc handlers.
5582 for (entry
= *tablep
; entry
->mode
; entry
++) {
5584 sd_free_ctl_entry(&entry
->child
);
5585 if (entry
->proc_handler
== NULL
)
5586 kfree(entry
->procname
);
5594 set_table_entry(struct ctl_table
*entry
,
5595 const char *procname
, void *data
, int maxlen
,
5596 mode_t mode
, proc_handler
*proc_handler
)
5598 entry
->procname
= procname
;
5600 entry
->maxlen
= maxlen
;
5602 entry
->proc_handler
= proc_handler
;
5605 static struct ctl_table
*
5606 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
5608 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
5613 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
5614 sizeof(long), 0644, proc_doulongvec_minmax
);
5615 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
5616 sizeof(long), 0644, proc_doulongvec_minmax
);
5617 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
5618 sizeof(int), 0644, proc_dointvec_minmax
);
5619 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
5620 sizeof(int), 0644, proc_dointvec_minmax
);
5621 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
5622 sizeof(int), 0644, proc_dointvec_minmax
);
5623 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
5624 sizeof(int), 0644, proc_dointvec_minmax
);
5625 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
5626 sizeof(int), 0644, proc_dointvec_minmax
);
5627 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
5628 sizeof(int), 0644, proc_dointvec_minmax
);
5629 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
5630 sizeof(int), 0644, proc_dointvec_minmax
);
5631 set_table_entry(&table
[9], "cache_nice_tries",
5632 &sd
->cache_nice_tries
,
5633 sizeof(int), 0644, proc_dointvec_minmax
);
5634 set_table_entry(&table
[10], "flags", &sd
->flags
,
5635 sizeof(int), 0644, proc_dointvec_minmax
);
5636 set_table_entry(&table
[11], "name", sd
->name
,
5637 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
5638 /* &table[12] is terminator */
5643 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
5645 struct ctl_table
*entry
, *table
;
5646 struct sched_domain
*sd
;
5647 int domain_num
= 0, i
;
5650 for_each_domain(cpu
, sd
)
5652 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
5657 for_each_domain(cpu
, sd
) {
5658 snprintf(buf
, 32, "domain%d", i
);
5659 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5661 entry
->child
= sd_alloc_ctl_domain_table(sd
);
5668 static struct ctl_table_header
*sd_sysctl_header
;
5669 static void register_sched_domain_sysctl(void)
5671 int i
, cpu_num
= num_possible_cpus();
5672 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
5675 WARN_ON(sd_ctl_dir
[0].child
);
5676 sd_ctl_dir
[0].child
= entry
;
5681 for_each_possible_cpu(i
) {
5682 snprintf(buf
, 32, "cpu%d", i
);
5683 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
5685 entry
->child
= sd_alloc_ctl_cpu_table(i
);
5689 WARN_ON(sd_sysctl_header
);
5690 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
5693 /* may be called multiple times per register */
5694 static void unregister_sched_domain_sysctl(void)
5696 if (sd_sysctl_header
)
5697 unregister_sysctl_table(sd_sysctl_header
);
5698 sd_sysctl_header
= NULL
;
5699 if (sd_ctl_dir
[0].child
)
5700 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
5703 static void register_sched_domain_sysctl(void)
5706 static void unregister_sched_domain_sysctl(void)
5711 static void set_rq_online(struct rq
*rq
)
5714 const struct sched_class
*class;
5716 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
5719 for_each_class(class) {
5720 if (class->rq_online
)
5721 class->rq_online(rq
);
5726 static void set_rq_offline(struct rq
*rq
)
5729 const struct sched_class
*class;
5731 for_each_class(class) {
5732 if (class->rq_offline
)
5733 class->rq_offline(rq
);
5736 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
5742 * migration_call - callback that gets triggered when a CPU is added.
5743 * Here we can start up the necessary migration thread for the new CPU.
5745 static int __cpuinit
5746 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5748 int cpu
= (long)hcpu
;
5749 unsigned long flags
;
5750 struct rq
*rq
= cpu_rq(cpu
);
5754 case CPU_UP_PREPARE
:
5755 case CPU_UP_PREPARE_FROZEN
:
5756 rq
->calc_load_update
= calc_load_update
;
5760 case CPU_ONLINE_FROZEN
:
5761 /* Update our root-domain */
5762 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5764 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5768 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5771 #ifdef CONFIG_HOTPLUG_CPU
5773 case CPU_DEAD_FROZEN
:
5774 migrate_live_tasks(cpu
);
5775 /* Idle task back to normal (off runqueue, low prio) */
5776 raw_spin_lock_irq(&rq
->lock
);
5777 deactivate_task(rq
, rq
->idle
, 0);
5778 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
5779 rq
->idle
->sched_class
= &idle_sched_class
;
5780 migrate_dead_tasks(cpu
);
5781 raw_spin_unlock_irq(&rq
->lock
);
5782 migrate_nr_uninterruptible(rq
);
5783 BUG_ON(rq
->nr_running
!= 0);
5784 calc_global_load_remove(rq
);
5788 case CPU_DYING_FROZEN
:
5789 /* Update our root-domain */
5790 raw_spin_lock_irqsave(&rq
->lock
, flags
);
5792 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
5795 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
5803 * Register at high priority so that task migration (migrate_all_tasks)
5804 * happens before everything else. This has to be lower priority than
5805 * the notifier in the perf_event subsystem, though.
5807 static struct notifier_block __cpuinitdata migration_notifier
= {
5808 .notifier_call
= migration_call
,
5812 static int __init
migration_init(void)
5814 void *cpu
= (void *)(long)smp_processor_id();
5817 /* Start one for the boot CPU: */
5818 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5819 BUG_ON(err
== NOTIFY_BAD
);
5820 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5821 register_cpu_notifier(&migration_notifier
);
5825 early_initcall(migration_init
);
5830 #ifdef CONFIG_SCHED_DEBUG
5832 static __read_mostly
int sched_domain_debug_enabled
;
5834 static int __init
sched_domain_debug_setup(char *str
)
5836 sched_domain_debug_enabled
= 1;
5840 early_param("sched_debug", sched_domain_debug_setup
);
5842 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
5843 struct cpumask
*groupmask
)
5845 struct sched_group
*group
= sd
->groups
;
5848 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
5849 cpumask_clear(groupmask
);
5851 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
5853 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5854 printk("does not load-balance\n");
5856 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5861 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
5863 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
5864 printk(KERN_ERR
"ERROR: domain->span does not contain "
5867 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
5868 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5872 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
5876 printk(KERN_ERR
"ERROR: group is NULL\n");
5880 if (!group
->cpu_power
) {
5881 printk(KERN_CONT
"\n");
5882 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5887 if (!cpumask_weight(sched_group_cpus(group
))) {
5888 printk(KERN_CONT
"\n");
5889 printk(KERN_ERR
"ERROR: empty group\n");
5893 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
5894 printk(KERN_CONT
"\n");
5895 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5899 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
5901 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
5903 printk(KERN_CONT
" %s", str
);
5904 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
5905 printk(KERN_CONT
" (cpu_power = %d)",
5909 group
= group
->next
;
5910 } while (group
!= sd
->groups
);
5911 printk(KERN_CONT
"\n");
5913 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
5914 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
5917 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
5918 printk(KERN_ERR
"ERROR: parent span is not a superset "
5919 "of domain->span\n");
5923 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5925 cpumask_var_t groupmask
;
5928 if (!sched_domain_debug_enabled
)
5932 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5936 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5938 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
5939 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
5944 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
5951 free_cpumask_var(groupmask
);
5953 #else /* !CONFIG_SCHED_DEBUG */
5954 # define sched_domain_debug(sd, cpu) do { } while (0)
5955 #endif /* CONFIG_SCHED_DEBUG */
5957 static int sd_degenerate(struct sched_domain
*sd
)
5959 if (cpumask_weight(sched_domain_span(sd
)) == 1)
5962 /* Following flags need at least 2 groups */
5963 if (sd
->flags
& (SD_LOAD_BALANCE
|
5964 SD_BALANCE_NEWIDLE
|
5968 SD_SHARE_PKG_RESOURCES
)) {
5969 if (sd
->groups
!= sd
->groups
->next
)
5973 /* Following flags don't use groups */
5974 if (sd
->flags
& (SD_WAKE_AFFINE
))
5981 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5983 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5985 if (sd_degenerate(parent
))
5988 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
5991 /* Flags needing groups don't count if only 1 group in parent */
5992 if (parent
->groups
== parent
->groups
->next
) {
5993 pflags
&= ~(SD_LOAD_BALANCE
|
5994 SD_BALANCE_NEWIDLE
|
5998 SD_SHARE_PKG_RESOURCES
);
5999 if (nr_node_ids
== 1)
6000 pflags
&= ~SD_SERIALIZE
;
6002 if (~cflags
& pflags
)
6008 static void free_rootdomain(struct root_domain
*rd
)
6010 synchronize_sched();
6012 cpupri_cleanup(&rd
->cpupri
);
6014 free_cpumask_var(rd
->rto_mask
);
6015 free_cpumask_var(rd
->online
);
6016 free_cpumask_var(rd
->span
);
6020 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6022 struct root_domain
*old_rd
= NULL
;
6023 unsigned long flags
;
6025 raw_spin_lock_irqsave(&rq
->lock
, flags
);
6030 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
6033 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
6036 * If we dont want to free the old_rt yet then
6037 * set old_rd to NULL to skip the freeing later
6040 if (!atomic_dec_and_test(&old_rd
->refcount
))
6044 atomic_inc(&rd
->refcount
);
6047 cpumask_set_cpu(rq
->cpu
, rd
->span
);
6048 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
6051 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
6054 free_rootdomain(old_rd
);
6057 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
6059 gfp_t gfp
= GFP_KERNEL
;
6061 memset(rd
, 0, sizeof(*rd
));
6066 if (!alloc_cpumask_var(&rd
->span
, gfp
))
6068 if (!alloc_cpumask_var(&rd
->online
, gfp
))
6070 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
6073 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
6078 free_cpumask_var(rd
->rto_mask
);
6080 free_cpumask_var(rd
->online
);
6082 free_cpumask_var(rd
->span
);
6087 static void init_defrootdomain(void)
6089 init_rootdomain(&def_root_domain
, true);
6091 atomic_set(&def_root_domain
.refcount
, 1);
6094 static struct root_domain
*alloc_rootdomain(void)
6096 struct root_domain
*rd
;
6098 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6102 if (init_rootdomain(rd
, false) != 0) {
6111 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6112 * hold the hotplug lock.
6115 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6117 struct rq
*rq
= cpu_rq(cpu
);
6118 struct sched_domain
*tmp
;
6120 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
)
6121 tmp
->span_weight
= cpumask_weight(sched_domain_span(tmp
));
6123 /* Remove the sched domains which do not contribute to scheduling. */
6124 for (tmp
= sd
; tmp
; ) {
6125 struct sched_domain
*parent
= tmp
->parent
;
6129 if (sd_parent_degenerate(tmp
, parent
)) {
6130 tmp
->parent
= parent
->parent
;
6132 parent
->parent
->child
= tmp
;
6137 if (sd
&& sd_degenerate(sd
)) {
6143 sched_domain_debug(sd
, cpu
);
6145 rq_attach_root(rq
, rd
);
6146 rcu_assign_pointer(rq
->sd
, sd
);
6149 /* cpus with isolated domains */
6150 static cpumask_var_t cpu_isolated_map
;
6152 /* Setup the mask of cpus configured for isolated domains */
6153 static int __init
isolated_cpu_setup(char *str
)
6155 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
6156 cpulist_parse(str
, cpu_isolated_map
);
6160 __setup("isolcpus=", isolated_cpu_setup
);
6163 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6164 * to a function which identifies what group(along with sched group) a CPU
6165 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
6166 * (due to the fact that we keep track of groups covered with a struct cpumask).
6168 * init_sched_build_groups will build a circular linked list of the groups
6169 * covered by the given span, and will set each group's ->cpumask correctly,
6170 * and ->cpu_power to 0.
6173 init_sched_build_groups(const struct cpumask
*span
,
6174 const struct cpumask
*cpu_map
,
6175 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
6176 struct sched_group
**sg
,
6177 struct cpumask
*tmpmask
),
6178 struct cpumask
*covered
, struct cpumask
*tmpmask
)
6180 struct sched_group
*first
= NULL
, *last
= NULL
;
6183 cpumask_clear(covered
);
6185 for_each_cpu(i
, span
) {
6186 struct sched_group
*sg
;
6187 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6190 if (cpumask_test_cpu(i
, covered
))
6193 cpumask_clear(sched_group_cpus(sg
));
6196 for_each_cpu(j
, span
) {
6197 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6200 cpumask_set_cpu(j
, covered
);
6201 cpumask_set_cpu(j
, sched_group_cpus(sg
));
6212 #define SD_NODES_PER_DOMAIN 16
6217 * find_next_best_node - find the next node to include in a sched_domain
6218 * @node: node whose sched_domain we're building
6219 * @used_nodes: nodes already in the sched_domain
6221 * Find the next node to include in a given scheduling domain. Simply
6222 * finds the closest node not already in the @used_nodes map.
6224 * Should use nodemask_t.
6226 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6228 int i
, n
, val
, min_val
, best_node
= 0;
6232 for (i
= 0; i
< nr_node_ids
; i
++) {
6233 /* Start at @node */
6234 n
= (node
+ i
) % nr_node_ids
;
6236 if (!nr_cpus_node(n
))
6239 /* Skip already used nodes */
6240 if (node_isset(n
, *used_nodes
))
6243 /* Simple min distance search */
6244 val
= node_distance(node
, n
);
6246 if (val
< min_val
) {
6252 node_set(best_node
, *used_nodes
);
6257 * sched_domain_node_span - get a cpumask for a node's sched_domain
6258 * @node: node whose cpumask we're constructing
6259 * @span: resulting cpumask
6261 * Given a node, construct a good cpumask for its sched_domain to span. It
6262 * should be one that prevents unnecessary balancing, but also spreads tasks
6265 static void sched_domain_node_span(int node
, struct cpumask
*span
)
6267 nodemask_t used_nodes
;
6270 cpumask_clear(span
);
6271 nodes_clear(used_nodes
);
6273 cpumask_or(span
, span
, cpumask_of_node(node
));
6274 node_set(node
, used_nodes
);
6276 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6277 int next_node
= find_next_best_node(node
, &used_nodes
);
6279 cpumask_or(span
, span
, cpumask_of_node(next_node
));
6282 #endif /* CONFIG_NUMA */
6284 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6287 * The cpus mask in sched_group and sched_domain hangs off the end.
6289 * ( See the the comments in include/linux/sched.h:struct sched_group
6290 * and struct sched_domain. )
6292 struct static_sched_group
{
6293 struct sched_group sg
;
6294 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
6297 struct static_sched_domain
{
6298 struct sched_domain sd
;
6299 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
6305 cpumask_var_t domainspan
;
6306 cpumask_var_t covered
;
6307 cpumask_var_t notcovered
;
6309 cpumask_var_t nodemask
;
6310 cpumask_var_t this_sibling_map
;
6311 cpumask_var_t this_core_map
;
6312 cpumask_var_t send_covered
;
6313 cpumask_var_t tmpmask
;
6314 struct sched_group
**sched_group_nodes
;
6315 struct root_domain
*rd
;
6319 sa_sched_groups
= 0,
6324 sa_this_sibling_map
,
6326 sa_sched_group_nodes
,
6336 * SMT sched-domains:
6338 #ifdef CONFIG_SCHED_SMT
6339 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
6340 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
6343 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
6344 struct sched_group
**sg
, struct cpumask
*unused
)
6347 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
6350 #endif /* CONFIG_SCHED_SMT */
6353 * multi-core sched-domains:
6355 #ifdef CONFIG_SCHED_MC
6356 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
6357 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
6358 #endif /* CONFIG_SCHED_MC */
6360 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6362 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6363 struct sched_group
**sg
, struct cpumask
*mask
)
6367 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6368 group
= cpumask_first(mask
);
6370 *sg
= &per_cpu(sched_group_core
, group
).sg
;
6373 #elif defined(CONFIG_SCHED_MC)
6375 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
6376 struct sched_group
**sg
, struct cpumask
*unused
)
6379 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
6384 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
6385 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
6388 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
6389 struct sched_group
**sg
, struct cpumask
*mask
)
6392 #ifdef CONFIG_SCHED_MC
6393 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
6394 group
= cpumask_first(mask
);
6395 #elif defined(CONFIG_SCHED_SMT)
6396 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
6397 group
= cpumask_first(mask
);
6402 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
6408 * The init_sched_build_groups can't handle what we want to do with node
6409 * groups, so roll our own. Now each node has its own list of groups which
6410 * gets dynamically allocated.
6412 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
6413 static struct sched_group
***sched_group_nodes_bycpu
;
6415 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
6416 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
6418 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
6419 struct sched_group
**sg
,
6420 struct cpumask
*nodemask
)
6424 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
6425 group
= cpumask_first(nodemask
);
6428 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
6432 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6434 struct sched_group
*sg
= group_head
;
6440 for_each_cpu(j
, sched_group_cpus(sg
)) {
6441 struct sched_domain
*sd
;
6443 sd
= &per_cpu(phys_domains
, j
).sd
;
6444 if (j
!= group_first_cpu(sd
->groups
)) {
6446 * Only add "power" once for each
6452 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6455 } while (sg
!= group_head
);
6458 static int build_numa_sched_groups(struct s_data
*d
,
6459 const struct cpumask
*cpu_map
, int num
)
6461 struct sched_domain
*sd
;
6462 struct sched_group
*sg
, *prev
;
6465 cpumask_clear(d
->covered
);
6466 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
6467 if (cpumask_empty(d
->nodemask
)) {
6468 d
->sched_group_nodes
[num
] = NULL
;
6472 sched_domain_node_span(num
, d
->domainspan
);
6473 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
6475 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6478 printk(KERN_WARNING
"Can not alloc domain group for node %d\n",
6482 d
->sched_group_nodes
[num
] = sg
;
6484 for_each_cpu(j
, d
->nodemask
) {
6485 sd
= &per_cpu(node_domains
, j
).sd
;
6490 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
6492 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
6495 for (j
= 0; j
< nr_node_ids
; j
++) {
6496 n
= (num
+ j
) % nr_node_ids
;
6497 cpumask_complement(d
->notcovered
, d
->covered
);
6498 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
6499 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
6500 if (cpumask_empty(d
->tmpmask
))
6502 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
6503 if (cpumask_empty(d
->tmpmask
))
6505 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
6509 "Can not alloc domain group for node %d\n", j
);
6513 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
6514 sg
->next
= prev
->next
;
6515 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
6522 #endif /* CONFIG_NUMA */
6525 /* Free memory allocated for various sched_group structures */
6526 static void free_sched_groups(const struct cpumask
*cpu_map
,
6527 struct cpumask
*nodemask
)
6531 for_each_cpu(cpu
, cpu_map
) {
6532 struct sched_group
**sched_group_nodes
6533 = sched_group_nodes_bycpu
[cpu
];
6535 if (!sched_group_nodes
)
6538 for (i
= 0; i
< nr_node_ids
; i
++) {
6539 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6541 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
6542 if (cpumask_empty(nodemask
))
6552 if (oldsg
!= sched_group_nodes
[i
])
6555 kfree(sched_group_nodes
);
6556 sched_group_nodes_bycpu
[cpu
] = NULL
;
6559 #else /* !CONFIG_NUMA */
6560 static void free_sched_groups(const struct cpumask
*cpu_map
,
6561 struct cpumask
*nodemask
)
6564 #endif /* CONFIG_NUMA */
6567 * Initialize sched groups cpu_power.
6569 * cpu_power indicates the capacity of sched group, which is used while
6570 * distributing the load between different sched groups in a sched domain.
6571 * Typically cpu_power for all the groups in a sched domain will be same unless
6572 * there are asymmetries in the topology. If there are asymmetries, group
6573 * having more cpu_power will pickup more load compared to the group having
6576 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6578 struct sched_domain
*child
;
6579 struct sched_group
*group
;
6583 WARN_ON(!sd
|| !sd
->groups
);
6585 if (cpu
!= group_first_cpu(sd
->groups
))
6590 sd
->groups
->cpu_power
= 0;
6593 power
= SCHED_LOAD_SCALE
;
6594 weight
= cpumask_weight(sched_domain_span(sd
));
6596 * SMT siblings share the power of a single core.
6597 * Usually multiple threads get a better yield out of
6598 * that one core than a single thread would have,
6599 * reflect that in sd->smt_gain.
6601 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
6602 power
*= sd
->smt_gain
;
6604 power
>>= SCHED_LOAD_SHIFT
;
6606 sd
->groups
->cpu_power
+= power
;
6611 * Add cpu_power of each child group to this groups cpu_power.
6613 group
= child
->groups
;
6615 sd
->groups
->cpu_power
+= group
->cpu_power
;
6616 group
= group
->next
;
6617 } while (group
!= child
->groups
);
6621 * Initializers for schedule domains
6622 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6625 #ifdef CONFIG_SCHED_DEBUG
6626 # define SD_INIT_NAME(sd, type) sd->name = #type
6628 # define SD_INIT_NAME(sd, type) do { } while (0)
6631 #define SD_INIT(sd, type) sd_init_##type(sd)
6633 #define SD_INIT_FUNC(type) \
6634 static noinline void sd_init_##type(struct sched_domain *sd) \
6636 memset(sd, 0, sizeof(*sd)); \
6637 *sd = SD_##type##_INIT; \
6638 sd->level = SD_LV_##type; \
6639 SD_INIT_NAME(sd, type); \
6644 SD_INIT_FUNC(ALLNODES
)
6647 #ifdef CONFIG_SCHED_SMT
6648 SD_INIT_FUNC(SIBLING
)
6650 #ifdef CONFIG_SCHED_MC
6654 static int default_relax_domain_level
= -1;
6656 static int __init
setup_relax_domain_level(char *str
)
6660 val
= simple_strtoul(str
, NULL
, 0);
6661 if (val
< SD_LV_MAX
)
6662 default_relax_domain_level
= val
;
6666 __setup("relax_domain_level=", setup_relax_domain_level
);
6668 static void set_domain_attribute(struct sched_domain
*sd
,
6669 struct sched_domain_attr
*attr
)
6673 if (!attr
|| attr
->relax_domain_level
< 0) {
6674 if (default_relax_domain_level
< 0)
6677 request
= default_relax_domain_level
;
6679 request
= attr
->relax_domain_level
;
6680 if (request
< sd
->level
) {
6681 /* turn off idle balance on this domain */
6682 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6684 /* turn on idle balance on this domain */
6685 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
6689 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
6690 const struct cpumask
*cpu_map
)
6693 case sa_sched_groups
:
6694 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
6695 d
->sched_group_nodes
= NULL
;
6697 free_rootdomain(d
->rd
); /* fall through */
6699 free_cpumask_var(d
->tmpmask
); /* fall through */
6700 case sa_send_covered
:
6701 free_cpumask_var(d
->send_covered
); /* fall through */
6702 case sa_this_core_map
:
6703 free_cpumask_var(d
->this_core_map
); /* fall through */
6704 case sa_this_sibling_map
:
6705 free_cpumask_var(d
->this_sibling_map
); /* fall through */
6707 free_cpumask_var(d
->nodemask
); /* fall through */
6708 case sa_sched_group_nodes
:
6710 kfree(d
->sched_group_nodes
); /* fall through */
6712 free_cpumask_var(d
->notcovered
); /* fall through */
6714 free_cpumask_var(d
->covered
); /* fall through */
6716 free_cpumask_var(d
->domainspan
); /* fall through */
6723 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
6724 const struct cpumask
*cpu_map
)
6727 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
6729 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
6730 return sa_domainspan
;
6731 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
6733 /* Allocate the per-node list of sched groups */
6734 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
6735 sizeof(struct sched_group
*), GFP_KERNEL
);
6736 if (!d
->sched_group_nodes
) {
6737 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6738 return sa_notcovered
;
6740 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
6742 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
6743 return sa_sched_group_nodes
;
6744 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
6746 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
6747 return sa_this_sibling_map
;
6748 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
6749 return sa_this_core_map
;
6750 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
6751 return sa_send_covered
;
6752 d
->rd
= alloc_rootdomain();
6754 printk(KERN_WARNING
"Cannot alloc root domain\n");
6757 return sa_rootdomain
;
6760 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
6761 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
6763 struct sched_domain
*sd
= NULL
;
6765 struct sched_domain
*parent
;
6768 if (cpumask_weight(cpu_map
) >
6769 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
6770 sd
= &per_cpu(allnodes_domains
, i
).sd
;
6771 SD_INIT(sd
, ALLNODES
);
6772 set_domain_attribute(sd
, attr
);
6773 cpumask_copy(sched_domain_span(sd
), cpu_map
);
6774 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6779 sd
= &per_cpu(node_domains
, i
).sd
;
6781 set_domain_attribute(sd
, attr
);
6782 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
6783 sd
->parent
= parent
;
6786 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
6791 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
6792 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6793 struct sched_domain
*parent
, int i
)
6795 struct sched_domain
*sd
;
6796 sd
= &per_cpu(phys_domains
, i
).sd
;
6798 set_domain_attribute(sd
, attr
);
6799 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
6800 sd
->parent
= parent
;
6803 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6807 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
6808 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6809 struct sched_domain
*parent
, int i
)
6811 struct sched_domain
*sd
= parent
;
6812 #ifdef CONFIG_SCHED_MC
6813 sd
= &per_cpu(core_domains
, i
).sd
;
6815 set_domain_attribute(sd
, attr
);
6816 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
6817 sd
->parent
= parent
;
6819 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6824 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
6825 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
6826 struct sched_domain
*parent
, int i
)
6828 struct sched_domain
*sd
= parent
;
6829 #ifdef CONFIG_SCHED_SMT
6830 sd
= &per_cpu(cpu_domains
, i
).sd
;
6831 SD_INIT(sd
, SIBLING
);
6832 set_domain_attribute(sd
, attr
);
6833 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
6834 sd
->parent
= parent
;
6836 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
6841 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
6842 const struct cpumask
*cpu_map
, int cpu
)
6845 #ifdef CONFIG_SCHED_SMT
6846 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
6847 cpumask_and(d
->this_sibling_map
, cpu_map
,
6848 topology_thread_cpumask(cpu
));
6849 if (cpu
== cpumask_first(d
->this_sibling_map
))
6850 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
6852 d
->send_covered
, d
->tmpmask
);
6855 #ifdef CONFIG_SCHED_MC
6856 case SD_LV_MC
: /* set up multi-core groups */
6857 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
6858 if (cpu
== cpumask_first(d
->this_core_map
))
6859 init_sched_build_groups(d
->this_core_map
, cpu_map
,
6861 d
->send_covered
, d
->tmpmask
);
6864 case SD_LV_CPU
: /* set up physical groups */
6865 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
6866 if (!cpumask_empty(d
->nodemask
))
6867 init_sched_build_groups(d
->nodemask
, cpu_map
,
6869 d
->send_covered
, d
->tmpmask
);
6872 case SD_LV_ALLNODES
:
6873 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
6874 d
->send_covered
, d
->tmpmask
);
6883 * Build sched domains for a given set of cpus and attach the sched domains
6884 * to the individual cpus
6886 static int __build_sched_domains(const struct cpumask
*cpu_map
,
6887 struct sched_domain_attr
*attr
)
6889 enum s_alloc alloc_state
= sa_none
;
6891 struct sched_domain
*sd
;
6897 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
6898 if (alloc_state
!= sa_rootdomain
)
6900 alloc_state
= sa_sched_groups
;
6903 * Set up domains for cpus specified by the cpu_map.
6905 for_each_cpu(i
, cpu_map
) {
6906 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
6909 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
6910 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6911 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6912 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
6915 for_each_cpu(i
, cpu_map
) {
6916 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
6917 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
6920 /* Set up physical groups */
6921 for (i
= 0; i
< nr_node_ids
; i
++)
6922 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
6925 /* Set up node groups */
6927 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
6929 for (i
= 0; i
< nr_node_ids
; i
++)
6930 if (build_numa_sched_groups(&d
, cpu_map
, i
))
6934 /* Calculate CPU power for physical packages and nodes */
6935 #ifdef CONFIG_SCHED_SMT
6936 for_each_cpu(i
, cpu_map
) {
6937 sd
= &per_cpu(cpu_domains
, i
).sd
;
6938 init_sched_groups_power(i
, sd
);
6941 #ifdef CONFIG_SCHED_MC
6942 for_each_cpu(i
, cpu_map
) {
6943 sd
= &per_cpu(core_domains
, i
).sd
;
6944 init_sched_groups_power(i
, sd
);
6948 for_each_cpu(i
, cpu_map
) {
6949 sd
= &per_cpu(phys_domains
, i
).sd
;
6950 init_sched_groups_power(i
, sd
);
6954 for (i
= 0; i
< nr_node_ids
; i
++)
6955 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
6957 if (d
.sd_allnodes
) {
6958 struct sched_group
*sg
;
6960 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
6962 init_numa_sched_groups_power(sg
);
6966 /* Attach the domains */
6967 for_each_cpu(i
, cpu_map
) {
6968 #ifdef CONFIG_SCHED_SMT
6969 sd
= &per_cpu(cpu_domains
, i
).sd
;
6970 #elif defined(CONFIG_SCHED_MC)
6971 sd
= &per_cpu(core_domains
, i
).sd
;
6973 sd
= &per_cpu(phys_domains
, i
).sd
;
6975 cpu_attach_domain(sd
, d
.rd
, i
);
6978 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
6979 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
6983 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
6987 static int build_sched_domains(const struct cpumask
*cpu_map
)
6989 return __build_sched_domains(cpu_map
, NULL
);
6992 static cpumask_var_t
*doms_cur
; /* current sched domains */
6993 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
6994 static struct sched_domain_attr
*dattr_cur
;
6995 /* attribues of custom domains in 'doms_cur' */
6998 * Special case: If a kmalloc of a doms_cur partition (array of
6999 * cpumask) fails, then fallback to a single sched domain,
7000 * as determined by the single cpumask fallback_doms.
7002 static cpumask_var_t fallback_doms
;
7005 * arch_update_cpu_topology lets virtualized architectures update the
7006 * cpu core maps. It is supposed to return 1 if the topology changed
7007 * or 0 if it stayed the same.
7009 int __attribute__((weak
)) arch_update_cpu_topology(void)
7014 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
7017 cpumask_var_t
*doms
;
7019 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
7022 for (i
= 0; i
< ndoms
; i
++) {
7023 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
7024 free_sched_domains(doms
, i
);
7031 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
7034 for (i
= 0; i
< ndoms
; i
++)
7035 free_cpumask_var(doms
[i
]);
7040 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7041 * For now this just excludes isolated cpus, but could be used to
7042 * exclude other special cases in the future.
7044 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
7048 arch_update_cpu_topology();
7050 doms_cur
= alloc_sched_domains(ndoms_cur
);
7052 doms_cur
= &fallback_doms
;
7053 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
7055 err
= build_sched_domains(doms_cur
[0]);
7056 register_sched_domain_sysctl();
7061 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
7062 struct cpumask
*tmpmask
)
7064 free_sched_groups(cpu_map
, tmpmask
);
7068 * Detach sched domains from a group of cpus specified in cpu_map
7069 * These cpus will now be attached to the NULL domain
7071 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
7073 /* Save because hotplug lock held. */
7074 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
7077 for_each_cpu(i
, cpu_map
)
7078 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7079 synchronize_sched();
7080 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
7083 /* handle null as "default" */
7084 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7085 struct sched_domain_attr
*new, int idx_new
)
7087 struct sched_domain_attr tmp
;
7094 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7095 new ? (new + idx_new
) : &tmp
,
7096 sizeof(struct sched_domain_attr
));
7100 * Partition sched domains as specified by the 'ndoms_new'
7101 * cpumasks in the array doms_new[] of cpumasks. This compares
7102 * doms_new[] to the current sched domain partitioning, doms_cur[].
7103 * It destroys each deleted domain and builds each new domain.
7105 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
7106 * The masks don't intersect (don't overlap.) We should setup one
7107 * sched domain for each mask. CPUs not in any of the cpumasks will
7108 * not be load balanced. If the same cpumask appears both in the
7109 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7112 * The passed in 'doms_new' should be allocated using
7113 * alloc_sched_domains. This routine takes ownership of it and will
7114 * free_sched_domains it when done with it. If the caller failed the
7115 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
7116 * and partition_sched_domains() will fallback to the single partition
7117 * 'fallback_doms', it also forces the domains to be rebuilt.
7119 * If doms_new == NULL it will be replaced with cpu_online_mask.
7120 * ndoms_new == 0 is a special case for destroying existing domains,
7121 * and it will not create the default domain.
7123 * Call with hotplug lock held
7125 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
7126 struct sched_domain_attr
*dattr_new
)
7131 mutex_lock(&sched_domains_mutex
);
7133 /* always unregister in case we don't destroy any domains */
7134 unregister_sched_domain_sysctl();
7136 /* Let architecture update cpu core mappings. */
7137 new_topology
= arch_update_cpu_topology();
7139 n
= doms_new
? ndoms_new
: 0;
7141 /* Destroy deleted domains */
7142 for (i
= 0; i
< ndoms_cur
; i
++) {
7143 for (j
= 0; j
< n
&& !new_topology
; j
++) {
7144 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
7145 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7148 /* no match - a current sched domain not in new doms_new[] */
7149 detach_destroy_domains(doms_cur
[i
]);
7154 if (doms_new
== NULL
) {
7156 doms_new
= &fallback_doms
;
7157 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
7158 WARN_ON_ONCE(dattr_new
);
7161 /* Build new domains */
7162 for (i
= 0; i
< ndoms_new
; i
++) {
7163 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
7164 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
7165 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7168 /* no match - add a new doms_new */
7169 __build_sched_domains(doms_new
[i
],
7170 dattr_new
? dattr_new
+ i
: NULL
);
7175 /* Remember the new sched domains */
7176 if (doms_cur
!= &fallback_doms
)
7177 free_sched_domains(doms_cur
, ndoms_cur
);
7178 kfree(dattr_cur
); /* kfree(NULL) is safe */
7179 doms_cur
= doms_new
;
7180 dattr_cur
= dattr_new
;
7181 ndoms_cur
= ndoms_new
;
7183 register_sched_domain_sysctl();
7185 mutex_unlock(&sched_domains_mutex
);
7188 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7189 static void arch_reinit_sched_domains(void)
7193 /* Destroy domains first to force the rebuild */
7194 partition_sched_domains(0, NULL
, NULL
);
7196 rebuild_sched_domains();
7200 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7202 unsigned int level
= 0;
7204 if (sscanf(buf
, "%u", &level
) != 1)
7208 * level is always be positive so don't check for
7209 * level < POWERSAVINGS_BALANCE_NONE which is 0
7210 * What happens on 0 or 1 byte write,
7211 * need to check for count as well?
7214 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
7218 sched_smt_power_savings
= level
;
7220 sched_mc_power_savings
= level
;
7222 arch_reinit_sched_domains();
7227 #ifdef CONFIG_SCHED_MC
7228 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7229 struct sysdev_class_attribute
*attr
,
7232 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7234 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7235 struct sysdev_class_attribute
*attr
,
7236 const char *buf
, size_t count
)
7238 return sched_power_savings_store(buf
, count
, 0);
7240 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7241 sched_mc_power_savings_show
,
7242 sched_mc_power_savings_store
);
7245 #ifdef CONFIG_SCHED_SMT
7246 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7247 struct sysdev_class_attribute
*attr
,
7250 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7252 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7253 struct sysdev_class_attribute
*attr
,
7254 const char *buf
, size_t count
)
7256 return sched_power_savings_store(buf
, count
, 1);
7258 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7259 sched_smt_power_savings_show
,
7260 sched_smt_power_savings_store
);
7263 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7267 #ifdef CONFIG_SCHED_SMT
7269 err
= sysfs_create_file(&cls
->kset
.kobj
,
7270 &attr_sched_smt_power_savings
.attr
);
7272 #ifdef CONFIG_SCHED_MC
7273 if (!err
&& mc_capable())
7274 err
= sysfs_create_file(&cls
->kset
.kobj
,
7275 &attr_sched_mc_power_savings
.attr
);
7279 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7281 #ifndef CONFIG_CPUSETS
7283 * Add online and remove offline CPUs from the scheduler domains.
7284 * When cpusets are enabled they take over this function.
7286 static int update_sched_domains(struct notifier_block
*nfb
,
7287 unsigned long action
, void *hcpu
)
7291 case CPU_ONLINE_FROZEN
:
7292 case CPU_DOWN_PREPARE
:
7293 case CPU_DOWN_PREPARE_FROZEN
:
7294 case CPU_DOWN_FAILED
:
7295 case CPU_DOWN_FAILED_FROZEN
:
7296 partition_sched_domains(1, NULL
, NULL
);
7305 static int update_runtime(struct notifier_block
*nfb
,
7306 unsigned long action
, void *hcpu
)
7308 int cpu
= (int)(long)hcpu
;
7311 case CPU_DOWN_PREPARE
:
7312 case CPU_DOWN_PREPARE_FROZEN
:
7313 disable_runtime(cpu_rq(cpu
));
7316 case CPU_DOWN_FAILED
:
7317 case CPU_DOWN_FAILED_FROZEN
:
7319 case CPU_ONLINE_FROZEN
:
7320 enable_runtime(cpu_rq(cpu
));
7328 void __init
sched_init_smp(void)
7330 cpumask_var_t non_isolated_cpus
;
7332 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
7333 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
7335 #if defined(CONFIG_NUMA)
7336 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7338 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7341 mutex_lock(&sched_domains_mutex
);
7342 arch_init_sched_domains(cpu_active_mask
);
7343 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
7344 if (cpumask_empty(non_isolated_cpus
))
7345 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
7346 mutex_unlock(&sched_domains_mutex
);
7349 #ifndef CONFIG_CPUSETS
7350 /* XXX: Theoretical race here - CPU may be hotplugged now */
7351 hotcpu_notifier(update_sched_domains
, 0);
7354 /* RT runtime code needs to handle some hotplug events */
7355 hotcpu_notifier(update_runtime
, 0);
7359 /* Move init over to a non-isolated CPU */
7360 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
7362 sched_init_granularity();
7363 free_cpumask_var(non_isolated_cpus
);
7365 init_sched_rt_class();
7368 void __init
sched_init_smp(void)
7370 sched_init_granularity();
7372 #endif /* CONFIG_SMP */
7374 const_debug
unsigned int sysctl_timer_migration
= 1;
7376 int in_sched_functions(unsigned long addr
)
7378 return in_lock_functions(addr
) ||
7379 (addr
>= (unsigned long)__sched_text_start
7380 && addr
< (unsigned long)__sched_text_end
);
7383 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7385 cfs_rq
->tasks_timeline
= RB_ROOT
;
7386 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7387 #ifdef CONFIG_FAIR_GROUP_SCHED
7390 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7393 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7395 struct rt_prio_array
*array
;
7398 array
= &rt_rq
->active
;
7399 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7400 INIT_LIST_HEAD(array
->queue
+ i
);
7401 __clear_bit(i
, array
->bitmap
);
7403 /* delimiter for bitsearch: */
7404 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7406 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7407 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
7409 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
7413 rt_rq
->rt_nr_migratory
= 0;
7414 rt_rq
->overloaded
= 0;
7415 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
7419 rt_rq
->rt_throttled
= 0;
7420 rt_rq
->rt_runtime
= 0;
7421 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
7423 #ifdef CONFIG_RT_GROUP_SCHED
7424 rt_rq
->rt_nr_boosted
= 0;
7429 #ifdef CONFIG_FAIR_GROUP_SCHED
7430 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7431 struct sched_entity
*se
, int cpu
, int add
,
7432 struct sched_entity
*parent
)
7434 struct rq
*rq
= cpu_rq(cpu
);
7435 tg
->cfs_rq
[cpu
] = cfs_rq
;
7436 init_cfs_rq(cfs_rq
, rq
);
7439 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7442 /* se could be NULL for init_task_group */
7447 se
->cfs_rq
= &rq
->cfs
;
7449 se
->cfs_rq
= parent
->my_q
;
7452 se
->load
.weight
= tg
->shares
;
7453 se
->load
.inv_weight
= 0;
7454 se
->parent
= parent
;
7458 #ifdef CONFIG_RT_GROUP_SCHED
7459 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7460 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7461 struct sched_rt_entity
*parent
)
7463 struct rq
*rq
= cpu_rq(cpu
);
7465 tg
->rt_rq
[cpu
] = rt_rq
;
7466 init_rt_rq(rt_rq
, rq
);
7468 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7470 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7472 tg
->rt_se
[cpu
] = rt_se
;
7477 rt_se
->rt_rq
= &rq
->rt
;
7479 rt_se
->rt_rq
= parent
->my_q
;
7481 rt_se
->my_q
= rt_rq
;
7482 rt_se
->parent
= parent
;
7483 INIT_LIST_HEAD(&rt_se
->run_list
);
7487 void __init
sched_init(void)
7490 unsigned long alloc_size
= 0, ptr
;
7492 #ifdef CONFIG_FAIR_GROUP_SCHED
7493 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7495 #ifdef CONFIG_RT_GROUP_SCHED
7496 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7498 #ifdef CONFIG_CPUMASK_OFFSTACK
7499 alloc_size
+= num_possible_cpus() * cpumask_size();
7502 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
7504 #ifdef CONFIG_FAIR_GROUP_SCHED
7505 init_task_group
.se
= (struct sched_entity
**)ptr
;
7506 ptr
+= nr_cpu_ids
* sizeof(void **);
7508 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7509 ptr
+= nr_cpu_ids
* sizeof(void **);
7511 #endif /* CONFIG_FAIR_GROUP_SCHED */
7512 #ifdef CONFIG_RT_GROUP_SCHED
7513 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7514 ptr
+= nr_cpu_ids
* sizeof(void **);
7516 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7517 ptr
+= nr_cpu_ids
* sizeof(void **);
7519 #endif /* CONFIG_RT_GROUP_SCHED */
7520 #ifdef CONFIG_CPUMASK_OFFSTACK
7521 for_each_possible_cpu(i
) {
7522 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
7523 ptr
+= cpumask_size();
7525 #endif /* CONFIG_CPUMASK_OFFSTACK */
7529 init_defrootdomain();
7532 init_rt_bandwidth(&def_rt_bandwidth
,
7533 global_rt_period(), global_rt_runtime());
7535 #ifdef CONFIG_RT_GROUP_SCHED
7536 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7537 global_rt_period(), global_rt_runtime());
7538 #endif /* CONFIG_RT_GROUP_SCHED */
7540 #ifdef CONFIG_CGROUP_SCHED
7541 list_add(&init_task_group
.list
, &task_groups
);
7542 INIT_LIST_HEAD(&init_task_group
.children
);
7544 #endif /* CONFIG_CGROUP_SCHED */
7546 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
7547 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
7548 __alignof__(unsigned long));
7550 for_each_possible_cpu(i
) {
7554 raw_spin_lock_init(&rq
->lock
);
7556 rq
->calc_load_active
= 0;
7557 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
7558 init_cfs_rq(&rq
->cfs
, rq
);
7559 init_rt_rq(&rq
->rt
, rq
);
7560 #ifdef CONFIG_FAIR_GROUP_SCHED
7561 init_task_group
.shares
= init_task_group_load
;
7562 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7563 #ifdef CONFIG_CGROUP_SCHED
7565 * How much cpu bandwidth does init_task_group get?
7567 * In case of task-groups formed thr' the cgroup filesystem, it
7568 * gets 100% of the cpu resources in the system. This overall
7569 * system cpu resource is divided among the tasks of
7570 * init_task_group and its child task-groups in a fair manner,
7571 * based on each entity's (task or task-group's) weight
7572 * (se->load.weight).
7574 * In other words, if init_task_group has 10 tasks of weight
7575 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7576 * then A0's share of the cpu resource is:
7578 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7580 * We achieve this by letting init_task_group's tasks sit
7581 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7583 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7585 #endif /* CONFIG_FAIR_GROUP_SCHED */
7587 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7588 #ifdef CONFIG_RT_GROUP_SCHED
7589 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7590 #ifdef CONFIG_CGROUP_SCHED
7591 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7595 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7596 rq
->cpu_load
[j
] = 0;
7600 rq
->cpu_power
= SCHED_LOAD_SCALE
;
7601 rq
->post_schedule
= 0;
7602 rq
->active_balance
= 0;
7603 rq
->next_balance
= jiffies
;
7608 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
7609 rq_attach_root(rq
, &def_root_domain
);
7612 atomic_set(&rq
->nr_iowait
, 0);
7615 set_load_weight(&init_task
);
7617 #ifdef CONFIG_PREEMPT_NOTIFIERS
7618 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7622 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
7625 #ifdef CONFIG_RT_MUTEXES
7626 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7630 * The boot idle thread does lazy MMU switching as well:
7632 atomic_inc(&init_mm
.mm_count
);
7633 enter_lazy_tlb(&init_mm
, current
);
7636 * Make us the idle thread. Technically, schedule() should not be
7637 * called from this thread, however somewhere below it might be,
7638 * but because we are the idle thread, we just pick up running again
7639 * when this runqueue becomes "idle".
7641 init_idle(current
, smp_processor_id());
7643 calc_load_update
= jiffies
+ LOAD_FREQ
;
7646 * During early bootup we pretend to be a normal task:
7648 current
->sched_class
= &fair_sched_class
;
7650 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
7651 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
7654 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
7655 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
7657 /* May be allocated at isolcpus cmdline parse time */
7658 if (cpu_isolated_map
== NULL
)
7659 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
7664 scheduler_running
= 1;
7667 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7668 static inline int preempt_count_equals(int preempt_offset
)
7670 int nested
= (preempt_count() & ~PREEMPT_ACTIVE
) + rcu_preempt_depth();
7672 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
7675 void __might_sleep(const char *file
, int line
, int preempt_offset
)
7678 static unsigned long prev_jiffy
; /* ratelimiting */
7680 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
7681 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
7683 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7685 prev_jiffy
= jiffies
;
7688 "BUG: sleeping function called from invalid context at %s:%d\n",
7691 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
7692 in_atomic(), irqs_disabled(),
7693 current
->pid
, current
->comm
);
7695 debug_show_held_locks(current
);
7696 if (irqs_disabled())
7697 print_irqtrace_events(current
);
7701 EXPORT_SYMBOL(__might_sleep
);
7704 #ifdef CONFIG_MAGIC_SYSRQ
7705 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7709 on_rq
= p
->se
.on_rq
;
7711 deactivate_task(rq
, p
, 0);
7712 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7714 activate_task(rq
, p
, 0);
7715 resched_task(rq
->curr
);
7719 void normalize_rt_tasks(void)
7721 struct task_struct
*g
, *p
;
7722 unsigned long flags
;
7725 read_lock_irqsave(&tasklist_lock
, flags
);
7726 do_each_thread(g
, p
) {
7728 * Only normalize user tasks:
7733 p
->se
.exec_start
= 0;
7734 #ifdef CONFIG_SCHEDSTATS
7735 p
->se
.statistics
.wait_start
= 0;
7736 p
->se
.statistics
.sleep_start
= 0;
7737 p
->se
.statistics
.block_start
= 0;
7742 * Renice negative nice level userspace
7745 if (TASK_NICE(p
) < 0 && p
->mm
)
7746 set_user_nice(p
, 0);
7750 raw_spin_lock(&p
->pi_lock
);
7751 rq
= __task_rq_lock(p
);
7753 normalize_task(rq
, p
);
7755 __task_rq_unlock(rq
);
7756 raw_spin_unlock(&p
->pi_lock
);
7757 } while_each_thread(g
, p
);
7759 read_unlock_irqrestore(&tasklist_lock
, flags
);
7762 #endif /* CONFIG_MAGIC_SYSRQ */
7764 #if defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB)
7766 * These functions are only useful for the IA64 MCA handling, or kdb.
7768 * They can only be called when the whole system has been
7769 * stopped - every CPU needs to be quiescent, and no scheduling
7770 * activity can take place. Using them for anything else would
7771 * be a serious bug, and as a result, they aren't even visible
7772 * under any other configuration.
7776 * curr_task - return the current task for a given cpu.
7777 * @cpu: the processor in question.
7779 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7781 struct task_struct
*curr_task(int cpu
)
7783 return cpu_curr(cpu
);
7786 #endif /* defined(CONFIG_IA64) || defined(CONFIG_KGDB_KDB) */
7790 * set_curr_task - set the current task for a given cpu.
7791 * @cpu: the processor in question.
7792 * @p: the task pointer to set.
7794 * Description: This function must only be used when non-maskable interrupts
7795 * are serviced on a separate stack. It allows the architecture to switch the
7796 * notion of the current task on a cpu in a non-blocking manner. This function
7797 * must be called with all CPU's synchronized, and interrupts disabled, the
7798 * and caller must save the original value of the current task (see
7799 * curr_task() above) and restore that value before reenabling interrupts and
7800 * re-starting the system.
7802 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7804 void set_curr_task(int cpu
, struct task_struct
*p
)
7811 #ifdef CONFIG_FAIR_GROUP_SCHED
7812 static void free_fair_sched_group(struct task_group
*tg
)
7816 for_each_possible_cpu(i
) {
7818 kfree(tg
->cfs_rq
[i
]);
7828 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7830 struct cfs_rq
*cfs_rq
;
7831 struct sched_entity
*se
;
7835 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7838 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
7842 tg
->shares
= NICE_0_LOAD
;
7844 for_each_possible_cpu(i
) {
7847 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
7848 GFP_KERNEL
, cpu_to_node(i
));
7852 se
= kzalloc_node(sizeof(struct sched_entity
),
7853 GFP_KERNEL
, cpu_to_node(i
));
7857 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
7868 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7870 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
7871 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
7874 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7876 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
7878 #else /* !CONFG_FAIR_GROUP_SCHED */
7879 static inline void free_fair_sched_group(struct task_group
*tg
)
7884 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7889 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
7893 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
7896 #endif /* CONFIG_FAIR_GROUP_SCHED */
7898 #ifdef CONFIG_RT_GROUP_SCHED
7899 static void free_rt_sched_group(struct task_group
*tg
)
7903 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
7905 for_each_possible_cpu(i
) {
7907 kfree(tg
->rt_rq
[i
]);
7909 kfree(tg
->rt_se
[i
]);
7917 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7919 struct rt_rq
*rt_rq
;
7920 struct sched_rt_entity
*rt_se
;
7924 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
7927 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
7931 init_rt_bandwidth(&tg
->rt_bandwidth
,
7932 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
7934 for_each_possible_cpu(i
) {
7937 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
7938 GFP_KERNEL
, cpu_to_node(i
));
7942 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
7943 GFP_KERNEL
, cpu_to_node(i
));
7947 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
7958 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7960 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
7961 &cpu_rq(cpu
)->leaf_rt_rq_list
);
7964 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7966 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
7968 #else /* !CONFIG_RT_GROUP_SCHED */
7969 static inline void free_rt_sched_group(struct task_group
*tg
)
7974 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
7979 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
7983 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
7986 #endif /* CONFIG_RT_GROUP_SCHED */
7988 #ifdef CONFIG_CGROUP_SCHED
7989 static void free_sched_group(struct task_group
*tg
)
7991 free_fair_sched_group(tg
);
7992 free_rt_sched_group(tg
);
7996 /* allocate runqueue etc for a new task group */
7997 struct task_group
*sched_create_group(struct task_group
*parent
)
7999 struct task_group
*tg
;
8000 unsigned long flags
;
8003 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8005 return ERR_PTR(-ENOMEM
);
8007 if (!alloc_fair_sched_group(tg
, parent
))
8010 if (!alloc_rt_sched_group(tg
, parent
))
8013 spin_lock_irqsave(&task_group_lock
, flags
);
8014 for_each_possible_cpu(i
) {
8015 register_fair_sched_group(tg
, i
);
8016 register_rt_sched_group(tg
, i
);
8018 list_add_rcu(&tg
->list
, &task_groups
);
8020 WARN_ON(!parent
); /* root should already exist */
8022 tg
->parent
= parent
;
8023 INIT_LIST_HEAD(&tg
->children
);
8024 list_add_rcu(&tg
->siblings
, &parent
->children
);
8025 spin_unlock_irqrestore(&task_group_lock
, flags
);
8030 free_sched_group(tg
);
8031 return ERR_PTR(-ENOMEM
);
8034 /* rcu callback to free various structures associated with a task group */
8035 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8037 /* now it should be safe to free those cfs_rqs */
8038 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8041 /* Destroy runqueue etc associated with a task group */
8042 void sched_destroy_group(struct task_group
*tg
)
8044 unsigned long flags
;
8047 spin_lock_irqsave(&task_group_lock
, flags
);
8048 for_each_possible_cpu(i
) {
8049 unregister_fair_sched_group(tg
, i
);
8050 unregister_rt_sched_group(tg
, i
);
8052 list_del_rcu(&tg
->list
);
8053 list_del_rcu(&tg
->siblings
);
8054 spin_unlock_irqrestore(&task_group_lock
, flags
);
8056 /* wait for possible concurrent references to cfs_rqs complete */
8057 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8060 /* change task's runqueue when it moves between groups.
8061 * The caller of this function should have put the task in its new group
8062 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8063 * reflect its new group.
8065 void sched_move_task(struct task_struct
*tsk
)
8068 unsigned long flags
;
8071 rq
= task_rq_lock(tsk
, &flags
);
8073 running
= task_current(rq
, tsk
);
8074 on_rq
= tsk
->se
.on_rq
;
8077 dequeue_task(rq
, tsk
, 0);
8078 if (unlikely(running
))
8079 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8081 set_task_rq(tsk
, task_cpu(tsk
));
8083 #ifdef CONFIG_FAIR_GROUP_SCHED
8084 if (tsk
->sched_class
->moved_group
)
8085 tsk
->sched_class
->moved_group(tsk
, on_rq
);
8088 if (unlikely(running
))
8089 tsk
->sched_class
->set_curr_task(rq
);
8091 enqueue_task(rq
, tsk
, 0);
8093 task_rq_unlock(rq
, &flags
);
8095 #endif /* CONFIG_CGROUP_SCHED */
8097 #ifdef CONFIG_FAIR_GROUP_SCHED
8098 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8100 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8105 dequeue_entity(cfs_rq
, se
, 0);
8107 se
->load
.weight
= shares
;
8108 se
->load
.inv_weight
= 0;
8111 enqueue_entity(cfs_rq
, se
, 0);
8114 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8116 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8117 struct rq
*rq
= cfs_rq
->rq
;
8118 unsigned long flags
;
8120 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8121 __set_se_shares(se
, shares
);
8122 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8125 static DEFINE_MUTEX(shares_mutex
);
8127 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8130 unsigned long flags
;
8133 * We can't change the weight of the root cgroup.
8138 if (shares
< MIN_SHARES
)
8139 shares
= MIN_SHARES
;
8140 else if (shares
> MAX_SHARES
)
8141 shares
= MAX_SHARES
;
8143 mutex_lock(&shares_mutex
);
8144 if (tg
->shares
== shares
)
8147 spin_lock_irqsave(&task_group_lock
, flags
);
8148 for_each_possible_cpu(i
)
8149 unregister_fair_sched_group(tg
, i
);
8150 list_del_rcu(&tg
->siblings
);
8151 spin_unlock_irqrestore(&task_group_lock
, flags
);
8153 /* wait for any ongoing reference to this group to finish */
8154 synchronize_sched();
8157 * Now we are free to modify the group's share on each cpu
8158 * w/o tripping rebalance_share or load_balance_fair.
8160 tg
->shares
= shares
;
8161 for_each_possible_cpu(i
) {
8165 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8166 set_se_shares(tg
->se
[i
], shares
);
8170 * Enable load balance activity on this group, by inserting it back on
8171 * each cpu's rq->leaf_cfs_rq_list.
8173 spin_lock_irqsave(&task_group_lock
, flags
);
8174 for_each_possible_cpu(i
)
8175 register_fair_sched_group(tg
, i
);
8176 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8177 spin_unlock_irqrestore(&task_group_lock
, flags
);
8179 mutex_unlock(&shares_mutex
);
8183 unsigned long sched_group_shares(struct task_group
*tg
)
8189 #ifdef CONFIG_RT_GROUP_SCHED
8191 * Ensure that the real time constraints are schedulable.
8193 static DEFINE_MUTEX(rt_constraints_mutex
);
8195 static unsigned long to_ratio(u64 period
, u64 runtime
)
8197 if (runtime
== RUNTIME_INF
)
8200 return div64_u64(runtime
<< 20, period
);
8203 /* Must be called with tasklist_lock held */
8204 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8206 struct task_struct
*g
, *p
;
8208 do_each_thread(g
, p
) {
8209 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8211 } while_each_thread(g
, p
);
8216 struct rt_schedulable_data
{
8217 struct task_group
*tg
;
8222 static int tg_schedulable(struct task_group
*tg
, void *data
)
8224 struct rt_schedulable_data
*d
= data
;
8225 struct task_group
*child
;
8226 unsigned long total
, sum
= 0;
8227 u64 period
, runtime
;
8229 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8230 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8233 period
= d
->rt_period
;
8234 runtime
= d
->rt_runtime
;
8238 * Cannot have more runtime than the period.
8240 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8244 * Ensure we don't starve existing RT tasks.
8246 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8249 total
= to_ratio(period
, runtime
);
8252 * Nobody can have more than the global setting allows.
8254 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8258 * The sum of our children's runtime should not exceed our own.
8260 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8261 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8262 runtime
= child
->rt_bandwidth
.rt_runtime
;
8264 if (child
== d
->tg
) {
8265 period
= d
->rt_period
;
8266 runtime
= d
->rt_runtime
;
8269 sum
+= to_ratio(period
, runtime
);
8278 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8280 struct rt_schedulable_data data
= {
8282 .rt_period
= period
,
8283 .rt_runtime
= runtime
,
8286 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8289 static int tg_set_bandwidth(struct task_group
*tg
,
8290 u64 rt_period
, u64 rt_runtime
)
8294 mutex_lock(&rt_constraints_mutex
);
8295 read_lock(&tasklist_lock
);
8296 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8300 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8301 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8302 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8304 for_each_possible_cpu(i
) {
8305 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8307 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8308 rt_rq
->rt_runtime
= rt_runtime
;
8309 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8311 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8313 read_unlock(&tasklist_lock
);
8314 mutex_unlock(&rt_constraints_mutex
);
8319 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8321 u64 rt_runtime
, rt_period
;
8323 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8324 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8325 if (rt_runtime_us
< 0)
8326 rt_runtime
= RUNTIME_INF
;
8328 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8331 long sched_group_rt_runtime(struct task_group
*tg
)
8335 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8338 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8339 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8340 return rt_runtime_us
;
8343 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8345 u64 rt_runtime
, rt_period
;
8347 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8348 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8353 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8356 long sched_group_rt_period(struct task_group
*tg
)
8360 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8361 do_div(rt_period_us
, NSEC_PER_USEC
);
8362 return rt_period_us
;
8365 static int sched_rt_global_constraints(void)
8367 u64 runtime
, period
;
8370 if (sysctl_sched_rt_period
<= 0)
8373 runtime
= global_rt_runtime();
8374 period
= global_rt_period();
8377 * Sanity check on the sysctl variables.
8379 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8382 mutex_lock(&rt_constraints_mutex
);
8383 read_lock(&tasklist_lock
);
8384 ret
= __rt_schedulable(NULL
, 0, 0);
8385 read_unlock(&tasklist_lock
);
8386 mutex_unlock(&rt_constraints_mutex
);
8391 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
8393 /* Don't accept realtime tasks when there is no way for them to run */
8394 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
8400 #else /* !CONFIG_RT_GROUP_SCHED */
8401 static int sched_rt_global_constraints(void)
8403 unsigned long flags
;
8406 if (sysctl_sched_rt_period
<= 0)
8410 * There's always some RT tasks in the root group
8411 * -- migration, kstopmachine etc..
8413 if (sysctl_sched_rt_runtime
== 0)
8416 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8417 for_each_possible_cpu(i
) {
8418 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8420 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
8421 rt_rq
->rt_runtime
= global_rt_runtime();
8422 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
8424 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8428 #endif /* CONFIG_RT_GROUP_SCHED */
8430 int sched_rt_handler(struct ctl_table
*table
, int write
,
8431 void __user
*buffer
, size_t *lenp
,
8435 int old_period
, old_runtime
;
8436 static DEFINE_MUTEX(mutex
);
8439 old_period
= sysctl_sched_rt_period
;
8440 old_runtime
= sysctl_sched_rt_runtime
;
8442 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
8444 if (!ret
&& write
) {
8445 ret
= sched_rt_global_constraints();
8447 sysctl_sched_rt_period
= old_period
;
8448 sysctl_sched_rt_runtime
= old_runtime
;
8450 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8451 def_rt_bandwidth
.rt_period
=
8452 ns_to_ktime(global_rt_period());
8455 mutex_unlock(&mutex
);
8460 #ifdef CONFIG_CGROUP_SCHED
8462 /* return corresponding task_group object of a cgroup */
8463 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8465 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8466 struct task_group
, css
);
8469 static struct cgroup_subsys_state
*
8470 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8472 struct task_group
*tg
, *parent
;
8474 if (!cgrp
->parent
) {
8475 /* This is early initialization for the top cgroup */
8476 return &init_task_group
.css
;
8479 parent
= cgroup_tg(cgrp
->parent
);
8480 tg
= sched_create_group(parent
);
8482 return ERR_PTR(-ENOMEM
);
8488 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8490 struct task_group
*tg
= cgroup_tg(cgrp
);
8492 sched_destroy_group(tg
);
8496 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
8498 #ifdef CONFIG_RT_GROUP_SCHED
8499 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
8502 /* We don't support RT-tasks being in separate groups */
8503 if (tsk
->sched_class
!= &fair_sched_class
)
8510 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8511 struct task_struct
*tsk
, bool threadgroup
)
8513 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
8517 struct task_struct
*c
;
8519 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8520 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
8532 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8533 struct cgroup
*old_cont
, struct task_struct
*tsk
,
8536 sched_move_task(tsk
);
8538 struct task_struct
*c
;
8540 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
8547 #ifdef CONFIG_FAIR_GROUP_SCHED
8548 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8551 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8554 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8556 struct task_group
*tg
= cgroup_tg(cgrp
);
8558 return (u64
) tg
->shares
;
8560 #endif /* CONFIG_FAIR_GROUP_SCHED */
8562 #ifdef CONFIG_RT_GROUP_SCHED
8563 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8566 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8569 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8571 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8574 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8577 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8580 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8582 return sched_group_rt_period(cgroup_tg(cgrp
));
8584 #endif /* CONFIG_RT_GROUP_SCHED */
8586 static struct cftype cpu_files
[] = {
8587 #ifdef CONFIG_FAIR_GROUP_SCHED
8590 .read_u64
= cpu_shares_read_u64
,
8591 .write_u64
= cpu_shares_write_u64
,
8594 #ifdef CONFIG_RT_GROUP_SCHED
8596 .name
= "rt_runtime_us",
8597 .read_s64
= cpu_rt_runtime_read
,
8598 .write_s64
= cpu_rt_runtime_write
,
8601 .name
= "rt_period_us",
8602 .read_u64
= cpu_rt_period_read_uint
,
8603 .write_u64
= cpu_rt_period_write_uint
,
8608 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8610 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8613 struct cgroup_subsys cpu_cgroup_subsys
= {
8615 .create
= cpu_cgroup_create
,
8616 .destroy
= cpu_cgroup_destroy
,
8617 .can_attach
= cpu_cgroup_can_attach
,
8618 .attach
= cpu_cgroup_attach
,
8619 .populate
= cpu_cgroup_populate
,
8620 .subsys_id
= cpu_cgroup_subsys_id
,
8624 #endif /* CONFIG_CGROUP_SCHED */
8626 #ifdef CONFIG_CGROUP_CPUACCT
8629 * CPU accounting code for task groups.
8631 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8632 * (balbir@in.ibm.com).
8635 /* track cpu usage of a group of tasks and its child groups */
8637 struct cgroup_subsys_state css
;
8638 /* cpuusage holds pointer to a u64-type object on every cpu */
8639 u64 __percpu
*cpuusage
;
8640 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
8641 struct cpuacct
*parent
;
8644 struct cgroup_subsys cpuacct_subsys
;
8646 /* return cpu accounting group corresponding to this container */
8647 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8649 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8650 struct cpuacct
, css
);
8653 /* return cpu accounting group to which this task belongs */
8654 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8656 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8657 struct cpuacct
, css
);
8660 /* create a new cpu accounting group */
8661 static struct cgroup_subsys_state
*cpuacct_create(
8662 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8664 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8670 ca
->cpuusage
= alloc_percpu(u64
);
8674 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8675 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
8676 goto out_free_counters
;
8679 ca
->parent
= cgroup_ca(cgrp
->parent
);
8685 percpu_counter_destroy(&ca
->cpustat
[i
]);
8686 free_percpu(ca
->cpuusage
);
8690 return ERR_PTR(-ENOMEM
);
8693 /* destroy an existing cpu accounting group */
8695 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8697 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8700 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
8701 percpu_counter_destroy(&ca
->cpustat
[i
]);
8702 free_percpu(ca
->cpuusage
);
8706 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
8708 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8711 #ifndef CONFIG_64BIT
8713 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
8715 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8717 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8725 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
8727 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8729 #ifndef CONFIG_64BIT
8731 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
8733 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
8735 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
8741 /* return total cpu usage (in nanoseconds) of a group */
8742 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8744 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8745 u64 totalcpuusage
= 0;
8748 for_each_present_cpu(i
)
8749 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
8751 return totalcpuusage
;
8754 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8757 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8766 for_each_present_cpu(i
)
8767 cpuacct_cpuusage_write(ca
, i
, 0);
8773 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
8776 struct cpuacct
*ca
= cgroup_ca(cgroup
);
8780 for_each_present_cpu(i
) {
8781 percpu
= cpuacct_cpuusage_read(ca
, i
);
8782 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
8784 seq_printf(m
, "\n");
8788 static const char *cpuacct_stat_desc
[] = {
8789 [CPUACCT_STAT_USER
] = "user",
8790 [CPUACCT_STAT_SYSTEM
] = "system",
8793 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
8794 struct cgroup_map_cb
*cb
)
8796 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8799 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
8800 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
8801 val
= cputime64_to_clock_t(val
);
8802 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
8807 static struct cftype files
[] = {
8810 .read_u64
= cpuusage_read
,
8811 .write_u64
= cpuusage_write
,
8814 .name
= "usage_percpu",
8815 .read_seq_string
= cpuacct_percpu_seq_read
,
8819 .read_map
= cpuacct_stats_show
,
8823 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8825 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8829 * charge this task's execution time to its accounting group.
8831 * called with rq->lock held.
8833 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8838 if (unlikely(!cpuacct_subsys
.active
))
8841 cpu
= task_cpu(tsk
);
8847 for (; ca
; ca
= ca
->parent
) {
8848 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
8849 *cpuusage
+= cputime
;
8856 * When CONFIG_VIRT_CPU_ACCOUNTING is enabled one jiffy can be very large
8857 * in cputime_t units. As a result, cpuacct_update_stats calls
8858 * percpu_counter_add with values large enough to always overflow the
8859 * per cpu batch limit causing bad SMP scalability.
8861 * To fix this we scale percpu_counter_batch by cputime_one_jiffy so we
8862 * batch the same amount of time with CONFIG_VIRT_CPU_ACCOUNTING disabled
8863 * and enabled. We cap it at INT_MAX which is the largest allowed batch value.
8866 #define CPUACCT_BATCH \
8867 min_t(long, percpu_counter_batch * cputime_one_jiffy, INT_MAX)
8869 #define CPUACCT_BATCH 0
8873 * Charge the system/user time to the task's accounting group.
8875 static void cpuacct_update_stats(struct task_struct
*tsk
,
8876 enum cpuacct_stat_index idx
, cputime_t val
)
8879 int batch
= CPUACCT_BATCH
;
8881 if (unlikely(!cpuacct_subsys
.active
))
8888 __percpu_counter_add(&ca
->cpustat
[idx
], val
, batch
);
8894 struct cgroup_subsys cpuacct_subsys
= {
8896 .create
= cpuacct_create
,
8897 .destroy
= cpuacct_destroy
,
8898 .populate
= cpuacct_populate
,
8899 .subsys_id
= cpuacct_subsys_id
,
8901 #endif /* CONFIG_CGROUP_CPUACCT */
8905 void synchronize_sched_expedited(void)
8909 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
8911 #else /* #ifndef CONFIG_SMP */
8913 static atomic_t synchronize_sched_expedited_count
= ATOMIC_INIT(0);
8915 static int synchronize_sched_expedited_cpu_stop(void *data
)
8918 * There must be a full memory barrier on each affected CPU
8919 * between the time that try_stop_cpus() is called and the
8920 * time that it returns.
8922 * In the current initial implementation of cpu_stop, the
8923 * above condition is already met when the control reaches
8924 * this point and the following smp_mb() is not strictly
8925 * necessary. Do smp_mb() anyway for documentation and
8926 * robustness against future implementation changes.
8928 smp_mb(); /* See above comment block. */
8933 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
8934 * approach to force grace period to end quickly. This consumes
8935 * significant time on all CPUs, and is thus not recommended for
8936 * any sort of common-case code.
8938 * Note that it is illegal to call this function while holding any
8939 * lock that is acquired by a CPU-hotplug notifier. Failing to
8940 * observe this restriction will result in deadlock.
8942 void synchronize_sched_expedited(void)
8944 int snap
, trycount
= 0;
8946 smp_mb(); /* ensure prior mod happens before capturing snap. */
8947 snap
= atomic_read(&synchronize_sched_expedited_count
) + 1;
8949 while (try_stop_cpus(cpu_online_mask
,
8950 synchronize_sched_expedited_cpu_stop
,
8953 if (trycount
++ < 10)
8954 udelay(trycount
* num_online_cpus());
8956 synchronize_sched();
8959 if (atomic_read(&synchronize_sched_expedited_count
) - snap
> 0) {
8960 smp_mb(); /* ensure test happens before caller kfree */
8965 atomic_inc(&synchronize_sched_expedited_count
);
8966 smp_mb__after_atomic_inc(); /* ensure post-GP actions seen after GP. */
8969 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
8971 #endif /* #else #ifndef CONFIG_SMP */