4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
7 * Copyright (C) 2004 Red Hat, Inc., Ingo Molnar <mingo@redhat.com>
9 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
10 * make semaphores SMP safe
11 * 1998-11-19 Implemented schedule_timeout() and related stuff
13 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
14 * hybrid priority-list and round-robin design with
15 * an array-switch method of distributing timeslices
16 * and per-CPU runqueues. Cleanups and useful suggestions
17 * by Davide Libenzi, preemptible kernel bits by Robert Love.
18 * 2003-09-03 Interactivity tuning by Con Kolivas.
19 * 2004-04-02 Scheduler domains code by Nick Piggin
20 * 2004-10-13 Real-Time Preemption support by Ingo Molnar
21 * 2007-04-15 Work begun on replacing all interactivity tuning with a
22 * fair scheduling design by Con Kolivas.
23 * 2007-05-05 Load balancing (smp-nice) and other improvements
25 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
26 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
27 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
28 * Thomas Gleixner, Mike Kravetz
32 #include <linux/module.h>
33 #include <linux/nmi.h>
34 #include <linux/init.h>
35 #include <linux/uaccess.h>
36 #include <linux/highmem.h>
37 #include <linux/smp_lock.h>
38 #include <asm/mmu_context.h>
39 #include <linux/interrupt.h>
40 #include <linux/capability.h>
41 #include <linux/completion.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/debug_locks.h>
44 #include <linux/security.h>
45 #include <linux/notifier.h>
46 #include <linux/profile.h>
47 #include <linux/freezer.h>
48 #include <linux/vmalloc.h>
49 #include <linux/blkdev.h>
50 #include <linux/delay.h>
51 #include <linux/pid_namespace.h>
52 #include <linux/smp.h>
53 #include <linux/threads.h>
54 #include <linux/timer.h>
55 #include <linux/rcupdate.h>
56 #include <linux/cpu.h>
57 #include <linux/cpuset.h>
58 #include <linux/percpu.h>
59 #include <linux/kthread.h>
60 #include <linux/proc_fs.h>
61 #include <linux/seq_file.h>
62 #include <linux/sysctl.h>
63 #include <linux/syscalls.h>
64 #include <linux/times.h>
65 #include <linux/kallsyms.h>
66 #include <linux/tsacct_kern.h>
67 #include <linux/kprobes.h>
68 #include <linux/delayacct.h>
69 #include <linux/reciprocal_div.h>
70 #include <linux/unistd.h>
71 #include <linux/pagemap.h>
72 #include <linux/hrtimer.h>
73 #include <linux/tick.h>
74 #include <linux/bootmem.h>
75 #include <linux/debugfs.h>
76 #include <linux/ctype.h>
77 #include <linux/ftrace.h>
78 #include <trace/sched.h>
81 #include <asm/irq_regs.h>
83 #include "sched_cpupri.h"
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
111 #if (BITS_PER_LONG < 64)
112 #define JIFFIES_TO_NS64(TIME) \
113 ((unsigned long long)(TIME) * ((unsigned long) (1000000000 / HZ)))
115 #define NS64_TO_JIFFIES(TIME) \
116 ((((unsigned long long)((TIME)) >> BITS_PER_LONG) * \
117 (1 + NS_TO_JIFFIES(~0UL))) + NS_TO_JIFFIES((unsigned long)(TIME)))
118 #else /* BITS_PER_LONG < 64 */
120 #define NS64_TO_JIFFIES(TIME) NS_TO_JIFFIES(TIME)
121 #define JIFFIES_TO_NS64(TIME) JIFFIES_TO_NS(TIME)
123 #endif /* BITS_PER_LONG < 64 */
126 * These are the 'tuning knobs' of the scheduler:
128 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
129 * Timeslices get refilled after they expire.
131 #define DEF_TIMESLICE (100 * HZ / 1000)
134 * single value that denotes runtime == period, ie unlimited time.
136 #define RUNTIME_INF ((u64)~0ULL)
138 DEFINE_TRACE(sched_wait_task
);
139 DEFINE_TRACE(sched_wakeup
);
140 DEFINE_TRACE(sched_wakeup_new
);
141 DEFINE_TRACE(sched_switch
);
142 DEFINE_TRACE(sched_migrate_task
);
143 DEFINE_TRACE(sched_task_setprio
);
147 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
150 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
151 * Since cpu_power is a 'constant', we can use a reciprocal divide.
153 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
155 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
159 * Each time a sched group cpu_power is changed,
160 * we must compute its reciprocal value
162 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
164 sg
->__cpu_power
+= val
;
165 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
169 #define TASK_PREEMPTS_CURR(p, rq) \
170 ((p)->prio < (rq)->curr->prio)
177 struct task_struct
* const ___current
= &init_task
;
178 struct task_struct
** const current_ptr
= (struct task_struct
** const)&___current
;
179 struct thread_info
* const current_ti
= &init_thread_union
.thread_info
;
180 struct thread_info
** const current_ti_ptr
= (struct thread_info
** const)¤t_ti
;
182 EXPORT_SYMBOL(___current
);
183 EXPORT_SYMBOL(current_ti
);
186 * The scheduler itself doesnt want 'current' to be cached
187 * during context-switches:
190 # define current __current()
191 # undef current_thread_info
192 # define current_thread_info() __current_thread_info()
195 static inline int rt_policy(int policy
)
197 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
202 static inline int task_has_rt_policy(struct task_struct
*p
)
204 return rt_policy(p
->policy
);
208 * This is the priority-queue data structure of the RT scheduling class:
210 struct rt_prio_array
{
211 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
212 struct list_head queue
[MAX_RT_PRIO
];
215 struct rt_bandwidth
{
216 /* nests inside the rq lock: */
217 raw_spinlock_t rt_runtime_lock
;
220 struct hrtimer rt_period_timer
;
223 static struct rt_bandwidth def_rt_bandwidth
;
225 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
227 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
229 struct rt_bandwidth
*rt_b
=
230 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
236 now
= hrtimer_cb_get_time(timer
);
237 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
242 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
245 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
249 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
251 rt_b
->rt_period
= ns_to_ktime(period
);
252 rt_b
->rt_runtime
= runtime
;
254 spin_lock_init(&rt_b
->rt_runtime_lock
);
256 hrtimer_init(&rt_b
->rt_period_timer
,
257 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
258 rt_b
->rt_period_timer
.irqsafe
= 1;
259 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
262 static inline int rt_bandwidth_enabled(void)
264 return sysctl_sched_rt_runtime
>= 0;
267 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
271 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
274 if (hrtimer_active(&rt_b
->rt_period_timer
))
277 spin_lock(&rt_b
->rt_runtime_lock
);
282 if (hrtimer_active(&rt_b
->rt_period_timer
))
285 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
286 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
288 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
289 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
290 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
291 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
292 HRTIMER_MODE_ABS
, 0);
294 spin_unlock(&rt_b
->rt_runtime_lock
);
297 #ifdef CONFIG_RT_GROUP_SCHED
298 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
300 hrtimer_cancel(&rt_b
->rt_period_timer
);
305 * sched_domains_mutex serializes calls to arch_init_sched_domains,
306 * detach_destroy_domains and partition_sched_domains.
308 static DEFINE_MUTEX(sched_domains_mutex
);
310 #ifdef CONFIG_GROUP_SCHED
312 #include <linux/cgroup.h>
316 static LIST_HEAD(task_groups
);
318 /* task group related information */
320 #ifdef CONFIG_CGROUP_SCHED
321 struct cgroup_subsys_state css
;
324 #ifdef CONFIG_USER_SCHED
328 #ifdef CONFIG_FAIR_GROUP_SCHED
329 /* schedulable entities of this group on each cpu */
330 struct sched_entity
**se
;
331 /* runqueue "owned" by this group on each cpu */
332 struct cfs_rq
**cfs_rq
;
333 unsigned long shares
;
336 #ifdef CONFIG_RT_GROUP_SCHED
337 struct sched_rt_entity
**rt_se
;
338 struct rt_rq
**rt_rq
;
340 struct rt_bandwidth rt_bandwidth
;
344 struct list_head list
;
346 struct task_group
*parent
;
347 struct list_head siblings
;
348 struct list_head children
;
351 #ifdef CONFIG_USER_SCHED
353 /* Helper function to pass uid information to create_sched_user() */
354 void set_tg_uid(struct user_struct
*user
)
356 user
->tg
->uid
= user
->uid
;
361 * Every UID task group (including init_task_group aka UID-0) will
362 * be a child to this group.
364 struct task_group root_task_group
;
366 #ifdef CONFIG_FAIR_GROUP_SCHED
367 /* Default task group's sched entity on each cpu */
368 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
369 /* Default task group's cfs_rq on each cpu */
370 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
371 #endif /* CONFIG_FAIR_GROUP_SCHED */
373 #ifdef CONFIG_RT_GROUP_SCHED
374 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
375 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
376 #endif /* CONFIG_RT_GROUP_SCHED */
377 #else /* !CONFIG_USER_SCHED */
378 #define root_task_group init_task_group
379 #endif /* CONFIG_USER_SCHED */
381 /* task_group_lock serializes add/remove of task groups and also changes to
382 * a task group's cpu shares.
384 static DEFINE_SPINLOCK(task_group_lock
);
387 static int root_task_group_empty(void)
389 return list_empty(&root_task_group
.children
);
393 #ifdef CONFIG_FAIR_GROUP_SCHED
394 #ifdef CONFIG_USER_SCHED
395 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
396 #else /* !CONFIG_USER_SCHED */
397 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
398 #endif /* CONFIG_USER_SCHED */
401 * A weight of 0 or 1 can cause arithmetics problems.
402 * A weight of a cfs_rq is the sum of weights of which entities
403 * are queued on this cfs_rq, so a weight of a entity should not be
404 * too large, so as the shares value of a task group.
405 * (The default weight is 1024 - so there's no practical
406 * limitation from this.)
409 #define MAX_SHARES (1UL << 18)
411 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
414 /* Default task group.
415 * Every task in system belong to this group at bootup.
417 struct task_group init_task_group
;
419 /* return group to which a task belongs */
420 static inline struct task_group
*task_group(struct task_struct
*p
)
422 struct task_group
*tg
;
424 #ifdef CONFIG_USER_SCHED
426 tg
= __task_cred(p
)->user
->tg
;
428 #elif defined(CONFIG_CGROUP_SCHED)
429 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
430 struct task_group
, css
);
432 tg
= &init_task_group
;
437 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
438 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
440 #ifdef CONFIG_FAIR_GROUP_SCHED
441 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
442 p
->se
.parent
= task_group(p
)->se
[cpu
];
445 #ifdef CONFIG_RT_GROUP_SCHED
446 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
447 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
454 static int root_task_group_empty(void)
460 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
461 static inline struct task_group
*task_group(struct task_struct
*p
)
466 #endif /* CONFIG_GROUP_SCHED */
468 /* CFS-related fields in a runqueue */
470 struct load_weight load
;
471 unsigned long nr_running
;
476 struct rb_root tasks_timeline
;
477 struct rb_node
*rb_leftmost
;
479 struct list_head tasks
;
480 struct list_head
*balance_iterator
;
483 * 'curr' points to currently running entity on this cfs_rq.
484 * It is set to NULL otherwise (i.e when none are currently running).
486 struct sched_entity
*curr
, *next
, *last
;
488 unsigned int nr_spread_over
;
490 #ifdef CONFIG_FAIR_GROUP_SCHED
491 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
494 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
495 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
496 * (like users, containers etc.)
498 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
499 * list is used during load balance.
501 struct list_head leaf_cfs_rq_list
;
502 struct task_group
*tg
; /* group that "owns" this runqueue */
506 * the part of load.weight contributed by tasks
508 unsigned long task_weight
;
511 * h_load = weight * f(tg)
513 * Where f(tg) is the recursive weight fraction assigned to
516 unsigned long h_load
;
519 * this cpu's part of tg->shares
521 unsigned long shares
;
524 * load.weight at the time we set shares
526 unsigned long rq_weight
;
531 /* Real-Time classes' related field in a runqueue: */
533 struct rt_prio_array active
;
534 unsigned long rt_nr_running
;
535 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
537 int curr
; /* highest queued rt task prio */
539 int next
; /* next highest */
544 unsigned long rt_nr_migratory
;
546 struct plist_head pushable_tasks
;
548 unsigned long rt_nr_uninterruptible
;
552 /* Nests inside the rq lock: */
553 raw_spinlock_t rt_runtime_lock
;
555 #ifdef CONFIG_RT_GROUP_SCHED
556 unsigned long rt_nr_boosted
;
559 struct list_head leaf_rt_rq_list
;
560 struct task_group
*tg
;
561 struct sched_rt_entity
*rt_se
;
568 * We add the notion of a root-domain which will be used to define per-domain
569 * variables. Each exclusive cpuset essentially defines an island domain by
570 * fully partitioning the member cpus from any other cpuset. Whenever a new
571 * exclusive cpuset is created, we also create and attach a new root-domain
578 cpumask_var_t online
;
581 * The "RT overload" flag: it gets set if a CPU has more than
582 * one runnable RT task.
584 cpumask_var_t rto_mask
;
587 struct cpupri cpupri
;
589 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
591 * Preferred wake up cpu nominated by sched_mc balance that will be
592 * used when most cpus are idle in the system indicating overall very
593 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
595 unsigned int sched_mc_preferred_wakeup_cpu
;
600 * By default the system creates a single root-domain with all cpus as
601 * members (mimicking the global state we have today).
603 static struct root_domain def_root_domain
;
608 * This is the main, per-CPU runqueue data structure.
610 * Locking rule: those places that want to lock multiple runqueues
611 * (such as the load balancing or the thread migration code), lock
612 * acquire operations must be ordered by ascending &runqueue.
619 * nr_running and cpu_load should be in the same cacheline because
620 * remote CPUs use both these fields when doing load calculation.
622 unsigned long nr_running
;
623 #define CPU_LOAD_IDX_MAX 5
624 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
626 unsigned long last_tick_seen
;
627 unsigned char in_nohz_recently
;
629 /* capture load from *all* tasks on this cpu: */
630 struct load_weight load
;
631 unsigned long nr_load_updates
;
633 u64 nr_migrations_in
;
638 #ifdef CONFIG_FAIR_GROUP_SCHED
639 /* list of leaf cfs_rq on this cpu: */
640 struct list_head leaf_cfs_rq_list
;
642 #ifdef CONFIG_RT_GROUP_SCHED
643 struct list_head leaf_rt_rq_list
;
647 * This is part of a global counter where only the total sum
648 * over all CPUs matters. A task can increase this counter on
649 * one CPU and if it got migrated afterwards it may decrease
650 * it on another CPU. Always updated under the runqueue lock:
652 unsigned long nr_uninterruptible
;
654 unsigned long switch_timestamp
;
655 unsigned long slice_avg
;
656 struct task_struct
*curr
, *idle
;
657 unsigned long next_balance
;
658 struct mm_struct
*prev_mm
;
665 struct root_domain
*rd
;
666 struct sched_domain
*sd
;
668 unsigned char idle_at_tick
;
669 /* For active balancing */
672 /* cpu of this runqueue: */
676 unsigned long avg_load_per_task
;
678 struct task_struct
*migration_thread
;
679 struct list_head migration_queue
;
682 /* calc_load related fields */
683 unsigned long calc_load_update
;
684 long calc_load_active
;
686 #ifdef CONFIG_SCHED_HRTICK
688 int hrtick_csd_pending
;
689 struct call_single_data hrtick_csd
;
691 struct hrtimer hrtick_timer
;
694 #ifdef CONFIG_SCHEDSTATS
696 struct sched_info rq_sched_info
;
697 unsigned long long rq_cpu_time
;
698 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
700 /* sys_sched_yield() stats */
701 unsigned int yld_count
;
703 /* schedule() stats */
704 unsigned int sched_switch
;
705 unsigned int sched_count
;
706 unsigned int sched_goidle
;
708 /* try_to_wake_up() stats */
709 unsigned int ttwu_count
;
710 unsigned int ttwu_local
;
713 unsigned int bkl_count
;
715 /* RT-overload stats: */
716 unsigned long rto_schedule
;
717 unsigned long rto_schedule_tail
;
718 unsigned long rto_wakeup
;
719 unsigned long rto_pulled
;
720 unsigned long rto_pushed
;
724 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
726 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
728 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
731 static inline int cpu_of(struct rq
*rq
)
741 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
742 * See detach_destroy_domains: synchronize_sched for details.
744 * The domain tree of any CPU may only be accessed from within
745 * preempt-disabled sections.
747 #define for_each_domain(cpu, __sd) \
748 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
750 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
751 #define this_rq() (&__get_cpu_var(runqueues))
752 #define task_rq(p) cpu_rq(task_cpu(p))
753 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
755 inline void update_rq_clock(struct rq
*rq
)
757 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
761 int task_is_current(struct task_struct
*task
)
763 return task_rq(task
)->curr
== task
;
768 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
770 #ifdef CONFIG_SCHED_DEBUG
771 # define const_debug __read_mostly
773 # define const_debug static const
779 * Returns true if the current cpu runqueue is locked.
780 * This interface allows printk to be called with the runqueue lock
781 * held and know whether or not it is OK to wake up the klogd.
783 int runqueue_is_locked(void)
786 struct rq
*rq
= cpu_rq(cpu
);
789 ret
= spin_is_locked(&rq
->lock
);
795 * Debugging: various feature bits
798 #define SCHED_FEAT(name, enabled) \
799 __SCHED_FEAT_##name ,
802 #include "sched_features.h"
807 #define SCHED_FEAT(name, enabled) \
808 (1UL << __SCHED_FEAT_##name) * enabled |
810 const_debug
unsigned int sysctl_sched_features
=
811 #include "sched_features.h"
816 #ifdef CONFIG_SCHED_DEBUG
817 #define SCHED_FEAT(name, enabled) \
820 static __read_mostly
char *sched_feat_names
[] = {
821 #include "sched_features.h"
827 static int sched_feat_show(struct seq_file
*m
, void *v
)
831 for (i
= 0; sched_feat_names
[i
]; i
++) {
832 if (!(sysctl_sched_features
& (1UL << i
)))
834 seq_printf(m
, "%s ", sched_feat_names
[i
]);
842 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
843 size_t cnt
, loff_t
*ppos
)
853 if (copy_from_user(&buf
, ubuf
, cnt
))
858 if (strncmp(buf
, "NO_", 3) == 0) {
863 for (i
= 0; sched_feat_names
[i
]; i
++) {
864 int len
= strlen(sched_feat_names
[i
]);
866 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
868 sysctl_sched_features
&= ~(1UL << i
);
870 sysctl_sched_features
|= (1UL << i
);
875 if (!sched_feat_names
[i
])
883 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
885 return single_open(filp
, sched_feat_show
, NULL
);
888 static struct file_operations sched_feat_fops
= {
889 .open
= sched_feat_open
,
890 .write
= sched_feat_write
,
893 .release
= single_release
,
896 static __init
int sched_init_debug(void)
898 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
903 late_initcall(sched_init_debug
);
907 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
910 * Number of tasks to iterate in a single balance run.
911 * Limited because this is done with IRQs disabled.
913 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
916 * ratelimit for updating the group shares.
919 unsigned int sysctl_sched_shares_ratelimit
= 250000;
922 * Inject some fuzzyness into changing the per-cpu group shares
923 * this avoids remote rq-locks at the expense of fairness.
926 unsigned int sysctl_sched_shares_thresh
= 4;
929 * period over which we measure -rt task cpu usage in us.
932 unsigned int sysctl_sched_rt_period
= 1000000;
934 static __read_mostly
int scheduler_running
;
937 * part of the period that we allow rt tasks to run in us.
940 int sysctl_sched_rt_runtime
= 950000;
942 static inline u64
global_rt_period(void)
944 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
947 static inline u64
global_rt_runtime(void)
949 if (sysctl_sched_rt_runtime
< 0)
952 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
956 * We really dont want to do anything complex within switch_to()
957 * on PREEMPT_RT - this check enforces this.
959 #ifdef prepare_arch_switch
960 # ifdef CONFIG_PREEMPT_RT
963 # define _finish_arch_switch finish_arch_switch
967 #ifndef prepare_arch_switch
968 # define prepare_arch_switch(next) do { } while (0)
970 #ifndef finish_arch_switch
971 # define _finish_arch_switch(prev) do { } while (0)
974 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
976 return rq
->curr
== p
;
979 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
984 return task_current(rq
, p
);
988 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
989 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
993 * We can optimise this out completely for !SMP, because the
994 * SMP rebalancing from interrupt is the only thing that cares
1001 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1005 * After ->oncpu is cleared, the task can be moved to a different CPU.
1006 * We must ensure this doesn't happen until the switch is completely
1012 #ifdef CONFIG_DEBUG_SPINLOCK
1013 /* this is a valid case when another task releases the spinlock */
1014 rq
->lock
.owner
= current
;
1017 * If we are tracking spinlock dependencies then we have to
1018 * fix up the runqueue lock - which gets 'carried over' from
1019 * prev into current:
1021 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
1023 spin_unlock(&rq
->lock
);
1026 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1028 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1032 * We can optimise this out completely for !SMP, because the
1033 * SMP rebalancing from interrupt is the only thing that cares
1038 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1039 spin_unlock_irq(&rq
->lock
);
1041 spin_unlock(&rq
->lock
);
1045 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1049 * After ->oncpu is cleared, the task can be moved to a different CPU.
1050 * We must ensure this doesn't happen until the switch is completely
1056 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1057 local_irq_disable();
1060 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1063 * __task_rq_lock - lock the runqueue a given task resides on.
1064 * Must be called interrupts disabled.
1066 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
1067 __acquires(rq
->lock
)
1070 struct rq
*rq
= task_rq(p
);
1071 spin_lock(&rq
->lock
);
1072 if (likely(rq
== task_rq(p
)))
1074 spin_unlock(&rq
->lock
);
1079 * task_rq_lock - lock the runqueue a given task resides on and disable
1080 * interrupts. Note the ordering: we can safely lookup the task_rq without
1081 * explicitly disabling preemption.
1083 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1084 __acquires(rq
->lock
)
1089 local_irq_save(*flags
);
1091 spin_lock(&rq
->lock
);
1092 if (likely(rq
== task_rq(p
)))
1094 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1098 void curr_rq_lock_irq_save(unsigned long *flags
)
1099 __acquires(rq
->lock
)
1103 local_irq_save(*flags
);
1104 rq
= cpu_rq(smp_processor_id());
1105 spin_lock(&rq
->lock
);
1108 void curr_rq_unlock_irq_restore(unsigned long *flags
)
1109 __releases(rq
->lock
)
1113 rq
= cpu_rq(smp_processor_id());
1114 spin_unlock(&rq
->lock
);
1115 local_irq_restore(*flags
);
1118 void task_rq_unlock_wait(struct task_struct
*p
)
1120 struct rq
*rq
= task_rq(p
);
1122 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1123 spin_unlock_wait(&rq
->lock
);
1126 static void __task_rq_unlock(struct rq
*rq
)
1127 __releases(rq
->lock
)
1129 spin_unlock(&rq
->lock
);
1132 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1133 __releases(rq
->lock
)
1135 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1139 * this_rq_lock - lock this runqueue and disable interrupts.
1141 static struct rq
*this_rq_lock(void)
1142 __acquires(rq
->lock
)
1146 local_irq_disable();
1148 spin_lock(&rq
->lock
);
1153 #ifdef CONFIG_SCHED_HRTICK
1155 * Use HR-timers to deliver accurate preemption points.
1157 * Its all a bit involved since we cannot program an hrt while holding the
1158 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1161 * When we get rescheduled we reprogram the hrtick_timer outside of the
1167 * - enabled by features
1168 * - hrtimer is actually high res
1170 static inline int hrtick_enabled(struct rq
*rq
)
1172 if (!sched_feat(HRTICK
))
1174 if (!cpu_active(cpu_of(rq
)))
1176 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1179 static void hrtick_clear(struct rq
*rq
)
1181 if (hrtimer_active(&rq
->hrtick_timer
))
1182 hrtimer_cancel(&rq
->hrtick_timer
);
1186 * High-resolution timer tick.
1187 * Runs from hardirq context with interrupts disabled.
1189 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1191 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1193 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1195 spin_lock(&rq
->lock
);
1196 update_rq_clock(rq
);
1197 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1198 spin_unlock(&rq
->lock
);
1200 return HRTIMER_NORESTART
;
1205 * called from hardirq (IPI) context
1207 static void __hrtick_start(void *arg
)
1209 struct rq
*rq
= arg
;
1211 spin_lock(&rq
->lock
);
1212 hrtimer_restart(&rq
->hrtick_timer
);
1213 rq
->hrtick_csd_pending
= 0;
1214 spin_unlock(&rq
->lock
);
1218 * Called to set the hrtick timer state.
1220 * called with rq->lock held and irqs disabled
1222 static void hrtick_start(struct rq
*rq
, u64 delay
)
1224 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1225 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1227 hrtimer_set_expires(timer
, time
);
1229 if (rq
== this_rq()) {
1230 hrtimer_restart(timer
);
1231 } else if (!rq
->hrtick_csd_pending
) {
1232 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1233 rq
->hrtick_csd_pending
= 1;
1238 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1240 int cpu
= (int)(long)hcpu
;
1243 case CPU_UP_CANCELED
:
1244 case CPU_UP_CANCELED_FROZEN
:
1245 case CPU_DOWN_PREPARE
:
1246 case CPU_DOWN_PREPARE_FROZEN
:
1248 case CPU_DEAD_FROZEN
:
1249 hrtick_clear(cpu_rq(cpu
));
1256 static __init
void init_hrtick(void)
1258 hotcpu_notifier(hotplug_hrtick
, 0);
1262 * Called to set the hrtick timer state.
1264 * called with rq->lock held and irqs disabled
1266 static void hrtick_start(struct rq
*rq
, u64 delay
)
1268 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1269 HRTIMER_MODE_REL
, 0);
1272 static inline void init_hrtick(void)
1275 #endif /* CONFIG_SMP */
1277 static void init_rq_hrtick(struct rq
*rq
)
1280 rq
->hrtick_csd_pending
= 0;
1282 rq
->hrtick_csd
.flags
= 0;
1283 rq
->hrtick_csd
.func
= __hrtick_start
;
1284 rq
->hrtick_csd
.info
= rq
;
1287 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1288 rq
->hrtick_timer
.function
= hrtick
;
1289 rq
->hrtick_timer
.irqsafe
= 1;
1291 #else /* CONFIG_SCHED_HRTICK */
1292 static inline void hrtick_clear(struct rq
*rq
)
1296 static inline void init_rq_hrtick(struct rq
*rq
)
1300 static inline void init_hrtick(void)
1303 #endif /* CONFIG_SCHED_HRTICK */
1306 * resched_task - mark a task 'to be rescheduled now'.
1308 * On UP this means the setting of the need_resched flag, on SMP it
1309 * might also involve a cross-CPU call to trigger the scheduler on
1314 #ifndef tsk_is_polling
1315 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1318 static void resched_task(struct task_struct
*p
)
1322 assert_spin_locked(&task_rq(p
)->lock
);
1324 if (test_tsk_need_resched(p
))
1327 set_tsk_need_resched(p
);
1330 if (cpu
== smp_processor_id())
1333 /* NEED_RESCHED must be visible before we test polling */
1335 if (!tsk_is_polling(p
))
1336 smp_send_reschedule(cpu
);
1339 static void resched_cpu(int cpu
)
1341 struct rq
*rq
= cpu_rq(cpu
);
1342 unsigned long flags
;
1344 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1346 resched_task(cpu_curr(cpu
));
1347 spin_unlock_irqrestore(&rq
->lock
, flags
);
1352 * When add_timer_on() enqueues a timer into the timer wheel of an
1353 * idle CPU then this timer might expire before the next timer event
1354 * which is scheduled to wake up that CPU. In case of a completely
1355 * idle system the next event might even be infinite time into the
1356 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1357 * leaves the inner idle loop so the newly added timer is taken into
1358 * account when the CPU goes back to idle and evaluates the timer
1359 * wheel for the next timer event.
1361 void wake_up_idle_cpu(int cpu
)
1363 struct rq
*rq
= cpu_rq(cpu
);
1365 if (cpu
== raw_smp_processor_id())
1369 * This is safe, as this function is called with the timer
1370 * wheel base lock of (cpu) held. When the CPU is on the way
1371 * to idle and has not yet set rq->curr to idle then it will
1372 * be serialized on the timer wheel base lock and take the new
1373 * timer into account automatically.
1375 if (rq
->curr
!= rq
->idle
)
1379 * We can set TIF_RESCHED on the idle task of the other CPU
1380 * lockless. The worst case is that the other CPU runs the
1381 * idle task through an additional NOOP schedule()
1383 set_tsk_need_resched(rq
->idle
);
1385 /* NEED_RESCHED must be visible before we test polling */
1387 if (!tsk_is_polling(rq
->idle
))
1388 smp_send_reschedule(cpu
);
1390 #endif /* CONFIG_NO_HZ */
1392 #else /* !CONFIG_SMP */
1393 static void resched_task(struct task_struct
*p
)
1395 assert_spin_locked(&task_rq(p
)->lock
);
1396 set_tsk_need_resched(p
);
1398 #endif /* CONFIG_SMP */
1400 #if BITS_PER_LONG == 32
1401 # define WMULT_CONST (~0UL)
1403 # define WMULT_CONST (1UL << 32)
1406 #define WMULT_SHIFT 32
1409 * Shift right and round:
1411 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1414 * delta *= weight / lw
1416 static unsigned long
1417 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1418 struct load_weight
*lw
)
1422 if (!lw
->inv_weight
) {
1423 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1426 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1430 tmp
= (u64
)delta_exec
* weight
;
1432 * Check whether we'd overflow the 64-bit multiplication:
1434 if (unlikely(tmp
> WMULT_CONST
))
1435 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1438 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1440 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1443 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1449 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1456 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1457 * of tasks with abnormal "nice" values across CPUs the contribution that
1458 * each task makes to its run queue's load is weighted according to its
1459 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1460 * scaled version of the new time slice allocation that they receive on time
1464 #define WEIGHT_IDLEPRIO 3
1465 #define WMULT_IDLEPRIO 1431655765
1468 * Nice levels are multiplicative, with a gentle 10% change for every
1469 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1470 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1471 * that remained on nice 0.
1473 * The "10% effect" is relative and cumulative: from _any_ nice level,
1474 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1475 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1476 * If a task goes up by ~10% and another task goes down by ~10% then
1477 * the relative distance between them is ~25%.)
1479 static const int prio_to_weight
[40] = {
1480 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1481 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1482 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1483 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1484 /* 0 */ 1024, 820, 655, 526, 423,
1485 /* 5 */ 335, 272, 215, 172, 137,
1486 /* 10 */ 110, 87, 70, 56, 45,
1487 /* 15 */ 36, 29, 23, 18, 15,
1491 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1493 * In cases where the weight does not change often, we can use the
1494 * precalculated inverse to speed up arithmetics by turning divisions
1495 * into multiplications:
1497 static const u32 prio_to_wmult
[40] = {
1498 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1499 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1500 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1501 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1502 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1503 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1504 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1505 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1508 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1511 * runqueue iterator, to support SMP load-balancing between different
1512 * scheduling classes, without having to expose their internal data
1513 * structures to the load-balancing proper:
1515 struct rq_iterator
{
1517 struct task_struct
*(*start
)(void *);
1518 struct task_struct
*(*next
)(void *);
1522 static unsigned long
1523 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1524 unsigned long max_load_move
, struct sched_domain
*sd
,
1525 enum cpu_idle_type idle
, int *all_pinned
,
1526 int *this_best_prio
, struct rq_iterator
*iterator
);
1529 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1530 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1531 struct rq_iterator
*iterator
);
1534 #ifdef CONFIG_CGROUP_CPUACCT
1535 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1537 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1540 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1542 update_load_add(&rq
->load
, load
);
1545 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1547 update_load_sub(&rq
->load
, load
);
1550 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1551 typedef int (*tg_visitor
)(struct task_group
*, void *);
1554 * Iterate the full tree, calling @down when first entering a node and @up when
1555 * leaving it for the final time.
1557 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1559 struct task_group
*parent
, *child
;
1563 parent
= &root_task_group
;
1565 ret
= (*down
)(parent
, data
);
1568 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1575 ret
= (*up
)(parent
, data
);
1580 parent
= parent
->parent
;
1589 static int tg_nop(struct task_group
*tg
, void *data
)
1596 static unsigned long source_load(int cpu
, int type
);
1597 static unsigned long target_load(int cpu
, int type
);
1598 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1600 static unsigned long cpu_avg_load_per_task(int cpu
)
1602 struct rq
*rq
= cpu_rq(cpu
);
1603 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1606 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1608 rq
->avg_load_per_task
= 0;
1610 return rq
->avg_load_per_task
;
1613 #ifdef CONFIG_FAIR_GROUP_SCHED
1615 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1618 * Calculate and set the cpu's group shares.
1621 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1622 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1624 unsigned long shares
;
1625 unsigned long rq_weight
;
1630 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1633 * \Sum shares * rq_weight
1634 * shares = -----------------------
1638 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1639 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1641 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1642 sysctl_sched_shares_thresh
) {
1643 struct rq
*rq
= cpu_rq(cpu
);
1644 unsigned long flags
;
1646 spin_lock_irqsave(&rq
->lock
, flags
);
1647 tg
->cfs_rq
[cpu
]->shares
= shares
;
1649 __set_se_shares(tg
->se
[cpu
], shares
);
1650 spin_unlock_irqrestore(&rq
->lock
, flags
);
1655 * Re-compute the task group their per cpu shares over the given domain.
1656 * This needs to be done in a bottom-up fashion because the rq weight of a
1657 * parent group depends on the shares of its child groups.
1659 static int tg_shares_up(struct task_group
*tg
, void *data
)
1661 unsigned long weight
, rq_weight
= 0;
1662 unsigned long shares
= 0;
1663 struct sched_domain
*sd
= data
;
1666 for_each_cpu(i
, sched_domain_span(sd
)) {
1668 * If there are currently no tasks on the cpu pretend there
1669 * is one of average load so that when a new task gets to
1670 * run here it will not get delayed by group starvation.
1672 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1674 weight
= NICE_0_LOAD
;
1676 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1677 rq_weight
+= weight
;
1678 shares
+= tg
->cfs_rq
[i
]->shares
;
1681 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1682 shares
= tg
->shares
;
1684 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1685 shares
= tg
->shares
;
1687 for_each_cpu(i
, sched_domain_span(sd
))
1688 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1694 * Compute the cpu's hierarchical load factor for each task group.
1695 * This needs to be done in a top-down fashion because the load of a child
1696 * group is a fraction of its parents load.
1698 static int tg_load_down(struct task_group
*tg
, void *data
)
1701 long cpu
= (long)data
;
1704 load
= cpu_rq(cpu
)->load
.weight
;
1706 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1707 load
*= tg
->cfs_rq
[cpu
]->shares
;
1708 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1711 tg
->cfs_rq
[cpu
]->h_load
= load
;
1716 static void update_shares(struct sched_domain
*sd
)
1718 u64 now
= cpu_clock(raw_smp_processor_id());
1719 s64 elapsed
= now
- sd
->last_update
;
1721 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1722 sd
->last_update
= now
;
1723 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1727 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1729 spin_unlock(&rq
->lock
);
1731 spin_lock(&rq
->lock
);
1734 static void update_h_load(long cpu
)
1736 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1741 static inline void update_shares(struct sched_domain
*sd
)
1745 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1751 #ifdef CONFIG_PREEMPT
1754 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1755 * way at the expense of forcing extra atomic operations in all
1756 * invocations. This assures that the double_lock is acquired using the
1757 * same underlying policy as the spinlock_t on this architecture, which
1758 * reduces latency compared to the unfair variant below. However, it
1759 * also adds more overhead and therefore may reduce throughput.
1761 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1762 __releases(this_rq
->lock
)
1763 __acquires(busiest
->lock
)
1764 __acquires(this_rq
->lock
)
1766 spin_unlock(&this_rq
->lock
);
1767 double_rq_lock(this_rq
, busiest
);
1774 * Unfair double_lock_balance: Optimizes throughput at the expense of
1775 * latency by eliminating extra atomic operations when the locks are
1776 * already in proper order on entry. This favors lower cpu-ids and will
1777 * grant the double lock to lower cpus over higher ids under contention,
1778 * regardless of entry order into the function.
1780 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1781 __releases(this_rq
->lock
)
1782 __acquires(busiest
->lock
)
1783 __acquires(this_rq
->lock
)
1787 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1788 if (busiest
< this_rq
) {
1789 spin_unlock(&this_rq
->lock
);
1790 spin_lock(&busiest
->lock
);
1791 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1794 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1799 #endif /* CONFIG_PREEMPT */
1802 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1804 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1806 if (unlikely(!irqs_disabled())) {
1807 /* printk() doesn't work good under rq->lock */
1808 spin_unlock(&this_rq
->lock
);
1812 return _double_lock_balance(this_rq
, busiest
);
1815 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1816 __releases(busiest
->lock
)
1818 spin_unlock(&busiest
->lock
);
1819 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1823 #ifdef CONFIG_FAIR_GROUP_SCHED
1824 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1827 cfs_rq
->shares
= shares
;
1832 static void calc_load_account_active(struct rq
*this_rq
);
1834 #include "sched_stats.h"
1835 #include "sched_idletask.c"
1836 #include "sched_fair.c"
1837 #include "sched_rt.c"
1838 #ifdef CONFIG_SCHED_DEBUG
1839 # include "sched_debug.c"
1842 #define sched_class_highest (&rt_sched_class)
1843 #define for_each_class(class) \
1844 for (class = sched_class_highest; class; class = class->next)
1846 static void inc_nr_running(struct rq
*rq
)
1851 static void dec_nr_running(struct rq
*rq
)
1856 static void set_load_weight(struct task_struct
*p
)
1858 if (task_has_rt_policy(p
)) {
1859 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1860 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1865 * SCHED_IDLE tasks get minimal weight:
1867 if (p
->policy
== SCHED_IDLE
) {
1868 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1869 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1873 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1874 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1877 static void update_avg(u64
*avg
, u64 sample
)
1879 s64 diff
= sample
- *avg
;
1883 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1886 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1888 sched_info_queued(p
);
1889 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1893 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1896 if (p
->se
.last_wakeup
) {
1897 update_avg(&p
->se
.avg_overlap
,
1898 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1899 p
->se
.last_wakeup
= 0;
1901 update_avg(&p
->se
.avg_wakeup
,
1902 sysctl_sched_wakeup_granularity
);
1906 sched_info_dequeued(p
);
1907 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1912 * __normal_prio - return the priority that is based on the static prio
1914 static inline int __normal_prio(struct task_struct
*p
)
1916 return p
->static_prio
;
1920 * Calculate the expected normal priority: i.e. priority
1921 * without taking RT-inheritance into account. Might be
1922 * boosted by interactivity modifiers. Changes upon fork,
1923 * setprio syscalls, and whenever the interactivity
1924 * estimator recalculates.
1926 static inline int normal_prio(struct task_struct
*p
)
1930 if (task_has_rt_policy(p
))
1931 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1933 prio
= __normal_prio(p
);
1935 // trace_special_pid(p->pid, PRIO(p), __PRIO(prio));
1940 * Calculate the current priority, i.e. the priority
1941 * taken into account by the scheduler. This value might
1942 * be boosted by RT tasks, or might be boosted by
1943 * interactivity modifiers. Will be RT if the task got
1944 * RT-boosted. If not then it returns p->normal_prio.
1946 static int effective_prio(struct task_struct
*p
)
1948 p
->normal_prio
= normal_prio(p
);
1950 * If we are RT tasks or we were boosted to RT priority,
1951 * keep the priority unchanged. Otherwise, update priority
1952 * to the normal priority:
1954 if (!rt_prio(p
->prio
))
1955 return p
->normal_prio
;
1960 * activate_task - move a task to the runqueue.
1962 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1964 if (task_contributes_to_load(p
))
1965 rq
->nr_uninterruptible
--;
1967 enqueue_task(rq
, p
, wakeup
);
1972 * deactivate_task - remove a task from the runqueue.
1974 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1976 if (task_contributes_to_load(p
))
1977 rq
->nr_uninterruptible
++;
1979 dequeue_task(rq
, p
, sleep
);
1984 * task_curr - is this task currently executing on a CPU?
1985 * @p: the task in question.
1987 inline int task_curr(const struct task_struct
*p
)
1989 return cpu_curr(task_cpu(p
)) == p
;
1992 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1994 set_task_rq(p
, cpu
);
1997 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1998 * successfuly executed on another CPU. We must ensure that updates of
1999 * per-task data have been completed by this moment.
2002 task_thread_info(p
)->cpu
= cpu
;
2006 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2007 const struct sched_class
*prev_class
,
2008 int oldprio
, int running
)
2010 if (prev_class
!= p
->sched_class
) {
2011 if (prev_class
->switched_from
)
2012 prev_class
->switched_from(rq
, p
, running
);
2013 p
->sched_class
->switched_to(rq
, p
, running
);
2015 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2020 /* Used instead of source_load when we know the type == 0 */
2021 static unsigned long weighted_cpuload(const int cpu
)
2023 return cpu_rq(cpu
)->load
.weight
;
2027 * Is this task likely cache-hot:
2030 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2035 * Buddy candidates are cache hot:
2037 if (sched_feat(CACHE_HOT_BUDDY
) &&
2038 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2039 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2042 if (p
->sched_class
!= &fair_sched_class
)
2045 if (sysctl_sched_migration_cost
== -1)
2047 if (sysctl_sched_migration_cost
== 0)
2050 delta
= now
- p
->se
.exec_start
;
2052 return delta
< (s64
)sysctl_sched_migration_cost
;
2056 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2058 int old_cpu
= task_cpu(p
);
2059 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2060 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2061 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2064 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2066 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
2068 #ifdef CONFIG_SCHEDSTATS
2069 if (p
->se
.wait_start
)
2070 p
->se
.wait_start
-= clock_offset
;
2071 if (p
->se
.sleep_start
)
2072 p
->se
.sleep_start
-= clock_offset
;
2073 if (p
->se
.block_start
)
2074 p
->se
.block_start
-= clock_offset
;
2076 if (old_cpu
!= new_cpu
) {
2077 p
->se
.nr_migrations
++;
2078 new_rq
->nr_migrations_in
++;
2079 #ifdef CONFIG_SCHEDSTATS
2080 if (task_hot(p
, old_rq
->clock
, NULL
))
2081 schedstat_inc(p
, se
.nr_forced2_migrations
);
2084 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2085 new_cfsrq
->min_vruntime
;
2087 __set_task_cpu(p
, new_cpu
);
2090 struct migration_req
{
2091 struct list_head list
;
2093 struct task_struct
*task
;
2096 struct completion done
;
2100 * The task's runqueue lock must be held.
2101 * Returns true if you have to wait for migration thread.
2104 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2106 struct rq
*rq
= task_rq(p
);
2109 * If the task is not on a runqueue (and not running), then
2110 * it is sufficient to simply update the task's cpu field.
2112 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2113 set_task_cpu(p
, dest_cpu
);
2117 init_completion(&req
->done
);
2119 req
->dest_cpu
= dest_cpu
;
2120 list_add(&req
->list
, &rq
->migration_queue
);
2126 * wait_task_inactive - wait for a thread to unschedule.
2128 * If @match_state is nonzero, it's the @p->state value just checked and
2129 * not expected to change. If it changes, i.e. @p might have woken up,
2130 * then return zero. When we succeed in waiting for @p to be off its CPU,
2131 * we return a positive number (its total switch count). If a second call
2132 * a short while later returns the same number, the caller can be sure that
2133 * @p has remained unscheduled the whole time.
2135 * The caller must ensure that the task *will* unschedule sometime soon,
2136 * else this function might spin for a *long* time. This function can't
2137 * be called with interrupts off, or it may introduce deadlock with
2138 * smp_call_function() if an IPI is sent by the same process we are
2139 * waiting to become inactive.
2141 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2143 unsigned long flags
;
2150 * We do the initial early heuristics without holding
2151 * any task-queue locks at all. We'll only try to get
2152 * the runqueue lock when things look like they will
2158 * If the task is actively running on another CPU
2159 * still, just relax and busy-wait without holding
2162 * NOTE! Since we don't hold any locks, it's not
2163 * even sure that "rq" stays as the right runqueue!
2164 * But we don't care, since "task_running()" will
2165 * return false if the runqueue has changed and p
2166 * is actually now running somewhere else!
2168 while (task_running(rq
, p
)) {
2169 if (match_state
&& unlikely(p
->state
!= match_state
))
2175 * Ok, time to look more closely! We need the rq
2176 * lock now, to be *sure*. If we're wrong, we'll
2177 * just go back and repeat.
2179 rq
= task_rq_lock(p
, &flags
);
2180 trace_sched_wait_task(rq
, p
);
2181 running
= task_running(rq
, p
);
2182 on_rq
= p
->se
.on_rq
;
2184 if (!match_state
|| p
->state
== match_state
)
2185 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2186 task_rq_unlock(rq
, &flags
);
2189 * If it changed from the expected state, bail out now.
2191 if (unlikely(!ncsw
))
2195 * Was it really running after all now that we
2196 * checked with the proper locks actually held?
2198 * Oops. Go back and try again..
2200 if (unlikely(running
)) {
2206 * It's not enough that it's not actively running,
2207 * it must be off the runqueue _entirely_, and not
2210 * So if it was still runnable (but just not actively
2211 * running right now), it's preempted, and we should
2212 * yield - it could be a while.
2214 if (unlikely(on_rq
)) {
2215 schedule_timeout_uninterruptible(1);
2220 * Ahh, all good. It wasn't running, and it wasn't
2221 * runnable, which means that it will never become
2222 * running in the future either. We're all done!
2231 * kick_process - kick a running thread to enter/exit the kernel
2232 * @p: the to-be-kicked thread
2234 * Cause a process which is running on another CPU to enter
2235 * kernel-mode, without any delay. (to get signals handled.)
2237 * NOTE: this function doesnt have to take the runqueue lock,
2238 * because all it wants to ensure is that the remote task enters
2239 * the kernel. If the IPI races and the task has been migrated
2240 * to another CPU then no harm is done and the purpose has been
2243 void kick_process(struct task_struct
*p
)
2249 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2250 smp_send_reschedule(cpu
);
2255 * Return a low guess at the load of a migration-source cpu weighted
2256 * according to the scheduling class and "nice" value.
2258 * We want to under-estimate the load of migration sources, to
2259 * balance conservatively.
2261 static unsigned long source_load(int cpu
, int type
)
2263 struct rq
*rq
= cpu_rq(cpu
);
2264 unsigned long total
= weighted_cpuload(cpu
);
2266 if (type
== 0 || !sched_feat(LB_BIAS
))
2269 return min(rq
->cpu_load
[type
-1], total
);
2273 * Return a high guess at the load of a migration-target cpu weighted
2274 * according to the scheduling class and "nice" value.
2276 static unsigned long target_load(int cpu
, int type
)
2278 struct rq
*rq
= cpu_rq(cpu
);
2279 unsigned long total
= weighted_cpuload(cpu
);
2281 if (type
== 0 || !sched_feat(LB_BIAS
))
2284 return max(rq
->cpu_load
[type
-1], total
);
2288 * find_idlest_group finds and returns the least busy CPU group within the
2291 static struct sched_group
*
2292 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2294 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2295 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2296 int load_idx
= sd
->forkexec_idx
;
2297 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2300 unsigned long load
, avg_load
;
2304 /* Skip over this group if it has no CPUs allowed */
2305 if (!cpumask_intersects(sched_group_cpus(group
),
2309 local_group
= cpumask_test_cpu(this_cpu
,
2310 sched_group_cpus(group
));
2312 /* Tally up the load of all CPUs in the group */
2315 for_each_cpu(i
, sched_group_cpus(group
)) {
2316 /* Bias balancing toward cpus of our domain */
2318 load
= source_load(i
, load_idx
);
2320 load
= target_load(i
, load_idx
);
2325 /* Adjust by relative CPU power of the group */
2326 avg_load
= sg_div_cpu_power(group
,
2327 avg_load
* SCHED_LOAD_SCALE
);
2330 this_load
= avg_load
;
2332 } else if (avg_load
< min_load
) {
2333 min_load
= avg_load
;
2336 } while (group
= group
->next
, group
!= sd
->groups
);
2338 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2344 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2347 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2349 unsigned long load
, min_load
= ULONG_MAX
;
2353 /* Traverse only the allowed CPUs */
2354 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2355 load
= weighted_cpuload(i
);
2357 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2367 * sched_balance_self: balance the current task (running on cpu) in domains
2368 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2371 * Balance, ie. select the least loaded group.
2373 * Returns the target CPU number, or the same CPU if no balancing is needed.
2375 * preempt must be disabled.
2377 static int sched_balance_self(int cpu
, int flag
)
2379 struct task_struct
*t
= current
;
2380 struct sched_domain
*tmp
, *sd
= NULL
;
2382 for_each_domain(cpu
, tmp
) {
2384 * If power savings logic is enabled for a domain, stop there.
2386 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2388 if (tmp
->flags
& flag
)
2396 struct sched_group
*group
;
2397 int new_cpu
, weight
;
2399 if (!(sd
->flags
& flag
)) {
2404 group
= find_idlest_group(sd
, t
, cpu
);
2410 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2411 if (new_cpu
== -1 || new_cpu
== cpu
) {
2412 /* Now try balancing at a lower domain level of cpu */
2417 /* Now try balancing at a lower domain level of new_cpu */
2419 weight
= cpumask_weight(sched_domain_span(sd
));
2421 for_each_domain(cpu
, tmp
) {
2422 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2424 if (tmp
->flags
& flag
)
2427 /* while loop will break here if sd == NULL */
2433 #endif /* CONFIG_SMP */
2435 #ifdef CONFIG_DEBUG_PREEMPT
2436 void notrace
preempt_enable_no_resched(void)
2438 static int once
= 1;
2441 dec_preempt_count();
2443 if (once
&& !preempt_count()) {
2445 printk(KERN_ERR
"BUG: %s:%d task might have lost a preemption check!\n",
2446 current
->comm
, current
->pid
);
2451 EXPORT_SYMBOL(preempt_enable_no_resched
);
2456 * task_oncpu_function_call - call a function on the cpu on which a task runs
2457 * @p: the task to evaluate
2458 * @func: the function to be called
2459 * @info: the function call argument
2461 * Calls the function @func when the task is currently running. This might
2462 * be on the current CPU, which just calls the function directly
2464 void task_oncpu_function_call(struct task_struct
*p
,
2465 void (*func
) (void *info
), void *info
)
2472 smp_call_function_single(cpu
, func
, info
, 1);
2477 * try_to_wake_up - wake up a thread
2478 * @p: the to-be-woken-up thread
2479 * @state: the mask of task states that can be woken
2480 * @sync: do a synchronous wakeup?
2482 * Put it on the run-queue if it's not already there. The "current"
2483 * thread is always on the run-queue (except when the actual
2484 * re-schedule is in progress), and as such you're allowed to do
2485 * the simpler "current->state = TASK_RUNNING" to mark yourself
2486 * runnable without the overhead of this.
2488 * returns failure only if the task is already active.
2491 try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
, int mutex
)
2493 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2494 unsigned long flags
;
2498 if (!sched_feat(SYNC_WAKEUPS
))
2502 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2503 struct sched_domain
*sd
;
2505 this_cpu
= raw_smp_processor_id();
2508 for_each_domain(this_cpu
, sd
) {
2509 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2517 #ifdef CONFIG_PREEMPT_RT
2519 * sync wakeups can increase wakeup latencies:
2525 rq
= task_rq_lock(p
, &flags
);
2526 update_rq_clock(rq
);
2527 old_state
= p
->state
;
2528 if (!(old_state
& state
))
2536 this_cpu
= smp_processor_id();
2539 if (unlikely(task_running(rq
, p
)))
2542 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2543 if (cpu
!= orig_cpu
) {
2544 set_task_cpu(p
, cpu
);
2545 task_rq_unlock(rq
, &flags
);
2546 /* might preempt at this point */
2547 rq
= task_rq_lock(p
, &flags
);
2548 old_state
= p
->state
;
2549 if (!(old_state
& state
))
2554 this_cpu
= smp_processor_id();
2558 #ifdef CONFIG_SCHEDSTATS
2559 schedstat_inc(rq
, ttwu_count
);
2560 if (cpu
== this_cpu
)
2561 schedstat_inc(rq
, ttwu_local
);
2563 struct sched_domain
*sd
;
2564 for_each_domain(this_cpu
, sd
) {
2565 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2566 schedstat_inc(sd
, ttwu_wake_remote
);
2571 #endif /* CONFIG_SCHEDSTATS */
2574 #endif /* CONFIG_SMP */
2575 schedstat_inc(p
, se
.nr_wakeups
);
2577 schedstat_inc(p
, se
.nr_wakeups_sync
);
2578 if (orig_cpu
!= cpu
)
2579 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2580 if (cpu
== this_cpu
)
2581 schedstat_inc(p
, se
.nr_wakeups_local
);
2583 schedstat_inc(p
, se
.nr_wakeups_remote
);
2584 activate_task(rq
, p
, 1);
2588 * Only attribute actual wakeups done by this task.
2590 if (!in_interrupt()) {
2591 struct sched_entity
*se
= ¤t
->se
;
2592 u64 sample
= se
->sum_exec_runtime
;
2594 if (se
->last_wakeup
)
2595 sample
-= se
->last_wakeup
;
2597 sample
-= se
->start_runtime
;
2598 update_avg(&se
->avg_wakeup
, sample
);
2600 se
->last_wakeup
= se
->sum_exec_runtime
;
2604 trace_sched_wakeup(rq
, p
, success
);
2605 check_preempt_curr(rq
, p
, sync
);
2608 * For a mutex wakeup we or TASK_RUNNING_MUTEX to the task
2609 * state to preserve the original state, so a real wakeup
2610 * still can see the (UN)INTERRUPTIBLE bits in the state check
2611 * above. We dont have to worry about the | TASK_RUNNING_MUTEX
2612 * here. The waiter is serialized by the mutex lock and nobody
2613 * else can fiddle with p->state as we hold rq lock.
2616 p
->state
|= TASK_RUNNING_MUTEX
;
2618 p
->state
= TASK_RUNNING
;
2620 if (p
->sched_class
->task_wake_up
)
2621 p
->sched_class
->task_wake_up(rq
, p
);
2624 task_rq_unlock(rq
, &flags
);
2629 int wake_up_process(struct task_struct
*p
)
2631 return try_to_wake_up(p
, TASK_ALL
, 0, 0);
2633 EXPORT_SYMBOL(wake_up_process
);
2635 int wake_up_process_sync(struct task_struct
* p
)
2637 return try_to_wake_up(p
, TASK_ALL
, 1, 0);
2639 EXPORT_SYMBOL(wake_up_process_sync
);
2641 int wake_up_process_mutex(struct task_struct
* p
)
2643 return try_to_wake_up(p
, TASK_ALL
, 0, 1);
2645 EXPORT_SYMBOL(wake_up_process_mutex
);
2647 int wake_up_process_mutex_sync(struct task_struct
* p
)
2649 return try_to_wake_up(p
, TASK_ALL
, 1, 1);
2651 EXPORT_SYMBOL(wake_up_process_mutex_sync
);
2653 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2655 return try_to_wake_up(p
, state
, 0, 0);
2659 * Perform scheduler related setup for a newly forked process p.
2660 * p is forked by current.
2662 * __sched_fork() is basic setup used by init_idle() too:
2664 static void __sched_fork(struct task_struct
*p
)
2666 p
->se
.exec_start
= 0;
2667 p
->se
.sum_exec_runtime
= 0;
2668 p
->se
.prev_sum_exec_runtime
= 0;
2669 p
->se
.nr_migrations
= 0;
2670 p
->se
.last_wakeup
= 0;
2671 p
->se
.avg_overlap
= 0;
2672 p
->se
.start_runtime
= 0;
2673 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2675 #ifdef CONFIG_SCHEDSTATS
2676 p
->se
.wait_start
= 0;
2677 p
->se
.sum_sleep_runtime
= 0;
2678 p
->se
.sleep_start
= 0;
2679 p
->se
.block_start
= 0;
2680 p
->se
.sleep_max
= 0;
2681 p
->se
.block_max
= 0;
2683 p
->se
.slice_max
= 0;
2687 INIT_LIST_HEAD(&p
->rt
.run_list
);
2689 INIT_LIST_HEAD(&p
->se
.group_node
);
2691 #ifdef CONFIG_PREEMPT_NOTIFIERS
2692 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2696 * We mark the process as running here, but have not actually
2697 * inserted it onto the runqueue yet. This guarantees that
2698 * nobody will actually run it, and a signal or other external
2699 * event cannot wake it up and insert it on the runqueue either.
2701 p
->state
= TASK_RUNNING
;
2705 * fork()/clone()-time setup:
2707 void sched_fork(struct task_struct
*p
, int clone_flags
)
2709 int cpu
= get_cpu();
2714 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2716 set_task_cpu(p
, cpu
);
2719 * Make sure we do not leak PI boosting priority to the child:
2721 p
->prio
= current
->normal_prio
;
2722 if (!rt_prio(p
->prio
))
2723 p
->sched_class
= &fair_sched_class
;
2725 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2726 if (likely(sched_info_on()))
2727 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2729 #if defined(CONFIG_SMP)
2732 #ifdef CONFIG_PREEMPT
2733 /* Want to start with kernel preemption disabled. */
2734 task_thread_info(p
)->preempt_count
= 1;
2736 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2742 * wake_up_new_task - wake up a newly created task for the first time.
2744 * This function will do some initial scheduler statistics housekeeping
2745 * that must be done for every newly created context, then puts the task
2746 * on the runqueue and wakes it.
2748 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2750 unsigned long flags
;
2753 rq
= task_rq_lock(p
, &flags
);
2754 BUG_ON(p
->state
!= TASK_RUNNING
);
2755 update_rq_clock(rq
);
2757 p
->prio
= effective_prio(p
);
2759 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2760 activate_task(rq
, p
, 0);
2763 * Let the scheduling class do new task startup
2764 * management (if any):
2766 p
->sched_class
->task_new(rq
, p
);
2769 trace_sched_wakeup_new(rq
, p
, 1);
2770 check_preempt_curr(rq
, p
, 0);
2772 if (p
->sched_class
->task_wake_up
)
2773 p
->sched_class
->task_wake_up(rq
, p
);
2775 task_rq_unlock(rq
, &flags
);
2778 #ifdef CONFIG_PREEMPT_NOTIFIERS
2781 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2782 * @notifier: notifier struct to register
2784 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2786 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2788 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2791 * preempt_notifier_unregister - no longer interested in preemption notifications
2792 * @notifier: notifier struct to unregister
2794 * This is safe to call from within a preemption notifier.
2796 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2798 hlist_del(¬ifier
->link
);
2800 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2802 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2804 struct preempt_notifier
*notifier
;
2805 struct hlist_node
*node
;
2807 if (hlist_empty(&curr
->preempt_notifiers
))
2811 * The KVM sched in notifier expects to be called with
2812 * interrupts enabled.
2815 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2816 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2817 local_irq_disable();
2821 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2822 struct task_struct
*next
)
2824 struct preempt_notifier
*notifier
;
2825 struct hlist_node
*node
;
2827 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2828 notifier
->ops
->sched_out(notifier
, next
);
2831 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2833 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2838 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2839 struct task_struct
*next
)
2843 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2846 * prepare_task_switch - prepare to switch tasks
2847 * @rq: the runqueue preparing to switch
2848 * @prev: the current task that is being switched out
2849 * @next: the task we are going to switch to.
2851 * This is called with the rq lock held and interrupts off. It must
2852 * be paired with a subsequent finish_task_switch after the context
2855 * prepare_task_switch sets up locking and calls architecture specific
2859 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2860 struct task_struct
*next
)
2862 fire_sched_out_preempt_notifiers(prev
, next
);
2863 prepare_lock_switch(rq
, next
);
2864 prepare_arch_switch(next
);
2868 * finish_task_switch - clean up after a task-switch
2869 * @rq: runqueue associated with task-switch
2870 * @prev: the thread we just switched away from.
2872 * finish_task_switch must be called after the context switch, paired
2873 * with a prepare_task_switch call before the context switch.
2874 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2875 * and do any other architecture-specific cleanup actions.
2877 * Note that we may have delayed dropping an mm in context_switch(). If
2878 * so, we finish that here outside of the runqueue lock. (Doing it
2879 * with the lock held can cause deadlocks; see schedule() for
2882 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2883 __releases(rq
->lock
)
2885 struct mm_struct
*mm
= rq
->prev_mm
;
2888 int post_schedule
= 0;
2890 if (current
->sched_class
->needs_post_schedule
)
2891 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2897 * A task struct has one reference for the use as "current".
2898 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2899 * schedule one last time. The schedule call will never return, and
2900 * the scheduled task must drop that reference.
2901 * The test for TASK_DEAD must occur while the runqueue locks are
2902 * still held, otherwise prev could be scheduled on another cpu, die
2903 * there before we look at prev->state, and then the reference would
2905 * Manfred Spraul <manfred@colorfullife.com>
2907 prev_state
= prev
->state
;
2908 _finish_arch_switch(prev
);
2909 perf_counter_task_sched_in(current
, cpu_of(rq
));
2910 finish_lock_switch(rq
, prev
);
2913 current
->sched_class
->post_schedule(rq
);
2916 fire_sched_in_preempt_notifiers(current
);
2918 * Delay the final freeing of the mm or task, so that we dont have
2919 * to do complex work from within the scheduler:
2923 if (unlikely(prev_state
== TASK_DEAD
)) {
2925 * Remove function-return probe instances associated with this
2926 * task and put them back on the free list.
2928 kprobe_flush_task(prev
);
2929 put_task_struct(prev
);
2934 * schedule_tail - first thing a freshly forked thread must call.
2935 * @prev: the thread we just switched away from.
2937 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2938 __releases(rq
->lock
)
2941 finish_task_switch(this_rq(), prev
);
2942 __preempt_enable_no_resched();
2944 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2945 /* In this case, finish_task_switch does not reenable preemption */
2948 preempt_check_resched();
2950 if (current
->set_child_tid
)
2951 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2955 * context_switch - switch to the new MM and the new
2956 * thread's register state.
2959 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2960 struct task_struct
*next
)
2962 struct mm_struct
*mm
, *oldmm
;
2964 prepare_task_switch(rq
, prev
, next
);
2965 trace_sched_switch(rq
, prev
, next
);
2967 oldmm
= prev
->active_mm
;
2969 * For paravirt, this is coupled with an exit in switch_to to
2970 * combine the page table reload and the switch backend into
2973 arch_enter_lazy_cpu_mode();
2975 if (unlikely(!mm
)) {
2976 next
->active_mm
= oldmm
;
2977 atomic_inc(&oldmm
->mm_count
);
2978 enter_lazy_tlb(oldmm
, next
);
2980 switch_mm(oldmm
, mm
, next
);
2982 if (unlikely(!prev
->mm
)) {
2983 prev
->active_mm
= NULL
;
2984 rq
->prev_mm
= oldmm
;
2987 * Since the runqueue lock will be released by the next
2988 * task (which is an invalid locking op but in the case
2989 * of the scheduler it's an obvious special-case), so we
2990 * do an early lockdep release here:
2992 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2993 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2998 *current_ptr
= next
;
2999 *current_ti_ptr
= next
->thread_info
;
3001 /* Here we just switch the register state and the stack. */
3002 switch_to(prev
, next
, prev
);
3006 * this_rq must be evaluated again because prev may have moved
3007 * CPUs since it called schedule(), thus the 'rq' on its stack
3008 * frame will be invalid.
3010 finish_task_switch(this_rq(), prev
);
3014 * nr_running, nr_uninterruptible and nr_context_switches:
3016 * externally visible scheduler statistics: current number of runnable
3017 * threads, current number of uninterruptible-sleeping threads, total
3018 * number of context switches performed since bootup.
3020 unsigned long nr_running(void)
3022 unsigned long i
, sum
= 0;
3024 for_each_online_cpu(i
)
3025 sum
+= cpu_rq(i
)->nr_running
;
3030 unsigned long nr_uninterruptible(void)
3032 unsigned long i
, sum
= 0;
3034 for_each_possible_cpu(i
)
3035 sum
+= cpu_rq(i
)->nr_uninterruptible
;
3038 * Since we read the counters lockless, it might be slightly
3039 * inaccurate. Do not allow it to go below zero though:
3041 if (unlikely((long)sum
< 0))
3047 unsigned long nr_uninterruptible_cpu(int cpu
)
3049 return cpu_rq(cpu
)->nr_uninterruptible
;
3052 unsigned long long nr_context_switches(void)
3055 unsigned long long sum
= 0;
3057 for_each_possible_cpu(i
)
3058 sum
+= cpu_rq(i
)->nr_switches
;
3063 unsigned long nr_iowait(void)
3065 unsigned long i
, sum
= 0;
3067 for_each_possible_cpu(i
)
3068 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3071 * Since we read the counters lockless, it might be slightly
3072 * inaccurate. Do not allow it to go below zero though:
3074 if (unlikely((long)sum
< 0))
3080 /* Variables and functions for calc_load */
3081 static atomic_long_t calc_load_tasks
;
3082 static unsigned long calc_load_update
;
3083 unsigned long avenrun
[3];
3084 EXPORT_SYMBOL(avenrun
);
3087 * get_avenrun - get the load average array
3088 * @loads: pointer to dest load array
3089 * @offset: offset to add
3090 * @shift: shift count to shift the result left
3092 * These values are estimates at best, so no need for locking.
3094 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3096 loads
[0] = (avenrun
[0] + offset
) << shift
;
3097 loads
[1] = (avenrun
[1] + offset
) << shift
;
3098 loads
[2] = (avenrun
[2] + offset
) << shift
;
3101 static unsigned long
3102 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3105 load
+= active
* (FIXED_1
- exp
);
3106 return load
>> FSHIFT
;
3110 * calc_load - update the avenrun load estimates 10 ticks after the
3111 * CPUs have updated calc_load_tasks.
3113 void calc_global_load(void)
3115 unsigned long upd
= calc_load_update
+ 10;
3118 if (time_before(jiffies
, upd
))
3121 active
= atomic_long_read(&calc_load_tasks
);
3122 active
= active
> 0 ? active
* FIXED_1
: 0;
3124 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3125 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3126 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3128 calc_load_update
+= LOAD_FREQ
;
3132 * Either called from update_cpu_load() or from a cpu going idle
3134 static void calc_load_account_active(struct rq
*this_rq
)
3136 long nr_active
, delta
;
3138 nr_active
= this_rq
->nr_running
;
3139 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3141 if (nr_active
!= this_rq
->calc_load_active
) {
3142 delta
= nr_active
- this_rq
->calc_load_active
;
3143 this_rq
->calc_load_active
= nr_active
;
3144 atomic_long_add(delta
, &calc_load_tasks
);
3149 * Externally visible per-cpu scheduler statistics:
3150 * cpu_nr_migrations(cpu) - number of migrations into that cpu
3152 u64
cpu_nr_migrations(int cpu
)
3154 return cpu_rq(cpu
)->nr_migrations_in
;
3158 * Update rq->cpu_load[] statistics. This function is usually called every
3159 * scheduler tick (TICK_NSEC).
3161 static void update_cpu_load(struct rq
*this_rq
)
3163 unsigned long this_load
= this_rq
->load
.weight
;
3166 this_rq
->nr_load_updates
++;
3168 /* Update our load: */
3169 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3170 unsigned long old_load
, new_load
;
3172 /* scale is effectively 1 << i now, and >> i divides by scale */
3174 old_load
= this_rq
->cpu_load
[i
];
3175 new_load
= this_load
;
3177 * Round up the averaging division if load is increasing. This
3178 * prevents us from getting stuck on 9 if the load is 10, for
3181 if (new_load
> old_load
)
3182 new_load
+= scale
-1;
3183 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3186 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3187 this_rq
->calc_load_update
+= LOAD_FREQ
;
3188 calc_load_account_active(this_rq
);
3195 * double_rq_lock - safely lock two runqueues
3197 * Note this does not disable interrupts like task_rq_lock,
3198 * you need to do so manually before calling.
3200 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3201 __acquires(rq1
->lock
)
3202 __acquires(rq2
->lock
)
3204 BUG_ON(!irqs_disabled());
3206 spin_lock(&rq1
->lock
);
3207 __acquire(rq2
->lock
); /* Fake it out ;) */
3210 spin_lock(&rq1
->lock
);
3211 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3213 spin_lock(&rq2
->lock
);
3214 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3217 update_rq_clock(rq1
);
3218 update_rq_clock(rq2
);
3222 * double_rq_unlock - safely unlock two runqueues
3224 * Note this does not restore interrupts like task_rq_unlock,
3225 * you need to do so manually after calling.
3227 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3228 __releases(rq1
->lock
)
3229 __releases(rq2
->lock
)
3231 spin_unlock(&rq1
->lock
);
3233 spin_unlock(&rq2
->lock
);
3235 __release(rq2
->lock
);
3239 * If dest_cpu is allowed for this process, migrate the task to it.
3240 * This is accomplished by forcing the cpu_allowed mask to only
3241 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3242 * the cpu_allowed mask is restored.
3244 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3246 struct migration_req req
;
3247 unsigned long flags
;
3250 rq
= task_rq_lock(p
, &flags
);
3251 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3252 || unlikely(!cpu_active(dest_cpu
)))
3255 /* force the process onto the specified CPU */
3256 if (migrate_task(p
, dest_cpu
, &req
)) {
3257 /* Need to wait for migration thread (might exit: take ref). */
3258 struct task_struct
*mt
= rq
->migration_thread
;
3260 get_task_struct(mt
);
3261 task_rq_unlock(rq
, &flags
);
3262 wake_up_process(mt
);
3263 put_task_struct(mt
);
3264 wait_for_completion(&req
.done
);
3269 task_rq_unlock(rq
, &flags
);
3273 * sched_exec - execve() is a valuable balancing opportunity, because at
3274 * this point the task has the smallest effective memory and cache footprint.
3276 void sched_exec(void)
3278 int new_cpu
, this_cpu
= get_cpu();
3279 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3281 if (new_cpu
!= this_cpu
)
3282 sched_migrate_task(current
, new_cpu
);
3286 * pull_task - move a task from a remote runqueue to the local runqueue.
3287 * Both runqueues must be locked.
3289 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3290 struct rq
*this_rq
, int this_cpu
)
3292 deactivate_task(src_rq
, p
, 0);
3293 set_task_cpu(p
, this_cpu
);
3294 activate_task(this_rq
, p
, 0);
3296 * Note that idle threads have a prio of MAX_PRIO, for this test
3297 * to be always true for them.
3299 check_preempt_curr(this_rq
, p
, 0);
3303 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3306 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3307 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3310 int tsk_cache_hot
= 0;
3312 * We do not migrate tasks that are:
3313 * 1) running (obviously), or
3314 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3315 * 3) are cache-hot on their current CPU.
3317 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3318 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3323 if (task_running(rq
, p
)) {
3324 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3329 * Aggressive migration if:
3330 * 1) task is cache cold, or
3331 * 2) too many balance attempts have failed.
3334 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3335 if (!tsk_cache_hot
||
3336 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3337 #ifdef CONFIG_SCHEDSTATS
3338 if (tsk_cache_hot
) {
3339 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3340 schedstat_inc(p
, se
.nr_forced_migrations
);
3346 if (tsk_cache_hot
) {
3347 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3353 static unsigned long
3354 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3355 unsigned long max_load_move
, struct sched_domain
*sd
,
3356 enum cpu_idle_type idle
, int *all_pinned
,
3357 int *this_best_prio
, struct rq_iterator
*iterator
)
3359 int loops
= 0, pulled
= 0, pinned
= 0;
3360 struct task_struct
*p
;
3361 long rem_load_move
= max_load_move
;
3363 if (max_load_move
== 0)
3369 * Start the load-balancing iterator:
3371 p
= iterator
->start(iterator
->arg
);
3373 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3376 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3377 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3378 p
= iterator
->next(iterator
->arg
);
3382 pull_task(busiest
, p
, this_rq
, this_cpu
);
3384 rem_load_move
-= p
->se
.load
.weight
;
3386 #ifdef CONFIG_PREEMPT
3388 * NEWIDLE balancing is a source of latency, so preemptible kernels
3389 * will stop after the first task is pulled to minimize the critical
3392 if (idle
== CPU_NEWLY_IDLE
)
3397 * We only want to steal up to the prescribed amount of weighted load.
3399 if (rem_load_move
> 0) {
3400 if (p
->prio
< *this_best_prio
)
3401 *this_best_prio
= p
->prio
;
3402 p
= iterator
->next(iterator
->arg
);
3407 * Right now, this is one of only two places pull_task() is called,
3408 * so we can safely collect pull_task() stats here rather than
3409 * inside pull_task().
3411 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3414 *all_pinned
= pinned
;
3416 return max_load_move
- rem_load_move
;
3420 * move_tasks tries to move up to max_load_move weighted load from busiest to
3421 * this_rq, as part of a balancing operation within domain "sd".
3422 * Returns 1 if successful and 0 otherwise.
3424 * Called with both runqueues locked.
3426 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3427 unsigned long max_load_move
,
3428 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3431 const struct sched_class
*class = sched_class_highest
;
3432 unsigned long total_load_moved
= 0;
3433 int this_best_prio
= this_rq
->curr
->prio
;
3437 class->load_balance(this_rq
, this_cpu
, busiest
,
3438 max_load_move
- total_load_moved
,
3439 sd
, idle
, all_pinned
, &this_best_prio
);
3440 class = class->next
;
3442 #ifdef CONFIG_PREEMPT
3444 * NEWIDLE balancing is a source of latency, so preemptible
3445 * kernels will stop after the first task is pulled to minimize
3446 * the critical section.
3448 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3451 } while (class && max_load_move
> total_load_moved
);
3453 return total_load_moved
> 0;
3457 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3458 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3459 struct rq_iterator
*iterator
)
3461 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3465 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3466 pull_task(busiest
, p
, this_rq
, this_cpu
);
3468 * Right now, this is only the second place pull_task()
3469 * is called, so we can safely collect pull_task()
3470 * stats here rather than inside pull_task().
3472 schedstat_inc(sd
, lb_gained
[idle
]);
3476 p
= iterator
->next(iterator
->arg
);
3483 * move_one_task tries to move exactly one task from busiest to this_rq, as
3484 * part of active balancing operations within "domain".
3485 * Returns 1 if successful and 0 otherwise.
3487 * Called with both runqueues locked.
3489 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3490 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3492 const struct sched_class
*class;
3494 for (class = sched_class_highest
; class; class = class->next
)
3495 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3500 /********** Helpers for find_busiest_group ************************/
3502 * sd_lb_stats - Structure to store the statistics of a sched_domain
3503 * during load balancing.
3505 struct sd_lb_stats
{
3506 struct sched_group
*busiest
; /* Busiest group in this sd */
3507 struct sched_group
*this; /* Local group in this sd */
3508 unsigned long total_load
; /* Total load of all groups in sd */
3509 unsigned long total_pwr
; /* Total power of all groups in sd */
3510 unsigned long avg_load
; /* Average load across all groups in sd */
3512 /** Statistics of this group */
3513 unsigned long this_load
;
3514 unsigned long this_load_per_task
;
3515 unsigned long this_nr_running
;
3517 /* Statistics of the busiest group */
3518 unsigned long max_load
;
3519 unsigned long busiest_load_per_task
;
3520 unsigned long busiest_nr_running
;
3522 int group_imb
; /* Is there imbalance in this sd */
3523 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3524 int power_savings_balance
; /* Is powersave balance needed for this sd */
3525 struct sched_group
*group_min
; /* Least loaded group in sd */
3526 struct sched_group
*group_leader
; /* Group which relieves group_min */
3527 unsigned long min_load_per_task
; /* load_per_task in group_min */
3528 unsigned long leader_nr_running
; /* Nr running of group_leader */
3529 unsigned long min_nr_running
; /* Nr running of group_min */
3534 * sg_lb_stats - stats of a sched_group required for load_balancing
3536 struct sg_lb_stats
{
3537 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3538 unsigned long group_load
; /* Total load over the CPUs of the group */
3539 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3540 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3541 unsigned long group_capacity
;
3542 int group_imb
; /* Is there an imbalance in the group ? */
3546 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3547 * @group: The group whose first cpu is to be returned.
3549 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3551 return cpumask_first(sched_group_cpus(group
));
3555 * get_sd_load_idx - Obtain the load index for a given sched domain.
3556 * @sd: The sched_domain whose load_idx is to be obtained.
3557 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3559 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3560 enum cpu_idle_type idle
)
3566 load_idx
= sd
->busy_idx
;
3569 case CPU_NEWLY_IDLE
:
3570 load_idx
= sd
->newidle_idx
;
3573 load_idx
= sd
->idle_idx
;
3581 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3583 * init_sd_power_savings_stats - Initialize power savings statistics for
3584 * the given sched_domain, during load balancing.
3586 * @sd: Sched domain whose power-savings statistics are to be initialized.
3587 * @sds: Variable containing the statistics for sd.
3588 * @idle: Idle status of the CPU at which we're performing load-balancing.
3590 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3591 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3594 * Busy processors will not participate in power savings
3597 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3598 sds
->power_savings_balance
= 0;
3600 sds
->power_savings_balance
= 1;
3601 sds
->min_nr_running
= ULONG_MAX
;
3602 sds
->leader_nr_running
= 0;
3607 * update_sd_power_savings_stats - Update the power saving stats for a
3608 * sched_domain while performing load balancing.
3610 * @group: sched_group belonging to the sched_domain under consideration.
3611 * @sds: Variable containing the statistics of the sched_domain
3612 * @local_group: Does group contain the CPU for which we're performing
3614 * @sgs: Variable containing the statistics of the group.
3616 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3617 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3620 if (!sds
->power_savings_balance
)
3624 * If the local group is idle or completely loaded
3625 * no need to do power savings balance at this domain
3627 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3628 !sds
->this_nr_running
))
3629 sds
->power_savings_balance
= 0;
3632 * If a group is already running at full capacity or idle,
3633 * don't include that group in power savings calculations
3635 if (!sds
->power_savings_balance
||
3636 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3637 !sgs
->sum_nr_running
)
3641 * Calculate the group which has the least non-idle load.
3642 * This is the group from where we need to pick up the load
3645 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3646 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3647 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3648 sds
->group_min
= group
;
3649 sds
->min_nr_running
= sgs
->sum_nr_running
;
3650 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3651 sgs
->sum_nr_running
;
3655 * Calculate the group which is almost near its
3656 * capacity but still has some space to pick up some load
3657 * from other group and save more power
3659 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3662 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3663 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3664 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3665 sds
->group_leader
= group
;
3666 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3671 * check_power_save_busiest_group - Check if we have potential to perform
3672 * some power-savings balance. If yes, set the busiest group to be
3673 * the least loaded group in the sched_domain, so that it's CPUs can
3676 * @sds: Variable containing the statistics of the sched_domain
3677 * under consideration.
3678 * @this_cpu: Cpu at which we're currently performing load-balancing.
3679 * @imbalance: Variable to store the imbalance.
3681 * Returns 1 if there is potential to perform power-savings balance.
3684 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3685 int this_cpu
, unsigned long *imbalance
)
3687 if (!sds
->power_savings_balance
)
3690 if (sds
->this != sds
->group_leader
||
3691 sds
->group_leader
== sds
->group_min
)
3694 *imbalance
= sds
->min_load_per_task
;
3695 sds
->busiest
= sds
->group_min
;
3697 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3698 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3699 group_first_cpu(sds
->group_leader
);
3705 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3706 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3707 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3712 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3713 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3718 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3719 int this_cpu
, unsigned long *imbalance
)
3723 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3727 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3728 * @group: sched_group whose statistics are to be updated.
3729 * @this_cpu: Cpu for which load balance is currently performed.
3730 * @idle: Idle status of this_cpu
3731 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3732 * @sd_idle: Idle status of the sched_domain containing group.
3733 * @local_group: Does group contain this_cpu.
3734 * @cpus: Set of cpus considered for load balancing.
3735 * @balance: Should we balance.
3736 * @sgs: variable to hold the statistics for this group.
3738 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3739 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3740 int local_group
, const struct cpumask
*cpus
,
3741 int *balance
, struct sg_lb_stats
*sgs
)
3743 unsigned long load
, max_cpu_load
, min_cpu_load
;
3745 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3746 unsigned long sum_avg_load_per_task
;
3747 unsigned long avg_load_per_task
;
3750 balance_cpu
= group_first_cpu(group
);
3752 /* Tally up the load of all CPUs in the group */
3753 sum_avg_load_per_task
= avg_load_per_task
= 0;
3755 min_cpu_load
= ~0UL;
3757 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3758 struct rq
*rq
= cpu_rq(i
);
3760 if (*sd_idle
&& rq
->nr_running
)
3763 /* Bias balancing toward cpus of our domain */
3765 if (idle_cpu(i
) && !first_idle_cpu
) {
3770 load
= target_load(i
, load_idx
);
3772 load
= source_load(i
, load_idx
);
3773 if (load
> max_cpu_load
)
3774 max_cpu_load
= load
;
3775 if (min_cpu_load
> load
)
3776 min_cpu_load
= load
;
3779 sgs
->group_load
+= load
;
3780 sgs
->sum_nr_running
+= rq
->nr_running
;
3781 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3783 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3787 * First idle cpu or the first cpu(busiest) in this sched group
3788 * is eligible for doing load balancing at this and above
3789 * domains. In the newly idle case, we will allow all the cpu's
3790 * to do the newly idle load balance.
3792 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3793 balance_cpu
!= this_cpu
&& balance
) {
3798 /* Adjust by relative CPU power of the group */
3799 sgs
->avg_load
= sg_div_cpu_power(group
,
3800 sgs
->group_load
* SCHED_LOAD_SCALE
);
3804 * Consider the group unbalanced when the imbalance is larger
3805 * than the average weight of two tasks.
3807 * APZ: with cgroup the avg task weight can vary wildly and
3808 * might not be a suitable number - should we keep a
3809 * normalized nr_running number somewhere that negates
3812 avg_load_per_task
= sg_div_cpu_power(group
,
3813 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3815 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3818 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3823 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3824 * @sd: sched_domain whose statistics are to be updated.
3825 * @this_cpu: Cpu for which load balance is currently performed.
3826 * @idle: Idle status of this_cpu
3827 * @sd_idle: Idle status of the sched_domain containing group.
3828 * @cpus: Set of cpus considered for load balancing.
3829 * @balance: Should we balance.
3830 * @sds: variable to hold the statistics for this sched_domain.
3832 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3833 enum cpu_idle_type idle
, int *sd_idle
,
3834 const struct cpumask
*cpus
, int *balance
,
3835 struct sd_lb_stats
*sds
)
3837 struct sched_group
*group
= sd
->groups
;
3838 struct sg_lb_stats sgs
;
3841 init_sd_power_savings_stats(sd
, sds
, idle
);
3842 load_idx
= get_sd_load_idx(sd
, idle
);
3847 local_group
= cpumask_test_cpu(this_cpu
,
3848 sched_group_cpus(group
));
3849 memset(&sgs
, 0, sizeof(sgs
));
3850 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3851 local_group
, cpus
, balance
, &sgs
);
3853 if (local_group
&& balance
&& !(*balance
))
3856 sds
->total_load
+= sgs
.group_load
;
3857 sds
->total_pwr
+= group
->__cpu_power
;
3860 sds
->this_load
= sgs
.avg_load
;
3862 sds
->this_nr_running
= sgs
.sum_nr_running
;
3863 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3864 } else if (sgs
.avg_load
> sds
->max_load
&&
3865 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3867 sds
->max_load
= sgs
.avg_load
;
3868 sds
->busiest
= group
;
3869 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3870 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3871 sds
->group_imb
= sgs
.group_imb
;
3874 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3875 group
= group
->next
;
3876 } while (group
!= sd
->groups
);
3881 * fix_small_imbalance - Calculate the minor imbalance that exists
3882 * amongst the groups of a sched_domain, during
3884 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3885 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3886 * @imbalance: Variable to store the imbalance.
3888 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3889 int this_cpu
, unsigned long *imbalance
)
3891 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3892 unsigned int imbn
= 2;
3894 if (sds
->this_nr_running
) {
3895 sds
->this_load_per_task
/= sds
->this_nr_running
;
3896 if (sds
->busiest_load_per_task
>
3897 sds
->this_load_per_task
)
3900 sds
->this_load_per_task
=
3901 cpu_avg_load_per_task(this_cpu
);
3903 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3904 sds
->busiest_load_per_task
* imbn
) {
3905 *imbalance
= sds
->busiest_load_per_task
;
3910 * OK, we don't have enough imbalance to justify moving tasks,
3911 * however we may be able to increase total CPU power used by
3915 pwr_now
+= sds
->busiest
->__cpu_power
*
3916 min(sds
->busiest_load_per_task
, sds
->max_load
);
3917 pwr_now
+= sds
->this->__cpu_power
*
3918 min(sds
->this_load_per_task
, sds
->this_load
);
3919 pwr_now
/= SCHED_LOAD_SCALE
;
3921 /* Amount of load we'd subtract */
3922 tmp
= sg_div_cpu_power(sds
->busiest
,
3923 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3924 if (sds
->max_load
> tmp
)
3925 pwr_move
+= sds
->busiest
->__cpu_power
*
3926 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3928 /* Amount of load we'd add */
3929 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3930 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3931 tmp
= sg_div_cpu_power(sds
->this,
3932 sds
->max_load
* sds
->busiest
->__cpu_power
);
3934 tmp
= sg_div_cpu_power(sds
->this,
3935 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3936 pwr_move
+= sds
->this->__cpu_power
*
3937 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3938 pwr_move
/= SCHED_LOAD_SCALE
;
3940 /* Move if we gain throughput */
3941 if (pwr_move
> pwr_now
)
3942 *imbalance
= sds
->busiest_load_per_task
;
3946 * calculate_imbalance - Calculate the amount of imbalance present within the
3947 * groups of a given sched_domain during load balance.
3948 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3949 * @this_cpu: Cpu for which currently load balance is being performed.
3950 * @imbalance: The variable to store the imbalance.
3952 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3953 unsigned long *imbalance
)
3955 unsigned long max_pull
;
3957 * In the presence of smp nice balancing, certain scenarios can have
3958 * max load less than avg load(as we skip the groups at or below
3959 * its cpu_power, while calculating max_load..)
3961 if (sds
->max_load
< sds
->avg_load
) {
3963 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3966 /* Don't want to pull so many tasks that a group would go idle */
3967 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3968 sds
->max_load
- sds
->busiest_load_per_task
);
3970 /* How much load to actually move to equalise the imbalance */
3971 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3972 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3976 * if *imbalance is less than the average load per runnable task
3977 * there is no gaurantee that any tasks will be moved so we'll have
3978 * a think about bumping its value to force at least one task to be
3981 if (*imbalance
< sds
->busiest_load_per_task
)
3982 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3985 /******* find_busiest_group() helpers end here *********************/
3988 * find_busiest_group - Returns the busiest group within the sched_domain
3989 * if there is an imbalance. If there isn't an imbalance, and
3990 * the user has opted for power-savings, it returns a group whose
3991 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3992 * such a group exists.
3994 * Also calculates the amount of weighted load which should be moved
3995 * to restore balance.
3997 * @sd: The sched_domain whose busiest group is to be returned.
3998 * @this_cpu: The cpu for which load balancing is currently being performed.
3999 * @imbalance: Variable which stores amount of weighted load which should
4000 * be moved to restore balance/put a group to idle.
4001 * @idle: The idle status of this_cpu.
4002 * @sd_idle: The idleness of sd
4003 * @cpus: The set of CPUs under consideration for load-balancing.
4004 * @balance: Pointer to a variable indicating if this_cpu
4005 * is the appropriate cpu to perform load balancing at this_level.
4007 * Returns: - the busiest group if imbalance exists.
4008 * - If no imbalance and user has opted for power-savings balance,
4009 * return the least loaded group whose CPUs can be
4010 * put to idle by rebalancing its tasks onto our group.
4012 static struct sched_group
*
4013 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4014 unsigned long *imbalance
, enum cpu_idle_type idle
,
4015 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4017 struct sd_lb_stats sds
;
4019 memset(&sds
, 0, sizeof(sds
));
4022 * Compute the various statistics relavent for load balancing at
4025 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4028 /* Cases where imbalance does not exist from POV of this_cpu */
4029 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4031 * 2) There is no busy sibling group to pull from.
4032 * 3) This group is the busiest group.
4033 * 4) This group is more busy than the avg busieness at this
4035 * 5) The imbalance is within the specified limit.
4036 * 6) Any rebalance would lead to ping-pong
4038 if (balance
&& !(*balance
))
4041 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4044 if (sds
.this_load
>= sds
.max_load
)
4047 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4049 if (sds
.this_load
>= sds
.avg_load
)
4052 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4055 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4057 sds
.busiest_load_per_task
=
4058 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4061 * We're trying to get all the cpus to the average_load, so we don't
4062 * want to push ourselves above the average load, nor do we wish to
4063 * reduce the max loaded cpu below the average load, as either of these
4064 * actions would just result in more rebalancing later, and ping-pong
4065 * tasks around. Thus we look for the minimum possible imbalance.
4066 * Negative imbalances (*we* are more loaded than anyone else) will
4067 * be counted as no imbalance for these purposes -- we can't fix that
4068 * by pulling tasks to us. Be careful of negative numbers as they'll
4069 * appear as very large values with unsigned longs.
4071 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4074 /* Looks like there is an imbalance. Compute it */
4075 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4080 * There is no obvious imbalance. But check if we can do some balancing
4083 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4091 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4094 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4095 unsigned long imbalance
, const struct cpumask
*cpus
)
4097 struct rq
*busiest
= NULL
, *rq
;
4098 unsigned long max_load
= 0;
4101 for_each_cpu(i
, sched_group_cpus(group
)) {
4104 if (!cpumask_test_cpu(i
, cpus
))
4108 wl
= weighted_cpuload(i
);
4110 if (rq
->nr_running
== 1 && wl
> imbalance
)
4113 if (wl
> max_load
) {
4123 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4124 * so long as it is large enough.
4126 #define MAX_PINNED_INTERVAL 512
4128 /* Working cpumask for load_balance and load_balance_newidle. */
4129 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4132 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4133 * tasks if there is an imbalance.
4135 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4136 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4139 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4140 struct sched_group
*group
;
4141 unsigned long imbalance
;
4143 unsigned long flags
;
4144 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4146 cpumask_setall(cpus
);
4149 * When power savings policy is enabled for the parent domain, idle
4150 * sibling can pick up load irrespective of busy siblings. In this case,
4151 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4152 * portraying it as CPU_NOT_IDLE.
4154 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4155 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4158 schedstat_inc(sd
, lb_count
[idle
]);
4162 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4169 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4173 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4175 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4179 BUG_ON(busiest
== this_rq
);
4181 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4184 if (busiest
->nr_running
> 1) {
4186 * Attempt to move tasks. If find_busiest_group has found
4187 * an imbalance but busiest->nr_running <= 1, the group is
4188 * still unbalanced. ld_moved simply stays zero, so it is
4189 * correctly treated as an imbalance.
4191 local_irq_save(flags
);
4192 double_rq_lock(this_rq
, busiest
);
4193 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4194 imbalance
, sd
, idle
, &all_pinned
);
4195 double_rq_unlock(this_rq
, busiest
);
4196 local_irq_restore(flags
);
4199 * some other cpu did the load balance for us.
4201 if (ld_moved
&& this_cpu
!= smp_processor_id())
4202 resched_cpu(this_cpu
);
4204 /* All tasks on this runqueue were pinned by CPU affinity */
4205 if (unlikely(all_pinned
)) {
4206 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4207 if (!cpumask_empty(cpus
))
4214 schedstat_inc(sd
, lb_failed
[idle
]);
4215 sd
->nr_balance_failed
++;
4217 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4219 spin_lock_irqsave(&busiest
->lock
, flags
);
4221 /* don't kick the migration_thread, if the curr
4222 * task on busiest cpu can't be moved to this_cpu
4224 if (!cpumask_test_cpu(this_cpu
,
4225 &busiest
->curr
->cpus_allowed
)) {
4226 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4228 goto out_one_pinned
;
4231 if (!busiest
->active_balance
) {
4232 busiest
->active_balance
= 1;
4233 busiest
->push_cpu
= this_cpu
;
4236 spin_unlock_irqrestore(&busiest
->lock
, flags
);
4238 wake_up_process(busiest
->migration_thread
);
4241 * We've kicked active balancing, reset the failure
4244 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4247 sd
->nr_balance_failed
= 0;
4249 if (likely(!active_balance
)) {
4250 /* We were unbalanced, so reset the balancing interval */
4251 sd
->balance_interval
= sd
->min_interval
;
4254 * If we've begun active balancing, start to back off. This
4255 * case may not be covered by the all_pinned logic if there
4256 * is only 1 task on the busy runqueue (because we don't call
4259 if (sd
->balance_interval
< sd
->max_interval
)
4260 sd
->balance_interval
*= 2;
4263 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4264 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4270 schedstat_inc(sd
, lb_balanced
[idle
]);
4272 sd
->nr_balance_failed
= 0;
4275 /* tune up the balancing interval */
4276 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4277 (sd
->balance_interval
< sd
->max_interval
))
4278 sd
->balance_interval
*= 2;
4280 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4281 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4292 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4293 * tasks if there is an imbalance.
4295 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4296 * this_rq is locked.
4299 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4301 struct sched_group
*group
;
4302 struct rq
*busiest
= NULL
;
4303 unsigned long imbalance
;
4307 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4309 cpumask_setall(cpus
);
4312 * When power savings policy is enabled for the parent domain, idle
4313 * sibling can pick up load irrespective of busy siblings. In this case,
4314 * let the state of idle sibling percolate up as IDLE, instead of
4315 * portraying it as CPU_NOT_IDLE.
4317 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4318 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4321 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4323 update_shares_locked(this_rq
, sd
);
4324 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4325 &sd_idle
, cpus
, NULL
);
4327 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4331 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4333 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4337 BUG_ON(busiest
== this_rq
);
4339 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4342 if (busiest
->nr_running
> 1) {
4343 /* Attempt to move tasks */
4344 double_lock_balance(this_rq
, busiest
);
4345 /* this_rq->clock is already updated */
4346 update_rq_clock(busiest
);
4347 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4348 imbalance
, sd
, CPU_NEWLY_IDLE
,
4350 double_unlock_balance(this_rq
, busiest
);
4352 if (unlikely(all_pinned
)) {
4353 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4354 if (!cpumask_empty(cpus
))
4360 int active_balance
= 0;
4362 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4363 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4364 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4367 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4370 if (sd
->nr_balance_failed
++ < 2)
4374 * The only task running in a non-idle cpu can be moved to this
4375 * cpu in an attempt to completely freeup the other CPU
4376 * package. The same method used to move task in load_balance()
4377 * have been extended for load_balance_newidle() to speedup
4378 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4380 * The package power saving logic comes from
4381 * find_busiest_group(). If there are no imbalance, then
4382 * f_b_g() will return NULL. However when sched_mc={1,2} then
4383 * f_b_g() will select a group from which a running task may be
4384 * pulled to this cpu in order to make the other package idle.
4385 * If there is no opportunity to make a package idle and if
4386 * there are no imbalance, then f_b_g() will return NULL and no
4387 * action will be taken in load_balance_newidle().
4389 * Under normal task pull operation due to imbalance, there
4390 * will be more than one task in the source run queue and
4391 * move_tasks() will succeed. ld_moved will be true and this
4392 * active balance code will not be triggered.
4395 /* Lock busiest in correct order while this_rq is held */
4396 double_lock_balance(this_rq
, busiest
);
4399 * don't kick the migration_thread, if the curr
4400 * task on busiest cpu can't be moved to this_cpu
4402 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4403 double_unlock_balance(this_rq
, busiest
);
4408 if (!busiest
->active_balance
) {
4409 busiest
->active_balance
= 1;
4410 busiest
->push_cpu
= this_cpu
;
4414 double_unlock_balance(this_rq
, busiest
);
4416 * Should not call ttwu while holding a rq->lock
4418 spin_unlock(&this_rq
->lock
);
4420 wake_up_process(busiest
->migration_thread
);
4421 spin_lock(&this_rq
->lock
);
4424 sd
->nr_balance_failed
= 0;
4426 update_shares_locked(this_rq
, sd
);
4430 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4431 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4432 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4434 sd
->nr_balance_failed
= 0;
4440 * idle_balance is called by schedule() if this_cpu is about to become
4441 * idle. Attempts to pull tasks from other CPUs.
4443 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4445 struct sched_domain
*sd
;
4446 int pulled_task
= 0;
4447 unsigned long next_balance
= jiffies
+ HZ
;
4449 for_each_domain(this_cpu
, sd
) {
4450 unsigned long interval
;
4452 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4455 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4456 /* If we've pulled tasks over stop searching: */
4457 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4460 interval
= msecs_to_jiffies(sd
->balance_interval
);
4461 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4462 next_balance
= sd
->last_balance
+ interval
;
4466 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4468 * We are going idle. next_balance may be set based on
4469 * a busy processor. So reset next_balance.
4471 this_rq
->next_balance
= next_balance
;
4476 * active_load_balance is run by migration threads. It pushes running tasks
4477 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4478 * running on each physical CPU where possible, and avoids physical /
4479 * logical imbalances.
4481 * Called with busiest_rq locked.
4483 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4485 int target_cpu
= busiest_rq
->push_cpu
;
4486 struct sched_domain
*sd
;
4487 struct rq
*target_rq
;
4489 /* Is there any task to move? */
4490 if (busiest_rq
->nr_running
<= 1)
4493 target_rq
= cpu_rq(target_cpu
);
4496 * This condition is "impossible", if it occurs
4497 * we need to fix it. Originally reported by
4498 * Bjorn Helgaas on a 128-cpu setup.
4500 BUG_ON(busiest_rq
== target_rq
);
4502 /* move a task from busiest_rq to target_rq */
4503 double_lock_balance(busiest_rq
, target_rq
);
4504 update_rq_clock(busiest_rq
);
4505 update_rq_clock(target_rq
);
4507 /* Search for an sd spanning us and the target CPU. */
4508 for_each_domain(target_cpu
, sd
) {
4509 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4510 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4515 schedstat_inc(sd
, alb_count
);
4517 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4519 schedstat_inc(sd
, alb_pushed
);
4521 schedstat_inc(sd
, alb_failed
);
4523 double_unlock_balance(busiest_rq
, target_rq
);
4528 atomic_t load_balancer
;
4529 cpumask_var_t cpu_mask
;
4530 } nohz ____cacheline_aligned
= {
4531 .load_balancer
= ATOMIC_INIT(-1),
4535 * This routine will try to nominate the ilb (idle load balancing)
4536 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4537 * load balancing on behalf of all those cpus. If all the cpus in the system
4538 * go into this tickless mode, then there will be no ilb owner (as there is
4539 * no need for one) and all the cpus will sleep till the next wakeup event
4542 * For the ilb owner, tick is not stopped. And this tick will be used
4543 * for idle load balancing. ilb owner will still be part of
4546 * While stopping the tick, this cpu will become the ilb owner if there
4547 * is no other owner. And will be the owner till that cpu becomes busy
4548 * or if all cpus in the system stop their ticks at which point
4549 * there is no need for ilb owner.
4551 * When the ilb owner becomes busy, it nominates another owner, during the
4552 * next busy scheduler_tick()
4554 int select_nohz_load_balancer(int stop_tick
)
4556 int cpu
= smp_processor_id();
4559 cpu_rq(cpu
)->in_nohz_recently
= 1;
4561 if (!cpu_active(cpu
)) {
4562 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4566 * If we are going offline and still the leader,
4569 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4575 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4577 /* time for ilb owner also to sleep */
4578 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4579 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4580 atomic_set(&nohz
.load_balancer
, -1);
4584 if (atomic_read(&nohz
.load_balancer
) == -1) {
4585 /* make me the ilb owner */
4586 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4588 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4591 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4594 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4596 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4597 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4604 static DEFINE_SPINLOCK(balancing
);
4607 * It checks each scheduling domain to see if it is due to be balanced,
4608 * and initiates a balancing operation if so.
4610 * Balancing parameters are set up in arch_init_sched_domains.
4612 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4615 struct rq
*rq
= cpu_rq(cpu
);
4616 unsigned long interval
;
4617 struct sched_domain
*sd
;
4618 /* Earliest time when we have to do rebalance again */
4619 unsigned long next_balance
= jiffies
+ 60*HZ
;
4620 int update_next_balance
= 0;
4623 for_each_domain(cpu
, sd
) {
4624 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4627 interval
= sd
->balance_interval
;
4628 if (idle
!= CPU_IDLE
)
4629 interval
*= sd
->busy_factor
;
4631 /* scale ms to jiffies */
4632 interval
= msecs_to_jiffies(interval
);
4633 if (unlikely(!interval
))
4635 if (interval
> HZ
*NR_CPUS
/10)
4636 interval
= HZ
*NR_CPUS
/10;
4638 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4640 if (need_serialize
) {
4641 if (!spin_trylock(&balancing
))
4645 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4646 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4648 * We've pulled tasks over so either we're no
4649 * longer idle, or one of our SMT siblings is
4652 idle
= CPU_NOT_IDLE
;
4654 sd
->last_balance
= jiffies
;
4657 spin_unlock(&balancing
);
4659 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4660 next_balance
= sd
->last_balance
+ interval
;
4661 update_next_balance
= 1;
4665 * Stop the load balance at this level. There is another
4666 * CPU in our sched group which is doing load balancing more
4674 * next_balance will be updated only when there is a need.
4675 * When the cpu is attached to null domain for ex, it will not be
4678 if (likely(update_next_balance
))
4679 rq
->next_balance
= next_balance
;
4683 * run_rebalance_domains is triggered when needed from the scheduler tick.
4684 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4685 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4687 static void run_rebalance_domains(struct softirq_action
*h
)
4689 int this_cpu
= raw_smp_processor_id();
4690 struct rq
*this_rq
= cpu_rq(this_cpu
);
4691 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4692 CPU_IDLE
: CPU_NOT_IDLE
;
4694 rebalance_domains(this_cpu
, idle
);
4698 * If this cpu is the owner for idle load balancing, then do the
4699 * balancing on behalf of the other idle cpus whose ticks are
4702 if (this_rq
->idle_at_tick
&&
4703 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4707 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4708 if (balance_cpu
== this_cpu
)
4712 * If this cpu gets work to do, stop the load balancing
4713 * work being done for other cpus. Next load
4714 * balancing owner will pick it up.
4719 rebalance_domains(balance_cpu
, CPU_IDLE
);
4721 rq
= cpu_rq(balance_cpu
);
4722 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4723 this_rq
->next_balance
= rq
->next_balance
;
4729 static inline int on_null_domain(int cpu
)
4731 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4735 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4737 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4738 * idle load balancing owner or decide to stop the periodic load balancing,
4739 * if the whole system is idle.
4741 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4745 * If we were in the nohz mode recently and busy at the current
4746 * scheduler tick, then check if we need to nominate new idle
4749 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4750 rq
->in_nohz_recently
= 0;
4752 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4753 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4754 atomic_set(&nohz
.load_balancer
, -1);
4757 if (atomic_read(&nohz
.load_balancer
) == -1) {
4759 * simple selection for now: Nominate the
4760 * first cpu in the nohz list to be the next
4763 * TBD: Traverse the sched domains and nominate
4764 * the nearest cpu in the nohz.cpu_mask.
4766 int ilb
= cpumask_first(nohz
.cpu_mask
);
4768 if (ilb
< nr_cpu_ids
)
4774 * If this cpu is idle and doing idle load balancing for all the
4775 * cpus with ticks stopped, is it time for that to stop?
4777 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4778 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4784 * If this cpu is idle and the idle load balancing is done by
4785 * someone else, then no need raise the SCHED_SOFTIRQ
4787 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4788 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4791 /* Don't need to rebalance while attached to NULL domain */
4792 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4793 likely(!on_null_domain(cpu
)))
4794 raise_softirq(SCHED_SOFTIRQ
);
4797 #else /* CONFIG_SMP */
4800 * on UP we do not need to balance between CPUs:
4802 static inline void idle_balance(int cpu
, struct rq
*rq
)
4808 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4810 EXPORT_PER_CPU_SYMBOL(kstat
);
4813 * Return any ns on the sched_clock that have not yet been accounted in
4814 * @p in case that task is currently running.
4816 * Called with task_rq_lock() held on @rq.
4818 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4822 if (task_current(rq
, p
)) {
4823 update_rq_clock(rq
);
4824 ns
= rq
->clock
- p
->se
.exec_start
;
4832 unsigned long long __task_delta_exec(struct task_struct
*p
, int update
)
4838 WARN_ON_ONCE(!runqueue_is_locked());
4839 WARN_ON_ONCE(!task_current(rq
, p
));
4842 update_rq_clock(rq
);
4844 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4846 WARN_ON_ONCE(delta_exec
< 0);
4852 * Return any ns on the sched_clock that have not yet been banked in
4853 * @p in case that task is currently running.
4855 unsigned long long task_delta_exec(struct task_struct
*p
)
4857 unsigned long flags
;
4861 rq
= task_rq_lock(p
, &flags
);
4862 ns
= do_task_delta_exec(p
, rq
);
4863 task_rq_unlock(rq
, &flags
);
4869 * Return accounted runtime for the task.
4870 * In case the task is currently running, return the runtime plus current's
4871 * pending runtime that have not been accounted yet.
4873 unsigned long long task_sched_runtime(struct task_struct
*p
)
4875 unsigned long flags
;
4879 rq
= task_rq_lock(p
, &flags
);
4880 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4881 task_rq_unlock(rq
, &flags
);
4887 * Return sum_exec_runtime for the thread group.
4888 * In case the task is currently running, return the sum plus current's
4889 * pending runtime that have not been accounted yet.
4891 * Note that the thread group might have other running tasks as well,
4892 * so the return value not includes other pending runtime that other
4893 * running tasks might have.
4895 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4897 struct task_cputime totals
;
4898 unsigned long flags
;
4902 rq
= task_rq_lock(p
, &flags
);
4903 thread_group_cputime(p
, &totals
);
4904 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4905 task_rq_unlock(rq
, &flags
);
4911 * Account user cpu time to a process.
4912 * @p: the process that the cpu time gets accounted to
4913 * @cputime: the cpu time spent in user space since the last update
4914 * @cputime_scaled: cputime scaled by cpu frequency
4916 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4917 cputime_t cputime_scaled
)
4919 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4922 /* Add user time to process. */
4923 p
->utime
= cputime_add(p
->utime
, cputime
);
4924 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4925 account_group_user_time(p
, cputime
);
4927 /* Add user time to cpustat. */
4928 tmp
= cputime_to_cputime64(cputime
);
4930 cpustat
->user_rt
= cputime64_add(cpustat
->user_rt
, tmp
);
4931 else if (TASK_NICE(p
) > 0)
4932 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4934 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4935 /* Account for user time used */
4936 acct_update_integrals(p
);
4940 * Account guest cpu time to a process.
4941 * @p: the process that the cpu time gets accounted to
4942 * @cputime: the cpu time spent in virtual machine since the last update
4943 * @cputime_scaled: cputime scaled by cpu frequency
4945 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4946 cputime_t cputime_scaled
)
4949 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4951 tmp
= cputime_to_cputime64(cputime
);
4953 /* Add guest time to process. */
4954 p
->utime
= cputime_add(p
->utime
, cputime
);
4955 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4956 account_group_user_time(p
, cputime
);
4957 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4959 /* Add guest time to cpustat. */
4960 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4961 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4965 * Account system cpu time to a process.
4966 * @p: the process that the cpu time gets accounted to
4967 * @hardirq_offset: the offset to subtract from hardirq_count()
4968 * @cputime: the cpu time spent in kernel space since the last update
4969 * @cputime_scaled: cputime scaled by cpu frequency
4971 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4972 cputime_t cputime
, cputime_t cputime_scaled
)
4974 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4977 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4978 account_guest_time(p
, cputime
, cputime_scaled
);
4982 /* Add system time to process. */
4983 p
->stime
= cputime_add(p
->stime
, cputime
);
4984 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4985 account_group_system_time(p
, cputime
);
4987 /* Add system time to cpustat. */
4988 tmp
= cputime_to_cputime64(cputime
);
4989 if (hardirq_count() - hardirq_offset
|| (p
->flags
& PF_HARDIRQ
))
4990 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4991 else if (softirq_count() || (p
->flags
& PF_SOFTIRQ
))
4992 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4993 else if (rt_task(p
))
4994 cpustat
->system_rt
= cputime64_add(cpustat
->system_rt
, tmp
);
4996 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4998 /* Account for system time used */
4999 acct_update_integrals(p
);
5003 * Account for involuntary wait time.
5004 * @steal: the cpu time spent in involuntary wait
5006 void account_steal_time(cputime_t cputime
)
5008 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5009 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5011 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5015 * Account for idle time.
5016 * @cputime: the cpu time spent in idle wait
5018 void account_idle_time(cputime_t cputime
)
5020 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5021 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5022 struct rq
*rq
= this_rq();
5024 if (atomic_read(&rq
->nr_iowait
) > 0)
5025 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5027 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5030 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5033 * Account a single tick of cpu time.
5034 * @p: the process that the cpu time gets accounted to
5035 * @user_tick: indicates if the tick is a user or a system tick
5037 void account_process_tick(struct task_struct
*p
, int user_tick
)
5039 cputime_t one_jiffy
= jiffies_to_cputime(1);
5040 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
5041 struct rq
*rq
= this_rq();
5044 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
5045 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5046 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
5049 account_idle_time(one_jiffy
);
5053 * Account multiple ticks of steal time.
5054 * @p: the process from which the cpu time has been stolen
5055 * @ticks: number of stolen ticks
5057 void account_steal_ticks(unsigned long ticks
)
5059 account_steal_time(jiffies_to_cputime(ticks
));
5063 * Account multiple ticks of idle time.
5064 * @ticks: number of stolen ticks
5066 void account_idle_ticks(unsigned long ticks
)
5068 account_idle_time(jiffies_to_cputime(ticks
));
5074 * Use precise platform statistics if available:
5076 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5077 cputime_t
task_utime(struct task_struct
*p
)
5082 cputime_t
task_stime(struct task_struct
*p
)
5087 cputime_t
task_utime(struct task_struct
*p
)
5089 clock_t utime
= cputime_to_clock_t(p
->utime
),
5090 total
= utime
+ cputime_to_clock_t(p
->stime
);
5094 * Use CFS's precise accounting:
5096 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
5100 do_div(temp
, total
);
5102 utime
= (clock_t)temp
;
5104 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
5105 return p
->prev_utime
;
5108 cputime_t
task_stime(struct task_struct
*p
)
5113 * Use CFS's precise accounting. (we subtract utime from
5114 * the total, to make sure the total observed by userspace
5115 * grows monotonically - apps rely on that):
5117 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
5118 cputime_to_clock_t(task_utime(p
));
5121 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
5123 return p
->prev_stime
;
5127 inline cputime_t
task_gtime(struct task_struct
*p
)
5133 * This function gets called by the timer code, with HZ frequency.
5134 * We call it with interrupts disabled.
5136 * It also gets called by the fork code, when changing the parent's
5139 void scheduler_tick(void)
5141 int cpu
= smp_processor_id();
5142 struct rq
*rq
= cpu_rq(cpu
);
5143 struct task_struct
*curr
= rq
->curr
;
5147 BUG_ON(!irqs_disabled());
5149 spin_lock(&rq
->lock
);
5150 update_rq_clock(rq
);
5151 update_cpu_load(rq
);
5152 if (curr
!= rq
->idle
&& curr
->se
.on_rq
)
5153 curr
->sched_class
->task_tick(rq
, curr
, 0);
5154 perf_counter_task_tick(curr
, cpu
);
5155 spin_unlock(&rq
->lock
);
5158 rq
->idle_at_tick
= idle_cpu(cpu
);
5159 trigger_load_balance(rq
, cpu
);
5163 unsigned long notrace
get_parent_ip(unsigned long addr
)
5165 if (in_lock_functions(addr
)) {
5166 addr
= CALLER_ADDR2
;
5167 if (in_lock_functions(addr
))
5168 addr
= CALLER_ADDR3
;
5173 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5174 defined(CONFIG_PREEMPT_TRACER))
5176 void __kprobes
add_preempt_count(int val
)
5178 #ifdef CONFIG_DEBUG_PREEMPT
5182 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5185 preempt_count() += val
;
5186 #ifdef CONFIG_DEBUG_PREEMPT
5188 * Spinlock count overflowing soon?
5190 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5193 if (preempt_count() == val
)
5194 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5196 EXPORT_SYMBOL(add_preempt_count
);
5198 void __kprobes
sub_preempt_count(int val
)
5200 #ifdef CONFIG_DEBUG_PREEMPT
5204 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5207 * Is the spinlock portion underflowing?
5209 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5210 !(preempt_count() & PREEMPT_MASK
)))
5214 if (preempt_count() == val
)
5215 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5216 preempt_count() -= val
;
5218 EXPORT_SYMBOL(sub_preempt_count
);
5223 * Print scheduling while atomic bug:
5225 static noinline
void __schedule_bug(struct task_struct
*prev
)
5227 struct pt_regs
*regs
= get_irq_regs();
5229 printk(KERN_ERR
"BUG: scheduling while atomic: %s/0x%08x/%d, CPU#%d\n",
5230 prev
->comm
, preempt_count(), prev
->pid
, smp_processor_id());
5232 debug_show_held_locks(prev
);
5234 if (irqs_disabled())
5235 print_irqtrace_events(prev
);
5244 * Various schedule()-time debugging checks and statistics:
5246 static inline void schedule_debug(struct task_struct
*prev
)
5248 // WARN_ON(system_state == SYSTEM_BOOTING);
5251 * Test if we are atomic. Since do_exit() needs to call into
5252 * schedule() atomically, we ignore that path for now.
5253 * Otherwise, whine if we are scheduling when we should not be.
5255 if (unlikely(in_atomic() && !prev
->exit_state
))
5256 __schedule_bug(prev
);
5258 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5260 schedstat_inc(this_rq(), sched_count
);
5261 #ifdef CONFIG_SCHEDSTATS
5262 if (unlikely(prev
->lock_depth
>= 0)) {
5263 schedstat_inc(this_rq(), bkl_count
);
5264 schedstat_inc(prev
, sched_info
.bkl_count
);
5269 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5271 if (prev
->state
== TASK_RUNNING
) {
5272 u64 runtime
= prev
->se
.sum_exec_runtime
;
5274 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5275 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5278 * In order to avoid avg_overlap growing stale when we are
5279 * indeed overlapping and hence not getting put to sleep, grow
5280 * the avg_overlap on preemption.
5282 * We use the average preemption runtime because that
5283 * correlates to the amount of cache footprint a task can
5286 update_avg(&prev
->se
.avg_overlap
, runtime
);
5288 prev
->sched_class
->put_prev_task(rq
, prev
);
5292 * Pick up the highest-prio task:
5294 static inline struct task_struct
*
5295 pick_next_task(struct rq
*rq
)
5297 const struct sched_class
*class;
5298 struct task_struct
*p
;
5301 * Optimization: we know that if all tasks are in
5302 * the fair class we can call that function directly:
5304 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5305 p
= fair_sched_class
.pick_next_task(rq
);
5310 class = sched_class_highest
;
5312 p
= class->pick_next_task(rq
);
5316 * Will never be NULL as the idle class always
5317 * returns a non-NULL p:
5319 class = class->next
;
5324 * schedule() is the main scheduler function.
5326 asmlinkage
void __sched
__schedule(void)
5328 struct task_struct
*prev
, *next
;
5329 unsigned long *switch_count
;
5333 cpu
= smp_processor_id();
5337 switch_count
= &prev
->nivcsw
;
5339 release_kernel_lock(prev
);
5341 schedule_debug(prev
);
5345 if (sched_feat(HRTICK
))
5348 spin_lock_irq(&rq
->lock
);
5349 update_rq_clock(rq
);
5350 clear_tsk_need_resched(prev
);
5352 if (!(prev
->state
& TASK_RUNNING_MUTEX
) && prev
->state
&&
5353 !(preempt_count() & PREEMPT_ACTIVE
)) {
5354 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5355 prev
->state
= TASK_RUNNING
;
5357 touch_softlockup_watchdog();
5358 deactivate_task(rq
, prev
, 1);
5360 switch_count
= &prev
->nvcsw
;
5363 if (preempt_count() & PREEMPT_ACTIVE
)
5364 sub_preempt_count(PREEMPT_ACTIVE
);
5367 if (prev
->sched_class
->pre_schedule
)
5368 prev
->sched_class
->pre_schedule(rq
, prev
);
5371 if (unlikely(!rq
->nr_running
))
5372 idle_balance(cpu
, rq
);
5374 put_prev_task(rq
, prev
);
5375 next
= pick_next_task(rq
);
5377 if (likely(prev
!= next
)) {
5378 sched_info_switch(prev
, next
);
5379 perf_counter_task_sched_out(prev
, cpu
);
5385 context_switch(rq
, prev
, next
); /* unlocks the rq */
5387 * the context switch might have flipped the stack from under
5388 * us, hence refresh the local variables.
5390 cpu
= smp_processor_id();
5392 __preempt_enable_no_resched();
5394 __preempt_enable_no_resched();
5395 spin_unlock(&rq
->lock
);
5398 reacquire_kernel_lock(current
);
5401 asmlinkage
void __sched
schedule(void)
5404 local_irq_disable();
5408 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5411 EXPORT_SYMBOL(schedule
);
5413 #if defined(CONFIG_SMP) && !defined(CONFIG_PREEMPT_RT)
5415 * Look out! "owner" is an entirely speculative pointer
5416 * access and not reliable.
5418 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5423 if (!sched_feat(OWNER_SPIN
))
5426 #ifdef CONFIG_DEBUG_PAGEALLOC
5428 * Need to access the cpu field knowing that
5429 * DEBUG_PAGEALLOC could have unmapped it if
5430 * the mutex owner just released it and exited.
5432 if (probe_kernel_address(&owner
->cpu
, cpu
))
5439 * Even if the access succeeded (likely case),
5440 * the cpu field may no longer be valid.
5442 if (cpu
>= nr_cpumask_bits
)
5446 * We need to validate that we can do a
5447 * get_cpu() and that we have the percpu area.
5449 if (!cpu_online(cpu
))
5456 * Owner changed, break to re-assess state.
5458 if (lock
->owner
!= owner
)
5462 * Is that owner really running on that cpu?
5464 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5474 #ifdef CONFIG_PREEMPT
5477 * Global flag to turn preemption off on a CONFIG_PREEMPT kernel:
5479 int kernel_preemption
= 1;
5481 static int __init
preempt_setup (char *str
)
5483 if (!strncmp(str
, "off", 3)) {
5484 if (kernel_preemption
) {
5485 printk(KERN_INFO
"turning off kernel preemption!\n");
5486 kernel_preemption
= 0;
5490 if (!strncmp(str
, "on", 2)) {
5491 if (!kernel_preemption
) {
5492 printk(KERN_INFO
"turning on kernel preemption!\n");
5493 kernel_preemption
= 1;
5497 get_option(&str
, &kernel_preemption
);
5502 __setup("preempt=", preempt_setup
);
5505 * this is the entry point to schedule() from in-kernel preemption
5506 * off of preempt_enable. Kernel preemptions off return from interrupt
5507 * occur there and call schedule directly.
5509 asmlinkage
void __sched
preempt_schedule(void)
5511 struct thread_info
*ti
= current_thread_info();
5512 struct task_struct
*task
= current
;
5513 int saved_lock_depth
;
5515 if (!kernel_preemption
)
5518 * If there is a non-zero preempt_count or interrupts are disabled,
5519 * we do not want to preempt the current task. Just return..
5521 if (likely(ti
->preempt_count
|| irqs_disabled()))
5525 local_irq_disable();
5526 add_preempt_count(PREEMPT_ACTIVE
);
5529 * We keep the big kernel semaphore locked, but we
5530 * clear ->lock_depth so that schedule() doesnt
5531 * auto-release the semaphore:
5533 saved_lock_depth
= task
->lock_depth
;
5534 task
->lock_depth
= -1;
5536 task
->lock_depth
= saved_lock_depth
;
5540 * Check again in case we missed a preemption opportunity
5541 * between schedule and now.
5544 } while (need_resched());
5546 EXPORT_SYMBOL(preempt_schedule
);
5549 * this is is the entry point for the IRQ return path. Called with
5550 * interrupts disabled. To avoid infinite irq-entry recursion problems
5551 * with fast-paced IRQ sources we do all of this carefully to never
5552 * enable interrupts again.
5554 asmlinkage
void __sched
preempt_schedule_irq(void)
5556 struct thread_info
*ti
= current_thread_info();
5557 struct task_struct
*task
= current
;
5558 int saved_lock_depth
;
5560 if (!kernel_preemption
)
5563 * If there is a non-zero preempt_count then just return.
5564 * (interrupts are disabled)
5566 if (unlikely(ti
->preempt_count
))
5570 local_irq_disable();
5571 add_preempt_count(PREEMPT_ACTIVE
);
5574 * We keep the big kernel semaphore locked, but we
5575 * clear ->lock_depth so that schedule() doesnt
5576 * auto-release the semaphore:
5578 saved_lock_depth
= task
->lock_depth
;
5579 task
->lock_depth
= -1;
5581 local_irq_disable();
5582 task
->lock_depth
= saved_lock_depth
;
5585 * Check again in case we missed a preemption opportunity
5586 * between schedule and now.
5589 } while (need_resched());
5592 #endif /* CONFIG_PREEMPT */
5594 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5597 return try_to_wake_up(curr
->private, mode
, sync
, 0);
5599 EXPORT_SYMBOL(default_wake_function
);
5602 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5603 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5604 * number) then we wake all the non-exclusive tasks and one exclusive task.
5606 * There are circumstances in which we can try to wake a task which has already
5607 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5608 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5610 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5611 int nr_exclusive
, int sync
, void *key
)
5613 wait_queue_t
*curr
, *next
;
5615 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5616 unsigned flags
= curr
->flags
;
5618 if (curr
->func(curr
, mode
, sync
, key
) &&
5619 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5625 * __wake_up - wake up threads blocked on a waitqueue.
5627 * @mode: which threads
5628 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5629 * @key: is directly passed to the wakeup function
5631 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5632 int nr_exclusive
, void *key
)
5634 unsigned long flags
;
5636 spin_lock_irqsave(&q
->lock
, flags
);
5637 __wake_up_common(q
, mode
, nr_exclusive
, 1, key
);
5638 spin_unlock_irqrestore(&q
->lock
, flags
);
5640 EXPORT_SYMBOL(__wake_up
);
5643 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5645 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5647 __wake_up_common(q
, mode
, 1, 0, NULL
);
5651 * __wake_up_sync - wake up threads blocked on a waitqueue.
5653 * @mode: which threads
5654 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5656 * The sync wakeup differs that the waker knows that it will schedule
5657 * away soon, so while the target thread will be woken up, it will not
5658 * be migrated to another CPU - ie. the two threads are 'synchronized'
5659 * with each other. This can prevent needless bouncing between CPUs.
5661 * On UP it can prevent extra preemption.
5664 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5666 unsigned long flags
;
5672 if (unlikely(!nr_exclusive
))
5675 spin_lock_irqsave(&q
->lock
, flags
);
5676 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
5677 spin_unlock_irqrestore(&q
->lock
, flags
);
5679 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5682 * complete: - signals a single thread waiting on this completion
5683 * @x: holds the state of this particular completion
5685 * This will wake up a single thread waiting on this completion. Threads will be
5686 * awakened in the same order in which they were queued.
5688 * See also complete_all(), wait_for_completion() and related routines.
5690 void complete(struct completion
*x
)
5692 unsigned long flags
;
5694 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5696 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 1, NULL
);
5697 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5699 EXPORT_SYMBOL(complete
);
5702 * complete_all: - signals all threads waiting on this completion
5703 * @x: holds the state of this particular completion
5705 * This will wake up all threads waiting on this particular completion event.
5707 void complete_all(struct completion
*x
)
5709 unsigned long flags
;
5711 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5712 x
->done
+= UINT_MAX
/2;
5713 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 1, NULL
);
5714 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5716 EXPORT_SYMBOL(complete_all
);
5718 static inline long __sched
5719 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5722 DECLARE_WAITQUEUE(wait
, current
);
5724 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5725 __add_wait_queue_tail(&x
->wait
, &wait
);
5727 if (signal_pending_state(state
, current
)) {
5728 timeout
= -ERESTARTSYS
;
5731 __set_current_state(state
);
5732 spin_unlock_irq(&x
->wait
.lock
);
5733 timeout
= schedule_timeout(timeout
);
5734 spin_lock_irq(&x
->wait
.lock
);
5735 } while (!x
->done
&& timeout
);
5736 __remove_wait_queue(&x
->wait
, &wait
);
5741 return timeout
?: 1;
5745 wait_for_common(struct completion
*x
, long timeout
, int state
)
5749 spin_lock_irq(&x
->wait
.lock
);
5750 timeout
= do_wait_for_common(x
, timeout
, state
);
5751 spin_unlock_irq(&x
->wait
.lock
);
5756 * wait_for_completion: - waits for completion of a task
5757 * @x: holds the state of this particular completion
5759 * This waits to be signaled for completion of a specific task. It is NOT
5760 * interruptible and there is no timeout.
5762 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5763 * and interrupt capability. Also see complete().
5765 void __sched
wait_for_completion(struct completion
*x
)
5767 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5769 EXPORT_SYMBOL(wait_for_completion
);
5772 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5773 * @x: holds the state of this particular completion
5774 * @timeout: timeout value in jiffies
5776 * This waits for either a completion of a specific task to be signaled or for a
5777 * specified timeout to expire. The timeout is in jiffies. It is not
5780 unsigned long __sched
5781 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5783 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5785 EXPORT_SYMBOL(wait_for_completion_timeout
);
5788 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5789 * @x: holds the state of this particular completion
5791 * This waits for completion of a specific task to be signaled. It is
5794 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5796 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5797 if (t
== -ERESTARTSYS
)
5801 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5804 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5805 * @x: holds the state of this particular completion
5806 * @timeout: timeout value in jiffies
5808 * This waits for either a completion of a specific task to be signaled or for a
5809 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5811 unsigned long __sched
5812 wait_for_completion_interruptible_timeout(struct completion
*x
,
5813 unsigned long timeout
)
5815 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5817 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5820 * wait_for_completion_killable: - waits for completion of a task (killable)
5821 * @x: holds the state of this particular completion
5823 * This waits to be signaled for completion of a specific task. It can be
5824 * interrupted by a kill signal.
5826 int __sched
wait_for_completion_killable(struct completion
*x
)
5828 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5829 if (t
== -ERESTARTSYS
)
5833 EXPORT_SYMBOL(wait_for_completion_killable
);
5836 * try_wait_for_completion - try to decrement a completion without blocking
5837 * @x: completion structure
5839 * Returns: 0 if a decrement cannot be done without blocking
5840 * 1 if a decrement succeeded.
5842 * If a completion is being used as a counting completion,
5843 * attempt to decrement the counter without blocking. This
5844 * enables us to avoid waiting if the resource the completion
5845 * is protecting is not available.
5847 bool try_wait_for_completion(struct completion
*x
)
5851 spin_lock_irq(&x
->wait
.lock
);
5856 spin_unlock_irq(&x
->wait
.lock
);
5859 EXPORT_SYMBOL(try_wait_for_completion
);
5862 * completion_done - Test to see if a completion has any waiters
5863 * @x: completion structure
5865 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5866 * 1 if there are no waiters.
5869 bool completion_done(struct completion
*x
)
5873 spin_lock_irq(&x
->wait
.lock
);
5876 spin_unlock_irq(&x
->wait
.lock
);
5879 EXPORT_SYMBOL(completion_done
);
5882 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5884 unsigned long flags
;
5887 init_waitqueue_entry(&wait
, current
);
5889 __set_current_state(state
);
5891 spin_lock_irqsave(&q
->lock
, flags
);
5892 __add_wait_queue(q
, &wait
);
5893 spin_unlock(&q
->lock
);
5894 timeout
= schedule_timeout(timeout
);
5895 spin_lock_irq(&q
->lock
);
5896 __remove_wait_queue(q
, &wait
);
5897 spin_unlock_irqrestore(&q
->lock
, flags
);
5902 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5904 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5906 EXPORT_SYMBOL(interruptible_sleep_on
);
5909 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5911 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5913 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5915 void __sched
sleep_on(wait_queue_head_t
*q
)
5917 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5919 EXPORT_SYMBOL(sleep_on
);
5921 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5923 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5925 EXPORT_SYMBOL(sleep_on_timeout
);
5928 * task_setprio - set the current priority of a task
5930 * @prio: prio value (kernel-internal form)
5932 * This function changes the 'effective' priority of a task. It does
5933 * not touch ->normal_prio like __setscheduler().
5935 * Used by the rt_mutex code to implement priority inheritance logic
5936 * and by rcupreempt-boost to boost priorities of tasks sleeping
5939 void task_setprio(struct task_struct
*p
, int prio
)
5941 unsigned long flags
;
5942 int oldprio
, on_rq
, running
;
5944 const struct sched_class
*prev_class
= p
->sched_class
;
5946 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5948 rq
= task_rq_lock(p
, &flags
);
5951 * Idle task boosting is a nono in general. There is one
5952 * exception, when NOHZ is active:
5954 * The idle task calls get_next_timer_interrupt() and holds
5955 * the timer wheel base->lock on the CPU and another CPU wants
5956 * to access the timer (probably to cancel it). We can safely
5957 * ignore the boosting request, as the idle CPU runs this code
5958 * with interrupts disabled and will complete the lock
5959 * protected section without being interrupted. So there is no
5960 * real need to boost.
5962 if (unlikely(p
== rq
->idle
)) {
5963 WARN_ON(p
!= rq
->curr
);
5964 WARN_ON(p
->pi_blocked_on
);
5968 update_rq_clock(rq
);
5971 on_rq
= p
->se
.on_rq
;
5972 running
= task_current(rq
, p
);
5974 dequeue_task(rq
, p
, 0);
5976 p
->sched_class
->put_prev_task(rq
, p
);
5979 p
->sched_class
= &rt_sched_class
;
5981 p
->sched_class
= &fair_sched_class
;
5985 trace_sched_task_setprio(rq
, p
, oldprio
);
5988 p
->sched_class
->set_curr_task(rq
);
5990 enqueue_task(rq
, p
, 0);
5992 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5996 task_rq_unlock(rq
, &flags
);
5999 void set_user_nice(struct task_struct
*p
, long nice
)
6001 int old_prio
, delta
, on_rq
;
6002 unsigned long flags
;
6005 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6008 * We have to be careful, if called from sys_setpriority(),
6009 * the task might be in the middle of scheduling on another CPU.
6011 rq
= task_rq_lock(p
, &flags
);
6012 update_rq_clock(rq
);
6014 * The RT priorities are set via sched_setscheduler(), but we still
6015 * allow the 'normal' nice value to be set - but as expected
6016 * it wont have any effect on scheduling until the task is
6017 * SCHED_FIFO/SCHED_RR:
6019 if (task_has_rt_policy(p
)) {
6020 p
->static_prio
= NICE_TO_PRIO(nice
);
6023 on_rq
= p
->se
.on_rq
;
6025 dequeue_task(rq
, p
, 0);
6027 p
->static_prio
= NICE_TO_PRIO(nice
);
6030 p
->prio
= effective_prio(p
);
6031 delta
= p
->prio
- old_prio
;
6034 enqueue_task(rq
, p
, 0);
6036 * If the task increased its priority or is running and
6037 * lowered its priority, then reschedule its CPU:
6039 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6040 resched_task(rq
->curr
);
6043 task_rq_unlock(rq
, &flags
);
6045 EXPORT_SYMBOL(set_user_nice
);
6048 * can_nice - check if a task can reduce its nice value
6052 int can_nice(const struct task_struct
*p
, const int nice
)
6054 /* convert nice value [19,-20] to rlimit style value [1,40] */
6055 int nice_rlim
= 20 - nice
;
6057 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6058 capable(CAP_SYS_NICE
));
6061 #ifdef __ARCH_WANT_SYS_NICE
6064 * sys_nice - change the priority of the current process.
6065 * @increment: priority increment
6067 * sys_setpriority is a more generic, but much slower function that
6068 * does similar things.
6070 SYSCALL_DEFINE1(nice
, int, increment
)
6075 * Setpriority might change our priority at the same moment.
6076 * We don't have to worry. Conceptually one call occurs first
6077 * and we have a single winner.
6079 if (increment
< -40)
6084 nice
= TASK_NICE(current
) + increment
;
6090 if (increment
< 0 && !can_nice(current
, nice
))
6093 retval
= security_task_setnice(current
, nice
);
6097 set_user_nice(current
, nice
);
6104 * task_prio - return the priority value of a given task.
6105 * @p: the task in question.
6107 * This is the priority value as seen by users in /proc.
6108 * RT tasks are offset by -200. Normal tasks are centered
6109 * around 0, value goes from -16 to +15.
6111 int task_prio(const struct task_struct
*p
)
6113 return p
->prio
- MAX_RT_PRIO
;
6117 * task_nice - return the nice value of a given task.
6118 * @p: the task in question.
6120 int task_nice(const struct task_struct
*p
)
6122 return TASK_NICE(p
);
6124 EXPORT_SYMBOL(task_nice
);
6127 * idle_cpu - is a given cpu idle currently?
6128 * @cpu: the processor in question.
6130 int idle_cpu(int cpu
)
6132 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6136 * idle_task - return the idle task for a given cpu.
6137 * @cpu: the processor in question.
6139 struct task_struct
*idle_task(int cpu
)
6141 return cpu_rq(cpu
)->idle
;
6145 * find_process_by_pid - find a process with a matching PID value.
6146 * @pid: the pid in question.
6148 static struct task_struct
*find_process_by_pid(pid_t pid
)
6150 return pid
? find_task_by_vpid(pid
) : current
;
6153 /* Actually do priority change: must hold rq lock. */
6155 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6157 BUG_ON(p
->se
.on_rq
);
6160 switch (p
->policy
) {
6164 p
->sched_class
= &fair_sched_class
;
6168 p
->sched_class
= &rt_sched_class
;
6172 p
->rt_priority
= prio
;
6173 p
->normal_prio
= normal_prio(p
);
6174 /* we are holding p->pi_lock already */
6175 p
->prio
= rt_mutex_getprio(p
);
6180 * check the target process has a UID that matches the current process's
6182 static bool check_same_owner(struct task_struct
*p
)
6184 const struct cred
*cred
= current_cred(), *pcred
;
6188 pcred
= __task_cred(p
);
6189 match
= (cred
->euid
== pcred
->euid
||
6190 cred
->euid
== pcred
->uid
);
6195 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6196 struct sched_param
*param
, bool user
)
6198 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6199 unsigned long flags
;
6200 const struct sched_class
*prev_class
= p
->sched_class
;
6203 /* may grab non-irq protected spin_locks */
6204 BUG_ON(in_interrupt());
6206 /* double check policy once rq lock held */
6208 policy
= oldpolicy
= p
->policy
;
6209 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6210 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6211 policy
!= SCHED_IDLE
)
6214 * Valid priorities for SCHED_FIFO and SCHED_RR are
6215 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6216 * SCHED_BATCH and SCHED_IDLE is 0.
6218 if (param
->sched_priority
< 0 ||
6219 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6220 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6222 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6226 * Allow unprivileged RT tasks to decrease priority:
6228 if (user
&& !capable(CAP_SYS_NICE
)) {
6229 if (rt_policy(policy
)) {
6230 unsigned long rlim_rtprio
;
6232 if (!lock_task_sighand(p
, &flags
))
6234 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6235 unlock_task_sighand(p
, &flags
);
6237 /* can't set/change the rt policy */
6238 if (policy
!= p
->policy
&& !rlim_rtprio
)
6241 /* can't increase priority */
6242 if (param
->sched_priority
> p
->rt_priority
&&
6243 param
->sched_priority
> rlim_rtprio
)
6247 * Like positive nice levels, dont allow tasks to
6248 * move out of SCHED_IDLE either:
6250 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6253 /* can't change other user's priorities */
6254 if (!check_same_owner(p
))
6259 #ifdef CONFIG_RT_GROUP_SCHED
6261 * Do not allow realtime tasks into groups that have no runtime
6264 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6265 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6269 retval
= security_task_setscheduler(p
, policy
, param
);
6275 * make sure no PI-waiters arrive (or leave) while we are
6276 * changing the priority of the task:
6278 spin_lock_irqsave(&p
->pi_lock
, flags
);
6280 * To be able to change p->policy safely, the apropriate
6281 * runqueue lock must be held.
6283 rq
= __task_rq_lock(p
);
6284 /* recheck policy now with rq lock held */
6285 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6286 policy
= oldpolicy
= -1;
6287 __task_rq_unlock(rq
);
6288 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6291 update_rq_clock(rq
);
6292 on_rq
= p
->se
.on_rq
;
6293 running
= task_current(rq
, p
);
6295 deactivate_task(rq
, p
, 0);
6297 p
->sched_class
->put_prev_task(rq
, p
);
6300 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6303 p
->sched_class
->set_curr_task(rq
);
6305 activate_task(rq
, p
, 0);
6307 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6309 __task_rq_unlock(rq
);
6310 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6312 rt_mutex_adjust_pi(p
);
6318 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6319 * @p: the task in question.
6320 * @policy: new policy.
6321 * @param: structure containing the new RT priority.
6323 * NOTE that the task may be already dead.
6325 int sched_setscheduler(struct task_struct
*p
, int policy
,
6326 struct sched_param
*param
)
6328 return __sched_setscheduler(p
, policy
, param
, true);
6330 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6333 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6334 * @p: the task in question.
6335 * @policy: new policy.
6336 * @param: structure containing the new RT priority.
6338 * Just like sched_setscheduler, only don't bother checking if the
6339 * current context has permission. For example, this is needed in
6340 * stop_machine(): we create temporary high priority worker threads,
6341 * but our caller might not have that capability.
6343 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6344 struct sched_param
*param
)
6346 return __sched_setscheduler(p
, policy
, param
, false);
6350 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6352 struct sched_param lparam
;
6353 struct task_struct
*p
;
6356 if (!param
|| pid
< 0)
6358 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6363 p
= find_process_by_pid(pid
);
6365 retval
= sched_setscheduler(p
, policy
, &lparam
);
6372 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6373 * @pid: the pid in question.
6374 * @policy: new policy.
6375 * @param: structure containing the new RT priority.
6377 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6378 struct sched_param __user
*, param
)
6380 /* negative values for policy are not valid */
6384 return do_sched_setscheduler(pid
, policy
, param
);
6388 * sys_sched_setparam - set/change the RT priority of a thread
6389 * @pid: the pid in question.
6390 * @param: structure containing the new RT priority.
6392 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6394 return do_sched_setscheduler(pid
, -1, param
);
6398 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6399 * @pid: the pid in question.
6401 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6403 struct task_struct
*p
;
6410 read_lock(&tasklist_lock
);
6411 p
= find_process_by_pid(pid
);
6413 retval
= security_task_getscheduler(p
);
6417 read_unlock(&tasklist_lock
);
6422 * sys_sched_getscheduler - get the RT priority of a thread
6423 * @pid: the pid in question.
6424 * @param: structure containing the RT priority.
6426 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6428 struct sched_param lp
;
6429 struct task_struct
*p
;
6432 if (!param
|| pid
< 0)
6435 read_lock(&tasklist_lock
);
6436 p
= find_process_by_pid(pid
);
6441 retval
= security_task_getscheduler(p
);
6445 lp
.sched_priority
= p
->rt_priority
;
6446 read_unlock(&tasklist_lock
);
6449 * This one might sleep, we cannot do it with a spinlock held ...
6451 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6456 read_unlock(&tasklist_lock
);
6460 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6462 cpumask_var_t cpus_allowed
, new_mask
;
6463 struct task_struct
*p
;
6467 read_lock(&tasklist_lock
);
6469 p
= find_process_by_pid(pid
);
6471 read_unlock(&tasklist_lock
);
6477 * It is not safe to call set_cpus_allowed with the
6478 * tasklist_lock held. We will bump the task_struct's
6479 * usage count and then drop tasklist_lock.
6482 read_unlock(&tasklist_lock
);
6484 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6488 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6490 goto out_free_cpus_allowed
;
6493 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6496 retval
= security_task_setscheduler(p
, 0, NULL
);
6500 cpuset_cpus_allowed(p
, cpus_allowed
);
6501 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6503 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6506 cpuset_cpus_allowed(p
, cpus_allowed
);
6507 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6509 * We must have raced with a concurrent cpuset
6510 * update. Just reset the cpus_allowed to the
6511 * cpuset's cpus_allowed
6513 cpumask_copy(new_mask
, cpus_allowed
);
6518 free_cpumask_var(new_mask
);
6519 out_free_cpus_allowed
:
6520 free_cpumask_var(cpus_allowed
);
6527 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6528 struct cpumask
*new_mask
)
6530 if (len
< cpumask_size())
6531 cpumask_clear(new_mask
);
6532 else if (len
> cpumask_size())
6533 len
= cpumask_size();
6535 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6539 * sys_sched_setaffinity - set the cpu affinity of a process
6540 * @pid: pid of the process
6541 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6542 * @user_mask_ptr: user-space pointer to the new cpu mask
6544 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6545 unsigned long __user
*, user_mask_ptr
)
6547 cpumask_var_t new_mask
;
6550 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6553 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6555 retval
= sched_setaffinity(pid
, new_mask
);
6556 free_cpumask_var(new_mask
);
6560 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6562 struct task_struct
*p
;
6566 read_lock(&tasklist_lock
);
6569 p
= find_process_by_pid(pid
);
6573 retval
= security_task_getscheduler(p
);
6577 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6580 read_unlock(&tasklist_lock
);
6587 * sys_sched_getaffinity - get the cpu affinity of a process
6588 * @pid: pid of the process
6589 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6590 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6592 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6593 unsigned long __user
*, user_mask_ptr
)
6598 if (len
< cpumask_size())
6601 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6604 ret
= sched_getaffinity(pid
, mask
);
6606 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6609 ret
= cpumask_size();
6611 free_cpumask_var(mask
);
6617 * sys_sched_yield - yield the current processor to other threads.
6619 * This function yields the current CPU to other tasks. If there are no
6620 * other threads running on this CPU then this function will return.
6622 SYSCALL_DEFINE0(sched_yield
)
6624 struct rq
*rq
= this_rq_lock();
6626 schedstat_inc(rq
, yld_count
);
6627 current
->sched_class
->yield_task(rq
);
6630 * Since we are going to call schedule() anyway, there's
6631 * no need to preempt or enable interrupts:
6633 spin_unlock_no_resched(&rq
->lock
);
6638 preempt_check_resched();
6643 #if defined(CONFIG_DEBUG_SPINLOCK_SLEEP) || defined(CONFIG_DEBUG_PREEMPT)
6644 void __might_sleep(char *file
, int line
)
6647 static unsigned long prev_jiffy
; /* ratelimiting */
6649 if ((!in_atomic() && !irqs_disabled()) ||
6650 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
6653 if (debug_direct_keyboard
&& hardirq_count())
6656 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6658 prev_jiffy
= jiffies
;
6661 "BUG: sleeping function called from invalid context at %s:%d\n",
6664 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
6665 in_atomic(), irqs_disabled(),
6666 current
->pid
, current
->comm
);
6668 debug_show_held_locks(current
);
6669 if (irqs_disabled())
6670 print_irqtrace_events(current
);
6674 EXPORT_SYMBOL(__might_sleep
);
6677 static void __cond_resched(void)
6679 #if defined(CONFIG_DEBUG_SPINLOCK_SLEEP) || defined(CONFIG_DEBUG_PREEMPT)
6680 __might_sleep(__FILE__
, __LINE__
);
6683 * The BKS might be reacquired before we have dropped
6684 * PREEMPT_ACTIVE, which could trigger a second
6685 * cond_resched() call.
6688 local_irq_disable();
6689 add_preempt_count(PREEMPT_ACTIVE
);
6691 } while (need_resched());
6695 int __sched
_cond_resched(void)
6697 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6698 system_state
== SYSTEM_RUNNING
) {
6704 EXPORT_SYMBOL(_cond_resched
);
6707 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6708 * call schedule, and on return reacquire the lock.
6710 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6711 * operations here to prevent schedule() from being called twice (once via
6712 * spin_unlock(), once by hand).
6714 int __cond_resched_raw_spinlock(raw_spinlock_t
*lock
)
6716 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6719 if (spin_needbreak(lock
) || resched
) {
6720 spin_unlock_no_resched(lock
);
6721 if (resched
&& need_resched())
6730 EXPORT_SYMBOL(__cond_resched_raw_spinlock
);
6732 #ifdef CONFIG_PREEMPT_RT
6734 int __cond_resched_spinlock(spinlock_t
*lock
)
6736 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6738 if (spin_needbreak(lock
) || resched
) {
6739 spin_unlock_no_resched(lock
);
6746 EXPORT_SYMBOL(__cond_resched_spinlock
);
6751 * Voluntarily preempt a process context that has softirqs disabled:
6753 int __sched
cond_resched_softirq(void)
6755 #ifndef CONFIG_PREEMPT_SOFTIRQS
6756 WARN_ON_ONCE(!in_softirq());
6760 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6768 EXPORT_SYMBOL(cond_resched_softirq
);
6771 * Voluntarily preempt a softirq context (possible with softirq threading):
6773 int __sched
cond_resched_softirq_context(void)
6775 WARN_ON_ONCE(!in_softirq());
6777 if (softirq_need_resched() && system_state
== SYSTEM_RUNNING
) {
6778 raw_local_irq_disable();
6780 raw_local_irq_enable();
6787 EXPORT_SYMBOL(cond_resched_softirq_context
);
6790 * Preempt a hardirq context if necessary (possible with hardirq threading):
6792 int cond_resched_hardirq_context(void)
6794 WARN_ON_ONCE(!in_irq());
6795 WARN_ON_ONCE(!irqs_disabled());
6797 if (hardirq_need_resched()) {
6798 #ifndef CONFIG_PREEMPT_RT
6803 #ifndef CONFIG_PREEMPT_RT
6804 local_irq_disable();
6812 EXPORT_SYMBOL(cond_resched_hardirq_context
);
6814 #ifdef CONFIG_PREEMPT_VOLUNTARY
6816 int voluntary_preemption
= 1;
6818 EXPORT_SYMBOL(voluntary_preemption
);
6820 static int __init
voluntary_preempt_setup (char *str
)
6822 if (!strncmp(str
, "off", 3))
6823 voluntary_preemption
= 0;
6825 get_option(&str
, &voluntary_preemption
);
6826 if (!voluntary_preemption
)
6827 printk("turning off voluntary preemption!\n");
6832 __setup("voluntary-preempt=", voluntary_preempt_setup
);
6837 * yield - yield the current processor to other threads.
6839 * This is a shortcut for kernel-space yielding - it marks the
6840 * thread runnable and calls sys_sched_yield().
6842 void __sched
__yield(void)
6844 set_current_state(TASK_RUNNING
);
6848 void __sched
yield(void)
6850 static int once
= 1;
6853 * it's a bug to rely on yield() with RT priorities. We print
6854 * the first occurance after bootup ... this will still give
6855 * us an idea about the scope of the problem, without spamming
6858 if (once
&& rt_task(current
)) {
6860 printk(KERN_ERR
"BUG: %s:%d RT task yield()-ing!\n",
6861 current
->comm
, current
->pid
);
6866 EXPORT_SYMBOL(yield
);
6869 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6870 * that process accounting knows that this is a task in IO wait state.
6872 * But don't do that if it is a deliberate, throttling IO wait (this task
6873 * has set its backing_dev_info: the queue against which it should throttle)
6875 void __sched
io_schedule(void)
6877 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6879 delayacct_blkio_start();
6880 atomic_inc(&rq
->nr_iowait
);
6882 atomic_dec(&rq
->nr_iowait
);
6883 delayacct_blkio_end();
6885 EXPORT_SYMBOL(io_schedule
);
6887 long __sched
io_schedule_timeout(long timeout
)
6889 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6892 delayacct_blkio_start();
6893 atomic_inc(&rq
->nr_iowait
);
6894 ret
= schedule_timeout(timeout
);
6895 atomic_dec(&rq
->nr_iowait
);
6896 delayacct_blkio_end();
6901 * sys_sched_get_priority_max - return maximum RT priority.
6902 * @policy: scheduling class.
6904 * this syscall returns the maximum rt_priority that can be used
6905 * by a given scheduling class.
6907 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6914 ret
= MAX_USER_RT_PRIO
-1;
6926 * sys_sched_get_priority_min - return minimum RT priority.
6927 * @policy: scheduling class.
6929 * this syscall returns the minimum rt_priority that can be used
6930 * by a given scheduling class.
6932 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6950 * sys_sched_rr_get_interval - return the default timeslice of a process.
6951 * @pid: pid of the process.
6952 * @interval: userspace pointer to the timeslice value.
6954 * this syscall writes the default timeslice value of a given process
6955 * into the user-space timespec buffer. A value of '0' means infinity.
6957 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6958 struct timespec __user
*, interval
)
6960 struct task_struct
*p
;
6961 unsigned int time_slice
;
6969 read_lock(&tasklist_lock
);
6970 p
= find_process_by_pid(pid
);
6974 retval
= security_task_getscheduler(p
);
6979 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6980 * tasks that are on an otherwise idle runqueue:
6983 if (p
->policy
== SCHED_RR
) {
6984 time_slice
= DEF_TIMESLICE
;
6985 } else if (p
->policy
!= SCHED_FIFO
) {
6986 struct sched_entity
*se
= &p
->se
;
6987 unsigned long flags
;
6990 rq
= task_rq_lock(p
, &flags
);
6991 if (rq
->cfs
.load
.weight
)
6992 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6993 task_rq_unlock(rq
, &flags
);
6995 read_unlock(&tasklist_lock
);
6996 jiffies_to_timespec(time_slice
, &t
);
6997 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
7001 read_unlock(&tasklist_lock
);
7005 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
7007 void sched_show_task(struct task_struct
*p
)
7009 unsigned long free
= 0;
7012 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
7013 printk("%-13.13s %c (%03lx) [%p]", p
->comm
,
7014 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?',
7015 (unsigned long) p
->state
, p
);
7016 #if BITS_PER_LONG == 32
7017 if (0 && (state
== TASK_RUNNING
))
7018 printk(KERN_CONT
" running ");
7020 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
7022 if (0 && (state
== TASK_RUNNING
))
7023 printk(KERN_CONT
" running task ");
7025 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
7029 else if (p
->se
.on_rq
)
7030 printk("[on rq #%d] ", task_cpu(p
));
7031 #ifdef CONFIG_DEBUG_STACK_USAGE
7032 free
= stack_not_used(p
);
7034 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
7035 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
7037 show_stack(p
, NULL
);
7040 void show_state_filter(unsigned long state_filter
)
7042 struct task_struct
*g
, *p
;
7045 #if BITS_PER_LONG == 32
7047 " task PC stack pid father\n");
7050 " task PC stack pid father\n");
7052 #ifdef CONFIG_PREEMPT_RT
7053 if (!read_trylock(&tasklist_lock
)) {
7054 printk("hm, tasklist_lock write-locked.\n");
7055 printk("ignoring ...\n");
7059 read_lock(&tasklist_lock
);
7062 do_each_thread(g
, p
) {
7064 * reset the NMI-timeout, listing all files on a slow
7065 * console might take alot of time:
7067 touch_nmi_watchdog();
7068 if (!state_filter
|| (p
->state
& state_filter
))
7070 } while_each_thread(g
, p
);
7072 touch_all_softlockup_watchdogs();
7074 #ifdef CONFIG_SCHED_DEBUG
7075 sysrq_sched_debug_show();
7078 read_unlock(&tasklist_lock
);
7080 * Only show locks if all tasks are dumped:
7082 if (state_filter
== -1)
7083 debug_show_all_locks();
7086 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
7088 idle
->sched_class
= &idle_sched_class
;
7092 * init_idle - set up an idle thread for a given CPU
7093 * @idle: task in question
7094 * @cpu: cpu the idle task belongs to
7096 * NOTE: this function does not set the idle thread's NEED_RESCHED
7097 * flag, to make booting more robust.
7099 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7101 struct rq
*rq
= cpu_rq(cpu
);
7102 unsigned long flags
;
7104 spin_lock_irqsave(&rq
->lock
, flags
);
7107 idle
->se
.exec_start
= sched_clock();
7109 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
7110 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7111 __set_task_cpu(idle
, cpu
);
7113 rq
->curr
= rq
->idle
= idle
;
7114 #if defined(CONFIG_SMP)
7117 spin_unlock_irqrestore(&rq
->lock
, flags
);
7119 /* Set the preempt count _outside_ the spinlocks! */
7120 task_thread_info(idle
)->preempt_count
= 0;
7123 * The idle tasks have their own, simple scheduling class:
7125 idle
->sched_class
= &idle_sched_class
;
7126 ftrace_graph_init_task(idle
);
7130 * In a system that switches off the HZ timer nohz_cpu_mask
7131 * indicates which cpus entered this state. This is used
7132 * in the rcu update to wait only for active cpus. For system
7133 * which do not switch off the HZ timer nohz_cpu_mask should
7134 * always be CPU_BITS_NONE.
7136 cpumask_var_t nohz_cpu_mask
;
7139 * Increase the granularity value when there are more CPUs,
7140 * because with more CPUs the 'effective latency' as visible
7141 * to users decreases. But the relationship is not linear,
7142 * so pick a second-best guess by going with the log2 of the
7145 * This idea comes from the SD scheduler of Con Kolivas:
7147 static inline void sched_init_granularity(void)
7149 unsigned int factor
= 1 + ilog2(num_online_cpus());
7150 const unsigned long limit
= 200000000;
7152 sysctl_sched_min_granularity
*= factor
;
7153 if (sysctl_sched_min_granularity
> limit
)
7154 sysctl_sched_min_granularity
= limit
;
7156 sysctl_sched_latency
*= factor
;
7157 if (sysctl_sched_latency
> limit
)
7158 sysctl_sched_latency
= limit
;
7160 sysctl_sched_wakeup_granularity
*= factor
;
7162 sysctl_sched_shares_ratelimit
*= factor
;
7167 * This is how migration works:
7169 * 1) we queue a struct migration_req structure in the source CPU's
7170 * runqueue and wake up that CPU's migration thread.
7171 * 2) we down() the locked semaphore => thread blocks.
7172 * 3) migration thread wakes up (implicitly it forces the migrated
7173 * thread off the CPU)
7174 * 4) it gets the migration request and checks whether the migrated
7175 * task is still in the wrong runqueue.
7176 * 5) if it's in the wrong runqueue then the migration thread removes
7177 * it and puts it into the right queue.
7178 * 6) migration thread up()s the semaphore.
7179 * 7) we wake up and the migration is done.
7183 * Change a given task's CPU affinity. Migrate the thread to a
7184 * proper CPU and schedule it away if the CPU it's executing on
7185 * is removed from the allowed bitmask.
7187 * NOTE: the caller must have a valid reference to the task, the
7188 * task must not exit() & deallocate itself prematurely. The
7189 * call is not atomic; no spinlocks may be held.
7191 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7193 struct migration_req req
;
7194 unsigned long flags
;
7198 rq
= task_rq_lock(p
, &flags
);
7199 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
7204 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7205 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7210 if (p
->sched_class
->set_cpus_allowed
)
7211 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7213 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7214 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7217 /* Can the task run on the task's current CPU? If so, we're done */
7218 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7221 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
7222 /* Need help from migration thread: drop lock and wait. */
7223 task_rq_unlock(rq
, &flags
);
7224 wake_up_process(rq
->migration_thread
);
7225 wait_for_completion(&req
.done
);
7226 tlb_migrate_finish(p
->mm
);
7230 task_rq_unlock(rq
, &flags
);
7234 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7237 * Move (not current) task off this cpu, onto dest cpu. We're doing
7238 * this because either it can't run here any more (set_cpus_allowed()
7239 * away from this CPU, or CPU going down), or because we're
7240 * attempting to rebalance this task on exec (sched_exec).
7242 * So we race with normal scheduler movements, but that's OK, as long
7243 * as the task is no longer on this CPU.
7245 * Returns non-zero if task was successfully migrated.
7247 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7249 struct rq
*rq_dest
, *rq_src
;
7250 unsigned long flags
;
7253 if (unlikely(!cpu_active(dest_cpu
)))
7257 * PREEMPT_RT: this relies on write_lock_irq(&tasklist_lock)
7258 * disabling interrupts - which on PREEMPT_RT does not do:
7260 local_irq_save(flags
);
7262 rq_src
= cpu_rq(src_cpu
);
7263 rq_dest
= cpu_rq(dest_cpu
);
7265 double_rq_lock(rq_src
, rq_dest
);
7266 /* Already moved. */
7267 if (task_cpu(p
) != src_cpu
)
7269 /* Affinity changed (again). */
7270 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7273 on_rq
= p
->se
.on_rq
;
7275 deactivate_task(rq_src
, p
, 0);
7277 set_task_cpu(p
, dest_cpu
);
7279 activate_task(rq_dest
, p
, 0);
7280 check_preempt_curr(rq_dest
, p
, 0);
7285 double_rq_unlock(rq_src
, rq_dest
);
7286 local_irq_restore(flags
);
7292 * migration_thread - this is a highprio system thread that performs
7293 * thread migration by bumping thread off CPU then 'pushing' onto
7296 static int migration_thread(void *data
)
7298 int cpu
= (long)data
;
7302 BUG_ON(rq
->migration_thread
!= current
);
7304 set_current_state(TASK_INTERRUPTIBLE
);
7305 while (!kthread_should_stop()) {
7306 struct migration_req
*req
;
7307 struct list_head
*head
;
7309 spin_lock_irq(&rq
->lock
);
7311 if (cpu_is_offline(cpu
)) {
7312 spin_unlock_irq(&rq
->lock
);
7316 if (rq
->active_balance
) {
7317 active_load_balance(rq
, cpu
);
7318 rq
->active_balance
= 0;
7321 head
= &rq
->migration_queue
;
7323 if (list_empty(head
)) {
7324 spin_unlock_irq(&rq
->lock
);
7326 set_current_state(TASK_INTERRUPTIBLE
);
7329 req
= list_entry(head
->next
, struct migration_req
, list
);
7330 list_del_init(head
->next
);
7332 spin_unlock(&rq
->lock
);
7333 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7336 complete(&req
->done
);
7338 __set_current_state(TASK_RUNNING
);
7342 /* Wait for kthread_stop */
7343 set_current_state(TASK_INTERRUPTIBLE
);
7344 while (!kthread_should_stop()) {
7346 set_current_state(TASK_INTERRUPTIBLE
);
7348 __set_current_state(TASK_RUNNING
);
7352 #ifdef CONFIG_HOTPLUG_CPU
7354 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7358 local_irq_disable();
7359 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7365 * Figure out where task on dead CPU should go, use force if necessary.
7367 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7370 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7373 /* Look for allowed, online CPU in same node. */
7374 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
7375 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7378 /* Any allowed, online CPU? */
7379 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
7380 if (dest_cpu
< nr_cpu_ids
)
7383 /* No more Mr. Nice Guy. */
7384 if (dest_cpu
>= nr_cpu_ids
) {
7385 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7386 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
7389 * Don't tell them about moving exiting tasks or
7390 * kernel threads (both mm NULL), since they never
7393 if (p
->mm
&& printk_ratelimit()) {
7394 printk(KERN_INFO
"process %d (%s) no "
7395 "longer affine to cpu%d\n",
7396 task_pid_nr(p
), p
->comm
, dead_cpu
);
7401 /* It can have affinity changed while we were choosing. */
7402 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7407 * While a dead CPU has no uninterruptible tasks queued at this point,
7408 * it might still have a nonzero ->nr_uninterruptible counter, because
7409 * for performance reasons the counter is not stricly tracking tasks to
7410 * their home CPUs. So we just add the counter to another CPU's counter,
7411 * to keep the global sum constant after CPU-down:
7413 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7415 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
7416 unsigned long flags
;
7418 local_irq_save(flags
);
7419 double_rq_lock(rq_src
, rq_dest
);
7420 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7421 rq_src
->nr_uninterruptible
= 0;
7422 double_rq_unlock(rq_src
, rq_dest
);
7423 local_irq_restore(flags
);
7426 /* Run through task list and migrate tasks from the dead cpu. */
7427 static void migrate_live_tasks(int src_cpu
)
7429 struct task_struct
*p
, *t
;
7431 read_lock(&tasklist_lock
);
7433 do_each_thread(t
, p
) {
7437 if (task_cpu(p
) == src_cpu
)
7438 move_task_off_dead_cpu(src_cpu
, p
);
7439 } while_each_thread(t
, p
);
7441 read_unlock(&tasklist_lock
);
7445 * Schedules idle task to be the next runnable task on current CPU.
7446 * It does so by boosting its priority to highest possible.
7447 * Used by CPU offline code.
7449 void sched_idle_next(void)
7451 int this_cpu
= smp_processor_id();
7452 struct rq
*rq
= cpu_rq(this_cpu
);
7453 struct task_struct
*p
= rq
->idle
;
7454 unsigned long flags
;
7456 /* cpu has to be offline */
7457 BUG_ON(cpu_online(this_cpu
));
7460 * Strictly not necessary since rest of the CPUs are stopped by now
7461 * and interrupts disabled on the current cpu.
7463 spin_lock_irqsave(&rq
->lock
, flags
);
7465 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7467 update_rq_clock(rq
);
7468 activate_task(rq
, p
, 0);
7470 spin_unlock_irqrestore(&rq
->lock
, flags
);
7474 * Ensures that the idle task is using init_mm right before its cpu goes
7477 void idle_task_exit(void)
7479 struct mm_struct
*mm
= current
->active_mm
;
7481 BUG_ON(cpu_online(smp_processor_id()));
7484 switch_mm(mm
, &init_mm
, current
);
7485 #ifdef CONFIG_PREEMPT_RT
7492 /* called under rq->lock with disabled interrupts */
7493 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7495 struct rq
*rq
= cpu_rq(dead_cpu
);
7497 /* Must be exiting, otherwise would be on tasklist. */
7498 BUG_ON(!p
->exit_state
);
7500 /* Cannot have done final schedule yet: would have vanished. */
7501 BUG_ON(p
->state
== TASK_DEAD
);
7506 * Drop lock around migration; if someone else moves it,
7507 * that's OK. No task can be added to this CPU, so iteration is
7510 spin_unlock_irq(&rq
->lock
);
7511 move_task_off_dead_cpu(dead_cpu
, p
);
7512 spin_lock_irq(&rq
->lock
);
7517 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7518 static void migrate_dead_tasks(unsigned int dead_cpu
)
7520 struct rq
*rq
= cpu_rq(dead_cpu
);
7521 struct task_struct
*next
;
7524 if (!rq
->nr_running
)
7526 update_rq_clock(rq
);
7527 next
= pick_next_task(rq
);
7530 next
->sched_class
->put_prev_task(rq
, next
);
7531 migrate_dead(dead_cpu
, next
);
7537 * remove the tasks which were accounted by rq from calc_load_tasks.
7539 static void calc_global_load_remove(struct rq
*rq
)
7541 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7543 #endif /* CONFIG_HOTPLUG_CPU */
7545 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7547 static struct ctl_table sd_ctl_dir
[] = {
7549 .procname
= "sched_domain",
7555 static struct ctl_table sd_ctl_root
[] = {
7557 .ctl_name
= CTL_KERN
,
7558 .procname
= "kernel",
7560 .child
= sd_ctl_dir
,
7565 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7567 struct ctl_table
*entry
=
7568 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7573 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7575 struct ctl_table
*entry
;
7578 * In the intermediate directories, both the child directory and
7579 * procname are dynamically allocated and could fail but the mode
7580 * will always be set. In the lowest directory the names are
7581 * static strings and all have proc handlers.
7583 for (entry
= *tablep
; entry
->mode
; entry
++) {
7585 sd_free_ctl_entry(&entry
->child
);
7586 if (entry
->proc_handler
== NULL
)
7587 kfree(entry
->procname
);
7595 set_table_entry(struct ctl_table
*entry
,
7596 const char *procname
, void *data
, int maxlen
,
7597 mode_t mode
, proc_handler
*proc_handler
)
7599 entry
->procname
= procname
;
7601 entry
->maxlen
= maxlen
;
7603 entry
->proc_handler
= proc_handler
;
7606 static struct ctl_table
*
7607 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7609 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7614 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7615 sizeof(long), 0644, proc_doulongvec_minmax
);
7616 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7617 sizeof(long), 0644, proc_doulongvec_minmax
);
7618 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7619 sizeof(int), 0644, proc_dointvec_minmax
);
7620 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7621 sizeof(int), 0644, proc_dointvec_minmax
);
7622 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7623 sizeof(int), 0644, proc_dointvec_minmax
);
7624 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7625 sizeof(int), 0644, proc_dointvec_minmax
);
7626 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7627 sizeof(int), 0644, proc_dointvec_minmax
);
7628 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7629 sizeof(int), 0644, proc_dointvec_minmax
);
7630 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7631 sizeof(int), 0644, proc_dointvec_minmax
);
7632 set_table_entry(&table
[9], "cache_nice_tries",
7633 &sd
->cache_nice_tries
,
7634 sizeof(int), 0644, proc_dointvec_minmax
);
7635 set_table_entry(&table
[10], "flags", &sd
->flags
,
7636 sizeof(int), 0644, proc_dointvec_minmax
);
7637 set_table_entry(&table
[11], "name", sd
->name
,
7638 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7639 /* &table[12] is terminator */
7644 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7646 struct ctl_table
*entry
, *table
;
7647 struct sched_domain
*sd
;
7648 int domain_num
= 0, i
;
7651 for_each_domain(cpu
, sd
)
7653 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7658 for_each_domain(cpu
, sd
) {
7659 snprintf(buf
, 32, "domain%d", i
);
7660 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7662 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7669 static struct ctl_table_header
*sd_sysctl_header
;
7670 static void register_sched_domain_sysctl(void)
7672 int i
, cpu_num
= num_online_cpus();
7673 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7676 WARN_ON(sd_ctl_dir
[0].child
);
7677 sd_ctl_dir
[0].child
= entry
;
7682 for_each_online_cpu(i
) {
7683 snprintf(buf
, 32, "cpu%d", i
);
7684 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7686 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7690 WARN_ON(sd_sysctl_header
);
7691 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7694 /* may be called multiple times per register */
7695 static void unregister_sched_domain_sysctl(void)
7697 if (sd_sysctl_header
)
7698 unregister_sysctl_table(sd_sysctl_header
);
7699 sd_sysctl_header
= NULL
;
7700 if (sd_ctl_dir
[0].child
)
7701 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7704 static void register_sched_domain_sysctl(void)
7707 static void unregister_sched_domain_sysctl(void)
7712 static void set_rq_online(struct rq
*rq
)
7715 const struct sched_class
*class;
7717 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7720 for_each_class(class) {
7721 if (class->rq_online
)
7722 class->rq_online(rq
);
7727 static void set_rq_offline(struct rq
*rq
)
7730 const struct sched_class
*class;
7732 for_each_class(class) {
7733 if (class->rq_offline
)
7734 class->rq_offline(rq
);
7737 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7743 * migration_call - callback that gets triggered when a CPU is added.
7744 * Here we can start up the necessary migration thread for the new CPU.
7746 static int __cpuinit
7747 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7749 struct task_struct
*p
;
7750 int cpu
= (long)hcpu
;
7751 unsigned long flags
;
7756 case CPU_UP_PREPARE
:
7757 case CPU_UP_PREPARE_FROZEN
:
7758 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7761 kthread_bind(p
, cpu
);
7762 /* Must be high prio: stop_machine expects to yield to it. */
7763 rq
= task_rq_lock(p
, &flags
);
7764 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7765 task_rq_unlock(rq
, &flags
);
7766 cpu_rq(cpu
)->migration_thread
= p
;
7770 case CPU_ONLINE_FROZEN
:
7771 /* Strictly unnecessary, as first user will wake it. */
7772 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7774 /* Update our root-domain */
7776 spin_lock_irqsave(&rq
->lock
, flags
);
7777 rq
->calc_load_update
= calc_load_update
;
7778 rq
->calc_load_active
= 0;
7780 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7784 spin_unlock_irqrestore(&rq
->lock
, flags
);
7787 #ifdef CONFIG_HOTPLUG_CPU
7788 case CPU_UP_CANCELED
:
7789 case CPU_UP_CANCELED_FROZEN
:
7790 if (!cpu_rq(cpu
)->migration_thread
)
7792 /* Unbind it from offline cpu so it can run. Fall thru. */
7793 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7794 cpumask_any(cpu_online_mask
));
7795 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7796 cpu_rq(cpu
)->migration_thread
= NULL
;
7800 case CPU_DEAD_FROZEN
:
7801 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7802 migrate_live_tasks(cpu
);
7804 kthread_stop(rq
->migration_thread
);
7805 rq
->migration_thread
= NULL
;
7806 /* Idle task back to normal (off runqueue, low prio) */
7807 spin_lock_irq(&rq
->lock
);
7808 update_rq_clock(rq
);
7809 deactivate_task(rq
, rq
->idle
, 0);
7810 rq
->idle
->static_prio
= MAX_PRIO
;
7811 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7812 rq
->idle
->sched_class
= &idle_sched_class
;
7813 migrate_dead_tasks(cpu
);
7814 spin_unlock_irq(&rq
->lock
);
7816 migrate_nr_uninterruptible(rq
);
7817 BUG_ON(rq
->nr_running
!= 0);
7818 calc_global_load_remove(rq
);
7820 * No need to migrate the tasks: it was best-effort if
7821 * they didn't take sched_hotcpu_mutex. Just wake up
7824 spin_lock_irq(&rq
->lock
);
7825 while (!list_empty(&rq
->migration_queue
)) {
7826 struct migration_req
*req
;
7828 req
= list_entry(rq
->migration_queue
.next
,
7829 struct migration_req
, list
);
7830 list_del_init(&req
->list
);
7831 spin_unlock_irq(&rq
->lock
);
7832 complete(&req
->done
);
7833 spin_lock_irq(&rq
->lock
);
7835 spin_unlock_irq(&rq
->lock
);
7839 case CPU_DYING_FROZEN
:
7840 /* Update our root-domain */
7842 spin_lock_irqsave(&rq
->lock
, flags
);
7844 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7847 spin_unlock_irqrestore(&rq
->lock
, flags
);
7854 /* Register at highest priority so that task migration (migrate_all_tasks)
7855 * happens before everything else.
7857 static struct notifier_block __cpuinitdata migration_notifier
= {
7858 .notifier_call
= migration_call
,
7862 static int __init
migration_init(void)
7864 void *cpu
= (void *)(long)smp_processor_id();
7867 /* Start one for the boot CPU: */
7868 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7869 BUG_ON(err
== NOTIFY_BAD
);
7870 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7871 register_cpu_notifier(&migration_notifier
);
7875 early_initcall(migration_init
);
7880 #ifdef CONFIG_SCHED_DEBUG
7882 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7883 struct cpumask
*groupmask
)
7885 struct sched_group
*group
= sd
->groups
;
7888 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7889 cpumask_clear(groupmask
);
7891 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7893 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7894 printk("does not load-balance\n");
7896 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7901 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7903 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7904 printk(KERN_ERR
"ERROR: domain->span does not contain "
7907 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7908 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7912 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7916 printk(KERN_ERR
"ERROR: group is NULL\n");
7920 if (!group
->__cpu_power
) {
7921 printk(KERN_CONT
"\n");
7922 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7927 if (!cpumask_weight(sched_group_cpus(group
))) {
7928 printk(KERN_CONT
"\n");
7929 printk(KERN_ERR
"ERROR: empty group\n");
7933 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7934 printk(KERN_CONT
"\n");
7935 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7939 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7941 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7942 printk(KERN_CONT
" %s", str
);
7944 group
= group
->next
;
7945 } while (group
!= sd
->groups
);
7946 printk(KERN_CONT
"\n");
7948 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7949 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7952 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7953 printk(KERN_ERR
"ERROR: parent span is not a superset "
7954 "of domain->span\n");
7958 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7960 cpumask_var_t groupmask
;
7964 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7968 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7970 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7971 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7976 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7983 free_cpumask_var(groupmask
);
7985 #else /* !CONFIG_SCHED_DEBUG */
7986 # define sched_domain_debug(sd, cpu) do { } while (0)
7987 #endif /* CONFIG_SCHED_DEBUG */
7989 static int sd_degenerate(struct sched_domain
*sd
)
7991 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7994 /* Following flags need at least 2 groups */
7995 if (sd
->flags
& (SD_LOAD_BALANCE
|
7996 SD_BALANCE_NEWIDLE
|
8000 SD_SHARE_PKG_RESOURCES
)) {
8001 if (sd
->groups
!= sd
->groups
->next
)
8005 /* Following flags don't use groups */
8006 if (sd
->flags
& (SD_WAKE_IDLE
|
8015 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
8017 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
8019 if (sd_degenerate(parent
))
8022 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
8025 /* Does parent contain flags not in child? */
8026 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
8027 if (cflags
& SD_WAKE_AFFINE
)
8028 pflags
&= ~SD_WAKE_BALANCE
;
8029 /* Flags needing groups don't count if only 1 group in parent */
8030 if (parent
->groups
== parent
->groups
->next
) {
8031 pflags
&= ~(SD_LOAD_BALANCE
|
8032 SD_BALANCE_NEWIDLE
|
8036 SD_SHARE_PKG_RESOURCES
);
8037 if (nr_node_ids
== 1)
8038 pflags
&= ~SD_SERIALIZE
;
8040 if (~cflags
& pflags
)
8046 static void free_rootdomain(struct root_domain
*rd
)
8048 cpupri_cleanup(&rd
->cpupri
);
8050 free_cpumask_var(rd
->rto_mask
);
8051 free_cpumask_var(rd
->online
);
8052 free_cpumask_var(rd
->span
);
8056 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8058 struct root_domain
*old_rd
= NULL
;
8059 unsigned long flags
;
8061 spin_lock_irqsave(&rq
->lock
, flags
);
8066 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8069 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8072 * If we dont want to free the old_rt yet then
8073 * set old_rd to NULL to skip the freeing later
8076 if (!atomic_dec_and_test(&old_rd
->refcount
))
8080 atomic_inc(&rd
->refcount
);
8083 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8084 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
8087 spin_unlock_irqrestore(&rq
->lock
, flags
);
8090 free_rootdomain(old_rd
);
8093 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8095 memset(rd
, 0, sizeof(*rd
));
8098 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
8099 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
8100 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
8101 cpupri_init(&rd
->cpupri
, true);
8105 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
8107 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
8109 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
8112 if (cpupri_init(&rd
->cpupri
, false) != 0)
8117 free_cpumask_var(rd
->rto_mask
);
8119 free_cpumask_var(rd
->online
);
8121 free_cpumask_var(rd
->span
);
8126 static void init_defrootdomain(void)
8128 init_rootdomain(&def_root_domain
, true);
8130 atomic_set(&def_root_domain
.refcount
, 1);
8133 static struct root_domain
*alloc_rootdomain(void)
8135 struct root_domain
*rd
;
8137 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8141 if (init_rootdomain(rd
, false) != 0) {
8150 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8151 * hold the hotplug lock.
8154 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8156 struct rq
*rq
= cpu_rq(cpu
);
8157 struct sched_domain
*tmp
;
8159 /* Remove the sched domains which do not contribute to scheduling. */
8160 for (tmp
= sd
; tmp
; ) {
8161 struct sched_domain
*parent
= tmp
->parent
;
8165 if (sd_parent_degenerate(tmp
, parent
)) {
8166 tmp
->parent
= parent
->parent
;
8168 parent
->parent
->child
= tmp
;
8173 if (sd
&& sd_degenerate(sd
)) {
8179 sched_domain_debug(sd
, cpu
);
8181 rq_attach_root(rq
, rd
);
8182 rcu_assign_pointer(rq
->sd
, sd
);
8185 /* cpus with isolated domains */
8186 static cpumask_var_t cpu_isolated_map
;
8188 /* Setup the mask of cpus configured for isolated domains */
8189 static int __init
isolated_cpu_setup(char *str
)
8191 cpulist_parse(str
, cpu_isolated_map
);
8195 __setup("isolcpus=", isolated_cpu_setup
);
8198 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8199 * to a function which identifies what group(along with sched group) a CPU
8200 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8201 * (due to the fact that we keep track of groups covered with a struct cpumask).
8203 * init_sched_build_groups will build a circular linked list of the groups
8204 * covered by the given span, and will set each group's ->cpumask correctly,
8205 * and ->cpu_power to 0.
8208 init_sched_build_groups(const struct cpumask
*span
,
8209 const struct cpumask
*cpu_map
,
8210 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8211 struct sched_group
**sg
,
8212 struct cpumask
*tmpmask
),
8213 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8215 struct sched_group
*first
= NULL
, *last
= NULL
;
8218 cpumask_clear(covered
);
8220 for_each_cpu(i
, span
) {
8221 struct sched_group
*sg
;
8222 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8225 if (cpumask_test_cpu(i
, covered
))
8228 cpumask_clear(sched_group_cpus(sg
));
8229 sg
->__cpu_power
= 0;
8231 for_each_cpu(j
, span
) {
8232 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8235 cpumask_set_cpu(j
, covered
);
8236 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8247 #define SD_NODES_PER_DOMAIN 16
8252 * find_next_best_node - find the next node to include in a sched_domain
8253 * @node: node whose sched_domain we're building
8254 * @used_nodes: nodes already in the sched_domain
8256 * Find the next node to include in a given scheduling domain. Simply
8257 * finds the closest node not already in the @used_nodes map.
8259 * Should use nodemask_t.
8261 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8263 int i
, n
, val
, min_val
, best_node
= 0;
8267 for (i
= 0; i
< nr_node_ids
; i
++) {
8268 /* Start at @node */
8269 n
= (node
+ i
) % nr_node_ids
;
8271 if (!nr_cpus_node(n
))
8274 /* Skip already used nodes */
8275 if (node_isset(n
, *used_nodes
))
8278 /* Simple min distance search */
8279 val
= node_distance(node
, n
);
8281 if (val
< min_val
) {
8287 node_set(best_node
, *used_nodes
);
8292 * sched_domain_node_span - get a cpumask for a node's sched_domain
8293 * @node: node whose cpumask we're constructing
8294 * @span: resulting cpumask
8296 * Given a node, construct a good cpumask for its sched_domain to span. It
8297 * should be one that prevents unnecessary balancing, but also spreads tasks
8300 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8302 nodemask_t used_nodes
;
8305 cpumask_clear(span
);
8306 nodes_clear(used_nodes
);
8308 cpumask_or(span
, span
, cpumask_of_node(node
));
8309 node_set(node
, used_nodes
);
8311 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8312 int next_node
= find_next_best_node(node
, &used_nodes
);
8314 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8317 #endif /* CONFIG_NUMA */
8319 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8322 * The cpus mask in sched_group and sched_domain hangs off the end.
8323 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
8324 * for nr_cpu_ids < CONFIG_NR_CPUS.
8326 struct static_sched_group
{
8327 struct sched_group sg
;
8328 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8331 struct static_sched_domain
{
8332 struct sched_domain sd
;
8333 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8337 * SMT sched-domains:
8339 #ifdef CONFIG_SCHED_SMT
8340 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8341 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
8344 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8345 struct sched_group
**sg
, struct cpumask
*unused
)
8348 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
8351 #endif /* CONFIG_SCHED_SMT */
8354 * multi-core sched-domains:
8356 #ifdef CONFIG_SCHED_MC
8357 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8358 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8359 #endif /* CONFIG_SCHED_MC */
8361 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8363 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8364 struct sched_group
**sg
, struct cpumask
*mask
)
8368 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8369 group
= cpumask_first(mask
);
8371 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8374 #elif defined(CONFIG_SCHED_MC)
8376 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8377 struct sched_group
**sg
, struct cpumask
*unused
)
8380 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8385 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8386 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8389 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8390 struct sched_group
**sg
, struct cpumask
*mask
)
8393 #ifdef CONFIG_SCHED_MC
8394 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8395 group
= cpumask_first(mask
);
8396 #elif defined(CONFIG_SCHED_SMT)
8397 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8398 group
= cpumask_first(mask
);
8403 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8409 * The init_sched_build_groups can't handle what we want to do with node
8410 * groups, so roll our own. Now each node has its own list of groups which
8411 * gets dynamically allocated.
8413 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8414 static struct sched_group
***sched_group_nodes_bycpu
;
8416 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8417 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8419 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8420 struct sched_group
**sg
,
8421 struct cpumask
*nodemask
)
8425 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8426 group
= cpumask_first(nodemask
);
8429 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8433 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8435 struct sched_group
*sg
= group_head
;
8441 for_each_cpu(j
, sched_group_cpus(sg
)) {
8442 struct sched_domain
*sd
;
8444 sd
= &per_cpu(phys_domains
, j
).sd
;
8445 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
8447 * Only add "power" once for each
8453 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
8456 } while (sg
!= group_head
);
8458 #endif /* CONFIG_NUMA */
8461 /* Free memory allocated for various sched_group structures */
8462 static void free_sched_groups(const struct cpumask
*cpu_map
,
8463 struct cpumask
*nodemask
)
8467 for_each_cpu(cpu
, cpu_map
) {
8468 struct sched_group
**sched_group_nodes
8469 = sched_group_nodes_bycpu
[cpu
];
8471 if (!sched_group_nodes
)
8474 for (i
= 0; i
< nr_node_ids
; i
++) {
8475 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8477 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8478 if (cpumask_empty(nodemask
))
8488 if (oldsg
!= sched_group_nodes
[i
])
8491 kfree(sched_group_nodes
);
8492 sched_group_nodes_bycpu
[cpu
] = NULL
;
8495 #else /* !CONFIG_NUMA */
8496 static void free_sched_groups(const struct cpumask
*cpu_map
,
8497 struct cpumask
*nodemask
)
8500 #endif /* CONFIG_NUMA */
8503 * Initialize sched groups cpu_power.
8505 * cpu_power indicates the capacity of sched group, which is used while
8506 * distributing the load between different sched groups in a sched domain.
8507 * Typically cpu_power for all the groups in a sched domain will be same unless
8508 * there are asymmetries in the topology. If there are asymmetries, group
8509 * having more cpu_power will pickup more load compared to the group having
8512 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
8513 * the maximum number of tasks a group can handle in the presence of other idle
8514 * or lightly loaded groups in the same sched domain.
8516 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8518 struct sched_domain
*child
;
8519 struct sched_group
*group
;
8521 WARN_ON(!sd
|| !sd
->groups
);
8523 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
8528 sd
->groups
->__cpu_power
= 0;
8531 * For perf policy, if the groups in child domain share resources
8532 * (for example cores sharing some portions of the cache hierarchy
8533 * or SMT), then set this domain groups cpu_power such that each group
8534 * can handle only one task, when there are other idle groups in the
8535 * same sched domain.
8537 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8539 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8540 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8545 * add cpu_power of each child group to this groups cpu_power
8547 group
= child
->groups
;
8549 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8550 group
= group
->next
;
8551 } while (group
!= child
->groups
);
8555 * Initializers for schedule domains
8556 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8559 #ifdef CONFIG_SCHED_DEBUG
8560 # define SD_INIT_NAME(sd, type) sd->name = #type
8562 # define SD_INIT_NAME(sd, type) do { } while (0)
8565 #define SD_INIT(sd, type) sd_init_##type(sd)
8567 #define SD_INIT_FUNC(type) \
8568 static noinline void sd_init_##type(struct sched_domain *sd) \
8570 memset(sd, 0, sizeof(*sd)); \
8571 *sd = SD_##type##_INIT; \
8572 sd->level = SD_LV_##type; \
8573 SD_INIT_NAME(sd, type); \
8578 SD_INIT_FUNC(ALLNODES
)
8581 #ifdef CONFIG_SCHED_SMT
8582 SD_INIT_FUNC(SIBLING
)
8584 #ifdef CONFIG_SCHED_MC
8588 static int default_relax_domain_level
= -1;
8590 static int __init
setup_relax_domain_level(char *str
)
8594 val
= simple_strtoul(str
, NULL
, 0);
8595 if (val
< SD_LV_MAX
)
8596 default_relax_domain_level
= val
;
8600 __setup("relax_domain_level=", setup_relax_domain_level
);
8602 static void set_domain_attribute(struct sched_domain
*sd
,
8603 struct sched_domain_attr
*attr
)
8607 if (!attr
|| attr
->relax_domain_level
< 0) {
8608 if (default_relax_domain_level
< 0)
8611 request
= default_relax_domain_level
;
8613 request
= attr
->relax_domain_level
;
8614 if (request
< sd
->level
) {
8615 /* turn off idle balance on this domain */
8616 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8618 /* turn on idle balance on this domain */
8619 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8624 * Build sched domains for a given set of cpus and attach the sched domains
8625 * to the individual cpus
8627 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8628 struct sched_domain_attr
*attr
)
8630 int i
, err
= -ENOMEM
;
8631 struct root_domain
*rd
;
8632 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8635 cpumask_var_t domainspan
, covered
, notcovered
;
8636 struct sched_group
**sched_group_nodes
= NULL
;
8637 int sd_allnodes
= 0;
8639 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8641 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8642 goto free_domainspan
;
8643 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8647 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8648 goto free_notcovered
;
8649 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8651 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8652 goto free_this_sibling_map
;
8653 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8654 goto free_this_core_map
;
8655 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8656 goto free_send_covered
;
8660 * Allocate the per-node list of sched groups
8662 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8664 if (!sched_group_nodes
) {
8665 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8670 rd
= alloc_rootdomain();
8672 printk(KERN_WARNING
"Cannot alloc root domain\n");
8673 goto free_sched_groups
;
8677 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8681 * Set up domains for cpus specified by the cpu_map.
8683 for_each_cpu(i
, cpu_map
) {
8684 struct sched_domain
*sd
= NULL
, *p
;
8686 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8689 if (cpumask_weight(cpu_map
) >
8690 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8691 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8692 SD_INIT(sd
, ALLNODES
);
8693 set_domain_attribute(sd
, attr
);
8694 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8695 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8701 sd
= &per_cpu(node_domains
, i
).sd
;
8703 set_domain_attribute(sd
, attr
);
8704 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8708 cpumask_and(sched_domain_span(sd
),
8709 sched_domain_span(sd
), cpu_map
);
8713 sd
= &per_cpu(phys_domains
, i
).sd
;
8715 set_domain_attribute(sd
, attr
);
8716 cpumask_copy(sched_domain_span(sd
), nodemask
);
8720 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8722 #ifdef CONFIG_SCHED_MC
8724 sd
= &per_cpu(core_domains
, i
).sd
;
8726 set_domain_attribute(sd
, attr
);
8727 cpumask_and(sched_domain_span(sd
), cpu_map
,
8728 cpu_coregroup_mask(i
));
8731 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8734 #ifdef CONFIG_SCHED_SMT
8736 sd
= &per_cpu(cpu_domains
, i
).sd
;
8737 SD_INIT(sd
, SIBLING
);
8738 set_domain_attribute(sd
, attr
);
8739 cpumask_and(sched_domain_span(sd
),
8740 topology_thread_cpumask(i
), cpu_map
);
8743 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8747 #ifdef CONFIG_SCHED_SMT
8748 /* Set up CPU (sibling) groups */
8749 for_each_cpu(i
, cpu_map
) {
8750 cpumask_and(this_sibling_map
,
8751 topology_thread_cpumask(i
), cpu_map
);
8752 if (i
!= cpumask_first(this_sibling_map
))
8755 init_sched_build_groups(this_sibling_map
, cpu_map
,
8757 send_covered
, tmpmask
);
8761 #ifdef CONFIG_SCHED_MC
8762 /* Set up multi-core groups */
8763 for_each_cpu(i
, cpu_map
) {
8764 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8765 if (i
!= cpumask_first(this_core_map
))
8768 init_sched_build_groups(this_core_map
, cpu_map
,
8770 send_covered
, tmpmask
);
8774 /* Set up physical groups */
8775 for (i
= 0; i
< nr_node_ids
; i
++) {
8776 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8777 if (cpumask_empty(nodemask
))
8780 init_sched_build_groups(nodemask
, cpu_map
,
8782 send_covered
, tmpmask
);
8786 /* Set up node groups */
8788 init_sched_build_groups(cpu_map
, cpu_map
,
8789 &cpu_to_allnodes_group
,
8790 send_covered
, tmpmask
);
8793 for (i
= 0; i
< nr_node_ids
; i
++) {
8794 /* Set up node groups */
8795 struct sched_group
*sg
, *prev
;
8798 cpumask_clear(covered
);
8799 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8800 if (cpumask_empty(nodemask
)) {
8801 sched_group_nodes
[i
] = NULL
;
8805 sched_domain_node_span(i
, domainspan
);
8806 cpumask_and(domainspan
, domainspan
, cpu_map
);
8808 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8811 printk(KERN_WARNING
"Can not alloc domain group for "
8815 sched_group_nodes
[i
] = sg
;
8816 for_each_cpu(j
, nodemask
) {
8817 struct sched_domain
*sd
;
8819 sd
= &per_cpu(node_domains
, j
).sd
;
8822 sg
->__cpu_power
= 0;
8823 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8825 cpumask_or(covered
, covered
, nodemask
);
8828 for (j
= 0; j
< nr_node_ids
; j
++) {
8829 int n
= (i
+ j
) % nr_node_ids
;
8831 cpumask_complement(notcovered
, covered
);
8832 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8833 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8834 if (cpumask_empty(tmpmask
))
8837 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8838 if (cpumask_empty(tmpmask
))
8841 sg
= kmalloc_node(sizeof(struct sched_group
) +
8846 "Can not alloc domain group for node %d\n", j
);
8849 sg
->__cpu_power
= 0;
8850 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8851 sg
->next
= prev
->next
;
8852 cpumask_or(covered
, covered
, tmpmask
);
8859 /* Calculate CPU power for physical packages and nodes */
8860 #ifdef CONFIG_SCHED_SMT
8861 for_each_cpu(i
, cpu_map
) {
8862 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8864 init_sched_groups_power(i
, sd
);
8867 #ifdef CONFIG_SCHED_MC
8868 for_each_cpu(i
, cpu_map
) {
8869 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8871 init_sched_groups_power(i
, sd
);
8875 for_each_cpu(i
, cpu_map
) {
8876 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8878 init_sched_groups_power(i
, sd
);
8882 for (i
= 0; i
< nr_node_ids
; i
++)
8883 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8886 struct sched_group
*sg
;
8888 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8890 init_numa_sched_groups_power(sg
);
8894 /* Attach the domains */
8895 for_each_cpu(i
, cpu_map
) {
8896 struct sched_domain
*sd
;
8897 #ifdef CONFIG_SCHED_SMT
8898 sd
= &per_cpu(cpu_domains
, i
).sd
;
8899 #elif defined(CONFIG_SCHED_MC)
8900 sd
= &per_cpu(core_domains
, i
).sd
;
8902 sd
= &per_cpu(phys_domains
, i
).sd
;
8904 cpu_attach_domain(sd
, rd
, i
);
8910 free_cpumask_var(tmpmask
);
8912 free_cpumask_var(send_covered
);
8914 free_cpumask_var(this_core_map
);
8915 free_this_sibling_map
:
8916 free_cpumask_var(this_sibling_map
);
8918 free_cpumask_var(nodemask
);
8921 free_cpumask_var(notcovered
);
8923 free_cpumask_var(covered
);
8925 free_cpumask_var(domainspan
);
8932 kfree(sched_group_nodes
);
8938 free_sched_groups(cpu_map
, tmpmask
);
8939 free_rootdomain(rd
);
8944 static int build_sched_domains(const struct cpumask
*cpu_map
)
8946 return __build_sched_domains(cpu_map
, NULL
);
8949 static struct cpumask
*doms_cur
; /* current sched domains */
8950 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8951 static struct sched_domain_attr
*dattr_cur
;
8952 /* attribues of custom domains in 'doms_cur' */
8955 * Special case: If a kmalloc of a doms_cur partition (array of
8956 * cpumask) fails, then fallback to a single sched domain,
8957 * as determined by the single cpumask fallback_doms.
8959 static cpumask_var_t fallback_doms
;
8962 * arch_update_cpu_topology lets virtualized architectures update the
8963 * cpu core maps. It is supposed to return 1 if the topology changed
8964 * or 0 if it stayed the same.
8966 int __attribute__((weak
)) arch_update_cpu_topology(void)
8972 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8973 * For now this just excludes isolated cpus, but could be used to
8974 * exclude other special cases in the future.
8976 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8980 arch_update_cpu_topology();
8982 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8984 doms_cur
= fallback_doms
;
8985 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8987 err
= build_sched_domains(doms_cur
);
8988 register_sched_domain_sysctl();
8993 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8994 struct cpumask
*tmpmask
)
8996 free_sched_groups(cpu_map
, tmpmask
);
9000 * Detach sched domains from a group of cpus specified in cpu_map
9001 * These cpus will now be attached to the NULL domain
9003 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9005 /* Save because hotplug lock held. */
9006 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9009 for_each_cpu(i
, cpu_map
)
9010 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9011 synchronize_sched();
9012 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9015 /* handle null as "default" */
9016 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9017 struct sched_domain_attr
*new, int idx_new
)
9019 struct sched_domain_attr tmp
;
9026 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9027 new ? (new + idx_new
) : &tmp
,
9028 sizeof(struct sched_domain_attr
));
9032 * Partition sched domains as specified by the 'ndoms_new'
9033 * cpumasks in the array doms_new[] of cpumasks. This compares
9034 * doms_new[] to the current sched domain partitioning, doms_cur[].
9035 * It destroys each deleted domain and builds each new domain.
9037 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
9038 * The masks don't intersect (don't overlap.) We should setup one
9039 * sched domain for each mask. CPUs not in any of the cpumasks will
9040 * not be load balanced. If the same cpumask appears both in the
9041 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9044 * The passed in 'doms_new' should be kmalloc'd. This routine takes
9045 * ownership of it and will kfree it when done with it. If the caller
9046 * failed the kmalloc call, then it can pass in doms_new == NULL &&
9047 * ndoms_new == 1, and partition_sched_domains() will fallback to
9048 * the single partition 'fallback_doms', it also forces the domains
9051 * If doms_new == NULL it will be replaced with cpu_online_mask.
9052 * ndoms_new == 0 is a special case for destroying existing domains,
9053 * and it will not create the default domain.
9055 * Call with hotplug lock held
9057 /* FIXME: Change to struct cpumask *doms_new[] */
9058 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
9059 struct sched_domain_attr
*dattr_new
)
9064 mutex_lock(&sched_domains_mutex
);
9066 /* always unregister in case we don't destroy any domains */
9067 unregister_sched_domain_sysctl();
9069 /* Let architecture update cpu core mappings. */
9070 new_topology
= arch_update_cpu_topology();
9072 n
= doms_new
? ndoms_new
: 0;
9074 /* Destroy deleted domains */
9075 for (i
= 0; i
< ndoms_cur
; i
++) {
9076 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9077 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
9078 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9081 /* no match - a current sched domain not in new doms_new[] */
9082 detach_destroy_domains(doms_cur
+ i
);
9087 if (doms_new
== NULL
) {
9089 doms_new
= fallback_doms
;
9090 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
9091 WARN_ON_ONCE(dattr_new
);
9094 /* Build new domains */
9095 for (i
= 0; i
< ndoms_new
; i
++) {
9096 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9097 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
9098 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9101 /* no match - add a new doms_new */
9102 __build_sched_domains(doms_new
+ i
,
9103 dattr_new
? dattr_new
+ i
: NULL
);
9108 /* Remember the new sched domains */
9109 if (doms_cur
!= fallback_doms
)
9111 kfree(dattr_cur
); /* kfree(NULL) is safe */
9112 doms_cur
= doms_new
;
9113 dattr_cur
= dattr_new
;
9114 ndoms_cur
= ndoms_new
;
9116 register_sched_domain_sysctl();
9118 mutex_unlock(&sched_domains_mutex
);
9121 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9122 static void arch_reinit_sched_domains(void)
9126 /* Destroy domains first to force the rebuild */
9127 partition_sched_domains(0, NULL
, NULL
);
9129 rebuild_sched_domains();
9133 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9135 unsigned int level
= 0;
9137 if (sscanf(buf
, "%u", &level
) != 1)
9141 * level is always be positive so don't check for
9142 * level < POWERSAVINGS_BALANCE_NONE which is 0
9143 * What happens on 0 or 1 byte write,
9144 * need to check for count as well?
9147 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9151 sched_smt_power_savings
= level
;
9153 sched_mc_power_savings
= level
;
9155 arch_reinit_sched_domains();
9160 #ifdef CONFIG_SCHED_MC
9161 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9164 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9166 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9167 const char *buf
, size_t count
)
9169 return sched_power_savings_store(buf
, count
, 0);
9171 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9172 sched_mc_power_savings_show
,
9173 sched_mc_power_savings_store
);
9176 #ifdef CONFIG_SCHED_SMT
9177 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9180 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9182 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9183 const char *buf
, size_t count
)
9185 return sched_power_savings_store(buf
, count
, 1);
9187 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9188 sched_smt_power_savings_show
,
9189 sched_smt_power_savings_store
);
9192 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9196 #ifdef CONFIG_SCHED_SMT
9198 err
= sysfs_create_file(&cls
->kset
.kobj
,
9199 &attr_sched_smt_power_savings
.attr
);
9201 #ifdef CONFIG_SCHED_MC
9202 if (!err
&& mc_capable())
9203 err
= sysfs_create_file(&cls
->kset
.kobj
,
9204 &attr_sched_mc_power_savings
.attr
);
9208 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9210 #ifndef CONFIG_CPUSETS
9212 * Add online and remove offline CPUs from the scheduler domains.
9213 * When cpusets are enabled they take over this function.
9215 static int update_sched_domains(struct notifier_block
*nfb
,
9216 unsigned long action
, void *hcpu
)
9220 case CPU_ONLINE_FROZEN
:
9222 case CPU_DEAD_FROZEN
:
9223 partition_sched_domains(1, NULL
, NULL
);
9232 static int update_runtime(struct notifier_block
*nfb
,
9233 unsigned long action
, void *hcpu
)
9235 int cpu
= (int)(long)hcpu
;
9238 case CPU_DOWN_PREPARE
:
9239 case CPU_DOWN_PREPARE_FROZEN
:
9240 disable_runtime(cpu_rq(cpu
));
9243 case CPU_DOWN_FAILED
:
9244 case CPU_DOWN_FAILED_FROZEN
:
9246 case CPU_ONLINE_FROZEN
:
9247 enable_runtime(cpu_rq(cpu
));
9255 void __init
sched_init_smp(void)
9257 cpumask_var_t non_isolated_cpus
;
9259 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9261 #if defined(CONFIG_NUMA)
9262 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9264 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9267 mutex_lock(&sched_domains_mutex
);
9268 arch_init_sched_domains(cpu_online_mask
);
9269 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9270 if (cpumask_empty(non_isolated_cpus
))
9271 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9272 mutex_unlock(&sched_domains_mutex
);
9275 #ifndef CONFIG_CPUSETS
9276 /* XXX: Theoretical race here - CPU may be hotplugged now */
9277 hotcpu_notifier(update_sched_domains
, 0);
9280 /* RT runtime code needs to handle some hotplug events */
9281 hotcpu_notifier(update_runtime
, 0);
9285 /* Move init over to a non-isolated CPU */
9286 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9288 sched_init_granularity();
9289 free_cpumask_var(non_isolated_cpus
);
9291 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9292 init_sched_rt_class();
9295 void __init
sched_init_smp(void)
9297 sched_init_granularity();
9299 #endif /* CONFIG_SMP */
9301 int in_sched_functions(unsigned long addr
)
9303 return in_lock_functions(addr
) ||
9304 (addr
>= (unsigned long)__sched_text_start
9305 && addr
< (unsigned long)__sched_text_end
);
9308 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9310 cfs_rq
->tasks_timeline
= RB_ROOT
;
9311 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9312 #ifdef CONFIG_FAIR_GROUP_SCHED
9315 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9318 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9320 struct rt_prio_array
*array
;
9323 array
= &rt_rq
->active
;
9324 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9325 INIT_LIST_HEAD(array
->queue
+ i
);
9326 __clear_bit(i
, array
->bitmap
);
9328 /* delimiter for bitsearch: */
9329 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9331 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9332 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9334 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9338 rt_rq
->rt_nr_migratory
= 0;
9339 rt_rq
->overloaded
= 0;
9340 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
9344 rt_rq
->rt_throttled
= 0;
9345 rt_rq
->rt_runtime
= 0;
9346 spin_lock_init(&rt_rq
->rt_runtime_lock
);
9348 #ifdef CONFIG_RT_GROUP_SCHED
9349 rt_rq
->rt_nr_boosted
= 0;
9354 #ifdef CONFIG_FAIR_GROUP_SCHED
9355 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9356 struct sched_entity
*se
, int cpu
, int add
,
9357 struct sched_entity
*parent
)
9359 struct rq
*rq
= cpu_rq(cpu
);
9360 tg
->cfs_rq
[cpu
] = cfs_rq
;
9361 init_cfs_rq(cfs_rq
, rq
);
9364 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9367 /* se could be NULL for init_task_group */
9372 se
->cfs_rq
= &rq
->cfs
;
9374 se
->cfs_rq
= parent
->my_q
;
9377 se
->load
.weight
= tg
->shares
;
9378 se
->load
.inv_weight
= 0;
9379 se
->parent
= parent
;
9383 #ifdef CONFIG_RT_GROUP_SCHED
9384 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9385 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9386 struct sched_rt_entity
*parent
)
9388 struct rq
*rq
= cpu_rq(cpu
);
9390 tg
->rt_rq
[cpu
] = rt_rq
;
9391 init_rt_rq(rt_rq
, rq
);
9393 rt_rq
->rt_se
= rt_se
;
9394 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9396 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9398 tg
->rt_se
[cpu
] = rt_se
;
9403 rt_se
->rt_rq
= &rq
->rt
;
9405 rt_se
->rt_rq
= parent
->my_q
;
9407 rt_se
->my_q
= rt_rq
;
9408 rt_se
->parent
= parent
;
9409 INIT_LIST_HEAD(&rt_se
->run_list
);
9413 void __init
sched_init(void)
9416 unsigned long alloc_size
= 0, ptr
;
9418 #ifdef CONFIG_FAIR_GROUP_SCHED
9419 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9421 #ifdef CONFIG_RT_GROUP_SCHED
9422 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9424 #ifdef CONFIG_USER_SCHED
9427 #ifdef CONFIG_CPUMASK_OFFSTACK
9428 alloc_size
+= num_possible_cpus() * cpumask_size();
9431 * As sched_init() is called before page_alloc is setup,
9432 * we use alloc_bootmem().
9435 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
9437 #ifdef CONFIG_FAIR_GROUP_SCHED
9438 init_task_group
.se
= (struct sched_entity
**)ptr
;
9439 ptr
+= nr_cpu_ids
* sizeof(void **);
9441 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9442 ptr
+= nr_cpu_ids
* sizeof(void **);
9444 #ifdef CONFIG_USER_SCHED
9445 root_task_group
.se
= (struct sched_entity
**)ptr
;
9446 ptr
+= nr_cpu_ids
* sizeof(void **);
9448 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9449 ptr
+= nr_cpu_ids
* sizeof(void **);
9450 #endif /* CONFIG_USER_SCHED */
9451 #endif /* CONFIG_FAIR_GROUP_SCHED */
9452 #ifdef CONFIG_RT_GROUP_SCHED
9453 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9454 ptr
+= nr_cpu_ids
* sizeof(void **);
9456 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9457 ptr
+= nr_cpu_ids
* sizeof(void **);
9459 #ifdef CONFIG_USER_SCHED
9460 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9461 ptr
+= nr_cpu_ids
* sizeof(void **);
9463 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9464 ptr
+= nr_cpu_ids
* sizeof(void **);
9465 #endif /* CONFIG_USER_SCHED */
9466 #endif /* CONFIG_RT_GROUP_SCHED */
9467 #ifdef CONFIG_CPUMASK_OFFSTACK
9468 for_each_possible_cpu(i
) {
9469 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9470 ptr
+= cpumask_size();
9472 #endif /* CONFIG_CPUMASK_OFFSTACK */
9476 init_defrootdomain();
9479 init_rt_bandwidth(&def_rt_bandwidth
,
9480 global_rt_period(), global_rt_runtime());
9482 #ifdef CONFIG_RT_GROUP_SCHED
9483 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9484 global_rt_period(), global_rt_runtime());
9485 #ifdef CONFIG_USER_SCHED
9486 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9487 global_rt_period(), RUNTIME_INF
);
9488 #endif /* CONFIG_USER_SCHED */
9489 #endif /* CONFIG_RT_GROUP_SCHED */
9491 #ifdef CONFIG_GROUP_SCHED
9492 list_add(&init_task_group
.list
, &task_groups
);
9493 INIT_LIST_HEAD(&init_task_group
.children
);
9495 #ifdef CONFIG_USER_SCHED
9496 INIT_LIST_HEAD(&root_task_group
.children
);
9497 init_task_group
.parent
= &root_task_group
;
9498 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9499 #endif /* CONFIG_USER_SCHED */
9500 #endif /* CONFIG_GROUP_SCHED */
9502 for_each_possible_cpu(i
) {
9506 spin_lock_init(&rq
->lock
);
9508 rq
->calc_load_active
= 0;
9509 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9510 init_cfs_rq(&rq
->cfs
, rq
);
9511 init_rt_rq(&rq
->rt
, rq
);
9512 #ifdef CONFIG_FAIR_GROUP_SCHED
9513 init_task_group
.shares
= init_task_group_load
;
9514 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9515 #ifdef CONFIG_CGROUP_SCHED
9517 * How much cpu bandwidth does init_task_group get?
9519 * In case of task-groups formed thr' the cgroup filesystem, it
9520 * gets 100% of the cpu resources in the system. This overall
9521 * system cpu resource is divided among the tasks of
9522 * init_task_group and its child task-groups in a fair manner,
9523 * based on each entity's (task or task-group's) weight
9524 * (se->load.weight).
9526 * In other words, if init_task_group has 10 tasks of weight
9527 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9528 * then A0's share of the cpu resource is:
9530 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9532 * We achieve this by letting init_task_group's tasks sit
9533 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9535 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9536 #elif defined CONFIG_USER_SCHED
9537 root_task_group
.shares
= NICE_0_LOAD
;
9538 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9540 * In case of task-groups formed thr' the user id of tasks,
9541 * init_task_group represents tasks belonging to root user.
9542 * Hence it forms a sibling of all subsequent groups formed.
9543 * In this case, init_task_group gets only a fraction of overall
9544 * system cpu resource, based on the weight assigned to root
9545 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9546 * by letting tasks of init_task_group sit in a separate cfs_rq
9547 * (init_cfs_rq) and having one entity represent this group of
9548 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9550 init_tg_cfs_entry(&init_task_group
,
9551 &per_cpu(init_cfs_rq
, i
),
9552 &per_cpu(init_sched_entity
, i
), i
, 1,
9553 root_task_group
.se
[i
]);
9556 #endif /* CONFIG_FAIR_GROUP_SCHED */
9558 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9559 #ifdef CONFIG_RT_GROUP_SCHED
9560 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9561 #ifdef CONFIG_CGROUP_SCHED
9562 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9563 #elif defined CONFIG_USER_SCHED
9564 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9565 init_tg_rt_entry(&init_task_group
,
9566 &per_cpu(init_rt_rq
, i
),
9567 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9568 root_task_group
.rt_se
[i
]);
9572 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9573 rq
->cpu_load
[j
] = 0;
9577 rq
->active_balance
= 0;
9578 rq
->next_balance
= jiffies
;
9582 rq
->migration_thread
= NULL
;
9583 INIT_LIST_HEAD(&rq
->migration_queue
);
9584 rq_attach_root(rq
, &def_root_domain
);
9587 atomic_set(&rq
->nr_iowait
, 0);
9590 set_load_weight(&init_task
);
9592 #ifdef CONFIG_PREEMPT_NOTIFIERS
9593 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9597 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9600 #ifdef CONFIG_RT_MUTEXES
9601 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9605 * The boot idle thread does lazy MMU switching as well:
9607 atomic_inc(&init_mm
.mm_count
);
9608 enter_lazy_tlb(&init_mm
, current
);
9610 #ifdef CONFIG_PREEMPT_RT
9611 printk("Real-Time Preemption Support (C) 2004-2007 Ingo Molnar\n");
9614 * Make us the idle thread. Technically, schedule() should not be
9615 * called from this thread, however somewhere below it might be,
9616 * but because we are the idle thread, we just pick up running again
9617 * when this runqueue becomes "idle".
9619 init_idle(current
, smp_processor_id());
9621 calc_load_update
= jiffies
+ LOAD_FREQ
;
9624 * During early bootup we pretend to be a normal task:
9626 current
->sched_class
= &fair_sched_class
;
9628 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9629 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9632 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9634 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9637 scheduler_running
= 1;
9640 #ifdef CONFIG_MAGIC_SYSRQ
9641 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9645 update_rq_clock(rq
);
9646 on_rq
= p
->se
.on_rq
;
9648 deactivate_task(rq
, p
, 0);
9649 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9651 activate_task(rq
, p
, 0);
9652 resched_task(rq
->curr
);
9656 void normalize_rt_tasks(void)
9658 struct task_struct
*g
, *p
;
9659 unsigned long flags
;
9662 read_lock_irqsave(&tasklist_lock
, flags
);
9663 do_each_thread(g
, p
) {
9665 * Only normalize user tasks:
9670 p
->se
.exec_start
= 0;
9671 #ifdef CONFIG_SCHEDSTATS
9672 p
->se
.wait_start
= 0;
9673 p
->se
.sleep_start
= 0;
9674 p
->se
.block_start
= 0;
9679 * Renice negative nice level userspace
9682 if (TASK_NICE(p
) < 0 && p
->mm
)
9683 set_user_nice(p
, 0);
9687 spin_lock(&p
->pi_lock
);
9688 rq
= __task_rq_lock(p
);
9690 normalize_task(rq
, p
);
9692 __task_rq_unlock(rq
);
9693 spin_unlock(&p
->pi_lock
);
9694 } while_each_thread(g
, p
);
9696 read_unlock_irqrestore(&tasklist_lock
, flags
);
9699 #endif /* CONFIG_MAGIC_SYSRQ */
9703 * These functions are only useful for the IA64 MCA handling.
9705 * They can only be called when the whole system has been
9706 * stopped - every CPU needs to be quiescent, and no scheduling
9707 * activity can take place. Using them for anything else would
9708 * be a serious bug, and as a result, they aren't even visible
9709 * under any other configuration.
9713 * curr_task - return the current task for a given cpu.
9714 * @cpu: the processor in question.
9716 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9718 struct task_struct
*curr_task(int cpu
)
9720 return cpu_curr(cpu
);
9724 * set_curr_task - set the current task for a given cpu.
9725 * @cpu: the processor in question.
9726 * @p: the task pointer to set.
9728 * Description: This function must only be used when non-maskable interrupts
9729 * are serviced on a separate stack. It allows the architecture to switch the
9730 * notion of the current task on a cpu in a non-blocking manner. This function
9731 * must be called with all CPU's synchronized, and interrupts disabled, the
9732 * and caller must save the original value of the current task (see
9733 * curr_task() above) and restore that value before reenabling interrupts and
9734 * re-starting the system.
9736 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9738 void set_curr_task(int cpu
, struct task_struct
*p
)
9745 #ifdef CONFIG_FAIR_GROUP_SCHED
9746 static void free_fair_sched_group(struct task_group
*tg
)
9750 for_each_possible_cpu(i
) {
9752 kfree(tg
->cfs_rq
[i
]);
9762 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9764 struct cfs_rq
*cfs_rq
;
9765 struct sched_entity
*se
;
9769 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9772 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9776 tg
->shares
= NICE_0_LOAD
;
9778 for_each_possible_cpu(i
) {
9781 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9782 GFP_KERNEL
, cpu_to_node(i
));
9786 se
= kzalloc_node(sizeof(struct sched_entity
),
9787 GFP_KERNEL
, cpu_to_node(i
));
9791 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9800 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9802 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9803 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9806 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9808 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9810 #else /* !CONFG_FAIR_GROUP_SCHED */
9811 static inline void free_fair_sched_group(struct task_group
*tg
)
9816 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9821 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9825 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9828 #endif /* CONFIG_FAIR_GROUP_SCHED */
9830 #ifdef CONFIG_RT_GROUP_SCHED
9831 static void free_rt_sched_group(struct task_group
*tg
)
9835 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9837 for_each_possible_cpu(i
) {
9839 kfree(tg
->rt_rq
[i
]);
9841 kfree(tg
->rt_se
[i
]);
9849 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9851 struct rt_rq
*rt_rq
;
9852 struct sched_rt_entity
*rt_se
;
9856 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9859 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9863 init_rt_bandwidth(&tg
->rt_bandwidth
,
9864 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9866 for_each_possible_cpu(i
) {
9869 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9870 GFP_KERNEL
, cpu_to_node(i
));
9874 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9875 GFP_KERNEL
, cpu_to_node(i
));
9879 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9888 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9890 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9891 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9894 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9896 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9898 #else /* !CONFIG_RT_GROUP_SCHED */
9899 static inline void free_rt_sched_group(struct task_group
*tg
)
9904 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9909 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9913 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9916 #endif /* CONFIG_RT_GROUP_SCHED */
9918 #ifdef CONFIG_GROUP_SCHED
9919 static void free_sched_group(struct task_group
*tg
)
9921 free_fair_sched_group(tg
);
9922 free_rt_sched_group(tg
);
9926 /* allocate runqueue etc for a new task group */
9927 struct task_group
*sched_create_group(struct task_group
*parent
)
9929 struct task_group
*tg
;
9930 unsigned long flags
;
9933 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9935 return ERR_PTR(-ENOMEM
);
9937 if (!alloc_fair_sched_group(tg
, parent
))
9940 if (!alloc_rt_sched_group(tg
, parent
))
9943 spin_lock_irqsave(&task_group_lock
, flags
);
9944 for_each_possible_cpu(i
) {
9945 register_fair_sched_group(tg
, i
);
9946 register_rt_sched_group(tg
, i
);
9948 list_add_rcu(&tg
->list
, &task_groups
);
9950 WARN_ON(!parent
); /* root should already exist */
9952 tg
->parent
= parent
;
9953 INIT_LIST_HEAD(&tg
->children
);
9954 list_add_rcu(&tg
->siblings
, &parent
->children
);
9955 spin_unlock_irqrestore(&task_group_lock
, flags
);
9960 free_sched_group(tg
);
9961 return ERR_PTR(-ENOMEM
);
9964 /* rcu callback to free various structures associated with a task group */
9965 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9967 /* now it should be safe to free those cfs_rqs */
9968 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9971 /* Destroy runqueue etc associated with a task group */
9972 void sched_destroy_group(struct task_group
*tg
)
9974 unsigned long flags
;
9977 spin_lock_irqsave(&task_group_lock
, flags
);
9978 for_each_possible_cpu(i
) {
9979 unregister_fair_sched_group(tg
, i
);
9980 unregister_rt_sched_group(tg
, i
);
9982 list_del_rcu(&tg
->list
);
9983 list_del_rcu(&tg
->siblings
);
9984 spin_unlock_irqrestore(&task_group_lock
, flags
);
9986 /* wait for possible concurrent references to cfs_rqs complete */
9987 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9990 /* change task's runqueue when it moves between groups.
9991 * The caller of this function should have put the task in its new group
9992 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9993 * reflect its new group.
9995 void sched_move_task(struct task_struct
*tsk
)
9998 unsigned long flags
;
10001 rq
= task_rq_lock(tsk
, &flags
);
10003 update_rq_clock(rq
);
10005 running
= task_current(rq
, tsk
);
10006 on_rq
= tsk
->se
.on_rq
;
10009 dequeue_task(rq
, tsk
, 0);
10010 if (unlikely(running
))
10011 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10013 set_task_rq(tsk
, task_cpu(tsk
));
10015 #ifdef CONFIG_FAIR_GROUP_SCHED
10016 if (tsk
->sched_class
->moved_group
)
10017 tsk
->sched_class
->moved_group(tsk
);
10020 if (unlikely(running
))
10021 tsk
->sched_class
->set_curr_task(rq
);
10023 enqueue_task(rq
, tsk
, 0);
10025 task_rq_unlock(rq
, &flags
);
10027 #endif /* CONFIG_GROUP_SCHED */
10029 #ifdef CONFIG_FAIR_GROUP_SCHED
10030 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10032 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10037 dequeue_entity(cfs_rq
, se
, 0);
10039 se
->load
.weight
= shares
;
10040 se
->load
.inv_weight
= 0;
10043 enqueue_entity(cfs_rq
, se
, 0);
10046 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10048 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10049 struct rq
*rq
= cfs_rq
->rq
;
10050 unsigned long flags
;
10052 spin_lock_irqsave(&rq
->lock
, flags
);
10053 __set_se_shares(se
, shares
);
10054 spin_unlock_irqrestore(&rq
->lock
, flags
);
10057 static DEFINE_MUTEX(shares_mutex
);
10059 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10062 unsigned long flags
;
10065 * We can't change the weight of the root cgroup.
10070 if (shares
< MIN_SHARES
)
10071 shares
= MIN_SHARES
;
10072 else if (shares
> MAX_SHARES
)
10073 shares
= MAX_SHARES
;
10075 mutex_lock(&shares_mutex
);
10076 if (tg
->shares
== shares
)
10079 spin_lock_irqsave(&task_group_lock
, flags
);
10080 for_each_possible_cpu(i
)
10081 unregister_fair_sched_group(tg
, i
);
10082 list_del_rcu(&tg
->siblings
);
10083 spin_unlock_irqrestore(&task_group_lock
, flags
);
10085 /* wait for any ongoing reference to this group to finish */
10086 synchronize_sched();
10089 * Now we are free to modify the group's share on each cpu
10090 * w/o tripping rebalance_share or load_balance_fair.
10092 tg
->shares
= shares
;
10093 for_each_possible_cpu(i
) {
10095 * force a rebalance
10097 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10098 set_se_shares(tg
->se
[i
], shares
);
10102 * Enable load balance activity on this group, by inserting it back on
10103 * each cpu's rq->leaf_cfs_rq_list.
10105 spin_lock_irqsave(&task_group_lock
, flags
);
10106 for_each_possible_cpu(i
)
10107 register_fair_sched_group(tg
, i
);
10108 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10109 spin_unlock_irqrestore(&task_group_lock
, flags
);
10111 mutex_unlock(&shares_mutex
);
10115 unsigned long sched_group_shares(struct task_group
*tg
)
10121 #ifdef CONFIG_RT_GROUP_SCHED
10123 * Ensure that the real time constraints are schedulable.
10125 static DEFINE_MUTEX(rt_constraints_mutex
);
10127 static unsigned long to_ratio(u64 period
, u64 runtime
)
10129 if (runtime
== RUNTIME_INF
)
10132 return div64_u64(runtime
<< 20, period
);
10135 /* Must be called with tasklist_lock held */
10136 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10138 struct task_struct
*g
, *p
;
10140 do_each_thread(g
, p
) {
10141 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10143 } while_each_thread(g
, p
);
10148 struct rt_schedulable_data
{
10149 struct task_group
*tg
;
10154 static int tg_schedulable(struct task_group
*tg
, void *data
)
10156 struct rt_schedulable_data
*d
= data
;
10157 struct task_group
*child
;
10158 unsigned long total
, sum
= 0;
10159 u64 period
, runtime
;
10161 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10162 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10165 period
= d
->rt_period
;
10166 runtime
= d
->rt_runtime
;
10169 #ifdef CONFIG_USER_SCHED
10170 if (tg
== &root_task_group
) {
10171 period
= global_rt_period();
10172 runtime
= global_rt_runtime();
10177 * Cannot have more runtime than the period.
10179 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10183 * Ensure we don't starve existing RT tasks.
10185 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10188 total
= to_ratio(period
, runtime
);
10191 * Nobody can have more than the global setting allows.
10193 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10197 * The sum of our children's runtime should not exceed our own.
10199 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10200 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10201 runtime
= child
->rt_bandwidth
.rt_runtime
;
10203 if (child
== d
->tg
) {
10204 period
= d
->rt_period
;
10205 runtime
= d
->rt_runtime
;
10208 sum
+= to_ratio(period
, runtime
);
10217 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10219 struct rt_schedulable_data data
= {
10221 .rt_period
= period
,
10222 .rt_runtime
= runtime
,
10225 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10228 static int tg_set_bandwidth(struct task_group
*tg
,
10229 u64 rt_period
, u64 rt_runtime
)
10233 mutex_lock(&rt_constraints_mutex
);
10234 read_lock(&tasklist_lock
);
10235 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10239 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10240 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10241 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10243 for_each_possible_cpu(i
) {
10244 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10246 spin_lock(&rt_rq
->rt_runtime_lock
);
10247 rt_rq
->rt_runtime
= rt_runtime
;
10248 spin_unlock(&rt_rq
->rt_runtime_lock
);
10250 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10252 read_unlock(&tasklist_lock
);
10253 mutex_unlock(&rt_constraints_mutex
);
10258 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10260 u64 rt_runtime
, rt_period
;
10262 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10263 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10264 if (rt_runtime_us
< 0)
10265 rt_runtime
= RUNTIME_INF
;
10267 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10270 long sched_group_rt_runtime(struct task_group
*tg
)
10274 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10277 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10278 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10279 return rt_runtime_us
;
10282 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10284 u64 rt_runtime
, rt_period
;
10286 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10287 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10289 if (rt_period
== 0)
10292 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10295 long sched_group_rt_period(struct task_group
*tg
)
10299 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10300 do_div(rt_period_us
, NSEC_PER_USEC
);
10301 return rt_period_us
;
10304 static int sched_rt_global_constraints(void)
10306 u64 runtime
, period
;
10309 if (sysctl_sched_rt_period
<= 0)
10312 runtime
= global_rt_runtime();
10313 period
= global_rt_period();
10316 * Sanity check on the sysctl variables.
10318 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10321 mutex_lock(&rt_constraints_mutex
);
10322 read_lock(&tasklist_lock
);
10323 ret
= __rt_schedulable(NULL
, 0, 0);
10324 read_unlock(&tasklist_lock
);
10325 mutex_unlock(&rt_constraints_mutex
);
10330 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10332 /* Don't accept realtime tasks when there is no way for them to run */
10333 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10339 #else /* !CONFIG_RT_GROUP_SCHED */
10340 static int sched_rt_global_constraints(void)
10342 unsigned long flags
;
10345 if (sysctl_sched_rt_period
<= 0)
10348 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10349 for_each_possible_cpu(i
) {
10350 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10352 spin_lock(&rt_rq
->rt_runtime_lock
);
10353 rt_rq
->rt_runtime
= global_rt_runtime();
10354 spin_unlock(&rt_rq
->rt_runtime_lock
);
10356 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10360 #endif /* CONFIG_RT_GROUP_SCHED */
10362 int sched_rt_handler(struct ctl_table
*table
, int write
,
10363 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
10367 int old_period
, old_runtime
;
10368 static DEFINE_MUTEX(mutex
);
10370 mutex_lock(&mutex
);
10371 old_period
= sysctl_sched_rt_period
;
10372 old_runtime
= sysctl_sched_rt_runtime
;
10374 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
10376 if (!ret
&& write
) {
10377 ret
= sched_rt_global_constraints();
10379 sysctl_sched_rt_period
= old_period
;
10380 sysctl_sched_rt_runtime
= old_runtime
;
10382 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10383 def_rt_bandwidth
.rt_period
=
10384 ns_to_ktime(global_rt_period());
10387 mutex_unlock(&mutex
);
10392 #ifdef CONFIG_CGROUP_SCHED
10394 /* return corresponding task_group object of a cgroup */
10395 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10397 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10398 struct task_group
, css
);
10401 static struct cgroup_subsys_state
*
10402 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10404 struct task_group
*tg
, *parent
;
10406 if (!cgrp
->parent
) {
10407 /* This is early initialization for the top cgroup */
10408 return &init_task_group
.css
;
10411 parent
= cgroup_tg(cgrp
->parent
);
10412 tg
= sched_create_group(parent
);
10414 return ERR_PTR(-ENOMEM
);
10420 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10422 struct task_group
*tg
= cgroup_tg(cgrp
);
10424 sched_destroy_group(tg
);
10428 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10429 struct task_struct
*tsk
)
10431 #ifdef CONFIG_RT_GROUP_SCHED
10432 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10435 /* We don't support RT-tasks being in separate groups */
10436 if (tsk
->sched_class
!= &fair_sched_class
)
10444 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10445 struct cgroup
*old_cont
, struct task_struct
*tsk
)
10447 sched_move_task(tsk
);
10450 #ifdef CONFIG_FAIR_GROUP_SCHED
10451 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10454 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10457 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10459 struct task_group
*tg
= cgroup_tg(cgrp
);
10461 return (u64
) tg
->shares
;
10463 #endif /* CONFIG_FAIR_GROUP_SCHED */
10465 #ifdef CONFIG_RT_GROUP_SCHED
10466 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10469 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10472 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10474 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10477 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10480 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10483 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10485 return sched_group_rt_period(cgroup_tg(cgrp
));
10487 #endif /* CONFIG_RT_GROUP_SCHED */
10489 static struct cftype cpu_files
[] = {
10490 #ifdef CONFIG_FAIR_GROUP_SCHED
10493 .read_u64
= cpu_shares_read_u64
,
10494 .write_u64
= cpu_shares_write_u64
,
10497 #ifdef CONFIG_RT_GROUP_SCHED
10499 .name
= "rt_runtime_us",
10500 .read_s64
= cpu_rt_runtime_read
,
10501 .write_s64
= cpu_rt_runtime_write
,
10504 .name
= "rt_period_us",
10505 .read_u64
= cpu_rt_period_read_uint
,
10506 .write_u64
= cpu_rt_period_write_uint
,
10511 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10513 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10516 struct cgroup_subsys cpu_cgroup_subsys
= {
10518 .create
= cpu_cgroup_create
,
10519 .destroy
= cpu_cgroup_destroy
,
10520 .can_attach
= cpu_cgroup_can_attach
,
10521 .attach
= cpu_cgroup_attach
,
10522 .populate
= cpu_cgroup_populate
,
10523 .subsys_id
= cpu_cgroup_subsys_id
,
10527 #endif /* CONFIG_CGROUP_SCHED */
10529 #ifdef CONFIG_CGROUP_CPUACCT
10532 * CPU accounting code for task groups.
10534 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10535 * (balbir@in.ibm.com).
10538 /* track cpu usage of a group of tasks and its child groups */
10540 struct cgroup_subsys_state css
;
10541 /* cpuusage holds pointer to a u64-type object on every cpu */
10543 struct cpuacct
*parent
;
10546 struct cgroup_subsys cpuacct_subsys
;
10548 /* return cpu accounting group corresponding to this container */
10549 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10551 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10552 struct cpuacct
, css
);
10555 /* return cpu accounting group to which this task belongs */
10556 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10558 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10559 struct cpuacct
, css
);
10562 /* create a new cpu accounting group */
10563 static struct cgroup_subsys_state
*cpuacct_create(
10564 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10566 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10569 return ERR_PTR(-ENOMEM
);
10571 ca
->cpuusage
= alloc_percpu(u64
);
10572 if (!ca
->cpuusage
) {
10574 return ERR_PTR(-ENOMEM
);
10578 ca
->parent
= cgroup_ca(cgrp
->parent
);
10583 /* destroy an existing cpu accounting group */
10585 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10587 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10589 free_percpu(ca
->cpuusage
);
10593 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10595 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10598 #ifndef CONFIG_64BIT
10600 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10602 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10604 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10612 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10614 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10616 #ifndef CONFIG_64BIT
10618 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10620 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10622 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10628 /* return total cpu usage (in nanoseconds) of a group */
10629 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10631 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10632 u64 totalcpuusage
= 0;
10635 for_each_present_cpu(i
)
10636 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10638 return totalcpuusage
;
10641 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10644 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10653 for_each_present_cpu(i
)
10654 cpuacct_cpuusage_write(ca
, i
, 0);
10660 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10661 struct seq_file
*m
)
10663 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10667 for_each_present_cpu(i
) {
10668 percpu
= cpuacct_cpuusage_read(ca
, i
);
10669 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10671 seq_printf(m
, "\n");
10675 static struct cftype files
[] = {
10678 .read_u64
= cpuusage_read
,
10679 .write_u64
= cpuusage_write
,
10682 .name
= "usage_percpu",
10683 .read_seq_string
= cpuacct_percpu_seq_read
,
10688 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10690 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10694 * charge this task's execution time to its accounting group.
10696 * called with rq->lock held.
10698 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10700 struct cpuacct
*ca
;
10703 if (unlikely(!cpuacct_subsys
.active
))
10706 cpu
= task_cpu(tsk
);
10712 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10713 *cpuusage
+= cputime
;
10719 struct cgroup_subsys cpuacct_subsys
= {
10721 .create
= cpuacct_create
,
10722 .destroy
= cpuacct_destroy
,
10723 .populate
= cpuacct_populate
,
10724 .subsys_id
= cpuacct_subsys_id
,
10726 #endif /* CONFIG_CGROUP_CPUACCT */