4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
121 DEFINE_TRACE(sched_wait_task
);
122 DEFINE_TRACE(sched_wakeup
);
123 DEFINE_TRACE(sched_wakeup_new
);
124 DEFINE_TRACE(sched_switch
);
125 DEFINE_TRACE(sched_migrate_task
);
129 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
132 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
133 * Since cpu_power is a 'constant', we can use a reciprocal divide.
135 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
137 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
141 * Each time a sched group cpu_power is changed,
142 * we must compute its reciprocal value
144 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
146 sg
->__cpu_power
+= val
;
147 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
151 static inline int rt_policy(int policy
)
153 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
158 static inline int task_has_rt_policy(struct task_struct
*p
)
160 return rt_policy(p
->policy
);
164 * This is the priority-queue data structure of the RT scheduling class:
166 struct rt_prio_array
{
167 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
168 struct list_head queue
[MAX_RT_PRIO
];
171 struct rt_bandwidth
{
172 /* nests inside the rq lock: */
173 spinlock_t rt_runtime_lock
;
176 struct hrtimer rt_period_timer
;
179 static struct rt_bandwidth def_rt_bandwidth
;
181 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
183 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
185 struct rt_bandwidth
*rt_b
=
186 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
192 now
= hrtimer_cb_get_time(timer
);
193 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
198 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
201 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
205 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
207 rt_b
->rt_period
= ns_to_ktime(period
);
208 rt_b
->rt_runtime
= runtime
;
210 spin_lock_init(&rt_b
->rt_runtime_lock
);
212 hrtimer_init(&rt_b
->rt_period_timer
,
213 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
214 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
217 static inline int rt_bandwidth_enabled(void)
219 return sysctl_sched_rt_runtime
>= 0;
222 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
226 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
229 if (hrtimer_active(&rt_b
->rt_period_timer
))
232 spin_lock(&rt_b
->rt_runtime_lock
);
237 if (hrtimer_active(&rt_b
->rt_period_timer
))
240 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
241 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
243 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
244 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
245 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
246 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
247 HRTIMER_MODE_ABS
, 0);
249 spin_unlock(&rt_b
->rt_runtime_lock
);
252 #ifdef CONFIG_RT_GROUP_SCHED
253 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
255 hrtimer_cancel(&rt_b
->rt_period_timer
);
260 * sched_domains_mutex serializes calls to arch_init_sched_domains,
261 * detach_destroy_domains and partition_sched_domains.
263 static DEFINE_MUTEX(sched_domains_mutex
);
265 #ifdef CONFIG_GROUP_SCHED
267 #include <linux/cgroup.h>
271 static LIST_HEAD(task_groups
);
273 /* task group related information */
275 #ifdef CONFIG_CGROUP_SCHED
276 struct cgroup_subsys_state css
;
279 #ifdef CONFIG_USER_SCHED
283 #ifdef CONFIG_FAIR_GROUP_SCHED
284 /* schedulable entities of this group on each cpu */
285 struct sched_entity
**se
;
286 /* runqueue "owned" by this group on each cpu */
287 struct cfs_rq
**cfs_rq
;
288 unsigned long shares
;
291 #ifdef CONFIG_RT_GROUP_SCHED
292 struct sched_rt_entity
**rt_se
;
293 struct rt_rq
**rt_rq
;
295 struct rt_bandwidth rt_bandwidth
;
299 struct list_head list
;
301 struct task_group
*parent
;
302 struct list_head siblings
;
303 struct list_head children
;
306 #ifdef CONFIG_USER_SCHED
308 /* Helper function to pass uid information to create_sched_user() */
309 void set_tg_uid(struct user_struct
*user
)
311 user
->tg
->uid
= user
->uid
;
316 * Every UID task group (including init_task_group aka UID-0) will
317 * be a child to this group.
319 struct task_group root_task_group
;
321 #ifdef CONFIG_FAIR_GROUP_SCHED
322 /* Default task group's sched entity on each cpu */
323 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
324 /* Default task group's cfs_rq on each cpu */
325 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
326 #endif /* CONFIG_FAIR_GROUP_SCHED */
328 #ifdef CONFIG_RT_GROUP_SCHED
329 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
330 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
331 #endif /* CONFIG_RT_GROUP_SCHED */
332 #else /* !CONFIG_USER_SCHED */
333 #define root_task_group init_task_group
334 #endif /* CONFIG_USER_SCHED */
336 /* task_group_lock serializes add/remove of task groups and also changes to
337 * a task group's cpu shares.
339 static DEFINE_SPINLOCK(task_group_lock
);
342 static int root_task_group_empty(void)
344 return list_empty(&root_task_group
.children
);
348 #ifdef CONFIG_FAIR_GROUP_SCHED
349 #ifdef CONFIG_USER_SCHED
350 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
351 #else /* !CONFIG_USER_SCHED */
352 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
353 #endif /* CONFIG_USER_SCHED */
356 * A weight of 0 or 1 can cause arithmetics problems.
357 * A weight of a cfs_rq is the sum of weights of which entities
358 * are queued on this cfs_rq, so a weight of a entity should not be
359 * too large, so as the shares value of a task group.
360 * (The default weight is 1024 - so there's no practical
361 * limitation from this.)
364 #define MAX_SHARES (1UL << 18)
366 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
369 /* Default task group.
370 * Every task in system belong to this group at bootup.
372 struct task_group init_task_group
;
374 /* return group to which a task belongs */
375 static inline struct task_group
*task_group(struct task_struct
*p
)
377 struct task_group
*tg
;
379 #ifdef CONFIG_USER_SCHED
381 tg
= __task_cred(p
)->user
->tg
;
383 #elif defined(CONFIG_CGROUP_SCHED)
384 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
385 struct task_group
, css
);
387 tg
= &init_task_group
;
392 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
393 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
395 #ifdef CONFIG_FAIR_GROUP_SCHED
396 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
397 p
->se
.parent
= task_group(p
)->se
[cpu
];
400 #ifdef CONFIG_RT_GROUP_SCHED
401 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
402 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
409 static int root_task_group_empty(void)
415 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
416 static inline struct task_group
*task_group(struct task_struct
*p
)
421 #endif /* CONFIG_GROUP_SCHED */
423 /* CFS-related fields in a runqueue */
425 struct load_weight load
;
426 unsigned long nr_running
;
431 struct rb_root tasks_timeline
;
432 struct rb_node
*rb_leftmost
;
434 struct list_head tasks
;
435 struct list_head
*balance_iterator
;
438 * 'curr' points to currently running entity on this cfs_rq.
439 * It is set to NULL otherwise (i.e when none are currently running).
441 struct sched_entity
*curr
, *next
, *last
;
443 unsigned int nr_spread_over
;
445 #ifdef CONFIG_FAIR_GROUP_SCHED
446 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
449 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
450 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
451 * (like users, containers etc.)
453 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
454 * list is used during load balance.
456 struct list_head leaf_cfs_rq_list
;
457 struct task_group
*tg
; /* group that "owns" this runqueue */
461 * the part of load.weight contributed by tasks
463 unsigned long task_weight
;
466 * h_load = weight * f(tg)
468 * Where f(tg) is the recursive weight fraction assigned to
471 unsigned long h_load
;
474 * this cpu's part of tg->shares
476 unsigned long shares
;
479 * load.weight at the time we set shares
481 unsigned long rq_weight
;
486 /* Real-Time classes' related field in a runqueue: */
488 struct rt_prio_array active
;
489 unsigned long rt_nr_running
;
490 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
492 int curr
; /* highest queued rt task prio */
494 int next
; /* next highest */
499 unsigned long rt_nr_migratory
;
500 unsigned long rt_nr_total
;
502 struct plist_head pushable_tasks
;
507 /* Nests inside the rq lock: */
508 spinlock_t rt_runtime_lock
;
510 #ifdef CONFIG_RT_GROUP_SCHED
511 unsigned long rt_nr_boosted
;
514 struct list_head leaf_rt_rq_list
;
515 struct task_group
*tg
;
516 struct sched_rt_entity
*rt_se
;
523 * We add the notion of a root-domain which will be used to define per-domain
524 * variables. Each exclusive cpuset essentially defines an island domain by
525 * fully partitioning the member cpus from any other cpuset. Whenever a new
526 * exclusive cpuset is created, we also create and attach a new root-domain
533 cpumask_var_t online
;
536 * The "RT overload" flag: it gets set if a CPU has more than
537 * one runnable RT task.
539 cpumask_var_t rto_mask
;
542 struct cpupri cpupri
;
544 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
546 * Preferred wake up cpu nominated by sched_mc balance that will be
547 * used when most cpus are idle in the system indicating overall very
548 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
550 unsigned int sched_mc_preferred_wakeup_cpu
;
555 * By default the system creates a single root-domain with all cpus as
556 * members (mimicking the global state we have today).
558 static struct root_domain def_root_domain
;
563 * This is the main, per-CPU runqueue data structure.
565 * Locking rule: those places that want to lock multiple runqueues
566 * (such as the load balancing or the thread migration code), lock
567 * acquire operations must be ordered by ascending &runqueue.
574 * nr_running and cpu_load should be in the same cacheline because
575 * remote CPUs use both these fields when doing load calculation.
577 unsigned long nr_running
;
578 #define CPU_LOAD_IDX_MAX 5
579 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
581 unsigned long last_tick_seen
;
582 unsigned char in_nohz_recently
;
584 /* capture load from *all* tasks on this cpu: */
585 struct load_weight load
;
586 unsigned long nr_load_updates
;
592 #ifdef CONFIG_FAIR_GROUP_SCHED
593 /* list of leaf cfs_rq on this cpu: */
594 struct list_head leaf_cfs_rq_list
;
596 #ifdef CONFIG_RT_GROUP_SCHED
597 struct list_head leaf_rt_rq_list
;
601 * This is part of a global counter where only the total sum
602 * over all CPUs matters. A task can increase this counter on
603 * one CPU and if it got migrated afterwards it may decrease
604 * it on another CPU. Always updated under the runqueue lock:
606 unsigned long nr_uninterruptible
;
608 struct task_struct
*curr
, *idle
;
609 unsigned long next_balance
;
610 struct mm_struct
*prev_mm
;
617 struct root_domain
*rd
;
618 struct sched_domain
*sd
;
620 unsigned char idle_at_tick
;
621 /* For active balancing */
624 /* cpu of this runqueue: */
628 unsigned long avg_load_per_task
;
630 struct task_struct
*migration_thread
;
631 struct list_head migration_queue
;
634 #ifdef CONFIG_SCHED_HRTICK
636 int hrtick_csd_pending
;
637 struct call_single_data hrtick_csd
;
639 struct hrtimer hrtick_timer
;
642 #ifdef CONFIG_SCHEDSTATS
644 struct sched_info rq_sched_info
;
645 unsigned long long rq_cpu_time
;
646 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
648 /* sys_sched_yield() stats */
649 unsigned int yld_count
;
651 /* schedule() stats */
652 unsigned int sched_switch
;
653 unsigned int sched_count
;
654 unsigned int sched_goidle
;
656 /* try_to_wake_up() stats */
657 unsigned int ttwu_count
;
658 unsigned int ttwu_local
;
661 unsigned int bkl_count
;
665 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
667 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
669 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
672 static inline int cpu_of(struct rq
*rq
)
682 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
683 * See detach_destroy_domains: synchronize_sched for details.
685 * The domain tree of any CPU may only be accessed from within
686 * preempt-disabled sections.
688 #define for_each_domain(cpu, __sd) \
689 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
691 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
692 #define this_rq() (&__get_cpu_var(runqueues))
693 #define task_rq(p) cpu_rq(task_cpu(p))
694 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
696 static inline void update_rq_clock(struct rq
*rq
)
698 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
702 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
704 #ifdef CONFIG_SCHED_DEBUG
705 # define const_debug __read_mostly
707 # define const_debug static const
713 * Returns true if the current cpu runqueue is locked.
714 * This interface allows printk to be called with the runqueue lock
715 * held and know whether or not it is OK to wake up the klogd.
717 int runqueue_is_locked(void)
720 struct rq
*rq
= cpu_rq(cpu
);
723 ret
= spin_is_locked(&rq
->lock
);
729 * Debugging: various feature bits
732 #define SCHED_FEAT(name, enabled) \
733 __SCHED_FEAT_##name ,
736 #include "sched_features.h"
741 #define SCHED_FEAT(name, enabled) \
742 (1UL << __SCHED_FEAT_##name) * enabled |
744 const_debug
unsigned int sysctl_sched_features
=
745 #include "sched_features.h"
750 #ifdef CONFIG_SCHED_DEBUG
751 #define SCHED_FEAT(name, enabled) \
754 static __read_mostly
char *sched_feat_names
[] = {
755 #include "sched_features.h"
761 static int sched_feat_show(struct seq_file
*m
, void *v
)
765 for (i
= 0; sched_feat_names
[i
]; i
++) {
766 if (!(sysctl_sched_features
& (1UL << i
)))
768 seq_printf(m
, "%s ", sched_feat_names
[i
]);
776 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
777 size_t cnt
, loff_t
*ppos
)
787 if (copy_from_user(&buf
, ubuf
, cnt
))
792 if (strncmp(buf
, "NO_", 3) == 0) {
797 for (i
= 0; sched_feat_names
[i
]; i
++) {
798 int len
= strlen(sched_feat_names
[i
]);
800 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
802 sysctl_sched_features
&= ~(1UL << i
);
804 sysctl_sched_features
|= (1UL << i
);
809 if (!sched_feat_names
[i
])
817 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
819 return single_open(filp
, sched_feat_show
, NULL
);
822 static struct file_operations sched_feat_fops
= {
823 .open
= sched_feat_open
,
824 .write
= sched_feat_write
,
827 .release
= single_release
,
830 static __init
int sched_init_debug(void)
832 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
837 late_initcall(sched_init_debug
);
841 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
844 * Number of tasks to iterate in a single balance run.
845 * Limited because this is done with IRQs disabled.
847 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
850 * ratelimit for updating the group shares.
853 unsigned int sysctl_sched_shares_ratelimit
= 250000;
856 * Inject some fuzzyness into changing the per-cpu group shares
857 * this avoids remote rq-locks at the expense of fairness.
860 unsigned int sysctl_sched_shares_thresh
= 4;
863 * period over which we measure -rt task cpu usage in us.
866 unsigned int sysctl_sched_rt_period
= 1000000;
868 static __read_mostly
int scheduler_running
;
871 * part of the period that we allow rt tasks to run in us.
874 int sysctl_sched_rt_runtime
= 950000;
876 static inline u64
global_rt_period(void)
878 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
881 static inline u64
global_rt_runtime(void)
883 if (sysctl_sched_rt_runtime
< 0)
886 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
889 #ifndef prepare_arch_switch
890 # define prepare_arch_switch(next) do { } while (0)
892 #ifndef finish_arch_switch
893 # define finish_arch_switch(prev) do { } while (0)
896 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
898 return rq
->curr
== p
;
901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
902 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
904 return task_current(rq
, p
);
907 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
911 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
913 #ifdef CONFIG_DEBUG_SPINLOCK
914 /* this is a valid case when another task releases the spinlock */
915 rq
->lock
.owner
= current
;
918 * If we are tracking spinlock dependencies then we have to
919 * fix up the runqueue lock - which gets 'carried over' from
922 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
924 spin_unlock_irq(&rq
->lock
);
927 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
928 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
933 return task_current(rq
, p
);
937 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
941 * We can optimise this out completely for !SMP, because the
942 * SMP rebalancing from interrupt is the only thing that cares
947 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
948 spin_unlock_irq(&rq
->lock
);
950 spin_unlock(&rq
->lock
);
954 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
958 * After ->oncpu is cleared, the task can be moved to a different CPU.
959 * We must ensure this doesn't happen until the switch is completely
965 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
969 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
972 * __task_rq_lock - lock the runqueue a given task resides on.
973 * Must be called interrupts disabled.
975 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
979 struct rq
*rq
= task_rq(p
);
980 spin_lock(&rq
->lock
);
981 if (likely(rq
== task_rq(p
)))
983 spin_unlock(&rq
->lock
);
988 * task_rq_lock - lock the runqueue a given task resides on and disable
989 * interrupts. Note the ordering: we can safely lookup the task_rq without
990 * explicitly disabling preemption.
992 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
998 local_irq_save(*flags
);
1000 spin_lock(&rq
->lock
);
1001 if (likely(rq
== task_rq(p
)))
1003 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1007 void task_rq_unlock_wait(struct task_struct
*p
)
1009 struct rq
*rq
= task_rq(p
);
1011 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
1012 spin_unlock_wait(&rq
->lock
);
1015 static void __task_rq_unlock(struct rq
*rq
)
1016 __releases(rq
->lock
)
1018 spin_unlock(&rq
->lock
);
1021 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1022 __releases(rq
->lock
)
1024 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1028 * this_rq_lock - lock this runqueue and disable interrupts.
1030 static struct rq
*this_rq_lock(void)
1031 __acquires(rq
->lock
)
1035 local_irq_disable();
1037 spin_lock(&rq
->lock
);
1042 #ifdef CONFIG_SCHED_HRTICK
1044 * Use HR-timers to deliver accurate preemption points.
1046 * Its all a bit involved since we cannot program an hrt while holding the
1047 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1050 * When we get rescheduled we reprogram the hrtick_timer outside of the
1056 * - enabled by features
1057 * - hrtimer is actually high res
1059 static inline int hrtick_enabled(struct rq
*rq
)
1061 if (!sched_feat(HRTICK
))
1063 if (!cpu_active(cpu_of(rq
)))
1065 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1068 static void hrtick_clear(struct rq
*rq
)
1070 if (hrtimer_active(&rq
->hrtick_timer
))
1071 hrtimer_cancel(&rq
->hrtick_timer
);
1075 * High-resolution timer tick.
1076 * Runs from hardirq context with interrupts disabled.
1078 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1080 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1082 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1084 spin_lock(&rq
->lock
);
1085 update_rq_clock(rq
);
1086 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1087 spin_unlock(&rq
->lock
);
1089 return HRTIMER_NORESTART
;
1094 * called from hardirq (IPI) context
1096 static void __hrtick_start(void *arg
)
1098 struct rq
*rq
= arg
;
1100 spin_lock(&rq
->lock
);
1101 hrtimer_restart(&rq
->hrtick_timer
);
1102 rq
->hrtick_csd_pending
= 0;
1103 spin_unlock(&rq
->lock
);
1107 * Called to set the hrtick timer state.
1109 * called with rq->lock held and irqs disabled
1111 static void hrtick_start(struct rq
*rq
, u64 delay
)
1113 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1114 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1116 hrtimer_set_expires(timer
, time
);
1118 if (rq
== this_rq()) {
1119 hrtimer_restart(timer
);
1120 } else if (!rq
->hrtick_csd_pending
) {
1121 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1122 rq
->hrtick_csd_pending
= 1;
1127 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1129 int cpu
= (int)(long)hcpu
;
1132 case CPU_UP_CANCELED
:
1133 case CPU_UP_CANCELED_FROZEN
:
1134 case CPU_DOWN_PREPARE
:
1135 case CPU_DOWN_PREPARE_FROZEN
:
1137 case CPU_DEAD_FROZEN
:
1138 hrtick_clear(cpu_rq(cpu
));
1145 static __init
void init_hrtick(void)
1147 hotcpu_notifier(hotplug_hrtick
, 0);
1151 * Called to set the hrtick timer state.
1153 * called with rq->lock held and irqs disabled
1155 static void hrtick_start(struct rq
*rq
, u64 delay
)
1157 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1158 HRTIMER_MODE_REL
, 0);
1161 static inline void init_hrtick(void)
1164 #endif /* CONFIG_SMP */
1166 static void init_rq_hrtick(struct rq
*rq
)
1169 rq
->hrtick_csd_pending
= 0;
1171 rq
->hrtick_csd
.flags
= 0;
1172 rq
->hrtick_csd
.func
= __hrtick_start
;
1173 rq
->hrtick_csd
.info
= rq
;
1176 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1177 rq
->hrtick_timer
.function
= hrtick
;
1179 #else /* CONFIG_SCHED_HRTICK */
1180 static inline void hrtick_clear(struct rq
*rq
)
1184 static inline void init_rq_hrtick(struct rq
*rq
)
1188 static inline void init_hrtick(void)
1191 #endif /* CONFIG_SCHED_HRTICK */
1194 * resched_task - mark a task 'to be rescheduled now'.
1196 * On UP this means the setting of the need_resched flag, on SMP it
1197 * might also involve a cross-CPU call to trigger the scheduler on
1202 #ifndef tsk_is_polling
1203 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1206 static void resched_task(struct task_struct
*p
)
1210 assert_spin_locked(&task_rq(p
)->lock
);
1212 if (test_tsk_need_resched(p
))
1215 set_tsk_need_resched(p
);
1218 if (cpu
== smp_processor_id())
1221 /* NEED_RESCHED must be visible before we test polling */
1223 if (!tsk_is_polling(p
))
1224 smp_send_reschedule(cpu
);
1227 static void resched_cpu(int cpu
)
1229 struct rq
*rq
= cpu_rq(cpu
);
1230 unsigned long flags
;
1232 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1234 resched_task(cpu_curr(cpu
));
1235 spin_unlock_irqrestore(&rq
->lock
, flags
);
1240 * When add_timer_on() enqueues a timer into the timer wheel of an
1241 * idle CPU then this timer might expire before the next timer event
1242 * which is scheduled to wake up that CPU. In case of a completely
1243 * idle system the next event might even be infinite time into the
1244 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1245 * leaves the inner idle loop so the newly added timer is taken into
1246 * account when the CPU goes back to idle and evaluates the timer
1247 * wheel for the next timer event.
1249 void wake_up_idle_cpu(int cpu
)
1251 struct rq
*rq
= cpu_rq(cpu
);
1253 if (cpu
== smp_processor_id())
1257 * This is safe, as this function is called with the timer
1258 * wheel base lock of (cpu) held. When the CPU is on the way
1259 * to idle and has not yet set rq->curr to idle then it will
1260 * be serialized on the timer wheel base lock and take the new
1261 * timer into account automatically.
1263 if (rq
->curr
!= rq
->idle
)
1267 * We can set TIF_RESCHED on the idle task of the other CPU
1268 * lockless. The worst case is that the other CPU runs the
1269 * idle task through an additional NOOP schedule()
1271 set_tsk_need_resched(rq
->idle
);
1273 /* NEED_RESCHED must be visible before we test polling */
1275 if (!tsk_is_polling(rq
->idle
))
1276 smp_send_reschedule(cpu
);
1278 #endif /* CONFIG_NO_HZ */
1280 #else /* !CONFIG_SMP */
1281 static void resched_task(struct task_struct
*p
)
1283 assert_spin_locked(&task_rq(p
)->lock
);
1284 set_tsk_need_resched(p
);
1286 #endif /* CONFIG_SMP */
1288 #if BITS_PER_LONG == 32
1289 # define WMULT_CONST (~0UL)
1291 # define WMULT_CONST (1UL << 32)
1294 #define WMULT_SHIFT 32
1297 * Shift right and round:
1299 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1302 * delta *= weight / lw
1304 static unsigned long
1305 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1306 struct load_weight
*lw
)
1310 if (!lw
->inv_weight
) {
1311 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1314 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1318 tmp
= (u64
)delta_exec
* weight
;
1320 * Check whether we'd overflow the 64-bit multiplication:
1322 if (unlikely(tmp
> WMULT_CONST
))
1323 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1326 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1328 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1331 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1337 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1344 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1345 * of tasks with abnormal "nice" values across CPUs the contribution that
1346 * each task makes to its run queue's load is weighted according to its
1347 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1348 * scaled version of the new time slice allocation that they receive on time
1352 #define WEIGHT_IDLEPRIO 3
1353 #define WMULT_IDLEPRIO 1431655765
1356 * Nice levels are multiplicative, with a gentle 10% change for every
1357 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1358 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1359 * that remained on nice 0.
1361 * The "10% effect" is relative and cumulative: from _any_ nice level,
1362 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1363 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1364 * If a task goes up by ~10% and another task goes down by ~10% then
1365 * the relative distance between them is ~25%.)
1367 static const int prio_to_weight
[40] = {
1368 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1369 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1370 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1371 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1372 /* 0 */ 1024, 820, 655, 526, 423,
1373 /* 5 */ 335, 272, 215, 172, 137,
1374 /* 10 */ 110, 87, 70, 56, 45,
1375 /* 15 */ 36, 29, 23, 18, 15,
1379 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1381 * In cases where the weight does not change often, we can use the
1382 * precalculated inverse to speed up arithmetics by turning divisions
1383 * into multiplications:
1385 static const u32 prio_to_wmult
[40] = {
1386 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1387 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1388 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1389 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1390 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1391 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1392 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1393 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1396 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1399 * runqueue iterator, to support SMP load-balancing between different
1400 * scheduling classes, without having to expose their internal data
1401 * structures to the load-balancing proper:
1403 struct rq_iterator
{
1405 struct task_struct
*(*start
)(void *);
1406 struct task_struct
*(*next
)(void *);
1410 static unsigned long
1411 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1412 unsigned long max_load_move
, struct sched_domain
*sd
,
1413 enum cpu_idle_type idle
, int *all_pinned
,
1414 int *this_best_prio
, struct rq_iterator
*iterator
);
1417 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1418 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1419 struct rq_iterator
*iterator
);
1422 /* Time spent by the tasks of the cpu accounting group executing in ... */
1423 enum cpuacct_stat_index
{
1424 CPUACCT_STAT_USER
, /* ... user mode */
1425 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1427 CPUACCT_STAT_NSTATS
,
1430 #ifdef CONFIG_CGROUP_CPUACCT
1431 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1432 static void cpuacct_update_stats(struct task_struct
*tsk
,
1433 enum cpuacct_stat_index idx
, cputime_t val
);
1435 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1436 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1437 enum cpuacct_stat_index idx
, cputime_t val
) {}
1440 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1442 update_load_add(&rq
->load
, load
);
1445 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1447 update_load_sub(&rq
->load
, load
);
1450 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1451 typedef int (*tg_visitor
)(struct task_group
*, void *);
1454 * Iterate the full tree, calling @down when first entering a node and @up when
1455 * leaving it for the final time.
1457 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1459 struct task_group
*parent
, *child
;
1463 parent
= &root_task_group
;
1465 ret
= (*down
)(parent
, data
);
1468 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1475 ret
= (*up
)(parent
, data
);
1480 parent
= parent
->parent
;
1489 static int tg_nop(struct task_group
*tg
, void *data
)
1496 static unsigned long source_load(int cpu
, int type
);
1497 static unsigned long target_load(int cpu
, int type
);
1498 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1500 static unsigned long cpu_avg_load_per_task(int cpu
)
1502 struct rq
*rq
= cpu_rq(cpu
);
1503 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1506 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1508 rq
->avg_load_per_task
= 0;
1510 return rq
->avg_load_per_task
;
1513 #ifdef CONFIG_FAIR_GROUP_SCHED
1515 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1518 * Calculate and set the cpu's group shares.
1521 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1522 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1524 unsigned long shares
;
1525 unsigned long rq_weight
;
1530 rq_weight
= tg
->cfs_rq
[cpu
]->rq_weight
;
1533 * \Sum shares * rq_weight
1534 * shares = -----------------------
1538 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1539 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1541 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1542 sysctl_sched_shares_thresh
) {
1543 struct rq
*rq
= cpu_rq(cpu
);
1544 unsigned long flags
;
1546 spin_lock_irqsave(&rq
->lock
, flags
);
1547 tg
->cfs_rq
[cpu
]->shares
= shares
;
1549 __set_se_shares(tg
->se
[cpu
], shares
);
1550 spin_unlock_irqrestore(&rq
->lock
, flags
);
1555 * Re-compute the task group their per cpu shares over the given domain.
1556 * This needs to be done in a bottom-up fashion because the rq weight of a
1557 * parent group depends on the shares of its child groups.
1559 static int tg_shares_up(struct task_group
*tg
, void *data
)
1561 unsigned long weight
, rq_weight
= 0;
1562 unsigned long shares
= 0;
1563 struct sched_domain
*sd
= data
;
1566 for_each_cpu(i
, sched_domain_span(sd
)) {
1568 * If there are currently no tasks on the cpu pretend there
1569 * is one of average load so that when a new task gets to
1570 * run here it will not get delayed by group starvation.
1572 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1574 weight
= NICE_0_LOAD
;
1576 tg
->cfs_rq
[i
]->rq_weight
= weight
;
1577 rq_weight
+= weight
;
1578 shares
+= tg
->cfs_rq
[i
]->shares
;
1581 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1582 shares
= tg
->shares
;
1584 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1585 shares
= tg
->shares
;
1587 for_each_cpu(i
, sched_domain_span(sd
))
1588 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1594 * Compute the cpu's hierarchical load factor for each task group.
1595 * This needs to be done in a top-down fashion because the load of a child
1596 * group is a fraction of its parents load.
1598 static int tg_load_down(struct task_group
*tg
, void *data
)
1601 long cpu
= (long)data
;
1604 load
= cpu_rq(cpu
)->load
.weight
;
1606 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1607 load
*= tg
->cfs_rq
[cpu
]->shares
;
1608 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1611 tg
->cfs_rq
[cpu
]->h_load
= load
;
1616 static void update_shares(struct sched_domain
*sd
)
1618 u64 now
= cpu_clock(raw_smp_processor_id());
1619 s64 elapsed
= now
- sd
->last_update
;
1621 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1622 sd
->last_update
= now
;
1623 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1627 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1629 spin_unlock(&rq
->lock
);
1631 spin_lock(&rq
->lock
);
1634 static void update_h_load(long cpu
)
1636 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1641 static inline void update_shares(struct sched_domain
*sd
)
1645 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1651 #ifdef CONFIG_PREEMPT
1654 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1655 * way at the expense of forcing extra atomic operations in all
1656 * invocations. This assures that the double_lock is acquired using the
1657 * same underlying policy as the spinlock_t on this architecture, which
1658 * reduces latency compared to the unfair variant below. However, it
1659 * also adds more overhead and therefore may reduce throughput.
1661 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1662 __releases(this_rq
->lock
)
1663 __acquires(busiest
->lock
)
1664 __acquires(this_rq
->lock
)
1666 spin_unlock(&this_rq
->lock
);
1667 double_rq_lock(this_rq
, busiest
);
1674 * Unfair double_lock_balance: Optimizes throughput at the expense of
1675 * latency by eliminating extra atomic operations when the locks are
1676 * already in proper order on entry. This favors lower cpu-ids and will
1677 * grant the double lock to lower cpus over higher ids under contention,
1678 * regardless of entry order into the function.
1680 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1681 __releases(this_rq
->lock
)
1682 __acquires(busiest
->lock
)
1683 __acquires(this_rq
->lock
)
1687 if (unlikely(!spin_trylock(&busiest
->lock
))) {
1688 if (busiest
< this_rq
) {
1689 spin_unlock(&this_rq
->lock
);
1690 spin_lock(&busiest
->lock
);
1691 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
1694 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
1699 #endif /* CONFIG_PREEMPT */
1702 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1704 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1706 if (unlikely(!irqs_disabled())) {
1707 /* printk() doesn't work good under rq->lock */
1708 spin_unlock(&this_rq
->lock
);
1712 return _double_lock_balance(this_rq
, busiest
);
1715 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1716 __releases(busiest
->lock
)
1718 spin_unlock(&busiest
->lock
);
1719 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1723 #ifdef CONFIG_FAIR_GROUP_SCHED
1724 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1727 cfs_rq
->shares
= shares
;
1732 #include "sched_stats.h"
1733 #include "sched_idletask.c"
1734 #include "sched_fair.c"
1735 #include "sched_rt.c"
1736 #ifdef CONFIG_SCHED_DEBUG
1737 # include "sched_debug.c"
1740 #define sched_class_highest (&rt_sched_class)
1741 #define for_each_class(class) \
1742 for (class = sched_class_highest; class; class = class->next)
1744 static void inc_nr_running(struct rq
*rq
)
1749 static void dec_nr_running(struct rq
*rq
)
1754 static void set_load_weight(struct task_struct
*p
)
1756 if (task_has_rt_policy(p
)) {
1757 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1758 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1763 * SCHED_IDLE tasks get minimal weight:
1765 if (p
->policy
== SCHED_IDLE
) {
1766 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1767 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1771 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1772 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1775 static void update_avg(u64
*avg
, u64 sample
)
1777 s64 diff
= sample
- *avg
;
1781 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1784 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1786 sched_info_queued(p
);
1787 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1791 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1794 if (p
->se
.last_wakeup
) {
1795 update_avg(&p
->se
.avg_overlap
,
1796 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1797 p
->se
.last_wakeup
= 0;
1799 update_avg(&p
->se
.avg_wakeup
,
1800 sysctl_sched_wakeup_granularity
);
1804 sched_info_dequeued(p
);
1805 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1810 * __normal_prio - return the priority that is based on the static prio
1812 static inline int __normal_prio(struct task_struct
*p
)
1814 return p
->static_prio
;
1818 * Calculate the expected normal priority: i.e. priority
1819 * without taking RT-inheritance into account. Might be
1820 * boosted by interactivity modifiers. Changes upon fork,
1821 * setprio syscalls, and whenever the interactivity
1822 * estimator recalculates.
1824 static inline int normal_prio(struct task_struct
*p
)
1828 if (task_has_rt_policy(p
))
1829 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1831 prio
= __normal_prio(p
);
1836 * Calculate the current priority, i.e. the priority
1837 * taken into account by the scheduler. This value might
1838 * be boosted by RT tasks, or might be boosted by
1839 * interactivity modifiers. Will be RT if the task got
1840 * RT-boosted. If not then it returns p->normal_prio.
1842 static int effective_prio(struct task_struct
*p
)
1844 p
->normal_prio
= normal_prio(p
);
1846 * If we are RT tasks or we were boosted to RT priority,
1847 * keep the priority unchanged. Otherwise, update priority
1848 * to the normal priority:
1850 if (!rt_prio(p
->prio
))
1851 return p
->normal_prio
;
1856 * activate_task - move a task to the runqueue.
1858 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1860 if (task_contributes_to_load(p
))
1861 rq
->nr_uninterruptible
--;
1863 enqueue_task(rq
, p
, wakeup
);
1868 * deactivate_task - remove a task from the runqueue.
1870 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1872 if (task_contributes_to_load(p
))
1873 rq
->nr_uninterruptible
++;
1875 dequeue_task(rq
, p
, sleep
);
1880 * task_curr - is this task currently executing on a CPU?
1881 * @p: the task in question.
1883 inline int task_curr(const struct task_struct
*p
)
1885 return cpu_curr(task_cpu(p
)) == p
;
1888 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1890 set_task_rq(p
, cpu
);
1893 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1894 * successfuly executed on another CPU. We must ensure that updates of
1895 * per-task data have been completed by this moment.
1898 task_thread_info(p
)->cpu
= cpu
;
1902 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1903 const struct sched_class
*prev_class
,
1904 int oldprio
, int running
)
1906 if (prev_class
!= p
->sched_class
) {
1907 if (prev_class
->switched_from
)
1908 prev_class
->switched_from(rq
, p
, running
);
1909 p
->sched_class
->switched_to(rq
, p
, running
);
1911 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1916 /* Used instead of source_load when we know the type == 0 */
1917 static unsigned long weighted_cpuload(const int cpu
)
1919 return cpu_rq(cpu
)->load
.weight
;
1923 * Is this task likely cache-hot:
1926 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1931 * Buddy candidates are cache hot:
1933 if (sched_feat(CACHE_HOT_BUDDY
) &&
1934 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1935 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1938 if (p
->sched_class
!= &fair_sched_class
)
1941 if (sysctl_sched_migration_cost
== -1)
1943 if (sysctl_sched_migration_cost
== 0)
1946 delta
= now
- p
->se
.exec_start
;
1948 return delta
< (s64
)sysctl_sched_migration_cost
;
1952 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1954 int old_cpu
= task_cpu(p
);
1955 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1956 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1957 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1960 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1962 trace_sched_migrate_task(p
, task_cpu(p
), new_cpu
);
1964 #ifdef CONFIG_SCHEDSTATS
1965 if (p
->se
.wait_start
)
1966 p
->se
.wait_start
-= clock_offset
;
1967 if (p
->se
.sleep_start
)
1968 p
->se
.sleep_start
-= clock_offset
;
1969 if (p
->se
.block_start
)
1970 p
->se
.block_start
-= clock_offset
;
1971 if (old_cpu
!= new_cpu
) {
1972 schedstat_inc(p
, se
.nr_migrations
);
1973 if (task_hot(p
, old_rq
->clock
, NULL
))
1974 schedstat_inc(p
, se
.nr_forced2_migrations
);
1977 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1978 new_cfsrq
->min_vruntime
;
1980 __set_task_cpu(p
, new_cpu
);
1983 struct migration_req
{
1984 struct list_head list
;
1986 struct task_struct
*task
;
1989 struct completion done
;
1993 * The task's runqueue lock must be held.
1994 * Returns true if you have to wait for migration thread.
1997 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1999 struct rq
*rq
= task_rq(p
);
2002 * If the task is not on a runqueue (and not running), then
2003 * it is sufficient to simply update the task's cpu field.
2005 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2006 set_task_cpu(p
, dest_cpu
);
2010 init_completion(&req
->done
);
2012 req
->dest_cpu
= dest_cpu
;
2013 list_add(&req
->list
, &rq
->migration_queue
);
2019 * wait_task_inactive - wait for a thread to unschedule.
2021 * If @match_state is nonzero, it's the @p->state value just checked and
2022 * not expected to change. If it changes, i.e. @p might have woken up,
2023 * then return zero. When we succeed in waiting for @p to be off its CPU,
2024 * we return a positive number (its total switch count). If a second call
2025 * a short while later returns the same number, the caller can be sure that
2026 * @p has remained unscheduled the whole time.
2028 * The caller must ensure that the task *will* unschedule sometime soon,
2029 * else this function might spin for a *long* time. This function can't
2030 * be called with interrupts off, or it may introduce deadlock with
2031 * smp_call_function() if an IPI is sent by the same process we are
2032 * waiting to become inactive.
2034 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2036 unsigned long flags
;
2043 * We do the initial early heuristics without holding
2044 * any task-queue locks at all. We'll only try to get
2045 * the runqueue lock when things look like they will
2051 * If the task is actively running on another CPU
2052 * still, just relax and busy-wait without holding
2055 * NOTE! Since we don't hold any locks, it's not
2056 * even sure that "rq" stays as the right runqueue!
2057 * But we don't care, since "task_running()" will
2058 * return false if the runqueue has changed and p
2059 * is actually now running somewhere else!
2061 while (task_running(rq
, p
)) {
2062 if (match_state
&& unlikely(p
->state
!= match_state
))
2068 * Ok, time to look more closely! We need the rq
2069 * lock now, to be *sure*. If we're wrong, we'll
2070 * just go back and repeat.
2072 rq
= task_rq_lock(p
, &flags
);
2073 trace_sched_wait_task(rq
, p
);
2074 running
= task_running(rq
, p
);
2075 on_rq
= p
->se
.on_rq
;
2077 if (!match_state
|| p
->state
== match_state
)
2078 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2079 task_rq_unlock(rq
, &flags
);
2082 * If it changed from the expected state, bail out now.
2084 if (unlikely(!ncsw
))
2088 * Was it really running after all now that we
2089 * checked with the proper locks actually held?
2091 * Oops. Go back and try again..
2093 if (unlikely(running
)) {
2099 * It's not enough that it's not actively running,
2100 * it must be off the runqueue _entirely_, and not
2103 * So if it was still runnable (but just not actively
2104 * running right now), it's preempted, and we should
2105 * yield - it could be a while.
2107 if (unlikely(on_rq
)) {
2108 schedule_timeout_uninterruptible(1);
2113 * Ahh, all good. It wasn't running, and it wasn't
2114 * runnable, which means that it will never become
2115 * running in the future either. We're all done!
2124 * kick_process - kick a running thread to enter/exit the kernel
2125 * @p: the to-be-kicked thread
2127 * Cause a process which is running on another CPU to enter
2128 * kernel-mode, without any delay. (to get signals handled.)
2130 * NOTE: this function doesnt have to take the runqueue lock,
2131 * because all it wants to ensure is that the remote task enters
2132 * the kernel. If the IPI races and the task has been migrated
2133 * to another CPU then no harm is done and the purpose has been
2136 void kick_process(struct task_struct
*p
)
2142 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2143 smp_send_reschedule(cpu
);
2148 * Return a low guess at the load of a migration-source cpu weighted
2149 * according to the scheduling class and "nice" value.
2151 * We want to under-estimate the load of migration sources, to
2152 * balance conservatively.
2154 static unsigned long source_load(int cpu
, int type
)
2156 struct rq
*rq
= cpu_rq(cpu
);
2157 unsigned long total
= weighted_cpuload(cpu
);
2159 if (type
== 0 || !sched_feat(LB_BIAS
))
2162 return min(rq
->cpu_load
[type
-1], total
);
2166 * Return a high guess at the load of a migration-target cpu weighted
2167 * according to the scheduling class and "nice" value.
2169 static unsigned long target_load(int cpu
, int type
)
2171 struct rq
*rq
= cpu_rq(cpu
);
2172 unsigned long total
= weighted_cpuload(cpu
);
2174 if (type
== 0 || !sched_feat(LB_BIAS
))
2177 return max(rq
->cpu_load
[type
-1], total
);
2181 * find_idlest_group finds and returns the least busy CPU group within the
2184 static struct sched_group
*
2185 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2187 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2188 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2189 int load_idx
= sd
->forkexec_idx
;
2190 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2193 unsigned long load
, avg_load
;
2197 /* Skip over this group if it has no CPUs allowed */
2198 if (!cpumask_intersects(sched_group_cpus(group
),
2202 local_group
= cpumask_test_cpu(this_cpu
,
2203 sched_group_cpus(group
));
2205 /* Tally up the load of all CPUs in the group */
2208 for_each_cpu(i
, sched_group_cpus(group
)) {
2209 /* Bias balancing toward cpus of our domain */
2211 load
= source_load(i
, load_idx
);
2213 load
= target_load(i
, load_idx
);
2218 /* Adjust by relative CPU power of the group */
2219 avg_load
= sg_div_cpu_power(group
,
2220 avg_load
* SCHED_LOAD_SCALE
);
2223 this_load
= avg_load
;
2225 } else if (avg_load
< min_load
) {
2226 min_load
= avg_load
;
2229 } while (group
= group
->next
, group
!= sd
->groups
);
2231 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2237 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2240 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2242 unsigned long load
, min_load
= ULONG_MAX
;
2246 /* Traverse only the allowed CPUs */
2247 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2248 load
= weighted_cpuload(i
);
2250 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2260 * sched_balance_self: balance the current task (running on cpu) in domains
2261 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2264 * Balance, ie. select the least loaded group.
2266 * Returns the target CPU number, or the same CPU if no balancing is needed.
2268 * preempt must be disabled.
2270 static int sched_balance_self(int cpu
, int flag
)
2272 struct task_struct
*t
= current
;
2273 struct sched_domain
*tmp
, *sd
= NULL
;
2275 for_each_domain(cpu
, tmp
) {
2277 * If power savings logic is enabled for a domain, stop there.
2279 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2281 if (tmp
->flags
& flag
)
2289 struct sched_group
*group
;
2290 int new_cpu
, weight
;
2292 if (!(sd
->flags
& flag
)) {
2297 group
= find_idlest_group(sd
, t
, cpu
);
2303 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2304 if (new_cpu
== -1 || new_cpu
== cpu
) {
2305 /* Now try balancing at a lower domain level of cpu */
2310 /* Now try balancing at a lower domain level of new_cpu */
2312 weight
= cpumask_weight(sched_domain_span(sd
));
2314 for_each_domain(cpu
, tmp
) {
2315 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2317 if (tmp
->flags
& flag
)
2320 /* while loop will break here if sd == NULL */
2326 #endif /* CONFIG_SMP */
2329 * try_to_wake_up - wake up a thread
2330 * @p: the to-be-woken-up thread
2331 * @state: the mask of task states that can be woken
2332 * @sync: do a synchronous wakeup?
2334 * Put it on the run-queue if it's not already there. The "current"
2335 * thread is always on the run-queue (except when the actual
2336 * re-schedule is in progress), and as such you're allowed to do
2337 * the simpler "current->state = TASK_RUNNING" to mark yourself
2338 * runnable without the overhead of this.
2340 * returns failure only if the task is already active.
2342 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2344 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2345 unsigned long flags
;
2349 if (!sched_feat(SYNC_WAKEUPS
))
2353 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2354 struct sched_domain
*sd
;
2356 this_cpu
= raw_smp_processor_id();
2359 for_each_domain(this_cpu
, sd
) {
2360 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2369 rq
= task_rq_lock(p
, &flags
);
2370 update_rq_clock(rq
);
2371 old_state
= p
->state
;
2372 if (!(old_state
& state
))
2380 this_cpu
= smp_processor_id();
2383 if (unlikely(task_running(rq
, p
)))
2386 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2387 if (cpu
!= orig_cpu
) {
2388 set_task_cpu(p
, cpu
);
2389 task_rq_unlock(rq
, &flags
);
2390 /* might preempt at this point */
2391 rq
= task_rq_lock(p
, &flags
);
2392 old_state
= p
->state
;
2393 if (!(old_state
& state
))
2398 this_cpu
= smp_processor_id();
2402 #ifdef CONFIG_SCHEDSTATS
2403 schedstat_inc(rq
, ttwu_count
);
2404 if (cpu
== this_cpu
)
2405 schedstat_inc(rq
, ttwu_local
);
2407 struct sched_domain
*sd
;
2408 for_each_domain(this_cpu
, sd
) {
2409 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2410 schedstat_inc(sd
, ttwu_wake_remote
);
2415 #endif /* CONFIG_SCHEDSTATS */
2418 #endif /* CONFIG_SMP */
2419 schedstat_inc(p
, se
.nr_wakeups
);
2421 schedstat_inc(p
, se
.nr_wakeups_sync
);
2422 if (orig_cpu
!= cpu
)
2423 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2424 if (cpu
== this_cpu
)
2425 schedstat_inc(p
, se
.nr_wakeups_local
);
2427 schedstat_inc(p
, se
.nr_wakeups_remote
);
2428 activate_task(rq
, p
, 1);
2432 * Only attribute actual wakeups done by this task.
2434 if (!in_interrupt()) {
2435 struct sched_entity
*se
= ¤t
->se
;
2436 u64 sample
= se
->sum_exec_runtime
;
2438 if (se
->last_wakeup
)
2439 sample
-= se
->last_wakeup
;
2441 sample
-= se
->start_runtime
;
2442 update_avg(&se
->avg_wakeup
, sample
);
2444 se
->last_wakeup
= se
->sum_exec_runtime
;
2448 trace_sched_wakeup(rq
, p
, success
);
2449 check_preempt_curr(rq
, p
, sync
);
2451 p
->state
= TASK_RUNNING
;
2453 if (p
->sched_class
->task_wake_up
)
2454 p
->sched_class
->task_wake_up(rq
, p
);
2457 task_rq_unlock(rq
, &flags
);
2462 int wake_up_process(struct task_struct
*p
)
2464 return try_to_wake_up(p
, TASK_ALL
, 0);
2466 EXPORT_SYMBOL(wake_up_process
);
2468 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2470 return try_to_wake_up(p
, state
, 0);
2474 * Perform scheduler related setup for a newly forked process p.
2475 * p is forked by current.
2477 * __sched_fork() is basic setup used by init_idle() too:
2479 static void __sched_fork(struct task_struct
*p
)
2481 p
->se
.exec_start
= 0;
2482 p
->se
.sum_exec_runtime
= 0;
2483 p
->se
.prev_sum_exec_runtime
= 0;
2484 p
->se
.last_wakeup
= 0;
2485 p
->se
.avg_overlap
= 0;
2486 p
->se
.start_runtime
= 0;
2487 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2489 #ifdef CONFIG_SCHEDSTATS
2490 p
->se
.wait_start
= 0;
2491 p
->se
.sum_sleep_runtime
= 0;
2492 p
->se
.sleep_start
= 0;
2493 p
->se
.block_start
= 0;
2494 p
->se
.sleep_max
= 0;
2495 p
->se
.block_max
= 0;
2497 p
->se
.slice_max
= 0;
2501 INIT_LIST_HEAD(&p
->rt
.run_list
);
2503 INIT_LIST_HEAD(&p
->se
.group_node
);
2505 #ifdef CONFIG_PREEMPT_NOTIFIERS
2506 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2510 * We mark the process as running here, but have not actually
2511 * inserted it onto the runqueue yet. This guarantees that
2512 * nobody will actually run it, and a signal or other external
2513 * event cannot wake it up and insert it on the runqueue either.
2515 p
->state
= TASK_RUNNING
;
2519 * fork()/clone()-time setup:
2521 void sched_fork(struct task_struct
*p
, int clone_flags
)
2523 int cpu
= get_cpu();
2528 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2530 set_task_cpu(p
, cpu
);
2533 * Make sure we do not leak PI boosting priority to the child:
2535 p
->prio
= current
->normal_prio
;
2536 if (!rt_prio(p
->prio
))
2537 p
->sched_class
= &fair_sched_class
;
2539 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2540 if (likely(sched_info_on()))
2541 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2543 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2546 #ifdef CONFIG_PREEMPT
2547 /* Want to start with kernel preemption disabled. */
2548 task_thread_info(p
)->preempt_count
= 1;
2550 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2556 * wake_up_new_task - wake up a newly created task for the first time.
2558 * This function will do some initial scheduler statistics housekeeping
2559 * that must be done for every newly created context, then puts the task
2560 * on the runqueue and wakes it.
2562 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2564 unsigned long flags
;
2567 rq
= task_rq_lock(p
, &flags
);
2568 BUG_ON(p
->state
!= TASK_RUNNING
);
2569 update_rq_clock(rq
);
2571 p
->prio
= effective_prio(p
);
2573 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2574 activate_task(rq
, p
, 0);
2577 * Let the scheduling class do new task startup
2578 * management (if any):
2580 p
->sched_class
->task_new(rq
, p
);
2583 trace_sched_wakeup_new(rq
, p
, 1);
2584 check_preempt_curr(rq
, p
, 0);
2586 if (p
->sched_class
->task_wake_up
)
2587 p
->sched_class
->task_wake_up(rq
, p
);
2589 task_rq_unlock(rq
, &flags
);
2592 #ifdef CONFIG_PREEMPT_NOTIFIERS
2595 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2596 * @notifier: notifier struct to register
2598 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2600 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2602 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2605 * preempt_notifier_unregister - no longer interested in preemption notifications
2606 * @notifier: notifier struct to unregister
2608 * This is safe to call from within a preemption notifier.
2610 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2612 hlist_del(¬ifier
->link
);
2614 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2616 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2618 struct preempt_notifier
*notifier
;
2619 struct hlist_node
*node
;
2621 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2622 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2626 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2627 struct task_struct
*next
)
2629 struct preempt_notifier
*notifier
;
2630 struct hlist_node
*node
;
2632 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2633 notifier
->ops
->sched_out(notifier
, next
);
2636 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2638 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2643 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2644 struct task_struct
*next
)
2648 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2651 * prepare_task_switch - prepare to switch tasks
2652 * @rq: the runqueue preparing to switch
2653 * @prev: the current task that is being switched out
2654 * @next: the task we are going to switch to.
2656 * This is called with the rq lock held and interrupts off. It must
2657 * be paired with a subsequent finish_task_switch after the context
2660 * prepare_task_switch sets up locking and calls architecture specific
2664 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2665 struct task_struct
*next
)
2667 fire_sched_out_preempt_notifiers(prev
, next
);
2668 prepare_lock_switch(rq
, next
);
2669 prepare_arch_switch(next
);
2673 * finish_task_switch - clean up after a task-switch
2674 * @rq: runqueue associated with task-switch
2675 * @prev: the thread we just switched away from.
2677 * finish_task_switch must be called after the context switch, paired
2678 * with a prepare_task_switch call before the context switch.
2679 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2680 * and do any other architecture-specific cleanup actions.
2682 * Note that we may have delayed dropping an mm in context_switch(). If
2683 * so, we finish that here outside of the runqueue lock. (Doing it
2684 * with the lock held can cause deadlocks; see schedule() for
2687 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2688 __releases(rq
->lock
)
2690 struct mm_struct
*mm
= rq
->prev_mm
;
2693 int post_schedule
= 0;
2695 if (current
->sched_class
->needs_post_schedule
)
2696 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2702 * A task struct has one reference for the use as "current".
2703 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2704 * schedule one last time. The schedule call will never return, and
2705 * the scheduled task must drop that reference.
2706 * The test for TASK_DEAD must occur while the runqueue locks are
2707 * still held, otherwise prev could be scheduled on another cpu, die
2708 * there before we look at prev->state, and then the reference would
2710 * Manfred Spraul <manfred@colorfullife.com>
2712 prev_state
= prev
->state
;
2713 finish_arch_switch(prev
);
2714 finish_lock_switch(rq
, prev
);
2717 current
->sched_class
->post_schedule(rq
);
2720 fire_sched_in_preempt_notifiers(current
);
2723 if (unlikely(prev_state
== TASK_DEAD
)) {
2725 * Remove function-return probe instances associated with this
2726 * task and put them back on the free list.
2728 kprobe_flush_task(prev
);
2729 put_task_struct(prev
);
2734 * schedule_tail - first thing a freshly forked thread must call.
2735 * @prev: the thread we just switched away from.
2737 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2738 __releases(rq
->lock
)
2740 struct rq
*rq
= this_rq();
2742 finish_task_switch(rq
, prev
);
2743 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2744 /* In this case, finish_task_switch does not reenable preemption */
2747 if (current
->set_child_tid
)
2748 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2752 * context_switch - switch to the new MM and the new
2753 * thread's register state.
2756 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2757 struct task_struct
*next
)
2759 struct mm_struct
*mm
, *oldmm
;
2761 prepare_task_switch(rq
, prev
, next
);
2762 trace_sched_switch(rq
, prev
, next
);
2764 oldmm
= prev
->active_mm
;
2766 * For paravirt, this is coupled with an exit in switch_to to
2767 * combine the page table reload and the switch backend into
2770 arch_enter_lazy_cpu_mode();
2772 if (unlikely(!mm
)) {
2773 next
->active_mm
= oldmm
;
2774 atomic_inc(&oldmm
->mm_count
);
2775 enter_lazy_tlb(oldmm
, next
);
2777 switch_mm(oldmm
, mm
, next
);
2779 if (unlikely(!prev
->mm
)) {
2780 prev
->active_mm
= NULL
;
2781 rq
->prev_mm
= oldmm
;
2784 * Since the runqueue lock will be released by the next
2785 * task (which is an invalid locking op but in the case
2786 * of the scheduler it's an obvious special-case), so we
2787 * do an early lockdep release here:
2789 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2790 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2793 /* Here we just switch the register state and the stack. */
2794 switch_to(prev
, next
, prev
);
2798 * this_rq must be evaluated again because prev may have moved
2799 * CPUs since it called schedule(), thus the 'rq' on its stack
2800 * frame will be invalid.
2802 finish_task_switch(this_rq(), prev
);
2806 * nr_running, nr_uninterruptible and nr_context_switches:
2808 * externally visible scheduler statistics: current number of runnable
2809 * threads, current number of uninterruptible-sleeping threads, total
2810 * number of context switches performed since bootup.
2812 unsigned long nr_running(void)
2814 unsigned long i
, sum
= 0;
2816 for_each_online_cpu(i
)
2817 sum
+= cpu_rq(i
)->nr_running
;
2822 unsigned long nr_uninterruptible(void)
2824 unsigned long i
, sum
= 0;
2826 for_each_possible_cpu(i
)
2827 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2830 * Since we read the counters lockless, it might be slightly
2831 * inaccurate. Do not allow it to go below zero though:
2833 if (unlikely((long)sum
< 0))
2839 unsigned long long nr_context_switches(void)
2842 unsigned long long sum
= 0;
2844 for_each_possible_cpu(i
)
2845 sum
+= cpu_rq(i
)->nr_switches
;
2850 unsigned long nr_iowait(void)
2852 unsigned long i
, sum
= 0;
2854 for_each_possible_cpu(i
)
2855 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2860 unsigned long nr_active(void)
2862 unsigned long i
, running
= 0, uninterruptible
= 0;
2864 for_each_online_cpu(i
) {
2865 running
+= cpu_rq(i
)->nr_running
;
2866 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2869 if (unlikely((long)uninterruptible
< 0))
2870 uninterruptible
= 0;
2872 return running
+ uninterruptible
;
2876 * Update rq->cpu_load[] statistics. This function is usually called every
2877 * scheduler tick (TICK_NSEC).
2879 static void update_cpu_load(struct rq
*this_rq
)
2881 unsigned long this_load
= this_rq
->load
.weight
;
2884 this_rq
->nr_load_updates
++;
2886 /* Update our load: */
2887 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2888 unsigned long old_load
, new_load
;
2890 /* scale is effectively 1 << i now, and >> i divides by scale */
2892 old_load
= this_rq
->cpu_load
[i
];
2893 new_load
= this_load
;
2895 * Round up the averaging division if load is increasing. This
2896 * prevents us from getting stuck on 9 if the load is 10, for
2899 if (new_load
> old_load
)
2900 new_load
+= scale
-1;
2901 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2908 * double_rq_lock - safely lock two runqueues
2910 * Note this does not disable interrupts like task_rq_lock,
2911 * you need to do so manually before calling.
2913 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2914 __acquires(rq1
->lock
)
2915 __acquires(rq2
->lock
)
2917 BUG_ON(!irqs_disabled());
2919 spin_lock(&rq1
->lock
);
2920 __acquire(rq2
->lock
); /* Fake it out ;) */
2923 spin_lock(&rq1
->lock
);
2924 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2926 spin_lock(&rq2
->lock
);
2927 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2930 update_rq_clock(rq1
);
2931 update_rq_clock(rq2
);
2935 * double_rq_unlock - safely unlock two runqueues
2937 * Note this does not restore interrupts like task_rq_unlock,
2938 * you need to do so manually after calling.
2940 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2941 __releases(rq1
->lock
)
2942 __releases(rq2
->lock
)
2944 spin_unlock(&rq1
->lock
);
2946 spin_unlock(&rq2
->lock
);
2948 __release(rq2
->lock
);
2952 * If dest_cpu is allowed for this process, migrate the task to it.
2953 * This is accomplished by forcing the cpu_allowed mask to only
2954 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2955 * the cpu_allowed mask is restored.
2957 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2959 struct migration_req req
;
2960 unsigned long flags
;
2963 rq
= task_rq_lock(p
, &flags
);
2964 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
2965 || unlikely(!cpu_active(dest_cpu
)))
2968 /* force the process onto the specified CPU */
2969 if (migrate_task(p
, dest_cpu
, &req
)) {
2970 /* Need to wait for migration thread (might exit: take ref). */
2971 struct task_struct
*mt
= rq
->migration_thread
;
2973 get_task_struct(mt
);
2974 task_rq_unlock(rq
, &flags
);
2975 wake_up_process(mt
);
2976 put_task_struct(mt
);
2977 wait_for_completion(&req
.done
);
2982 task_rq_unlock(rq
, &flags
);
2986 * sched_exec - execve() is a valuable balancing opportunity, because at
2987 * this point the task has the smallest effective memory and cache footprint.
2989 void sched_exec(void)
2991 int new_cpu
, this_cpu
= get_cpu();
2992 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2994 if (new_cpu
!= this_cpu
)
2995 sched_migrate_task(current
, new_cpu
);
2999 * pull_task - move a task from a remote runqueue to the local runqueue.
3000 * Both runqueues must be locked.
3002 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3003 struct rq
*this_rq
, int this_cpu
)
3005 deactivate_task(src_rq
, p
, 0);
3006 set_task_cpu(p
, this_cpu
);
3007 activate_task(this_rq
, p
, 0);
3009 * Note that idle threads have a prio of MAX_PRIO, for this test
3010 * to be always true for them.
3012 check_preempt_curr(this_rq
, p
, 0);
3016 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3019 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3020 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3023 int tsk_cache_hot
= 0;
3025 * We do not migrate tasks that are:
3026 * 1) running (obviously), or
3027 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3028 * 3) are cache-hot on their current CPU.
3030 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3031 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3036 if (task_running(rq
, p
)) {
3037 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3042 * Aggressive migration if:
3043 * 1) task is cache cold, or
3044 * 2) too many balance attempts have failed.
3047 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3048 if (!tsk_cache_hot
||
3049 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3050 #ifdef CONFIG_SCHEDSTATS
3051 if (tsk_cache_hot
) {
3052 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3053 schedstat_inc(p
, se
.nr_forced_migrations
);
3059 if (tsk_cache_hot
) {
3060 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3066 static unsigned long
3067 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3068 unsigned long max_load_move
, struct sched_domain
*sd
,
3069 enum cpu_idle_type idle
, int *all_pinned
,
3070 int *this_best_prio
, struct rq_iterator
*iterator
)
3072 int loops
= 0, pulled
= 0, pinned
= 0;
3073 struct task_struct
*p
;
3074 long rem_load_move
= max_load_move
;
3076 if (max_load_move
== 0)
3082 * Start the load-balancing iterator:
3084 p
= iterator
->start(iterator
->arg
);
3086 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3089 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3090 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3091 p
= iterator
->next(iterator
->arg
);
3095 pull_task(busiest
, p
, this_rq
, this_cpu
);
3097 rem_load_move
-= p
->se
.load
.weight
;
3099 #ifdef CONFIG_PREEMPT
3101 * NEWIDLE balancing is a source of latency, so preemptible kernels
3102 * will stop after the first task is pulled to minimize the critical
3105 if (idle
== CPU_NEWLY_IDLE
)
3110 * We only want to steal up to the prescribed amount of weighted load.
3112 if (rem_load_move
> 0) {
3113 if (p
->prio
< *this_best_prio
)
3114 *this_best_prio
= p
->prio
;
3115 p
= iterator
->next(iterator
->arg
);
3120 * Right now, this is one of only two places pull_task() is called,
3121 * so we can safely collect pull_task() stats here rather than
3122 * inside pull_task().
3124 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3127 *all_pinned
= pinned
;
3129 return max_load_move
- rem_load_move
;
3133 * move_tasks tries to move up to max_load_move weighted load from busiest to
3134 * this_rq, as part of a balancing operation within domain "sd".
3135 * Returns 1 if successful and 0 otherwise.
3137 * Called with both runqueues locked.
3139 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3140 unsigned long max_load_move
,
3141 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3144 const struct sched_class
*class = sched_class_highest
;
3145 unsigned long total_load_moved
= 0;
3146 int this_best_prio
= this_rq
->curr
->prio
;
3150 class->load_balance(this_rq
, this_cpu
, busiest
,
3151 max_load_move
- total_load_moved
,
3152 sd
, idle
, all_pinned
, &this_best_prio
);
3153 class = class->next
;
3155 #ifdef CONFIG_PREEMPT
3157 * NEWIDLE balancing is a source of latency, so preemptible
3158 * kernels will stop after the first task is pulled to minimize
3159 * the critical section.
3161 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3164 } while (class && max_load_move
> total_load_moved
);
3166 return total_load_moved
> 0;
3170 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3171 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3172 struct rq_iterator
*iterator
)
3174 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3178 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3179 pull_task(busiest
, p
, this_rq
, this_cpu
);
3181 * Right now, this is only the second place pull_task()
3182 * is called, so we can safely collect pull_task()
3183 * stats here rather than inside pull_task().
3185 schedstat_inc(sd
, lb_gained
[idle
]);
3189 p
= iterator
->next(iterator
->arg
);
3196 * move_one_task tries to move exactly one task from busiest to this_rq, as
3197 * part of active balancing operations within "domain".
3198 * Returns 1 if successful and 0 otherwise.
3200 * Called with both runqueues locked.
3202 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3203 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3205 const struct sched_class
*class;
3207 for (class = sched_class_highest
; class; class = class->next
)
3208 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3213 /********** Helpers for find_busiest_group ************************/
3215 * sd_lb_stats - Structure to store the statistics of a sched_domain
3216 * during load balancing.
3218 struct sd_lb_stats
{
3219 struct sched_group
*busiest
; /* Busiest group in this sd */
3220 struct sched_group
*this; /* Local group in this sd */
3221 unsigned long total_load
; /* Total load of all groups in sd */
3222 unsigned long total_pwr
; /* Total power of all groups in sd */
3223 unsigned long avg_load
; /* Average load across all groups in sd */
3225 /** Statistics of this group */
3226 unsigned long this_load
;
3227 unsigned long this_load_per_task
;
3228 unsigned long this_nr_running
;
3230 /* Statistics of the busiest group */
3231 unsigned long max_load
;
3232 unsigned long busiest_load_per_task
;
3233 unsigned long busiest_nr_running
;
3235 int group_imb
; /* Is there imbalance in this sd */
3236 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3237 int power_savings_balance
; /* Is powersave balance needed for this sd */
3238 struct sched_group
*group_min
; /* Least loaded group in sd */
3239 struct sched_group
*group_leader
; /* Group which relieves group_min */
3240 unsigned long min_load_per_task
; /* load_per_task in group_min */
3241 unsigned long leader_nr_running
; /* Nr running of group_leader */
3242 unsigned long min_nr_running
; /* Nr running of group_min */
3247 * sg_lb_stats - stats of a sched_group required for load_balancing
3249 struct sg_lb_stats
{
3250 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3251 unsigned long group_load
; /* Total load over the CPUs of the group */
3252 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3253 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3254 unsigned long group_capacity
;
3255 int group_imb
; /* Is there an imbalance in the group ? */
3259 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3260 * @group: The group whose first cpu is to be returned.
3262 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3264 return cpumask_first(sched_group_cpus(group
));
3268 * get_sd_load_idx - Obtain the load index for a given sched domain.
3269 * @sd: The sched_domain whose load_idx is to be obtained.
3270 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3272 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3273 enum cpu_idle_type idle
)
3279 load_idx
= sd
->busy_idx
;
3282 case CPU_NEWLY_IDLE
:
3283 load_idx
= sd
->newidle_idx
;
3286 load_idx
= sd
->idle_idx
;
3294 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3296 * init_sd_power_savings_stats - Initialize power savings statistics for
3297 * the given sched_domain, during load balancing.
3299 * @sd: Sched domain whose power-savings statistics are to be initialized.
3300 * @sds: Variable containing the statistics for sd.
3301 * @idle: Idle status of the CPU at which we're performing load-balancing.
3303 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3304 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3307 * Busy processors will not participate in power savings
3310 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3311 sds
->power_savings_balance
= 0;
3313 sds
->power_savings_balance
= 1;
3314 sds
->min_nr_running
= ULONG_MAX
;
3315 sds
->leader_nr_running
= 0;
3320 * update_sd_power_savings_stats - Update the power saving stats for a
3321 * sched_domain while performing load balancing.
3323 * @group: sched_group belonging to the sched_domain under consideration.
3324 * @sds: Variable containing the statistics of the sched_domain
3325 * @local_group: Does group contain the CPU for which we're performing
3327 * @sgs: Variable containing the statistics of the group.
3329 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3330 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3333 if (!sds
->power_savings_balance
)
3337 * If the local group is idle or completely loaded
3338 * no need to do power savings balance at this domain
3340 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3341 !sds
->this_nr_running
))
3342 sds
->power_savings_balance
= 0;
3345 * If a group is already running at full capacity or idle,
3346 * don't include that group in power savings calculations
3348 if (!sds
->power_savings_balance
||
3349 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3350 !sgs
->sum_nr_running
)
3354 * Calculate the group which has the least non-idle load.
3355 * This is the group from where we need to pick up the load
3358 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3359 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3360 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3361 sds
->group_min
= group
;
3362 sds
->min_nr_running
= sgs
->sum_nr_running
;
3363 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3364 sgs
->sum_nr_running
;
3368 * Calculate the group which is almost near its
3369 * capacity but still has some space to pick up some load
3370 * from other group and save more power
3372 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3375 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3376 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3377 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3378 sds
->group_leader
= group
;
3379 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3384 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3385 * @sds: Variable containing the statistics of the sched_domain
3386 * under consideration.
3387 * @this_cpu: Cpu at which we're currently performing load-balancing.
3388 * @imbalance: Variable to store the imbalance.
3391 * Check if we have potential to perform some power-savings balance.
3392 * If yes, set the busiest group to be the least loaded group in the
3393 * sched_domain, so that it's CPUs can be put to idle.
3395 * Returns 1 if there is potential to perform power-savings balance.
3398 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3399 int this_cpu
, unsigned long *imbalance
)
3401 if (!sds
->power_savings_balance
)
3404 if (sds
->this != sds
->group_leader
||
3405 sds
->group_leader
== sds
->group_min
)
3408 *imbalance
= sds
->min_load_per_task
;
3409 sds
->busiest
= sds
->group_min
;
3411 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3412 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3413 group_first_cpu(sds
->group_leader
);
3419 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3420 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3421 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3426 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3427 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3432 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3433 int this_cpu
, unsigned long *imbalance
)
3437 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3441 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3442 * @group: sched_group whose statistics are to be updated.
3443 * @this_cpu: Cpu for which load balance is currently performed.
3444 * @idle: Idle status of this_cpu
3445 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3446 * @sd_idle: Idle status of the sched_domain containing group.
3447 * @local_group: Does group contain this_cpu.
3448 * @cpus: Set of cpus considered for load balancing.
3449 * @balance: Should we balance.
3450 * @sgs: variable to hold the statistics for this group.
3452 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3453 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3454 int local_group
, const struct cpumask
*cpus
,
3455 int *balance
, struct sg_lb_stats
*sgs
)
3457 unsigned long load
, max_cpu_load
, min_cpu_load
;
3459 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3460 unsigned long sum_avg_load_per_task
;
3461 unsigned long avg_load_per_task
;
3464 balance_cpu
= group_first_cpu(group
);
3466 /* Tally up the load of all CPUs in the group */
3467 sum_avg_load_per_task
= avg_load_per_task
= 0;
3469 min_cpu_load
= ~0UL;
3471 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3472 struct rq
*rq
= cpu_rq(i
);
3474 if (*sd_idle
&& rq
->nr_running
)
3477 /* Bias balancing toward cpus of our domain */
3479 if (idle_cpu(i
) && !first_idle_cpu
) {
3484 load
= target_load(i
, load_idx
);
3486 load
= source_load(i
, load_idx
);
3487 if (load
> max_cpu_load
)
3488 max_cpu_load
= load
;
3489 if (min_cpu_load
> load
)
3490 min_cpu_load
= load
;
3493 sgs
->group_load
+= load
;
3494 sgs
->sum_nr_running
+= rq
->nr_running
;
3495 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3497 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3501 * First idle cpu or the first cpu(busiest) in this sched group
3502 * is eligible for doing load balancing at this and above
3503 * domains. In the newly idle case, we will allow all the cpu's
3504 * to do the newly idle load balance.
3506 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3507 balance_cpu
!= this_cpu
&& balance
) {
3512 /* Adjust by relative CPU power of the group */
3513 sgs
->avg_load
= sg_div_cpu_power(group
,
3514 sgs
->group_load
* SCHED_LOAD_SCALE
);
3518 * Consider the group unbalanced when the imbalance is larger
3519 * than the average weight of two tasks.
3521 * APZ: with cgroup the avg task weight can vary wildly and
3522 * might not be a suitable number - should we keep a
3523 * normalized nr_running number somewhere that negates
3526 avg_load_per_task
= sg_div_cpu_power(group
,
3527 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3529 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3532 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3537 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3538 * @sd: sched_domain whose statistics are to be updated.
3539 * @this_cpu: Cpu for which load balance is currently performed.
3540 * @idle: Idle status of this_cpu
3541 * @sd_idle: Idle status of the sched_domain containing group.
3542 * @cpus: Set of cpus considered for load balancing.
3543 * @balance: Should we balance.
3544 * @sds: variable to hold the statistics for this sched_domain.
3546 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3547 enum cpu_idle_type idle
, int *sd_idle
,
3548 const struct cpumask
*cpus
, int *balance
,
3549 struct sd_lb_stats
*sds
)
3551 struct sched_group
*group
= sd
->groups
;
3552 struct sg_lb_stats sgs
;
3555 init_sd_power_savings_stats(sd
, sds
, idle
);
3556 load_idx
= get_sd_load_idx(sd
, idle
);
3561 local_group
= cpumask_test_cpu(this_cpu
,
3562 sched_group_cpus(group
));
3563 memset(&sgs
, 0, sizeof(sgs
));
3564 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3565 local_group
, cpus
, balance
, &sgs
);
3567 if (local_group
&& balance
&& !(*balance
))
3570 sds
->total_load
+= sgs
.group_load
;
3571 sds
->total_pwr
+= group
->__cpu_power
;
3574 sds
->this_load
= sgs
.avg_load
;
3576 sds
->this_nr_running
= sgs
.sum_nr_running
;
3577 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3578 } else if (sgs
.avg_load
> sds
->max_load
&&
3579 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3581 sds
->max_load
= sgs
.avg_load
;
3582 sds
->busiest
= group
;
3583 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3584 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3585 sds
->group_imb
= sgs
.group_imb
;
3588 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3589 group
= group
->next
;
3590 } while (group
!= sd
->groups
);
3595 * fix_small_imbalance - Calculate the minor imbalance that exists
3596 * amongst the groups of a sched_domain, during
3598 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3599 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3600 * @imbalance: Variable to store the imbalance.
3602 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3603 int this_cpu
, unsigned long *imbalance
)
3605 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3606 unsigned int imbn
= 2;
3608 if (sds
->this_nr_running
) {
3609 sds
->this_load_per_task
/= sds
->this_nr_running
;
3610 if (sds
->busiest_load_per_task
>
3611 sds
->this_load_per_task
)
3614 sds
->this_load_per_task
=
3615 cpu_avg_load_per_task(this_cpu
);
3617 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3618 sds
->busiest_load_per_task
* imbn
) {
3619 *imbalance
= sds
->busiest_load_per_task
;
3624 * OK, we don't have enough imbalance to justify moving tasks,
3625 * however we may be able to increase total CPU power used by
3629 pwr_now
+= sds
->busiest
->__cpu_power
*
3630 min(sds
->busiest_load_per_task
, sds
->max_load
);
3631 pwr_now
+= sds
->this->__cpu_power
*
3632 min(sds
->this_load_per_task
, sds
->this_load
);
3633 pwr_now
/= SCHED_LOAD_SCALE
;
3635 /* Amount of load we'd subtract */
3636 tmp
= sg_div_cpu_power(sds
->busiest
,
3637 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3638 if (sds
->max_load
> tmp
)
3639 pwr_move
+= sds
->busiest
->__cpu_power
*
3640 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3642 /* Amount of load we'd add */
3643 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3644 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3645 tmp
= sg_div_cpu_power(sds
->this,
3646 sds
->max_load
* sds
->busiest
->__cpu_power
);
3648 tmp
= sg_div_cpu_power(sds
->this,
3649 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3650 pwr_move
+= sds
->this->__cpu_power
*
3651 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3652 pwr_move
/= SCHED_LOAD_SCALE
;
3654 /* Move if we gain throughput */
3655 if (pwr_move
> pwr_now
)
3656 *imbalance
= sds
->busiest_load_per_task
;
3660 * calculate_imbalance - Calculate the amount of imbalance present within the
3661 * groups of a given sched_domain during load balance.
3662 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3663 * @this_cpu: Cpu for which currently load balance is being performed.
3664 * @imbalance: The variable to store the imbalance.
3666 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3667 unsigned long *imbalance
)
3669 unsigned long max_pull
;
3671 * In the presence of smp nice balancing, certain scenarios can have
3672 * max load less than avg load(as we skip the groups at or below
3673 * its cpu_power, while calculating max_load..)
3675 if (sds
->max_load
< sds
->avg_load
) {
3677 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3680 /* Don't want to pull so many tasks that a group would go idle */
3681 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3682 sds
->max_load
- sds
->busiest_load_per_task
);
3684 /* How much load to actually move to equalise the imbalance */
3685 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3686 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3690 * if *imbalance is less than the average load per runnable task
3691 * there is no gaurantee that any tasks will be moved so we'll have
3692 * a think about bumping its value to force at least one task to be
3695 if (*imbalance
< sds
->busiest_load_per_task
)
3696 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3699 /******* find_busiest_group() helpers end here *********************/
3702 * find_busiest_group - Returns the busiest group within the sched_domain
3703 * if there is an imbalance. If there isn't an imbalance, and
3704 * the user has opted for power-savings, it returns a group whose
3705 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3706 * such a group exists.
3708 * Also calculates the amount of weighted load which should be moved
3709 * to restore balance.
3711 * @sd: The sched_domain whose busiest group is to be returned.
3712 * @this_cpu: The cpu for which load balancing is currently being performed.
3713 * @imbalance: Variable which stores amount of weighted load which should
3714 * be moved to restore balance/put a group to idle.
3715 * @idle: The idle status of this_cpu.
3716 * @sd_idle: The idleness of sd
3717 * @cpus: The set of CPUs under consideration for load-balancing.
3718 * @balance: Pointer to a variable indicating if this_cpu
3719 * is the appropriate cpu to perform load balancing at this_level.
3721 * Returns: - the busiest group if imbalance exists.
3722 * - If no imbalance and user has opted for power-savings balance,
3723 * return the least loaded group whose CPUs can be
3724 * put to idle by rebalancing its tasks onto our group.
3726 static struct sched_group
*
3727 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3728 unsigned long *imbalance
, enum cpu_idle_type idle
,
3729 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3731 struct sd_lb_stats sds
;
3733 memset(&sds
, 0, sizeof(sds
));
3736 * Compute the various statistics relavent for load balancing at
3739 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3742 /* Cases where imbalance does not exist from POV of this_cpu */
3743 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3745 * 2) There is no busy sibling group to pull from.
3746 * 3) This group is the busiest group.
3747 * 4) This group is more busy than the avg busieness at this
3749 * 5) The imbalance is within the specified limit.
3750 * 6) Any rebalance would lead to ping-pong
3752 if (balance
&& !(*balance
))
3755 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3758 if (sds
.this_load
>= sds
.max_load
)
3761 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3763 if (sds
.this_load
>= sds
.avg_load
)
3766 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3769 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3771 sds
.busiest_load_per_task
=
3772 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3775 * We're trying to get all the cpus to the average_load, so we don't
3776 * want to push ourselves above the average load, nor do we wish to
3777 * reduce the max loaded cpu below the average load, as either of these
3778 * actions would just result in more rebalancing later, and ping-pong
3779 * tasks around. Thus we look for the minimum possible imbalance.
3780 * Negative imbalances (*we* are more loaded than anyone else) will
3781 * be counted as no imbalance for these purposes -- we can't fix that
3782 * by pulling tasks to us. Be careful of negative numbers as they'll
3783 * appear as very large values with unsigned longs.
3785 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3788 /* Looks like there is an imbalance. Compute it */
3789 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3794 * There is no obvious imbalance. But check if we can do some balancing
3797 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3805 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3808 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3809 unsigned long imbalance
, const struct cpumask
*cpus
)
3811 struct rq
*busiest
= NULL
, *rq
;
3812 unsigned long max_load
= 0;
3815 for_each_cpu(i
, sched_group_cpus(group
)) {
3818 if (!cpumask_test_cpu(i
, cpus
))
3822 wl
= weighted_cpuload(i
);
3824 if (rq
->nr_running
== 1 && wl
> imbalance
)
3827 if (wl
> max_load
) {
3837 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3838 * so long as it is large enough.
3840 #define MAX_PINNED_INTERVAL 512
3842 /* Working cpumask for load_balance and load_balance_newidle. */
3843 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3846 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3847 * tasks if there is an imbalance.
3849 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3850 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3853 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3854 struct sched_group
*group
;
3855 unsigned long imbalance
;
3857 unsigned long flags
;
3858 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3860 cpumask_setall(cpus
);
3863 * When power savings policy is enabled for the parent domain, idle
3864 * sibling can pick up load irrespective of busy siblings. In this case,
3865 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3866 * portraying it as CPU_NOT_IDLE.
3868 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3869 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3872 schedstat_inc(sd
, lb_count
[idle
]);
3876 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3883 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3887 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3889 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3893 BUG_ON(busiest
== this_rq
);
3895 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3898 if (busiest
->nr_running
> 1) {
3900 * Attempt to move tasks. If find_busiest_group has found
3901 * an imbalance but busiest->nr_running <= 1, the group is
3902 * still unbalanced. ld_moved simply stays zero, so it is
3903 * correctly treated as an imbalance.
3905 local_irq_save(flags
);
3906 double_rq_lock(this_rq
, busiest
);
3907 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3908 imbalance
, sd
, idle
, &all_pinned
);
3909 double_rq_unlock(this_rq
, busiest
);
3910 local_irq_restore(flags
);
3913 * some other cpu did the load balance for us.
3915 if (ld_moved
&& this_cpu
!= smp_processor_id())
3916 resched_cpu(this_cpu
);
3918 /* All tasks on this runqueue were pinned by CPU affinity */
3919 if (unlikely(all_pinned
)) {
3920 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3921 if (!cpumask_empty(cpus
))
3928 schedstat_inc(sd
, lb_failed
[idle
]);
3929 sd
->nr_balance_failed
++;
3931 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3933 spin_lock_irqsave(&busiest
->lock
, flags
);
3935 /* don't kick the migration_thread, if the curr
3936 * task on busiest cpu can't be moved to this_cpu
3938 if (!cpumask_test_cpu(this_cpu
,
3939 &busiest
->curr
->cpus_allowed
)) {
3940 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3942 goto out_one_pinned
;
3945 if (!busiest
->active_balance
) {
3946 busiest
->active_balance
= 1;
3947 busiest
->push_cpu
= this_cpu
;
3950 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3952 wake_up_process(busiest
->migration_thread
);
3955 * We've kicked active balancing, reset the failure
3958 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3961 sd
->nr_balance_failed
= 0;
3963 if (likely(!active_balance
)) {
3964 /* We were unbalanced, so reset the balancing interval */
3965 sd
->balance_interval
= sd
->min_interval
;
3968 * If we've begun active balancing, start to back off. This
3969 * case may not be covered by the all_pinned logic if there
3970 * is only 1 task on the busy runqueue (because we don't call
3973 if (sd
->balance_interval
< sd
->max_interval
)
3974 sd
->balance_interval
*= 2;
3977 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3978 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3984 schedstat_inc(sd
, lb_balanced
[idle
]);
3986 sd
->nr_balance_failed
= 0;
3989 /* tune up the balancing interval */
3990 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3991 (sd
->balance_interval
< sd
->max_interval
))
3992 sd
->balance_interval
*= 2;
3994 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3995 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4006 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4007 * tasks if there is an imbalance.
4009 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4010 * this_rq is locked.
4013 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4015 struct sched_group
*group
;
4016 struct rq
*busiest
= NULL
;
4017 unsigned long imbalance
;
4021 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4023 cpumask_setall(cpus
);
4026 * When power savings policy is enabled for the parent domain, idle
4027 * sibling can pick up load irrespective of busy siblings. In this case,
4028 * let the state of idle sibling percolate up as IDLE, instead of
4029 * portraying it as CPU_NOT_IDLE.
4031 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4032 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4035 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4037 update_shares_locked(this_rq
, sd
);
4038 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4039 &sd_idle
, cpus
, NULL
);
4041 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4045 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4047 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4051 BUG_ON(busiest
== this_rq
);
4053 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4056 if (busiest
->nr_running
> 1) {
4057 /* Attempt to move tasks */
4058 double_lock_balance(this_rq
, busiest
);
4059 /* this_rq->clock is already updated */
4060 update_rq_clock(busiest
);
4061 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4062 imbalance
, sd
, CPU_NEWLY_IDLE
,
4064 double_unlock_balance(this_rq
, busiest
);
4066 if (unlikely(all_pinned
)) {
4067 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4068 if (!cpumask_empty(cpus
))
4074 int active_balance
= 0;
4076 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4077 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4078 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4081 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4084 if (sd
->nr_balance_failed
++ < 2)
4088 * The only task running in a non-idle cpu can be moved to this
4089 * cpu in an attempt to completely freeup the other CPU
4090 * package. The same method used to move task in load_balance()
4091 * have been extended for load_balance_newidle() to speedup
4092 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4094 * The package power saving logic comes from
4095 * find_busiest_group(). If there are no imbalance, then
4096 * f_b_g() will return NULL. However when sched_mc={1,2} then
4097 * f_b_g() will select a group from which a running task may be
4098 * pulled to this cpu in order to make the other package idle.
4099 * If there is no opportunity to make a package idle and if
4100 * there are no imbalance, then f_b_g() will return NULL and no
4101 * action will be taken in load_balance_newidle().
4103 * Under normal task pull operation due to imbalance, there
4104 * will be more than one task in the source run queue and
4105 * move_tasks() will succeed. ld_moved will be true and this
4106 * active balance code will not be triggered.
4109 /* Lock busiest in correct order while this_rq is held */
4110 double_lock_balance(this_rq
, busiest
);
4113 * don't kick the migration_thread, if the curr
4114 * task on busiest cpu can't be moved to this_cpu
4116 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4117 double_unlock_balance(this_rq
, busiest
);
4122 if (!busiest
->active_balance
) {
4123 busiest
->active_balance
= 1;
4124 busiest
->push_cpu
= this_cpu
;
4128 double_unlock_balance(this_rq
, busiest
);
4130 * Should not call ttwu while holding a rq->lock
4132 spin_unlock(&this_rq
->lock
);
4134 wake_up_process(busiest
->migration_thread
);
4135 spin_lock(&this_rq
->lock
);
4138 sd
->nr_balance_failed
= 0;
4140 update_shares_locked(this_rq
, sd
);
4144 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4145 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4146 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4148 sd
->nr_balance_failed
= 0;
4154 * idle_balance is called by schedule() if this_cpu is about to become
4155 * idle. Attempts to pull tasks from other CPUs.
4157 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4159 struct sched_domain
*sd
;
4160 int pulled_task
= 0;
4161 unsigned long next_balance
= jiffies
+ HZ
;
4163 for_each_domain(this_cpu
, sd
) {
4164 unsigned long interval
;
4166 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4169 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4170 /* If we've pulled tasks over stop searching: */
4171 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4174 interval
= msecs_to_jiffies(sd
->balance_interval
);
4175 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4176 next_balance
= sd
->last_balance
+ interval
;
4180 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4182 * We are going idle. next_balance may be set based on
4183 * a busy processor. So reset next_balance.
4185 this_rq
->next_balance
= next_balance
;
4190 * active_load_balance is run by migration threads. It pushes running tasks
4191 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4192 * running on each physical CPU where possible, and avoids physical /
4193 * logical imbalances.
4195 * Called with busiest_rq locked.
4197 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4199 int target_cpu
= busiest_rq
->push_cpu
;
4200 struct sched_domain
*sd
;
4201 struct rq
*target_rq
;
4203 /* Is there any task to move? */
4204 if (busiest_rq
->nr_running
<= 1)
4207 target_rq
= cpu_rq(target_cpu
);
4210 * This condition is "impossible", if it occurs
4211 * we need to fix it. Originally reported by
4212 * Bjorn Helgaas on a 128-cpu setup.
4214 BUG_ON(busiest_rq
== target_rq
);
4216 /* move a task from busiest_rq to target_rq */
4217 double_lock_balance(busiest_rq
, target_rq
);
4218 update_rq_clock(busiest_rq
);
4219 update_rq_clock(target_rq
);
4221 /* Search for an sd spanning us and the target CPU. */
4222 for_each_domain(target_cpu
, sd
) {
4223 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4224 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4229 schedstat_inc(sd
, alb_count
);
4231 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4233 schedstat_inc(sd
, alb_pushed
);
4235 schedstat_inc(sd
, alb_failed
);
4237 double_unlock_balance(busiest_rq
, target_rq
);
4242 atomic_t load_balancer
;
4243 cpumask_var_t cpu_mask
;
4244 } nohz ____cacheline_aligned
= {
4245 .load_balancer
= ATOMIC_INIT(-1),
4249 * This routine will try to nominate the ilb (idle load balancing)
4250 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4251 * load balancing on behalf of all those cpus. If all the cpus in the system
4252 * go into this tickless mode, then there will be no ilb owner (as there is
4253 * no need for one) and all the cpus will sleep till the next wakeup event
4256 * For the ilb owner, tick is not stopped. And this tick will be used
4257 * for idle load balancing. ilb owner will still be part of
4260 * While stopping the tick, this cpu will become the ilb owner if there
4261 * is no other owner. And will be the owner till that cpu becomes busy
4262 * or if all cpus in the system stop their ticks at which point
4263 * there is no need for ilb owner.
4265 * When the ilb owner becomes busy, it nominates another owner, during the
4266 * next busy scheduler_tick()
4268 int select_nohz_load_balancer(int stop_tick
)
4270 int cpu
= smp_processor_id();
4273 cpu_rq(cpu
)->in_nohz_recently
= 1;
4275 if (!cpu_active(cpu
)) {
4276 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4280 * If we are going offline and still the leader,
4283 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4289 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4291 /* time for ilb owner also to sleep */
4292 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4293 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4294 atomic_set(&nohz
.load_balancer
, -1);
4298 if (atomic_read(&nohz
.load_balancer
) == -1) {
4299 /* make me the ilb owner */
4300 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4302 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4305 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4308 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4310 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4311 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4318 static DEFINE_SPINLOCK(balancing
);
4321 * It checks each scheduling domain to see if it is due to be balanced,
4322 * and initiates a balancing operation if so.
4324 * Balancing parameters are set up in arch_init_sched_domains.
4326 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4329 struct rq
*rq
= cpu_rq(cpu
);
4330 unsigned long interval
;
4331 struct sched_domain
*sd
;
4332 /* Earliest time when we have to do rebalance again */
4333 unsigned long next_balance
= jiffies
+ 60*HZ
;
4334 int update_next_balance
= 0;
4337 for_each_domain(cpu
, sd
) {
4338 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4341 interval
= sd
->balance_interval
;
4342 if (idle
!= CPU_IDLE
)
4343 interval
*= sd
->busy_factor
;
4345 /* scale ms to jiffies */
4346 interval
= msecs_to_jiffies(interval
);
4347 if (unlikely(!interval
))
4349 if (interval
> HZ
*NR_CPUS
/10)
4350 interval
= HZ
*NR_CPUS
/10;
4352 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4354 if (need_serialize
) {
4355 if (!spin_trylock(&balancing
))
4359 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4360 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4362 * We've pulled tasks over so either we're no
4363 * longer idle, or one of our SMT siblings is
4366 idle
= CPU_NOT_IDLE
;
4368 sd
->last_balance
= jiffies
;
4371 spin_unlock(&balancing
);
4373 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4374 next_balance
= sd
->last_balance
+ interval
;
4375 update_next_balance
= 1;
4379 * Stop the load balance at this level. There is another
4380 * CPU in our sched group which is doing load balancing more
4388 * next_balance will be updated only when there is a need.
4389 * When the cpu is attached to null domain for ex, it will not be
4392 if (likely(update_next_balance
))
4393 rq
->next_balance
= next_balance
;
4397 * run_rebalance_domains is triggered when needed from the scheduler tick.
4398 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4399 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4401 static void run_rebalance_domains(struct softirq_action
*h
)
4403 int this_cpu
= smp_processor_id();
4404 struct rq
*this_rq
= cpu_rq(this_cpu
);
4405 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4406 CPU_IDLE
: CPU_NOT_IDLE
;
4408 rebalance_domains(this_cpu
, idle
);
4412 * If this cpu is the owner for idle load balancing, then do the
4413 * balancing on behalf of the other idle cpus whose ticks are
4416 if (this_rq
->idle_at_tick
&&
4417 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4421 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4422 if (balance_cpu
== this_cpu
)
4426 * If this cpu gets work to do, stop the load balancing
4427 * work being done for other cpus. Next load
4428 * balancing owner will pick it up.
4433 rebalance_domains(balance_cpu
, CPU_IDLE
);
4435 rq
= cpu_rq(balance_cpu
);
4436 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4437 this_rq
->next_balance
= rq
->next_balance
;
4443 static inline int on_null_domain(int cpu
)
4445 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4449 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4451 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4452 * idle load balancing owner or decide to stop the periodic load balancing,
4453 * if the whole system is idle.
4455 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4459 * If we were in the nohz mode recently and busy at the current
4460 * scheduler tick, then check if we need to nominate new idle
4463 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4464 rq
->in_nohz_recently
= 0;
4466 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4467 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4468 atomic_set(&nohz
.load_balancer
, -1);
4471 if (atomic_read(&nohz
.load_balancer
) == -1) {
4473 * simple selection for now: Nominate the
4474 * first cpu in the nohz list to be the next
4477 * TBD: Traverse the sched domains and nominate
4478 * the nearest cpu in the nohz.cpu_mask.
4480 int ilb
= cpumask_first(nohz
.cpu_mask
);
4482 if (ilb
< nr_cpu_ids
)
4488 * If this cpu is idle and doing idle load balancing for all the
4489 * cpus with ticks stopped, is it time for that to stop?
4491 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4492 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4498 * If this cpu is idle and the idle load balancing is done by
4499 * someone else, then no need raise the SCHED_SOFTIRQ
4501 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4502 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4505 /* Don't need to rebalance while attached to NULL domain */
4506 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4507 likely(!on_null_domain(cpu
)))
4508 raise_softirq(SCHED_SOFTIRQ
);
4511 #else /* CONFIG_SMP */
4514 * on UP we do not need to balance between CPUs:
4516 static inline void idle_balance(int cpu
, struct rq
*rq
)
4522 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4524 EXPORT_PER_CPU_SYMBOL(kstat
);
4527 * Return any ns on the sched_clock that have not yet been accounted in
4528 * @p in case that task is currently running.
4530 * Called with task_rq_lock() held on @rq.
4532 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4536 if (task_current(rq
, p
)) {
4537 update_rq_clock(rq
);
4538 ns
= rq
->clock
- p
->se
.exec_start
;
4546 unsigned long long task_delta_exec(struct task_struct
*p
)
4548 unsigned long flags
;
4552 rq
= task_rq_lock(p
, &flags
);
4553 ns
= do_task_delta_exec(p
, rq
);
4554 task_rq_unlock(rq
, &flags
);
4560 * Return accounted runtime for the task.
4561 * In case the task is currently running, return the runtime plus current's
4562 * pending runtime that have not been accounted yet.
4564 unsigned long long task_sched_runtime(struct task_struct
*p
)
4566 unsigned long flags
;
4570 rq
= task_rq_lock(p
, &flags
);
4571 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4572 task_rq_unlock(rq
, &flags
);
4578 * Return sum_exec_runtime for the thread group.
4579 * In case the task is currently running, return the sum plus current's
4580 * pending runtime that have not been accounted yet.
4582 * Note that the thread group might have other running tasks as well,
4583 * so the return value not includes other pending runtime that other
4584 * running tasks might have.
4586 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4588 struct task_cputime totals
;
4589 unsigned long flags
;
4593 rq
= task_rq_lock(p
, &flags
);
4594 thread_group_cputime(p
, &totals
);
4595 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4596 task_rq_unlock(rq
, &flags
);
4602 * Account user cpu time to a process.
4603 * @p: the process that the cpu time gets accounted to
4604 * @cputime: the cpu time spent in user space since the last update
4605 * @cputime_scaled: cputime scaled by cpu frequency
4607 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4608 cputime_t cputime_scaled
)
4610 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4613 /* Add user time to process. */
4614 p
->utime
= cputime_add(p
->utime
, cputime
);
4615 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4616 account_group_user_time(p
, cputime
);
4618 /* Add user time to cpustat. */
4619 tmp
= cputime_to_cputime64(cputime
);
4620 if (TASK_NICE(p
) > 0)
4621 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4623 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4625 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4626 /* Account for user time used */
4627 acct_update_integrals(p
);
4631 * Account guest cpu time to a process.
4632 * @p: the process that the cpu time gets accounted to
4633 * @cputime: the cpu time spent in virtual machine since the last update
4634 * @cputime_scaled: cputime scaled by cpu frequency
4636 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4637 cputime_t cputime_scaled
)
4640 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4642 tmp
= cputime_to_cputime64(cputime
);
4644 /* Add guest time to process. */
4645 p
->utime
= cputime_add(p
->utime
, cputime
);
4646 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4647 account_group_user_time(p
, cputime
);
4648 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4650 /* Add guest time to cpustat. */
4651 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4652 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4656 * Account system cpu time to a process.
4657 * @p: the process that the cpu time gets accounted to
4658 * @hardirq_offset: the offset to subtract from hardirq_count()
4659 * @cputime: the cpu time spent in kernel space since the last update
4660 * @cputime_scaled: cputime scaled by cpu frequency
4662 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4663 cputime_t cputime
, cputime_t cputime_scaled
)
4665 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4668 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4669 account_guest_time(p
, cputime
, cputime_scaled
);
4673 /* Add system time to process. */
4674 p
->stime
= cputime_add(p
->stime
, cputime
);
4675 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4676 account_group_system_time(p
, cputime
);
4678 /* Add system time to cpustat. */
4679 tmp
= cputime_to_cputime64(cputime
);
4680 if (hardirq_count() - hardirq_offset
)
4681 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4682 else if (softirq_count())
4683 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4685 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4687 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4689 /* Account for system time used */
4690 acct_update_integrals(p
);
4694 * Account for involuntary wait time.
4695 * @steal: the cpu time spent in involuntary wait
4697 void account_steal_time(cputime_t cputime
)
4699 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4700 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4702 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4706 * Account for idle time.
4707 * @cputime: the cpu time spent in idle wait
4709 void account_idle_time(cputime_t cputime
)
4711 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4712 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4713 struct rq
*rq
= this_rq();
4715 if (atomic_read(&rq
->nr_iowait
) > 0)
4716 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4718 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4721 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4724 * Account a single tick of cpu time.
4725 * @p: the process that the cpu time gets accounted to
4726 * @user_tick: indicates if the tick is a user or a system tick
4728 void account_process_tick(struct task_struct
*p
, int user_tick
)
4730 cputime_t one_jiffy
= jiffies_to_cputime(1);
4731 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4732 struct rq
*rq
= this_rq();
4735 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4736 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
4737 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4740 account_idle_time(one_jiffy
);
4744 * Account multiple ticks of steal time.
4745 * @p: the process from which the cpu time has been stolen
4746 * @ticks: number of stolen ticks
4748 void account_steal_ticks(unsigned long ticks
)
4750 account_steal_time(jiffies_to_cputime(ticks
));
4754 * Account multiple ticks of idle time.
4755 * @ticks: number of stolen ticks
4757 void account_idle_ticks(unsigned long ticks
)
4759 account_idle_time(jiffies_to_cputime(ticks
));
4765 * Use precise platform statistics if available:
4767 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4768 cputime_t
task_utime(struct task_struct
*p
)
4773 cputime_t
task_stime(struct task_struct
*p
)
4778 cputime_t
task_utime(struct task_struct
*p
)
4780 clock_t utime
= cputime_to_clock_t(p
->utime
),
4781 total
= utime
+ cputime_to_clock_t(p
->stime
);
4785 * Use CFS's precise accounting:
4787 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4791 do_div(temp
, total
);
4793 utime
= (clock_t)temp
;
4795 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4796 return p
->prev_utime
;
4799 cputime_t
task_stime(struct task_struct
*p
)
4804 * Use CFS's precise accounting. (we subtract utime from
4805 * the total, to make sure the total observed by userspace
4806 * grows monotonically - apps rely on that):
4808 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4809 cputime_to_clock_t(task_utime(p
));
4812 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4814 return p
->prev_stime
;
4818 inline cputime_t
task_gtime(struct task_struct
*p
)
4824 * This function gets called by the timer code, with HZ frequency.
4825 * We call it with interrupts disabled.
4827 * It also gets called by the fork code, when changing the parent's
4830 void scheduler_tick(void)
4832 int cpu
= smp_processor_id();
4833 struct rq
*rq
= cpu_rq(cpu
);
4834 struct task_struct
*curr
= rq
->curr
;
4838 spin_lock(&rq
->lock
);
4839 update_rq_clock(rq
);
4840 update_cpu_load(rq
);
4841 curr
->sched_class
->task_tick(rq
, curr
, 0);
4842 spin_unlock(&rq
->lock
);
4845 rq
->idle_at_tick
= idle_cpu(cpu
);
4846 trigger_load_balance(rq
, cpu
);
4850 notrace
unsigned long get_parent_ip(unsigned long addr
)
4852 if (in_lock_functions(addr
)) {
4853 addr
= CALLER_ADDR2
;
4854 if (in_lock_functions(addr
))
4855 addr
= CALLER_ADDR3
;
4860 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4861 defined(CONFIG_PREEMPT_TRACER))
4863 void __kprobes
add_preempt_count(int val
)
4865 #ifdef CONFIG_DEBUG_PREEMPT
4869 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4872 preempt_count() += val
;
4873 #ifdef CONFIG_DEBUG_PREEMPT
4875 * Spinlock count overflowing soon?
4877 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4880 if (preempt_count() == val
)
4881 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4883 EXPORT_SYMBOL(add_preempt_count
);
4885 void __kprobes
sub_preempt_count(int val
)
4887 #ifdef CONFIG_DEBUG_PREEMPT
4891 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4894 * Is the spinlock portion underflowing?
4896 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4897 !(preempt_count() & PREEMPT_MASK
)))
4901 if (preempt_count() == val
)
4902 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4903 preempt_count() -= val
;
4905 EXPORT_SYMBOL(sub_preempt_count
);
4910 * Print scheduling while atomic bug:
4912 static noinline
void __schedule_bug(struct task_struct
*prev
)
4914 struct pt_regs
*regs
= get_irq_regs();
4916 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4917 prev
->comm
, prev
->pid
, preempt_count());
4919 debug_show_held_locks(prev
);
4921 if (irqs_disabled())
4922 print_irqtrace_events(prev
);
4931 * Various schedule()-time debugging checks and statistics:
4933 static inline void schedule_debug(struct task_struct
*prev
)
4936 * Test if we are atomic. Since do_exit() needs to call into
4937 * schedule() atomically, we ignore that path for now.
4938 * Otherwise, whine if we are scheduling when we should not be.
4940 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4941 __schedule_bug(prev
);
4943 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4945 schedstat_inc(this_rq(), sched_count
);
4946 #ifdef CONFIG_SCHEDSTATS
4947 if (unlikely(prev
->lock_depth
>= 0)) {
4948 schedstat_inc(this_rq(), bkl_count
);
4949 schedstat_inc(prev
, sched_info
.bkl_count
);
4954 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4956 if (prev
->state
== TASK_RUNNING
) {
4957 u64 runtime
= prev
->se
.sum_exec_runtime
;
4959 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4960 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
4963 * In order to avoid avg_overlap growing stale when we are
4964 * indeed overlapping and hence not getting put to sleep, grow
4965 * the avg_overlap on preemption.
4967 * We use the average preemption runtime because that
4968 * correlates to the amount of cache footprint a task can
4971 update_avg(&prev
->se
.avg_overlap
, runtime
);
4973 prev
->sched_class
->put_prev_task(rq
, prev
);
4977 * Pick up the highest-prio task:
4979 static inline struct task_struct
*
4980 pick_next_task(struct rq
*rq
)
4982 const struct sched_class
*class;
4983 struct task_struct
*p
;
4986 * Optimization: we know that if all tasks are in
4987 * the fair class we can call that function directly:
4989 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4990 p
= fair_sched_class
.pick_next_task(rq
);
4995 class = sched_class_highest
;
4997 p
= class->pick_next_task(rq
);
5001 * Will never be NULL as the idle class always
5002 * returns a non-NULL p:
5004 class = class->next
;
5009 * schedule() is the main scheduler function.
5011 asmlinkage
void __sched
__schedule(void)
5013 struct task_struct
*prev
, *next
;
5014 unsigned long *switch_count
;
5018 cpu
= smp_processor_id();
5022 switch_count
= &prev
->nivcsw
;
5024 release_kernel_lock(prev
);
5025 need_resched_nonpreemptible
:
5027 schedule_debug(prev
);
5029 if (sched_feat(HRTICK
))
5032 spin_lock_irq(&rq
->lock
);
5033 update_rq_clock(rq
);
5034 clear_tsk_need_resched(prev
);
5036 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5037 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5038 prev
->state
= TASK_RUNNING
;
5040 deactivate_task(rq
, prev
, 1);
5041 switch_count
= &prev
->nvcsw
;
5045 if (prev
->sched_class
->pre_schedule
)
5046 prev
->sched_class
->pre_schedule(rq
, prev
);
5049 if (unlikely(!rq
->nr_running
))
5050 idle_balance(cpu
, rq
);
5052 put_prev_task(rq
, prev
);
5053 next
= pick_next_task(rq
);
5055 if (likely(prev
!= next
)) {
5056 sched_info_switch(prev
, next
);
5062 context_switch(rq
, prev
, next
); /* unlocks the rq */
5064 * the context switch might have flipped the stack from under
5065 * us, hence refresh the local variables.
5067 cpu
= smp_processor_id();
5070 spin_unlock_irq(&rq
->lock
);
5072 if (unlikely(reacquire_kernel_lock(current
) < 0))
5073 goto need_resched_nonpreemptible
;
5076 asmlinkage
void __sched
schedule(void)
5081 preempt_enable_no_resched();
5082 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5085 EXPORT_SYMBOL(schedule
);
5089 * Look out! "owner" is an entirely speculative pointer
5090 * access and not reliable.
5092 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5097 if (!sched_feat(OWNER_SPIN
))
5100 #ifdef CONFIG_DEBUG_PAGEALLOC
5102 * Need to access the cpu field knowing that
5103 * DEBUG_PAGEALLOC could have unmapped it if
5104 * the mutex owner just released it and exited.
5106 if (probe_kernel_address(&owner
->cpu
, cpu
))
5113 * Even if the access succeeded (likely case),
5114 * the cpu field may no longer be valid.
5116 if (cpu
>= nr_cpumask_bits
)
5120 * We need to validate that we can do a
5121 * get_cpu() and that we have the percpu area.
5123 if (!cpu_online(cpu
))
5130 * Owner changed, break to re-assess state.
5132 if (lock
->owner
!= owner
)
5136 * Is that owner really running on that cpu?
5138 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5148 #ifdef CONFIG_PREEMPT
5150 * this is the entry point to schedule() from in-kernel preemption
5151 * off of preempt_enable. Kernel preemptions off return from interrupt
5152 * occur there and call schedule directly.
5154 asmlinkage
void __sched
preempt_schedule(void)
5156 struct thread_info
*ti
= current_thread_info();
5159 * If there is a non-zero preempt_count or interrupts are disabled,
5160 * we do not want to preempt the current task. Just return..
5162 if (likely(ti
->preempt_count
|| irqs_disabled()))
5166 add_preempt_count(PREEMPT_ACTIVE
);
5168 sub_preempt_count(PREEMPT_ACTIVE
);
5171 * Check again in case we missed a preemption opportunity
5172 * between schedule and now.
5175 } while (need_resched());
5177 EXPORT_SYMBOL(preempt_schedule
);
5180 * this is the entry point to schedule() from kernel preemption
5181 * off of irq context.
5182 * Note, that this is called and return with irqs disabled. This will
5183 * protect us against recursive calling from irq.
5185 asmlinkage
void __sched
preempt_schedule_irq(void)
5187 struct thread_info
*ti
= current_thread_info();
5189 /* Catch callers which need to be fixed */
5190 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5193 add_preempt_count(PREEMPT_ACTIVE
);
5196 local_irq_disable();
5197 sub_preempt_count(PREEMPT_ACTIVE
);
5200 * Check again in case we missed a preemption opportunity
5201 * between schedule and now.
5204 } while (need_resched());
5207 #endif /* CONFIG_PREEMPT */
5209 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5212 return try_to_wake_up(curr
->private, mode
, sync
);
5214 EXPORT_SYMBOL(default_wake_function
);
5217 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5218 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5219 * number) then we wake all the non-exclusive tasks and one exclusive task.
5221 * There are circumstances in which we can try to wake a task which has already
5222 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5223 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5225 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5226 int nr_exclusive
, int sync
, void *key
)
5228 wait_queue_t
*curr
, *next
;
5230 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5231 unsigned flags
= curr
->flags
;
5233 if (curr
->func(curr
, mode
, sync
, key
) &&
5234 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5240 * __wake_up - wake up threads blocked on a waitqueue.
5242 * @mode: which threads
5243 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5244 * @key: is directly passed to the wakeup function
5246 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5247 int nr_exclusive
, void *key
)
5249 unsigned long flags
;
5251 spin_lock_irqsave(&q
->lock
, flags
);
5252 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5253 spin_unlock_irqrestore(&q
->lock
, flags
);
5255 EXPORT_SYMBOL(__wake_up
);
5258 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5260 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5262 __wake_up_common(q
, mode
, 1, 0, NULL
);
5265 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5267 __wake_up_common(q
, mode
, 1, 0, key
);
5271 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5273 * @mode: which threads
5274 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5275 * @key: opaque value to be passed to wakeup targets
5277 * The sync wakeup differs that the waker knows that it will schedule
5278 * away soon, so while the target thread will be woken up, it will not
5279 * be migrated to another CPU - ie. the two threads are 'synchronized'
5280 * with each other. This can prevent needless bouncing between CPUs.
5282 * On UP it can prevent extra preemption.
5284 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5285 int nr_exclusive
, void *key
)
5287 unsigned long flags
;
5293 if (unlikely(!nr_exclusive
))
5296 spin_lock_irqsave(&q
->lock
, flags
);
5297 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5298 spin_unlock_irqrestore(&q
->lock
, flags
);
5300 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5303 * __wake_up_sync - see __wake_up_sync_key()
5305 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5307 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5309 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5312 * complete: - signals a single thread waiting on this completion
5313 * @x: holds the state of this particular completion
5315 * This will wake up a single thread waiting on this completion. Threads will be
5316 * awakened in the same order in which they were queued.
5318 * See also complete_all(), wait_for_completion() and related routines.
5320 void complete(struct completion
*x
)
5322 unsigned long flags
;
5324 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5326 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5327 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5329 EXPORT_SYMBOL(complete
);
5332 * complete_all: - signals all threads waiting on this completion
5333 * @x: holds the state of this particular completion
5335 * This will wake up all threads waiting on this particular completion event.
5337 void complete_all(struct completion
*x
)
5339 unsigned long flags
;
5341 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5342 x
->done
+= UINT_MAX
/2;
5343 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5344 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5346 EXPORT_SYMBOL(complete_all
);
5348 static inline long __sched
5349 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5352 DECLARE_WAITQUEUE(wait
, current
);
5354 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5355 __add_wait_queue_tail(&x
->wait
, &wait
);
5357 if (signal_pending_state(state
, current
)) {
5358 timeout
= -ERESTARTSYS
;
5361 __set_current_state(state
);
5362 spin_unlock_irq(&x
->wait
.lock
);
5363 timeout
= schedule_timeout(timeout
);
5364 spin_lock_irq(&x
->wait
.lock
);
5365 } while (!x
->done
&& timeout
);
5366 __remove_wait_queue(&x
->wait
, &wait
);
5371 return timeout
?: 1;
5375 wait_for_common(struct completion
*x
, long timeout
, int state
)
5379 spin_lock_irq(&x
->wait
.lock
);
5380 timeout
= do_wait_for_common(x
, timeout
, state
);
5381 spin_unlock_irq(&x
->wait
.lock
);
5386 * wait_for_completion: - waits for completion of a task
5387 * @x: holds the state of this particular completion
5389 * This waits to be signaled for completion of a specific task. It is NOT
5390 * interruptible and there is no timeout.
5392 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5393 * and interrupt capability. Also see complete().
5395 void __sched
wait_for_completion(struct completion
*x
)
5397 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5399 EXPORT_SYMBOL(wait_for_completion
);
5402 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5403 * @x: holds the state of this particular completion
5404 * @timeout: timeout value in jiffies
5406 * This waits for either a completion of a specific task to be signaled or for a
5407 * specified timeout to expire. The timeout is in jiffies. It is not
5410 unsigned long __sched
5411 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5413 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5415 EXPORT_SYMBOL(wait_for_completion_timeout
);
5418 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5419 * @x: holds the state of this particular completion
5421 * This waits for completion of a specific task to be signaled. It is
5424 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5426 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5427 if (t
== -ERESTARTSYS
)
5431 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5434 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5435 * @x: holds the state of this particular completion
5436 * @timeout: timeout value in jiffies
5438 * This waits for either a completion of a specific task to be signaled or for a
5439 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5441 unsigned long __sched
5442 wait_for_completion_interruptible_timeout(struct completion
*x
,
5443 unsigned long timeout
)
5445 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5447 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5450 * wait_for_completion_killable: - waits for completion of a task (killable)
5451 * @x: holds the state of this particular completion
5453 * This waits to be signaled for completion of a specific task. It can be
5454 * interrupted by a kill signal.
5456 int __sched
wait_for_completion_killable(struct completion
*x
)
5458 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5459 if (t
== -ERESTARTSYS
)
5463 EXPORT_SYMBOL(wait_for_completion_killable
);
5466 * try_wait_for_completion - try to decrement a completion without blocking
5467 * @x: completion structure
5469 * Returns: 0 if a decrement cannot be done without blocking
5470 * 1 if a decrement succeeded.
5472 * If a completion is being used as a counting completion,
5473 * attempt to decrement the counter without blocking. This
5474 * enables us to avoid waiting if the resource the completion
5475 * is protecting is not available.
5477 bool try_wait_for_completion(struct completion
*x
)
5481 spin_lock_irq(&x
->wait
.lock
);
5486 spin_unlock_irq(&x
->wait
.lock
);
5489 EXPORT_SYMBOL(try_wait_for_completion
);
5492 * completion_done - Test to see if a completion has any waiters
5493 * @x: completion structure
5495 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5496 * 1 if there are no waiters.
5499 bool completion_done(struct completion
*x
)
5503 spin_lock_irq(&x
->wait
.lock
);
5506 spin_unlock_irq(&x
->wait
.lock
);
5509 EXPORT_SYMBOL(completion_done
);
5512 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5514 unsigned long flags
;
5517 init_waitqueue_entry(&wait
, current
);
5519 __set_current_state(state
);
5521 spin_lock_irqsave(&q
->lock
, flags
);
5522 __add_wait_queue(q
, &wait
);
5523 spin_unlock(&q
->lock
);
5524 timeout
= schedule_timeout(timeout
);
5525 spin_lock_irq(&q
->lock
);
5526 __remove_wait_queue(q
, &wait
);
5527 spin_unlock_irqrestore(&q
->lock
, flags
);
5532 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5534 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5536 EXPORT_SYMBOL(interruptible_sleep_on
);
5539 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5541 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5543 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5545 void __sched
sleep_on(wait_queue_head_t
*q
)
5547 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5549 EXPORT_SYMBOL(sleep_on
);
5551 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5553 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5555 EXPORT_SYMBOL(sleep_on_timeout
);
5557 #ifdef CONFIG_RT_MUTEXES
5560 * rt_mutex_setprio - set the current priority of a task
5562 * @prio: prio value (kernel-internal form)
5564 * This function changes the 'effective' priority of a task. It does
5565 * not touch ->normal_prio like __setscheduler().
5567 * Used by the rt_mutex code to implement priority inheritance logic.
5569 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5571 unsigned long flags
;
5572 int oldprio
, on_rq
, running
;
5574 const struct sched_class
*prev_class
= p
->sched_class
;
5576 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5578 rq
= task_rq_lock(p
, &flags
);
5579 update_rq_clock(rq
);
5582 on_rq
= p
->se
.on_rq
;
5583 running
= task_current(rq
, p
);
5585 dequeue_task(rq
, p
, 0);
5587 p
->sched_class
->put_prev_task(rq
, p
);
5590 p
->sched_class
= &rt_sched_class
;
5592 p
->sched_class
= &fair_sched_class
;
5597 p
->sched_class
->set_curr_task(rq
);
5599 enqueue_task(rq
, p
, 0);
5601 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5603 task_rq_unlock(rq
, &flags
);
5608 void set_user_nice(struct task_struct
*p
, long nice
)
5610 int old_prio
, delta
, on_rq
;
5611 unsigned long flags
;
5614 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5617 * We have to be careful, if called from sys_setpriority(),
5618 * the task might be in the middle of scheduling on another CPU.
5620 rq
= task_rq_lock(p
, &flags
);
5621 update_rq_clock(rq
);
5623 * The RT priorities are set via sched_setscheduler(), but we still
5624 * allow the 'normal' nice value to be set - but as expected
5625 * it wont have any effect on scheduling until the task is
5626 * SCHED_FIFO/SCHED_RR:
5628 if (task_has_rt_policy(p
)) {
5629 p
->static_prio
= NICE_TO_PRIO(nice
);
5632 on_rq
= p
->se
.on_rq
;
5634 dequeue_task(rq
, p
, 0);
5636 p
->static_prio
= NICE_TO_PRIO(nice
);
5639 p
->prio
= effective_prio(p
);
5640 delta
= p
->prio
- old_prio
;
5643 enqueue_task(rq
, p
, 0);
5645 * If the task increased its priority or is running and
5646 * lowered its priority, then reschedule its CPU:
5648 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5649 resched_task(rq
->curr
);
5652 task_rq_unlock(rq
, &flags
);
5654 EXPORT_SYMBOL(set_user_nice
);
5657 * can_nice - check if a task can reduce its nice value
5661 int can_nice(const struct task_struct
*p
, const int nice
)
5663 /* convert nice value [19,-20] to rlimit style value [1,40] */
5664 int nice_rlim
= 20 - nice
;
5666 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5667 capable(CAP_SYS_NICE
));
5670 #ifdef __ARCH_WANT_SYS_NICE
5673 * sys_nice - change the priority of the current process.
5674 * @increment: priority increment
5676 * sys_setpriority is a more generic, but much slower function that
5677 * does similar things.
5679 SYSCALL_DEFINE1(nice
, int, increment
)
5684 * Setpriority might change our priority at the same moment.
5685 * We don't have to worry. Conceptually one call occurs first
5686 * and we have a single winner.
5688 if (increment
< -40)
5693 nice
= TASK_NICE(current
) + increment
;
5699 if (increment
< 0 && !can_nice(current
, nice
))
5702 retval
= security_task_setnice(current
, nice
);
5706 set_user_nice(current
, nice
);
5713 * task_prio - return the priority value of a given task.
5714 * @p: the task in question.
5716 * This is the priority value as seen by users in /proc.
5717 * RT tasks are offset by -200. Normal tasks are centered
5718 * around 0, value goes from -16 to +15.
5720 int task_prio(const struct task_struct
*p
)
5722 return p
->prio
- MAX_RT_PRIO
;
5726 * task_nice - return the nice value of a given task.
5727 * @p: the task in question.
5729 int task_nice(const struct task_struct
*p
)
5731 return TASK_NICE(p
);
5733 EXPORT_SYMBOL(task_nice
);
5736 * idle_cpu - is a given cpu idle currently?
5737 * @cpu: the processor in question.
5739 int idle_cpu(int cpu
)
5741 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5745 * idle_task - return the idle task for a given cpu.
5746 * @cpu: the processor in question.
5748 struct task_struct
*idle_task(int cpu
)
5750 return cpu_rq(cpu
)->idle
;
5754 * find_process_by_pid - find a process with a matching PID value.
5755 * @pid: the pid in question.
5757 static struct task_struct
*find_process_by_pid(pid_t pid
)
5759 return pid
? find_task_by_vpid(pid
) : current
;
5762 /* Actually do priority change: must hold rq lock. */
5764 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5766 BUG_ON(p
->se
.on_rq
);
5769 switch (p
->policy
) {
5773 p
->sched_class
= &fair_sched_class
;
5777 p
->sched_class
= &rt_sched_class
;
5781 p
->rt_priority
= prio
;
5782 p
->normal_prio
= normal_prio(p
);
5783 /* we are holding p->pi_lock already */
5784 p
->prio
= rt_mutex_getprio(p
);
5789 * check the target process has a UID that matches the current process's
5791 static bool check_same_owner(struct task_struct
*p
)
5793 const struct cred
*cred
= current_cred(), *pcred
;
5797 pcred
= __task_cred(p
);
5798 match
= (cred
->euid
== pcred
->euid
||
5799 cred
->euid
== pcred
->uid
);
5804 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5805 struct sched_param
*param
, bool user
)
5807 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5808 unsigned long flags
;
5809 const struct sched_class
*prev_class
= p
->sched_class
;
5812 /* may grab non-irq protected spin_locks */
5813 BUG_ON(in_interrupt());
5815 /* double check policy once rq lock held */
5817 policy
= oldpolicy
= p
->policy
;
5818 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5819 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5820 policy
!= SCHED_IDLE
)
5823 * Valid priorities for SCHED_FIFO and SCHED_RR are
5824 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5825 * SCHED_BATCH and SCHED_IDLE is 0.
5827 if (param
->sched_priority
< 0 ||
5828 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5829 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5831 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5835 * Allow unprivileged RT tasks to decrease priority:
5837 if (user
&& !capable(CAP_SYS_NICE
)) {
5838 if (rt_policy(policy
)) {
5839 unsigned long rlim_rtprio
;
5841 if (!lock_task_sighand(p
, &flags
))
5843 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5844 unlock_task_sighand(p
, &flags
);
5846 /* can't set/change the rt policy */
5847 if (policy
!= p
->policy
&& !rlim_rtprio
)
5850 /* can't increase priority */
5851 if (param
->sched_priority
> p
->rt_priority
&&
5852 param
->sched_priority
> rlim_rtprio
)
5856 * Like positive nice levels, dont allow tasks to
5857 * move out of SCHED_IDLE either:
5859 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5862 /* can't change other user's priorities */
5863 if (!check_same_owner(p
))
5868 #ifdef CONFIG_RT_GROUP_SCHED
5870 * Do not allow realtime tasks into groups that have no runtime
5873 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5874 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5878 retval
= security_task_setscheduler(p
, policy
, param
);
5884 * make sure no PI-waiters arrive (or leave) while we are
5885 * changing the priority of the task:
5887 spin_lock_irqsave(&p
->pi_lock
, flags
);
5889 * To be able to change p->policy safely, the apropriate
5890 * runqueue lock must be held.
5892 rq
= __task_rq_lock(p
);
5893 /* recheck policy now with rq lock held */
5894 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5895 policy
= oldpolicy
= -1;
5896 __task_rq_unlock(rq
);
5897 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5900 update_rq_clock(rq
);
5901 on_rq
= p
->se
.on_rq
;
5902 running
= task_current(rq
, p
);
5904 deactivate_task(rq
, p
, 0);
5906 p
->sched_class
->put_prev_task(rq
, p
);
5909 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5912 p
->sched_class
->set_curr_task(rq
);
5914 activate_task(rq
, p
, 0);
5916 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5918 __task_rq_unlock(rq
);
5919 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5921 rt_mutex_adjust_pi(p
);
5927 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5928 * @p: the task in question.
5929 * @policy: new policy.
5930 * @param: structure containing the new RT priority.
5932 * NOTE that the task may be already dead.
5934 int sched_setscheduler(struct task_struct
*p
, int policy
,
5935 struct sched_param
*param
)
5937 return __sched_setscheduler(p
, policy
, param
, true);
5939 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5942 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5943 * @p: the task in question.
5944 * @policy: new policy.
5945 * @param: structure containing the new RT priority.
5947 * Just like sched_setscheduler, only don't bother checking if the
5948 * current context has permission. For example, this is needed in
5949 * stop_machine(): we create temporary high priority worker threads,
5950 * but our caller might not have that capability.
5952 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5953 struct sched_param
*param
)
5955 return __sched_setscheduler(p
, policy
, param
, false);
5959 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5961 struct sched_param lparam
;
5962 struct task_struct
*p
;
5965 if (!param
|| pid
< 0)
5967 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5972 p
= find_process_by_pid(pid
);
5974 retval
= sched_setscheduler(p
, policy
, &lparam
);
5981 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5982 * @pid: the pid in question.
5983 * @policy: new policy.
5984 * @param: structure containing the new RT priority.
5986 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
5987 struct sched_param __user
*, param
)
5989 /* negative values for policy are not valid */
5993 return do_sched_setscheduler(pid
, policy
, param
);
5997 * sys_sched_setparam - set/change the RT priority of a thread
5998 * @pid: the pid in question.
5999 * @param: structure containing the new RT priority.
6001 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6003 return do_sched_setscheduler(pid
, -1, param
);
6007 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6008 * @pid: the pid in question.
6010 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6012 struct task_struct
*p
;
6019 read_lock(&tasklist_lock
);
6020 p
= find_process_by_pid(pid
);
6022 retval
= security_task_getscheduler(p
);
6026 read_unlock(&tasklist_lock
);
6031 * sys_sched_getscheduler - get the RT priority of a thread
6032 * @pid: the pid in question.
6033 * @param: structure containing the RT priority.
6035 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6037 struct sched_param lp
;
6038 struct task_struct
*p
;
6041 if (!param
|| pid
< 0)
6044 read_lock(&tasklist_lock
);
6045 p
= find_process_by_pid(pid
);
6050 retval
= security_task_getscheduler(p
);
6054 lp
.sched_priority
= p
->rt_priority
;
6055 read_unlock(&tasklist_lock
);
6058 * This one might sleep, we cannot do it with a spinlock held ...
6060 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6065 read_unlock(&tasklist_lock
);
6069 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6071 cpumask_var_t cpus_allowed
, new_mask
;
6072 struct task_struct
*p
;
6076 read_lock(&tasklist_lock
);
6078 p
= find_process_by_pid(pid
);
6080 read_unlock(&tasklist_lock
);
6086 * It is not safe to call set_cpus_allowed with the
6087 * tasklist_lock held. We will bump the task_struct's
6088 * usage count and then drop tasklist_lock.
6091 read_unlock(&tasklist_lock
);
6093 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6097 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6099 goto out_free_cpus_allowed
;
6102 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6105 retval
= security_task_setscheduler(p
, 0, NULL
);
6109 cpuset_cpus_allowed(p
, cpus_allowed
);
6110 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6112 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6115 cpuset_cpus_allowed(p
, cpus_allowed
);
6116 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6118 * We must have raced with a concurrent cpuset
6119 * update. Just reset the cpus_allowed to the
6120 * cpuset's cpus_allowed
6122 cpumask_copy(new_mask
, cpus_allowed
);
6127 free_cpumask_var(new_mask
);
6128 out_free_cpus_allowed
:
6129 free_cpumask_var(cpus_allowed
);
6136 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6137 struct cpumask
*new_mask
)
6139 if (len
< cpumask_size())
6140 cpumask_clear(new_mask
);
6141 else if (len
> cpumask_size())
6142 len
= cpumask_size();
6144 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6148 * sys_sched_setaffinity - set the cpu affinity of a process
6149 * @pid: pid of the process
6150 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6151 * @user_mask_ptr: user-space pointer to the new cpu mask
6153 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6154 unsigned long __user
*, user_mask_ptr
)
6156 cpumask_var_t new_mask
;
6159 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6162 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6164 retval
= sched_setaffinity(pid
, new_mask
);
6165 free_cpumask_var(new_mask
);
6169 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6171 struct task_struct
*p
;
6175 read_lock(&tasklist_lock
);
6178 p
= find_process_by_pid(pid
);
6182 retval
= security_task_getscheduler(p
);
6186 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6189 read_unlock(&tasklist_lock
);
6196 * sys_sched_getaffinity - get the cpu affinity of a process
6197 * @pid: pid of the process
6198 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6199 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6201 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6202 unsigned long __user
*, user_mask_ptr
)
6207 if (len
< cpumask_size())
6210 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6213 ret
= sched_getaffinity(pid
, mask
);
6215 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6218 ret
= cpumask_size();
6220 free_cpumask_var(mask
);
6226 * sys_sched_yield - yield the current processor to other threads.
6228 * This function yields the current CPU to other tasks. If there are no
6229 * other threads running on this CPU then this function will return.
6231 SYSCALL_DEFINE0(sched_yield
)
6233 struct rq
*rq
= this_rq_lock();
6235 schedstat_inc(rq
, yld_count
);
6236 current
->sched_class
->yield_task(rq
);
6239 * Since we are going to call schedule() anyway, there's
6240 * no need to preempt or enable interrupts:
6242 __release(rq
->lock
);
6243 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6244 _raw_spin_unlock(&rq
->lock
);
6245 preempt_enable_no_resched();
6252 static void __cond_resched(void)
6254 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6255 __might_sleep(__FILE__
, __LINE__
);
6258 * The BKS might be reacquired before we have dropped
6259 * PREEMPT_ACTIVE, which could trigger a second
6260 * cond_resched() call.
6263 add_preempt_count(PREEMPT_ACTIVE
);
6265 sub_preempt_count(PREEMPT_ACTIVE
);
6266 } while (need_resched());
6269 int __sched
_cond_resched(void)
6271 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6272 system_state
== SYSTEM_RUNNING
) {
6278 EXPORT_SYMBOL(_cond_resched
);
6281 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6282 * call schedule, and on return reacquire the lock.
6284 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6285 * operations here to prevent schedule() from being called twice (once via
6286 * spin_unlock(), once by hand).
6288 int cond_resched_lock(spinlock_t
*lock
)
6290 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6293 if (spin_needbreak(lock
) || resched
) {
6295 if (resched
&& need_resched())
6304 EXPORT_SYMBOL(cond_resched_lock
);
6306 int __sched
cond_resched_softirq(void)
6308 BUG_ON(!in_softirq());
6310 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6318 EXPORT_SYMBOL(cond_resched_softirq
);
6321 * yield - yield the current processor to other threads.
6323 * This is a shortcut for kernel-space yielding - it marks the
6324 * thread runnable and calls sys_sched_yield().
6326 void __sched
yield(void)
6328 set_current_state(TASK_RUNNING
);
6331 EXPORT_SYMBOL(yield
);
6334 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6335 * that process accounting knows that this is a task in IO wait state.
6337 * But don't do that if it is a deliberate, throttling IO wait (this task
6338 * has set its backing_dev_info: the queue against which it should throttle)
6340 void __sched
io_schedule(void)
6342 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6344 delayacct_blkio_start();
6345 atomic_inc(&rq
->nr_iowait
);
6347 atomic_dec(&rq
->nr_iowait
);
6348 delayacct_blkio_end();
6350 EXPORT_SYMBOL(io_schedule
);
6352 long __sched
io_schedule_timeout(long timeout
)
6354 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6357 delayacct_blkio_start();
6358 atomic_inc(&rq
->nr_iowait
);
6359 ret
= schedule_timeout(timeout
);
6360 atomic_dec(&rq
->nr_iowait
);
6361 delayacct_blkio_end();
6366 * sys_sched_get_priority_max - return maximum RT priority.
6367 * @policy: scheduling class.
6369 * this syscall returns the maximum rt_priority that can be used
6370 * by a given scheduling class.
6372 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6379 ret
= MAX_USER_RT_PRIO
-1;
6391 * sys_sched_get_priority_min - return minimum RT priority.
6392 * @policy: scheduling class.
6394 * this syscall returns the minimum rt_priority that can be used
6395 * by a given scheduling class.
6397 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6415 * sys_sched_rr_get_interval - return the default timeslice of a process.
6416 * @pid: pid of the process.
6417 * @interval: userspace pointer to the timeslice value.
6419 * this syscall writes the default timeslice value of a given process
6420 * into the user-space timespec buffer. A value of '0' means infinity.
6422 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6423 struct timespec __user
*, interval
)
6425 struct task_struct
*p
;
6426 unsigned int time_slice
;
6434 read_lock(&tasklist_lock
);
6435 p
= find_process_by_pid(pid
);
6439 retval
= security_task_getscheduler(p
);
6444 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6445 * tasks that are on an otherwise idle runqueue:
6448 if (p
->policy
== SCHED_RR
) {
6449 time_slice
= DEF_TIMESLICE
;
6450 } else if (p
->policy
!= SCHED_FIFO
) {
6451 struct sched_entity
*se
= &p
->se
;
6452 unsigned long flags
;
6455 rq
= task_rq_lock(p
, &flags
);
6456 if (rq
->cfs
.load
.weight
)
6457 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6458 task_rq_unlock(rq
, &flags
);
6460 read_unlock(&tasklist_lock
);
6461 jiffies_to_timespec(time_slice
, &t
);
6462 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6466 read_unlock(&tasklist_lock
);
6470 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6472 void sched_show_task(struct task_struct
*p
)
6474 unsigned long free
= 0;
6477 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6478 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6479 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6480 #if BITS_PER_LONG == 32
6481 if (state
== TASK_RUNNING
)
6482 printk(KERN_CONT
" running ");
6484 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6486 if (state
== TASK_RUNNING
)
6487 printk(KERN_CONT
" running task ");
6489 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6491 #ifdef CONFIG_DEBUG_STACK_USAGE
6492 free
= stack_not_used(p
);
6494 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6495 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6497 show_stack(p
, NULL
);
6500 void show_state_filter(unsigned long state_filter
)
6502 struct task_struct
*g
, *p
;
6504 #if BITS_PER_LONG == 32
6506 " task PC stack pid father\n");
6509 " task PC stack pid father\n");
6511 read_lock(&tasklist_lock
);
6512 do_each_thread(g
, p
) {
6514 * reset the NMI-timeout, listing all files on a slow
6515 * console might take alot of time:
6517 touch_nmi_watchdog();
6518 if (!state_filter
|| (p
->state
& state_filter
))
6520 } while_each_thread(g
, p
);
6522 touch_all_softlockup_watchdogs();
6524 #ifdef CONFIG_SCHED_DEBUG
6525 sysrq_sched_debug_show();
6527 read_unlock(&tasklist_lock
);
6529 * Only show locks if all tasks are dumped:
6531 if (state_filter
== -1)
6532 debug_show_all_locks();
6535 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6537 idle
->sched_class
= &idle_sched_class
;
6541 * init_idle - set up an idle thread for a given CPU
6542 * @idle: task in question
6543 * @cpu: cpu the idle task belongs to
6545 * NOTE: this function does not set the idle thread's NEED_RESCHED
6546 * flag, to make booting more robust.
6548 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6550 struct rq
*rq
= cpu_rq(cpu
);
6551 unsigned long flags
;
6553 spin_lock_irqsave(&rq
->lock
, flags
);
6556 idle
->se
.exec_start
= sched_clock();
6558 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6559 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6560 __set_task_cpu(idle
, cpu
);
6562 rq
->curr
= rq
->idle
= idle
;
6563 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6566 spin_unlock_irqrestore(&rq
->lock
, flags
);
6568 /* Set the preempt count _outside_ the spinlocks! */
6569 #if defined(CONFIG_PREEMPT)
6570 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6572 task_thread_info(idle
)->preempt_count
= 0;
6575 * The idle tasks have their own, simple scheduling class:
6577 idle
->sched_class
= &idle_sched_class
;
6578 ftrace_graph_init_task(idle
);
6582 * In a system that switches off the HZ timer nohz_cpu_mask
6583 * indicates which cpus entered this state. This is used
6584 * in the rcu update to wait only for active cpus. For system
6585 * which do not switch off the HZ timer nohz_cpu_mask should
6586 * always be CPU_BITS_NONE.
6588 cpumask_var_t nohz_cpu_mask
;
6591 * Increase the granularity value when there are more CPUs,
6592 * because with more CPUs the 'effective latency' as visible
6593 * to users decreases. But the relationship is not linear,
6594 * so pick a second-best guess by going with the log2 of the
6597 * This idea comes from the SD scheduler of Con Kolivas:
6599 static inline void sched_init_granularity(void)
6601 unsigned int factor
= 1 + ilog2(num_online_cpus());
6602 const unsigned long limit
= 200000000;
6604 sysctl_sched_min_granularity
*= factor
;
6605 if (sysctl_sched_min_granularity
> limit
)
6606 sysctl_sched_min_granularity
= limit
;
6608 sysctl_sched_latency
*= factor
;
6609 if (sysctl_sched_latency
> limit
)
6610 sysctl_sched_latency
= limit
;
6612 sysctl_sched_wakeup_granularity
*= factor
;
6614 sysctl_sched_shares_ratelimit
*= factor
;
6619 * This is how migration works:
6621 * 1) we queue a struct migration_req structure in the source CPU's
6622 * runqueue and wake up that CPU's migration thread.
6623 * 2) we down() the locked semaphore => thread blocks.
6624 * 3) migration thread wakes up (implicitly it forces the migrated
6625 * thread off the CPU)
6626 * 4) it gets the migration request and checks whether the migrated
6627 * task is still in the wrong runqueue.
6628 * 5) if it's in the wrong runqueue then the migration thread removes
6629 * it and puts it into the right queue.
6630 * 6) migration thread up()s the semaphore.
6631 * 7) we wake up and the migration is done.
6635 * Change a given task's CPU affinity. Migrate the thread to a
6636 * proper CPU and schedule it away if the CPU it's executing on
6637 * is removed from the allowed bitmask.
6639 * NOTE: the caller must have a valid reference to the task, the
6640 * task must not exit() & deallocate itself prematurely. The
6641 * call is not atomic; no spinlocks may be held.
6643 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6645 struct migration_req req
;
6646 unsigned long flags
;
6650 rq
= task_rq_lock(p
, &flags
);
6651 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6656 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6657 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6662 if (p
->sched_class
->set_cpus_allowed
)
6663 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6665 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6666 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6669 /* Can the task run on the task's current CPU? If so, we're done */
6670 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6673 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6674 /* Need help from migration thread: drop lock and wait. */
6675 task_rq_unlock(rq
, &flags
);
6676 wake_up_process(rq
->migration_thread
);
6677 wait_for_completion(&req
.done
);
6678 tlb_migrate_finish(p
->mm
);
6682 task_rq_unlock(rq
, &flags
);
6686 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6689 * Move (not current) task off this cpu, onto dest cpu. We're doing
6690 * this because either it can't run here any more (set_cpus_allowed()
6691 * away from this CPU, or CPU going down), or because we're
6692 * attempting to rebalance this task on exec (sched_exec).
6694 * So we race with normal scheduler movements, but that's OK, as long
6695 * as the task is no longer on this CPU.
6697 * Returns non-zero if task was successfully migrated.
6699 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6701 struct rq
*rq_dest
, *rq_src
;
6704 if (unlikely(!cpu_active(dest_cpu
)))
6707 rq_src
= cpu_rq(src_cpu
);
6708 rq_dest
= cpu_rq(dest_cpu
);
6710 double_rq_lock(rq_src
, rq_dest
);
6711 /* Already moved. */
6712 if (task_cpu(p
) != src_cpu
)
6714 /* Affinity changed (again). */
6715 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6718 on_rq
= p
->se
.on_rq
;
6720 deactivate_task(rq_src
, p
, 0);
6722 set_task_cpu(p
, dest_cpu
);
6724 activate_task(rq_dest
, p
, 0);
6725 check_preempt_curr(rq_dest
, p
, 0);
6730 double_rq_unlock(rq_src
, rq_dest
);
6735 * migration_thread - this is a highprio system thread that performs
6736 * thread migration by bumping thread off CPU then 'pushing' onto
6739 static int migration_thread(void *data
)
6741 int cpu
= (long)data
;
6745 BUG_ON(rq
->migration_thread
!= current
);
6747 set_current_state(TASK_INTERRUPTIBLE
);
6748 while (!kthread_should_stop()) {
6749 struct migration_req
*req
;
6750 struct list_head
*head
;
6752 spin_lock_irq(&rq
->lock
);
6754 if (cpu_is_offline(cpu
)) {
6755 spin_unlock_irq(&rq
->lock
);
6759 if (rq
->active_balance
) {
6760 active_load_balance(rq
, cpu
);
6761 rq
->active_balance
= 0;
6764 head
= &rq
->migration_queue
;
6766 if (list_empty(head
)) {
6767 spin_unlock_irq(&rq
->lock
);
6769 set_current_state(TASK_INTERRUPTIBLE
);
6772 req
= list_entry(head
->next
, struct migration_req
, list
);
6773 list_del_init(head
->next
);
6775 spin_unlock(&rq
->lock
);
6776 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6779 complete(&req
->done
);
6781 __set_current_state(TASK_RUNNING
);
6785 /* Wait for kthread_stop */
6786 set_current_state(TASK_INTERRUPTIBLE
);
6787 while (!kthread_should_stop()) {
6789 set_current_state(TASK_INTERRUPTIBLE
);
6791 __set_current_state(TASK_RUNNING
);
6795 #ifdef CONFIG_HOTPLUG_CPU
6797 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6801 local_irq_disable();
6802 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6808 * Figure out where task on dead CPU should go, use force if necessary.
6810 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6813 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6816 /* Look for allowed, online CPU in same node. */
6817 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6818 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6821 /* Any allowed, online CPU? */
6822 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6823 if (dest_cpu
< nr_cpu_ids
)
6826 /* No more Mr. Nice Guy. */
6827 if (dest_cpu
>= nr_cpu_ids
) {
6828 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6829 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6832 * Don't tell them about moving exiting tasks or
6833 * kernel threads (both mm NULL), since they never
6836 if (p
->mm
&& printk_ratelimit()) {
6837 printk(KERN_INFO
"process %d (%s) no "
6838 "longer affine to cpu%d\n",
6839 task_pid_nr(p
), p
->comm
, dead_cpu
);
6844 /* It can have affinity changed while we were choosing. */
6845 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6850 * While a dead CPU has no uninterruptible tasks queued at this point,
6851 * it might still have a nonzero ->nr_uninterruptible counter, because
6852 * for performance reasons the counter is not stricly tracking tasks to
6853 * their home CPUs. So we just add the counter to another CPU's counter,
6854 * to keep the global sum constant after CPU-down:
6856 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6858 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6859 unsigned long flags
;
6861 local_irq_save(flags
);
6862 double_rq_lock(rq_src
, rq_dest
);
6863 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6864 rq_src
->nr_uninterruptible
= 0;
6865 double_rq_unlock(rq_src
, rq_dest
);
6866 local_irq_restore(flags
);
6869 /* Run through task list and migrate tasks from the dead cpu. */
6870 static void migrate_live_tasks(int src_cpu
)
6872 struct task_struct
*p
, *t
;
6874 read_lock(&tasklist_lock
);
6876 do_each_thread(t
, p
) {
6880 if (task_cpu(p
) == src_cpu
)
6881 move_task_off_dead_cpu(src_cpu
, p
);
6882 } while_each_thread(t
, p
);
6884 read_unlock(&tasklist_lock
);
6888 * Schedules idle task to be the next runnable task on current CPU.
6889 * It does so by boosting its priority to highest possible.
6890 * Used by CPU offline code.
6892 void sched_idle_next(void)
6894 int this_cpu
= smp_processor_id();
6895 struct rq
*rq
= cpu_rq(this_cpu
);
6896 struct task_struct
*p
= rq
->idle
;
6897 unsigned long flags
;
6899 /* cpu has to be offline */
6900 BUG_ON(cpu_online(this_cpu
));
6903 * Strictly not necessary since rest of the CPUs are stopped by now
6904 * and interrupts disabled on the current cpu.
6906 spin_lock_irqsave(&rq
->lock
, flags
);
6908 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6910 update_rq_clock(rq
);
6911 activate_task(rq
, p
, 0);
6913 spin_unlock_irqrestore(&rq
->lock
, flags
);
6917 * Ensures that the idle task is using init_mm right before its cpu goes
6920 void idle_task_exit(void)
6922 struct mm_struct
*mm
= current
->active_mm
;
6924 BUG_ON(cpu_online(smp_processor_id()));
6927 switch_mm(mm
, &init_mm
, current
);
6931 /* called under rq->lock with disabled interrupts */
6932 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6934 struct rq
*rq
= cpu_rq(dead_cpu
);
6936 /* Must be exiting, otherwise would be on tasklist. */
6937 BUG_ON(!p
->exit_state
);
6939 /* Cannot have done final schedule yet: would have vanished. */
6940 BUG_ON(p
->state
== TASK_DEAD
);
6945 * Drop lock around migration; if someone else moves it,
6946 * that's OK. No task can be added to this CPU, so iteration is
6949 spin_unlock_irq(&rq
->lock
);
6950 move_task_off_dead_cpu(dead_cpu
, p
);
6951 spin_lock_irq(&rq
->lock
);
6956 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6957 static void migrate_dead_tasks(unsigned int dead_cpu
)
6959 struct rq
*rq
= cpu_rq(dead_cpu
);
6960 struct task_struct
*next
;
6963 if (!rq
->nr_running
)
6965 update_rq_clock(rq
);
6966 next
= pick_next_task(rq
);
6969 next
->sched_class
->put_prev_task(rq
, next
);
6970 migrate_dead(dead_cpu
, next
);
6974 #endif /* CONFIG_HOTPLUG_CPU */
6976 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6978 static struct ctl_table sd_ctl_dir
[] = {
6980 .procname
= "sched_domain",
6986 static struct ctl_table sd_ctl_root
[] = {
6988 .ctl_name
= CTL_KERN
,
6989 .procname
= "kernel",
6991 .child
= sd_ctl_dir
,
6996 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6998 struct ctl_table
*entry
=
6999 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7004 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7006 struct ctl_table
*entry
;
7009 * In the intermediate directories, both the child directory and
7010 * procname are dynamically allocated and could fail but the mode
7011 * will always be set. In the lowest directory the names are
7012 * static strings and all have proc handlers.
7014 for (entry
= *tablep
; entry
->mode
; entry
++) {
7016 sd_free_ctl_entry(&entry
->child
);
7017 if (entry
->proc_handler
== NULL
)
7018 kfree(entry
->procname
);
7026 set_table_entry(struct ctl_table
*entry
,
7027 const char *procname
, void *data
, int maxlen
,
7028 mode_t mode
, proc_handler
*proc_handler
)
7030 entry
->procname
= procname
;
7032 entry
->maxlen
= maxlen
;
7034 entry
->proc_handler
= proc_handler
;
7037 static struct ctl_table
*
7038 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7040 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7045 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7046 sizeof(long), 0644, proc_doulongvec_minmax
);
7047 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7048 sizeof(long), 0644, proc_doulongvec_minmax
);
7049 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7050 sizeof(int), 0644, proc_dointvec_minmax
);
7051 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7052 sizeof(int), 0644, proc_dointvec_minmax
);
7053 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7054 sizeof(int), 0644, proc_dointvec_minmax
);
7055 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7056 sizeof(int), 0644, proc_dointvec_minmax
);
7057 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7058 sizeof(int), 0644, proc_dointvec_minmax
);
7059 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7060 sizeof(int), 0644, proc_dointvec_minmax
);
7061 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7062 sizeof(int), 0644, proc_dointvec_minmax
);
7063 set_table_entry(&table
[9], "cache_nice_tries",
7064 &sd
->cache_nice_tries
,
7065 sizeof(int), 0644, proc_dointvec_minmax
);
7066 set_table_entry(&table
[10], "flags", &sd
->flags
,
7067 sizeof(int), 0644, proc_dointvec_minmax
);
7068 set_table_entry(&table
[11], "name", sd
->name
,
7069 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7070 /* &table[12] is terminator */
7075 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7077 struct ctl_table
*entry
, *table
;
7078 struct sched_domain
*sd
;
7079 int domain_num
= 0, i
;
7082 for_each_domain(cpu
, sd
)
7084 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7089 for_each_domain(cpu
, sd
) {
7090 snprintf(buf
, 32, "domain%d", i
);
7091 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7093 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7100 static struct ctl_table_header
*sd_sysctl_header
;
7101 static void register_sched_domain_sysctl(void)
7103 int i
, cpu_num
= num_online_cpus();
7104 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7107 WARN_ON(sd_ctl_dir
[0].child
);
7108 sd_ctl_dir
[0].child
= entry
;
7113 for_each_online_cpu(i
) {
7114 snprintf(buf
, 32, "cpu%d", i
);
7115 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7117 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7121 WARN_ON(sd_sysctl_header
);
7122 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7125 /* may be called multiple times per register */
7126 static void unregister_sched_domain_sysctl(void)
7128 if (sd_sysctl_header
)
7129 unregister_sysctl_table(sd_sysctl_header
);
7130 sd_sysctl_header
= NULL
;
7131 if (sd_ctl_dir
[0].child
)
7132 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7135 static void register_sched_domain_sysctl(void)
7138 static void unregister_sched_domain_sysctl(void)
7143 static void set_rq_online(struct rq
*rq
)
7146 const struct sched_class
*class;
7148 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7151 for_each_class(class) {
7152 if (class->rq_online
)
7153 class->rq_online(rq
);
7158 static void set_rq_offline(struct rq
*rq
)
7161 const struct sched_class
*class;
7163 for_each_class(class) {
7164 if (class->rq_offline
)
7165 class->rq_offline(rq
);
7168 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7174 * migration_call - callback that gets triggered when a CPU is added.
7175 * Here we can start up the necessary migration thread for the new CPU.
7177 static int __cpuinit
7178 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7180 struct task_struct
*p
;
7181 int cpu
= (long)hcpu
;
7182 unsigned long flags
;
7187 case CPU_UP_PREPARE
:
7188 case CPU_UP_PREPARE_FROZEN
:
7189 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7192 kthread_bind(p
, cpu
);
7193 /* Must be high prio: stop_machine expects to yield to it. */
7194 rq
= task_rq_lock(p
, &flags
);
7195 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7196 task_rq_unlock(rq
, &flags
);
7197 cpu_rq(cpu
)->migration_thread
= p
;
7201 case CPU_ONLINE_FROZEN
:
7202 /* Strictly unnecessary, as first user will wake it. */
7203 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7205 /* Update our root-domain */
7207 spin_lock_irqsave(&rq
->lock
, flags
);
7209 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7213 spin_unlock_irqrestore(&rq
->lock
, flags
);
7216 #ifdef CONFIG_HOTPLUG_CPU
7217 case CPU_UP_CANCELED
:
7218 case CPU_UP_CANCELED_FROZEN
:
7219 if (!cpu_rq(cpu
)->migration_thread
)
7221 /* Unbind it from offline cpu so it can run. Fall thru. */
7222 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7223 cpumask_any(cpu_online_mask
));
7224 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7225 cpu_rq(cpu
)->migration_thread
= NULL
;
7229 case CPU_DEAD_FROZEN
:
7230 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7231 migrate_live_tasks(cpu
);
7233 kthread_stop(rq
->migration_thread
);
7234 rq
->migration_thread
= NULL
;
7235 /* Idle task back to normal (off runqueue, low prio) */
7236 spin_lock_irq(&rq
->lock
);
7237 update_rq_clock(rq
);
7238 deactivate_task(rq
, rq
->idle
, 0);
7239 rq
->idle
->static_prio
= MAX_PRIO
;
7240 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7241 rq
->idle
->sched_class
= &idle_sched_class
;
7242 migrate_dead_tasks(cpu
);
7243 spin_unlock_irq(&rq
->lock
);
7245 migrate_nr_uninterruptible(rq
);
7246 BUG_ON(rq
->nr_running
!= 0);
7249 * No need to migrate the tasks: it was best-effort if
7250 * they didn't take sched_hotcpu_mutex. Just wake up
7253 spin_lock_irq(&rq
->lock
);
7254 while (!list_empty(&rq
->migration_queue
)) {
7255 struct migration_req
*req
;
7257 req
= list_entry(rq
->migration_queue
.next
,
7258 struct migration_req
, list
);
7259 list_del_init(&req
->list
);
7260 spin_unlock_irq(&rq
->lock
);
7261 complete(&req
->done
);
7262 spin_lock_irq(&rq
->lock
);
7264 spin_unlock_irq(&rq
->lock
);
7268 case CPU_DYING_FROZEN
:
7269 /* Update our root-domain */
7271 spin_lock_irqsave(&rq
->lock
, flags
);
7273 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7276 spin_unlock_irqrestore(&rq
->lock
, flags
);
7283 /* Register at highest priority so that task migration (migrate_all_tasks)
7284 * happens before everything else.
7286 static struct notifier_block __cpuinitdata migration_notifier
= {
7287 .notifier_call
= migration_call
,
7291 static int __init
migration_init(void)
7293 void *cpu
= (void *)(long)smp_processor_id();
7296 /* Start one for the boot CPU: */
7297 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7298 BUG_ON(err
== NOTIFY_BAD
);
7299 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7300 register_cpu_notifier(&migration_notifier
);
7304 early_initcall(migration_init
);
7309 #ifdef CONFIG_SCHED_DEBUG
7311 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7312 struct cpumask
*groupmask
)
7314 struct sched_group
*group
= sd
->groups
;
7317 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7318 cpumask_clear(groupmask
);
7320 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7322 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7323 printk("does not load-balance\n");
7325 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7330 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7332 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7333 printk(KERN_ERR
"ERROR: domain->span does not contain "
7336 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7337 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7341 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7345 printk(KERN_ERR
"ERROR: group is NULL\n");
7349 if (!group
->__cpu_power
) {
7350 printk(KERN_CONT
"\n");
7351 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7356 if (!cpumask_weight(sched_group_cpus(group
))) {
7357 printk(KERN_CONT
"\n");
7358 printk(KERN_ERR
"ERROR: empty group\n");
7362 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7363 printk(KERN_CONT
"\n");
7364 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7368 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7370 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7372 printk(KERN_CONT
" %s", str
);
7373 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7374 printk(KERN_CONT
" (__cpu_power = %d)",
7375 group
->__cpu_power
);
7378 group
= group
->next
;
7379 } while (group
!= sd
->groups
);
7380 printk(KERN_CONT
"\n");
7382 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7383 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7386 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7387 printk(KERN_ERR
"ERROR: parent span is not a superset "
7388 "of domain->span\n");
7392 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7394 cpumask_var_t groupmask
;
7398 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7402 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7404 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7405 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7410 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7417 free_cpumask_var(groupmask
);
7419 #else /* !CONFIG_SCHED_DEBUG */
7420 # define sched_domain_debug(sd, cpu) do { } while (0)
7421 #endif /* CONFIG_SCHED_DEBUG */
7423 static int sd_degenerate(struct sched_domain
*sd
)
7425 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7428 /* Following flags need at least 2 groups */
7429 if (sd
->flags
& (SD_LOAD_BALANCE
|
7430 SD_BALANCE_NEWIDLE
|
7434 SD_SHARE_PKG_RESOURCES
)) {
7435 if (sd
->groups
!= sd
->groups
->next
)
7439 /* Following flags don't use groups */
7440 if (sd
->flags
& (SD_WAKE_IDLE
|
7449 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7451 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7453 if (sd_degenerate(parent
))
7456 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7459 /* Does parent contain flags not in child? */
7460 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7461 if (cflags
& SD_WAKE_AFFINE
)
7462 pflags
&= ~SD_WAKE_BALANCE
;
7463 /* Flags needing groups don't count if only 1 group in parent */
7464 if (parent
->groups
== parent
->groups
->next
) {
7465 pflags
&= ~(SD_LOAD_BALANCE
|
7466 SD_BALANCE_NEWIDLE
|
7470 SD_SHARE_PKG_RESOURCES
);
7471 if (nr_node_ids
== 1)
7472 pflags
&= ~SD_SERIALIZE
;
7474 if (~cflags
& pflags
)
7480 static void free_rootdomain(struct root_domain
*rd
)
7482 cpupri_cleanup(&rd
->cpupri
);
7484 free_cpumask_var(rd
->rto_mask
);
7485 free_cpumask_var(rd
->online
);
7486 free_cpumask_var(rd
->span
);
7490 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7492 struct root_domain
*old_rd
= NULL
;
7493 unsigned long flags
;
7495 spin_lock_irqsave(&rq
->lock
, flags
);
7500 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7503 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7506 * If we dont want to free the old_rt yet then
7507 * set old_rd to NULL to skip the freeing later
7510 if (!atomic_dec_and_test(&old_rd
->refcount
))
7514 atomic_inc(&rd
->refcount
);
7517 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7518 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7521 spin_unlock_irqrestore(&rq
->lock
, flags
);
7524 free_rootdomain(old_rd
);
7527 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7529 memset(rd
, 0, sizeof(*rd
));
7532 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7533 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7534 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7535 cpupri_init(&rd
->cpupri
, true);
7539 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7541 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7543 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7546 if (cpupri_init(&rd
->cpupri
, false) != 0)
7551 free_cpumask_var(rd
->rto_mask
);
7553 free_cpumask_var(rd
->online
);
7555 free_cpumask_var(rd
->span
);
7560 static void init_defrootdomain(void)
7562 init_rootdomain(&def_root_domain
, true);
7564 atomic_set(&def_root_domain
.refcount
, 1);
7567 static struct root_domain
*alloc_rootdomain(void)
7569 struct root_domain
*rd
;
7571 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7575 if (init_rootdomain(rd
, false) != 0) {
7584 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7585 * hold the hotplug lock.
7588 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7590 struct rq
*rq
= cpu_rq(cpu
);
7591 struct sched_domain
*tmp
;
7593 /* Remove the sched domains which do not contribute to scheduling. */
7594 for (tmp
= sd
; tmp
; ) {
7595 struct sched_domain
*parent
= tmp
->parent
;
7599 if (sd_parent_degenerate(tmp
, parent
)) {
7600 tmp
->parent
= parent
->parent
;
7602 parent
->parent
->child
= tmp
;
7607 if (sd
&& sd_degenerate(sd
)) {
7613 sched_domain_debug(sd
, cpu
);
7615 rq_attach_root(rq
, rd
);
7616 rcu_assign_pointer(rq
->sd
, sd
);
7619 /* cpus with isolated domains */
7620 static cpumask_var_t cpu_isolated_map
;
7622 /* Setup the mask of cpus configured for isolated domains */
7623 static int __init
isolated_cpu_setup(char *str
)
7625 cpulist_parse(str
, cpu_isolated_map
);
7629 __setup("isolcpus=", isolated_cpu_setup
);
7632 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7633 * to a function which identifies what group(along with sched group) a CPU
7634 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7635 * (due to the fact that we keep track of groups covered with a struct cpumask).
7637 * init_sched_build_groups will build a circular linked list of the groups
7638 * covered by the given span, and will set each group's ->cpumask correctly,
7639 * and ->cpu_power to 0.
7642 init_sched_build_groups(const struct cpumask
*span
,
7643 const struct cpumask
*cpu_map
,
7644 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7645 struct sched_group
**sg
,
7646 struct cpumask
*tmpmask
),
7647 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7649 struct sched_group
*first
= NULL
, *last
= NULL
;
7652 cpumask_clear(covered
);
7654 for_each_cpu(i
, span
) {
7655 struct sched_group
*sg
;
7656 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7659 if (cpumask_test_cpu(i
, covered
))
7662 cpumask_clear(sched_group_cpus(sg
));
7663 sg
->__cpu_power
= 0;
7665 for_each_cpu(j
, span
) {
7666 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7669 cpumask_set_cpu(j
, covered
);
7670 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7681 #define SD_NODES_PER_DOMAIN 16
7686 * find_next_best_node - find the next node to include in a sched_domain
7687 * @node: node whose sched_domain we're building
7688 * @used_nodes: nodes already in the sched_domain
7690 * Find the next node to include in a given scheduling domain. Simply
7691 * finds the closest node not already in the @used_nodes map.
7693 * Should use nodemask_t.
7695 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7697 int i
, n
, val
, min_val
, best_node
= 0;
7701 for (i
= 0; i
< nr_node_ids
; i
++) {
7702 /* Start at @node */
7703 n
= (node
+ i
) % nr_node_ids
;
7705 if (!nr_cpus_node(n
))
7708 /* Skip already used nodes */
7709 if (node_isset(n
, *used_nodes
))
7712 /* Simple min distance search */
7713 val
= node_distance(node
, n
);
7715 if (val
< min_val
) {
7721 node_set(best_node
, *used_nodes
);
7726 * sched_domain_node_span - get a cpumask for a node's sched_domain
7727 * @node: node whose cpumask we're constructing
7728 * @span: resulting cpumask
7730 * Given a node, construct a good cpumask for its sched_domain to span. It
7731 * should be one that prevents unnecessary balancing, but also spreads tasks
7734 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7736 nodemask_t used_nodes
;
7739 cpumask_clear(span
);
7740 nodes_clear(used_nodes
);
7742 cpumask_or(span
, span
, cpumask_of_node(node
));
7743 node_set(node
, used_nodes
);
7745 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7746 int next_node
= find_next_best_node(node
, &used_nodes
);
7748 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7751 #endif /* CONFIG_NUMA */
7753 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7756 * The cpus mask in sched_group and sched_domain hangs off the end.
7757 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7758 * for nr_cpu_ids < CONFIG_NR_CPUS.
7760 struct static_sched_group
{
7761 struct sched_group sg
;
7762 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7765 struct static_sched_domain
{
7766 struct sched_domain sd
;
7767 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7771 * SMT sched-domains:
7773 #ifdef CONFIG_SCHED_SMT
7774 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7775 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7778 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7779 struct sched_group
**sg
, struct cpumask
*unused
)
7782 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7785 #endif /* CONFIG_SCHED_SMT */
7788 * multi-core sched-domains:
7790 #ifdef CONFIG_SCHED_MC
7791 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7792 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7793 #endif /* CONFIG_SCHED_MC */
7795 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7797 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7798 struct sched_group
**sg
, struct cpumask
*mask
)
7802 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7803 group
= cpumask_first(mask
);
7805 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7808 #elif defined(CONFIG_SCHED_MC)
7810 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7811 struct sched_group
**sg
, struct cpumask
*unused
)
7814 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7819 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7820 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7823 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7824 struct sched_group
**sg
, struct cpumask
*mask
)
7827 #ifdef CONFIG_SCHED_MC
7828 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7829 group
= cpumask_first(mask
);
7830 #elif defined(CONFIG_SCHED_SMT)
7831 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7832 group
= cpumask_first(mask
);
7837 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7843 * The init_sched_build_groups can't handle what we want to do with node
7844 * groups, so roll our own. Now each node has its own list of groups which
7845 * gets dynamically allocated.
7847 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7848 static struct sched_group
***sched_group_nodes_bycpu
;
7850 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7851 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7853 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7854 struct sched_group
**sg
,
7855 struct cpumask
*nodemask
)
7859 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7860 group
= cpumask_first(nodemask
);
7863 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7867 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7869 struct sched_group
*sg
= group_head
;
7875 for_each_cpu(j
, sched_group_cpus(sg
)) {
7876 struct sched_domain
*sd
;
7878 sd
= &per_cpu(phys_domains
, j
).sd
;
7879 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7881 * Only add "power" once for each
7887 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7890 } while (sg
!= group_head
);
7892 #endif /* CONFIG_NUMA */
7895 /* Free memory allocated for various sched_group structures */
7896 static void free_sched_groups(const struct cpumask
*cpu_map
,
7897 struct cpumask
*nodemask
)
7901 for_each_cpu(cpu
, cpu_map
) {
7902 struct sched_group
**sched_group_nodes
7903 = sched_group_nodes_bycpu
[cpu
];
7905 if (!sched_group_nodes
)
7908 for (i
= 0; i
< nr_node_ids
; i
++) {
7909 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7911 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7912 if (cpumask_empty(nodemask
))
7922 if (oldsg
!= sched_group_nodes
[i
])
7925 kfree(sched_group_nodes
);
7926 sched_group_nodes_bycpu
[cpu
] = NULL
;
7929 #else /* !CONFIG_NUMA */
7930 static void free_sched_groups(const struct cpumask
*cpu_map
,
7931 struct cpumask
*nodemask
)
7934 #endif /* CONFIG_NUMA */
7937 * Initialize sched groups cpu_power.
7939 * cpu_power indicates the capacity of sched group, which is used while
7940 * distributing the load between different sched groups in a sched domain.
7941 * Typically cpu_power for all the groups in a sched domain will be same unless
7942 * there are asymmetries in the topology. If there are asymmetries, group
7943 * having more cpu_power will pickup more load compared to the group having
7946 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7947 * the maximum number of tasks a group can handle in the presence of other idle
7948 * or lightly loaded groups in the same sched domain.
7950 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7952 struct sched_domain
*child
;
7953 struct sched_group
*group
;
7955 WARN_ON(!sd
|| !sd
->groups
);
7957 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7962 sd
->groups
->__cpu_power
= 0;
7965 * For perf policy, if the groups in child domain share resources
7966 * (for example cores sharing some portions of the cache hierarchy
7967 * or SMT), then set this domain groups cpu_power such that each group
7968 * can handle only one task, when there are other idle groups in the
7969 * same sched domain.
7971 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7973 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7974 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7979 * add cpu_power of each child group to this groups cpu_power
7981 group
= child
->groups
;
7983 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7984 group
= group
->next
;
7985 } while (group
!= child
->groups
);
7989 * Initializers for schedule domains
7990 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7993 #ifdef CONFIG_SCHED_DEBUG
7994 # define SD_INIT_NAME(sd, type) sd->name = #type
7996 # define SD_INIT_NAME(sd, type) do { } while (0)
7999 #define SD_INIT(sd, type) sd_init_##type(sd)
8001 #define SD_INIT_FUNC(type) \
8002 static noinline void sd_init_##type(struct sched_domain *sd) \
8004 memset(sd, 0, sizeof(*sd)); \
8005 *sd = SD_##type##_INIT; \
8006 sd->level = SD_LV_##type; \
8007 SD_INIT_NAME(sd, type); \
8012 SD_INIT_FUNC(ALLNODES
)
8015 #ifdef CONFIG_SCHED_SMT
8016 SD_INIT_FUNC(SIBLING
)
8018 #ifdef CONFIG_SCHED_MC
8022 static int default_relax_domain_level
= -1;
8024 static int __init
setup_relax_domain_level(char *str
)
8028 val
= simple_strtoul(str
, NULL
, 0);
8029 if (val
< SD_LV_MAX
)
8030 default_relax_domain_level
= val
;
8034 __setup("relax_domain_level=", setup_relax_domain_level
);
8036 static void set_domain_attribute(struct sched_domain
*sd
,
8037 struct sched_domain_attr
*attr
)
8041 if (!attr
|| attr
->relax_domain_level
< 0) {
8042 if (default_relax_domain_level
< 0)
8045 request
= default_relax_domain_level
;
8047 request
= attr
->relax_domain_level
;
8048 if (request
< sd
->level
) {
8049 /* turn off idle balance on this domain */
8050 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8052 /* turn on idle balance on this domain */
8053 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8058 * Build sched domains for a given set of cpus and attach the sched domains
8059 * to the individual cpus
8061 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8062 struct sched_domain_attr
*attr
)
8064 int i
, err
= -ENOMEM
;
8065 struct root_domain
*rd
;
8066 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8069 cpumask_var_t domainspan
, covered
, notcovered
;
8070 struct sched_group
**sched_group_nodes
= NULL
;
8071 int sd_allnodes
= 0;
8073 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8075 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8076 goto free_domainspan
;
8077 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8081 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8082 goto free_notcovered
;
8083 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8085 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8086 goto free_this_sibling_map
;
8087 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8088 goto free_this_core_map
;
8089 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8090 goto free_send_covered
;
8094 * Allocate the per-node list of sched groups
8096 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8098 if (!sched_group_nodes
) {
8099 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8104 rd
= alloc_rootdomain();
8106 printk(KERN_WARNING
"Cannot alloc root domain\n");
8107 goto free_sched_groups
;
8111 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8115 * Set up domains for cpus specified by the cpu_map.
8117 for_each_cpu(i
, cpu_map
) {
8118 struct sched_domain
*sd
= NULL
, *p
;
8120 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8123 if (cpumask_weight(cpu_map
) >
8124 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8125 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8126 SD_INIT(sd
, ALLNODES
);
8127 set_domain_attribute(sd
, attr
);
8128 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8129 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8135 sd
= &per_cpu(node_domains
, i
).sd
;
8137 set_domain_attribute(sd
, attr
);
8138 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8142 cpumask_and(sched_domain_span(sd
),
8143 sched_domain_span(sd
), cpu_map
);
8147 sd
= &per_cpu(phys_domains
, i
).sd
;
8149 set_domain_attribute(sd
, attr
);
8150 cpumask_copy(sched_domain_span(sd
), nodemask
);
8154 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8156 #ifdef CONFIG_SCHED_MC
8158 sd
= &per_cpu(core_domains
, i
).sd
;
8160 set_domain_attribute(sd
, attr
);
8161 cpumask_and(sched_domain_span(sd
), cpu_map
,
8162 cpu_coregroup_mask(i
));
8165 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8168 #ifdef CONFIG_SCHED_SMT
8170 sd
= &per_cpu(cpu_domains
, i
).sd
;
8171 SD_INIT(sd
, SIBLING
);
8172 set_domain_attribute(sd
, attr
);
8173 cpumask_and(sched_domain_span(sd
),
8174 topology_thread_cpumask(i
), cpu_map
);
8177 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8181 #ifdef CONFIG_SCHED_SMT
8182 /* Set up CPU (sibling) groups */
8183 for_each_cpu(i
, cpu_map
) {
8184 cpumask_and(this_sibling_map
,
8185 topology_thread_cpumask(i
), cpu_map
);
8186 if (i
!= cpumask_first(this_sibling_map
))
8189 init_sched_build_groups(this_sibling_map
, cpu_map
,
8191 send_covered
, tmpmask
);
8195 #ifdef CONFIG_SCHED_MC
8196 /* Set up multi-core groups */
8197 for_each_cpu(i
, cpu_map
) {
8198 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8199 if (i
!= cpumask_first(this_core_map
))
8202 init_sched_build_groups(this_core_map
, cpu_map
,
8204 send_covered
, tmpmask
);
8208 /* Set up physical groups */
8209 for (i
= 0; i
< nr_node_ids
; i
++) {
8210 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8211 if (cpumask_empty(nodemask
))
8214 init_sched_build_groups(nodemask
, cpu_map
,
8216 send_covered
, tmpmask
);
8220 /* Set up node groups */
8222 init_sched_build_groups(cpu_map
, cpu_map
,
8223 &cpu_to_allnodes_group
,
8224 send_covered
, tmpmask
);
8227 for (i
= 0; i
< nr_node_ids
; i
++) {
8228 /* Set up node groups */
8229 struct sched_group
*sg
, *prev
;
8232 cpumask_clear(covered
);
8233 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8234 if (cpumask_empty(nodemask
)) {
8235 sched_group_nodes
[i
] = NULL
;
8239 sched_domain_node_span(i
, domainspan
);
8240 cpumask_and(domainspan
, domainspan
, cpu_map
);
8242 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8245 printk(KERN_WARNING
"Can not alloc domain group for "
8249 sched_group_nodes
[i
] = sg
;
8250 for_each_cpu(j
, nodemask
) {
8251 struct sched_domain
*sd
;
8253 sd
= &per_cpu(node_domains
, j
).sd
;
8256 sg
->__cpu_power
= 0;
8257 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8259 cpumask_or(covered
, covered
, nodemask
);
8262 for (j
= 0; j
< nr_node_ids
; j
++) {
8263 int n
= (i
+ j
) % nr_node_ids
;
8265 cpumask_complement(notcovered
, covered
);
8266 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8267 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8268 if (cpumask_empty(tmpmask
))
8271 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8272 if (cpumask_empty(tmpmask
))
8275 sg
= kmalloc_node(sizeof(struct sched_group
) +
8280 "Can not alloc domain group for node %d\n", j
);
8283 sg
->__cpu_power
= 0;
8284 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8285 sg
->next
= prev
->next
;
8286 cpumask_or(covered
, covered
, tmpmask
);
8293 /* Calculate CPU power for physical packages and nodes */
8294 #ifdef CONFIG_SCHED_SMT
8295 for_each_cpu(i
, cpu_map
) {
8296 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8298 init_sched_groups_power(i
, sd
);
8301 #ifdef CONFIG_SCHED_MC
8302 for_each_cpu(i
, cpu_map
) {
8303 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8305 init_sched_groups_power(i
, sd
);
8309 for_each_cpu(i
, cpu_map
) {
8310 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8312 init_sched_groups_power(i
, sd
);
8316 for (i
= 0; i
< nr_node_ids
; i
++)
8317 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8320 struct sched_group
*sg
;
8322 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8324 init_numa_sched_groups_power(sg
);
8328 /* Attach the domains */
8329 for_each_cpu(i
, cpu_map
) {
8330 struct sched_domain
*sd
;
8331 #ifdef CONFIG_SCHED_SMT
8332 sd
= &per_cpu(cpu_domains
, i
).sd
;
8333 #elif defined(CONFIG_SCHED_MC)
8334 sd
= &per_cpu(core_domains
, i
).sd
;
8336 sd
= &per_cpu(phys_domains
, i
).sd
;
8338 cpu_attach_domain(sd
, rd
, i
);
8344 free_cpumask_var(tmpmask
);
8346 free_cpumask_var(send_covered
);
8348 free_cpumask_var(this_core_map
);
8349 free_this_sibling_map
:
8350 free_cpumask_var(this_sibling_map
);
8352 free_cpumask_var(nodemask
);
8355 free_cpumask_var(notcovered
);
8357 free_cpumask_var(covered
);
8359 free_cpumask_var(domainspan
);
8366 kfree(sched_group_nodes
);
8372 free_sched_groups(cpu_map
, tmpmask
);
8373 free_rootdomain(rd
);
8378 static int build_sched_domains(const struct cpumask
*cpu_map
)
8380 return __build_sched_domains(cpu_map
, NULL
);
8383 static struct cpumask
*doms_cur
; /* current sched domains */
8384 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8385 static struct sched_domain_attr
*dattr_cur
;
8386 /* attribues of custom domains in 'doms_cur' */
8389 * Special case: If a kmalloc of a doms_cur partition (array of
8390 * cpumask) fails, then fallback to a single sched domain,
8391 * as determined by the single cpumask fallback_doms.
8393 static cpumask_var_t fallback_doms
;
8396 * arch_update_cpu_topology lets virtualized architectures update the
8397 * cpu core maps. It is supposed to return 1 if the topology changed
8398 * or 0 if it stayed the same.
8400 int __attribute__((weak
)) arch_update_cpu_topology(void)
8406 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8407 * For now this just excludes isolated cpus, but could be used to
8408 * exclude other special cases in the future.
8410 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8414 arch_update_cpu_topology();
8416 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8418 doms_cur
= fallback_doms
;
8419 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8421 err
= build_sched_domains(doms_cur
);
8422 register_sched_domain_sysctl();
8427 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8428 struct cpumask
*tmpmask
)
8430 free_sched_groups(cpu_map
, tmpmask
);
8434 * Detach sched domains from a group of cpus specified in cpu_map
8435 * These cpus will now be attached to the NULL domain
8437 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8439 /* Save because hotplug lock held. */
8440 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8443 for_each_cpu(i
, cpu_map
)
8444 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8445 synchronize_sched();
8446 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8449 /* handle null as "default" */
8450 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8451 struct sched_domain_attr
*new, int idx_new
)
8453 struct sched_domain_attr tmp
;
8460 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8461 new ? (new + idx_new
) : &tmp
,
8462 sizeof(struct sched_domain_attr
));
8466 * Partition sched domains as specified by the 'ndoms_new'
8467 * cpumasks in the array doms_new[] of cpumasks. This compares
8468 * doms_new[] to the current sched domain partitioning, doms_cur[].
8469 * It destroys each deleted domain and builds each new domain.
8471 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8472 * The masks don't intersect (don't overlap.) We should setup one
8473 * sched domain for each mask. CPUs not in any of the cpumasks will
8474 * not be load balanced. If the same cpumask appears both in the
8475 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8478 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8479 * ownership of it and will kfree it when done with it. If the caller
8480 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8481 * ndoms_new == 1, and partition_sched_domains() will fallback to
8482 * the single partition 'fallback_doms', it also forces the domains
8485 * If doms_new == NULL it will be replaced with cpu_online_mask.
8486 * ndoms_new == 0 is a special case for destroying existing domains,
8487 * and it will not create the default domain.
8489 * Call with hotplug lock held
8491 /* FIXME: Change to struct cpumask *doms_new[] */
8492 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8493 struct sched_domain_attr
*dattr_new
)
8498 mutex_lock(&sched_domains_mutex
);
8500 /* always unregister in case we don't destroy any domains */
8501 unregister_sched_domain_sysctl();
8503 /* Let architecture update cpu core mappings. */
8504 new_topology
= arch_update_cpu_topology();
8506 n
= doms_new
? ndoms_new
: 0;
8508 /* Destroy deleted domains */
8509 for (i
= 0; i
< ndoms_cur
; i
++) {
8510 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8511 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8512 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8515 /* no match - a current sched domain not in new doms_new[] */
8516 detach_destroy_domains(doms_cur
+ i
);
8521 if (doms_new
== NULL
) {
8523 doms_new
= fallback_doms
;
8524 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8525 WARN_ON_ONCE(dattr_new
);
8528 /* Build new domains */
8529 for (i
= 0; i
< ndoms_new
; i
++) {
8530 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8531 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8532 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8535 /* no match - add a new doms_new */
8536 __build_sched_domains(doms_new
+ i
,
8537 dattr_new
? dattr_new
+ i
: NULL
);
8542 /* Remember the new sched domains */
8543 if (doms_cur
!= fallback_doms
)
8545 kfree(dattr_cur
); /* kfree(NULL) is safe */
8546 doms_cur
= doms_new
;
8547 dattr_cur
= dattr_new
;
8548 ndoms_cur
= ndoms_new
;
8550 register_sched_domain_sysctl();
8552 mutex_unlock(&sched_domains_mutex
);
8555 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8556 static void arch_reinit_sched_domains(void)
8560 /* Destroy domains first to force the rebuild */
8561 partition_sched_domains(0, NULL
, NULL
);
8563 rebuild_sched_domains();
8567 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8569 unsigned int level
= 0;
8571 if (sscanf(buf
, "%u", &level
) != 1)
8575 * level is always be positive so don't check for
8576 * level < POWERSAVINGS_BALANCE_NONE which is 0
8577 * What happens on 0 or 1 byte write,
8578 * need to check for count as well?
8581 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8585 sched_smt_power_savings
= level
;
8587 sched_mc_power_savings
= level
;
8589 arch_reinit_sched_domains();
8594 #ifdef CONFIG_SCHED_MC
8595 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8598 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8600 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8601 const char *buf
, size_t count
)
8603 return sched_power_savings_store(buf
, count
, 0);
8605 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8606 sched_mc_power_savings_show
,
8607 sched_mc_power_savings_store
);
8610 #ifdef CONFIG_SCHED_SMT
8611 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8614 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8616 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8617 const char *buf
, size_t count
)
8619 return sched_power_savings_store(buf
, count
, 1);
8621 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8622 sched_smt_power_savings_show
,
8623 sched_smt_power_savings_store
);
8626 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8630 #ifdef CONFIG_SCHED_SMT
8632 err
= sysfs_create_file(&cls
->kset
.kobj
,
8633 &attr_sched_smt_power_savings
.attr
);
8635 #ifdef CONFIG_SCHED_MC
8636 if (!err
&& mc_capable())
8637 err
= sysfs_create_file(&cls
->kset
.kobj
,
8638 &attr_sched_mc_power_savings
.attr
);
8642 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8644 #ifndef CONFIG_CPUSETS
8646 * Add online and remove offline CPUs from the scheduler domains.
8647 * When cpusets are enabled they take over this function.
8649 static int update_sched_domains(struct notifier_block
*nfb
,
8650 unsigned long action
, void *hcpu
)
8654 case CPU_ONLINE_FROZEN
:
8656 case CPU_DEAD_FROZEN
:
8657 partition_sched_domains(1, NULL
, NULL
);
8666 static int update_runtime(struct notifier_block
*nfb
,
8667 unsigned long action
, void *hcpu
)
8669 int cpu
= (int)(long)hcpu
;
8672 case CPU_DOWN_PREPARE
:
8673 case CPU_DOWN_PREPARE_FROZEN
:
8674 disable_runtime(cpu_rq(cpu
));
8677 case CPU_DOWN_FAILED
:
8678 case CPU_DOWN_FAILED_FROZEN
:
8680 case CPU_ONLINE_FROZEN
:
8681 enable_runtime(cpu_rq(cpu
));
8689 void __init
sched_init_smp(void)
8691 cpumask_var_t non_isolated_cpus
;
8693 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8695 #if defined(CONFIG_NUMA)
8696 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8698 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8701 mutex_lock(&sched_domains_mutex
);
8702 arch_init_sched_domains(cpu_online_mask
);
8703 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8704 if (cpumask_empty(non_isolated_cpus
))
8705 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8706 mutex_unlock(&sched_domains_mutex
);
8709 #ifndef CONFIG_CPUSETS
8710 /* XXX: Theoretical race here - CPU may be hotplugged now */
8711 hotcpu_notifier(update_sched_domains
, 0);
8714 /* RT runtime code needs to handle some hotplug events */
8715 hotcpu_notifier(update_runtime
, 0);
8719 /* Move init over to a non-isolated CPU */
8720 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8722 sched_init_granularity();
8723 free_cpumask_var(non_isolated_cpus
);
8725 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8726 init_sched_rt_class();
8729 void __init
sched_init_smp(void)
8731 sched_init_granularity();
8733 #endif /* CONFIG_SMP */
8735 int in_sched_functions(unsigned long addr
)
8737 return in_lock_functions(addr
) ||
8738 (addr
>= (unsigned long)__sched_text_start
8739 && addr
< (unsigned long)__sched_text_end
);
8742 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8744 cfs_rq
->tasks_timeline
= RB_ROOT
;
8745 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8746 #ifdef CONFIG_FAIR_GROUP_SCHED
8749 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8752 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8754 struct rt_prio_array
*array
;
8757 array
= &rt_rq
->active
;
8758 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8759 INIT_LIST_HEAD(array
->queue
+ i
);
8760 __clear_bit(i
, array
->bitmap
);
8762 /* delimiter for bitsearch: */
8763 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8765 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8766 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8768 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8772 rt_rq
->rt_nr_migratory
= 0;
8773 rt_rq
->overloaded
= 0;
8774 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8778 rt_rq
->rt_throttled
= 0;
8779 rt_rq
->rt_runtime
= 0;
8780 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8782 #ifdef CONFIG_RT_GROUP_SCHED
8783 rt_rq
->rt_nr_boosted
= 0;
8788 #ifdef CONFIG_FAIR_GROUP_SCHED
8789 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8790 struct sched_entity
*se
, int cpu
, int add
,
8791 struct sched_entity
*parent
)
8793 struct rq
*rq
= cpu_rq(cpu
);
8794 tg
->cfs_rq
[cpu
] = cfs_rq
;
8795 init_cfs_rq(cfs_rq
, rq
);
8798 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8801 /* se could be NULL for init_task_group */
8806 se
->cfs_rq
= &rq
->cfs
;
8808 se
->cfs_rq
= parent
->my_q
;
8811 se
->load
.weight
= tg
->shares
;
8812 se
->load
.inv_weight
= 0;
8813 se
->parent
= parent
;
8817 #ifdef CONFIG_RT_GROUP_SCHED
8818 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8819 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8820 struct sched_rt_entity
*parent
)
8822 struct rq
*rq
= cpu_rq(cpu
);
8824 tg
->rt_rq
[cpu
] = rt_rq
;
8825 init_rt_rq(rt_rq
, rq
);
8827 rt_rq
->rt_se
= rt_se
;
8828 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8830 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8832 tg
->rt_se
[cpu
] = rt_se
;
8837 rt_se
->rt_rq
= &rq
->rt
;
8839 rt_se
->rt_rq
= parent
->my_q
;
8841 rt_se
->my_q
= rt_rq
;
8842 rt_se
->parent
= parent
;
8843 INIT_LIST_HEAD(&rt_se
->run_list
);
8847 void __init
sched_init(void)
8850 unsigned long alloc_size
= 0, ptr
;
8852 #ifdef CONFIG_FAIR_GROUP_SCHED
8853 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8855 #ifdef CONFIG_RT_GROUP_SCHED
8856 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8858 #ifdef CONFIG_USER_SCHED
8861 #ifdef CONFIG_CPUMASK_OFFSTACK
8862 alloc_size
+= num_possible_cpus() * cpumask_size();
8865 * As sched_init() is called before page_alloc is setup,
8866 * we use alloc_bootmem().
8869 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8871 #ifdef CONFIG_FAIR_GROUP_SCHED
8872 init_task_group
.se
= (struct sched_entity
**)ptr
;
8873 ptr
+= nr_cpu_ids
* sizeof(void **);
8875 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8876 ptr
+= nr_cpu_ids
* sizeof(void **);
8878 #ifdef CONFIG_USER_SCHED
8879 root_task_group
.se
= (struct sched_entity
**)ptr
;
8880 ptr
+= nr_cpu_ids
* sizeof(void **);
8882 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8883 ptr
+= nr_cpu_ids
* sizeof(void **);
8884 #endif /* CONFIG_USER_SCHED */
8885 #endif /* CONFIG_FAIR_GROUP_SCHED */
8886 #ifdef CONFIG_RT_GROUP_SCHED
8887 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8888 ptr
+= nr_cpu_ids
* sizeof(void **);
8890 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8891 ptr
+= nr_cpu_ids
* sizeof(void **);
8893 #ifdef CONFIG_USER_SCHED
8894 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8895 ptr
+= nr_cpu_ids
* sizeof(void **);
8897 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8898 ptr
+= nr_cpu_ids
* sizeof(void **);
8899 #endif /* CONFIG_USER_SCHED */
8900 #endif /* CONFIG_RT_GROUP_SCHED */
8901 #ifdef CONFIG_CPUMASK_OFFSTACK
8902 for_each_possible_cpu(i
) {
8903 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8904 ptr
+= cpumask_size();
8906 #endif /* CONFIG_CPUMASK_OFFSTACK */
8910 init_defrootdomain();
8913 init_rt_bandwidth(&def_rt_bandwidth
,
8914 global_rt_period(), global_rt_runtime());
8916 #ifdef CONFIG_RT_GROUP_SCHED
8917 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8918 global_rt_period(), global_rt_runtime());
8919 #ifdef CONFIG_USER_SCHED
8920 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8921 global_rt_period(), RUNTIME_INF
);
8922 #endif /* CONFIG_USER_SCHED */
8923 #endif /* CONFIG_RT_GROUP_SCHED */
8925 #ifdef CONFIG_GROUP_SCHED
8926 list_add(&init_task_group
.list
, &task_groups
);
8927 INIT_LIST_HEAD(&init_task_group
.children
);
8929 #ifdef CONFIG_USER_SCHED
8930 INIT_LIST_HEAD(&root_task_group
.children
);
8931 init_task_group
.parent
= &root_task_group
;
8932 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8933 #endif /* CONFIG_USER_SCHED */
8934 #endif /* CONFIG_GROUP_SCHED */
8936 for_each_possible_cpu(i
) {
8940 spin_lock_init(&rq
->lock
);
8942 init_cfs_rq(&rq
->cfs
, rq
);
8943 init_rt_rq(&rq
->rt
, rq
);
8944 #ifdef CONFIG_FAIR_GROUP_SCHED
8945 init_task_group
.shares
= init_task_group_load
;
8946 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8947 #ifdef CONFIG_CGROUP_SCHED
8949 * How much cpu bandwidth does init_task_group get?
8951 * In case of task-groups formed thr' the cgroup filesystem, it
8952 * gets 100% of the cpu resources in the system. This overall
8953 * system cpu resource is divided among the tasks of
8954 * init_task_group and its child task-groups in a fair manner,
8955 * based on each entity's (task or task-group's) weight
8956 * (se->load.weight).
8958 * In other words, if init_task_group has 10 tasks of weight
8959 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8960 * then A0's share of the cpu resource is:
8962 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8964 * We achieve this by letting init_task_group's tasks sit
8965 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8967 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8968 #elif defined CONFIG_USER_SCHED
8969 root_task_group
.shares
= NICE_0_LOAD
;
8970 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8972 * In case of task-groups formed thr' the user id of tasks,
8973 * init_task_group represents tasks belonging to root user.
8974 * Hence it forms a sibling of all subsequent groups formed.
8975 * In this case, init_task_group gets only a fraction of overall
8976 * system cpu resource, based on the weight assigned to root
8977 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8978 * by letting tasks of init_task_group sit in a separate cfs_rq
8979 * (init_cfs_rq) and having one entity represent this group of
8980 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8982 init_tg_cfs_entry(&init_task_group
,
8983 &per_cpu(init_cfs_rq
, i
),
8984 &per_cpu(init_sched_entity
, i
), i
, 1,
8985 root_task_group
.se
[i
]);
8988 #endif /* CONFIG_FAIR_GROUP_SCHED */
8990 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8991 #ifdef CONFIG_RT_GROUP_SCHED
8992 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8993 #ifdef CONFIG_CGROUP_SCHED
8994 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8995 #elif defined CONFIG_USER_SCHED
8996 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8997 init_tg_rt_entry(&init_task_group
,
8998 &per_cpu(init_rt_rq
, i
),
8999 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9000 root_task_group
.rt_se
[i
]);
9004 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9005 rq
->cpu_load
[j
] = 0;
9009 rq
->active_balance
= 0;
9010 rq
->next_balance
= jiffies
;
9014 rq
->migration_thread
= NULL
;
9015 INIT_LIST_HEAD(&rq
->migration_queue
);
9016 rq_attach_root(rq
, &def_root_domain
);
9019 atomic_set(&rq
->nr_iowait
, 0);
9022 set_load_weight(&init_task
);
9024 #ifdef CONFIG_PREEMPT_NOTIFIERS
9025 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9029 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9032 #ifdef CONFIG_RT_MUTEXES
9033 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9037 * The boot idle thread does lazy MMU switching as well:
9039 atomic_inc(&init_mm
.mm_count
);
9040 enter_lazy_tlb(&init_mm
, current
);
9043 * Make us the idle thread. Technically, schedule() should not be
9044 * called from this thread, however somewhere below it might be,
9045 * but because we are the idle thread, we just pick up running again
9046 * when this runqueue becomes "idle".
9048 init_idle(current
, smp_processor_id());
9050 * During early bootup we pretend to be a normal task:
9052 current
->sched_class
= &fair_sched_class
;
9054 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9055 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9058 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9060 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9063 scheduler_running
= 1;
9066 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9067 void __might_sleep(char *file
, int line
)
9070 static unsigned long prev_jiffy
; /* ratelimiting */
9072 if ((!in_atomic() && !irqs_disabled()) ||
9073 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9075 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9077 prev_jiffy
= jiffies
;
9080 "BUG: sleeping function called from invalid context at %s:%d\n",
9083 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9084 in_atomic(), irqs_disabled(),
9085 current
->pid
, current
->comm
);
9087 debug_show_held_locks(current
);
9088 if (irqs_disabled())
9089 print_irqtrace_events(current
);
9093 EXPORT_SYMBOL(__might_sleep
);
9096 #ifdef CONFIG_MAGIC_SYSRQ
9097 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9101 update_rq_clock(rq
);
9102 on_rq
= p
->se
.on_rq
;
9104 deactivate_task(rq
, p
, 0);
9105 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9107 activate_task(rq
, p
, 0);
9108 resched_task(rq
->curr
);
9112 void normalize_rt_tasks(void)
9114 struct task_struct
*g
, *p
;
9115 unsigned long flags
;
9118 read_lock_irqsave(&tasklist_lock
, flags
);
9119 do_each_thread(g
, p
) {
9121 * Only normalize user tasks:
9126 p
->se
.exec_start
= 0;
9127 #ifdef CONFIG_SCHEDSTATS
9128 p
->se
.wait_start
= 0;
9129 p
->se
.sleep_start
= 0;
9130 p
->se
.block_start
= 0;
9135 * Renice negative nice level userspace
9138 if (TASK_NICE(p
) < 0 && p
->mm
)
9139 set_user_nice(p
, 0);
9143 spin_lock(&p
->pi_lock
);
9144 rq
= __task_rq_lock(p
);
9146 normalize_task(rq
, p
);
9148 __task_rq_unlock(rq
);
9149 spin_unlock(&p
->pi_lock
);
9150 } while_each_thread(g
, p
);
9152 read_unlock_irqrestore(&tasklist_lock
, flags
);
9155 #endif /* CONFIG_MAGIC_SYSRQ */
9159 * These functions are only useful for the IA64 MCA handling.
9161 * They can only be called when the whole system has been
9162 * stopped - every CPU needs to be quiescent, and no scheduling
9163 * activity can take place. Using them for anything else would
9164 * be a serious bug, and as a result, they aren't even visible
9165 * under any other configuration.
9169 * curr_task - return the current task for a given cpu.
9170 * @cpu: the processor in question.
9172 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9174 struct task_struct
*curr_task(int cpu
)
9176 return cpu_curr(cpu
);
9180 * set_curr_task - set the current task for a given cpu.
9181 * @cpu: the processor in question.
9182 * @p: the task pointer to set.
9184 * Description: This function must only be used when non-maskable interrupts
9185 * are serviced on a separate stack. It allows the architecture to switch the
9186 * notion of the current task on a cpu in a non-blocking manner. This function
9187 * must be called with all CPU's synchronized, and interrupts disabled, the
9188 * and caller must save the original value of the current task (see
9189 * curr_task() above) and restore that value before reenabling interrupts and
9190 * re-starting the system.
9192 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9194 void set_curr_task(int cpu
, struct task_struct
*p
)
9201 #ifdef CONFIG_FAIR_GROUP_SCHED
9202 static void free_fair_sched_group(struct task_group
*tg
)
9206 for_each_possible_cpu(i
) {
9208 kfree(tg
->cfs_rq
[i
]);
9218 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9220 struct cfs_rq
*cfs_rq
;
9221 struct sched_entity
*se
;
9225 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9228 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9232 tg
->shares
= NICE_0_LOAD
;
9234 for_each_possible_cpu(i
) {
9237 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9238 GFP_KERNEL
, cpu_to_node(i
));
9242 se
= kzalloc_node(sizeof(struct sched_entity
),
9243 GFP_KERNEL
, cpu_to_node(i
));
9247 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9256 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9258 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9259 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9262 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9264 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9266 #else /* !CONFG_FAIR_GROUP_SCHED */
9267 static inline void free_fair_sched_group(struct task_group
*tg
)
9272 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9277 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9281 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9284 #endif /* CONFIG_FAIR_GROUP_SCHED */
9286 #ifdef CONFIG_RT_GROUP_SCHED
9287 static void free_rt_sched_group(struct task_group
*tg
)
9291 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9293 for_each_possible_cpu(i
) {
9295 kfree(tg
->rt_rq
[i
]);
9297 kfree(tg
->rt_se
[i
]);
9305 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9307 struct rt_rq
*rt_rq
;
9308 struct sched_rt_entity
*rt_se
;
9312 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9315 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9319 init_rt_bandwidth(&tg
->rt_bandwidth
,
9320 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9322 for_each_possible_cpu(i
) {
9325 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9326 GFP_KERNEL
, cpu_to_node(i
));
9330 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9331 GFP_KERNEL
, cpu_to_node(i
));
9335 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9344 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9346 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9347 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9350 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9352 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9354 #else /* !CONFIG_RT_GROUP_SCHED */
9355 static inline void free_rt_sched_group(struct task_group
*tg
)
9360 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9365 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9369 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9372 #endif /* CONFIG_RT_GROUP_SCHED */
9374 #ifdef CONFIG_GROUP_SCHED
9375 static void free_sched_group(struct task_group
*tg
)
9377 free_fair_sched_group(tg
);
9378 free_rt_sched_group(tg
);
9382 /* allocate runqueue etc for a new task group */
9383 struct task_group
*sched_create_group(struct task_group
*parent
)
9385 struct task_group
*tg
;
9386 unsigned long flags
;
9389 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9391 return ERR_PTR(-ENOMEM
);
9393 if (!alloc_fair_sched_group(tg
, parent
))
9396 if (!alloc_rt_sched_group(tg
, parent
))
9399 spin_lock_irqsave(&task_group_lock
, flags
);
9400 for_each_possible_cpu(i
) {
9401 register_fair_sched_group(tg
, i
);
9402 register_rt_sched_group(tg
, i
);
9404 list_add_rcu(&tg
->list
, &task_groups
);
9406 WARN_ON(!parent
); /* root should already exist */
9408 tg
->parent
= parent
;
9409 INIT_LIST_HEAD(&tg
->children
);
9410 list_add_rcu(&tg
->siblings
, &parent
->children
);
9411 spin_unlock_irqrestore(&task_group_lock
, flags
);
9416 free_sched_group(tg
);
9417 return ERR_PTR(-ENOMEM
);
9420 /* rcu callback to free various structures associated with a task group */
9421 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9423 /* now it should be safe to free those cfs_rqs */
9424 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9427 /* Destroy runqueue etc associated with a task group */
9428 void sched_destroy_group(struct task_group
*tg
)
9430 unsigned long flags
;
9433 spin_lock_irqsave(&task_group_lock
, flags
);
9434 for_each_possible_cpu(i
) {
9435 unregister_fair_sched_group(tg
, i
);
9436 unregister_rt_sched_group(tg
, i
);
9438 list_del_rcu(&tg
->list
);
9439 list_del_rcu(&tg
->siblings
);
9440 spin_unlock_irqrestore(&task_group_lock
, flags
);
9442 /* wait for possible concurrent references to cfs_rqs complete */
9443 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9446 /* change task's runqueue when it moves between groups.
9447 * The caller of this function should have put the task in its new group
9448 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9449 * reflect its new group.
9451 void sched_move_task(struct task_struct
*tsk
)
9454 unsigned long flags
;
9457 rq
= task_rq_lock(tsk
, &flags
);
9459 update_rq_clock(rq
);
9461 running
= task_current(rq
, tsk
);
9462 on_rq
= tsk
->se
.on_rq
;
9465 dequeue_task(rq
, tsk
, 0);
9466 if (unlikely(running
))
9467 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9469 set_task_rq(tsk
, task_cpu(tsk
));
9471 #ifdef CONFIG_FAIR_GROUP_SCHED
9472 if (tsk
->sched_class
->moved_group
)
9473 tsk
->sched_class
->moved_group(tsk
);
9476 if (unlikely(running
))
9477 tsk
->sched_class
->set_curr_task(rq
);
9479 enqueue_task(rq
, tsk
, 0);
9481 task_rq_unlock(rq
, &flags
);
9483 #endif /* CONFIG_GROUP_SCHED */
9485 #ifdef CONFIG_FAIR_GROUP_SCHED
9486 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9488 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9493 dequeue_entity(cfs_rq
, se
, 0);
9495 se
->load
.weight
= shares
;
9496 se
->load
.inv_weight
= 0;
9499 enqueue_entity(cfs_rq
, se
, 0);
9502 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9504 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9505 struct rq
*rq
= cfs_rq
->rq
;
9506 unsigned long flags
;
9508 spin_lock_irqsave(&rq
->lock
, flags
);
9509 __set_se_shares(se
, shares
);
9510 spin_unlock_irqrestore(&rq
->lock
, flags
);
9513 static DEFINE_MUTEX(shares_mutex
);
9515 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9518 unsigned long flags
;
9521 * We can't change the weight of the root cgroup.
9526 if (shares
< MIN_SHARES
)
9527 shares
= MIN_SHARES
;
9528 else if (shares
> MAX_SHARES
)
9529 shares
= MAX_SHARES
;
9531 mutex_lock(&shares_mutex
);
9532 if (tg
->shares
== shares
)
9535 spin_lock_irqsave(&task_group_lock
, flags
);
9536 for_each_possible_cpu(i
)
9537 unregister_fair_sched_group(tg
, i
);
9538 list_del_rcu(&tg
->siblings
);
9539 spin_unlock_irqrestore(&task_group_lock
, flags
);
9541 /* wait for any ongoing reference to this group to finish */
9542 synchronize_sched();
9545 * Now we are free to modify the group's share on each cpu
9546 * w/o tripping rebalance_share or load_balance_fair.
9548 tg
->shares
= shares
;
9549 for_each_possible_cpu(i
) {
9553 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9554 set_se_shares(tg
->se
[i
], shares
);
9558 * Enable load balance activity on this group, by inserting it back on
9559 * each cpu's rq->leaf_cfs_rq_list.
9561 spin_lock_irqsave(&task_group_lock
, flags
);
9562 for_each_possible_cpu(i
)
9563 register_fair_sched_group(tg
, i
);
9564 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9565 spin_unlock_irqrestore(&task_group_lock
, flags
);
9567 mutex_unlock(&shares_mutex
);
9571 unsigned long sched_group_shares(struct task_group
*tg
)
9577 #ifdef CONFIG_RT_GROUP_SCHED
9579 * Ensure that the real time constraints are schedulable.
9581 static DEFINE_MUTEX(rt_constraints_mutex
);
9583 static unsigned long to_ratio(u64 period
, u64 runtime
)
9585 if (runtime
== RUNTIME_INF
)
9588 return div64_u64(runtime
<< 20, period
);
9591 /* Must be called with tasklist_lock held */
9592 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9594 struct task_struct
*g
, *p
;
9596 do_each_thread(g
, p
) {
9597 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9599 } while_each_thread(g
, p
);
9604 struct rt_schedulable_data
{
9605 struct task_group
*tg
;
9610 static int tg_schedulable(struct task_group
*tg
, void *data
)
9612 struct rt_schedulable_data
*d
= data
;
9613 struct task_group
*child
;
9614 unsigned long total
, sum
= 0;
9615 u64 period
, runtime
;
9617 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9618 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9621 period
= d
->rt_period
;
9622 runtime
= d
->rt_runtime
;
9625 #ifdef CONFIG_USER_SCHED
9626 if (tg
== &root_task_group
) {
9627 period
= global_rt_period();
9628 runtime
= global_rt_runtime();
9633 * Cannot have more runtime than the period.
9635 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9639 * Ensure we don't starve existing RT tasks.
9641 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9644 total
= to_ratio(period
, runtime
);
9647 * Nobody can have more than the global setting allows.
9649 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9653 * The sum of our children's runtime should not exceed our own.
9655 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9656 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9657 runtime
= child
->rt_bandwidth
.rt_runtime
;
9659 if (child
== d
->tg
) {
9660 period
= d
->rt_period
;
9661 runtime
= d
->rt_runtime
;
9664 sum
+= to_ratio(period
, runtime
);
9673 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9675 struct rt_schedulable_data data
= {
9677 .rt_period
= period
,
9678 .rt_runtime
= runtime
,
9681 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9684 static int tg_set_bandwidth(struct task_group
*tg
,
9685 u64 rt_period
, u64 rt_runtime
)
9689 mutex_lock(&rt_constraints_mutex
);
9690 read_lock(&tasklist_lock
);
9691 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9695 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9696 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9697 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9699 for_each_possible_cpu(i
) {
9700 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9702 spin_lock(&rt_rq
->rt_runtime_lock
);
9703 rt_rq
->rt_runtime
= rt_runtime
;
9704 spin_unlock(&rt_rq
->rt_runtime_lock
);
9706 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9708 read_unlock(&tasklist_lock
);
9709 mutex_unlock(&rt_constraints_mutex
);
9714 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9716 u64 rt_runtime
, rt_period
;
9718 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9719 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9720 if (rt_runtime_us
< 0)
9721 rt_runtime
= RUNTIME_INF
;
9723 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9726 long sched_group_rt_runtime(struct task_group
*tg
)
9730 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9733 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9734 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9735 return rt_runtime_us
;
9738 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9740 u64 rt_runtime
, rt_period
;
9742 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9743 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9748 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9751 long sched_group_rt_period(struct task_group
*tg
)
9755 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9756 do_div(rt_period_us
, NSEC_PER_USEC
);
9757 return rt_period_us
;
9760 static int sched_rt_global_constraints(void)
9762 u64 runtime
, period
;
9765 if (sysctl_sched_rt_period
<= 0)
9768 runtime
= global_rt_runtime();
9769 period
= global_rt_period();
9772 * Sanity check on the sysctl variables.
9774 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9777 mutex_lock(&rt_constraints_mutex
);
9778 read_lock(&tasklist_lock
);
9779 ret
= __rt_schedulable(NULL
, 0, 0);
9780 read_unlock(&tasklist_lock
);
9781 mutex_unlock(&rt_constraints_mutex
);
9786 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9788 /* Don't accept realtime tasks when there is no way for them to run */
9789 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9795 #else /* !CONFIG_RT_GROUP_SCHED */
9796 static int sched_rt_global_constraints(void)
9798 unsigned long flags
;
9801 if (sysctl_sched_rt_period
<= 0)
9804 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9805 for_each_possible_cpu(i
) {
9806 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9808 spin_lock(&rt_rq
->rt_runtime_lock
);
9809 rt_rq
->rt_runtime
= global_rt_runtime();
9810 spin_unlock(&rt_rq
->rt_runtime_lock
);
9812 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9816 #endif /* CONFIG_RT_GROUP_SCHED */
9818 int sched_rt_handler(struct ctl_table
*table
, int write
,
9819 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9823 int old_period
, old_runtime
;
9824 static DEFINE_MUTEX(mutex
);
9827 old_period
= sysctl_sched_rt_period
;
9828 old_runtime
= sysctl_sched_rt_runtime
;
9830 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9832 if (!ret
&& write
) {
9833 ret
= sched_rt_global_constraints();
9835 sysctl_sched_rt_period
= old_period
;
9836 sysctl_sched_rt_runtime
= old_runtime
;
9838 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9839 def_rt_bandwidth
.rt_period
=
9840 ns_to_ktime(global_rt_period());
9843 mutex_unlock(&mutex
);
9848 #ifdef CONFIG_CGROUP_SCHED
9850 /* return corresponding task_group object of a cgroup */
9851 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9853 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9854 struct task_group
, css
);
9857 static struct cgroup_subsys_state
*
9858 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9860 struct task_group
*tg
, *parent
;
9862 if (!cgrp
->parent
) {
9863 /* This is early initialization for the top cgroup */
9864 return &init_task_group
.css
;
9867 parent
= cgroup_tg(cgrp
->parent
);
9868 tg
= sched_create_group(parent
);
9870 return ERR_PTR(-ENOMEM
);
9876 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9878 struct task_group
*tg
= cgroup_tg(cgrp
);
9880 sched_destroy_group(tg
);
9884 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9885 struct task_struct
*tsk
)
9887 #ifdef CONFIG_RT_GROUP_SCHED
9888 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9891 /* We don't support RT-tasks being in separate groups */
9892 if (tsk
->sched_class
!= &fair_sched_class
)
9900 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9901 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9903 sched_move_task(tsk
);
9906 #ifdef CONFIG_FAIR_GROUP_SCHED
9907 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9910 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9913 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9915 struct task_group
*tg
= cgroup_tg(cgrp
);
9917 return (u64
) tg
->shares
;
9919 #endif /* CONFIG_FAIR_GROUP_SCHED */
9921 #ifdef CONFIG_RT_GROUP_SCHED
9922 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9925 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9928 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9930 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9933 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9936 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9939 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9941 return sched_group_rt_period(cgroup_tg(cgrp
));
9943 #endif /* CONFIG_RT_GROUP_SCHED */
9945 static struct cftype cpu_files
[] = {
9946 #ifdef CONFIG_FAIR_GROUP_SCHED
9949 .read_u64
= cpu_shares_read_u64
,
9950 .write_u64
= cpu_shares_write_u64
,
9953 #ifdef CONFIG_RT_GROUP_SCHED
9955 .name
= "rt_runtime_us",
9956 .read_s64
= cpu_rt_runtime_read
,
9957 .write_s64
= cpu_rt_runtime_write
,
9960 .name
= "rt_period_us",
9961 .read_u64
= cpu_rt_period_read_uint
,
9962 .write_u64
= cpu_rt_period_write_uint
,
9967 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9969 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9972 struct cgroup_subsys cpu_cgroup_subsys
= {
9974 .create
= cpu_cgroup_create
,
9975 .destroy
= cpu_cgroup_destroy
,
9976 .can_attach
= cpu_cgroup_can_attach
,
9977 .attach
= cpu_cgroup_attach
,
9978 .populate
= cpu_cgroup_populate
,
9979 .subsys_id
= cpu_cgroup_subsys_id
,
9983 #endif /* CONFIG_CGROUP_SCHED */
9985 #ifdef CONFIG_CGROUP_CPUACCT
9988 * CPU accounting code for task groups.
9990 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9991 * (balbir@in.ibm.com).
9994 /* track cpu usage of a group of tasks and its child groups */
9996 struct cgroup_subsys_state css
;
9997 /* cpuusage holds pointer to a u64-type object on every cpu */
9999 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10000 struct cpuacct
*parent
;
10003 struct cgroup_subsys cpuacct_subsys
;
10005 /* return cpu accounting group corresponding to this container */
10006 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10008 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10009 struct cpuacct
, css
);
10012 /* return cpu accounting group to which this task belongs */
10013 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10015 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10016 struct cpuacct
, css
);
10019 /* create a new cpu accounting group */
10020 static struct cgroup_subsys_state
*cpuacct_create(
10021 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10023 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10029 ca
->cpuusage
= alloc_percpu(u64
);
10033 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10034 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10035 goto out_free_counters
;
10038 ca
->parent
= cgroup_ca(cgrp
->parent
);
10044 percpu_counter_destroy(&ca
->cpustat
[i
]);
10045 free_percpu(ca
->cpuusage
);
10049 return ERR_PTR(-ENOMEM
);
10052 /* destroy an existing cpu accounting group */
10054 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10056 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10059 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10060 percpu_counter_destroy(&ca
->cpustat
[i
]);
10061 free_percpu(ca
->cpuusage
);
10065 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10067 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10070 #ifndef CONFIG_64BIT
10072 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10074 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10076 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10084 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10086 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10088 #ifndef CONFIG_64BIT
10090 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10092 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10094 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10100 /* return total cpu usage (in nanoseconds) of a group */
10101 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10103 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10104 u64 totalcpuusage
= 0;
10107 for_each_present_cpu(i
)
10108 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10110 return totalcpuusage
;
10113 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10116 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10125 for_each_present_cpu(i
)
10126 cpuacct_cpuusage_write(ca
, i
, 0);
10132 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10133 struct seq_file
*m
)
10135 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10139 for_each_present_cpu(i
) {
10140 percpu
= cpuacct_cpuusage_read(ca
, i
);
10141 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10143 seq_printf(m
, "\n");
10147 static const char *cpuacct_stat_desc
[] = {
10148 [CPUACCT_STAT_USER
] = "user",
10149 [CPUACCT_STAT_SYSTEM
] = "system",
10152 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10153 struct cgroup_map_cb
*cb
)
10155 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10158 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10159 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10160 val
= cputime64_to_clock_t(val
);
10161 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10166 static struct cftype files
[] = {
10169 .read_u64
= cpuusage_read
,
10170 .write_u64
= cpuusage_write
,
10173 .name
= "usage_percpu",
10174 .read_seq_string
= cpuacct_percpu_seq_read
,
10178 .read_map
= cpuacct_stats_show
,
10182 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10184 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10188 * charge this task's execution time to its accounting group.
10190 * called with rq->lock held.
10192 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10194 struct cpuacct
*ca
;
10197 if (unlikely(!cpuacct_subsys
.active
))
10200 cpu
= task_cpu(tsk
);
10206 for (; ca
; ca
= ca
->parent
) {
10207 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10208 *cpuusage
+= cputime
;
10215 * Charge the system/user time to the task's accounting group.
10217 static void cpuacct_update_stats(struct task_struct
*tsk
,
10218 enum cpuacct_stat_index idx
, cputime_t val
)
10220 struct cpuacct
*ca
;
10222 if (unlikely(!cpuacct_subsys
.active
))
10229 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10235 struct cgroup_subsys cpuacct_subsys
= {
10237 .create
= cpuacct_create
,
10238 .destroy
= cpuacct_destroy
,
10239 .populate
= cpuacct_populate
,
10240 .subsys_id
= cpuacct_subsys_id
,
10242 #endif /* CONFIG_CGROUP_CPUACCT */