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
;
501 struct plist_head pushable_tasks
;
506 /* Nests inside the rq lock: */
507 spinlock_t rt_runtime_lock
;
509 #ifdef CONFIG_RT_GROUP_SCHED
510 unsigned long rt_nr_boosted
;
513 struct list_head leaf_rt_rq_list
;
514 struct task_group
*tg
;
515 struct sched_rt_entity
*rt_se
;
522 * We add the notion of a root-domain which will be used to define per-domain
523 * variables. Each exclusive cpuset essentially defines an island domain by
524 * fully partitioning the member cpus from any other cpuset. Whenever a new
525 * exclusive cpuset is created, we also create and attach a new root-domain
532 cpumask_var_t online
;
535 * The "RT overload" flag: it gets set if a CPU has more than
536 * one runnable RT task.
538 cpumask_var_t rto_mask
;
541 struct cpupri cpupri
;
543 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
545 * Preferred wake up cpu nominated by sched_mc balance that will be
546 * used when most cpus are idle in the system indicating overall very
547 * low system utilisation. Triggered at POWERSAVINGS_BALANCE_WAKEUP(2)
549 unsigned int sched_mc_preferred_wakeup_cpu
;
554 * By default the system creates a single root-domain with all cpus as
555 * members (mimicking the global state we have today).
557 static struct root_domain def_root_domain
;
562 * This is the main, per-CPU runqueue data structure.
564 * Locking rule: those places that want to lock multiple runqueues
565 * (such as the load balancing or the thread migration code), lock
566 * acquire operations must be ordered by ascending &runqueue.
573 * nr_running and cpu_load should be in the same cacheline because
574 * remote CPUs use both these fields when doing load calculation.
576 unsigned long nr_running
;
577 #define CPU_LOAD_IDX_MAX 5
578 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
580 unsigned long last_tick_seen
;
581 unsigned char in_nohz_recently
;
583 /* capture load from *all* tasks on this cpu: */
584 struct load_weight load
;
585 unsigned long nr_load_updates
;
587 u64 nr_migrations_in
;
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 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
;
1972 if (old_cpu
!= new_cpu
) {
1973 p
->se
.nr_migrations
++;
1974 new_rq
->nr_migrations_in
++;
1975 #ifdef CONFIG_SCHEDSTATS
1976 if (task_hot(p
, old_rq
->clock
, NULL
))
1977 schedstat_inc(p
, se
.nr_forced2_migrations
);
1980 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1981 new_cfsrq
->min_vruntime
;
1983 __set_task_cpu(p
, new_cpu
);
1986 struct migration_req
{
1987 struct list_head list
;
1989 struct task_struct
*task
;
1992 struct completion done
;
1996 * The task's runqueue lock must be held.
1997 * Returns true if you have to wait for migration thread.
2000 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2002 struct rq
*rq
= task_rq(p
);
2005 * If the task is not on a runqueue (and not running), then
2006 * it is sufficient to simply update the task's cpu field.
2008 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2009 set_task_cpu(p
, dest_cpu
);
2013 init_completion(&req
->done
);
2015 req
->dest_cpu
= dest_cpu
;
2016 list_add(&req
->list
, &rq
->migration_queue
);
2022 * wait_task_inactive - wait for a thread to unschedule.
2024 * If @match_state is nonzero, it's the @p->state value just checked and
2025 * not expected to change. If it changes, i.e. @p might have woken up,
2026 * then return zero. When we succeed in waiting for @p to be off its CPU,
2027 * we return a positive number (its total switch count). If a second call
2028 * a short while later returns the same number, the caller can be sure that
2029 * @p has remained unscheduled the whole time.
2031 * The caller must ensure that the task *will* unschedule sometime soon,
2032 * else this function might spin for a *long* time. This function can't
2033 * be called with interrupts off, or it may introduce deadlock with
2034 * smp_call_function() if an IPI is sent by the same process we are
2035 * waiting to become inactive.
2037 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2039 unsigned long flags
;
2046 * We do the initial early heuristics without holding
2047 * any task-queue locks at all. We'll only try to get
2048 * the runqueue lock when things look like they will
2054 * If the task is actively running on another CPU
2055 * still, just relax and busy-wait without holding
2058 * NOTE! Since we don't hold any locks, it's not
2059 * even sure that "rq" stays as the right runqueue!
2060 * But we don't care, since "task_running()" will
2061 * return false if the runqueue has changed and p
2062 * is actually now running somewhere else!
2064 while (task_running(rq
, p
)) {
2065 if (match_state
&& unlikely(p
->state
!= match_state
))
2071 * Ok, time to look more closely! We need the rq
2072 * lock now, to be *sure*. If we're wrong, we'll
2073 * just go back and repeat.
2075 rq
= task_rq_lock(p
, &flags
);
2076 trace_sched_wait_task(rq
, p
);
2077 running
= task_running(rq
, p
);
2078 on_rq
= p
->se
.on_rq
;
2080 if (!match_state
|| p
->state
== match_state
)
2081 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2082 task_rq_unlock(rq
, &flags
);
2085 * If it changed from the expected state, bail out now.
2087 if (unlikely(!ncsw
))
2091 * Was it really running after all now that we
2092 * checked with the proper locks actually held?
2094 * Oops. Go back and try again..
2096 if (unlikely(running
)) {
2102 * It's not enough that it's not actively running,
2103 * it must be off the runqueue _entirely_, and not
2106 * So if it was still runnable (but just not actively
2107 * running right now), it's preempted, and we should
2108 * yield - it could be a while.
2110 if (unlikely(on_rq
)) {
2111 schedule_timeout_uninterruptible(1);
2116 * Ahh, all good. It wasn't running, and it wasn't
2117 * runnable, which means that it will never become
2118 * running in the future either. We're all done!
2127 * kick_process - kick a running thread to enter/exit the kernel
2128 * @p: the to-be-kicked thread
2130 * Cause a process which is running on another CPU to enter
2131 * kernel-mode, without any delay. (to get signals handled.)
2133 * NOTE: this function doesnt have to take the runqueue lock,
2134 * because all it wants to ensure is that the remote task enters
2135 * the kernel. If the IPI races and the task has been migrated
2136 * to another CPU then no harm is done and the purpose has been
2139 void kick_process(struct task_struct
*p
)
2145 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2146 smp_send_reschedule(cpu
);
2151 * Return a low guess at the load of a migration-source cpu weighted
2152 * according to the scheduling class and "nice" value.
2154 * We want to under-estimate the load of migration sources, to
2155 * balance conservatively.
2157 static unsigned long source_load(int cpu
, int type
)
2159 struct rq
*rq
= cpu_rq(cpu
);
2160 unsigned long total
= weighted_cpuload(cpu
);
2162 if (type
== 0 || !sched_feat(LB_BIAS
))
2165 return min(rq
->cpu_load
[type
-1], total
);
2169 * Return a high guess at the load of a migration-target cpu weighted
2170 * according to the scheduling class and "nice" value.
2172 static unsigned long target_load(int cpu
, int type
)
2174 struct rq
*rq
= cpu_rq(cpu
);
2175 unsigned long total
= weighted_cpuload(cpu
);
2177 if (type
== 0 || !sched_feat(LB_BIAS
))
2180 return max(rq
->cpu_load
[type
-1], total
);
2184 * find_idlest_group finds and returns the least busy CPU group within the
2187 static struct sched_group
*
2188 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2190 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2191 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2192 int load_idx
= sd
->forkexec_idx
;
2193 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2196 unsigned long load
, avg_load
;
2200 /* Skip over this group if it has no CPUs allowed */
2201 if (!cpumask_intersects(sched_group_cpus(group
),
2205 local_group
= cpumask_test_cpu(this_cpu
,
2206 sched_group_cpus(group
));
2208 /* Tally up the load of all CPUs in the group */
2211 for_each_cpu(i
, sched_group_cpus(group
)) {
2212 /* Bias balancing toward cpus of our domain */
2214 load
= source_load(i
, load_idx
);
2216 load
= target_load(i
, load_idx
);
2221 /* Adjust by relative CPU power of the group */
2222 avg_load
= sg_div_cpu_power(group
,
2223 avg_load
* SCHED_LOAD_SCALE
);
2226 this_load
= avg_load
;
2228 } else if (avg_load
< min_load
) {
2229 min_load
= avg_load
;
2232 } while (group
= group
->next
, group
!= sd
->groups
);
2234 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2240 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2243 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
2245 unsigned long load
, min_load
= ULONG_MAX
;
2249 /* Traverse only the allowed CPUs */
2250 for_each_cpu_and(i
, sched_group_cpus(group
), &p
->cpus_allowed
) {
2251 load
= weighted_cpuload(i
);
2253 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2263 * sched_balance_self: balance the current task (running on cpu) in domains
2264 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2267 * Balance, ie. select the least loaded group.
2269 * Returns the target CPU number, or the same CPU if no balancing is needed.
2271 * preempt must be disabled.
2273 static int sched_balance_self(int cpu
, int flag
)
2275 struct task_struct
*t
= current
;
2276 struct sched_domain
*tmp
, *sd
= NULL
;
2278 for_each_domain(cpu
, tmp
) {
2280 * If power savings logic is enabled for a domain, stop there.
2282 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2284 if (tmp
->flags
& flag
)
2292 struct sched_group
*group
;
2293 int new_cpu
, weight
;
2295 if (!(sd
->flags
& flag
)) {
2300 group
= find_idlest_group(sd
, t
, cpu
);
2306 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
2307 if (new_cpu
== -1 || new_cpu
== cpu
) {
2308 /* Now try balancing at a lower domain level of cpu */
2313 /* Now try balancing at a lower domain level of new_cpu */
2315 weight
= cpumask_weight(sched_domain_span(sd
));
2317 for_each_domain(cpu
, tmp
) {
2318 if (weight
<= cpumask_weight(sched_domain_span(tmp
)))
2320 if (tmp
->flags
& flag
)
2323 /* while loop will break here if sd == NULL */
2329 #endif /* CONFIG_SMP */
2332 * task_oncpu_function_call - call a function on the cpu on which a task runs
2333 * @p: the task to evaluate
2334 * @func: the function to be called
2335 * @info: the function call argument
2337 * Calls the function @func when the task is currently running. This might
2338 * be on the current CPU, which just calls the function directly
2340 void task_oncpu_function_call(struct task_struct
*p
,
2341 void (*func
) (void *info
), void *info
)
2348 smp_call_function_single(cpu
, func
, info
, 1);
2353 * try_to_wake_up - wake up a thread
2354 * @p: the to-be-woken-up thread
2355 * @state: the mask of task states that can be woken
2356 * @sync: do a synchronous wakeup?
2358 * Put it on the run-queue if it's not already there. The "current"
2359 * thread is always on the run-queue (except when the actual
2360 * re-schedule is in progress), and as such you're allowed to do
2361 * the simpler "current->state = TASK_RUNNING" to mark yourself
2362 * runnable without the overhead of this.
2364 * returns failure only if the task is already active.
2366 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2368 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2369 unsigned long flags
;
2373 if (!sched_feat(SYNC_WAKEUPS
))
2377 if (sched_feat(LB_WAKEUP_UPDATE
) && !root_task_group_empty()) {
2378 struct sched_domain
*sd
;
2380 this_cpu
= raw_smp_processor_id();
2383 for_each_domain(this_cpu
, sd
) {
2384 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2393 rq
= task_rq_lock(p
, &flags
);
2394 update_rq_clock(rq
);
2395 old_state
= p
->state
;
2396 if (!(old_state
& state
))
2404 this_cpu
= smp_processor_id();
2407 if (unlikely(task_running(rq
, p
)))
2410 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2411 if (cpu
!= orig_cpu
) {
2412 set_task_cpu(p
, cpu
);
2413 task_rq_unlock(rq
, &flags
);
2414 /* might preempt at this point */
2415 rq
= task_rq_lock(p
, &flags
);
2416 old_state
= p
->state
;
2417 if (!(old_state
& state
))
2422 this_cpu
= smp_processor_id();
2426 #ifdef CONFIG_SCHEDSTATS
2427 schedstat_inc(rq
, ttwu_count
);
2428 if (cpu
== this_cpu
)
2429 schedstat_inc(rq
, ttwu_local
);
2431 struct sched_domain
*sd
;
2432 for_each_domain(this_cpu
, sd
) {
2433 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2434 schedstat_inc(sd
, ttwu_wake_remote
);
2439 #endif /* CONFIG_SCHEDSTATS */
2442 #endif /* CONFIG_SMP */
2443 schedstat_inc(p
, se
.nr_wakeups
);
2445 schedstat_inc(p
, se
.nr_wakeups_sync
);
2446 if (orig_cpu
!= cpu
)
2447 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2448 if (cpu
== this_cpu
)
2449 schedstat_inc(p
, se
.nr_wakeups_local
);
2451 schedstat_inc(p
, se
.nr_wakeups_remote
);
2452 activate_task(rq
, p
, 1);
2456 * Only attribute actual wakeups done by this task.
2458 if (!in_interrupt()) {
2459 struct sched_entity
*se
= ¤t
->se
;
2460 u64 sample
= se
->sum_exec_runtime
;
2462 if (se
->last_wakeup
)
2463 sample
-= se
->last_wakeup
;
2465 sample
-= se
->start_runtime
;
2466 update_avg(&se
->avg_wakeup
, sample
);
2468 se
->last_wakeup
= se
->sum_exec_runtime
;
2472 trace_sched_wakeup(rq
, p
, success
);
2473 check_preempt_curr(rq
, p
, sync
);
2475 p
->state
= TASK_RUNNING
;
2477 if (p
->sched_class
->task_wake_up
)
2478 p
->sched_class
->task_wake_up(rq
, p
);
2481 task_rq_unlock(rq
, &flags
);
2486 int wake_up_process(struct task_struct
*p
)
2488 return try_to_wake_up(p
, TASK_ALL
, 0);
2490 EXPORT_SYMBOL(wake_up_process
);
2492 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2494 return try_to_wake_up(p
, state
, 0);
2498 * Perform scheduler related setup for a newly forked process p.
2499 * p is forked by current.
2501 * __sched_fork() is basic setup used by init_idle() too:
2503 static void __sched_fork(struct task_struct
*p
)
2505 p
->se
.exec_start
= 0;
2506 p
->se
.sum_exec_runtime
= 0;
2507 p
->se
.prev_sum_exec_runtime
= 0;
2508 p
->se
.nr_migrations
= 0;
2509 p
->se
.last_wakeup
= 0;
2510 p
->se
.avg_overlap
= 0;
2511 p
->se
.start_runtime
= 0;
2512 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2514 #ifdef CONFIG_SCHEDSTATS
2515 p
->se
.wait_start
= 0;
2516 p
->se
.sum_sleep_runtime
= 0;
2517 p
->se
.sleep_start
= 0;
2518 p
->se
.block_start
= 0;
2519 p
->se
.sleep_max
= 0;
2520 p
->se
.block_max
= 0;
2522 p
->se
.slice_max
= 0;
2526 INIT_LIST_HEAD(&p
->rt
.run_list
);
2528 INIT_LIST_HEAD(&p
->se
.group_node
);
2530 #ifdef CONFIG_PREEMPT_NOTIFIERS
2531 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2535 * We mark the process as running here, but have not actually
2536 * inserted it onto the runqueue yet. This guarantees that
2537 * nobody will actually run it, and a signal or other external
2538 * event cannot wake it up and insert it on the runqueue either.
2540 p
->state
= TASK_RUNNING
;
2544 * fork()/clone()-time setup:
2546 void sched_fork(struct task_struct
*p
, int clone_flags
)
2548 int cpu
= get_cpu();
2553 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2555 set_task_cpu(p
, cpu
);
2558 * Make sure we do not leak PI boosting priority to the child:
2560 p
->prio
= current
->normal_prio
;
2561 if (!rt_prio(p
->prio
))
2562 p
->sched_class
= &fair_sched_class
;
2564 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2565 if (likely(sched_info_on()))
2566 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2568 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2571 #ifdef CONFIG_PREEMPT
2572 /* Want to start with kernel preemption disabled. */
2573 task_thread_info(p
)->preempt_count
= 1;
2575 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2581 * wake_up_new_task - wake up a newly created task for the first time.
2583 * This function will do some initial scheduler statistics housekeeping
2584 * that must be done for every newly created context, then puts the task
2585 * on the runqueue and wakes it.
2587 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2589 unsigned long flags
;
2592 rq
= task_rq_lock(p
, &flags
);
2593 BUG_ON(p
->state
!= TASK_RUNNING
);
2594 update_rq_clock(rq
);
2596 p
->prio
= effective_prio(p
);
2598 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2599 activate_task(rq
, p
, 0);
2602 * Let the scheduling class do new task startup
2603 * management (if any):
2605 p
->sched_class
->task_new(rq
, p
);
2608 trace_sched_wakeup_new(rq
, p
, 1);
2609 check_preempt_curr(rq
, p
, 0);
2611 if (p
->sched_class
->task_wake_up
)
2612 p
->sched_class
->task_wake_up(rq
, p
);
2614 task_rq_unlock(rq
, &flags
);
2617 #ifdef CONFIG_PREEMPT_NOTIFIERS
2620 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2621 * @notifier: notifier struct to register
2623 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2625 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2627 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2630 * preempt_notifier_unregister - no longer interested in preemption notifications
2631 * @notifier: notifier struct to unregister
2633 * This is safe to call from within a preemption notifier.
2635 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2637 hlist_del(¬ifier
->link
);
2639 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2641 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2643 struct preempt_notifier
*notifier
;
2644 struct hlist_node
*node
;
2646 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2647 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2651 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2652 struct task_struct
*next
)
2654 struct preempt_notifier
*notifier
;
2655 struct hlist_node
*node
;
2657 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2658 notifier
->ops
->sched_out(notifier
, next
);
2661 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2663 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2668 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2669 struct task_struct
*next
)
2673 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2676 * prepare_task_switch - prepare to switch tasks
2677 * @rq: the runqueue preparing to switch
2678 * @prev: the current task that is being switched out
2679 * @next: the task we are going to switch to.
2681 * This is called with the rq lock held and interrupts off. It must
2682 * be paired with a subsequent finish_task_switch after the context
2685 * prepare_task_switch sets up locking and calls architecture specific
2689 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2690 struct task_struct
*next
)
2692 fire_sched_out_preempt_notifiers(prev
, next
);
2693 prepare_lock_switch(rq
, next
);
2694 prepare_arch_switch(next
);
2698 * finish_task_switch - clean up after a task-switch
2699 * @rq: runqueue associated with task-switch
2700 * @prev: the thread we just switched away from.
2702 * finish_task_switch must be called after the context switch, paired
2703 * with a prepare_task_switch call before the context switch.
2704 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2705 * and do any other architecture-specific cleanup actions.
2707 * Note that we may have delayed dropping an mm in context_switch(). If
2708 * so, we finish that here outside of the runqueue lock. (Doing it
2709 * with the lock held can cause deadlocks; see schedule() for
2712 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2713 __releases(rq
->lock
)
2715 struct mm_struct
*mm
= rq
->prev_mm
;
2718 int post_schedule
= 0;
2720 if (current
->sched_class
->needs_post_schedule
)
2721 post_schedule
= current
->sched_class
->needs_post_schedule(rq
);
2727 * A task struct has one reference for the use as "current".
2728 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2729 * schedule one last time. The schedule call will never return, and
2730 * the scheduled task must drop that reference.
2731 * The test for TASK_DEAD must occur while the runqueue locks are
2732 * still held, otherwise prev could be scheduled on another cpu, die
2733 * there before we look at prev->state, and then the reference would
2735 * Manfred Spraul <manfred@colorfullife.com>
2737 prev_state
= prev
->state
;
2738 finish_arch_switch(prev
);
2739 perf_counter_task_sched_in(current
, cpu_of(rq
));
2740 finish_lock_switch(rq
, prev
);
2743 current
->sched_class
->post_schedule(rq
);
2746 fire_sched_in_preempt_notifiers(current
);
2749 if (unlikely(prev_state
== TASK_DEAD
)) {
2751 * Remove function-return probe instances associated with this
2752 * task and put them back on the free list.
2754 kprobe_flush_task(prev
);
2755 put_task_struct(prev
);
2760 * schedule_tail - first thing a freshly forked thread must call.
2761 * @prev: the thread we just switched away from.
2763 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2764 __releases(rq
->lock
)
2766 struct rq
*rq
= this_rq();
2768 finish_task_switch(rq
, prev
);
2769 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2770 /* In this case, finish_task_switch does not reenable preemption */
2773 if (current
->set_child_tid
)
2774 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2778 * context_switch - switch to the new MM and the new
2779 * thread's register state.
2782 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2783 struct task_struct
*next
)
2785 struct mm_struct
*mm
, *oldmm
;
2787 prepare_task_switch(rq
, prev
, next
);
2788 trace_sched_switch(rq
, prev
, next
);
2790 oldmm
= prev
->active_mm
;
2792 * For paravirt, this is coupled with an exit in switch_to to
2793 * combine the page table reload and the switch backend into
2796 arch_enter_lazy_cpu_mode();
2798 if (unlikely(!mm
)) {
2799 next
->active_mm
= oldmm
;
2800 atomic_inc(&oldmm
->mm_count
);
2801 enter_lazy_tlb(oldmm
, next
);
2803 switch_mm(oldmm
, mm
, next
);
2805 if (unlikely(!prev
->mm
)) {
2806 prev
->active_mm
= NULL
;
2807 rq
->prev_mm
= oldmm
;
2810 * Since the runqueue lock will be released by the next
2811 * task (which is an invalid locking op but in the case
2812 * of the scheduler it's an obvious special-case), so we
2813 * do an early lockdep release here:
2815 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2816 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2819 /* Here we just switch the register state and the stack. */
2820 switch_to(prev
, next
, prev
);
2824 * this_rq must be evaluated again because prev may have moved
2825 * CPUs since it called schedule(), thus the 'rq' on its stack
2826 * frame will be invalid.
2828 finish_task_switch(this_rq(), prev
);
2832 * nr_running, nr_uninterruptible and nr_context_switches:
2834 * externally visible scheduler statistics: current number of runnable
2835 * threads, current number of uninterruptible-sleeping threads, total
2836 * number of context switches performed since bootup.
2838 unsigned long nr_running(void)
2840 unsigned long i
, sum
= 0;
2842 for_each_online_cpu(i
)
2843 sum
+= cpu_rq(i
)->nr_running
;
2848 unsigned long nr_uninterruptible(void)
2850 unsigned long i
, sum
= 0;
2852 for_each_possible_cpu(i
)
2853 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2856 * Since we read the counters lockless, it might be slightly
2857 * inaccurate. Do not allow it to go below zero though:
2859 if (unlikely((long)sum
< 0))
2865 unsigned long long nr_context_switches(void)
2868 unsigned long long sum
= 0;
2870 for_each_possible_cpu(i
)
2871 sum
+= cpu_rq(i
)->nr_switches
;
2876 unsigned long nr_iowait(void)
2878 unsigned long i
, sum
= 0;
2880 for_each_possible_cpu(i
)
2881 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2886 unsigned long nr_active(void)
2888 unsigned long i
, running
= 0, uninterruptible
= 0;
2890 for_each_online_cpu(i
) {
2891 running
+= cpu_rq(i
)->nr_running
;
2892 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2895 if (unlikely((long)uninterruptible
< 0))
2896 uninterruptible
= 0;
2898 return running
+ uninterruptible
;
2902 * Externally visible per-cpu scheduler statistics:
2903 * cpu_nr_migrations(cpu) - number of migrations into that cpu
2905 u64
cpu_nr_migrations(int cpu
)
2907 return cpu_rq(cpu
)->nr_migrations_in
;
2911 * Update rq->cpu_load[] statistics. This function is usually called every
2912 * scheduler tick (TICK_NSEC).
2914 static void update_cpu_load(struct rq
*this_rq
)
2916 unsigned long this_load
= this_rq
->load
.weight
;
2919 this_rq
->nr_load_updates
++;
2921 /* Update our load: */
2922 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2923 unsigned long old_load
, new_load
;
2925 /* scale is effectively 1 << i now, and >> i divides by scale */
2927 old_load
= this_rq
->cpu_load
[i
];
2928 new_load
= this_load
;
2930 * Round up the averaging division if load is increasing. This
2931 * prevents us from getting stuck on 9 if the load is 10, for
2934 if (new_load
> old_load
)
2935 new_load
+= scale
-1;
2936 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2943 * double_rq_lock - safely lock two runqueues
2945 * Note this does not disable interrupts like task_rq_lock,
2946 * you need to do so manually before calling.
2948 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2949 __acquires(rq1
->lock
)
2950 __acquires(rq2
->lock
)
2952 BUG_ON(!irqs_disabled());
2954 spin_lock(&rq1
->lock
);
2955 __acquire(rq2
->lock
); /* Fake it out ;) */
2958 spin_lock(&rq1
->lock
);
2959 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2961 spin_lock(&rq2
->lock
);
2962 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2965 update_rq_clock(rq1
);
2966 update_rq_clock(rq2
);
2970 * double_rq_unlock - safely unlock two runqueues
2972 * Note this does not restore interrupts like task_rq_unlock,
2973 * you need to do so manually after calling.
2975 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2976 __releases(rq1
->lock
)
2977 __releases(rq2
->lock
)
2979 spin_unlock(&rq1
->lock
);
2981 spin_unlock(&rq2
->lock
);
2983 __release(rq2
->lock
);
2987 * If dest_cpu is allowed for this process, migrate the task to it.
2988 * This is accomplished by forcing the cpu_allowed mask to only
2989 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2990 * the cpu_allowed mask is restored.
2992 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2994 struct migration_req req
;
2995 unsigned long flags
;
2998 rq
= task_rq_lock(p
, &flags
);
2999 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3000 || unlikely(!cpu_active(dest_cpu
)))
3003 /* force the process onto the specified CPU */
3004 if (migrate_task(p
, dest_cpu
, &req
)) {
3005 /* Need to wait for migration thread (might exit: take ref). */
3006 struct task_struct
*mt
= rq
->migration_thread
;
3008 get_task_struct(mt
);
3009 task_rq_unlock(rq
, &flags
);
3010 wake_up_process(mt
);
3011 put_task_struct(mt
);
3012 wait_for_completion(&req
.done
);
3017 task_rq_unlock(rq
, &flags
);
3021 * sched_exec - execve() is a valuable balancing opportunity, because at
3022 * this point the task has the smallest effective memory and cache footprint.
3024 void sched_exec(void)
3026 int new_cpu
, this_cpu
= get_cpu();
3027 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3029 if (new_cpu
!= this_cpu
)
3030 sched_migrate_task(current
, new_cpu
);
3034 * pull_task - move a task from a remote runqueue to the local runqueue.
3035 * Both runqueues must be locked.
3037 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3038 struct rq
*this_rq
, int this_cpu
)
3040 deactivate_task(src_rq
, p
, 0);
3041 set_task_cpu(p
, this_cpu
);
3042 activate_task(this_rq
, p
, 0);
3044 * Note that idle threads have a prio of MAX_PRIO, for this test
3045 * to be always true for them.
3047 check_preempt_curr(this_rq
, p
, 0);
3051 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3054 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3055 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3058 int tsk_cache_hot
= 0;
3060 * We do not migrate tasks that are:
3061 * 1) running (obviously), or
3062 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3063 * 3) are cache-hot on their current CPU.
3065 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3066 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3071 if (task_running(rq
, p
)) {
3072 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3077 * Aggressive migration if:
3078 * 1) task is cache cold, or
3079 * 2) too many balance attempts have failed.
3082 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3083 if (!tsk_cache_hot
||
3084 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3085 #ifdef CONFIG_SCHEDSTATS
3086 if (tsk_cache_hot
) {
3087 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3088 schedstat_inc(p
, se
.nr_forced_migrations
);
3094 if (tsk_cache_hot
) {
3095 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3101 static unsigned long
3102 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3103 unsigned long max_load_move
, struct sched_domain
*sd
,
3104 enum cpu_idle_type idle
, int *all_pinned
,
3105 int *this_best_prio
, struct rq_iterator
*iterator
)
3107 int loops
= 0, pulled
= 0, pinned
= 0;
3108 struct task_struct
*p
;
3109 long rem_load_move
= max_load_move
;
3111 if (max_load_move
== 0)
3117 * Start the load-balancing iterator:
3119 p
= iterator
->start(iterator
->arg
);
3121 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3124 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3125 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3126 p
= iterator
->next(iterator
->arg
);
3130 pull_task(busiest
, p
, this_rq
, this_cpu
);
3132 rem_load_move
-= p
->se
.load
.weight
;
3134 #ifdef CONFIG_PREEMPT
3136 * NEWIDLE balancing is a source of latency, so preemptible kernels
3137 * will stop after the first task is pulled to minimize the critical
3140 if (idle
== CPU_NEWLY_IDLE
)
3145 * We only want to steal up to the prescribed amount of weighted load.
3147 if (rem_load_move
> 0) {
3148 if (p
->prio
< *this_best_prio
)
3149 *this_best_prio
= p
->prio
;
3150 p
= iterator
->next(iterator
->arg
);
3155 * Right now, this is one of only two places pull_task() is called,
3156 * so we can safely collect pull_task() stats here rather than
3157 * inside pull_task().
3159 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3162 *all_pinned
= pinned
;
3164 return max_load_move
- rem_load_move
;
3168 * move_tasks tries to move up to max_load_move weighted load from busiest to
3169 * this_rq, as part of a balancing operation within domain "sd".
3170 * Returns 1 if successful and 0 otherwise.
3172 * Called with both runqueues locked.
3174 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3175 unsigned long max_load_move
,
3176 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3179 const struct sched_class
*class = sched_class_highest
;
3180 unsigned long total_load_moved
= 0;
3181 int this_best_prio
= this_rq
->curr
->prio
;
3185 class->load_balance(this_rq
, this_cpu
, busiest
,
3186 max_load_move
- total_load_moved
,
3187 sd
, idle
, all_pinned
, &this_best_prio
);
3188 class = class->next
;
3190 #ifdef CONFIG_PREEMPT
3192 * NEWIDLE balancing is a source of latency, so preemptible
3193 * kernels will stop after the first task is pulled to minimize
3194 * the critical section.
3196 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3199 } while (class && max_load_move
> total_load_moved
);
3201 return total_load_moved
> 0;
3205 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3206 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3207 struct rq_iterator
*iterator
)
3209 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3213 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3214 pull_task(busiest
, p
, this_rq
, this_cpu
);
3216 * Right now, this is only the second place pull_task()
3217 * is called, so we can safely collect pull_task()
3218 * stats here rather than inside pull_task().
3220 schedstat_inc(sd
, lb_gained
[idle
]);
3224 p
= iterator
->next(iterator
->arg
);
3231 * move_one_task tries to move exactly one task from busiest to this_rq, as
3232 * part of active balancing operations within "domain".
3233 * Returns 1 if successful and 0 otherwise.
3235 * Called with both runqueues locked.
3237 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3238 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3240 const struct sched_class
*class;
3242 for (class = sched_class_highest
; class; class = class->next
)
3243 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3248 /********** Helpers for find_busiest_group ************************/
3250 * sd_lb_stats - Structure to store the statistics of a sched_domain
3251 * during load balancing.
3253 struct sd_lb_stats
{
3254 struct sched_group
*busiest
; /* Busiest group in this sd */
3255 struct sched_group
*this; /* Local group in this sd */
3256 unsigned long total_load
; /* Total load of all groups in sd */
3257 unsigned long total_pwr
; /* Total power of all groups in sd */
3258 unsigned long avg_load
; /* Average load across all groups in sd */
3260 /** Statistics of this group */
3261 unsigned long this_load
;
3262 unsigned long this_load_per_task
;
3263 unsigned long this_nr_running
;
3265 /* Statistics of the busiest group */
3266 unsigned long max_load
;
3267 unsigned long busiest_load_per_task
;
3268 unsigned long busiest_nr_running
;
3270 int group_imb
; /* Is there imbalance in this sd */
3271 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3272 int power_savings_balance
; /* Is powersave balance needed for this sd */
3273 struct sched_group
*group_min
; /* Least loaded group in sd */
3274 struct sched_group
*group_leader
; /* Group which relieves group_min */
3275 unsigned long min_load_per_task
; /* load_per_task in group_min */
3276 unsigned long leader_nr_running
; /* Nr running of group_leader */
3277 unsigned long min_nr_running
; /* Nr running of group_min */
3282 * sg_lb_stats - stats of a sched_group required for load_balancing
3284 struct sg_lb_stats
{
3285 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3286 unsigned long group_load
; /* Total load over the CPUs of the group */
3287 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3288 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3289 unsigned long group_capacity
;
3290 int group_imb
; /* Is there an imbalance in the group ? */
3294 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3295 * @group: The group whose first cpu is to be returned.
3297 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3299 return cpumask_first(sched_group_cpus(group
));
3303 * get_sd_load_idx - Obtain the load index for a given sched domain.
3304 * @sd: The sched_domain whose load_idx is to be obtained.
3305 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3307 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3308 enum cpu_idle_type idle
)
3314 load_idx
= sd
->busy_idx
;
3317 case CPU_NEWLY_IDLE
:
3318 load_idx
= sd
->newidle_idx
;
3321 load_idx
= sd
->idle_idx
;
3329 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3331 * init_sd_power_savings_stats - Initialize power savings statistics for
3332 * the given sched_domain, during load balancing.
3334 * @sd: Sched domain whose power-savings statistics are to be initialized.
3335 * @sds: Variable containing the statistics for sd.
3336 * @idle: Idle status of the CPU at which we're performing load-balancing.
3338 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3339 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3342 * Busy processors will not participate in power savings
3345 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3346 sds
->power_savings_balance
= 0;
3348 sds
->power_savings_balance
= 1;
3349 sds
->min_nr_running
= ULONG_MAX
;
3350 sds
->leader_nr_running
= 0;
3355 * update_sd_power_savings_stats - Update the power saving stats for a
3356 * sched_domain while performing load balancing.
3358 * @group: sched_group belonging to the sched_domain under consideration.
3359 * @sds: Variable containing the statistics of the sched_domain
3360 * @local_group: Does group contain the CPU for which we're performing
3362 * @sgs: Variable containing the statistics of the group.
3364 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3365 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3368 if (!sds
->power_savings_balance
)
3372 * If the local group is idle or completely loaded
3373 * no need to do power savings balance at this domain
3375 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3376 !sds
->this_nr_running
))
3377 sds
->power_savings_balance
= 0;
3380 * If a group is already running at full capacity or idle,
3381 * don't include that group in power savings calculations
3383 if (!sds
->power_savings_balance
||
3384 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3385 !sgs
->sum_nr_running
)
3389 * Calculate the group which has the least non-idle load.
3390 * This is the group from where we need to pick up the load
3393 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3394 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3395 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3396 sds
->group_min
= group
;
3397 sds
->min_nr_running
= sgs
->sum_nr_running
;
3398 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3399 sgs
->sum_nr_running
;
3403 * Calculate the group which is almost near its
3404 * capacity but still has some space to pick up some load
3405 * from other group and save more power
3407 if (sgs
->sum_nr_running
> sgs
->group_capacity
- 1)
3410 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3411 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3412 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3413 sds
->group_leader
= group
;
3414 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3419 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3420 * @sds: Variable containing the statistics of the sched_domain
3421 * under consideration.
3422 * @this_cpu: Cpu at which we're currently performing load-balancing.
3423 * @imbalance: Variable to store the imbalance.
3426 * Check if we have potential to perform some power-savings balance.
3427 * If yes, set the busiest group to be the least loaded group in the
3428 * sched_domain, so that it's CPUs can be put to idle.
3430 * Returns 1 if there is potential to perform power-savings balance.
3433 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3434 int this_cpu
, unsigned long *imbalance
)
3436 if (!sds
->power_savings_balance
)
3439 if (sds
->this != sds
->group_leader
||
3440 sds
->group_leader
== sds
->group_min
)
3443 *imbalance
= sds
->min_load_per_task
;
3444 sds
->busiest
= sds
->group_min
;
3446 if (sched_mc_power_savings
>= POWERSAVINGS_BALANCE_WAKEUP
) {
3447 cpu_rq(this_cpu
)->rd
->sched_mc_preferred_wakeup_cpu
=
3448 group_first_cpu(sds
->group_leader
);
3454 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3455 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3456 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3461 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3462 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3467 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3468 int this_cpu
, unsigned long *imbalance
)
3472 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3476 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3477 * @group: sched_group whose statistics are to be updated.
3478 * @this_cpu: Cpu for which load balance is currently performed.
3479 * @idle: Idle status of this_cpu
3480 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3481 * @sd_idle: Idle status of the sched_domain containing group.
3482 * @local_group: Does group contain this_cpu.
3483 * @cpus: Set of cpus considered for load balancing.
3484 * @balance: Should we balance.
3485 * @sgs: variable to hold the statistics for this group.
3487 static inline void update_sg_lb_stats(struct sched_group
*group
, int this_cpu
,
3488 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3489 int local_group
, const struct cpumask
*cpus
,
3490 int *balance
, struct sg_lb_stats
*sgs
)
3492 unsigned long load
, max_cpu_load
, min_cpu_load
;
3494 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3495 unsigned long sum_avg_load_per_task
;
3496 unsigned long avg_load_per_task
;
3499 balance_cpu
= group_first_cpu(group
);
3501 /* Tally up the load of all CPUs in the group */
3502 sum_avg_load_per_task
= avg_load_per_task
= 0;
3504 min_cpu_load
= ~0UL;
3506 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3507 struct rq
*rq
= cpu_rq(i
);
3509 if (*sd_idle
&& rq
->nr_running
)
3512 /* Bias balancing toward cpus of our domain */
3514 if (idle_cpu(i
) && !first_idle_cpu
) {
3519 load
= target_load(i
, load_idx
);
3521 load
= source_load(i
, load_idx
);
3522 if (load
> max_cpu_load
)
3523 max_cpu_load
= load
;
3524 if (min_cpu_load
> load
)
3525 min_cpu_load
= load
;
3528 sgs
->group_load
+= load
;
3529 sgs
->sum_nr_running
+= rq
->nr_running
;
3530 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3532 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3536 * First idle cpu or the first cpu(busiest) in this sched group
3537 * is eligible for doing load balancing at this and above
3538 * domains. In the newly idle case, we will allow all the cpu's
3539 * to do the newly idle load balance.
3541 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3542 balance_cpu
!= this_cpu
&& balance
) {
3547 /* Adjust by relative CPU power of the group */
3548 sgs
->avg_load
= sg_div_cpu_power(group
,
3549 sgs
->group_load
* SCHED_LOAD_SCALE
);
3553 * Consider the group unbalanced when the imbalance is larger
3554 * than the average weight of two tasks.
3556 * APZ: with cgroup the avg task weight can vary wildly and
3557 * might not be a suitable number - should we keep a
3558 * normalized nr_running number somewhere that negates
3561 avg_load_per_task
= sg_div_cpu_power(group
,
3562 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3564 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3567 sgs
->group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3572 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3573 * @sd: sched_domain whose statistics are to be updated.
3574 * @this_cpu: Cpu for which load balance is currently performed.
3575 * @idle: Idle status of this_cpu
3576 * @sd_idle: Idle status of the sched_domain containing group.
3577 * @cpus: Set of cpus considered for load balancing.
3578 * @balance: Should we balance.
3579 * @sds: variable to hold the statistics for this sched_domain.
3581 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3582 enum cpu_idle_type idle
, int *sd_idle
,
3583 const struct cpumask
*cpus
, int *balance
,
3584 struct sd_lb_stats
*sds
)
3586 struct sched_group
*group
= sd
->groups
;
3587 struct sg_lb_stats sgs
;
3590 init_sd_power_savings_stats(sd
, sds
, idle
);
3591 load_idx
= get_sd_load_idx(sd
, idle
);
3596 local_group
= cpumask_test_cpu(this_cpu
,
3597 sched_group_cpus(group
));
3598 memset(&sgs
, 0, sizeof(sgs
));
3599 update_sg_lb_stats(group
, this_cpu
, idle
, load_idx
, sd_idle
,
3600 local_group
, cpus
, balance
, &sgs
);
3602 if (local_group
&& balance
&& !(*balance
))
3605 sds
->total_load
+= sgs
.group_load
;
3606 sds
->total_pwr
+= group
->__cpu_power
;
3609 sds
->this_load
= sgs
.avg_load
;
3611 sds
->this_nr_running
= sgs
.sum_nr_running
;
3612 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3613 } else if (sgs
.avg_load
> sds
->max_load
&&
3614 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3616 sds
->max_load
= sgs
.avg_load
;
3617 sds
->busiest
= group
;
3618 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3619 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3620 sds
->group_imb
= sgs
.group_imb
;
3623 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3624 group
= group
->next
;
3625 } while (group
!= sd
->groups
);
3630 * fix_small_imbalance - Calculate the minor imbalance that exists
3631 * amongst the groups of a sched_domain, during
3633 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3634 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3635 * @imbalance: Variable to store the imbalance.
3637 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3638 int this_cpu
, unsigned long *imbalance
)
3640 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3641 unsigned int imbn
= 2;
3643 if (sds
->this_nr_running
) {
3644 sds
->this_load_per_task
/= sds
->this_nr_running
;
3645 if (sds
->busiest_load_per_task
>
3646 sds
->this_load_per_task
)
3649 sds
->this_load_per_task
=
3650 cpu_avg_load_per_task(this_cpu
);
3652 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3653 sds
->busiest_load_per_task
* imbn
) {
3654 *imbalance
= sds
->busiest_load_per_task
;
3659 * OK, we don't have enough imbalance to justify moving tasks,
3660 * however we may be able to increase total CPU power used by
3664 pwr_now
+= sds
->busiest
->__cpu_power
*
3665 min(sds
->busiest_load_per_task
, sds
->max_load
);
3666 pwr_now
+= sds
->this->__cpu_power
*
3667 min(sds
->this_load_per_task
, sds
->this_load
);
3668 pwr_now
/= SCHED_LOAD_SCALE
;
3670 /* Amount of load we'd subtract */
3671 tmp
= sg_div_cpu_power(sds
->busiest
,
3672 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3673 if (sds
->max_load
> tmp
)
3674 pwr_move
+= sds
->busiest
->__cpu_power
*
3675 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3677 /* Amount of load we'd add */
3678 if (sds
->max_load
* sds
->busiest
->__cpu_power
<
3679 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3680 tmp
= sg_div_cpu_power(sds
->this,
3681 sds
->max_load
* sds
->busiest
->__cpu_power
);
3683 tmp
= sg_div_cpu_power(sds
->this,
3684 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
);
3685 pwr_move
+= sds
->this->__cpu_power
*
3686 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3687 pwr_move
/= SCHED_LOAD_SCALE
;
3689 /* Move if we gain throughput */
3690 if (pwr_move
> pwr_now
)
3691 *imbalance
= sds
->busiest_load_per_task
;
3695 * calculate_imbalance - Calculate the amount of imbalance present within the
3696 * groups of a given sched_domain during load balance.
3697 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3698 * @this_cpu: Cpu for which currently load balance is being performed.
3699 * @imbalance: The variable to store the imbalance.
3701 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3702 unsigned long *imbalance
)
3704 unsigned long max_pull
;
3706 * In the presence of smp nice balancing, certain scenarios can have
3707 * max load less than avg load(as we skip the groups at or below
3708 * its cpu_power, while calculating max_load..)
3710 if (sds
->max_load
< sds
->avg_load
) {
3712 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3715 /* Don't want to pull so many tasks that a group would go idle */
3716 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3717 sds
->max_load
- sds
->busiest_load_per_task
);
3719 /* How much load to actually move to equalise the imbalance */
3720 *imbalance
= min(max_pull
* sds
->busiest
->__cpu_power
,
3721 (sds
->avg_load
- sds
->this_load
) * sds
->this->__cpu_power
)
3725 * if *imbalance is less than the average load per runnable task
3726 * there is no gaurantee that any tasks will be moved so we'll have
3727 * a think about bumping its value to force at least one task to be
3730 if (*imbalance
< sds
->busiest_load_per_task
)
3731 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3734 /******* find_busiest_group() helpers end here *********************/
3737 * find_busiest_group - Returns the busiest group within the sched_domain
3738 * if there is an imbalance. If there isn't an imbalance, and
3739 * the user has opted for power-savings, it returns a group whose
3740 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3741 * such a group exists.
3743 * Also calculates the amount of weighted load which should be moved
3744 * to restore balance.
3746 * @sd: The sched_domain whose busiest group is to be returned.
3747 * @this_cpu: The cpu for which load balancing is currently being performed.
3748 * @imbalance: Variable which stores amount of weighted load which should
3749 * be moved to restore balance/put a group to idle.
3750 * @idle: The idle status of this_cpu.
3751 * @sd_idle: The idleness of sd
3752 * @cpus: The set of CPUs under consideration for load-balancing.
3753 * @balance: Pointer to a variable indicating if this_cpu
3754 * is the appropriate cpu to perform load balancing at this_level.
3756 * Returns: - the busiest group if imbalance exists.
3757 * - If no imbalance and user has opted for power-savings balance,
3758 * return the least loaded group whose CPUs can be
3759 * put to idle by rebalancing its tasks onto our group.
3761 static struct sched_group
*
3762 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3763 unsigned long *imbalance
, enum cpu_idle_type idle
,
3764 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
3766 struct sd_lb_stats sds
;
3768 memset(&sds
, 0, sizeof(sds
));
3771 * Compute the various statistics relavent for load balancing at
3774 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
3777 /* Cases where imbalance does not exist from POV of this_cpu */
3778 /* 1) this_cpu is not the appropriate cpu to perform load balancing
3780 * 2) There is no busy sibling group to pull from.
3781 * 3) This group is the busiest group.
3782 * 4) This group is more busy than the avg busieness at this
3784 * 5) The imbalance is within the specified limit.
3785 * 6) Any rebalance would lead to ping-pong
3787 if (balance
&& !(*balance
))
3790 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
3793 if (sds
.this_load
>= sds
.max_load
)
3796 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
3798 if (sds
.this_load
>= sds
.avg_load
)
3801 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
3804 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
3806 sds
.busiest_load_per_task
=
3807 min(sds
.busiest_load_per_task
, sds
.avg_load
);
3810 * We're trying to get all the cpus to the average_load, so we don't
3811 * want to push ourselves above the average load, nor do we wish to
3812 * reduce the max loaded cpu below the average load, as either of these
3813 * actions would just result in more rebalancing later, and ping-pong
3814 * tasks around. Thus we look for the minimum possible imbalance.
3815 * Negative imbalances (*we* are more loaded than anyone else) will
3816 * be counted as no imbalance for these purposes -- we can't fix that
3817 * by pulling tasks to us. Be careful of negative numbers as they'll
3818 * appear as very large values with unsigned longs.
3820 if (sds
.max_load
<= sds
.busiest_load_per_task
)
3823 /* Looks like there is an imbalance. Compute it */
3824 calculate_imbalance(&sds
, this_cpu
, imbalance
);
3829 * There is no obvious imbalance. But check if we can do some balancing
3832 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
3840 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3843 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3844 unsigned long imbalance
, const struct cpumask
*cpus
)
3846 struct rq
*busiest
= NULL
, *rq
;
3847 unsigned long max_load
= 0;
3850 for_each_cpu(i
, sched_group_cpus(group
)) {
3853 if (!cpumask_test_cpu(i
, cpus
))
3857 wl
= weighted_cpuload(i
);
3859 if (rq
->nr_running
== 1 && wl
> imbalance
)
3862 if (wl
> max_load
) {
3872 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3873 * so long as it is large enough.
3875 #define MAX_PINNED_INTERVAL 512
3877 /* Working cpumask for load_balance and load_balance_newidle. */
3878 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
3881 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3882 * tasks if there is an imbalance.
3884 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3885 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3888 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3889 struct sched_group
*group
;
3890 unsigned long imbalance
;
3892 unsigned long flags
;
3893 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
3895 cpumask_setall(cpus
);
3898 * When power savings policy is enabled for the parent domain, idle
3899 * sibling can pick up load irrespective of busy siblings. In this case,
3900 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3901 * portraying it as CPU_NOT_IDLE.
3903 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3904 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3907 schedstat_inc(sd
, lb_count
[idle
]);
3911 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3918 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3922 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3924 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3928 BUG_ON(busiest
== this_rq
);
3930 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3933 if (busiest
->nr_running
> 1) {
3935 * Attempt to move tasks. If find_busiest_group has found
3936 * an imbalance but busiest->nr_running <= 1, the group is
3937 * still unbalanced. ld_moved simply stays zero, so it is
3938 * correctly treated as an imbalance.
3940 local_irq_save(flags
);
3941 double_rq_lock(this_rq
, busiest
);
3942 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3943 imbalance
, sd
, idle
, &all_pinned
);
3944 double_rq_unlock(this_rq
, busiest
);
3945 local_irq_restore(flags
);
3948 * some other cpu did the load balance for us.
3950 if (ld_moved
&& this_cpu
!= smp_processor_id())
3951 resched_cpu(this_cpu
);
3953 /* All tasks on this runqueue were pinned by CPU affinity */
3954 if (unlikely(all_pinned
)) {
3955 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
3956 if (!cpumask_empty(cpus
))
3963 schedstat_inc(sd
, lb_failed
[idle
]);
3964 sd
->nr_balance_failed
++;
3966 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3968 spin_lock_irqsave(&busiest
->lock
, flags
);
3970 /* don't kick the migration_thread, if the curr
3971 * task on busiest cpu can't be moved to this_cpu
3973 if (!cpumask_test_cpu(this_cpu
,
3974 &busiest
->curr
->cpus_allowed
)) {
3975 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3977 goto out_one_pinned
;
3980 if (!busiest
->active_balance
) {
3981 busiest
->active_balance
= 1;
3982 busiest
->push_cpu
= this_cpu
;
3985 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3987 wake_up_process(busiest
->migration_thread
);
3990 * We've kicked active balancing, reset the failure
3993 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3996 sd
->nr_balance_failed
= 0;
3998 if (likely(!active_balance
)) {
3999 /* We were unbalanced, so reset the balancing interval */
4000 sd
->balance_interval
= sd
->min_interval
;
4003 * If we've begun active balancing, start to back off. This
4004 * case may not be covered by the all_pinned logic if there
4005 * is only 1 task on the busy runqueue (because we don't call
4008 if (sd
->balance_interval
< sd
->max_interval
)
4009 sd
->balance_interval
*= 2;
4012 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4013 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4019 schedstat_inc(sd
, lb_balanced
[idle
]);
4021 sd
->nr_balance_failed
= 0;
4024 /* tune up the balancing interval */
4025 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4026 (sd
->balance_interval
< sd
->max_interval
))
4027 sd
->balance_interval
*= 2;
4029 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4030 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4041 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4042 * tasks if there is an imbalance.
4044 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4045 * this_rq is locked.
4048 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4050 struct sched_group
*group
;
4051 struct rq
*busiest
= NULL
;
4052 unsigned long imbalance
;
4056 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4058 cpumask_setall(cpus
);
4061 * When power savings policy is enabled for the parent domain, idle
4062 * sibling can pick up load irrespective of busy siblings. In this case,
4063 * let the state of idle sibling percolate up as IDLE, instead of
4064 * portraying it as CPU_NOT_IDLE.
4066 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4067 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4070 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4072 update_shares_locked(this_rq
, sd
);
4073 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4074 &sd_idle
, cpus
, NULL
);
4076 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4080 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4082 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4086 BUG_ON(busiest
== this_rq
);
4088 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4091 if (busiest
->nr_running
> 1) {
4092 /* Attempt to move tasks */
4093 double_lock_balance(this_rq
, busiest
);
4094 /* this_rq->clock is already updated */
4095 update_rq_clock(busiest
);
4096 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4097 imbalance
, sd
, CPU_NEWLY_IDLE
,
4099 double_unlock_balance(this_rq
, busiest
);
4101 if (unlikely(all_pinned
)) {
4102 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4103 if (!cpumask_empty(cpus
))
4109 int active_balance
= 0;
4111 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4112 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4113 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4116 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4119 if (sd
->nr_balance_failed
++ < 2)
4123 * The only task running in a non-idle cpu can be moved to this
4124 * cpu in an attempt to completely freeup the other CPU
4125 * package. The same method used to move task in load_balance()
4126 * have been extended for load_balance_newidle() to speedup
4127 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4129 * The package power saving logic comes from
4130 * find_busiest_group(). If there are no imbalance, then
4131 * f_b_g() will return NULL. However when sched_mc={1,2} then
4132 * f_b_g() will select a group from which a running task may be
4133 * pulled to this cpu in order to make the other package idle.
4134 * If there is no opportunity to make a package idle and if
4135 * there are no imbalance, then f_b_g() will return NULL and no
4136 * action will be taken in load_balance_newidle().
4138 * Under normal task pull operation due to imbalance, there
4139 * will be more than one task in the source run queue and
4140 * move_tasks() will succeed. ld_moved will be true and this
4141 * active balance code will not be triggered.
4144 /* Lock busiest in correct order while this_rq is held */
4145 double_lock_balance(this_rq
, busiest
);
4148 * don't kick the migration_thread, if the curr
4149 * task on busiest cpu can't be moved to this_cpu
4151 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4152 double_unlock_balance(this_rq
, busiest
);
4157 if (!busiest
->active_balance
) {
4158 busiest
->active_balance
= 1;
4159 busiest
->push_cpu
= this_cpu
;
4163 double_unlock_balance(this_rq
, busiest
);
4165 * Should not call ttwu while holding a rq->lock
4167 spin_unlock(&this_rq
->lock
);
4169 wake_up_process(busiest
->migration_thread
);
4170 spin_lock(&this_rq
->lock
);
4173 sd
->nr_balance_failed
= 0;
4175 update_shares_locked(this_rq
, sd
);
4179 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4180 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4181 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4183 sd
->nr_balance_failed
= 0;
4189 * idle_balance is called by schedule() if this_cpu is about to become
4190 * idle. Attempts to pull tasks from other CPUs.
4192 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4194 struct sched_domain
*sd
;
4195 int pulled_task
= 0;
4196 unsigned long next_balance
= jiffies
+ HZ
;
4198 for_each_domain(this_cpu
, sd
) {
4199 unsigned long interval
;
4201 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4204 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4205 /* If we've pulled tasks over stop searching: */
4206 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4209 interval
= msecs_to_jiffies(sd
->balance_interval
);
4210 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4211 next_balance
= sd
->last_balance
+ interval
;
4215 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4217 * We are going idle. next_balance may be set based on
4218 * a busy processor. So reset next_balance.
4220 this_rq
->next_balance
= next_balance
;
4225 * active_load_balance is run by migration threads. It pushes running tasks
4226 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4227 * running on each physical CPU where possible, and avoids physical /
4228 * logical imbalances.
4230 * Called with busiest_rq locked.
4232 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4234 int target_cpu
= busiest_rq
->push_cpu
;
4235 struct sched_domain
*sd
;
4236 struct rq
*target_rq
;
4238 /* Is there any task to move? */
4239 if (busiest_rq
->nr_running
<= 1)
4242 target_rq
= cpu_rq(target_cpu
);
4245 * This condition is "impossible", if it occurs
4246 * we need to fix it. Originally reported by
4247 * Bjorn Helgaas on a 128-cpu setup.
4249 BUG_ON(busiest_rq
== target_rq
);
4251 /* move a task from busiest_rq to target_rq */
4252 double_lock_balance(busiest_rq
, target_rq
);
4253 update_rq_clock(busiest_rq
);
4254 update_rq_clock(target_rq
);
4256 /* Search for an sd spanning us and the target CPU. */
4257 for_each_domain(target_cpu
, sd
) {
4258 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4259 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4264 schedstat_inc(sd
, alb_count
);
4266 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4268 schedstat_inc(sd
, alb_pushed
);
4270 schedstat_inc(sd
, alb_failed
);
4272 double_unlock_balance(busiest_rq
, target_rq
);
4277 atomic_t load_balancer
;
4278 cpumask_var_t cpu_mask
;
4279 } nohz ____cacheline_aligned
= {
4280 .load_balancer
= ATOMIC_INIT(-1),
4284 * This routine will try to nominate the ilb (idle load balancing)
4285 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4286 * load balancing on behalf of all those cpus. If all the cpus in the system
4287 * go into this tickless mode, then there will be no ilb owner (as there is
4288 * no need for one) and all the cpus will sleep till the next wakeup event
4291 * For the ilb owner, tick is not stopped. And this tick will be used
4292 * for idle load balancing. ilb owner will still be part of
4295 * While stopping the tick, this cpu will become the ilb owner if there
4296 * is no other owner. And will be the owner till that cpu becomes busy
4297 * or if all cpus in the system stop their ticks at which point
4298 * there is no need for ilb owner.
4300 * When the ilb owner becomes busy, it nominates another owner, during the
4301 * next busy scheduler_tick()
4303 int select_nohz_load_balancer(int stop_tick
)
4305 int cpu
= smp_processor_id();
4308 cpu_rq(cpu
)->in_nohz_recently
= 1;
4310 if (!cpu_active(cpu
)) {
4311 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4315 * If we are going offline and still the leader,
4318 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4324 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4326 /* time for ilb owner also to sleep */
4327 if (cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4328 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4329 atomic_set(&nohz
.load_balancer
, -1);
4333 if (atomic_read(&nohz
.load_balancer
) == -1) {
4334 /* make me the ilb owner */
4335 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4337 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4340 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4343 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4345 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4346 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4353 static DEFINE_SPINLOCK(balancing
);
4356 * It checks each scheduling domain to see if it is due to be balanced,
4357 * and initiates a balancing operation if so.
4359 * Balancing parameters are set up in arch_init_sched_domains.
4361 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4364 struct rq
*rq
= cpu_rq(cpu
);
4365 unsigned long interval
;
4366 struct sched_domain
*sd
;
4367 /* Earliest time when we have to do rebalance again */
4368 unsigned long next_balance
= jiffies
+ 60*HZ
;
4369 int update_next_balance
= 0;
4372 for_each_domain(cpu
, sd
) {
4373 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4376 interval
= sd
->balance_interval
;
4377 if (idle
!= CPU_IDLE
)
4378 interval
*= sd
->busy_factor
;
4380 /* scale ms to jiffies */
4381 interval
= msecs_to_jiffies(interval
);
4382 if (unlikely(!interval
))
4384 if (interval
> HZ
*NR_CPUS
/10)
4385 interval
= HZ
*NR_CPUS
/10;
4387 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4389 if (need_serialize
) {
4390 if (!spin_trylock(&balancing
))
4394 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4395 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4397 * We've pulled tasks over so either we're no
4398 * longer idle, or one of our SMT siblings is
4401 idle
= CPU_NOT_IDLE
;
4403 sd
->last_balance
= jiffies
;
4406 spin_unlock(&balancing
);
4408 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4409 next_balance
= sd
->last_balance
+ interval
;
4410 update_next_balance
= 1;
4414 * Stop the load balance at this level. There is another
4415 * CPU in our sched group which is doing load balancing more
4423 * next_balance will be updated only when there is a need.
4424 * When the cpu is attached to null domain for ex, it will not be
4427 if (likely(update_next_balance
))
4428 rq
->next_balance
= next_balance
;
4432 * run_rebalance_domains is triggered when needed from the scheduler tick.
4433 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4434 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4436 static void run_rebalance_domains(struct softirq_action
*h
)
4438 int this_cpu
= smp_processor_id();
4439 struct rq
*this_rq
= cpu_rq(this_cpu
);
4440 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4441 CPU_IDLE
: CPU_NOT_IDLE
;
4443 rebalance_domains(this_cpu
, idle
);
4447 * If this cpu is the owner for idle load balancing, then do the
4448 * balancing on behalf of the other idle cpus whose ticks are
4451 if (this_rq
->idle_at_tick
&&
4452 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4456 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4457 if (balance_cpu
== this_cpu
)
4461 * If this cpu gets work to do, stop the load balancing
4462 * work being done for other cpus. Next load
4463 * balancing owner will pick it up.
4468 rebalance_domains(balance_cpu
, CPU_IDLE
);
4470 rq
= cpu_rq(balance_cpu
);
4471 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4472 this_rq
->next_balance
= rq
->next_balance
;
4478 static inline int on_null_domain(int cpu
)
4480 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4484 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4486 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4487 * idle load balancing owner or decide to stop the periodic load balancing,
4488 * if the whole system is idle.
4490 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4494 * If we were in the nohz mode recently and busy at the current
4495 * scheduler tick, then check if we need to nominate new idle
4498 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4499 rq
->in_nohz_recently
= 0;
4501 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4502 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4503 atomic_set(&nohz
.load_balancer
, -1);
4506 if (atomic_read(&nohz
.load_balancer
) == -1) {
4508 * simple selection for now: Nominate the
4509 * first cpu in the nohz list to be the next
4512 * TBD: Traverse the sched domains and nominate
4513 * the nearest cpu in the nohz.cpu_mask.
4515 int ilb
= cpumask_first(nohz
.cpu_mask
);
4517 if (ilb
< nr_cpu_ids
)
4523 * If this cpu is idle and doing idle load balancing for all the
4524 * cpus with ticks stopped, is it time for that to stop?
4526 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4527 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4533 * If this cpu is idle and the idle load balancing is done by
4534 * someone else, then no need raise the SCHED_SOFTIRQ
4536 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4537 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4540 /* Don't need to rebalance while attached to NULL domain */
4541 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4542 likely(!on_null_domain(cpu
)))
4543 raise_softirq(SCHED_SOFTIRQ
);
4546 #else /* CONFIG_SMP */
4549 * on UP we do not need to balance between CPUs:
4551 static inline void idle_balance(int cpu
, struct rq
*rq
)
4557 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4559 EXPORT_PER_CPU_SYMBOL(kstat
);
4562 * Return any ns on the sched_clock that have not yet been accounted in
4563 * @p in case that task is currently running.
4565 * Called with task_rq_lock() held on @rq.
4567 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4571 if (task_current(rq
, p
)) {
4572 update_rq_clock(rq
);
4573 ns
= rq
->clock
- p
->se
.exec_start
;
4581 unsigned long long task_delta_exec(struct task_struct
*p
)
4583 unsigned long flags
;
4587 rq
= task_rq_lock(p
, &flags
);
4588 ns
= do_task_delta_exec(p
, rq
);
4589 task_rq_unlock(rq
, &flags
);
4595 * Return accounted runtime for the task.
4596 * In case the task is currently running, return the runtime plus current's
4597 * pending runtime that have not been accounted yet.
4599 unsigned long long task_sched_runtime(struct task_struct
*p
)
4601 unsigned long flags
;
4605 rq
= task_rq_lock(p
, &flags
);
4606 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4607 task_rq_unlock(rq
, &flags
);
4613 * Return sum_exec_runtime for the thread group.
4614 * In case the task is currently running, return the sum plus current's
4615 * pending runtime that have not been accounted yet.
4617 * Note that the thread group might have other running tasks as well,
4618 * so the return value not includes other pending runtime that other
4619 * running tasks might have.
4621 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
4623 struct task_cputime totals
;
4624 unsigned long flags
;
4628 rq
= task_rq_lock(p
, &flags
);
4629 thread_group_cputime(p
, &totals
);
4630 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4631 task_rq_unlock(rq
, &flags
);
4637 * Account user cpu time to a process.
4638 * @p: the process that the cpu time gets accounted to
4639 * @cputime: the cpu time spent in user space since the last update
4640 * @cputime_scaled: cputime scaled by cpu frequency
4642 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
4643 cputime_t cputime_scaled
)
4645 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4648 /* Add user time to process. */
4649 p
->utime
= cputime_add(p
->utime
, cputime
);
4650 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4651 account_group_user_time(p
, cputime
);
4653 /* Add user time to cpustat. */
4654 tmp
= cputime_to_cputime64(cputime
);
4655 if (TASK_NICE(p
) > 0)
4656 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4658 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4660 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
4661 /* Account for user time used */
4662 acct_update_integrals(p
);
4666 * Account guest cpu time to a process.
4667 * @p: the process that the cpu time gets accounted to
4668 * @cputime: the cpu time spent in virtual machine since the last update
4669 * @cputime_scaled: cputime scaled by cpu frequency
4671 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
4672 cputime_t cputime_scaled
)
4675 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4677 tmp
= cputime_to_cputime64(cputime
);
4679 /* Add guest time to process. */
4680 p
->utime
= cputime_add(p
->utime
, cputime
);
4681 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
4682 account_group_user_time(p
, cputime
);
4683 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4685 /* Add guest time to cpustat. */
4686 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4687 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4691 * Account system cpu time to a process.
4692 * @p: the process that the cpu time gets accounted to
4693 * @hardirq_offset: the offset to subtract from hardirq_count()
4694 * @cputime: the cpu time spent in kernel space since the last update
4695 * @cputime_scaled: cputime scaled by cpu frequency
4697 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4698 cputime_t cputime
, cputime_t cputime_scaled
)
4700 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4703 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4704 account_guest_time(p
, cputime
, cputime_scaled
);
4708 /* Add system time to process. */
4709 p
->stime
= cputime_add(p
->stime
, cputime
);
4710 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
4711 account_group_system_time(p
, cputime
);
4713 /* Add system time to cpustat. */
4714 tmp
= cputime_to_cputime64(cputime
);
4715 if (hardirq_count() - hardirq_offset
)
4716 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4717 else if (softirq_count())
4718 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4720 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4722 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
4724 /* Account for system time used */
4725 acct_update_integrals(p
);
4729 * Account for involuntary wait time.
4730 * @steal: the cpu time spent in involuntary wait
4732 void account_steal_time(cputime_t cputime
)
4734 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4735 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4737 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
4741 * Account for idle time.
4742 * @cputime: the cpu time spent in idle wait
4744 void account_idle_time(cputime_t cputime
)
4746 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4747 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
4748 struct rq
*rq
= this_rq();
4750 if (atomic_read(&rq
->nr_iowait
) > 0)
4751 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
4753 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
4756 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
4759 * Account a single tick of cpu time.
4760 * @p: the process that the cpu time gets accounted to
4761 * @user_tick: indicates if the tick is a user or a system tick
4763 void account_process_tick(struct task_struct
*p
, int user_tick
)
4765 cputime_t one_jiffy
= jiffies_to_cputime(1);
4766 cputime_t one_jiffy_scaled
= cputime_to_scaled(one_jiffy
);
4767 struct rq
*rq
= this_rq();
4770 account_user_time(p
, one_jiffy
, one_jiffy_scaled
);
4771 else if (p
!= rq
->idle
)
4772 account_system_time(p
, HARDIRQ_OFFSET
, one_jiffy
,
4775 account_idle_time(one_jiffy
);
4779 * Account multiple ticks of steal time.
4780 * @p: the process from which the cpu time has been stolen
4781 * @ticks: number of stolen ticks
4783 void account_steal_ticks(unsigned long ticks
)
4785 account_steal_time(jiffies_to_cputime(ticks
));
4789 * Account multiple ticks of idle time.
4790 * @ticks: number of stolen ticks
4792 void account_idle_ticks(unsigned long ticks
)
4794 account_idle_time(jiffies_to_cputime(ticks
));
4800 * Use precise platform statistics if available:
4802 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4803 cputime_t
task_utime(struct task_struct
*p
)
4808 cputime_t
task_stime(struct task_struct
*p
)
4813 cputime_t
task_utime(struct task_struct
*p
)
4815 clock_t utime
= cputime_to_clock_t(p
->utime
),
4816 total
= utime
+ cputime_to_clock_t(p
->stime
);
4820 * Use CFS's precise accounting:
4822 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4826 do_div(temp
, total
);
4828 utime
= (clock_t)temp
;
4830 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4831 return p
->prev_utime
;
4834 cputime_t
task_stime(struct task_struct
*p
)
4839 * Use CFS's precise accounting. (we subtract utime from
4840 * the total, to make sure the total observed by userspace
4841 * grows monotonically - apps rely on that):
4843 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4844 cputime_to_clock_t(task_utime(p
));
4847 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4849 return p
->prev_stime
;
4853 inline cputime_t
task_gtime(struct task_struct
*p
)
4859 * This function gets called by the timer code, with HZ frequency.
4860 * We call it with interrupts disabled.
4862 * It also gets called by the fork code, when changing the parent's
4865 void scheduler_tick(void)
4867 int cpu
= smp_processor_id();
4868 struct rq
*rq
= cpu_rq(cpu
);
4869 struct task_struct
*curr
= rq
->curr
;
4873 spin_lock(&rq
->lock
);
4874 update_rq_clock(rq
);
4875 update_cpu_load(rq
);
4876 curr
->sched_class
->task_tick(rq
, curr
, 0);
4877 perf_counter_task_tick(curr
, cpu
);
4878 spin_unlock(&rq
->lock
);
4881 rq
->idle_at_tick
= idle_cpu(cpu
);
4882 trigger_load_balance(rq
, cpu
);
4886 notrace
unsigned long get_parent_ip(unsigned long addr
)
4888 if (in_lock_functions(addr
)) {
4889 addr
= CALLER_ADDR2
;
4890 if (in_lock_functions(addr
))
4891 addr
= CALLER_ADDR3
;
4896 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4897 defined(CONFIG_PREEMPT_TRACER))
4899 void __kprobes
add_preempt_count(int val
)
4901 #ifdef CONFIG_DEBUG_PREEMPT
4905 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4908 preempt_count() += val
;
4909 #ifdef CONFIG_DEBUG_PREEMPT
4911 * Spinlock count overflowing soon?
4913 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4916 if (preempt_count() == val
)
4917 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4919 EXPORT_SYMBOL(add_preempt_count
);
4921 void __kprobes
sub_preempt_count(int val
)
4923 #ifdef CONFIG_DEBUG_PREEMPT
4927 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4930 * Is the spinlock portion underflowing?
4932 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4933 !(preempt_count() & PREEMPT_MASK
)))
4937 if (preempt_count() == val
)
4938 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4939 preempt_count() -= val
;
4941 EXPORT_SYMBOL(sub_preempt_count
);
4946 * Print scheduling while atomic bug:
4948 static noinline
void __schedule_bug(struct task_struct
*prev
)
4950 struct pt_regs
*regs
= get_irq_regs();
4952 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4953 prev
->comm
, prev
->pid
, preempt_count());
4955 debug_show_held_locks(prev
);
4957 if (irqs_disabled())
4958 print_irqtrace_events(prev
);
4967 * Various schedule()-time debugging checks and statistics:
4969 static inline void schedule_debug(struct task_struct
*prev
)
4972 * Test if we are atomic. Since do_exit() needs to call into
4973 * schedule() atomically, we ignore that path for now.
4974 * Otherwise, whine if we are scheduling when we should not be.
4976 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4977 __schedule_bug(prev
);
4979 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4981 schedstat_inc(this_rq(), sched_count
);
4982 #ifdef CONFIG_SCHEDSTATS
4983 if (unlikely(prev
->lock_depth
>= 0)) {
4984 schedstat_inc(this_rq(), bkl_count
);
4985 schedstat_inc(prev
, sched_info
.bkl_count
);
4990 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
4992 if (prev
->state
== TASK_RUNNING
) {
4993 u64 runtime
= prev
->se
.sum_exec_runtime
;
4995 runtime
-= prev
->se
.prev_sum_exec_runtime
;
4996 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
4999 * In order to avoid avg_overlap growing stale when we are
5000 * indeed overlapping and hence not getting put to sleep, grow
5001 * the avg_overlap on preemption.
5003 * We use the average preemption runtime because that
5004 * correlates to the amount of cache footprint a task can
5007 update_avg(&prev
->se
.avg_overlap
, runtime
);
5009 prev
->sched_class
->put_prev_task(rq
, prev
);
5013 * Pick up the highest-prio task:
5015 static inline struct task_struct
*
5016 pick_next_task(struct rq
*rq
)
5018 const struct sched_class
*class;
5019 struct task_struct
*p
;
5022 * Optimization: we know that if all tasks are in
5023 * the fair class we can call that function directly:
5025 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5026 p
= fair_sched_class
.pick_next_task(rq
);
5031 class = sched_class_highest
;
5033 p
= class->pick_next_task(rq
);
5037 * Will never be NULL as the idle class always
5038 * returns a non-NULL p:
5040 class = class->next
;
5045 * schedule() is the main scheduler function.
5047 asmlinkage
void __sched
__schedule(void)
5049 struct task_struct
*prev
, *next
;
5050 unsigned long *switch_count
;
5054 cpu
= smp_processor_id();
5058 switch_count
= &prev
->nivcsw
;
5060 release_kernel_lock(prev
);
5061 need_resched_nonpreemptible
:
5063 schedule_debug(prev
);
5065 if (sched_feat(HRTICK
))
5068 spin_lock_irq(&rq
->lock
);
5069 update_rq_clock(rq
);
5070 clear_tsk_need_resched(prev
);
5072 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5073 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5074 prev
->state
= TASK_RUNNING
;
5076 deactivate_task(rq
, prev
, 1);
5077 switch_count
= &prev
->nvcsw
;
5081 if (prev
->sched_class
->pre_schedule
)
5082 prev
->sched_class
->pre_schedule(rq
, prev
);
5085 if (unlikely(!rq
->nr_running
))
5086 idle_balance(cpu
, rq
);
5088 put_prev_task(rq
, prev
);
5089 next
= pick_next_task(rq
);
5091 if (likely(prev
!= next
)) {
5092 sched_info_switch(prev
, next
);
5093 perf_counter_task_sched_out(prev
, cpu
);
5099 context_switch(rq
, prev
, next
); /* unlocks the rq */
5101 * the context switch might have flipped the stack from under
5102 * us, hence refresh the local variables.
5104 cpu
= smp_processor_id();
5107 spin_unlock_irq(&rq
->lock
);
5109 if (unlikely(reacquire_kernel_lock(current
) < 0))
5110 goto need_resched_nonpreemptible
;
5113 asmlinkage
void __sched
schedule(void)
5118 preempt_enable_no_resched();
5119 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
5122 EXPORT_SYMBOL(schedule
);
5126 * Look out! "owner" is an entirely speculative pointer
5127 * access and not reliable.
5129 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5134 if (!sched_feat(OWNER_SPIN
))
5137 #ifdef CONFIG_DEBUG_PAGEALLOC
5139 * Need to access the cpu field knowing that
5140 * DEBUG_PAGEALLOC could have unmapped it if
5141 * the mutex owner just released it and exited.
5143 if (probe_kernel_address(&owner
->cpu
, cpu
))
5150 * Even if the access succeeded (likely case),
5151 * the cpu field may no longer be valid.
5153 if (cpu
>= nr_cpumask_bits
)
5157 * We need to validate that we can do a
5158 * get_cpu() and that we have the percpu area.
5160 if (!cpu_online(cpu
))
5167 * Owner changed, break to re-assess state.
5169 if (lock
->owner
!= owner
)
5173 * Is that owner really running on that cpu?
5175 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5185 #ifdef CONFIG_PREEMPT
5187 * this is the entry point to schedule() from in-kernel preemption
5188 * off of preempt_enable. Kernel preemptions off return from interrupt
5189 * occur there and call schedule directly.
5191 asmlinkage
void __sched
preempt_schedule(void)
5193 struct thread_info
*ti
= current_thread_info();
5196 * If there is a non-zero preempt_count or interrupts are disabled,
5197 * we do not want to preempt the current task. Just return..
5199 if (likely(ti
->preempt_count
|| irqs_disabled()))
5203 add_preempt_count(PREEMPT_ACTIVE
);
5205 sub_preempt_count(PREEMPT_ACTIVE
);
5208 * Check again in case we missed a preemption opportunity
5209 * between schedule and now.
5212 } while (need_resched());
5214 EXPORT_SYMBOL(preempt_schedule
);
5217 * this is the entry point to schedule() from kernel preemption
5218 * off of irq context.
5219 * Note, that this is called and return with irqs disabled. This will
5220 * protect us against recursive calling from irq.
5222 asmlinkage
void __sched
preempt_schedule_irq(void)
5224 struct thread_info
*ti
= current_thread_info();
5226 /* Catch callers which need to be fixed */
5227 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5230 add_preempt_count(PREEMPT_ACTIVE
);
5233 local_irq_disable();
5234 sub_preempt_count(PREEMPT_ACTIVE
);
5237 * Check again in case we missed a preemption opportunity
5238 * between schedule and now.
5241 } while (need_resched());
5244 #endif /* CONFIG_PREEMPT */
5246 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
5249 return try_to_wake_up(curr
->private, mode
, sync
);
5251 EXPORT_SYMBOL(default_wake_function
);
5254 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5255 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5256 * number) then we wake all the non-exclusive tasks and one exclusive task.
5258 * There are circumstances in which we can try to wake a task which has already
5259 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5260 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5262 void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5263 int nr_exclusive
, int sync
, void *key
)
5265 wait_queue_t
*curr
, *next
;
5267 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5268 unsigned flags
= curr
->flags
;
5270 if (curr
->func(curr
, mode
, sync
, key
) &&
5271 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5277 * __wake_up - wake up threads blocked on a waitqueue.
5279 * @mode: which threads
5280 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5281 * @key: is directly passed to the wakeup function
5283 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5284 int nr_exclusive
, void *key
)
5286 unsigned long flags
;
5288 spin_lock_irqsave(&q
->lock
, flags
);
5289 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5290 spin_unlock_irqrestore(&q
->lock
, flags
);
5292 EXPORT_SYMBOL(__wake_up
);
5295 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5297 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5299 __wake_up_common(q
, mode
, 1, 0, NULL
);
5302 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5304 __wake_up_common(q
, mode
, 1, 0, key
);
5308 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5310 * @mode: which threads
5311 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5312 * @key: opaque value to be passed to wakeup targets
5314 * The sync wakeup differs that the waker knows that it will schedule
5315 * away soon, so while the target thread will be woken up, it will not
5316 * be migrated to another CPU - ie. the two threads are 'synchronized'
5317 * with each other. This can prevent needless bouncing between CPUs.
5319 * On UP it can prevent extra preemption.
5321 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5322 int nr_exclusive
, void *key
)
5324 unsigned long flags
;
5330 if (unlikely(!nr_exclusive
))
5333 spin_lock_irqsave(&q
->lock
, flags
);
5334 __wake_up_common(q
, mode
, nr_exclusive
, sync
, key
);
5335 spin_unlock_irqrestore(&q
->lock
, flags
);
5337 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5340 * __wake_up_sync - see __wake_up_sync_key()
5342 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5344 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5346 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5349 * complete: - signals a single thread waiting on this completion
5350 * @x: holds the state of this particular completion
5352 * This will wake up a single thread waiting on this completion. Threads will be
5353 * awakened in the same order in which they were queued.
5355 * See also complete_all(), wait_for_completion() and related routines.
5357 void complete(struct completion
*x
)
5359 unsigned long flags
;
5361 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5363 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5364 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5366 EXPORT_SYMBOL(complete
);
5369 * complete_all: - signals all threads waiting on this completion
5370 * @x: holds the state of this particular completion
5372 * This will wake up all threads waiting on this particular completion event.
5374 void complete_all(struct completion
*x
)
5376 unsigned long flags
;
5378 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5379 x
->done
+= UINT_MAX
/2;
5380 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5381 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5383 EXPORT_SYMBOL(complete_all
);
5385 static inline long __sched
5386 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5389 DECLARE_WAITQUEUE(wait
, current
);
5391 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5392 __add_wait_queue_tail(&x
->wait
, &wait
);
5394 if (signal_pending_state(state
, current
)) {
5395 timeout
= -ERESTARTSYS
;
5398 __set_current_state(state
);
5399 spin_unlock_irq(&x
->wait
.lock
);
5400 timeout
= schedule_timeout(timeout
);
5401 spin_lock_irq(&x
->wait
.lock
);
5402 } while (!x
->done
&& timeout
);
5403 __remove_wait_queue(&x
->wait
, &wait
);
5408 return timeout
?: 1;
5412 wait_for_common(struct completion
*x
, long timeout
, int state
)
5416 spin_lock_irq(&x
->wait
.lock
);
5417 timeout
= do_wait_for_common(x
, timeout
, state
);
5418 spin_unlock_irq(&x
->wait
.lock
);
5423 * wait_for_completion: - waits for completion of a task
5424 * @x: holds the state of this particular completion
5426 * This waits to be signaled for completion of a specific task. It is NOT
5427 * interruptible and there is no timeout.
5429 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5430 * and interrupt capability. Also see complete().
5432 void __sched
wait_for_completion(struct completion
*x
)
5434 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5436 EXPORT_SYMBOL(wait_for_completion
);
5439 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5440 * @x: holds the state of this particular completion
5441 * @timeout: timeout value in jiffies
5443 * This waits for either a completion of a specific task to be signaled or for a
5444 * specified timeout to expire. The timeout is in jiffies. It is not
5447 unsigned long __sched
5448 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5450 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5452 EXPORT_SYMBOL(wait_for_completion_timeout
);
5455 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5456 * @x: holds the state of this particular completion
5458 * This waits for completion of a specific task to be signaled. It is
5461 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5463 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5464 if (t
== -ERESTARTSYS
)
5468 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5471 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5472 * @x: holds the state of this particular completion
5473 * @timeout: timeout value in jiffies
5475 * This waits for either a completion of a specific task to be signaled or for a
5476 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5478 unsigned long __sched
5479 wait_for_completion_interruptible_timeout(struct completion
*x
,
5480 unsigned long timeout
)
5482 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5484 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5487 * wait_for_completion_killable: - waits for completion of a task (killable)
5488 * @x: holds the state of this particular completion
5490 * This waits to be signaled for completion of a specific task. It can be
5491 * interrupted by a kill signal.
5493 int __sched
wait_for_completion_killable(struct completion
*x
)
5495 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5496 if (t
== -ERESTARTSYS
)
5500 EXPORT_SYMBOL(wait_for_completion_killable
);
5503 * try_wait_for_completion - try to decrement a completion without blocking
5504 * @x: completion structure
5506 * Returns: 0 if a decrement cannot be done without blocking
5507 * 1 if a decrement succeeded.
5509 * If a completion is being used as a counting completion,
5510 * attempt to decrement the counter without blocking. This
5511 * enables us to avoid waiting if the resource the completion
5512 * is protecting is not available.
5514 bool try_wait_for_completion(struct completion
*x
)
5518 spin_lock_irq(&x
->wait
.lock
);
5523 spin_unlock_irq(&x
->wait
.lock
);
5526 EXPORT_SYMBOL(try_wait_for_completion
);
5529 * completion_done - Test to see if a completion has any waiters
5530 * @x: completion structure
5532 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5533 * 1 if there are no waiters.
5536 bool completion_done(struct completion
*x
)
5540 spin_lock_irq(&x
->wait
.lock
);
5543 spin_unlock_irq(&x
->wait
.lock
);
5546 EXPORT_SYMBOL(completion_done
);
5549 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5551 unsigned long flags
;
5554 init_waitqueue_entry(&wait
, current
);
5556 __set_current_state(state
);
5558 spin_lock_irqsave(&q
->lock
, flags
);
5559 __add_wait_queue(q
, &wait
);
5560 spin_unlock(&q
->lock
);
5561 timeout
= schedule_timeout(timeout
);
5562 spin_lock_irq(&q
->lock
);
5563 __remove_wait_queue(q
, &wait
);
5564 spin_unlock_irqrestore(&q
->lock
, flags
);
5569 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
5571 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5573 EXPORT_SYMBOL(interruptible_sleep_on
);
5576 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5578 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5580 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5582 void __sched
sleep_on(wait_queue_head_t
*q
)
5584 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5586 EXPORT_SYMBOL(sleep_on
);
5588 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5590 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5592 EXPORT_SYMBOL(sleep_on_timeout
);
5594 #ifdef CONFIG_RT_MUTEXES
5597 * rt_mutex_setprio - set the current priority of a task
5599 * @prio: prio value (kernel-internal form)
5601 * This function changes the 'effective' priority of a task. It does
5602 * not touch ->normal_prio like __setscheduler().
5604 * Used by the rt_mutex code to implement priority inheritance logic.
5606 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5608 unsigned long flags
;
5609 int oldprio
, on_rq
, running
;
5611 const struct sched_class
*prev_class
= p
->sched_class
;
5613 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5615 rq
= task_rq_lock(p
, &flags
);
5616 update_rq_clock(rq
);
5619 on_rq
= p
->se
.on_rq
;
5620 running
= task_current(rq
, p
);
5622 dequeue_task(rq
, p
, 0);
5624 p
->sched_class
->put_prev_task(rq
, p
);
5627 p
->sched_class
= &rt_sched_class
;
5629 p
->sched_class
= &fair_sched_class
;
5634 p
->sched_class
->set_curr_task(rq
);
5636 enqueue_task(rq
, p
, 0);
5638 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5640 task_rq_unlock(rq
, &flags
);
5645 void set_user_nice(struct task_struct
*p
, long nice
)
5647 int old_prio
, delta
, on_rq
;
5648 unsigned long flags
;
5651 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5654 * We have to be careful, if called from sys_setpriority(),
5655 * the task might be in the middle of scheduling on another CPU.
5657 rq
= task_rq_lock(p
, &flags
);
5658 update_rq_clock(rq
);
5660 * The RT priorities are set via sched_setscheduler(), but we still
5661 * allow the 'normal' nice value to be set - but as expected
5662 * it wont have any effect on scheduling until the task is
5663 * SCHED_FIFO/SCHED_RR:
5665 if (task_has_rt_policy(p
)) {
5666 p
->static_prio
= NICE_TO_PRIO(nice
);
5669 on_rq
= p
->se
.on_rq
;
5671 dequeue_task(rq
, p
, 0);
5673 p
->static_prio
= NICE_TO_PRIO(nice
);
5676 p
->prio
= effective_prio(p
);
5677 delta
= p
->prio
- old_prio
;
5680 enqueue_task(rq
, p
, 0);
5682 * If the task increased its priority or is running and
5683 * lowered its priority, then reschedule its CPU:
5685 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5686 resched_task(rq
->curr
);
5689 task_rq_unlock(rq
, &flags
);
5691 EXPORT_SYMBOL(set_user_nice
);
5694 * can_nice - check if a task can reduce its nice value
5698 int can_nice(const struct task_struct
*p
, const int nice
)
5700 /* convert nice value [19,-20] to rlimit style value [1,40] */
5701 int nice_rlim
= 20 - nice
;
5703 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5704 capable(CAP_SYS_NICE
));
5707 #ifdef __ARCH_WANT_SYS_NICE
5710 * sys_nice - change the priority of the current process.
5711 * @increment: priority increment
5713 * sys_setpriority is a more generic, but much slower function that
5714 * does similar things.
5716 SYSCALL_DEFINE1(nice
, int, increment
)
5721 * Setpriority might change our priority at the same moment.
5722 * We don't have to worry. Conceptually one call occurs first
5723 * and we have a single winner.
5725 if (increment
< -40)
5730 nice
= TASK_NICE(current
) + increment
;
5736 if (increment
< 0 && !can_nice(current
, nice
))
5739 retval
= security_task_setnice(current
, nice
);
5743 set_user_nice(current
, nice
);
5750 * task_prio - return the priority value of a given task.
5751 * @p: the task in question.
5753 * This is the priority value as seen by users in /proc.
5754 * RT tasks are offset by -200. Normal tasks are centered
5755 * around 0, value goes from -16 to +15.
5757 int task_prio(const struct task_struct
*p
)
5759 return p
->prio
- MAX_RT_PRIO
;
5763 * task_nice - return the nice value of a given task.
5764 * @p: the task in question.
5766 int task_nice(const struct task_struct
*p
)
5768 return TASK_NICE(p
);
5770 EXPORT_SYMBOL(task_nice
);
5773 * idle_cpu - is a given cpu idle currently?
5774 * @cpu: the processor in question.
5776 int idle_cpu(int cpu
)
5778 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5782 * idle_task - return the idle task for a given cpu.
5783 * @cpu: the processor in question.
5785 struct task_struct
*idle_task(int cpu
)
5787 return cpu_rq(cpu
)->idle
;
5791 * find_process_by_pid - find a process with a matching PID value.
5792 * @pid: the pid in question.
5794 static struct task_struct
*find_process_by_pid(pid_t pid
)
5796 return pid
? find_task_by_vpid(pid
) : current
;
5799 /* Actually do priority change: must hold rq lock. */
5801 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5803 BUG_ON(p
->se
.on_rq
);
5806 switch (p
->policy
) {
5810 p
->sched_class
= &fair_sched_class
;
5814 p
->sched_class
= &rt_sched_class
;
5818 p
->rt_priority
= prio
;
5819 p
->normal_prio
= normal_prio(p
);
5820 /* we are holding p->pi_lock already */
5821 p
->prio
= rt_mutex_getprio(p
);
5826 * check the target process has a UID that matches the current process's
5828 static bool check_same_owner(struct task_struct
*p
)
5830 const struct cred
*cred
= current_cred(), *pcred
;
5834 pcred
= __task_cred(p
);
5835 match
= (cred
->euid
== pcred
->euid
||
5836 cred
->euid
== pcred
->uid
);
5841 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5842 struct sched_param
*param
, bool user
)
5844 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5845 unsigned long flags
;
5846 const struct sched_class
*prev_class
= p
->sched_class
;
5849 /* may grab non-irq protected spin_locks */
5850 BUG_ON(in_interrupt());
5852 /* double check policy once rq lock held */
5854 policy
= oldpolicy
= p
->policy
;
5855 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5856 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5857 policy
!= SCHED_IDLE
)
5860 * Valid priorities for SCHED_FIFO and SCHED_RR are
5861 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5862 * SCHED_BATCH and SCHED_IDLE is 0.
5864 if (param
->sched_priority
< 0 ||
5865 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5866 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5868 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5872 * Allow unprivileged RT tasks to decrease priority:
5874 if (user
&& !capable(CAP_SYS_NICE
)) {
5875 if (rt_policy(policy
)) {
5876 unsigned long rlim_rtprio
;
5878 if (!lock_task_sighand(p
, &flags
))
5880 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5881 unlock_task_sighand(p
, &flags
);
5883 /* can't set/change the rt policy */
5884 if (policy
!= p
->policy
&& !rlim_rtprio
)
5887 /* can't increase priority */
5888 if (param
->sched_priority
> p
->rt_priority
&&
5889 param
->sched_priority
> rlim_rtprio
)
5893 * Like positive nice levels, dont allow tasks to
5894 * move out of SCHED_IDLE either:
5896 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5899 /* can't change other user's priorities */
5900 if (!check_same_owner(p
))
5905 #ifdef CONFIG_RT_GROUP_SCHED
5907 * Do not allow realtime tasks into groups that have no runtime
5910 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5911 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5915 retval
= security_task_setscheduler(p
, policy
, param
);
5921 * make sure no PI-waiters arrive (or leave) while we are
5922 * changing the priority of the task:
5924 spin_lock_irqsave(&p
->pi_lock
, flags
);
5926 * To be able to change p->policy safely, the apropriate
5927 * runqueue lock must be held.
5929 rq
= __task_rq_lock(p
);
5930 /* recheck policy now with rq lock held */
5931 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5932 policy
= oldpolicy
= -1;
5933 __task_rq_unlock(rq
);
5934 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5937 update_rq_clock(rq
);
5938 on_rq
= p
->se
.on_rq
;
5939 running
= task_current(rq
, p
);
5941 deactivate_task(rq
, p
, 0);
5943 p
->sched_class
->put_prev_task(rq
, p
);
5946 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5949 p
->sched_class
->set_curr_task(rq
);
5951 activate_task(rq
, p
, 0);
5953 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5955 __task_rq_unlock(rq
);
5956 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5958 rt_mutex_adjust_pi(p
);
5964 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5965 * @p: the task in question.
5966 * @policy: new policy.
5967 * @param: structure containing the new RT priority.
5969 * NOTE that the task may be already dead.
5971 int sched_setscheduler(struct task_struct
*p
, int policy
,
5972 struct sched_param
*param
)
5974 return __sched_setscheduler(p
, policy
, param
, true);
5976 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5979 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5980 * @p: the task in question.
5981 * @policy: new policy.
5982 * @param: structure containing the new RT priority.
5984 * Just like sched_setscheduler, only don't bother checking if the
5985 * current context has permission. For example, this is needed in
5986 * stop_machine(): we create temporary high priority worker threads,
5987 * but our caller might not have that capability.
5989 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5990 struct sched_param
*param
)
5992 return __sched_setscheduler(p
, policy
, param
, false);
5996 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5998 struct sched_param lparam
;
5999 struct task_struct
*p
;
6002 if (!param
|| pid
< 0)
6004 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6009 p
= find_process_by_pid(pid
);
6011 retval
= sched_setscheduler(p
, policy
, &lparam
);
6018 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6019 * @pid: the pid in question.
6020 * @policy: new policy.
6021 * @param: structure containing the new RT priority.
6023 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6024 struct sched_param __user
*, param
)
6026 /* negative values for policy are not valid */
6030 return do_sched_setscheduler(pid
, policy
, param
);
6034 * sys_sched_setparam - set/change the RT priority of a thread
6035 * @pid: the pid in question.
6036 * @param: structure containing the new RT priority.
6038 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6040 return do_sched_setscheduler(pid
, -1, param
);
6044 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6045 * @pid: the pid in question.
6047 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6049 struct task_struct
*p
;
6056 read_lock(&tasklist_lock
);
6057 p
= find_process_by_pid(pid
);
6059 retval
= security_task_getscheduler(p
);
6063 read_unlock(&tasklist_lock
);
6068 * sys_sched_getscheduler - get the RT priority of a thread
6069 * @pid: the pid in question.
6070 * @param: structure containing the RT priority.
6072 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6074 struct sched_param lp
;
6075 struct task_struct
*p
;
6078 if (!param
|| pid
< 0)
6081 read_lock(&tasklist_lock
);
6082 p
= find_process_by_pid(pid
);
6087 retval
= security_task_getscheduler(p
);
6091 lp
.sched_priority
= p
->rt_priority
;
6092 read_unlock(&tasklist_lock
);
6095 * This one might sleep, we cannot do it with a spinlock held ...
6097 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6102 read_unlock(&tasklist_lock
);
6106 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6108 cpumask_var_t cpus_allowed
, new_mask
;
6109 struct task_struct
*p
;
6113 read_lock(&tasklist_lock
);
6115 p
= find_process_by_pid(pid
);
6117 read_unlock(&tasklist_lock
);
6123 * It is not safe to call set_cpus_allowed with the
6124 * tasklist_lock held. We will bump the task_struct's
6125 * usage count and then drop tasklist_lock.
6128 read_unlock(&tasklist_lock
);
6130 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6134 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6136 goto out_free_cpus_allowed
;
6139 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6142 retval
= security_task_setscheduler(p
, 0, NULL
);
6146 cpuset_cpus_allowed(p
, cpus_allowed
);
6147 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6149 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6152 cpuset_cpus_allowed(p
, cpus_allowed
);
6153 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6155 * We must have raced with a concurrent cpuset
6156 * update. Just reset the cpus_allowed to the
6157 * cpuset's cpus_allowed
6159 cpumask_copy(new_mask
, cpus_allowed
);
6164 free_cpumask_var(new_mask
);
6165 out_free_cpus_allowed
:
6166 free_cpumask_var(cpus_allowed
);
6173 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6174 struct cpumask
*new_mask
)
6176 if (len
< cpumask_size())
6177 cpumask_clear(new_mask
);
6178 else if (len
> cpumask_size())
6179 len
= cpumask_size();
6181 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6185 * sys_sched_setaffinity - set the cpu affinity of a process
6186 * @pid: pid of the process
6187 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6188 * @user_mask_ptr: user-space pointer to the new cpu mask
6190 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6191 unsigned long __user
*, user_mask_ptr
)
6193 cpumask_var_t new_mask
;
6196 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6199 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6201 retval
= sched_setaffinity(pid
, new_mask
);
6202 free_cpumask_var(new_mask
);
6206 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6208 struct task_struct
*p
;
6212 read_lock(&tasklist_lock
);
6215 p
= find_process_by_pid(pid
);
6219 retval
= security_task_getscheduler(p
);
6223 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6226 read_unlock(&tasklist_lock
);
6233 * sys_sched_getaffinity - get the cpu affinity of a process
6234 * @pid: pid of the process
6235 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6236 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6238 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6239 unsigned long __user
*, user_mask_ptr
)
6244 if (len
< cpumask_size())
6247 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6250 ret
= sched_getaffinity(pid
, mask
);
6252 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6255 ret
= cpumask_size();
6257 free_cpumask_var(mask
);
6263 * sys_sched_yield - yield the current processor to other threads.
6265 * This function yields the current CPU to other tasks. If there are no
6266 * other threads running on this CPU then this function will return.
6268 SYSCALL_DEFINE0(sched_yield
)
6270 struct rq
*rq
= this_rq_lock();
6272 schedstat_inc(rq
, yld_count
);
6273 current
->sched_class
->yield_task(rq
);
6276 * Since we are going to call schedule() anyway, there's
6277 * no need to preempt or enable interrupts:
6279 __release(rq
->lock
);
6280 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6281 _raw_spin_unlock(&rq
->lock
);
6282 preempt_enable_no_resched();
6289 static void __cond_resched(void)
6291 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6292 __might_sleep(__FILE__
, __LINE__
);
6295 * The BKS might be reacquired before we have dropped
6296 * PREEMPT_ACTIVE, which could trigger a second
6297 * cond_resched() call.
6300 add_preempt_count(PREEMPT_ACTIVE
);
6302 sub_preempt_count(PREEMPT_ACTIVE
);
6303 } while (need_resched());
6306 int __sched
_cond_resched(void)
6308 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
6309 system_state
== SYSTEM_RUNNING
) {
6315 EXPORT_SYMBOL(_cond_resched
);
6318 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
6319 * call schedule, and on return reacquire the lock.
6321 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6322 * operations here to prevent schedule() from being called twice (once via
6323 * spin_unlock(), once by hand).
6325 int cond_resched_lock(spinlock_t
*lock
)
6327 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
6330 if (spin_needbreak(lock
) || resched
) {
6332 if (resched
&& need_resched())
6341 EXPORT_SYMBOL(cond_resched_lock
);
6343 int __sched
cond_resched_softirq(void)
6345 BUG_ON(!in_softirq());
6347 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
6355 EXPORT_SYMBOL(cond_resched_softirq
);
6358 * yield - yield the current processor to other threads.
6360 * This is a shortcut for kernel-space yielding - it marks the
6361 * thread runnable and calls sys_sched_yield().
6363 void __sched
yield(void)
6365 set_current_state(TASK_RUNNING
);
6368 EXPORT_SYMBOL(yield
);
6371 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6372 * that process accounting knows that this is a task in IO wait state.
6374 * But don't do that if it is a deliberate, throttling IO wait (this task
6375 * has set its backing_dev_info: the queue against which it should throttle)
6377 void __sched
io_schedule(void)
6379 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6381 delayacct_blkio_start();
6382 atomic_inc(&rq
->nr_iowait
);
6384 atomic_dec(&rq
->nr_iowait
);
6385 delayacct_blkio_end();
6387 EXPORT_SYMBOL(io_schedule
);
6389 long __sched
io_schedule_timeout(long timeout
)
6391 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
6394 delayacct_blkio_start();
6395 atomic_inc(&rq
->nr_iowait
);
6396 ret
= schedule_timeout(timeout
);
6397 atomic_dec(&rq
->nr_iowait
);
6398 delayacct_blkio_end();
6403 * sys_sched_get_priority_max - return maximum RT priority.
6404 * @policy: scheduling class.
6406 * this syscall returns the maximum rt_priority that can be used
6407 * by a given scheduling class.
6409 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6416 ret
= MAX_USER_RT_PRIO
-1;
6428 * sys_sched_get_priority_min - return minimum RT priority.
6429 * @policy: scheduling class.
6431 * this syscall returns the minimum rt_priority that can be used
6432 * by a given scheduling class.
6434 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6452 * sys_sched_rr_get_interval - return the default timeslice of a process.
6453 * @pid: pid of the process.
6454 * @interval: userspace pointer to the timeslice value.
6456 * this syscall writes the default timeslice value of a given process
6457 * into the user-space timespec buffer. A value of '0' means infinity.
6459 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6460 struct timespec __user
*, interval
)
6462 struct task_struct
*p
;
6463 unsigned int time_slice
;
6471 read_lock(&tasklist_lock
);
6472 p
= find_process_by_pid(pid
);
6476 retval
= security_task_getscheduler(p
);
6481 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
6482 * tasks that are on an otherwise idle runqueue:
6485 if (p
->policy
== SCHED_RR
) {
6486 time_slice
= DEF_TIMESLICE
;
6487 } else if (p
->policy
!= SCHED_FIFO
) {
6488 struct sched_entity
*se
= &p
->se
;
6489 unsigned long flags
;
6492 rq
= task_rq_lock(p
, &flags
);
6493 if (rq
->cfs
.load
.weight
)
6494 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
6495 task_rq_unlock(rq
, &flags
);
6497 read_unlock(&tasklist_lock
);
6498 jiffies_to_timespec(time_slice
, &t
);
6499 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6503 read_unlock(&tasklist_lock
);
6507 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6509 void sched_show_task(struct task_struct
*p
)
6511 unsigned long free
= 0;
6514 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6515 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
6516 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6517 #if BITS_PER_LONG == 32
6518 if (state
== TASK_RUNNING
)
6519 printk(KERN_CONT
" running ");
6521 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
6523 if (state
== TASK_RUNNING
)
6524 printk(KERN_CONT
" running task ");
6526 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
6528 #ifdef CONFIG_DEBUG_STACK_USAGE
6529 free
= stack_not_used(p
);
6531 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
6532 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
6534 show_stack(p
, NULL
);
6537 void show_state_filter(unsigned long state_filter
)
6539 struct task_struct
*g
, *p
;
6541 #if BITS_PER_LONG == 32
6543 " task PC stack pid father\n");
6546 " task PC stack pid father\n");
6548 read_lock(&tasklist_lock
);
6549 do_each_thread(g
, p
) {
6551 * reset the NMI-timeout, listing all files on a slow
6552 * console might take alot of time:
6554 touch_nmi_watchdog();
6555 if (!state_filter
|| (p
->state
& state_filter
))
6557 } while_each_thread(g
, p
);
6559 touch_all_softlockup_watchdogs();
6561 #ifdef CONFIG_SCHED_DEBUG
6562 sysrq_sched_debug_show();
6564 read_unlock(&tasklist_lock
);
6566 * Only show locks if all tasks are dumped:
6568 if (state_filter
== -1)
6569 debug_show_all_locks();
6572 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6574 idle
->sched_class
= &idle_sched_class
;
6578 * init_idle - set up an idle thread for a given CPU
6579 * @idle: task in question
6580 * @cpu: cpu the idle task belongs to
6582 * NOTE: this function does not set the idle thread's NEED_RESCHED
6583 * flag, to make booting more robust.
6585 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
6587 struct rq
*rq
= cpu_rq(cpu
);
6588 unsigned long flags
;
6590 spin_lock_irqsave(&rq
->lock
, flags
);
6593 idle
->se
.exec_start
= sched_clock();
6595 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
6596 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
6597 __set_task_cpu(idle
, cpu
);
6599 rq
->curr
= rq
->idle
= idle
;
6600 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
6603 spin_unlock_irqrestore(&rq
->lock
, flags
);
6605 /* Set the preempt count _outside_ the spinlocks! */
6606 #if defined(CONFIG_PREEMPT)
6607 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
6609 task_thread_info(idle
)->preempt_count
= 0;
6612 * The idle tasks have their own, simple scheduling class:
6614 idle
->sched_class
= &idle_sched_class
;
6615 ftrace_graph_init_task(idle
);
6619 * In a system that switches off the HZ timer nohz_cpu_mask
6620 * indicates which cpus entered this state. This is used
6621 * in the rcu update to wait only for active cpus. For system
6622 * which do not switch off the HZ timer nohz_cpu_mask should
6623 * always be CPU_BITS_NONE.
6625 cpumask_var_t nohz_cpu_mask
;
6628 * Increase the granularity value when there are more CPUs,
6629 * because with more CPUs the 'effective latency' as visible
6630 * to users decreases. But the relationship is not linear,
6631 * so pick a second-best guess by going with the log2 of the
6634 * This idea comes from the SD scheduler of Con Kolivas:
6636 static inline void sched_init_granularity(void)
6638 unsigned int factor
= 1 + ilog2(num_online_cpus());
6639 const unsigned long limit
= 200000000;
6641 sysctl_sched_min_granularity
*= factor
;
6642 if (sysctl_sched_min_granularity
> limit
)
6643 sysctl_sched_min_granularity
= limit
;
6645 sysctl_sched_latency
*= factor
;
6646 if (sysctl_sched_latency
> limit
)
6647 sysctl_sched_latency
= limit
;
6649 sysctl_sched_wakeup_granularity
*= factor
;
6651 sysctl_sched_shares_ratelimit
*= factor
;
6656 * This is how migration works:
6658 * 1) we queue a struct migration_req structure in the source CPU's
6659 * runqueue and wake up that CPU's migration thread.
6660 * 2) we down() the locked semaphore => thread blocks.
6661 * 3) migration thread wakes up (implicitly it forces the migrated
6662 * thread off the CPU)
6663 * 4) it gets the migration request and checks whether the migrated
6664 * task is still in the wrong runqueue.
6665 * 5) if it's in the wrong runqueue then the migration thread removes
6666 * it and puts it into the right queue.
6667 * 6) migration thread up()s the semaphore.
6668 * 7) we wake up and the migration is done.
6672 * Change a given task's CPU affinity. Migrate the thread to a
6673 * proper CPU and schedule it away if the CPU it's executing on
6674 * is removed from the allowed bitmask.
6676 * NOTE: the caller must have a valid reference to the task, the
6677 * task must not exit() & deallocate itself prematurely. The
6678 * call is not atomic; no spinlocks may be held.
6680 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
6682 struct migration_req req
;
6683 unsigned long flags
;
6687 rq
= task_rq_lock(p
, &flags
);
6688 if (!cpumask_intersects(new_mask
, cpu_online_mask
)) {
6693 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
6694 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
6699 if (p
->sched_class
->set_cpus_allowed
)
6700 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6702 cpumask_copy(&p
->cpus_allowed
, new_mask
);
6703 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
6706 /* Can the task run on the task's current CPU? If so, we're done */
6707 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
6710 if (migrate_task(p
, cpumask_any_and(cpu_online_mask
, new_mask
), &req
)) {
6711 /* Need help from migration thread: drop lock and wait. */
6712 task_rq_unlock(rq
, &flags
);
6713 wake_up_process(rq
->migration_thread
);
6714 wait_for_completion(&req
.done
);
6715 tlb_migrate_finish(p
->mm
);
6719 task_rq_unlock(rq
, &flags
);
6723 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6726 * Move (not current) task off this cpu, onto dest cpu. We're doing
6727 * this because either it can't run here any more (set_cpus_allowed()
6728 * away from this CPU, or CPU going down), or because we're
6729 * attempting to rebalance this task on exec (sched_exec).
6731 * So we race with normal scheduler movements, but that's OK, as long
6732 * as the task is no longer on this CPU.
6734 * Returns non-zero if task was successfully migrated.
6736 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6738 struct rq
*rq_dest
, *rq_src
;
6741 if (unlikely(!cpu_active(dest_cpu
)))
6744 rq_src
= cpu_rq(src_cpu
);
6745 rq_dest
= cpu_rq(dest_cpu
);
6747 double_rq_lock(rq_src
, rq_dest
);
6748 /* Already moved. */
6749 if (task_cpu(p
) != src_cpu
)
6751 /* Affinity changed (again). */
6752 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6755 on_rq
= p
->se
.on_rq
;
6757 deactivate_task(rq_src
, p
, 0);
6759 set_task_cpu(p
, dest_cpu
);
6761 activate_task(rq_dest
, p
, 0);
6762 check_preempt_curr(rq_dest
, p
, 0);
6767 double_rq_unlock(rq_src
, rq_dest
);
6772 * migration_thread - this is a highprio system thread that performs
6773 * thread migration by bumping thread off CPU then 'pushing' onto
6776 static int migration_thread(void *data
)
6778 int cpu
= (long)data
;
6782 BUG_ON(rq
->migration_thread
!= current
);
6784 set_current_state(TASK_INTERRUPTIBLE
);
6785 while (!kthread_should_stop()) {
6786 struct migration_req
*req
;
6787 struct list_head
*head
;
6789 spin_lock_irq(&rq
->lock
);
6791 if (cpu_is_offline(cpu
)) {
6792 spin_unlock_irq(&rq
->lock
);
6796 if (rq
->active_balance
) {
6797 active_load_balance(rq
, cpu
);
6798 rq
->active_balance
= 0;
6801 head
= &rq
->migration_queue
;
6803 if (list_empty(head
)) {
6804 spin_unlock_irq(&rq
->lock
);
6806 set_current_state(TASK_INTERRUPTIBLE
);
6809 req
= list_entry(head
->next
, struct migration_req
, list
);
6810 list_del_init(head
->next
);
6812 spin_unlock(&rq
->lock
);
6813 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6816 complete(&req
->done
);
6818 __set_current_state(TASK_RUNNING
);
6822 /* Wait for kthread_stop */
6823 set_current_state(TASK_INTERRUPTIBLE
);
6824 while (!kthread_should_stop()) {
6826 set_current_state(TASK_INTERRUPTIBLE
);
6828 __set_current_state(TASK_RUNNING
);
6832 #ifdef CONFIG_HOTPLUG_CPU
6834 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6838 local_irq_disable();
6839 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6845 * Figure out where task on dead CPU should go, use force if necessary.
6847 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6850 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
6853 /* Look for allowed, online CPU in same node. */
6854 for_each_cpu_and(dest_cpu
, nodemask
, cpu_online_mask
)
6855 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
6858 /* Any allowed, online CPU? */
6859 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_online_mask
);
6860 if (dest_cpu
< nr_cpu_ids
)
6863 /* No more Mr. Nice Guy. */
6864 if (dest_cpu
>= nr_cpu_ids
) {
6865 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
6866 dest_cpu
= cpumask_any_and(cpu_online_mask
, &p
->cpus_allowed
);
6869 * Don't tell them about moving exiting tasks or
6870 * kernel threads (both mm NULL), since they never
6873 if (p
->mm
&& printk_ratelimit()) {
6874 printk(KERN_INFO
"process %d (%s) no "
6875 "longer affine to cpu%d\n",
6876 task_pid_nr(p
), p
->comm
, dead_cpu
);
6881 /* It can have affinity changed while we were choosing. */
6882 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
6887 * While a dead CPU has no uninterruptible tasks queued at this point,
6888 * it might still have a nonzero ->nr_uninterruptible counter, because
6889 * for performance reasons the counter is not stricly tracking tasks to
6890 * their home CPUs. So we just add the counter to another CPU's counter,
6891 * to keep the global sum constant after CPU-down:
6893 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6895 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_online_mask
));
6896 unsigned long flags
;
6898 local_irq_save(flags
);
6899 double_rq_lock(rq_src
, rq_dest
);
6900 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6901 rq_src
->nr_uninterruptible
= 0;
6902 double_rq_unlock(rq_src
, rq_dest
);
6903 local_irq_restore(flags
);
6906 /* Run through task list and migrate tasks from the dead cpu. */
6907 static void migrate_live_tasks(int src_cpu
)
6909 struct task_struct
*p
, *t
;
6911 read_lock(&tasklist_lock
);
6913 do_each_thread(t
, p
) {
6917 if (task_cpu(p
) == src_cpu
)
6918 move_task_off_dead_cpu(src_cpu
, p
);
6919 } while_each_thread(t
, p
);
6921 read_unlock(&tasklist_lock
);
6925 * Schedules idle task to be the next runnable task on current CPU.
6926 * It does so by boosting its priority to highest possible.
6927 * Used by CPU offline code.
6929 void sched_idle_next(void)
6931 int this_cpu
= smp_processor_id();
6932 struct rq
*rq
= cpu_rq(this_cpu
);
6933 struct task_struct
*p
= rq
->idle
;
6934 unsigned long flags
;
6936 /* cpu has to be offline */
6937 BUG_ON(cpu_online(this_cpu
));
6940 * Strictly not necessary since rest of the CPUs are stopped by now
6941 * and interrupts disabled on the current cpu.
6943 spin_lock_irqsave(&rq
->lock
, flags
);
6945 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6947 update_rq_clock(rq
);
6948 activate_task(rq
, p
, 0);
6950 spin_unlock_irqrestore(&rq
->lock
, flags
);
6954 * Ensures that the idle task is using init_mm right before its cpu goes
6957 void idle_task_exit(void)
6959 struct mm_struct
*mm
= current
->active_mm
;
6961 BUG_ON(cpu_online(smp_processor_id()));
6964 switch_mm(mm
, &init_mm
, current
);
6968 /* called under rq->lock with disabled interrupts */
6969 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6971 struct rq
*rq
= cpu_rq(dead_cpu
);
6973 /* Must be exiting, otherwise would be on tasklist. */
6974 BUG_ON(!p
->exit_state
);
6976 /* Cannot have done final schedule yet: would have vanished. */
6977 BUG_ON(p
->state
== TASK_DEAD
);
6982 * Drop lock around migration; if someone else moves it,
6983 * that's OK. No task can be added to this CPU, so iteration is
6986 spin_unlock_irq(&rq
->lock
);
6987 move_task_off_dead_cpu(dead_cpu
, p
);
6988 spin_lock_irq(&rq
->lock
);
6993 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6994 static void migrate_dead_tasks(unsigned int dead_cpu
)
6996 struct rq
*rq
= cpu_rq(dead_cpu
);
6997 struct task_struct
*next
;
7000 if (!rq
->nr_running
)
7002 update_rq_clock(rq
);
7003 next
= pick_next_task(rq
);
7006 next
->sched_class
->put_prev_task(rq
, next
);
7007 migrate_dead(dead_cpu
, next
);
7011 #endif /* CONFIG_HOTPLUG_CPU */
7013 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7015 static struct ctl_table sd_ctl_dir
[] = {
7017 .procname
= "sched_domain",
7023 static struct ctl_table sd_ctl_root
[] = {
7025 .ctl_name
= CTL_KERN
,
7026 .procname
= "kernel",
7028 .child
= sd_ctl_dir
,
7033 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7035 struct ctl_table
*entry
=
7036 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7041 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7043 struct ctl_table
*entry
;
7046 * In the intermediate directories, both the child directory and
7047 * procname are dynamically allocated and could fail but the mode
7048 * will always be set. In the lowest directory the names are
7049 * static strings and all have proc handlers.
7051 for (entry
= *tablep
; entry
->mode
; entry
++) {
7053 sd_free_ctl_entry(&entry
->child
);
7054 if (entry
->proc_handler
== NULL
)
7055 kfree(entry
->procname
);
7063 set_table_entry(struct ctl_table
*entry
,
7064 const char *procname
, void *data
, int maxlen
,
7065 mode_t mode
, proc_handler
*proc_handler
)
7067 entry
->procname
= procname
;
7069 entry
->maxlen
= maxlen
;
7071 entry
->proc_handler
= proc_handler
;
7074 static struct ctl_table
*
7075 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7077 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7082 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7083 sizeof(long), 0644, proc_doulongvec_minmax
);
7084 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7085 sizeof(long), 0644, proc_doulongvec_minmax
);
7086 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7087 sizeof(int), 0644, proc_dointvec_minmax
);
7088 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7089 sizeof(int), 0644, proc_dointvec_minmax
);
7090 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7091 sizeof(int), 0644, proc_dointvec_minmax
);
7092 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7093 sizeof(int), 0644, proc_dointvec_minmax
);
7094 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7095 sizeof(int), 0644, proc_dointvec_minmax
);
7096 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7097 sizeof(int), 0644, proc_dointvec_minmax
);
7098 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7099 sizeof(int), 0644, proc_dointvec_minmax
);
7100 set_table_entry(&table
[9], "cache_nice_tries",
7101 &sd
->cache_nice_tries
,
7102 sizeof(int), 0644, proc_dointvec_minmax
);
7103 set_table_entry(&table
[10], "flags", &sd
->flags
,
7104 sizeof(int), 0644, proc_dointvec_minmax
);
7105 set_table_entry(&table
[11], "name", sd
->name
,
7106 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7107 /* &table[12] is terminator */
7112 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7114 struct ctl_table
*entry
, *table
;
7115 struct sched_domain
*sd
;
7116 int domain_num
= 0, i
;
7119 for_each_domain(cpu
, sd
)
7121 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7126 for_each_domain(cpu
, sd
) {
7127 snprintf(buf
, 32, "domain%d", i
);
7128 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7130 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7137 static struct ctl_table_header
*sd_sysctl_header
;
7138 static void register_sched_domain_sysctl(void)
7140 int i
, cpu_num
= num_online_cpus();
7141 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7144 WARN_ON(sd_ctl_dir
[0].child
);
7145 sd_ctl_dir
[0].child
= entry
;
7150 for_each_online_cpu(i
) {
7151 snprintf(buf
, 32, "cpu%d", i
);
7152 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7154 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7158 WARN_ON(sd_sysctl_header
);
7159 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7162 /* may be called multiple times per register */
7163 static void unregister_sched_domain_sysctl(void)
7165 if (sd_sysctl_header
)
7166 unregister_sysctl_table(sd_sysctl_header
);
7167 sd_sysctl_header
= NULL
;
7168 if (sd_ctl_dir
[0].child
)
7169 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7172 static void register_sched_domain_sysctl(void)
7175 static void unregister_sched_domain_sysctl(void)
7180 static void set_rq_online(struct rq
*rq
)
7183 const struct sched_class
*class;
7185 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7188 for_each_class(class) {
7189 if (class->rq_online
)
7190 class->rq_online(rq
);
7195 static void set_rq_offline(struct rq
*rq
)
7198 const struct sched_class
*class;
7200 for_each_class(class) {
7201 if (class->rq_offline
)
7202 class->rq_offline(rq
);
7205 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7211 * migration_call - callback that gets triggered when a CPU is added.
7212 * Here we can start up the necessary migration thread for the new CPU.
7214 static int __cpuinit
7215 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7217 struct task_struct
*p
;
7218 int cpu
= (long)hcpu
;
7219 unsigned long flags
;
7224 case CPU_UP_PREPARE
:
7225 case CPU_UP_PREPARE_FROZEN
:
7226 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7229 kthread_bind(p
, cpu
);
7230 /* Must be high prio: stop_machine expects to yield to it. */
7231 rq
= task_rq_lock(p
, &flags
);
7232 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7233 task_rq_unlock(rq
, &flags
);
7234 cpu_rq(cpu
)->migration_thread
= p
;
7238 case CPU_ONLINE_FROZEN
:
7239 /* Strictly unnecessary, as first user will wake it. */
7240 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7242 /* Update our root-domain */
7244 spin_lock_irqsave(&rq
->lock
, flags
);
7246 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7250 spin_unlock_irqrestore(&rq
->lock
, flags
);
7253 #ifdef CONFIG_HOTPLUG_CPU
7254 case CPU_UP_CANCELED
:
7255 case CPU_UP_CANCELED_FROZEN
:
7256 if (!cpu_rq(cpu
)->migration_thread
)
7258 /* Unbind it from offline cpu so it can run. Fall thru. */
7259 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7260 cpumask_any(cpu_online_mask
));
7261 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7262 cpu_rq(cpu
)->migration_thread
= NULL
;
7266 case CPU_DEAD_FROZEN
:
7267 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7268 migrate_live_tasks(cpu
);
7270 kthread_stop(rq
->migration_thread
);
7271 rq
->migration_thread
= NULL
;
7272 /* Idle task back to normal (off runqueue, low prio) */
7273 spin_lock_irq(&rq
->lock
);
7274 update_rq_clock(rq
);
7275 deactivate_task(rq
, rq
->idle
, 0);
7276 rq
->idle
->static_prio
= MAX_PRIO
;
7277 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7278 rq
->idle
->sched_class
= &idle_sched_class
;
7279 migrate_dead_tasks(cpu
);
7280 spin_unlock_irq(&rq
->lock
);
7282 migrate_nr_uninterruptible(rq
);
7283 BUG_ON(rq
->nr_running
!= 0);
7286 * No need to migrate the tasks: it was best-effort if
7287 * they didn't take sched_hotcpu_mutex. Just wake up
7290 spin_lock_irq(&rq
->lock
);
7291 while (!list_empty(&rq
->migration_queue
)) {
7292 struct migration_req
*req
;
7294 req
= list_entry(rq
->migration_queue
.next
,
7295 struct migration_req
, list
);
7296 list_del_init(&req
->list
);
7297 spin_unlock_irq(&rq
->lock
);
7298 complete(&req
->done
);
7299 spin_lock_irq(&rq
->lock
);
7301 spin_unlock_irq(&rq
->lock
);
7305 case CPU_DYING_FROZEN
:
7306 /* Update our root-domain */
7308 spin_lock_irqsave(&rq
->lock
, flags
);
7310 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7313 spin_unlock_irqrestore(&rq
->lock
, flags
);
7320 /* Register at highest priority so that task migration (migrate_all_tasks)
7321 * happens before everything else.
7323 static struct notifier_block __cpuinitdata migration_notifier
= {
7324 .notifier_call
= migration_call
,
7328 static int __init
migration_init(void)
7330 void *cpu
= (void *)(long)smp_processor_id();
7333 /* Start one for the boot CPU: */
7334 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7335 BUG_ON(err
== NOTIFY_BAD
);
7336 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7337 register_cpu_notifier(&migration_notifier
);
7341 early_initcall(migration_init
);
7346 #ifdef CONFIG_SCHED_DEBUG
7348 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7349 struct cpumask
*groupmask
)
7351 struct sched_group
*group
= sd
->groups
;
7354 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7355 cpumask_clear(groupmask
);
7357 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7359 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7360 printk("does not load-balance\n");
7362 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
7367 printk(KERN_CONT
"span %s level %s\n", str
, sd
->name
);
7369 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7370 printk(KERN_ERR
"ERROR: domain->span does not contain "
7373 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7374 printk(KERN_ERR
"ERROR: domain->groups does not contain"
7378 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7382 printk(KERN_ERR
"ERROR: group is NULL\n");
7386 if (!group
->__cpu_power
) {
7387 printk(KERN_CONT
"\n");
7388 printk(KERN_ERR
"ERROR: domain->cpu_power not "
7393 if (!cpumask_weight(sched_group_cpus(group
))) {
7394 printk(KERN_CONT
"\n");
7395 printk(KERN_ERR
"ERROR: empty group\n");
7399 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7400 printk(KERN_CONT
"\n");
7401 printk(KERN_ERR
"ERROR: repeated CPUs\n");
7405 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7407 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7409 printk(KERN_CONT
" %s", str
);
7410 if (group
->__cpu_power
!= SCHED_LOAD_SCALE
) {
7411 printk(KERN_CONT
" (__cpu_power = %d)",
7412 group
->__cpu_power
);
7415 group
= group
->next
;
7416 } while (group
!= sd
->groups
);
7417 printk(KERN_CONT
"\n");
7419 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7420 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
7423 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7424 printk(KERN_ERR
"ERROR: parent span is not a superset "
7425 "of domain->span\n");
7429 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7431 cpumask_var_t groupmask
;
7435 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7439 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7441 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7442 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7447 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7454 free_cpumask_var(groupmask
);
7456 #else /* !CONFIG_SCHED_DEBUG */
7457 # define sched_domain_debug(sd, cpu) do { } while (0)
7458 #endif /* CONFIG_SCHED_DEBUG */
7460 static int sd_degenerate(struct sched_domain
*sd
)
7462 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7465 /* Following flags need at least 2 groups */
7466 if (sd
->flags
& (SD_LOAD_BALANCE
|
7467 SD_BALANCE_NEWIDLE
|
7471 SD_SHARE_PKG_RESOURCES
)) {
7472 if (sd
->groups
!= sd
->groups
->next
)
7476 /* Following flags don't use groups */
7477 if (sd
->flags
& (SD_WAKE_IDLE
|
7486 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7488 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7490 if (sd_degenerate(parent
))
7493 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7496 /* Does parent contain flags not in child? */
7497 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
7498 if (cflags
& SD_WAKE_AFFINE
)
7499 pflags
&= ~SD_WAKE_BALANCE
;
7500 /* Flags needing groups don't count if only 1 group in parent */
7501 if (parent
->groups
== parent
->groups
->next
) {
7502 pflags
&= ~(SD_LOAD_BALANCE
|
7503 SD_BALANCE_NEWIDLE
|
7507 SD_SHARE_PKG_RESOURCES
);
7508 if (nr_node_ids
== 1)
7509 pflags
&= ~SD_SERIALIZE
;
7511 if (~cflags
& pflags
)
7517 static void free_rootdomain(struct root_domain
*rd
)
7519 cpupri_cleanup(&rd
->cpupri
);
7521 free_cpumask_var(rd
->rto_mask
);
7522 free_cpumask_var(rd
->online
);
7523 free_cpumask_var(rd
->span
);
7527 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
7529 struct root_domain
*old_rd
= NULL
;
7530 unsigned long flags
;
7532 spin_lock_irqsave(&rq
->lock
, flags
);
7537 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
7540 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
7543 * If we dont want to free the old_rt yet then
7544 * set old_rd to NULL to skip the freeing later
7547 if (!atomic_dec_and_test(&old_rd
->refcount
))
7551 atomic_inc(&rd
->refcount
);
7554 cpumask_set_cpu(rq
->cpu
, rd
->span
);
7555 if (cpumask_test_cpu(rq
->cpu
, cpu_online_mask
))
7558 spin_unlock_irqrestore(&rq
->lock
, flags
);
7561 free_rootdomain(old_rd
);
7564 static int __init_refok
init_rootdomain(struct root_domain
*rd
, bool bootmem
)
7566 memset(rd
, 0, sizeof(*rd
));
7569 alloc_bootmem_cpumask_var(&def_root_domain
.span
);
7570 alloc_bootmem_cpumask_var(&def_root_domain
.online
);
7571 alloc_bootmem_cpumask_var(&def_root_domain
.rto_mask
);
7572 cpupri_init(&rd
->cpupri
, true);
7576 if (!alloc_cpumask_var(&rd
->span
, GFP_KERNEL
))
7578 if (!alloc_cpumask_var(&rd
->online
, GFP_KERNEL
))
7580 if (!alloc_cpumask_var(&rd
->rto_mask
, GFP_KERNEL
))
7583 if (cpupri_init(&rd
->cpupri
, false) != 0)
7588 free_cpumask_var(rd
->rto_mask
);
7590 free_cpumask_var(rd
->online
);
7592 free_cpumask_var(rd
->span
);
7597 static void init_defrootdomain(void)
7599 init_rootdomain(&def_root_domain
, true);
7601 atomic_set(&def_root_domain
.refcount
, 1);
7604 static struct root_domain
*alloc_rootdomain(void)
7606 struct root_domain
*rd
;
7608 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
7612 if (init_rootdomain(rd
, false) != 0) {
7621 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
7622 * hold the hotplug lock.
7625 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
7627 struct rq
*rq
= cpu_rq(cpu
);
7628 struct sched_domain
*tmp
;
7630 /* Remove the sched domains which do not contribute to scheduling. */
7631 for (tmp
= sd
; tmp
; ) {
7632 struct sched_domain
*parent
= tmp
->parent
;
7636 if (sd_parent_degenerate(tmp
, parent
)) {
7637 tmp
->parent
= parent
->parent
;
7639 parent
->parent
->child
= tmp
;
7644 if (sd
&& sd_degenerate(sd
)) {
7650 sched_domain_debug(sd
, cpu
);
7652 rq_attach_root(rq
, rd
);
7653 rcu_assign_pointer(rq
->sd
, sd
);
7656 /* cpus with isolated domains */
7657 static cpumask_var_t cpu_isolated_map
;
7659 /* Setup the mask of cpus configured for isolated domains */
7660 static int __init
isolated_cpu_setup(char *str
)
7662 cpulist_parse(str
, cpu_isolated_map
);
7666 __setup("isolcpus=", isolated_cpu_setup
);
7669 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
7670 * to a function which identifies what group(along with sched group) a CPU
7671 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
7672 * (due to the fact that we keep track of groups covered with a struct cpumask).
7674 * init_sched_build_groups will build a circular linked list of the groups
7675 * covered by the given span, and will set each group's ->cpumask correctly,
7676 * and ->cpu_power to 0.
7679 init_sched_build_groups(const struct cpumask
*span
,
7680 const struct cpumask
*cpu_map
,
7681 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
7682 struct sched_group
**sg
,
7683 struct cpumask
*tmpmask
),
7684 struct cpumask
*covered
, struct cpumask
*tmpmask
)
7686 struct sched_group
*first
= NULL
, *last
= NULL
;
7689 cpumask_clear(covered
);
7691 for_each_cpu(i
, span
) {
7692 struct sched_group
*sg
;
7693 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
7696 if (cpumask_test_cpu(i
, covered
))
7699 cpumask_clear(sched_group_cpus(sg
));
7700 sg
->__cpu_power
= 0;
7702 for_each_cpu(j
, span
) {
7703 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7706 cpumask_set_cpu(j
, covered
);
7707 cpumask_set_cpu(j
, sched_group_cpus(sg
));
7718 #define SD_NODES_PER_DOMAIN 16
7723 * find_next_best_node - find the next node to include in a sched_domain
7724 * @node: node whose sched_domain we're building
7725 * @used_nodes: nodes already in the sched_domain
7727 * Find the next node to include in a given scheduling domain. Simply
7728 * finds the closest node not already in the @used_nodes map.
7730 * Should use nodemask_t.
7732 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7734 int i
, n
, val
, min_val
, best_node
= 0;
7738 for (i
= 0; i
< nr_node_ids
; i
++) {
7739 /* Start at @node */
7740 n
= (node
+ i
) % nr_node_ids
;
7742 if (!nr_cpus_node(n
))
7745 /* Skip already used nodes */
7746 if (node_isset(n
, *used_nodes
))
7749 /* Simple min distance search */
7750 val
= node_distance(node
, n
);
7752 if (val
< min_val
) {
7758 node_set(best_node
, *used_nodes
);
7763 * sched_domain_node_span - get a cpumask for a node's sched_domain
7764 * @node: node whose cpumask we're constructing
7765 * @span: resulting cpumask
7767 * Given a node, construct a good cpumask for its sched_domain to span. It
7768 * should be one that prevents unnecessary balancing, but also spreads tasks
7771 static void sched_domain_node_span(int node
, struct cpumask
*span
)
7773 nodemask_t used_nodes
;
7776 cpumask_clear(span
);
7777 nodes_clear(used_nodes
);
7779 cpumask_or(span
, span
, cpumask_of_node(node
));
7780 node_set(node
, used_nodes
);
7782 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7783 int next_node
= find_next_best_node(node
, &used_nodes
);
7785 cpumask_or(span
, span
, cpumask_of_node(next_node
));
7788 #endif /* CONFIG_NUMA */
7790 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7793 * The cpus mask in sched_group and sched_domain hangs off the end.
7794 * FIXME: use cpumask_var_t or dynamic percpu alloc to avoid wasting space
7795 * for nr_cpu_ids < CONFIG_NR_CPUS.
7797 struct static_sched_group
{
7798 struct sched_group sg
;
7799 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
7802 struct static_sched_domain
{
7803 struct sched_domain sd
;
7804 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
7808 * SMT sched-domains:
7810 #ifdef CONFIG_SCHED_SMT
7811 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
7812 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_cpus
);
7815 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
7816 struct sched_group
**sg
, struct cpumask
*unused
)
7819 *sg
= &per_cpu(sched_group_cpus
, cpu
).sg
;
7822 #endif /* CONFIG_SCHED_SMT */
7825 * multi-core sched-domains:
7827 #ifdef CONFIG_SCHED_MC
7828 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
7829 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
7830 #endif /* CONFIG_SCHED_MC */
7832 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7834 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7835 struct sched_group
**sg
, struct cpumask
*mask
)
7839 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7840 group
= cpumask_first(mask
);
7842 *sg
= &per_cpu(sched_group_core
, group
).sg
;
7845 #elif defined(CONFIG_SCHED_MC)
7847 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
7848 struct sched_group
**sg
, struct cpumask
*unused
)
7851 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
7856 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
7857 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
7860 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
7861 struct sched_group
**sg
, struct cpumask
*mask
)
7864 #ifdef CONFIG_SCHED_MC
7865 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
7866 group
= cpumask_first(mask
);
7867 #elif defined(CONFIG_SCHED_SMT)
7868 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
7869 group
= cpumask_first(mask
);
7874 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
7880 * The init_sched_build_groups can't handle what we want to do with node
7881 * groups, so roll our own. Now each node has its own list of groups which
7882 * gets dynamically allocated.
7884 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
7885 static struct sched_group
***sched_group_nodes_bycpu
;
7887 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
7888 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
7890 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
7891 struct sched_group
**sg
,
7892 struct cpumask
*nodemask
)
7896 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
7897 group
= cpumask_first(nodemask
);
7900 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
7904 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7906 struct sched_group
*sg
= group_head
;
7912 for_each_cpu(j
, sched_group_cpus(sg
)) {
7913 struct sched_domain
*sd
;
7915 sd
= &per_cpu(phys_domains
, j
).sd
;
7916 if (j
!= cpumask_first(sched_group_cpus(sd
->groups
))) {
7918 * Only add "power" once for each
7924 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7927 } while (sg
!= group_head
);
7929 #endif /* CONFIG_NUMA */
7932 /* Free memory allocated for various sched_group structures */
7933 static void free_sched_groups(const struct cpumask
*cpu_map
,
7934 struct cpumask
*nodemask
)
7938 for_each_cpu(cpu
, cpu_map
) {
7939 struct sched_group
**sched_group_nodes
7940 = sched_group_nodes_bycpu
[cpu
];
7942 if (!sched_group_nodes
)
7945 for (i
= 0; i
< nr_node_ids
; i
++) {
7946 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7948 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
7949 if (cpumask_empty(nodemask
))
7959 if (oldsg
!= sched_group_nodes
[i
])
7962 kfree(sched_group_nodes
);
7963 sched_group_nodes_bycpu
[cpu
] = NULL
;
7966 #else /* !CONFIG_NUMA */
7967 static void free_sched_groups(const struct cpumask
*cpu_map
,
7968 struct cpumask
*nodemask
)
7971 #endif /* CONFIG_NUMA */
7974 * Initialize sched groups cpu_power.
7976 * cpu_power indicates the capacity of sched group, which is used while
7977 * distributing the load between different sched groups in a sched domain.
7978 * Typically cpu_power for all the groups in a sched domain will be same unless
7979 * there are asymmetries in the topology. If there are asymmetries, group
7980 * having more cpu_power will pickup more load compared to the group having
7983 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7984 * the maximum number of tasks a group can handle in the presence of other idle
7985 * or lightly loaded groups in the same sched domain.
7987 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7989 struct sched_domain
*child
;
7990 struct sched_group
*group
;
7992 WARN_ON(!sd
|| !sd
->groups
);
7994 if (cpu
!= cpumask_first(sched_group_cpus(sd
->groups
)))
7999 sd
->groups
->__cpu_power
= 0;
8002 * For perf policy, if the groups in child domain share resources
8003 * (for example cores sharing some portions of the cache hierarchy
8004 * or SMT), then set this domain groups cpu_power such that each group
8005 * can handle only one task, when there are other idle groups in the
8006 * same sched domain.
8008 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
8010 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
8011 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
8016 * add cpu_power of each child group to this groups cpu_power
8018 group
= child
->groups
;
8020 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
8021 group
= group
->next
;
8022 } while (group
!= child
->groups
);
8026 * Initializers for schedule domains
8027 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8030 #ifdef CONFIG_SCHED_DEBUG
8031 # define SD_INIT_NAME(sd, type) sd->name = #type
8033 # define SD_INIT_NAME(sd, type) do { } while (0)
8036 #define SD_INIT(sd, type) sd_init_##type(sd)
8038 #define SD_INIT_FUNC(type) \
8039 static noinline void sd_init_##type(struct sched_domain *sd) \
8041 memset(sd, 0, sizeof(*sd)); \
8042 *sd = SD_##type##_INIT; \
8043 sd->level = SD_LV_##type; \
8044 SD_INIT_NAME(sd, type); \
8049 SD_INIT_FUNC(ALLNODES
)
8052 #ifdef CONFIG_SCHED_SMT
8053 SD_INIT_FUNC(SIBLING
)
8055 #ifdef CONFIG_SCHED_MC
8059 static int default_relax_domain_level
= -1;
8061 static int __init
setup_relax_domain_level(char *str
)
8065 val
= simple_strtoul(str
, NULL
, 0);
8066 if (val
< SD_LV_MAX
)
8067 default_relax_domain_level
= val
;
8071 __setup("relax_domain_level=", setup_relax_domain_level
);
8073 static void set_domain_attribute(struct sched_domain
*sd
,
8074 struct sched_domain_attr
*attr
)
8078 if (!attr
|| attr
->relax_domain_level
< 0) {
8079 if (default_relax_domain_level
< 0)
8082 request
= default_relax_domain_level
;
8084 request
= attr
->relax_domain_level
;
8085 if (request
< sd
->level
) {
8086 /* turn off idle balance on this domain */
8087 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
8089 /* turn on idle balance on this domain */
8090 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
8095 * Build sched domains for a given set of cpus and attach the sched domains
8096 * to the individual cpus
8098 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8099 struct sched_domain_attr
*attr
)
8101 int i
, err
= -ENOMEM
;
8102 struct root_domain
*rd
;
8103 cpumask_var_t nodemask
, this_sibling_map
, this_core_map
, send_covered
,
8106 cpumask_var_t domainspan
, covered
, notcovered
;
8107 struct sched_group
**sched_group_nodes
= NULL
;
8108 int sd_allnodes
= 0;
8110 if (!alloc_cpumask_var(&domainspan
, GFP_KERNEL
))
8112 if (!alloc_cpumask_var(&covered
, GFP_KERNEL
))
8113 goto free_domainspan
;
8114 if (!alloc_cpumask_var(¬covered
, GFP_KERNEL
))
8118 if (!alloc_cpumask_var(&nodemask
, GFP_KERNEL
))
8119 goto free_notcovered
;
8120 if (!alloc_cpumask_var(&this_sibling_map
, GFP_KERNEL
))
8122 if (!alloc_cpumask_var(&this_core_map
, GFP_KERNEL
))
8123 goto free_this_sibling_map
;
8124 if (!alloc_cpumask_var(&send_covered
, GFP_KERNEL
))
8125 goto free_this_core_map
;
8126 if (!alloc_cpumask_var(&tmpmask
, GFP_KERNEL
))
8127 goto free_send_covered
;
8131 * Allocate the per-node list of sched groups
8133 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
8135 if (!sched_group_nodes
) {
8136 printk(KERN_WARNING
"Can not alloc sched group node list\n");
8141 rd
= alloc_rootdomain();
8143 printk(KERN_WARNING
"Cannot alloc root domain\n");
8144 goto free_sched_groups
;
8148 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = sched_group_nodes
;
8152 * Set up domains for cpus specified by the cpu_map.
8154 for_each_cpu(i
, cpu_map
) {
8155 struct sched_domain
*sd
= NULL
, *p
;
8157 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(i
)), cpu_map
);
8160 if (cpumask_weight(cpu_map
) >
8161 SD_NODES_PER_DOMAIN
*cpumask_weight(nodemask
)) {
8162 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8163 SD_INIT(sd
, ALLNODES
);
8164 set_domain_attribute(sd
, attr
);
8165 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8166 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8172 sd
= &per_cpu(node_domains
, i
).sd
;
8174 set_domain_attribute(sd
, attr
);
8175 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8179 cpumask_and(sched_domain_span(sd
),
8180 sched_domain_span(sd
), cpu_map
);
8184 sd
= &per_cpu(phys_domains
, i
).sd
;
8186 set_domain_attribute(sd
, attr
);
8187 cpumask_copy(sched_domain_span(sd
), nodemask
);
8191 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8193 #ifdef CONFIG_SCHED_MC
8195 sd
= &per_cpu(core_domains
, i
).sd
;
8197 set_domain_attribute(sd
, attr
);
8198 cpumask_and(sched_domain_span(sd
), cpu_map
,
8199 cpu_coregroup_mask(i
));
8202 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8205 #ifdef CONFIG_SCHED_SMT
8207 sd
= &per_cpu(cpu_domains
, i
).sd
;
8208 SD_INIT(sd
, SIBLING
);
8209 set_domain_attribute(sd
, attr
);
8210 cpumask_and(sched_domain_span(sd
),
8211 topology_thread_cpumask(i
), cpu_map
);
8214 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
8218 #ifdef CONFIG_SCHED_SMT
8219 /* Set up CPU (sibling) groups */
8220 for_each_cpu(i
, cpu_map
) {
8221 cpumask_and(this_sibling_map
,
8222 topology_thread_cpumask(i
), cpu_map
);
8223 if (i
!= cpumask_first(this_sibling_map
))
8226 init_sched_build_groups(this_sibling_map
, cpu_map
,
8228 send_covered
, tmpmask
);
8232 #ifdef CONFIG_SCHED_MC
8233 /* Set up multi-core groups */
8234 for_each_cpu(i
, cpu_map
) {
8235 cpumask_and(this_core_map
, cpu_coregroup_mask(i
), cpu_map
);
8236 if (i
!= cpumask_first(this_core_map
))
8239 init_sched_build_groups(this_core_map
, cpu_map
,
8241 send_covered
, tmpmask
);
8245 /* Set up physical groups */
8246 for (i
= 0; i
< nr_node_ids
; i
++) {
8247 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8248 if (cpumask_empty(nodemask
))
8251 init_sched_build_groups(nodemask
, cpu_map
,
8253 send_covered
, tmpmask
);
8257 /* Set up node groups */
8259 init_sched_build_groups(cpu_map
, cpu_map
,
8260 &cpu_to_allnodes_group
,
8261 send_covered
, tmpmask
);
8264 for (i
= 0; i
< nr_node_ids
; i
++) {
8265 /* Set up node groups */
8266 struct sched_group
*sg
, *prev
;
8269 cpumask_clear(covered
);
8270 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8271 if (cpumask_empty(nodemask
)) {
8272 sched_group_nodes
[i
] = NULL
;
8276 sched_domain_node_span(i
, domainspan
);
8277 cpumask_and(domainspan
, domainspan
, cpu_map
);
8279 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8282 printk(KERN_WARNING
"Can not alloc domain group for "
8286 sched_group_nodes
[i
] = sg
;
8287 for_each_cpu(j
, nodemask
) {
8288 struct sched_domain
*sd
;
8290 sd
= &per_cpu(node_domains
, j
).sd
;
8293 sg
->__cpu_power
= 0;
8294 cpumask_copy(sched_group_cpus(sg
), nodemask
);
8296 cpumask_or(covered
, covered
, nodemask
);
8299 for (j
= 0; j
< nr_node_ids
; j
++) {
8300 int n
= (i
+ j
) % nr_node_ids
;
8302 cpumask_complement(notcovered
, covered
);
8303 cpumask_and(tmpmask
, notcovered
, cpu_map
);
8304 cpumask_and(tmpmask
, tmpmask
, domainspan
);
8305 if (cpumask_empty(tmpmask
))
8308 cpumask_and(tmpmask
, tmpmask
, cpumask_of_node(n
));
8309 if (cpumask_empty(tmpmask
))
8312 sg
= kmalloc_node(sizeof(struct sched_group
) +
8317 "Can not alloc domain group for node %d\n", j
);
8320 sg
->__cpu_power
= 0;
8321 cpumask_copy(sched_group_cpus(sg
), tmpmask
);
8322 sg
->next
= prev
->next
;
8323 cpumask_or(covered
, covered
, tmpmask
);
8330 /* Calculate CPU power for physical packages and nodes */
8331 #ifdef CONFIG_SCHED_SMT
8332 for_each_cpu(i
, cpu_map
) {
8333 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
).sd
;
8335 init_sched_groups_power(i
, sd
);
8338 #ifdef CONFIG_SCHED_MC
8339 for_each_cpu(i
, cpu_map
) {
8340 struct sched_domain
*sd
= &per_cpu(core_domains
, i
).sd
;
8342 init_sched_groups_power(i
, sd
);
8346 for_each_cpu(i
, cpu_map
) {
8347 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
).sd
;
8349 init_sched_groups_power(i
, sd
);
8353 for (i
= 0; i
< nr_node_ids
; i
++)
8354 init_numa_sched_groups_power(sched_group_nodes
[i
]);
8357 struct sched_group
*sg
;
8359 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8361 init_numa_sched_groups_power(sg
);
8365 /* Attach the domains */
8366 for_each_cpu(i
, cpu_map
) {
8367 struct sched_domain
*sd
;
8368 #ifdef CONFIG_SCHED_SMT
8369 sd
= &per_cpu(cpu_domains
, i
).sd
;
8370 #elif defined(CONFIG_SCHED_MC)
8371 sd
= &per_cpu(core_domains
, i
).sd
;
8373 sd
= &per_cpu(phys_domains
, i
).sd
;
8375 cpu_attach_domain(sd
, rd
, i
);
8381 free_cpumask_var(tmpmask
);
8383 free_cpumask_var(send_covered
);
8385 free_cpumask_var(this_core_map
);
8386 free_this_sibling_map
:
8387 free_cpumask_var(this_sibling_map
);
8389 free_cpumask_var(nodemask
);
8392 free_cpumask_var(notcovered
);
8394 free_cpumask_var(covered
);
8396 free_cpumask_var(domainspan
);
8403 kfree(sched_group_nodes
);
8409 free_sched_groups(cpu_map
, tmpmask
);
8410 free_rootdomain(rd
);
8415 static int build_sched_domains(const struct cpumask
*cpu_map
)
8417 return __build_sched_domains(cpu_map
, NULL
);
8420 static struct cpumask
*doms_cur
; /* current sched domains */
8421 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8422 static struct sched_domain_attr
*dattr_cur
;
8423 /* attribues of custom domains in 'doms_cur' */
8426 * Special case: If a kmalloc of a doms_cur partition (array of
8427 * cpumask) fails, then fallback to a single sched domain,
8428 * as determined by the single cpumask fallback_doms.
8430 static cpumask_var_t fallback_doms
;
8433 * arch_update_cpu_topology lets virtualized architectures update the
8434 * cpu core maps. It is supposed to return 1 if the topology changed
8435 * or 0 if it stayed the same.
8437 int __attribute__((weak
)) arch_update_cpu_topology(void)
8443 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
8444 * For now this just excludes isolated cpus, but could be used to
8445 * exclude other special cases in the future.
8447 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
8451 arch_update_cpu_topology();
8453 doms_cur
= kmalloc(cpumask_size(), GFP_KERNEL
);
8455 doms_cur
= fallback_doms
;
8456 cpumask_andnot(doms_cur
, cpu_map
, cpu_isolated_map
);
8458 err
= build_sched_domains(doms_cur
);
8459 register_sched_domain_sysctl();
8464 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
8465 struct cpumask
*tmpmask
)
8467 free_sched_groups(cpu_map
, tmpmask
);
8471 * Detach sched domains from a group of cpus specified in cpu_map
8472 * These cpus will now be attached to the NULL domain
8474 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
8476 /* Save because hotplug lock held. */
8477 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
8480 for_each_cpu(i
, cpu_map
)
8481 cpu_attach_domain(NULL
, &def_root_domain
, i
);
8482 synchronize_sched();
8483 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
8486 /* handle null as "default" */
8487 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
8488 struct sched_domain_attr
*new, int idx_new
)
8490 struct sched_domain_attr tmp
;
8497 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
8498 new ? (new + idx_new
) : &tmp
,
8499 sizeof(struct sched_domain_attr
));
8503 * Partition sched domains as specified by the 'ndoms_new'
8504 * cpumasks in the array doms_new[] of cpumasks. This compares
8505 * doms_new[] to the current sched domain partitioning, doms_cur[].
8506 * It destroys each deleted domain and builds each new domain.
8508 * 'doms_new' is an array of cpumask's of length 'ndoms_new'.
8509 * The masks don't intersect (don't overlap.) We should setup one
8510 * sched domain for each mask. CPUs not in any of the cpumasks will
8511 * not be load balanced. If the same cpumask appears both in the
8512 * current 'doms_cur' domains and in the new 'doms_new', we can leave
8515 * The passed in 'doms_new' should be kmalloc'd. This routine takes
8516 * ownership of it and will kfree it when done with it. If the caller
8517 * failed the kmalloc call, then it can pass in doms_new == NULL &&
8518 * ndoms_new == 1, and partition_sched_domains() will fallback to
8519 * the single partition 'fallback_doms', it also forces the domains
8522 * If doms_new == NULL it will be replaced with cpu_online_mask.
8523 * ndoms_new == 0 is a special case for destroying existing domains,
8524 * and it will not create the default domain.
8526 * Call with hotplug lock held
8528 /* FIXME: Change to struct cpumask *doms_new[] */
8529 void partition_sched_domains(int ndoms_new
, struct cpumask
*doms_new
,
8530 struct sched_domain_attr
*dattr_new
)
8535 mutex_lock(&sched_domains_mutex
);
8537 /* always unregister in case we don't destroy any domains */
8538 unregister_sched_domain_sysctl();
8540 /* Let architecture update cpu core mappings. */
8541 new_topology
= arch_update_cpu_topology();
8543 n
= doms_new
? ndoms_new
: 0;
8545 /* Destroy deleted domains */
8546 for (i
= 0; i
< ndoms_cur
; i
++) {
8547 for (j
= 0; j
< n
&& !new_topology
; j
++) {
8548 if (cpumask_equal(&doms_cur
[i
], &doms_new
[j
])
8549 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
8552 /* no match - a current sched domain not in new doms_new[] */
8553 detach_destroy_domains(doms_cur
+ i
);
8558 if (doms_new
== NULL
) {
8560 doms_new
= fallback_doms
;
8561 cpumask_andnot(&doms_new
[0], cpu_online_mask
, cpu_isolated_map
);
8562 WARN_ON_ONCE(dattr_new
);
8565 /* Build new domains */
8566 for (i
= 0; i
< ndoms_new
; i
++) {
8567 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
8568 if (cpumask_equal(&doms_new
[i
], &doms_cur
[j
])
8569 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
8572 /* no match - add a new doms_new */
8573 __build_sched_domains(doms_new
+ i
,
8574 dattr_new
? dattr_new
+ i
: NULL
);
8579 /* Remember the new sched domains */
8580 if (doms_cur
!= fallback_doms
)
8582 kfree(dattr_cur
); /* kfree(NULL) is safe */
8583 doms_cur
= doms_new
;
8584 dattr_cur
= dattr_new
;
8585 ndoms_cur
= ndoms_new
;
8587 register_sched_domain_sysctl();
8589 mutex_unlock(&sched_domains_mutex
);
8592 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
8593 static void arch_reinit_sched_domains(void)
8597 /* Destroy domains first to force the rebuild */
8598 partition_sched_domains(0, NULL
, NULL
);
8600 rebuild_sched_domains();
8604 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
8606 unsigned int level
= 0;
8608 if (sscanf(buf
, "%u", &level
) != 1)
8612 * level is always be positive so don't check for
8613 * level < POWERSAVINGS_BALANCE_NONE which is 0
8614 * What happens on 0 or 1 byte write,
8615 * need to check for count as well?
8618 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
8622 sched_smt_power_savings
= level
;
8624 sched_mc_power_savings
= level
;
8626 arch_reinit_sched_domains();
8631 #ifdef CONFIG_SCHED_MC
8632 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
8635 return sprintf(page
, "%u\n", sched_mc_power_savings
);
8637 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
8638 const char *buf
, size_t count
)
8640 return sched_power_savings_store(buf
, count
, 0);
8642 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
8643 sched_mc_power_savings_show
,
8644 sched_mc_power_savings_store
);
8647 #ifdef CONFIG_SCHED_SMT
8648 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
8651 return sprintf(page
, "%u\n", sched_smt_power_savings
);
8653 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
8654 const char *buf
, size_t count
)
8656 return sched_power_savings_store(buf
, count
, 1);
8658 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
8659 sched_smt_power_savings_show
,
8660 sched_smt_power_savings_store
);
8663 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
8667 #ifdef CONFIG_SCHED_SMT
8669 err
= sysfs_create_file(&cls
->kset
.kobj
,
8670 &attr_sched_smt_power_savings
.attr
);
8672 #ifdef CONFIG_SCHED_MC
8673 if (!err
&& mc_capable())
8674 err
= sysfs_create_file(&cls
->kset
.kobj
,
8675 &attr_sched_mc_power_savings
.attr
);
8679 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
8681 #ifndef CONFIG_CPUSETS
8683 * Add online and remove offline CPUs from the scheduler domains.
8684 * When cpusets are enabled they take over this function.
8686 static int update_sched_domains(struct notifier_block
*nfb
,
8687 unsigned long action
, void *hcpu
)
8691 case CPU_ONLINE_FROZEN
:
8693 case CPU_DEAD_FROZEN
:
8694 partition_sched_domains(1, NULL
, NULL
);
8703 static int update_runtime(struct notifier_block
*nfb
,
8704 unsigned long action
, void *hcpu
)
8706 int cpu
= (int)(long)hcpu
;
8709 case CPU_DOWN_PREPARE
:
8710 case CPU_DOWN_PREPARE_FROZEN
:
8711 disable_runtime(cpu_rq(cpu
));
8714 case CPU_DOWN_FAILED
:
8715 case CPU_DOWN_FAILED_FROZEN
:
8717 case CPU_ONLINE_FROZEN
:
8718 enable_runtime(cpu_rq(cpu
));
8726 void __init
sched_init_smp(void)
8728 cpumask_var_t non_isolated_cpus
;
8730 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
8732 #if defined(CONFIG_NUMA)
8733 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8735 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8738 mutex_lock(&sched_domains_mutex
);
8739 arch_init_sched_domains(cpu_online_mask
);
8740 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
8741 if (cpumask_empty(non_isolated_cpus
))
8742 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
8743 mutex_unlock(&sched_domains_mutex
);
8746 #ifndef CONFIG_CPUSETS
8747 /* XXX: Theoretical race here - CPU may be hotplugged now */
8748 hotcpu_notifier(update_sched_domains
, 0);
8751 /* RT runtime code needs to handle some hotplug events */
8752 hotcpu_notifier(update_runtime
, 0);
8756 /* Move init over to a non-isolated CPU */
8757 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
8759 sched_init_granularity();
8760 free_cpumask_var(non_isolated_cpus
);
8762 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
8763 init_sched_rt_class();
8766 void __init
sched_init_smp(void)
8768 sched_init_granularity();
8770 #endif /* CONFIG_SMP */
8772 int in_sched_functions(unsigned long addr
)
8774 return in_lock_functions(addr
) ||
8775 (addr
>= (unsigned long)__sched_text_start
8776 && addr
< (unsigned long)__sched_text_end
);
8779 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8781 cfs_rq
->tasks_timeline
= RB_ROOT
;
8782 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8783 #ifdef CONFIG_FAIR_GROUP_SCHED
8786 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8789 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8791 struct rt_prio_array
*array
;
8794 array
= &rt_rq
->active
;
8795 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8796 INIT_LIST_HEAD(array
->queue
+ i
);
8797 __clear_bit(i
, array
->bitmap
);
8799 /* delimiter for bitsearch: */
8800 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8802 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8803 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
8805 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
8809 rt_rq
->rt_nr_migratory
= 0;
8810 rt_rq
->overloaded
= 0;
8811 plist_head_init(&rq
->rt
.pushable_tasks
, &rq
->lock
);
8815 rt_rq
->rt_throttled
= 0;
8816 rt_rq
->rt_runtime
= 0;
8817 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8819 #ifdef CONFIG_RT_GROUP_SCHED
8820 rt_rq
->rt_nr_boosted
= 0;
8825 #ifdef CONFIG_FAIR_GROUP_SCHED
8826 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8827 struct sched_entity
*se
, int cpu
, int add
,
8828 struct sched_entity
*parent
)
8830 struct rq
*rq
= cpu_rq(cpu
);
8831 tg
->cfs_rq
[cpu
] = cfs_rq
;
8832 init_cfs_rq(cfs_rq
, rq
);
8835 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8838 /* se could be NULL for init_task_group */
8843 se
->cfs_rq
= &rq
->cfs
;
8845 se
->cfs_rq
= parent
->my_q
;
8848 se
->load
.weight
= tg
->shares
;
8849 se
->load
.inv_weight
= 0;
8850 se
->parent
= parent
;
8854 #ifdef CONFIG_RT_GROUP_SCHED
8855 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8856 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8857 struct sched_rt_entity
*parent
)
8859 struct rq
*rq
= cpu_rq(cpu
);
8861 tg
->rt_rq
[cpu
] = rt_rq
;
8862 init_rt_rq(rt_rq
, rq
);
8864 rt_rq
->rt_se
= rt_se
;
8865 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8867 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8869 tg
->rt_se
[cpu
] = rt_se
;
8874 rt_se
->rt_rq
= &rq
->rt
;
8876 rt_se
->rt_rq
= parent
->my_q
;
8878 rt_se
->my_q
= rt_rq
;
8879 rt_se
->parent
= parent
;
8880 INIT_LIST_HEAD(&rt_se
->run_list
);
8884 void __init
sched_init(void)
8887 unsigned long alloc_size
= 0, ptr
;
8889 #ifdef CONFIG_FAIR_GROUP_SCHED
8890 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8892 #ifdef CONFIG_RT_GROUP_SCHED
8893 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8895 #ifdef CONFIG_USER_SCHED
8898 #ifdef CONFIG_CPUMASK_OFFSTACK
8899 alloc_size
+= num_possible_cpus() * cpumask_size();
8902 * As sched_init() is called before page_alloc is setup,
8903 * we use alloc_bootmem().
8906 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8908 #ifdef CONFIG_FAIR_GROUP_SCHED
8909 init_task_group
.se
= (struct sched_entity
**)ptr
;
8910 ptr
+= nr_cpu_ids
* sizeof(void **);
8912 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8913 ptr
+= nr_cpu_ids
* sizeof(void **);
8915 #ifdef CONFIG_USER_SCHED
8916 root_task_group
.se
= (struct sched_entity
**)ptr
;
8917 ptr
+= nr_cpu_ids
* sizeof(void **);
8919 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8920 ptr
+= nr_cpu_ids
* sizeof(void **);
8921 #endif /* CONFIG_USER_SCHED */
8922 #endif /* CONFIG_FAIR_GROUP_SCHED */
8923 #ifdef CONFIG_RT_GROUP_SCHED
8924 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8925 ptr
+= nr_cpu_ids
* sizeof(void **);
8927 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8928 ptr
+= nr_cpu_ids
* sizeof(void **);
8930 #ifdef CONFIG_USER_SCHED
8931 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8932 ptr
+= nr_cpu_ids
* sizeof(void **);
8934 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8935 ptr
+= nr_cpu_ids
* sizeof(void **);
8936 #endif /* CONFIG_USER_SCHED */
8937 #endif /* CONFIG_RT_GROUP_SCHED */
8938 #ifdef CONFIG_CPUMASK_OFFSTACK
8939 for_each_possible_cpu(i
) {
8940 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
8941 ptr
+= cpumask_size();
8943 #endif /* CONFIG_CPUMASK_OFFSTACK */
8947 init_defrootdomain();
8950 init_rt_bandwidth(&def_rt_bandwidth
,
8951 global_rt_period(), global_rt_runtime());
8953 #ifdef CONFIG_RT_GROUP_SCHED
8954 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8955 global_rt_period(), global_rt_runtime());
8956 #ifdef CONFIG_USER_SCHED
8957 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8958 global_rt_period(), RUNTIME_INF
);
8959 #endif /* CONFIG_USER_SCHED */
8960 #endif /* CONFIG_RT_GROUP_SCHED */
8962 #ifdef CONFIG_GROUP_SCHED
8963 list_add(&init_task_group
.list
, &task_groups
);
8964 INIT_LIST_HEAD(&init_task_group
.children
);
8966 #ifdef CONFIG_USER_SCHED
8967 INIT_LIST_HEAD(&root_task_group
.children
);
8968 init_task_group
.parent
= &root_task_group
;
8969 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8970 #endif /* CONFIG_USER_SCHED */
8971 #endif /* CONFIG_GROUP_SCHED */
8973 for_each_possible_cpu(i
) {
8977 spin_lock_init(&rq
->lock
);
8979 init_cfs_rq(&rq
->cfs
, rq
);
8980 init_rt_rq(&rq
->rt
, rq
);
8981 #ifdef CONFIG_FAIR_GROUP_SCHED
8982 init_task_group
.shares
= init_task_group_load
;
8983 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8984 #ifdef CONFIG_CGROUP_SCHED
8986 * How much cpu bandwidth does init_task_group get?
8988 * In case of task-groups formed thr' the cgroup filesystem, it
8989 * gets 100% of the cpu resources in the system. This overall
8990 * system cpu resource is divided among the tasks of
8991 * init_task_group and its child task-groups in a fair manner,
8992 * based on each entity's (task or task-group's) weight
8993 * (se->load.weight).
8995 * In other words, if init_task_group has 10 tasks of weight
8996 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8997 * then A0's share of the cpu resource is:
8999 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9001 * We achieve this by letting init_task_group's tasks sit
9002 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9004 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9005 #elif defined CONFIG_USER_SCHED
9006 root_task_group
.shares
= NICE_0_LOAD
;
9007 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9009 * In case of task-groups formed thr' the user id of tasks,
9010 * init_task_group represents tasks belonging to root user.
9011 * Hence it forms a sibling of all subsequent groups formed.
9012 * In this case, init_task_group gets only a fraction of overall
9013 * system cpu resource, based on the weight assigned to root
9014 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9015 * by letting tasks of init_task_group sit in a separate cfs_rq
9016 * (init_cfs_rq) and having one entity represent this group of
9017 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9019 init_tg_cfs_entry(&init_task_group
,
9020 &per_cpu(init_cfs_rq
, i
),
9021 &per_cpu(init_sched_entity
, i
), i
, 1,
9022 root_task_group
.se
[i
]);
9025 #endif /* CONFIG_FAIR_GROUP_SCHED */
9027 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9028 #ifdef CONFIG_RT_GROUP_SCHED
9029 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9030 #ifdef CONFIG_CGROUP_SCHED
9031 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9032 #elif defined CONFIG_USER_SCHED
9033 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9034 init_tg_rt_entry(&init_task_group
,
9035 &per_cpu(init_rt_rq
, i
),
9036 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9037 root_task_group
.rt_se
[i
]);
9041 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9042 rq
->cpu_load
[j
] = 0;
9046 rq
->active_balance
= 0;
9047 rq
->next_balance
= jiffies
;
9051 rq
->migration_thread
= NULL
;
9052 INIT_LIST_HEAD(&rq
->migration_queue
);
9053 rq_attach_root(rq
, &def_root_domain
);
9056 atomic_set(&rq
->nr_iowait
, 0);
9059 set_load_weight(&init_task
);
9061 #ifdef CONFIG_PREEMPT_NOTIFIERS
9062 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9066 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9069 #ifdef CONFIG_RT_MUTEXES
9070 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9074 * The boot idle thread does lazy MMU switching as well:
9076 atomic_inc(&init_mm
.mm_count
);
9077 enter_lazy_tlb(&init_mm
, current
);
9080 * Make us the idle thread. Technically, schedule() should not be
9081 * called from this thread, however somewhere below it might be,
9082 * but because we are the idle thread, we just pick up running again
9083 * when this runqueue becomes "idle".
9085 init_idle(current
, smp_processor_id());
9087 * During early bootup we pretend to be a normal task:
9089 current
->sched_class
= &fair_sched_class
;
9091 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9092 alloc_bootmem_cpumask_var(&nohz_cpu_mask
);
9095 alloc_bootmem_cpumask_var(&nohz
.cpu_mask
);
9097 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
9100 scheduler_running
= 1;
9103 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9104 void __might_sleep(char *file
, int line
)
9107 static unsigned long prev_jiffy
; /* ratelimiting */
9109 if ((!in_atomic() && !irqs_disabled()) ||
9110 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9112 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9114 prev_jiffy
= jiffies
;
9117 "BUG: sleeping function called from invalid context at %s:%d\n",
9120 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9121 in_atomic(), irqs_disabled(),
9122 current
->pid
, current
->comm
);
9124 debug_show_held_locks(current
);
9125 if (irqs_disabled())
9126 print_irqtrace_events(current
);
9130 EXPORT_SYMBOL(__might_sleep
);
9133 #ifdef CONFIG_MAGIC_SYSRQ
9134 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9138 update_rq_clock(rq
);
9139 on_rq
= p
->se
.on_rq
;
9141 deactivate_task(rq
, p
, 0);
9142 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9144 activate_task(rq
, p
, 0);
9145 resched_task(rq
->curr
);
9149 void normalize_rt_tasks(void)
9151 struct task_struct
*g
, *p
;
9152 unsigned long flags
;
9155 read_lock_irqsave(&tasklist_lock
, flags
);
9156 do_each_thread(g
, p
) {
9158 * Only normalize user tasks:
9163 p
->se
.exec_start
= 0;
9164 #ifdef CONFIG_SCHEDSTATS
9165 p
->se
.wait_start
= 0;
9166 p
->se
.sleep_start
= 0;
9167 p
->se
.block_start
= 0;
9172 * Renice negative nice level userspace
9175 if (TASK_NICE(p
) < 0 && p
->mm
)
9176 set_user_nice(p
, 0);
9180 spin_lock(&p
->pi_lock
);
9181 rq
= __task_rq_lock(p
);
9183 normalize_task(rq
, p
);
9185 __task_rq_unlock(rq
);
9186 spin_unlock(&p
->pi_lock
);
9187 } while_each_thread(g
, p
);
9189 read_unlock_irqrestore(&tasklist_lock
, flags
);
9192 #endif /* CONFIG_MAGIC_SYSRQ */
9196 * These functions are only useful for the IA64 MCA handling.
9198 * They can only be called when the whole system has been
9199 * stopped - every CPU needs to be quiescent, and no scheduling
9200 * activity can take place. Using them for anything else would
9201 * be a serious bug, and as a result, they aren't even visible
9202 * under any other configuration.
9206 * curr_task - return the current task for a given cpu.
9207 * @cpu: the processor in question.
9209 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9211 struct task_struct
*curr_task(int cpu
)
9213 return cpu_curr(cpu
);
9217 * set_curr_task - set the current task for a given cpu.
9218 * @cpu: the processor in question.
9219 * @p: the task pointer to set.
9221 * Description: This function must only be used when non-maskable interrupts
9222 * are serviced on a separate stack. It allows the architecture to switch the
9223 * notion of the current task on a cpu in a non-blocking manner. This function
9224 * must be called with all CPU's synchronized, and interrupts disabled, the
9225 * and caller must save the original value of the current task (see
9226 * curr_task() above) and restore that value before reenabling interrupts and
9227 * re-starting the system.
9229 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9231 void set_curr_task(int cpu
, struct task_struct
*p
)
9238 #ifdef CONFIG_FAIR_GROUP_SCHED
9239 static void free_fair_sched_group(struct task_group
*tg
)
9243 for_each_possible_cpu(i
) {
9245 kfree(tg
->cfs_rq
[i
]);
9255 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9257 struct cfs_rq
*cfs_rq
;
9258 struct sched_entity
*se
;
9262 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9265 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9269 tg
->shares
= NICE_0_LOAD
;
9271 for_each_possible_cpu(i
) {
9274 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9275 GFP_KERNEL
, cpu_to_node(i
));
9279 se
= kzalloc_node(sizeof(struct sched_entity
),
9280 GFP_KERNEL
, cpu_to_node(i
));
9284 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9293 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9295 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9296 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9299 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9301 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9303 #else /* !CONFG_FAIR_GROUP_SCHED */
9304 static inline void free_fair_sched_group(struct task_group
*tg
)
9309 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9314 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9318 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9321 #endif /* CONFIG_FAIR_GROUP_SCHED */
9323 #ifdef CONFIG_RT_GROUP_SCHED
9324 static void free_rt_sched_group(struct task_group
*tg
)
9328 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9330 for_each_possible_cpu(i
) {
9332 kfree(tg
->rt_rq
[i
]);
9334 kfree(tg
->rt_se
[i
]);
9342 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9344 struct rt_rq
*rt_rq
;
9345 struct sched_rt_entity
*rt_se
;
9349 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9352 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9356 init_rt_bandwidth(&tg
->rt_bandwidth
,
9357 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9359 for_each_possible_cpu(i
) {
9362 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9363 GFP_KERNEL
, cpu_to_node(i
));
9367 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9368 GFP_KERNEL
, cpu_to_node(i
));
9372 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9381 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9383 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9384 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9387 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9389 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
9391 #else /* !CONFIG_RT_GROUP_SCHED */
9392 static inline void free_rt_sched_group(struct task_group
*tg
)
9397 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9402 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9406 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
9409 #endif /* CONFIG_RT_GROUP_SCHED */
9411 #ifdef CONFIG_GROUP_SCHED
9412 static void free_sched_group(struct task_group
*tg
)
9414 free_fair_sched_group(tg
);
9415 free_rt_sched_group(tg
);
9419 /* allocate runqueue etc for a new task group */
9420 struct task_group
*sched_create_group(struct task_group
*parent
)
9422 struct task_group
*tg
;
9423 unsigned long flags
;
9426 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
9428 return ERR_PTR(-ENOMEM
);
9430 if (!alloc_fair_sched_group(tg
, parent
))
9433 if (!alloc_rt_sched_group(tg
, parent
))
9436 spin_lock_irqsave(&task_group_lock
, flags
);
9437 for_each_possible_cpu(i
) {
9438 register_fair_sched_group(tg
, i
);
9439 register_rt_sched_group(tg
, i
);
9441 list_add_rcu(&tg
->list
, &task_groups
);
9443 WARN_ON(!parent
); /* root should already exist */
9445 tg
->parent
= parent
;
9446 INIT_LIST_HEAD(&tg
->children
);
9447 list_add_rcu(&tg
->siblings
, &parent
->children
);
9448 spin_unlock_irqrestore(&task_group_lock
, flags
);
9453 free_sched_group(tg
);
9454 return ERR_PTR(-ENOMEM
);
9457 /* rcu callback to free various structures associated with a task group */
9458 static void free_sched_group_rcu(struct rcu_head
*rhp
)
9460 /* now it should be safe to free those cfs_rqs */
9461 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
9464 /* Destroy runqueue etc associated with a task group */
9465 void sched_destroy_group(struct task_group
*tg
)
9467 unsigned long flags
;
9470 spin_lock_irqsave(&task_group_lock
, flags
);
9471 for_each_possible_cpu(i
) {
9472 unregister_fair_sched_group(tg
, i
);
9473 unregister_rt_sched_group(tg
, i
);
9475 list_del_rcu(&tg
->list
);
9476 list_del_rcu(&tg
->siblings
);
9477 spin_unlock_irqrestore(&task_group_lock
, flags
);
9479 /* wait for possible concurrent references to cfs_rqs complete */
9480 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
9483 /* change task's runqueue when it moves between groups.
9484 * The caller of this function should have put the task in its new group
9485 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
9486 * reflect its new group.
9488 void sched_move_task(struct task_struct
*tsk
)
9491 unsigned long flags
;
9494 rq
= task_rq_lock(tsk
, &flags
);
9496 update_rq_clock(rq
);
9498 running
= task_current(rq
, tsk
);
9499 on_rq
= tsk
->se
.on_rq
;
9502 dequeue_task(rq
, tsk
, 0);
9503 if (unlikely(running
))
9504 tsk
->sched_class
->put_prev_task(rq
, tsk
);
9506 set_task_rq(tsk
, task_cpu(tsk
));
9508 #ifdef CONFIG_FAIR_GROUP_SCHED
9509 if (tsk
->sched_class
->moved_group
)
9510 tsk
->sched_class
->moved_group(tsk
);
9513 if (unlikely(running
))
9514 tsk
->sched_class
->set_curr_task(rq
);
9516 enqueue_task(rq
, tsk
, 0);
9518 task_rq_unlock(rq
, &flags
);
9520 #endif /* CONFIG_GROUP_SCHED */
9522 #ifdef CONFIG_FAIR_GROUP_SCHED
9523 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9525 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9530 dequeue_entity(cfs_rq
, se
, 0);
9532 se
->load
.weight
= shares
;
9533 se
->load
.inv_weight
= 0;
9536 enqueue_entity(cfs_rq
, se
, 0);
9539 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
9541 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
9542 struct rq
*rq
= cfs_rq
->rq
;
9543 unsigned long flags
;
9545 spin_lock_irqsave(&rq
->lock
, flags
);
9546 __set_se_shares(se
, shares
);
9547 spin_unlock_irqrestore(&rq
->lock
, flags
);
9550 static DEFINE_MUTEX(shares_mutex
);
9552 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
9555 unsigned long flags
;
9558 * We can't change the weight of the root cgroup.
9563 if (shares
< MIN_SHARES
)
9564 shares
= MIN_SHARES
;
9565 else if (shares
> MAX_SHARES
)
9566 shares
= MAX_SHARES
;
9568 mutex_lock(&shares_mutex
);
9569 if (tg
->shares
== shares
)
9572 spin_lock_irqsave(&task_group_lock
, flags
);
9573 for_each_possible_cpu(i
)
9574 unregister_fair_sched_group(tg
, i
);
9575 list_del_rcu(&tg
->siblings
);
9576 spin_unlock_irqrestore(&task_group_lock
, flags
);
9578 /* wait for any ongoing reference to this group to finish */
9579 synchronize_sched();
9582 * Now we are free to modify the group's share on each cpu
9583 * w/o tripping rebalance_share or load_balance_fair.
9585 tg
->shares
= shares
;
9586 for_each_possible_cpu(i
) {
9590 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
9591 set_se_shares(tg
->se
[i
], shares
);
9595 * Enable load balance activity on this group, by inserting it back on
9596 * each cpu's rq->leaf_cfs_rq_list.
9598 spin_lock_irqsave(&task_group_lock
, flags
);
9599 for_each_possible_cpu(i
)
9600 register_fair_sched_group(tg
, i
);
9601 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
9602 spin_unlock_irqrestore(&task_group_lock
, flags
);
9604 mutex_unlock(&shares_mutex
);
9608 unsigned long sched_group_shares(struct task_group
*tg
)
9614 #ifdef CONFIG_RT_GROUP_SCHED
9616 * Ensure that the real time constraints are schedulable.
9618 static DEFINE_MUTEX(rt_constraints_mutex
);
9620 static unsigned long to_ratio(u64 period
, u64 runtime
)
9622 if (runtime
== RUNTIME_INF
)
9625 return div64_u64(runtime
<< 20, period
);
9628 /* Must be called with tasklist_lock held */
9629 static inline int tg_has_rt_tasks(struct task_group
*tg
)
9631 struct task_struct
*g
, *p
;
9633 do_each_thread(g
, p
) {
9634 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
9636 } while_each_thread(g
, p
);
9641 struct rt_schedulable_data
{
9642 struct task_group
*tg
;
9647 static int tg_schedulable(struct task_group
*tg
, void *data
)
9649 struct rt_schedulable_data
*d
= data
;
9650 struct task_group
*child
;
9651 unsigned long total
, sum
= 0;
9652 u64 period
, runtime
;
9654 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9655 runtime
= tg
->rt_bandwidth
.rt_runtime
;
9658 period
= d
->rt_period
;
9659 runtime
= d
->rt_runtime
;
9662 #ifdef CONFIG_USER_SCHED
9663 if (tg
== &root_task_group
) {
9664 period
= global_rt_period();
9665 runtime
= global_rt_runtime();
9670 * Cannot have more runtime than the period.
9672 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9676 * Ensure we don't starve existing RT tasks.
9678 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
9681 total
= to_ratio(period
, runtime
);
9684 * Nobody can have more than the global setting allows.
9686 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
9690 * The sum of our children's runtime should not exceed our own.
9692 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
9693 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
9694 runtime
= child
->rt_bandwidth
.rt_runtime
;
9696 if (child
== d
->tg
) {
9697 period
= d
->rt_period
;
9698 runtime
= d
->rt_runtime
;
9701 sum
+= to_ratio(period
, runtime
);
9710 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
9712 struct rt_schedulable_data data
= {
9714 .rt_period
= period
,
9715 .rt_runtime
= runtime
,
9718 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
9721 static int tg_set_bandwidth(struct task_group
*tg
,
9722 u64 rt_period
, u64 rt_runtime
)
9726 mutex_lock(&rt_constraints_mutex
);
9727 read_lock(&tasklist_lock
);
9728 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
9732 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9733 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
9734 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
9736 for_each_possible_cpu(i
) {
9737 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
9739 spin_lock(&rt_rq
->rt_runtime_lock
);
9740 rt_rq
->rt_runtime
= rt_runtime
;
9741 spin_unlock(&rt_rq
->rt_runtime_lock
);
9743 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
9745 read_unlock(&tasklist_lock
);
9746 mutex_unlock(&rt_constraints_mutex
);
9751 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9753 u64 rt_runtime
, rt_period
;
9755 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9756 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9757 if (rt_runtime_us
< 0)
9758 rt_runtime
= RUNTIME_INF
;
9760 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9763 long sched_group_rt_runtime(struct task_group
*tg
)
9767 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9770 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9771 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9772 return rt_runtime_us
;
9775 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9777 u64 rt_runtime
, rt_period
;
9779 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9780 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9785 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9788 long sched_group_rt_period(struct task_group
*tg
)
9792 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9793 do_div(rt_period_us
, NSEC_PER_USEC
);
9794 return rt_period_us
;
9797 static int sched_rt_global_constraints(void)
9799 u64 runtime
, period
;
9802 if (sysctl_sched_rt_period
<= 0)
9805 runtime
= global_rt_runtime();
9806 period
= global_rt_period();
9809 * Sanity check on the sysctl variables.
9811 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9814 mutex_lock(&rt_constraints_mutex
);
9815 read_lock(&tasklist_lock
);
9816 ret
= __rt_schedulable(NULL
, 0, 0);
9817 read_unlock(&tasklist_lock
);
9818 mutex_unlock(&rt_constraints_mutex
);
9823 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
9825 /* Don't accept realtime tasks when there is no way for them to run */
9826 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
9832 #else /* !CONFIG_RT_GROUP_SCHED */
9833 static int sched_rt_global_constraints(void)
9835 unsigned long flags
;
9838 if (sysctl_sched_rt_period
<= 0)
9841 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9842 for_each_possible_cpu(i
) {
9843 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9845 spin_lock(&rt_rq
->rt_runtime_lock
);
9846 rt_rq
->rt_runtime
= global_rt_runtime();
9847 spin_unlock(&rt_rq
->rt_runtime_lock
);
9849 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9853 #endif /* CONFIG_RT_GROUP_SCHED */
9855 int sched_rt_handler(struct ctl_table
*table
, int write
,
9856 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9860 int old_period
, old_runtime
;
9861 static DEFINE_MUTEX(mutex
);
9864 old_period
= sysctl_sched_rt_period
;
9865 old_runtime
= sysctl_sched_rt_runtime
;
9867 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9869 if (!ret
&& write
) {
9870 ret
= sched_rt_global_constraints();
9872 sysctl_sched_rt_period
= old_period
;
9873 sysctl_sched_rt_runtime
= old_runtime
;
9875 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9876 def_rt_bandwidth
.rt_period
=
9877 ns_to_ktime(global_rt_period());
9880 mutex_unlock(&mutex
);
9885 #ifdef CONFIG_CGROUP_SCHED
9887 /* return corresponding task_group object of a cgroup */
9888 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9890 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9891 struct task_group
, css
);
9894 static struct cgroup_subsys_state
*
9895 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9897 struct task_group
*tg
, *parent
;
9899 if (!cgrp
->parent
) {
9900 /* This is early initialization for the top cgroup */
9901 return &init_task_group
.css
;
9904 parent
= cgroup_tg(cgrp
->parent
);
9905 tg
= sched_create_group(parent
);
9907 return ERR_PTR(-ENOMEM
);
9913 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9915 struct task_group
*tg
= cgroup_tg(cgrp
);
9917 sched_destroy_group(tg
);
9921 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9922 struct task_struct
*tsk
)
9924 #ifdef CONFIG_RT_GROUP_SCHED
9925 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
9928 /* We don't support RT-tasks being in separate groups */
9929 if (tsk
->sched_class
!= &fair_sched_class
)
9937 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9938 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9940 sched_move_task(tsk
);
9943 #ifdef CONFIG_FAIR_GROUP_SCHED
9944 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9947 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9950 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9952 struct task_group
*tg
= cgroup_tg(cgrp
);
9954 return (u64
) tg
->shares
;
9956 #endif /* CONFIG_FAIR_GROUP_SCHED */
9958 #ifdef CONFIG_RT_GROUP_SCHED
9959 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9962 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9965 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9967 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9970 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9973 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9976 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9978 return sched_group_rt_period(cgroup_tg(cgrp
));
9980 #endif /* CONFIG_RT_GROUP_SCHED */
9982 static struct cftype cpu_files
[] = {
9983 #ifdef CONFIG_FAIR_GROUP_SCHED
9986 .read_u64
= cpu_shares_read_u64
,
9987 .write_u64
= cpu_shares_write_u64
,
9990 #ifdef CONFIG_RT_GROUP_SCHED
9992 .name
= "rt_runtime_us",
9993 .read_s64
= cpu_rt_runtime_read
,
9994 .write_s64
= cpu_rt_runtime_write
,
9997 .name
= "rt_period_us",
9998 .read_u64
= cpu_rt_period_read_uint
,
9999 .write_u64
= cpu_rt_period_write_uint
,
10004 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10006 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10009 struct cgroup_subsys cpu_cgroup_subsys
= {
10011 .create
= cpu_cgroup_create
,
10012 .destroy
= cpu_cgroup_destroy
,
10013 .can_attach
= cpu_cgroup_can_attach
,
10014 .attach
= cpu_cgroup_attach
,
10015 .populate
= cpu_cgroup_populate
,
10016 .subsys_id
= cpu_cgroup_subsys_id
,
10020 #endif /* CONFIG_CGROUP_SCHED */
10022 #ifdef CONFIG_CGROUP_CPUACCT
10025 * CPU accounting code for task groups.
10027 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10028 * (balbir@in.ibm.com).
10031 /* track cpu usage of a group of tasks and its child groups */
10033 struct cgroup_subsys_state css
;
10034 /* cpuusage holds pointer to a u64-type object on every cpu */
10036 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10037 struct cpuacct
*parent
;
10040 struct cgroup_subsys cpuacct_subsys
;
10042 /* return cpu accounting group corresponding to this container */
10043 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10045 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10046 struct cpuacct
, css
);
10049 /* return cpu accounting group to which this task belongs */
10050 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10052 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10053 struct cpuacct
, css
);
10056 /* create a new cpu accounting group */
10057 static struct cgroup_subsys_state
*cpuacct_create(
10058 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10060 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10066 ca
->cpuusage
= alloc_percpu(u64
);
10070 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10071 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10072 goto out_free_counters
;
10075 ca
->parent
= cgroup_ca(cgrp
->parent
);
10081 percpu_counter_destroy(&ca
->cpustat
[i
]);
10082 free_percpu(ca
->cpuusage
);
10086 return ERR_PTR(-ENOMEM
);
10089 /* destroy an existing cpu accounting group */
10091 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10093 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10096 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10097 percpu_counter_destroy(&ca
->cpustat
[i
]);
10098 free_percpu(ca
->cpuusage
);
10102 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10104 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10107 #ifndef CONFIG_64BIT
10109 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10111 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10113 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10121 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10123 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10125 #ifndef CONFIG_64BIT
10127 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10129 spin_lock_irq(&cpu_rq(cpu
)->lock
);
10131 spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10137 /* return total cpu usage (in nanoseconds) of a group */
10138 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10140 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10141 u64 totalcpuusage
= 0;
10144 for_each_present_cpu(i
)
10145 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10147 return totalcpuusage
;
10150 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10153 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10162 for_each_present_cpu(i
)
10163 cpuacct_cpuusage_write(ca
, i
, 0);
10169 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10170 struct seq_file
*m
)
10172 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10176 for_each_present_cpu(i
) {
10177 percpu
= cpuacct_cpuusage_read(ca
, i
);
10178 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10180 seq_printf(m
, "\n");
10184 static const char *cpuacct_stat_desc
[] = {
10185 [CPUACCT_STAT_USER
] = "user",
10186 [CPUACCT_STAT_SYSTEM
] = "system",
10189 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10190 struct cgroup_map_cb
*cb
)
10192 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10195 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10196 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10197 val
= cputime64_to_clock_t(val
);
10198 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10203 static struct cftype files
[] = {
10206 .read_u64
= cpuusage_read
,
10207 .write_u64
= cpuusage_write
,
10210 .name
= "usage_percpu",
10211 .read_seq_string
= cpuacct_percpu_seq_read
,
10215 .read_map
= cpuacct_stats_show
,
10219 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10221 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10225 * charge this task's execution time to its accounting group.
10227 * called with rq->lock held.
10229 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10231 struct cpuacct
*ca
;
10234 if (unlikely(!cpuacct_subsys
.active
))
10237 cpu
= task_cpu(tsk
);
10243 for (; ca
; ca
= ca
->parent
) {
10244 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10245 *cpuusage
+= cputime
;
10252 * Charge the system/user time to the task's accounting group.
10254 static void cpuacct_update_stats(struct task_struct
*tsk
,
10255 enum cpuacct_stat_index idx
, cputime_t val
)
10257 struct cpuacct
*ca
;
10259 if (unlikely(!cpuacct_subsys
.active
))
10266 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10272 struct cgroup_subsys cpuacct_subsys
= {
10274 .create
= cpuacct_create
,
10275 .destroy
= cpuacct_destroy
,
10276 .populate
= cpuacct_populate
,
10277 .subsys_id
= cpuacct_subsys_id
,
10279 #endif /* CONFIG_CGROUP_CPUACCT */