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 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
130 * Since cpu_power is a 'constant', we can use a reciprocal divide.
132 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
134 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
138 * Each time a sched group cpu_power is changed,
139 * we must compute its reciprocal value
141 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
143 sg
->__cpu_power
+= val
;
144 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
148 static inline int rt_policy(int policy
)
150 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
155 static inline int task_has_rt_policy(struct task_struct
*p
)
157 return rt_policy(p
->policy
);
161 * This is the priority-queue data structure of the RT scheduling class:
163 struct rt_prio_array
{
164 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
165 struct list_head queue
[MAX_RT_PRIO
];
168 struct rt_bandwidth
{
169 /* nests inside the rq lock: */
170 spinlock_t rt_runtime_lock
;
173 struct hrtimer rt_period_timer
;
176 static struct rt_bandwidth def_rt_bandwidth
;
178 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
180 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
182 struct rt_bandwidth
*rt_b
=
183 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
189 now
= hrtimer_cb_get_time(timer
);
190 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
195 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
198 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
202 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
204 rt_b
->rt_period
= ns_to_ktime(period
);
205 rt_b
->rt_runtime
= runtime
;
207 spin_lock_init(&rt_b
->rt_runtime_lock
);
209 hrtimer_init(&rt_b
->rt_period_timer
,
210 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
211 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
212 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
215 static inline int rt_bandwidth_enabled(void)
217 return sysctl_sched_rt_runtime
>= 0;
220 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
224 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
227 if (hrtimer_active(&rt_b
->rt_period_timer
))
230 spin_lock(&rt_b
->rt_runtime_lock
);
232 if (hrtimer_active(&rt_b
->rt_period_timer
))
235 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
236 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
237 hrtimer_start_expires(&rt_b
->rt_period_timer
,
240 spin_unlock(&rt_b
->rt_runtime_lock
);
243 #ifdef CONFIG_RT_GROUP_SCHED
244 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
246 hrtimer_cancel(&rt_b
->rt_period_timer
);
251 * sched_domains_mutex serializes calls to arch_init_sched_domains,
252 * detach_destroy_domains and partition_sched_domains.
254 static DEFINE_MUTEX(sched_domains_mutex
);
256 #ifdef CONFIG_GROUP_SCHED
258 #include <linux/cgroup.h>
262 static LIST_HEAD(task_groups
);
264 /* task group related information */
266 #ifdef CONFIG_CGROUP_SCHED
267 struct cgroup_subsys_state css
;
270 #ifdef CONFIG_FAIR_GROUP_SCHED
271 /* schedulable entities of this group on each cpu */
272 struct sched_entity
**se
;
273 /* runqueue "owned" by this group on each cpu */
274 struct cfs_rq
**cfs_rq
;
275 unsigned long shares
;
278 #ifdef CONFIG_RT_GROUP_SCHED
279 struct sched_rt_entity
**rt_se
;
280 struct rt_rq
**rt_rq
;
282 struct rt_bandwidth rt_bandwidth
;
286 struct list_head list
;
288 struct task_group
*parent
;
289 struct list_head siblings
;
290 struct list_head children
;
293 #ifdef CONFIG_USER_SCHED
297 * Every UID task group (including init_task_group aka UID-0) will
298 * be a child to this group.
300 struct task_group root_task_group
;
302 #ifdef CONFIG_FAIR_GROUP_SCHED
303 /* Default task group's sched entity on each cpu */
304 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
305 /* Default task group's cfs_rq on each cpu */
306 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
307 #endif /* CONFIG_FAIR_GROUP_SCHED */
309 #ifdef CONFIG_RT_GROUP_SCHED
310 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
311 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
312 #endif /* CONFIG_RT_GROUP_SCHED */
313 #else /* !CONFIG_USER_SCHED */
314 #define root_task_group init_task_group
315 #endif /* CONFIG_USER_SCHED */
317 /* task_group_lock serializes add/remove of task groups and also changes to
318 * a task group's cpu shares.
320 static DEFINE_SPINLOCK(task_group_lock
);
322 #ifdef CONFIG_FAIR_GROUP_SCHED
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group
;
348 /* return group to which a task belongs */
349 static inline struct task_group
*task_group(struct task_struct
*p
)
351 struct task_group
*tg
;
353 #ifdef CONFIG_USER_SCHED
355 #elif defined(CONFIG_CGROUP_SCHED)
356 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
357 struct task_group
, css
);
359 tg
= &init_task_group
;
364 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
365 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
367 #ifdef CONFIG_FAIR_GROUP_SCHED
368 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
369 p
->se
.parent
= task_group(p
)->se
[cpu
];
372 #ifdef CONFIG_RT_GROUP_SCHED
373 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
374 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
380 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
381 static inline struct task_group
*task_group(struct task_struct
*p
)
386 #endif /* CONFIG_GROUP_SCHED */
388 /* CFS-related fields in a runqueue */
390 struct load_weight load
;
391 unsigned long nr_running
;
396 struct rb_root tasks_timeline
;
397 struct rb_node
*rb_leftmost
;
399 struct list_head tasks
;
400 struct list_head
*balance_iterator
;
403 * 'curr' points to currently running entity on this cfs_rq.
404 * It is set to NULL otherwise (i.e when none are currently running).
406 struct sched_entity
*curr
, *next
, *last
;
408 unsigned int nr_spread_over
;
410 #ifdef CONFIG_FAIR_GROUP_SCHED
411 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
414 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
415 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
416 * (like users, containers etc.)
418 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
419 * list is used during load balance.
421 struct list_head leaf_cfs_rq_list
;
422 struct task_group
*tg
; /* group that "owns" this runqueue */
426 * the part of load.weight contributed by tasks
428 unsigned long task_weight
;
431 * h_load = weight * f(tg)
433 * Where f(tg) is the recursive weight fraction assigned to
436 unsigned long h_load
;
439 * this cpu's part of tg->shares
441 unsigned long shares
;
444 * load.weight at the time we set shares
446 unsigned long rq_weight
;
451 /* Real-Time classes' related field in a runqueue: */
453 struct rt_prio_array active
;
454 unsigned long rt_nr_running
;
455 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
456 int highest_prio
; /* highest queued rt task prio */
459 unsigned long rt_nr_migratory
;
465 /* Nests inside the rq lock: */
466 spinlock_t rt_runtime_lock
;
468 #ifdef CONFIG_RT_GROUP_SCHED
469 unsigned long rt_nr_boosted
;
472 struct list_head leaf_rt_rq_list
;
473 struct task_group
*tg
;
474 struct sched_rt_entity
*rt_se
;
481 * We add the notion of a root-domain which will be used to define per-domain
482 * variables. Each exclusive cpuset essentially defines an island domain by
483 * fully partitioning the member cpus from any other cpuset. Whenever a new
484 * exclusive cpuset is created, we also create and attach a new root-domain
494 * The "RT overload" flag: it gets set if a CPU has more than
495 * one runnable RT task.
500 struct cpupri cpupri
;
505 * By default the system creates a single root-domain with all cpus as
506 * members (mimicking the global state we have today).
508 static struct root_domain def_root_domain
;
513 * This is the main, per-CPU runqueue data structure.
515 * Locking rule: those places that want to lock multiple runqueues
516 * (such as the load balancing or the thread migration code), lock
517 * acquire operations must be ordered by ascending &runqueue.
524 * nr_running and cpu_load should be in the same cacheline because
525 * remote CPUs use both these fields when doing load calculation.
527 unsigned long nr_running
;
528 #define CPU_LOAD_IDX_MAX 5
529 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
530 unsigned char idle_at_tick
;
532 unsigned long last_tick_seen
;
533 unsigned char in_nohz_recently
;
535 /* capture load from *all* tasks on this cpu: */
536 struct load_weight load
;
537 unsigned long nr_load_updates
;
543 #ifdef CONFIG_FAIR_GROUP_SCHED
544 /* list of leaf cfs_rq on this cpu: */
545 struct list_head leaf_cfs_rq_list
;
547 #ifdef CONFIG_RT_GROUP_SCHED
548 struct list_head leaf_rt_rq_list
;
552 * This is part of a global counter where only the total sum
553 * over all CPUs matters. A task can increase this counter on
554 * one CPU and if it got migrated afterwards it may decrease
555 * it on another CPU. Always updated under the runqueue lock:
557 unsigned long nr_uninterruptible
;
559 struct task_struct
*curr
, *idle
;
560 unsigned long next_balance
;
561 struct mm_struct
*prev_mm
;
568 struct root_domain
*rd
;
569 struct sched_domain
*sd
;
571 /* For active balancing */
574 /* cpu of this runqueue: */
578 unsigned long avg_load_per_task
;
580 struct task_struct
*migration_thread
;
581 struct list_head migration_queue
;
584 #ifdef CONFIG_SCHED_HRTICK
586 int hrtick_csd_pending
;
587 struct call_single_data hrtick_csd
;
589 struct hrtimer hrtick_timer
;
592 #ifdef CONFIG_SCHEDSTATS
594 struct sched_info rq_sched_info
;
596 /* sys_sched_yield() stats */
597 unsigned int yld_exp_empty
;
598 unsigned int yld_act_empty
;
599 unsigned int yld_both_empty
;
600 unsigned int yld_count
;
602 /* schedule() stats */
603 unsigned int sched_switch
;
604 unsigned int sched_count
;
605 unsigned int sched_goidle
;
607 /* try_to_wake_up() stats */
608 unsigned int ttwu_count
;
609 unsigned int ttwu_local
;
612 unsigned int bkl_count
;
616 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
618 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
620 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
623 static inline int cpu_of(struct rq
*rq
)
633 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
634 * See detach_destroy_domains: synchronize_sched for details.
636 * The domain tree of any CPU may only be accessed from within
637 * preempt-disabled sections.
639 #define for_each_domain(cpu, __sd) \
640 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
642 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
643 #define this_rq() (&__get_cpu_var(runqueues))
644 #define task_rq(p) cpu_rq(task_cpu(p))
645 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
647 static inline void update_rq_clock(struct rq
*rq
)
649 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
653 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
655 #ifdef CONFIG_SCHED_DEBUG
656 # define const_debug __read_mostly
658 # define const_debug static const
664 * Returns true if the current cpu runqueue is locked.
665 * This interface allows printk to be called with the runqueue lock
666 * held and know whether or not it is OK to wake up the klogd.
668 int runqueue_is_locked(void)
671 struct rq
*rq
= cpu_rq(cpu
);
674 ret
= spin_is_locked(&rq
->lock
);
680 * Debugging: various feature bits
683 #define SCHED_FEAT(name, enabled) \
684 __SCHED_FEAT_##name ,
687 #include "sched_features.h"
692 #define SCHED_FEAT(name, enabled) \
693 (1UL << __SCHED_FEAT_##name) * enabled |
695 const_debug
unsigned int sysctl_sched_features
=
696 #include "sched_features.h"
701 #ifdef CONFIG_SCHED_DEBUG
702 #define SCHED_FEAT(name, enabled) \
705 static __read_mostly
char *sched_feat_names
[] = {
706 #include "sched_features.h"
712 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
714 filp
->private_data
= inode
->i_private
;
719 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
720 size_t cnt
, loff_t
*ppos
)
727 for (i
= 0; sched_feat_names
[i
]; i
++) {
728 len
+= strlen(sched_feat_names
[i
]);
732 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
736 for (i
= 0; sched_feat_names
[i
]; i
++) {
737 if (sysctl_sched_features
& (1UL << i
))
738 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
740 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
743 r
+= sprintf(buf
+ r
, "\n");
744 WARN_ON(r
>= len
+ 2);
746 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
754 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
755 size_t cnt
, loff_t
*ppos
)
765 if (copy_from_user(&buf
, ubuf
, cnt
))
770 if (strncmp(buf
, "NO_", 3) == 0) {
775 for (i
= 0; sched_feat_names
[i
]; i
++) {
776 int len
= strlen(sched_feat_names
[i
]);
778 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
780 sysctl_sched_features
&= ~(1UL << i
);
782 sysctl_sched_features
|= (1UL << i
);
787 if (!sched_feat_names
[i
])
795 static struct file_operations sched_feat_fops
= {
796 .open
= sched_feat_open
,
797 .read
= sched_feat_read
,
798 .write
= sched_feat_write
,
801 static __init
int sched_init_debug(void)
803 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
808 late_initcall(sched_init_debug
);
812 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
815 * Number of tasks to iterate in a single balance run.
816 * Limited because this is done with IRQs disabled.
818 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
821 * ratelimit for updating the group shares.
824 unsigned int sysctl_sched_shares_ratelimit
= 250000;
827 * Inject some fuzzyness into changing the per-cpu group shares
828 * this avoids remote rq-locks at the expense of fairness.
831 unsigned int sysctl_sched_shares_thresh
= 4;
834 * period over which we measure -rt task cpu usage in us.
837 unsigned int sysctl_sched_rt_period
= 1000000;
839 static __read_mostly
int scheduler_running
;
842 * part of the period that we allow rt tasks to run in us.
845 int sysctl_sched_rt_runtime
= 950000;
847 static inline u64
global_rt_period(void)
849 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
852 static inline u64
global_rt_runtime(void)
854 if (sysctl_sched_rt_runtime
< 0)
857 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
860 #ifndef prepare_arch_switch
861 # define prepare_arch_switch(next) do { } while (0)
863 #ifndef finish_arch_switch
864 # define finish_arch_switch(prev) do { } while (0)
867 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
869 return rq
->curr
== p
;
872 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
873 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
875 return task_current(rq
, p
);
878 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
882 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
884 #ifdef CONFIG_DEBUG_SPINLOCK
885 /* this is a valid case when another task releases the spinlock */
886 rq
->lock
.owner
= current
;
889 * If we are tracking spinlock dependencies then we have to
890 * fix up the runqueue lock - which gets 'carried over' from
893 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
895 spin_unlock_irq(&rq
->lock
);
898 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
899 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
904 return task_current(rq
, p
);
908 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
912 * We can optimise this out completely for !SMP, because the
913 * SMP rebalancing from interrupt is the only thing that cares
918 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
919 spin_unlock_irq(&rq
->lock
);
921 spin_unlock(&rq
->lock
);
925 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
929 * After ->oncpu is cleared, the task can be moved to a different CPU.
930 * We must ensure this doesn't happen until the switch is completely
936 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
940 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
943 * __task_rq_lock - lock the runqueue a given task resides on.
944 * Must be called interrupts disabled.
946 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
950 struct rq
*rq
= task_rq(p
);
951 spin_lock(&rq
->lock
);
952 if (likely(rq
== task_rq(p
)))
954 spin_unlock(&rq
->lock
);
959 * task_rq_lock - lock the runqueue a given task resides on and disable
960 * interrupts. Note the ordering: we can safely lookup the task_rq without
961 * explicitly disabling preemption.
963 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
969 local_irq_save(*flags
);
971 spin_lock(&rq
->lock
);
972 if (likely(rq
== task_rq(p
)))
974 spin_unlock_irqrestore(&rq
->lock
, *flags
);
978 void task_rq_unlock_wait(struct task_struct
*p
)
980 struct rq
*rq
= task_rq(p
);
982 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
983 spin_unlock_wait(&rq
->lock
);
986 static void __task_rq_unlock(struct rq
*rq
)
989 spin_unlock(&rq
->lock
);
992 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
995 spin_unlock_irqrestore(&rq
->lock
, *flags
);
999 * this_rq_lock - lock this runqueue and disable interrupts.
1001 static struct rq
*this_rq_lock(void)
1002 __acquires(rq
->lock
)
1006 local_irq_disable();
1008 spin_lock(&rq
->lock
);
1013 #ifdef CONFIG_SCHED_HRTICK
1015 * Use HR-timers to deliver accurate preemption points.
1017 * Its all a bit involved since we cannot program an hrt while holding the
1018 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1021 * When we get rescheduled we reprogram the hrtick_timer outside of the
1027 * - enabled by features
1028 * - hrtimer is actually high res
1030 static inline int hrtick_enabled(struct rq
*rq
)
1032 if (!sched_feat(HRTICK
))
1034 if (!cpu_active(cpu_of(rq
)))
1036 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1039 static void hrtick_clear(struct rq
*rq
)
1041 if (hrtimer_active(&rq
->hrtick_timer
))
1042 hrtimer_cancel(&rq
->hrtick_timer
);
1046 * High-resolution timer tick.
1047 * Runs from hardirq context with interrupts disabled.
1049 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1051 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1053 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1055 spin_lock(&rq
->lock
);
1056 update_rq_clock(rq
);
1057 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1058 spin_unlock(&rq
->lock
);
1060 return HRTIMER_NORESTART
;
1065 * called from hardirq (IPI) context
1067 static void __hrtick_start(void *arg
)
1069 struct rq
*rq
= arg
;
1071 spin_lock(&rq
->lock
);
1072 hrtimer_restart(&rq
->hrtick_timer
);
1073 rq
->hrtick_csd_pending
= 0;
1074 spin_unlock(&rq
->lock
);
1078 * Called to set the hrtick timer state.
1080 * called with rq->lock held and irqs disabled
1082 static void hrtick_start(struct rq
*rq
, u64 delay
)
1084 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1085 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1087 hrtimer_set_expires(timer
, time
);
1089 if (rq
== this_rq()) {
1090 hrtimer_restart(timer
);
1091 } else if (!rq
->hrtick_csd_pending
) {
1092 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1093 rq
->hrtick_csd_pending
= 1;
1098 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1100 int cpu
= (int)(long)hcpu
;
1103 case CPU_UP_CANCELED
:
1104 case CPU_UP_CANCELED_FROZEN
:
1105 case CPU_DOWN_PREPARE
:
1106 case CPU_DOWN_PREPARE_FROZEN
:
1108 case CPU_DEAD_FROZEN
:
1109 hrtick_clear(cpu_rq(cpu
));
1116 static __init
void init_hrtick(void)
1118 hotcpu_notifier(hotplug_hrtick
, 0);
1122 * Called to set the hrtick timer state.
1124 * called with rq->lock held and irqs disabled
1126 static void hrtick_start(struct rq
*rq
, u64 delay
)
1128 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1131 static inline void init_hrtick(void)
1134 #endif /* CONFIG_SMP */
1136 static void init_rq_hrtick(struct rq
*rq
)
1139 rq
->hrtick_csd_pending
= 0;
1141 rq
->hrtick_csd
.flags
= 0;
1142 rq
->hrtick_csd
.func
= __hrtick_start
;
1143 rq
->hrtick_csd
.info
= rq
;
1146 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1147 rq
->hrtick_timer
.function
= hrtick
;
1148 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1150 #else /* CONFIG_SCHED_HRTICK */
1151 static inline void hrtick_clear(struct rq
*rq
)
1155 static inline void init_rq_hrtick(struct rq
*rq
)
1159 static inline void init_hrtick(void)
1162 #endif /* CONFIG_SCHED_HRTICK */
1165 * resched_task - mark a task 'to be rescheduled now'.
1167 * On UP this means the setting of the need_resched flag, on SMP it
1168 * might also involve a cross-CPU call to trigger the scheduler on
1173 #ifndef tsk_is_polling
1174 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1177 static void resched_task(struct task_struct
*p
)
1181 assert_spin_locked(&task_rq(p
)->lock
);
1183 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1186 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1189 if (cpu
== smp_processor_id())
1192 /* NEED_RESCHED must be visible before we test polling */
1194 if (!tsk_is_polling(p
))
1195 smp_send_reschedule(cpu
);
1198 static void resched_cpu(int cpu
)
1200 struct rq
*rq
= cpu_rq(cpu
);
1201 unsigned long flags
;
1203 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1205 resched_task(cpu_curr(cpu
));
1206 spin_unlock_irqrestore(&rq
->lock
, flags
);
1211 * When add_timer_on() enqueues a timer into the timer wheel of an
1212 * idle CPU then this timer might expire before the next timer event
1213 * which is scheduled to wake up that CPU. In case of a completely
1214 * idle system the next event might even be infinite time into the
1215 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1216 * leaves the inner idle loop so the newly added timer is taken into
1217 * account when the CPU goes back to idle and evaluates the timer
1218 * wheel for the next timer event.
1220 void wake_up_idle_cpu(int cpu
)
1222 struct rq
*rq
= cpu_rq(cpu
);
1224 if (cpu
== smp_processor_id())
1228 * This is safe, as this function is called with the timer
1229 * wheel base lock of (cpu) held. When the CPU is on the way
1230 * to idle and has not yet set rq->curr to idle then it will
1231 * be serialized on the timer wheel base lock and take the new
1232 * timer into account automatically.
1234 if (rq
->curr
!= rq
->idle
)
1238 * We can set TIF_RESCHED on the idle task of the other CPU
1239 * lockless. The worst case is that the other CPU runs the
1240 * idle task through an additional NOOP schedule()
1242 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1244 /* NEED_RESCHED must be visible before we test polling */
1246 if (!tsk_is_polling(rq
->idle
))
1247 smp_send_reschedule(cpu
);
1249 #endif /* CONFIG_NO_HZ */
1251 #else /* !CONFIG_SMP */
1252 static void resched_task(struct task_struct
*p
)
1254 assert_spin_locked(&task_rq(p
)->lock
);
1255 set_tsk_need_resched(p
);
1257 #endif /* CONFIG_SMP */
1259 #if BITS_PER_LONG == 32
1260 # define WMULT_CONST (~0UL)
1262 # define WMULT_CONST (1UL << 32)
1265 #define WMULT_SHIFT 32
1268 * Shift right and round:
1270 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1273 * delta *= weight / lw
1275 static unsigned long
1276 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1277 struct load_weight
*lw
)
1281 if (!lw
->inv_weight
) {
1282 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1285 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1289 tmp
= (u64
)delta_exec
* weight
;
1291 * Check whether we'd overflow the 64-bit multiplication:
1293 if (unlikely(tmp
> WMULT_CONST
))
1294 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1297 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1299 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1302 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1308 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1315 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1316 * of tasks with abnormal "nice" values across CPUs the contribution that
1317 * each task makes to its run queue's load is weighted according to its
1318 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1319 * scaled version of the new time slice allocation that they receive on time
1323 #define WEIGHT_IDLEPRIO 2
1324 #define WMULT_IDLEPRIO (1 << 31)
1327 * Nice levels are multiplicative, with a gentle 10% change for every
1328 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1329 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1330 * that remained on nice 0.
1332 * The "10% effect" is relative and cumulative: from _any_ nice level,
1333 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1334 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1335 * If a task goes up by ~10% and another task goes down by ~10% then
1336 * the relative distance between them is ~25%.)
1338 static const int prio_to_weight
[40] = {
1339 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1340 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1341 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1342 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1343 /* 0 */ 1024, 820, 655, 526, 423,
1344 /* 5 */ 335, 272, 215, 172, 137,
1345 /* 10 */ 110, 87, 70, 56, 45,
1346 /* 15 */ 36, 29, 23, 18, 15,
1350 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1352 * In cases where the weight does not change often, we can use the
1353 * precalculated inverse to speed up arithmetics by turning divisions
1354 * into multiplications:
1356 static const u32 prio_to_wmult
[40] = {
1357 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1358 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1359 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1360 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1361 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1362 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1363 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1364 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1367 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1370 * runqueue iterator, to support SMP load-balancing between different
1371 * scheduling classes, without having to expose their internal data
1372 * structures to the load-balancing proper:
1374 struct rq_iterator
{
1376 struct task_struct
*(*start
)(void *);
1377 struct task_struct
*(*next
)(void *);
1381 static unsigned long
1382 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1383 unsigned long max_load_move
, struct sched_domain
*sd
,
1384 enum cpu_idle_type idle
, int *all_pinned
,
1385 int *this_best_prio
, struct rq_iterator
*iterator
);
1388 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1389 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1390 struct rq_iterator
*iterator
);
1393 #ifdef CONFIG_CGROUP_CPUACCT
1394 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1396 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1399 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1401 update_load_add(&rq
->load
, load
);
1404 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1406 update_load_sub(&rq
->load
, load
);
1409 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1410 typedef int (*tg_visitor
)(struct task_group
*, void *);
1413 * Iterate the full tree, calling @down when first entering a node and @up when
1414 * leaving it for the final time.
1416 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1418 struct task_group
*parent
, *child
;
1422 parent
= &root_task_group
;
1424 ret
= (*down
)(parent
, data
);
1427 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1434 ret
= (*up
)(parent
, data
);
1439 parent
= parent
->parent
;
1448 static int tg_nop(struct task_group
*tg
, void *data
)
1455 static unsigned long source_load(int cpu
, int type
);
1456 static unsigned long target_load(int cpu
, int type
);
1457 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1459 static unsigned long cpu_avg_load_per_task(int cpu
)
1461 struct rq
*rq
= cpu_rq(cpu
);
1464 rq
->avg_load_per_task
= rq
->load
.weight
/ rq
->nr_running
;
1466 return rq
->avg_load_per_task
;
1469 #ifdef CONFIG_FAIR_GROUP_SCHED
1471 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1474 * Calculate and set the cpu's group shares.
1477 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1478 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1481 unsigned long shares
;
1482 unsigned long rq_weight
;
1487 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1490 * If there are currently no tasks on the cpu pretend there is one of
1491 * average load so that when a new task gets to run here it will not
1492 * get delayed by group starvation.
1496 rq_weight
= NICE_0_LOAD
;
1499 if (unlikely(rq_weight
> sd_rq_weight
))
1500 rq_weight
= sd_rq_weight
;
1503 * \Sum shares * rq_weight
1504 * shares = -----------------------
1508 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1509 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1511 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1512 sysctl_sched_shares_thresh
) {
1513 struct rq
*rq
= cpu_rq(cpu
);
1514 unsigned long flags
;
1516 spin_lock_irqsave(&rq
->lock
, flags
);
1518 * record the actual number of shares, not the boosted amount.
1520 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1521 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1523 __set_se_shares(tg
->se
[cpu
], shares
);
1524 spin_unlock_irqrestore(&rq
->lock
, flags
);
1529 * Re-compute the task group their per cpu shares over the given domain.
1530 * This needs to be done in a bottom-up fashion because the rq weight of a
1531 * parent group depends on the shares of its child groups.
1533 static int tg_shares_up(struct task_group
*tg
, void *data
)
1535 unsigned long rq_weight
= 0;
1536 unsigned long shares
= 0;
1537 struct sched_domain
*sd
= data
;
1540 for_each_cpu_mask(i
, sd
->span
) {
1541 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1542 shares
+= tg
->cfs_rq
[i
]->shares
;
1545 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1546 shares
= tg
->shares
;
1548 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1549 shares
= tg
->shares
;
1552 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1554 for_each_cpu_mask(i
, sd
->span
)
1555 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1561 * Compute the cpu's hierarchical load factor for each task group.
1562 * This needs to be done in a top-down fashion because the load of a child
1563 * group is a fraction of its parents load.
1565 static int tg_load_down(struct task_group
*tg
, void *data
)
1568 long cpu
= (long)data
;
1571 load
= cpu_rq(cpu
)->load
.weight
;
1573 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1574 load
*= tg
->cfs_rq
[cpu
]->shares
;
1575 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1578 tg
->cfs_rq
[cpu
]->h_load
= load
;
1583 static void update_shares(struct sched_domain
*sd
)
1585 u64 now
= cpu_clock(raw_smp_processor_id());
1586 s64 elapsed
= now
- sd
->last_update
;
1588 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1589 sd
->last_update
= now
;
1590 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1594 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1596 spin_unlock(&rq
->lock
);
1598 spin_lock(&rq
->lock
);
1601 static void update_h_load(long cpu
)
1603 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1608 static inline void update_shares(struct sched_domain
*sd
)
1612 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1620 #ifdef CONFIG_FAIR_GROUP_SCHED
1621 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1624 cfs_rq
->shares
= shares
;
1629 #include "sched_stats.h"
1630 #include "sched_idletask.c"
1631 #include "sched_fair.c"
1632 #include "sched_rt.c"
1633 #ifdef CONFIG_SCHED_DEBUG
1634 # include "sched_debug.c"
1637 #define sched_class_highest (&rt_sched_class)
1638 #define for_each_class(class) \
1639 for (class = sched_class_highest; class; class = class->next)
1641 static void inc_nr_running(struct rq
*rq
)
1646 static void dec_nr_running(struct rq
*rq
)
1651 static void set_load_weight(struct task_struct
*p
)
1653 if (task_has_rt_policy(p
)) {
1654 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1655 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1660 * SCHED_IDLE tasks get minimal weight:
1662 if (p
->policy
== SCHED_IDLE
) {
1663 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1664 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1668 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1669 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1672 static void update_avg(u64
*avg
, u64 sample
)
1674 s64 diff
= sample
- *avg
;
1678 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1680 sched_info_queued(p
);
1681 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1685 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1687 if (sleep
&& p
->se
.last_wakeup
) {
1688 update_avg(&p
->se
.avg_overlap
,
1689 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1690 p
->se
.last_wakeup
= 0;
1693 sched_info_dequeued(p
);
1694 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1699 * __normal_prio - return the priority that is based on the static prio
1701 static inline int __normal_prio(struct task_struct
*p
)
1703 return p
->static_prio
;
1707 * Calculate the expected normal priority: i.e. priority
1708 * without taking RT-inheritance into account. Might be
1709 * boosted by interactivity modifiers. Changes upon fork,
1710 * setprio syscalls, and whenever the interactivity
1711 * estimator recalculates.
1713 static inline int normal_prio(struct task_struct
*p
)
1717 if (task_has_rt_policy(p
))
1718 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1720 prio
= __normal_prio(p
);
1725 * Calculate the current priority, i.e. the priority
1726 * taken into account by the scheduler. This value might
1727 * be boosted by RT tasks, or might be boosted by
1728 * interactivity modifiers. Will be RT if the task got
1729 * RT-boosted. If not then it returns p->normal_prio.
1731 static int effective_prio(struct task_struct
*p
)
1733 p
->normal_prio
= normal_prio(p
);
1735 * If we are RT tasks or we were boosted to RT priority,
1736 * keep the priority unchanged. Otherwise, update priority
1737 * to the normal priority:
1739 if (!rt_prio(p
->prio
))
1740 return p
->normal_prio
;
1745 * activate_task - move a task to the runqueue.
1747 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1749 if (task_contributes_to_load(p
))
1750 rq
->nr_uninterruptible
--;
1752 enqueue_task(rq
, p
, wakeup
);
1757 * deactivate_task - remove a task from the runqueue.
1759 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1761 if (task_contributes_to_load(p
))
1762 rq
->nr_uninterruptible
++;
1764 dequeue_task(rq
, p
, sleep
);
1769 * task_curr - is this task currently executing on a CPU?
1770 * @p: the task in question.
1772 inline int task_curr(const struct task_struct
*p
)
1774 return cpu_curr(task_cpu(p
)) == p
;
1777 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1779 set_task_rq(p
, cpu
);
1782 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1783 * successfuly executed on another CPU. We must ensure that updates of
1784 * per-task data have been completed by this moment.
1787 task_thread_info(p
)->cpu
= cpu
;
1791 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1792 const struct sched_class
*prev_class
,
1793 int oldprio
, int running
)
1795 if (prev_class
!= p
->sched_class
) {
1796 if (prev_class
->switched_from
)
1797 prev_class
->switched_from(rq
, p
, running
);
1798 p
->sched_class
->switched_to(rq
, p
, running
);
1800 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1805 /* Used instead of source_load when we know the type == 0 */
1806 static unsigned long weighted_cpuload(const int cpu
)
1808 return cpu_rq(cpu
)->load
.weight
;
1812 * Is this task likely cache-hot:
1815 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1820 * Buddy candidates are cache hot:
1822 if (sched_feat(CACHE_HOT_BUDDY
) &&
1823 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1824 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1827 if (p
->sched_class
!= &fair_sched_class
)
1830 if (sysctl_sched_migration_cost
== -1)
1832 if (sysctl_sched_migration_cost
== 0)
1835 delta
= now
- p
->se
.exec_start
;
1837 return delta
< (s64
)sysctl_sched_migration_cost
;
1841 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1843 int old_cpu
= task_cpu(p
);
1844 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1845 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1846 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1849 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1851 #ifdef CONFIG_SCHEDSTATS
1852 if (p
->se
.wait_start
)
1853 p
->se
.wait_start
-= clock_offset
;
1854 if (p
->se
.sleep_start
)
1855 p
->se
.sleep_start
-= clock_offset
;
1856 if (p
->se
.block_start
)
1857 p
->se
.block_start
-= clock_offset
;
1858 if (old_cpu
!= new_cpu
) {
1859 schedstat_inc(p
, se
.nr_migrations
);
1860 if (task_hot(p
, old_rq
->clock
, NULL
))
1861 schedstat_inc(p
, se
.nr_forced2_migrations
);
1864 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1865 new_cfsrq
->min_vruntime
;
1867 __set_task_cpu(p
, new_cpu
);
1870 struct migration_req
{
1871 struct list_head list
;
1873 struct task_struct
*task
;
1876 struct completion done
;
1880 * The task's runqueue lock must be held.
1881 * Returns true if you have to wait for migration thread.
1884 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1886 struct rq
*rq
= task_rq(p
);
1889 * If the task is not on a runqueue (and not running), then
1890 * it is sufficient to simply update the task's cpu field.
1892 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1893 set_task_cpu(p
, dest_cpu
);
1897 init_completion(&req
->done
);
1899 req
->dest_cpu
= dest_cpu
;
1900 list_add(&req
->list
, &rq
->migration_queue
);
1906 * wait_task_inactive - wait for a thread to unschedule.
1908 * If @match_state is nonzero, it's the @p->state value just checked and
1909 * not expected to change. If it changes, i.e. @p might have woken up,
1910 * then return zero. When we succeed in waiting for @p to be off its CPU,
1911 * we return a positive number (its total switch count). If a second call
1912 * a short while later returns the same number, the caller can be sure that
1913 * @p has remained unscheduled the whole time.
1915 * The caller must ensure that the task *will* unschedule sometime soon,
1916 * else this function might spin for a *long* time. This function can't
1917 * be called with interrupts off, or it may introduce deadlock with
1918 * smp_call_function() if an IPI is sent by the same process we are
1919 * waiting to become inactive.
1921 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1923 unsigned long flags
;
1930 * We do the initial early heuristics without holding
1931 * any task-queue locks at all. We'll only try to get
1932 * the runqueue lock when things look like they will
1938 * If the task is actively running on another CPU
1939 * still, just relax and busy-wait without holding
1942 * NOTE! Since we don't hold any locks, it's not
1943 * even sure that "rq" stays as the right runqueue!
1944 * But we don't care, since "task_running()" will
1945 * return false if the runqueue has changed and p
1946 * is actually now running somewhere else!
1948 while (task_running(rq
, p
)) {
1949 if (match_state
&& unlikely(p
->state
!= match_state
))
1955 * Ok, time to look more closely! We need the rq
1956 * lock now, to be *sure*. If we're wrong, we'll
1957 * just go back and repeat.
1959 rq
= task_rq_lock(p
, &flags
);
1960 trace_sched_wait_task(rq
, p
);
1961 running
= task_running(rq
, p
);
1962 on_rq
= p
->se
.on_rq
;
1964 if (!match_state
|| p
->state
== match_state
)
1965 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1966 task_rq_unlock(rq
, &flags
);
1969 * If it changed from the expected state, bail out now.
1971 if (unlikely(!ncsw
))
1975 * Was it really running after all now that we
1976 * checked with the proper locks actually held?
1978 * Oops. Go back and try again..
1980 if (unlikely(running
)) {
1986 * It's not enough that it's not actively running,
1987 * it must be off the runqueue _entirely_, and not
1990 * So if it wa still runnable (but just not actively
1991 * running right now), it's preempted, and we should
1992 * yield - it could be a while.
1994 if (unlikely(on_rq
)) {
1995 schedule_timeout_uninterruptible(1);
2000 * Ahh, all good. It wasn't running, and it wasn't
2001 * runnable, which means that it will never become
2002 * running in the future either. We're all done!
2011 * kick_process - kick a running thread to enter/exit the kernel
2012 * @p: the to-be-kicked thread
2014 * Cause a process which is running on another CPU to enter
2015 * kernel-mode, without any delay. (to get signals handled.)
2017 * NOTE: this function doesnt have to take the runqueue lock,
2018 * because all it wants to ensure is that the remote task enters
2019 * the kernel. If the IPI races and the task has been migrated
2020 * to another CPU then no harm is done and the purpose has been
2023 void kick_process(struct task_struct
*p
)
2029 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2030 smp_send_reschedule(cpu
);
2035 * Return a low guess at the load of a migration-source cpu weighted
2036 * according to the scheduling class and "nice" value.
2038 * We want to under-estimate the load of migration sources, to
2039 * balance conservatively.
2041 static unsigned long source_load(int cpu
, int type
)
2043 struct rq
*rq
= cpu_rq(cpu
);
2044 unsigned long total
= weighted_cpuload(cpu
);
2046 if (type
== 0 || !sched_feat(LB_BIAS
))
2049 return min(rq
->cpu_load
[type
-1], total
);
2053 * Return a high guess at the load of a migration-target cpu weighted
2054 * according to the scheduling class and "nice" value.
2056 static unsigned long target_load(int cpu
, int type
)
2058 struct rq
*rq
= cpu_rq(cpu
);
2059 unsigned long total
= weighted_cpuload(cpu
);
2061 if (type
== 0 || !sched_feat(LB_BIAS
))
2064 return max(rq
->cpu_load
[type
-1], total
);
2068 * find_idlest_group finds and returns the least busy CPU group within the
2071 static struct sched_group
*
2072 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2074 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2075 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2076 int load_idx
= sd
->forkexec_idx
;
2077 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2080 unsigned long load
, avg_load
;
2084 /* Skip over this group if it has no CPUs allowed */
2085 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2088 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2090 /* Tally up the load of all CPUs in the group */
2093 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2094 /* Bias balancing toward cpus of our domain */
2096 load
= source_load(i
, load_idx
);
2098 load
= target_load(i
, load_idx
);
2103 /* Adjust by relative CPU power of the group */
2104 avg_load
= sg_div_cpu_power(group
,
2105 avg_load
* SCHED_LOAD_SCALE
);
2108 this_load
= avg_load
;
2110 } else if (avg_load
< min_load
) {
2111 min_load
= avg_load
;
2114 } while (group
= group
->next
, group
!= sd
->groups
);
2116 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2122 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2125 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2128 unsigned long load
, min_load
= ULONG_MAX
;
2132 /* Traverse only the allowed CPUs */
2133 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2135 for_each_cpu_mask_nr(i
, *tmp
) {
2136 load
= weighted_cpuload(i
);
2138 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2148 * sched_balance_self: balance the current task (running on cpu) in domains
2149 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2152 * Balance, ie. select the least loaded group.
2154 * Returns the target CPU number, or the same CPU if no balancing is needed.
2156 * preempt must be disabled.
2158 static int sched_balance_self(int cpu
, int flag
)
2160 struct task_struct
*t
= current
;
2161 struct sched_domain
*tmp
, *sd
= NULL
;
2163 for_each_domain(cpu
, tmp
) {
2165 * If power savings logic is enabled for a domain, stop there.
2167 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2169 if (tmp
->flags
& flag
)
2177 cpumask_t span
, tmpmask
;
2178 struct sched_group
*group
;
2179 int new_cpu
, weight
;
2181 if (!(sd
->flags
& flag
)) {
2187 group
= find_idlest_group(sd
, t
, cpu
);
2193 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2194 if (new_cpu
== -1 || new_cpu
== cpu
) {
2195 /* Now try balancing at a lower domain level of cpu */
2200 /* Now try balancing at a lower domain level of new_cpu */
2203 weight
= cpus_weight(span
);
2204 for_each_domain(cpu
, tmp
) {
2205 if (weight
<= cpus_weight(tmp
->span
))
2207 if (tmp
->flags
& flag
)
2210 /* while loop will break here if sd == NULL */
2216 #endif /* CONFIG_SMP */
2219 * try_to_wake_up - wake up a thread
2220 * @p: the to-be-woken-up thread
2221 * @state: the mask of task states that can be woken
2222 * @sync: do a synchronous wakeup?
2224 * Put it on the run-queue if it's not already there. The "current"
2225 * thread is always on the run-queue (except when the actual
2226 * re-schedule is in progress), and as such you're allowed to do
2227 * the simpler "current->state = TASK_RUNNING" to mark yourself
2228 * runnable without the overhead of this.
2230 * returns failure only if the task is already active.
2232 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2234 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2235 unsigned long flags
;
2239 if (!sched_feat(SYNC_WAKEUPS
))
2243 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2244 struct sched_domain
*sd
;
2246 this_cpu
= raw_smp_processor_id();
2249 for_each_domain(this_cpu
, sd
) {
2250 if (cpu_isset(cpu
, sd
->span
)) {
2259 rq
= task_rq_lock(p
, &flags
);
2260 old_state
= p
->state
;
2261 if (!(old_state
& state
))
2269 this_cpu
= smp_processor_id();
2272 if (unlikely(task_running(rq
, p
)))
2275 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2276 if (cpu
!= orig_cpu
) {
2277 set_task_cpu(p
, cpu
);
2278 task_rq_unlock(rq
, &flags
);
2279 /* might preempt at this point */
2280 rq
= task_rq_lock(p
, &flags
);
2281 old_state
= p
->state
;
2282 if (!(old_state
& state
))
2287 this_cpu
= smp_processor_id();
2291 #ifdef CONFIG_SCHEDSTATS
2292 schedstat_inc(rq
, ttwu_count
);
2293 if (cpu
== this_cpu
)
2294 schedstat_inc(rq
, ttwu_local
);
2296 struct sched_domain
*sd
;
2297 for_each_domain(this_cpu
, sd
) {
2298 if (cpu_isset(cpu
, sd
->span
)) {
2299 schedstat_inc(sd
, ttwu_wake_remote
);
2304 #endif /* CONFIG_SCHEDSTATS */
2307 #endif /* CONFIG_SMP */
2308 schedstat_inc(p
, se
.nr_wakeups
);
2310 schedstat_inc(p
, se
.nr_wakeups_sync
);
2311 if (orig_cpu
!= cpu
)
2312 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2313 if (cpu
== this_cpu
)
2314 schedstat_inc(p
, se
.nr_wakeups_local
);
2316 schedstat_inc(p
, se
.nr_wakeups_remote
);
2317 update_rq_clock(rq
);
2318 activate_task(rq
, p
, 1);
2322 trace_sched_wakeup(rq
, p
);
2323 check_preempt_curr(rq
, p
, sync
);
2325 p
->state
= TASK_RUNNING
;
2327 if (p
->sched_class
->task_wake_up
)
2328 p
->sched_class
->task_wake_up(rq
, p
);
2331 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2333 task_rq_unlock(rq
, &flags
);
2338 int wake_up_process(struct task_struct
*p
)
2340 return try_to_wake_up(p
, TASK_ALL
, 0);
2342 EXPORT_SYMBOL(wake_up_process
);
2344 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2346 return try_to_wake_up(p
, state
, 0);
2350 * Perform scheduler related setup for a newly forked process p.
2351 * p is forked by current.
2353 * __sched_fork() is basic setup used by init_idle() too:
2355 static void __sched_fork(struct task_struct
*p
)
2357 p
->se
.exec_start
= 0;
2358 p
->se
.sum_exec_runtime
= 0;
2359 p
->se
.prev_sum_exec_runtime
= 0;
2360 p
->se
.last_wakeup
= 0;
2361 p
->se
.avg_overlap
= 0;
2363 #ifdef CONFIG_SCHEDSTATS
2364 p
->se
.wait_start
= 0;
2365 p
->se
.sum_sleep_runtime
= 0;
2366 p
->se
.sleep_start
= 0;
2367 p
->se
.block_start
= 0;
2368 p
->se
.sleep_max
= 0;
2369 p
->se
.block_max
= 0;
2371 p
->se
.slice_max
= 0;
2375 INIT_LIST_HEAD(&p
->rt
.run_list
);
2377 INIT_LIST_HEAD(&p
->se
.group_node
);
2379 #ifdef CONFIG_PREEMPT_NOTIFIERS
2380 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2384 * We mark the process as running here, but have not actually
2385 * inserted it onto the runqueue yet. This guarantees that
2386 * nobody will actually run it, and a signal or other external
2387 * event cannot wake it up and insert it on the runqueue either.
2389 p
->state
= TASK_RUNNING
;
2393 * fork()/clone()-time setup:
2395 void sched_fork(struct task_struct
*p
, int clone_flags
)
2397 int cpu
= get_cpu();
2402 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2404 set_task_cpu(p
, cpu
);
2407 * Make sure we do not leak PI boosting priority to the child:
2409 p
->prio
= current
->normal_prio
;
2410 if (!rt_prio(p
->prio
))
2411 p
->sched_class
= &fair_sched_class
;
2413 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2414 if (likely(sched_info_on()))
2415 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2417 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2420 #ifdef CONFIG_PREEMPT
2421 /* Want to start with kernel preemption disabled. */
2422 task_thread_info(p
)->preempt_count
= 1;
2428 * wake_up_new_task - wake up a newly created task for the first time.
2430 * This function will do some initial scheduler statistics housekeeping
2431 * that must be done for every newly created context, then puts the task
2432 * on the runqueue and wakes it.
2434 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2436 unsigned long flags
;
2439 rq
= task_rq_lock(p
, &flags
);
2440 BUG_ON(p
->state
!= TASK_RUNNING
);
2441 update_rq_clock(rq
);
2443 p
->prio
= effective_prio(p
);
2445 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2446 activate_task(rq
, p
, 0);
2449 * Let the scheduling class do new task startup
2450 * management (if any):
2452 p
->sched_class
->task_new(rq
, p
);
2455 trace_sched_wakeup_new(rq
, p
);
2456 check_preempt_curr(rq
, p
, 0);
2458 if (p
->sched_class
->task_wake_up
)
2459 p
->sched_class
->task_wake_up(rq
, p
);
2461 task_rq_unlock(rq
, &flags
);
2464 #ifdef CONFIG_PREEMPT_NOTIFIERS
2467 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2468 * @notifier: notifier struct to register
2470 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2472 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2474 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2477 * preempt_notifier_unregister - no longer interested in preemption notifications
2478 * @notifier: notifier struct to unregister
2480 * This is safe to call from within a preemption notifier.
2482 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2484 hlist_del(¬ifier
->link
);
2486 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2488 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2490 struct preempt_notifier
*notifier
;
2491 struct hlist_node
*node
;
2493 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2494 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2498 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2499 struct task_struct
*next
)
2501 struct preempt_notifier
*notifier
;
2502 struct hlist_node
*node
;
2504 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2505 notifier
->ops
->sched_out(notifier
, next
);
2508 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2510 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2515 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2516 struct task_struct
*next
)
2520 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2523 * prepare_task_switch - prepare to switch tasks
2524 * @rq: the runqueue preparing to switch
2525 * @prev: the current task that is being switched out
2526 * @next: the task we are going to switch to.
2528 * This is called with the rq lock held and interrupts off. It must
2529 * be paired with a subsequent finish_task_switch after the context
2532 * prepare_task_switch sets up locking and calls architecture specific
2536 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2537 struct task_struct
*next
)
2539 fire_sched_out_preempt_notifiers(prev
, next
);
2540 prepare_lock_switch(rq
, next
);
2541 prepare_arch_switch(next
);
2545 * finish_task_switch - clean up after a task-switch
2546 * @rq: runqueue associated with task-switch
2547 * @prev: the thread we just switched away from.
2549 * finish_task_switch must be called after the context switch, paired
2550 * with a prepare_task_switch call before the context switch.
2551 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2552 * and do any other architecture-specific cleanup actions.
2554 * Note that we may have delayed dropping an mm in context_switch(). If
2555 * so, we finish that here outside of the runqueue lock. (Doing it
2556 * with the lock held can cause deadlocks; see schedule() for
2559 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2560 __releases(rq
->lock
)
2562 struct mm_struct
*mm
= rq
->prev_mm
;
2568 * A task struct has one reference for the use as "current".
2569 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2570 * schedule one last time. The schedule call will never return, and
2571 * the scheduled task must drop that reference.
2572 * The test for TASK_DEAD must occur while the runqueue locks are
2573 * still held, otherwise prev could be scheduled on another cpu, die
2574 * there before we look at prev->state, and then the reference would
2576 * Manfred Spraul <manfred@colorfullife.com>
2578 prev_state
= prev
->state
;
2579 finish_arch_switch(prev
);
2580 finish_lock_switch(rq
, prev
);
2582 if (current
->sched_class
->post_schedule
)
2583 current
->sched_class
->post_schedule(rq
);
2586 fire_sched_in_preempt_notifiers(current
);
2589 if (unlikely(prev_state
== TASK_DEAD
)) {
2591 * Remove function-return probe instances associated with this
2592 * task and put them back on the free list.
2594 kprobe_flush_task(prev
);
2595 put_task_struct(prev
);
2600 * schedule_tail - first thing a freshly forked thread must call.
2601 * @prev: the thread we just switched away from.
2603 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2604 __releases(rq
->lock
)
2606 struct rq
*rq
= this_rq();
2608 finish_task_switch(rq
, prev
);
2609 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2610 /* In this case, finish_task_switch does not reenable preemption */
2613 if (current
->set_child_tid
)
2614 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2618 * context_switch - switch to the new MM and the new
2619 * thread's register state.
2622 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2623 struct task_struct
*next
)
2625 struct mm_struct
*mm
, *oldmm
;
2627 prepare_task_switch(rq
, prev
, next
);
2628 trace_sched_switch(rq
, prev
, next
);
2630 oldmm
= prev
->active_mm
;
2632 * For paravirt, this is coupled with an exit in switch_to to
2633 * combine the page table reload and the switch backend into
2636 arch_enter_lazy_cpu_mode();
2638 if (unlikely(!mm
)) {
2639 next
->active_mm
= oldmm
;
2640 atomic_inc(&oldmm
->mm_count
);
2641 enter_lazy_tlb(oldmm
, next
);
2643 switch_mm(oldmm
, mm
, next
);
2645 if (unlikely(!prev
->mm
)) {
2646 prev
->active_mm
= NULL
;
2647 rq
->prev_mm
= oldmm
;
2650 * Since the runqueue lock will be released by the next
2651 * task (which is an invalid locking op but in the case
2652 * of the scheduler it's an obvious special-case), so we
2653 * do an early lockdep release here:
2655 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2656 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2659 /* Here we just switch the register state and the stack. */
2660 switch_to(prev
, next
, prev
);
2664 * this_rq must be evaluated again because prev may have moved
2665 * CPUs since it called schedule(), thus the 'rq' on its stack
2666 * frame will be invalid.
2668 finish_task_switch(this_rq(), prev
);
2672 * nr_running, nr_uninterruptible and nr_context_switches:
2674 * externally visible scheduler statistics: current number of runnable
2675 * threads, current number of uninterruptible-sleeping threads, total
2676 * number of context switches performed since bootup.
2678 unsigned long nr_running(void)
2680 unsigned long i
, sum
= 0;
2682 for_each_online_cpu(i
)
2683 sum
+= cpu_rq(i
)->nr_running
;
2688 unsigned long nr_uninterruptible(void)
2690 unsigned long i
, sum
= 0;
2692 for_each_possible_cpu(i
)
2693 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2696 * Since we read the counters lockless, it might be slightly
2697 * inaccurate. Do not allow it to go below zero though:
2699 if (unlikely((long)sum
< 0))
2705 unsigned long long nr_context_switches(void)
2708 unsigned long long sum
= 0;
2710 for_each_possible_cpu(i
)
2711 sum
+= cpu_rq(i
)->nr_switches
;
2716 unsigned long nr_iowait(void)
2718 unsigned long i
, sum
= 0;
2720 for_each_possible_cpu(i
)
2721 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2726 unsigned long nr_active(void)
2728 unsigned long i
, running
= 0, uninterruptible
= 0;
2730 for_each_online_cpu(i
) {
2731 running
+= cpu_rq(i
)->nr_running
;
2732 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2735 if (unlikely((long)uninterruptible
< 0))
2736 uninterruptible
= 0;
2738 return running
+ uninterruptible
;
2742 * Update rq->cpu_load[] statistics. This function is usually called every
2743 * scheduler tick (TICK_NSEC).
2745 static void update_cpu_load(struct rq
*this_rq
)
2747 unsigned long this_load
= this_rq
->load
.weight
;
2750 this_rq
->nr_load_updates
++;
2752 /* Update our load: */
2753 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2754 unsigned long old_load
, new_load
;
2756 /* scale is effectively 1 << i now, and >> i divides by scale */
2758 old_load
= this_rq
->cpu_load
[i
];
2759 new_load
= this_load
;
2761 * Round up the averaging division if load is increasing. This
2762 * prevents us from getting stuck on 9 if the load is 10, for
2765 if (new_load
> old_load
)
2766 new_load
+= scale
-1;
2767 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2774 * double_rq_lock - safely lock two runqueues
2776 * Note this does not disable interrupts like task_rq_lock,
2777 * you need to do so manually before calling.
2779 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2780 __acquires(rq1
->lock
)
2781 __acquires(rq2
->lock
)
2783 BUG_ON(!irqs_disabled());
2785 spin_lock(&rq1
->lock
);
2786 __acquire(rq2
->lock
); /* Fake it out ;) */
2789 spin_lock(&rq1
->lock
);
2790 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2792 spin_lock(&rq2
->lock
);
2793 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2796 update_rq_clock(rq1
);
2797 update_rq_clock(rq2
);
2801 * double_rq_unlock - safely unlock two runqueues
2803 * Note this does not restore interrupts like task_rq_unlock,
2804 * you need to do so manually after calling.
2806 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2807 __releases(rq1
->lock
)
2808 __releases(rq2
->lock
)
2810 spin_unlock(&rq1
->lock
);
2812 spin_unlock(&rq2
->lock
);
2814 __release(rq2
->lock
);
2818 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2820 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2821 __releases(this_rq
->lock
)
2822 __acquires(busiest
->lock
)
2823 __acquires(this_rq
->lock
)
2827 if (unlikely(!irqs_disabled())) {
2828 /* printk() doesn't work good under rq->lock */
2829 spin_unlock(&this_rq
->lock
);
2832 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2833 if (busiest
< this_rq
) {
2834 spin_unlock(&this_rq
->lock
);
2835 spin_lock(&busiest
->lock
);
2836 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2839 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2844 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2845 __releases(busiest
->lock
)
2847 spin_unlock(&busiest
->lock
);
2848 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2852 * If dest_cpu is allowed for this process, migrate the task to it.
2853 * This is accomplished by forcing the cpu_allowed mask to only
2854 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2855 * the cpu_allowed mask is restored.
2857 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2859 struct migration_req req
;
2860 unsigned long flags
;
2863 rq
= task_rq_lock(p
, &flags
);
2864 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2865 || unlikely(!cpu_active(dest_cpu
)))
2868 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2869 /* force the process onto the specified CPU */
2870 if (migrate_task(p
, dest_cpu
, &req
)) {
2871 /* Need to wait for migration thread (might exit: take ref). */
2872 struct task_struct
*mt
= rq
->migration_thread
;
2874 get_task_struct(mt
);
2875 task_rq_unlock(rq
, &flags
);
2876 wake_up_process(mt
);
2877 put_task_struct(mt
);
2878 wait_for_completion(&req
.done
);
2883 task_rq_unlock(rq
, &flags
);
2887 * sched_exec - execve() is a valuable balancing opportunity, because at
2888 * this point the task has the smallest effective memory and cache footprint.
2890 void sched_exec(void)
2892 int new_cpu
, this_cpu
= get_cpu();
2893 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2895 if (new_cpu
!= this_cpu
)
2896 sched_migrate_task(current
, new_cpu
);
2900 * pull_task - move a task from a remote runqueue to the local runqueue.
2901 * Both runqueues must be locked.
2903 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2904 struct rq
*this_rq
, int this_cpu
)
2906 deactivate_task(src_rq
, p
, 0);
2907 set_task_cpu(p
, this_cpu
);
2908 activate_task(this_rq
, p
, 0);
2910 * Note that idle threads have a prio of MAX_PRIO, for this test
2911 * to be always true for them.
2913 check_preempt_curr(this_rq
, p
, 0);
2917 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2920 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2921 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2925 * We do not migrate tasks that are:
2926 * 1) running (obviously), or
2927 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2928 * 3) are cache-hot on their current CPU.
2930 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2931 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2936 if (task_running(rq
, p
)) {
2937 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2942 * Aggressive migration if:
2943 * 1) task is cache cold, or
2944 * 2) too many balance attempts have failed.
2947 if (!task_hot(p
, rq
->clock
, sd
) ||
2948 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2949 #ifdef CONFIG_SCHEDSTATS
2950 if (task_hot(p
, rq
->clock
, sd
)) {
2951 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2952 schedstat_inc(p
, se
.nr_forced_migrations
);
2958 if (task_hot(p
, rq
->clock
, sd
)) {
2959 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2965 static unsigned long
2966 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2967 unsigned long max_load_move
, struct sched_domain
*sd
,
2968 enum cpu_idle_type idle
, int *all_pinned
,
2969 int *this_best_prio
, struct rq_iterator
*iterator
)
2971 int loops
= 0, pulled
= 0, pinned
= 0;
2972 struct task_struct
*p
;
2973 long rem_load_move
= max_load_move
;
2975 if (max_load_move
== 0)
2981 * Start the load-balancing iterator:
2983 p
= iterator
->start(iterator
->arg
);
2985 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2988 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2989 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2990 p
= iterator
->next(iterator
->arg
);
2994 pull_task(busiest
, p
, this_rq
, this_cpu
);
2996 rem_load_move
-= p
->se
.load
.weight
;
2999 * We only want to steal up to the prescribed amount of weighted load.
3001 if (rem_load_move
> 0) {
3002 if (p
->prio
< *this_best_prio
)
3003 *this_best_prio
= p
->prio
;
3004 p
= iterator
->next(iterator
->arg
);
3009 * Right now, this is one of only two places pull_task() is called,
3010 * so we can safely collect pull_task() stats here rather than
3011 * inside pull_task().
3013 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3016 *all_pinned
= pinned
;
3018 return max_load_move
- rem_load_move
;
3022 * move_tasks tries to move up to max_load_move weighted load from busiest to
3023 * this_rq, as part of a balancing operation within domain "sd".
3024 * Returns 1 if successful and 0 otherwise.
3026 * Called with both runqueues locked.
3028 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3029 unsigned long max_load_move
,
3030 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3033 const struct sched_class
*class = sched_class_highest
;
3034 unsigned long total_load_moved
= 0;
3035 int this_best_prio
= this_rq
->curr
->prio
;
3039 class->load_balance(this_rq
, this_cpu
, busiest
,
3040 max_load_move
- total_load_moved
,
3041 sd
, idle
, all_pinned
, &this_best_prio
);
3042 class = class->next
;
3044 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3047 } while (class && max_load_move
> total_load_moved
);
3049 return total_load_moved
> 0;
3053 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3054 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3055 struct rq_iterator
*iterator
)
3057 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3061 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3062 pull_task(busiest
, p
, this_rq
, this_cpu
);
3064 * Right now, this is only the second place pull_task()
3065 * is called, so we can safely collect pull_task()
3066 * stats here rather than inside pull_task().
3068 schedstat_inc(sd
, lb_gained
[idle
]);
3072 p
= iterator
->next(iterator
->arg
);
3079 * move_one_task tries to move exactly one task from busiest to this_rq, as
3080 * part of active balancing operations within "domain".
3081 * Returns 1 if successful and 0 otherwise.
3083 * Called with both runqueues locked.
3085 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3086 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3088 const struct sched_class
*class;
3090 for (class = sched_class_highest
; class; class = class->next
)
3091 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3098 * find_busiest_group finds and returns the busiest CPU group within the
3099 * domain. It calculates and returns the amount of weighted load which
3100 * should be moved to restore balance via the imbalance parameter.
3102 static struct sched_group
*
3103 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3104 unsigned long *imbalance
, enum cpu_idle_type idle
,
3105 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3107 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3108 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3109 unsigned long max_pull
;
3110 unsigned long busiest_load_per_task
, busiest_nr_running
;
3111 unsigned long this_load_per_task
, this_nr_running
;
3112 int load_idx
, group_imb
= 0;
3113 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3114 int power_savings_balance
= 1;
3115 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3116 unsigned long min_nr_running
= ULONG_MAX
;
3117 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3120 max_load
= this_load
= total_load
= total_pwr
= 0;
3121 busiest_load_per_task
= busiest_nr_running
= 0;
3122 this_load_per_task
= this_nr_running
= 0;
3124 if (idle
== CPU_NOT_IDLE
)
3125 load_idx
= sd
->busy_idx
;
3126 else if (idle
== CPU_NEWLY_IDLE
)
3127 load_idx
= sd
->newidle_idx
;
3129 load_idx
= sd
->idle_idx
;
3132 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3135 int __group_imb
= 0;
3136 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3137 unsigned long sum_nr_running
, sum_weighted_load
;
3138 unsigned long sum_avg_load_per_task
;
3139 unsigned long avg_load_per_task
;
3141 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3144 balance_cpu
= first_cpu(group
->cpumask
);
3146 /* Tally up the load of all CPUs in the group */
3147 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3148 sum_avg_load_per_task
= avg_load_per_task
= 0;
3151 min_cpu_load
= ~0UL;
3153 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3156 if (!cpu_isset(i
, *cpus
))
3161 if (*sd_idle
&& rq
->nr_running
)
3164 /* Bias balancing toward cpus of our domain */
3166 if (idle_cpu(i
) && !first_idle_cpu
) {
3171 load
= target_load(i
, load_idx
);
3173 load
= source_load(i
, load_idx
);
3174 if (load
> max_cpu_load
)
3175 max_cpu_load
= load
;
3176 if (min_cpu_load
> load
)
3177 min_cpu_load
= load
;
3181 sum_nr_running
+= rq
->nr_running
;
3182 sum_weighted_load
+= weighted_cpuload(i
);
3184 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3188 * First idle cpu or the first cpu(busiest) in this sched group
3189 * is eligible for doing load balancing at this and above
3190 * domains. In the newly idle case, we will allow all the cpu's
3191 * to do the newly idle load balance.
3193 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3194 balance_cpu
!= this_cpu
&& balance
) {
3199 total_load
+= avg_load
;
3200 total_pwr
+= group
->__cpu_power
;
3202 /* Adjust by relative CPU power of the group */
3203 avg_load
= sg_div_cpu_power(group
,
3204 avg_load
* SCHED_LOAD_SCALE
);
3208 * Consider the group unbalanced when the imbalance is larger
3209 * than the average weight of two tasks.
3211 * APZ: with cgroup the avg task weight can vary wildly and
3212 * might not be a suitable number - should we keep a
3213 * normalized nr_running number somewhere that negates
3216 avg_load_per_task
= sg_div_cpu_power(group
,
3217 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3219 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3222 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3225 this_load
= avg_load
;
3227 this_nr_running
= sum_nr_running
;
3228 this_load_per_task
= sum_weighted_load
;
3229 } else if (avg_load
> max_load
&&
3230 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3231 max_load
= avg_load
;
3233 busiest_nr_running
= sum_nr_running
;
3234 busiest_load_per_task
= sum_weighted_load
;
3235 group_imb
= __group_imb
;
3238 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3240 * Busy processors will not participate in power savings
3243 if (idle
== CPU_NOT_IDLE
||
3244 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3248 * If the local group is idle or completely loaded
3249 * no need to do power savings balance at this domain
3251 if (local_group
&& (this_nr_running
>= group_capacity
||
3253 power_savings_balance
= 0;
3256 * If a group is already running at full capacity or idle,
3257 * don't include that group in power savings calculations
3259 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3264 * Calculate the group which has the least non-idle load.
3265 * This is the group from where we need to pick up the load
3268 if ((sum_nr_running
< min_nr_running
) ||
3269 (sum_nr_running
== min_nr_running
&&
3270 first_cpu(group
->cpumask
) <
3271 first_cpu(group_min
->cpumask
))) {
3273 min_nr_running
= sum_nr_running
;
3274 min_load_per_task
= sum_weighted_load
/
3279 * Calculate the group which is almost near its
3280 * capacity but still has some space to pick up some load
3281 * from other group and save more power
3283 if (sum_nr_running
<= group_capacity
- 1) {
3284 if (sum_nr_running
> leader_nr_running
||
3285 (sum_nr_running
== leader_nr_running
&&
3286 first_cpu(group
->cpumask
) >
3287 first_cpu(group_leader
->cpumask
))) {
3288 group_leader
= group
;
3289 leader_nr_running
= sum_nr_running
;
3294 group
= group
->next
;
3295 } while (group
!= sd
->groups
);
3297 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3300 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3302 if (this_load
>= avg_load
||
3303 100*max_load
<= sd
->imbalance_pct
*this_load
)
3306 busiest_load_per_task
/= busiest_nr_running
;
3308 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3311 * We're trying to get all the cpus to the average_load, so we don't
3312 * want to push ourselves above the average load, nor do we wish to
3313 * reduce the max loaded cpu below the average load, as either of these
3314 * actions would just result in more rebalancing later, and ping-pong
3315 * tasks around. Thus we look for the minimum possible imbalance.
3316 * Negative imbalances (*we* are more loaded than anyone else) will
3317 * be counted as no imbalance for these purposes -- we can't fix that
3318 * by pulling tasks to us. Be careful of negative numbers as they'll
3319 * appear as very large values with unsigned longs.
3321 if (max_load
<= busiest_load_per_task
)
3325 * In the presence of smp nice balancing, certain scenarios can have
3326 * max load less than avg load(as we skip the groups at or below
3327 * its cpu_power, while calculating max_load..)
3329 if (max_load
< avg_load
) {
3331 goto small_imbalance
;
3334 /* Don't want to pull so many tasks that a group would go idle */
3335 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3337 /* How much load to actually move to equalise the imbalance */
3338 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3339 (avg_load
- this_load
) * this->__cpu_power
)
3343 * if *imbalance is less than the average load per runnable task
3344 * there is no gaurantee that any tasks will be moved so we'll have
3345 * a think about bumping its value to force at least one task to be
3348 if (*imbalance
< busiest_load_per_task
) {
3349 unsigned long tmp
, pwr_now
, pwr_move
;
3353 pwr_move
= pwr_now
= 0;
3355 if (this_nr_running
) {
3356 this_load_per_task
/= this_nr_running
;
3357 if (busiest_load_per_task
> this_load_per_task
)
3360 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3362 if (max_load
- this_load
+ busiest_load_per_task
>=
3363 busiest_load_per_task
* imbn
) {
3364 *imbalance
= busiest_load_per_task
;
3369 * OK, we don't have enough imbalance to justify moving tasks,
3370 * however we may be able to increase total CPU power used by
3374 pwr_now
+= busiest
->__cpu_power
*
3375 min(busiest_load_per_task
, max_load
);
3376 pwr_now
+= this->__cpu_power
*
3377 min(this_load_per_task
, this_load
);
3378 pwr_now
/= SCHED_LOAD_SCALE
;
3380 /* Amount of load we'd subtract */
3381 tmp
= sg_div_cpu_power(busiest
,
3382 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3384 pwr_move
+= busiest
->__cpu_power
*
3385 min(busiest_load_per_task
, max_load
- tmp
);
3387 /* Amount of load we'd add */
3388 if (max_load
* busiest
->__cpu_power
<
3389 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3390 tmp
= sg_div_cpu_power(this,
3391 max_load
* busiest
->__cpu_power
);
3393 tmp
= sg_div_cpu_power(this,
3394 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3395 pwr_move
+= this->__cpu_power
*
3396 min(this_load_per_task
, this_load
+ tmp
);
3397 pwr_move
/= SCHED_LOAD_SCALE
;
3399 /* Move if we gain throughput */
3400 if (pwr_move
> pwr_now
)
3401 *imbalance
= busiest_load_per_task
;
3407 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3408 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3411 if (this == group_leader
&& group_leader
!= group_min
) {
3412 *imbalance
= min_load_per_task
;
3422 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3425 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3426 unsigned long imbalance
, const cpumask_t
*cpus
)
3428 struct rq
*busiest
= NULL
, *rq
;
3429 unsigned long max_load
= 0;
3432 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3435 if (!cpu_isset(i
, *cpus
))
3439 wl
= weighted_cpuload(i
);
3441 if (rq
->nr_running
== 1 && wl
> imbalance
)
3444 if (wl
> max_load
) {
3454 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3455 * so long as it is large enough.
3457 #define MAX_PINNED_INTERVAL 512
3460 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3461 * tasks if there is an imbalance.
3463 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3464 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3465 int *balance
, cpumask_t
*cpus
)
3467 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3468 struct sched_group
*group
;
3469 unsigned long imbalance
;
3471 unsigned long flags
;
3476 * When power savings policy is enabled for the parent domain, idle
3477 * sibling can pick up load irrespective of busy siblings. In this case,
3478 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3479 * portraying it as CPU_NOT_IDLE.
3481 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3482 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3485 schedstat_inc(sd
, lb_count
[idle
]);
3489 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3496 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3500 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3502 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3506 BUG_ON(busiest
== this_rq
);
3508 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3511 if (busiest
->nr_running
> 1) {
3513 * Attempt to move tasks. If find_busiest_group has found
3514 * an imbalance but busiest->nr_running <= 1, the group is
3515 * still unbalanced. ld_moved simply stays zero, so it is
3516 * correctly treated as an imbalance.
3518 local_irq_save(flags
);
3519 double_rq_lock(this_rq
, busiest
);
3520 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3521 imbalance
, sd
, idle
, &all_pinned
);
3522 double_rq_unlock(this_rq
, busiest
);
3523 local_irq_restore(flags
);
3526 * some other cpu did the load balance for us.
3528 if (ld_moved
&& this_cpu
!= smp_processor_id())
3529 resched_cpu(this_cpu
);
3531 /* All tasks on this runqueue were pinned by CPU affinity */
3532 if (unlikely(all_pinned
)) {
3533 cpu_clear(cpu_of(busiest
), *cpus
);
3534 if (!cpus_empty(*cpus
))
3541 schedstat_inc(sd
, lb_failed
[idle
]);
3542 sd
->nr_balance_failed
++;
3544 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3546 spin_lock_irqsave(&busiest
->lock
, flags
);
3548 /* don't kick the migration_thread, if the curr
3549 * task on busiest cpu can't be moved to this_cpu
3551 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3552 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3554 goto out_one_pinned
;
3557 if (!busiest
->active_balance
) {
3558 busiest
->active_balance
= 1;
3559 busiest
->push_cpu
= this_cpu
;
3562 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3564 wake_up_process(busiest
->migration_thread
);
3567 * We've kicked active balancing, reset the failure
3570 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3573 sd
->nr_balance_failed
= 0;
3575 if (likely(!active_balance
)) {
3576 /* We were unbalanced, so reset the balancing interval */
3577 sd
->balance_interval
= sd
->min_interval
;
3580 * If we've begun active balancing, start to back off. This
3581 * case may not be covered by the all_pinned logic if there
3582 * is only 1 task on the busy runqueue (because we don't call
3585 if (sd
->balance_interval
< sd
->max_interval
)
3586 sd
->balance_interval
*= 2;
3589 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3590 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3596 schedstat_inc(sd
, lb_balanced
[idle
]);
3598 sd
->nr_balance_failed
= 0;
3601 /* tune up the balancing interval */
3602 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3603 (sd
->balance_interval
< sd
->max_interval
))
3604 sd
->balance_interval
*= 2;
3606 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3607 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3618 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3619 * tasks if there is an imbalance.
3621 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3622 * this_rq is locked.
3625 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3628 struct sched_group
*group
;
3629 struct rq
*busiest
= NULL
;
3630 unsigned long imbalance
;
3638 * When power savings policy is enabled for the parent domain, idle
3639 * sibling can pick up load irrespective of busy siblings. In this case,
3640 * let the state of idle sibling percolate up as IDLE, instead of
3641 * portraying it as CPU_NOT_IDLE.
3643 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3644 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3647 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3649 update_shares_locked(this_rq
, sd
);
3650 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3651 &sd_idle
, cpus
, NULL
);
3653 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3657 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3659 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3663 BUG_ON(busiest
== this_rq
);
3665 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3668 if (busiest
->nr_running
> 1) {
3669 /* Attempt to move tasks */
3670 double_lock_balance(this_rq
, busiest
);
3671 /* this_rq->clock is already updated */
3672 update_rq_clock(busiest
);
3673 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3674 imbalance
, sd
, CPU_NEWLY_IDLE
,
3676 double_unlock_balance(this_rq
, busiest
);
3678 if (unlikely(all_pinned
)) {
3679 cpu_clear(cpu_of(busiest
), *cpus
);
3680 if (!cpus_empty(*cpus
))
3686 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3687 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3688 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3691 sd
->nr_balance_failed
= 0;
3693 update_shares_locked(this_rq
, sd
);
3697 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3698 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3699 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3701 sd
->nr_balance_failed
= 0;
3707 * idle_balance is called by schedule() if this_cpu is about to become
3708 * idle. Attempts to pull tasks from other CPUs.
3710 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3712 struct sched_domain
*sd
;
3713 int pulled_task
= -1;
3714 unsigned long next_balance
= jiffies
+ HZ
;
3717 for_each_domain(this_cpu
, sd
) {
3718 unsigned long interval
;
3720 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3723 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3724 /* If we've pulled tasks over stop searching: */
3725 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3728 interval
= msecs_to_jiffies(sd
->balance_interval
);
3729 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3730 next_balance
= sd
->last_balance
+ interval
;
3734 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3736 * We are going idle. next_balance may be set based on
3737 * a busy processor. So reset next_balance.
3739 this_rq
->next_balance
= next_balance
;
3744 * active_load_balance is run by migration threads. It pushes running tasks
3745 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3746 * running on each physical CPU where possible, and avoids physical /
3747 * logical imbalances.
3749 * Called with busiest_rq locked.
3751 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3753 int target_cpu
= busiest_rq
->push_cpu
;
3754 struct sched_domain
*sd
;
3755 struct rq
*target_rq
;
3757 /* Is there any task to move? */
3758 if (busiest_rq
->nr_running
<= 1)
3761 target_rq
= cpu_rq(target_cpu
);
3764 * This condition is "impossible", if it occurs
3765 * we need to fix it. Originally reported by
3766 * Bjorn Helgaas on a 128-cpu setup.
3768 BUG_ON(busiest_rq
== target_rq
);
3770 /* move a task from busiest_rq to target_rq */
3771 double_lock_balance(busiest_rq
, target_rq
);
3772 update_rq_clock(busiest_rq
);
3773 update_rq_clock(target_rq
);
3775 /* Search for an sd spanning us and the target CPU. */
3776 for_each_domain(target_cpu
, sd
) {
3777 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3778 cpu_isset(busiest_cpu
, sd
->span
))
3783 schedstat_inc(sd
, alb_count
);
3785 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3787 schedstat_inc(sd
, alb_pushed
);
3789 schedstat_inc(sd
, alb_failed
);
3791 double_unlock_balance(busiest_rq
, target_rq
);
3796 atomic_t load_balancer
;
3798 } nohz ____cacheline_aligned
= {
3799 .load_balancer
= ATOMIC_INIT(-1),
3800 .cpu_mask
= CPU_MASK_NONE
,
3804 * This routine will try to nominate the ilb (idle load balancing)
3805 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3806 * load balancing on behalf of all those cpus. If all the cpus in the system
3807 * go into this tickless mode, then there will be no ilb owner (as there is
3808 * no need for one) and all the cpus will sleep till the next wakeup event
3811 * For the ilb owner, tick is not stopped. And this tick will be used
3812 * for idle load balancing. ilb owner will still be part of
3815 * While stopping the tick, this cpu will become the ilb owner if there
3816 * is no other owner. And will be the owner till that cpu becomes busy
3817 * or if all cpus in the system stop their ticks at which point
3818 * there is no need for ilb owner.
3820 * When the ilb owner becomes busy, it nominates another owner, during the
3821 * next busy scheduler_tick()
3823 int select_nohz_load_balancer(int stop_tick
)
3825 int cpu
= smp_processor_id();
3828 cpu_set(cpu
, nohz
.cpu_mask
);
3829 cpu_rq(cpu
)->in_nohz_recently
= 1;
3832 * If we are going offline and still the leader, give up!
3834 if (!cpu_active(cpu
) &&
3835 atomic_read(&nohz
.load_balancer
) == cpu
) {
3836 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3841 /* time for ilb owner also to sleep */
3842 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3843 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3844 atomic_set(&nohz
.load_balancer
, -1);
3848 if (atomic_read(&nohz
.load_balancer
) == -1) {
3849 /* make me the ilb owner */
3850 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3852 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3855 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3858 cpu_clear(cpu
, nohz
.cpu_mask
);
3860 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3861 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3868 static DEFINE_SPINLOCK(balancing
);
3871 * It checks each scheduling domain to see if it is due to be balanced,
3872 * and initiates a balancing operation if so.
3874 * Balancing parameters are set up in arch_init_sched_domains.
3876 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3879 struct rq
*rq
= cpu_rq(cpu
);
3880 unsigned long interval
;
3881 struct sched_domain
*sd
;
3882 /* Earliest time when we have to do rebalance again */
3883 unsigned long next_balance
= jiffies
+ 60*HZ
;
3884 int update_next_balance
= 0;
3888 for_each_domain(cpu
, sd
) {
3889 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3892 interval
= sd
->balance_interval
;
3893 if (idle
!= CPU_IDLE
)
3894 interval
*= sd
->busy_factor
;
3896 /* scale ms to jiffies */
3897 interval
= msecs_to_jiffies(interval
);
3898 if (unlikely(!interval
))
3900 if (interval
> HZ
*NR_CPUS
/10)
3901 interval
= HZ
*NR_CPUS
/10;
3903 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3905 if (need_serialize
) {
3906 if (!spin_trylock(&balancing
))
3910 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3911 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3913 * We've pulled tasks over so either we're no
3914 * longer idle, or one of our SMT siblings is
3917 idle
= CPU_NOT_IDLE
;
3919 sd
->last_balance
= jiffies
;
3922 spin_unlock(&balancing
);
3924 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3925 next_balance
= sd
->last_balance
+ interval
;
3926 update_next_balance
= 1;
3930 * Stop the load balance at this level. There is another
3931 * CPU in our sched group which is doing load balancing more
3939 * next_balance will be updated only when there is a need.
3940 * When the cpu is attached to null domain for ex, it will not be
3943 if (likely(update_next_balance
))
3944 rq
->next_balance
= next_balance
;
3948 * run_rebalance_domains is triggered when needed from the scheduler tick.
3949 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3950 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3952 static void run_rebalance_domains(struct softirq_action
*h
)
3954 int this_cpu
= smp_processor_id();
3955 struct rq
*this_rq
= cpu_rq(this_cpu
);
3956 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3957 CPU_IDLE
: CPU_NOT_IDLE
;
3959 rebalance_domains(this_cpu
, idle
);
3963 * If this cpu is the owner for idle load balancing, then do the
3964 * balancing on behalf of the other idle cpus whose ticks are
3967 if (this_rq
->idle_at_tick
&&
3968 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3969 cpumask_t cpus
= nohz
.cpu_mask
;
3973 cpu_clear(this_cpu
, cpus
);
3974 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3976 * If this cpu gets work to do, stop the load balancing
3977 * work being done for other cpus. Next load
3978 * balancing owner will pick it up.
3983 rebalance_domains(balance_cpu
, CPU_IDLE
);
3985 rq
= cpu_rq(balance_cpu
);
3986 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3987 this_rq
->next_balance
= rq
->next_balance
;
3994 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3996 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3997 * idle load balancing owner or decide to stop the periodic load balancing,
3998 * if the whole system is idle.
4000 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4004 * If we were in the nohz mode recently and busy at the current
4005 * scheduler tick, then check if we need to nominate new idle
4008 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4009 rq
->in_nohz_recently
= 0;
4011 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4012 cpu_clear(cpu
, nohz
.cpu_mask
);
4013 atomic_set(&nohz
.load_balancer
, -1);
4016 if (atomic_read(&nohz
.load_balancer
) == -1) {
4018 * simple selection for now: Nominate the
4019 * first cpu in the nohz list to be the next
4022 * TBD: Traverse the sched domains and nominate
4023 * the nearest cpu in the nohz.cpu_mask.
4025 int ilb
= first_cpu(nohz
.cpu_mask
);
4027 if (ilb
< nr_cpu_ids
)
4033 * If this cpu is idle and doing idle load balancing for all the
4034 * cpus with ticks stopped, is it time for that to stop?
4036 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4037 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4043 * If this cpu is idle and the idle load balancing is done by
4044 * someone else, then no need raise the SCHED_SOFTIRQ
4046 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4047 cpu_isset(cpu
, nohz
.cpu_mask
))
4050 if (time_after_eq(jiffies
, rq
->next_balance
))
4051 raise_softirq(SCHED_SOFTIRQ
);
4054 #else /* CONFIG_SMP */
4057 * on UP we do not need to balance between CPUs:
4059 static inline void idle_balance(int cpu
, struct rq
*rq
)
4065 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4067 EXPORT_PER_CPU_SYMBOL(kstat
);
4070 * Return any ns on the sched_clock that have not yet been banked in
4071 * @p in case that task is currently running.
4073 unsigned long long task_delta_exec(struct task_struct
*p
)
4075 unsigned long flags
;
4079 rq
= task_rq_lock(p
, &flags
);
4081 if (task_current(rq
, p
)) {
4084 update_rq_clock(rq
);
4085 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4086 if ((s64
)delta_exec
> 0)
4090 task_rq_unlock(rq
, &flags
);
4096 * Account user cpu time to a process.
4097 * @p: the process that the cpu time gets accounted to
4098 * @cputime: the cpu time spent in user space since the last update
4100 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4102 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4105 p
->utime
= cputime_add(p
->utime
, cputime
);
4106 account_group_user_time(p
, cputime
);
4108 /* Add user time to cpustat. */
4109 tmp
= cputime_to_cputime64(cputime
);
4110 if (TASK_NICE(p
) > 0)
4111 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4113 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4114 /* Account for user time used */
4115 acct_update_integrals(p
);
4119 * Account guest cpu time to a process.
4120 * @p: the process that the cpu time gets accounted to
4121 * @cputime: the cpu time spent in virtual machine since the last update
4123 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4126 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4128 tmp
= cputime_to_cputime64(cputime
);
4130 p
->utime
= cputime_add(p
->utime
, cputime
);
4131 account_group_user_time(p
, cputime
);
4132 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4134 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4135 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4139 * Account scaled user cpu time to a process.
4140 * @p: the process that the cpu time gets accounted to
4141 * @cputime: the cpu time spent in user space since the last update
4143 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4145 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4149 * Account system cpu time to a process.
4150 * @p: the process that the cpu time gets accounted to
4151 * @hardirq_offset: the offset to subtract from hardirq_count()
4152 * @cputime: the cpu time spent in kernel space since the last update
4154 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4157 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4158 struct rq
*rq
= this_rq();
4161 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4162 account_guest_time(p
, cputime
);
4166 p
->stime
= cputime_add(p
->stime
, cputime
);
4167 account_group_system_time(p
, cputime
);
4169 /* Add system time to cpustat. */
4170 tmp
= cputime_to_cputime64(cputime
);
4171 if (hardirq_count() - hardirq_offset
)
4172 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4173 else if (softirq_count())
4174 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4175 else if (p
!= rq
->idle
)
4176 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4177 else if (atomic_read(&rq
->nr_iowait
) > 0)
4178 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4180 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4181 /* Account for system time used */
4182 acct_update_integrals(p
);
4186 * Account scaled system cpu time to a process.
4187 * @p: the process that the cpu time gets accounted to
4188 * @hardirq_offset: the offset to subtract from hardirq_count()
4189 * @cputime: the cpu time spent in kernel space since the last update
4191 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4193 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4197 * Account for involuntary wait time.
4198 * @p: the process from which the cpu time has been stolen
4199 * @steal: the cpu time spent in involuntary wait
4201 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4203 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4204 cputime64_t tmp
= cputime_to_cputime64(steal
);
4205 struct rq
*rq
= this_rq();
4207 if (p
== rq
->idle
) {
4208 p
->stime
= cputime_add(p
->stime
, steal
);
4209 account_group_system_time(p
, steal
);
4210 if (atomic_read(&rq
->nr_iowait
) > 0)
4211 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4213 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4215 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4219 * Use precise platform statistics if available:
4221 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4222 cputime_t
task_utime(struct task_struct
*p
)
4227 cputime_t
task_stime(struct task_struct
*p
)
4232 cputime_t
task_utime(struct task_struct
*p
)
4234 clock_t utime
= cputime_to_clock_t(p
->utime
),
4235 total
= utime
+ cputime_to_clock_t(p
->stime
);
4239 * Use CFS's precise accounting:
4241 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4245 do_div(temp
, total
);
4247 utime
= (clock_t)temp
;
4249 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4250 return p
->prev_utime
;
4253 cputime_t
task_stime(struct task_struct
*p
)
4258 * Use CFS's precise accounting. (we subtract utime from
4259 * the total, to make sure the total observed by userspace
4260 * grows monotonically - apps rely on that):
4262 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4263 cputime_to_clock_t(task_utime(p
));
4266 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4268 return p
->prev_stime
;
4272 inline cputime_t
task_gtime(struct task_struct
*p
)
4278 * This function gets called by the timer code, with HZ frequency.
4279 * We call it with interrupts disabled.
4281 * It also gets called by the fork code, when changing the parent's
4284 void scheduler_tick(void)
4286 int cpu
= smp_processor_id();
4287 struct rq
*rq
= cpu_rq(cpu
);
4288 struct task_struct
*curr
= rq
->curr
;
4292 spin_lock(&rq
->lock
);
4293 update_rq_clock(rq
);
4294 update_cpu_load(rq
);
4295 curr
->sched_class
->task_tick(rq
, curr
, 0);
4296 spin_unlock(&rq
->lock
);
4299 rq
->idle_at_tick
= idle_cpu(cpu
);
4300 trigger_load_balance(rq
, cpu
);
4304 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4305 defined(CONFIG_PREEMPT_TRACER))
4307 static inline unsigned long get_parent_ip(unsigned long addr
)
4309 if (in_lock_functions(addr
)) {
4310 addr
= CALLER_ADDR2
;
4311 if (in_lock_functions(addr
))
4312 addr
= CALLER_ADDR3
;
4317 void __kprobes
add_preempt_count(int val
)
4319 #ifdef CONFIG_DEBUG_PREEMPT
4323 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4326 preempt_count() += val
;
4327 #ifdef CONFIG_DEBUG_PREEMPT
4329 * Spinlock count overflowing soon?
4331 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4334 if (preempt_count() == val
)
4335 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4337 EXPORT_SYMBOL(add_preempt_count
);
4339 void __kprobes
sub_preempt_count(int val
)
4341 #ifdef CONFIG_DEBUG_PREEMPT
4345 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4348 * Is the spinlock portion underflowing?
4350 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4351 !(preempt_count() & PREEMPT_MASK
)))
4355 if (preempt_count() == val
)
4356 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4357 preempt_count() -= val
;
4359 EXPORT_SYMBOL(sub_preempt_count
);
4364 * Print scheduling while atomic bug:
4366 static noinline
void __schedule_bug(struct task_struct
*prev
)
4368 struct pt_regs
*regs
= get_irq_regs();
4370 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4371 prev
->comm
, prev
->pid
, preempt_count());
4373 debug_show_held_locks(prev
);
4375 if (irqs_disabled())
4376 print_irqtrace_events(prev
);
4385 * Various schedule()-time debugging checks and statistics:
4387 static inline void schedule_debug(struct task_struct
*prev
)
4390 * Test if we are atomic. Since do_exit() needs to call into
4391 * schedule() atomically, we ignore that path for now.
4392 * Otherwise, whine if we are scheduling when we should not be.
4394 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4395 __schedule_bug(prev
);
4397 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4399 schedstat_inc(this_rq(), sched_count
);
4400 #ifdef CONFIG_SCHEDSTATS
4401 if (unlikely(prev
->lock_depth
>= 0)) {
4402 schedstat_inc(this_rq(), bkl_count
);
4403 schedstat_inc(prev
, sched_info
.bkl_count
);
4409 * Pick up the highest-prio task:
4411 static inline struct task_struct
*
4412 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4414 const struct sched_class
*class;
4415 struct task_struct
*p
;
4418 * Optimization: we know that if all tasks are in
4419 * the fair class we can call that function directly:
4421 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4422 p
= fair_sched_class
.pick_next_task(rq
);
4427 class = sched_class_highest
;
4429 p
= class->pick_next_task(rq
);
4433 * Will never be NULL as the idle class always
4434 * returns a non-NULL p:
4436 class = class->next
;
4441 * schedule() is the main scheduler function.
4443 asmlinkage
void __sched
schedule(void)
4445 struct task_struct
*prev
, *next
;
4446 unsigned long *switch_count
;
4452 cpu
= smp_processor_id();
4456 switch_count
= &prev
->nivcsw
;
4458 release_kernel_lock(prev
);
4459 need_resched_nonpreemptible
:
4461 schedule_debug(prev
);
4463 if (sched_feat(HRTICK
))
4466 spin_lock_irq(&rq
->lock
);
4467 update_rq_clock(rq
);
4468 clear_tsk_need_resched(prev
);
4470 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4471 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4472 prev
->state
= TASK_RUNNING
;
4474 deactivate_task(rq
, prev
, 1);
4475 switch_count
= &prev
->nvcsw
;
4479 if (prev
->sched_class
->pre_schedule
)
4480 prev
->sched_class
->pre_schedule(rq
, prev
);
4483 if (unlikely(!rq
->nr_running
))
4484 idle_balance(cpu
, rq
);
4486 prev
->sched_class
->put_prev_task(rq
, prev
);
4487 next
= pick_next_task(rq
, prev
);
4489 if (likely(prev
!= next
)) {
4490 sched_info_switch(prev
, next
);
4496 context_switch(rq
, prev
, next
); /* unlocks the rq */
4498 * the context switch might have flipped the stack from under
4499 * us, hence refresh the local variables.
4501 cpu
= smp_processor_id();
4504 spin_unlock_irq(&rq
->lock
);
4506 if (unlikely(reacquire_kernel_lock(current
) < 0))
4507 goto need_resched_nonpreemptible
;
4509 preempt_enable_no_resched();
4510 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4513 EXPORT_SYMBOL(schedule
);
4515 #ifdef CONFIG_PREEMPT
4517 * this is the entry point to schedule() from in-kernel preemption
4518 * off of preempt_enable. Kernel preemptions off return from interrupt
4519 * occur there and call schedule directly.
4521 asmlinkage
void __sched
preempt_schedule(void)
4523 struct thread_info
*ti
= current_thread_info();
4526 * If there is a non-zero preempt_count or interrupts are disabled,
4527 * we do not want to preempt the current task. Just return..
4529 if (likely(ti
->preempt_count
|| irqs_disabled()))
4533 add_preempt_count(PREEMPT_ACTIVE
);
4535 sub_preempt_count(PREEMPT_ACTIVE
);
4538 * Check again in case we missed a preemption opportunity
4539 * between schedule and now.
4542 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4544 EXPORT_SYMBOL(preempt_schedule
);
4547 * this is the entry point to schedule() from kernel preemption
4548 * off of irq context.
4549 * Note, that this is called and return with irqs disabled. This will
4550 * protect us against recursive calling from irq.
4552 asmlinkage
void __sched
preempt_schedule_irq(void)
4554 struct thread_info
*ti
= current_thread_info();
4556 /* Catch callers which need to be fixed */
4557 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4560 add_preempt_count(PREEMPT_ACTIVE
);
4563 local_irq_disable();
4564 sub_preempt_count(PREEMPT_ACTIVE
);
4567 * Check again in case we missed a preemption opportunity
4568 * between schedule and now.
4571 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4574 #endif /* CONFIG_PREEMPT */
4576 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4579 return try_to_wake_up(curr
->private, mode
, sync
);
4581 EXPORT_SYMBOL(default_wake_function
);
4584 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4585 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4586 * number) then we wake all the non-exclusive tasks and one exclusive task.
4588 * There are circumstances in which we can try to wake a task which has already
4589 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4590 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4592 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4593 int nr_exclusive
, int sync
, void *key
)
4595 wait_queue_t
*curr
, *next
;
4597 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4598 unsigned flags
= curr
->flags
;
4600 if (curr
->func(curr
, mode
, sync
, key
) &&
4601 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4607 * __wake_up - wake up threads blocked on a waitqueue.
4609 * @mode: which threads
4610 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4611 * @key: is directly passed to the wakeup function
4613 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4614 int nr_exclusive
, void *key
)
4616 unsigned long flags
;
4618 spin_lock_irqsave(&q
->lock
, flags
);
4619 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4620 spin_unlock_irqrestore(&q
->lock
, flags
);
4622 EXPORT_SYMBOL(__wake_up
);
4625 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4627 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4629 __wake_up_common(q
, mode
, 1, 0, NULL
);
4633 * __wake_up_sync - wake up threads blocked on a waitqueue.
4635 * @mode: which threads
4636 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4638 * The sync wakeup differs that the waker knows that it will schedule
4639 * away soon, so while the target thread will be woken up, it will not
4640 * be migrated to another CPU - ie. the two threads are 'synchronized'
4641 * with each other. This can prevent needless bouncing between CPUs.
4643 * On UP it can prevent extra preemption.
4646 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4648 unsigned long flags
;
4654 if (unlikely(!nr_exclusive
))
4657 spin_lock_irqsave(&q
->lock
, flags
);
4658 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4659 spin_unlock_irqrestore(&q
->lock
, flags
);
4661 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4664 * complete: - signals a single thread waiting on this completion
4665 * @x: holds the state of this particular completion
4667 * This will wake up a single thread waiting on this completion. Threads will be
4668 * awakened in the same order in which they were queued.
4670 * See also complete_all(), wait_for_completion() and related routines.
4672 void complete(struct completion
*x
)
4674 unsigned long flags
;
4676 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4678 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4679 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4681 EXPORT_SYMBOL(complete
);
4684 * complete_all: - signals all threads waiting on this completion
4685 * @x: holds the state of this particular completion
4687 * This will wake up all threads waiting on this particular completion event.
4689 void complete_all(struct completion
*x
)
4691 unsigned long flags
;
4693 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4694 x
->done
+= UINT_MAX
/2;
4695 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4696 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4698 EXPORT_SYMBOL(complete_all
);
4700 static inline long __sched
4701 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4704 DECLARE_WAITQUEUE(wait
, current
);
4706 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4707 __add_wait_queue_tail(&x
->wait
, &wait
);
4709 if (signal_pending_state(state
, current
)) {
4710 timeout
= -ERESTARTSYS
;
4713 __set_current_state(state
);
4714 spin_unlock_irq(&x
->wait
.lock
);
4715 timeout
= schedule_timeout(timeout
);
4716 spin_lock_irq(&x
->wait
.lock
);
4717 } while (!x
->done
&& timeout
);
4718 __remove_wait_queue(&x
->wait
, &wait
);
4723 return timeout
?: 1;
4727 wait_for_common(struct completion
*x
, long timeout
, int state
)
4731 spin_lock_irq(&x
->wait
.lock
);
4732 timeout
= do_wait_for_common(x
, timeout
, state
);
4733 spin_unlock_irq(&x
->wait
.lock
);
4738 * wait_for_completion: - waits for completion of a task
4739 * @x: holds the state of this particular completion
4741 * This waits to be signaled for completion of a specific task. It is NOT
4742 * interruptible and there is no timeout.
4744 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4745 * and interrupt capability. Also see complete().
4747 void __sched
wait_for_completion(struct completion
*x
)
4749 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4751 EXPORT_SYMBOL(wait_for_completion
);
4754 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4755 * @x: holds the state of this particular completion
4756 * @timeout: timeout value in jiffies
4758 * This waits for either a completion of a specific task to be signaled or for a
4759 * specified timeout to expire. The timeout is in jiffies. It is not
4762 unsigned long __sched
4763 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4765 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4767 EXPORT_SYMBOL(wait_for_completion_timeout
);
4770 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4771 * @x: holds the state of this particular completion
4773 * This waits for completion of a specific task to be signaled. It is
4776 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4778 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4779 if (t
== -ERESTARTSYS
)
4783 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4786 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4787 * @x: holds the state of this particular completion
4788 * @timeout: timeout value in jiffies
4790 * This waits for either a completion of a specific task to be signaled or for a
4791 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4793 unsigned long __sched
4794 wait_for_completion_interruptible_timeout(struct completion
*x
,
4795 unsigned long timeout
)
4797 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4799 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4802 * wait_for_completion_killable: - waits for completion of a task (killable)
4803 * @x: holds the state of this particular completion
4805 * This waits to be signaled for completion of a specific task. It can be
4806 * interrupted by a kill signal.
4808 int __sched
wait_for_completion_killable(struct completion
*x
)
4810 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4811 if (t
== -ERESTARTSYS
)
4815 EXPORT_SYMBOL(wait_for_completion_killable
);
4818 * try_wait_for_completion - try to decrement a completion without blocking
4819 * @x: completion structure
4821 * Returns: 0 if a decrement cannot be done without blocking
4822 * 1 if a decrement succeeded.
4824 * If a completion is being used as a counting completion,
4825 * attempt to decrement the counter without blocking. This
4826 * enables us to avoid waiting if the resource the completion
4827 * is protecting is not available.
4829 bool try_wait_for_completion(struct completion
*x
)
4833 spin_lock_irq(&x
->wait
.lock
);
4838 spin_unlock_irq(&x
->wait
.lock
);
4841 EXPORT_SYMBOL(try_wait_for_completion
);
4844 * completion_done - Test to see if a completion has any waiters
4845 * @x: completion structure
4847 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4848 * 1 if there are no waiters.
4851 bool completion_done(struct completion
*x
)
4855 spin_lock_irq(&x
->wait
.lock
);
4858 spin_unlock_irq(&x
->wait
.lock
);
4861 EXPORT_SYMBOL(completion_done
);
4864 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4866 unsigned long flags
;
4869 init_waitqueue_entry(&wait
, current
);
4871 __set_current_state(state
);
4873 spin_lock_irqsave(&q
->lock
, flags
);
4874 __add_wait_queue(q
, &wait
);
4875 spin_unlock(&q
->lock
);
4876 timeout
= schedule_timeout(timeout
);
4877 spin_lock_irq(&q
->lock
);
4878 __remove_wait_queue(q
, &wait
);
4879 spin_unlock_irqrestore(&q
->lock
, flags
);
4884 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4886 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4888 EXPORT_SYMBOL(interruptible_sleep_on
);
4891 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4893 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4895 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4897 void __sched
sleep_on(wait_queue_head_t
*q
)
4899 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4901 EXPORT_SYMBOL(sleep_on
);
4903 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4905 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4907 EXPORT_SYMBOL(sleep_on_timeout
);
4909 #ifdef CONFIG_RT_MUTEXES
4912 * rt_mutex_setprio - set the current priority of a task
4914 * @prio: prio value (kernel-internal form)
4916 * This function changes the 'effective' priority of a task. It does
4917 * not touch ->normal_prio like __setscheduler().
4919 * Used by the rt_mutex code to implement priority inheritance logic.
4921 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4923 unsigned long flags
;
4924 int oldprio
, on_rq
, running
;
4926 const struct sched_class
*prev_class
= p
->sched_class
;
4928 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4930 rq
= task_rq_lock(p
, &flags
);
4931 update_rq_clock(rq
);
4934 on_rq
= p
->se
.on_rq
;
4935 running
= task_current(rq
, p
);
4937 dequeue_task(rq
, p
, 0);
4939 p
->sched_class
->put_prev_task(rq
, p
);
4942 p
->sched_class
= &rt_sched_class
;
4944 p
->sched_class
= &fair_sched_class
;
4949 p
->sched_class
->set_curr_task(rq
);
4951 enqueue_task(rq
, p
, 0);
4953 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4955 task_rq_unlock(rq
, &flags
);
4960 void set_user_nice(struct task_struct
*p
, long nice
)
4962 int old_prio
, delta
, on_rq
;
4963 unsigned long flags
;
4966 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4969 * We have to be careful, if called from sys_setpriority(),
4970 * the task might be in the middle of scheduling on another CPU.
4972 rq
= task_rq_lock(p
, &flags
);
4973 update_rq_clock(rq
);
4975 * The RT priorities are set via sched_setscheduler(), but we still
4976 * allow the 'normal' nice value to be set - but as expected
4977 * it wont have any effect on scheduling until the task is
4978 * SCHED_FIFO/SCHED_RR:
4980 if (task_has_rt_policy(p
)) {
4981 p
->static_prio
= NICE_TO_PRIO(nice
);
4984 on_rq
= p
->se
.on_rq
;
4986 dequeue_task(rq
, p
, 0);
4988 p
->static_prio
= NICE_TO_PRIO(nice
);
4991 p
->prio
= effective_prio(p
);
4992 delta
= p
->prio
- old_prio
;
4995 enqueue_task(rq
, p
, 0);
4997 * If the task increased its priority or is running and
4998 * lowered its priority, then reschedule its CPU:
5000 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5001 resched_task(rq
->curr
);
5004 task_rq_unlock(rq
, &flags
);
5006 EXPORT_SYMBOL(set_user_nice
);
5009 * can_nice - check if a task can reduce its nice value
5013 int can_nice(const struct task_struct
*p
, const int nice
)
5015 /* convert nice value [19,-20] to rlimit style value [1,40] */
5016 int nice_rlim
= 20 - nice
;
5018 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5019 capable(CAP_SYS_NICE
));
5022 #ifdef __ARCH_WANT_SYS_NICE
5025 * sys_nice - change the priority of the current process.
5026 * @increment: priority increment
5028 * sys_setpriority is a more generic, but much slower function that
5029 * does similar things.
5031 asmlinkage
long sys_nice(int increment
)
5036 * Setpriority might change our priority at the same moment.
5037 * We don't have to worry. Conceptually one call occurs first
5038 * and we have a single winner.
5040 if (increment
< -40)
5045 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5051 if (increment
< 0 && !can_nice(current
, nice
))
5054 retval
= security_task_setnice(current
, nice
);
5058 set_user_nice(current
, nice
);
5065 * task_prio - return the priority value of a given task.
5066 * @p: the task in question.
5068 * This is the priority value as seen by users in /proc.
5069 * RT tasks are offset by -200. Normal tasks are centered
5070 * around 0, value goes from -16 to +15.
5072 int task_prio(const struct task_struct
*p
)
5074 return p
->prio
- MAX_RT_PRIO
;
5078 * task_nice - return the nice value of a given task.
5079 * @p: the task in question.
5081 int task_nice(const struct task_struct
*p
)
5083 return TASK_NICE(p
);
5085 EXPORT_SYMBOL(task_nice
);
5088 * idle_cpu - is a given cpu idle currently?
5089 * @cpu: the processor in question.
5091 int idle_cpu(int cpu
)
5093 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5097 * idle_task - return the idle task for a given cpu.
5098 * @cpu: the processor in question.
5100 struct task_struct
*idle_task(int cpu
)
5102 return cpu_rq(cpu
)->idle
;
5106 * find_process_by_pid - find a process with a matching PID value.
5107 * @pid: the pid in question.
5109 static struct task_struct
*find_process_by_pid(pid_t pid
)
5111 return pid
? find_task_by_vpid(pid
) : current
;
5114 /* Actually do priority change: must hold rq lock. */
5116 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5118 BUG_ON(p
->se
.on_rq
);
5121 switch (p
->policy
) {
5125 p
->sched_class
= &fair_sched_class
;
5129 p
->sched_class
= &rt_sched_class
;
5133 p
->rt_priority
= prio
;
5134 p
->normal_prio
= normal_prio(p
);
5135 /* we are holding p->pi_lock already */
5136 p
->prio
= rt_mutex_getprio(p
);
5140 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5141 struct sched_param
*param
, bool user
)
5143 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5144 unsigned long flags
;
5145 const struct sched_class
*prev_class
= p
->sched_class
;
5148 /* may grab non-irq protected spin_locks */
5149 BUG_ON(in_interrupt());
5151 /* double check policy once rq lock held */
5153 policy
= oldpolicy
= p
->policy
;
5154 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5155 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5156 policy
!= SCHED_IDLE
)
5159 * Valid priorities for SCHED_FIFO and SCHED_RR are
5160 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5161 * SCHED_BATCH and SCHED_IDLE is 0.
5163 if (param
->sched_priority
< 0 ||
5164 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5165 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5167 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5171 * Allow unprivileged RT tasks to decrease priority:
5173 if (user
&& !capable(CAP_SYS_NICE
)) {
5174 if (rt_policy(policy
)) {
5175 unsigned long rlim_rtprio
;
5177 if (!lock_task_sighand(p
, &flags
))
5179 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5180 unlock_task_sighand(p
, &flags
);
5182 /* can't set/change the rt policy */
5183 if (policy
!= p
->policy
&& !rlim_rtprio
)
5186 /* can't increase priority */
5187 if (param
->sched_priority
> p
->rt_priority
&&
5188 param
->sched_priority
> rlim_rtprio
)
5192 * Like positive nice levels, dont allow tasks to
5193 * move out of SCHED_IDLE either:
5195 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5198 /* can't change other user's priorities */
5199 if ((current
->euid
!= p
->euid
) &&
5200 (current
->euid
!= p
->uid
))
5205 #ifdef CONFIG_RT_GROUP_SCHED
5207 * Do not allow realtime tasks into groups that have no runtime
5210 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5211 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5215 retval
= security_task_setscheduler(p
, policy
, param
);
5221 * make sure no PI-waiters arrive (or leave) while we are
5222 * changing the priority of the task:
5224 spin_lock_irqsave(&p
->pi_lock
, flags
);
5226 * To be able to change p->policy safely, the apropriate
5227 * runqueue lock must be held.
5229 rq
= __task_rq_lock(p
);
5230 /* recheck policy now with rq lock held */
5231 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5232 policy
= oldpolicy
= -1;
5233 __task_rq_unlock(rq
);
5234 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5237 update_rq_clock(rq
);
5238 on_rq
= p
->se
.on_rq
;
5239 running
= task_current(rq
, p
);
5241 deactivate_task(rq
, p
, 0);
5243 p
->sched_class
->put_prev_task(rq
, p
);
5246 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5249 p
->sched_class
->set_curr_task(rq
);
5251 activate_task(rq
, p
, 0);
5253 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5255 __task_rq_unlock(rq
);
5256 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5258 rt_mutex_adjust_pi(p
);
5264 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5265 * @p: the task in question.
5266 * @policy: new policy.
5267 * @param: structure containing the new RT priority.
5269 * NOTE that the task may be already dead.
5271 int sched_setscheduler(struct task_struct
*p
, int policy
,
5272 struct sched_param
*param
)
5274 return __sched_setscheduler(p
, policy
, param
, true);
5276 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5279 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5280 * @p: the task in question.
5281 * @policy: new policy.
5282 * @param: structure containing the new RT priority.
5284 * Just like sched_setscheduler, only don't bother checking if the
5285 * current context has permission. For example, this is needed in
5286 * stop_machine(): we create temporary high priority worker threads,
5287 * but our caller might not have that capability.
5289 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5290 struct sched_param
*param
)
5292 return __sched_setscheduler(p
, policy
, param
, false);
5296 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5298 struct sched_param lparam
;
5299 struct task_struct
*p
;
5302 if (!param
|| pid
< 0)
5304 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5309 p
= find_process_by_pid(pid
);
5311 retval
= sched_setscheduler(p
, policy
, &lparam
);
5318 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5319 * @pid: the pid in question.
5320 * @policy: new policy.
5321 * @param: structure containing the new RT priority.
5324 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5326 /* negative values for policy are not valid */
5330 return do_sched_setscheduler(pid
, policy
, param
);
5334 * sys_sched_setparam - set/change the RT priority of a thread
5335 * @pid: the pid in question.
5336 * @param: structure containing the new RT priority.
5338 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5340 return do_sched_setscheduler(pid
, -1, param
);
5344 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5345 * @pid: the pid in question.
5347 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5349 struct task_struct
*p
;
5356 read_lock(&tasklist_lock
);
5357 p
= find_process_by_pid(pid
);
5359 retval
= security_task_getscheduler(p
);
5363 read_unlock(&tasklist_lock
);
5368 * sys_sched_getscheduler - get the RT priority of a thread
5369 * @pid: the pid in question.
5370 * @param: structure containing the RT priority.
5372 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5374 struct sched_param lp
;
5375 struct task_struct
*p
;
5378 if (!param
|| pid
< 0)
5381 read_lock(&tasklist_lock
);
5382 p
= find_process_by_pid(pid
);
5387 retval
= security_task_getscheduler(p
);
5391 lp
.sched_priority
= p
->rt_priority
;
5392 read_unlock(&tasklist_lock
);
5395 * This one might sleep, we cannot do it with a spinlock held ...
5397 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5402 read_unlock(&tasklist_lock
);
5406 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5408 cpumask_t cpus_allowed
;
5409 cpumask_t new_mask
= *in_mask
;
5410 struct task_struct
*p
;
5414 read_lock(&tasklist_lock
);
5416 p
= find_process_by_pid(pid
);
5418 read_unlock(&tasklist_lock
);
5424 * It is not safe to call set_cpus_allowed with the
5425 * tasklist_lock held. We will bump the task_struct's
5426 * usage count and then drop tasklist_lock.
5429 read_unlock(&tasklist_lock
);
5432 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5433 !capable(CAP_SYS_NICE
))
5436 retval
= security_task_setscheduler(p
, 0, NULL
);
5440 cpuset_cpus_allowed(p
, &cpus_allowed
);
5441 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5443 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5446 cpuset_cpus_allowed(p
, &cpus_allowed
);
5447 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5449 * We must have raced with a concurrent cpuset
5450 * update. Just reset the cpus_allowed to the
5451 * cpuset's cpus_allowed
5453 new_mask
= cpus_allowed
;
5463 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5464 cpumask_t
*new_mask
)
5466 if (len
< sizeof(cpumask_t
)) {
5467 memset(new_mask
, 0, sizeof(cpumask_t
));
5468 } else if (len
> sizeof(cpumask_t
)) {
5469 len
= sizeof(cpumask_t
);
5471 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5475 * sys_sched_setaffinity - set the cpu affinity of a process
5476 * @pid: pid of the process
5477 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5478 * @user_mask_ptr: user-space pointer to the new cpu mask
5480 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5481 unsigned long __user
*user_mask_ptr
)
5486 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5490 return sched_setaffinity(pid
, &new_mask
);
5493 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5495 struct task_struct
*p
;
5499 read_lock(&tasklist_lock
);
5502 p
= find_process_by_pid(pid
);
5506 retval
= security_task_getscheduler(p
);
5510 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5513 read_unlock(&tasklist_lock
);
5520 * sys_sched_getaffinity - get the cpu affinity of a process
5521 * @pid: pid of the process
5522 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5523 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5525 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5526 unsigned long __user
*user_mask_ptr
)
5531 if (len
< sizeof(cpumask_t
))
5534 ret
= sched_getaffinity(pid
, &mask
);
5538 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5541 return sizeof(cpumask_t
);
5545 * sys_sched_yield - yield the current processor to other threads.
5547 * This function yields the current CPU to other tasks. If there are no
5548 * other threads running on this CPU then this function will return.
5550 asmlinkage
long sys_sched_yield(void)
5552 struct rq
*rq
= this_rq_lock();
5554 schedstat_inc(rq
, yld_count
);
5555 current
->sched_class
->yield_task(rq
);
5558 * Since we are going to call schedule() anyway, there's
5559 * no need to preempt or enable interrupts:
5561 __release(rq
->lock
);
5562 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5563 _raw_spin_unlock(&rq
->lock
);
5564 preempt_enable_no_resched();
5571 static void __cond_resched(void)
5573 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5574 __might_sleep(__FILE__
, __LINE__
);
5577 * The BKS might be reacquired before we have dropped
5578 * PREEMPT_ACTIVE, which could trigger a second
5579 * cond_resched() call.
5582 add_preempt_count(PREEMPT_ACTIVE
);
5584 sub_preempt_count(PREEMPT_ACTIVE
);
5585 } while (need_resched());
5588 int __sched
_cond_resched(void)
5590 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5591 system_state
== SYSTEM_RUNNING
) {
5597 EXPORT_SYMBOL(_cond_resched
);
5600 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5601 * call schedule, and on return reacquire the lock.
5603 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5604 * operations here to prevent schedule() from being called twice (once via
5605 * spin_unlock(), once by hand).
5607 int cond_resched_lock(spinlock_t
*lock
)
5609 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5612 if (spin_needbreak(lock
) || resched
) {
5614 if (resched
&& need_resched())
5623 EXPORT_SYMBOL(cond_resched_lock
);
5625 int __sched
cond_resched_softirq(void)
5627 BUG_ON(!in_softirq());
5629 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5637 EXPORT_SYMBOL(cond_resched_softirq
);
5640 * yield - yield the current processor to other threads.
5642 * This is a shortcut for kernel-space yielding - it marks the
5643 * thread runnable and calls sys_sched_yield().
5645 void __sched
yield(void)
5647 set_current_state(TASK_RUNNING
);
5650 EXPORT_SYMBOL(yield
);
5653 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5654 * that process accounting knows that this is a task in IO wait state.
5656 * But don't do that if it is a deliberate, throttling IO wait (this task
5657 * has set its backing_dev_info: the queue against which it should throttle)
5659 void __sched
io_schedule(void)
5661 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5663 delayacct_blkio_start();
5664 atomic_inc(&rq
->nr_iowait
);
5666 atomic_dec(&rq
->nr_iowait
);
5667 delayacct_blkio_end();
5669 EXPORT_SYMBOL(io_schedule
);
5671 long __sched
io_schedule_timeout(long timeout
)
5673 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5676 delayacct_blkio_start();
5677 atomic_inc(&rq
->nr_iowait
);
5678 ret
= schedule_timeout(timeout
);
5679 atomic_dec(&rq
->nr_iowait
);
5680 delayacct_blkio_end();
5685 * sys_sched_get_priority_max - return maximum RT priority.
5686 * @policy: scheduling class.
5688 * this syscall returns the maximum rt_priority that can be used
5689 * by a given scheduling class.
5691 asmlinkage
long sys_sched_get_priority_max(int policy
)
5698 ret
= MAX_USER_RT_PRIO
-1;
5710 * sys_sched_get_priority_min - return minimum RT priority.
5711 * @policy: scheduling class.
5713 * this syscall returns the minimum rt_priority that can be used
5714 * by a given scheduling class.
5716 asmlinkage
long sys_sched_get_priority_min(int policy
)
5734 * sys_sched_rr_get_interval - return the default timeslice of a process.
5735 * @pid: pid of the process.
5736 * @interval: userspace pointer to the timeslice value.
5738 * this syscall writes the default timeslice value of a given process
5739 * into the user-space timespec buffer. A value of '0' means infinity.
5742 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5744 struct task_struct
*p
;
5745 unsigned int time_slice
;
5753 read_lock(&tasklist_lock
);
5754 p
= find_process_by_pid(pid
);
5758 retval
= security_task_getscheduler(p
);
5763 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5764 * tasks that are on an otherwise idle runqueue:
5767 if (p
->policy
== SCHED_RR
) {
5768 time_slice
= DEF_TIMESLICE
;
5769 } else if (p
->policy
!= SCHED_FIFO
) {
5770 struct sched_entity
*se
= &p
->se
;
5771 unsigned long flags
;
5774 rq
= task_rq_lock(p
, &flags
);
5775 if (rq
->cfs
.load
.weight
)
5776 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5777 task_rq_unlock(rq
, &flags
);
5779 read_unlock(&tasklist_lock
);
5780 jiffies_to_timespec(time_slice
, &t
);
5781 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5785 read_unlock(&tasklist_lock
);
5789 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5791 void sched_show_task(struct task_struct
*p
)
5793 unsigned long free
= 0;
5796 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5797 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5798 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5799 #if BITS_PER_LONG == 32
5800 if (state
== TASK_RUNNING
)
5801 printk(KERN_CONT
" running ");
5803 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5805 if (state
== TASK_RUNNING
)
5806 printk(KERN_CONT
" running task ");
5808 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5810 #ifdef CONFIG_DEBUG_STACK_USAGE
5812 unsigned long *n
= end_of_stack(p
);
5815 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5818 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5819 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5821 show_stack(p
, NULL
);
5824 void show_state_filter(unsigned long state_filter
)
5826 struct task_struct
*g
, *p
;
5828 #if BITS_PER_LONG == 32
5830 " task PC stack pid father\n");
5833 " task PC stack pid father\n");
5835 read_lock(&tasklist_lock
);
5836 do_each_thread(g
, p
) {
5838 * reset the NMI-timeout, listing all files on a slow
5839 * console might take alot of time:
5841 touch_nmi_watchdog();
5842 if (!state_filter
|| (p
->state
& state_filter
))
5844 } while_each_thread(g
, p
);
5846 touch_all_softlockup_watchdogs();
5848 #ifdef CONFIG_SCHED_DEBUG
5849 sysrq_sched_debug_show();
5851 read_unlock(&tasklist_lock
);
5853 * Only show locks if all tasks are dumped:
5855 if (state_filter
== -1)
5856 debug_show_all_locks();
5859 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5861 idle
->sched_class
= &idle_sched_class
;
5865 * init_idle - set up an idle thread for a given CPU
5866 * @idle: task in question
5867 * @cpu: cpu the idle task belongs to
5869 * NOTE: this function does not set the idle thread's NEED_RESCHED
5870 * flag, to make booting more robust.
5872 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5874 struct rq
*rq
= cpu_rq(cpu
);
5875 unsigned long flags
;
5878 idle
->se
.exec_start
= sched_clock();
5880 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5881 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5882 __set_task_cpu(idle
, cpu
);
5884 spin_lock_irqsave(&rq
->lock
, flags
);
5885 rq
->curr
= rq
->idle
= idle
;
5886 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5889 spin_unlock_irqrestore(&rq
->lock
, flags
);
5891 /* Set the preempt count _outside_ the spinlocks! */
5892 #if defined(CONFIG_PREEMPT)
5893 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5895 task_thread_info(idle
)->preempt_count
= 0;
5898 * The idle tasks have their own, simple scheduling class:
5900 idle
->sched_class
= &idle_sched_class
;
5904 * In a system that switches off the HZ timer nohz_cpu_mask
5905 * indicates which cpus entered this state. This is used
5906 * in the rcu update to wait only for active cpus. For system
5907 * which do not switch off the HZ timer nohz_cpu_mask should
5908 * always be CPU_MASK_NONE.
5910 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5913 * Increase the granularity value when there are more CPUs,
5914 * because with more CPUs the 'effective latency' as visible
5915 * to users decreases. But the relationship is not linear,
5916 * so pick a second-best guess by going with the log2 of the
5919 * This idea comes from the SD scheduler of Con Kolivas:
5921 static inline void sched_init_granularity(void)
5923 unsigned int factor
= 1 + ilog2(num_online_cpus());
5924 const unsigned long limit
= 200000000;
5926 sysctl_sched_min_granularity
*= factor
;
5927 if (sysctl_sched_min_granularity
> limit
)
5928 sysctl_sched_min_granularity
= limit
;
5930 sysctl_sched_latency
*= factor
;
5931 if (sysctl_sched_latency
> limit
)
5932 sysctl_sched_latency
= limit
;
5934 sysctl_sched_wakeup_granularity
*= factor
;
5936 sysctl_sched_shares_ratelimit
*= factor
;
5941 * This is how migration works:
5943 * 1) we queue a struct migration_req structure in the source CPU's
5944 * runqueue and wake up that CPU's migration thread.
5945 * 2) we down() the locked semaphore => thread blocks.
5946 * 3) migration thread wakes up (implicitly it forces the migrated
5947 * thread off the CPU)
5948 * 4) it gets the migration request and checks whether the migrated
5949 * task is still in the wrong runqueue.
5950 * 5) if it's in the wrong runqueue then the migration thread removes
5951 * it and puts it into the right queue.
5952 * 6) migration thread up()s the semaphore.
5953 * 7) we wake up and the migration is done.
5957 * Change a given task's CPU affinity. Migrate the thread to a
5958 * proper CPU and schedule it away if the CPU it's executing on
5959 * is removed from the allowed bitmask.
5961 * NOTE: the caller must have a valid reference to the task, the
5962 * task must not exit() & deallocate itself prematurely. The
5963 * call is not atomic; no spinlocks may be held.
5965 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5967 struct migration_req req
;
5968 unsigned long flags
;
5972 rq
= task_rq_lock(p
, &flags
);
5973 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5978 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5979 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5984 if (p
->sched_class
->set_cpus_allowed
)
5985 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5987 p
->cpus_allowed
= *new_mask
;
5988 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5991 /* Can the task run on the task's current CPU? If so, we're done */
5992 if (cpu_isset(task_cpu(p
), *new_mask
))
5995 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5996 /* Need help from migration thread: drop lock and wait. */
5997 task_rq_unlock(rq
, &flags
);
5998 wake_up_process(rq
->migration_thread
);
5999 wait_for_completion(&req
.done
);
6000 tlb_migrate_finish(p
->mm
);
6004 task_rq_unlock(rq
, &flags
);
6008 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6011 * Move (not current) task off this cpu, onto dest cpu. We're doing
6012 * this because either it can't run here any more (set_cpus_allowed()
6013 * away from this CPU, or CPU going down), or because we're
6014 * attempting to rebalance this task on exec (sched_exec).
6016 * So we race with normal scheduler movements, but that's OK, as long
6017 * as the task is no longer on this CPU.
6019 * Returns non-zero if task was successfully migrated.
6021 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6023 struct rq
*rq_dest
, *rq_src
;
6026 if (unlikely(!cpu_active(dest_cpu
)))
6029 rq_src
= cpu_rq(src_cpu
);
6030 rq_dest
= cpu_rq(dest_cpu
);
6032 double_rq_lock(rq_src
, rq_dest
);
6033 /* Already moved. */
6034 if (task_cpu(p
) != src_cpu
)
6036 /* Affinity changed (again). */
6037 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6040 on_rq
= p
->se
.on_rq
;
6042 deactivate_task(rq_src
, p
, 0);
6044 set_task_cpu(p
, dest_cpu
);
6046 activate_task(rq_dest
, p
, 0);
6047 check_preempt_curr(rq_dest
, p
, 0);
6052 double_rq_unlock(rq_src
, rq_dest
);
6057 * migration_thread - this is a highprio system thread that performs
6058 * thread migration by bumping thread off CPU then 'pushing' onto
6061 static int migration_thread(void *data
)
6063 int cpu
= (long)data
;
6067 BUG_ON(rq
->migration_thread
!= current
);
6069 set_current_state(TASK_INTERRUPTIBLE
);
6070 while (!kthread_should_stop()) {
6071 struct migration_req
*req
;
6072 struct list_head
*head
;
6074 spin_lock_irq(&rq
->lock
);
6076 if (cpu_is_offline(cpu
)) {
6077 spin_unlock_irq(&rq
->lock
);
6081 if (rq
->active_balance
) {
6082 active_load_balance(rq
, cpu
);
6083 rq
->active_balance
= 0;
6086 head
= &rq
->migration_queue
;
6088 if (list_empty(head
)) {
6089 spin_unlock_irq(&rq
->lock
);
6091 set_current_state(TASK_INTERRUPTIBLE
);
6094 req
= list_entry(head
->next
, struct migration_req
, list
);
6095 list_del_init(head
->next
);
6097 spin_unlock(&rq
->lock
);
6098 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6101 complete(&req
->done
);
6103 __set_current_state(TASK_RUNNING
);
6107 /* Wait for kthread_stop */
6108 set_current_state(TASK_INTERRUPTIBLE
);
6109 while (!kthread_should_stop()) {
6111 set_current_state(TASK_INTERRUPTIBLE
);
6113 __set_current_state(TASK_RUNNING
);
6117 #ifdef CONFIG_HOTPLUG_CPU
6119 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6123 local_irq_disable();
6124 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6130 * Figure out where task on dead CPU should go, use force if necessary.
6131 * NOTE: interrupts should be disabled by the caller
6133 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6135 unsigned long flags
;
6142 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6143 cpus_and(mask
, mask
, p
->cpus_allowed
);
6144 dest_cpu
= any_online_cpu(mask
);
6146 /* On any allowed CPU? */
6147 if (dest_cpu
>= nr_cpu_ids
)
6148 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6150 /* No more Mr. Nice Guy. */
6151 if (dest_cpu
>= nr_cpu_ids
) {
6152 cpumask_t cpus_allowed
;
6154 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6156 * Try to stay on the same cpuset, where the
6157 * current cpuset may be a subset of all cpus.
6158 * The cpuset_cpus_allowed_locked() variant of
6159 * cpuset_cpus_allowed() will not block. It must be
6160 * called within calls to cpuset_lock/cpuset_unlock.
6162 rq
= task_rq_lock(p
, &flags
);
6163 p
->cpus_allowed
= cpus_allowed
;
6164 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6165 task_rq_unlock(rq
, &flags
);
6168 * Don't tell them about moving exiting tasks or
6169 * kernel threads (both mm NULL), since they never
6172 if (p
->mm
&& printk_ratelimit()) {
6173 printk(KERN_INFO
"process %d (%s) no "
6174 "longer affine to cpu%d\n",
6175 task_pid_nr(p
), p
->comm
, dead_cpu
);
6178 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6182 * While a dead CPU has no uninterruptible tasks queued at this point,
6183 * it might still have a nonzero ->nr_uninterruptible counter, because
6184 * for performance reasons the counter is not stricly tracking tasks to
6185 * their home CPUs. So we just add the counter to another CPU's counter,
6186 * to keep the global sum constant after CPU-down:
6188 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6190 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6191 unsigned long flags
;
6193 local_irq_save(flags
);
6194 double_rq_lock(rq_src
, rq_dest
);
6195 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6196 rq_src
->nr_uninterruptible
= 0;
6197 double_rq_unlock(rq_src
, rq_dest
);
6198 local_irq_restore(flags
);
6201 /* Run through task list and migrate tasks from the dead cpu. */
6202 static void migrate_live_tasks(int src_cpu
)
6204 struct task_struct
*p
, *t
;
6206 read_lock(&tasklist_lock
);
6208 do_each_thread(t
, p
) {
6212 if (task_cpu(p
) == src_cpu
)
6213 move_task_off_dead_cpu(src_cpu
, p
);
6214 } while_each_thread(t
, p
);
6216 read_unlock(&tasklist_lock
);
6220 * Schedules idle task to be the next runnable task on current CPU.
6221 * It does so by boosting its priority to highest possible.
6222 * Used by CPU offline code.
6224 void sched_idle_next(void)
6226 int this_cpu
= smp_processor_id();
6227 struct rq
*rq
= cpu_rq(this_cpu
);
6228 struct task_struct
*p
= rq
->idle
;
6229 unsigned long flags
;
6231 /* cpu has to be offline */
6232 BUG_ON(cpu_online(this_cpu
));
6235 * Strictly not necessary since rest of the CPUs are stopped by now
6236 * and interrupts disabled on the current cpu.
6238 spin_lock_irqsave(&rq
->lock
, flags
);
6240 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6242 update_rq_clock(rq
);
6243 activate_task(rq
, p
, 0);
6245 spin_unlock_irqrestore(&rq
->lock
, flags
);
6249 * Ensures that the idle task is using init_mm right before its cpu goes
6252 void idle_task_exit(void)
6254 struct mm_struct
*mm
= current
->active_mm
;
6256 BUG_ON(cpu_online(smp_processor_id()));
6259 switch_mm(mm
, &init_mm
, current
);
6263 /* called under rq->lock with disabled interrupts */
6264 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6266 struct rq
*rq
= cpu_rq(dead_cpu
);
6268 /* Must be exiting, otherwise would be on tasklist. */
6269 BUG_ON(!p
->exit_state
);
6271 /* Cannot have done final schedule yet: would have vanished. */
6272 BUG_ON(p
->state
== TASK_DEAD
);
6277 * Drop lock around migration; if someone else moves it,
6278 * that's OK. No task can be added to this CPU, so iteration is
6281 spin_unlock_irq(&rq
->lock
);
6282 move_task_off_dead_cpu(dead_cpu
, p
);
6283 spin_lock_irq(&rq
->lock
);
6288 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6289 static void migrate_dead_tasks(unsigned int dead_cpu
)
6291 struct rq
*rq
= cpu_rq(dead_cpu
);
6292 struct task_struct
*next
;
6295 if (!rq
->nr_running
)
6297 update_rq_clock(rq
);
6298 next
= pick_next_task(rq
, rq
->curr
);
6301 next
->sched_class
->put_prev_task(rq
, next
);
6302 migrate_dead(dead_cpu
, next
);
6306 #endif /* CONFIG_HOTPLUG_CPU */
6308 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6310 static struct ctl_table sd_ctl_dir
[] = {
6312 .procname
= "sched_domain",
6318 static struct ctl_table sd_ctl_root
[] = {
6320 .ctl_name
= CTL_KERN
,
6321 .procname
= "kernel",
6323 .child
= sd_ctl_dir
,
6328 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6330 struct ctl_table
*entry
=
6331 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6336 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6338 struct ctl_table
*entry
;
6341 * In the intermediate directories, both the child directory and
6342 * procname are dynamically allocated and could fail but the mode
6343 * will always be set. In the lowest directory the names are
6344 * static strings and all have proc handlers.
6346 for (entry
= *tablep
; entry
->mode
; entry
++) {
6348 sd_free_ctl_entry(&entry
->child
);
6349 if (entry
->proc_handler
== NULL
)
6350 kfree(entry
->procname
);
6358 set_table_entry(struct ctl_table
*entry
,
6359 const char *procname
, void *data
, int maxlen
,
6360 mode_t mode
, proc_handler
*proc_handler
)
6362 entry
->procname
= procname
;
6364 entry
->maxlen
= maxlen
;
6366 entry
->proc_handler
= proc_handler
;
6369 static struct ctl_table
*
6370 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6372 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6377 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6378 sizeof(long), 0644, proc_doulongvec_minmax
);
6379 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6380 sizeof(long), 0644, proc_doulongvec_minmax
);
6381 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6382 sizeof(int), 0644, proc_dointvec_minmax
);
6383 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6384 sizeof(int), 0644, proc_dointvec_minmax
);
6385 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6386 sizeof(int), 0644, proc_dointvec_minmax
);
6387 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6388 sizeof(int), 0644, proc_dointvec_minmax
);
6389 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6390 sizeof(int), 0644, proc_dointvec_minmax
);
6391 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6392 sizeof(int), 0644, proc_dointvec_minmax
);
6393 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6394 sizeof(int), 0644, proc_dointvec_minmax
);
6395 set_table_entry(&table
[9], "cache_nice_tries",
6396 &sd
->cache_nice_tries
,
6397 sizeof(int), 0644, proc_dointvec_minmax
);
6398 set_table_entry(&table
[10], "flags", &sd
->flags
,
6399 sizeof(int), 0644, proc_dointvec_minmax
);
6400 set_table_entry(&table
[11], "name", sd
->name
,
6401 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6402 /* &table[12] is terminator */
6407 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6409 struct ctl_table
*entry
, *table
;
6410 struct sched_domain
*sd
;
6411 int domain_num
= 0, i
;
6414 for_each_domain(cpu
, sd
)
6416 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6421 for_each_domain(cpu
, sd
) {
6422 snprintf(buf
, 32, "domain%d", i
);
6423 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6425 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6432 static struct ctl_table_header
*sd_sysctl_header
;
6433 static void register_sched_domain_sysctl(void)
6435 int i
, cpu_num
= num_online_cpus();
6436 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6439 WARN_ON(sd_ctl_dir
[0].child
);
6440 sd_ctl_dir
[0].child
= entry
;
6445 for_each_online_cpu(i
) {
6446 snprintf(buf
, 32, "cpu%d", i
);
6447 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6449 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6453 WARN_ON(sd_sysctl_header
);
6454 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6457 /* may be called multiple times per register */
6458 static void unregister_sched_domain_sysctl(void)
6460 if (sd_sysctl_header
)
6461 unregister_sysctl_table(sd_sysctl_header
);
6462 sd_sysctl_header
= NULL
;
6463 if (sd_ctl_dir
[0].child
)
6464 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6467 static void register_sched_domain_sysctl(void)
6470 static void unregister_sched_domain_sysctl(void)
6475 static void set_rq_online(struct rq
*rq
)
6478 const struct sched_class
*class;
6480 cpu_set(rq
->cpu
, rq
->rd
->online
);
6483 for_each_class(class) {
6484 if (class->rq_online
)
6485 class->rq_online(rq
);
6490 static void set_rq_offline(struct rq
*rq
)
6493 const struct sched_class
*class;
6495 for_each_class(class) {
6496 if (class->rq_offline
)
6497 class->rq_offline(rq
);
6500 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6506 * migration_call - callback that gets triggered when a CPU is added.
6507 * Here we can start up the necessary migration thread for the new CPU.
6509 static int __cpuinit
6510 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6512 struct task_struct
*p
;
6513 int cpu
= (long)hcpu
;
6514 unsigned long flags
;
6519 case CPU_UP_PREPARE
:
6520 case CPU_UP_PREPARE_FROZEN
:
6521 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6524 kthread_bind(p
, cpu
);
6525 /* Must be high prio: stop_machine expects to yield to it. */
6526 rq
= task_rq_lock(p
, &flags
);
6527 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6528 task_rq_unlock(rq
, &flags
);
6529 cpu_rq(cpu
)->migration_thread
= p
;
6533 case CPU_ONLINE_FROZEN
:
6534 /* Strictly unnecessary, as first user will wake it. */
6535 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6537 /* Update our root-domain */
6539 spin_lock_irqsave(&rq
->lock
, flags
);
6541 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6545 spin_unlock_irqrestore(&rq
->lock
, flags
);
6548 #ifdef CONFIG_HOTPLUG_CPU
6549 case CPU_UP_CANCELED
:
6550 case CPU_UP_CANCELED_FROZEN
:
6551 if (!cpu_rq(cpu
)->migration_thread
)
6553 /* Unbind it from offline cpu so it can run. Fall thru. */
6554 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6555 any_online_cpu(cpu_online_map
));
6556 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6557 cpu_rq(cpu
)->migration_thread
= NULL
;
6561 case CPU_DEAD_FROZEN
:
6562 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6563 migrate_live_tasks(cpu
);
6565 kthread_stop(rq
->migration_thread
);
6566 rq
->migration_thread
= NULL
;
6567 /* Idle task back to normal (off runqueue, low prio) */
6568 spin_lock_irq(&rq
->lock
);
6569 update_rq_clock(rq
);
6570 deactivate_task(rq
, rq
->idle
, 0);
6571 rq
->idle
->static_prio
= MAX_PRIO
;
6572 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6573 rq
->idle
->sched_class
= &idle_sched_class
;
6574 migrate_dead_tasks(cpu
);
6575 spin_unlock_irq(&rq
->lock
);
6577 migrate_nr_uninterruptible(rq
);
6578 BUG_ON(rq
->nr_running
!= 0);
6581 * No need to migrate the tasks: it was best-effort if
6582 * they didn't take sched_hotcpu_mutex. Just wake up
6585 spin_lock_irq(&rq
->lock
);
6586 while (!list_empty(&rq
->migration_queue
)) {
6587 struct migration_req
*req
;
6589 req
= list_entry(rq
->migration_queue
.next
,
6590 struct migration_req
, list
);
6591 list_del_init(&req
->list
);
6592 complete(&req
->done
);
6594 spin_unlock_irq(&rq
->lock
);
6598 case CPU_DYING_FROZEN
:
6599 /* Update our root-domain */
6601 spin_lock_irqsave(&rq
->lock
, flags
);
6603 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6606 spin_unlock_irqrestore(&rq
->lock
, flags
);
6613 /* Register at highest priority so that task migration (migrate_all_tasks)
6614 * happens before everything else.
6616 static struct notifier_block __cpuinitdata migration_notifier
= {
6617 .notifier_call
= migration_call
,
6621 static int __init
migration_init(void)
6623 void *cpu
= (void *)(long)smp_processor_id();
6626 /* Start one for the boot CPU: */
6627 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6628 BUG_ON(err
== NOTIFY_BAD
);
6629 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6630 register_cpu_notifier(&migration_notifier
);
6634 early_initcall(migration_init
);
6639 #ifdef CONFIG_SCHED_DEBUG
6641 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6654 case SD_LV_ALLNODES
:
6663 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6664 cpumask_t
*groupmask
)
6666 struct sched_group
*group
= sd
->groups
;
6669 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6670 cpus_clear(*groupmask
);
6672 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6674 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6675 printk("does not load-balance\n");
6677 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6682 printk(KERN_CONT
"span %s level %s\n",
6683 str
, sd_level_to_string(sd
->level
));
6685 if (!cpu_isset(cpu
, sd
->span
)) {
6686 printk(KERN_ERR
"ERROR: domain->span does not contain "
6689 if (!cpu_isset(cpu
, group
->cpumask
)) {
6690 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6694 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6698 printk(KERN_ERR
"ERROR: group is NULL\n");
6702 if (!group
->__cpu_power
) {
6703 printk(KERN_CONT
"\n");
6704 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6709 if (!cpus_weight(group
->cpumask
)) {
6710 printk(KERN_CONT
"\n");
6711 printk(KERN_ERR
"ERROR: empty group\n");
6715 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6716 printk(KERN_CONT
"\n");
6717 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6721 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6723 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6724 printk(KERN_CONT
" %s", str
);
6726 group
= group
->next
;
6727 } while (group
!= sd
->groups
);
6728 printk(KERN_CONT
"\n");
6730 if (!cpus_equal(sd
->span
, *groupmask
))
6731 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6733 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6734 printk(KERN_ERR
"ERROR: parent span is not a superset "
6735 "of domain->span\n");
6739 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6741 cpumask_t
*groupmask
;
6745 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6749 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6751 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6753 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6758 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6767 #else /* !CONFIG_SCHED_DEBUG */
6768 # define sched_domain_debug(sd, cpu) do { } while (0)
6769 #endif /* CONFIG_SCHED_DEBUG */
6771 static int sd_degenerate(struct sched_domain
*sd
)
6773 if (cpus_weight(sd
->span
) == 1)
6776 /* Following flags need at least 2 groups */
6777 if (sd
->flags
& (SD_LOAD_BALANCE
|
6778 SD_BALANCE_NEWIDLE
|
6782 SD_SHARE_PKG_RESOURCES
)) {
6783 if (sd
->groups
!= sd
->groups
->next
)
6787 /* Following flags don't use groups */
6788 if (sd
->flags
& (SD_WAKE_IDLE
|
6797 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6799 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6801 if (sd_degenerate(parent
))
6804 if (!cpus_equal(sd
->span
, parent
->span
))
6807 /* Does parent contain flags not in child? */
6808 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6809 if (cflags
& SD_WAKE_AFFINE
)
6810 pflags
&= ~SD_WAKE_BALANCE
;
6811 /* Flags needing groups don't count if only 1 group in parent */
6812 if (parent
->groups
== parent
->groups
->next
) {
6813 pflags
&= ~(SD_LOAD_BALANCE
|
6814 SD_BALANCE_NEWIDLE
|
6818 SD_SHARE_PKG_RESOURCES
);
6820 if (~cflags
& pflags
)
6826 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6828 unsigned long flags
;
6830 spin_lock_irqsave(&rq
->lock
, flags
);
6833 struct root_domain
*old_rd
= rq
->rd
;
6835 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6838 cpu_clear(rq
->cpu
, old_rd
->span
);
6840 if (atomic_dec_and_test(&old_rd
->refcount
))
6844 atomic_inc(&rd
->refcount
);
6847 cpu_set(rq
->cpu
, rd
->span
);
6848 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6851 spin_unlock_irqrestore(&rq
->lock
, flags
);
6854 static void init_rootdomain(struct root_domain
*rd
)
6856 memset(rd
, 0, sizeof(*rd
));
6858 cpus_clear(rd
->span
);
6859 cpus_clear(rd
->online
);
6861 cpupri_init(&rd
->cpupri
);
6864 static void init_defrootdomain(void)
6866 init_rootdomain(&def_root_domain
);
6867 atomic_set(&def_root_domain
.refcount
, 1);
6870 static struct root_domain
*alloc_rootdomain(void)
6872 struct root_domain
*rd
;
6874 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6878 init_rootdomain(rd
);
6884 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6885 * hold the hotplug lock.
6888 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6890 struct rq
*rq
= cpu_rq(cpu
);
6891 struct sched_domain
*tmp
;
6893 /* Remove the sched domains which do not contribute to scheduling. */
6894 for (tmp
= sd
; tmp
; ) {
6895 struct sched_domain
*parent
= tmp
->parent
;
6899 if (sd_parent_degenerate(tmp
, parent
)) {
6900 tmp
->parent
= parent
->parent
;
6902 parent
->parent
->child
= tmp
;
6907 if (sd
&& sd_degenerate(sd
)) {
6913 sched_domain_debug(sd
, cpu
);
6915 rq_attach_root(rq
, rd
);
6916 rcu_assign_pointer(rq
->sd
, sd
);
6919 /* cpus with isolated domains */
6920 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6922 /* Setup the mask of cpus configured for isolated domains */
6923 static int __init
isolated_cpu_setup(char *str
)
6925 static int __initdata ints
[NR_CPUS
];
6928 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6929 cpus_clear(cpu_isolated_map
);
6930 for (i
= 1; i
<= ints
[0]; i
++)
6931 if (ints
[i
] < NR_CPUS
)
6932 cpu_set(ints
[i
], cpu_isolated_map
);
6936 __setup("isolcpus=", isolated_cpu_setup
);
6939 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6940 * to a function which identifies what group(along with sched group) a CPU
6941 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6942 * (due to the fact that we keep track of groups covered with a cpumask_t).
6944 * init_sched_build_groups will build a circular linked list of the groups
6945 * covered by the given span, and will set each group's ->cpumask correctly,
6946 * and ->cpu_power to 0.
6949 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6950 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6951 struct sched_group
**sg
,
6952 cpumask_t
*tmpmask
),
6953 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6955 struct sched_group
*first
= NULL
, *last
= NULL
;
6958 cpus_clear(*covered
);
6960 for_each_cpu_mask_nr(i
, *span
) {
6961 struct sched_group
*sg
;
6962 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6965 if (cpu_isset(i
, *covered
))
6968 cpus_clear(sg
->cpumask
);
6969 sg
->__cpu_power
= 0;
6971 for_each_cpu_mask_nr(j
, *span
) {
6972 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6975 cpu_set(j
, *covered
);
6976 cpu_set(j
, sg
->cpumask
);
6987 #define SD_NODES_PER_DOMAIN 16
6992 * find_next_best_node - find the next node to include in a sched_domain
6993 * @node: node whose sched_domain we're building
6994 * @used_nodes: nodes already in the sched_domain
6996 * Find the next node to include in a given scheduling domain. Simply
6997 * finds the closest node not already in the @used_nodes map.
6999 * Should use nodemask_t.
7001 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7003 int i
, n
, val
, min_val
, best_node
= 0;
7007 for (i
= 0; i
< nr_node_ids
; i
++) {
7008 /* Start at @node */
7009 n
= (node
+ i
) % nr_node_ids
;
7011 if (!nr_cpus_node(n
))
7014 /* Skip already used nodes */
7015 if (node_isset(n
, *used_nodes
))
7018 /* Simple min distance search */
7019 val
= node_distance(node
, n
);
7021 if (val
< min_val
) {
7027 node_set(best_node
, *used_nodes
);
7032 * sched_domain_node_span - get a cpumask for a node's sched_domain
7033 * @node: node whose cpumask we're constructing
7034 * @span: resulting cpumask
7036 * Given a node, construct a good cpumask for its sched_domain to span. It
7037 * should be one that prevents unnecessary balancing, but also spreads tasks
7040 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7042 nodemask_t used_nodes
;
7043 node_to_cpumask_ptr(nodemask
, node
);
7047 nodes_clear(used_nodes
);
7049 cpus_or(*span
, *span
, *nodemask
);
7050 node_set(node
, used_nodes
);
7052 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7053 int next_node
= find_next_best_node(node
, &used_nodes
);
7055 node_to_cpumask_ptr_next(nodemask
, next_node
);
7056 cpus_or(*span
, *span
, *nodemask
);
7059 #endif /* CONFIG_NUMA */
7061 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7064 * SMT sched-domains:
7066 #ifdef CONFIG_SCHED_SMT
7067 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7068 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7071 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7075 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7078 #endif /* CONFIG_SCHED_SMT */
7081 * multi-core sched-domains:
7083 #ifdef CONFIG_SCHED_MC
7084 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7085 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7086 #endif /* CONFIG_SCHED_MC */
7088 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7090 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7095 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7096 cpus_and(*mask
, *mask
, *cpu_map
);
7097 group
= first_cpu(*mask
);
7099 *sg
= &per_cpu(sched_group_core
, group
);
7102 #elif defined(CONFIG_SCHED_MC)
7104 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7108 *sg
= &per_cpu(sched_group_core
, cpu
);
7113 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7114 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7117 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7121 #ifdef CONFIG_SCHED_MC
7122 *mask
= cpu_coregroup_map(cpu
);
7123 cpus_and(*mask
, *mask
, *cpu_map
);
7124 group
= first_cpu(*mask
);
7125 #elif defined(CONFIG_SCHED_SMT)
7126 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7127 cpus_and(*mask
, *mask
, *cpu_map
);
7128 group
= first_cpu(*mask
);
7133 *sg
= &per_cpu(sched_group_phys
, group
);
7139 * The init_sched_build_groups can't handle what we want to do with node
7140 * groups, so roll our own. Now each node has its own list of groups which
7141 * gets dynamically allocated.
7143 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7144 static struct sched_group
***sched_group_nodes_bycpu
;
7146 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7147 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7149 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7150 struct sched_group
**sg
, cpumask_t
*nodemask
)
7154 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7155 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7156 group
= first_cpu(*nodemask
);
7159 *sg
= &per_cpu(sched_group_allnodes
, group
);
7163 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7165 struct sched_group
*sg
= group_head
;
7171 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7172 struct sched_domain
*sd
;
7174 sd
= &per_cpu(phys_domains
, j
);
7175 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7177 * Only add "power" once for each
7183 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7186 } while (sg
!= group_head
);
7188 #endif /* CONFIG_NUMA */
7191 /* Free memory allocated for various sched_group structures */
7192 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7196 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7197 struct sched_group
**sched_group_nodes
7198 = sched_group_nodes_bycpu
[cpu
];
7200 if (!sched_group_nodes
)
7203 for (i
= 0; i
< nr_node_ids
; i
++) {
7204 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7206 *nodemask
= node_to_cpumask(i
);
7207 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7208 if (cpus_empty(*nodemask
))
7218 if (oldsg
!= sched_group_nodes
[i
])
7221 kfree(sched_group_nodes
);
7222 sched_group_nodes_bycpu
[cpu
] = NULL
;
7225 #else /* !CONFIG_NUMA */
7226 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7229 #endif /* CONFIG_NUMA */
7232 * Initialize sched groups cpu_power.
7234 * cpu_power indicates the capacity of sched group, which is used while
7235 * distributing the load between different sched groups in a sched domain.
7236 * Typically cpu_power for all the groups in a sched domain will be same unless
7237 * there are asymmetries in the topology. If there are asymmetries, group
7238 * having more cpu_power will pickup more load compared to the group having
7241 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7242 * the maximum number of tasks a group can handle in the presence of other idle
7243 * or lightly loaded groups in the same sched domain.
7245 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7247 struct sched_domain
*child
;
7248 struct sched_group
*group
;
7250 WARN_ON(!sd
|| !sd
->groups
);
7252 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7257 sd
->groups
->__cpu_power
= 0;
7260 * For perf policy, if the groups in child domain share resources
7261 * (for example cores sharing some portions of the cache hierarchy
7262 * or SMT), then set this domain groups cpu_power such that each group
7263 * can handle only one task, when there are other idle groups in the
7264 * same sched domain.
7266 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7268 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7269 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7274 * add cpu_power of each child group to this groups cpu_power
7276 group
= child
->groups
;
7278 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7279 group
= group
->next
;
7280 } while (group
!= child
->groups
);
7284 * Initializers for schedule domains
7285 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7288 #ifdef CONFIG_SCHED_DEBUG
7289 # define SD_INIT_NAME(sd, type) sd->name = #type
7291 # define SD_INIT_NAME(sd, type) do { } while (0)
7294 #define SD_INIT(sd, type) sd_init_##type(sd)
7296 #define SD_INIT_FUNC(type) \
7297 static noinline void sd_init_##type(struct sched_domain *sd) \
7299 memset(sd, 0, sizeof(*sd)); \
7300 *sd = SD_##type##_INIT; \
7301 sd->level = SD_LV_##type; \
7302 SD_INIT_NAME(sd, type); \
7307 SD_INIT_FUNC(ALLNODES
)
7310 #ifdef CONFIG_SCHED_SMT
7311 SD_INIT_FUNC(SIBLING
)
7313 #ifdef CONFIG_SCHED_MC
7318 * To minimize stack usage kmalloc room for cpumasks and share the
7319 * space as the usage in build_sched_domains() dictates. Used only
7320 * if the amount of space is significant.
7323 cpumask_t tmpmask
; /* make this one first */
7326 cpumask_t this_sibling_map
;
7327 cpumask_t this_core_map
;
7329 cpumask_t send_covered
;
7332 cpumask_t domainspan
;
7334 cpumask_t notcovered
;
7339 #define SCHED_CPUMASK_ALLOC 1
7340 #define SCHED_CPUMASK_FREE(v) kfree(v)
7341 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7343 #define SCHED_CPUMASK_ALLOC 0
7344 #define SCHED_CPUMASK_FREE(v)
7345 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7348 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7349 ((unsigned long)(a) + offsetof(struct allmasks, v))
7351 static int default_relax_domain_level
= -1;
7353 static int __init
setup_relax_domain_level(char *str
)
7357 val
= simple_strtoul(str
, NULL
, 0);
7358 if (val
< SD_LV_MAX
)
7359 default_relax_domain_level
= val
;
7363 __setup("relax_domain_level=", setup_relax_domain_level
);
7365 static void set_domain_attribute(struct sched_domain
*sd
,
7366 struct sched_domain_attr
*attr
)
7370 if (!attr
|| attr
->relax_domain_level
< 0) {
7371 if (default_relax_domain_level
< 0)
7374 request
= default_relax_domain_level
;
7376 request
= attr
->relax_domain_level
;
7377 if (request
< sd
->level
) {
7378 /* turn off idle balance on this domain */
7379 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7381 /* turn on idle balance on this domain */
7382 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7387 * Build sched domains for a given set of cpus and attach the sched domains
7388 * to the individual cpus
7390 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7391 struct sched_domain_attr
*attr
)
7394 struct root_domain
*rd
;
7395 SCHED_CPUMASK_DECLARE(allmasks
);
7398 struct sched_group
**sched_group_nodes
= NULL
;
7399 int sd_allnodes
= 0;
7402 * Allocate the per-node list of sched groups
7404 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7406 if (!sched_group_nodes
) {
7407 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7412 rd
= alloc_rootdomain();
7414 printk(KERN_WARNING
"Cannot alloc root domain\n");
7416 kfree(sched_group_nodes
);
7421 #if SCHED_CPUMASK_ALLOC
7422 /* get space for all scratch cpumask variables */
7423 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7425 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7428 kfree(sched_group_nodes
);
7433 tmpmask
= (cpumask_t
*)allmasks
;
7437 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7441 * Set up domains for cpus specified by the cpu_map.
7443 for_each_cpu_mask_nr(i
, *cpu_map
) {
7444 struct sched_domain
*sd
= NULL
, *p
;
7445 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7447 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7448 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7451 if (cpus_weight(*cpu_map
) >
7452 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7453 sd
= &per_cpu(allnodes_domains
, i
);
7454 SD_INIT(sd
, ALLNODES
);
7455 set_domain_attribute(sd
, attr
);
7456 sd
->span
= *cpu_map
;
7457 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7463 sd
= &per_cpu(node_domains
, i
);
7465 set_domain_attribute(sd
, attr
);
7466 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7470 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7474 sd
= &per_cpu(phys_domains
, i
);
7476 set_domain_attribute(sd
, attr
);
7477 sd
->span
= *nodemask
;
7481 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7483 #ifdef CONFIG_SCHED_MC
7485 sd
= &per_cpu(core_domains
, i
);
7487 set_domain_attribute(sd
, attr
);
7488 sd
->span
= cpu_coregroup_map(i
);
7489 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7492 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7495 #ifdef CONFIG_SCHED_SMT
7497 sd
= &per_cpu(cpu_domains
, i
);
7498 SD_INIT(sd
, SIBLING
);
7499 set_domain_attribute(sd
, attr
);
7500 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7501 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7504 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7508 #ifdef CONFIG_SCHED_SMT
7509 /* Set up CPU (sibling) groups */
7510 for_each_cpu_mask_nr(i
, *cpu_map
) {
7511 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7512 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7514 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7515 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7516 if (i
!= first_cpu(*this_sibling_map
))
7519 init_sched_build_groups(this_sibling_map
, cpu_map
,
7521 send_covered
, tmpmask
);
7525 #ifdef CONFIG_SCHED_MC
7526 /* Set up multi-core groups */
7527 for_each_cpu_mask_nr(i
, *cpu_map
) {
7528 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7529 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7531 *this_core_map
= cpu_coregroup_map(i
);
7532 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7533 if (i
!= first_cpu(*this_core_map
))
7536 init_sched_build_groups(this_core_map
, cpu_map
,
7538 send_covered
, tmpmask
);
7542 /* Set up physical groups */
7543 for (i
= 0; i
< nr_node_ids
; i
++) {
7544 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7545 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7547 *nodemask
= node_to_cpumask(i
);
7548 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7549 if (cpus_empty(*nodemask
))
7552 init_sched_build_groups(nodemask
, cpu_map
,
7554 send_covered
, tmpmask
);
7558 /* Set up node groups */
7560 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7562 init_sched_build_groups(cpu_map
, cpu_map
,
7563 &cpu_to_allnodes_group
,
7564 send_covered
, tmpmask
);
7567 for (i
= 0; i
< nr_node_ids
; i
++) {
7568 /* Set up node groups */
7569 struct sched_group
*sg
, *prev
;
7570 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7571 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7572 SCHED_CPUMASK_VAR(covered
, allmasks
);
7575 *nodemask
= node_to_cpumask(i
);
7576 cpus_clear(*covered
);
7578 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7579 if (cpus_empty(*nodemask
)) {
7580 sched_group_nodes
[i
] = NULL
;
7584 sched_domain_node_span(i
, domainspan
);
7585 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7587 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7589 printk(KERN_WARNING
"Can not alloc domain group for "
7593 sched_group_nodes
[i
] = sg
;
7594 for_each_cpu_mask_nr(j
, *nodemask
) {
7595 struct sched_domain
*sd
;
7597 sd
= &per_cpu(node_domains
, j
);
7600 sg
->__cpu_power
= 0;
7601 sg
->cpumask
= *nodemask
;
7603 cpus_or(*covered
, *covered
, *nodemask
);
7606 for (j
= 0; j
< nr_node_ids
; j
++) {
7607 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7608 int n
= (i
+ j
) % nr_node_ids
;
7609 node_to_cpumask_ptr(pnodemask
, n
);
7611 cpus_complement(*notcovered
, *covered
);
7612 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7613 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7614 if (cpus_empty(*tmpmask
))
7617 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7618 if (cpus_empty(*tmpmask
))
7621 sg
= kmalloc_node(sizeof(struct sched_group
),
7625 "Can not alloc domain group for node %d\n", j
);
7628 sg
->__cpu_power
= 0;
7629 sg
->cpumask
= *tmpmask
;
7630 sg
->next
= prev
->next
;
7631 cpus_or(*covered
, *covered
, *tmpmask
);
7638 /* Calculate CPU power for physical packages and nodes */
7639 #ifdef CONFIG_SCHED_SMT
7640 for_each_cpu_mask_nr(i
, *cpu_map
) {
7641 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7643 init_sched_groups_power(i
, sd
);
7646 #ifdef CONFIG_SCHED_MC
7647 for_each_cpu_mask_nr(i
, *cpu_map
) {
7648 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7650 init_sched_groups_power(i
, sd
);
7654 for_each_cpu_mask_nr(i
, *cpu_map
) {
7655 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7657 init_sched_groups_power(i
, sd
);
7661 for (i
= 0; i
< nr_node_ids
; i
++)
7662 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7665 struct sched_group
*sg
;
7667 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7669 init_numa_sched_groups_power(sg
);
7673 /* Attach the domains */
7674 for_each_cpu_mask_nr(i
, *cpu_map
) {
7675 struct sched_domain
*sd
;
7676 #ifdef CONFIG_SCHED_SMT
7677 sd
= &per_cpu(cpu_domains
, i
);
7678 #elif defined(CONFIG_SCHED_MC)
7679 sd
= &per_cpu(core_domains
, i
);
7681 sd
= &per_cpu(phys_domains
, i
);
7683 cpu_attach_domain(sd
, rd
, i
);
7686 SCHED_CPUMASK_FREE((void *)allmasks
);
7691 free_sched_groups(cpu_map
, tmpmask
);
7692 SCHED_CPUMASK_FREE((void *)allmasks
);
7698 static int build_sched_domains(const cpumask_t
*cpu_map
)
7700 return __build_sched_domains(cpu_map
, NULL
);
7703 static cpumask_t
*doms_cur
; /* current sched domains */
7704 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7705 static struct sched_domain_attr
*dattr_cur
;
7706 /* attribues of custom domains in 'doms_cur' */
7709 * Special case: If a kmalloc of a doms_cur partition (array of
7710 * cpumask_t) fails, then fallback to a single sched domain,
7711 * as determined by the single cpumask_t fallback_doms.
7713 static cpumask_t fallback_doms
;
7715 void __attribute__((weak
)) arch_update_cpu_topology(void)
7720 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7721 * For now this just excludes isolated cpus, but could be used to
7722 * exclude other special cases in the future.
7724 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7728 arch_update_cpu_topology();
7730 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7732 doms_cur
= &fallback_doms
;
7733 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7735 err
= build_sched_domains(doms_cur
);
7736 register_sched_domain_sysctl();
7741 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7744 free_sched_groups(cpu_map
, tmpmask
);
7748 * Detach sched domains from a group of cpus specified in cpu_map
7749 * These cpus will now be attached to the NULL domain
7751 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7756 unregister_sched_domain_sysctl();
7758 for_each_cpu_mask_nr(i
, *cpu_map
)
7759 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7760 synchronize_sched();
7761 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7764 /* handle null as "default" */
7765 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7766 struct sched_domain_attr
*new, int idx_new
)
7768 struct sched_domain_attr tmp
;
7775 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7776 new ? (new + idx_new
) : &tmp
,
7777 sizeof(struct sched_domain_attr
));
7781 * Partition sched domains as specified by the 'ndoms_new'
7782 * cpumasks in the array doms_new[] of cpumasks. This compares
7783 * doms_new[] to the current sched domain partitioning, doms_cur[].
7784 * It destroys each deleted domain and builds each new domain.
7786 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7787 * The masks don't intersect (don't overlap.) We should setup one
7788 * sched domain for each mask. CPUs not in any of the cpumasks will
7789 * not be load balanced. If the same cpumask appears both in the
7790 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7793 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7794 * ownership of it and will kfree it when done with it. If the caller
7795 * failed the kmalloc call, then it can pass in doms_new == NULL,
7796 * and partition_sched_domains() will fallback to the single partition
7797 * 'fallback_doms', it also forces the domains to be rebuilt.
7799 * If doms_new==NULL it will be replaced with cpu_online_map.
7800 * ndoms_new==0 is a special case for destroying existing domains.
7801 * It will not create the default domain.
7803 * Call with hotplug lock held
7805 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7806 struct sched_domain_attr
*dattr_new
)
7810 mutex_lock(&sched_domains_mutex
);
7812 /* always unregister in case we don't destroy any domains */
7813 unregister_sched_domain_sysctl();
7815 n
= doms_new
? ndoms_new
: 0;
7817 /* Destroy deleted domains */
7818 for (i
= 0; i
< ndoms_cur
; i
++) {
7819 for (j
= 0; j
< n
; j
++) {
7820 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7821 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7824 /* no match - a current sched domain not in new doms_new[] */
7825 detach_destroy_domains(doms_cur
+ i
);
7830 if (doms_new
== NULL
) {
7832 doms_new
= &fallback_doms
;
7833 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7837 /* Build new domains */
7838 for (i
= 0; i
< ndoms_new
; i
++) {
7839 for (j
= 0; j
< ndoms_cur
; j
++) {
7840 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7841 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7844 /* no match - add a new doms_new */
7845 __build_sched_domains(doms_new
+ i
,
7846 dattr_new
? dattr_new
+ i
: NULL
);
7851 /* Remember the new sched domains */
7852 if (doms_cur
!= &fallback_doms
)
7854 kfree(dattr_cur
); /* kfree(NULL) is safe */
7855 doms_cur
= doms_new
;
7856 dattr_cur
= dattr_new
;
7857 ndoms_cur
= ndoms_new
;
7859 register_sched_domain_sysctl();
7861 mutex_unlock(&sched_domains_mutex
);
7864 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7865 int arch_reinit_sched_domains(void)
7869 /* Destroy domains first to force the rebuild */
7870 partition_sched_domains(0, NULL
, NULL
);
7872 rebuild_sched_domains();
7878 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7882 if (buf
[0] != '0' && buf
[0] != '1')
7886 sched_smt_power_savings
= (buf
[0] == '1');
7888 sched_mc_power_savings
= (buf
[0] == '1');
7890 ret
= arch_reinit_sched_domains();
7892 return ret
? ret
: count
;
7895 #ifdef CONFIG_SCHED_MC
7896 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7899 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7901 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7902 const char *buf
, size_t count
)
7904 return sched_power_savings_store(buf
, count
, 0);
7906 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7907 sched_mc_power_savings_show
,
7908 sched_mc_power_savings_store
);
7911 #ifdef CONFIG_SCHED_SMT
7912 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7915 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7917 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7918 const char *buf
, size_t count
)
7920 return sched_power_savings_store(buf
, count
, 1);
7922 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7923 sched_smt_power_savings_show
,
7924 sched_smt_power_savings_store
);
7927 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7931 #ifdef CONFIG_SCHED_SMT
7933 err
= sysfs_create_file(&cls
->kset
.kobj
,
7934 &attr_sched_smt_power_savings
.attr
);
7936 #ifdef CONFIG_SCHED_MC
7937 if (!err
&& mc_capable())
7938 err
= sysfs_create_file(&cls
->kset
.kobj
,
7939 &attr_sched_mc_power_savings
.attr
);
7943 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7945 #ifndef CONFIG_CPUSETS
7947 * Add online and remove offline CPUs from the scheduler domains.
7948 * When cpusets are enabled they take over this function.
7950 static int update_sched_domains(struct notifier_block
*nfb
,
7951 unsigned long action
, void *hcpu
)
7955 case CPU_ONLINE_FROZEN
:
7957 case CPU_DEAD_FROZEN
:
7958 partition_sched_domains(1, NULL
, NULL
);
7967 static int update_runtime(struct notifier_block
*nfb
,
7968 unsigned long action
, void *hcpu
)
7970 int cpu
= (int)(long)hcpu
;
7973 case CPU_DOWN_PREPARE
:
7974 case CPU_DOWN_PREPARE_FROZEN
:
7975 disable_runtime(cpu_rq(cpu
));
7978 case CPU_DOWN_FAILED
:
7979 case CPU_DOWN_FAILED_FROZEN
:
7981 case CPU_ONLINE_FROZEN
:
7982 enable_runtime(cpu_rq(cpu
));
7990 void __init
sched_init_smp(void)
7992 cpumask_t non_isolated_cpus
;
7994 #if defined(CONFIG_NUMA)
7995 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7997 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8000 mutex_lock(&sched_domains_mutex
);
8001 arch_init_sched_domains(&cpu_online_map
);
8002 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
8003 if (cpus_empty(non_isolated_cpus
))
8004 cpu_set(smp_processor_id(), non_isolated_cpus
);
8005 mutex_unlock(&sched_domains_mutex
);
8008 #ifndef CONFIG_CPUSETS
8009 /* XXX: Theoretical race here - CPU may be hotplugged now */
8010 hotcpu_notifier(update_sched_domains
, 0);
8013 /* RT runtime code needs to handle some hotplug events */
8014 hotcpu_notifier(update_runtime
, 0);
8018 /* Move init over to a non-isolated CPU */
8019 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
8021 sched_init_granularity();
8024 void __init
sched_init_smp(void)
8026 sched_init_granularity();
8028 #endif /* CONFIG_SMP */
8030 int in_sched_functions(unsigned long addr
)
8032 return in_lock_functions(addr
) ||
8033 (addr
>= (unsigned long)__sched_text_start
8034 && addr
< (unsigned long)__sched_text_end
);
8037 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8039 cfs_rq
->tasks_timeline
= RB_ROOT
;
8040 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8041 #ifdef CONFIG_FAIR_GROUP_SCHED
8044 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8047 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8049 struct rt_prio_array
*array
;
8052 array
= &rt_rq
->active
;
8053 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8054 INIT_LIST_HEAD(array
->queue
+ i
);
8055 __clear_bit(i
, array
->bitmap
);
8057 /* delimiter for bitsearch: */
8058 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8060 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8061 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8064 rt_rq
->rt_nr_migratory
= 0;
8065 rt_rq
->overloaded
= 0;
8069 rt_rq
->rt_throttled
= 0;
8070 rt_rq
->rt_runtime
= 0;
8071 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8073 #ifdef CONFIG_RT_GROUP_SCHED
8074 rt_rq
->rt_nr_boosted
= 0;
8079 #ifdef CONFIG_FAIR_GROUP_SCHED
8080 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8081 struct sched_entity
*se
, int cpu
, int add
,
8082 struct sched_entity
*parent
)
8084 struct rq
*rq
= cpu_rq(cpu
);
8085 tg
->cfs_rq
[cpu
] = cfs_rq
;
8086 init_cfs_rq(cfs_rq
, rq
);
8089 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8092 /* se could be NULL for init_task_group */
8097 se
->cfs_rq
= &rq
->cfs
;
8099 se
->cfs_rq
= parent
->my_q
;
8102 se
->load
.weight
= tg
->shares
;
8103 se
->load
.inv_weight
= 0;
8104 se
->parent
= parent
;
8108 #ifdef CONFIG_RT_GROUP_SCHED
8109 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8110 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8111 struct sched_rt_entity
*parent
)
8113 struct rq
*rq
= cpu_rq(cpu
);
8115 tg
->rt_rq
[cpu
] = rt_rq
;
8116 init_rt_rq(rt_rq
, rq
);
8118 rt_rq
->rt_se
= rt_se
;
8119 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8121 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8123 tg
->rt_se
[cpu
] = rt_se
;
8128 rt_se
->rt_rq
= &rq
->rt
;
8130 rt_se
->rt_rq
= parent
->my_q
;
8132 rt_se
->my_q
= rt_rq
;
8133 rt_se
->parent
= parent
;
8134 INIT_LIST_HEAD(&rt_se
->run_list
);
8138 void __init
sched_init(void)
8141 unsigned long alloc_size
= 0, ptr
;
8143 #ifdef CONFIG_FAIR_GROUP_SCHED
8144 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8146 #ifdef CONFIG_RT_GROUP_SCHED
8147 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8149 #ifdef CONFIG_USER_SCHED
8153 * As sched_init() is called before page_alloc is setup,
8154 * we use alloc_bootmem().
8157 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8159 #ifdef CONFIG_FAIR_GROUP_SCHED
8160 init_task_group
.se
= (struct sched_entity
**)ptr
;
8161 ptr
+= nr_cpu_ids
* sizeof(void **);
8163 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8164 ptr
+= nr_cpu_ids
* sizeof(void **);
8166 #ifdef CONFIG_USER_SCHED
8167 root_task_group
.se
= (struct sched_entity
**)ptr
;
8168 ptr
+= nr_cpu_ids
* sizeof(void **);
8170 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8171 ptr
+= nr_cpu_ids
* sizeof(void **);
8172 #endif /* CONFIG_USER_SCHED */
8173 #endif /* CONFIG_FAIR_GROUP_SCHED */
8174 #ifdef CONFIG_RT_GROUP_SCHED
8175 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8176 ptr
+= nr_cpu_ids
* sizeof(void **);
8178 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8179 ptr
+= nr_cpu_ids
* sizeof(void **);
8181 #ifdef CONFIG_USER_SCHED
8182 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8183 ptr
+= nr_cpu_ids
* sizeof(void **);
8185 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8186 ptr
+= nr_cpu_ids
* sizeof(void **);
8187 #endif /* CONFIG_USER_SCHED */
8188 #endif /* CONFIG_RT_GROUP_SCHED */
8192 init_defrootdomain();
8195 init_rt_bandwidth(&def_rt_bandwidth
,
8196 global_rt_period(), global_rt_runtime());
8198 #ifdef CONFIG_RT_GROUP_SCHED
8199 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8200 global_rt_period(), global_rt_runtime());
8201 #ifdef CONFIG_USER_SCHED
8202 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8203 global_rt_period(), RUNTIME_INF
);
8204 #endif /* CONFIG_USER_SCHED */
8205 #endif /* CONFIG_RT_GROUP_SCHED */
8207 #ifdef CONFIG_GROUP_SCHED
8208 list_add(&init_task_group
.list
, &task_groups
);
8209 INIT_LIST_HEAD(&init_task_group
.children
);
8211 #ifdef CONFIG_USER_SCHED
8212 INIT_LIST_HEAD(&root_task_group
.children
);
8213 init_task_group
.parent
= &root_task_group
;
8214 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8215 #endif /* CONFIG_USER_SCHED */
8216 #endif /* CONFIG_GROUP_SCHED */
8218 for_each_possible_cpu(i
) {
8222 spin_lock_init(&rq
->lock
);
8224 init_cfs_rq(&rq
->cfs
, rq
);
8225 init_rt_rq(&rq
->rt
, rq
);
8226 #ifdef CONFIG_FAIR_GROUP_SCHED
8227 init_task_group
.shares
= init_task_group_load
;
8228 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8229 #ifdef CONFIG_CGROUP_SCHED
8231 * How much cpu bandwidth does init_task_group get?
8233 * In case of task-groups formed thr' the cgroup filesystem, it
8234 * gets 100% of the cpu resources in the system. This overall
8235 * system cpu resource is divided among the tasks of
8236 * init_task_group and its child task-groups in a fair manner,
8237 * based on each entity's (task or task-group's) weight
8238 * (se->load.weight).
8240 * In other words, if init_task_group has 10 tasks of weight
8241 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8242 * then A0's share of the cpu resource is:
8244 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8246 * We achieve this by letting init_task_group's tasks sit
8247 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8249 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8250 #elif defined CONFIG_USER_SCHED
8251 root_task_group
.shares
= NICE_0_LOAD
;
8252 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8254 * In case of task-groups formed thr' the user id of tasks,
8255 * init_task_group represents tasks belonging to root user.
8256 * Hence it forms a sibling of all subsequent groups formed.
8257 * In this case, init_task_group gets only a fraction of overall
8258 * system cpu resource, based on the weight assigned to root
8259 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8260 * by letting tasks of init_task_group sit in a separate cfs_rq
8261 * (init_cfs_rq) and having one entity represent this group of
8262 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8264 init_tg_cfs_entry(&init_task_group
,
8265 &per_cpu(init_cfs_rq
, i
),
8266 &per_cpu(init_sched_entity
, i
), i
, 1,
8267 root_task_group
.se
[i
]);
8270 #endif /* CONFIG_FAIR_GROUP_SCHED */
8272 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8273 #ifdef CONFIG_RT_GROUP_SCHED
8274 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8275 #ifdef CONFIG_CGROUP_SCHED
8276 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8277 #elif defined CONFIG_USER_SCHED
8278 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8279 init_tg_rt_entry(&init_task_group
,
8280 &per_cpu(init_rt_rq
, i
),
8281 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8282 root_task_group
.rt_se
[i
]);
8286 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8287 rq
->cpu_load
[j
] = 0;
8291 rq
->active_balance
= 0;
8292 rq
->next_balance
= jiffies
;
8296 rq
->migration_thread
= NULL
;
8297 INIT_LIST_HEAD(&rq
->migration_queue
);
8298 rq_attach_root(rq
, &def_root_domain
);
8301 atomic_set(&rq
->nr_iowait
, 0);
8304 set_load_weight(&init_task
);
8306 #ifdef CONFIG_PREEMPT_NOTIFIERS
8307 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8311 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8314 #ifdef CONFIG_RT_MUTEXES
8315 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8319 * The boot idle thread does lazy MMU switching as well:
8321 atomic_inc(&init_mm
.mm_count
);
8322 enter_lazy_tlb(&init_mm
, current
);
8325 * Make us the idle thread. Technically, schedule() should not be
8326 * called from this thread, however somewhere below it might be,
8327 * but because we are the idle thread, we just pick up running again
8328 * when this runqueue becomes "idle".
8330 init_idle(current
, smp_processor_id());
8332 * During early bootup we pretend to be a normal task:
8334 current
->sched_class
= &fair_sched_class
;
8336 scheduler_running
= 1;
8339 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8340 void __might_sleep(char *file
, int line
)
8343 static unsigned long prev_jiffy
; /* ratelimiting */
8345 if ((!in_atomic() && !irqs_disabled()) ||
8346 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8348 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8350 prev_jiffy
= jiffies
;
8353 "BUG: sleeping function called from invalid context at %s:%d\n",
8356 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8357 in_atomic(), irqs_disabled(),
8358 current
->pid
, current
->comm
);
8360 debug_show_held_locks(current
);
8361 if (irqs_disabled())
8362 print_irqtrace_events(current
);
8366 EXPORT_SYMBOL(__might_sleep
);
8369 #ifdef CONFIG_MAGIC_SYSRQ
8370 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8374 update_rq_clock(rq
);
8375 on_rq
= p
->se
.on_rq
;
8377 deactivate_task(rq
, p
, 0);
8378 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8380 activate_task(rq
, p
, 0);
8381 resched_task(rq
->curr
);
8385 void normalize_rt_tasks(void)
8387 struct task_struct
*g
, *p
;
8388 unsigned long flags
;
8391 read_lock_irqsave(&tasklist_lock
, flags
);
8392 do_each_thread(g
, p
) {
8394 * Only normalize user tasks:
8399 p
->se
.exec_start
= 0;
8400 #ifdef CONFIG_SCHEDSTATS
8401 p
->se
.wait_start
= 0;
8402 p
->se
.sleep_start
= 0;
8403 p
->se
.block_start
= 0;
8408 * Renice negative nice level userspace
8411 if (TASK_NICE(p
) < 0 && p
->mm
)
8412 set_user_nice(p
, 0);
8416 spin_lock(&p
->pi_lock
);
8417 rq
= __task_rq_lock(p
);
8419 normalize_task(rq
, p
);
8421 __task_rq_unlock(rq
);
8422 spin_unlock(&p
->pi_lock
);
8423 } while_each_thread(g
, p
);
8425 read_unlock_irqrestore(&tasklist_lock
, flags
);
8428 #endif /* CONFIG_MAGIC_SYSRQ */
8432 * These functions are only useful for the IA64 MCA handling.
8434 * They can only be called when the whole system has been
8435 * stopped - every CPU needs to be quiescent, and no scheduling
8436 * activity can take place. Using them for anything else would
8437 * be a serious bug, and as a result, they aren't even visible
8438 * under any other configuration.
8442 * curr_task - return the current task for a given cpu.
8443 * @cpu: the processor in question.
8445 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8447 struct task_struct
*curr_task(int cpu
)
8449 return cpu_curr(cpu
);
8453 * set_curr_task - set the current task for a given cpu.
8454 * @cpu: the processor in question.
8455 * @p: the task pointer to set.
8457 * Description: This function must only be used when non-maskable interrupts
8458 * are serviced on a separate stack. It allows the architecture to switch the
8459 * notion of the current task on a cpu in a non-blocking manner. This function
8460 * must be called with all CPU's synchronized, and interrupts disabled, the
8461 * and caller must save the original value of the current task (see
8462 * curr_task() above) and restore that value before reenabling interrupts and
8463 * re-starting the system.
8465 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8467 void set_curr_task(int cpu
, struct task_struct
*p
)
8474 #ifdef CONFIG_FAIR_GROUP_SCHED
8475 static void free_fair_sched_group(struct task_group
*tg
)
8479 for_each_possible_cpu(i
) {
8481 kfree(tg
->cfs_rq
[i
]);
8491 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8493 struct cfs_rq
*cfs_rq
;
8494 struct sched_entity
*se
, *parent_se
;
8498 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8501 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8505 tg
->shares
= NICE_0_LOAD
;
8507 for_each_possible_cpu(i
) {
8510 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8511 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8515 se
= kmalloc_node(sizeof(struct sched_entity
),
8516 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8520 parent_se
= parent
? parent
->se
[i
] : NULL
;
8521 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8530 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8532 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8533 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8536 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8538 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8540 #else /* !CONFG_FAIR_GROUP_SCHED */
8541 static inline void free_fair_sched_group(struct task_group
*tg
)
8546 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8551 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8555 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8558 #endif /* CONFIG_FAIR_GROUP_SCHED */
8560 #ifdef CONFIG_RT_GROUP_SCHED
8561 static void free_rt_sched_group(struct task_group
*tg
)
8565 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8567 for_each_possible_cpu(i
) {
8569 kfree(tg
->rt_rq
[i
]);
8571 kfree(tg
->rt_se
[i
]);
8579 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8581 struct rt_rq
*rt_rq
;
8582 struct sched_rt_entity
*rt_se
, *parent_se
;
8586 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8589 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8593 init_rt_bandwidth(&tg
->rt_bandwidth
,
8594 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8596 for_each_possible_cpu(i
) {
8599 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8600 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8604 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8605 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8609 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8610 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8619 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8621 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8622 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8625 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8627 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8629 #else /* !CONFIG_RT_GROUP_SCHED */
8630 static inline void free_rt_sched_group(struct task_group
*tg
)
8635 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8640 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8644 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8647 #endif /* CONFIG_RT_GROUP_SCHED */
8649 #ifdef CONFIG_GROUP_SCHED
8650 static void free_sched_group(struct task_group
*tg
)
8652 free_fair_sched_group(tg
);
8653 free_rt_sched_group(tg
);
8657 /* allocate runqueue etc for a new task group */
8658 struct task_group
*sched_create_group(struct task_group
*parent
)
8660 struct task_group
*tg
;
8661 unsigned long flags
;
8664 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8666 return ERR_PTR(-ENOMEM
);
8668 if (!alloc_fair_sched_group(tg
, parent
))
8671 if (!alloc_rt_sched_group(tg
, parent
))
8674 spin_lock_irqsave(&task_group_lock
, flags
);
8675 for_each_possible_cpu(i
) {
8676 register_fair_sched_group(tg
, i
);
8677 register_rt_sched_group(tg
, i
);
8679 list_add_rcu(&tg
->list
, &task_groups
);
8681 WARN_ON(!parent
); /* root should already exist */
8683 tg
->parent
= parent
;
8684 INIT_LIST_HEAD(&tg
->children
);
8685 list_add_rcu(&tg
->siblings
, &parent
->children
);
8686 spin_unlock_irqrestore(&task_group_lock
, flags
);
8691 free_sched_group(tg
);
8692 return ERR_PTR(-ENOMEM
);
8695 /* rcu callback to free various structures associated with a task group */
8696 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8698 /* now it should be safe to free those cfs_rqs */
8699 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8702 /* Destroy runqueue etc associated with a task group */
8703 void sched_destroy_group(struct task_group
*tg
)
8705 unsigned long flags
;
8708 spin_lock_irqsave(&task_group_lock
, flags
);
8709 for_each_possible_cpu(i
) {
8710 unregister_fair_sched_group(tg
, i
);
8711 unregister_rt_sched_group(tg
, i
);
8713 list_del_rcu(&tg
->list
);
8714 list_del_rcu(&tg
->siblings
);
8715 spin_unlock_irqrestore(&task_group_lock
, flags
);
8717 /* wait for possible concurrent references to cfs_rqs complete */
8718 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8721 /* change task's runqueue when it moves between groups.
8722 * The caller of this function should have put the task in its new group
8723 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8724 * reflect its new group.
8726 void sched_move_task(struct task_struct
*tsk
)
8729 unsigned long flags
;
8732 rq
= task_rq_lock(tsk
, &flags
);
8734 update_rq_clock(rq
);
8736 running
= task_current(rq
, tsk
);
8737 on_rq
= tsk
->se
.on_rq
;
8740 dequeue_task(rq
, tsk
, 0);
8741 if (unlikely(running
))
8742 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8744 set_task_rq(tsk
, task_cpu(tsk
));
8746 #ifdef CONFIG_FAIR_GROUP_SCHED
8747 if (tsk
->sched_class
->moved_group
)
8748 tsk
->sched_class
->moved_group(tsk
);
8751 if (unlikely(running
))
8752 tsk
->sched_class
->set_curr_task(rq
);
8754 enqueue_task(rq
, tsk
, 0);
8756 task_rq_unlock(rq
, &flags
);
8758 #endif /* CONFIG_GROUP_SCHED */
8760 #ifdef CONFIG_FAIR_GROUP_SCHED
8761 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8763 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8768 dequeue_entity(cfs_rq
, se
, 0);
8770 se
->load
.weight
= shares
;
8771 se
->load
.inv_weight
= 0;
8774 enqueue_entity(cfs_rq
, se
, 0);
8777 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8779 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8780 struct rq
*rq
= cfs_rq
->rq
;
8781 unsigned long flags
;
8783 spin_lock_irqsave(&rq
->lock
, flags
);
8784 __set_se_shares(se
, shares
);
8785 spin_unlock_irqrestore(&rq
->lock
, flags
);
8788 static DEFINE_MUTEX(shares_mutex
);
8790 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8793 unsigned long flags
;
8796 * We can't change the weight of the root cgroup.
8801 if (shares
< MIN_SHARES
)
8802 shares
= MIN_SHARES
;
8803 else if (shares
> MAX_SHARES
)
8804 shares
= MAX_SHARES
;
8806 mutex_lock(&shares_mutex
);
8807 if (tg
->shares
== shares
)
8810 spin_lock_irqsave(&task_group_lock
, flags
);
8811 for_each_possible_cpu(i
)
8812 unregister_fair_sched_group(tg
, i
);
8813 list_del_rcu(&tg
->siblings
);
8814 spin_unlock_irqrestore(&task_group_lock
, flags
);
8816 /* wait for any ongoing reference to this group to finish */
8817 synchronize_sched();
8820 * Now we are free to modify the group's share on each cpu
8821 * w/o tripping rebalance_share or load_balance_fair.
8823 tg
->shares
= shares
;
8824 for_each_possible_cpu(i
) {
8828 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8829 set_se_shares(tg
->se
[i
], shares
);
8833 * Enable load balance activity on this group, by inserting it back on
8834 * each cpu's rq->leaf_cfs_rq_list.
8836 spin_lock_irqsave(&task_group_lock
, flags
);
8837 for_each_possible_cpu(i
)
8838 register_fair_sched_group(tg
, i
);
8839 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8840 spin_unlock_irqrestore(&task_group_lock
, flags
);
8842 mutex_unlock(&shares_mutex
);
8846 unsigned long sched_group_shares(struct task_group
*tg
)
8852 #ifdef CONFIG_RT_GROUP_SCHED
8854 * Ensure that the real time constraints are schedulable.
8856 static DEFINE_MUTEX(rt_constraints_mutex
);
8858 static unsigned long to_ratio(u64 period
, u64 runtime
)
8860 if (runtime
== RUNTIME_INF
)
8863 return div64_u64(runtime
<< 20, period
);
8866 /* Must be called with tasklist_lock held */
8867 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8869 struct task_struct
*g
, *p
;
8871 do_each_thread(g
, p
) {
8872 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8874 } while_each_thread(g
, p
);
8879 struct rt_schedulable_data
{
8880 struct task_group
*tg
;
8885 static int tg_schedulable(struct task_group
*tg
, void *data
)
8887 struct rt_schedulable_data
*d
= data
;
8888 struct task_group
*child
;
8889 unsigned long total
, sum
= 0;
8890 u64 period
, runtime
;
8892 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8893 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8896 period
= d
->rt_period
;
8897 runtime
= d
->rt_runtime
;
8901 * Cannot have more runtime than the period.
8903 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8907 * Ensure we don't starve existing RT tasks.
8909 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8912 total
= to_ratio(period
, runtime
);
8915 * Nobody can have more than the global setting allows.
8917 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8921 * The sum of our children's runtime should not exceed our own.
8923 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8924 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8925 runtime
= child
->rt_bandwidth
.rt_runtime
;
8927 if (child
== d
->tg
) {
8928 period
= d
->rt_period
;
8929 runtime
= d
->rt_runtime
;
8932 sum
+= to_ratio(period
, runtime
);
8941 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8943 struct rt_schedulable_data data
= {
8945 .rt_period
= period
,
8946 .rt_runtime
= runtime
,
8949 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8952 static int tg_set_bandwidth(struct task_group
*tg
,
8953 u64 rt_period
, u64 rt_runtime
)
8957 mutex_lock(&rt_constraints_mutex
);
8958 read_lock(&tasklist_lock
);
8959 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8963 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8964 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8965 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8967 for_each_possible_cpu(i
) {
8968 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8970 spin_lock(&rt_rq
->rt_runtime_lock
);
8971 rt_rq
->rt_runtime
= rt_runtime
;
8972 spin_unlock(&rt_rq
->rt_runtime_lock
);
8974 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8976 read_unlock(&tasklist_lock
);
8977 mutex_unlock(&rt_constraints_mutex
);
8982 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8984 u64 rt_runtime
, rt_period
;
8986 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8987 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8988 if (rt_runtime_us
< 0)
8989 rt_runtime
= RUNTIME_INF
;
8991 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8994 long sched_group_rt_runtime(struct task_group
*tg
)
8998 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9001 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9002 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9003 return rt_runtime_us
;
9006 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9008 u64 rt_runtime
, rt_period
;
9010 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9011 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9016 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9019 long sched_group_rt_period(struct task_group
*tg
)
9023 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9024 do_div(rt_period_us
, NSEC_PER_USEC
);
9025 return rt_period_us
;
9028 static int sched_rt_global_constraints(void)
9030 u64 runtime
, period
;
9033 if (sysctl_sched_rt_period
<= 0)
9036 runtime
= global_rt_runtime();
9037 period
= global_rt_period();
9040 * Sanity check on the sysctl variables.
9042 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9045 mutex_lock(&rt_constraints_mutex
);
9046 read_lock(&tasklist_lock
);
9047 ret
= __rt_schedulable(NULL
, 0, 0);
9048 read_unlock(&tasklist_lock
);
9049 mutex_unlock(&rt_constraints_mutex
);
9053 #else /* !CONFIG_RT_GROUP_SCHED */
9054 static int sched_rt_global_constraints(void)
9056 unsigned long flags
;
9059 if (sysctl_sched_rt_period
<= 0)
9062 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9063 for_each_possible_cpu(i
) {
9064 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9066 spin_lock(&rt_rq
->rt_runtime_lock
);
9067 rt_rq
->rt_runtime
= global_rt_runtime();
9068 spin_unlock(&rt_rq
->rt_runtime_lock
);
9070 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9074 #endif /* CONFIG_RT_GROUP_SCHED */
9076 int sched_rt_handler(struct ctl_table
*table
, int write
,
9077 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9081 int old_period
, old_runtime
;
9082 static DEFINE_MUTEX(mutex
);
9085 old_period
= sysctl_sched_rt_period
;
9086 old_runtime
= sysctl_sched_rt_runtime
;
9088 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9090 if (!ret
&& write
) {
9091 ret
= sched_rt_global_constraints();
9093 sysctl_sched_rt_period
= old_period
;
9094 sysctl_sched_rt_runtime
= old_runtime
;
9096 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9097 def_rt_bandwidth
.rt_period
=
9098 ns_to_ktime(global_rt_period());
9101 mutex_unlock(&mutex
);
9106 #ifdef CONFIG_CGROUP_SCHED
9108 /* return corresponding task_group object of a cgroup */
9109 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9111 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9112 struct task_group
, css
);
9115 static struct cgroup_subsys_state
*
9116 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9118 struct task_group
*tg
, *parent
;
9120 if (!cgrp
->parent
) {
9121 /* This is early initialization for the top cgroup */
9122 return &init_task_group
.css
;
9125 parent
= cgroup_tg(cgrp
->parent
);
9126 tg
= sched_create_group(parent
);
9128 return ERR_PTR(-ENOMEM
);
9134 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9136 struct task_group
*tg
= cgroup_tg(cgrp
);
9138 sched_destroy_group(tg
);
9142 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9143 struct task_struct
*tsk
)
9145 #ifdef CONFIG_RT_GROUP_SCHED
9146 /* Don't accept realtime tasks when there is no way for them to run */
9147 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9150 /* We don't support RT-tasks being in separate groups */
9151 if (tsk
->sched_class
!= &fair_sched_class
)
9159 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9160 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9162 sched_move_task(tsk
);
9165 #ifdef CONFIG_FAIR_GROUP_SCHED
9166 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9169 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9172 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9174 struct task_group
*tg
= cgroup_tg(cgrp
);
9176 return (u64
) tg
->shares
;
9178 #endif /* CONFIG_FAIR_GROUP_SCHED */
9180 #ifdef CONFIG_RT_GROUP_SCHED
9181 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9184 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9187 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9189 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9192 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9195 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9198 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9200 return sched_group_rt_period(cgroup_tg(cgrp
));
9202 #endif /* CONFIG_RT_GROUP_SCHED */
9204 static struct cftype cpu_files
[] = {
9205 #ifdef CONFIG_FAIR_GROUP_SCHED
9208 .read_u64
= cpu_shares_read_u64
,
9209 .write_u64
= cpu_shares_write_u64
,
9212 #ifdef CONFIG_RT_GROUP_SCHED
9214 .name
= "rt_runtime_us",
9215 .read_s64
= cpu_rt_runtime_read
,
9216 .write_s64
= cpu_rt_runtime_write
,
9219 .name
= "rt_period_us",
9220 .read_u64
= cpu_rt_period_read_uint
,
9221 .write_u64
= cpu_rt_period_write_uint
,
9226 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9228 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9231 struct cgroup_subsys cpu_cgroup_subsys
= {
9233 .create
= cpu_cgroup_create
,
9234 .destroy
= cpu_cgroup_destroy
,
9235 .can_attach
= cpu_cgroup_can_attach
,
9236 .attach
= cpu_cgroup_attach
,
9237 .populate
= cpu_cgroup_populate
,
9238 .subsys_id
= cpu_cgroup_subsys_id
,
9242 #endif /* CONFIG_CGROUP_SCHED */
9244 #ifdef CONFIG_CGROUP_CPUACCT
9247 * CPU accounting code for task groups.
9249 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9250 * (balbir@in.ibm.com).
9253 /* track cpu usage of a group of tasks */
9255 struct cgroup_subsys_state css
;
9256 /* cpuusage holds pointer to a u64-type object on every cpu */
9260 struct cgroup_subsys cpuacct_subsys
;
9262 /* return cpu accounting group corresponding to this container */
9263 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9265 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9266 struct cpuacct
, css
);
9269 /* return cpu accounting group to which this task belongs */
9270 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9272 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9273 struct cpuacct
, css
);
9276 /* create a new cpu accounting group */
9277 static struct cgroup_subsys_state
*cpuacct_create(
9278 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9280 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9283 return ERR_PTR(-ENOMEM
);
9285 ca
->cpuusage
= alloc_percpu(u64
);
9286 if (!ca
->cpuusage
) {
9288 return ERR_PTR(-ENOMEM
);
9294 /* destroy an existing cpu accounting group */
9296 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9298 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9300 free_percpu(ca
->cpuusage
);
9304 /* return total cpu usage (in nanoseconds) of a group */
9305 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9307 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9308 u64 totalcpuusage
= 0;
9311 for_each_possible_cpu(i
) {
9312 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9315 * Take rq->lock to make 64-bit addition safe on 32-bit
9318 spin_lock_irq(&cpu_rq(i
)->lock
);
9319 totalcpuusage
+= *cpuusage
;
9320 spin_unlock_irq(&cpu_rq(i
)->lock
);
9323 return totalcpuusage
;
9326 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9329 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9338 for_each_possible_cpu(i
) {
9339 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9341 spin_lock_irq(&cpu_rq(i
)->lock
);
9343 spin_unlock_irq(&cpu_rq(i
)->lock
);
9349 static struct cftype files
[] = {
9352 .read_u64
= cpuusage_read
,
9353 .write_u64
= cpuusage_write
,
9357 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9359 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9363 * charge this task's execution time to its accounting group.
9365 * called with rq->lock held.
9367 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9371 if (!cpuacct_subsys
.active
)
9376 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9378 *cpuusage
+= cputime
;
9382 struct cgroup_subsys cpuacct_subsys
= {
9384 .create
= cpuacct_create
,
9385 .destroy
= cpuacct_destroy
,
9386 .populate
= cpuacct_populate
,
9387 .subsys_id
= cpuacct_subsys_id
,
9389 #endif /* CONFIG_CGROUP_CPUACCT */