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/seq_file.h>
59 #include <linux/sysctl.h>
60 #include <linux/syscalls.h>
61 #include <linux/times.h>
62 #include <linux/tsacct_kern.h>
63 #include <linux/kprobes.h>
64 #include <linux/delayacct.h>
65 #include <linux/reciprocal_div.h>
66 #include <linux/unistd.h>
67 #include <linux/pagemap.h>
68 #include <linux/hrtimer.h>
69 #include <linux/tick.h>
70 #include <linux/bootmem.h>
71 #include <linux/debugfs.h>
72 #include <linux/ctype.h>
75 #include <asm/irq_regs.h>
78 * Scheduler clock - returns current time in nanosec units.
79 * This is default implementation.
80 * Architectures and sub-architectures can override this.
82 unsigned long long __attribute__((weak
)) sched_clock(void)
84 return (unsigned long long)jiffies
* (NSEC_PER_SEC
/ HZ
);
88 * Convert user-nice values [ -20 ... 0 ... 19 ]
89 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
92 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
93 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
94 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
97 * 'User priority' is the nice value converted to something we
98 * can work with better when scaling various scheduler parameters,
99 * it's a [ 0 ... 39 ] range.
101 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
102 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
103 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
106 * Helpers for converting nanosecond timing to jiffy resolution
108 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
110 #define NICE_0_LOAD SCHED_LOAD_SCALE
111 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
114 * These are the 'tuning knobs' of the scheduler:
116 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
117 * Timeslices get refilled after they expire.
119 #define DEF_TIMESLICE (100 * HZ / 1000)
122 * single value that denotes runtime == period, ie unlimited time.
124 #define RUNTIME_INF ((u64)~0ULL)
128 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
129 * Since cpu_power is a 'constant', we can use a reciprocal divide.
131 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
133 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
137 * Each time a sched group cpu_power is changed,
138 * we must compute its reciprocal value
140 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
142 sg
->__cpu_power
+= val
;
143 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
147 static inline int rt_policy(int policy
)
149 if (unlikely(policy
== SCHED_FIFO
) || unlikely(policy
== SCHED_RR
))
154 static inline int task_has_rt_policy(struct task_struct
*p
)
156 return rt_policy(p
->policy
);
160 * This is the priority-queue data structure of the RT scheduling class:
162 struct rt_prio_array
{
163 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
164 struct list_head queue
[MAX_RT_PRIO
];
167 struct rt_bandwidth
{
168 /* nests inside the rq lock: */
169 spinlock_t rt_runtime_lock
;
172 struct hrtimer rt_period_timer
;
175 static struct rt_bandwidth def_rt_bandwidth
;
177 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
179 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
181 struct rt_bandwidth
*rt_b
=
182 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
188 now
= hrtimer_cb_get_time(timer
);
189 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
194 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
197 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
201 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
203 rt_b
->rt_period
= ns_to_ktime(period
);
204 rt_b
->rt_runtime
= runtime
;
206 spin_lock_init(&rt_b
->rt_runtime_lock
);
208 hrtimer_init(&rt_b
->rt_period_timer
,
209 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
210 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
211 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
214 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
218 if (rt_b
->rt_runtime
== RUNTIME_INF
)
221 if (hrtimer_active(&rt_b
->rt_period_timer
))
224 spin_lock(&rt_b
->rt_runtime_lock
);
226 if (hrtimer_active(&rt_b
->rt_period_timer
))
229 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
230 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
231 hrtimer_start(&rt_b
->rt_period_timer
,
232 rt_b
->rt_period_timer
.expires
,
235 spin_unlock(&rt_b
->rt_runtime_lock
);
238 #ifdef CONFIG_RT_GROUP_SCHED
239 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
241 hrtimer_cancel(&rt_b
->rt_period_timer
);
246 * sched_domains_mutex serializes calls to arch_init_sched_domains,
247 * detach_destroy_domains and partition_sched_domains.
249 static DEFINE_MUTEX(sched_domains_mutex
);
251 #ifdef CONFIG_GROUP_SCHED
253 #include <linux/cgroup.h>
257 static LIST_HEAD(task_groups
);
259 /* task group related information */
261 #ifdef CONFIG_CGROUP_SCHED
262 struct cgroup_subsys_state css
;
265 #ifdef CONFIG_FAIR_GROUP_SCHED
266 /* schedulable entities of this group on each cpu */
267 struct sched_entity
**se
;
268 /* runqueue "owned" by this group on each cpu */
269 struct cfs_rq
**cfs_rq
;
270 unsigned long shares
;
273 #ifdef CONFIG_RT_GROUP_SCHED
274 struct sched_rt_entity
**rt_se
;
275 struct rt_rq
**rt_rq
;
277 struct rt_bandwidth rt_bandwidth
;
281 struct list_head list
;
283 struct task_group
*parent
;
284 struct list_head siblings
;
285 struct list_head children
;
288 #ifdef CONFIG_USER_SCHED
292 * Every UID task group (including init_task_group aka UID-0) will
293 * be a child to this group.
295 struct task_group root_task_group
;
297 #ifdef CONFIG_FAIR_GROUP_SCHED
298 /* Default task group's sched entity on each cpu */
299 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
300 /* Default task group's cfs_rq on each cpu */
301 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
304 #ifdef CONFIG_RT_GROUP_SCHED
305 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
306 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
309 #define root_task_group init_task_group
312 /* task_group_lock serializes add/remove of task groups and also changes to
313 * a task group's cpu shares.
315 static DEFINE_SPINLOCK(task_group_lock
);
317 #ifdef CONFIG_FAIR_GROUP_SCHED
318 #ifdef CONFIG_USER_SCHED
319 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
321 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
325 * A weight of 0, 1 or ULONG_MAX can cause arithmetics problems.
326 * (The default weight is 1024 - so there's no practical
327 * limitation from this.)
330 #define MAX_SHARES (ULONG_MAX - 1)
332 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
335 /* Default task group.
336 * Every task in system belong to this group at bootup.
338 struct task_group init_task_group
;
340 /* return group to which a task belongs */
341 static inline struct task_group
*task_group(struct task_struct
*p
)
343 struct task_group
*tg
;
345 #ifdef CONFIG_USER_SCHED
347 #elif defined(CONFIG_CGROUP_SCHED)
348 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
349 struct task_group
, css
);
351 tg
= &init_task_group
;
356 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
357 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
359 #ifdef CONFIG_FAIR_GROUP_SCHED
360 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
361 p
->se
.parent
= task_group(p
)->se
[cpu
];
364 #ifdef CONFIG_RT_GROUP_SCHED
365 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
366 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
372 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
374 #endif /* CONFIG_GROUP_SCHED */
376 /* CFS-related fields in a runqueue */
378 struct load_weight load
;
379 unsigned long nr_running
;
384 struct rb_root tasks_timeline
;
385 struct rb_node
*rb_leftmost
;
387 struct list_head tasks
;
388 struct list_head
*balance_iterator
;
391 * 'curr' points to currently running entity on this cfs_rq.
392 * It is set to NULL otherwise (i.e when none are currently running).
394 struct sched_entity
*curr
, *next
;
396 unsigned long nr_spread_over
;
398 #ifdef CONFIG_FAIR_GROUP_SCHED
399 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
402 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
403 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
404 * (like users, containers etc.)
406 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
407 * list is used during load balance.
409 struct list_head leaf_cfs_rq_list
;
410 struct task_group
*tg
; /* group that "owns" this runqueue */
413 unsigned long task_weight
;
414 unsigned long shares
;
416 * We need space to build a sched_domain wide view of the full task
417 * group tree, in order to avoid depending on dynamic memory allocation
418 * during the load balancing we place this in the per cpu task group
419 * hierarchy. This limits the load balancing to one instance per cpu,
420 * but more should not be needed anyway.
422 struct aggregate_struct
{
424 * load = weight(cpus) * f(tg)
426 * Where f(tg) is the recursive weight fraction assigned to
432 * part of the group weight distributed to this span.
434 unsigned long shares
;
437 * The sum of all runqueue weights within this span.
439 unsigned long rq_weight
;
442 * Weight contributed by tasks; this is the part we can
443 * influence by moving tasks around.
445 unsigned long task_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.
502 * By default the system creates a single root-domain with all cpus as
503 * members (mimicking the global state we have today).
505 static struct root_domain def_root_domain
;
510 * This is the main, per-CPU runqueue data structure.
512 * Locking rule: those places that want to lock multiple runqueues
513 * (such as the load balancing or the thread migration code), lock
514 * acquire operations must be ordered by ascending &runqueue.
521 * nr_running and cpu_load should be in the same cacheline because
522 * remote CPUs use both these fields when doing load calculation.
524 unsigned long nr_running
;
525 #define CPU_LOAD_IDX_MAX 5
526 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
527 unsigned char idle_at_tick
;
529 unsigned long last_tick_seen
;
530 unsigned char in_nohz_recently
;
532 /* capture load from *all* tasks on this cpu: */
533 struct load_weight load
;
534 unsigned long nr_load_updates
;
540 #ifdef CONFIG_FAIR_GROUP_SCHED
541 /* list of leaf cfs_rq on this cpu: */
542 struct list_head leaf_cfs_rq_list
;
544 #ifdef CONFIG_RT_GROUP_SCHED
545 struct list_head leaf_rt_rq_list
;
549 * This is part of a global counter where only the total sum
550 * over all CPUs matters. A task can increase this counter on
551 * one CPU and if it got migrated afterwards it may decrease
552 * it on another CPU. Always updated under the runqueue lock:
554 unsigned long nr_uninterruptible
;
556 struct task_struct
*curr
, *idle
;
557 unsigned long next_balance
;
558 struct mm_struct
*prev_mm
;
560 u64 clock
, prev_clock_raw
;
563 unsigned int clock_warps
, clock_overflows
, clock_underflows
;
565 unsigned int clock_deep_idle_events
;
571 struct root_domain
*rd
;
572 struct sched_domain
*sd
;
574 /* For active balancing */
577 /* cpu of this runqueue: */
580 struct task_struct
*migration_thread
;
581 struct list_head migration_queue
;
584 #ifdef CONFIG_SCHED_HRTICK
585 unsigned long hrtick_flags
;
586 ktime_t hrtick_expire
;
587 struct hrtimer hrtick_timer
;
590 #ifdef CONFIG_SCHEDSTATS
592 struct sched_info rq_sched_info
;
594 /* sys_sched_yield() stats */
595 unsigned int yld_exp_empty
;
596 unsigned int yld_act_empty
;
597 unsigned int yld_both_empty
;
598 unsigned int yld_count
;
600 /* schedule() stats */
601 unsigned int sched_switch
;
602 unsigned int sched_count
;
603 unsigned int sched_goidle
;
605 /* try_to_wake_up() stats */
606 unsigned int ttwu_count
;
607 unsigned int ttwu_local
;
610 unsigned int bkl_count
;
612 struct lock_class_key rq_lock_key
;
615 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
617 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
619 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
622 static inline int cpu_of(struct rq
*rq
)
632 static inline bool nohz_on(int cpu
)
634 return tick_get_tick_sched(cpu
)->nohz_mode
!= NOHZ_MODE_INACTIVE
;
637 static inline u64
max_skipped_ticks(struct rq
*rq
)
639 return nohz_on(cpu_of(rq
)) ? jiffies
- rq
->last_tick_seen
+ 2 : 1;
642 static inline void update_last_tick_seen(struct rq
*rq
)
644 rq
->last_tick_seen
= jiffies
;
647 static inline u64
max_skipped_ticks(struct rq
*rq
)
652 static inline void update_last_tick_seen(struct rq
*rq
)
658 * Update the per-runqueue clock, as finegrained as the platform can give
659 * us, but without assuming monotonicity, etc.:
661 static void __update_rq_clock(struct rq
*rq
)
663 u64 prev_raw
= rq
->prev_clock_raw
;
664 u64 now
= sched_clock();
665 s64 delta
= now
- prev_raw
;
666 u64 clock
= rq
->clock
;
668 #ifdef CONFIG_SCHED_DEBUG
669 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
672 * Protect against sched_clock() occasionally going backwards:
674 if (unlikely(delta
< 0)) {
679 * Catch too large forward jumps too:
681 u64 max_jump
= max_skipped_ticks(rq
) * TICK_NSEC
;
682 u64 max_time
= rq
->tick_timestamp
+ max_jump
;
684 if (unlikely(clock
+ delta
> max_time
)) {
685 if (clock
< max_time
)
689 rq
->clock_overflows
++;
691 if (unlikely(delta
> rq
->clock_max_delta
))
692 rq
->clock_max_delta
= delta
;
697 rq
->prev_clock_raw
= now
;
701 static void update_rq_clock(struct rq
*rq
)
703 if (likely(smp_processor_id() == cpu_of(rq
)))
704 __update_rq_clock(rq
);
708 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
709 * See detach_destroy_domains: synchronize_sched for details.
711 * The domain tree of any CPU may only be accessed from within
712 * preempt-disabled sections.
714 #define for_each_domain(cpu, __sd) \
715 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
717 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
718 #define this_rq() (&__get_cpu_var(runqueues))
719 #define task_rq(p) cpu_rq(task_cpu(p))
720 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
723 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
725 #ifdef CONFIG_SCHED_DEBUG
726 # define const_debug __read_mostly
728 # define const_debug static const
732 * Debugging: various feature bits
735 #define SCHED_FEAT(name, enabled) \
736 __SCHED_FEAT_##name ,
739 #include "sched_features.h"
744 #define SCHED_FEAT(name, enabled) \
745 (1UL << __SCHED_FEAT_##name) * enabled |
747 const_debug
unsigned int sysctl_sched_features
=
748 #include "sched_features.h"
753 #ifdef CONFIG_SCHED_DEBUG
754 #define SCHED_FEAT(name, enabled) \
757 static __read_mostly
char *sched_feat_names
[] = {
758 #include "sched_features.h"
764 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
766 filp
->private_data
= inode
->i_private
;
771 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
772 size_t cnt
, loff_t
*ppos
)
779 for (i
= 0; sched_feat_names
[i
]; i
++) {
780 len
+= strlen(sched_feat_names
[i
]);
784 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
788 for (i
= 0; sched_feat_names
[i
]; i
++) {
789 if (sysctl_sched_features
& (1UL << i
))
790 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
792 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
795 r
+= sprintf(buf
+ r
, "\n");
796 WARN_ON(r
>= len
+ 2);
798 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
806 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
807 size_t cnt
, loff_t
*ppos
)
817 if (copy_from_user(&buf
, ubuf
, cnt
))
822 if (strncmp(buf
, "NO_", 3) == 0) {
827 for (i
= 0; sched_feat_names
[i
]; i
++) {
828 int len
= strlen(sched_feat_names
[i
]);
830 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
832 sysctl_sched_features
&= ~(1UL << i
);
834 sysctl_sched_features
|= (1UL << i
);
839 if (!sched_feat_names
[i
])
847 static struct file_operations sched_feat_fops
= {
848 .open
= sched_feat_open
,
849 .read
= sched_feat_read
,
850 .write
= sched_feat_write
,
853 static __init
int sched_init_debug(void)
855 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
860 late_initcall(sched_init_debug
);
864 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
867 * Number of tasks to iterate in a single balance run.
868 * Limited because this is done with IRQs disabled.
870 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
873 * period over which we measure -rt task cpu usage in us.
876 unsigned int sysctl_sched_rt_period
= 1000000;
878 static __read_mostly
int scheduler_running
;
881 * part of the period that we allow rt tasks to run in us.
884 int sysctl_sched_rt_runtime
= 950000;
886 static inline u64
global_rt_period(void)
888 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
891 static inline u64
global_rt_runtime(void)
893 if (sysctl_sched_rt_period
< 0)
896 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
899 unsigned long long time_sync_thresh
= 100000;
901 static DEFINE_PER_CPU(unsigned long long, time_offset
);
902 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time
);
905 * Global lock which we take every now and then to synchronize
906 * the CPUs time. This method is not warp-safe, but it's good
907 * enough to synchronize slowly diverging time sources and thus
908 * it's good enough for tracing:
910 static DEFINE_SPINLOCK(time_sync_lock
);
911 static unsigned long long prev_global_time
;
913 static unsigned long long __sync_cpu_clock(cycles_t time
, int cpu
)
917 spin_lock_irqsave(&time_sync_lock
, flags
);
919 if (time
< prev_global_time
) {
920 per_cpu(time_offset
, cpu
) += prev_global_time
- time
;
921 time
= prev_global_time
;
923 prev_global_time
= time
;
926 spin_unlock_irqrestore(&time_sync_lock
, flags
);
931 static unsigned long long __cpu_clock(int cpu
)
933 unsigned long long now
;
938 * Only call sched_clock() if the scheduler has already been
939 * initialized (some code might call cpu_clock() very early):
941 if (unlikely(!scheduler_running
))
944 local_irq_save(flags
);
948 local_irq_restore(flags
);
954 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
955 * clock constructed from sched_clock():
957 unsigned long long cpu_clock(int cpu
)
959 unsigned long long prev_cpu_time
, time
, delta_time
;
961 prev_cpu_time
= per_cpu(prev_cpu_time
, cpu
);
962 time
= __cpu_clock(cpu
) + per_cpu(time_offset
, cpu
);
963 delta_time
= time
-prev_cpu_time
;
965 if (unlikely(delta_time
> time_sync_thresh
))
966 time
= __sync_cpu_clock(time
, cpu
);
970 EXPORT_SYMBOL_GPL(cpu_clock
);
972 #ifndef prepare_arch_switch
973 # define prepare_arch_switch(next) do { } while (0)
975 #ifndef finish_arch_switch
976 # define finish_arch_switch(prev) do { } while (0)
979 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
981 return rq
->curr
== p
;
984 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
985 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
987 return task_current(rq
, p
);
990 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
994 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
996 #ifdef CONFIG_DEBUG_SPINLOCK
997 /* this is a valid case when another task releases the spinlock */
998 rq
->lock
.owner
= current
;
1001 * If we are tracking spinlock dependencies then we have to
1002 * fix up the runqueue lock - which gets 'carried over' from
1003 * prev into current:
1005 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
1007 spin_unlock_irq(&rq
->lock
);
1010 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1011 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
1016 return task_current(rq
, p
);
1020 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1024 * We can optimise this out completely for !SMP, because the
1025 * SMP rebalancing from interrupt is the only thing that cares
1030 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1031 spin_unlock_irq(&rq
->lock
);
1033 spin_unlock(&rq
->lock
);
1037 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1041 * After ->oncpu is cleared, the task can be moved to a different CPU.
1042 * We must ensure this doesn't happen until the switch is completely
1048 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1052 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1055 * __task_rq_lock - lock the runqueue a given task resides on.
1056 * Must be called interrupts disabled.
1058 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
1059 __acquires(rq
->lock
)
1062 struct rq
*rq
= task_rq(p
);
1063 spin_lock(&rq
->lock
);
1064 if (likely(rq
== task_rq(p
)))
1066 spin_unlock(&rq
->lock
);
1071 * task_rq_lock - lock the runqueue a given task resides on and disable
1072 * interrupts. Note the ordering: we can safely lookup the task_rq without
1073 * explicitly disabling preemption.
1075 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1076 __acquires(rq
->lock
)
1081 local_irq_save(*flags
);
1083 spin_lock(&rq
->lock
);
1084 if (likely(rq
== task_rq(p
)))
1086 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1090 static void __task_rq_unlock(struct rq
*rq
)
1091 __releases(rq
->lock
)
1093 spin_unlock(&rq
->lock
);
1096 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1097 __releases(rq
->lock
)
1099 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1103 * this_rq_lock - lock this runqueue and disable interrupts.
1105 static struct rq
*this_rq_lock(void)
1106 __acquires(rq
->lock
)
1110 local_irq_disable();
1112 spin_lock(&rq
->lock
);
1118 * We are going deep-idle (irqs are disabled):
1120 void sched_clock_idle_sleep_event(void)
1122 struct rq
*rq
= cpu_rq(smp_processor_id());
1124 WARN_ON(!irqs_disabled());
1125 spin_lock(&rq
->lock
);
1126 __update_rq_clock(rq
);
1127 spin_unlock(&rq
->lock
);
1128 rq
->clock_deep_idle_events
++;
1130 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
1133 * We just idled delta nanoseconds (called with irqs disabled):
1135 void sched_clock_idle_wakeup_event(u64 delta_ns
)
1137 struct rq
*rq
= cpu_rq(smp_processor_id());
1138 u64 now
= sched_clock();
1140 WARN_ON(!irqs_disabled());
1141 rq
->idle_clock
+= delta_ns
;
1143 * Override the previous timestamp and ignore all
1144 * sched_clock() deltas that occured while we idled,
1145 * and use the PM-provided delta_ns to advance the
1148 spin_lock(&rq
->lock
);
1149 rq
->prev_clock_raw
= now
;
1150 rq
->clock
+= delta_ns
;
1151 spin_unlock(&rq
->lock
);
1152 touch_softlockup_watchdog();
1154 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
1156 static void __resched_task(struct task_struct
*p
, int tif_bit
);
1158 static inline void resched_task(struct task_struct
*p
)
1160 __resched_task(p
, TIF_NEED_RESCHED
);
1163 #ifdef CONFIG_SCHED_HRTICK
1165 * Use HR-timers to deliver accurate preemption points.
1167 * Its all a bit involved since we cannot program an hrt while holding the
1168 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1171 * When we get rescheduled we reprogram the hrtick_timer outside of the
1174 static inline void resched_hrt(struct task_struct
*p
)
1176 __resched_task(p
, TIF_HRTICK_RESCHED
);
1179 static inline void resched_rq(struct rq
*rq
)
1181 unsigned long flags
;
1183 spin_lock_irqsave(&rq
->lock
, flags
);
1184 resched_task(rq
->curr
);
1185 spin_unlock_irqrestore(&rq
->lock
, flags
);
1189 HRTICK_SET
, /* re-programm hrtick_timer */
1190 HRTICK_RESET
, /* not a new slice */
1191 HRTICK_BLOCK
, /* stop hrtick operations */
1196 * - enabled by features
1197 * - hrtimer is actually high res
1199 static inline int hrtick_enabled(struct rq
*rq
)
1201 if (!sched_feat(HRTICK
))
1203 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1205 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1209 * Called to set the hrtick timer state.
1211 * called with rq->lock held and irqs disabled
1213 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1215 assert_spin_locked(&rq
->lock
);
1218 * preempt at: now + delay
1221 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1223 * indicate we need to program the timer
1225 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1227 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1230 * New slices are called from the schedule path and don't need a
1231 * forced reschedule.
1234 resched_hrt(rq
->curr
);
1237 static void hrtick_clear(struct rq
*rq
)
1239 if (hrtimer_active(&rq
->hrtick_timer
))
1240 hrtimer_cancel(&rq
->hrtick_timer
);
1244 * Update the timer from the possible pending state.
1246 static void hrtick_set(struct rq
*rq
)
1250 unsigned long flags
;
1252 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1254 spin_lock_irqsave(&rq
->lock
, flags
);
1255 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1256 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1257 time
= rq
->hrtick_expire
;
1258 clear_thread_flag(TIF_HRTICK_RESCHED
);
1259 spin_unlock_irqrestore(&rq
->lock
, flags
);
1262 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1263 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1270 * High-resolution timer tick.
1271 * Runs from hardirq context with interrupts disabled.
1273 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1275 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1277 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1279 spin_lock(&rq
->lock
);
1280 __update_rq_clock(rq
);
1281 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1282 spin_unlock(&rq
->lock
);
1284 return HRTIMER_NORESTART
;
1287 static void hotplug_hrtick_disable(int cpu
)
1289 struct rq
*rq
= cpu_rq(cpu
);
1290 unsigned long flags
;
1292 spin_lock_irqsave(&rq
->lock
, flags
);
1293 rq
->hrtick_flags
= 0;
1294 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1295 spin_unlock_irqrestore(&rq
->lock
, flags
);
1300 static void hotplug_hrtick_enable(int cpu
)
1302 struct rq
*rq
= cpu_rq(cpu
);
1303 unsigned long flags
;
1305 spin_lock_irqsave(&rq
->lock
, flags
);
1306 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1307 spin_unlock_irqrestore(&rq
->lock
, flags
);
1311 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1313 int cpu
= (int)(long)hcpu
;
1316 case CPU_UP_CANCELED
:
1317 case CPU_UP_CANCELED_FROZEN
:
1318 case CPU_DOWN_PREPARE
:
1319 case CPU_DOWN_PREPARE_FROZEN
:
1321 case CPU_DEAD_FROZEN
:
1322 hotplug_hrtick_disable(cpu
);
1325 case CPU_UP_PREPARE
:
1326 case CPU_UP_PREPARE_FROZEN
:
1327 case CPU_DOWN_FAILED
:
1328 case CPU_DOWN_FAILED_FROZEN
:
1330 case CPU_ONLINE_FROZEN
:
1331 hotplug_hrtick_enable(cpu
);
1338 static void init_hrtick(void)
1340 hotcpu_notifier(hotplug_hrtick
, 0);
1343 static void init_rq_hrtick(struct rq
*rq
)
1345 rq
->hrtick_flags
= 0;
1346 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1347 rq
->hrtick_timer
.function
= hrtick
;
1348 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1351 void hrtick_resched(void)
1354 unsigned long flags
;
1356 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1359 local_irq_save(flags
);
1360 rq
= cpu_rq(smp_processor_id());
1362 local_irq_restore(flags
);
1365 static inline void hrtick_clear(struct rq
*rq
)
1369 static inline void hrtick_set(struct rq
*rq
)
1373 static inline void init_rq_hrtick(struct rq
*rq
)
1377 void hrtick_resched(void)
1381 static inline void init_hrtick(void)
1387 * resched_task - mark a task 'to be rescheduled now'.
1389 * On UP this means the setting of the need_resched flag, on SMP it
1390 * might also involve a cross-CPU call to trigger the scheduler on
1395 #ifndef tsk_is_polling
1396 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1399 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1403 assert_spin_locked(&task_rq(p
)->lock
);
1405 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1408 set_tsk_thread_flag(p
, tif_bit
);
1411 if (cpu
== smp_processor_id())
1414 /* NEED_RESCHED must be visible before we test polling */
1416 if (!tsk_is_polling(p
))
1417 smp_send_reschedule(cpu
);
1420 static void resched_cpu(int cpu
)
1422 struct rq
*rq
= cpu_rq(cpu
);
1423 unsigned long flags
;
1425 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1427 resched_task(cpu_curr(cpu
));
1428 spin_unlock_irqrestore(&rq
->lock
, flags
);
1433 * When add_timer_on() enqueues a timer into the timer wheel of an
1434 * idle CPU then this timer might expire before the next timer event
1435 * which is scheduled to wake up that CPU. In case of a completely
1436 * idle system the next event might even be infinite time into the
1437 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1438 * leaves the inner idle loop so the newly added timer is taken into
1439 * account when the CPU goes back to idle and evaluates the timer
1440 * wheel for the next timer event.
1442 void wake_up_idle_cpu(int cpu
)
1444 struct rq
*rq
= cpu_rq(cpu
);
1446 if (cpu
== smp_processor_id())
1450 * This is safe, as this function is called with the timer
1451 * wheel base lock of (cpu) held. When the CPU is on the way
1452 * to idle and has not yet set rq->curr to idle then it will
1453 * be serialized on the timer wheel base lock and take the new
1454 * timer into account automatically.
1456 if (rq
->curr
!= rq
->idle
)
1460 * We can set TIF_RESCHED on the idle task of the other CPU
1461 * lockless. The worst case is that the other CPU runs the
1462 * idle task through an additional NOOP schedule()
1464 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1466 /* NEED_RESCHED must be visible before we test polling */
1468 if (!tsk_is_polling(rq
->idle
))
1469 smp_send_reschedule(cpu
);
1474 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1476 assert_spin_locked(&task_rq(p
)->lock
);
1477 set_tsk_thread_flag(p
, tif_bit
);
1481 #if BITS_PER_LONG == 32
1482 # define WMULT_CONST (~0UL)
1484 # define WMULT_CONST (1UL << 32)
1487 #define WMULT_SHIFT 32
1490 * Shift right and round:
1492 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1495 * delta *= weight / lw
1497 static unsigned long
1498 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1499 struct load_weight
*lw
)
1503 if (!lw
->inv_weight
)
1504 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)/(lw
->weight
+1);
1506 tmp
= (u64
)delta_exec
* weight
;
1508 * Check whether we'd overflow the 64-bit multiplication:
1510 if (unlikely(tmp
> WMULT_CONST
))
1511 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1514 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1516 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1519 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1525 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1532 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1533 * of tasks with abnormal "nice" values across CPUs the contribution that
1534 * each task makes to its run queue's load is weighted according to its
1535 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1536 * scaled version of the new time slice allocation that they receive on time
1540 #define WEIGHT_IDLEPRIO 2
1541 #define WMULT_IDLEPRIO (1 << 31)
1544 * Nice levels are multiplicative, with a gentle 10% change for every
1545 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1546 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1547 * that remained on nice 0.
1549 * The "10% effect" is relative and cumulative: from _any_ nice level,
1550 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1551 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1552 * If a task goes up by ~10% and another task goes down by ~10% then
1553 * the relative distance between them is ~25%.)
1555 static const int prio_to_weight
[40] = {
1556 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1557 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1558 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1559 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1560 /* 0 */ 1024, 820, 655, 526, 423,
1561 /* 5 */ 335, 272, 215, 172, 137,
1562 /* 10 */ 110, 87, 70, 56, 45,
1563 /* 15 */ 36, 29, 23, 18, 15,
1567 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1569 * In cases where the weight does not change often, we can use the
1570 * precalculated inverse to speed up arithmetics by turning divisions
1571 * into multiplications:
1573 static const u32 prio_to_wmult
[40] = {
1574 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1575 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1576 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1577 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1578 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1579 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1580 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1581 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1584 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1587 * runqueue iterator, to support SMP load-balancing between different
1588 * scheduling classes, without having to expose their internal data
1589 * structures to the load-balancing proper:
1591 struct rq_iterator
{
1593 struct task_struct
*(*start
)(void *);
1594 struct task_struct
*(*next
)(void *);
1598 static unsigned long
1599 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1600 unsigned long max_load_move
, struct sched_domain
*sd
,
1601 enum cpu_idle_type idle
, int *all_pinned
,
1602 int *this_best_prio
, struct rq_iterator
*iterator
);
1605 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1606 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1607 struct rq_iterator
*iterator
);
1610 #ifdef CONFIG_CGROUP_CPUACCT
1611 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1613 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1616 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1618 update_load_add(&rq
->load
, load
);
1621 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1623 update_load_sub(&rq
->load
, load
);
1627 static unsigned long source_load(int cpu
, int type
);
1628 static unsigned long target_load(int cpu
, int type
);
1629 static unsigned long cpu_avg_load_per_task(int cpu
);
1630 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1632 #ifdef CONFIG_FAIR_GROUP_SCHED
1635 * Group load balancing.
1637 * We calculate a few balance domain wide aggregate numbers; load and weight.
1638 * Given the pictures below, and assuming each item has equal weight:
1649 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1650 * which equals 1/9-th of the total load.
1653 * The weight of this group on the selected cpus.
1656 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1660 * Part of the rq_weight contributed by tasks; all groups except B would
1664 static inline struct aggregate_struct
*
1665 aggregate(struct task_group
*tg
, struct sched_domain
*sd
)
1667 return &tg
->cfs_rq
[sd
->first_cpu
]->aggregate
;
1670 typedef void (*aggregate_func
)(struct task_group
*, struct sched_domain
*);
1673 * Iterate the full tree, calling @down when first entering a node and @up when
1674 * leaving it for the final time.
1677 void aggregate_walk_tree(aggregate_func down
, aggregate_func up
,
1678 struct sched_domain
*sd
)
1680 struct task_group
*parent
, *child
;
1683 parent
= &root_task_group
;
1685 (*down
)(parent
, sd
);
1686 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1696 parent
= parent
->parent
;
1703 * Calculate the aggregate runqueue weight.
1706 void aggregate_group_weight(struct task_group
*tg
, struct sched_domain
*sd
)
1708 unsigned long rq_weight
= 0;
1709 unsigned long task_weight
= 0;
1712 for_each_cpu_mask(i
, sd
->span
) {
1713 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1714 task_weight
+= tg
->cfs_rq
[i
]->task_weight
;
1717 aggregate(tg
, sd
)->rq_weight
= rq_weight
;
1718 aggregate(tg
, sd
)->task_weight
= task_weight
;
1722 * Compute the weight of this group on the given cpus.
1725 void aggregate_group_shares(struct task_group
*tg
, struct sched_domain
*sd
)
1727 unsigned long shares
= 0;
1730 for_each_cpu_mask(i
, sd
->span
)
1731 shares
+= tg
->cfs_rq
[i
]->shares
;
1733 if ((!shares
&& aggregate(tg
, sd
)->rq_weight
) || shares
> tg
->shares
)
1734 shares
= tg
->shares
;
1736 aggregate(tg
, sd
)->shares
= shares
;
1740 * Compute the load fraction assigned to this group, relies on the aggregate
1741 * weight and this group's parent's load, i.e. top-down.
1744 void aggregate_group_load(struct task_group
*tg
, struct sched_domain
*sd
)
1752 for_each_cpu_mask(i
, sd
->span
)
1753 load
+= cpu_rq(i
)->load
.weight
;
1756 load
= aggregate(tg
->parent
, sd
)->load
;
1759 * shares is our weight in the parent's rq so
1760 * shares/parent->rq_weight gives our fraction of the load
1762 load
*= aggregate(tg
, sd
)->shares
;
1763 load
/= aggregate(tg
->parent
, sd
)->rq_weight
+ 1;
1766 aggregate(tg
, sd
)->load
= load
;
1769 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1772 * Calculate and set the cpu's group shares.
1775 __update_group_shares_cpu(struct task_group
*tg
, struct sched_domain
*sd
,
1779 unsigned long shares
;
1780 unsigned long rq_weight
;
1785 rq_weight
= tg
->cfs_rq
[tcpu
]->load
.weight
;
1788 * If there are currently no tasks on the cpu pretend there is one of
1789 * average load so that when a new task gets to run here it will not
1790 * get delayed by group starvation.
1794 rq_weight
= NICE_0_LOAD
;
1798 * \Sum shares * rq_weight
1799 * shares = -----------------------
1803 shares
= aggregate(tg
, sd
)->shares
* rq_weight
;
1804 shares
/= aggregate(tg
, sd
)->rq_weight
+ 1;
1807 * record the actual number of shares, not the boosted amount.
1809 tg
->cfs_rq
[tcpu
]->shares
= boost
? 0 : shares
;
1811 if (shares
< MIN_SHARES
)
1812 shares
= MIN_SHARES
;
1813 else if (shares
> MAX_SHARES
)
1814 shares
= MAX_SHARES
;
1816 __set_se_shares(tg
->se
[tcpu
], shares
);
1820 * Re-adjust the weights on the cpu the task came from and on the cpu the
1824 __move_group_shares(struct task_group
*tg
, struct sched_domain
*sd
,
1827 unsigned long shares
;
1829 shares
= tg
->cfs_rq
[scpu
]->shares
+ tg
->cfs_rq
[dcpu
]->shares
;
1831 __update_group_shares_cpu(tg
, sd
, scpu
);
1832 __update_group_shares_cpu(tg
, sd
, dcpu
);
1835 * ensure we never loose shares due to rounding errors in the
1836 * above redistribution.
1838 shares
-= tg
->cfs_rq
[scpu
]->shares
+ tg
->cfs_rq
[dcpu
]->shares
;
1840 tg
->cfs_rq
[dcpu
]->shares
+= shares
;
1844 * Because changing a group's shares changes the weight of the super-group
1845 * we need to walk up the tree and change all shares until we hit the root.
1848 move_group_shares(struct task_group
*tg
, struct sched_domain
*sd
,
1852 __move_group_shares(tg
, sd
, scpu
, dcpu
);
1858 void aggregate_group_set_shares(struct task_group
*tg
, struct sched_domain
*sd
)
1860 unsigned long shares
= aggregate(tg
, sd
)->shares
;
1863 for_each_cpu_mask(i
, sd
->span
) {
1864 struct rq
*rq
= cpu_rq(i
);
1865 unsigned long flags
;
1867 spin_lock_irqsave(&rq
->lock
, flags
);
1868 __update_group_shares_cpu(tg
, sd
, i
);
1869 spin_unlock_irqrestore(&rq
->lock
, flags
);
1872 aggregate_group_shares(tg
, sd
);
1875 * ensure we never loose shares due to rounding errors in the
1876 * above redistribution.
1878 shares
-= aggregate(tg
, sd
)->shares
;
1880 tg
->cfs_rq
[sd
->first_cpu
]->shares
+= shares
;
1881 aggregate(tg
, sd
)->shares
+= shares
;
1886 * Calculate the accumulative weight and recursive load of each task group
1887 * while walking down the tree.
1890 void aggregate_get_down(struct task_group
*tg
, struct sched_domain
*sd
)
1892 aggregate_group_weight(tg
, sd
);
1893 aggregate_group_shares(tg
, sd
);
1894 aggregate_group_load(tg
, sd
);
1898 * Rebalance the cpu shares while walking back up the tree.
1901 void aggregate_get_up(struct task_group
*tg
, struct sched_domain
*sd
)
1903 aggregate_group_set_shares(tg
, sd
);
1906 static DEFINE_PER_CPU(spinlock_t
, aggregate_lock
);
1908 static void __init
init_aggregate(void)
1912 for_each_possible_cpu(i
)
1913 spin_lock_init(&per_cpu(aggregate_lock
, i
));
1916 static int get_aggregate(struct sched_domain
*sd
)
1918 if (!spin_trylock(&per_cpu(aggregate_lock
, sd
->first_cpu
)))
1921 aggregate_walk_tree(aggregate_get_down
, aggregate_get_up
, sd
);
1925 static void put_aggregate(struct sched_domain
*sd
)
1927 spin_unlock(&per_cpu(aggregate_lock
, sd
->first_cpu
));
1930 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1932 cfs_rq
->shares
= shares
;
1937 static inline void init_aggregate(void)
1941 static inline int get_aggregate(struct sched_domain
*sd
)
1946 static inline void put_aggregate(struct sched_domain
*sd
)
1951 #else /* CONFIG_SMP */
1953 #ifdef CONFIG_FAIR_GROUP_SCHED
1954 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1959 #endif /* CONFIG_SMP */
1961 #include "sched_stats.h"
1962 #include "sched_idletask.c"
1963 #include "sched_fair.c"
1964 #include "sched_rt.c"
1965 #ifdef CONFIG_SCHED_DEBUG
1966 # include "sched_debug.c"
1969 #define sched_class_highest (&rt_sched_class)
1971 static void inc_nr_running(struct rq
*rq
)
1976 static void dec_nr_running(struct rq
*rq
)
1981 static void set_load_weight(struct task_struct
*p
)
1983 if (task_has_rt_policy(p
)) {
1984 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1985 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1990 * SCHED_IDLE tasks get minimal weight:
1992 if (p
->policy
== SCHED_IDLE
) {
1993 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1994 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1998 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1999 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
2002 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
2004 sched_info_queued(p
);
2005 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
2009 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
2011 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
2016 * __normal_prio - return the priority that is based on the static prio
2018 static inline int __normal_prio(struct task_struct
*p
)
2020 return p
->static_prio
;
2024 * Calculate the expected normal priority: i.e. priority
2025 * without taking RT-inheritance into account. Might be
2026 * boosted by interactivity modifiers. Changes upon fork,
2027 * setprio syscalls, and whenever the interactivity
2028 * estimator recalculates.
2030 static inline int normal_prio(struct task_struct
*p
)
2034 if (task_has_rt_policy(p
))
2035 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2037 prio
= __normal_prio(p
);
2042 * Calculate the current priority, i.e. the priority
2043 * taken into account by the scheduler. This value might
2044 * be boosted by RT tasks, or might be boosted by
2045 * interactivity modifiers. Will be RT if the task got
2046 * RT-boosted. If not then it returns p->normal_prio.
2048 static int effective_prio(struct task_struct
*p
)
2050 p
->normal_prio
= normal_prio(p
);
2052 * If we are RT tasks or we were boosted to RT priority,
2053 * keep the priority unchanged. Otherwise, update priority
2054 * to the normal priority:
2056 if (!rt_prio(p
->prio
))
2057 return p
->normal_prio
;
2062 * activate_task - move a task to the runqueue.
2064 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
2066 if (task_contributes_to_load(p
))
2067 rq
->nr_uninterruptible
--;
2069 enqueue_task(rq
, p
, wakeup
);
2074 * deactivate_task - remove a task from the runqueue.
2076 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
2078 if (task_contributes_to_load(p
))
2079 rq
->nr_uninterruptible
++;
2081 dequeue_task(rq
, p
, sleep
);
2086 * task_curr - is this task currently executing on a CPU?
2087 * @p: the task in question.
2089 inline int task_curr(const struct task_struct
*p
)
2091 return cpu_curr(task_cpu(p
)) == p
;
2094 /* Used instead of source_load when we know the type == 0 */
2095 unsigned long weighted_cpuload(const int cpu
)
2097 return cpu_rq(cpu
)->load
.weight
;
2100 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
2102 set_task_rq(p
, cpu
);
2105 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2106 * successfuly executed on another CPU. We must ensure that updates of
2107 * per-task data have been completed by this moment.
2110 task_thread_info(p
)->cpu
= cpu
;
2114 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2115 const struct sched_class
*prev_class
,
2116 int oldprio
, int running
)
2118 if (prev_class
!= p
->sched_class
) {
2119 if (prev_class
->switched_from
)
2120 prev_class
->switched_from(rq
, p
, running
);
2121 p
->sched_class
->switched_to(rq
, p
, running
);
2123 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2129 * Is this task likely cache-hot:
2132 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2137 * Buddy candidates are cache hot:
2139 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
2142 if (p
->sched_class
!= &fair_sched_class
)
2145 if (sysctl_sched_migration_cost
== -1)
2147 if (sysctl_sched_migration_cost
== 0)
2150 delta
= now
- p
->se
.exec_start
;
2152 return delta
< (s64
)sysctl_sched_migration_cost
;
2156 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2158 int old_cpu
= task_cpu(p
);
2159 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2160 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2161 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2164 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2166 #ifdef CONFIG_SCHEDSTATS
2167 if (p
->se
.wait_start
)
2168 p
->se
.wait_start
-= clock_offset
;
2169 if (p
->se
.sleep_start
)
2170 p
->se
.sleep_start
-= clock_offset
;
2171 if (p
->se
.block_start
)
2172 p
->se
.block_start
-= clock_offset
;
2173 if (old_cpu
!= new_cpu
) {
2174 schedstat_inc(p
, se
.nr_migrations
);
2175 if (task_hot(p
, old_rq
->clock
, NULL
))
2176 schedstat_inc(p
, se
.nr_forced2_migrations
);
2179 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2180 new_cfsrq
->min_vruntime
;
2182 __set_task_cpu(p
, new_cpu
);
2185 struct migration_req
{
2186 struct list_head list
;
2188 struct task_struct
*task
;
2191 struct completion done
;
2195 * The task's runqueue lock must be held.
2196 * Returns true if you have to wait for migration thread.
2199 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2201 struct rq
*rq
= task_rq(p
);
2204 * If the task is not on a runqueue (and not running), then
2205 * it is sufficient to simply update the task's cpu field.
2207 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2208 set_task_cpu(p
, dest_cpu
);
2212 init_completion(&req
->done
);
2214 req
->dest_cpu
= dest_cpu
;
2215 list_add(&req
->list
, &rq
->migration_queue
);
2221 * wait_task_inactive - wait for a thread to unschedule.
2223 * The caller must ensure that the task *will* unschedule sometime soon,
2224 * else this function might spin for a *long* time. This function can't
2225 * be called with interrupts off, or it may introduce deadlock with
2226 * smp_call_function() if an IPI is sent by the same process we are
2227 * waiting to become inactive.
2229 void wait_task_inactive(struct task_struct
*p
)
2231 unsigned long flags
;
2237 * We do the initial early heuristics without holding
2238 * any task-queue locks at all. We'll only try to get
2239 * the runqueue lock when things look like they will
2245 * If the task is actively running on another CPU
2246 * still, just relax and busy-wait without holding
2249 * NOTE! Since we don't hold any locks, it's not
2250 * even sure that "rq" stays as the right runqueue!
2251 * But we don't care, since "task_running()" will
2252 * return false if the runqueue has changed and p
2253 * is actually now running somewhere else!
2255 while (task_running(rq
, p
))
2259 * Ok, time to look more closely! We need the rq
2260 * lock now, to be *sure*. If we're wrong, we'll
2261 * just go back and repeat.
2263 rq
= task_rq_lock(p
, &flags
);
2264 running
= task_running(rq
, p
);
2265 on_rq
= p
->se
.on_rq
;
2266 task_rq_unlock(rq
, &flags
);
2269 * Was it really running after all now that we
2270 * checked with the proper locks actually held?
2272 * Oops. Go back and try again..
2274 if (unlikely(running
)) {
2280 * It's not enough that it's not actively running,
2281 * it must be off the runqueue _entirely_, and not
2284 * So if it wa still runnable (but just not actively
2285 * running right now), it's preempted, and we should
2286 * yield - it could be a while.
2288 if (unlikely(on_rq
)) {
2289 schedule_timeout_uninterruptible(1);
2294 * Ahh, all good. It wasn't running, and it wasn't
2295 * runnable, which means that it will never become
2296 * running in the future either. We're all done!
2303 * kick_process - kick a running thread to enter/exit the kernel
2304 * @p: the to-be-kicked thread
2306 * Cause a process which is running on another CPU to enter
2307 * kernel-mode, without any delay. (to get signals handled.)
2309 * NOTE: this function doesnt have to take the runqueue lock,
2310 * because all it wants to ensure is that the remote task enters
2311 * the kernel. If the IPI races and the task has been migrated
2312 * to another CPU then no harm is done and the purpose has been
2315 void kick_process(struct task_struct
*p
)
2321 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2322 smp_send_reschedule(cpu
);
2327 * Return a low guess at the load of a migration-source cpu weighted
2328 * according to the scheduling class and "nice" value.
2330 * We want to under-estimate the load of migration sources, to
2331 * balance conservatively.
2333 static unsigned long source_load(int cpu
, int type
)
2335 struct rq
*rq
= cpu_rq(cpu
);
2336 unsigned long total
= weighted_cpuload(cpu
);
2341 return min(rq
->cpu_load
[type
-1], total
);
2345 * Return a high guess at the load of a migration-target cpu weighted
2346 * according to the scheduling class and "nice" value.
2348 static unsigned long target_load(int cpu
, int type
)
2350 struct rq
*rq
= cpu_rq(cpu
);
2351 unsigned long total
= weighted_cpuload(cpu
);
2356 return max(rq
->cpu_load
[type
-1], total
);
2360 * Return the average load per task on the cpu's run queue
2362 static unsigned long cpu_avg_load_per_task(int cpu
)
2364 struct rq
*rq
= cpu_rq(cpu
);
2365 unsigned long total
= weighted_cpuload(cpu
);
2366 unsigned long n
= rq
->nr_running
;
2368 return n
? total
/ n
: SCHED_LOAD_SCALE
;
2372 * find_idlest_group finds and returns the least busy CPU group within the
2375 static struct sched_group
*
2376 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2378 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2379 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2380 int load_idx
= sd
->forkexec_idx
;
2381 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2384 unsigned long load
, avg_load
;
2388 /* Skip over this group if it has no CPUs allowed */
2389 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2392 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2394 /* Tally up the load of all CPUs in the group */
2397 for_each_cpu_mask(i
, group
->cpumask
) {
2398 /* Bias balancing toward cpus of our domain */
2400 load
= source_load(i
, load_idx
);
2402 load
= target_load(i
, load_idx
);
2407 /* Adjust by relative CPU power of the group */
2408 avg_load
= sg_div_cpu_power(group
,
2409 avg_load
* SCHED_LOAD_SCALE
);
2412 this_load
= avg_load
;
2414 } else if (avg_load
< min_load
) {
2415 min_load
= avg_load
;
2418 } while (group
= group
->next
, group
!= sd
->groups
);
2420 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2426 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2429 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2432 unsigned long load
, min_load
= ULONG_MAX
;
2436 /* Traverse only the allowed CPUs */
2437 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2439 for_each_cpu_mask(i
, *tmp
) {
2440 load
= weighted_cpuload(i
);
2442 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2452 * sched_balance_self: balance the current task (running on cpu) in domains
2453 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2456 * Balance, ie. select the least loaded group.
2458 * Returns the target CPU number, or the same CPU if no balancing is needed.
2460 * preempt must be disabled.
2462 static int sched_balance_self(int cpu
, int flag
)
2464 struct task_struct
*t
= current
;
2465 struct sched_domain
*tmp
, *sd
= NULL
;
2467 for_each_domain(cpu
, tmp
) {
2469 * If power savings logic is enabled for a domain, stop there.
2471 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2473 if (tmp
->flags
& flag
)
2478 cpumask_t span
, tmpmask
;
2479 struct sched_group
*group
;
2480 int new_cpu
, weight
;
2482 if (!(sd
->flags
& flag
)) {
2488 group
= find_idlest_group(sd
, t
, cpu
);
2494 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2495 if (new_cpu
== -1 || new_cpu
== cpu
) {
2496 /* Now try balancing at a lower domain level of cpu */
2501 /* Now try balancing at a lower domain level of new_cpu */
2504 weight
= cpus_weight(span
);
2505 for_each_domain(cpu
, tmp
) {
2506 if (weight
<= cpus_weight(tmp
->span
))
2508 if (tmp
->flags
& flag
)
2511 /* while loop will break here if sd == NULL */
2517 #endif /* CONFIG_SMP */
2520 * try_to_wake_up - wake up a thread
2521 * @p: the to-be-woken-up thread
2522 * @state: the mask of task states that can be woken
2523 * @sync: do a synchronous wakeup?
2525 * Put it on the run-queue if it's not already there. The "current"
2526 * thread is always on the run-queue (except when the actual
2527 * re-schedule is in progress), and as such you're allowed to do
2528 * the simpler "current->state = TASK_RUNNING" to mark yourself
2529 * runnable without the overhead of this.
2531 * returns failure only if the task is already active.
2533 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2535 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2536 unsigned long flags
;
2540 if (!sched_feat(SYNC_WAKEUPS
))
2544 rq
= task_rq_lock(p
, &flags
);
2545 old_state
= p
->state
;
2546 if (!(old_state
& state
))
2554 this_cpu
= smp_processor_id();
2557 if (unlikely(task_running(rq
, p
)))
2560 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2561 if (cpu
!= orig_cpu
) {
2562 set_task_cpu(p
, cpu
);
2563 task_rq_unlock(rq
, &flags
);
2564 /* might preempt at this point */
2565 rq
= task_rq_lock(p
, &flags
);
2566 old_state
= p
->state
;
2567 if (!(old_state
& state
))
2572 this_cpu
= smp_processor_id();
2576 #ifdef CONFIG_SCHEDSTATS
2577 schedstat_inc(rq
, ttwu_count
);
2578 if (cpu
== this_cpu
)
2579 schedstat_inc(rq
, ttwu_local
);
2581 struct sched_domain
*sd
;
2582 for_each_domain(this_cpu
, sd
) {
2583 if (cpu_isset(cpu
, sd
->span
)) {
2584 schedstat_inc(sd
, ttwu_wake_remote
);
2592 #endif /* CONFIG_SMP */
2593 schedstat_inc(p
, se
.nr_wakeups
);
2595 schedstat_inc(p
, se
.nr_wakeups_sync
);
2596 if (orig_cpu
!= cpu
)
2597 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2598 if (cpu
== this_cpu
)
2599 schedstat_inc(p
, se
.nr_wakeups_local
);
2601 schedstat_inc(p
, se
.nr_wakeups_remote
);
2602 update_rq_clock(rq
);
2603 activate_task(rq
, p
, 1);
2607 check_preempt_curr(rq
, p
);
2609 p
->state
= TASK_RUNNING
;
2611 if (p
->sched_class
->task_wake_up
)
2612 p
->sched_class
->task_wake_up(rq
, p
);
2615 task_rq_unlock(rq
, &flags
);
2620 int wake_up_process(struct task_struct
*p
)
2622 return try_to_wake_up(p
, TASK_ALL
, 0);
2624 EXPORT_SYMBOL(wake_up_process
);
2626 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2628 return try_to_wake_up(p
, state
, 0);
2632 * Perform scheduler related setup for a newly forked process p.
2633 * p is forked by current.
2635 * __sched_fork() is basic setup used by init_idle() too:
2637 static void __sched_fork(struct task_struct
*p
)
2639 p
->se
.exec_start
= 0;
2640 p
->se
.sum_exec_runtime
= 0;
2641 p
->se
.prev_sum_exec_runtime
= 0;
2642 p
->se
.last_wakeup
= 0;
2643 p
->se
.avg_overlap
= 0;
2645 #ifdef CONFIG_SCHEDSTATS
2646 p
->se
.wait_start
= 0;
2647 p
->se
.sum_sleep_runtime
= 0;
2648 p
->se
.sleep_start
= 0;
2649 p
->se
.block_start
= 0;
2650 p
->se
.sleep_max
= 0;
2651 p
->se
.block_max
= 0;
2653 p
->se
.slice_max
= 0;
2657 INIT_LIST_HEAD(&p
->rt
.run_list
);
2659 INIT_LIST_HEAD(&p
->se
.group_node
);
2661 #ifdef CONFIG_PREEMPT_NOTIFIERS
2662 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2666 * We mark the process as running here, but have not actually
2667 * inserted it onto the runqueue yet. This guarantees that
2668 * nobody will actually run it, and a signal or other external
2669 * event cannot wake it up and insert it on the runqueue either.
2671 p
->state
= TASK_RUNNING
;
2675 * fork()/clone()-time setup:
2677 void sched_fork(struct task_struct
*p
, int clone_flags
)
2679 int cpu
= get_cpu();
2684 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2686 set_task_cpu(p
, cpu
);
2689 * Make sure we do not leak PI boosting priority to the child:
2691 p
->prio
= current
->normal_prio
;
2692 if (!rt_prio(p
->prio
))
2693 p
->sched_class
= &fair_sched_class
;
2695 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2696 if (likely(sched_info_on()))
2697 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2699 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2702 #ifdef CONFIG_PREEMPT
2703 /* Want to start with kernel preemption disabled. */
2704 task_thread_info(p
)->preempt_count
= 1;
2710 * wake_up_new_task - wake up a newly created task for the first time.
2712 * This function will do some initial scheduler statistics housekeeping
2713 * that must be done for every newly created context, then puts the task
2714 * on the runqueue and wakes it.
2716 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2718 unsigned long flags
;
2721 rq
= task_rq_lock(p
, &flags
);
2722 BUG_ON(p
->state
!= TASK_RUNNING
);
2723 update_rq_clock(rq
);
2725 p
->prio
= effective_prio(p
);
2727 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2728 activate_task(rq
, p
, 0);
2731 * Let the scheduling class do new task startup
2732 * management (if any):
2734 p
->sched_class
->task_new(rq
, p
);
2737 check_preempt_curr(rq
, p
);
2739 if (p
->sched_class
->task_wake_up
)
2740 p
->sched_class
->task_wake_up(rq
, p
);
2742 task_rq_unlock(rq
, &flags
);
2745 #ifdef CONFIG_PREEMPT_NOTIFIERS
2748 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2749 * @notifier: notifier struct to register
2751 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2753 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2755 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2758 * preempt_notifier_unregister - no longer interested in preemption notifications
2759 * @notifier: notifier struct to unregister
2761 * This is safe to call from within a preemption notifier.
2763 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2765 hlist_del(¬ifier
->link
);
2767 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2769 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2771 struct preempt_notifier
*notifier
;
2772 struct hlist_node
*node
;
2774 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2775 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2779 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2780 struct task_struct
*next
)
2782 struct preempt_notifier
*notifier
;
2783 struct hlist_node
*node
;
2785 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2786 notifier
->ops
->sched_out(notifier
, next
);
2791 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2796 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2797 struct task_struct
*next
)
2804 * prepare_task_switch - prepare to switch tasks
2805 * @rq: the runqueue preparing to switch
2806 * @prev: the current task that is being switched out
2807 * @next: the task we are going to switch to.
2809 * This is called with the rq lock held and interrupts off. It must
2810 * be paired with a subsequent finish_task_switch after the context
2813 * prepare_task_switch sets up locking and calls architecture specific
2817 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2818 struct task_struct
*next
)
2820 fire_sched_out_preempt_notifiers(prev
, next
);
2821 prepare_lock_switch(rq
, next
);
2822 prepare_arch_switch(next
);
2826 * finish_task_switch - clean up after a task-switch
2827 * @rq: runqueue associated with task-switch
2828 * @prev: the thread we just switched away from.
2830 * finish_task_switch must be called after the context switch, paired
2831 * with a prepare_task_switch call before the context switch.
2832 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2833 * and do any other architecture-specific cleanup actions.
2835 * Note that we may have delayed dropping an mm in context_switch(). If
2836 * so, we finish that here outside of the runqueue lock. (Doing it
2837 * with the lock held can cause deadlocks; see schedule() for
2840 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2841 __releases(rq
->lock
)
2843 struct mm_struct
*mm
= rq
->prev_mm
;
2849 * A task struct has one reference for the use as "current".
2850 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2851 * schedule one last time. The schedule call will never return, and
2852 * the scheduled task must drop that reference.
2853 * The test for TASK_DEAD must occur while the runqueue locks are
2854 * still held, otherwise prev could be scheduled on another cpu, die
2855 * there before we look at prev->state, and then the reference would
2857 * Manfred Spraul <manfred@colorfullife.com>
2859 prev_state
= prev
->state
;
2860 finish_arch_switch(prev
);
2861 finish_lock_switch(rq
, prev
);
2863 if (current
->sched_class
->post_schedule
)
2864 current
->sched_class
->post_schedule(rq
);
2867 fire_sched_in_preempt_notifiers(current
);
2870 if (unlikely(prev_state
== TASK_DEAD
)) {
2872 * Remove function-return probe instances associated with this
2873 * task and put them back on the free list.
2875 kprobe_flush_task(prev
);
2876 put_task_struct(prev
);
2881 * schedule_tail - first thing a freshly forked thread must call.
2882 * @prev: the thread we just switched away from.
2884 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2885 __releases(rq
->lock
)
2887 struct rq
*rq
= this_rq();
2889 finish_task_switch(rq
, prev
);
2890 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2891 /* In this case, finish_task_switch does not reenable preemption */
2894 if (current
->set_child_tid
)
2895 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2899 * context_switch - switch to the new MM and the new
2900 * thread's register state.
2903 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2904 struct task_struct
*next
)
2906 struct mm_struct
*mm
, *oldmm
;
2908 prepare_task_switch(rq
, prev
, next
);
2910 oldmm
= prev
->active_mm
;
2912 * For paravirt, this is coupled with an exit in switch_to to
2913 * combine the page table reload and the switch backend into
2916 arch_enter_lazy_cpu_mode();
2918 if (unlikely(!mm
)) {
2919 next
->active_mm
= oldmm
;
2920 atomic_inc(&oldmm
->mm_count
);
2921 enter_lazy_tlb(oldmm
, next
);
2923 switch_mm(oldmm
, mm
, next
);
2925 if (unlikely(!prev
->mm
)) {
2926 prev
->active_mm
= NULL
;
2927 rq
->prev_mm
= oldmm
;
2930 * Since the runqueue lock will be released by the next
2931 * task (which is an invalid locking op but in the case
2932 * of the scheduler it's an obvious special-case), so we
2933 * do an early lockdep release here:
2935 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2936 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2939 /* Here we just switch the register state and the stack. */
2940 switch_to(prev
, next
, prev
);
2944 * this_rq must be evaluated again because prev may have moved
2945 * CPUs since it called schedule(), thus the 'rq' on its stack
2946 * frame will be invalid.
2948 finish_task_switch(this_rq(), prev
);
2952 * nr_running, nr_uninterruptible and nr_context_switches:
2954 * externally visible scheduler statistics: current number of runnable
2955 * threads, current number of uninterruptible-sleeping threads, total
2956 * number of context switches performed since bootup.
2958 unsigned long nr_running(void)
2960 unsigned long i
, sum
= 0;
2962 for_each_online_cpu(i
)
2963 sum
+= cpu_rq(i
)->nr_running
;
2968 unsigned long nr_uninterruptible(void)
2970 unsigned long i
, sum
= 0;
2972 for_each_possible_cpu(i
)
2973 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2976 * Since we read the counters lockless, it might be slightly
2977 * inaccurate. Do not allow it to go below zero though:
2979 if (unlikely((long)sum
< 0))
2985 unsigned long long nr_context_switches(void)
2988 unsigned long long sum
= 0;
2990 for_each_possible_cpu(i
)
2991 sum
+= cpu_rq(i
)->nr_switches
;
2996 unsigned long nr_iowait(void)
2998 unsigned long i
, sum
= 0;
3000 for_each_possible_cpu(i
)
3001 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3006 unsigned long nr_active(void)
3008 unsigned long i
, running
= 0, uninterruptible
= 0;
3010 for_each_online_cpu(i
) {
3011 running
+= cpu_rq(i
)->nr_running
;
3012 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
3015 if (unlikely((long)uninterruptible
< 0))
3016 uninterruptible
= 0;
3018 return running
+ uninterruptible
;
3022 * Update rq->cpu_load[] statistics. This function is usually called every
3023 * scheduler tick (TICK_NSEC).
3025 static void update_cpu_load(struct rq
*this_rq
)
3027 unsigned long this_load
= this_rq
->load
.weight
;
3030 this_rq
->nr_load_updates
++;
3032 /* Update our load: */
3033 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3034 unsigned long old_load
, new_load
;
3036 /* scale is effectively 1 << i now, and >> i divides by scale */
3038 old_load
= this_rq
->cpu_load
[i
];
3039 new_load
= this_load
;
3041 * Round up the averaging division if load is increasing. This
3042 * prevents us from getting stuck on 9 if the load is 10, for
3045 if (new_load
> old_load
)
3046 new_load
+= scale
-1;
3047 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3054 * double_rq_lock - safely lock two runqueues
3056 * Note this does not disable interrupts like task_rq_lock,
3057 * you need to do so manually before calling.
3059 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3060 __acquires(rq1
->lock
)
3061 __acquires(rq2
->lock
)
3063 BUG_ON(!irqs_disabled());
3065 spin_lock(&rq1
->lock
);
3066 __acquire(rq2
->lock
); /* Fake it out ;) */
3069 spin_lock(&rq1
->lock
);
3070 spin_lock(&rq2
->lock
);
3072 spin_lock(&rq2
->lock
);
3073 spin_lock(&rq1
->lock
);
3076 update_rq_clock(rq1
);
3077 update_rq_clock(rq2
);
3081 * double_rq_unlock - safely unlock two runqueues
3083 * Note this does not restore interrupts like task_rq_unlock,
3084 * you need to do so manually after calling.
3086 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3087 __releases(rq1
->lock
)
3088 __releases(rq2
->lock
)
3090 spin_unlock(&rq1
->lock
);
3092 spin_unlock(&rq2
->lock
);
3094 __release(rq2
->lock
);
3098 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3100 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
3101 __releases(this_rq
->lock
)
3102 __acquires(busiest
->lock
)
3103 __acquires(this_rq
->lock
)
3107 if (unlikely(!irqs_disabled())) {
3108 /* printk() doesn't work good under rq->lock */
3109 spin_unlock(&this_rq
->lock
);
3112 if (unlikely(!spin_trylock(&busiest
->lock
))) {
3113 if (busiest
< this_rq
) {
3114 spin_unlock(&this_rq
->lock
);
3115 spin_lock(&busiest
->lock
);
3116 spin_lock(&this_rq
->lock
);
3119 spin_lock(&busiest
->lock
);
3125 * If dest_cpu is allowed for this process, migrate the task to it.
3126 * This is accomplished by forcing the cpu_allowed mask to only
3127 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3128 * the cpu_allowed mask is restored.
3130 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3132 struct migration_req req
;
3133 unsigned long flags
;
3136 rq
= task_rq_lock(p
, &flags
);
3137 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
3138 || unlikely(cpu_is_offline(dest_cpu
)))
3141 /* force the process onto the specified CPU */
3142 if (migrate_task(p
, dest_cpu
, &req
)) {
3143 /* Need to wait for migration thread (might exit: take ref). */
3144 struct task_struct
*mt
= rq
->migration_thread
;
3146 get_task_struct(mt
);
3147 task_rq_unlock(rq
, &flags
);
3148 wake_up_process(mt
);
3149 put_task_struct(mt
);
3150 wait_for_completion(&req
.done
);
3155 task_rq_unlock(rq
, &flags
);
3159 * sched_exec - execve() is a valuable balancing opportunity, because at
3160 * this point the task has the smallest effective memory and cache footprint.
3162 void sched_exec(void)
3164 int new_cpu
, this_cpu
= get_cpu();
3165 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3167 if (new_cpu
!= this_cpu
)
3168 sched_migrate_task(current
, new_cpu
);
3172 * pull_task - move a task from a remote runqueue to the local runqueue.
3173 * Both runqueues must be locked.
3175 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3176 struct rq
*this_rq
, int this_cpu
)
3178 deactivate_task(src_rq
, p
, 0);
3179 set_task_cpu(p
, this_cpu
);
3180 activate_task(this_rq
, p
, 0);
3182 * Note that idle threads have a prio of MAX_PRIO, for this test
3183 * to be always true for them.
3185 check_preempt_curr(this_rq
, p
);
3189 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3192 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3193 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3197 * We do not migrate tasks that are:
3198 * 1) running (obviously), or
3199 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3200 * 3) are cache-hot on their current CPU.
3202 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
3203 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3208 if (task_running(rq
, p
)) {
3209 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3214 * Aggressive migration if:
3215 * 1) task is cache cold, or
3216 * 2) too many balance attempts have failed.
3219 if (!task_hot(p
, rq
->clock
, sd
) ||
3220 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3221 #ifdef CONFIG_SCHEDSTATS
3222 if (task_hot(p
, rq
->clock
, sd
)) {
3223 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3224 schedstat_inc(p
, se
.nr_forced_migrations
);
3230 if (task_hot(p
, rq
->clock
, sd
)) {
3231 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3237 static unsigned long
3238 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3239 unsigned long max_load_move
, struct sched_domain
*sd
,
3240 enum cpu_idle_type idle
, int *all_pinned
,
3241 int *this_best_prio
, struct rq_iterator
*iterator
)
3243 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
3244 struct task_struct
*p
;
3245 long rem_load_move
= max_load_move
;
3247 if (max_load_move
== 0)
3253 * Start the load-balancing iterator:
3255 p
= iterator
->start(iterator
->arg
);
3257 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3260 * To help distribute high priority tasks across CPUs we don't
3261 * skip a task if it will be the highest priority task (i.e. smallest
3262 * prio value) on its new queue regardless of its load weight
3264 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
3265 SCHED_LOAD_SCALE_FUZZ
;
3266 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
3267 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3268 p
= iterator
->next(iterator
->arg
);
3272 pull_task(busiest
, p
, this_rq
, this_cpu
);
3274 rem_load_move
-= p
->se
.load
.weight
;
3277 * We only want to steal up to the prescribed amount of weighted load.
3279 if (rem_load_move
> 0) {
3280 if (p
->prio
< *this_best_prio
)
3281 *this_best_prio
= p
->prio
;
3282 p
= iterator
->next(iterator
->arg
);
3287 * Right now, this is one of only two places pull_task() is called,
3288 * so we can safely collect pull_task() stats here rather than
3289 * inside pull_task().
3291 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3294 *all_pinned
= pinned
;
3296 return max_load_move
- rem_load_move
;
3300 * move_tasks tries to move up to max_load_move weighted load from busiest to
3301 * this_rq, as part of a balancing operation within domain "sd".
3302 * Returns 1 if successful and 0 otherwise.
3304 * Called with both runqueues locked.
3306 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3307 unsigned long max_load_move
,
3308 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3311 const struct sched_class
*class = sched_class_highest
;
3312 unsigned long total_load_moved
= 0;
3313 int this_best_prio
= this_rq
->curr
->prio
;
3317 class->load_balance(this_rq
, this_cpu
, busiest
,
3318 max_load_move
- total_load_moved
,
3319 sd
, idle
, all_pinned
, &this_best_prio
);
3320 class = class->next
;
3321 } while (class && max_load_move
> total_load_moved
);
3323 return total_load_moved
> 0;
3327 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3328 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3329 struct rq_iterator
*iterator
)
3331 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3335 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3336 pull_task(busiest
, p
, this_rq
, this_cpu
);
3338 * Right now, this is only the second place pull_task()
3339 * is called, so we can safely collect pull_task()
3340 * stats here rather than inside pull_task().
3342 schedstat_inc(sd
, lb_gained
[idle
]);
3346 p
= iterator
->next(iterator
->arg
);
3353 * move_one_task tries to move exactly one task from busiest to this_rq, as
3354 * part of active balancing operations within "domain".
3355 * Returns 1 if successful and 0 otherwise.
3357 * Called with both runqueues locked.
3359 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3360 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3362 const struct sched_class
*class;
3364 for (class = sched_class_highest
; class; class = class->next
)
3365 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3372 * find_busiest_group finds and returns the busiest CPU group within the
3373 * domain. It calculates and returns the amount of weighted load which
3374 * should be moved to restore balance via the imbalance parameter.
3376 static struct sched_group
*
3377 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3378 unsigned long *imbalance
, enum cpu_idle_type idle
,
3379 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3381 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3382 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3383 unsigned long max_pull
;
3384 unsigned long busiest_load_per_task
, busiest_nr_running
;
3385 unsigned long this_load_per_task
, this_nr_running
;
3386 int load_idx
, group_imb
= 0;
3387 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3388 int power_savings_balance
= 1;
3389 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3390 unsigned long min_nr_running
= ULONG_MAX
;
3391 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3394 max_load
= this_load
= total_load
= total_pwr
= 0;
3395 busiest_load_per_task
= busiest_nr_running
= 0;
3396 this_load_per_task
= this_nr_running
= 0;
3397 if (idle
== CPU_NOT_IDLE
)
3398 load_idx
= sd
->busy_idx
;
3399 else if (idle
== CPU_NEWLY_IDLE
)
3400 load_idx
= sd
->newidle_idx
;
3402 load_idx
= sd
->idle_idx
;
3405 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3408 int __group_imb
= 0;
3409 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3410 unsigned long sum_nr_running
, sum_weighted_load
;
3412 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3415 balance_cpu
= first_cpu(group
->cpumask
);
3417 /* Tally up the load of all CPUs in the group */
3418 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3420 min_cpu_load
= ~0UL;
3422 for_each_cpu_mask(i
, group
->cpumask
) {
3425 if (!cpu_isset(i
, *cpus
))
3430 if (*sd_idle
&& rq
->nr_running
)
3433 /* Bias balancing toward cpus of our domain */
3435 if (idle_cpu(i
) && !first_idle_cpu
) {
3440 load
= target_load(i
, load_idx
);
3442 load
= source_load(i
, load_idx
);
3443 if (load
> max_cpu_load
)
3444 max_cpu_load
= load
;
3445 if (min_cpu_load
> load
)
3446 min_cpu_load
= load
;
3450 sum_nr_running
+= rq
->nr_running
;
3451 sum_weighted_load
+= weighted_cpuload(i
);
3455 * First idle cpu or the first cpu(busiest) in this sched group
3456 * is eligible for doing load balancing at this and above
3457 * domains. In the newly idle case, we will allow all the cpu's
3458 * to do the newly idle load balance.
3460 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3461 balance_cpu
!= this_cpu
&& balance
) {
3466 total_load
+= avg_load
;
3467 total_pwr
+= group
->__cpu_power
;
3469 /* Adjust by relative CPU power of the group */
3470 avg_load
= sg_div_cpu_power(group
,
3471 avg_load
* SCHED_LOAD_SCALE
);
3473 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3476 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3479 this_load
= avg_load
;
3481 this_nr_running
= sum_nr_running
;
3482 this_load_per_task
= sum_weighted_load
;
3483 } else if (avg_load
> max_load
&&
3484 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3485 max_load
= avg_load
;
3487 busiest_nr_running
= sum_nr_running
;
3488 busiest_load_per_task
= sum_weighted_load
;
3489 group_imb
= __group_imb
;
3492 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3494 * Busy processors will not participate in power savings
3497 if (idle
== CPU_NOT_IDLE
||
3498 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3502 * If the local group is idle or completely loaded
3503 * no need to do power savings balance at this domain
3505 if (local_group
&& (this_nr_running
>= group_capacity
||
3507 power_savings_balance
= 0;
3510 * If a group is already running at full capacity or idle,
3511 * don't include that group in power savings calculations
3513 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3518 * Calculate the group which has the least non-idle load.
3519 * This is the group from where we need to pick up the load
3522 if ((sum_nr_running
< min_nr_running
) ||
3523 (sum_nr_running
== min_nr_running
&&
3524 first_cpu(group
->cpumask
) <
3525 first_cpu(group_min
->cpumask
))) {
3527 min_nr_running
= sum_nr_running
;
3528 min_load_per_task
= sum_weighted_load
/
3533 * Calculate the group which is almost near its
3534 * capacity but still has some space to pick up some load
3535 * from other group and save more power
3537 if (sum_nr_running
<= group_capacity
- 1) {
3538 if (sum_nr_running
> leader_nr_running
||
3539 (sum_nr_running
== leader_nr_running
&&
3540 first_cpu(group
->cpumask
) >
3541 first_cpu(group_leader
->cpumask
))) {
3542 group_leader
= group
;
3543 leader_nr_running
= sum_nr_running
;
3548 group
= group
->next
;
3549 } while (group
!= sd
->groups
);
3551 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3554 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3556 if (this_load
>= avg_load
||
3557 100*max_load
<= sd
->imbalance_pct
*this_load
)
3560 busiest_load_per_task
/= busiest_nr_running
;
3562 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3565 * We're trying to get all the cpus to the average_load, so we don't
3566 * want to push ourselves above the average load, nor do we wish to
3567 * reduce the max loaded cpu below the average load, as either of these
3568 * actions would just result in more rebalancing later, and ping-pong
3569 * tasks around. Thus we look for the minimum possible imbalance.
3570 * Negative imbalances (*we* are more loaded than anyone else) will
3571 * be counted as no imbalance for these purposes -- we can't fix that
3572 * by pulling tasks to us. Be careful of negative numbers as they'll
3573 * appear as very large values with unsigned longs.
3575 if (max_load
<= busiest_load_per_task
)
3579 * In the presence of smp nice balancing, certain scenarios can have
3580 * max load less than avg load(as we skip the groups at or below
3581 * its cpu_power, while calculating max_load..)
3583 if (max_load
< avg_load
) {
3585 goto small_imbalance
;
3588 /* Don't want to pull so many tasks that a group would go idle */
3589 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3591 /* How much load to actually move to equalise the imbalance */
3592 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3593 (avg_load
- this_load
) * this->__cpu_power
)
3597 * if *imbalance is less than the average load per runnable task
3598 * there is no gaurantee that any tasks will be moved so we'll have
3599 * a think about bumping its value to force at least one task to be
3602 if (*imbalance
< busiest_load_per_task
) {
3603 unsigned long tmp
, pwr_now
, pwr_move
;
3607 pwr_move
= pwr_now
= 0;
3609 if (this_nr_running
) {
3610 this_load_per_task
/= this_nr_running
;
3611 if (busiest_load_per_task
> this_load_per_task
)
3614 this_load_per_task
= SCHED_LOAD_SCALE
;
3616 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3617 busiest_load_per_task
* imbn
) {
3618 *imbalance
= busiest_load_per_task
;
3623 * OK, we don't have enough imbalance to justify moving tasks,
3624 * however we may be able to increase total CPU power used by
3628 pwr_now
+= busiest
->__cpu_power
*
3629 min(busiest_load_per_task
, max_load
);
3630 pwr_now
+= this->__cpu_power
*
3631 min(this_load_per_task
, this_load
);
3632 pwr_now
/= SCHED_LOAD_SCALE
;
3634 /* Amount of load we'd subtract */
3635 tmp
= sg_div_cpu_power(busiest
,
3636 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3638 pwr_move
+= busiest
->__cpu_power
*
3639 min(busiest_load_per_task
, max_load
- tmp
);
3641 /* Amount of load we'd add */
3642 if (max_load
* busiest
->__cpu_power
<
3643 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3644 tmp
= sg_div_cpu_power(this,
3645 max_load
* busiest
->__cpu_power
);
3647 tmp
= sg_div_cpu_power(this,
3648 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3649 pwr_move
+= this->__cpu_power
*
3650 min(this_load_per_task
, this_load
+ tmp
);
3651 pwr_move
/= SCHED_LOAD_SCALE
;
3653 /* Move if we gain throughput */
3654 if (pwr_move
> pwr_now
)
3655 *imbalance
= busiest_load_per_task
;
3661 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3662 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3665 if (this == group_leader
&& group_leader
!= group_min
) {
3666 *imbalance
= min_load_per_task
;
3676 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3679 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3680 unsigned long imbalance
, const cpumask_t
*cpus
)
3682 struct rq
*busiest
= NULL
, *rq
;
3683 unsigned long max_load
= 0;
3686 for_each_cpu_mask(i
, group
->cpumask
) {
3689 if (!cpu_isset(i
, *cpus
))
3693 wl
= weighted_cpuload(i
);
3695 if (rq
->nr_running
== 1 && wl
> imbalance
)
3698 if (wl
> max_load
) {
3708 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3709 * so long as it is large enough.
3711 #define MAX_PINNED_INTERVAL 512
3714 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3715 * tasks if there is an imbalance.
3717 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3718 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3719 int *balance
, cpumask_t
*cpus
)
3721 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3722 struct sched_group
*group
;
3723 unsigned long imbalance
;
3725 unsigned long flags
;
3726 int unlock_aggregate
;
3730 unlock_aggregate
= get_aggregate(sd
);
3733 * When power savings policy is enabled for the parent domain, idle
3734 * sibling can pick up load irrespective of busy siblings. In this case,
3735 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3736 * portraying it as CPU_NOT_IDLE.
3738 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3739 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3742 schedstat_inc(sd
, lb_count
[idle
]);
3745 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3752 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3756 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3758 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3762 BUG_ON(busiest
== this_rq
);
3764 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3767 if (busiest
->nr_running
> 1) {
3769 * Attempt to move tasks. If find_busiest_group has found
3770 * an imbalance but busiest->nr_running <= 1, the group is
3771 * still unbalanced. ld_moved simply stays zero, so it is
3772 * correctly treated as an imbalance.
3774 local_irq_save(flags
);
3775 double_rq_lock(this_rq
, busiest
);
3776 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3777 imbalance
, sd
, idle
, &all_pinned
);
3778 double_rq_unlock(this_rq
, busiest
);
3779 local_irq_restore(flags
);
3782 * some other cpu did the load balance for us.
3784 if (ld_moved
&& this_cpu
!= smp_processor_id())
3785 resched_cpu(this_cpu
);
3787 /* All tasks on this runqueue were pinned by CPU affinity */
3788 if (unlikely(all_pinned
)) {
3789 cpu_clear(cpu_of(busiest
), *cpus
);
3790 if (!cpus_empty(*cpus
))
3797 schedstat_inc(sd
, lb_failed
[idle
]);
3798 sd
->nr_balance_failed
++;
3800 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3802 spin_lock_irqsave(&busiest
->lock
, flags
);
3804 /* don't kick the migration_thread, if the curr
3805 * task on busiest cpu can't be moved to this_cpu
3807 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3808 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3810 goto out_one_pinned
;
3813 if (!busiest
->active_balance
) {
3814 busiest
->active_balance
= 1;
3815 busiest
->push_cpu
= this_cpu
;
3818 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3820 wake_up_process(busiest
->migration_thread
);
3823 * We've kicked active balancing, reset the failure
3826 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3829 sd
->nr_balance_failed
= 0;
3831 if (likely(!active_balance
)) {
3832 /* We were unbalanced, so reset the balancing interval */
3833 sd
->balance_interval
= sd
->min_interval
;
3836 * If we've begun active balancing, start to back off. This
3837 * case may not be covered by the all_pinned logic if there
3838 * is only 1 task on the busy runqueue (because we don't call
3841 if (sd
->balance_interval
< sd
->max_interval
)
3842 sd
->balance_interval
*= 2;
3845 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3846 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3852 schedstat_inc(sd
, lb_balanced
[idle
]);
3854 sd
->nr_balance_failed
= 0;
3857 /* tune up the balancing interval */
3858 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3859 (sd
->balance_interval
< sd
->max_interval
))
3860 sd
->balance_interval
*= 2;
3862 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3863 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3868 if (unlock_aggregate
)
3874 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3875 * tasks if there is an imbalance.
3877 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3878 * this_rq is locked.
3881 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3884 struct sched_group
*group
;
3885 struct rq
*busiest
= NULL
;
3886 unsigned long imbalance
;
3894 * When power savings policy is enabled for the parent domain, idle
3895 * sibling can pick up load irrespective of busy siblings. In this case,
3896 * let the state of idle sibling percolate up as IDLE, instead of
3897 * portraying it as CPU_NOT_IDLE.
3899 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3900 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3903 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3905 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3906 &sd_idle
, cpus
, NULL
);
3908 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3912 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3914 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3918 BUG_ON(busiest
== this_rq
);
3920 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3923 if (busiest
->nr_running
> 1) {
3924 /* Attempt to move tasks */
3925 double_lock_balance(this_rq
, busiest
);
3926 /* this_rq->clock is already updated */
3927 update_rq_clock(busiest
);
3928 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3929 imbalance
, sd
, CPU_NEWLY_IDLE
,
3931 spin_unlock(&busiest
->lock
);
3933 if (unlikely(all_pinned
)) {
3934 cpu_clear(cpu_of(busiest
), *cpus
);
3935 if (!cpus_empty(*cpus
))
3941 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3942 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3943 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3946 sd
->nr_balance_failed
= 0;
3951 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3952 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3953 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3955 sd
->nr_balance_failed
= 0;
3961 * idle_balance is called by schedule() if this_cpu is about to become
3962 * idle. Attempts to pull tasks from other CPUs.
3964 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3966 struct sched_domain
*sd
;
3967 int pulled_task
= -1;
3968 unsigned long next_balance
= jiffies
+ HZ
;
3971 for_each_domain(this_cpu
, sd
) {
3972 unsigned long interval
;
3974 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3977 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3978 /* If we've pulled tasks over stop searching: */
3979 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3982 interval
= msecs_to_jiffies(sd
->balance_interval
);
3983 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3984 next_balance
= sd
->last_balance
+ interval
;
3988 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3990 * We are going idle. next_balance may be set based on
3991 * a busy processor. So reset next_balance.
3993 this_rq
->next_balance
= next_balance
;
3998 * active_load_balance is run by migration threads. It pushes running tasks
3999 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4000 * running on each physical CPU where possible, and avoids physical /
4001 * logical imbalances.
4003 * Called with busiest_rq locked.
4005 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4007 int target_cpu
= busiest_rq
->push_cpu
;
4008 struct sched_domain
*sd
;
4009 struct rq
*target_rq
;
4011 /* Is there any task to move? */
4012 if (busiest_rq
->nr_running
<= 1)
4015 target_rq
= cpu_rq(target_cpu
);
4018 * This condition is "impossible", if it occurs
4019 * we need to fix it. Originally reported by
4020 * Bjorn Helgaas on a 128-cpu setup.
4022 BUG_ON(busiest_rq
== target_rq
);
4024 /* move a task from busiest_rq to target_rq */
4025 double_lock_balance(busiest_rq
, target_rq
);
4026 update_rq_clock(busiest_rq
);
4027 update_rq_clock(target_rq
);
4029 /* Search for an sd spanning us and the target CPU. */
4030 for_each_domain(target_cpu
, sd
) {
4031 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4032 cpu_isset(busiest_cpu
, sd
->span
))
4037 schedstat_inc(sd
, alb_count
);
4039 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4041 schedstat_inc(sd
, alb_pushed
);
4043 schedstat_inc(sd
, alb_failed
);
4045 spin_unlock(&target_rq
->lock
);
4050 atomic_t load_balancer
;
4052 } nohz ____cacheline_aligned
= {
4053 .load_balancer
= ATOMIC_INIT(-1),
4054 .cpu_mask
= CPU_MASK_NONE
,
4058 * This routine will try to nominate the ilb (idle load balancing)
4059 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4060 * load balancing on behalf of all those cpus. If all the cpus in the system
4061 * go into this tickless mode, then there will be no ilb owner (as there is
4062 * no need for one) and all the cpus will sleep till the next wakeup event
4065 * For the ilb owner, tick is not stopped. And this tick will be used
4066 * for idle load balancing. ilb owner will still be part of
4069 * While stopping the tick, this cpu will become the ilb owner if there
4070 * is no other owner. And will be the owner till that cpu becomes busy
4071 * or if all cpus in the system stop their ticks at which point
4072 * there is no need for ilb owner.
4074 * When the ilb owner becomes busy, it nominates another owner, during the
4075 * next busy scheduler_tick()
4077 int select_nohz_load_balancer(int stop_tick
)
4079 int cpu
= smp_processor_id();
4082 cpu_set(cpu
, nohz
.cpu_mask
);
4083 cpu_rq(cpu
)->in_nohz_recently
= 1;
4086 * If we are going offline and still the leader, give up!
4088 if (cpu_is_offline(cpu
) &&
4089 atomic_read(&nohz
.load_balancer
) == cpu
) {
4090 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4095 /* time for ilb owner also to sleep */
4096 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4097 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4098 atomic_set(&nohz
.load_balancer
, -1);
4102 if (atomic_read(&nohz
.load_balancer
) == -1) {
4103 /* make me the ilb owner */
4104 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4106 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4109 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
4112 cpu_clear(cpu
, nohz
.cpu_mask
);
4114 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4115 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4122 static DEFINE_SPINLOCK(balancing
);
4125 * It checks each scheduling domain to see if it is due to be balanced,
4126 * and initiates a balancing operation if so.
4128 * Balancing parameters are set up in arch_init_sched_domains.
4130 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4133 struct rq
*rq
= cpu_rq(cpu
);
4134 unsigned long interval
;
4135 struct sched_domain
*sd
;
4136 /* Earliest time when we have to do rebalance again */
4137 unsigned long next_balance
= jiffies
+ 60*HZ
;
4138 int update_next_balance
= 0;
4141 for_each_domain(cpu
, sd
) {
4142 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4145 interval
= sd
->balance_interval
;
4146 if (idle
!= CPU_IDLE
)
4147 interval
*= sd
->busy_factor
;
4149 /* scale ms to jiffies */
4150 interval
= msecs_to_jiffies(interval
);
4151 if (unlikely(!interval
))
4153 if (interval
> HZ
*NR_CPUS
/10)
4154 interval
= HZ
*NR_CPUS
/10;
4157 if (sd
->flags
& SD_SERIALIZE
) {
4158 if (!spin_trylock(&balancing
))
4162 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4163 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
4165 * We've pulled tasks over so either we're no
4166 * longer idle, or one of our SMT siblings is
4169 idle
= CPU_NOT_IDLE
;
4171 sd
->last_balance
= jiffies
;
4173 if (sd
->flags
& SD_SERIALIZE
)
4174 spin_unlock(&balancing
);
4176 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4177 next_balance
= sd
->last_balance
+ interval
;
4178 update_next_balance
= 1;
4182 * Stop the load balance at this level. There is another
4183 * CPU in our sched group which is doing load balancing more
4191 * next_balance will be updated only when there is a need.
4192 * When the cpu is attached to null domain for ex, it will not be
4195 if (likely(update_next_balance
))
4196 rq
->next_balance
= next_balance
;
4200 * run_rebalance_domains is triggered when needed from the scheduler tick.
4201 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4202 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4204 static void run_rebalance_domains(struct softirq_action
*h
)
4206 int this_cpu
= smp_processor_id();
4207 struct rq
*this_rq
= cpu_rq(this_cpu
);
4208 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4209 CPU_IDLE
: CPU_NOT_IDLE
;
4211 rebalance_domains(this_cpu
, idle
);
4215 * If this cpu is the owner for idle load balancing, then do the
4216 * balancing on behalf of the other idle cpus whose ticks are
4219 if (this_rq
->idle_at_tick
&&
4220 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4221 cpumask_t cpus
= nohz
.cpu_mask
;
4225 cpu_clear(this_cpu
, cpus
);
4226 for_each_cpu_mask(balance_cpu
, cpus
) {
4228 * If this cpu gets work to do, stop the load balancing
4229 * work being done for other cpus. Next load
4230 * balancing owner will pick it up.
4235 rebalance_domains(balance_cpu
, CPU_IDLE
);
4237 rq
= cpu_rq(balance_cpu
);
4238 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4239 this_rq
->next_balance
= rq
->next_balance
;
4246 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4248 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4249 * idle load balancing owner or decide to stop the periodic load balancing,
4250 * if the whole system is idle.
4252 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4256 * If we were in the nohz mode recently and busy at the current
4257 * scheduler tick, then check if we need to nominate new idle
4260 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4261 rq
->in_nohz_recently
= 0;
4263 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4264 cpu_clear(cpu
, nohz
.cpu_mask
);
4265 atomic_set(&nohz
.load_balancer
, -1);
4268 if (atomic_read(&nohz
.load_balancer
) == -1) {
4270 * simple selection for now: Nominate the
4271 * first cpu in the nohz list to be the next
4274 * TBD: Traverse the sched domains and nominate
4275 * the nearest cpu in the nohz.cpu_mask.
4277 int ilb
= first_cpu(nohz
.cpu_mask
);
4279 if (ilb
< nr_cpu_ids
)
4285 * If this cpu is idle and doing idle load balancing for all the
4286 * cpus with ticks stopped, is it time for that to stop?
4288 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4289 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4295 * If this cpu is idle and the idle load balancing is done by
4296 * someone else, then no need raise the SCHED_SOFTIRQ
4298 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4299 cpu_isset(cpu
, nohz
.cpu_mask
))
4302 if (time_after_eq(jiffies
, rq
->next_balance
))
4303 raise_softirq(SCHED_SOFTIRQ
);
4306 #else /* CONFIG_SMP */
4309 * on UP we do not need to balance between CPUs:
4311 static inline void idle_balance(int cpu
, struct rq
*rq
)
4317 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4319 EXPORT_PER_CPU_SYMBOL(kstat
);
4322 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4323 * that have not yet been banked in case the task is currently running.
4325 unsigned long long task_sched_runtime(struct task_struct
*p
)
4327 unsigned long flags
;
4331 rq
= task_rq_lock(p
, &flags
);
4332 ns
= p
->se
.sum_exec_runtime
;
4333 if (task_current(rq
, p
)) {
4334 update_rq_clock(rq
);
4335 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4336 if ((s64
)delta_exec
> 0)
4339 task_rq_unlock(rq
, &flags
);
4345 * Account user cpu time to a process.
4346 * @p: the process that the cpu time gets accounted to
4347 * @cputime: the cpu time spent in user space since the last update
4349 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4351 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4354 p
->utime
= cputime_add(p
->utime
, cputime
);
4356 /* Add user time to cpustat. */
4357 tmp
= cputime_to_cputime64(cputime
);
4358 if (TASK_NICE(p
) > 0)
4359 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4361 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4365 * Account guest cpu time to a process.
4366 * @p: the process that the cpu time gets accounted to
4367 * @cputime: the cpu time spent in virtual machine since the last update
4369 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4372 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4374 tmp
= cputime_to_cputime64(cputime
);
4376 p
->utime
= cputime_add(p
->utime
, cputime
);
4377 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4379 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4380 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4384 * Account scaled user cpu time to a process.
4385 * @p: the process that the cpu time gets accounted to
4386 * @cputime: the cpu time spent in user space since the last update
4388 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4390 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4394 * Account system cpu time to a process.
4395 * @p: the process that the cpu time gets accounted to
4396 * @hardirq_offset: the offset to subtract from hardirq_count()
4397 * @cputime: the cpu time spent in kernel space since the last update
4399 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4402 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4403 struct rq
*rq
= this_rq();
4406 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4407 account_guest_time(p
, cputime
);
4411 p
->stime
= cputime_add(p
->stime
, cputime
);
4413 /* Add system time to cpustat. */
4414 tmp
= cputime_to_cputime64(cputime
);
4415 if (hardirq_count() - hardirq_offset
)
4416 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4417 else if (softirq_count())
4418 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4419 else if (p
!= rq
->idle
)
4420 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4421 else if (atomic_read(&rq
->nr_iowait
) > 0)
4422 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4424 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4425 /* Account for system time used */
4426 acct_update_integrals(p
);
4430 * Account scaled system cpu time to a process.
4431 * @p: the process that the cpu time gets accounted to
4432 * @hardirq_offset: the offset to subtract from hardirq_count()
4433 * @cputime: the cpu time spent in kernel space since the last update
4435 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4437 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4441 * Account for involuntary wait time.
4442 * @p: the process from which the cpu time has been stolen
4443 * @steal: the cpu time spent in involuntary wait
4445 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4447 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4448 cputime64_t tmp
= cputime_to_cputime64(steal
);
4449 struct rq
*rq
= this_rq();
4451 if (p
== rq
->idle
) {
4452 p
->stime
= cputime_add(p
->stime
, steal
);
4453 if (atomic_read(&rq
->nr_iowait
) > 0)
4454 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4456 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4458 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4462 * This function gets called by the timer code, with HZ frequency.
4463 * We call it with interrupts disabled.
4465 * It also gets called by the fork code, when changing the parent's
4468 void scheduler_tick(void)
4470 int cpu
= smp_processor_id();
4471 struct rq
*rq
= cpu_rq(cpu
);
4472 struct task_struct
*curr
= rq
->curr
;
4473 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
4475 spin_lock(&rq
->lock
);
4476 __update_rq_clock(rq
);
4478 * Let rq->clock advance by at least TICK_NSEC:
4480 if (unlikely(rq
->clock
< next_tick
)) {
4481 rq
->clock
= next_tick
;
4482 rq
->clock_underflows
++;
4484 rq
->tick_timestamp
= rq
->clock
;
4485 update_last_tick_seen(rq
);
4486 update_cpu_load(rq
);
4487 curr
->sched_class
->task_tick(rq
, curr
, 0);
4488 spin_unlock(&rq
->lock
);
4491 rq
->idle_at_tick
= idle_cpu(cpu
);
4492 trigger_load_balance(rq
, cpu
);
4496 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4498 void __kprobes
add_preempt_count(int val
)
4503 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4505 preempt_count() += val
;
4507 * Spinlock count overflowing soon?
4509 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4512 EXPORT_SYMBOL(add_preempt_count
);
4514 void __kprobes
sub_preempt_count(int val
)
4519 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4522 * Is the spinlock portion underflowing?
4524 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4525 !(preempt_count() & PREEMPT_MASK
)))
4528 preempt_count() -= val
;
4530 EXPORT_SYMBOL(sub_preempt_count
);
4535 * Print scheduling while atomic bug:
4537 static noinline
void __schedule_bug(struct task_struct
*prev
)
4539 struct pt_regs
*regs
= get_irq_regs();
4541 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4542 prev
->comm
, prev
->pid
, preempt_count());
4544 debug_show_held_locks(prev
);
4545 if (irqs_disabled())
4546 print_irqtrace_events(prev
);
4555 * Various schedule()-time debugging checks and statistics:
4557 static inline void schedule_debug(struct task_struct
*prev
)
4560 * Test if we are atomic. Since do_exit() needs to call into
4561 * schedule() atomically, we ignore that path for now.
4562 * Otherwise, whine if we are scheduling when we should not be.
4564 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
4565 __schedule_bug(prev
);
4567 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4569 schedstat_inc(this_rq(), sched_count
);
4570 #ifdef CONFIG_SCHEDSTATS
4571 if (unlikely(prev
->lock_depth
>= 0)) {
4572 schedstat_inc(this_rq(), bkl_count
);
4573 schedstat_inc(prev
, sched_info
.bkl_count
);
4579 * Pick up the highest-prio task:
4581 static inline struct task_struct
*
4582 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4584 const struct sched_class
*class;
4585 struct task_struct
*p
;
4588 * Optimization: we know that if all tasks are in
4589 * the fair class we can call that function directly:
4591 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4592 p
= fair_sched_class
.pick_next_task(rq
);
4597 class = sched_class_highest
;
4599 p
= class->pick_next_task(rq
);
4603 * Will never be NULL as the idle class always
4604 * returns a non-NULL p:
4606 class = class->next
;
4611 * schedule() is the main scheduler function.
4613 asmlinkage
void __sched
schedule(void)
4615 struct task_struct
*prev
, *next
;
4616 unsigned long *switch_count
;
4622 cpu
= smp_processor_id();
4626 switch_count
= &prev
->nivcsw
;
4628 release_kernel_lock(prev
);
4629 need_resched_nonpreemptible
:
4631 schedule_debug(prev
);
4636 * Do the rq-clock update outside the rq lock:
4638 local_irq_disable();
4639 __update_rq_clock(rq
);
4640 spin_lock(&rq
->lock
);
4641 clear_tsk_need_resched(prev
);
4643 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4644 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
4645 signal_pending(prev
))) {
4646 prev
->state
= TASK_RUNNING
;
4648 deactivate_task(rq
, prev
, 1);
4650 switch_count
= &prev
->nvcsw
;
4654 if (prev
->sched_class
->pre_schedule
)
4655 prev
->sched_class
->pre_schedule(rq
, prev
);
4658 if (unlikely(!rq
->nr_running
))
4659 idle_balance(cpu
, rq
);
4661 prev
->sched_class
->put_prev_task(rq
, prev
);
4662 next
= pick_next_task(rq
, prev
);
4664 if (likely(prev
!= next
)) {
4665 sched_info_switch(prev
, next
);
4671 context_switch(rq
, prev
, next
); /* unlocks the rq */
4673 * the context switch might have flipped the stack from under
4674 * us, hence refresh the local variables.
4676 cpu
= smp_processor_id();
4679 spin_unlock_irq(&rq
->lock
);
4683 if (unlikely(reacquire_kernel_lock(current
) < 0))
4684 goto need_resched_nonpreemptible
;
4686 preempt_enable_no_resched();
4687 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4690 EXPORT_SYMBOL(schedule
);
4692 #ifdef CONFIG_PREEMPT
4694 * this is the entry point to schedule() from in-kernel preemption
4695 * off of preempt_enable. Kernel preemptions off return from interrupt
4696 * occur there and call schedule directly.
4698 asmlinkage
void __sched
preempt_schedule(void)
4700 struct thread_info
*ti
= current_thread_info();
4701 struct task_struct
*task
= current
;
4702 int saved_lock_depth
;
4705 * If there is a non-zero preempt_count or interrupts are disabled,
4706 * we do not want to preempt the current task. Just return..
4708 if (likely(ti
->preempt_count
|| irqs_disabled()))
4712 add_preempt_count(PREEMPT_ACTIVE
);
4715 * We keep the big kernel semaphore locked, but we
4716 * clear ->lock_depth so that schedule() doesnt
4717 * auto-release the semaphore:
4719 saved_lock_depth
= task
->lock_depth
;
4720 task
->lock_depth
= -1;
4722 task
->lock_depth
= saved_lock_depth
;
4723 sub_preempt_count(PREEMPT_ACTIVE
);
4726 * Check again in case we missed a preemption opportunity
4727 * between schedule and now.
4730 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4732 EXPORT_SYMBOL(preempt_schedule
);
4735 * this is the entry point to schedule() from kernel preemption
4736 * off of irq context.
4737 * Note, that this is called and return with irqs disabled. This will
4738 * protect us against recursive calling from irq.
4740 asmlinkage
void __sched
preempt_schedule_irq(void)
4742 struct thread_info
*ti
= current_thread_info();
4743 struct task_struct
*task
= current
;
4744 int saved_lock_depth
;
4746 /* Catch callers which need to be fixed */
4747 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4750 add_preempt_count(PREEMPT_ACTIVE
);
4753 * We keep the big kernel semaphore locked, but we
4754 * clear ->lock_depth so that schedule() doesnt
4755 * auto-release the semaphore:
4757 saved_lock_depth
= task
->lock_depth
;
4758 task
->lock_depth
= -1;
4761 local_irq_disable();
4762 task
->lock_depth
= saved_lock_depth
;
4763 sub_preempt_count(PREEMPT_ACTIVE
);
4766 * Check again in case we missed a preemption opportunity
4767 * between schedule and now.
4770 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4773 #endif /* CONFIG_PREEMPT */
4775 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4778 return try_to_wake_up(curr
->private, mode
, sync
);
4780 EXPORT_SYMBOL(default_wake_function
);
4783 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4784 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4785 * number) then we wake all the non-exclusive tasks and one exclusive task.
4787 * There are circumstances in which we can try to wake a task which has already
4788 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4789 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4791 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4792 int nr_exclusive
, int sync
, void *key
)
4794 wait_queue_t
*curr
, *next
;
4796 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4797 unsigned flags
= curr
->flags
;
4799 if (curr
->func(curr
, mode
, sync
, key
) &&
4800 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4806 * __wake_up - wake up threads blocked on a waitqueue.
4808 * @mode: which threads
4809 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4810 * @key: is directly passed to the wakeup function
4812 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4813 int nr_exclusive
, void *key
)
4815 unsigned long flags
;
4817 spin_lock_irqsave(&q
->lock
, flags
);
4818 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4819 spin_unlock_irqrestore(&q
->lock
, flags
);
4821 EXPORT_SYMBOL(__wake_up
);
4824 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4826 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4828 __wake_up_common(q
, mode
, 1, 0, NULL
);
4832 * __wake_up_sync - wake up threads blocked on a waitqueue.
4834 * @mode: which threads
4835 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4837 * The sync wakeup differs that the waker knows that it will schedule
4838 * away soon, so while the target thread will be woken up, it will not
4839 * be migrated to another CPU - ie. the two threads are 'synchronized'
4840 * with each other. This can prevent needless bouncing between CPUs.
4842 * On UP it can prevent extra preemption.
4845 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4847 unsigned long flags
;
4853 if (unlikely(!nr_exclusive
))
4856 spin_lock_irqsave(&q
->lock
, flags
);
4857 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4858 spin_unlock_irqrestore(&q
->lock
, flags
);
4860 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4862 void complete(struct completion
*x
)
4864 unsigned long flags
;
4866 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4868 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4869 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4871 EXPORT_SYMBOL(complete
);
4873 void complete_all(struct completion
*x
)
4875 unsigned long flags
;
4877 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4878 x
->done
+= UINT_MAX
/2;
4879 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4880 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4882 EXPORT_SYMBOL(complete_all
);
4884 static inline long __sched
4885 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4888 DECLARE_WAITQUEUE(wait
, current
);
4890 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4891 __add_wait_queue_tail(&x
->wait
, &wait
);
4893 if ((state
== TASK_INTERRUPTIBLE
&&
4894 signal_pending(current
)) ||
4895 (state
== TASK_KILLABLE
&&
4896 fatal_signal_pending(current
))) {
4897 __remove_wait_queue(&x
->wait
, &wait
);
4898 return -ERESTARTSYS
;
4900 __set_current_state(state
);
4901 spin_unlock_irq(&x
->wait
.lock
);
4902 timeout
= schedule_timeout(timeout
);
4903 spin_lock_irq(&x
->wait
.lock
);
4905 __remove_wait_queue(&x
->wait
, &wait
);
4909 __remove_wait_queue(&x
->wait
, &wait
);
4916 wait_for_common(struct completion
*x
, long timeout
, int state
)
4920 spin_lock_irq(&x
->wait
.lock
);
4921 timeout
= do_wait_for_common(x
, timeout
, state
);
4922 spin_unlock_irq(&x
->wait
.lock
);
4926 void __sched
wait_for_completion(struct completion
*x
)
4928 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4930 EXPORT_SYMBOL(wait_for_completion
);
4932 unsigned long __sched
4933 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4935 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4937 EXPORT_SYMBOL(wait_for_completion_timeout
);
4939 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4941 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4942 if (t
== -ERESTARTSYS
)
4946 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4948 unsigned long __sched
4949 wait_for_completion_interruptible_timeout(struct completion
*x
,
4950 unsigned long timeout
)
4952 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4954 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4956 int __sched
wait_for_completion_killable(struct completion
*x
)
4958 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4959 if (t
== -ERESTARTSYS
)
4963 EXPORT_SYMBOL(wait_for_completion_killable
);
4966 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4968 unsigned long flags
;
4971 init_waitqueue_entry(&wait
, current
);
4973 __set_current_state(state
);
4975 spin_lock_irqsave(&q
->lock
, flags
);
4976 __add_wait_queue(q
, &wait
);
4977 spin_unlock(&q
->lock
);
4978 timeout
= schedule_timeout(timeout
);
4979 spin_lock_irq(&q
->lock
);
4980 __remove_wait_queue(q
, &wait
);
4981 spin_unlock_irqrestore(&q
->lock
, flags
);
4986 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4988 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4990 EXPORT_SYMBOL(interruptible_sleep_on
);
4993 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4995 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4997 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4999 void __sched
sleep_on(wait_queue_head_t
*q
)
5001 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5003 EXPORT_SYMBOL(sleep_on
);
5005 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5007 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5009 EXPORT_SYMBOL(sleep_on_timeout
);
5011 #ifdef CONFIG_RT_MUTEXES
5014 * rt_mutex_setprio - set the current priority of a task
5016 * @prio: prio value (kernel-internal form)
5018 * This function changes the 'effective' priority of a task. It does
5019 * not touch ->normal_prio like __setscheduler().
5021 * Used by the rt_mutex code to implement priority inheritance logic.
5023 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5025 unsigned long flags
;
5026 int oldprio
, on_rq
, running
;
5028 const struct sched_class
*prev_class
= p
->sched_class
;
5030 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5032 rq
= task_rq_lock(p
, &flags
);
5033 update_rq_clock(rq
);
5036 on_rq
= p
->se
.on_rq
;
5037 running
= task_current(rq
, p
);
5039 dequeue_task(rq
, p
, 0);
5041 p
->sched_class
->put_prev_task(rq
, p
);
5044 p
->sched_class
= &rt_sched_class
;
5046 p
->sched_class
= &fair_sched_class
;
5051 p
->sched_class
->set_curr_task(rq
);
5053 enqueue_task(rq
, p
, 0);
5055 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5057 task_rq_unlock(rq
, &flags
);
5062 void set_user_nice(struct task_struct
*p
, long nice
)
5064 int old_prio
, delta
, on_rq
;
5065 unsigned long flags
;
5068 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5071 * We have to be careful, if called from sys_setpriority(),
5072 * the task might be in the middle of scheduling on another CPU.
5074 rq
= task_rq_lock(p
, &flags
);
5075 update_rq_clock(rq
);
5077 * The RT priorities are set via sched_setscheduler(), but we still
5078 * allow the 'normal' nice value to be set - but as expected
5079 * it wont have any effect on scheduling until the task is
5080 * SCHED_FIFO/SCHED_RR:
5082 if (task_has_rt_policy(p
)) {
5083 p
->static_prio
= NICE_TO_PRIO(nice
);
5086 on_rq
= p
->se
.on_rq
;
5088 dequeue_task(rq
, p
, 0);
5090 p
->static_prio
= NICE_TO_PRIO(nice
);
5093 p
->prio
= effective_prio(p
);
5094 delta
= p
->prio
- old_prio
;
5097 enqueue_task(rq
, p
, 0);
5099 * If the task increased its priority or is running and
5100 * lowered its priority, then reschedule its CPU:
5102 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5103 resched_task(rq
->curr
);
5106 task_rq_unlock(rq
, &flags
);
5108 EXPORT_SYMBOL(set_user_nice
);
5111 * can_nice - check if a task can reduce its nice value
5115 int can_nice(const struct task_struct
*p
, const int nice
)
5117 /* convert nice value [19,-20] to rlimit style value [1,40] */
5118 int nice_rlim
= 20 - nice
;
5120 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5121 capable(CAP_SYS_NICE
));
5124 #ifdef __ARCH_WANT_SYS_NICE
5127 * sys_nice - change the priority of the current process.
5128 * @increment: priority increment
5130 * sys_setpriority is a more generic, but much slower function that
5131 * does similar things.
5133 asmlinkage
long sys_nice(int increment
)
5138 * Setpriority might change our priority at the same moment.
5139 * We don't have to worry. Conceptually one call occurs first
5140 * and we have a single winner.
5142 if (increment
< -40)
5147 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5153 if (increment
< 0 && !can_nice(current
, nice
))
5156 retval
= security_task_setnice(current
, nice
);
5160 set_user_nice(current
, nice
);
5167 * task_prio - return the priority value of a given task.
5168 * @p: the task in question.
5170 * This is the priority value as seen by users in /proc.
5171 * RT tasks are offset by -200. Normal tasks are centered
5172 * around 0, value goes from -16 to +15.
5174 int task_prio(const struct task_struct
*p
)
5176 return p
->prio
- MAX_RT_PRIO
;
5180 * task_nice - return the nice value of a given task.
5181 * @p: the task in question.
5183 int task_nice(const struct task_struct
*p
)
5185 return TASK_NICE(p
);
5187 EXPORT_SYMBOL(task_nice
);
5190 * idle_cpu - is a given cpu idle currently?
5191 * @cpu: the processor in question.
5193 int idle_cpu(int cpu
)
5195 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5199 * idle_task - return the idle task for a given cpu.
5200 * @cpu: the processor in question.
5202 struct task_struct
*idle_task(int cpu
)
5204 return cpu_rq(cpu
)->idle
;
5208 * find_process_by_pid - find a process with a matching PID value.
5209 * @pid: the pid in question.
5211 static struct task_struct
*find_process_by_pid(pid_t pid
)
5213 return pid
? find_task_by_vpid(pid
) : current
;
5216 /* Actually do priority change: must hold rq lock. */
5218 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5220 BUG_ON(p
->se
.on_rq
);
5223 switch (p
->policy
) {
5227 p
->sched_class
= &fair_sched_class
;
5231 p
->sched_class
= &rt_sched_class
;
5235 p
->rt_priority
= prio
;
5236 p
->normal_prio
= normal_prio(p
);
5237 /* we are holding p->pi_lock already */
5238 p
->prio
= rt_mutex_getprio(p
);
5243 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5244 * @p: the task in question.
5245 * @policy: new policy.
5246 * @param: structure containing the new RT priority.
5248 * NOTE that the task may be already dead.
5250 int sched_setscheduler(struct task_struct
*p
, int policy
,
5251 struct sched_param
*param
)
5253 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5254 unsigned long flags
;
5255 const struct sched_class
*prev_class
= p
->sched_class
;
5258 /* may grab non-irq protected spin_locks */
5259 BUG_ON(in_interrupt());
5261 /* double check policy once rq lock held */
5263 policy
= oldpolicy
= p
->policy
;
5264 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5265 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5266 policy
!= SCHED_IDLE
)
5269 * Valid priorities for SCHED_FIFO and SCHED_RR are
5270 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5271 * SCHED_BATCH and SCHED_IDLE is 0.
5273 if (param
->sched_priority
< 0 ||
5274 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5275 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5277 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5281 * Allow unprivileged RT tasks to decrease priority:
5283 if (!capable(CAP_SYS_NICE
)) {
5284 if (rt_policy(policy
)) {
5285 unsigned long rlim_rtprio
;
5287 if (!lock_task_sighand(p
, &flags
))
5289 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5290 unlock_task_sighand(p
, &flags
);
5292 /* can't set/change the rt policy */
5293 if (policy
!= p
->policy
&& !rlim_rtprio
)
5296 /* can't increase priority */
5297 if (param
->sched_priority
> p
->rt_priority
&&
5298 param
->sched_priority
> rlim_rtprio
)
5302 * Like positive nice levels, dont allow tasks to
5303 * move out of SCHED_IDLE either:
5305 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5308 /* can't change other user's priorities */
5309 if ((current
->euid
!= p
->euid
) &&
5310 (current
->euid
!= p
->uid
))
5314 #ifdef CONFIG_RT_GROUP_SCHED
5316 * Do not allow realtime tasks into groups that have no runtime
5319 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5323 retval
= security_task_setscheduler(p
, policy
, param
);
5327 * make sure no PI-waiters arrive (or leave) while we are
5328 * changing the priority of the task:
5330 spin_lock_irqsave(&p
->pi_lock
, flags
);
5332 * To be able to change p->policy safely, the apropriate
5333 * runqueue lock must be held.
5335 rq
= __task_rq_lock(p
);
5336 /* recheck policy now with rq lock held */
5337 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5338 policy
= oldpolicy
= -1;
5339 __task_rq_unlock(rq
);
5340 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5343 update_rq_clock(rq
);
5344 on_rq
= p
->se
.on_rq
;
5345 running
= task_current(rq
, p
);
5347 deactivate_task(rq
, p
, 0);
5349 p
->sched_class
->put_prev_task(rq
, p
);
5352 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5355 p
->sched_class
->set_curr_task(rq
);
5357 activate_task(rq
, p
, 0);
5359 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5361 __task_rq_unlock(rq
);
5362 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5364 rt_mutex_adjust_pi(p
);
5368 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5371 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5373 struct sched_param lparam
;
5374 struct task_struct
*p
;
5377 if (!param
|| pid
< 0)
5379 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5384 p
= find_process_by_pid(pid
);
5386 retval
= sched_setscheduler(p
, policy
, &lparam
);
5393 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5394 * @pid: the pid in question.
5395 * @policy: new policy.
5396 * @param: structure containing the new RT priority.
5399 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5401 /* negative values for policy are not valid */
5405 return do_sched_setscheduler(pid
, policy
, param
);
5409 * sys_sched_setparam - set/change the RT priority of a thread
5410 * @pid: the pid in question.
5411 * @param: structure containing the new RT priority.
5413 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5415 return do_sched_setscheduler(pid
, -1, param
);
5419 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5420 * @pid: the pid in question.
5422 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5424 struct task_struct
*p
;
5431 read_lock(&tasklist_lock
);
5432 p
= find_process_by_pid(pid
);
5434 retval
= security_task_getscheduler(p
);
5438 read_unlock(&tasklist_lock
);
5443 * sys_sched_getscheduler - get the RT priority of a thread
5444 * @pid: the pid in question.
5445 * @param: structure containing the RT priority.
5447 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5449 struct sched_param lp
;
5450 struct task_struct
*p
;
5453 if (!param
|| pid
< 0)
5456 read_lock(&tasklist_lock
);
5457 p
= find_process_by_pid(pid
);
5462 retval
= security_task_getscheduler(p
);
5466 lp
.sched_priority
= p
->rt_priority
;
5467 read_unlock(&tasklist_lock
);
5470 * This one might sleep, we cannot do it with a spinlock held ...
5472 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5477 read_unlock(&tasklist_lock
);
5481 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5483 cpumask_t cpus_allowed
;
5484 cpumask_t new_mask
= *in_mask
;
5485 struct task_struct
*p
;
5489 read_lock(&tasklist_lock
);
5491 p
= find_process_by_pid(pid
);
5493 read_unlock(&tasklist_lock
);
5499 * It is not safe to call set_cpus_allowed with the
5500 * tasklist_lock held. We will bump the task_struct's
5501 * usage count and then drop tasklist_lock.
5504 read_unlock(&tasklist_lock
);
5507 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5508 !capable(CAP_SYS_NICE
))
5511 retval
= security_task_setscheduler(p
, 0, NULL
);
5515 cpuset_cpus_allowed(p
, &cpus_allowed
);
5516 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5518 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5521 cpuset_cpus_allowed(p
, &cpus_allowed
);
5522 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5524 * We must have raced with a concurrent cpuset
5525 * update. Just reset the cpus_allowed to the
5526 * cpuset's cpus_allowed
5528 new_mask
= cpus_allowed
;
5538 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5539 cpumask_t
*new_mask
)
5541 if (len
< sizeof(cpumask_t
)) {
5542 memset(new_mask
, 0, sizeof(cpumask_t
));
5543 } else if (len
> sizeof(cpumask_t
)) {
5544 len
= sizeof(cpumask_t
);
5546 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5550 * sys_sched_setaffinity - set the cpu affinity of a process
5551 * @pid: pid of the process
5552 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5553 * @user_mask_ptr: user-space pointer to the new cpu mask
5555 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5556 unsigned long __user
*user_mask_ptr
)
5561 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5565 return sched_setaffinity(pid
, &new_mask
);
5569 * Represents all cpu's present in the system
5570 * In systems capable of hotplug, this map could dynamically grow
5571 * as new cpu's are detected in the system via any platform specific
5572 * method, such as ACPI for e.g.
5575 cpumask_t cpu_present_map __read_mostly
;
5576 EXPORT_SYMBOL(cpu_present_map
);
5579 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
5580 EXPORT_SYMBOL(cpu_online_map
);
5582 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
5583 EXPORT_SYMBOL(cpu_possible_map
);
5586 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5588 struct task_struct
*p
;
5592 read_lock(&tasklist_lock
);
5595 p
= find_process_by_pid(pid
);
5599 retval
= security_task_getscheduler(p
);
5603 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5606 read_unlock(&tasklist_lock
);
5613 * sys_sched_getaffinity - get the cpu affinity of a process
5614 * @pid: pid of the process
5615 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5616 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5618 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5619 unsigned long __user
*user_mask_ptr
)
5624 if (len
< sizeof(cpumask_t
))
5627 ret
= sched_getaffinity(pid
, &mask
);
5631 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5634 return sizeof(cpumask_t
);
5638 * sys_sched_yield - yield the current processor to other threads.
5640 * This function yields the current CPU to other tasks. If there are no
5641 * other threads running on this CPU then this function will return.
5643 asmlinkage
long sys_sched_yield(void)
5645 struct rq
*rq
= this_rq_lock();
5647 schedstat_inc(rq
, yld_count
);
5648 current
->sched_class
->yield_task(rq
);
5651 * Since we are going to call schedule() anyway, there's
5652 * no need to preempt or enable interrupts:
5654 __release(rq
->lock
);
5655 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5656 _raw_spin_unlock(&rq
->lock
);
5657 preempt_enable_no_resched();
5664 static void __cond_resched(void)
5666 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5667 __might_sleep(__FILE__
, __LINE__
);
5670 * The BKS might be reacquired before we have dropped
5671 * PREEMPT_ACTIVE, which could trigger a second
5672 * cond_resched() call.
5675 add_preempt_count(PREEMPT_ACTIVE
);
5677 sub_preempt_count(PREEMPT_ACTIVE
);
5678 } while (need_resched());
5681 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5682 int __sched
_cond_resched(void)
5684 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5685 system_state
== SYSTEM_RUNNING
) {
5691 EXPORT_SYMBOL(_cond_resched
);
5695 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5696 * call schedule, and on return reacquire the lock.
5698 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5699 * operations here to prevent schedule() from being called twice (once via
5700 * spin_unlock(), once by hand).
5702 int cond_resched_lock(spinlock_t
*lock
)
5704 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5707 if (spin_needbreak(lock
) || resched
) {
5709 if (resched
&& need_resched())
5718 EXPORT_SYMBOL(cond_resched_lock
);
5720 int __sched
cond_resched_softirq(void)
5722 BUG_ON(!in_softirq());
5724 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5732 EXPORT_SYMBOL(cond_resched_softirq
);
5735 * yield - yield the current processor to other threads.
5737 * This is a shortcut for kernel-space yielding - it marks the
5738 * thread runnable and calls sys_sched_yield().
5740 void __sched
yield(void)
5742 set_current_state(TASK_RUNNING
);
5745 EXPORT_SYMBOL(yield
);
5748 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5749 * that process accounting knows that this is a task in IO wait state.
5751 * But don't do that if it is a deliberate, throttling IO wait (this task
5752 * has set its backing_dev_info: the queue against which it should throttle)
5754 void __sched
io_schedule(void)
5756 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5758 delayacct_blkio_start();
5759 atomic_inc(&rq
->nr_iowait
);
5761 atomic_dec(&rq
->nr_iowait
);
5762 delayacct_blkio_end();
5764 EXPORT_SYMBOL(io_schedule
);
5766 long __sched
io_schedule_timeout(long timeout
)
5768 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5771 delayacct_blkio_start();
5772 atomic_inc(&rq
->nr_iowait
);
5773 ret
= schedule_timeout(timeout
);
5774 atomic_dec(&rq
->nr_iowait
);
5775 delayacct_blkio_end();
5780 * sys_sched_get_priority_max - return maximum RT priority.
5781 * @policy: scheduling class.
5783 * this syscall returns the maximum rt_priority that can be used
5784 * by a given scheduling class.
5786 asmlinkage
long sys_sched_get_priority_max(int policy
)
5793 ret
= MAX_USER_RT_PRIO
-1;
5805 * sys_sched_get_priority_min - return minimum RT priority.
5806 * @policy: scheduling class.
5808 * this syscall returns the minimum rt_priority that can be used
5809 * by a given scheduling class.
5811 asmlinkage
long sys_sched_get_priority_min(int policy
)
5829 * sys_sched_rr_get_interval - return the default timeslice of a process.
5830 * @pid: pid of the process.
5831 * @interval: userspace pointer to the timeslice value.
5833 * this syscall writes the default timeslice value of a given process
5834 * into the user-space timespec buffer. A value of '0' means infinity.
5837 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5839 struct task_struct
*p
;
5840 unsigned int time_slice
;
5848 read_lock(&tasklist_lock
);
5849 p
= find_process_by_pid(pid
);
5853 retval
= security_task_getscheduler(p
);
5858 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5859 * tasks that are on an otherwise idle runqueue:
5862 if (p
->policy
== SCHED_RR
) {
5863 time_slice
= DEF_TIMESLICE
;
5864 } else if (p
->policy
!= SCHED_FIFO
) {
5865 struct sched_entity
*se
= &p
->se
;
5866 unsigned long flags
;
5869 rq
= task_rq_lock(p
, &flags
);
5870 if (rq
->cfs
.load
.weight
)
5871 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5872 task_rq_unlock(rq
, &flags
);
5874 read_unlock(&tasklist_lock
);
5875 jiffies_to_timespec(time_slice
, &t
);
5876 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5880 read_unlock(&tasklist_lock
);
5884 static const char stat_nam
[] = "RSDTtZX";
5886 void sched_show_task(struct task_struct
*p
)
5888 unsigned long free
= 0;
5891 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5892 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5893 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5894 #if BITS_PER_LONG == 32
5895 if (state
== TASK_RUNNING
)
5896 printk(KERN_CONT
" running ");
5898 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5900 if (state
== TASK_RUNNING
)
5901 printk(KERN_CONT
" running task ");
5903 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5905 #ifdef CONFIG_DEBUG_STACK_USAGE
5907 unsigned long *n
= end_of_stack(p
);
5910 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5913 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5914 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5916 show_stack(p
, NULL
);
5919 void show_state_filter(unsigned long state_filter
)
5921 struct task_struct
*g
, *p
;
5923 #if BITS_PER_LONG == 32
5925 " task PC stack pid father\n");
5928 " task PC stack pid father\n");
5930 read_lock(&tasklist_lock
);
5931 do_each_thread(g
, p
) {
5933 * reset the NMI-timeout, listing all files on a slow
5934 * console might take alot of time:
5936 touch_nmi_watchdog();
5937 if (!state_filter
|| (p
->state
& state_filter
))
5939 } while_each_thread(g
, p
);
5941 touch_all_softlockup_watchdogs();
5943 #ifdef CONFIG_SCHED_DEBUG
5944 sysrq_sched_debug_show();
5946 read_unlock(&tasklist_lock
);
5948 * Only show locks if all tasks are dumped:
5950 if (state_filter
== -1)
5951 debug_show_all_locks();
5954 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5956 idle
->sched_class
= &idle_sched_class
;
5960 * init_idle - set up an idle thread for a given CPU
5961 * @idle: task in question
5962 * @cpu: cpu the idle task belongs to
5964 * NOTE: this function does not set the idle thread's NEED_RESCHED
5965 * flag, to make booting more robust.
5967 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5969 struct rq
*rq
= cpu_rq(cpu
);
5970 unsigned long flags
;
5973 idle
->se
.exec_start
= sched_clock();
5975 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5976 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5977 __set_task_cpu(idle
, cpu
);
5979 spin_lock_irqsave(&rq
->lock
, flags
);
5980 rq
->curr
= rq
->idle
= idle
;
5981 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5984 spin_unlock_irqrestore(&rq
->lock
, flags
);
5986 /* Set the preempt count _outside_ the spinlocks! */
5987 task_thread_info(idle
)->preempt_count
= 0;
5990 * The idle tasks have their own, simple scheduling class:
5992 idle
->sched_class
= &idle_sched_class
;
5996 * In a system that switches off the HZ timer nohz_cpu_mask
5997 * indicates which cpus entered this state. This is used
5998 * in the rcu update to wait only for active cpus. For system
5999 * which do not switch off the HZ timer nohz_cpu_mask should
6000 * always be CPU_MASK_NONE.
6002 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
6005 * Increase the granularity value when there are more CPUs,
6006 * because with more CPUs the 'effective latency' as visible
6007 * to users decreases. But the relationship is not linear,
6008 * so pick a second-best guess by going with the log2 of the
6011 * This idea comes from the SD scheduler of Con Kolivas:
6013 static inline void sched_init_granularity(void)
6015 unsigned int factor
= 1 + ilog2(num_online_cpus());
6016 const unsigned long limit
= 200000000;
6018 sysctl_sched_min_granularity
*= factor
;
6019 if (sysctl_sched_min_granularity
> limit
)
6020 sysctl_sched_min_granularity
= limit
;
6022 sysctl_sched_latency
*= factor
;
6023 if (sysctl_sched_latency
> limit
)
6024 sysctl_sched_latency
= limit
;
6026 sysctl_sched_wakeup_granularity
*= factor
;
6031 * This is how migration works:
6033 * 1) we queue a struct migration_req structure in the source CPU's
6034 * runqueue and wake up that CPU's migration thread.
6035 * 2) we down() the locked semaphore => thread blocks.
6036 * 3) migration thread wakes up (implicitly it forces the migrated
6037 * thread off the CPU)
6038 * 4) it gets the migration request and checks whether the migrated
6039 * task is still in the wrong runqueue.
6040 * 5) if it's in the wrong runqueue then the migration thread removes
6041 * it and puts it into the right queue.
6042 * 6) migration thread up()s the semaphore.
6043 * 7) we wake up and the migration is done.
6047 * Change a given task's CPU affinity. Migrate the thread to a
6048 * proper CPU and schedule it away if the CPU it's executing on
6049 * is removed from the allowed bitmask.
6051 * NOTE: the caller must have a valid reference to the task, the
6052 * task must not exit() & deallocate itself prematurely. The
6053 * call is not atomic; no spinlocks may be held.
6055 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
6057 struct migration_req req
;
6058 unsigned long flags
;
6062 rq
= task_rq_lock(p
, &flags
);
6063 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
6068 if (p
->sched_class
->set_cpus_allowed
)
6069 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6071 p
->cpus_allowed
= *new_mask
;
6072 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
6075 /* Can the task run on the task's current CPU? If so, we're done */
6076 if (cpu_isset(task_cpu(p
), *new_mask
))
6079 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
6080 /* Need help from migration thread: drop lock and wait. */
6081 task_rq_unlock(rq
, &flags
);
6082 wake_up_process(rq
->migration_thread
);
6083 wait_for_completion(&req
.done
);
6084 tlb_migrate_finish(p
->mm
);
6088 task_rq_unlock(rq
, &flags
);
6092 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6095 * Move (not current) task off this cpu, onto dest cpu. We're doing
6096 * this because either it can't run here any more (set_cpus_allowed()
6097 * away from this CPU, or CPU going down), or because we're
6098 * attempting to rebalance this task on exec (sched_exec).
6100 * So we race with normal scheduler movements, but that's OK, as long
6101 * as the task is no longer on this CPU.
6103 * Returns non-zero if task was successfully migrated.
6105 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6107 struct rq
*rq_dest
, *rq_src
;
6110 if (unlikely(cpu_is_offline(dest_cpu
)))
6113 rq_src
= cpu_rq(src_cpu
);
6114 rq_dest
= cpu_rq(dest_cpu
);
6116 double_rq_lock(rq_src
, rq_dest
);
6117 /* Already moved. */
6118 if (task_cpu(p
) != src_cpu
)
6120 /* Affinity changed (again). */
6121 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6124 on_rq
= p
->se
.on_rq
;
6126 deactivate_task(rq_src
, p
, 0);
6128 set_task_cpu(p
, dest_cpu
);
6130 activate_task(rq_dest
, p
, 0);
6131 check_preempt_curr(rq_dest
, p
);
6135 double_rq_unlock(rq_src
, rq_dest
);
6140 * migration_thread - this is a highprio system thread that performs
6141 * thread migration by bumping thread off CPU then 'pushing' onto
6144 static int migration_thread(void *data
)
6146 int cpu
= (long)data
;
6150 BUG_ON(rq
->migration_thread
!= current
);
6152 set_current_state(TASK_INTERRUPTIBLE
);
6153 while (!kthread_should_stop()) {
6154 struct migration_req
*req
;
6155 struct list_head
*head
;
6157 spin_lock_irq(&rq
->lock
);
6159 if (cpu_is_offline(cpu
)) {
6160 spin_unlock_irq(&rq
->lock
);
6164 if (rq
->active_balance
) {
6165 active_load_balance(rq
, cpu
);
6166 rq
->active_balance
= 0;
6169 head
= &rq
->migration_queue
;
6171 if (list_empty(head
)) {
6172 spin_unlock_irq(&rq
->lock
);
6174 set_current_state(TASK_INTERRUPTIBLE
);
6177 req
= list_entry(head
->next
, struct migration_req
, list
);
6178 list_del_init(head
->next
);
6180 spin_unlock(&rq
->lock
);
6181 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6184 complete(&req
->done
);
6186 __set_current_state(TASK_RUNNING
);
6190 /* Wait for kthread_stop */
6191 set_current_state(TASK_INTERRUPTIBLE
);
6192 while (!kthread_should_stop()) {
6194 set_current_state(TASK_INTERRUPTIBLE
);
6196 __set_current_state(TASK_RUNNING
);
6200 #ifdef CONFIG_HOTPLUG_CPU
6202 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6206 local_irq_disable();
6207 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6213 * Figure out where task on dead CPU should go, use force if necessary.
6214 * NOTE: interrupts should be disabled by the caller
6216 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6218 unsigned long flags
;
6225 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6226 cpus_and(mask
, mask
, p
->cpus_allowed
);
6227 dest_cpu
= any_online_cpu(mask
);
6229 /* On any allowed CPU? */
6230 if (dest_cpu
>= nr_cpu_ids
)
6231 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6233 /* No more Mr. Nice Guy. */
6234 if (dest_cpu
>= nr_cpu_ids
) {
6235 cpumask_t cpus_allowed
;
6237 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6239 * Try to stay on the same cpuset, where the
6240 * current cpuset may be a subset of all cpus.
6241 * The cpuset_cpus_allowed_locked() variant of
6242 * cpuset_cpus_allowed() will not block. It must be
6243 * called within calls to cpuset_lock/cpuset_unlock.
6245 rq
= task_rq_lock(p
, &flags
);
6246 p
->cpus_allowed
= cpus_allowed
;
6247 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6248 task_rq_unlock(rq
, &flags
);
6251 * Don't tell them about moving exiting tasks or
6252 * kernel threads (both mm NULL), since they never
6255 if (p
->mm
&& printk_ratelimit()) {
6256 printk(KERN_INFO
"process %d (%s) no "
6257 "longer affine to cpu%d\n",
6258 task_pid_nr(p
), p
->comm
, dead_cpu
);
6261 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6265 * While a dead CPU has no uninterruptible tasks queued at this point,
6266 * it might still have a nonzero ->nr_uninterruptible counter, because
6267 * for performance reasons the counter is not stricly tracking tasks to
6268 * their home CPUs. So we just add the counter to another CPU's counter,
6269 * to keep the global sum constant after CPU-down:
6271 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6273 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6274 unsigned long flags
;
6276 local_irq_save(flags
);
6277 double_rq_lock(rq_src
, rq_dest
);
6278 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6279 rq_src
->nr_uninterruptible
= 0;
6280 double_rq_unlock(rq_src
, rq_dest
);
6281 local_irq_restore(flags
);
6284 /* Run through task list and migrate tasks from the dead cpu. */
6285 static void migrate_live_tasks(int src_cpu
)
6287 struct task_struct
*p
, *t
;
6289 read_lock(&tasklist_lock
);
6291 do_each_thread(t
, p
) {
6295 if (task_cpu(p
) == src_cpu
)
6296 move_task_off_dead_cpu(src_cpu
, p
);
6297 } while_each_thread(t
, p
);
6299 read_unlock(&tasklist_lock
);
6303 * Schedules idle task to be the next runnable task on current CPU.
6304 * It does so by boosting its priority to highest possible.
6305 * Used by CPU offline code.
6307 void sched_idle_next(void)
6309 int this_cpu
= smp_processor_id();
6310 struct rq
*rq
= cpu_rq(this_cpu
);
6311 struct task_struct
*p
= rq
->idle
;
6312 unsigned long flags
;
6314 /* cpu has to be offline */
6315 BUG_ON(cpu_online(this_cpu
));
6318 * Strictly not necessary since rest of the CPUs are stopped by now
6319 * and interrupts disabled on the current cpu.
6321 spin_lock_irqsave(&rq
->lock
, flags
);
6323 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6325 update_rq_clock(rq
);
6326 activate_task(rq
, p
, 0);
6328 spin_unlock_irqrestore(&rq
->lock
, flags
);
6332 * Ensures that the idle task is using init_mm right before its cpu goes
6335 void idle_task_exit(void)
6337 struct mm_struct
*mm
= current
->active_mm
;
6339 BUG_ON(cpu_online(smp_processor_id()));
6342 switch_mm(mm
, &init_mm
, current
);
6346 /* called under rq->lock with disabled interrupts */
6347 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6349 struct rq
*rq
= cpu_rq(dead_cpu
);
6351 /* Must be exiting, otherwise would be on tasklist. */
6352 BUG_ON(!p
->exit_state
);
6354 /* Cannot have done final schedule yet: would have vanished. */
6355 BUG_ON(p
->state
== TASK_DEAD
);
6360 * Drop lock around migration; if someone else moves it,
6361 * that's OK. No task can be added to this CPU, so iteration is
6364 spin_unlock_irq(&rq
->lock
);
6365 move_task_off_dead_cpu(dead_cpu
, p
);
6366 spin_lock_irq(&rq
->lock
);
6371 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6372 static void migrate_dead_tasks(unsigned int dead_cpu
)
6374 struct rq
*rq
= cpu_rq(dead_cpu
);
6375 struct task_struct
*next
;
6378 if (!rq
->nr_running
)
6380 update_rq_clock(rq
);
6381 next
= pick_next_task(rq
, rq
->curr
);
6384 migrate_dead(dead_cpu
, next
);
6388 #endif /* CONFIG_HOTPLUG_CPU */
6390 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6392 static struct ctl_table sd_ctl_dir
[] = {
6394 .procname
= "sched_domain",
6400 static struct ctl_table sd_ctl_root
[] = {
6402 .ctl_name
= CTL_KERN
,
6403 .procname
= "kernel",
6405 .child
= sd_ctl_dir
,
6410 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6412 struct ctl_table
*entry
=
6413 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6418 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6420 struct ctl_table
*entry
;
6423 * In the intermediate directories, both the child directory and
6424 * procname are dynamically allocated and could fail but the mode
6425 * will always be set. In the lowest directory the names are
6426 * static strings and all have proc handlers.
6428 for (entry
= *tablep
; entry
->mode
; entry
++) {
6430 sd_free_ctl_entry(&entry
->child
);
6431 if (entry
->proc_handler
== NULL
)
6432 kfree(entry
->procname
);
6440 set_table_entry(struct ctl_table
*entry
,
6441 const char *procname
, void *data
, int maxlen
,
6442 mode_t mode
, proc_handler
*proc_handler
)
6444 entry
->procname
= procname
;
6446 entry
->maxlen
= maxlen
;
6448 entry
->proc_handler
= proc_handler
;
6451 static struct ctl_table
*
6452 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6454 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6459 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6460 sizeof(long), 0644, proc_doulongvec_minmax
);
6461 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6462 sizeof(long), 0644, proc_doulongvec_minmax
);
6463 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6464 sizeof(int), 0644, proc_dointvec_minmax
);
6465 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6466 sizeof(int), 0644, proc_dointvec_minmax
);
6467 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6468 sizeof(int), 0644, proc_dointvec_minmax
);
6469 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6470 sizeof(int), 0644, proc_dointvec_minmax
);
6471 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6472 sizeof(int), 0644, proc_dointvec_minmax
);
6473 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6474 sizeof(int), 0644, proc_dointvec_minmax
);
6475 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6476 sizeof(int), 0644, proc_dointvec_minmax
);
6477 set_table_entry(&table
[9], "cache_nice_tries",
6478 &sd
->cache_nice_tries
,
6479 sizeof(int), 0644, proc_dointvec_minmax
);
6480 set_table_entry(&table
[10], "flags", &sd
->flags
,
6481 sizeof(int), 0644, proc_dointvec_minmax
);
6482 /* &table[11] is terminator */
6487 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6489 struct ctl_table
*entry
, *table
;
6490 struct sched_domain
*sd
;
6491 int domain_num
= 0, i
;
6494 for_each_domain(cpu
, sd
)
6496 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6501 for_each_domain(cpu
, sd
) {
6502 snprintf(buf
, 32, "domain%d", i
);
6503 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6505 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6512 static struct ctl_table_header
*sd_sysctl_header
;
6513 static void register_sched_domain_sysctl(void)
6515 int i
, cpu_num
= num_online_cpus();
6516 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6519 WARN_ON(sd_ctl_dir
[0].child
);
6520 sd_ctl_dir
[0].child
= entry
;
6525 for_each_online_cpu(i
) {
6526 snprintf(buf
, 32, "cpu%d", i
);
6527 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6529 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6533 WARN_ON(sd_sysctl_header
);
6534 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6537 /* may be called multiple times per register */
6538 static void unregister_sched_domain_sysctl(void)
6540 if (sd_sysctl_header
)
6541 unregister_sysctl_table(sd_sysctl_header
);
6542 sd_sysctl_header
= NULL
;
6543 if (sd_ctl_dir
[0].child
)
6544 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6547 static void register_sched_domain_sysctl(void)
6550 static void unregister_sched_domain_sysctl(void)
6556 * migration_call - callback that gets triggered when a CPU is added.
6557 * Here we can start up the necessary migration thread for the new CPU.
6559 static int __cpuinit
6560 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6562 struct task_struct
*p
;
6563 int cpu
= (long)hcpu
;
6564 unsigned long flags
;
6569 case CPU_UP_PREPARE
:
6570 case CPU_UP_PREPARE_FROZEN
:
6571 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6574 kthread_bind(p
, cpu
);
6575 /* Must be high prio: stop_machine expects to yield to it. */
6576 rq
= task_rq_lock(p
, &flags
);
6577 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6578 task_rq_unlock(rq
, &flags
);
6579 cpu_rq(cpu
)->migration_thread
= p
;
6583 case CPU_ONLINE_FROZEN
:
6584 /* Strictly unnecessary, as first user will wake it. */
6585 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6587 /* Update our root-domain */
6589 spin_lock_irqsave(&rq
->lock
, flags
);
6591 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6592 cpu_set(cpu
, rq
->rd
->online
);
6594 spin_unlock_irqrestore(&rq
->lock
, flags
);
6597 #ifdef CONFIG_HOTPLUG_CPU
6598 case CPU_UP_CANCELED
:
6599 case CPU_UP_CANCELED_FROZEN
:
6600 if (!cpu_rq(cpu
)->migration_thread
)
6602 /* Unbind it from offline cpu so it can run. Fall thru. */
6603 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6604 any_online_cpu(cpu_online_map
));
6605 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6606 cpu_rq(cpu
)->migration_thread
= NULL
;
6610 case CPU_DEAD_FROZEN
:
6611 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6612 migrate_live_tasks(cpu
);
6614 kthread_stop(rq
->migration_thread
);
6615 rq
->migration_thread
= NULL
;
6616 /* Idle task back to normal (off runqueue, low prio) */
6617 spin_lock_irq(&rq
->lock
);
6618 update_rq_clock(rq
);
6619 deactivate_task(rq
, rq
->idle
, 0);
6620 rq
->idle
->static_prio
= MAX_PRIO
;
6621 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6622 rq
->idle
->sched_class
= &idle_sched_class
;
6623 migrate_dead_tasks(cpu
);
6624 spin_unlock_irq(&rq
->lock
);
6626 migrate_nr_uninterruptible(rq
);
6627 BUG_ON(rq
->nr_running
!= 0);
6630 * No need to migrate the tasks: it was best-effort if
6631 * they didn't take sched_hotcpu_mutex. Just wake up
6634 spin_lock_irq(&rq
->lock
);
6635 while (!list_empty(&rq
->migration_queue
)) {
6636 struct migration_req
*req
;
6638 req
= list_entry(rq
->migration_queue
.next
,
6639 struct migration_req
, list
);
6640 list_del_init(&req
->list
);
6641 complete(&req
->done
);
6643 spin_unlock_irq(&rq
->lock
);
6647 case CPU_DYING_FROZEN
:
6648 /* Update our root-domain */
6650 spin_lock_irqsave(&rq
->lock
, flags
);
6652 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6653 cpu_clear(cpu
, rq
->rd
->online
);
6655 spin_unlock_irqrestore(&rq
->lock
, flags
);
6662 /* Register at highest priority so that task migration (migrate_all_tasks)
6663 * happens before everything else.
6665 static struct notifier_block __cpuinitdata migration_notifier
= {
6666 .notifier_call
= migration_call
,
6670 void __init
migration_init(void)
6672 void *cpu
= (void *)(long)smp_processor_id();
6675 /* Start one for the boot CPU: */
6676 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6677 BUG_ON(err
== NOTIFY_BAD
);
6678 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6679 register_cpu_notifier(&migration_notifier
);
6685 #ifdef CONFIG_SCHED_DEBUG
6687 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6688 cpumask_t
*groupmask
)
6690 struct sched_group
*group
= sd
->groups
;
6693 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6694 cpus_clear(*groupmask
);
6696 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6698 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6699 printk("does not load-balance\n");
6701 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6706 printk(KERN_CONT
"span %s\n", str
);
6708 if (!cpu_isset(cpu
, sd
->span
)) {
6709 printk(KERN_ERR
"ERROR: domain->span does not contain "
6712 if (!cpu_isset(cpu
, group
->cpumask
)) {
6713 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6717 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6721 printk(KERN_ERR
"ERROR: group is NULL\n");
6725 if (!group
->__cpu_power
) {
6726 printk(KERN_CONT
"\n");
6727 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6732 if (!cpus_weight(group
->cpumask
)) {
6733 printk(KERN_CONT
"\n");
6734 printk(KERN_ERR
"ERROR: empty group\n");
6738 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6739 printk(KERN_CONT
"\n");
6740 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6744 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6746 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6747 printk(KERN_CONT
" %s", str
);
6749 group
= group
->next
;
6750 } while (group
!= sd
->groups
);
6751 printk(KERN_CONT
"\n");
6753 if (!cpus_equal(sd
->span
, *groupmask
))
6754 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6756 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6757 printk(KERN_ERR
"ERROR: parent span is not a superset "
6758 "of domain->span\n");
6762 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6764 cpumask_t
*groupmask
;
6768 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6772 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6774 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6776 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6781 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6791 # define sched_domain_debug(sd, cpu) do { } while (0)
6794 static int sd_degenerate(struct sched_domain
*sd
)
6796 if (cpus_weight(sd
->span
) == 1)
6799 /* Following flags need at least 2 groups */
6800 if (sd
->flags
& (SD_LOAD_BALANCE
|
6801 SD_BALANCE_NEWIDLE
|
6805 SD_SHARE_PKG_RESOURCES
)) {
6806 if (sd
->groups
!= sd
->groups
->next
)
6810 /* Following flags don't use groups */
6811 if (sd
->flags
& (SD_WAKE_IDLE
|
6820 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6822 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6824 if (sd_degenerate(parent
))
6827 if (!cpus_equal(sd
->span
, parent
->span
))
6830 /* Does parent contain flags not in child? */
6831 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6832 if (cflags
& SD_WAKE_AFFINE
)
6833 pflags
&= ~SD_WAKE_BALANCE
;
6834 /* Flags needing groups don't count if only 1 group in parent */
6835 if (parent
->groups
== parent
->groups
->next
) {
6836 pflags
&= ~(SD_LOAD_BALANCE
|
6837 SD_BALANCE_NEWIDLE
|
6841 SD_SHARE_PKG_RESOURCES
);
6843 if (~cflags
& pflags
)
6849 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6851 unsigned long flags
;
6852 const struct sched_class
*class;
6854 spin_lock_irqsave(&rq
->lock
, flags
);
6857 struct root_domain
*old_rd
= rq
->rd
;
6859 for (class = sched_class_highest
; class; class = class->next
) {
6860 if (class->leave_domain
)
6861 class->leave_domain(rq
);
6864 cpu_clear(rq
->cpu
, old_rd
->span
);
6865 cpu_clear(rq
->cpu
, old_rd
->online
);
6867 if (atomic_dec_and_test(&old_rd
->refcount
))
6871 atomic_inc(&rd
->refcount
);
6874 cpu_set(rq
->cpu
, rd
->span
);
6875 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6876 cpu_set(rq
->cpu
, rd
->online
);
6878 for (class = sched_class_highest
; class; class = class->next
) {
6879 if (class->join_domain
)
6880 class->join_domain(rq
);
6883 spin_unlock_irqrestore(&rq
->lock
, flags
);
6886 static void init_rootdomain(struct root_domain
*rd
)
6888 memset(rd
, 0, sizeof(*rd
));
6890 cpus_clear(rd
->span
);
6891 cpus_clear(rd
->online
);
6894 static void init_defrootdomain(void)
6896 init_rootdomain(&def_root_domain
);
6897 atomic_set(&def_root_domain
.refcount
, 1);
6900 static struct root_domain
*alloc_rootdomain(void)
6902 struct root_domain
*rd
;
6904 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6908 init_rootdomain(rd
);
6914 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6915 * hold the hotplug lock.
6918 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6920 struct rq
*rq
= cpu_rq(cpu
);
6921 struct sched_domain
*tmp
;
6923 /* Remove the sched domains which do not contribute to scheduling. */
6924 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6925 struct sched_domain
*parent
= tmp
->parent
;
6928 if (sd_parent_degenerate(tmp
, parent
)) {
6929 tmp
->parent
= parent
->parent
;
6931 parent
->parent
->child
= tmp
;
6935 if (sd
&& sd_degenerate(sd
)) {
6941 sched_domain_debug(sd
, cpu
);
6943 rq_attach_root(rq
, rd
);
6944 rcu_assign_pointer(rq
->sd
, sd
);
6947 /* cpus with isolated domains */
6948 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6950 /* Setup the mask of cpus configured for isolated domains */
6951 static int __init
isolated_cpu_setup(char *str
)
6953 int ints
[NR_CPUS
], i
;
6955 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6956 cpus_clear(cpu_isolated_map
);
6957 for (i
= 1; i
<= ints
[0]; i
++)
6958 if (ints
[i
] < NR_CPUS
)
6959 cpu_set(ints
[i
], cpu_isolated_map
);
6963 __setup("isolcpus=", isolated_cpu_setup
);
6966 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6967 * to a function which identifies what group(along with sched group) a CPU
6968 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6969 * (due to the fact that we keep track of groups covered with a cpumask_t).
6971 * init_sched_build_groups will build a circular linked list of the groups
6972 * covered by the given span, and will set each group's ->cpumask correctly,
6973 * and ->cpu_power to 0.
6976 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6977 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6978 struct sched_group
**sg
,
6979 cpumask_t
*tmpmask
),
6980 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6982 struct sched_group
*first
= NULL
, *last
= NULL
;
6985 cpus_clear(*covered
);
6987 for_each_cpu_mask(i
, *span
) {
6988 struct sched_group
*sg
;
6989 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6992 if (cpu_isset(i
, *covered
))
6995 cpus_clear(sg
->cpumask
);
6996 sg
->__cpu_power
= 0;
6998 for_each_cpu_mask(j
, *span
) {
6999 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7002 cpu_set(j
, *covered
);
7003 cpu_set(j
, sg
->cpumask
);
7014 #define SD_NODES_PER_DOMAIN 16
7019 * find_next_best_node - find the next node to include in a sched_domain
7020 * @node: node whose sched_domain we're building
7021 * @used_nodes: nodes already in the sched_domain
7023 * Find the next node to include in a given scheduling domain. Simply
7024 * finds the closest node not already in the @used_nodes map.
7026 * Should use nodemask_t.
7028 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7030 int i
, n
, val
, min_val
, best_node
= 0;
7034 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7035 /* Start at @node */
7036 n
= (node
+ i
) % MAX_NUMNODES
;
7038 if (!nr_cpus_node(n
))
7041 /* Skip already used nodes */
7042 if (node_isset(n
, *used_nodes
))
7045 /* Simple min distance search */
7046 val
= node_distance(node
, n
);
7048 if (val
< min_val
) {
7054 node_set(best_node
, *used_nodes
);
7059 * sched_domain_node_span - get a cpumask for a node's sched_domain
7060 * @node: node whose cpumask we're constructing
7061 * @span: resulting cpumask
7063 * Given a node, construct a good cpumask for its sched_domain to span. It
7064 * should be one that prevents unnecessary balancing, but also spreads tasks
7067 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7069 nodemask_t used_nodes
;
7070 node_to_cpumask_ptr(nodemask
, node
);
7074 nodes_clear(used_nodes
);
7076 cpus_or(*span
, *span
, *nodemask
);
7077 node_set(node
, used_nodes
);
7079 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7080 int next_node
= find_next_best_node(node
, &used_nodes
);
7082 node_to_cpumask_ptr_next(nodemask
, next_node
);
7083 cpus_or(*span
, *span
, *nodemask
);
7088 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7091 * SMT sched-domains:
7093 #ifdef CONFIG_SCHED_SMT
7094 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7095 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7098 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7102 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7108 * multi-core sched-domains:
7110 #ifdef CONFIG_SCHED_MC
7111 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7112 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7115 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7117 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7122 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7123 cpus_and(*mask
, *mask
, *cpu_map
);
7124 group
= first_cpu(*mask
);
7126 *sg
= &per_cpu(sched_group_core
, group
);
7129 #elif defined(CONFIG_SCHED_MC)
7131 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7135 *sg
= &per_cpu(sched_group_core
, cpu
);
7140 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7141 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7144 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7148 #ifdef CONFIG_SCHED_MC
7149 *mask
= cpu_coregroup_map(cpu
);
7150 cpus_and(*mask
, *mask
, *cpu_map
);
7151 group
= first_cpu(*mask
);
7152 #elif defined(CONFIG_SCHED_SMT)
7153 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7154 cpus_and(*mask
, *mask
, *cpu_map
);
7155 group
= first_cpu(*mask
);
7160 *sg
= &per_cpu(sched_group_phys
, group
);
7166 * The init_sched_build_groups can't handle what we want to do with node
7167 * groups, so roll our own. Now each node has its own list of groups which
7168 * gets dynamically allocated.
7170 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7171 static struct sched_group
***sched_group_nodes_bycpu
;
7173 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7174 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7176 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7177 struct sched_group
**sg
, cpumask_t
*nodemask
)
7181 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7182 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7183 group
= first_cpu(*nodemask
);
7186 *sg
= &per_cpu(sched_group_allnodes
, group
);
7190 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7192 struct sched_group
*sg
= group_head
;
7198 for_each_cpu_mask(j
, sg
->cpumask
) {
7199 struct sched_domain
*sd
;
7201 sd
= &per_cpu(phys_domains
, j
);
7202 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7204 * Only add "power" once for each
7210 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7213 } while (sg
!= group_head
);
7218 /* Free memory allocated for various sched_group structures */
7219 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7223 for_each_cpu_mask(cpu
, *cpu_map
) {
7224 struct sched_group
**sched_group_nodes
7225 = sched_group_nodes_bycpu
[cpu
];
7227 if (!sched_group_nodes
)
7230 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7231 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7233 *nodemask
= node_to_cpumask(i
);
7234 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7235 if (cpus_empty(*nodemask
))
7245 if (oldsg
!= sched_group_nodes
[i
])
7248 kfree(sched_group_nodes
);
7249 sched_group_nodes_bycpu
[cpu
] = NULL
;
7253 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7259 * Initialize sched groups cpu_power.
7261 * cpu_power indicates the capacity of sched group, which is used while
7262 * distributing the load between different sched groups in a sched domain.
7263 * Typically cpu_power for all the groups in a sched domain will be same unless
7264 * there are asymmetries in the topology. If there are asymmetries, group
7265 * having more cpu_power will pickup more load compared to the group having
7268 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7269 * the maximum number of tasks a group can handle in the presence of other idle
7270 * or lightly loaded groups in the same sched domain.
7272 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7274 struct sched_domain
*child
;
7275 struct sched_group
*group
;
7277 WARN_ON(!sd
|| !sd
->groups
);
7279 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7284 sd
->groups
->__cpu_power
= 0;
7287 * For perf policy, if the groups in child domain share resources
7288 * (for example cores sharing some portions of the cache hierarchy
7289 * or SMT), then set this domain groups cpu_power such that each group
7290 * can handle only one task, when there are other idle groups in the
7291 * same sched domain.
7293 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7295 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7296 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7301 * add cpu_power of each child group to this groups cpu_power
7303 group
= child
->groups
;
7305 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7306 group
= group
->next
;
7307 } while (group
!= child
->groups
);
7311 * Initializers for schedule domains
7312 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7315 #define SD_INIT(sd, type) sd_init_##type(sd)
7316 #define SD_INIT_FUNC(type) \
7317 static noinline void sd_init_##type(struct sched_domain *sd) \
7319 memset(sd, 0, sizeof(*sd)); \
7320 *sd = SD_##type##_INIT; \
7321 sd->level = SD_LV_##type; \
7326 SD_INIT_FUNC(ALLNODES
)
7329 #ifdef CONFIG_SCHED_SMT
7330 SD_INIT_FUNC(SIBLING
)
7332 #ifdef CONFIG_SCHED_MC
7337 * To minimize stack usage kmalloc room for cpumasks and share the
7338 * space as the usage in build_sched_domains() dictates. Used only
7339 * if the amount of space is significant.
7342 cpumask_t tmpmask
; /* make this one first */
7345 cpumask_t this_sibling_map
;
7346 cpumask_t this_core_map
;
7348 cpumask_t send_covered
;
7351 cpumask_t domainspan
;
7353 cpumask_t notcovered
;
7358 #define SCHED_CPUMASK_ALLOC 1
7359 #define SCHED_CPUMASK_FREE(v) kfree(v)
7360 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7362 #define SCHED_CPUMASK_ALLOC 0
7363 #define SCHED_CPUMASK_FREE(v)
7364 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7367 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7368 ((unsigned long)(a) + offsetof(struct allmasks, v))
7370 static int default_relax_domain_level
= -1;
7372 static int __init
setup_relax_domain_level(char *str
)
7374 default_relax_domain_level
= simple_strtoul(str
, NULL
, 0);
7377 __setup("relax_domain_level=", setup_relax_domain_level
);
7379 static void set_domain_attribute(struct sched_domain
*sd
,
7380 struct sched_domain_attr
*attr
)
7384 if (!attr
|| attr
->relax_domain_level
< 0) {
7385 if (default_relax_domain_level
< 0)
7388 request
= default_relax_domain_level
;
7390 request
= attr
->relax_domain_level
;
7391 if (request
< sd
->level
) {
7392 /* turn off idle balance on this domain */
7393 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7395 /* turn on idle balance on this domain */
7396 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7401 * Build sched domains for a given set of cpus and attach the sched domains
7402 * to the individual cpus
7404 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7405 struct sched_domain_attr
*attr
)
7408 struct root_domain
*rd
;
7409 SCHED_CPUMASK_DECLARE(allmasks
);
7412 struct sched_group
**sched_group_nodes
= NULL
;
7413 int sd_allnodes
= 0;
7416 * Allocate the per-node list of sched groups
7418 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7420 if (!sched_group_nodes
) {
7421 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7426 rd
= alloc_rootdomain();
7428 printk(KERN_WARNING
"Cannot alloc root domain\n");
7430 kfree(sched_group_nodes
);
7435 #if SCHED_CPUMASK_ALLOC
7436 /* get space for all scratch cpumask variables */
7437 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7439 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7442 kfree(sched_group_nodes
);
7447 tmpmask
= (cpumask_t
*)allmasks
;
7451 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7455 * Set up domains for cpus specified by the cpu_map.
7457 for_each_cpu_mask(i
, *cpu_map
) {
7458 struct sched_domain
*sd
= NULL
, *p
;
7459 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7461 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7462 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7465 if (cpus_weight(*cpu_map
) >
7466 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7467 sd
= &per_cpu(allnodes_domains
, i
);
7468 SD_INIT(sd
, ALLNODES
);
7469 set_domain_attribute(sd
, attr
);
7470 sd
->span
= *cpu_map
;
7471 sd
->first_cpu
= first_cpu(sd
->span
);
7472 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7478 sd
= &per_cpu(node_domains
, i
);
7480 set_domain_attribute(sd
, attr
);
7481 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7482 sd
->first_cpu
= first_cpu(sd
->span
);
7486 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7490 sd
= &per_cpu(phys_domains
, i
);
7492 set_domain_attribute(sd
, attr
);
7493 sd
->span
= *nodemask
;
7494 sd
->first_cpu
= first_cpu(sd
->span
);
7498 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7500 #ifdef CONFIG_SCHED_MC
7502 sd
= &per_cpu(core_domains
, i
);
7504 set_domain_attribute(sd
, attr
);
7505 sd
->span
= cpu_coregroup_map(i
);
7506 sd
->first_cpu
= first_cpu(sd
->span
);
7507 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7510 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7513 #ifdef CONFIG_SCHED_SMT
7515 sd
= &per_cpu(cpu_domains
, i
);
7516 SD_INIT(sd
, SIBLING
);
7517 set_domain_attribute(sd
, attr
);
7518 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7519 sd
->first_cpu
= first_cpu(sd
->span
);
7520 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7523 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7527 #ifdef CONFIG_SCHED_SMT
7528 /* Set up CPU (sibling) groups */
7529 for_each_cpu_mask(i
, *cpu_map
) {
7530 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7531 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7533 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7534 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7535 if (i
!= first_cpu(*this_sibling_map
))
7538 init_sched_build_groups(this_sibling_map
, cpu_map
,
7540 send_covered
, tmpmask
);
7544 #ifdef CONFIG_SCHED_MC
7545 /* Set up multi-core groups */
7546 for_each_cpu_mask(i
, *cpu_map
) {
7547 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7548 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7550 *this_core_map
= cpu_coregroup_map(i
);
7551 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7552 if (i
!= first_cpu(*this_core_map
))
7555 init_sched_build_groups(this_core_map
, cpu_map
,
7557 send_covered
, tmpmask
);
7561 /* Set up physical groups */
7562 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7563 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7564 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7566 *nodemask
= node_to_cpumask(i
);
7567 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7568 if (cpus_empty(*nodemask
))
7571 init_sched_build_groups(nodemask
, cpu_map
,
7573 send_covered
, tmpmask
);
7577 /* Set up node groups */
7579 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7581 init_sched_build_groups(cpu_map
, cpu_map
,
7582 &cpu_to_allnodes_group
,
7583 send_covered
, tmpmask
);
7586 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7587 /* Set up node groups */
7588 struct sched_group
*sg
, *prev
;
7589 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7590 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7591 SCHED_CPUMASK_VAR(covered
, allmasks
);
7594 *nodemask
= node_to_cpumask(i
);
7595 cpus_clear(*covered
);
7597 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7598 if (cpus_empty(*nodemask
)) {
7599 sched_group_nodes
[i
] = NULL
;
7603 sched_domain_node_span(i
, domainspan
);
7604 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7606 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7608 printk(KERN_WARNING
"Can not alloc domain group for "
7612 sched_group_nodes
[i
] = sg
;
7613 for_each_cpu_mask(j
, *nodemask
) {
7614 struct sched_domain
*sd
;
7616 sd
= &per_cpu(node_domains
, j
);
7619 sg
->__cpu_power
= 0;
7620 sg
->cpumask
= *nodemask
;
7622 cpus_or(*covered
, *covered
, *nodemask
);
7625 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7626 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7627 int n
= (i
+ j
) % MAX_NUMNODES
;
7628 node_to_cpumask_ptr(pnodemask
, n
);
7630 cpus_complement(*notcovered
, *covered
);
7631 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7632 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7633 if (cpus_empty(*tmpmask
))
7636 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7637 if (cpus_empty(*tmpmask
))
7640 sg
= kmalloc_node(sizeof(struct sched_group
),
7644 "Can not alloc domain group for node %d\n", j
);
7647 sg
->__cpu_power
= 0;
7648 sg
->cpumask
= *tmpmask
;
7649 sg
->next
= prev
->next
;
7650 cpus_or(*covered
, *covered
, *tmpmask
);
7657 /* Calculate CPU power for physical packages and nodes */
7658 #ifdef CONFIG_SCHED_SMT
7659 for_each_cpu_mask(i
, *cpu_map
) {
7660 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7662 init_sched_groups_power(i
, sd
);
7665 #ifdef CONFIG_SCHED_MC
7666 for_each_cpu_mask(i
, *cpu_map
) {
7667 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7669 init_sched_groups_power(i
, sd
);
7673 for_each_cpu_mask(i
, *cpu_map
) {
7674 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7676 init_sched_groups_power(i
, sd
);
7680 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7681 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7684 struct sched_group
*sg
;
7686 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7688 init_numa_sched_groups_power(sg
);
7692 /* Attach the domains */
7693 for_each_cpu_mask(i
, *cpu_map
) {
7694 struct sched_domain
*sd
;
7695 #ifdef CONFIG_SCHED_SMT
7696 sd
= &per_cpu(cpu_domains
, i
);
7697 #elif defined(CONFIG_SCHED_MC)
7698 sd
= &per_cpu(core_domains
, i
);
7700 sd
= &per_cpu(phys_domains
, i
);
7702 cpu_attach_domain(sd
, rd
, i
);
7705 SCHED_CPUMASK_FREE((void *)allmasks
);
7710 free_sched_groups(cpu_map
, tmpmask
);
7711 SCHED_CPUMASK_FREE((void *)allmasks
);
7716 static int build_sched_domains(const cpumask_t
*cpu_map
)
7718 return __build_sched_domains(cpu_map
, NULL
);
7721 static cpumask_t
*doms_cur
; /* current sched domains */
7722 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7723 static struct sched_domain_attr
*dattr_cur
; /* attribues of custom domains
7727 * Special case: If a kmalloc of a doms_cur partition (array of
7728 * cpumask_t) fails, then fallback to a single sched domain,
7729 * as determined by the single cpumask_t fallback_doms.
7731 static cpumask_t fallback_doms
;
7733 void __attribute__((weak
)) arch_update_cpu_topology(void)
7738 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7739 * For now this just excludes isolated cpus, but could be used to
7740 * exclude other special cases in the future.
7742 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7746 arch_update_cpu_topology();
7748 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7750 doms_cur
= &fallback_doms
;
7751 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7753 err
= build_sched_domains(doms_cur
);
7754 register_sched_domain_sysctl();
7759 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7762 free_sched_groups(cpu_map
, tmpmask
);
7766 * Detach sched domains from a group of cpus specified in cpu_map
7767 * These cpus will now be attached to the NULL domain
7769 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7774 unregister_sched_domain_sysctl();
7776 for_each_cpu_mask(i
, *cpu_map
)
7777 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7778 synchronize_sched();
7779 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7782 /* handle null as "default" */
7783 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7784 struct sched_domain_attr
*new, int idx_new
)
7786 struct sched_domain_attr tmp
;
7793 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7794 new ? (new + idx_new
) : &tmp
,
7795 sizeof(struct sched_domain_attr
));
7799 * Partition sched domains as specified by the 'ndoms_new'
7800 * cpumasks in the array doms_new[] of cpumasks. This compares
7801 * doms_new[] to the current sched domain partitioning, doms_cur[].
7802 * It destroys each deleted domain and builds each new domain.
7804 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7805 * The masks don't intersect (don't overlap.) We should setup one
7806 * sched domain for each mask. CPUs not in any of the cpumasks will
7807 * not be load balanced. If the same cpumask appears both in the
7808 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7811 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7812 * ownership of it and will kfree it when done with it. If the caller
7813 * failed the kmalloc call, then it can pass in doms_new == NULL,
7814 * and partition_sched_domains() will fallback to the single partition
7817 * Call with hotplug lock held
7819 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7820 struct sched_domain_attr
*dattr_new
)
7824 mutex_lock(&sched_domains_mutex
);
7826 /* always unregister in case we don't destroy any domains */
7827 unregister_sched_domain_sysctl();
7829 if (doms_new
== NULL
) {
7831 doms_new
= &fallback_doms
;
7832 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7836 /* Destroy deleted domains */
7837 for (i
= 0; i
< ndoms_cur
; i
++) {
7838 for (j
= 0; j
< ndoms_new
; j
++) {
7839 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7840 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7843 /* no match - a current sched domain not in new doms_new[] */
7844 detach_destroy_domains(doms_cur
+ i
);
7849 /* Build new domains */
7850 for (i
= 0; i
< ndoms_new
; i
++) {
7851 for (j
= 0; j
< ndoms_cur
; j
++) {
7852 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7853 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7856 /* no match - add a new doms_new */
7857 __build_sched_domains(doms_new
+ i
,
7858 dattr_new
? dattr_new
+ i
: NULL
);
7863 /* Remember the new sched domains */
7864 if (doms_cur
!= &fallback_doms
)
7866 kfree(dattr_cur
); /* kfree(NULL) is safe */
7867 doms_cur
= doms_new
;
7868 dattr_cur
= dattr_new
;
7869 ndoms_cur
= ndoms_new
;
7871 register_sched_domain_sysctl();
7873 mutex_unlock(&sched_domains_mutex
);
7876 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7877 int arch_reinit_sched_domains(void)
7882 mutex_lock(&sched_domains_mutex
);
7883 detach_destroy_domains(&cpu_online_map
);
7884 err
= arch_init_sched_domains(&cpu_online_map
);
7885 mutex_unlock(&sched_domains_mutex
);
7891 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7895 if (buf
[0] != '0' && buf
[0] != '1')
7899 sched_smt_power_savings
= (buf
[0] == '1');
7901 sched_mc_power_savings
= (buf
[0] == '1');
7903 ret
= arch_reinit_sched_domains();
7905 return ret
? ret
: count
;
7908 #ifdef CONFIG_SCHED_MC
7909 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7911 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7913 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7914 const char *buf
, size_t count
)
7916 return sched_power_savings_store(buf
, count
, 0);
7918 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7919 sched_mc_power_savings_store
);
7922 #ifdef CONFIG_SCHED_SMT
7923 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7925 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7927 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7928 const char *buf
, size_t count
)
7930 return sched_power_savings_store(buf
, count
, 1);
7932 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7933 sched_smt_power_savings_store
);
7936 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7940 #ifdef CONFIG_SCHED_SMT
7942 err
= sysfs_create_file(&cls
->kset
.kobj
,
7943 &attr_sched_smt_power_savings
.attr
);
7945 #ifdef CONFIG_SCHED_MC
7946 if (!err
&& mc_capable())
7947 err
= sysfs_create_file(&cls
->kset
.kobj
,
7948 &attr_sched_mc_power_savings
.attr
);
7955 * Force a reinitialization of the sched domains hierarchy. The domains
7956 * and groups cannot be updated in place without racing with the balancing
7957 * code, so we temporarily attach all running cpus to the NULL domain
7958 * which will prevent rebalancing while the sched domains are recalculated.
7960 static int update_sched_domains(struct notifier_block
*nfb
,
7961 unsigned long action
, void *hcpu
)
7964 case CPU_UP_PREPARE
:
7965 case CPU_UP_PREPARE_FROZEN
:
7966 case CPU_DOWN_PREPARE
:
7967 case CPU_DOWN_PREPARE_FROZEN
:
7968 detach_destroy_domains(&cpu_online_map
);
7971 case CPU_UP_CANCELED
:
7972 case CPU_UP_CANCELED_FROZEN
:
7973 case CPU_DOWN_FAILED
:
7974 case CPU_DOWN_FAILED_FROZEN
:
7976 case CPU_ONLINE_FROZEN
:
7978 case CPU_DEAD_FROZEN
:
7980 * Fall through and re-initialise the domains.
7987 /* The hotplug lock is already held by cpu_up/cpu_down */
7988 arch_init_sched_domains(&cpu_online_map
);
7993 void __init
sched_init_smp(void)
7995 cpumask_t non_isolated_cpus
;
7997 #if defined(CONFIG_NUMA)
7998 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8000 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8003 mutex_lock(&sched_domains_mutex
);
8004 arch_init_sched_domains(&cpu_online_map
);
8005 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
8006 if (cpus_empty(non_isolated_cpus
))
8007 cpu_set(smp_processor_id(), non_isolated_cpus
);
8008 mutex_unlock(&sched_domains_mutex
);
8010 /* XXX: Theoretical race here - CPU may be hotplugged now */
8011 hotcpu_notifier(update_sched_domains
, 0);
8014 /* Move init over to a non-isolated CPU */
8015 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
8017 sched_init_granularity();
8020 void __init
sched_init_smp(void)
8022 sched_init_granularity();
8024 #endif /* CONFIG_SMP */
8026 int in_sched_functions(unsigned long addr
)
8028 return in_lock_functions(addr
) ||
8029 (addr
>= (unsigned long)__sched_text_start
8030 && addr
< (unsigned long)__sched_text_end
);
8033 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8035 cfs_rq
->tasks_timeline
= RB_ROOT
;
8036 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8037 #ifdef CONFIG_FAIR_GROUP_SCHED
8040 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8043 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8045 struct rt_prio_array
*array
;
8048 array
= &rt_rq
->active
;
8049 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8050 INIT_LIST_HEAD(array
->queue
+ i
);
8051 __clear_bit(i
, array
->bitmap
);
8053 /* delimiter for bitsearch: */
8054 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8056 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8057 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8060 rt_rq
->rt_nr_migratory
= 0;
8061 rt_rq
->overloaded
= 0;
8065 rt_rq
->rt_throttled
= 0;
8066 rt_rq
->rt_runtime
= 0;
8067 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8069 #ifdef CONFIG_RT_GROUP_SCHED
8070 rt_rq
->rt_nr_boosted
= 0;
8075 #ifdef CONFIG_FAIR_GROUP_SCHED
8076 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8077 struct sched_entity
*se
, int cpu
, int add
,
8078 struct sched_entity
*parent
)
8080 struct rq
*rq
= cpu_rq(cpu
);
8081 tg
->cfs_rq
[cpu
] = cfs_rq
;
8082 init_cfs_rq(cfs_rq
, rq
);
8085 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8088 /* se could be NULL for init_task_group */
8093 se
->cfs_rq
= &rq
->cfs
;
8095 se
->cfs_rq
= parent
->my_q
;
8098 se
->load
.weight
= tg
->shares
;
8099 se
->load
.inv_weight
= 0;
8100 se
->parent
= parent
;
8104 #ifdef CONFIG_RT_GROUP_SCHED
8105 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8106 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8107 struct sched_rt_entity
*parent
)
8109 struct rq
*rq
= cpu_rq(cpu
);
8111 tg
->rt_rq
[cpu
] = rt_rq
;
8112 init_rt_rq(rt_rq
, rq
);
8114 rt_rq
->rt_se
= rt_se
;
8115 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8117 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8119 tg
->rt_se
[cpu
] = rt_se
;
8124 rt_se
->rt_rq
= &rq
->rt
;
8126 rt_se
->rt_rq
= parent
->my_q
;
8128 rt_se
->rt_rq
= &rq
->rt
;
8129 rt_se
->my_q
= rt_rq
;
8130 rt_se
->parent
= parent
;
8131 INIT_LIST_HEAD(&rt_se
->run_list
);
8135 void __init
sched_init(void)
8138 unsigned long alloc_size
= 0, ptr
;
8140 #ifdef CONFIG_FAIR_GROUP_SCHED
8141 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8143 #ifdef CONFIG_RT_GROUP_SCHED
8144 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8146 #ifdef CONFIG_USER_SCHED
8150 * As sched_init() is called before page_alloc is setup,
8151 * we use alloc_bootmem().
8154 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8156 #ifdef CONFIG_FAIR_GROUP_SCHED
8157 init_task_group
.se
= (struct sched_entity
**)ptr
;
8158 ptr
+= nr_cpu_ids
* sizeof(void **);
8160 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8161 ptr
+= nr_cpu_ids
* sizeof(void **);
8163 #ifdef CONFIG_USER_SCHED
8164 root_task_group
.se
= (struct sched_entity
**)ptr
;
8165 ptr
+= nr_cpu_ids
* sizeof(void **);
8167 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8168 ptr
+= nr_cpu_ids
* sizeof(void **);
8171 #ifdef CONFIG_RT_GROUP_SCHED
8172 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8173 ptr
+= nr_cpu_ids
* sizeof(void **);
8175 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8176 ptr
+= nr_cpu_ids
* sizeof(void **);
8178 #ifdef CONFIG_USER_SCHED
8179 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8180 ptr
+= nr_cpu_ids
* sizeof(void **);
8182 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8183 ptr
+= nr_cpu_ids
* sizeof(void **);
8190 init_defrootdomain();
8193 init_rt_bandwidth(&def_rt_bandwidth
,
8194 global_rt_period(), global_rt_runtime());
8196 #ifdef CONFIG_RT_GROUP_SCHED
8197 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8198 global_rt_period(), global_rt_runtime());
8199 #ifdef CONFIG_USER_SCHED
8200 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8201 global_rt_period(), RUNTIME_INF
);
8205 #ifdef CONFIG_GROUP_SCHED
8206 list_add(&init_task_group
.list
, &task_groups
);
8207 INIT_LIST_HEAD(&init_task_group
.children
);
8209 #ifdef CONFIG_USER_SCHED
8210 INIT_LIST_HEAD(&root_task_group
.children
);
8211 init_task_group
.parent
= &root_task_group
;
8212 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8216 for_each_possible_cpu(i
) {
8220 spin_lock_init(&rq
->lock
);
8221 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
8224 update_last_tick_seen(rq
);
8225 init_cfs_rq(&rq
->cfs
, rq
);
8226 init_rt_rq(&rq
->rt
, rq
);
8227 #ifdef CONFIG_FAIR_GROUP_SCHED
8228 init_task_group
.shares
= init_task_group_load
;
8229 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8230 #ifdef CONFIG_CGROUP_SCHED
8232 * How much cpu bandwidth does init_task_group get?
8234 * In case of task-groups formed thr' the cgroup filesystem, it
8235 * gets 100% of the cpu resources in the system. This overall
8236 * system cpu resource is divided among the tasks of
8237 * init_task_group and its child task-groups in a fair manner,
8238 * based on each entity's (task or task-group's) weight
8239 * (se->load.weight).
8241 * In other words, if init_task_group has 10 tasks of weight
8242 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8243 * then A0's share of the cpu resource is:
8245 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8247 * We achieve this by letting init_task_group's tasks sit
8248 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8250 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8251 #elif defined CONFIG_USER_SCHED
8252 root_task_group
.shares
= NICE_0_LOAD
;
8253 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8255 * In case of task-groups formed thr' the user id of tasks,
8256 * init_task_group represents tasks belonging to root user.
8257 * Hence it forms a sibling of all subsequent groups formed.
8258 * In this case, init_task_group gets only a fraction of overall
8259 * system cpu resource, based on the weight assigned to root
8260 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8261 * by letting tasks of init_task_group sit in a separate cfs_rq
8262 * (init_cfs_rq) and having one entity represent this group of
8263 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8265 init_tg_cfs_entry(&init_task_group
,
8266 &per_cpu(init_cfs_rq
, i
),
8267 &per_cpu(init_sched_entity
, i
), i
, 1,
8268 root_task_group
.se
[i
]);
8271 #endif /* CONFIG_FAIR_GROUP_SCHED */
8273 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8274 #ifdef CONFIG_RT_GROUP_SCHED
8275 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8276 #ifdef CONFIG_CGROUP_SCHED
8277 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8278 #elif defined CONFIG_USER_SCHED
8279 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8280 init_tg_rt_entry(&init_task_group
,
8281 &per_cpu(init_rt_rq
, i
),
8282 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8283 root_task_group
.rt_se
[i
]);
8287 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8288 rq
->cpu_load
[j
] = 0;
8292 rq
->active_balance
= 0;
8293 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
, NULL
);
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
) {
8347 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8349 prev_jiffy
= jiffies
;
8350 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8351 " context at %s:%d\n", file
, line
);
8352 printk("in_atomic():%d, irqs_disabled():%d\n",
8353 in_atomic(), irqs_disabled());
8354 debug_show_held_locks(current
);
8355 if (irqs_disabled())
8356 print_irqtrace_events(current
);
8361 EXPORT_SYMBOL(__might_sleep
);
8364 #ifdef CONFIG_MAGIC_SYSRQ
8365 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8368 update_rq_clock(rq
);
8369 on_rq
= p
->se
.on_rq
;
8371 deactivate_task(rq
, p
, 0);
8372 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8374 activate_task(rq
, p
, 0);
8375 resched_task(rq
->curr
);
8379 void normalize_rt_tasks(void)
8381 struct task_struct
*g
, *p
;
8382 unsigned long flags
;
8385 read_lock_irqsave(&tasklist_lock
, flags
);
8386 do_each_thread(g
, p
) {
8388 * Only normalize user tasks:
8393 p
->se
.exec_start
= 0;
8394 #ifdef CONFIG_SCHEDSTATS
8395 p
->se
.wait_start
= 0;
8396 p
->se
.sleep_start
= 0;
8397 p
->se
.block_start
= 0;
8399 task_rq(p
)->clock
= 0;
8403 * Renice negative nice level userspace
8406 if (TASK_NICE(p
) < 0 && p
->mm
)
8407 set_user_nice(p
, 0);
8411 spin_lock(&p
->pi_lock
);
8412 rq
= __task_rq_lock(p
);
8414 normalize_task(rq
, p
);
8416 __task_rq_unlock(rq
);
8417 spin_unlock(&p
->pi_lock
);
8418 } while_each_thread(g
, p
);
8420 read_unlock_irqrestore(&tasklist_lock
, flags
);
8423 #endif /* CONFIG_MAGIC_SYSRQ */
8427 * These functions are only useful for the IA64 MCA handling.
8429 * They can only be called when the whole system has been
8430 * stopped - every CPU needs to be quiescent, and no scheduling
8431 * activity can take place. Using them for anything else would
8432 * be a serious bug, and as a result, they aren't even visible
8433 * under any other configuration.
8437 * curr_task - return the current task for a given cpu.
8438 * @cpu: the processor in question.
8440 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8442 struct task_struct
*curr_task(int cpu
)
8444 return cpu_curr(cpu
);
8448 * set_curr_task - set the current task for a given cpu.
8449 * @cpu: the processor in question.
8450 * @p: the task pointer to set.
8452 * Description: This function must only be used when non-maskable interrupts
8453 * are serviced on a separate stack. It allows the architecture to switch the
8454 * notion of the current task on a cpu in a non-blocking manner. This function
8455 * must be called with all CPU's synchronized, and interrupts disabled, the
8456 * and caller must save the original value of the current task (see
8457 * curr_task() above) and restore that value before reenabling interrupts and
8458 * re-starting the system.
8460 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8462 void set_curr_task(int cpu
, struct task_struct
*p
)
8469 #ifdef CONFIG_FAIR_GROUP_SCHED
8470 static void free_fair_sched_group(struct task_group
*tg
)
8474 for_each_possible_cpu(i
) {
8476 kfree(tg
->cfs_rq
[i
]);
8486 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8488 struct cfs_rq
*cfs_rq
;
8489 struct sched_entity
*se
, *parent_se
;
8493 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8496 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8500 tg
->shares
= NICE_0_LOAD
;
8502 for_each_possible_cpu(i
) {
8505 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8506 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8510 se
= kmalloc_node(sizeof(struct sched_entity
),
8511 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8515 parent_se
= parent
? parent
->se
[i
] : NULL
;
8516 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8525 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8527 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8528 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8531 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8533 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8536 static inline void free_fair_sched_group(struct task_group
*tg
)
8541 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8546 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8550 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8555 #ifdef CONFIG_RT_GROUP_SCHED
8556 static void free_rt_sched_group(struct task_group
*tg
)
8560 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8562 for_each_possible_cpu(i
) {
8564 kfree(tg
->rt_rq
[i
]);
8566 kfree(tg
->rt_se
[i
]);
8574 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8576 struct rt_rq
*rt_rq
;
8577 struct sched_rt_entity
*rt_se
, *parent_se
;
8581 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8584 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8588 init_rt_bandwidth(&tg
->rt_bandwidth
,
8589 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8591 for_each_possible_cpu(i
) {
8594 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8595 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8599 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8600 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8604 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8605 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8614 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8616 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8617 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8620 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8622 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8625 static inline void free_rt_sched_group(struct task_group
*tg
)
8630 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8635 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8639 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8644 #ifdef CONFIG_GROUP_SCHED
8645 static void free_sched_group(struct task_group
*tg
)
8647 free_fair_sched_group(tg
);
8648 free_rt_sched_group(tg
);
8652 /* allocate runqueue etc for a new task group */
8653 struct task_group
*sched_create_group(struct task_group
*parent
)
8655 struct task_group
*tg
;
8656 unsigned long flags
;
8659 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8661 return ERR_PTR(-ENOMEM
);
8663 if (!alloc_fair_sched_group(tg
, parent
))
8666 if (!alloc_rt_sched_group(tg
, parent
))
8669 spin_lock_irqsave(&task_group_lock
, flags
);
8670 for_each_possible_cpu(i
) {
8671 register_fair_sched_group(tg
, i
);
8672 register_rt_sched_group(tg
, i
);
8674 list_add_rcu(&tg
->list
, &task_groups
);
8676 WARN_ON(!parent
); /* root should already exist */
8678 tg
->parent
= parent
;
8679 list_add_rcu(&tg
->siblings
, &parent
->children
);
8680 INIT_LIST_HEAD(&tg
->children
);
8681 spin_unlock_irqrestore(&task_group_lock
, flags
);
8686 free_sched_group(tg
);
8687 return ERR_PTR(-ENOMEM
);
8690 /* rcu callback to free various structures associated with a task group */
8691 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8693 /* now it should be safe to free those cfs_rqs */
8694 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8697 /* Destroy runqueue etc associated with a task group */
8698 void sched_destroy_group(struct task_group
*tg
)
8700 unsigned long flags
;
8703 spin_lock_irqsave(&task_group_lock
, flags
);
8704 for_each_possible_cpu(i
) {
8705 unregister_fair_sched_group(tg
, i
);
8706 unregister_rt_sched_group(tg
, i
);
8708 list_del_rcu(&tg
->list
);
8709 list_del_rcu(&tg
->siblings
);
8710 spin_unlock_irqrestore(&task_group_lock
, flags
);
8712 /* wait for possible concurrent references to cfs_rqs complete */
8713 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8716 /* change task's runqueue when it moves between groups.
8717 * The caller of this function should have put the task in its new group
8718 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8719 * reflect its new group.
8721 void sched_move_task(struct task_struct
*tsk
)
8724 unsigned long flags
;
8727 rq
= task_rq_lock(tsk
, &flags
);
8729 update_rq_clock(rq
);
8731 running
= task_current(rq
, tsk
);
8732 on_rq
= tsk
->se
.on_rq
;
8735 dequeue_task(rq
, tsk
, 0);
8736 if (unlikely(running
))
8737 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8739 set_task_rq(tsk
, task_cpu(tsk
));
8741 #ifdef CONFIG_FAIR_GROUP_SCHED
8742 if (tsk
->sched_class
->moved_group
)
8743 tsk
->sched_class
->moved_group(tsk
);
8746 if (unlikely(running
))
8747 tsk
->sched_class
->set_curr_task(rq
);
8749 enqueue_task(rq
, tsk
, 0);
8751 task_rq_unlock(rq
, &flags
);
8755 #ifdef CONFIG_FAIR_GROUP_SCHED
8756 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8758 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8763 dequeue_entity(cfs_rq
, se
, 0);
8765 se
->load
.weight
= shares
;
8766 se
->load
.inv_weight
= 0;
8769 enqueue_entity(cfs_rq
, se
, 0);
8772 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8774 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8775 struct rq
*rq
= cfs_rq
->rq
;
8776 unsigned long flags
;
8778 spin_lock_irqsave(&rq
->lock
, flags
);
8779 __set_se_shares(se
, shares
);
8780 spin_unlock_irqrestore(&rq
->lock
, flags
);
8783 static DEFINE_MUTEX(shares_mutex
);
8785 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8788 unsigned long flags
;
8791 * We can't change the weight of the root cgroup.
8796 if (shares
< MIN_SHARES
)
8797 shares
= MIN_SHARES
;
8798 else if (shares
> MAX_SHARES
)
8799 shares
= MAX_SHARES
;
8801 mutex_lock(&shares_mutex
);
8802 if (tg
->shares
== shares
)
8805 spin_lock_irqsave(&task_group_lock
, flags
);
8806 for_each_possible_cpu(i
)
8807 unregister_fair_sched_group(tg
, i
);
8808 list_del_rcu(&tg
->siblings
);
8809 spin_unlock_irqrestore(&task_group_lock
, flags
);
8811 /* wait for any ongoing reference to this group to finish */
8812 synchronize_sched();
8815 * Now we are free to modify the group's share on each cpu
8816 * w/o tripping rebalance_share or load_balance_fair.
8818 tg
->shares
= shares
;
8819 for_each_possible_cpu(i
) {
8823 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8824 set_se_shares(tg
->se
[i
], shares
);
8828 * Enable load balance activity on this group, by inserting it back on
8829 * each cpu's rq->leaf_cfs_rq_list.
8831 spin_lock_irqsave(&task_group_lock
, flags
);
8832 for_each_possible_cpu(i
)
8833 register_fair_sched_group(tg
, i
);
8834 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8835 spin_unlock_irqrestore(&task_group_lock
, flags
);
8837 mutex_unlock(&shares_mutex
);
8841 unsigned long sched_group_shares(struct task_group
*tg
)
8847 #ifdef CONFIG_RT_GROUP_SCHED
8849 * Ensure that the real time constraints are schedulable.
8851 static DEFINE_MUTEX(rt_constraints_mutex
);
8853 static unsigned long to_ratio(u64 period
, u64 runtime
)
8855 if (runtime
== RUNTIME_INF
)
8858 return div64_u64(runtime
<< 16, period
);
8861 #ifdef CONFIG_CGROUP_SCHED
8862 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8864 struct task_group
*tgi
, *parent
= tg
->parent
;
8865 unsigned long total
= 0;
8868 if (global_rt_period() < period
)
8871 return to_ratio(period
, runtime
) <
8872 to_ratio(global_rt_period(), global_rt_runtime());
8875 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8879 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8883 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8884 tgi
->rt_bandwidth
.rt_runtime
);
8888 return total
+ to_ratio(period
, runtime
) <
8889 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8890 parent
->rt_bandwidth
.rt_runtime
);
8892 #elif defined CONFIG_USER_SCHED
8893 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8895 struct task_group
*tgi
;
8896 unsigned long total
= 0;
8897 unsigned long global_ratio
=
8898 to_ratio(global_rt_period(), global_rt_runtime());
8901 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8905 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8906 tgi
->rt_bandwidth
.rt_runtime
);
8910 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8914 /* Must be called with tasklist_lock held */
8915 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8917 struct task_struct
*g
, *p
;
8918 do_each_thread(g
, p
) {
8919 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8921 } while_each_thread(g
, p
);
8925 static int tg_set_bandwidth(struct task_group
*tg
,
8926 u64 rt_period
, u64 rt_runtime
)
8930 mutex_lock(&rt_constraints_mutex
);
8931 read_lock(&tasklist_lock
);
8932 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8936 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8941 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8942 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8943 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8945 for_each_possible_cpu(i
) {
8946 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8948 spin_lock(&rt_rq
->rt_runtime_lock
);
8949 rt_rq
->rt_runtime
= rt_runtime
;
8950 spin_unlock(&rt_rq
->rt_runtime_lock
);
8952 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8954 read_unlock(&tasklist_lock
);
8955 mutex_unlock(&rt_constraints_mutex
);
8960 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8962 u64 rt_runtime
, rt_period
;
8964 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8965 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8966 if (rt_runtime_us
< 0)
8967 rt_runtime
= RUNTIME_INF
;
8969 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8972 long sched_group_rt_runtime(struct task_group
*tg
)
8976 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8979 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8980 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8981 return rt_runtime_us
;
8984 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8986 u64 rt_runtime
, rt_period
;
8988 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8989 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8991 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8994 long sched_group_rt_period(struct task_group
*tg
)
8998 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8999 do_div(rt_period_us
, NSEC_PER_USEC
);
9000 return rt_period_us
;
9003 static int sched_rt_global_constraints(void)
9007 mutex_lock(&rt_constraints_mutex
);
9008 if (!__rt_schedulable(NULL
, 1, 0))
9010 mutex_unlock(&rt_constraints_mutex
);
9015 static int sched_rt_global_constraints(void)
9017 unsigned long flags
;
9020 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9021 for_each_possible_cpu(i
) {
9022 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9024 spin_lock(&rt_rq
->rt_runtime_lock
);
9025 rt_rq
->rt_runtime
= global_rt_runtime();
9026 spin_unlock(&rt_rq
->rt_runtime_lock
);
9028 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9034 int sched_rt_handler(struct ctl_table
*table
, int write
,
9035 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9039 int old_period
, old_runtime
;
9040 static DEFINE_MUTEX(mutex
);
9043 old_period
= sysctl_sched_rt_period
;
9044 old_runtime
= sysctl_sched_rt_runtime
;
9046 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9048 if (!ret
&& write
) {
9049 ret
= sched_rt_global_constraints();
9051 sysctl_sched_rt_period
= old_period
;
9052 sysctl_sched_rt_runtime
= old_runtime
;
9054 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9055 def_rt_bandwidth
.rt_period
=
9056 ns_to_ktime(global_rt_period());
9059 mutex_unlock(&mutex
);
9064 #ifdef CONFIG_CGROUP_SCHED
9066 /* return corresponding task_group object of a cgroup */
9067 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9069 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9070 struct task_group
, css
);
9073 static struct cgroup_subsys_state
*
9074 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9076 struct task_group
*tg
, *parent
;
9078 if (!cgrp
->parent
) {
9079 /* This is early initialization for the top cgroup */
9080 init_task_group
.css
.cgroup
= cgrp
;
9081 return &init_task_group
.css
;
9084 parent
= cgroup_tg(cgrp
->parent
);
9085 tg
= sched_create_group(parent
);
9087 return ERR_PTR(-ENOMEM
);
9089 /* Bind the cgroup to task_group object we just created */
9090 tg
->css
.cgroup
= cgrp
;
9096 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9098 struct task_group
*tg
= cgroup_tg(cgrp
);
9100 sched_destroy_group(tg
);
9104 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9105 struct task_struct
*tsk
)
9107 #ifdef CONFIG_RT_GROUP_SCHED
9108 /* Don't accept realtime tasks when there is no way for them to run */
9109 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9112 /* We don't support RT-tasks being in separate groups */
9113 if (tsk
->sched_class
!= &fair_sched_class
)
9121 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9122 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9124 sched_move_task(tsk
);
9127 #ifdef CONFIG_FAIR_GROUP_SCHED
9128 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9131 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9134 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9136 struct task_group
*tg
= cgroup_tg(cgrp
);
9138 return (u64
) tg
->shares
;
9142 #ifdef CONFIG_RT_GROUP_SCHED
9143 static ssize_t
cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9146 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9149 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9151 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9154 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9157 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9160 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9162 return sched_group_rt_period(cgroup_tg(cgrp
));
9166 static struct cftype cpu_files
[] = {
9167 #ifdef CONFIG_FAIR_GROUP_SCHED
9170 .read_u64
= cpu_shares_read_u64
,
9171 .write_u64
= cpu_shares_write_u64
,
9174 #ifdef CONFIG_RT_GROUP_SCHED
9176 .name
= "rt_runtime_us",
9177 .read_s64
= cpu_rt_runtime_read
,
9178 .write_s64
= cpu_rt_runtime_write
,
9181 .name
= "rt_period_us",
9182 .read_u64
= cpu_rt_period_read_uint
,
9183 .write_u64
= cpu_rt_period_write_uint
,
9188 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9190 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9193 struct cgroup_subsys cpu_cgroup_subsys
= {
9195 .create
= cpu_cgroup_create
,
9196 .destroy
= cpu_cgroup_destroy
,
9197 .can_attach
= cpu_cgroup_can_attach
,
9198 .attach
= cpu_cgroup_attach
,
9199 .populate
= cpu_cgroup_populate
,
9200 .subsys_id
= cpu_cgroup_subsys_id
,
9204 #endif /* CONFIG_CGROUP_SCHED */
9206 #ifdef CONFIG_CGROUP_CPUACCT
9209 * CPU accounting code for task groups.
9211 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9212 * (balbir@in.ibm.com).
9215 /* track cpu usage of a group of tasks */
9217 struct cgroup_subsys_state css
;
9218 /* cpuusage holds pointer to a u64-type object on every cpu */
9222 struct cgroup_subsys cpuacct_subsys
;
9224 /* return cpu accounting group corresponding to this container */
9225 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9227 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9228 struct cpuacct
, css
);
9231 /* return cpu accounting group to which this task belongs */
9232 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9234 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9235 struct cpuacct
, css
);
9238 /* create a new cpu accounting group */
9239 static struct cgroup_subsys_state
*cpuacct_create(
9240 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9242 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9245 return ERR_PTR(-ENOMEM
);
9247 ca
->cpuusage
= alloc_percpu(u64
);
9248 if (!ca
->cpuusage
) {
9250 return ERR_PTR(-ENOMEM
);
9256 /* destroy an existing cpu accounting group */
9258 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9260 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9262 free_percpu(ca
->cpuusage
);
9266 /* return total cpu usage (in nanoseconds) of a group */
9267 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9269 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9270 u64 totalcpuusage
= 0;
9273 for_each_possible_cpu(i
) {
9274 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9277 * Take rq->lock to make 64-bit addition safe on 32-bit
9280 spin_lock_irq(&cpu_rq(i
)->lock
);
9281 totalcpuusage
+= *cpuusage
;
9282 spin_unlock_irq(&cpu_rq(i
)->lock
);
9285 return totalcpuusage
;
9288 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9291 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9300 for_each_possible_cpu(i
) {
9301 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9303 spin_lock_irq(&cpu_rq(i
)->lock
);
9305 spin_unlock_irq(&cpu_rq(i
)->lock
);
9311 static struct cftype files
[] = {
9314 .read_u64
= cpuusage_read
,
9315 .write_u64
= cpuusage_write
,
9319 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9321 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9325 * charge this task's execution time to its accounting group.
9327 * called with rq->lock held.
9329 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9333 if (!cpuacct_subsys
.active
)
9338 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9340 *cpuusage
+= cputime
;
9344 struct cgroup_subsys cpuacct_subsys
= {
9346 .create
= cpuacct_create
,
9347 .destroy
= cpuacct_destroy
,
9348 .populate
= cpuacct_populate
,
9349 .subsys_id
= cpuacct_subsys_id
,
9351 #endif /* CONFIG_CGROUP_CPUACCT */