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(unsigned long long time
, int cpu
)
916 * We want this inlined, to not get tracer function calls
917 * in this critical section:
919 spin_acquire(&time_sync_lock
.dep_map
, 0, 0, _THIS_IP_
);
920 __raw_spin_lock(&time_sync_lock
.raw_lock
);
922 if (time
< prev_global_time
) {
923 per_cpu(time_offset
, cpu
) += prev_global_time
- time
;
924 time
= prev_global_time
;
926 prev_global_time
= time
;
929 __raw_spin_unlock(&time_sync_lock
.raw_lock
);
930 spin_release(&time_sync_lock
.dep_map
, 1, _THIS_IP_
);
935 static unsigned long long __cpu_clock(int cpu
)
937 unsigned long long now
;
941 * Only call sched_clock() if the scheduler has already been
942 * initialized (some code might call cpu_clock() very early):
944 if (unlikely(!scheduler_running
))
955 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
956 * clock constructed from sched_clock():
958 unsigned long long cpu_clock(int cpu
)
960 unsigned long long prev_cpu_time
, time
, delta_time
;
963 local_irq_save(flags
);
964 prev_cpu_time
= per_cpu(prev_cpu_time
, cpu
);
965 time
= __cpu_clock(cpu
) + per_cpu(time_offset
, cpu
);
966 delta_time
= time
-prev_cpu_time
;
968 if (unlikely(delta_time
> time_sync_thresh
)) {
969 time
= __sync_cpu_clock(time
, cpu
);
970 per_cpu(prev_cpu_time
, cpu
) = time
;
972 local_irq_restore(flags
);
976 EXPORT_SYMBOL_GPL(cpu_clock
);
978 #ifndef prepare_arch_switch
979 # define prepare_arch_switch(next) do { } while (0)
981 #ifndef finish_arch_switch
982 # define finish_arch_switch(prev) do { } while (0)
985 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
987 return rq
->curr
== p
;
990 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
991 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
993 return task_current(rq
, p
);
996 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1000 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1002 #ifdef CONFIG_DEBUG_SPINLOCK
1003 /* this is a valid case when another task releases the spinlock */
1004 rq
->lock
.owner
= current
;
1007 * If we are tracking spinlock dependencies then we have to
1008 * fix up the runqueue lock - which gets 'carried over' from
1009 * prev into current:
1011 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
1013 spin_unlock_irq(&rq
->lock
);
1016 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
1017 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
1022 return task_current(rq
, p
);
1026 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
1030 * We can optimise this out completely for !SMP, because the
1031 * SMP rebalancing from interrupt is the only thing that cares
1036 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1037 spin_unlock_irq(&rq
->lock
);
1039 spin_unlock(&rq
->lock
);
1043 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
1047 * After ->oncpu is cleared, the task can be moved to a different CPU.
1048 * We must ensure this doesn't happen until the switch is completely
1054 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
1058 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
1061 * __task_rq_lock - lock the runqueue a given task resides on.
1062 * Must be called interrupts disabled.
1064 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
1065 __acquires(rq
->lock
)
1068 struct rq
*rq
= task_rq(p
);
1069 spin_lock(&rq
->lock
);
1070 if (likely(rq
== task_rq(p
)))
1072 spin_unlock(&rq
->lock
);
1077 * task_rq_lock - lock the runqueue a given task resides on and disable
1078 * interrupts. Note the ordering: we can safely lookup the task_rq without
1079 * explicitly disabling preemption.
1081 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
1082 __acquires(rq
->lock
)
1087 local_irq_save(*flags
);
1089 spin_lock(&rq
->lock
);
1090 if (likely(rq
== task_rq(p
)))
1092 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1096 static void __task_rq_unlock(struct rq
*rq
)
1097 __releases(rq
->lock
)
1099 spin_unlock(&rq
->lock
);
1102 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1103 __releases(rq
->lock
)
1105 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1109 * this_rq_lock - lock this runqueue and disable interrupts.
1111 static struct rq
*this_rq_lock(void)
1112 __acquires(rq
->lock
)
1116 local_irq_disable();
1118 spin_lock(&rq
->lock
);
1124 * We are going deep-idle (irqs are disabled):
1126 void sched_clock_idle_sleep_event(void)
1128 struct rq
*rq
= cpu_rq(smp_processor_id());
1130 WARN_ON(!irqs_disabled());
1131 spin_lock(&rq
->lock
);
1132 __update_rq_clock(rq
);
1133 spin_unlock(&rq
->lock
);
1134 rq
->clock_deep_idle_events
++;
1136 EXPORT_SYMBOL_GPL(sched_clock_idle_sleep_event
);
1139 * We just idled delta nanoseconds (called with irqs disabled):
1141 void sched_clock_idle_wakeup_event(u64 delta_ns
)
1143 struct rq
*rq
= cpu_rq(smp_processor_id());
1144 u64 now
= sched_clock();
1146 WARN_ON(!irqs_disabled());
1147 rq
->idle_clock
+= delta_ns
;
1149 * Override the previous timestamp and ignore all
1150 * sched_clock() deltas that occured while we idled,
1151 * and use the PM-provided delta_ns to advance the
1154 spin_lock(&rq
->lock
);
1155 rq
->prev_clock_raw
= now
;
1156 rq
->clock
+= delta_ns
;
1157 spin_unlock(&rq
->lock
);
1158 touch_softlockup_watchdog();
1160 EXPORT_SYMBOL_GPL(sched_clock_idle_wakeup_event
);
1162 static void __resched_task(struct task_struct
*p
, int tif_bit
);
1164 static inline void resched_task(struct task_struct
*p
)
1166 __resched_task(p
, TIF_NEED_RESCHED
);
1169 #ifdef CONFIG_SCHED_HRTICK
1171 * Use HR-timers to deliver accurate preemption points.
1173 * Its all a bit involved since we cannot program an hrt while holding the
1174 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1177 * When we get rescheduled we reprogram the hrtick_timer outside of the
1180 static inline void resched_hrt(struct task_struct
*p
)
1182 __resched_task(p
, TIF_HRTICK_RESCHED
);
1185 static inline void resched_rq(struct rq
*rq
)
1187 unsigned long flags
;
1189 spin_lock_irqsave(&rq
->lock
, flags
);
1190 resched_task(rq
->curr
);
1191 spin_unlock_irqrestore(&rq
->lock
, flags
);
1195 HRTICK_SET
, /* re-programm hrtick_timer */
1196 HRTICK_RESET
, /* not a new slice */
1197 HRTICK_BLOCK
, /* stop hrtick operations */
1202 * - enabled by features
1203 * - hrtimer is actually high res
1205 static inline int hrtick_enabled(struct rq
*rq
)
1207 if (!sched_feat(HRTICK
))
1209 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1211 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1215 * Called to set the hrtick timer state.
1217 * called with rq->lock held and irqs disabled
1219 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1221 assert_spin_locked(&rq
->lock
);
1224 * preempt at: now + delay
1227 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1229 * indicate we need to program the timer
1231 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1233 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1236 * New slices are called from the schedule path and don't need a
1237 * forced reschedule.
1240 resched_hrt(rq
->curr
);
1243 static void hrtick_clear(struct rq
*rq
)
1245 if (hrtimer_active(&rq
->hrtick_timer
))
1246 hrtimer_cancel(&rq
->hrtick_timer
);
1250 * Update the timer from the possible pending state.
1252 static void hrtick_set(struct rq
*rq
)
1256 unsigned long flags
;
1258 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1260 spin_lock_irqsave(&rq
->lock
, flags
);
1261 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1262 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1263 time
= rq
->hrtick_expire
;
1264 clear_thread_flag(TIF_HRTICK_RESCHED
);
1265 spin_unlock_irqrestore(&rq
->lock
, flags
);
1268 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1269 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1276 * High-resolution timer tick.
1277 * Runs from hardirq context with interrupts disabled.
1279 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1281 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1283 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1285 spin_lock(&rq
->lock
);
1286 __update_rq_clock(rq
);
1287 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1288 spin_unlock(&rq
->lock
);
1290 return HRTIMER_NORESTART
;
1293 static void hotplug_hrtick_disable(int cpu
)
1295 struct rq
*rq
= cpu_rq(cpu
);
1296 unsigned long flags
;
1298 spin_lock_irqsave(&rq
->lock
, flags
);
1299 rq
->hrtick_flags
= 0;
1300 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1301 spin_unlock_irqrestore(&rq
->lock
, flags
);
1306 static void hotplug_hrtick_enable(int cpu
)
1308 struct rq
*rq
= cpu_rq(cpu
);
1309 unsigned long flags
;
1311 spin_lock_irqsave(&rq
->lock
, flags
);
1312 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1313 spin_unlock_irqrestore(&rq
->lock
, flags
);
1317 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1319 int cpu
= (int)(long)hcpu
;
1322 case CPU_UP_CANCELED
:
1323 case CPU_UP_CANCELED_FROZEN
:
1324 case CPU_DOWN_PREPARE
:
1325 case CPU_DOWN_PREPARE_FROZEN
:
1327 case CPU_DEAD_FROZEN
:
1328 hotplug_hrtick_disable(cpu
);
1331 case CPU_UP_PREPARE
:
1332 case CPU_UP_PREPARE_FROZEN
:
1333 case CPU_DOWN_FAILED
:
1334 case CPU_DOWN_FAILED_FROZEN
:
1336 case CPU_ONLINE_FROZEN
:
1337 hotplug_hrtick_enable(cpu
);
1344 static void init_hrtick(void)
1346 hotcpu_notifier(hotplug_hrtick
, 0);
1349 static void init_rq_hrtick(struct rq
*rq
)
1351 rq
->hrtick_flags
= 0;
1352 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1353 rq
->hrtick_timer
.function
= hrtick
;
1354 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1357 void hrtick_resched(void)
1360 unsigned long flags
;
1362 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1365 local_irq_save(flags
);
1366 rq
= cpu_rq(smp_processor_id());
1368 local_irq_restore(flags
);
1371 static inline void hrtick_clear(struct rq
*rq
)
1375 static inline void hrtick_set(struct rq
*rq
)
1379 static inline void init_rq_hrtick(struct rq
*rq
)
1383 void hrtick_resched(void)
1387 static inline void init_hrtick(void)
1393 * resched_task - mark a task 'to be rescheduled now'.
1395 * On UP this means the setting of the need_resched flag, on SMP it
1396 * might also involve a cross-CPU call to trigger the scheduler on
1401 #ifndef tsk_is_polling
1402 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1405 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1409 assert_spin_locked(&task_rq(p
)->lock
);
1411 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1414 set_tsk_thread_flag(p
, tif_bit
);
1417 if (cpu
== smp_processor_id())
1420 /* NEED_RESCHED must be visible before we test polling */
1422 if (!tsk_is_polling(p
))
1423 smp_send_reschedule(cpu
);
1426 static void resched_cpu(int cpu
)
1428 struct rq
*rq
= cpu_rq(cpu
);
1429 unsigned long flags
;
1431 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1433 resched_task(cpu_curr(cpu
));
1434 spin_unlock_irqrestore(&rq
->lock
, flags
);
1439 * When add_timer_on() enqueues a timer into the timer wheel of an
1440 * idle CPU then this timer might expire before the next timer event
1441 * which is scheduled to wake up that CPU. In case of a completely
1442 * idle system the next event might even be infinite time into the
1443 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1444 * leaves the inner idle loop so the newly added timer is taken into
1445 * account when the CPU goes back to idle and evaluates the timer
1446 * wheel for the next timer event.
1448 void wake_up_idle_cpu(int cpu
)
1450 struct rq
*rq
= cpu_rq(cpu
);
1452 if (cpu
== smp_processor_id())
1456 * This is safe, as this function is called with the timer
1457 * wheel base lock of (cpu) held. When the CPU is on the way
1458 * to idle and has not yet set rq->curr to idle then it will
1459 * be serialized on the timer wheel base lock and take the new
1460 * timer into account automatically.
1462 if (rq
->curr
!= rq
->idle
)
1466 * We can set TIF_RESCHED on the idle task of the other CPU
1467 * lockless. The worst case is that the other CPU runs the
1468 * idle task through an additional NOOP schedule()
1470 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1472 /* NEED_RESCHED must be visible before we test polling */
1474 if (!tsk_is_polling(rq
->idle
))
1475 smp_send_reschedule(cpu
);
1480 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1482 assert_spin_locked(&task_rq(p
)->lock
);
1483 set_tsk_thread_flag(p
, tif_bit
);
1487 #if BITS_PER_LONG == 32
1488 # define WMULT_CONST (~0UL)
1490 # define WMULT_CONST (1UL << 32)
1493 #define WMULT_SHIFT 32
1496 * Shift right and round:
1498 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1501 * delta *= weight / lw
1503 static unsigned long
1504 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1505 struct load_weight
*lw
)
1509 if (!lw
->inv_weight
)
1510 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)/(lw
->weight
+1);
1512 tmp
= (u64
)delta_exec
* weight
;
1514 * Check whether we'd overflow the 64-bit multiplication:
1516 if (unlikely(tmp
> WMULT_CONST
))
1517 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1520 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1522 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1525 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1531 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1538 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1539 * of tasks with abnormal "nice" values across CPUs the contribution that
1540 * each task makes to its run queue's load is weighted according to its
1541 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1542 * scaled version of the new time slice allocation that they receive on time
1546 #define WEIGHT_IDLEPRIO 2
1547 #define WMULT_IDLEPRIO (1 << 31)
1550 * Nice levels are multiplicative, with a gentle 10% change for every
1551 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1552 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1553 * that remained on nice 0.
1555 * The "10% effect" is relative and cumulative: from _any_ nice level,
1556 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1557 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1558 * If a task goes up by ~10% and another task goes down by ~10% then
1559 * the relative distance between them is ~25%.)
1561 static const int prio_to_weight
[40] = {
1562 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1563 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1564 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1565 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1566 /* 0 */ 1024, 820, 655, 526, 423,
1567 /* 5 */ 335, 272, 215, 172, 137,
1568 /* 10 */ 110, 87, 70, 56, 45,
1569 /* 15 */ 36, 29, 23, 18, 15,
1573 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1575 * In cases where the weight does not change often, we can use the
1576 * precalculated inverse to speed up arithmetics by turning divisions
1577 * into multiplications:
1579 static const u32 prio_to_wmult
[40] = {
1580 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1581 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1582 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1583 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1584 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1585 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1586 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1587 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1590 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1593 * runqueue iterator, to support SMP load-balancing between different
1594 * scheduling classes, without having to expose their internal data
1595 * structures to the load-balancing proper:
1597 struct rq_iterator
{
1599 struct task_struct
*(*start
)(void *);
1600 struct task_struct
*(*next
)(void *);
1604 static unsigned long
1605 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1606 unsigned long max_load_move
, struct sched_domain
*sd
,
1607 enum cpu_idle_type idle
, int *all_pinned
,
1608 int *this_best_prio
, struct rq_iterator
*iterator
);
1611 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1612 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1613 struct rq_iterator
*iterator
);
1616 #ifdef CONFIG_CGROUP_CPUACCT
1617 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1619 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1622 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1624 update_load_add(&rq
->load
, load
);
1627 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1629 update_load_sub(&rq
->load
, load
);
1633 static unsigned long source_load(int cpu
, int type
);
1634 static unsigned long target_load(int cpu
, int type
);
1635 static unsigned long cpu_avg_load_per_task(int cpu
);
1636 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1638 #ifdef CONFIG_FAIR_GROUP_SCHED
1641 * Group load balancing.
1643 * We calculate a few balance domain wide aggregate numbers; load and weight.
1644 * Given the pictures below, and assuming each item has equal weight:
1655 * A and B get 1/3-rd of the total load. C and D get 1/3-rd of A's 1/3-rd,
1656 * which equals 1/9-th of the total load.
1659 * The weight of this group on the selected cpus.
1662 * Direct sum of all the cpu's their rq weight, e.g. A would get 3 while
1666 * Part of the rq_weight contributed by tasks; all groups except B would
1670 static inline struct aggregate_struct
*
1671 aggregate(struct task_group
*tg
, struct sched_domain
*sd
)
1673 return &tg
->cfs_rq
[sd
->first_cpu
]->aggregate
;
1676 typedef void (*aggregate_func
)(struct task_group
*, struct sched_domain
*);
1679 * Iterate the full tree, calling @down when first entering a node and @up when
1680 * leaving it for the final time.
1683 void aggregate_walk_tree(aggregate_func down
, aggregate_func up
,
1684 struct sched_domain
*sd
)
1686 struct task_group
*parent
, *child
;
1689 parent
= &root_task_group
;
1691 (*down
)(parent
, sd
);
1692 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1702 parent
= parent
->parent
;
1709 * Calculate the aggregate runqueue weight.
1712 void aggregate_group_weight(struct task_group
*tg
, struct sched_domain
*sd
)
1714 unsigned long rq_weight
= 0;
1715 unsigned long task_weight
= 0;
1718 for_each_cpu_mask(i
, sd
->span
) {
1719 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1720 task_weight
+= tg
->cfs_rq
[i
]->task_weight
;
1723 aggregate(tg
, sd
)->rq_weight
= rq_weight
;
1724 aggregate(tg
, sd
)->task_weight
= task_weight
;
1728 * Compute the weight of this group on the given cpus.
1731 void aggregate_group_shares(struct task_group
*tg
, struct sched_domain
*sd
)
1733 unsigned long shares
= 0;
1736 for_each_cpu_mask(i
, sd
->span
)
1737 shares
+= tg
->cfs_rq
[i
]->shares
;
1739 if ((!shares
&& aggregate(tg
, sd
)->rq_weight
) || shares
> tg
->shares
)
1740 shares
= tg
->shares
;
1742 aggregate(tg
, sd
)->shares
= shares
;
1746 * Compute the load fraction assigned to this group, relies on the aggregate
1747 * weight and this group's parent's load, i.e. top-down.
1750 void aggregate_group_load(struct task_group
*tg
, struct sched_domain
*sd
)
1758 for_each_cpu_mask(i
, sd
->span
)
1759 load
+= cpu_rq(i
)->load
.weight
;
1762 load
= aggregate(tg
->parent
, sd
)->load
;
1765 * shares is our weight in the parent's rq so
1766 * shares/parent->rq_weight gives our fraction of the load
1768 load
*= aggregate(tg
, sd
)->shares
;
1769 load
/= aggregate(tg
->parent
, sd
)->rq_weight
+ 1;
1772 aggregate(tg
, sd
)->load
= load
;
1775 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1778 * Calculate and set the cpu's group shares.
1781 __update_group_shares_cpu(struct task_group
*tg
, struct sched_domain
*sd
,
1785 unsigned long shares
;
1786 unsigned long rq_weight
;
1791 rq_weight
= tg
->cfs_rq
[tcpu
]->load
.weight
;
1794 * If there are currently no tasks on the cpu pretend there is one of
1795 * average load so that when a new task gets to run here it will not
1796 * get delayed by group starvation.
1800 rq_weight
= NICE_0_LOAD
;
1804 * \Sum shares * rq_weight
1805 * shares = -----------------------
1809 shares
= aggregate(tg
, sd
)->shares
* rq_weight
;
1810 shares
/= aggregate(tg
, sd
)->rq_weight
+ 1;
1813 * record the actual number of shares, not the boosted amount.
1815 tg
->cfs_rq
[tcpu
]->shares
= boost
? 0 : shares
;
1817 if (shares
< MIN_SHARES
)
1818 shares
= MIN_SHARES
;
1819 else if (shares
> MAX_SHARES
)
1820 shares
= MAX_SHARES
;
1822 __set_se_shares(tg
->se
[tcpu
], shares
);
1826 * Re-adjust the weights on the cpu the task came from and on the cpu the
1830 __move_group_shares(struct task_group
*tg
, struct sched_domain
*sd
,
1833 unsigned long shares
;
1835 shares
= tg
->cfs_rq
[scpu
]->shares
+ tg
->cfs_rq
[dcpu
]->shares
;
1837 __update_group_shares_cpu(tg
, sd
, scpu
);
1838 __update_group_shares_cpu(tg
, sd
, dcpu
);
1841 * ensure we never loose shares due to rounding errors in the
1842 * above redistribution.
1844 shares
-= tg
->cfs_rq
[scpu
]->shares
+ tg
->cfs_rq
[dcpu
]->shares
;
1846 tg
->cfs_rq
[dcpu
]->shares
+= shares
;
1850 * Because changing a group's shares changes the weight of the super-group
1851 * we need to walk up the tree and change all shares until we hit the root.
1854 move_group_shares(struct task_group
*tg
, struct sched_domain
*sd
,
1858 __move_group_shares(tg
, sd
, scpu
, dcpu
);
1864 void aggregate_group_set_shares(struct task_group
*tg
, struct sched_domain
*sd
)
1866 unsigned long shares
= aggregate(tg
, sd
)->shares
;
1869 for_each_cpu_mask(i
, sd
->span
) {
1870 struct rq
*rq
= cpu_rq(i
);
1871 unsigned long flags
;
1873 spin_lock_irqsave(&rq
->lock
, flags
);
1874 __update_group_shares_cpu(tg
, sd
, i
);
1875 spin_unlock_irqrestore(&rq
->lock
, flags
);
1878 aggregate_group_shares(tg
, sd
);
1881 * ensure we never loose shares due to rounding errors in the
1882 * above redistribution.
1884 shares
-= aggregate(tg
, sd
)->shares
;
1886 tg
->cfs_rq
[sd
->first_cpu
]->shares
+= shares
;
1887 aggregate(tg
, sd
)->shares
+= shares
;
1892 * Calculate the accumulative weight and recursive load of each task group
1893 * while walking down the tree.
1896 void aggregate_get_down(struct task_group
*tg
, struct sched_domain
*sd
)
1898 aggregate_group_weight(tg
, sd
);
1899 aggregate_group_shares(tg
, sd
);
1900 aggregate_group_load(tg
, sd
);
1904 * Rebalance the cpu shares while walking back up the tree.
1907 void aggregate_get_up(struct task_group
*tg
, struct sched_domain
*sd
)
1909 aggregate_group_set_shares(tg
, sd
);
1912 static DEFINE_PER_CPU(spinlock_t
, aggregate_lock
);
1914 static void __init
init_aggregate(void)
1918 for_each_possible_cpu(i
)
1919 spin_lock_init(&per_cpu(aggregate_lock
, i
));
1922 static int get_aggregate(struct sched_domain
*sd
)
1924 if (!spin_trylock(&per_cpu(aggregate_lock
, sd
->first_cpu
)))
1927 aggregate_walk_tree(aggregate_get_down
, aggregate_get_up
, sd
);
1931 static void put_aggregate(struct sched_domain
*sd
)
1933 spin_unlock(&per_cpu(aggregate_lock
, sd
->first_cpu
));
1936 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1938 cfs_rq
->shares
= shares
;
1943 static inline void init_aggregate(void)
1947 static inline int get_aggregate(struct sched_domain
*sd
)
1952 static inline void put_aggregate(struct sched_domain
*sd
)
1957 #else /* CONFIG_SMP */
1959 #ifdef CONFIG_FAIR_GROUP_SCHED
1960 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1965 #endif /* CONFIG_SMP */
1967 #include "sched_stats.h"
1968 #include "sched_idletask.c"
1969 #include "sched_fair.c"
1970 #include "sched_rt.c"
1971 #ifdef CONFIG_SCHED_DEBUG
1972 # include "sched_debug.c"
1975 #define sched_class_highest (&rt_sched_class)
1977 static void inc_nr_running(struct rq
*rq
)
1982 static void dec_nr_running(struct rq
*rq
)
1987 static void set_load_weight(struct task_struct
*p
)
1989 if (task_has_rt_policy(p
)) {
1990 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1991 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1996 * SCHED_IDLE tasks get minimal weight:
1998 if (p
->policy
== SCHED_IDLE
) {
1999 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
2000 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
2004 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
2005 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
2008 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
2010 sched_info_queued(p
);
2011 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
2015 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
2017 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
2022 * __normal_prio - return the priority that is based on the static prio
2024 static inline int __normal_prio(struct task_struct
*p
)
2026 return p
->static_prio
;
2030 * Calculate the expected normal priority: i.e. priority
2031 * without taking RT-inheritance into account. Might be
2032 * boosted by interactivity modifiers. Changes upon fork,
2033 * setprio syscalls, and whenever the interactivity
2034 * estimator recalculates.
2036 static inline int normal_prio(struct task_struct
*p
)
2040 if (task_has_rt_policy(p
))
2041 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
2043 prio
= __normal_prio(p
);
2048 * Calculate the current priority, i.e. the priority
2049 * taken into account by the scheduler. This value might
2050 * be boosted by RT tasks, or might be boosted by
2051 * interactivity modifiers. Will be RT if the task got
2052 * RT-boosted. If not then it returns p->normal_prio.
2054 static int effective_prio(struct task_struct
*p
)
2056 p
->normal_prio
= normal_prio(p
);
2058 * If we are RT tasks or we were boosted to RT priority,
2059 * keep the priority unchanged. Otherwise, update priority
2060 * to the normal priority:
2062 if (!rt_prio(p
->prio
))
2063 return p
->normal_prio
;
2068 * activate_task - move a task to the runqueue.
2070 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
2072 if (task_contributes_to_load(p
))
2073 rq
->nr_uninterruptible
--;
2075 enqueue_task(rq
, p
, wakeup
);
2080 * deactivate_task - remove a task from the runqueue.
2082 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
2084 if (task_contributes_to_load(p
))
2085 rq
->nr_uninterruptible
++;
2087 dequeue_task(rq
, p
, sleep
);
2092 * task_curr - is this task currently executing on a CPU?
2093 * @p: the task in question.
2095 inline int task_curr(const struct task_struct
*p
)
2097 return cpu_curr(task_cpu(p
)) == p
;
2100 /* Used instead of source_load when we know the type == 0 */
2101 unsigned long weighted_cpuload(const int cpu
)
2103 return cpu_rq(cpu
)->load
.weight
;
2106 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
2108 set_task_rq(p
, cpu
);
2111 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
2112 * successfuly executed on another CPU. We must ensure that updates of
2113 * per-task data have been completed by this moment.
2116 task_thread_info(p
)->cpu
= cpu
;
2120 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
2121 const struct sched_class
*prev_class
,
2122 int oldprio
, int running
)
2124 if (prev_class
!= p
->sched_class
) {
2125 if (prev_class
->switched_from
)
2126 prev_class
->switched_from(rq
, p
, running
);
2127 p
->sched_class
->switched_to(rq
, p
, running
);
2129 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2135 * Is this task likely cache-hot:
2138 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2143 * Buddy candidates are cache hot:
2145 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
2148 if (p
->sched_class
!= &fair_sched_class
)
2151 if (sysctl_sched_migration_cost
== -1)
2153 if (sysctl_sched_migration_cost
== 0)
2156 delta
= now
- p
->se
.exec_start
;
2158 return delta
< (s64
)sysctl_sched_migration_cost
;
2162 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2164 int old_cpu
= task_cpu(p
);
2165 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
2166 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2167 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2170 clock_offset
= old_rq
->clock
- new_rq
->clock
;
2172 #ifdef CONFIG_SCHEDSTATS
2173 if (p
->se
.wait_start
)
2174 p
->se
.wait_start
-= clock_offset
;
2175 if (p
->se
.sleep_start
)
2176 p
->se
.sleep_start
-= clock_offset
;
2177 if (p
->se
.block_start
)
2178 p
->se
.block_start
-= clock_offset
;
2179 if (old_cpu
!= new_cpu
) {
2180 schedstat_inc(p
, se
.nr_migrations
);
2181 if (task_hot(p
, old_rq
->clock
, NULL
))
2182 schedstat_inc(p
, se
.nr_forced2_migrations
);
2185 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2186 new_cfsrq
->min_vruntime
;
2188 __set_task_cpu(p
, new_cpu
);
2191 struct migration_req
{
2192 struct list_head list
;
2194 struct task_struct
*task
;
2197 struct completion done
;
2201 * The task's runqueue lock must be held.
2202 * Returns true if you have to wait for migration thread.
2205 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2207 struct rq
*rq
= task_rq(p
);
2210 * If the task is not on a runqueue (and not running), then
2211 * it is sufficient to simply update the task's cpu field.
2213 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
2214 set_task_cpu(p
, dest_cpu
);
2218 init_completion(&req
->done
);
2220 req
->dest_cpu
= dest_cpu
;
2221 list_add(&req
->list
, &rq
->migration_queue
);
2227 * wait_task_inactive - wait for a thread to unschedule.
2229 * The caller must ensure that the task *will* unschedule sometime soon,
2230 * else this function might spin for a *long* time. This function can't
2231 * be called with interrupts off, or it may introduce deadlock with
2232 * smp_call_function() if an IPI is sent by the same process we are
2233 * waiting to become inactive.
2235 void wait_task_inactive(struct task_struct
*p
)
2237 unsigned long flags
;
2243 * We do the initial early heuristics without holding
2244 * any task-queue locks at all. We'll only try to get
2245 * the runqueue lock when things look like they will
2251 * If the task is actively running on another CPU
2252 * still, just relax and busy-wait without holding
2255 * NOTE! Since we don't hold any locks, it's not
2256 * even sure that "rq" stays as the right runqueue!
2257 * But we don't care, since "task_running()" will
2258 * return false if the runqueue has changed and p
2259 * is actually now running somewhere else!
2261 while (task_running(rq
, p
))
2265 * Ok, time to look more closely! We need the rq
2266 * lock now, to be *sure*. If we're wrong, we'll
2267 * just go back and repeat.
2269 rq
= task_rq_lock(p
, &flags
);
2270 running
= task_running(rq
, p
);
2271 on_rq
= p
->se
.on_rq
;
2272 task_rq_unlock(rq
, &flags
);
2275 * Was it really running after all now that we
2276 * checked with the proper locks actually held?
2278 * Oops. Go back and try again..
2280 if (unlikely(running
)) {
2286 * It's not enough that it's not actively running,
2287 * it must be off the runqueue _entirely_, and not
2290 * So if it wa still runnable (but just not actively
2291 * running right now), it's preempted, and we should
2292 * yield - it could be a while.
2294 if (unlikely(on_rq
)) {
2295 schedule_timeout_uninterruptible(1);
2300 * Ahh, all good. It wasn't running, and it wasn't
2301 * runnable, which means that it will never become
2302 * running in the future either. We're all done!
2309 * kick_process - kick a running thread to enter/exit the kernel
2310 * @p: the to-be-kicked thread
2312 * Cause a process which is running on another CPU to enter
2313 * kernel-mode, without any delay. (to get signals handled.)
2315 * NOTE: this function doesnt have to take the runqueue lock,
2316 * because all it wants to ensure is that the remote task enters
2317 * the kernel. If the IPI races and the task has been migrated
2318 * to another CPU then no harm is done and the purpose has been
2321 void kick_process(struct task_struct
*p
)
2327 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2328 smp_send_reschedule(cpu
);
2333 * Return a low guess at the load of a migration-source cpu weighted
2334 * according to the scheduling class and "nice" value.
2336 * We want to under-estimate the load of migration sources, to
2337 * balance conservatively.
2339 static unsigned long source_load(int cpu
, int type
)
2341 struct rq
*rq
= cpu_rq(cpu
);
2342 unsigned long total
= weighted_cpuload(cpu
);
2347 return min(rq
->cpu_load
[type
-1], total
);
2351 * Return a high guess at the load of a migration-target cpu weighted
2352 * according to the scheduling class and "nice" value.
2354 static unsigned long target_load(int cpu
, int type
)
2356 struct rq
*rq
= cpu_rq(cpu
);
2357 unsigned long total
= weighted_cpuload(cpu
);
2362 return max(rq
->cpu_load
[type
-1], total
);
2366 * Return the average load per task on the cpu's run queue
2368 static unsigned long cpu_avg_load_per_task(int cpu
)
2370 struct rq
*rq
= cpu_rq(cpu
);
2371 unsigned long total
= weighted_cpuload(cpu
);
2372 unsigned long n
= rq
->nr_running
;
2374 return n
? total
/ n
: SCHED_LOAD_SCALE
;
2378 * find_idlest_group finds and returns the least busy CPU group within the
2381 static struct sched_group
*
2382 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2384 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2385 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2386 int load_idx
= sd
->forkexec_idx
;
2387 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2390 unsigned long load
, avg_load
;
2394 /* Skip over this group if it has no CPUs allowed */
2395 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2398 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2400 /* Tally up the load of all CPUs in the group */
2403 for_each_cpu_mask(i
, group
->cpumask
) {
2404 /* Bias balancing toward cpus of our domain */
2406 load
= source_load(i
, load_idx
);
2408 load
= target_load(i
, load_idx
);
2413 /* Adjust by relative CPU power of the group */
2414 avg_load
= sg_div_cpu_power(group
,
2415 avg_load
* SCHED_LOAD_SCALE
);
2418 this_load
= avg_load
;
2420 } else if (avg_load
< min_load
) {
2421 min_load
= avg_load
;
2424 } while (group
= group
->next
, group
!= sd
->groups
);
2426 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2432 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2435 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2438 unsigned long load
, min_load
= ULONG_MAX
;
2442 /* Traverse only the allowed CPUs */
2443 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2445 for_each_cpu_mask(i
, *tmp
) {
2446 load
= weighted_cpuload(i
);
2448 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2458 * sched_balance_self: balance the current task (running on cpu) in domains
2459 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2462 * Balance, ie. select the least loaded group.
2464 * Returns the target CPU number, or the same CPU if no balancing is needed.
2466 * preempt must be disabled.
2468 static int sched_balance_self(int cpu
, int flag
)
2470 struct task_struct
*t
= current
;
2471 struct sched_domain
*tmp
, *sd
= NULL
;
2473 for_each_domain(cpu
, tmp
) {
2475 * If power savings logic is enabled for a domain, stop there.
2477 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2479 if (tmp
->flags
& flag
)
2484 cpumask_t span
, tmpmask
;
2485 struct sched_group
*group
;
2486 int new_cpu
, weight
;
2488 if (!(sd
->flags
& flag
)) {
2494 group
= find_idlest_group(sd
, t
, cpu
);
2500 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2501 if (new_cpu
== -1 || new_cpu
== cpu
) {
2502 /* Now try balancing at a lower domain level of cpu */
2507 /* Now try balancing at a lower domain level of new_cpu */
2510 weight
= cpus_weight(span
);
2511 for_each_domain(cpu
, tmp
) {
2512 if (weight
<= cpus_weight(tmp
->span
))
2514 if (tmp
->flags
& flag
)
2517 /* while loop will break here if sd == NULL */
2523 #endif /* CONFIG_SMP */
2526 * try_to_wake_up - wake up a thread
2527 * @p: the to-be-woken-up thread
2528 * @state: the mask of task states that can be woken
2529 * @sync: do a synchronous wakeup?
2531 * Put it on the run-queue if it's not already there. The "current"
2532 * thread is always on the run-queue (except when the actual
2533 * re-schedule is in progress), and as such you're allowed to do
2534 * the simpler "current->state = TASK_RUNNING" to mark yourself
2535 * runnable without the overhead of this.
2537 * returns failure only if the task is already active.
2539 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2541 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2542 unsigned long flags
;
2546 if (!sched_feat(SYNC_WAKEUPS
))
2550 rq
= task_rq_lock(p
, &flags
);
2551 old_state
= p
->state
;
2552 if (!(old_state
& state
))
2560 this_cpu
= smp_processor_id();
2563 if (unlikely(task_running(rq
, p
)))
2566 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2567 if (cpu
!= orig_cpu
) {
2568 set_task_cpu(p
, cpu
);
2569 task_rq_unlock(rq
, &flags
);
2570 /* might preempt at this point */
2571 rq
= task_rq_lock(p
, &flags
);
2572 old_state
= p
->state
;
2573 if (!(old_state
& state
))
2578 this_cpu
= smp_processor_id();
2582 #ifdef CONFIG_SCHEDSTATS
2583 schedstat_inc(rq
, ttwu_count
);
2584 if (cpu
== this_cpu
)
2585 schedstat_inc(rq
, ttwu_local
);
2587 struct sched_domain
*sd
;
2588 for_each_domain(this_cpu
, sd
) {
2589 if (cpu_isset(cpu
, sd
->span
)) {
2590 schedstat_inc(sd
, ttwu_wake_remote
);
2598 #endif /* CONFIG_SMP */
2599 schedstat_inc(p
, se
.nr_wakeups
);
2601 schedstat_inc(p
, se
.nr_wakeups_sync
);
2602 if (orig_cpu
!= cpu
)
2603 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2604 if (cpu
== this_cpu
)
2605 schedstat_inc(p
, se
.nr_wakeups_local
);
2607 schedstat_inc(p
, se
.nr_wakeups_remote
);
2608 update_rq_clock(rq
);
2609 activate_task(rq
, p
, 1);
2613 check_preempt_curr(rq
, p
);
2615 p
->state
= TASK_RUNNING
;
2617 if (p
->sched_class
->task_wake_up
)
2618 p
->sched_class
->task_wake_up(rq
, p
);
2621 task_rq_unlock(rq
, &flags
);
2626 int wake_up_process(struct task_struct
*p
)
2628 return try_to_wake_up(p
, TASK_ALL
, 0);
2630 EXPORT_SYMBOL(wake_up_process
);
2632 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2634 return try_to_wake_up(p
, state
, 0);
2638 * Perform scheduler related setup for a newly forked process p.
2639 * p is forked by current.
2641 * __sched_fork() is basic setup used by init_idle() too:
2643 static void __sched_fork(struct task_struct
*p
)
2645 p
->se
.exec_start
= 0;
2646 p
->se
.sum_exec_runtime
= 0;
2647 p
->se
.prev_sum_exec_runtime
= 0;
2648 p
->se
.last_wakeup
= 0;
2649 p
->se
.avg_overlap
= 0;
2651 #ifdef CONFIG_SCHEDSTATS
2652 p
->se
.wait_start
= 0;
2653 p
->se
.sum_sleep_runtime
= 0;
2654 p
->se
.sleep_start
= 0;
2655 p
->se
.block_start
= 0;
2656 p
->se
.sleep_max
= 0;
2657 p
->se
.block_max
= 0;
2659 p
->se
.slice_max
= 0;
2663 INIT_LIST_HEAD(&p
->rt
.run_list
);
2665 INIT_LIST_HEAD(&p
->se
.group_node
);
2667 #ifdef CONFIG_PREEMPT_NOTIFIERS
2668 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2672 * We mark the process as running here, but have not actually
2673 * inserted it onto the runqueue yet. This guarantees that
2674 * nobody will actually run it, and a signal or other external
2675 * event cannot wake it up and insert it on the runqueue either.
2677 p
->state
= TASK_RUNNING
;
2681 * fork()/clone()-time setup:
2683 void sched_fork(struct task_struct
*p
, int clone_flags
)
2685 int cpu
= get_cpu();
2690 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2692 set_task_cpu(p
, cpu
);
2695 * Make sure we do not leak PI boosting priority to the child:
2697 p
->prio
= current
->normal_prio
;
2698 if (!rt_prio(p
->prio
))
2699 p
->sched_class
= &fair_sched_class
;
2701 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2702 if (likely(sched_info_on()))
2703 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2705 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2708 #ifdef CONFIG_PREEMPT
2709 /* Want to start with kernel preemption disabled. */
2710 task_thread_info(p
)->preempt_count
= 1;
2716 * wake_up_new_task - wake up a newly created task for the first time.
2718 * This function will do some initial scheduler statistics housekeeping
2719 * that must be done for every newly created context, then puts the task
2720 * on the runqueue and wakes it.
2722 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2724 unsigned long flags
;
2727 rq
= task_rq_lock(p
, &flags
);
2728 BUG_ON(p
->state
!= TASK_RUNNING
);
2729 update_rq_clock(rq
);
2731 p
->prio
= effective_prio(p
);
2733 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2734 activate_task(rq
, p
, 0);
2737 * Let the scheduling class do new task startup
2738 * management (if any):
2740 p
->sched_class
->task_new(rq
, p
);
2743 check_preempt_curr(rq
, p
);
2745 if (p
->sched_class
->task_wake_up
)
2746 p
->sched_class
->task_wake_up(rq
, p
);
2748 task_rq_unlock(rq
, &flags
);
2751 #ifdef CONFIG_PREEMPT_NOTIFIERS
2754 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2755 * @notifier: notifier struct to register
2757 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2759 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2761 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2764 * preempt_notifier_unregister - no longer interested in preemption notifications
2765 * @notifier: notifier struct to unregister
2767 * This is safe to call from within a preemption notifier.
2769 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2771 hlist_del(¬ifier
->link
);
2773 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2775 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2777 struct preempt_notifier
*notifier
;
2778 struct hlist_node
*node
;
2780 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2781 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2785 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2786 struct task_struct
*next
)
2788 struct preempt_notifier
*notifier
;
2789 struct hlist_node
*node
;
2791 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2792 notifier
->ops
->sched_out(notifier
, next
);
2797 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2802 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2803 struct task_struct
*next
)
2810 * prepare_task_switch - prepare to switch tasks
2811 * @rq: the runqueue preparing to switch
2812 * @prev: the current task that is being switched out
2813 * @next: the task we are going to switch to.
2815 * This is called with the rq lock held and interrupts off. It must
2816 * be paired with a subsequent finish_task_switch after the context
2819 * prepare_task_switch sets up locking and calls architecture specific
2823 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2824 struct task_struct
*next
)
2826 fire_sched_out_preempt_notifiers(prev
, next
);
2827 prepare_lock_switch(rq
, next
);
2828 prepare_arch_switch(next
);
2832 * finish_task_switch - clean up after a task-switch
2833 * @rq: runqueue associated with task-switch
2834 * @prev: the thread we just switched away from.
2836 * finish_task_switch must be called after the context switch, paired
2837 * with a prepare_task_switch call before the context switch.
2838 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2839 * and do any other architecture-specific cleanup actions.
2841 * Note that we may have delayed dropping an mm in context_switch(). If
2842 * so, we finish that here outside of the runqueue lock. (Doing it
2843 * with the lock held can cause deadlocks; see schedule() for
2846 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2847 __releases(rq
->lock
)
2849 struct mm_struct
*mm
= rq
->prev_mm
;
2855 * A task struct has one reference for the use as "current".
2856 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2857 * schedule one last time. The schedule call will never return, and
2858 * the scheduled task must drop that reference.
2859 * The test for TASK_DEAD must occur while the runqueue locks are
2860 * still held, otherwise prev could be scheduled on another cpu, die
2861 * there before we look at prev->state, and then the reference would
2863 * Manfred Spraul <manfred@colorfullife.com>
2865 prev_state
= prev
->state
;
2866 finish_arch_switch(prev
);
2867 finish_lock_switch(rq
, prev
);
2869 if (current
->sched_class
->post_schedule
)
2870 current
->sched_class
->post_schedule(rq
);
2873 fire_sched_in_preempt_notifiers(current
);
2876 if (unlikely(prev_state
== TASK_DEAD
)) {
2878 * Remove function-return probe instances associated with this
2879 * task and put them back on the free list.
2881 kprobe_flush_task(prev
);
2882 put_task_struct(prev
);
2887 * schedule_tail - first thing a freshly forked thread must call.
2888 * @prev: the thread we just switched away from.
2890 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2891 __releases(rq
->lock
)
2893 struct rq
*rq
= this_rq();
2895 finish_task_switch(rq
, prev
);
2896 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2897 /* In this case, finish_task_switch does not reenable preemption */
2900 if (current
->set_child_tid
)
2901 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2905 * context_switch - switch to the new MM and the new
2906 * thread's register state.
2909 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2910 struct task_struct
*next
)
2912 struct mm_struct
*mm
, *oldmm
;
2914 prepare_task_switch(rq
, prev
, next
);
2916 oldmm
= prev
->active_mm
;
2918 * For paravirt, this is coupled with an exit in switch_to to
2919 * combine the page table reload and the switch backend into
2922 arch_enter_lazy_cpu_mode();
2924 if (unlikely(!mm
)) {
2925 next
->active_mm
= oldmm
;
2926 atomic_inc(&oldmm
->mm_count
);
2927 enter_lazy_tlb(oldmm
, next
);
2929 switch_mm(oldmm
, mm
, next
);
2931 if (unlikely(!prev
->mm
)) {
2932 prev
->active_mm
= NULL
;
2933 rq
->prev_mm
= oldmm
;
2936 * Since the runqueue lock will be released by the next
2937 * task (which is an invalid locking op but in the case
2938 * of the scheduler it's an obvious special-case), so we
2939 * do an early lockdep release here:
2941 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2942 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2945 /* Here we just switch the register state and the stack. */
2946 switch_to(prev
, next
, prev
);
2950 * this_rq must be evaluated again because prev may have moved
2951 * CPUs since it called schedule(), thus the 'rq' on its stack
2952 * frame will be invalid.
2954 finish_task_switch(this_rq(), prev
);
2958 * nr_running, nr_uninterruptible and nr_context_switches:
2960 * externally visible scheduler statistics: current number of runnable
2961 * threads, current number of uninterruptible-sleeping threads, total
2962 * number of context switches performed since bootup.
2964 unsigned long nr_running(void)
2966 unsigned long i
, sum
= 0;
2968 for_each_online_cpu(i
)
2969 sum
+= cpu_rq(i
)->nr_running
;
2974 unsigned long nr_uninterruptible(void)
2976 unsigned long i
, sum
= 0;
2978 for_each_possible_cpu(i
)
2979 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2982 * Since we read the counters lockless, it might be slightly
2983 * inaccurate. Do not allow it to go below zero though:
2985 if (unlikely((long)sum
< 0))
2991 unsigned long long nr_context_switches(void)
2994 unsigned long long sum
= 0;
2996 for_each_possible_cpu(i
)
2997 sum
+= cpu_rq(i
)->nr_switches
;
3002 unsigned long nr_iowait(void)
3004 unsigned long i
, sum
= 0;
3006 for_each_possible_cpu(i
)
3007 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
3012 unsigned long nr_active(void)
3014 unsigned long i
, running
= 0, uninterruptible
= 0;
3016 for_each_online_cpu(i
) {
3017 running
+= cpu_rq(i
)->nr_running
;
3018 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
3021 if (unlikely((long)uninterruptible
< 0))
3022 uninterruptible
= 0;
3024 return running
+ uninterruptible
;
3028 * Update rq->cpu_load[] statistics. This function is usually called every
3029 * scheduler tick (TICK_NSEC).
3031 static void update_cpu_load(struct rq
*this_rq
)
3033 unsigned long this_load
= this_rq
->load
.weight
;
3036 this_rq
->nr_load_updates
++;
3038 /* Update our load: */
3039 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3040 unsigned long old_load
, new_load
;
3042 /* scale is effectively 1 << i now, and >> i divides by scale */
3044 old_load
= this_rq
->cpu_load
[i
];
3045 new_load
= this_load
;
3047 * Round up the averaging division if load is increasing. This
3048 * prevents us from getting stuck on 9 if the load is 10, for
3051 if (new_load
> old_load
)
3052 new_load
+= scale
-1;
3053 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3060 * double_rq_lock - safely lock two runqueues
3062 * Note this does not disable interrupts like task_rq_lock,
3063 * you need to do so manually before calling.
3065 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3066 __acquires(rq1
->lock
)
3067 __acquires(rq2
->lock
)
3069 BUG_ON(!irqs_disabled());
3071 spin_lock(&rq1
->lock
);
3072 __acquire(rq2
->lock
); /* Fake it out ;) */
3075 spin_lock(&rq1
->lock
);
3076 spin_lock(&rq2
->lock
);
3078 spin_lock(&rq2
->lock
);
3079 spin_lock(&rq1
->lock
);
3082 update_rq_clock(rq1
);
3083 update_rq_clock(rq2
);
3087 * double_rq_unlock - safely unlock two runqueues
3089 * Note this does not restore interrupts like task_rq_unlock,
3090 * you need to do so manually after calling.
3092 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3093 __releases(rq1
->lock
)
3094 __releases(rq2
->lock
)
3096 spin_unlock(&rq1
->lock
);
3098 spin_unlock(&rq2
->lock
);
3100 __release(rq2
->lock
);
3104 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
3106 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
3107 __releases(this_rq
->lock
)
3108 __acquires(busiest
->lock
)
3109 __acquires(this_rq
->lock
)
3113 if (unlikely(!irqs_disabled())) {
3114 /* printk() doesn't work good under rq->lock */
3115 spin_unlock(&this_rq
->lock
);
3118 if (unlikely(!spin_trylock(&busiest
->lock
))) {
3119 if (busiest
< this_rq
) {
3120 spin_unlock(&this_rq
->lock
);
3121 spin_lock(&busiest
->lock
);
3122 spin_lock(&this_rq
->lock
);
3125 spin_lock(&busiest
->lock
);
3131 * If dest_cpu is allowed for this process, migrate the task to it.
3132 * This is accomplished by forcing the cpu_allowed mask to only
3133 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3134 * the cpu_allowed mask is restored.
3136 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3138 struct migration_req req
;
3139 unsigned long flags
;
3142 rq
= task_rq_lock(p
, &flags
);
3143 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
3144 || unlikely(cpu_is_offline(dest_cpu
)))
3147 /* force the process onto the specified CPU */
3148 if (migrate_task(p
, dest_cpu
, &req
)) {
3149 /* Need to wait for migration thread (might exit: take ref). */
3150 struct task_struct
*mt
= rq
->migration_thread
;
3152 get_task_struct(mt
);
3153 task_rq_unlock(rq
, &flags
);
3154 wake_up_process(mt
);
3155 put_task_struct(mt
);
3156 wait_for_completion(&req
.done
);
3161 task_rq_unlock(rq
, &flags
);
3165 * sched_exec - execve() is a valuable balancing opportunity, because at
3166 * this point the task has the smallest effective memory and cache footprint.
3168 void sched_exec(void)
3170 int new_cpu
, this_cpu
= get_cpu();
3171 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
3173 if (new_cpu
!= this_cpu
)
3174 sched_migrate_task(current
, new_cpu
);
3178 * pull_task - move a task from a remote runqueue to the local runqueue.
3179 * Both runqueues must be locked.
3181 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3182 struct rq
*this_rq
, int this_cpu
)
3184 deactivate_task(src_rq
, p
, 0);
3185 set_task_cpu(p
, this_cpu
);
3186 activate_task(this_rq
, p
, 0);
3188 * Note that idle threads have a prio of MAX_PRIO, for this test
3189 * to be always true for them.
3191 check_preempt_curr(this_rq
, p
);
3195 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3198 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3199 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3203 * We do not migrate tasks that are:
3204 * 1) running (obviously), or
3205 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3206 * 3) are cache-hot on their current CPU.
3208 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
3209 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3214 if (task_running(rq
, p
)) {
3215 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3220 * Aggressive migration if:
3221 * 1) task is cache cold, or
3222 * 2) too many balance attempts have failed.
3225 if (!task_hot(p
, rq
->clock
, sd
) ||
3226 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3227 #ifdef CONFIG_SCHEDSTATS
3228 if (task_hot(p
, rq
->clock
, sd
)) {
3229 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3230 schedstat_inc(p
, se
.nr_forced_migrations
);
3236 if (task_hot(p
, rq
->clock
, sd
)) {
3237 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3243 static unsigned long
3244 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3245 unsigned long max_load_move
, struct sched_domain
*sd
,
3246 enum cpu_idle_type idle
, int *all_pinned
,
3247 int *this_best_prio
, struct rq_iterator
*iterator
)
3249 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
3250 struct task_struct
*p
;
3251 long rem_load_move
= max_load_move
;
3253 if (max_load_move
== 0)
3259 * Start the load-balancing iterator:
3261 p
= iterator
->start(iterator
->arg
);
3263 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3266 * To help distribute high priority tasks across CPUs we don't
3267 * skip a task if it will be the highest priority task (i.e. smallest
3268 * prio value) on its new queue regardless of its load weight
3270 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
3271 SCHED_LOAD_SCALE_FUZZ
;
3272 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
3273 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3274 p
= iterator
->next(iterator
->arg
);
3278 pull_task(busiest
, p
, this_rq
, this_cpu
);
3280 rem_load_move
-= p
->se
.load
.weight
;
3283 * We only want to steal up to the prescribed amount of weighted load.
3285 if (rem_load_move
> 0) {
3286 if (p
->prio
< *this_best_prio
)
3287 *this_best_prio
= p
->prio
;
3288 p
= iterator
->next(iterator
->arg
);
3293 * Right now, this is one of only two places pull_task() is called,
3294 * so we can safely collect pull_task() stats here rather than
3295 * inside pull_task().
3297 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3300 *all_pinned
= pinned
;
3302 return max_load_move
- rem_load_move
;
3306 * move_tasks tries to move up to max_load_move weighted load from busiest to
3307 * this_rq, as part of a balancing operation within domain "sd".
3308 * Returns 1 if successful and 0 otherwise.
3310 * Called with both runqueues locked.
3312 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3313 unsigned long max_load_move
,
3314 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3317 const struct sched_class
*class = sched_class_highest
;
3318 unsigned long total_load_moved
= 0;
3319 int this_best_prio
= this_rq
->curr
->prio
;
3323 class->load_balance(this_rq
, this_cpu
, busiest
,
3324 max_load_move
- total_load_moved
,
3325 sd
, idle
, all_pinned
, &this_best_prio
);
3326 class = class->next
;
3327 } while (class && max_load_move
> total_load_moved
);
3329 return total_load_moved
> 0;
3333 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3334 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3335 struct rq_iterator
*iterator
)
3337 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3341 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3342 pull_task(busiest
, p
, this_rq
, this_cpu
);
3344 * Right now, this is only the second place pull_task()
3345 * is called, so we can safely collect pull_task()
3346 * stats here rather than inside pull_task().
3348 schedstat_inc(sd
, lb_gained
[idle
]);
3352 p
= iterator
->next(iterator
->arg
);
3359 * move_one_task tries to move exactly one task from busiest to this_rq, as
3360 * part of active balancing operations within "domain".
3361 * Returns 1 if successful and 0 otherwise.
3363 * Called with both runqueues locked.
3365 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3366 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3368 const struct sched_class
*class;
3370 for (class = sched_class_highest
; class; class = class->next
)
3371 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3378 * find_busiest_group finds and returns the busiest CPU group within the
3379 * domain. It calculates and returns the amount of weighted load which
3380 * should be moved to restore balance via the imbalance parameter.
3382 static struct sched_group
*
3383 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3384 unsigned long *imbalance
, enum cpu_idle_type idle
,
3385 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3387 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3388 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3389 unsigned long max_pull
;
3390 unsigned long busiest_load_per_task
, busiest_nr_running
;
3391 unsigned long this_load_per_task
, this_nr_running
;
3392 int load_idx
, group_imb
= 0;
3393 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3394 int power_savings_balance
= 1;
3395 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3396 unsigned long min_nr_running
= ULONG_MAX
;
3397 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3400 max_load
= this_load
= total_load
= total_pwr
= 0;
3401 busiest_load_per_task
= busiest_nr_running
= 0;
3402 this_load_per_task
= this_nr_running
= 0;
3403 if (idle
== CPU_NOT_IDLE
)
3404 load_idx
= sd
->busy_idx
;
3405 else if (idle
== CPU_NEWLY_IDLE
)
3406 load_idx
= sd
->newidle_idx
;
3408 load_idx
= sd
->idle_idx
;
3411 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3414 int __group_imb
= 0;
3415 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3416 unsigned long sum_nr_running
, sum_weighted_load
;
3418 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3421 balance_cpu
= first_cpu(group
->cpumask
);
3423 /* Tally up the load of all CPUs in the group */
3424 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3426 min_cpu_load
= ~0UL;
3428 for_each_cpu_mask(i
, group
->cpumask
) {
3431 if (!cpu_isset(i
, *cpus
))
3436 if (*sd_idle
&& rq
->nr_running
)
3439 /* Bias balancing toward cpus of our domain */
3441 if (idle_cpu(i
) && !first_idle_cpu
) {
3446 load
= target_load(i
, load_idx
);
3448 load
= source_load(i
, load_idx
);
3449 if (load
> max_cpu_load
)
3450 max_cpu_load
= load
;
3451 if (min_cpu_load
> load
)
3452 min_cpu_load
= load
;
3456 sum_nr_running
+= rq
->nr_running
;
3457 sum_weighted_load
+= weighted_cpuload(i
);
3461 * First idle cpu or the first cpu(busiest) in this sched group
3462 * is eligible for doing load balancing at this and above
3463 * domains. In the newly idle case, we will allow all the cpu's
3464 * to do the newly idle load balance.
3466 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3467 balance_cpu
!= this_cpu
&& balance
) {
3472 total_load
+= avg_load
;
3473 total_pwr
+= group
->__cpu_power
;
3475 /* Adjust by relative CPU power of the group */
3476 avg_load
= sg_div_cpu_power(group
,
3477 avg_load
* SCHED_LOAD_SCALE
);
3479 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3482 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3485 this_load
= avg_load
;
3487 this_nr_running
= sum_nr_running
;
3488 this_load_per_task
= sum_weighted_load
;
3489 } else if (avg_load
> max_load
&&
3490 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3491 max_load
= avg_load
;
3493 busiest_nr_running
= sum_nr_running
;
3494 busiest_load_per_task
= sum_weighted_load
;
3495 group_imb
= __group_imb
;
3498 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3500 * Busy processors will not participate in power savings
3503 if (idle
== CPU_NOT_IDLE
||
3504 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3508 * If the local group is idle or completely loaded
3509 * no need to do power savings balance at this domain
3511 if (local_group
&& (this_nr_running
>= group_capacity
||
3513 power_savings_balance
= 0;
3516 * If a group is already running at full capacity or idle,
3517 * don't include that group in power savings calculations
3519 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3524 * Calculate the group which has the least non-idle load.
3525 * This is the group from where we need to pick up the load
3528 if ((sum_nr_running
< min_nr_running
) ||
3529 (sum_nr_running
== min_nr_running
&&
3530 first_cpu(group
->cpumask
) <
3531 first_cpu(group_min
->cpumask
))) {
3533 min_nr_running
= sum_nr_running
;
3534 min_load_per_task
= sum_weighted_load
/
3539 * Calculate the group which is almost near its
3540 * capacity but still has some space to pick up some load
3541 * from other group and save more power
3543 if (sum_nr_running
<= group_capacity
- 1) {
3544 if (sum_nr_running
> leader_nr_running
||
3545 (sum_nr_running
== leader_nr_running
&&
3546 first_cpu(group
->cpumask
) >
3547 first_cpu(group_leader
->cpumask
))) {
3548 group_leader
= group
;
3549 leader_nr_running
= sum_nr_running
;
3554 group
= group
->next
;
3555 } while (group
!= sd
->groups
);
3557 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3560 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3562 if (this_load
>= avg_load
||
3563 100*max_load
<= sd
->imbalance_pct
*this_load
)
3566 busiest_load_per_task
/= busiest_nr_running
;
3568 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3571 * We're trying to get all the cpus to the average_load, so we don't
3572 * want to push ourselves above the average load, nor do we wish to
3573 * reduce the max loaded cpu below the average load, as either of these
3574 * actions would just result in more rebalancing later, and ping-pong
3575 * tasks around. Thus we look for the minimum possible imbalance.
3576 * Negative imbalances (*we* are more loaded than anyone else) will
3577 * be counted as no imbalance for these purposes -- we can't fix that
3578 * by pulling tasks to us. Be careful of negative numbers as they'll
3579 * appear as very large values with unsigned longs.
3581 if (max_load
<= busiest_load_per_task
)
3585 * In the presence of smp nice balancing, certain scenarios can have
3586 * max load less than avg load(as we skip the groups at or below
3587 * its cpu_power, while calculating max_load..)
3589 if (max_load
< avg_load
) {
3591 goto small_imbalance
;
3594 /* Don't want to pull so many tasks that a group would go idle */
3595 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3597 /* How much load to actually move to equalise the imbalance */
3598 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3599 (avg_load
- this_load
) * this->__cpu_power
)
3603 * if *imbalance is less than the average load per runnable task
3604 * there is no gaurantee that any tasks will be moved so we'll have
3605 * a think about bumping its value to force at least one task to be
3608 if (*imbalance
< busiest_load_per_task
) {
3609 unsigned long tmp
, pwr_now
, pwr_move
;
3613 pwr_move
= pwr_now
= 0;
3615 if (this_nr_running
) {
3616 this_load_per_task
/= this_nr_running
;
3617 if (busiest_load_per_task
> this_load_per_task
)
3620 this_load_per_task
= SCHED_LOAD_SCALE
;
3622 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3623 busiest_load_per_task
* imbn
) {
3624 *imbalance
= busiest_load_per_task
;
3629 * OK, we don't have enough imbalance to justify moving tasks,
3630 * however we may be able to increase total CPU power used by
3634 pwr_now
+= busiest
->__cpu_power
*
3635 min(busiest_load_per_task
, max_load
);
3636 pwr_now
+= this->__cpu_power
*
3637 min(this_load_per_task
, this_load
);
3638 pwr_now
/= SCHED_LOAD_SCALE
;
3640 /* Amount of load we'd subtract */
3641 tmp
= sg_div_cpu_power(busiest
,
3642 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3644 pwr_move
+= busiest
->__cpu_power
*
3645 min(busiest_load_per_task
, max_load
- tmp
);
3647 /* Amount of load we'd add */
3648 if (max_load
* busiest
->__cpu_power
<
3649 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3650 tmp
= sg_div_cpu_power(this,
3651 max_load
* busiest
->__cpu_power
);
3653 tmp
= sg_div_cpu_power(this,
3654 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3655 pwr_move
+= this->__cpu_power
*
3656 min(this_load_per_task
, this_load
+ tmp
);
3657 pwr_move
/= SCHED_LOAD_SCALE
;
3659 /* Move if we gain throughput */
3660 if (pwr_move
> pwr_now
)
3661 *imbalance
= busiest_load_per_task
;
3667 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3668 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3671 if (this == group_leader
&& group_leader
!= group_min
) {
3672 *imbalance
= min_load_per_task
;
3682 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3685 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3686 unsigned long imbalance
, const cpumask_t
*cpus
)
3688 struct rq
*busiest
= NULL
, *rq
;
3689 unsigned long max_load
= 0;
3692 for_each_cpu_mask(i
, group
->cpumask
) {
3695 if (!cpu_isset(i
, *cpus
))
3699 wl
= weighted_cpuload(i
);
3701 if (rq
->nr_running
== 1 && wl
> imbalance
)
3704 if (wl
> max_load
) {
3714 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3715 * so long as it is large enough.
3717 #define MAX_PINNED_INTERVAL 512
3720 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3721 * tasks if there is an imbalance.
3723 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3724 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3725 int *balance
, cpumask_t
*cpus
)
3727 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3728 struct sched_group
*group
;
3729 unsigned long imbalance
;
3731 unsigned long flags
;
3732 int unlock_aggregate
;
3736 unlock_aggregate
= get_aggregate(sd
);
3739 * When power savings policy is enabled for the parent domain, idle
3740 * sibling can pick up load irrespective of busy siblings. In this case,
3741 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3742 * portraying it as CPU_NOT_IDLE.
3744 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3745 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3748 schedstat_inc(sd
, lb_count
[idle
]);
3751 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3758 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3762 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3764 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3768 BUG_ON(busiest
== this_rq
);
3770 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3773 if (busiest
->nr_running
> 1) {
3775 * Attempt to move tasks. If find_busiest_group has found
3776 * an imbalance but busiest->nr_running <= 1, the group is
3777 * still unbalanced. ld_moved simply stays zero, so it is
3778 * correctly treated as an imbalance.
3780 local_irq_save(flags
);
3781 double_rq_lock(this_rq
, busiest
);
3782 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3783 imbalance
, sd
, idle
, &all_pinned
);
3784 double_rq_unlock(this_rq
, busiest
);
3785 local_irq_restore(flags
);
3788 * some other cpu did the load balance for us.
3790 if (ld_moved
&& this_cpu
!= smp_processor_id())
3791 resched_cpu(this_cpu
);
3793 /* All tasks on this runqueue were pinned by CPU affinity */
3794 if (unlikely(all_pinned
)) {
3795 cpu_clear(cpu_of(busiest
), *cpus
);
3796 if (!cpus_empty(*cpus
))
3803 schedstat_inc(sd
, lb_failed
[idle
]);
3804 sd
->nr_balance_failed
++;
3806 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3808 spin_lock_irqsave(&busiest
->lock
, flags
);
3810 /* don't kick the migration_thread, if the curr
3811 * task on busiest cpu can't be moved to this_cpu
3813 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3814 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3816 goto out_one_pinned
;
3819 if (!busiest
->active_balance
) {
3820 busiest
->active_balance
= 1;
3821 busiest
->push_cpu
= this_cpu
;
3824 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3826 wake_up_process(busiest
->migration_thread
);
3829 * We've kicked active balancing, reset the failure
3832 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3835 sd
->nr_balance_failed
= 0;
3837 if (likely(!active_balance
)) {
3838 /* We were unbalanced, so reset the balancing interval */
3839 sd
->balance_interval
= sd
->min_interval
;
3842 * If we've begun active balancing, start to back off. This
3843 * case may not be covered by the all_pinned logic if there
3844 * is only 1 task on the busy runqueue (because we don't call
3847 if (sd
->balance_interval
< sd
->max_interval
)
3848 sd
->balance_interval
*= 2;
3851 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3852 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3858 schedstat_inc(sd
, lb_balanced
[idle
]);
3860 sd
->nr_balance_failed
= 0;
3863 /* tune up the balancing interval */
3864 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3865 (sd
->balance_interval
< sd
->max_interval
))
3866 sd
->balance_interval
*= 2;
3868 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3869 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3874 if (unlock_aggregate
)
3880 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3881 * tasks if there is an imbalance.
3883 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3884 * this_rq is locked.
3887 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3890 struct sched_group
*group
;
3891 struct rq
*busiest
= NULL
;
3892 unsigned long imbalance
;
3900 * When power savings policy is enabled for the parent domain, idle
3901 * sibling can pick up load irrespective of busy siblings. In this case,
3902 * let the state of idle sibling percolate up as IDLE, instead of
3903 * portraying it as CPU_NOT_IDLE.
3905 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3906 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3909 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3911 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3912 &sd_idle
, cpus
, NULL
);
3914 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3918 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3920 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3924 BUG_ON(busiest
== this_rq
);
3926 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3929 if (busiest
->nr_running
> 1) {
3930 /* Attempt to move tasks */
3931 double_lock_balance(this_rq
, busiest
);
3932 /* this_rq->clock is already updated */
3933 update_rq_clock(busiest
);
3934 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3935 imbalance
, sd
, CPU_NEWLY_IDLE
,
3937 spin_unlock(&busiest
->lock
);
3939 if (unlikely(all_pinned
)) {
3940 cpu_clear(cpu_of(busiest
), *cpus
);
3941 if (!cpus_empty(*cpus
))
3947 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3948 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3949 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3952 sd
->nr_balance_failed
= 0;
3957 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3958 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3959 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3961 sd
->nr_balance_failed
= 0;
3967 * idle_balance is called by schedule() if this_cpu is about to become
3968 * idle. Attempts to pull tasks from other CPUs.
3970 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3972 struct sched_domain
*sd
;
3973 int pulled_task
= -1;
3974 unsigned long next_balance
= jiffies
+ HZ
;
3977 for_each_domain(this_cpu
, sd
) {
3978 unsigned long interval
;
3980 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3983 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3984 /* If we've pulled tasks over stop searching: */
3985 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3988 interval
= msecs_to_jiffies(sd
->balance_interval
);
3989 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3990 next_balance
= sd
->last_balance
+ interval
;
3994 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3996 * We are going idle. next_balance may be set based on
3997 * a busy processor. So reset next_balance.
3999 this_rq
->next_balance
= next_balance
;
4004 * active_load_balance is run by migration threads. It pushes running tasks
4005 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4006 * running on each physical CPU where possible, and avoids physical /
4007 * logical imbalances.
4009 * Called with busiest_rq locked.
4011 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4013 int target_cpu
= busiest_rq
->push_cpu
;
4014 struct sched_domain
*sd
;
4015 struct rq
*target_rq
;
4017 /* Is there any task to move? */
4018 if (busiest_rq
->nr_running
<= 1)
4021 target_rq
= cpu_rq(target_cpu
);
4024 * This condition is "impossible", if it occurs
4025 * we need to fix it. Originally reported by
4026 * Bjorn Helgaas on a 128-cpu setup.
4028 BUG_ON(busiest_rq
== target_rq
);
4030 /* move a task from busiest_rq to target_rq */
4031 double_lock_balance(busiest_rq
, target_rq
);
4032 update_rq_clock(busiest_rq
);
4033 update_rq_clock(target_rq
);
4035 /* Search for an sd spanning us and the target CPU. */
4036 for_each_domain(target_cpu
, sd
) {
4037 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4038 cpu_isset(busiest_cpu
, sd
->span
))
4043 schedstat_inc(sd
, alb_count
);
4045 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4047 schedstat_inc(sd
, alb_pushed
);
4049 schedstat_inc(sd
, alb_failed
);
4051 spin_unlock(&target_rq
->lock
);
4056 atomic_t load_balancer
;
4058 } nohz ____cacheline_aligned
= {
4059 .load_balancer
= ATOMIC_INIT(-1),
4060 .cpu_mask
= CPU_MASK_NONE
,
4064 * This routine will try to nominate the ilb (idle load balancing)
4065 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4066 * load balancing on behalf of all those cpus. If all the cpus in the system
4067 * go into this tickless mode, then there will be no ilb owner (as there is
4068 * no need for one) and all the cpus will sleep till the next wakeup event
4071 * For the ilb owner, tick is not stopped. And this tick will be used
4072 * for idle load balancing. ilb owner will still be part of
4075 * While stopping the tick, this cpu will become the ilb owner if there
4076 * is no other owner. And will be the owner till that cpu becomes busy
4077 * or if all cpus in the system stop their ticks at which point
4078 * there is no need for ilb owner.
4080 * When the ilb owner becomes busy, it nominates another owner, during the
4081 * next busy scheduler_tick()
4083 int select_nohz_load_balancer(int stop_tick
)
4085 int cpu
= smp_processor_id();
4088 cpu_set(cpu
, nohz
.cpu_mask
);
4089 cpu_rq(cpu
)->in_nohz_recently
= 1;
4092 * If we are going offline and still the leader, give up!
4094 if (cpu_is_offline(cpu
) &&
4095 atomic_read(&nohz
.load_balancer
) == cpu
) {
4096 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4101 /* time for ilb owner also to sleep */
4102 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4103 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4104 atomic_set(&nohz
.load_balancer
, -1);
4108 if (atomic_read(&nohz
.load_balancer
) == -1) {
4109 /* make me the ilb owner */
4110 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4112 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
4115 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
4118 cpu_clear(cpu
, nohz
.cpu_mask
);
4120 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4121 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4128 static DEFINE_SPINLOCK(balancing
);
4131 * It checks each scheduling domain to see if it is due to be balanced,
4132 * and initiates a balancing operation if so.
4134 * Balancing parameters are set up in arch_init_sched_domains.
4136 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4139 struct rq
*rq
= cpu_rq(cpu
);
4140 unsigned long interval
;
4141 struct sched_domain
*sd
;
4142 /* Earliest time when we have to do rebalance again */
4143 unsigned long next_balance
= jiffies
+ 60*HZ
;
4144 int update_next_balance
= 0;
4147 for_each_domain(cpu
, sd
) {
4148 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4151 interval
= sd
->balance_interval
;
4152 if (idle
!= CPU_IDLE
)
4153 interval
*= sd
->busy_factor
;
4155 /* scale ms to jiffies */
4156 interval
= msecs_to_jiffies(interval
);
4157 if (unlikely(!interval
))
4159 if (interval
> HZ
*NR_CPUS
/10)
4160 interval
= HZ
*NR_CPUS
/10;
4163 if (sd
->flags
& SD_SERIALIZE
) {
4164 if (!spin_trylock(&balancing
))
4168 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4169 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
4171 * We've pulled tasks over so either we're no
4172 * longer idle, or one of our SMT siblings is
4175 idle
= CPU_NOT_IDLE
;
4177 sd
->last_balance
= jiffies
;
4179 if (sd
->flags
& SD_SERIALIZE
)
4180 spin_unlock(&balancing
);
4182 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4183 next_balance
= sd
->last_balance
+ interval
;
4184 update_next_balance
= 1;
4188 * Stop the load balance at this level. There is another
4189 * CPU in our sched group which is doing load balancing more
4197 * next_balance will be updated only when there is a need.
4198 * When the cpu is attached to null domain for ex, it will not be
4201 if (likely(update_next_balance
))
4202 rq
->next_balance
= next_balance
;
4206 * run_rebalance_domains is triggered when needed from the scheduler tick.
4207 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4208 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4210 static void run_rebalance_domains(struct softirq_action
*h
)
4212 int this_cpu
= smp_processor_id();
4213 struct rq
*this_rq
= cpu_rq(this_cpu
);
4214 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4215 CPU_IDLE
: CPU_NOT_IDLE
;
4217 rebalance_domains(this_cpu
, idle
);
4221 * If this cpu is the owner for idle load balancing, then do the
4222 * balancing on behalf of the other idle cpus whose ticks are
4225 if (this_rq
->idle_at_tick
&&
4226 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4227 cpumask_t cpus
= nohz
.cpu_mask
;
4231 cpu_clear(this_cpu
, cpus
);
4232 for_each_cpu_mask(balance_cpu
, cpus
) {
4234 * If this cpu gets work to do, stop the load balancing
4235 * work being done for other cpus. Next load
4236 * balancing owner will pick it up.
4241 rebalance_domains(balance_cpu
, CPU_IDLE
);
4243 rq
= cpu_rq(balance_cpu
);
4244 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4245 this_rq
->next_balance
= rq
->next_balance
;
4252 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4254 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4255 * idle load balancing owner or decide to stop the periodic load balancing,
4256 * if the whole system is idle.
4258 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4262 * If we were in the nohz mode recently and busy at the current
4263 * scheduler tick, then check if we need to nominate new idle
4266 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4267 rq
->in_nohz_recently
= 0;
4269 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4270 cpu_clear(cpu
, nohz
.cpu_mask
);
4271 atomic_set(&nohz
.load_balancer
, -1);
4274 if (atomic_read(&nohz
.load_balancer
) == -1) {
4276 * simple selection for now: Nominate the
4277 * first cpu in the nohz list to be the next
4280 * TBD: Traverse the sched domains and nominate
4281 * the nearest cpu in the nohz.cpu_mask.
4283 int ilb
= first_cpu(nohz
.cpu_mask
);
4285 if (ilb
< nr_cpu_ids
)
4291 * If this cpu is idle and doing idle load balancing for all the
4292 * cpus with ticks stopped, is it time for that to stop?
4294 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4295 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4301 * If this cpu is idle and the idle load balancing is done by
4302 * someone else, then no need raise the SCHED_SOFTIRQ
4304 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4305 cpu_isset(cpu
, nohz
.cpu_mask
))
4308 if (time_after_eq(jiffies
, rq
->next_balance
))
4309 raise_softirq(SCHED_SOFTIRQ
);
4312 #else /* CONFIG_SMP */
4315 * on UP we do not need to balance between CPUs:
4317 static inline void idle_balance(int cpu
, struct rq
*rq
)
4323 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4325 EXPORT_PER_CPU_SYMBOL(kstat
);
4328 * Return p->sum_exec_runtime plus any more ns on the sched_clock
4329 * that have not yet been banked in case the task is currently running.
4331 unsigned long long task_sched_runtime(struct task_struct
*p
)
4333 unsigned long flags
;
4337 rq
= task_rq_lock(p
, &flags
);
4338 ns
= p
->se
.sum_exec_runtime
;
4339 if (task_current(rq
, p
)) {
4340 update_rq_clock(rq
);
4341 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4342 if ((s64
)delta_exec
> 0)
4345 task_rq_unlock(rq
, &flags
);
4351 * Account user cpu time to a process.
4352 * @p: the process that the cpu time gets accounted to
4353 * @cputime: the cpu time spent in user space since the last update
4355 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4357 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4360 p
->utime
= cputime_add(p
->utime
, cputime
);
4362 /* Add user time to cpustat. */
4363 tmp
= cputime_to_cputime64(cputime
);
4364 if (TASK_NICE(p
) > 0)
4365 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4367 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4371 * Account guest cpu time to a process.
4372 * @p: the process that the cpu time gets accounted to
4373 * @cputime: the cpu time spent in virtual machine since the last update
4375 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4378 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4380 tmp
= cputime_to_cputime64(cputime
);
4382 p
->utime
= cputime_add(p
->utime
, cputime
);
4383 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4385 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4386 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4390 * Account scaled user cpu time to a process.
4391 * @p: the process that the cpu time gets accounted to
4392 * @cputime: the cpu time spent in user space since the last update
4394 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4396 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4400 * Account system cpu time to a process.
4401 * @p: the process that the cpu time gets accounted to
4402 * @hardirq_offset: the offset to subtract from hardirq_count()
4403 * @cputime: the cpu time spent in kernel space since the last update
4405 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4408 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4409 struct rq
*rq
= this_rq();
4412 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4413 account_guest_time(p
, cputime
);
4417 p
->stime
= cputime_add(p
->stime
, cputime
);
4419 /* Add system time to cpustat. */
4420 tmp
= cputime_to_cputime64(cputime
);
4421 if (hardirq_count() - hardirq_offset
)
4422 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4423 else if (softirq_count())
4424 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4425 else if (p
!= rq
->idle
)
4426 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4427 else if (atomic_read(&rq
->nr_iowait
) > 0)
4428 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4430 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4431 /* Account for system time used */
4432 acct_update_integrals(p
);
4436 * Account scaled system cpu time to a process.
4437 * @p: the process that the cpu time gets accounted to
4438 * @hardirq_offset: the offset to subtract from hardirq_count()
4439 * @cputime: the cpu time spent in kernel space since the last update
4441 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4443 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4447 * Account for involuntary wait time.
4448 * @p: the process from which the cpu time has been stolen
4449 * @steal: the cpu time spent in involuntary wait
4451 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4453 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4454 cputime64_t tmp
= cputime_to_cputime64(steal
);
4455 struct rq
*rq
= this_rq();
4457 if (p
== rq
->idle
) {
4458 p
->stime
= cputime_add(p
->stime
, steal
);
4459 if (atomic_read(&rq
->nr_iowait
) > 0)
4460 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4462 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4464 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4468 * This function gets called by the timer code, with HZ frequency.
4469 * We call it with interrupts disabled.
4471 * It also gets called by the fork code, when changing the parent's
4474 void scheduler_tick(void)
4476 int cpu
= smp_processor_id();
4477 struct rq
*rq
= cpu_rq(cpu
);
4478 struct task_struct
*curr
= rq
->curr
;
4479 u64 next_tick
= rq
->tick_timestamp
+ TICK_NSEC
;
4481 spin_lock(&rq
->lock
);
4482 __update_rq_clock(rq
);
4484 * Let rq->clock advance by at least TICK_NSEC:
4486 if (unlikely(rq
->clock
< next_tick
)) {
4487 rq
->clock
= next_tick
;
4488 rq
->clock_underflows
++;
4490 rq
->tick_timestamp
= rq
->clock
;
4491 update_last_tick_seen(rq
);
4492 update_cpu_load(rq
);
4493 curr
->sched_class
->task_tick(rq
, curr
, 0);
4494 spin_unlock(&rq
->lock
);
4497 rq
->idle_at_tick
= idle_cpu(cpu
);
4498 trigger_load_balance(rq
, cpu
);
4502 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
4504 void __kprobes
add_preempt_count(int val
)
4509 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4511 preempt_count() += val
;
4513 * Spinlock count overflowing soon?
4515 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4518 EXPORT_SYMBOL(add_preempt_count
);
4520 void __kprobes
sub_preempt_count(int val
)
4525 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4528 * Is the spinlock portion underflowing?
4530 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4531 !(preempt_count() & PREEMPT_MASK
)))
4534 preempt_count() -= val
;
4536 EXPORT_SYMBOL(sub_preempt_count
);
4541 * Print scheduling while atomic bug:
4543 static noinline
void __schedule_bug(struct task_struct
*prev
)
4545 struct pt_regs
*regs
= get_irq_regs();
4547 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4548 prev
->comm
, prev
->pid
, preempt_count());
4550 debug_show_held_locks(prev
);
4551 if (irqs_disabled())
4552 print_irqtrace_events(prev
);
4561 * Various schedule()-time debugging checks and statistics:
4563 static inline void schedule_debug(struct task_struct
*prev
)
4566 * Test if we are atomic. Since do_exit() needs to call into
4567 * schedule() atomically, we ignore that path for now.
4568 * Otherwise, whine if we are scheduling when we should not be.
4570 if (unlikely(in_atomic_preempt_off()) && unlikely(!prev
->exit_state
))
4571 __schedule_bug(prev
);
4573 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4575 schedstat_inc(this_rq(), sched_count
);
4576 #ifdef CONFIG_SCHEDSTATS
4577 if (unlikely(prev
->lock_depth
>= 0)) {
4578 schedstat_inc(this_rq(), bkl_count
);
4579 schedstat_inc(prev
, sched_info
.bkl_count
);
4585 * Pick up the highest-prio task:
4587 static inline struct task_struct
*
4588 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4590 const struct sched_class
*class;
4591 struct task_struct
*p
;
4594 * Optimization: we know that if all tasks are in
4595 * the fair class we can call that function directly:
4597 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4598 p
= fair_sched_class
.pick_next_task(rq
);
4603 class = sched_class_highest
;
4605 p
= class->pick_next_task(rq
);
4609 * Will never be NULL as the idle class always
4610 * returns a non-NULL p:
4612 class = class->next
;
4617 * schedule() is the main scheduler function.
4619 asmlinkage
void __sched
schedule(void)
4621 struct task_struct
*prev
, *next
;
4622 unsigned long *switch_count
;
4628 cpu
= smp_processor_id();
4632 switch_count
= &prev
->nivcsw
;
4634 release_kernel_lock(prev
);
4635 need_resched_nonpreemptible
:
4637 schedule_debug(prev
);
4642 * Do the rq-clock update outside the rq lock:
4644 local_irq_disable();
4645 __update_rq_clock(rq
);
4646 spin_lock(&rq
->lock
);
4647 clear_tsk_need_resched(prev
);
4649 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4650 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
4651 signal_pending(prev
))) {
4652 prev
->state
= TASK_RUNNING
;
4654 deactivate_task(rq
, prev
, 1);
4656 switch_count
= &prev
->nvcsw
;
4660 if (prev
->sched_class
->pre_schedule
)
4661 prev
->sched_class
->pre_schedule(rq
, prev
);
4664 if (unlikely(!rq
->nr_running
))
4665 idle_balance(cpu
, rq
);
4667 prev
->sched_class
->put_prev_task(rq
, prev
);
4668 next
= pick_next_task(rq
, prev
);
4670 if (likely(prev
!= next
)) {
4671 sched_info_switch(prev
, next
);
4677 context_switch(rq
, prev
, next
); /* unlocks the rq */
4679 * the context switch might have flipped the stack from under
4680 * us, hence refresh the local variables.
4682 cpu
= smp_processor_id();
4685 spin_unlock_irq(&rq
->lock
);
4689 if (unlikely(reacquire_kernel_lock(current
) < 0))
4690 goto need_resched_nonpreemptible
;
4692 preempt_enable_no_resched();
4693 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4696 EXPORT_SYMBOL(schedule
);
4698 #ifdef CONFIG_PREEMPT
4700 * this is the entry point to schedule() from in-kernel preemption
4701 * off of preempt_enable. Kernel preemptions off return from interrupt
4702 * occur there and call schedule directly.
4704 asmlinkage
void __sched
preempt_schedule(void)
4706 struct thread_info
*ti
= current_thread_info();
4707 struct task_struct
*task
= current
;
4708 int saved_lock_depth
;
4711 * If there is a non-zero preempt_count or interrupts are disabled,
4712 * we do not want to preempt the current task. Just return..
4714 if (likely(ti
->preempt_count
|| irqs_disabled()))
4718 add_preempt_count(PREEMPT_ACTIVE
);
4721 * We keep the big kernel semaphore locked, but we
4722 * clear ->lock_depth so that schedule() doesnt
4723 * auto-release the semaphore:
4725 saved_lock_depth
= task
->lock_depth
;
4726 task
->lock_depth
= -1;
4728 task
->lock_depth
= saved_lock_depth
;
4729 sub_preempt_count(PREEMPT_ACTIVE
);
4732 * Check again in case we missed a preemption opportunity
4733 * between schedule and now.
4736 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4738 EXPORT_SYMBOL(preempt_schedule
);
4741 * this is the entry point to schedule() from kernel preemption
4742 * off of irq context.
4743 * Note, that this is called and return with irqs disabled. This will
4744 * protect us against recursive calling from irq.
4746 asmlinkage
void __sched
preempt_schedule_irq(void)
4748 struct thread_info
*ti
= current_thread_info();
4749 struct task_struct
*task
= current
;
4750 int saved_lock_depth
;
4752 /* Catch callers which need to be fixed */
4753 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4756 add_preempt_count(PREEMPT_ACTIVE
);
4759 * We keep the big kernel semaphore locked, but we
4760 * clear ->lock_depth so that schedule() doesnt
4761 * auto-release the semaphore:
4763 saved_lock_depth
= task
->lock_depth
;
4764 task
->lock_depth
= -1;
4767 local_irq_disable();
4768 task
->lock_depth
= saved_lock_depth
;
4769 sub_preempt_count(PREEMPT_ACTIVE
);
4772 * Check again in case we missed a preemption opportunity
4773 * between schedule and now.
4776 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4779 #endif /* CONFIG_PREEMPT */
4781 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4784 return try_to_wake_up(curr
->private, mode
, sync
);
4786 EXPORT_SYMBOL(default_wake_function
);
4789 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4790 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4791 * number) then we wake all the non-exclusive tasks and one exclusive task.
4793 * There are circumstances in which we can try to wake a task which has already
4794 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4795 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4797 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4798 int nr_exclusive
, int sync
, void *key
)
4800 wait_queue_t
*curr
, *next
;
4802 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4803 unsigned flags
= curr
->flags
;
4805 if (curr
->func(curr
, mode
, sync
, key
) &&
4806 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4812 * __wake_up - wake up threads blocked on a waitqueue.
4814 * @mode: which threads
4815 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4816 * @key: is directly passed to the wakeup function
4818 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4819 int nr_exclusive
, void *key
)
4821 unsigned long flags
;
4823 spin_lock_irqsave(&q
->lock
, flags
);
4824 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4825 spin_unlock_irqrestore(&q
->lock
, flags
);
4827 EXPORT_SYMBOL(__wake_up
);
4830 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4832 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4834 __wake_up_common(q
, mode
, 1, 0, NULL
);
4838 * __wake_up_sync - wake up threads blocked on a waitqueue.
4840 * @mode: which threads
4841 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4843 * The sync wakeup differs that the waker knows that it will schedule
4844 * away soon, so while the target thread will be woken up, it will not
4845 * be migrated to another CPU - ie. the two threads are 'synchronized'
4846 * with each other. This can prevent needless bouncing between CPUs.
4848 * On UP it can prevent extra preemption.
4851 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4853 unsigned long flags
;
4859 if (unlikely(!nr_exclusive
))
4862 spin_lock_irqsave(&q
->lock
, flags
);
4863 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4864 spin_unlock_irqrestore(&q
->lock
, flags
);
4866 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4868 void complete(struct completion
*x
)
4870 unsigned long flags
;
4872 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4874 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4875 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4877 EXPORT_SYMBOL(complete
);
4879 void complete_all(struct completion
*x
)
4881 unsigned long flags
;
4883 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4884 x
->done
+= UINT_MAX
/2;
4885 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4886 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4888 EXPORT_SYMBOL(complete_all
);
4890 static inline long __sched
4891 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4894 DECLARE_WAITQUEUE(wait
, current
);
4896 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4897 __add_wait_queue_tail(&x
->wait
, &wait
);
4899 if ((state
== TASK_INTERRUPTIBLE
&&
4900 signal_pending(current
)) ||
4901 (state
== TASK_KILLABLE
&&
4902 fatal_signal_pending(current
))) {
4903 __remove_wait_queue(&x
->wait
, &wait
);
4904 return -ERESTARTSYS
;
4906 __set_current_state(state
);
4907 spin_unlock_irq(&x
->wait
.lock
);
4908 timeout
= schedule_timeout(timeout
);
4909 spin_lock_irq(&x
->wait
.lock
);
4911 __remove_wait_queue(&x
->wait
, &wait
);
4915 __remove_wait_queue(&x
->wait
, &wait
);
4922 wait_for_common(struct completion
*x
, long timeout
, int state
)
4926 spin_lock_irq(&x
->wait
.lock
);
4927 timeout
= do_wait_for_common(x
, timeout
, state
);
4928 spin_unlock_irq(&x
->wait
.lock
);
4932 void __sched
wait_for_completion(struct completion
*x
)
4934 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4936 EXPORT_SYMBOL(wait_for_completion
);
4938 unsigned long __sched
4939 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4941 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4943 EXPORT_SYMBOL(wait_for_completion_timeout
);
4945 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4947 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4948 if (t
== -ERESTARTSYS
)
4952 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4954 unsigned long __sched
4955 wait_for_completion_interruptible_timeout(struct completion
*x
,
4956 unsigned long timeout
)
4958 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4960 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4962 int __sched
wait_for_completion_killable(struct completion
*x
)
4964 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4965 if (t
== -ERESTARTSYS
)
4969 EXPORT_SYMBOL(wait_for_completion_killable
);
4972 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4974 unsigned long flags
;
4977 init_waitqueue_entry(&wait
, current
);
4979 __set_current_state(state
);
4981 spin_lock_irqsave(&q
->lock
, flags
);
4982 __add_wait_queue(q
, &wait
);
4983 spin_unlock(&q
->lock
);
4984 timeout
= schedule_timeout(timeout
);
4985 spin_lock_irq(&q
->lock
);
4986 __remove_wait_queue(q
, &wait
);
4987 spin_unlock_irqrestore(&q
->lock
, flags
);
4992 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4994 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4996 EXPORT_SYMBOL(interruptible_sleep_on
);
4999 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5001 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
5003 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
5005 void __sched
sleep_on(wait_queue_head_t
*q
)
5007 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
5009 EXPORT_SYMBOL(sleep_on
);
5011 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
5013 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
5015 EXPORT_SYMBOL(sleep_on_timeout
);
5017 #ifdef CONFIG_RT_MUTEXES
5020 * rt_mutex_setprio - set the current priority of a task
5022 * @prio: prio value (kernel-internal form)
5024 * This function changes the 'effective' priority of a task. It does
5025 * not touch ->normal_prio like __setscheduler().
5027 * Used by the rt_mutex code to implement priority inheritance logic.
5029 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
5031 unsigned long flags
;
5032 int oldprio
, on_rq
, running
;
5034 const struct sched_class
*prev_class
= p
->sched_class
;
5036 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
5038 rq
= task_rq_lock(p
, &flags
);
5039 update_rq_clock(rq
);
5042 on_rq
= p
->se
.on_rq
;
5043 running
= task_current(rq
, p
);
5045 dequeue_task(rq
, p
, 0);
5047 p
->sched_class
->put_prev_task(rq
, p
);
5050 p
->sched_class
= &rt_sched_class
;
5052 p
->sched_class
= &fair_sched_class
;
5057 p
->sched_class
->set_curr_task(rq
);
5059 enqueue_task(rq
, p
, 0);
5061 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5063 task_rq_unlock(rq
, &flags
);
5068 void set_user_nice(struct task_struct
*p
, long nice
)
5070 int old_prio
, delta
, on_rq
;
5071 unsigned long flags
;
5074 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
5077 * We have to be careful, if called from sys_setpriority(),
5078 * the task might be in the middle of scheduling on another CPU.
5080 rq
= task_rq_lock(p
, &flags
);
5081 update_rq_clock(rq
);
5083 * The RT priorities are set via sched_setscheduler(), but we still
5084 * allow the 'normal' nice value to be set - but as expected
5085 * it wont have any effect on scheduling until the task is
5086 * SCHED_FIFO/SCHED_RR:
5088 if (task_has_rt_policy(p
)) {
5089 p
->static_prio
= NICE_TO_PRIO(nice
);
5092 on_rq
= p
->se
.on_rq
;
5094 dequeue_task(rq
, p
, 0);
5096 p
->static_prio
= NICE_TO_PRIO(nice
);
5099 p
->prio
= effective_prio(p
);
5100 delta
= p
->prio
- old_prio
;
5103 enqueue_task(rq
, p
, 0);
5105 * If the task increased its priority or is running and
5106 * lowered its priority, then reschedule its CPU:
5108 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5109 resched_task(rq
->curr
);
5112 task_rq_unlock(rq
, &flags
);
5114 EXPORT_SYMBOL(set_user_nice
);
5117 * can_nice - check if a task can reduce its nice value
5121 int can_nice(const struct task_struct
*p
, const int nice
)
5123 /* convert nice value [19,-20] to rlimit style value [1,40] */
5124 int nice_rlim
= 20 - nice
;
5126 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5127 capable(CAP_SYS_NICE
));
5130 #ifdef __ARCH_WANT_SYS_NICE
5133 * sys_nice - change the priority of the current process.
5134 * @increment: priority increment
5136 * sys_setpriority is a more generic, but much slower function that
5137 * does similar things.
5139 asmlinkage
long sys_nice(int increment
)
5144 * Setpriority might change our priority at the same moment.
5145 * We don't have to worry. Conceptually one call occurs first
5146 * and we have a single winner.
5148 if (increment
< -40)
5153 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5159 if (increment
< 0 && !can_nice(current
, nice
))
5162 retval
= security_task_setnice(current
, nice
);
5166 set_user_nice(current
, nice
);
5173 * task_prio - return the priority value of a given task.
5174 * @p: the task in question.
5176 * This is the priority value as seen by users in /proc.
5177 * RT tasks are offset by -200. Normal tasks are centered
5178 * around 0, value goes from -16 to +15.
5180 int task_prio(const struct task_struct
*p
)
5182 return p
->prio
- MAX_RT_PRIO
;
5186 * task_nice - return the nice value of a given task.
5187 * @p: the task in question.
5189 int task_nice(const struct task_struct
*p
)
5191 return TASK_NICE(p
);
5193 EXPORT_SYMBOL(task_nice
);
5196 * idle_cpu - is a given cpu idle currently?
5197 * @cpu: the processor in question.
5199 int idle_cpu(int cpu
)
5201 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5205 * idle_task - return the idle task for a given cpu.
5206 * @cpu: the processor in question.
5208 struct task_struct
*idle_task(int cpu
)
5210 return cpu_rq(cpu
)->idle
;
5214 * find_process_by_pid - find a process with a matching PID value.
5215 * @pid: the pid in question.
5217 static struct task_struct
*find_process_by_pid(pid_t pid
)
5219 return pid
? find_task_by_vpid(pid
) : current
;
5222 /* Actually do priority change: must hold rq lock. */
5224 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5226 BUG_ON(p
->se
.on_rq
);
5229 switch (p
->policy
) {
5233 p
->sched_class
= &fair_sched_class
;
5237 p
->sched_class
= &rt_sched_class
;
5241 p
->rt_priority
= prio
;
5242 p
->normal_prio
= normal_prio(p
);
5243 /* we are holding p->pi_lock already */
5244 p
->prio
= rt_mutex_getprio(p
);
5249 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5250 * @p: the task in question.
5251 * @policy: new policy.
5252 * @param: structure containing the new RT priority.
5254 * NOTE that the task may be already dead.
5256 int sched_setscheduler(struct task_struct
*p
, int policy
,
5257 struct sched_param
*param
)
5259 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5260 unsigned long flags
;
5261 const struct sched_class
*prev_class
= p
->sched_class
;
5264 /* may grab non-irq protected spin_locks */
5265 BUG_ON(in_interrupt());
5267 /* double check policy once rq lock held */
5269 policy
= oldpolicy
= p
->policy
;
5270 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5271 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5272 policy
!= SCHED_IDLE
)
5275 * Valid priorities for SCHED_FIFO and SCHED_RR are
5276 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5277 * SCHED_BATCH and SCHED_IDLE is 0.
5279 if (param
->sched_priority
< 0 ||
5280 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5281 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5283 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5287 * Allow unprivileged RT tasks to decrease priority:
5289 if (!capable(CAP_SYS_NICE
)) {
5290 if (rt_policy(policy
)) {
5291 unsigned long rlim_rtprio
;
5293 if (!lock_task_sighand(p
, &flags
))
5295 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5296 unlock_task_sighand(p
, &flags
);
5298 /* can't set/change the rt policy */
5299 if (policy
!= p
->policy
&& !rlim_rtprio
)
5302 /* can't increase priority */
5303 if (param
->sched_priority
> p
->rt_priority
&&
5304 param
->sched_priority
> rlim_rtprio
)
5308 * Like positive nice levels, dont allow tasks to
5309 * move out of SCHED_IDLE either:
5311 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5314 /* can't change other user's priorities */
5315 if ((current
->euid
!= p
->euid
) &&
5316 (current
->euid
!= p
->uid
))
5320 #ifdef CONFIG_RT_GROUP_SCHED
5322 * Do not allow realtime tasks into groups that have no runtime
5325 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5329 retval
= security_task_setscheduler(p
, policy
, param
);
5333 * make sure no PI-waiters arrive (or leave) while we are
5334 * changing the priority of the task:
5336 spin_lock_irqsave(&p
->pi_lock
, flags
);
5338 * To be able to change p->policy safely, the apropriate
5339 * runqueue lock must be held.
5341 rq
= __task_rq_lock(p
);
5342 /* recheck policy now with rq lock held */
5343 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5344 policy
= oldpolicy
= -1;
5345 __task_rq_unlock(rq
);
5346 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5349 update_rq_clock(rq
);
5350 on_rq
= p
->se
.on_rq
;
5351 running
= task_current(rq
, p
);
5353 deactivate_task(rq
, p
, 0);
5355 p
->sched_class
->put_prev_task(rq
, p
);
5358 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5361 p
->sched_class
->set_curr_task(rq
);
5363 activate_task(rq
, p
, 0);
5365 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5367 __task_rq_unlock(rq
);
5368 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5370 rt_mutex_adjust_pi(p
);
5374 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5377 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5379 struct sched_param lparam
;
5380 struct task_struct
*p
;
5383 if (!param
|| pid
< 0)
5385 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5390 p
= find_process_by_pid(pid
);
5392 retval
= sched_setscheduler(p
, policy
, &lparam
);
5399 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5400 * @pid: the pid in question.
5401 * @policy: new policy.
5402 * @param: structure containing the new RT priority.
5405 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5407 /* negative values for policy are not valid */
5411 return do_sched_setscheduler(pid
, policy
, param
);
5415 * sys_sched_setparam - set/change the RT priority of a thread
5416 * @pid: the pid in question.
5417 * @param: structure containing the new RT priority.
5419 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5421 return do_sched_setscheduler(pid
, -1, param
);
5425 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5426 * @pid: the pid in question.
5428 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5430 struct task_struct
*p
;
5437 read_lock(&tasklist_lock
);
5438 p
= find_process_by_pid(pid
);
5440 retval
= security_task_getscheduler(p
);
5444 read_unlock(&tasklist_lock
);
5449 * sys_sched_getscheduler - get the RT priority of a thread
5450 * @pid: the pid in question.
5451 * @param: structure containing the RT priority.
5453 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5455 struct sched_param lp
;
5456 struct task_struct
*p
;
5459 if (!param
|| pid
< 0)
5462 read_lock(&tasklist_lock
);
5463 p
= find_process_by_pid(pid
);
5468 retval
= security_task_getscheduler(p
);
5472 lp
.sched_priority
= p
->rt_priority
;
5473 read_unlock(&tasklist_lock
);
5476 * This one might sleep, we cannot do it with a spinlock held ...
5478 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5483 read_unlock(&tasklist_lock
);
5487 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5489 cpumask_t cpus_allowed
;
5490 cpumask_t new_mask
= *in_mask
;
5491 struct task_struct
*p
;
5495 read_lock(&tasklist_lock
);
5497 p
= find_process_by_pid(pid
);
5499 read_unlock(&tasklist_lock
);
5505 * It is not safe to call set_cpus_allowed with the
5506 * tasklist_lock held. We will bump the task_struct's
5507 * usage count and then drop tasklist_lock.
5510 read_unlock(&tasklist_lock
);
5513 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5514 !capable(CAP_SYS_NICE
))
5517 retval
= security_task_setscheduler(p
, 0, NULL
);
5521 cpuset_cpus_allowed(p
, &cpus_allowed
);
5522 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5524 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5527 cpuset_cpus_allowed(p
, &cpus_allowed
);
5528 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5530 * We must have raced with a concurrent cpuset
5531 * update. Just reset the cpus_allowed to the
5532 * cpuset's cpus_allowed
5534 new_mask
= cpus_allowed
;
5544 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5545 cpumask_t
*new_mask
)
5547 if (len
< sizeof(cpumask_t
)) {
5548 memset(new_mask
, 0, sizeof(cpumask_t
));
5549 } else if (len
> sizeof(cpumask_t
)) {
5550 len
= sizeof(cpumask_t
);
5552 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5556 * sys_sched_setaffinity - set the cpu affinity of a process
5557 * @pid: pid of the process
5558 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5559 * @user_mask_ptr: user-space pointer to the new cpu mask
5561 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5562 unsigned long __user
*user_mask_ptr
)
5567 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5571 return sched_setaffinity(pid
, &new_mask
);
5575 * Represents all cpu's present in the system
5576 * In systems capable of hotplug, this map could dynamically grow
5577 * as new cpu's are detected in the system via any platform specific
5578 * method, such as ACPI for e.g.
5581 cpumask_t cpu_present_map __read_mostly
;
5582 EXPORT_SYMBOL(cpu_present_map
);
5585 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
5586 EXPORT_SYMBOL(cpu_online_map
);
5588 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
5589 EXPORT_SYMBOL(cpu_possible_map
);
5592 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5594 struct task_struct
*p
;
5598 read_lock(&tasklist_lock
);
5601 p
= find_process_by_pid(pid
);
5605 retval
= security_task_getscheduler(p
);
5609 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5612 read_unlock(&tasklist_lock
);
5619 * sys_sched_getaffinity - get the cpu affinity of a process
5620 * @pid: pid of the process
5621 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5622 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5624 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5625 unsigned long __user
*user_mask_ptr
)
5630 if (len
< sizeof(cpumask_t
))
5633 ret
= sched_getaffinity(pid
, &mask
);
5637 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5640 return sizeof(cpumask_t
);
5644 * sys_sched_yield - yield the current processor to other threads.
5646 * This function yields the current CPU to other tasks. If there are no
5647 * other threads running on this CPU then this function will return.
5649 asmlinkage
long sys_sched_yield(void)
5651 struct rq
*rq
= this_rq_lock();
5653 schedstat_inc(rq
, yld_count
);
5654 current
->sched_class
->yield_task(rq
);
5657 * Since we are going to call schedule() anyway, there's
5658 * no need to preempt or enable interrupts:
5660 __release(rq
->lock
);
5661 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5662 _raw_spin_unlock(&rq
->lock
);
5663 preempt_enable_no_resched();
5670 static void __cond_resched(void)
5672 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5673 __might_sleep(__FILE__
, __LINE__
);
5676 * The BKS might be reacquired before we have dropped
5677 * PREEMPT_ACTIVE, which could trigger a second
5678 * cond_resched() call.
5681 add_preempt_count(PREEMPT_ACTIVE
);
5683 sub_preempt_count(PREEMPT_ACTIVE
);
5684 } while (need_resched());
5687 #if !defined(CONFIG_PREEMPT) || defined(CONFIG_PREEMPT_VOLUNTARY)
5688 int __sched
_cond_resched(void)
5690 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5691 system_state
== SYSTEM_RUNNING
) {
5697 EXPORT_SYMBOL(_cond_resched
);
5701 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5702 * call schedule, and on return reacquire the lock.
5704 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5705 * operations here to prevent schedule() from being called twice (once via
5706 * spin_unlock(), once by hand).
5708 int cond_resched_lock(spinlock_t
*lock
)
5710 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5713 if (spin_needbreak(lock
) || resched
) {
5715 if (resched
&& need_resched())
5724 EXPORT_SYMBOL(cond_resched_lock
);
5726 int __sched
cond_resched_softirq(void)
5728 BUG_ON(!in_softirq());
5730 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5738 EXPORT_SYMBOL(cond_resched_softirq
);
5741 * yield - yield the current processor to other threads.
5743 * This is a shortcut for kernel-space yielding - it marks the
5744 * thread runnable and calls sys_sched_yield().
5746 void __sched
yield(void)
5748 set_current_state(TASK_RUNNING
);
5751 EXPORT_SYMBOL(yield
);
5754 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5755 * that process accounting knows that this is a task in IO wait state.
5757 * But don't do that if it is a deliberate, throttling IO wait (this task
5758 * has set its backing_dev_info: the queue against which it should throttle)
5760 void __sched
io_schedule(void)
5762 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5764 delayacct_blkio_start();
5765 atomic_inc(&rq
->nr_iowait
);
5767 atomic_dec(&rq
->nr_iowait
);
5768 delayacct_blkio_end();
5770 EXPORT_SYMBOL(io_schedule
);
5772 long __sched
io_schedule_timeout(long timeout
)
5774 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5777 delayacct_blkio_start();
5778 atomic_inc(&rq
->nr_iowait
);
5779 ret
= schedule_timeout(timeout
);
5780 atomic_dec(&rq
->nr_iowait
);
5781 delayacct_blkio_end();
5786 * sys_sched_get_priority_max - return maximum RT priority.
5787 * @policy: scheduling class.
5789 * this syscall returns the maximum rt_priority that can be used
5790 * by a given scheduling class.
5792 asmlinkage
long sys_sched_get_priority_max(int policy
)
5799 ret
= MAX_USER_RT_PRIO
-1;
5811 * sys_sched_get_priority_min - return minimum RT priority.
5812 * @policy: scheduling class.
5814 * this syscall returns the minimum rt_priority that can be used
5815 * by a given scheduling class.
5817 asmlinkage
long sys_sched_get_priority_min(int policy
)
5835 * sys_sched_rr_get_interval - return the default timeslice of a process.
5836 * @pid: pid of the process.
5837 * @interval: userspace pointer to the timeslice value.
5839 * this syscall writes the default timeslice value of a given process
5840 * into the user-space timespec buffer. A value of '0' means infinity.
5843 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5845 struct task_struct
*p
;
5846 unsigned int time_slice
;
5854 read_lock(&tasklist_lock
);
5855 p
= find_process_by_pid(pid
);
5859 retval
= security_task_getscheduler(p
);
5864 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5865 * tasks that are on an otherwise idle runqueue:
5868 if (p
->policy
== SCHED_RR
) {
5869 time_slice
= DEF_TIMESLICE
;
5870 } else if (p
->policy
!= SCHED_FIFO
) {
5871 struct sched_entity
*se
= &p
->se
;
5872 unsigned long flags
;
5875 rq
= task_rq_lock(p
, &flags
);
5876 if (rq
->cfs
.load
.weight
)
5877 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5878 task_rq_unlock(rq
, &flags
);
5880 read_unlock(&tasklist_lock
);
5881 jiffies_to_timespec(time_slice
, &t
);
5882 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5886 read_unlock(&tasklist_lock
);
5890 static const char stat_nam
[] = "RSDTtZX";
5892 void sched_show_task(struct task_struct
*p
)
5894 unsigned long free
= 0;
5897 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5898 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5899 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5900 #if BITS_PER_LONG == 32
5901 if (state
== TASK_RUNNING
)
5902 printk(KERN_CONT
" running ");
5904 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5906 if (state
== TASK_RUNNING
)
5907 printk(KERN_CONT
" running task ");
5909 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5911 #ifdef CONFIG_DEBUG_STACK_USAGE
5913 unsigned long *n
= end_of_stack(p
);
5916 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5919 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5920 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5922 show_stack(p
, NULL
);
5925 void show_state_filter(unsigned long state_filter
)
5927 struct task_struct
*g
, *p
;
5929 #if BITS_PER_LONG == 32
5931 " task PC stack pid father\n");
5934 " task PC stack pid father\n");
5936 read_lock(&tasklist_lock
);
5937 do_each_thread(g
, p
) {
5939 * reset the NMI-timeout, listing all files on a slow
5940 * console might take alot of time:
5942 touch_nmi_watchdog();
5943 if (!state_filter
|| (p
->state
& state_filter
))
5945 } while_each_thread(g
, p
);
5947 touch_all_softlockup_watchdogs();
5949 #ifdef CONFIG_SCHED_DEBUG
5950 sysrq_sched_debug_show();
5952 read_unlock(&tasklist_lock
);
5954 * Only show locks if all tasks are dumped:
5956 if (state_filter
== -1)
5957 debug_show_all_locks();
5960 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5962 idle
->sched_class
= &idle_sched_class
;
5966 * init_idle - set up an idle thread for a given CPU
5967 * @idle: task in question
5968 * @cpu: cpu the idle task belongs to
5970 * NOTE: this function does not set the idle thread's NEED_RESCHED
5971 * flag, to make booting more robust.
5973 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5975 struct rq
*rq
= cpu_rq(cpu
);
5976 unsigned long flags
;
5979 idle
->se
.exec_start
= sched_clock();
5981 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5982 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5983 __set_task_cpu(idle
, cpu
);
5985 spin_lock_irqsave(&rq
->lock
, flags
);
5986 rq
->curr
= rq
->idle
= idle
;
5987 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5990 spin_unlock_irqrestore(&rq
->lock
, flags
);
5992 /* Set the preempt count _outside_ the spinlocks! */
5993 task_thread_info(idle
)->preempt_count
= 0;
5996 * The idle tasks have their own, simple scheduling class:
5998 idle
->sched_class
= &idle_sched_class
;
6002 * In a system that switches off the HZ timer nohz_cpu_mask
6003 * indicates which cpus entered this state. This is used
6004 * in the rcu update to wait only for active cpus. For system
6005 * which do not switch off the HZ timer nohz_cpu_mask should
6006 * always be CPU_MASK_NONE.
6008 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
6011 * Increase the granularity value when there are more CPUs,
6012 * because with more CPUs the 'effective latency' as visible
6013 * to users decreases. But the relationship is not linear,
6014 * so pick a second-best guess by going with the log2 of the
6017 * This idea comes from the SD scheduler of Con Kolivas:
6019 static inline void sched_init_granularity(void)
6021 unsigned int factor
= 1 + ilog2(num_online_cpus());
6022 const unsigned long limit
= 200000000;
6024 sysctl_sched_min_granularity
*= factor
;
6025 if (sysctl_sched_min_granularity
> limit
)
6026 sysctl_sched_min_granularity
= limit
;
6028 sysctl_sched_latency
*= factor
;
6029 if (sysctl_sched_latency
> limit
)
6030 sysctl_sched_latency
= limit
;
6032 sysctl_sched_wakeup_granularity
*= factor
;
6037 * This is how migration works:
6039 * 1) we queue a struct migration_req structure in the source CPU's
6040 * runqueue and wake up that CPU's migration thread.
6041 * 2) we down() the locked semaphore => thread blocks.
6042 * 3) migration thread wakes up (implicitly it forces the migrated
6043 * thread off the CPU)
6044 * 4) it gets the migration request and checks whether the migrated
6045 * task is still in the wrong runqueue.
6046 * 5) if it's in the wrong runqueue then the migration thread removes
6047 * it and puts it into the right queue.
6048 * 6) migration thread up()s the semaphore.
6049 * 7) we wake up and the migration is done.
6053 * Change a given task's CPU affinity. Migrate the thread to a
6054 * proper CPU and schedule it away if the CPU it's executing on
6055 * is removed from the allowed bitmask.
6057 * NOTE: the caller must have a valid reference to the task, the
6058 * task must not exit() & deallocate itself prematurely. The
6059 * call is not atomic; no spinlocks may be held.
6061 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
6063 struct migration_req req
;
6064 unsigned long flags
;
6068 rq
= task_rq_lock(p
, &flags
);
6069 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
6074 if (p
->sched_class
->set_cpus_allowed
)
6075 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6077 p
->cpus_allowed
= *new_mask
;
6078 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
6081 /* Can the task run on the task's current CPU? If so, we're done */
6082 if (cpu_isset(task_cpu(p
), *new_mask
))
6085 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
6086 /* Need help from migration thread: drop lock and wait. */
6087 task_rq_unlock(rq
, &flags
);
6088 wake_up_process(rq
->migration_thread
);
6089 wait_for_completion(&req
.done
);
6090 tlb_migrate_finish(p
->mm
);
6094 task_rq_unlock(rq
, &flags
);
6098 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6101 * Move (not current) task off this cpu, onto dest cpu. We're doing
6102 * this because either it can't run here any more (set_cpus_allowed()
6103 * away from this CPU, or CPU going down), or because we're
6104 * attempting to rebalance this task on exec (sched_exec).
6106 * So we race with normal scheduler movements, but that's OK, as long
6107 * as the task is no longer on this CPU.
6109 * Returns non-zero if task was successfully migrated.
6111 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6113 struct rq
*rq_dest
, *rq_src
;
6116 if (unlikely(cpu_is_offline(dest_cpu
)))
6119 rq_src
= cpu_rq(src_cpu
);
6120 rq_dest
= cpu_rq(dest_cpu
);
6122 double_rq_lock(rq_src
, rq_dest
);
6123 /* Already moved. */
6124 if (task_cpu(p
) != src_cpu
)
6126 /* Affinity changed (again). */
6127 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6130 on_rq
= p
->se
.on_rq
;
6132 deactivate_task(rq_src
, p
, 0);
6134 set_task_cpu(p
, dest_cpu
);
6136 activate_task(rq_dest
, p
, 0);
6137 check_preempt_curr(rq_dest
, p
);
6141 double_rq_unlock(rq_src
, rq_dest
);
6146 * migration_thread - this is a highprio system thread that performs
6147 * thread migration by bumping thread off CPU then 'pushing' onto
6150 static int migration_thread(void *data
)
6152 int cpu
= (long)data
;
6156 BUG_ON(rq
->migration_thread
!= current
);
6158 set_current_state(TASK_INTERRUPTIBLE
);
6159 while (!kthread_should_stop()) {
6160 struct migration_req
*req
;
6161 struct list_head
*head
;
6163 spin_lock_irq(&rq
->lock
);
6165 if (cpu_is_offline(cpu
)) {
6166 spin_unlock_irq(&rq
->lock
);
6170 if (rq
->active_balance
) {
6171 active_load_balance(rq
, cpu
);
6172 rq
->active_balance
= 0;
6175 head
= &rq
->migration_queue
;
6177 if (list_empty(head
)) {
6178 spin_unlock_irq(&rq
->lock
);
6180 set_current_state(TASK_INTERRUPTIBLE
);
6183 req
= list_entry(head
->next
, struct migration_req
, list
);
6184 list_del_init(head
->next
);
6186 spin_unlock(&rq
->lock
);
6187 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6190 complete(&req
->done
);
6192 __set_current_state(TASK_RUNNING
);
6196 /* Wait for kthread_stop */
6197 set_current_state(TASK_INTERRUPTIBLE
);
6198 while (!kthread_should_stop()) {
6200 set_current_state(TASK_INTERRUPTIBLE
);
6202 __set_current_state(TASK_RUNNING
);
6206 #ifdef CONFIG_HOTPLUG_CPU
6208 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6212 local_irq_disable();
6213 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6219 * Figure out where task on dead CPU should go, use force if necessary.
6220 * NOTE: interrupts should be disabled by the caller
6222 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6224 unsigned long flags
;
6231 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6232 cpus_and(mask
, mask
, p
->cpus_allowed
);
6233 dest_cpu
= any_online_cpu(mask
);
6235 /* On any allowed CPU? */
6236 if (dest_cpu
>= nr_cpu_ids
)
6237 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6239 /* No more Mr. Nice Guy. */
6240 if (dest_cpu
>= nr_cpu_ids
) {
6241 cpumask_t cpus_allowed
;
6243 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6245 * Try to stay on the same cpuset, where the
6246 * current cpuset may be a subset of all cpus.
6247 * The cpuset_cpus_allowed_locked() variant of
6248 * cpuset_cpus_allowed() will not block. It must be
6249 * called within calls to cpuset_lock/cpuset_unlock.
6251 rq
= task_rq_lock(p
, &flags
);
6252 p
->cpus_allowed
= cpus_allowed
;
6253 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6254 task_rq_unlock(rq
, &flags
);
6257 * Don't tell them about moving exiting tasks or
6258 * kernel threads (both mm NULL), since they never
6261 if (p
->mm
&& printk_ratelimit()) {
6262 printk(KERN_INFO
"process %d (%s) no "
6263 "longer affine to cpu%d\n",
6264 task_pid_nr(p
), p
->comm
, dead_cpu
);
6267 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6271 * While a dead CPU has no uninterruptible tasks queued at this point,
6272 * it might still have a nonzero ->nr_uninterruptible counter, because
6273 * for performance reasons the counter is not stricly tracking tasks to
6274 * their home CPUs. So we just add the counter to another CPU's counter,
6275 * to keep the global sum constant after CPU-down:
6277 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6279 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6280 unsigned long flags
;
6282 local_irq_save(flags
);
6283 double_rq_lock(rq_src
, rq_dest
);
6284 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6285 rq_src
->nr_uninterruptible
= 0;
6286 double_rq_unlock(rq_src
, rq_dest
);
6287 local_irq_restore(flags
);
6290 /* Run through task list and migrate tasks from the dead cpu. */
6291 static void migrate_live_tasks(int src_cpu
)
6293 struct task_struct
*p
, *t
;
6295 read_lock(&tasklist_lock
);
6297 do_each_thread(t
, p
) {
6301 if (task_cpu(p
) == src_cpu
)
6302 move_task_off_dead_cpu(src_cpu
, p
);
6303 } while_each_thread(t
, p
);
6305 read_unlock(&tasklist_lock
);
6309 * Schedules idle task to be the next runnable task on current CPU.
6310 * It does so by boosting its priority to highest possible.
6311 * Used by CPU offline code.
6313 void sched_idle_next(void)
6315 int this_cpu
= smp_processor_id();
6316 struct rq
*rq
= cpu_rq(this_cpu
);
6317 struct task_struct
*p
= rq
->idle
;
6318 unsigned long flags
;
6320 /* cpu has to be offline */
6321 BUG_ON(cpu_online(this_cpu
));
6324 * Strictly not necessary since rest of the CPUs are stopped by now
6325 * and interrupts disabled on the current cpu.
6327 spin_lock_irqsave(&rq
->lock
, flags
);
6329 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6331 update_rq_clock(rq
);
6332 activate_task(rq
, p
, 0);
6334 spin_unlock_irqrestore(&rq
->lock
, flags
);
6338 * Ensures that the idle task is using init_mm right before its cpu goes
6341 void idle_task_exit(void)
6343 struct mm_struct
*mm
= current
->active_mm
;
6345 BUG_ON(cpu_online(smp_processor_id()));
6348 switch_mm(mm
, &init_mm
, current
);
6352 /* called under rq->lock with disabled interrupts */
6353 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6355 struct rq
*rq
= cpu_rq(dead_cpu
);
6357 /* Must be exiting, otherwise would be on tasklist. */
6358 BUG_ON(!p
->exit_state
);
6360 /* Cannot have done final schedule yet: would have vanished. */
6361 BUG_ON(p
->state
== TASK_DEAD
);
6366 * Drop lock around migration; if someone else moves it,
6367 * that's OK. No task can be added to this CPU, so iteration is
6370 spin_unlock_irq(&rq
->lock
);
6371 move_task_off_dead_cpu(dead_cpu
, p
);
6372 spin_lock_irq(&rq
->lock
);
6377 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6378 static void migrate_dead_tasks(unsigned int dead_cpu
)
6380 struct rq
*rq
= cpu_rq(dead_cpu
);
6381 struct task_struct
*next
;
6384 if (!rq
->nr_running
)
6386 update_rq_clock(rq
);
6387 next
= pick_next_task(rq
, rq
->curr
);
6390 migrate_dead(dead_cpu
, next
);
6394 #endif /* CONFIG_HOTPLUG_CPU */
6396 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6398 static struct ctl_table sd_ctl_dir
[] = {
6400 .procname
= "sched_domain",
6406 static struct ctl_table sd_ctl_root
[] = {
6408 .ctl_name
= CTL_KERN
,
6409 .procname
= "kernel",
6411 .child
= sd_ctl_dir
,
6416 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6418 struct ctl_table
*entry
=
6419 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6424 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6426 struct ctl_table
*entry
;
6429 * In the intermediate directories, both the child directory and
6430 * procname are dynamically allocated and could fail but the mode
6431 * will always be set. In the lowest directory the names are
6432 * static strings and all have proc handlers.
6434 for (entry
= *tablep
; entry
->mode
; entry
++) {
6436 sd_free_ctl_entry(&entry
->child
);
6437 if (entry
->proc_handler
== NULL
)
6438 kfree(entry
->procname
);
6446 set_table_entry(struct ctl_table
*entry
,
6447 const char *procname
, void *data
, int maxlen
,
6448 mode_t mode
, proc_handler
*proc_handler
)
6450 entry
->procname
= procname
;
6452 entry
->maxlen
= maxlen
;
6454 entry
->proc_handler
= proc_handler
;
6457 static struct ctl_table
*
6458 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6460 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6465 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6466 sizeof(long), 0644, proc_doulongvec_minmax
);
6467 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6468 sizeof(long), 0644, proc_doulongvec_minmax
);
6469 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6470 sizeof(int), 0644, proc_dointvec_minmax
);
6471 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6472 sizeof(int), 0644, proc_dointvec_minmax
);
6473 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6474 sizeof(int), 0644, proc_dointvec_minmax
);
6475 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6476 sizeof(int), 0644, proc_dointvec_minmax
);
6477 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6478 sizeof(int), 0644, proc_dointvec_minmax
);
6479 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6480 sizeof(int), 0644, proc_dointvec_minmax
);
6481 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6482 sizeof(int), 0644, proc_dointvec_minmax
);
6483 set_table_entry(&table
[9], "cache_nice_tries",
6484 &sd
->cache_nice_tries
,
6485 sizeof(int), 0644, proc_dointvec_minmax
);
6486 set_table_entry(&table
[10], "flags", &sd
->flags
,
6487 sizeof(int), 0644, proc_dointvec_minmax
);
6488 /* &table[11] is terminator */
6493 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6495 struct ctl_table
*entry
, *table
;
6496 struct sched_domain
*sd
;
6497 int domain_num
= 0, i
;
6500 for_each_domain(cpu
, sd
)
6502 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6507 for_each_domain(cpu
, sd
) {
6508 snprintf(buf
, 32, "domain%d", i
);
6509 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6511 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6518 static struct ctl_table_header
*sd_sysctl_header
;
6519 static void register_sched_domain_sysctl(void)
6521 int i
, cpu_num
= num_online_cpus();
6522 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6525 WARN_ON(sd_ctl_dir
[0].child
);
6526 sd_ctl_dir
[0].child
= entry
;
6531 for_each_online_cpu(i
) {
6532 snprintf(buf
, 32, "cpu%d", i
);
6533 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6535 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6539 WARN_ON(sd_sysctl_header
);
6540 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6543 /* may be called multiple times per register */
6544 static void unregister_sched_domain_sysctl(void)
6546 if (sd_sysctl_header
)
6547 unregister_sysctl_table(sd_sysctl_header
);
6548 sd_sysctl_header
= NULL
;
6549 if (sd_ctl_dir
[0].child
)
6550 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6553 static void register_sched_domain_sysctl(void)
6556 static void unregister_sched_domain_sysctl(void)
6562 * migration_call - callback that gets triggered when a CPU is added.
6563 * Here we can start up the necessary migration thread for the new CPU.
6565 static int __cpuinit
6566 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6568 struct task_struct
*p
;
6569 int cpu
= (long)hcpu
;
6570 unsigned long flags
;
6575 case CPU_UP_PREPARE
:
6576 case CPU_UP_PREPARE_FROZEN
:
6577 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6580 kthread_bind(p
, cpu
);
6581 /* Must be high prio: stop_machine expects to yield to it. */
6582 rq
= task_rq_lock(p
, &flags
);
6583 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6584 task_rq_unlock(rq
, &flags
);
6585 cpu_rq(cpu
)->migration_thread
= p
;
6589 case CPU_ONLINE_FROZEN
:
6590 /* Strictly unnecessary, as first user will wake it. */
6591 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6593 /* Update our root-domain */
6595 spin_lock_irqsave(&rq
->lock
, flags
);
6597 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6598 cpu_set(cpu
, rq
->rd
->online
);
6600 spin_unlock_irqrestore(&rq
->lock
, flags
);
6603 #ifdef CONFIG_HOTPLUG_CPU
6604 case CPU_UP_CANCELED
:
6605 case CPU_UP_CANCELED_FROZEN
:
6606 if (!cpu_rq(cpu
)->migration_thread
)
6608 /* Unbind it from offline cpu so it can run. Fall thru. */
6609 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6610 any_online_cpu(cpu_online_map
));
6611 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6612 cpu_rq(cpu
)->migration_thread
= NULL
;
6616 case CPU_DEAD_FROZEN
:
6617 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6618 migrate_live_tasks(cpu
);
6620 kthread_stop(rq
->migration_thread
);
6621 rq
->migration_thread
= NULL
;
6622 /* Idle task back to normal (off runqueue, low prio) */
6623 spin_lock_irq(&rq
->lock
);
6624 update_rq_clock(rq
);
6625 deactivate_task(rq
, rq
->idle
, 0);
6626 rq
->idle
->static_prio
= MAX_PRIO
;
6627 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6628 rq
->idle
->sched_class
= &idle_sched_class
;
6629 migrate_dead_tasks(cpu
);
6630 spin_unlock_irq(&rq
->lock
);
6632 migrate_nr_uninterruptible(rq
);
6633 BUG_ON(rq
->nr_running
!= 0);
6636 * No need to migrate the tasks: it was best-effort if
6637 * they didn't take sched_hotcpu_mutex. Just wake up
6640 spin_lock_irq(&rq
->lock
);
6641 while (!list_empty(&rq
->migration_queue
)) {
6642 struct migration_req
*req
;
6644 req
= list_entry(rq
->migration_queue
.next
,
6645 struct migration_req
, list
);
6646 list_del_init(&req
->list
);
6647 complete(&req
->done
);
6649 spin_unlock_irq(&rq
->lock
);
6653 case CPU_DYING_FROZEN
:
6654 /* Update our root-domain */
6656 spin_lock_irqsave(&rq
->lock
, flags
);
6658 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6659 cpu_clear(cpu
, rq
->rd
->online
);
6661 spin_unlock_irqrestore(&rq
->lock
, flags
);
6668 /* Register at highest priority so that task migration (migrate_all_tasks)
6669 * happens before everything else.
6671 static struct notifier_block __cpuinitdata migration_notifier
= {
6672 .notifier_call
= migration_call
,
6676 void __init
migration_init(void)
6678 void *cpu
= (void *)(long)smp_processor_id();
6681 /* Start one for the boot CPU: */
6682 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6683 BUG_ON(err
== NOTIFY_BAD
);
6684 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6685 register_cpu_notifier(&migration_notifier
);
6691 #ifdef CONFIG_SCHED_DEBUG
6693 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6694 cpumask_t
*groupmask
)
6696 struct sched_group
*group
= sd
->groups
;
6699 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6700 cpus_clear(*groupmask
);
6702 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6704 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6705 printk("does not load-balance\n");
6707 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6712 printk(KERN_CONT
"span %s\n", str
);
6714 if (!cpu_isset(cpu
, sd
->span
)) {
6715 printk(KERN_ERR
"ERROR: domain->span does not contain "
6718 if (!cpu_isset(cpu
, group
->cpumask
)) {
6719 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6723 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6727 printk(KERN_ERR
"ERROR: group is NULL\n");
6731 if (!group
->__cpu_power
) {
6732 printk(KERN_CONT
"\n");
6733 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6738 if (!cpus_weight(group
->cpumask
)) {
6739 printk(KERN_CONT
"\n");
6740 printk(KERN_ERR
"ERROR: empty group\n");
6744 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6745 printk(KERN_CONT
"\n");
6746 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6750 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6752 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6753 printk(KERN_CONT
" %s", str
);
6755 group
= group
->next
;
6756 } while (group
!= sd
->groups
);
6757 printk(KERN_CONT
"\n");
6759 if (!cpus_equal(sd
->span
, *groupmask
))
6760 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6762 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6763 printk(KERN_ERR
"ERROR: parent span is not a superset "
6764 "of domain->span\n");
6768 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6770 cpumask_t
*groupmask
;
6774 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6778 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6780 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6782 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6787 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6797 # define sched_domain_debug(sd, cpu) do { } while (0)
6800 static int sd_degenerate(struct sched_domain
*sd
)
6802 if (cpus_weight(sd
->span
) == 1)
6805 /* Following flags need at least 2 groups */
6806 if (sd
->flags
& (SD_LOAD_BALANCE
|
6807 SD_BALANCE_NEWIDLE
|
6811 SD_SHARE_PKG_RESOURCES
)) {
6812 if (sd
->groups
!= sd
->groups
->next
)
6816 /* Following flags don't use groups */
6817 if (sd
->flags
& (SD_WAKE_IDLE
|
6826 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6828 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6830 if (sd_degenerate(parent
))
6833 if (!cpus_equal(sd
->span
, parent
->span
))
6836 /* Does parent contain flags not in child? */
6837 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6838 if (cflags
& SD_WAKE_AFFINE
)
6839 pflags
&= ~SD_WAKE_BALANCE
;
6840 /* Flags needing groups don't count if only 1 group in parent */
6841 if (parent
->groups
== parent
->groups
->next
) {
6842 pflags
&= ~(SD_LOAD_BALANCE
|
6843 SD_BALANCE_NEWIDLE
|
6847 SD_SHARE_PKG_RESOURCES
);
6849 if (~cflags
& pflags
)
6855 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6857 unsigned long flags
;
6858 const struct sched_class
*class;
6860 spin_lock_irqsave(&rq
->lock
, flags
);
6863 struct root_domain
*old_rd
= rq
->rd
;
6865 for (class = sched_class_highest
; class; class = class->next
) {
6866 if (class->leave_domain
)
6867 class->leave_domain(rq
);
6870 cpu_clear(rq
->cpu
, old_rd
->span
);
6871 cpu_clear(rq
->cpu
, old_rd
->online
);
6873 if (atomic_dec_and_test(&old_rd
->refcount
))
6877 atomic_inc(&rd
->refcount
);
6880 cpu_set(rq
->cpu
, rd
->span
);
6881 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6882 cpu_set(rq
->cpu
, rd
->online
);
6884 for (class = sched_class_highest
; class; class = class->next
) {
6885 if (class->join_domain
)
6886 class->join_domain(rq
);
6889 spin_unlock_irqrestore(&rq
->lock
, flags
);
6892 static void init_rootdomain(struct root_domain
*rd
)
6894 memset(rd
, 0, sizeof(*rd
));
6896 cpus_clear(rd
->span
);
6897 cpus_clear(rd
->online
);
6900 static void init_defrootdomain(void)
6902 init_rootdomain(&def_root_domain
);
6903 atomic_set(&def_root_domain
.refcount
, 1);
6906 static struct root_domain
*alloc_rootdomain(void)
6908 struct root_domain
*rd
;
6910 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6914 init_rootdomain(rd
);
6920 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6921 * hold the hotplug lock.
6924 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6926 struct rq
*rq
= cpu_rq(cpu
);
6927 struct sched_domain
*tmp
;
6929 /* Remove the sched domains which do not contribute to scheduling. */
6930 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6931 struct sched_domain
*parent
= tmp
->parent
;
6934 if (sd_parent_degenerate(tmp
, parent
)) {
6935 tmp
->parent
= parent
->parent
;
6937 parent
->parent
->child
= tmp
;
6941 if (sd
&& sd_degenerate(sd
)) {
6947 sched_domain_debug(sd
, cpu
);
6949 rq_attach_root(rq
, rd
);
6950 rcu_assign_pointer(rq
->sd
, sd
);
6953 /* cpus with isolated domains */
6954 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6956 /* Setup the mask of cpus configured for isolated domains */
6957 static int __init
isolated_cpu_setup(char *str
)
6959 int ints
[NR_CPUS
], i
;
6961 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6962 cpus_clear(cpu_isolated_map
);
6963 for (i
= 1; i
<= ints
[0]; i
++)
6964 if (ints
[i
] < NR_CPUS
)
6965 cpu_set(ints
[i
], cpu_isolated_map
);
6969 __setup("isolcpus=", isolated_cpu_setup
);
6972 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6973 * to a function which identifies what group(along with sched group) a CPU
6974 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6975 * (due to the fact that we keep track of groups covered with a cpumask_t).
6977 * init_sched_build_groups will build a circular linked list of the groups
6978 * covered by the given span, and will set each group's ->cpumask correctly,
6979 * and ->cpu_power to 0.
6982 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6983 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6984 struct sched_group
**sg
,
6985 cpumask_t
*tmpmask
),
6986 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6988 struct sched_group
*first
= NULL
, *last
= NULL
;
6991 cpus_clear(*covered
);
6993 for_each_cpu_mask(i
, *span
) {
6994 struct sched_group
*sg
;
6995 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6998 if (cpu_isset(i
, *covered
))
7001 cpus_clear(sg
->cpumask
);
7002 sg
->__cpu_power
= 0;
7004 for_each_cpu_mask(j
, *span
) {
7005 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
7008 cpu_set(j
, *covered
);
7009 cpu_set(j
, sg
->cpumask
);
7020 #define SD_NODES_PER_DOMAIN 16
7025 * find_next_best_node - find the next node to include in a sched_domain
7026 * @node: node whose sched_domain we're building
7027 * @used_nodes: nodes already in the sched_domain
7029 * Find the next node to include in a given scheduling domain. Simply
7030 * finds the closest node not already in the @used_nodes map.
7032 * Should use nodemask_t.
7034 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7036 int i
, n
, val
, min_val
, best_node
= 0;
7040 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7041 /* Start at @node */
7042 n
= (node
+ i
) % MAX_NUMNODES
;
7044 if (!nr_cpus_node(n
))
7047 /* Skip already used nodes */
7048 if (node_isset(n
, *used_nodes
))
7051 /* Simple min distance search */
7052 val
= node_distance(node
, n
);
7054 if (val
< min_val
) {
7060 node_set(best_node
, *used_nodes
);
7065 * sched_domain_node_span - get a cpumask for a node's sched_domain
7066 * @node: node whose cpumask we're constructing
7067 * @span: resulting cpumask
7069 * Given a node, construct a good cpumask for its sched_domain to span. It
7070 * should be one that prevents unnecessary balancing, but also spreads tasks
7073 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7075 nodemask_t used_nodes
;
7076 node_to_cpumask_ptr(nodemask
, node
);
7080 nodes_clear(used_nodes
);
7082 cpus_or(*span
, *span
, *nodemask
);
7083 node_set(node
, used_nodes
);
7085 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7086 int next_node
= find_next_best_node(node
, &used_nodes
);
7088 node_to_cpumask_ptr_next(nodemask
, next_node
);
7089 cpus_or(*span
, *span
, *nodemask
);
7094 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7097 * SMT sched-domains:
7099 #ifdef CONFIG_SCHED_SMT
7100 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7101 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7104 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7108 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7114 * multi-core sched-domains:
7116 #ifdef CONFIG_SCHED_MC
7117 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7118 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7121 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7123 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7128 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7129 cpus_and(*mask
, *mask
, *cpu_map
);
7130 group
= first_cpu(*mask
);
7132 *sg
= &per_cpu(sched_group_core
, group
);
7135 #elif defined(CONFIG_SCHED_MC)
7137 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7141 *sg
= &per_cpu(sched_group_core
, cpu
);
7146 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7147 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7150 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7154 #ifdef CONFIG_SCHED_MC
7155 *mask
= cpu_coregroup_map(cpu
);
7156 cpus_and(*mask
, *mask
, *cpu_map
);
7157 group
= first_cpu(*mask
);
7158 #elif defined(CONFIG_SCHED_SMT)
7159 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7160 cpus_and(*mask
, *mask
, *cpu_map
);
7161 group
= first_cpu(*mask
);
7166 *sg
= &per_cpu(sched_group_phys
, group
);
7172 * The init_sched_build_groups can't handle what we want to do with node
7173 * groups, so roll our own. Now each node has its own list of groups which
7174 * gets dynamically allocated.
7176 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7177 static struct sched_group
***sched_group_nodes_bycpu
;
7179 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7180 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7182 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7183 struct sched_group
**sg
, cpumask_t
*nodemask
)
7187 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7188 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7189 group
= first_cpu(*nodemask
);
7192 *sg
= &per_cpu(sched_group_allnodes
, group
);
7196 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7198 struct sched_group
*sg
= group_head
;
7204 for_each_cpu_mask(j
, sg
->cpumask
) {
7205 struct sched_domain
*sd
;
7207 sd
= &per_cpu(phys_domains
, j
);
7208 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7210 * Only add "power" once for each
7216 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7219 } while (sg
!= group_head
);
7224 /* Free memory allocated for various sched_group structures */
7225 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7229 for_each_cpu_mask(cpu
, *cpu_map
) {
7230 struct sched_group
**sched_group_nodes
7231 = sched_group_nodes_bycpu
[cpu
];
7233 if (!sched_group_nodes
)
7236 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7237 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7239 *nodemask
= node_to_cpumask(i
);
7240 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7241 if (cpus_empty(*nodemask
))
7251 if (oldsg
!= sched_group_nodes
[i
])
7254 kfree(sched_group_nodes
);
7255 sched_group_nodes_bycpu
[cpu
] = NULL
;
7259 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7265 * Initialize sched groups cpu_power.
7267 * cpu_power indicates the capacity of sched group, which is used while
7268 * distributing the load between different sched groups in a sched domain.
7269 * Typically cpu_power for all the groups in a sched domain will be same unless
7270 * there are asymmetries in the topology. If there are asymmetries, group
7271 * having more cpu_power will pickup more load compared to the group having
7274 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7275 * the maximum number of tasks a group can handle in the presence of other idle
7276 * or lightly loaded groups in the same sched domain.
7278 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7280 struct sched_domain
*child
;
7281 struct sched_group
*group
;
7283 WARN_ON(!sd
|| !sd
->groups
);
7285 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7290 sd
->groups
->__cpu_power
= 0;
7293 * For perf policy, if the groups in child domain share resources
7294 * (for example cores sharing some portions of the cache hierarchy
7295 * or SMT), then set this domain groups cpu_power such that each group
7296 * can handle only one task, when there are other idle groups in the
7297 * same sched domain.
7299 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7301 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7302 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7307 * add cpu_power of each child group to this groups cpu_power
7309 group
= child
->groups
;
7311 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7312 group
= group
->next
;
7313 } while (group
!= child
->groups
);
7317 * Initializers for schedule domains
7318 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7321 #define SD_INIT(sd, type) sd_init_##type(sd)
7322 #define SD_INIT_FUNC(type) \
7323 static noinline void sd_init_##type(struct sched_domain *sd) \
7325 memset(sd, 0, sizeof(*sd)); \
7326 *sd = SD_##type##_INIT; \
7327 sd->level = SD_LV_##type; \
7332 SD_INIT_FUNC(ALLNODES
)
7335 #ifdef CONFIG_SCHED_SMT
7336 SD_INIT_FUNC(SIBLING
)
7338 #ifdef CONFIG_SCHED_MC
7343 * To minimize stack usage kmalloc room for cpumasks and share the
7344 * space as the usage in build_sched_domains() dictates. Used only
7345 * if the amount of space is significant.
7348 cpumask_t tmpmask
; /* make this one first */
7351 cpumask_t this_sibling_map
;
7352 cpumask_t this_core_map
;
7354 cpumask_t send_covered
;
7357 cpumask_t domainspan
;
7359 cpumask_t notcovered
;
7364 #define SCHED_CPUMASK_ALLOC 1
7365 #define SCHED_CPUMASK_FREE(v) kfree(v)
7366 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7368 #define SCHED_CPUMASK_ALLOC 0
7369 #define SCHED_CPUMASK_FREE(v)
7370 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7373 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7374 ((unsigned long)(a) + offsetof(struct allmasks, v))
7376 static int default_relax_domain_level
= -1;
7378 static int __init
setup_relax_domain_level(char *str
)
7380 default_relax_domain_level
= simple_strtoul(str
, NULL
, 0);
7383 __setup("relax_domain_level=", setup_relax_domain_level
);
7385 static void set_domain_attribute(struct sched_domain
*sd
,
7386 struct sched_domain_attr
*attr
)
7390 if (!attr
|| attr
->relax_domain_level
< 0) {
7391 if (default_relax_domain_level
< 0)
7394 request
= default_relax_domain_level
;
7396 request
= attr
->relax_domain_level
;
7397 if (request
< sd
->level
) {
7398 /* turn off idle balance on this domain */
7399 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7401 /* turn on idle balance on this domain */
7402 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7407 * Build sched domains for a given set of cpus and attach the sched domains
7408 * to the individual cpus
7410 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7411 struct sched_domain_attr
*attr
)
7414 struct root_domain
*rd
;
7415 SCHED_CPUMASK_DECLARE(allmasks
);
7418 struct sched_group
**sched_group_nodes
= NULL
;
7419 int sd_allnodes
= 0;
7422 * Allocate the per-node list of sched groups
7424 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
7426 if (!sched_group_nodes
) {
7427 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7432 rd
= alloc_rootdomain();
7434 printk(KERN_WARNING
"Cannot alloc root domain\n");
7436 kfree(sched_group_nodes
);
7441 #if SCHED_CPUMASK_ALLOC
7442 /* get space for all scratch cpumask variables */
7443 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7445 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7448 kfree(sched_group_nodes
);
7453 tmpmask
= (cpumask_t
*)allmasks
;
7457 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7461 * Set up domains for cpus specified by the cpu_map.
7463 for_each_cpu_mask(i
, *cpu_map
) {
7464 struct sched_domain
*sd
= NULL
, *p
;
7465 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7467 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7468 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7471 if (cpus_weight(*cpu_map
) >
7472 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7473 sd
= &per_cpu(allnodes_domains
, i
);
7474 SD_INIT(sd
, ALLNODES
);
7475 set_domain_attribute(sd
, attr
);
7476 sd
->span
= *cpu_map
;
7477 sd
->first_cpu
= first_cpu(sd
->span
);
7478 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7484 sd
= &per_cpu(node_domains
, i
);
7486 set_domain_attribute(sd
, attr
);
7487 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7488 sd
->first_cpu
= first_cpu(sd
->span
);
7492 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7496 sd
= &per_cpu(phys_domains
, i
);
7498 set_domain_attribute(sd
, attr
);
7499 sd
->span
= *nodemask
;
7500 sd
->first_cpu
= first_cpu(sd
->span
);
7504 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7506 #ifdef CONFIG_SCHED_MC
7508 sd
= &per_cpu(core_domains
, i
);
7510 set_domain_attribute(sd
, attr
);
7511 sd
->span
= cpu_coregroup_map(i
);
7512 sd
->first_cpu
= first_cpu(sd
->span
);
7513 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7516 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7519 #ifdef CONFIG_SCHED_SMT
7521 sd
= &per_cpu(cpu_domains
, i
);
7522 SD_INIT(sd
, SIBLING
);
7523 set_domain_attribute(sd
, attr
);
7524 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7525 sd
->first_cpu
= first_cpu(sd
->span
);
7526 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7529 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7533 #ifdef CONFIG_SCHED_SMT
7534 /* Set up CPU (sibling) groups */
7535 for_each_cpu_mask(i
, *cpu_map
) {
7536 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7537 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7539 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7540 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7541 if (i
!= first_cpu(*this_sibling_map
))
7544 init_sched_build_groups(this_sibling_map
, cpu_map
,
7546 send_covered
, tmpmask
);
7550 #ifdef CONFIG_SCHED_MC
7551 /* Set up multi-core groups */
7552 for_each_cpu_mask(i
, *cpu_map
) {
7553 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7554 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7556 *this_core_map
= cpu_coregroup_map(i
);
7557 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7558 if (i
!= first_cpu(*this_core_map
))
7561 init_sched_build_groups(this_core_map
, cpu_map
,
7563 send_covered
, tmpmask
);
7567 /* Set up physical groups */
7568 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7569 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7570 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7572 *nodemask
= node_to_cpumask(i
);
7573 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7574 if (cpus_empty(*nodemask
))
7577 init_sched_build_groups(nodemask
, cpu_map
,
7579 send_covered
, tmpmask
);
7583 /* Set up node groups */
7585 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7587 init_sched_build_groups(cpu_map
, cpu_map
,
7588 &cpu_to_allnodes_group
,
7589 send_covered
, tmpmask
);
7592 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7593 /* Set up node groups */
7594 struct sched_group
*sg
, *prev
;
7595 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7596 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7597 SCHED_CPUMASK_VAR(covered
, allmasks
);
7600 *nodemask
= node_to_cpumask(i
);
7601 cpus_clear(*covered
);
7603 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7604 if (cpus_empty(*nodemask
)) {
7605 sched_group_nodes
[i
] = NULL
;
7609 sched_domain_node_span(i
, domainspan
);
7610 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7612 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7614 printk(KERN_WARNING
"Can not alloc domain group for "
7618 sched_group_nodes
[i
] = sg
;
7619 for_each_cpu_mask(j
, *nodemask
) {
7620 struct sched_domain
*sd
;
7622 sd
= &per_cpu(node_domains
, j
);
7625 sg
->__cpu_power
= 0;
7626 sg
->cpumask
= *nodemask
;
7628 cpus_or(*covered
, *covered
, *nodemask
);
7631 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7632 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7633 int n
= (i
+ j
) % MAX_NUMNODES
;
7634 node_to_cpumask_ptr(pnodemask
, n
);
7636 cpus_complement(*notcovered
, *covered
);
7637 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7638 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7639 if (cpus_empty(*tmpmask
))
7642 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7643 if (cpus_empty(*tmpmask
))
7646 sg
= kmalloc_node(sizeof(struct sched_group
),
7650 "Can not alloc domain group for node %d\n", j
);
7653 sg
->__cpu_power
= 0;
7654 sg
->cpumask
= *tmpmask
;
7655 sg
->next
= prev
->next
;
7656 cpus_or(*covered
, *covered
, *tmpmask
);
7663 /* Calculate CPU power for physical packages and nodes */
7664 #ifdef CONFIG_SCHED_SMT
7665 for_each_cpu_mask(i
, *cpu_map
) {
7666 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7668 init_sched_groups_power(i
, sd
);
7671 #ifdef CONFIG_SCHED_MC
7672 for_each_cpu_mask(i
, *cpu_map
) {
7673 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7675 init_sched_groups_power(i
, sd
);
7679 for_each_cpu_mask(i
, *cpu_map
) {
7680 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7682 init_sched_groups_power(i
, sd
);
7686 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7687 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7690 struct sched_group
*sg
;
7692 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7694 init_numa_sched_groups_power(sg
);
7698 /* Attach the domains */
7699 for_each_cpu_mask(i
, *cpu_map
) {
7700 struct sched_domain
*sd
;
7701 #ifdef CONFIG_SCHED_SMT
7702 sd
= &per_cpu(cpu_domains
, i
);
7703 #elif defined(CONFIG_SCHED_MC)
7704 sd
= &per_cpu(core_domains
, i
);
7706 sd
= &per_cpu(phys_domains
, i
);
7708 cpu_attach_domain(sd
, rd
, i
);
7711 SCHED_CPUMASK_FREE((void *)allmasks
);
7716 free_sched_groups(cpu_map
, tmpmask
);
7717 SCHED_CPUMASK_FREE((void *)allmasks
);
7722 static int build_sched_domains(const cpumask_t
*cpu_map
)
7724 return __build_sched_domains(cpu_map
, NULL
);
7727 static cpumask_t
*doms_cur
; /* current sched domains */
7728 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7729 static struct sched_domain_attr
*dattr_cur
; /* attribues of custom domains
7733 * Special case: If a kmalloc of a doms_cur partition (array of
7734 * cpumask_t) fails, then fallback to a single sched domain,
7735 * as determined by the single cpumask_t fallback_doms.
7737 static cpumask_t fallback_doms
;
7739 void __attribute__((weak
)) arch_update_cpu_topology(void)
7744 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7745 * For now this just excludes isolated cpus, but could be used to
7746 * exclude other special cases in the future.
7748 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7752 arch_update_cpu_topology();
7754 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7756 doms_cur
= &fallback_doms
;
7757 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7759 err
= build_sched_domains(doms_cur
);
7760 register_sched_domain_sysctl();
7765 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7768 free_sched_groups(cpu_map
, tmpmask
);
7772 * Detach sched domains from a group of cpus specified in cpu_map
7773 * These cpus will now be attached to the NULL domain
7775 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7780 unregister_sched_domain_sysctl();
7782 for_each_cpu_mask(i
, *cpu_map
)
7783 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7784 synchronize_sched();
7785 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7788 /* handle null as "default" */
7789 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7790 struct sched_domain_attr
*new, int idx_new
)
7792 struct sched_domain_attr tmp
;
7799 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7800 new ? (new + idx_new
) : &tmp
,
7801 sizeof(struct sched_domain_attr
));
7805 * Partition sched domains as specified by the 'ndoms_new'
7806 * cpumasks in the array doms_new[] of cpumasks. This compares
7807 * doms_new[] to the current sched domain partitioning, doms_cur[].
7808 * It destroys each deleted domain and builds each new domain.
7810 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7811 * The masks don't intersect (don't overlap.) We should setup one
7812 * sched domain for each mask. CPUs not in any of the cpumasks will
7813 * not be load balanced. If the same cpumask appears both in the
7814 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7817 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7818 * ownership of it and will kfree it when done with it. If the caller
7819 * failed the kmalloc call, then it can pass in doms_new == NULL,
7820 * and partition_sched_domains() will fallback to the single partition
7823 * Call with hotplug lock held
7825 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7826 struct sched_domain_attr
*dattr_new
)
7830 mutex_lock(&sched_domains_mutex
);
7832 /* always unregister in case we don't destroy any domains */
7833 unregister_sched_domain_sysctl();
7835 if (doms_new
== NULL
) {
7837 doms_new
= &fallback_doms
;
7838 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7842 /* Destroy deleted domains */
7843 for (i
= 0; i
< ndoms_cur
; i
++) {
7844 for (j
= 0; j
< ndoms_new
; j
++) {
7845 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7846 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7849 /* no match - a current sched domain not in new doms_new[] */
7850 detach_destroy_domains(doms_cur
+ i
);
7855 /* Build new domains */
7856 for (i
= 0; i
< ndoms_new
; i
++) {
7857 for (j
= 0; j
< ndoms_cur
; j
++) {
7858 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7859 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7862 /* no match - add a new doms_new */
7863 __build_sched_domains(doms_new
+ i
,
7864 dattr_new
? dattr_new
+ i
: NULL
);
7869 /* Remember the new sched domains */
7870 if (doms_cur
!= &fallback_doms
)
7872 kfree(dattr_cur
); /* kfree(NULL) is safe */
7873 doms_cur
= doms_new
;
7874 dattr_cur
= dattr_new
;
7875 ndoms_cur
= ndoms_new
;
7877 register_sched_domain_sysctl();
7879 mutex_unlock(&sched_domains_mutex
);
7882 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7883 int arch_reinit_sched_domains(void)
7888 mutex_lock(&sched_domains_mutex
);
7889 detach_destroy_domains(&cpu_online_map
);
7890 err
= arch_init_sched_domains(&cpu_online_map
);
7891 mutex_unlock(&sched_domains_mutex
);
7897 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7901 if (buf
[0] != '0' && buf
[0] != '1')
7905 sched_smt_power_savings
= (buf
[0] == '1');
7907 sched_mc_power_savings
= (buf
[0] == '1');
7909 ret
= arch_reinit_sched_domains();
7911 return ret
? ret
: count
;
7914 #ifdef CONFIG_SCHED_MC
7915 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7917 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7919 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7920 const char *buf
, size_t count
)
7922 return sched_power_savings_store(buf
, count
, 0);
7924 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7925 sched_mc_power_savings_store
);
7928 #ifdef CONFIG_SCHED_SMT
7929 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7931 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7933 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7934 const char *buf
, size_t count
)
7936 return sched_power_savings_store(buf
, count
, 1);
7938 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7939 sched_smt_power_savings_store
);
7942 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7946 #ifdef CONFIG_SCHED_SMT
7948 err
= sysfs_create_file(&cls
->kset
.kobj
,
7949 &attr_sched_smt_power_savings
.attr
);
7951 #ifdef CONFIG_SCHED_MC
7952 if (!err
&& mc_capable())
7953 err
= sysfs_create_file(&cls
->kset
.kobj
,
7954 &attr_sched_mc_power_savings
.attr
);
7961 * Force a reinitialization of the sched domains hierarchy. The domains
7962 * and groups cannot be updated in place without racing with the balancing
7963 * code, so we temporarily attach all running cpus to the NULL domain
7964 * which will prevent rebalancing while the sched domains are recalculated.
7966 static int update_sched_domains(struct notifier_block
*nfb
,
7967 unsigned long action
, void *hcpu
)
7970 case CPU_UP_PREPARE
:
7971 case CPU_UP_PREPARE_FROZEN
:
7972 case CPU_DOWN_PREPARE
:
7973 case CPU_DOWN_PREPARE_FROZEN
:
7974 detach_destroy_domains(&cpu_online_map
);
7977 case CPU_UP_CANCELED
:
7978 case CPU_UP_CANCELED_FROZEN
:
7979 case CPU_DOWN_FAILED
:
7980 case CPU_DOWN_FAILED_FROZEN
:
7982 case CPU_ONLINE_FROZEN
:
7984 case CPU_DEAD_FROZEN
:
7986 * Fall through and re-initialise the domains.
7993 /* The hotplug lock is already held by cpu_up/cpu_down */
7994 arch_init_sched_domains(&cpu_online_map
);
7999 void __init
sched_init_smp(void)
8001 cpumask_t non_isolated_cpus
;
8003 #if defined(CONFIG_NUMA)
8004 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8006 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8009 mutex_lock(&sched_domains_mutex
);
8010 arch_init_sched_domains(&cpu_online_map
);
8011 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
8012 if (cpus_empty(non_isolated_cpus
))
8013 cpu_set(smp_processor_id(), non_isolated_cpus
);
8014 mutex_unlock(&sched_domains_mutex
);
8016 /* XXX: Theoretical race here - CPU may be hotplugged now */
8017 hotcpu_notifier(update_sched_domains
, 0);
8020 /* Move init over to a non-isolated CPU */
8021 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
8023 sched_init_granularity();
8026 void __init
sched_init_smp(void)
8028 sched_init_granularity();
8030 #endif /* CONFIG_SMP */
8032 int in_sched_functions(unsigned long addr
)
8034 return in_lock_functions(addr
) ||
8035 (addr
>= (unsigned long)__sched_text_start
8036 && addr
< (unsigned long)__sched_text_end
);
8039 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8041 cfs_rq
->tasks_timeline
= RB_ROOT
;
8042 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8043 #ifdef CONFIG_FAIR_GROUP_SCHED
8046 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8049 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8051 struct rt_prio_array
*array
;
8054 array
= &rt_rq
->active
;
8055 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8056 INIT_LIST_HEAD(array
->queue
+ i
);
8057 __clear_bit(i
, array
->bitmap
);
8059 /* delimiter for bitsearch: */
8060 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8062 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8063 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8066 rt_rq
->rt_nr_migratory
= 0;
8067 rt_rq
->overloaded
= 0;
8071 rt_rq
->rt_throttled
= 0;
8072 rt_rq
->rt_runtime
= 0;
8073 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8075 #ifdef CONFIG_RT_GROUP_SCHED
8076 rt_rq
->rt_nr_boosted
= 0;
8081 #ifdef CONFIG_FAIR_GROUP_SCHED
8082 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8083 struct sched_entity
*se
, int cpu
, int add
,
8084 struct sched_entity
*parent
)
8086 struct rq
*rq
= cpu_rq(cpu
);
8087 tg
->cfs_rq
[cpu
] = cfs_rq
;
8088 init_cfs_rq(cfs_rq
, rq
);
8091 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8094 /* se could be NULL for init_task_group */
8099 se
->cfs_rq
= &rq
->cfs
;
8101 se
->cfs_rq
= parent
->my_q
;
8104 se
->load
.weight
= tg
->shares
;
8105 se
->load
.inv_weight
= 0;
8106 se
->parent
= parent
;
8110 #ifdef CONFIG_RT_GROUP_SCHED
8111 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8112 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8113 struct sched_rt_entity
*parent
)
8115 struct rq
*rq
= cpu_rq(cpu
);
8117 tg
->rt_rq
[cpu
] = rt_rq
;
8118 init_rt_rq(rt_rq
, rq
);
8120 rt_rq
->rt_se
= rt_se
;
8121 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8123 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8125 tg
->rt_se
[cpu
] = rt_se
;
8130 rt_se
->rt_rq
= &rq
->rt
;
8132 rt_se
->rt_rq
= parent
->my_q
;
8134 rt_se
->rt_rq
= &rq
->rt
;
8135 rt_se
->my_q
= rt_rq
;
8136 rt_se
->parent
= parent
;
8137 INIT_LIST_HEAD(&rt_se
->run_list
);
8141 void __init
sched_init(void)
8144 unsigned long alloc_size
= 0, ptr
;
8146 #ifdef CONFIG_FAIR_GROUP_SCHED
8147 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8149 #ifdef CONFIG_RT_GROUP_SCHED
8150 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8152 #ifdef CONFIG_USER_SCHED
8156 * As sched_init() is called before page_alloc is setup,
8157 * we use alloc_bootmem().
8160 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8162 #ifdef CONFIG_FAIR_GROUP_SCHED
8163 init_task_group
.se
= (struct sched_entity
**)ptr
;
8164 ptr
+= nr_cpu_ids
* sizeof(void **);
8166 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8167 ptr
+= nr_cpu_ids
* sizeof(void **);
8169 #ifdef CONFIG_USER_SCHED
8170 root_task_group
.se
= (struct sched_entity
**)ptr
;
8171 ptr
+= nr_cpu_ids
* sizeof(void **);
8173 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8174 ptr
+= nr_cpu_ids
* sizeof(void **);
8177 #ifdef CONFIG_RT_GROUP_SCHED
8178 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8179 ptr
+= nr_cpu_ids
* sizeof(void **);
8181 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8182 ptr
+= nr_cpu_ids
* sizeof(void **);
8184 #ifdef CONFIG_USER_SCHED
8185 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8186 ptr
+= nr_cpu_ids
* sizeof(void **);
8188 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8189 ptr
+= nr_cpu_ids
* sizeof(void **);
8196 init_defrootdomain();
8199 init_rt_bandwidth(&def_rt_bandwidth
,
8200 global_rt_period(), global_rt_runtime());
8202 #ifdef CONFIG_RT_GROUP_SCHED
8203 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8204 global_rt_period(), global_rt_runtime());
8205 #ifdef CONFIG_USER_SCHED
8206 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8207 global_rt_period(), RUNTIME_INF
);
8211 #ifdef CONFIG_GROUP_SCHED
8212 list_add(&init_task_group
.list
, &task_groups
);
8213 INIT_LIST_HEAD(&init_task_group
.children
);
8215 #ifdef CONFIG_USER_SCHED
8216 INIT_LIST_HEAD(&root_task_group
.children
);
8217 init_task_group
.parent
= &root_task_group
;
8218 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8222 for_each_possible_cpu(i
) {
8226 spin_lock_init(&rq
->lock
);
8227 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
8230 update_last_tick_seen(rq
);
8231 init_cfs_rq(&rq
->cfs
, rq
);
8232 init_rt_rq(&rq
->rt
, rq
);
8233 #ifdef CONFIG_FAIR_GROUP_SCHED
8234 init_task_group
.shares
= init_task_group_load
;
8235 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8236 #ifdef CONFIG_CGROUP_SCHED
8238 * How much cpu bandwidth does init_task_group get?
8240 * In case of task-groups formed thr' the cgroup filesystem, it
8241 * gets 100% of the cpu resources in the system. This overall
8242 * system cpu resource is divided among the tasks of
8243 * init_task_group and its child task-groups in a fair manner,
8244 * based on each entity's (task or task-group's) weight
8245 * (se->load.weight).
8247 * In other words, if init_task_group has 10 tasks of weight
8248 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8249 * then A0's share of the cpu resource is:
8251 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8253 * We achieve this by letting init_task_group's tasks sit
8254 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8256 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8257 #elif defined CONFIG_USER_SCHED
8258 root_task_group
.shares
= NICE_0_LOAD
;
8259 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8261 * In case of task-groups formed thr' the user id of tasks,
8262 * init_task_group represents tasks belonging to root user.
8263 * Hence it forms a sibling of all subsequent groups formed.
8264 * In this case, init_task_group gets only a fraction of overall
8265 * system cpu resource, based on the weight assigned to root
8266 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8267 * by letting tasks of init_task_group sit in a separate cfs_rq
8268 * (init_cfs_rq) and having one entity represent this group of
8269 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8271 init_tg_cfs_entry(&init_task_group
,
8272 &per_cpu(init_cfs_rq
, i
),
8273 &per_cpu(init_sched_entity
, i
), i
, 1,
8274 root_task_group
.se
[i
]);
8277 #endif /* CONFIG_FAIR_GROUP_SCHED */
8279 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8280 #ifdef CONFIG_RT_GROUP_SCHED
8281 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8282 #ifdef CONFIG_CGROUP_SCHED
8283 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8284 #elif defined CONFIG_USER_SCHED
8285 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8286 init_tg_rt_entry(&init_task_group
,
8287 &per_cpu(init_rt_rq
, i
),
8288 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8289 root_task_group
.rt_se
[i
]);
8293 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8294 rq
->cpu_load
[j
] = 0;
8298 rq
->active_balance
= 0;
8299 rq
->next_balance
= jiffies
;
8302 rq
->migration_thread
= NULL
;
8303 INIT_LIST_HEAD(&rq
->migration_queue
);
8304 rq_attach_root(rq
, &def_root_domain
);
8307 atomic_set(&rq
->nr_iowait
, 0);
8310 set_load_weight(&init_task
);
8312 #ifdef CONFIG_PREEMPT_NOTIFIERS
8313 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8317 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
8320 #ifdef CONFIG_RT_MUTEXES
8321 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8325 * The boot idle thread does lazy MMU switching as well:
8327 atomic_inc(&init_mm
.mm_count
);
8328 enter_lazy_tlb(&init_mm
, current
);
8331 * Make us the idle thread. Technically, schedule() should not be
8332 * called from this thread, however somewhere below it might be,
8333 * but because we are the idle thread, we just pick up running again
8334 * when this runqueue becomes "idle".
8336 init_idle(current
, smp_processor_id());
8338 * During early bootup we pretend to be a normal task:
8340 current
->sched_class
= &fair_sched_class
;
8342 scheduler_running
= 1;
8345 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8346 void __might_sleep(char *file
, int line
)
8349 static unsigned long prev_jiffy
; /* ratelimiting */
8351 if ((in_atomic() || irqs_disabled()) &&
8352 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
8353 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8355 prev_jiffy
= jiffies
;
8356 printk(KERN_ERR
"BUG: sleeping function called from invalid"
8357 " context at %s:%d\n", file
, line
);
8358 printk("in_atomic():%d, irqs_disabled():%d\n",
8359 in_atomic(), irqs_disabled());
8360 debug_show_held_locks(current
);
8361 if (irqs_disabled())
8362 print_irqtrace_events(current
);
8367 EXPORT_SYMBOL(__might_sleep
);
8370 #ifdef CONFIG_MAGIC_SYSRQ
8371 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8374 update_rq_clock(rq
);
8375 on_rq
= p
->se
.on_rq
;
8377 deactivate_task(rq
, p
, 0);
8378 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8380 activate_task(rq
, p
, 0);
8381 resched_task(rq
->curr
);
8385 void normalize_rt_tasks(void)
8387 struct task_struct
*g
, *p
;
8388 unsigned long flags
;
8391 read_lock_irqsave(&tasklist_lock
, flags
);
8392 do_each_thread(g
, p
) {
8394 * Only normalize user tasks:
8399 p
->se
.exec_start
= 0;
8400 #ifdef CONFIG_SCHEDSTATS
8401 p
->se
.wait_start
= 0;
8402 p
->se
.sleep_start
= 0;
8403 p
->se
.block_start
= 0;
8405 task_rq(p
)->clock
= 0;
8409 * Renice negative nice level userspace
8412 if (TASK_NICE(p
) < 0 && p
->mm
)
8413 set_user_nice(p
, 0);
8417 spin_lock(&p
->pi_lock
);
8418 rq
= __task_rq_lock(p
);
8420 normalize_task(rq
, p
);
8422 __task_rq_unlock(rq
);
8423 spin_unlock(&p
->pi_lock
);
8424 } while_each_thread(g
, p
);
8426 read_unlock_irqrestore(&tasklist_lock
, flags
);
8429 #endif /* CONFIG_MAGIC_SYSRQ */
8433 * These functions are only useful for the IA64 MCA handling.
8435 * They can only be called when the whole system has been
8436 * stopped - every CPU needs to be quiescent, and no scheduling
8437 * activity can take place. Using them for anything else would
8438 * be a serious bug, and as a result, they aren't even visible
8439 * under any other configuration.
8443 * curr_task - return the current task for a given cpu.
8444 * @cpu: the processor in question.
8446 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8448 struct task_struct
*curr_task(int cpu
)
8450 return cpu_curr(cpu
);
8454 * set_curr_task - set the current task for a given cpu.
8455 * @cpu: the processor in question.
8456 * @p: the task pointer to set.
8458 * Description: This function must only be used when non-maskable interrupts
8459 * are serviced on a separate stack. It allows the architecture to switch the
8460 * notion of the current task on a cpu in a non-blocking manner. This function
8461 * must be called with all CPU's synchronized, and interrupts disabled, the
8462 * and caller must save the original value of the current task (see
8463 * curr_task() above) and restore that value before reenabling interrupts and
8464 * re-starting the system.
8466 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8468 void set_curr_task(int cpu
, struct task_struct
*p
)
8475 #ifdef CONFIG_FAIR_GROUP_SCHED
8476 static void free_fair_sched_group(struct task_group
*tg
)
8480 for_each_possible_cpu(i
) {
8482 kfree(tg
->cfs_rq
[i
]);
8492 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8494 struct cfs_rq
*cfs_rq
;
8495 struct sched_entity
*se
, *parent_se
;
8499 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8502 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8506 tg
->shares
= NICE_0_LOAD
;
8508 for_each_possible_cpu(i
) {
8511 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8512 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8516 se
= kmalloc_node(sizeof(struct sched_entity
),
8517 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8521 parent_se
= parent
? parent
->se
[i
] : NULL
;
8522 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8531 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8533 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8534 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8537 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8539 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8542 static inline void free_fair_sched_group(struct task_group
*tg
)
8547 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8552 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8556 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8561 #ifdef CONFIG_RT_GROUP_SCHED
8562 static void free_rt_sched_group(struct task_group
*tg
)
8566 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8568 for_each_possible_cpu(i
) {
8570 kfree(tg
->rt_rq
[i
]);
8572 kfree(tg
->rt_se
[i
]);
8580 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8582 struct rt_rq
*rt_rq
;
8583 struct sched_rt_entity
*rt_se
, *parent_se
;
8587 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8590 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8594 init_rt_bandwidth(&tg
->rt_bandwidth
,
8595 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8597 for_each_possible_cpu(i
) {
8600 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8601 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8605 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8606 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8610 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8611 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8620 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8622 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8623 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8626 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8628 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8631 static inline void free_rt_sched_group(struct task_group
*tg
)
8636 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8641 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8645 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8650 #ifdef CONFIG_GROUP_SCHED
8651 static void free_sched_group(struct task_group
*tg
)
8653 free_fair_sched_group(tg
);
8654 free_rt_sched_group(tg
);
8658 /* allocate runqueue etc for a new task group */
8659 struct task_group
*sched_create_group(struct task_group
*parent
)
8661 struct task_group
*tg
;
8662 unsigned long flags
;
8665 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8667 return ERR_PTR(-ENOMEM
);
8669 if (!alloc_fair_sched_group(tg
, parent
))
8672 if (!alloc_rt_sched_group(tg
, parent
))
8675 spin_lock_irqsave(&task_group_lock
, flags
);
8676 for_each_possible_cpu(i
) {
8677 register_fair_sched_group(tg
, i
);
8678 register_rt_sched_group(tg
, i
);
8680 list_add_rcu(&tg
->list
, &task_groups
);
8682 WARN_ON(!parent
); /* root should already exist */
8684 tg
->parent
= parent
;
8685 list_add_rcu(&tg
->siblings
, &parent
->children
);
8686 INIT_LIST_HEAD(&tg
->children
);
8687 spin_unlock_irqrestore(&task_group_lock
, flags
);
8692 free_sched_group(tg
);
8693 return ERR_PTR(-ENOMEM
);
8696 /* rcu callback to free various structures associated with a task group */
8697 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8699 /* now it should be safe to free those cfs_rqs */
8700 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8703 /* Destroy runqueue etc associated with a task group */
8704 void sched_destroy_group(struct task_group
*tg
)
8706 unsigned long flags
;
8709 spin_lock_irqsave(&task_group_lock
, flags
);
8710 for_each_possible_cpu(i
) {
8711 unregister_fair_sched_group(tg
, i
);
8712 unregister_rt_sched_group(tg
, i
);
8714 list_del_rcu(&tg
->list
);
8715 list_del_rcu(&tg
->siblings
);
8716 spin_unlock_irqrestore(&task_group_lock
, flags
);
8718 /* wait for possible concurrent references to cfs_rqs complete */
8719 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8722 /* change task's runqueue when it moves between groups.
8723 * The caller of this function should have put the task in its new group
8724 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8725 * reflect its new group.
8727 void sched_move_task(struct task_struct
*tsk
)
8730 unsigned long flags
;
8733 rq
= task_rq_lock(tsk
, &flags
);
8735 update_rq_clock(rq
);
8737 running
= task_current(rq
, tsk
);
8738 on_rq
= tsk
->se
.on_rq
;
8741 dequeue_task(rq
, tsk
, 0);
8742 if (unlikely(running
))
8743 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8745 set_task_rq(tsk
, task_cpu(tsk
));
8747 #ifdef CONFIG_FAIR_GROUP_SCHED
8748 if (tsk
->sched_class
->moved_group
)
8749 tsk
->sched_class
->moved_group(tsk
);
8752 if (unlikely(running
))
8753 tsk
->sched_class
->set_curr_task(rq
);
8755 enqueue_task(rq
, tsk
, 0);
8757 task_rq_unlock(rq
, &flags
);
8761 #ifdef CONFIG_FAIR_GROUP_SCHED
8762 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8764 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8769 dequeue_entity(cfs_rq
, se
, 0);
8771 se
->load
.weight
= shares
;
8772 se
->load
.inv_weight
= 0;
8775 enqueue_entity(cfs_rq
, se
, 0);
8778 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8780 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8781 struct rq
*rq
= cfs_rq
->rq
;
8782 unsigned long flags
;
8784 spin_lock_irqsave(&rq
->lock
, flags
);
8785 __set_se_shares(se
, shares
);
8786 spin_unlock_irqrestore(&rq
->lock
, flags
);
8789 static DEFINE_MUTEX(shares_mutex
);
8791 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8794 unsigned long flags
;
8797 * We can't change the weight of the root cgroup.
8802 if (shares
< MIN_SHARES
)
8803 shares
= MIN_SHARES
;
8804 else if (shares
> MAX_SHARES
)
8805 shares
= MAX_SHARES
;
8807 mutex_lock(&shares_mutex
);
8808 if (tg
->shares
== shares
)
8811 spin_lock_irqsave(&task_group_lock
, flags
);
8812 for_each_possible_cpu(i
)
8813 unregister_fair_sched_group(tg
, i
);
8814 list_del_rcu(&tg
->siblings
);
8815 spin_unlock_irqrestore(&task_group_lock
, flags
);
8817 /* wait for any ongoing reference to this group to finish */
8818 synchronize_sched();
8821 * Now we are free to modify the group's share on each cpu
8822 * w/o tripping rebalance_share or load_balance_fair.
8824 tg
->shares
= shares
;
8825 for_each_possible_cpu(i
) {
8829 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8830 set_se_shares(tg
->se
[i
], shares
);
8834 * Enable load balance activity on this group, by inserting it back on
8835 * each cpu's rq->leaf_cfs_rq_list.
8837 spin_lock_irqsave(&task_group_lock
, flags
);
8838 for_each_possible_cpu(i
)
8839 register_fair_sched_group(tg
, i
);
8840 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8841 spin_unlock_irqrestore(&task_group_lock
, flags
);
8843 mutex_unlock(&shares_mutex
);
8847 unsigned long sched_group_shares(struct task_group
*tg
)
8853 #ifdef CONFIG_RT_GROUP_SCHED
8855 * Ensure that the real time constraints are schedulable.
8857 static DEFINE_MUTEX(rt_constraints_mutex
);
8859 static unsigned long to_ratio(u64 period
, u64 runtime
)
8861 if (runtime
== RUNTIME_INF
)
8864 return div64_u64(runtime
<< 16, period
);
8867 #ifdef CONFIG_CGROUP_SCHED
8868 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8870 struct task_group
*tgi
, *parent
= tg
->parent
;
8871 unsigned long total
= 0;
8874 if (global_rt_period() < period
)
8877 return to_ratio(period
, runtime
) <
8878 to_ratio(global_rt_period(), global_rt_runtime());
8881 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8885 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8889 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8890 tgi
->rt_bandwidth
.rt_runtime
);
8894 return total
+ to_ratio(period
, runtime
) <
8895 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8896 parent
->rt_bandwidth
.rt_runtime
);
8898 #elif defined CONFIG_USER_SCHED
8899 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8901 struct task_group
*tgi
;
8902 unsigned long total
= 0;
8903 unsigned long global_ratio
=
8904 to_ratio(global_rt_period(), global_rt_runtime());
8907 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8911 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8912 tgi
->rt_bandwidth
.rt_runtime
);
8916 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8920 /* Must be called with tasklist_lock held */
8921 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8923 struct task_struct
*g
, *p
;
8924 do_each_thread(g
, p
) {
8925 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8927 } while_each_thread(g
, p
);
8931 static int tg_set_bandwidth(struct task_group
*tg
,
8932 u64 rt_period
, u64 rt_runtime
)
8936 mutex_lock(&rt_constraints_mutex
);
8937 read_lock(&tasklist_lock
);
8938 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8942 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8947 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8948 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8949 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8951 for_each_possible_cpu(i
) {
8952 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8954 spin_lock(&rt_rq
->rt_runtime_lock
);
8955 rt_rq
->rt_runtime
= rt_runtime
;
8956 spin_unlock(&rt_rq
->rt_runtime_lock
);
8958 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8960 read_unlock(&tasklist_lock
);
8961 mutex_unlock(&rt_constraints_mutex
);
8966 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8968 u64 rt_runtime
, rt_period
;
8970 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8971 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8972 if (rt_runtime_us
< 0)
8973 rt_runtime
= RUNTIME_INF
;
8975 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8978 long sched_group_rt_runtime(struct task_group
*tg
)
8982 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8985 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8986 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8987 return rt_runtime_us
;
8990 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8992 u64 rt_runtime
, rt_period
;
8994 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8995 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8997 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9000 long sched_group_rt_period(struct task_group
*tg
)
9004 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9005 do_div(rt_period_us
, NSEC_PER_USEC
);
9006 return rt_period_us
;
9009 static int sched_rt_global_constraints(void)
9013 mutex_lock(&rt_constraints_mutex
);
9014 if (!__rt_schedulable(NULL
, 1, 0))
9016 mutex_unlock(&rt_constraints_mutex
);
9021 static int sched_rt_global_constraints(void)
9023 unsigned long flags
;
9026 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9027 for_each_possible_cpu(i
) {
9028 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9030 spin_lock(&rt_rq
->rt_runtime_lock
);
9031 rt_rq
->rt_runtime
= global_rt_runtime();
9032 spin_unlock(&rt_rq
->rt_runtime_lock
);
9034 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9040 int sched_rt_handler(struct ctl_table
*table
, int write
,
9041 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9045 int old_period
, old_runtime
;
9046 static DEFINE_MUTEX(mutex
);
9049 old_period
= sysctl_sched_rt_period
;
9050 old_runtime
= sysctl_sched_rt_runtime
;
9052 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9054 if (!ret
&& write
) {
9055 ret
= sched_rt_global_constraints();
9057 sysctl_sched_rt_period
= old_period
;
9058 sysctl_sched_rt_runtime
= old_runtime
;
9060 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9061 def_rt_bandwidth
.rt_period
=
9062 ns_to_ktime(global_rt_period());
9065 mutex_unlock(&mutex
);
9070 #ifdef CONFIG_CGROUP_SCHED
9072 /* return corresponding task_group object of a cgroup */
9073 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9075 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9076 struct task_group
, css
);
9079 static struct cgroup_subsys_state
*
9080 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9082 struct task_group
*tg
, *parent
;
9084 if (!cgrp
->parent
) {
9085 /* This is early initialization for the top cgroup */
9086 init_task_group
.css
.cgroup
= cgrp
;
9087 return &init_task_group
.css
;
9090 parent
= cgroup_tg(cgrp
->parent
);
9091 tg
= sched_create_group(parent
);
9093 return ERR_PTR(-ENOMEM
);
9095 /* Bind the cgroup to task_group object we just created */
9096 tg
->css
.cgroup
= cgrp
;
9102 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9104 struct task_group
*tg
= cgroup_tg(cgrp
);
9106 sched_destroy_group(tg
);
9110 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9111 struct task_struct
*tsk
)
9113 #ifdef CONFIG_RT_GROUP_SCHED
9114 /* Don't accept realtime tasks when there is no way for them to run */
9115 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9118 /* We don't support RT-tasks being in separate groups */
9119 if (tsk
->sched_class
!= &fair_sched_class
)
9127 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9128 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9130 sched_move_task(tsk
);
9133 #ifdef CONFIG_FAIR_GROUP_SCHED
9134 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9137 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9140 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9142 struct task_group
*tg
= cgroup_tg(cgrp
);
9144 return (u64
) tg
->shares
;
9148 #ifdef CONFIG_RT_GROUP_SCHED
9149 static ssize_t
cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9152 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9155 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9157 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9160 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9163 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9166 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9168 return sched_group_rt_period(cgroup_tg(cgrp
));
9172 static struct cftype cpu_files
[] = {
9173 #ifdef CONFIG_FAIR_GROUP_SCHED
9176 .read_u64
= cpu_shares_read_u64
,
9177 .write_u64
= cpu_shares_write_u64
,
9180 #ifdef CONFIG_RT_GROUP_SCHED
9182 .name
= "rt_runtime_us",
9183 .read_s64
= cpu_rt_runtime_read
,
9184 .write_s64
= cpu_rt_runtime_write
,
9187 .name
= "rt_period_us",
9188 .read_u64
= cpu_rt_period_read_uint
,
9189 .write_u64
= cpu_rt_period_write_uint
,
9194 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9196 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9199 struct cgroup_subsys cpu_cgroup_subsys
= {
9201 .create
= cpu_cgroup_create
,
9202 .destroy
= cpu_cgroup_destroy
,
9203 .can_attach
= cpu_cgroup_can_attach
,
9204 .attach
= cpu_cgroup_attach
,
9205 .populate
= cpu_cgroup_populate
,
9206 .subsys_id
= cpu_cgroup_subsys_id
,
9210 #endif /* CONFIG_CGROUP_SCHED */
9212 #ifdef CONFIG_CGROUP_CPUACCT
9215 * CPU accounting code for task groups.
9217 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9218 * (balbir@in.ibm.com).
9221 /* track cpu usage of a group of tasks */
9223 struct cgroup_subsys_state css
;
9224 /* cpuusage holds pointer to a u64-type object on every cpu */
9228 struct cgroup_subsys cpuacct_subsys
;
9230 /* return cpu accounting group corresponding to this container */
9231 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9233 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9234 struct cpuacct
, css
);
9237 /* return cpu accounting group to which this task belongs */
9238 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9240 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9241 struct cpuacct
, css
);
9244 /* create a new cpu accounting group */
9245 static struct cgroup_subsys_state
*cpuacct_create(
9246 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9248 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9251 return ERR_PTR(-ENOMEM
);
9253 ca
->cpuusage
= alloc_percpu(u64
);
9254 if (!ca
->cpuusage
) {
9256 return ERR_PTR(-ENOMEM
);
9262 /* destroy an existing cpu accounting group */
9264 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9266 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9268 free_percpu(ca
->cpuusage
);
9272 /* return total cpu usage (in nanoseconds) of a group */
9273 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9275 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9276 u64 totalcpuusage
= 0;
9279 for_each_possible_cpu(i
) {
9280 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9283 * Take rq->lock to make 64-bit addition safe on 32-bit
9286 spin_lock_irq(&cpu_rq(i
)->lock
);
9287 totalcpuusage
+= *cpuusage
;
9288 spin_unlock_irq(&cpu_rq(i
)->lock
);
9291 return totalcpuusage
;
9294 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9297 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9306 for_each_possible_cpu(i
) {
9307 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9309 spin_lock_irq(&cpu_rq(i
)->lock
);
9311 spin_unlock_irq(&cpu_rq(i
)->lock
);
9317 static struct cftype files
[] = {
9320 .read_u64
= cpuusage_read
,
9321 .write_u64
= cpuusage_write
,
9325 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9327 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9331 * charge this task's execution time to its accounting group.
9333 * called with rq->lock held.
9335 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9339 if (!cpuacct_subsys
.active
)
9344 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9346 *cpuusage
+= cputime
;
9350 struct cgroup_subsys cpuacct_subsys
= {
9352 .create
= cpuacct_create
,
9353 .destroy
= cpuacct_destroy
,
9354 .populate
= cpuacct_populate
,
9355 .subsys_id
= cpuacct_subsys_id
,
9357 #endif /* CONFIG_CGROUP_CPUACCT */