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>
73 #include <linux/ftrace.h>
76 #include <asm/irq_regs.h>
79 * Convert user-nice values [ -20 ... 0 ... 19 ]
80 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
83 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
84 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
85 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
88 * 'User priority' is the nice value converted to something we
89 * can work with better when scaling various scheduler parameters,
90 * it's a [ 0 ... 39 ] range.
92 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
93 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
94 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
97 * Helpers for converting nanosecond timing to jiffy resolution
99 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
101 #define NICE_0_LOAD SCHED_LOAD_SCALE
102 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
105 * These are the 'tuning knobs' of the scheduler:
107 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
108 * Timeslices get refilled after they expire.
110 #define DEF_TIMESLICE (100 * HZ / 1000)
113 * single value that denotes runtime == period, ie unlimited time.
115 #define RUNTIME_INF ((u64)~0ULL)
119 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
120 * Since cpu_power is a 'constant', we can use a reciprocal divide.
122 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
124 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
128 * Each time a sched group cpu_power is changed,
129 * we must compute its reciprocal value
131 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
133 sg
->__cpu_power
+= val
;
134 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
138 static inline int rt_policy(int policy
)
140 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
145 static inline int task_has_rt_policy(struct task_struct
*p
)
147 return rt_policy(p
->policy
);
151 * This is the priority-queue data structure of the RT scheduling class:
153 struct rt_prio_array
{
154 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
155 struct list_head queue
[MAX_RT_PRIO
];
158 struct rt_bandwidth
{
159 /* nests inside the rq lock: */
160 spinlock_t rt_runtime_lock
;
163 struct hrtimer rt_period_timer
;
166 static struct rt_bandwidth def_rt_bandwidth
;
168 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
170 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
172 struct rt_bandwidth
*rt_b
=
173 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
179 now
= hrtimer_cb_get_time(timer
);
180 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
185 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
188 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
192 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
194 rt_b
->rt_period
= ns_to_ktime(period
);
195 rt_b
->rt_runtime
= runtime
;
197 spin_lock_init(&rt_b
->rt_runtime_lock
);
199 hrtimer_init(&rt_b
->rt_period_timer
,
200 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
201 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
202 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
205 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
209 if (rt_b
->rt_runtime
== RUNTIME_INF
)
212 if (hrtimer_active(&rt_b
->rt_period_timer
))
215 spin_lock(&rt_b
->rt_runtime_lock
);
217 if (hrtimer_active(&rt_b
->rt_period_timer
))
220 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
221 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
222 hrtimer_start(&rt_b
->rt_period_timer
,
223 rt_b
->rt_period_timer
.expires
,
226 spin_unlock(&rt_b
->rt_runtime_lock
);
229 #ifdef CONFIG_RT_GROUP_SCHED
230 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
232 hrtimer_cancel(&rt_b
->rt_period_timer
);
237 * sched_domains_mutex serializes calls to arch_init_sched_domains,
238 * detach_destroy_domains and partition_sched_domains.
240 static DEFINE_MUTEX(sched_domains_mutex
);
242 #ifdef CONFIG_GROUP_SCHED
244 #include <linux/cgroup.h>
248 static LIST_HEAD(task_groups
);
250 /* task group related information */
252 #ifdef CONFIG_CGROUP_SCHED
253 struct cgroup_subsys_state css
;
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity
**se
;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq
**cfs_rq
;
261 unsigned long shares
;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity
**rt_se
;
266 struct rt_rq
**rt_rq
;
268 struct rt_bandwidth rt_bandwidth
;
272 struct list_head list
;
274 struct task_group
*parent
;
275 struct list_head siblings
;
276 struct list_head children
;
279 #ifdef CONFIG_USER_SCHED
283 * Every UID task group (including init_task_group aka UID-0) will
284 * be a child to this group.
286 struct task_group root_task_group
;
288 #ifdef CONFIG_FAIR_GROUP_SCHED
289 /* Default task group's sched entity on each cpu */
290 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
291 /* Default task group's cfs_rq on each cpu */
292 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
295 #ifdef CONFIG_RT_GROUP_SCHED
296 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
297 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
300 #define root_task_group init_task_group
303 /* task_group_lock serializes add/remove of task groups and also changes to
304 * a task group's cpu shares.
306 static DEFINE_SPINLOCK(task_group_lock
);
308 #ifdef CONFIG_FAIR_GROUP_SCHED
309 #ifdef CONFIG_USER_SCHED
310 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
312 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
316 * A weight of 0 or 1 can cause arithmetics problems.
317 * A weight of a cfs_rq is the sum of weights of which entities
318 * are queued on this cfs_rq, so a weight of a entity should not be
319 * too large, so as the shares value of a task group.
320 * (The default weight is 1024 - so there's no practical
321 * limitation from this.)
324 #define MAX_SHARES (1UL << 18)
326 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
329 /* Default task group.
330 * Every task in system belong to this group at bootup.
332 struct task_group init_task_group
;
334 /* return group to which a task belongs */
335 static inline struct task_group
*task_group(struct task_struct
*p
)
337 struct task_group
*tg
;
339 #ifdef CONFIG_USER_SCHED
341 #elif defined(CONFIG_CGROUP_SCHED)
342 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
343 struct task_group
, css
);
345 tg
= &init_task_group
;
350 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
351 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
353 #ifdef CONFIG_FAIR_GROUP_SCHED
354 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
355 p
->se
.parent
= task_group(p
)->se
[cpu
];
358 #ifdef CONFIG_RT_GROUP_SCHED
359 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
360 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
366 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
368 #endif /* CONFIG_GROUP_SCHED */
370 /* CFS-related fields in a runqueue */
372 struct load_weight load
;
373 unsigned long nr_running
;
378 struct rb_root tasks_timeline
;
379 struct rb_node
*rb_leftmost
;
381 struct list_head tasks
;
382 struct list_head
*balance_iterator
;
385 * 'curr' points to currently running entity on this cfs_rq.
386 * It is set to NULL otherwise (i.e when none are currently running).
388 struct sched_entity
*curr
, *next
;
390 unsigned long nr_spread_over
;
392 #ifdef CONFIG_FAIR_GROUP_SCHED
393 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
396 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
397 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
398 * (like users, containers etc.)
400 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
401 * list is used during load balance.
403 struct list_head leaf_cfs_rq_list
;
404 struct task_group
*tg
; /* group that "owns" this runqueue */
408 /* Real-Time classes' related field in a runqueue: */
410 struct rt_prio_array active
;
411 unsigned long rt_nr_running
;
412 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
413 int highest_prio
; /* highest queued rt task prio */
416 unsigned long rt_nr_migratory
;
422 /* Nests inside the rq lock: */
423 spinlock_t rt_runtime_lock
;
425 #ifdef CONFIG_RT_GROUP_SCHED
426 unsigned long rt_nr_boosted
;
429 struct list_head leaf_rt_rq_list
;
430 struct task_group
*tg
;
431 struct sched_rt_entity
*rt_se
;
438 * We add the notion of a root-domain which will be used to define per-domain
439 * variables. Each exclusive cpuset essentially defines an island domain by
440 * fully partitioning the member cpus from any other cpuset. Whenever a new
441 * exclusive cpuset is created, we also create and attach a new root-domain
451 * The "RT overload" flag: it gets set if a CPU has more than
452 * one runnable RT task.
459 * By default the system creates a single root-domain with all cpus as
460 * members (mimicking the global state we have today).
462 static struct root_domain def_root_domain
;
467 * This is the main, per-CPU runqueue data structure.
469 * Locking rule: those places that want to lock multiple runqueues
470 * (such as the load balancing or the thread migration code), lock
471 * acquire operations must be ordered by ascending &runqueue.
478 * nr_running and cpu_load should be in the same cacheline because
479 * remote CPUs use both these fields when doing load calculation.
481 unsigned long nr_running
;
482 #define CPU_LOAD_IDX_MAX 5
483 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
484 unsigned char idle_at_tick
;
486 unsigned long last_tick_seen
;
487 unsigned char in_nohz_recently
;
489 /* capture load from *all* tasks on this cpu: */
490 struct load_weight load
;
491 unsigned long nr_load_updates
;
497 #ifdef CONFIG_FAIR_GROUP_SCHED
498 /* list of leaf cfs_rq on this cpu: */
499 struct list_head leaf_cfs_rq_list
;
501 #ifdef CONFIG_RT_GROUP_SCHED
502 struct list_head leaf_rt_rq_list
;
506 * This is part of a global counter where only the total sum
507 * over all CPUs matters. A task can increase this counter on
508 * one CPU and if it got migrated afterwards it may decrease
509 * it on another CPU. Always updated under the runqueue lock:
511 unsigned long nr_uninterruptible
;
513 struct task_struct
*curr
, *idle
;
514 unsigned long next_balance
;
515 struct mm_struct
*prev_mm
;
522 struct root_domain
*rd
;
523 struct sched_domain
*sd
;
525 /* For active balancing */
528 /* cpu of this runqueue: */
531 struct task_struct
*migration_thread
;
532 struct list_head migration_queue
;
535 #ifdef CONFIG_SCHED_HRTICK
536 unsigned long hrtick_flags
;
537 ktime_t hrtick_expire
;
538 struct hrtimer hrtick_timer
;
541 #ifdef CONFIG_SCHEDSTATS
543 struct sched_info rq_sched_info
;
545 /* sys_sched_yield() stats */
546 unsigned int yld_exp_empty
;
547 unsigned int yld_act_empty
;
548 unsigned int yld_both_empty
;
549 unsigned int yld_count
;
551 /* schedule() stats */
552 unsigned int sched_switch
;
553 unsigned int sched_count
;
554 unsigned int sched_goidle
;
556 /* try_to_wake_up() stats */
557 unsigned int ttwu_count
;
558 unsigned int ttwu_local
;
561 unsigned int bkl_count
;
563 struct lock_class_key rq_lock_key
;
566 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
568 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
)
570 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
);
573 static inline int cpu_of(struct rq
*rq
)
583 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
584 * See detach_destroy_domains: synchronize_sched for details.
586 * The domain tree of any CPU may only be accessed from within
587 * preempt-disabled sections.
589 #define for_each_domain(cpu, __sd) \
590 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
592 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
593 #define this_rq() (&__get_cpu_var(runqueues))
594 #define task_rq(p) cpu_rq(task_cpu(p))
595 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
597 static inline void update_rq_clock(struct rq
*rq
)
599 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
603 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
605 #ifdef CONFIG_SCHED_DEBUG
606 # define const_debug __read_mostly
608 # define const_debug static const
614 * Returns true if the current cpu runqueue is locked.
615 * This interface allows printk to be called with the runqueue lock
616 * held and know whether or not it is OK to wake up the klogd.
618 int runqueue_is_locked(void)
621 struct rq
*rq
= cpu_rq(cpu
);
624 ret
= spin_is_locked(&rq
->lock
);
630 * Debugging: various feature bits
633 #define SCHED_FEAT(name, enabled) \
634 __SCHED_FEAT_##name ,
637 #include "sched_features.h"
642 #define SCHED_FEAT(name, enabled) \
643 (1UL << __SCHED_FEAT_##name) * enabled |
645 const_debug
unsigned int sysctl_sched_features
=
646 #include "sched_features.h"
651 #ifdef CONFIG_SCHED_DEBUG
652 #define SCHED_FEAT(name, enabled) \
655 static __read_mostly
char *sched_feat_names
[] = {
656 #include "sched_features.h"
662 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
664 filp
->private_data
= inode
->i_private
;
669 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
670 size_t cnt
, loff_t
*ppos
)
677 for (i
= 0; sched_feat_names
[i
]; i
++) {
678 len
+= strlen(sched_feat_names
[i
]);
682 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
686 for (i
= 0; sched_feat_names
[i
]; i
++) {
687 if (sysctl_sched_features
& (1UL << i
))
688 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
690 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
693 r
+= sprintf(buf
+ r
, "\n");
694 WARN_ON(r
>= len
+ 2);
696 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
704 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
705 size_t cnt
, loff_t
*ppos
)
715 if (copy_from_user(&buf
, ubuf
, cnt
))
720 if (strncmp(buf
, "NO_", 3) == 0) {
725 for (i
= 0; sched_feat_names
[i
]; i
++) {
726 int len
= strlen(sched_feat_names
[i
]);
728 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
730 sysctl_sched_features
&= ~(1UL << i
);
732 sysctl_sched_features
|= (1UL << i
);
737 if (!sched_feat_names
[i
])
745 static struct file_operations sched_feat_fops
= {
746 .open
= sched_feat_open
,
747 .read
= sched_feat_read
,
748 .write
= sched_feat_write
,
751 static __init
int sched_init_debug(void)
753 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
758 late_initcall(sched_init_debug
);
762 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
765 * Number of tasks to iterate in a single balance run.
766 * Limited because this is done with IRQs disabled.
768 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
771 * period over which we measure -rt task cpu usage in us.
774 unsigned int sysctl_sched_rt_period
= 1000000;
776 static __read_mostly
int scheduler_running
;
779 * part of the period that we allow rt tasks to run in us.
782 int sysctl_sched_rt_runtime
= 950000;
784 static inline u64
global_rt_period(void)
786 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
789 static inline u64
global_rt_runtime(void)
791 if (sysctl_sched_rt_period
< 0)
794 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
797 unsigned long long time_sync_thresh
= 100000;
799 static DEFINE_PER_CPU(unsigned long long, time_offset
);
800 static DEFINE_PER_CPU(unsigned long long, prev_cpu_time
);
803 * Global lock which we take every now and then to synchronize
804 * the CPUs time. This method is not warp-safe, but it's good
805 * enough to synchronize slowly diverging time sources and thus
806 * it's good enough for tracing:
808 static DEFINE_SPINLOCK(time_sync_lock
);
809 static unsigned long long prev_global_time
;
811 static unsigned long long __sync_cpu_clock(unsigned long long time
, int cpu
)
814 * We want this inlined, to not get tracer function calls
815 * in this critical section:
817 spin_acquire(&time_sync_lock
.dep_map
, 0, 0, _THIS_IP_
);
818 __raw_spin_lock(&time_sync_lock
.raw_lock
);
820 if (time
< prev_global_time
) {
821 per_cpu(time_offset
, cpu
) += prev_global_time
- time
;
822 time
= prev_global_time
;
824 prev_global_time
= time
;
827 __raw_spin_unlock(&time_sync_lock
.raw_lock
);
828 spin_release(&time_sync_lock
.dep_map
, 1, _THIS_IP_
);
833 static unsigned long long __cpu_clock(int cpu
)
835 unsigned long long now
;
838 * Only call sched_clock() if the scheduler has already been
839 * initialized (some code might call cpu_clock() very early):
841 if (unlikely(!scheduler_running
))
844 now
= sched_clock_cpu(cpu
);
850 * For kernel-internal use: high-speed (but slightly incorrect) per-cpu
851 * clock constructed from sched_clock():
853 unsigned long long notrace
cpu_clock(int cpu
)
855 unsigned long long prev_cpu_time
, time
, delta_time
;
858 local_irq_save(flags
);
859 prev_cpu_time
= per_cpu(prev_cpu_time
, cpu
);
860 time
= __cpu_clock(cpu
) + per_cpu(time_offset
, cpu
);
861 delta_time
= time
-prev_cpu_time
;
863 if (unlikely(delta_time
> time_sync_thresh
)) {
864 time
= __sync_cpu_clock(time
, cpu
);
865 per_cpu(prev_cpu_time
, cpu
) = time
;
867 local_irq_restore(flags
);
871 EXPORT_SYMBOL_GPL(cpu_clock
);
873 #ifndef prepare_arch_switch
874 # define prepare_arch_switch(next) do { } while (0)
876 #ifndef finish_arch_switch
877 # define finish_arch_switch(prev) do { } while (0)
880 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
882 return rq
->curr
== p
;
885 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
886 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
888 return task_current(rq
, p
);
891 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
895 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
897 #ifdef CONFIG_DEBUG_SPINLOCK
898 /* this is a valid case when another task releases the spinlock */
899 rq
->lock
.owner
= current
;
902 * If we are tracking spinlock dependencies then we have to
903 * fix up the runqueue lock - which gets 'carried over' from
906 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
908 spin_unlock_irq(&rq
->lock
);
911 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
912 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
917 return task_current(rq
, p
);
921 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
925 * We can optimise this out completely for !SMP, because the
926 * SMP rebalancing from interrupt is the only thing that cares
931 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
932 spin_unlock_irq(&rq
->lock
);
934 spin_unlock(&rq
->lock
);
938 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
942 * After ->oncpu is cleared, the task can be moved to a different CPU.
943 * We must ensure this doesn't happen until the switch is completely
949 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
953 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
956 * __task_rq_lock - lock the runqueue a given task resides on.
957 * Must be called interrupts disabled.
959 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
963 struct rq
*rq
= task_rq(p
);
964 spin_lock(&rq
->lock
);
965 if (likely(rq
== task_rq(p
)))
967 spin_unlock(&rq
->lock
);
972 * task_rq_lock - lock the runqueue a given task resides on and disable
973 * interrupts. Note the ordering: we can safely lookup the task_rq without
974 * explicitly disabling preemption.
976 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
982 local_irq_save(*flags
);
984 spin_lock(&rq
->lock
);
985 if (likely(rq
== task_rq(p
)))
987 spin_unlock_irqrestore(&rq
->lock
, *flags
);
991 static void __task_rq_unlock(struct rq
*rq
)
994 spin_unlock(&rq
->lock
);
997 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
1000 spin_unlock_irqrestore(&rq
->lock
, *flags
);
1004 * this_rq_lock - lock this runqueue and disable interrupts.
1006 static struct rq
*this_rq_lock(void)
1007 __acquires(rq
->lock
)
1011 local_irq_disable();
1013 spin_lock(&rq
->lock
);
1018 static void __resched_task(struct task_struct
*p
, int tif_bit
);
1020 static inline void resched_task(struct task_struct
*p
)
1022 __resched_task(p
, TIF_NEED_RESCHED
);
1025 #ifdef CONFIG_SCHED_HRTICK
1027 * Use HR-timers to deliver accurate preemption points.
1029 * Its all a bit involved since we cannot program an hrt while holding the
1030 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1033 * When we get rescheduled we reprogram the hrtick_timer outside of the
1036 static inline void resched_hrt(struct task_struct
*p
)
1038 __resched_task(p
, TIF_HRTICK_RESCHED
);
1041 static inline void resched_rq(struct rq
*rq
)
1043 unsigned long flags
;
1045 spin_lock_irqsave(&rq
->lock
, flags
);
1046 resched_task(rq
->curr
);
1047 spin_unlock_irqrestore(&rq
->lock
, flags
);
1051 HRTICK_SET
, /* re-programm hrtick_timer */
1052 HRTICK_RESET
, /* not a new slice */
1053 HRTICK_BLOCK
, /* stop hrtick operations */
1058 * - enabled by features
1059 * - hrtimer is actually high res
1061 static inline int hrtick_enabled(struct rq
*rq
)
1063 if (!sched_feat(HRTICK
))
1065 if (unlikely(test_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
)))
1067 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1071 * Called to set the hrtick timer state.
1073 * called with rq->lock held and irqs disabled
1075 static void hrtick_start(struct rq
*rq
, u64 delay
, int reset
)
1077 assert_spin_locked(&rq
->lock
);
1080 * preempt at: now + delay
1083 ktime_add_ns(rq
->hrtick_timer
.base
->get_time(), delay
);
1085 * indicate we need to program the timer
1087 __set_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1089 __set_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1092 * New slices are called from the schedule path and don't need a
1093 * forced reschedule.
1096 resched_hrt(rq
->curr
);
1099 static void hrtick_clear(struct rq
*rq
)
1101 if (hrtimer_active(&rq
->hrtick_timer
))
1102 hrtimer_cancel(&rq
->hrtick_timer
);
1106 * Update the timer from the possible pending state.
1108 static void hrtick_set(struct rq
*rq
)
1112 unsigned long flags
;
1114 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1116 spin_lock_irqsave(&rq
->lock
, flags
);
1117 set
= __test_and_clear_bit(HRTICK_SET
, &rq
->hrtick_flags
);
1118 reset
= __test_and_clear_bit(HRTICK_RESET
, &rq
->hrtick_flags
);
1119 time
= rq
->hrtick_expire
;
1120 clear_thread_flag(TIF_HRTICK_RESCHED
);
1121 spin_unlock_irqrestore(&rq
->lock
, flags
);
1124 hrtimer_start(&rq
->hrtick_timer
, time
, HRTIMER_MODE_ABS
);
1125 if (reset
&& !hrtimer_active(&rq
->hrtick_timer
))
1132 * High-resolution timer tick.
1133 * Runs from hardirq context with interrupts disabled.
1135 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1137 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1139 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1141 spin_lock(&rq
->lock
);
1142 update_rq_clock(rq
);
1143 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1144 spin_unlock(&rq
->lock
);
1146 return HRTIMER_NORESTART
;
1150 static void hotplug_hrtick_disable(int cpu
)
1152 struct rq
*rq
= cpu_rq(cpu
);
1153 unsigned long flags
;
1155 spin_lock_irqsave(&rq
->lock
, flags
);
1156 rq
->hrtick_flags
= 0;
1157 __set_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1158 spin_unlock_irqrestore(&rq
->lock
, flags
);
1163 static void hotplug_hrtick_enable(int cpu
)
1165 struct rq
*rq
= cpu_rq(cpu
);
1166 unsigned long flags
;
1168 spin_lock_irqsave(&rq
->lock
, flags
);
1169 __clear_bit(HRTICK_BLOCK
, &rq
->hrtick_flags
);
1170 spin_unlock_irqrestore(&rq
->lock
, flags
);
1174 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1176 int cpu
= (int)(long)hcpu
;
1179 case CPU_UP_CANCELED
:
1180 case CPU_UP_CANCELED_FROZEN
:
1181 case CPU_DOWN_PREPARE
:
1182 case CPU_DOWN_PREPARE_FROZEN
:
1184 case CPU_DEAD_FROZEN
:
1185 hotplug_hrtick_disable(cpu
);
1188 case CPU_UP_PREPARE
:
1189 case CPU_UP_PREPARE_FROZEN
:
1190 case CPU_DOWN_FAILED
:
1191 case CPU_DOWN_FAILED_FROZEN
:
1193 case CPU_ONLINE_FROZEN
:
1194 hotplug_hrtick_enable(cpu
);
1201 static void init_hrtick(void)
1203 hotcpu_notifier(hotplug_hrtick
, 0);
1205 #endif /* CONFIG_SMP */
1207 static void init_rq_hrtick(struct rq
*rq
)
1209 rq
->hrtick_flags
= 0;
1210 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1211 rq
->hrtick_timer
.function
= hrtick
;
1212 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_NO_SOFTIRQ
;
1215 void hrtick_resched(void)
1218 unsigned long flags
;
1220 if (!test_thread_flag(TIF_HRTICK_RESCHED
))
1223 local_irq_save(flags
);
1224 rq
= cpu_rq(smp_processor_id());
1226 local_irq_restore(flags
);
1229 static inline void hrtick_clear(struct rq
*rq
)
1233 static inline void hrtick_set(struct rq
*rq
)
1237 static inline void init_rq_hrtick(struct rq
*rq
)
1241 void hrtick_resched(void)
1245 static inline void init_hrtick(void)
1251 * resched_task - mark a task 'to be rescheduled now'.
1253 * On UP this means the setting of the need_resched flag, on SMP it
1254 * might also involve a cross-CPU call to trigger the scheduler on
1259 #ifndef tsk_is_polling
1260 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1263 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1267 assert_spin_locked(&task_rq(p
)->lock
);
1269 if (unlikely(test_tsk_thread_flag(p
, tif_bit
)))
1272 set_tsk_thread_flag(p
, tif_bit
);
1275 if (cpu
== smp_processor_id())
1278 /* NEED_RESCHED must be visible before we test polling */
1280 if (!tsk_is_polling(p
))
1281 smp_send_reschedule(cpu
);
1284 static void resched_cpu(int cpu
)
1286 struct rq
*rq
= cpu_rq(cpu
);
1287 unsigned long flags
;
1289 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1291 resched_task(cpu_curr(cpu
));
1292 spin_unlock_irqrestore(&rq
->lock
, flags
);
1297 * When add_timer_on() enqueues a timer into the timer wheel of an
1298 * idle CPU then this timer might expire before the next timer event
1299 * which is scheduled to wake up that CPU. In case of a completely
1300 * idle system the next event might even be infinite time into the
1301 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1302 * leaves the inner idle loop so the newly added timer is taken into
1303 * account when the CPU goes back to idle and evaluates the timer
1304 * wheel for the next timer event.
1306 void wake_up_idle_cpu(int cpu
)
1308 struct rq
*rq
= cpu_rq(cpu
);
1310 if (cpu
== smp_processor_id())
1314 * This is safe, as this function is called with the timer
1315 * wheel base lock of (cpu) held. When the CPU is on the way
1316 * to idle and has not yet set rq->curr to idle then it will
1317 * be serialized on the timer wheel base lock and take the new
1318 * timer into account automatically.
1320 if (rq
->curr
!= rq
->idle
)
1324 * We can set TIF_RESCHED on the idle task of the other CPU
1325 * lockless. The worst case is that the other CPU runs the
1326 * idle task through an additional NOOP schedule()
1328 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1330 /* NEED_RESCHED must be visible before we test polling */
1332 if (!tsk_is_polling(rq
->idle
))
1333 smp_send_reschedule(cpu
);
1338 static void __resched_task(struct task_struct
*p
, int tif_bit
)
1340 assert_spin_locked(&task_rq(p
)->lock
);
1341 set_tsk_thread_flag(p
, tif_bit
);
1345 #if BITS_PER_LONG == 32
1346 # define WMULT_CONST (~0UL)
1348 # define WMULT_CONST (1UL << 32)
1351 #define WMULT_SHIFT 32
1354 * Shift right and round:
1356 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1358 static unsigned long
1359 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1360 struct load_weight
*lw
)
1364 if (!lw
->inv_weight
) {
1365 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1368 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1372 tmp
= (u64
)delta_exec
* weight
;
1374 * Check whether we'd overflow the 64-bit multiplication:
1376 if (unlikely(tmp
> WMULT_CONST
))
1377 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1380 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1382 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1385 static inline unsigned long
1386 calc_delta_fair(unsigned long delta_exec
, struct load_weight
*lw
)
1388 return calc_delta_mine(delta_exec
, NICE_0_LOAD
, lw
);
1391 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1397 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1404 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1405 * of tasks with abnormal "nice" values across CPUs the contribution that
1406 * each task makes to its run queue's load is weighted according to its
1407 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1408 * scaled version of the new time slice allocation that they receive on time
1412 #define WEIGHT_IDLEPRIO 2
1413 #define WMULT_IDLEPRIO (1 << 31)
1416 * Nice levels are multiplicative, with a gentle 10% change for every
1417 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1418 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1419 * that remained on nice 0.
1421 * The "10% effect" is relative and cumulative: from _any_ nice level,
1422 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1423 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1424 * If a task goes up by ~10% and another task goes down by ~10% then
1425 * the relative distance between them is ~25%.)
1427 static const int prio_to_weight
[40] = {
1428 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1429 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1430 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1431 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1432 /* 0 */ 1024, 820, 655, 526, 423,
1433 /* 5 */ 335, 272, 215, 172, 137,
1434 /* 10 */ 110, 87, 70, 56, 45,
1435 /* 15 */ 36, 29, 23, 18, 15,
1439 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1441 * In cases where the weight does not change often, we can use the
1442 * precalculated inverse to speed up arithmetics by turning divisions
1443 * into multiplications:
1445 static const u32 prio_to_wmult
[40] = {
1446 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1447 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1448 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1449 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1450 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1451 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1452 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1453 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1456 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1459 * runqueue iterator, to support SMP load-balancing between different
1460 * scheduling classes, without having to expose their internal data
1461 * structures to the load-balancing proper:
1463 struct rq_iterator
{
1465 struct task_struct
*(*start
)(void *);
1466 struct task_struct
*(*next
)(void *);
1470 static unsigned long
1471 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1472 unsigned long max_load_move
, struct sched_domain
*sd
,
1473 enum cpu_idle_type idle
, int *all_pinned
,
1474 int *this_best_prio
, struct rq_iterator
*iterator
);
1477 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1478 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1479 struct rq_iterator
*iterator
);
1482 #ifdef CONFIG_CGROUP_CPUACCT
1483 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1485 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1488 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1490 update_load_add(&rq
->load
, load
);
1493 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1495 update_load_sub(&rq
->load
, load
);
1499 static unsigned long source_load(int cpu
, int type
);
1500 static unsigned long target_load(int cpu
, int type
);
1501 static unsigned long cpu_avg_load_per_task(int cpu
);
1502 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1503 #else /* CONFIG_SMP */
1505 #ifdef CONFIG_FAIR_GROUP_SCHED
1506 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1511 #endif /* CONFIG_SMP */
1513 #include "sched_stats.h"
1514 #include "sched_idletask.c"
1515 #include "sched_fair.c"
1516 #include "sched_rt.c"
1517 #ifdef CONFIG_SCHED_DEBUG
1518 # include "sched_debug.c"
1521 #define sched_class_highest (&rt_sched_class)
1523 static inline void inc_load(struct rq
*rq
, const struct task_struct
*p
)
1525 update_load_add(&rq
->load
, p
->se
.load
.weight
);
1528 static inline void dec_load(struct rq
*rq
, const struct task_struct
*p
)
1530 update_load_sub(&rq
->load
, p
->se
.load
.weight
);
1533 static void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
1539 static void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
1545 static void set_load_weight(struct task_struct
*p
)
1547 if (task_has_rt_policy(p
)) {
1548 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1549 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1554 * SCHED_IDLE tasks get minimal weight:
1556 if (p
->policy
== SCHED_IDLE
) {
1557 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1558 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1562 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1563 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1566 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1568 sched_info_queued(p
);
1569 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1573 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1575 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1580 * __normal_prio - return the priority that is based on the static prio
1582 static inline int __normal_prio(struct task_struct
*p
)
1584 return p
->static_prio
;
1588 * Calculate the expected normal priority: i.e. priority
1589 * without taking RT-inheritance into account. Might be
1590 * boosted by interactivity modifiers. Changes upon fork,
1591 * setprio syscalls, and whenever the interactivity
1592 * estimator recalculates.
1594 static inline int normal_prio(struct task_struct
*p
)
1598 if (task_has_rt_policy(p
))
1599 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1601 prio
= __normal_prio(p
);
1606 * Calculate the current priority, i.e. the priority
1607 * taken into account by the scheduler. This value might
1608 * be boosted by RT tasks, or might be boosted by
1609 * interactivity modifiers. Will be RT if the task got
1610 * RT-boosted. If not then it returns p->normal_prio.
1612 static int effective_prio(struct task_struct
*p
)
1614 p
->normal_prio
= normal_prio(p
);
1616 * If we are RT tasks or we were boosted to RT priority,
1617 * keep the priority unchanged. Otherwise, update priority
1618 * to the normal priority:
1620 if (!rt_prio(p
->prio
))
1621 return p
->normal_prio
;
1626 * activate_task - move a task to the runqueue.
1628 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1630 if (task_contributes_to_load(p
))
1631 rq
->nr_uninterruptible
--;
1633 enqueue_task(rq
, p
, wakeup
);
1634 inc_nr_running(p
, rq
);
1638 * deactivate_task - remove a task from the runqueue.
1640 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1642 if (task_contributes_to_load(p
))
1643 rq
->nr_uninterruptible
++;
1645 dequeue_task(rq
, p
, sleep
);
1646 dec_nr_running(p
, rq
);
1650 * task_curr - is this task currently executing on a CPU?
1651 * @p: the task in question.
1653 inline int task_curr(const struct task_struct
*p
)
1655 return cpu_curr(task_cpu(p
)) == p
;
1658 /* Used instead of source_load when we know the type == 0 */
1659 unsigned long weighted_cpuload(const int cpu
)
1661 return cpu_rq(cpu
)->load
.weight
;
1664 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1666 set_task_rq(p
, cpu
);
1669 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1670 * successfuly executed on another CPU. We must ensure that updates of
1671 * per-task data have been completed by this moment.
1674 task_thread_info(p
)->cpu
= cpu
;
1678 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1679 const struct sched_class
*prev_class
,
1680 int oldprio
, int running
)
1682 if (prev_class
!= p
->sched_class
) {
1683 if (prev_class
->switched_from
)
1684 prev_class
->switched_from(rq
, p
, running
);
1685 p
->sched_class
->switched_to(rq
, p
, running
);
1687 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1693 * Is this task likely cache-hot:
1696 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1701 * Buddy candidates are cache hot:
1703 if (sched_feat(CACHE_HOT_BUDDY
) && (&p
->se
== cfs_rq_of(&p
->se
)->next
))
1706 if (p
->sched_class
!= &fair_sched_class
)
1709 if (sysctl_sched_migration_cost
== -1)
1711 if (sysctl_sched_migration_cost
== 0)
1714 delta
= now
- p
->se
.exec_start
;
1716 return delta
< (s64
)sysctl_sched_migration_cost
;
1720 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1722 int old_cpu
= task_cpu(p
);
1723 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1724 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1725 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1728 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1730 #ifdef CONFIG_SCHEDSTATS
1731 if (p
->se
.wait_start
)
1732 p
->se
.wait_start
-= clock_offset
;
1733 if (p
->se
.sleep_start
)
1734 p
->se
.sleep_start
-= clock_offset
;
1735 if (p
->se
.block_start
)
1736 p
->se
.block_start
-= clock_offset
;
1737 if (old_cpu
!= new_cpu
) {
1738 schedstat_inc(p
, se
.nr_migrations
);
1739 if (task_hot(p
, old_rq
->clock
, NULL
))
1740 schedstat_inc(p
, se
.nr_forced2_migrations
);
1743 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1744 new_cfsrq
->min_vruntime
;
1746 __set_task_cpu(p
, new_cpu
);
1749 struct migration_req
{
1750 struct list_head list
;
1752 struct task_struct
*task
;
1755 struct completion done
;
1759 * The task's runqueue lock must be held.
1760 * Returns true if you have to wait for migration thread.
1763 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1765 struct rq
*rq
= task_rq(p
);
1768 * If the task is not on a runqueue (and not running), then
1769 * it is sufficient to simply update the task's cpu field.
1771 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1772 set_task_cpu(p
, dest_cpu
);
1776 init_completion(&req
->done
);
1778 req
->dest_cpu
= dest_cpu
;
1779 list_add(&req
->list
, &rq
->migration_queue
);
1785 * wait_task_inactive - wait for a thread to unschedule.
1787 * The caller must ensure that the task *will* unschedule sometime soon,
1788 * else this function might spin for a *long* time. This function can't
1789 * be called with interrupts off, or it may introduce deadlock with
1790 * smp_call_function() if an IPI is sent by the same process we are
1791 * waiting to become inactive.
1793 void wait_task_inactive(struct task_struct
*p
)
1795 unsigned long flags
;
1801 * We do the initial early heuristics without holding
1802 * any task-queue locks at all. We'll only try to get
1803 * the runqueue lock when things look like they will
1809 * If the task is actively running on another CPU
1810 * still, just relax and busy-wait without holding
1813 * NOTE! Since we don't hold any locks, it's not
1814 * even sure that "rq" stays as the right runqueue!
1815 * But we don't care, since "task_running()" will
1816 * return false if the runqueue has changed and p
1817 * is actually now running somewhere else!
1819 while (task_running(rq
, p
))
1823 * Ok, time to look more closely! We need the rq
1824 * lock now, to be *sure*. If we're wrong, we'll
1825 * just go back and repeat.
1827 rq
= task_rq_lock(p
, &flags
);
1828 running
= task_running(rq
, p
);
1829 on_rq
= p
->se
.on_rq
;
1830 task_rq_unlock(rq
, &flags
);
1833 * Was it really running after all now that we
1834 * checked with the proper locks actually held?
1836 * Oops. Go back and try again..
1838 if (unlikely(running
)) {
1844 * It's not enough that it's not actively running,
1845 * it must be off the runqueue _entirely_, and not
1848 * So if it wa still runnable (but just not actively
1849 * running right now), it's preempted, and we should
1850 * yield - it could be a while.
1852 if (unlikely(on_rq
)) {
1853 schedule_timeout_uninterruptible(1);
1858 * Ahh, all good. It wasn't running, and it wasn't
1859 * runnable, which means that it will never become
1860 * running in the future either. We're all done!
1867 * kick_process - kick a running thread to enter/exit the kernel
1868 * @p: the to-be-kicked thread
1870 * Cause a process which is running on another CPU to enter
1871 * kernel-mode, without any delay. (to get signals handled.)
1873 * NOTE: this function doesnt have to take the runqueue lock,
1874 * because all it wants to ensure is that the remote task enters
1875 * the kernel. If the IPI races and the task has been migrated
1876 * to another CPU then no harm is done and the purpose has been
1879 void kick_process(struct task_struct
*p
)
1885 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1886 smp_send_reschedule(cpu
);
1891 * Return a low guess at the load of a migration-source cpu weighted
1892 * according to the scheduling class and "nice" value.
1894 * We want to under-estimate the load of migration sources, to
1895 * balance conservatively.
1897 static unsigned long source_load(int cpu
, int type
)
1899 struct rq
*rq
= cpu_rq(cpu
);
1900 unsigned long total
= weighted_cpuload(cpu
);
1905 return min(rq
->cpu_load
[type
-1], total
);
1909 * Return a high guess at the load of a migration-target cpu weighted
1910 * according to the scheduling class and "nice" value.
1912 static unsigned long target_load(int cpu
, int type
)
1914 struct rq
*rq
= cpu_rq(cpu
);
1915 unsigned long total
= weighted_cpuload(cpu
);
1920 return max(rq
->cpu_load
[type
-1], total
);
1924 * Return the average load per task on the cpu's run queue
1926 static unsigned long cpu_avg_load_per_task(int cpu
)
1928 struct rq
*rq
= cpu_rq(cpu
);
1929 unsigned long total
= weighted_cpuload(cpu
);
1930 unsigned long n
= rq
->nr_running
;
1932 return n
? total
/ n
: SCHED_LOAD_SCALE
;
1936 * find_idlest_group finds and returns the least busy CPU group within the
1939 static struct sched_group
*
1940 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1942 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1943 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1944 int load_idx
= sd
->forkexec_idx
;
1945 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1948 unsigned long load
, avg_load
;
1952 /* Skip over this group if it has no CPUs allowed */
1953 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1956 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1958 /* Tally up the load of all CPUs in the group */
1961 for_each_cpu_mask(i
, group
->cpumask
) {
1962 /* Bias balancing toward cpus of our domain */
1964 load
= source_load(i
, load_idx
);
1966 load
= target_load(i
, load_idx
);
1971 /* Adjust by relative CPU power of the group */
1972 avg_load
= sg_div_cpu_power(group
,
1973 avg_load
* SCHED_LOAD_SCALE
);
1976 this_load
= avg_load
;
1978 } else if (avg_load
< min_load
) {
1979 min_load
= avg_load
;
1982 } while (group
= group
->next
, group
!= sd
->groups
);
1984 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1990 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1993 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
1996 unsigned long load
, min_load
= ULONG_MAX
;
2000 /* Traverse only the allowed CPUs */
2001 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2003 for_each_cpu_mask(i
, *tmp
) {
2004 load
= weighted_cpuload(i
);
2006 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2016 * sched_balance_self: balance the current task (running on cpu) in domains
2017 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2020 * Balance, ie. select the least loaded group.
2022 * Returns the target CPU number, or the same CPU if no balancing is needed.
2024 * preempt must be disabled.
2026 static int sched_balance_self(int cpu
, int flag
)
2028 struct task_struct
*t
= current
;
2029 struct sched_domain
*tmp
, *sd
= NULL
;
2031 for_each_domain(cpu
, tmp
) {
2033 * If power savings logic is enabled for a domain, stop there.
2035 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2037 if (tmp
->flags
& flag
)
2042 cpumask_t span
, tmpmask
;
2043 struct sched_group
*group
;
2044 int new_cpu
, weight
;
2046 if (!(sd
->flags
& flag
)) {
2052 group
= find_idlest_group(sd
, t
, cpu
);
2058 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2059 if (new_cpu
== -1 || new_cpu
== cpu
) {
2060 /* Now try balancing at a lower domain level of cpu */
2065 /* Now try balancing at a lower domain level of new_cpu */
2068 weight
= cpus_weight(span
);
2069 for_each_domain(cpu
, tmp
) {
2070 if (weight
<= cpus_weight(tmp
->span
))
2072 if (tmp
->flags
& flag
)
2075 /* while loop will break here if sd == NULL */
2081 #endif /* CONFIG_SMP */
2084 * try_to_wake_up - wake up a thread
2085 * @p: the to-be-woken-up thread
2086 * @state: the mask of task states that can be woken
2087 * @sync: do a synchronous wakeup?
2089 * Put it on the run-queue if it's not already there. The "current"
2090 * thread is always on the run-queue (except when the actual
2091 * re-schedule is in progress), and as such you're allowed to do
2092 * the simpler "current->state = TASK_RUNNING" to mark yourself
2093 * runnable without the overhead of this.
2095 * returns failure only if the task is already active.
2097 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2099 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2100 unsigned long flags
;
2104 if (!sched_feat(SYNC_WAKEUPS
))
2108 rq
= task_rq_lock(p
, &flags
);
2109 old_state
= p
->state
;
2110 if (!(old_state
& state
))
2118 this_cpu
= smp_processor_id();
2121 if (unlikely(task_running(rq
, p
)))
2124 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2125 if (cpu
!= orig_cpu
) {
2126 set_task_cpu(p
, cpu
);
2127 task_rq_unlock(rq
, &flags
);
2128 /* might preempt at this point */
2129 rq
= task_rq_lock(p
, &flags
);
2130 old_state
= p
->state
;
2131 if (!(old_state
& state
))
2136 this_cpu
= smp_processor_id();
2140 #ifdef CONFIG_SCHEDSTATS
2141 schedstat_inc(rq
, ttwu_count
);
2142 if (cpu
== this_cpu
)
2143 schedstat_inc(rq
, ttwu_local
);
2145 struct sched_domain
*sd
;
2146 for_each_domain(this_cpu
, sd
) {
2147 if (cpu_isset(cpu
, sd
->span
)) {
2148 schedstat_inc(sd
, ttwu_wake_remote
);
2156 #endif /* CONFIG_SMP */
2157 schedstat_inc(p
, se
.nr_wakeups
);
2159 schedstat_inc(p
, se
.nr_wakeups_sync
);
2160 if (orig_cpu
!= cpu
)
2161 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2162 if (cpu
== this_cpu
)
2163 schedstat_inc(p
, se
.nr_wakeups_local
);
2165 schedstat_inc(p
, se
.nr_wakeups_remote
);
2166 update_rq_clock(rq
);
2167 activate_task(rq
, p
, 1);
2171 trace_mark(kernel_sched_wakeup
,
2172 "pid %d state %ld ## rq %p task %p rq->curr %p",
2173 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2174 check_preempt_curr(rq
, p
);
2176 p
->state
= TASK_RUNNING
;
2178 if (p
->sched_class
->task_wake_up
)
2179 p
->sched_class
->task_wake_up(rq
, p
);
2182 task_rq_unlock(rq
, &flags
);
2187 int wake_up_process(struct task_struct
*p
)
2189 return try_to_wake_up(p
, TASK_ALL
, 0);
2191 EXPORT_SYMBOL(wake_up_process
);
2193 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2195 return try_to_wake_up(p
, state
, 0);
2199 * Perform scheduler related setup for a newly forked process p.
2200 * p is forked by current.
2202 * __sched_fork() is basic setup used by init_idle() too:
2204 static void __sched_fork(struct task_struct
*p
)
2206 p
->se
.exec_start
= 0;
2207 p
->se
.sum_exec_runtime
= 0;
2208 p
->se
.prev_sum_exec_runtime
= 0;
2209 p
->se
.last_wakeup
= 0;
2210 p
->se
.avg_overlap
= 0;
2212 #ifdef CONFIG_SCHEDSTATS
2213 p
->se
.wait_start
= 0;
2214 p
->se
.sum_sleep_runtime
= 0;
2215 p
->se
.sleep_start
= 0;
2216 p
->se
.block_start
= 0;
2217 p
->se
.sleep_max
= 0;
2218 p
->se
.block_max
= 0;
2220 p
->se
.slice_max
= 0;
2224 INIT_LIST_HEAD(&p
->rt
.run_list
);
2226 INIT_LIST_HEAD(&p
->se
.group_node
);
2228 #ifdef CONFIG_PREEMPT_NOTIFIERS
2229 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2233 * We mark the process as running here, but have not actually
2234 * inserted it onto the runqueue yet. This guarantees that
2235 * nobody will actually run it, and a signal or other external
2236 * event cannot wake it up and insert it on the runqueue either.
2238 p
->state
= TASK_RUNNING
;
2242 * fork()/clone()-time setup:
2244 void sched_fork(struct task_struct
*p
, int clone_flags
)
2246 int cpu
= get_cpu();
2251 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2253 set_task_cpu(p
, cpu
);
2256 * Make sure we do not leak PI boosting priority to the child:
2258 p
->prio
= current
->normal_prio
;
2259 if (!rt_prio(p
->prio
))
2260 p
->sched_class
= &fair_sched_class
;
2262 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2263 if (likely(sched_info_on()))
2264 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2266 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2269 #ifdef CONFIG_PREEMPT
2270 /* Want to start with kernel preemption disabled. */
2271 task_thread_info(p
)->preempt_count
= 1;
2277 * wake_up_new_task - wake up a newly created task for the first time.
2279 * This function will do some initial scheduler statistics housekeeping
2280 * that must be done for every newly created context, then puts the task
2281 * on the runqueue and wakes it.
2283 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2285 unsigned long flags
;
2288 rq
= task_rq_lock(p
, &flags
);
2289 BUG_ON(p
->state
!= TASK_RUNNING
);
2290 update_rq_clock(rq
);
2292 p
->prio
= effective_prio(p
);
2294 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2295 activate_task(rq
, p
, 0);
2298 * Let the scheduling class do new task startup
2299 * management (if any):
2301 p
->sched_class
->task_new(rq
, p
);
2302 inc_nr_running(p
, rq
);
2304 trace_mark(kernel_sched_wakeup_new
,
2305 "pid %d state %ld ## rq %p task %p rq->curr %p",
2306 p
->pid
, p
->state
, rq
, p
, rq
->curr
);
2307 check_preempt_curr(rq
, p
);
2309 if (p
->sched_class
->task_wake_up
)
2310 p
->sched_class
->task_wake_up(rq
, p
);
2312 task_rq_unlock(rq
, &flags
);
2315 #ifdef CONFIG_PREEMPT_NOTIFIERS
2318 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2319 * @notifier: notifier struct to register
2321 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2323 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2325 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2328 * preempt_notifier_unregister - no longer interested in preemption notifications
2329 * @notifier: notifier struct to unregister
2331 * This is safe to call from within a preemption notifier.
2333 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2335 hlist_del(¬ifier
->link
);
2337 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2339 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2341 struct preempt_notifier
*notifier
;
2342 struct hlist_node
*node
;
2344 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2345 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2349 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2350 struct task_struct
*next
)
2352 struct preempt_notifier
*notifier
;
2353 struct hlist_node
*node
;
2355 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2356 notifier
->ops
->sched_out(notifier
, next
);
2361 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2366 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2367 struct task_struct
*next
)
2374 * prepare_task_switch - prepare to switch tasks
2375 * @rq: the runqueue preparing to switch
2376 * @prev: the current task that is being switched out
2377 * @next: the task we are going to switch to.
2379 * This is called with the rq lock held and interrupts off. It must
2380 * be paired with a subsequent finish_task_switch after the context
2383 * prepare_task_switch sets up locking and calls architecture specific
2387 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2388 struct task_struct
*next
)
2390 fire_sched_out_preempt_notifiers(prev
, next
);
2391 prepare_lock_switch(rq
, next
);
2392 prepare_arch_switch(next
);
2396 * finish_task_switch - clean up after a task-switch
2397 * @rq: runqueue associated with task-switch
2398 * @prev: the thread we just switched away from.
2400 * finish_task_switch must be called after the context switch, paired
2401 * with a prepare_task_switch call before the context switch.
2402 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2403 * and do any other architecture-specific cleanup actions.
2405 * Note that we may have delayed dropping an mm in context_switch(). If
2406 * so, we finish that here outside of the runqueue lock. (Doing it
2407 * with the lock held can cause deadlocks; see schedule() for
2410 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2411 __releases(rq
->lock
)
2413 struct mm_struct
*mm
= rq
->prev_mm
;
2419 * A task struct has one reference for the use as "current".
2420 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2421 * schedule one last time. The schedule call will never return, and
2422 * the scheduled task must drop that reference.
2423 * The test for TASK_DEAD must occur while the runqueue locks are
2424 * still held, otherwise prev could be scheduled on another cpu, die
2425 * there before we look at prev->state, and then the reference would
2427 * Manfred Spraul <manfred@colorfullife.com>
2429 prev_state
= prev
->state
;
2430 finish_arch_switch(prev
);
2431 finish_lock_switch(rq
, prev
);
2433 if (current
->sched_class
->post_schedule
)
2434 current
->sched_class
->post_schedule(rq
);
2437 fire_sched_in_preempt_notifiers(current
);
2440 if (unlikely(prev_state
== TASK_DEAD
)) {
2442 * Remove function-return probe instances associated with this
2443 * task and put them back on the free list.
2445 kprobe_flush_task(prev
);
2446 put_task_struct(prev
);
2451 * schedule_tail - first thing a freshly forked thread must call.
2452 * @prev: the thread we just switched away from.
2454 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2455 __releases(rq
->lock
)
2457 struct rq
*rq
= this_rq();
2459 finish_task_switch(rq
, prev
);
2460 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2461 /* In this case, finish_task_switch does not reenable preemption */
2464 if (current
->set_child_tid
)
2465 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2469 * context_switch - switch to the new MM and the new
2470 * thread's register state.
2473 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2474 struct task_struct
*next
)
2476 struct mm_struct
*mm
, *oldmm
;
2478 prepare_task_switch(rq
, prev
, next
);
2479 trace_mark(kernel_sched_schedule
,
2480 "prev_pid %d next_pid %d prev_state %ld "
2481 "## rq %p prev %p next %p",
2482 prev
->pid
, next
->pid
, prev
->state
,
2485 oldmm
= prev
->active_mm
;
2487 * For paravirt, this is coupled with an exit in switch_to to
2488 * combine the page table reload and the switch backend into
2491 arch_enter_lazy_cpu_mode();
2493 if (unlikely(!mm
)) {
2494 next
->active_mm
= oldmm
;
2495 atomic_inc(&oldmm
->mm_count
);
2496 enter_lazy_tlb(oldmm
, next
);
2498 switch_mm(oldmm
, mm
, next
);
2500 if (unlikely(!prev
->mm
)) {
2501 prev
->active_mm
= NULL
;
2502 rq
->prev_mm
= oldmm
;
2505 * Since the runqueue lock will be released by the next
2506 * task (which is an invalid locking op but in the case
2507 * of the scheduler it's an obvious special-case), so we
2508 * do an early lockdep release here:
2510 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2511 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2514 /* Here we just switch the register state and the stack. */
2515 switch_to(prev
, next
, prev
);
2519 * this_rq must be evaluated again because prev may have moved
2520 * CPUs since it called schedule(), thus the 'rq' on its stack
2521 * frame will be invalid.
2523 finish_task_switch(this_rq(), prev
);
2527 * nr_running, nr_uninterruptible and nr_context_switches:
2529 * externally visible scheduler statistics: current number of runnable
2530 * threads, current number of uninterruptible-sleeping threads, total
2531 * number of context switches performed since bootup.
2533 unsigned long nr_running(void)
2535 unsigned long i
, sum
= 0;
2537 for_each_online_cpu(i
)
2538 sum
+= cpu_rq(i
)->nr_running
;
2543 unsigned long nr_uninterruptible(void)
2545 unsigned long i
, sum
= 0;
2547 for_each_possible_cpu(i
)
2548 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2551 * Since we read the counters lockless, it might be slightly
2552 * inaccurate. Do not allow it to go below zero though:
2554 if (unlikely((long)sum
< 0))
2560 unsigned long long nr_context_switches(void)
2563 unsigned long long sum
= 0;
2565 for_each_possible_cpu(i
)
2566 sum
+= cpu_rq(i
)->nr_switches
;
2571 unsigned long nr_iowait(void)
2573 unsigned long i
, sum
= 0;
2575 for_each_possible_cpu(i
)
2576 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2581 unsigned long nr_active(void)
2583 unsigned long i
, running
= 0, uninterruptible
= 0;
2585 for_each_online_cpu(i
) {
2586 running
+= cpu_rq(i
)->nr_running
;
2587 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2590 if (unlikely((long)uninterruptible
< 0))
2591 uninterruptible
= 0;
2593 return running
+ uninterruptible
;
2597 * Update rq->cpu_load[] statistics. This function is usually called every
2598 * scheduler tick (TICK_NSEC).
2600 static void update_cpu_load(struct rq
*this_rq
)
2602 unsigned long this_load
= this_rq
->load
.weight
;
2605 this_rq
->nr_load_updates
++;
2607 /* Update our load: */
2608 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2609 unsigned long old_load
, new_load
;
2611 /* scale is effectively 1 << i now, and >> i divides by scale */
2613 old_load
= this_rq
->cpu_load
[i
];
2614 new_load
= this_load
;
2616 * Round up the averaging division if load is increasing. This
2617 * prevents us from getting stuck on 9 if the load is 10, for
2620 if (new_load
> old_load
)
2621 new_load
+= scale
-1;
2622 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2629 * double_rq_lock - safely lock two runqueues
2631 * Note this does not disable interrupts like task_rq_lock,
2632 * you need to do so manually before calling.
2634 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2635 __acquires(rq1
->lock
)
2636 __acquires(rq2
->lock
)
2638 BUG_ON(!irqs_disabled());
2640 spin_lock(&rq1
->lock
);
2641 __acquire(rq2
->lock
); /* Fake it out ;) */
2644 spin_lock(&rq1
->lock
);
2645 spin_lock(&rq2
->lock
);
2647 spin_lock(&rq2
->lock
);
2648 spin_lock(&rq1
->lock
);
2651 update_rq_clock(rq1
);
2652 update_rq_clock(rq2
);
2656 * double_rq_unlock - safely unlock two runqueues
2658 * Note this does not restore interrupts like task_rq_unlock,
2659 * you need to do so manually after calling.
2661 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2662 __releases(rq1
->lock
)
2663 __releases(rq2
->lock
)
2665 spin_unlock(&rq1
->lock
);
2667 spin_unlock(&rq2
->lock
);
2669 __release(rq2
->lock
);
2673 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2675 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2676 __releases(this_rq
->lock
)
2677 __acquires(busiest
->lock
)
2678 __acquires(this_rq
->lock
)
2682 if (unlikely(!irqs_disabled())) {
2683 /* printk() doesn't work good under rq->lock */
2684 spin_unlock(&this_rq
->lock
);
2687 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2688 if (busiest
< this_rq
) {
2689 spin_unlock(&this_rq
->lock
);
2690 spin_lock(&busiest
->lock
);
2691 spin_lock(&this_rq
->lock
);
2694 spin_lock(&busiest
->lock
);
2700 * If dest_cpu is allowed for this process, migrate the task to it.
2701 * This is accomplished by forcing the cpu_allowed mask to only
2702 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2703 * the cpu_allowed mask is restored.
2705 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2707 struct migration_req req
;
2708 unsigned long flags
;
2711 rq
= task_rq_lock(p
, &flags
);
2712 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2713 || unlikely(cpu_is_offline(dest_cpu
)))
2716 /* force the process onto the specified CPU */
2717 if (migrate_task(p
, dest_cpu
, &req
)) {
2718 /* Need to wait for migration thread (might exit: take ref). */
2719 struct task_struct
*mt
= rq
->migration_thread
;
2721 get_task_struct(mt
);
2722 task_rq_unlock(rq
, &flags
);
2723 wake_up_process(mt
);
2724 put_task_struct(mt
);
2725 wait_for_completion(&req
.done
);
2730 task_rq_unlock(rq
, &flags
);
2734 * sched_exec - execve() is a valuable balancing opportunity, because at
2735 * this point the task has the smallest effective memory and cache footprint.
2737 void sched_exec(void)
2739 int new_cpu
, this_cpu
= get_cpu();
2740 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2742 if (new_cpu
!= this_cpu
)
2743 sched_migrate_task(current
, new_cpu
);
2747 * pull_task - move a task from a remote runqueue to the local runqueue.
2748 * Both runqueues must be locked.
2750 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2751 struct rq
*this_rq
, int this_cpu
)
2753 deactivate_task(src_rq
, p
, 0);
2754 set_task_cpu(p
, this_cpu
);
2755 activate_task(this_rq
, p
, 0);
2757 * Note that idle threads have a prio of MAX_PRIO, for this test
2758 * to be always true for them.
2760 check_preempt_curr(this_rq
, p
);
2764 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2767 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2768 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2772 * We do not migrate tasks that are:
2773 * 1) running (obviously), or
2774 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2775 * 3) are cache-hot on their current CPU.
2777 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2778 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2783 if (task_running(rq
, p
)) {
2784 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2789 * Aggressive migration if:
2790 * 1) task is cache cold, or
2791 * 2) too many balance attempts have failed.
2794 if (!task_hot(p
, rq
->clock
, sd
) ||
2795 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2796 #ifdef CONFIG_SCHEDSTATS
2797 if (task_hot(p
, rq
->clock
, sd
)) {
2798 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2799 schedstat_inc(p
, se
.nr_forced_migrations
);
2805 if (task_hot(p
, rq
->clock
, sd
)) {
2806 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2812 static unsigned long
2813 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2814 unsigned long max_load_move
, struct sched_domain
*sd
,
2815 enum cpu_idle_type idle
, int *all_pinned
,
2816 int *this_best_prio
, struct rq_iterator
*iterator
)
2818 int loops
= 0, pulled
= 0, pinned
= 0, skip_for_load
;
2819 struct task_struct
*p
;
2820 long rem_load_move
= max_load_move
;
2822 if (max_load_move
== 0)
2828 * Start the load-balancing iterator:
2830 p
= iterator
->start(iterator
->arg
);
2832 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2835 * To help distribute high priority tasks across CPUs we don't
2836 * skip a task if it will be the highest priority task (i.e. smallest
2837 * prio value) on its new queue regardless of its load weight
2839 skip_for_load
= (p
->se
.load
.weight
>> 1) > rem_load_move
+
2840 SCHED_LOAD_SCALE_FUZZ
;
2841 if ((skip_for_load
&& p
->prio
>= *this_best_prio
) ||
2842 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2843 p
= iterator
->next(iterator
->arg
);
2847 pull_task(busiest
, p
, this_rq
, this_cpu
);
2849 rem_load_move
-= p
->se
.load
.weight
;
2852 * We only want to steal up to the prescribed amount of weighted load.
2854 if (rem_load_move
> 0) {
2855 if (p
->prio
< *this_best_prio
)
2856 *this_best_prio
= p
->prio
;
2857 p
= iterator
->next(iterator
->arg
);
2862 * Right now, this is one of only two places pull_task() is called,
2863 * so we can safely collect pull_task() stats here rather than
2864 * inside pull_task().
2866 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2869 *all_pinned
= pinned
;
2871 return max_load_move
- rem_load_move
;
2875 * move_tasks tries to move up to max_load_move weighted load from busiest to
2876 * this_rq, as part of a balancing operation within domain "sd".
2877 * Returns 1 if successful and 0 otherwise.
2879 * Called with both runqueues locked.
2881 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2882 unsigned long max_load_move
,
2883 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2886 const struct sched_class
*class = sched_class_highest
;
2887 unsigned long total_load_moved
= 0;
2888 int this_best_prio
= this_rq
->curr
->prio
;
2892 class->load_balance(this_rq
, this_cpu
, busiest
,
2893 max_load_move
- total_load_moved
,
2894 sd
, idle
, all_pinned
, &this_best_prio
);
2895 class = class->next
;
2896 } while (class && max_load_move
> total_load_moved
);
2898 return total_load_moved
> 0;
2902 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2903 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2904 struct rq_iterator
*iterator
)
2906 struct task_struct
*p
= iterator
->start(iterator
->arg
);
2910 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2911 pull_task(busiest
, p
, this_rq
, this_cpu
);
2913 * Right now, this is only the second place pull_task()
2914 * is called, so we can safely collect pull_task()
2915 * stats here rather than inside pull_task().
2917 schedstat_inc(sd
, lb_gained
[idle
]);
2921 p
= iterator
->next(iterator
->arg
);
2928 * move_one_task tries to move exactly one task from busiest to this_rq, as
2929 * part of active balancing operations within "domain".
2930 * Returns 1 if successful and 0 otherwise.
2932 * Called with both runqueues locked.
2934 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2935 struct sched_domain
*sd
, enum cpu_idle_type idle
)
2937 const struct sched_class
*class;
2939 for (class = sched_class_highest
; class; class = class->next
)
2940 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
2947 * find_busiest_group finds and returns the busiest CPU group within the
2948 * domain. It calculates and returns the amount of weighted load which
2949 * should be moved to restore balance via the imbalance parameter.
2951 static struct sched_group
*
2952 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2953 unsigned long *imbalance
, enum cpu_idle_type idle
,
2954 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
2956 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2957 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2958 unsigned long max_pull
;
2959 unsigned long busiest_load_per_task
, busiest_nr_running
;
2960 unsigned long this_load_per_task
, this_nr_running
;
2961 int load_idx
, group_imb
= 0;
2962 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2963 int power_savings_balance
= 1;
2964 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2965 unsigned long min_nr_running
= ULONG_MAX
;
2966 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2969 max_load
= this_load
= total_load
= total_pwr
= 0;
2970 busiest_load_per_task
= busiest_nr_running
= 0;
2971 this_load_per_task
= this_nr_running
= 0;
2972 if (idle
== CPU_NOT_IDLE
)
2973 load_idx
= sd
->busy_idx
;
2974 else if (idle
== CPU_NEWLY_IDLE
)
2975 load_idx
= sd
->newidle_idx
;
2977 load_idx
= sd
->idle_idx
;
2980 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
2983 int __group_imb
= 0;
2984 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2985 unsigned long sum_nr_running
, sum_weighted_load
;
2987 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2990 balance_cpu
= first_cpu(group
->cpumask
);
2992 /* Tally up the load of all CPUs in the group */
2993 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2995 min_cpu_load
= ~0UL;
2997 for_each_cpu_mask(i
, group
->cpumask
) {
3000 if (!cpu_isset(i
, *cpus
))
3005 if (*sd_idle
&& rq
->nr_running
)
3008 /* Bias balancing toward cpus of our domain */
3010 if (idle_cpu(i
) && !first_idle_cpu
) {
3015 load
= target_load(i
, load_idx
);
3017 load
= source_load(i
, load_idx
);
3018 if (load
> max_cpu_load
)
3019 max_cpu_load
= load
;
3020 if (min_cpu_load
> load
)
3021 min_cpu_load
= load
;
3025 sum_nr_running
+= rq
->nr_running
;
3026 sum_weighted_load
+= weighted_cpuload(i
);
3030 * First idle cpu or the first cpu(busiest) in this sched group
3031 * is eligible for doing load balancing at this and above
3032 * domains. In the newly idle case, we will allow all the cpu's
3033 * to do the newly idle load balance.
3035 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3036 balance_cpu
!= this_cpu
&& balance
) {
3041 total_load
+= avg_load
;
3042 total_pwr
+= group
->__cpu_power
;
3044 /* Adjust by relative CPU power of the group */
3045 avg_load
= sg_div_cpu_power(group
,
3046 avg_load
* SCHED_LOAD_SCALE
);
3048 if ((max_cpu_load
- min_cpu_load
) > SCHED_LOAD_SCALE
)
3051 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3054 this_load
= avg_load
;
3056 this_nr_running
= sum_nr_running
;
3057 this_load_per_task
= sum_weighted_load
;
3058 } else if (avg_load
> max_load
&&
3059 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3060 max_load
= avg_load
;
3062 busiest_nr_running
= sum_nr_running
;
3063 busiest_load_per_task
= sum_weighted_load
;
3064 group_imb
= __group_imb
;
3067 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3069 * Busy processors will not participate in power savings
3072 if (idle
== CPU_NOT_IDLE
||
3073 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3077 * If the local group is idle or completely loaded
3078 * no need to do power savings balance at this domain
3080 if (local_group
&& (this_nr_running
>= group_capacity
||
3082 power_savings_balance
= 0;
3085 * If a group is already running at full capacity or idle,
3086 * don't include that group in power savings calculations
3088 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3093 * Calculate the group which has the least non-idle load.
3094 * This is the group from where we need to pick up the load
3097 if ((sum_nr_running
< min_nr_running
) ||
3098 (sum_nr_running
== min_nr_running
&&
3099 first_cpu(group
->cpumask
) <
3100 first_cpu(group_min
->cpumask
))) {
3102 min_nr_running
= sum_nr_running
;
3103 min_load_per_task
= sum_weighted_load
/
3108 * Calculate the group which is almost near its
3109 * capacity but still has some space to pick up some load
3110 * from other group and save more power
3112 if (sum_nr_running
<= group_capacity
- 1) {
3113 if (sum_nr_running
> leader_nr_running
||
3114 (sum_nr_running
== leader_nr_running
&&
3115 first_cpu(group
->cpumask
) >
3116 first_cpu(group_leader
->cpumask
))) {
3117 group_leader
= group
;
3118 leader_nr_running
= sum_nr_running
;
3123 group
= group
->next
;
3124 } while (group
!= sd
->groups
);
3126 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3129 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3131 if (this_load
>= avg_load
||
3132 100*max_load
<= sd
->imbalance_pct
*this_load
)
3135 busiest_load_per_task
/= busiest_nr_running
;
3137 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3140 * We're trying to get all the cpus to the average_load, so we don't
3141 * want to push ourselves above the average load, nor do we wish to
3142 * reduce the max loaded cpu below the average load, as either of these
3143 * actions would just result in more rebalancing later, and ping-pong
3144 * tasks around. Thus we look for the minimum possible imbalance.
3145 * Negative imbalances (*we* are more loaded than anyone else) will
3146 * be counted as no imbalance for these purposes -- we can't fix that
3147 * by pulling tasks to us. Be careful of negative numbers as they'll
3148 * appear as very large values with unsigned longs.
3150 if (max_load
<= busiest_load_per_task
)
3154 * In the presence of smp nice balancing, certain scenarios can have
3155 * max load less than avg load(as we skip the groups at or below
3156 * its cpu_power, while calculating max_load..)
3158 if (max_load
< avg_load
) {
3160 goto small_imbalance
;
3163 /* Don't want to pull so many tasks that a group would go idle */
3164 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3166 /* How much load to actually move to equalise the imbalance */
3167 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3168 (avg_load
- this_load
) * this->__cpu_power
)
3172 * if *imbalance is less than the average load per runnable task
3173 * there is no gaurantee that any tasks will be moved so we'll have
3174 * a think about bumping its value to force at least one task to be
3177 if (*imbalance
< busiest_load_per_task
) {
3178 unsigned long tmp
, pwr_now
, pwr_move
;
3182 pwr_move
= pwr_now
= 0;
3184 if (this_nr_running
) {
3185 this_load_per_task
/= this_nr_running
;
3186 if (busiest_load_per_task
> this_load_per_task
)
3189 this_load_per_task
= SCHED_LOAD_SCALE
;
3191 if (max_load
- this_load
+ SCHED_LOAD_SCALE_FUZZ
>=
3192 busiest_load_per_task
* imbn
) {
3193 *imbalance
= busiest_load_per_task
;
3198 * OK, we don't have enough imbalance to justify moving tasks,
3199 * however we may be able to increase total CPU power used by
3203 pwr_now
+= busiest
->__cpu_power
*
3204 min(busiest_load_per_task
, max_load
);
3205 pwr_now
+= this->__cpu_power
*
3206 min(this_load_per_task
, this_load
);
3207 pwr_now
/= SCHED_LOAD_SCALE
;
3209 /* Amount of load we'd subtract */
3210 tmp
= sg_div_cpu_power(busiest
,
3211 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3213 pwr_move
+= busiest
->__cpu_power
*
3214 min(busiest_load_per_task
, max_load
- tmp
);
3216 /* Amount of load we'd add */
3217 if (max_load
* busiest
->__cpu_power
<
3218 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3219 tmp
= sg_div_cpu_power(this,
3220 max_load
* busiest
->__cpu_power
);
3222 tmp
= sg_div_cpu_power(this,
3223 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3224 pwr_move
+= this->__cpu_power
*
3225 min(this_load_per_task
, this_load
+ tmp
);
3226 pwr_move
/= SCHED_LOAD_SCALE
;
3228 /* Move if we gain throughput */
3229 if (pwr_move
> pwr_now
)
3230 *imbalance
= busiest_load_per_task
;
3236 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3237 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3240 if (this == group_leader
&& group_leader
!= group_min
) {
3241 *imbalance
= min_load_per_task
;
3251 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3254 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3255 unsigned long imbalance
, const cpumask_t
*cpus
)
3257 struct rq
*busiest
= NULL
, *rq
;
3258 unsigned long max_load
= 0;
3261 for_each_cpu_mask(i
, group
->cpumask
) {
3264 if (!cpu_isset(i
, *cpus
))
3268 wl
= weighted_cpuload(i
);
3270 if (rq
->nr_running
== 1 && wl
> imbalance
)
3273 if (wl
> max_load
) {
3283 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3284 * so long as it is large enough.
3286 #define MAX_PINNED_INTERVAL 512
3289 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3290 * tasks if there is an imbalance.
3292 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3293 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3294 int *balance
, cpumask_t
*cpus
)
3296 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3297 struct sched_group
*group
;
3298 unsigned long imbalance
;
3300 unsigned long flags
;
3305 * When power savings policy is enabled for the parent domain, idle
3306 * sibling can pick up load irrespective of busy siblings. In this case,
3307 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3308 * portraying it as CPU_NOT_IDLE.
3310 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3311 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3314 schedstat_inc(sd
, lb_count
[idle
]);
3317 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3324 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3328 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3330 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3334 BUG_ON(busiest
== this_rq
);
3336 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3339 if (busiest
->nr_running
> 1) {
3341 * Attempt to move tasks. If find_busiest_group has found
3342 * an imbalance but busiest->nr_running <= 1, the group is
3343 * still unbalanced. ld_moved simply stays zero, so it is
3344 * correctly treated as an imbalance.
3346 local_irq_save(flags
);
3347 double_rq_lock(this_rq
, busiest
);
3348 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3349 imbalance
, sd
, idle
, &all_pinned
);
3350 double_rq_unlock(this_rq
, busiest
);
3351 local_irq_restore(flags
);
3354 * some other cpu did the load balance for us.
3356 if (ld_moved
&& this_cpu
!= smp_processor_id())
3357 resched_cpu(this_cpu
);
3359 /* All tasks on this runqueue were pinned by CPU affinity */
3360 if (unlikely(all_pinned
)) {
3361 cpu_clear(cpu_of(busiest
), *cpus
);
3362 if (!cpus_empty(*cpus
))
3369 schedstat_inc(sd
, lb_failed
[idle
]);
3370 sd
->nr_balance_failed
++;
3372 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3374 spin_lock_irqsave(&busiest
->lock
, flags
);
3376 /* don't kick the migration_thread, if the curr
3377 * task on busiest cpu can't be moved to this_cpu
3379 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3380 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3382 goto out_one_pinned
;
3385 if (!busiest
->active_balance
) {
3386 busiest
->active_balance
= 1;
3387 busiest
->push_cpu
= this_cpu
;
3390 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3392 wake_up_process(busiest
->migration_thread
);
3395 * We've kicked active balancing, reset the failure
3398 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3401 sd
->nr_balance_failed
= 0;
3403 if (likely(!active_balance
)) {
3404 /* We were unbalanced, so reset the balancing interval */
3405 sd
->balance_interval
= sd
->min_interval
;
3408 * If we've begun active balancing, start to back off. This
3409 * case may not be covered by the all_pinned logic if there
3410 * is only 1 task on the busy runqueue (because we don't call
3413 if (sd
->balance_interval
< sd
->max_interval
)
3414 sd
->balance_interval
*= 2;
3417 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3418 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3423 schedstat_inc(sd
, lb_balanced
[idle
]);
3425 sd
->nr_balance_failed
= 0;
3428 /* tune up the balancing interval */
3429 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3430 (sd
->balance_interval
< sd
->max_interval
))
3431 sd
->balance_interval
*= 2;
3433 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3434 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3440 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3441 * tasks if there is an imbalance.
3443 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3444 * this_rq is locked.
3447 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3450 struct sched_group
*group
;
3451 struct rq
*busiest
= NULL
;
3452 unsigned long imbalance
;
3460 * When power savings policy is enabled for the parent domain, idle
3461 * sibling can pick up load irrespective of busy siblings. In this case,
3462 * let the state of idle sibling percolate up as IDLE, instead of
3463 * portraying it as CPU_NOT_IDLE.
3465 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3466 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3469 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3471 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3472 &sd_idle
, cpus
, NULL
);
3474 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3478 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3480 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3484 BUG_ON(busiest
== this_rq
);
3486 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3489 if (busiest
->nr_running
> 1) {
3490 /* Attempt to move tasks */
3491 double_lock_balance(this_rq
, busiest
);
3492 /* this_rq->clock is already updated */
3493 update_rq_clock(busiest
);
3494 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3495 imbalance
, sd
, CPU_NEWLY_IDLE
,
3497 spin_unlock(&busiest
->lock
);
3499 if (unlikely(all_pinned
)) {
3500 cpu_clear(cpu_of(busiest
), *cpus
);
3501 if (!cpus_empty(*cpus
))
3507 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3508 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3509 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3512 sd
->nr_balance_failed
= 0;
3517 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3518 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3519 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3521 sd
->nr_balance_failed
= 0;
3527 * idle_balance is called by schedule() if this_cpu is about to become
3528 * idle. Attempts to pull tasks from other CPUs.
3530 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3532 struct sched_domain
*sd
;
3533 int pulled_task
= -1;
3534 unsigned long next_balance
= jiffies
+ HZ
;
3537 for_each_domain(this_cpu
, sd
) {
3538 unsigned long interval
;
3540 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3543 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3544 /* If we've pulled tasks over stop searching: */
3545 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3548 interval
= msecs_to_jiffies(sd
->balance_interval
);
3549 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3550 next_balance
= sd
->last_balance
+ interval
;
3554 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3556 * We are going idle. next_balance may be set based on
3557 * a busy processor. So reset next_balance.
3559 this_rq
->next_balance
= next_balance
;
3564 * active_load_balance is run by migration threads. It pushes running tasks
3565 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3566 * running on each physical CPU where possible, and avoids physical /
3567 * logical imbalances.
3569 * Called with busiest_rq locked.
3571 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3573 int target_cpu
= busiest_rq
->push_cpu
;
3574 struct sched_domain
*sd
;
3575 struct rq
*target_rq
;
3577 /* Is there any task to move? */
3578 if (busiest_rq
->nr_running
<= 1)
3581 target_rq
= cpu_rq(target_cpu
);
3584 * This condition is "impossible", if it occurs
3585 * we need to fix it. Originally reported by
3586 * Bjorn Helgaas on a 128-cpu setup.
3588 BUG_ON(busiest_rq
== target_rq
);
3590 /* move a task from busiest_rq to target_rq */
3591 double_lock_balance(busiest_rq
, target_rq
);
3592 update_rq_clock(busiest_rq
);
3593 update_rq_clock(target_rq
);
3595 /* Search for an sd spanning us and the target CPU. */
3596 for_each_domain(target_cpu
, sd
) {
3597 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3598 cpu_isset(busiest_cpu
, sd
->span
))
3603 schedstat_inc(sd
, alb_count
);
3605 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3607 schedstat_inc(sd
, alb_pushed
);
3609 schedstat_inc(sd
, alb_failed
);
3611 spin_unlock(&target_rq
->lock
);
3616 atomic_t load_balancer
;
3618 } nohz ____cacheline_aligned
= {
3619 .load_balancer
= ATOMIC_INIT(-1),
3620 .cpu_mask
= CPU_MASK_NONE
,
3624 * This routine will try to nominate the ilb (idle load balancing)
3625 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3626 * load balancing on behalf of all those cpus. If all the cpus in the system
3627 * go into this tickless mode, then there will be no ilb owner (as there is
3628 * no need for one) and all the cpus will sleep till the next wakeup event
3631 * For the ilb owner, tick is not stopped. And this tick will be used
3632 * for idle load balancing. ilb owner will still be part of
3635 * While stopping the tick, this cpu will become the ilb owner if there
3636 * is no other owner. And will be the owner till that cpu becomes busy
3637 * or if all cpus in the system stop their ticks at which point
3638 * there is no need for ilb owner.
3640 * When the ilb owner becomes busy, it nominates another owner, during the
3641 * next busy scheduler_tick()
3643 int select_nohz_load_balancer(int stop_tick
)
3645 int cpu
= smp_processor_id();
3648 cpu_set(cpu
, nohz
.cpu_mask
);
3649 cpu_rq(cpu
)->in_nohz_recently
= 1;
3652 * If we are going offline and still the leader, give up!
3654 if (cpu_is_offline(cpu
) &&
3655 atomic_read(&nohz
.load_balancer
) == cpu
) {
3656 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3661 /* time for ilb owner also to sleep */
3662 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3663 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3664 atomic_set(&nohz
.load_balancer
, -1);
3668 if (atomic_read(&nohz
.load_balancer
) == -1) {
3669 /* make me the ilb owner */
3670 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3672 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3675 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3678 cpu_clear(cpu
, nohz
.cpu_mask
);
3680 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3681 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3688 static DEFINE_SPINLOCK(balancing
);
3691 * It checks each scheduling domain to see if it is due to be balanced,
3692 * and initiates a balancing operation if so.
3694 * Balancing parameters are set up in arch_init_sched_domains.
3696 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3699 struct rq
*rq
= cpu_rq(cpu
);
3700 unsigned long interval
;
3701 struct sched_domain
*sd
;
3702 /* Earliest time when we have to do rebalance again */
3703 unsigned long next_balance
= jiffies
+ 60*HZ
;
3704 int update_next_balance
= 0;
3707 for_each_domain(cpu
, sd
) {
3708 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3711 interval
= sd
->balance_interval
;
3712 if (idle
!= CPU_IDLE
)
3713 interval
*= sd
->busy_factor
;
3715 /* scale ms to jiffies */
3716 interval
= msecs_to_jiffies(interval
);
3717 if (unlikely(!interval
))
3719 if (interval
> HZ
*NR_CPUS
/10)
3720 interval
= HZ
*NR_CPUS
/10;
3723 if (sd
->flags
& SD_SERIALIZE
) {
3724 if (!spin_trylock(&balancing
))
3728 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3729 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3731 * We've pulled tasks over so either we're no
3732 * longer idle, or one of our SMT siblings is
3735 idle
= CPU_NOT_IDLE
;
3737 sd
->last_balance
= jiffies
;
3739 if (sd
->flags
& SD_SERIALIZE
)
3740 spin_unlock(&balancing
);
3742 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3743 next_balance
= sd
->last_balance
+ interval
;
3744 update_next_balance
= 1;
3748 * Stop the load balance at this level. There is another
3749 * CPU in our sched group which is doing load balancing more
3757 * next_balance will be updated only when there is a need.
3758 * When the cpu is attached to null domain for ex, it will not be
3761 if (likely(update_next_balance
))
3762 rq
->next_balance
= next_balance
;
3766 * run_rebalance_domains is triggered when needed from the scheduler tick.
3767 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3768 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3770 static void run_rebalance_domains(struct softirq_action
*h
)
3772 int this_cpu
= smp_processor_id();
3773 struct rq
*this_rq
= cpu_rq(this_cpu
);
3774 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3775 CPU_IDLE
: CPU_NOT_IDLE
;
3777 rebalance_domains(this_cpu
, idle
);
3781 * If this cpu is the owner for idle load balancing, then do the
3782 * balancing on behalf of the other idle cpus whose ticks are
3785 if (this_rq
->idle_at_tick
&&
3786 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3787 cpumask_t cpus
= nohz
.cpu_mask
;
3791 cpu_clear(this_cpu
, cpus
);
3792 for_each_cpu_mask(balance_cpu
, cpus
) {
3794 * If this cpu gets work to do, stop the load balancing
3795 * work being done for other cpus. Next load
3796 * balancing owner will pick it up.
3801 rebalance_domains(balance_cpu
, CPU_IDLE
);
3803 rq
= cpu_rq(balance_cpu
);
3804 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3805 this_rq
->next_balance
= rq
->next_balance
;
3812 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3814 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3815 * idle load balancing owner or decide to stop the periodic load balancing,
3816 * if the whole system is idle.
3818 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
3822 * If we were in the nohz mode recently and busy at the current
3823 * scheduler tick, then check if we need to nominate new idle
3826 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
3827 rq
->in_nohz_recently
= 0;
3829 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
3830 cpu_clear(cpu
, nohz
.cpu_mask
);
3831 atomic_set(&nohz
.load_balancer
, -1);
3834 if (atomic_read(&nohz
.load_balancer
) == -1) {
3836 * simple selection for now: Nominate the
3837 * first cpu in the nohz list to be the next
3840 * TBD: Traverse the sched domains and nominate
3841 * the nearest cpu in the nohz.cpu_mask.
3843 int ilb
= first_cpu(nohz
.cpu_mask
);
3845 if (ilb
< nr_cpu_ids
)
3851 * If this cpu is idle and doing idle load balancing for all the
3852 * cpus with ticks stopped, is it time for that to stop?
3854 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
3855 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3861 * If this cpu is idle and the idle load balancing is done by
3862 * someone else, then no need raise the SCHED_SOFTIRQ
3864 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
3865 cpu_isset(cpu
, nohz
.cpu_mask
))
3868 if (time_after_eq(jiffies
, rq
->next_balance
))
3869 raise_softirq(SCHED_SOFTIRQ
);
3872 #else /* CONFIG_SMP */
3875 * on UP we do not need to balance between CPUs:
3877 static inline void idle_balance(int cpu
, struct rq
*rq
)
3883 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3885 EXPORT_PER_CPU_SYMBOL(kstat
);
3888 * Return p->sum_exec_runtime plus any more ns on the sched_clock
3889 * that have not yet been banked in case the task is currently running.
3891 unsigned long long task_sched_runtime(struct task_struct
*p
)
3893 unsigned long flags
;
3897 rq
= task_rq_lock(p
, &flags
);
3898 ns
= p
->se
.sum_exec_runtime
;
3899 if (task_current(rq
, p
)) {
3900 update_rq_clock(rq
);
3901 delta_exec
= rq
->clock
- p
->se
.exec_start
;
3902 if ((s64
)delta_exec
> 0)
3905 task_rq_unlock(rq
, &flags
);
3911 * Account user cpu time to a process.
3912 * @p: the process that the cpu time gets accounted to
3913 * @cputime: the cpu time spent in user space since the last update
3915 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3917 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3920 p
->utime
= cputime_add(p
->utime
, cputime
);
3922 /* Add user time to cpustat. */
3923 tmp
= cputime_to_cputime64(cputime
);
3924 if (TASK_NICE(p
) > 0)
3925 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3927 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3931 * Account guest cpu time to a process.
3932 * @p: the process that the cpu time gets accounted to
3933 * @cputime: the cpu time spent in virtual machine since the last update
3935 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
3938 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3940 tmp
= cputime_to_cputime64(cputime
);
3942 p
->utime
= cputime_add(p
->utime
, cputime
);
3943 p
->gtime
= cputime_add(p
->gtime
, cputime
);
3945 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3946 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
3950 * Account scaled user cpu time to a process.
3951 * @p: the process that the cpu time gets accounted to
3952 * @cputime: the cpu time spent in user space since the last update
3954 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
3956 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
3960 * Account system cpu time to a process.
3961 * @p: the process that the cpu time gets accounted to
3962 * @hardirq_offset: the offset to subtract from hardirq_count()
3963 * @cputime: the cpu time spent in kernel space since the last update
3965 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3968 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3969 struct rq
*rq
= this_rq();
3972 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
3973 account_guest_time(p
, cputime
);
3977 p
->stime
= cputime_add(p
->stime
, cputime
);
3979 /* Add system time to cpustat. */
3980 tmp
= cputime_to_cputime64(cputime
);
3981 if (hardirq_count() - hardirq_offset
)
3982 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3983 else if (softirq_count())
3984 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3985 else if (p
!= rq
->idle
)
3986 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3987 else if (atomic_read(&rq
->nr_iowait
) > 0)
3988 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3990 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3991 /* Account for system time used */
3992 acct_update_integrals(p
);
3996 * Account scaled system cpu time to a process.
3997 * @p: the process that the cpu time gets accounted to
3998 * @hardirq_offset: the offset to subtract from hardirq_count()
3999 * @cputime: the cpu time spent in kernel space since the last update
4001 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4003 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4007 * Account for involuntary wait time.
4008 * @p: the process from which the cpu time has been stolen
4009 * @steal: the cpu time spent in involuntary wait
4011 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4013 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4014 cputime64_t tmp
= cputime_to_cputime64(steal
);
4015 struct rq
*rq
= this_rq();
4017 if (p
== rq
->idle
) {
4018 p
->stime
= cputime_add(p
->stime
, steal
);
4019 if (atomic_read(&rq
->nr_iowait
) > 0)
4020 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4022 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4024 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4028 * This function gets called by the timer code, with HZ frequency.
4029 * We call it with interrupts disabled.
4031 * It also gets called by the fork code, when changing the parent's
4034 void scheduler_tick(void)
4036 int cpu
= smp_processor_id();
4037 struct rq
*rq
= cpu_rq(cpu
);
4038 struct task_struct
*curr
= rq
->curr
;
4042 spin_lock(&rq
->lock
);
4043 update_rq_clock(rq
);
4044 update_cpu_load(rq
);
4045 curr
->sched_class
->task_tick(rq
, curr
, 0);
4046 spin_unlock(&rq
->lock
);
4049 rq
->idle_at_tick
= idle_cpu(cpu
);
4050 trigger_load_balance(rq
, cpu
);
4054 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4055 defined(CONFIG_PREEMPT_TRACER))
4057 static inline unsigned long get_parent_ip(unsigned long addr
)
4059 if (in_lock_functions(addr
)) {
4060 addr
= CALLER_ADDR2
;
4061 if (in_lock_functions(addr
))
4062 addr
= CALLER_ADDR3
;
4067 void __kprobes
add_preempt_count(int val
)
4069 #ifdef CONFIG_DEBUG_PREEMPT
4073 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4076 preempt_count() += val
;
4077 #ifdef CONFIG_DEBUG_PREEMPT
4079 * Spinlock count overflowing soon?
4081 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4084 if (preempt_count() == val
)
4085 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4087 EXPORT_SYMBOL(add_preempt_count
);
4089 void __kprobes
sub_preempt_count(int val
)
4091 #ifdef CONFIG_DEBUG_PREEMPT
4095 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4098 * Is the spinlock portion underflowing?
4100 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4101 !(preempt_count() & PREEMPT_MASK
)))
4105 if (preempt_count() == val
)
4106 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4107 preempt_count() -= val
;
4109 EXPORT_SYMBOL(sub_preempt_count
);
4114 * Print scheduling while atomic bug:
4116 static noinline
void __schedule_bug(struct task_struct
*prev
)
4118 struct pt_regs
*regs
= get_irq_regs();
4120 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4121 prev
->comm
, prev
->pid
, preempt_count());
4123 debug_show_held_locks(prev
);
4124 if (irqs_disabled())
4125 print_irqtrace_events(prev
);
4134 * Various schedule()-time debugging checks and statistics:
4136 static inline void schedule_debug(struct task_struct
*prev
)
4139 * Test if we are atomic. Since do_exit() needs to call into
4140 * schedule() atomically, we ignore that path for now.
4141 * Otherwise, whine if we are scheduling when we should not be.
4143 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4144 __schedule_bug(prev
);
4146 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4148 schedstat_inc(this_rq(), sched_count
);
4149 #ifdef CONFIG_SCHEDSTATS
4150 if (unlikely(prev
->lock_depth
>= 0)) {
4151 schedstat_inc(this_rq(), bkl_count
);
4152 schedstat_inc(prev
, sched_info
.bkl_count
);
4158 * Pick up the highest-prio task:
4160 static inline struct task_struct
*
4161 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4163 const struct sched_class
*class;
4164 struct task_struct
*p
;
4167 * Optimization: we know that if all tasks are in
4168 * the fair class we can call that function directly:
4170 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4171 p
= fair_sched_class
.pick_next_task(rq
);
4176 class = sched_class_highest
;
4178 p
= class->pick_next_task(rq
);
4182 * Will never be NULL as the idle class always
4183 * returns a non-NULL p:
4185 class = class->next
;
4190 * schedule() is the main scheduler function.
4192 asmlinkage
void __sched
schedule(void)
4194 struct task_struct
*prev
, *next
;
4195 unsigned long *switch_count
;
4201 cpu
= smp_processor_id();
4205 switch_count
= &prev
->nivcsw
;
4207 release_kernel_lock(prev
);
4208 need_resched_nonpreemptible
:
4210 schedule_debug(prev
);
4215 * Do the rq-clock update outside the rq lock:
4217 local_irq_disable();
4218 update_rq_clock(rq
);
4219 spin_lock(&rq
->lock
);
4220 clear_tsk_need_resched(prev
);
4222 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4223 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4224 prev
->state
= TASK_RUNNING
;
4226 deactivate_task(rq
, prev
, 1);
4227 switch_count
= &prev
->nvcsw
;
4231 if (prev
->sched_class
->pre_schedule
)
4232 prev
->sched_class
->pre_schedule(rq
, prev
);
4235 if (unlikely(!rq
->nr_running
))
4236 idle_balance(cpu
, rq
);
4238 prev
->sched_class
->put_prev_task(rq
, prev
);
4239 next
= pick_next_task(rq
, prev
);
4241 if (likely(prev
!= next
)) {
4242 sched_info_switch(prev
, next
);
4248 context_switch(rq
, prev
, next
); /* unlocks the rq */
4250 * the context switch might have flipped the stack from under
4251 * us, hence refresh the local variables.
4253 cpu
= smp_processor_id();
4256 spin_unlock_irq(&rq
->lock
);
4260 if (unlikely(reacquire_kernel_lock(current
) < 0))
4261 goto need_resched_nonpreemptible
;
4263 preempt_enable_no_resched();
4264 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4267 EXPORT_SYMBOL(schedule
);
4269 #ifdef CONFIG_PREEMPT
4271 * this is the entry point to schedule() from in-kernel preemption
4272 * off of preempt_enable. Kernel preemptions off return from interrupt
4273 * occur there and call schedule directly.
4275 asmlinkage
void __sched
preempt_schedule(void)
4277 struct thread_info
*ti
= current_thread_info();
4280 * If there is a non-zero preempt_count or interrupts are disabled,
4281 * we do not want to preempt the current task. Just return..
4283 if (likely(ti
->preempt_count
|| irqs_disabled()))
4287 add_preempt_count(PREEMPT_ACTIVE
);
4289 sub_preempt_count(PREEMPT_ACTIVE
);
4292 * Check again in case we missed a preemption opportunity
4293 * between schedule and now.
4296 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4298 EXPORT_SYMBOL(preempt_schedule
);
4301 * this is the entry point to schedule() from kernel preemption
4302 * off of irq context.
4303 * Note, that this is called and return with irqs disabled. This will
4304 * protect us against recursive calling from irq.
4306 asmlinkage
void __sched
preempt_schedule_irq(void)
4308 struct thread_info
*ti
= current_thread_info();
4310 /* Catch callers which need to be fixed */
4311 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4314 add_preempt_count(PREEMPT_ACTIVE
);
4317 local_irq_disable();
4318 sub_preempt_count(PREEMPT_ACTIVE
);
4321 * Check again in case we missed a preemption opportunity
4322 * between schedule and now.
4325 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4328 #endif /* CONFIG_PREEMPT */
4330 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4333 return try_to_wake_up(curr
->private, mode
, sync
);
4335 EXPORT_SYMBOL(default_wake_function
);
4338 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4339 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4340 * number) then we wake all the non-exclusive tasks and one exclusive task.
4342 * There are circumstances in which we can try to wake a task which has already
4343 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4344 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4346 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4347 int nr_exclusive
, int sync
, void *key
)
4349 wait_queue_t
*curr
, *next
;
4351 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4352 unsigned flags
= curr
->flags
;
4354 if (curr
->func(curr
, mode
, sync
, key
) &&
4355 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4361 * __wake_up - wake up threads blocked on a waitqueue.
4363 * @mode: which threads
4364 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4365 * @key: is directly passed to the wakeup function
4367 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4368 int nr_exclusive
, void *key
)
4370 unsigned long flags
;
4372 spin_lock_irqsave(&q
->lock
, flags
);
4373 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4374 spin_unlock_irqrestore(&q
->lock
, flags
);
4376 EXPORT_SYMBOL(__wake_up
);
4379 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4381 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4383 __wake_up_common(q
, mode
, 1, 0, NULL
);
4387 * __wake_up_sync - wake up threads blocked on a waitqueue.
4389 * @mode: which threads
4390 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4392 * The sync wakeup differs that the waker knows that it will schedule
4393 * away soon, so while the target thread will be woken up, it will not
4394 * be migrated to another CPU - ie. the two threads are 'synchronized'
4395 * with each other. This can prevent needless bouncing between CPUs.
4397 * On UP it can prevent extra preemption.
4400 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4402 unsigned long flags
;
4408 if (unlikely(!nr_exclusive
))
4411 spin_lock_irqsave(&q
->lock
, flags
);
4412 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4413 spin_unlock_irqrestore(&q
->lock
, flags
);
4415 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4417 void complete(struct completion
*x
)
4419 unsigned long flags
;
4421 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4423 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4424 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4426 EXPORT_SYMBOL(complete
);
4428 void complete_all(struct completion
*x
)
4430 unsigned long flags
;
4432 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4433 x
->done
+= UINT_MAX
/2;
4434 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4435 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4437 EXPORT_SYMBOL(complete_all
);
4439 static inline long __sched
4440 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4443 DECLARE_WAITQUEUE(wait
, current
);
4445 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4446 __add_wait_queue_tail(&x
->wait
, &wait
);
4448 if ((state
== TASK_INTERRUPTIBLE
&&
4449 signal_pending(current
)) ||
4450 (state
== TASK_KILLABLE
&&
4451 fatal_signal_pending(current
))) {
4452 __remove_wait_queue(&x
->wait
, &wait
);
4453 return -ERESTARTSYS
;
4455 __set_current_state(state
);
4456 spin_unlock_irq(&x
->wait
.lock
);
4457 timeout
= schedule_timeout(timeout
);
4458 spin_lock_irq(&x
->wait
.lock
);
4460 __remove_wait_queue(&x
->wait
, &wait
);
4464 __remove_wait_queue(&x
->wait
, &wait
);
4471 wait_for_common(struct completion
*x
, long timeout
, int state
)
4475 spin_lock_irq(&x
->wait
.lock
);
4476 timeout
= do_wait_for_common(x
, timeout
, state
);
4477 spin_unlock_irq(&x
->wait
.lock
);
4481 void __sched
wait_for_completion(struct completion
*x
)
4483 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4485 EXPORT_SYMBOL(wait_for_completion
);
4487 unsigned long __sched
4488 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4490 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4492 EXPORT_SYMBOL(wait_for_completion_timeout
);
4494 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4496 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4497 if (t
== -ERESTARTSYS
)
4501 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4503 unsigned long __sched
4504 wait_for_completion_interruptible_timeout(struct completion
*x
,
4505 unsigned long timeout
)
4507 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4509 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4511 int __sched
wait_for_completion_killable(struct completion
*x
)
4513 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4514 if (t
== -ERESTARTSYS
)
4518 EXPORT_SYMBOL(wait_for_completion_killable
);
4521 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4523 unsigned long flags
;
4526 init_waitqueue_entry(&wait
, current
);
4528 __set_current_state(state
);
4530 spin_lock_irqsave(&q
->lock
, flags
);
4531 __add_wait_queue(q
, &wait
);
4532 spin_unlock(&q
->lock
);
4533 timeout
= schedule_timeout(timeout
);
4534 spin_lock_irq(&q
->lock
);
4535 __remove_wait_queue(q
, &wait
);
4536 spin_unlock_irqrestore(&q
->lock
, flags
);
4541 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4543 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4545 EXPORT_SYMBOL(interruptible_sleep_on
);
4548 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4550 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4552 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4554 void __sched
sleep_on(wait_queue_head_t
*q
)
4556 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4558 EXPORT_SYMBOL(sleep_on
);
4560 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4562 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4564 EXPORT_SYMBOL(sleep_on_timeout
);
4566 #ifdef CONFIG_RT_MUTEXES
4569 * rt_mutex_setprio - set the current priority of a task
4571 * @prio: prio value (kernel-internal form)
4573 * This function changes the 'effective' priority of a task. It does
4574 * not touch ->normal_prio like __setscheduler().
4576 * Used by the rt_mutex code to implement priority inheritance logic.
4578 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4580 unsigned long flags
;
4581 int oldprio
, on_rq
, running
;
4583 const struct sched_class
*prev_class
= p
->sched_class
;
4585 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4587 rq
= task_rq_lock(p
, &flags
);
4588 update_rq_clock(rq
);
4591 on_rq
= p
->se
.on_rq
;
4592 running
= task_current(rq
, p
);
4594 dequeue_task(rq
, p
, 0);
4596 p
->sched_class
->put_prev_task(rq
, p
);
4599 p
->sched_class
= &rt_sched_class
;
4601 p
->sched_class
= &fair_sched_class
;
4606 p
->sched_class
->set_curr_task(rq
);
4608 enqueue_task(rq
, p
, 0);
4610 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4612 task_rq_unlock(rq
, &flags
);
4617 void set_user_nice(struct task_struct
*p
, long nice
)
4619 int old_prio
, delta
, on_rq
;
4620 unsigned long flags
;
4623 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4626 * We have to be careful, if called from sys_setpriority(),
4627 * the task might be in the middle of scheduling on another CPU.
4629 rq
= task_rq_lock(p
, &flags
);
4630 update_rq_clock(rq
);
4632 * The RT priorities are set via sched_setscheduler(), but we still
4633 * allow the 'normal' nice value to be set - but as expected
4634 * it wont have any effect on scheduling until the task is
4635 * SCHED_FIFO/SCHED_RR:
4637 if (task_has_rt_policy(p
)) {
4638 p
->static_prio
= NICE_TO_PRIO(nice
);
4641 on_rq
= p
->se
.on_rq
;
4643 dequeue_task(rq
, p
, 0);
4647 p
->static_prio
= NICE_TO_PRIO(nice
);
4650 p
->prio
= effective_prio(p
);
4651 delta
= p
->prio
- old_prio
;
4654 enqueue_task(rq
, p
, 0);
4657 * If the task increased its priority or is running and
4658 * lowered its priority, then reschedule its CPU:
4660 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4661 resched_task(rq
->curr
);
4664 task_rq_unlock(rq
, &flags
);
4666 EXPORT_SYMBOL(set_user_nice
);
4669 * can_nice - check if a task can reduce its nice value
4673 int can_nice(const struct task_struct
*p
, const int nice
)
4675 /* convert nice value [19,-20] to rlimit style value [1,40] */
4676 int nice_rlim
= 20 - nice
;
4678 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4679 capable(CAP_SYS_NICE
));
4682 #ifdef __ARCH_WANT_SYS_NICE
4685 * sys_nice - change the priority of the current process.
4686 * @increment: priority increment
4688 * sys_setpriority is a more generic, but much slower function that
4689 * does similar things.
4691 asmlinkage
long sys_nice(int increment
)
4696 * Setpriority might change our priority at the same moment.
4697 * We don't have to worry. Conceptually one call occurs first
4698 * and we have a single winner.
4700 if (increment
< -40)
4705 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4711 if (increment
< 0 && !can_nice(current
, nice
))
4714 retval
= security_task_setnice(current
, nice
);
4718 set_user_nice(current
, nice
);
4725 * task_prio - return the priority value of a given task.
4726 * @p: the task in question.
4728 * This is the priority value as seen by users in /proc.
4729 * RT tasks are offset by -200. Normal tasks are centered
4730 * around 0, value goes from -16 to +15.
4732 int task_prio(const struct task_struct
*p
)
4734 return p
->prio
- MAX_RT_PRIO
;
4738 * task_nice - return the nice value of a given task.
4739 * @p: the task in question.
4741 int task_nice(const struct task_struct
*p
)
4743 return TASK_NICE(p
);
4745 EXPORT_SYMBOL(task_nice
);
4748 * idle_cpu - is a given cpu idle currently?
4749 * @cpu: the processor in question.
4751 int idle_cpu(int cpu
)
4753 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4757 * idle_task - return the idle task for a given cpu.
4758 * @cpu: the processor in question.
4760 struct task_struct
*idle_task(int cpu
)
4762 return cpu_rq(cpu
)->idle
;
4766 * find_process_by_pid - find a process with a matching PID value.
4767 * @pid: the pid in question.
4769 static struct task_struct
*find_process_by_pid(pid_t pid
)
4771 return pid
? find_task_by_vpid(pid
) : current
;
4774 /* Actually do priority change: must hold rq lock. */
4776 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
4778 BUG_ON(p
->se
.on_rq
);
4781 switch (p
->policy
) {
4785 p
->sched_class
= &fair_sched_class
;
4789 p
->sched_class
= &rt_sched_class
;
4793 p
->rt_priority
= prio
;
4794 p
->normal_prio
= normal_prio(p
);
4795 /* we are holding p->pi_lock already */
4796 p
->prio
= rt_mutex_getprio(p
);
4801 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4802 * @p: the task in question.
4803 * @policy: new policy.
4804 * @param: structure containing the new RT priority.
4806 * NOTE that the task may be already dead.
4808 int sched_setscheduler(struct task_struct
*p
, int policy
,
4809 struct sched_param
*param
)
4811 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
4812 unsigned long flags
;
4813 const struct sched_class
*prev_class
= p
->sched_class
;
4816 /* may grab non-irq protected spin_locks */
4817 BUG_ON(in_interrupt());
4819 /* double check policy once rq lock held */
4821 policy
= oldpolicy
= p
->policy
;
4822 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4823 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
4824 policy
!= SCHED_IDLE
)
4827 * Valid priorities for SCHED_FIFO and SCHED_RR are
4828 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
4829 * SCHED_BATCH and SCHED_IDLE is 0.
4831 if (param
->sched_priority
< 0 ||
4832 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4833 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4835 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
4839 * Allow unprivileged RT tasks to decrease priority:
4841 if (!capable(CAP_SYS_NICE
)) {
4842 if (rt_policy(policy
)) {
4843 unsigned long rlim_rtprio
;
4845 if (!lock_task_sighand(p
, &flags
))
4847 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4848 unlock_task_sighand(p
, &flags
);
4850 /* can't set/change the rt policy */
4851 if (policy
!= p
->policy
&& !rlim_rtprio
)
4854 /* can't increase priority */
4855 if (param
->sched_priority
> p
->rt_priority
&&
4856 param
->sched_priority
> rlim_rtprio
)
4860 * Like positive nice levels, dont allow tasks to
4861 * move out of SCHED_IDLE either:
4863 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
4866 /* can't change other user's priorities */
4867 if ((current
->euid
!= p
->euid
) &&
4868 (current
->euid
!= p
->uid
))
4872 #ifdef CONFIG_RT_GROUP_SCHED
4874 * Do not allow realtime tasks into groups that have no runtime
4877 if (rt_policy(policy
) && task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
4881 retval
= security_task_setscheduler(p
, policy
, param
);
4885 * make sure no PI-waiters arrive (or leave) while we are
4886 * changing the priority of the task:
4888 spin_lock_irqsave(&p
->pi_lock
, flags
);
4890 * To be able to change p->policy safely, the apropriate
4891 * runqueue lock must be held.
4893 rq
= __task_rq_lock(p
);
4894 /* recheck policy now with rq lock held */
4895 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4896 policy
= oldpolicy
= -1;
4897 __task_rq_unlock(rq
);
4898 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4901 update_rq_clock(rq
);
4902 on_rq
= p
->se
.on_rq
;
4903 running
= task_current(rq
, p
);
4905 deactivate_task(rq
, p
, 0);
4907 p
->sched_class
->put_prev_task(rq
, p
);
4910 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
4913 p
->sched_class
->set_curr_task(rq
);
4915 activate_task(rq
, p
, 0);
4917 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4919 __task_rq_unlock(rq
);
4920 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4922 rt_mutex_adjust_pi(p
);
4926 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4929 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4931 struct sched_param lparam
;
4932 struct task_struct
*p
;
4935 if (!param
|| pid
< 0)
4937 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4942 p
= find_process_by_pid(pid
);
4944 retval
= sched_setscheduler(p
, policy
, &lparam
);
4951 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4952 * @pid: the pid in question.
4953 * @policy: new policy.
4954 * @param: structure containing the new RT priority.
4957 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4959 /* negative values for policy are not valid */
4963 return do_sched_setscheduler(pid
, policy
, param
);
4967 * sys_sched_setparam - set/change the RT priority of a thread
4968 * @pid: the pid in question.
4969 * @param: structure containing the new RT priority.
4971 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4973 return do_sched_setscheduler(pid
, -1, param
);
4977 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4978 * @pid: the pid in question.
4980 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4982 struct task_struct
*p
;
4989 read_lock(&tasklist_lock
);
4990 p
= find_process_by_pid(pid
);
4992 retval
= security_task_getscheduler(p
);
4996 read_unlock(&tasklist_lock
);
5001 * sys_sched_getscheduler - get the RT priority of a thread
5002 * @pid: the pid in question.
5003 * @param: structure containing the RT priority.
5005 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5007 struct sched_param lp
;
5008 struct task_struct
*p
;
5011 if (!param
|| pid
< 0)
5014 read_lock(&tasklist_lock
);
5015 p
= find_process_by_pid(pid
);
5020 retval
= security_task_getscheduler(p
);
5024 lp
.sched_priority
= p
->rt_priority
;
5025 read_unlock(&tasklist_lock
);
5028 * This one might sleep, we cannot do it with a spinlock held ...
5030 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5035 read_unlock(&tasklist_lock
);
5039 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5041 cpumask_t cpus_allowed
;
5042 cpumask_t new_mask
= *in_mask
;
5043 struct task_struct
*p
;
5047 read_lock(&tasklist_lock
);
5049 p
= find_process_by_pid(pid
);
5051 read_unlock(&tasklist_lock
);
5057 * It is not safe to call set_cpus_allowed with the
5058 * tasklist_lock held. We will bump the task_struct's
5059 * usage count and then drop tasklist_lock.
5062 read_unlock(&tasklist_lock
);
5065 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
5066 !capable(CAP_SYS_NICE
))
5069 retval
= security_task_setscheduler(p
, 0, NULL
);
5073 cpuset_cpus_allowed(p
, &cpus_allowed
);
5074 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5076 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5079 cpuset_cpus_allowed(p
, &cpus_allowed
);
5080 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5082 * We must have raced with a concurrent cpuset
5083 * update. Just reset the cpus_allowed to the
5084 * cpuset's cpus_allowed
5086 new_mask
= cpus_allowed
;
5096 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5097 cpumask_t
*new_mask
)
5099 if (len
< sizeof(cpumask_t
)) {
5100 memset(new_mask
, 0, sizeof(cpumask_t
));
5101 } else if (len
> sizeof(cpumask_t
)) {
5102 len
= sizeof(cpumask_t
);
5104 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5108 * sys_sched_setaffinity - set the cpu affinity of a process
5109 * @pid: pid of the process
5110 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5111 * @user_mask_ptr: user-space pointer to the new cpu mask
5113 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5114 unsigned long __user
*user_mask_ptr
)
5119 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5123 return sched_setaffinity(pid
, &new_mask
);
5127 * Represents all cpu's present in the system
5128 * In systems capable of hotplug, this map could dynamically grow
5129 * as new cpu's are detected in the system via any platform specific
5130 * method, such as ACPI for e.g.
5133 cpumask_t cpu_present_map __read_mostly
;
5134 EXPORT_SYMBOL(cpu_present_map
);
5137 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
5138 EXPORT_SYMBOL(cpu_online_map
);
5140 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
5141 EXPORT_SYMBOL(cpu_possible_map
);
5144 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5146 struct task_struct
*p
;
5150 read_lock(&tasklist_lock
);
5153 p
= find_process_by_pid(pid
);
5157 retval
= security_task_getscheduler(p
);
5161 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5164 read_unlock(&tasklist_lock
);
5171 * sys_sched_getaffinity - get the cpu affinity of a process
5172 * @pid: pid of the process
5173 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5174 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5176 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5177 unsigned long __user
*user_mask_ptr
)
5182 if (len
< sizeof(cpumask_t
))
5185 ret
= sched_getaffinity(pid
, &mask
);
5189 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5192 return sizeof(cpumask_t
);
5196 * sys_sched_yield - yield the current processor to other threads.
5198 * This function yields the current CPU to other tasks. If there are no
5199 * other threads running on this CPU then this function will return.
5201 asmlinkage
long sys_sched_yield(void)
5203 struct rq
*rq
= this_rq_lock();
5205 schedstat_inc(rq
, yld_count
);
5206 current
->sched_class
->yield_task(rq
);
5209 * Since we are going to call schedule() anyway, there's
5210 * no need to preempt or enable interrupts:
5212 __release(rq
->lock
);
5213 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5214 _raw_spin_unlock(&rq
->lock
);
5215 preempt_enable_no_resched();
5222 static void __cond_resched(void)
5224 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5225 __might_sleep(__FILE__
, __LINE__
);
5228 * The BKS might be reacquired before we have dropped
5229 * PREEMPT_ACTIVE, which could trigger a second
5230 * cond_resched() call.
5233 add_preempt_count(PREEMPT_ACTIVE
);
5235 sub_preempt_count(PREEMPT_ACTIVE
);
5236 } while (need_resched());
5239 int __sched
_cond_resched(void)
5241 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5242 system_state
== SYSTEM_RUNNING
) {
5248 EXPORT_SYMBOL(_cond_resched
);
5251 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5252 * call schedule, and on return reacquire the lock.
5254 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5255 * operations here to prevent schedule() from being called twice (once via
5256 * spin_unlock(), once by hand).
5258 int cond_resched_lock(spinlock_t
*lock
)
5260 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5263 if (spin_needbreak(lock
) || resched
) {
5265 if (resched
&& need_resched())
5274 EXPORT_SYMBOL(cond_resched_lock
);
5276 int __sched
cond_resched_softirq(void)
5278 BUG_ON(!in_softirq());
5280 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5288 EXPORT_SYMBOL(cond_resched_softirq
);
5291 * yield - yield the current processor to other threads.
5293 * This is a shortcut for kernel-space yielding - it marks the
5294 * thread runnable and calls sys_sched_yield().
5296 void __sched
yield(void)
5298 set_current_state(TASK_RUNNING
);
5301 EXPORT_SYMBOL(yield
);
5304 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5305 * that process accounting knows that this is a task in IO wait state.
5307 * But don't do that if it is a deliberate, throttling IO wait (this task
5308 * has set its backing_dev_info: the queue against which it should throttle)
5310 void __sched
io_schedule(void)
5312 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5314 delayacct_blkio_start();
5315 atomic_inc(&rq
->nr_iowait
);
5317 atomic_dec(&rq
->nr_iowait
);
5318 delayacct_blkio_end();
5320 EXPORT_SYMBOL(io_schedule
);
5322 long __sched
io_schedule_timeout(long timeout
)
5324 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5327 delayacct_blkio_start();
5328 atomic_inc(&rq
->nr_iowait
);
5329 ret
= schedule_timeout(timeout
);
5330 atomic_dec(&rq
->nr_iowait
);
5331 delayacct_blkio_end();
5336 * sys_sched_get_priority_max - return maximum RT priority.
5337 * @policy: scheduling class.
5339 * this syscall returns the maximum rt_priority that can be used
5340 * by a given scheduling class.
5342 asmlinkage
long sys_sched_get_priority_max(int policy
)
5349 ret
= MAX_USER_RT_PRIO
-1;
5361 * sys_sched_get_priority_min - return minimum RT priority.
5362 * @policy: scheduling class.
5364 * this syscall returns the minimum rt_priority that can be used
5365 * by a given scheduling class.
5367 asmlinkage
long sys_sched_get_priority_min(int policy
)
5385 * sys_sched_rr_get_interval - return the default timeslice of a process.
5386 * @pid: pid of the process.
5387 * @interval: userspace pointer to the timeslice value.
5389 * this syscall writes the default timeslice value of a given process
5390 * into the user-space timespec buffer. A value of '0' means infinity.
5393 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5395 struct task_struct
*p
;
5396 unsigned int time_slice
;
5404 read_lock(&tasklist_lock
);
5405 p
= find_process_by_pid(pid
);
5409 retval
= security_task_getscheduler(p
);
5414 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5415 * tasks that are on an otherwise idle runqueue:
5418 if (p
->policy
== SCHED_RR
) {
5419 time_slice
= DEF_TIMESLICE
;
5420 } else if (p
->policy
!= SCHED_FIFO
) {
5421 struct sched_entity
*se
= &p
->se
;
5422 unsigned long flags
;
5425 rq
= task_rq_lock(p
, &flags
);
5426 if (rq
->cfs
.load
.weight
)
5427 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5428 task_rq_unlock(rq
, &flags
);
5430 read_unlock(&tasklist_lock
);
5431 jiffies_to_timespec(time_slice
, &t
);
5432 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5436 read_unlock(&tasklist_lock
);
5440 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5442 void sched_show_task(struct task_struct
*p
)
5444 unsigned long free
= 0;
5447 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5448 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5449 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5450 #if BITS_PER_LONG == 32
5451 if (state
== TASK_RUNNING
)
5452 printk(KERN_CONT
" running ");
5454 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5456 if (state
== TASK_RUNNING
)
5457 printk(KERN_CONT
" running task ");
5459 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5461 #ifdef CONFIG_DEBUG_STACK_USAGE
5463 unsigned long *n
= end_of_stack(p
);
5466 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5469 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5470 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5472 show_stack(p
, NULL
);
5475 void show_state_filter(unsigned long state_filter
)
5477 struct task_struct
*g
, *p
;
5479 #if BITS_PER_LONG == 32
5481 " task PC stack pid father\n");
5484 " task PC stack pid father\n");
5486 read_lock(&tasklist_lock
);
5487 do_each_thread(g
, p
) {
5489 * reset the NMI-timeout, listing all files on a slow
5490 * console might take alot of time:
5492 touch_nmi_watchdog();
5493 if (!state_filter
|| (p
->state
& state_filter
))
5495 } while_each_thread(g
, p
);
5497 touch_all_softlockup_watchdogs();
5499 #ifdef CONFIG_SCHED_DEBUG
5500 sysrq_sched_debug_show();
5502 read_unlock(&tasklist_lock
);
5504 * Only show locks if all tasks are dumped:
5506 if (state_filter
== -1)
5507 debug_show_all_locks();
5510 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5512 idle
->sched_class
= &idle_sched_class
;
5516 * init_idle - set up an idle thread for a given CPU
5517 * @idle: task in question
5518 * @cpu: cpu the idle task belongs to
5520 * NOTE: this function does not set the idle thread's NEED_RESCHED
5521 * flag, to make booting more robust.
5523 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5525 struct rq
*rq
= cpu_rq(cpu
);
5526 unsigned long flags
;
5529 idle
->se
.exec_start
= sched_clock();
5531 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5532 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5533 __set_task_cpu(idle
, cpu
);
5535 spin_lock_irqsave(&rq
->lock
, flags
);
5536 rq
->curr
= rq
->idle
= idle
;
5537 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5540 spin_unlock_irqrestore(&rq
->lock
, flags
);
5542 /* Set the preempt count _outside_ the spinlocks! */
5543 #if defined(CONFIG_PREEMPT)
5544 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5546 task_thread_info(idle
)->preempt_count
= 0;
5549 * The idle tasks have their own, simple scheduling class:
5551 idle
->sched_class
= &idle_sched_class
;
5555 * In a system that switches off the HZ timer nohz_cpu_mask
5556 * indicates which cpus entered this state. This is used
5557 * in the rcu update to wait only for active cpus. For system
5558 * which do not switch off the HZ timer nohz_cpu_mask should
5559 * always be CPU_MASK_NONE.
5561 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5564 * Increase the granularity value when there are more CPUs,
5565 * because with more CPUs the 'effective latency' as visible
5566 * to users decreases. But the relationship is not linear,
5567 * so pick a second-best guess by going with the log2 of the
5570 * This idea comes from the SD scheduler of Con Kolivas:
5572 static inline void sched_init_granularity(void)
5574 unsigned int factor
= 1 + ilog2(num_online_cpus());
5575 const unsigned long limit
= 200000000;
5577 sysctl_sched_min_granularity
*= factor
;
5578 if (sysctl_sched_min_granularity
> limit
)
5579 sysctl_sched_min_granularity
= limit
;
5581 sysctl_sched_latency
*= factor
;
5582 if (sysctl_sched_latency
> limit
)
5583 sysctl_sched_latency
= limit
;
5585 sysctl_sched_wakeup_granularity
*= factor
;
5590 * This is how migration works:
5592 * 1) we queue a struct migration_req structure in the source CPU's
5593 * runqueue and wake up that CPU's migration thread.
5594 * 2) we down() the locked semaphore => thread blocks.
5595 * 3) migration thread wakes up (implicitly it forces the migrated
5596 * thread off the CPU)
5597 * 4) it gets the migration request and checks whether the migrated
5598 * task is still in the wrong runqueue.
5599 * 5) if it's in the wrong runqueue then the migration thread removes
5600 * it and puts it into the right queue.
5601 * 6) migration thread up()s the semaphore.
5602 * 7) we wake up and the migration is done.
5606 * Change a given task's CPU affinity. Migrate the thread to a
5607 * proper CPU and schedule it away if the CPU it's executing on
5608 * is removed from the allowed bitmask.
5610 * NOTE: the caller must have a valid reference to the task, the
5611 * task must not exit() & deallocate itself prematurely. The
5612 * call is not atomic; no spinlocks may be held.
5614 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5616 struct migration_req req
;
5617 unsigned long flags
;
5621 rq
= task_rq_lock(p
, &flags
);
5622 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5627 if (p
->sched_class
->set_cpus_allowed
)
5628 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
5630 p
->cpus_allowed
= *new_mask
;
5631 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
5634 /* Can the task run on the task's current CPU? If so, we're done */
5635 if (cpu_isset(task_cpu(p
), *new_mask
))
5638 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
5639 /* Need help from migration thread: drop lock and wait. */
5640 task_rq_unlock(rq
, &flags
);
5641 wake_up_process(rq
->migration_thread
);
5642 wait_for_completion(&req
.done
);
5643 tlb_migrate_finish(p
->mm
);
5647 task_rq_unlock(rq
, &flags
);
5651 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
5654 * Move (not current) task off this cpu, onto dest cpu. We're doing
5655 * this because either it can't run here any more (set_cpus_allowed()
5656 * away from this CPU, or CPU going down), or because we're
5657 * attempting to rebalance this task on exec (sched_exec).
5659 * So we race with normal scheduler movements, but that's OK, as long
5660 * as the task is no longer on this CPU.
5662 * Returns non-zero if task was successfully migrated.
5664 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5666 struct rq
*rq_dest
, *rq_src
;
5669 if (unlikely(cpu_is_offline(dest_cpu
)))
5672 rq_src
= cpu_rq(src_cpu
);
5673 rq_dest
= cpu_rq(dest_cpu
);
5675 double_rq_lock(rq_src
, rq_dest
);
5676 /* Already moved. */
5677 if (task_cpu(p
) != src_cpu
)
5679 /* Affinity changed (again). */
5680 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5683 on_rq
= p
->se
.on_rq
;
5685 deactivate_task(rq_src
, p
, 0);
5687 set_task_cpu(p
, dest_cpu
);
5689 activate_task(rq_dest
, p
, 0);
5690 check_preempt_curr(rq_dest
, p
);
5694 double_rq_unlock(rq_src
, rq_dest
);
5699 * migration_thread - this is a highprio system thread that performs
5700 * thread migration by bumping thread off CPU then 'pushing' onto
5703 static int migration_thread(void *data
)
5705 int cpu
= (long)data
;
5709 BUG_ON(rq
->migration_thread
!= current
);
5711 set_current_state(TASK_INTERRUPTIBLE
);
5712 while (!kthread_should_stop()) {
5713 struct migration_req
*req
;
5714 struct list_head
*head
;
5716 spin_lock_irq(&rq
->lock
);
5718 if (cpu_is_offline(cpu
)) {
5719 spin_unlock_irq(&rq
->lock
);
5723 if (rq
->active_balance
) {
5724 active_load_balance(rq
, cpu
);
5725 rq
->active_balance
= 0;
5728 head
= &rq
->migration_queue
;
5730 if (list_empty(head
)) {
5731 spin_unlock_irq(&rq
->lock
);
5733 set_current_state(TASK_INTERRUPTIBLE
);
5736 req
= list_entry(head
->next
, struct migration_req
, list
);
5737 list_del_init(head
->next
);
5739 spin_unlock(&rq
->lock
);
5740 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5743 complete(&req
->done
);
5745 __set_current_state(TASK_RUNNING
);
5749 /* Wait for kthread_stop */
5750 set_current_state(TASK_INTERRUPTIBLE
);
5751 while (!kthread_should_stop()) {
5753 set_current_state(TASK_INTERRUPTIBLE
);
5755 __set_current_state(TASK_RUNNING
);
5759 #ifdef CONFIG_HOTPLUG_CPU
5761 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5765 local_irq_disable();
5766 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
5772 * Figure out where task on dead CPU should go, use force if necessary.
5773 * NOTE: interrupts should be disabled by the caller
5775 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5777 unsigned long flags
;
5784 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5785 cpus_and(mask
, mask
, p
->cpus_allowed
);
5786 dest_cpu
= any_online_cpu(mask
);
5788 /* On any allowed CPU? */
5789 if (dest_cpu
>= nr_cpu_ids
)
5790 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5792 /* No more Mr. Nice Guy. */
5793 if (dest_cpu
>= nr_cpu_ids
) {
5794 cpumask_t cpus_allowed
;
5796 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
5798 * Try to stay on the same cpuset, where the
5799 * current cpuset may be a subset of all cpus.
5800 * The cpuset_cpus_allowed_locked() variant of
5801 * cpuset_cpus_allowed() will not block. It must be
5802 * called within calls to cpuset_lock/cpuset_unlock.
5804 rq
= task_rq_lock(p
, &flags
);
5805 p
->cpus_allowed
= cpus_allowed
;
5806 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5807 task_rq_unlock(rq
, &flags
);
5810 * Don't tell them about moving exiting tasks or
5811 * kernel threads (both mm NULL), since they never
5814 if (p
->mm
&& printk_ratelimit()) {
5815 printk(KERN_INFO
"process %d (%s) no "
5816 "longer affine to cpu%d\n",
5817 task_pid_nr(p
), p
->comm
, dead_cpu
);
5820 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
5824 * While a dead CPU has no uninterruptible tasks queued at this point,
5825 * it might still have a nonzero ->nr_uninterruptible counter, because
5826 * for performance reasons the counter is not stricly tracking tasks to
5827 * their home CPUs. So we just add the counter to another CPU's counter,
5828 * to keep the global sum constant after CPU-down:
5830 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5832 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
5833 unsigned long flags
;
5835 local_irq_save(flags
);
5836 double_rq_lock(rq_src
, rq_dest
);
5837 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5838 rq_src
->nr_uninterruptible
= 0;
5839 double_rq_unlock(rq_src
, rq_dest
);
5840 local_irq_restore(flags
);
5843 /* Run through task list and migrate tasks from the dead cpu. */
5844 static void migrate_live_tasks(int src_cpu
)
5846 struct task_struct
*p
, *t
;
5848 read_lock(&tasklist_lock
);
5850 do_each_thread(t
, p
) {
5854 if (task_cpu(p
) == src_cpu
)
5855 move_task_off_dead_cpu(src_cpu
, p
);
5856 } while_each_thread(t
, p
);
5858 read_unlock(&tasklist_lock
);
5862 * Schedules idle task to be the next runnable task on current CPU.
5863 * It does so by boosting its priority to highest possible.
5864 * Used by CPU offline code.
5866 void sched_idle_next(void)
5868 int this_cpu
= smp_processor_id();
5869 struct rq
*rq
= cpu_rq(this_cpu
);
5870 struct task_struct
*p
= rq
->idle
;
5871 unsigned long flags
;
5873 /* cpu has to be offline */
5874 BUG_ON(cpu_online(this_cpu
));
5877 * Strictly not necessary since rest of the CPUs are stopped by now
5878 * and interrupts disabled on the current cpu.
5880 spin_lock_irqsave(&rq
->lock
, flags
);
5882 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5884 update_rq_clock(rq
);
5885 activate_task(rq
, p
, 0);
5887 spin_unlock_irqrestore(&rq
->lock
, flags
);
5891 * Ensures that the idle task is using init_mm right before its cpu goes
5894 void idle_task_exit(void)
5896 struct mm_struct
*mm
= current
->active_mm
;
5898 BUG_ON(cpu_online(smp_processor_id()));
5901 switch_mm(mm
, &init_mm
, current
);
5905 /* called under rq->lock with disabled interrupts */
5906 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5908 struct rq
*rq
= cpu_rq(dead_cpu
);
5910 /* Must be exiting, otherwise would be on tasklist. */
5911 BUG_ON(!p
->exit_state
);
5913 /* Cannot have done final schedule yet: would have vanished. */
5914 BUG_ON(p
->state
== TASK_DEAD
);
5919 * Drop lock around migration; if someone else moves it,
5920 * that's OK. No task can be added to this CPU, so iteration is
5923 spin_unlock_irq(&rq
->lock
);
5924 move_task_off_dead_cpu(dead_cpu
, p
);
5925 spin_lock_irq(&rq
->lock
);
5930 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5931 static void migrate_dead_tasks(unsigned int dead_cpu
)
5933 struct rq
*rq
= cpu_rq(dead_cpu
);
5934 struct task_struct
*next
;
5937 if (!rq
->nr_running
)
5939 update_rq_clock(rq
);
5940 next
= pick_next_task(rq
, rq
->curr
);
5943 migrate_dead(dead_cpu
, next
);
5947 #endif /* CONFIG_HOTPLUG_CPU */
5949 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
5951 static struct ctl_table sd_ctl_dir
[] = {
5953 .procname
= "sched_domain",
5959 static struct ctl_table sd_ctl_root
[] = {
5961 .ctl_name
= CTL_KERN
,
5962 .procname
= "kernel",
5964 .child
= sd_ctl_dir
,
5969 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
5971 struct ctl_table
*entry
=
5972 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
5977 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
5979 struct ctl_table
*entry
;
5982 * In the intermediate directories, both the child directory and
5983 * procname are dynamically allocated and could fail but the mode
5984 * will always be set. In the lowest directory the names are
5985 * static strings and all have proc handlers.
5987 for (entry
= *tablep
; entry
->mode
; entry
++) {
5989 sd_free_ctl_entry(&entry
->child
);
5990 if (entry
->proc_handler
== NULL
)
5991 kfree(entry
->procname
);
5999 set_table_entry(struct ctl_table
*entry
,
6000 const char *procname
, void *data
, int maxlen
,
6001 mode_t mode
, proc_handler
*proc_handler
)
6003 entry
->procname
= procname
;
6005 entry
->maxlen
= maxlen
;
6007 entry
->proc_handler
= proc_handler
;
6010 static struct ctl_table
*
6011 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6013 struct ctl_table
*table
= sd_alloc_ctl_entry(12);
6018 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6019 sizeof(long), 0644, proc_doulongvec_minmax
);
6020 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6021 sizeof(long), 0644, proc_doulongvec_minmax
);
6022 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6023 sizeof(int), 0644, proc_dointvec_minmax
);
6024 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6025 sizeof(int), 0644, proc_dointvec_minmax
);
6026 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6027 sizeof(int), 0644, proc_dointvec_minmax
);
6028 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6029 sizeof(int), 0644, proc_dointvec_minmax
);
6030 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6031 sizeof(int), 0644, proc_dointvec_minmax
);
6032 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6033 sizeof(int), 0644, proc_dointvec_minmax
);
6034 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6035 sizeof(int), 0644, proc_dointvec_minmax
);
6036 set_table_entry(&table
[9], "cache_nice_tries",
6037 &sd
->cache_nice_tries
,
6038 sizeof(int), 0644, proc_dointvec_minmax
);
6039 set_table_entry(&table
[10], "flags", &sd
->flags
,
6040 sizeof(int), 0644, proc_dointvec_minmax
);
6041 /* &table[11] is terminator */
6046 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6048 struct ctl_table
*entry
, *table
;
6049 struct sched_domain
*sd
;
6050 int domain_num
= 0, i
;
6053 for_each_domain(cpu
, sd
)
6055 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6060 for_each_domain(cpu
, sd
) {
6061 snprintf(buf
, 32, "domain%d", i
);
6062 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6064 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6071 static struct ctl_table_header
*sd_sysctl_header
;
6072 static void register_sched_domain_sysctl(void)
6074 int i
, cpu_num
= num_online_cpus();
6075 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6078 WARN_ON(sd_ctl_dir
[0].child
);
6079 sd_ctl_dir
[0].child
= entry
;
6084 for_each_online_cpu(i
) {
6085 snprintf(buf
, 32, "cpu%d", i
);
6086 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6088 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6092 WARN_ON(sd_sysctl_header
);
6093 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6096 /* may be called multiple times per register */
6097 static void unregister_sched_domain_sysctl(void)
6099 if (sd_sysctl_header
)
6100 unregister_sysctl_table(sd_sysctl_header
);
6101 sd_sysctl_header
= NULL
;
6102 if (sd_ctl_dir
[0].child
)
6103 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6106 static void register_sched_domain_sysctl(void)
6109 static void unregister_sched_domain_sysctl(void)
6115 * migration_call - callback that gets triggered when a CPU is added.
6116 * Here we can start up the necessary migration thread for the new CPU.
6118 static int __cpuinit
6119 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6121 struct task_struct
*p
;
6122 int cpu
= (long)hcpu
;
6123 unsigned long flags
;
6128 case CPU_UP_PREPARE
:
6129 case CPU_UP_PREPARE_FROZEN
:
6130 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6133 kthread_bind(p
, cpu
);
6134 /* Must be high prio: stop_machine expects to yield to it. */
6135 rq
= task_rq_lock(p
, &flags
);
6136 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6137 task_rq_unlock(rq
, &flags
);
6138 cpu_rq(cpu
)->migration_thread
= p
;
6142 case CPU_ONLINE_FROZEN
:
6143 /* Strictly unnecessary, as first user will wake it. */
6144 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6146 /* Update our root-domain */
6148 spin_lock_irqsave(&rq
->lock
, flags
);
6150 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6151 cpu_set(cpu
, rq
->rd
->online
);
6153 spin_unlock_irqrestore(&rq
->lock
, flags
);
6156 #ifdef CONFIG_HOTPLUG_CPU
6157 case CPU_UP_CANCELED
:
6158 case CPU_UP_CANCELED_FROZEN
:
6159 if (!cpu_rq(cpu
)->migration_thread
)
6161 /* Unbind it from offline cpu so it can run. Fall thru. */
6162 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6163 any_online_cpu(cpu_online_map
));
6164 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6165 cpu_rq(cpu
)->migration_thread
= NULL
;
6169 case CPU_DEAD_FROZEN
:
6170 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6171 migrate_live_tasks(cpu
);
6173 kthread_stop(rq
->migration_thread
);
6174 rq
->migration_thread
= NULL
;
6175 /* Idle task back to normal (off runqueue, low prio) */
6176 spin_lock_irq(&rq
->lock
);
6177 update_rq_clock(rq
);
6178 deactivate_task(rq
, rq
->idle
, 0);
6179 rq
->idle
->static_prio
= MAX_PRIO
;
6180 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6181 rq
->idle
->sched_class
= &idle_sched_class
;
6182 migrate_dead_tasks(cpu
);
6183 spin_unlock_irq(&rq
->lock
);
6185 migrate_nr_uninterruptible(rq
);
6186 BUG_ON(rq
->nr_running
!= 0);
6189 * No need to migrate the tasks: it was best-effort if
6190 * they didn't take sched_hotcpu_mutex. Just wake up
6193 spin_lock_irq(&rq
->lock
);
6194 while (!list_empty(&rq
->migration_queue
)) {
6195 struct migration_req
*req
;
6197 req
= list_entry(rq
->migration_queue
.next
,
6198 struct migration_req
, list
);
6199 list_del_init(&req
->list
);
6200 complete(&req
->done
);
6202 spin_unlock_irq(&rq
->lock
);
6206 case CPU_DYING_FROZEN
:
6207 /* Update our root-domain */
6209 spin_lock_irqsave(&rq
->lock
, flags
);
6211 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6212 cpu_clear(cpu
, rq
->rd
->online
);
6214 spin_unlock_irqrestore(&rq
->lock
, flags
);
6221 /* Register at highest priority so that task migration (migrate_all_tasks)
6222 * happens before everything else.
6224 static struct notifier_block __cpuinitdata migration_notifier
= {
6225 .notifier_call
= migration_call
,
6229 void __init
migration_init(void)
6231 void *cpu
= (void *)(long)smp_processor_id();
6234 /* Start one for the boot CPU: */
6235 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6236 BUG_ON(err
== NOTIFY_BAD
);
6237 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6238 register_cpu_notifier(&migration_notifier
);
6244 #ifdef CONFIG_SCHED_DEBUG
6246 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6247 cpumask_t
*groupmask
)
6249 struct sched_group
*group
= sd
->groups
;
6252 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6253 cpus_clear(*groupmask
);
6255 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6257 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6258 printk("does not load-balance\n");
6260 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6265 printk(KERN_CONT
"span %s\n", str
);
6267 if (!cpu_isset(cpu
, sd
->span
)) {
6268 printk(KERN_ERR
"ERROR: domain->span does not contain "
6271 if (!cpu_isset(cpu
, group
->cpumask
)) {
6272 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6276 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6280 printk(KERN_ERR
"ERROR: group is NULL\n");
6284 if (!group
->__cpu_power
) {
6285 printk(KERN_CONT
"\n");
6286 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6291 if (!cpus_weight(group
->cpumask
)) {
6292 printk(KERN_CONT
"\n");
6293 printk(KERN_ERR
"ERROR: empty group\n");
6297 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6298 printk(KERN_CONT
"\n");
6299 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6303 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6305 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6306 printk(KERN_CONT
" %s", str
);
6308 group
= group
->next
;
6309 } while (group
!= sd
->groups
);
6310 printk(KERN_CONT
"\n");
6312 if (!cpus_equal(sd
->span
, *groupmask
))
6313 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6315 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6316 printk(KERN_ERR
"ERROR: parent span is not a superset "
6317 "of domain->span\n");
6321 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6323 cpumask_t
*groupmask
;
6327 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6331 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6333 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6335 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6340 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6350 # define sched_domain_debug(sd, cpu) do { } while (0)
6353 static int sd_degenerate(struct sched_domain
*sd
)
6355 if (cpus_weight(sd
->span
) == 1)
6358 /* Following flags need at least 2 groups */
6359 if (sd
->flags
& (SD_LOAD_BALANCE
|
6360 SD_BALANCE_NEWIDLE
|
6364 SD_SHARE_PKG_RESOURCES
)) {
6365 if (sd
->groups
!= sd
->groups
->next
)
6369 /* Following flags don't use groups */
6370 if (sd
->flags
& (SD_WAKE_IDLE
|
6379 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6381 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6383 if (sd_degenerate(parent
))
6386 if (!cpus_equal(sd
->span
, parent
->span
))
6389 /* Does parent contain flags not in child? */
6390 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6391 if (cflags
& SD_WAKE_AFFINE
)
6392 pflags
&= ~SD_WAKE_BALANCE
;
6393 /* Flags needing groups don't count if only 1 group in parent */
6394 if (parent
->groups
== parent
->groups
->next
) {
6395 pflags
&= ~(SD_LOAD_BALANCE
|
6396 SD_BALANCE_NEWIDLE
|
6400 SD_SHARE_PKG_RESOURCES
);
6402 if (~cflags
& pflags
)
6408 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6410 unsigned long flags
;
6411 const struct sched_class
*class;
6413 spin_lock_irqsave(&rq
->lock
, flags
);
6416 struct root_domain
*old_rd
= rq
->rd
;
6418 for (class = sched_class_highest
; class; class = class->next
) {
6419 if (class->leave_domain
)
6420 class->leave_domain(rq
);
6423 cpu_clear(rq
->cpu
, old_rd
->span
);
6424 cpu_clear(rq
->cpu
, old_rd
->online
);
6426 if (atomic_dec_and_test(&old_rd
->refcount
))
6430 atomic_inc(&rd
->refcount
);
6433 cpu_set(rq
->cpu
, rd
->span
);
6434 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6435 cpu_set(rq
->cpu
, rd
->online
);
6437 for (class = sched_class_highest
; class; class = class->next
) {
6438 if (class->join_domain
)
6439 class->join_domain(rq
);
6442 spin_unlock_irqrestore(&rq
->lock
, flags
);
6445 static void init_rootdomain(struct root_domain
*rd
)
6447 memset(rd
, 0, sizeof(*rd
));
6449 cpus_clear(rd
->span
);
6450 cpus_clear(rd
->online
);
6453 static void init_defrootdomain(void)
6455 init_rootdomain(&def_root_domain
);
6456 atomic_set(&def_root_domain
.refcount
, 1);
6459 static struct root_domain
*alloc_rootdomain(void)
6461 struct root_domain
*rd
;
6463 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6467 init_rootdomain(rd
);
6473 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6474 * hold the hotplug lock.
6477 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6479 struct rq
*rq
= cpu_rq(cpu
);
6480 struct sched_domain
*tmp
;
6482 /* Remove the sched domains which do not contribute to scheduling. */
6483 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
6484 struct sched_domain
*parent
= tmp
->parent
;
6487 if (sd_parent_degenerate(tmp
, parent
)) {
6488 tmp
->parent
= parent
->parent
;
6490 parent
->parent
->child
= tmp
;
6494 if (sd
&& sd_degenerate(sd
)) {
6500 sched_domain_debug(sd
, cpu
);
6502 rq_attach_root(rq
, rd
);
6503 rcu_assign_pointer(rq
->sd
, sd
);
6506 /* cpus with isolated domains */
6507 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6509 /* Setup the mask of cpus configured for isolated domains */
6510 static int __init
isolated_cpu_setup(char *str
)
6512 int ints
[NR_CPUS
], i
;
6514 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6515 cpus_clear(cpu_isolated_map
);
6516 for (i
= 1; i
<= ints
[0]; i
++)
6517 if (ints
[i
] < NR_CPUS
)
6518 cpu_set(ints
[i
], cpu_isolated_map
);
6522 __setup("isolcpus=", isolated_cpu_setup
);
6525 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6526 * to a function which identifies what group(along with sched group) a CPU
6527 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6528 * (due to the fact that we keep track of groups covered with a cpumask_t).
6530 * init_sched_build_groups will build a circular linked list of the groups
6531 * covered by the given span, and will set each group's ->cpumask correctly,
6532 * and ->cpu_power to 0.
6535 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6536 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6537 struct sched_group
**sg
,
6538 cpumask_t
*tmpmask
),
6539 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6541 struct sched_group
*first
= NULL
, *last
= NULL
;
6544 cpus_clear(*covered
);
6546 for_each_cpu_mask(i
, *span
) {
6547 struct sched_group
*sg
;
6548 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6551 if (cpu_isset(i
, *covered
))
6554 cpus_clear(sg
->cpumask
);
6555 sg
->__cpu_power
= 0;
6557 for_each_cpu_mask(j
, *span
) {
6558 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6561 cpu_set(j
, *covered
);
6562 cpu_set(j
, sg
->cpumask
);
6573 #define SD_NODES_PER_DOMAIN 16
6578 * find_next_best_node - find the next node to include in a sched_domain
6579 * @node: node whose sched_domain we're building
6580 * @used_nodes: nodes already in the sched_domain
6582 * Find the next node to include in a given scheduling domain. Simply
6583 * finds the closest node not already in the @used_nodes map.
6585 * Should use nodemask_t.
6587 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
6589 int i
, n
, val
, min_val
, best_node
= 0;
6593 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6594 /* Start at @node */
6595 n
= (node
+ i
) % MAX_NUMNODES
;
6597 if (!nr_cpus_node(n
))
6600 /* Skip already used nodes */
6601 if (node_isset(n
, *used_nodes
))
6604 /* Simple min distance search */
6605 val
= node_distance(node
, n
);
6607 if (val
< min_val
) {
6613 node_set(best_node
, *used_nodes
);
6618 * sched_domain_node_span - get a cpumask for a node's sched_domain
6619 * @node: node whose cpumask we're constructing
6620 * @span: resulting cpumask
6622 * Given a node, construct a good cpumask for its sched_domain to span. It
6623 * should be one that prevents unnecessary balancing, but also spreads tasks
6626 static void sched_domain_node_span(int node
, cpumask_t
*span
)
6628 nodemask_t used_nodes
;
6629 node_to_cpumask_ptr(nodemask
, node
);
6633 nodes_clear(used_nodes
);
6635 cpus_or(*span
, *span
, *nodemask
);
6636 node_set(node
, used_nodes
);
6638 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6639 int next_node
= find_next_best_node(node
, &used_nodes
);
6641 node_to_cpumask_ptr_next(nodemask
, next_node
);
6642 cpus_or(*span
, *span
, *nodemask
);
6647 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6650 * SMT sched-domains:
6652 #ifdef CONFIG_SCHED_SMT
6653 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6654 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6657 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6661 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6667 * multi-core sched-domains:
6669 #ifdef CONFIG_SCHED_MC
6670 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6671 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6674 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6676 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6681 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6682 cpus_and(*mask
, *mask
, *cpu_map
);
6683 group
= first_cpu(*mask
);
6685 *sg
= &per_cpu(sched_group_core
, group
);
6688 #elif defined(CONFIG_SCHED_MC)
6690 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6694 *sg
= &per_cpu(sched_group_core
, cpu
);
6699 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6700 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6703 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
6707 #ifdef CONFIG_SCHED_MC
6708 *mask
= cpu_coregroup_map(cpu
);
6709 cpus_and(*mask
, *mask
, *cpu_map
);
6710 group
= first_cpu(*mask
);
6711 #elif defined(CONFIG_SCHED_SMT)
6712 *mask
= per_cpu(cpu_sibling_map
, cpu
);
6713 cpus_and(*mask
, *mask
, *cpu_map
);
6714 group
= first_cpu(*mask
);
6719 *sg
= &per_cpu(sched_group_phys
, group
);
6725 * The init_sched_build_groups can't handle what we want to do with node
6726 * groups, so roll our own. Now each node has its own list of groups which
6727 * gets dynamically allocated.
6729 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6730 static struct sched_group
***sched_group_nodes_bycpu
;
6732 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6733 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6735 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6736 struct sched_group
**sg
, cpumask_t
*nodemask
)
6740 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6741 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6742 group
= first_cpu(*nodemask
);
6745 *sg
= &per_cpu(sched_group_allnodes
, group
);
6749 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6751 struct sched_group
*sg
= group_head
;
6757 for_each_cpu_mask(j
, sg
->cpumask
) {
6758 struct sched_domain
*sd
;
6760 sd
= &per_cpu(phys_domains
, j
);
6761 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6763 * Only add "power" once for each
6769 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
6772 } while (sg
!= group_head
);
6777 /* Free memory allocated for various sched_group structures */
6778 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6782 for_each_cpu_mask(cpu
, *cpu_map
) {
6783 struct sched_group
**sched_group_nodes
6784 = sched_group_nodes_bycpu
[cpu
];
6786 if (!sched_group_nodes
)
6789 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6790 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6792 *nodemask
= node_to_cpumask(i
);
6793 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
6794 if (cpus_empty(*nodemask
))
6804 if (oldsg
!= sched_group_nodes
[i
])
6807 kfree(sched_group_nodes
);
6808 sched_group_nodes_bycpu
[cpu
] = NULL
;
6812 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
6818 * Initialize sched groups cpu_power.
6820 * cpu_power indicates the capacity of sched group, which is used while
6821 * distributing the load between different sched groups in a sched domain.
6822 * Typically cpu_power for all the groups in a sched domain will be same unless
6823 * there are asymmetries in the topology. If there are asymmetries, group
6824 * having more cpu_power will pickup more load compared to the group having
6827 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6828 * the maximum number of tasks a group can handle in the presence of other idle
6829 * or lightly loaded groups in the same sched domain.
6831 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6833 struct sched_domain
*child
;
6834 struct sched_group
*group
;
6836 WARN_ON(!sd
|| !sd
->groups
);
6838 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6843 sd
->groups
->__cpu_power
= 0;
6846 * For perf policy, if the groups in child domain share resources
6847 * (for example cores sharing some portions of the cache hierarchy
6848 * or SMT), then set this domain groups cpu_power such that each group
6849 * can handle only one task, when there are other idle groups in the
6850 * same sched domain.
6852 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6854 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6855 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
6860 * add cpu_power of each child group to this groups cpu_power
6862 group
= child
->groups
;
6864 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
6865 group
= group
->next
;
6866 } while (group
!= child
->groups
);
6870 * Initializers for schedule domains
6871 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
6874 #define SD_INIT(sd, type) sd_init_##type(sd)
6875 #define SD_INIT_FUNC(type) \
6876 static noinline void sd_init_##type(struct sched_domain *sd) \
6878 memset(sd, 0, sizeof(*sd)); \
6879 *sd = SD_##type##_INIT; \
6880 sd->level = SD_LV_##type; \
6885 SD_INIT_FUNC(ALLNODES
)
6888 #ifdef CONFIG_SCHED_SMT
6889 SD_INIT_FUNC(SIBLING
)
6891 #ifdef CONFIG_SCHED_MC
6896 * To minimize stack usage kmalloc room for cpumasks and share the
6897 * space as the usage in build_sched_domains() dictates. Used only
6898 * if the amount of space is significant.
6901 cpumask_t tmpmask
; /* make this one first */
6904 cpumask_t this_sibling_map
;
6905 cpumask_t this_core_map
;
6907 cpumask_t send_covered
;
6910 cpumask_t domainspan
;
6912 cpumask_t notcovered
;
6917 #define SCHED_CPUMASK_ALLOC 1
6918 #define SCHED_CPUMASK_FREE(v) kfree(v)
6919 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
6921 #define SCHED_CPUMASK_ALLOC 0
6922 #define SCHED_CPUMASK_FREE(v)
6923 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
6926 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
6927 ((unsigned long)(a) + offsetof(struct allmasks, v))
6929 static int default_relax_domain_level
= -1;
6931 static int __init
setup_relax_domain_level(char *str
)
6935 val
= simple_strtoul(str
, NULL
, 0);
6936 if (val
< SD_LV_MAX
)
6937 default_relax_domain_level
= val
;
6941 __setup("relax_domain_level=", setup_relax_domain_level
);
6943 static void set_domain_attribute(struct sched_domain
*sd
,
6944 struct sched_domain_attr
*attr
)
6948 if (!attr
|| attr
->relax_domain_level
< 0) {
6949 if (default_relax_domain_level
< 0)
6952 request
= default_relax_domain_level
;
6954 request
= attr
->relax_domain_level
;
6955 if (request
< sd
->level
) {
6956 /* turn off idle balance on this domain */
6957 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
6959 /* turn on idle balance on this domain */
6960 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
6965 * Build sched domains for a given set of cpus and attach the sched domains
6966 * to the individual cpus
6968 static int __build_sched_domains(const cpumask_t
*cpu_map
,
6969 struct sched_domain_attr
*attr
)
6972 struct root_domain
*rd
;
6973 SCHED_CPUMASK_DECLARE(allmasks
);
6976 struct sched_group
**sched_group_nodes
= NULL
;
6977 int sd_allnodes
= 0;
6980 * Allocate the per-node list of sched groups
6982 sched_group_nodes
= kcalloc(MAX_NUMNODES
, sizeof(struct sched_group
*),
6984 if (!sched_group_nodes
) {
6985 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6990 rd
= alloc_rootdomain();
6992 printk(KERN_WARNING
"Cannot alloc root domain\n");
6994 kfree(sched_group_nodes
);
6999 #if SCHED_CPUMASK_ALLOC
7000 /* get space for all scratch cpumask variables */
7001 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7003 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7006 kfree(sched_group_nodes
);
7011 tmpmask
= (cpumask_t
*)allmasks
;
7015 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7019 * Set up domains for cpus specified by the cpu_map.
7021 for_each_cpu_mask(i
, *cpu_map
) {
7022 struct sched_domain
*sd
= NULL
, *p
;
7023 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7025 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7026 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7029 if (cpus_weight(*cpu_map
) >
7030 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7031 sd
= &per_cpu(allnodes_domains
, i
);
7032 SD_INIT(sd
, ALLNODES
);
7033 set_domain_attribute(sd
, attr
);
7034 sd
->span
= *cpu_map
;
7035 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7041 sd
= &per_cpu(node_domains
, i
);
7043 set_domain_attribute(sd
, attr
);
7044 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7048 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7052 sd
= &per_cpu(phys_domains
, i
);
7054 set_domain_attribute(sd
, attr
);
7055 sd
->span
= *nodemask
;
7059 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7061 #ifdef CONFIG_SCHED_MC
7063 sd
= &per_cpu(core_domains
, i
);
7065 set_domain_attribute(sd
, attr
);
7066 sd
->span
= cpu_coregroup_map(i
);
7067 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7070 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7073 #ifdef CONFIG_SCHED_SMT
7075 sd
= &per_cpu(cpu_domains
, i
);
7076 SD_INIT(sd
, SIBLING
);
7077 set_domain_attribute(sd
, attr
);
7078 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7079 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7082 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7086 #ifdef CONFIG_SCHED_SMT
7087 /* Set up CPU (sibling) groups */
7088 for_each_cpu_mask(i
, *cpu_map
) {
7089 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7090 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7092 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7093 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7094 if (i
!= first_cpu(*this_sibling_map
))
7097 init_sched_build_groups(this_sibling_map
, cpu_map
,
7099 send_covered
, tmpmask
);
7103 #ifdef CONFIG_SCHED_MC
7104 /* Set up multi-core groups */
7105 for_each_cpu_mask(i
, *cpu_map
) {
7106 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7107 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7109 *this_core_map
= cpu_coregroup_map(i
);
7110 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7111 if (i
!= first_cpu(*this_core_map
))
7114 init_sched_build_groups(this_core_map
, cpu_map
,
7116 send_covered
, tmpmask
);
7120 /* Set up physical groups */
7121 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7122 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7123 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7125 *nodemask
= node_to_cpumask(i
);
7126 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7127 if (cpus_empty(*nodemask
))
7130 init_sched_build_groups(nodemask
, cpu_map
,
7132 send_covered
, tmpmask
);
7136 /* Set up node groups */
7138 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7140 init_sched_build_groups(cpu_map
, cpu_map
,
7141 &cpu_to_allnodes_group
,
7142 send_covered
, tmpmask
);
7145 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
7146 /* Set up node groups */
7147 struct sched_group
*sg
, *prev
;
7148 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7149 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7150 SCHED_CPUMASK_VAR(covered
, allmasks
);
7153 *nodemask
= node_to_cpumask(i
);
7154 cpus_clear(*covered
);
7156 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7157 if (cpus_empty(*nodemask
)) {
7158 sched_group_nodes
[i
] = NULL
;
7162 sched_domain_node_span(i
, domainspan
);
7163 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7165 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7167 printk(KERN_WARNING
"Can not alloc domain group for "
7171 sched_group_nodes
[i
] = sg
;
7172 for_each_cpu_mask(j
, *nodemask
) {
7173 struct sched_domain
*sd
;
7175 sd
= &per_cpu(node_domains
, j
);
7178 sg
->__cpu_power
= 0;
7179 sg
->cpumask
= *nodemask
;
7181 cpus_or(*covered
, *covered
, *nodemask
);
7184 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
7185 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7186 int n
= (i
+ j
) % MAX_NUMNODES
;
7187 node_to_cpumask_ptr(pnodemask
, n
);
7189 cpus_complement(*notcovered
, *covered
);
7190 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7191 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7192 if (cpus_empty(*tmpmask
))
7195 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7196 if (cpus_empty(*tmpmask
))
7199 sg
= kmalloc_node(sizeof(struct sched_group
),
7203 "Can not alloc domain group for node %d\n", j
);
7206 sg
->__cpu_power
= 0;
7207 sg
->cpumask
= *tmpmask
;
7208 sg
->next
= prev
->next
;
7209 cpus_or(*covered
, *covered
, *tmpmask
);
7216 /* Calculate CPU power for physical packages and nodes */
7217 #ifdef CONFIG_SCHED_SMT
7218 for_each_cpu_mask(i
, *cpu_map
) {
7219 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7221 init_sched_groups_power(i
, sd
);
7224 #ifdef CONFIG_SCHED_MC
7225 for_each_cpu_mask(i
, *cpu_map
) {
7226 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7228 init_sched_groups_power(i
, sd
);
7232 for_each_cpu_mask(i
, *cpu_map
) {
7233 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7235 init_sched_groups_power(i
, sd
);
7239 for (i
= 0; i
< MAX_NUMNODES
; i
++)
7240 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7243 struct sched_group
*sg
;
7245 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7247 init_numa_sched_groups_power(sg
);
7251 /* Attach the domains */
7252 for_each_cpu_mask(i
, *cpu_map
) {
7253 struct sched_domain
*sd
;
7254 #ifdef CONFIG_SCHED_SMT
7255 sd
= &per_cpu(cpu_domains
, i
);
7256 #elif defined(CONFIG_SCHED_MC)
7257 sd
= &per_cpu(core_domains
, i
);
7259 sd
= &per_cpu(phys_domains
, i
);
7261 cpu_attach_domain(sd
, rd
, i
);
7264 SCHED_CPUMASK_FREE((void *)allmasks
);
7269 free_sched_groups(cpu_map
, tmpmask
);
7270 SCHED_CPUMASK_FREE((void *)allmasks
);
7275 static int build_sched_domains(const cpumask_t
*cpu_map
)
7277 return __build_sched_domains(cpu_map
, NULL
);
7280 static cpumask_t
*doms_cur
; /* current sched domains */
7281 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7282 static struct sched_domain_attr
*dattr_cur
;
7283 /* attribues of custom domains in 'doms_cur' */
7286 * Special case: If a kmalloc of a doms_cur partition (array of
7287 * cpumask_t) fails, then fallback to a single sched domain,
7288 * as determined by the single cpumask_t fallback_doms.
7290 static cpumask_t fallback_doms
;
7292 void __attribute__((weak
)) arch_update_cpu_topology(void)
7297 * Free current domain masks.
7298 * Called after all cpus are attached to NULL domain.
7300 static void free_sched_domains(void)
7303 if (doms_cur
!= &fallback_doms
)
7305 doms_cur
= &fallback_doms
;
7309 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7310 * For now this just excludes isolated cpus, but could be used to
7311 * exclude other special cases in the future.
7313 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7317 arch_update_cpu_topology();
7319 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7321 doms_cur
= &fallback_doms
;
7322 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7324 err
= build_sched_domains(doms_cur
);
7325 register_sched_domain_sysctl();
7330 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7333 free_sched_groups(cpu_map
, tmpmask
);
7337 * Detach sched domains from a group of cpus specified in cpu_map
7338 * These cpus will now be attached to the NULL domain
7340 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7345 unregister_sched_domain_sysctl();
7347 for_each_cpu_mask(i
, *cpu_map
)
7348 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7349 synchronize_sched();
7350 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7353 /* handle null as "default" */
7354 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7355 struct sched_domain_attr
*new, int idx_new
)
7357 struct sched_domain_attr tmp
;
7364 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7365 new ? (new + idx_new
) : &tmp
,
7366 sizeof(struct sched_domain_attr
));
7370 * Partition sched domains as specified by the 'ndoms_new'
7371 * cpumasks in the array doms_new[] of cpumasks. This compares
7372 * doms_new[] to the current sched domain partitioning, doms_cur[].
7373 * It destroys each deleted domain and builds each new domain.
7375 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7376 * The masks don't intersect (don't overlap.) We should setup one
7377 * sched domain for each mask. CPUs not in any of the cpumasks will
7378 * not be load balanced. If the same cpumask appears both in the
7379 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7382 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7383 * ownership of it and will kfree it when done with it. If the caller
7384 * failed the kmalloc call, then it can pass in doms_new == NULL,
7385 * and partition_sched_domains() will fallback to the single partition
7388 * Call with hotplug lock held
7390 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7391 struct sched_domain_attr
*dattr_new
)
7395 mutex_lock(&sched_domains_mutex
);
7397 /* always unregister in case we don't destroy any domains */
7398 unregister_sched_domain_sysctl();
7400 if (doms_new
== NULL
) {
7402 doms_new
= &fallback_doms
;
7403 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7407 /* Destroy deleted domains */
7408 for (i
= 0; i
< ndoms_cur
; i
++) {
7409 for (j
= 0; j
< ndoms_new
; j
++) {
7410 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7411 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7414 /* no match - a current sched domain not in new doms_new[] */
7415 detach_destroy_domains(doms_cur
+ i
);
7420 /* Build new domains */
7421 for (i
= 0; i
< ndoms_new
; i
++) {
7422 for (j
= 0; j
< ndoms_cur
; j
++) {
7423 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7424 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7427 /* no match - add a new doms_new */
7428 __build_sched_domains(doms_new
+ i
,
7429 dattr_new
? dattr_new
+ i
: NULL
);
7434 /* Remember the new sched domains */
7435 if (doms_cur
!= &fallback_doms
)
7437 kfree(dattr_cur
); /* kfree(NULL) is safe */
7438 doms_cur
= doms_new
;
7439 dattr_cur
= dattr_new
;
7440 ndoms_cur
= ndoms_new
;
7442 register_sched_domain_sysctl();
7444 mutex_unlock(&sched_domains_mutex
);
7447 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7448 int arch_reinit_sched_domains(void)
7453 mutex_lock(&sched_domains_mutex
);
7454 detach_destroy_domains(&cpu_online_map
);
7455 free_sched_domains();
7456 err
= arch_init_sched_domains(&cpu_online_map
);
7457 mutex_unlock(&sched_domains_mutex
);
7463 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7467 if (buf
[0] != '0' && buf
[0] != '1')
7471 sched_smt_power_savings
= (buf
[0] == '1');
7473 sched_mc_power_savings
= (buf
[0] == '1');
7475 ret
= arch_reinit_sched_domains();
7477 return ret
? ret
: count
;
7480 #ifdef CONFIG_SCHED_MC
7481 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
7483 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7485 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
7486 const char *buf
, size_t count
)
7488 return sched_power_savings_store(buf
, count
, 0);
7490 static SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
7491 sched_mc_power_savings_store
);
7494 #ifdef CONFIG_SCHED_SMT
7495 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
7497 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7499 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
7500 const char *buf
, size_t count
)
7502 return sched_power_savings_store(buf
, count
, 1);
7504 static SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
7505 sched_smt_power_savings_store
);
7508 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7512 #ifdef CONFIG_SCHED_SMT
7514 err
= sysfs_create_file(&cls
->kset
.kobj
,
7515 &attr_sched_smt_power_savings
.attr
);
7517 #ifdef CONFIG_SCHED_MC
7518 if (!err
&& mc_capable())
7519 err
= sysfs_create_file(&cls
->kset
.kobj
,
7520 &attr_sched_mc_power_savings
.attr
);
7527 * Force a reinitialization of the sched domains hierarchy. The domains
7528 * and groups cannot be updated in place without racing with the balancing
7529 * code, so we temporarily attach all running cpus to the NULL domain
7530 * which will prevent rebalancing while the sched domains are recalculated.
7532 static int update_sched_domains(struct notifier_block
*nfb
,
7533 unsigned long action
, void *hcpu
)
7536 case CPU_UP_PREPARE
:
7537 case CPU_UP_PREPARE_FROZEN
:
7538 case CPU_DOWN_PREPARE
:
7539 case CPU_DOWN_PREPARE_FROZEN
:
7540 detach_destroy_domains(&cpu_online_map
);
7541 free_sched_domains();
7544 case CPU_UP_CANCELED
:
7545 case CPU_UP_CANCELED_FROZEN
:
7546 case CPU_DOWN_FAILED
:
7547 case CPU_DOWN_FAILED_FROZEN
:
7549 case CPU_ONLINE_FROZEN
:
7551 case CPU_DEAD_FROZEN
:
7553 * Fall through and re-initialise the domains.
7560 #ifndef CONFIG_CPUSETS
7562 * Create default domain partitioning if cpusets are disabled.
7563 * Otherwise we let cpusets rebuild the domains based on the
7567 /* The hotplug lock is already held by cpu_up/cpu_down */
7568 arch_init_sched_domains(&cpu_online_map
);
7574 void __init
sched_init_smp(void)
7576 cpumask_t non_isolated_cpus
;
7578 #if defined(CONFIG_NUMA)
7579 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
7581 BUG_ON(sched_group_nodes_bycpu
== NULL
);
7584 mutex_lock(&sched_domains_mutex
);
7585 arch_init_sched_domains(&cpu_online_map
);
7586 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
7587 if (cpus_empty(non_isolated_cpus
))
7588 cpu_set(smp_processor_id(), non_isolated_cpus
);
7589 mutex_unlock(&sched_domains_mutex
);
7591 /* XXX: Theoretical race here - CPU may be hotplugged now */
7592 hotcpu_notifier(update_sched_domains
, 0);
7595 /* Move init over to a non-isolated CPU */
7596 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
7598 sched_init_granularity();
7601 void __init
sched_init_smp(void)
7603 sched_init_granularity();
7605 #endif /* CONFIG_SMP */
7607 int in_sched_functions(unsigned long addr
)
7609 return in_lock_functions(addr
) ||
7610 (addr
>= (unsigned long)__sched_text_start
7611 && addr
< (unsigned long)__sched_text_end
);
7614 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
7616 cfs_rq
->tasks_timeline
= RB_ROOT
;
7617 INIT_LIST_HEAD(&cfs_rq
->tasks
);
7618 #ifdef CONFIG_FAIR_GROUP_SCHED
7621 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
7624 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
7626 struct rt_prio_array
*array
;
7629 array
= &rt_rq
->active
;
7630 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
7631 INIT_LIST_HEAD(array
->queue
+ i
);
7632 __clear_bit(i
, array
->bitmap
);
7634 /* delimiter for bitsearch: */
7635 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
7637 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
7638 rt_rq
->highest_prio
= MAX_RT_PRIO
;
7641 rt_rq
->rt_nr_migratory
= 0;
7642 rt_rq
->overloaded
= 0;
7646 rt_rq
->rt_throttled
= 0;
7647 rt_rq
->rt_runtime
= 0;
7648 spin_lock_init(&rt_rq
->rt_runtime_lock
);
7650 #ifdef CONFIG_RT_GROUP_SCHED
7651 rt_rq
->rt_nr_boosted
= 0;
7656 #ifdef CONFIG_FAIR_GROUP_SCHED
7657 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
7658 struct sched_entity
*se
, int cpu
, int add
,
7659 struct sched_entity
*parent
)
7661 struct rq
*rq
= cpu_rq(cpu
);
7662 tg
->cfs_rq
[cpu
] = cfs_rq
;
7663 init_cfs_rq(cfs_rq
, rq
);
7666 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
7669 /* se could be NULL for init_task_group */
7674 se
->cfs_rq
= &rq
->cfs
;
7676 se
->cfs_rq
= parent
->my_q
;
7679 se
->load
.weight
= tg
->shares
;
7680 se
->load
.inv_weight
= 0;
7681 se
->parent
= parent
;
7685 #ifdef CONFIG_RT_GROUP_SCHED
7686 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
7687 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
7688 struct sched_rt_entity
*parent
)
7690 struct rq
*rq
= cpu_rq(cpu
);
7692 tg
->rt_rq
[cpu
] = rt_rq
;
7693 init_rt_rq(rt_rq
, rq
);
7695 rt_rq
->rt_se
= rt_se
;
7696 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
7698 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
7700 tg
->rt_se
[cpu
] = rt_se
;
7705 rt_se
->rt_rq
= &rq
->rt
;
7707 rt_se
->rt_rq
= parent
->my_q
;
7709 rt_se
->my_q
= rt_rq
;
7710 rt_se
->parent
= parent
;
7711 INIT_LIST_HEAD(&rt_se
->run_list
);
7715 void __init
sched_init(void)
7718 unsigned long alloc_size
= 0, ptr
;
7720 #ifdef CONFIG_FAIR_GROUP_SCHED
7721 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7723 #ifdef CONFIG_RT_GROUP_SCHED
7724 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
7726 #ifdef CONFIG_USER_SCHED
7730 * As sched_init() is called before page_alloc is setup,
7731 * we use alloc_bootmem().
7734 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
7736 #ifdef CONFIG_FAIR_GROUP_SCHED
7737 init_task_group
.se
= (struct sched_entity
**)ptr
;
7738 ptr
+= nr_cpu_ids
* sizeof(void **);
7740 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7741 ptr
+= nr_cpu_ids
* sizeof(void **);
7743 #ifdef CONFIG_USER_SCHED
7744 root_task_group
.se
= (struct sched_entity
**)ptr
;
7745 ptr
+= nr_cpu_ids
* sizeof(void **);
7747 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
7748 ptr
+= nr_cpu_ids
* sizeof(void **);
7751 #ifdef CONFIG_RT_GROUP_SCHED
7752 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7753 ptr
+= nr_cpu_ids
* sizeof(void **);
7755 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7756 ptr
+= nr_cpu_ids
* sizeof(void **);
7758 #ifdef CONFIG_USER_SCHED
7759 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
7760 ptr
+= nr_cpu_ids
* sizeof(void **);
7762 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
7763 ptr
+= nr_cpu_ids
* sizeof(void **);
7769 init_defrootdomain();
7772 init_rt_bandwidth(&def_rt_bandwidth
,
7773 global_rt_period(), global_rt_runtime());
7775 #ifdef CONFIG_RT_GROUP_SCHED
7776 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
7777 global_rt_period(), global_rt_runtime());
7778 #ifdef CONFIG_USER_SCHED
7779 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
7780 global_rt_period(), RUNTIME_INF
);
7784 #ifdef CONFIG_GROUP_SCHED
7785 list_add(&init_task_group
.list
, &task_groups
);
7786 INIT_LIST_HEAD(&init_task_group
.children
);
7788 #ifdef CONFIG_USER_SCHED
7789 INIT_LIST_HEAD(&root_task_group
.children
);
7790 init_task_group
.parent
= &root_task_group
;
7791 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
7795 for_each_possible_cpu(i
) {
7799 spin_lock_init(&rq
->lock
);
7800 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
7802 init_cfs_rq(&rq
->cfs
, rq
);
7803 init_rt_rq(&rq
->rt
, rq
);
7804 #ifdef CONFIG_FAIR_GROUP_SCHED
7805 init_task_group
.shares
= init_task_group_load
;
7806 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
7807 #ifdef CONFIG_CGROUP_SCHED
7809 * How much cpu bandwidth does init_task_group get?
7811 * In case of task-groups formed thr' the cgroup filesystem, it
7812 * gets 100% of the cpu resources in the system. This overall
7813 * system cpu resource is divided among the tasks of
7814 * init_task_group and its child task-groups in a fair manner,
7815 * based on each entity's (task or task-group's) weight
7816 * (se->load.weight).
7818 * In other words, if init_task_group has 10 tasks of weight
7819 * 1024) and two child groups A0 and A1 (of weight 1024 each),
7820 * then A0's share of the cpu resource is:
7822 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
7824 * We achieve this by letting init_task_group's tasks sit
7825 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
7827 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
7828 #elif defined CONFIG_USER_SCHED
7829 root_task_group
.shares
= NICE_0_LOAD
;
7830 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
7832 * In case of task-groups formed thr' the user id of tasks,
7833 * init_task_group represents tasks belonging to root user.
7834 * Hence it forms a sibling of all subsequent groups formed.
7835 * In this case, init_task_group gets only a fraction of overall
7836 * system cpu resource, based on the weight assigned to root
7837 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
7838 * by letting tasks of init_task_group sit in a separate cfs_rq
7839 * (init_cfs_rq) and having one entity represent this group of
7840 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
7842 init_tg_cfs_entry(&init_task_group
,
7843 &per_cpu(init_cfs_rq
, i
),
7844 &per_cpu(init_sched_entity
, i
), i
, 1,
7845 root_task_group
.se
[i
]);
7848 #endif /* CONFIG_FAIR_GROUP_SCHED */
7850 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
7851 #ifdef CONFIG_RT_GROUP_SCHED
7852 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
7853 #ifdef CONFIG_CGROUP_SCHED
7854 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
7855 #elif defined CONFIG_USER_SCHED
7856 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
7857 init_tg_rt_entry(&init_task_group
,
7858 &per_cpu(init_rt_rq
, i
),
7859 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
7860 root_task_group
.rt_se
[i
]);
7864 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
7865 rq
->cpu_load
[j
] = 0;
7869 rq
->active_balance
= 0;
7870 rq
->next_balance
= jiffies
;
7873 rq
->migration_thread
= NULL
;
7874 INIT_LIST_HEAD(&rq
->migration_queue
);
7875 rq_attach_root(rq
, &def_root_domain
);
7878 atomic_set(&rq
->nr_iowait
, 0);
7881 set_load_weight(&init_task
);
7883 #ifdef CONFIG_PREEMPT_NOTIFIERS
7884 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
7888 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
7891 #ifdef CONFIG_RT_MUTEXES
7892 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
7896 * The boot idle thread does lazy MMU switching as well:
7898 atomic_inc(&init_mm
.mm_count
);
7899 enter_lazy_tlb(&init_mm
, current
);
7902 * Make us the idle thread. Technically, schedule() should not be
7903 * called from this thread, however somewhere below it might be,
7904 * but because we are the idle thread, we just pick up running again
7905 * when this runqueue becomes "idle".
7907 init_idle(current
, smp_processor_id());
7909 * During early bootup we pretend to be a normal task:
7911 current
->sched_class
= &fair_sched_class
;
7913 scheduler_running
= 1;
7916 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
7917 void __might_sleep(char *file
, int line
)
7920 static unsigned long prev_jiffy
; /* ratelimiting */
7922 if ((in_atomic() || irqs_disabled()) &&
7923 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
7924 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
7926 prev_jiffy
= jiffies
;
7927 printk(KERN_ERR
"BUG: sleeping function called from invalid"
7928 " context at %s:%d\n", file
, line
);
7929 printk("in_atomic():%d, irqs_disabled():%d\n",
7930 in_atomic(), irqs_disabled());
7931 debug_show_held_locks(current
);
7932 if (irqs_disabled())
7933 print_irqtrace_events(current
);
7938 EXPORT_SYMBOL(__might_sleep
);
7941 #ifdef CONFIG_MAGIC_SYSRQ
7942 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
7946 update_rq_clock(rq
);
7947 on_rq
= p
->se
.on_rq
;
7949 deactivate_task(rq
, p
, 0);
7950 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
7952 activate_task(rq
, p
, 0);
7953 resched_task(rq
->curr
);
7957 void normalize_rt_tasks(void)
7959 struct task_struct
*g
, *p
;
7960 unsigned long flags
;
7963 read_lock_irqsave(&tasklist_lock
, flags
);
7964 do_each_thread(g
, p
) {
7966 * Only normalize user tasks:
7971 p
->se
.exec_start
= 0;
7972 #ifdef CONFIG_SCHEDSTATS
7973 p
->se
.wait_start
= 0;
7974 p
->se
.sleep_start
= 0;
7975 p
->se
.block_start
= 0;
7980 * Renice negative nice level userspace
7983 if (TASK_NICE(p
) < 0 && p
->mm
)
7984 set_user_nice(p
, 0);
7988 spin_lock(&p
->pi_lock
);
7989 rq
= __task_rq_lock(p
);
7991 normalize_task(rq
, p
);
7993 __task_rq_unlock(rq
);
7994 spin_unlock(&p
->pi_lock
);
7995 } while_each_thread(g
, p
);
7997 read_unlock_irqrestore(&tasklist_lock
, flags
);
8000 #endif /* CONFIG_MAGIC_SYSRQ */
8004 * These functions are only useful for the IA64 MCA handling.
8006 * They can only be called when the whole system has been
8007 * stopped - every CPU needs to be quiescent, and no scheduling
8008 * activity can take place. Using them for anything else would
8009 * be a serious bug, and as a result, they aren't even visible
8010 * under any other configuration.
8014 * curr_task - return the current task for a given cpu.
8015 * @cpu: the processor in question.
8017 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8019 struct task_struct
*curr_task(int cpu
)
8021 return cpu_curr(cpu
);
8025 * set_curr_task - set the current task for a given cpu.
8026 * @cpu: the processor in question.
8027 * @p: the task pointer to set.
8029 * Description: This function must only be used when non-maskable interrupts
8030 * are serviced on a separate stack. It allows the architecture to switch the
8031 * notion of the current task on a cpu in a non-blocking manner. This function
8032 * must be called with all CPU's synchronized, and interrupts disabled, the
8033 * and caller must save the original value of the current task (see
8034 * curr_task() above) and restore that value before reenabling interrupts and
8035 * re-starting the system.
8037 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8039 void set_curr_task(int cpu
, struct task_struct
*p
)
8046 #ifdef CONFIG_FAIR_GROUP_SCHED
8047 static void free_fair_sched_group(struct task_group
*tg
)
8051 for_each_possible_cpu(i
) {
8053 kfree(tg
->cfs_rq
[i
]);
8063 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8065 struct cfs_rq
*cfs_rq
;
8066 struct sched_entity
*se
, *parent_se
;
8070 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8073 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8077 tg
->shares
= NICE_0_LOAD
;
8079 for_each_possible_cpu(i
) {
8082 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8083 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8087 se
= kmalloc_node(sizeof(struct sched_entity
),
8088 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8092 parent_se
= parent
? parent
->se
[i
] : NULL
;
8093 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8102 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8104 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8105 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8108 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8110 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8113 static inline void free_fair_sched_group(struct task_group
*tg
)
8118 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8123 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8127 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8132 #ifdef CONFIG_RT_GROUP_SCHED
8133 static void free_rt_sched_group(struct task_group
*tg
)
8137 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8139 for_each_possible_cpu(i
) {
8141 kfree(tg
->rt_rq
[i
]);
8143 kfree(tg
->rt_se
[i
]);
8151 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8153 struct rt_rq
*rt_rq
;
8154 struct sched_rt_entity
*rt_se
, *parent_se
;
8158 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8161 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8165 init_rt_bandwidth(&tg
->rt_bandwidth
,
8166 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8168 for_each_possible_cpu(i
) {
8171 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8172 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8176 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8177 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8181 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8182 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8191 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8193 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8194 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8197 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8199 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8202 static inline void free_rt_sched_group(struct task_group
*tg
)
8207 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8212 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8216 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8221 #ifdef CONFIG_GROUP_SCHED
8222 static void free_sched_group(struct task_group
*tg
)
8224 free_fair_sched_group(tg
);
8225 free_rt_sched_group(tg
);
8229 /* allocate runqueue etc for a new task group */
8230 struct task_group
*sched_create_group(struct task_group
*parent
)
8232 struct task_group
*tg
;
8233 unsigned long flags
;
8236 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8238 return ERR_PTR(-ENOMEM
);
8240 if (!alloc_fair_sched_group(tg
, parent
))
8243 if (!alloc_rt_sched_group(tg
, parent
))
8246 spin_lock_irqsave(&task_group_lock
, flags
);
8247 for_each_possible_cpu(i
) {
8248 register_fair_sched_group(tg
, i
);
8249 register_rt_sched_group(tg
, i
);
8251 list_add_rcu(&tg
->list
, &task_groups
);
8253 WARN_ON(!parent
); /* root should already exist */
8255 tg
->parent
= parent
;
8256 list_add_rcu(&tg
->siblings
, &parent
->children
);
8257 INIT_LIST_HEAD(&tg
->children
);
8258 spin_unlock_irqrestore(&task_group_lock
, flags
);
8263 free_sched_group(tg
);
8264 return ERR_PTR(-ENOMEM
);
8267 /* rcu callback to free various structures associated with a task group */
8268 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8270 /* now it should be safe to free those cfs_rqs */
8271 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8274 /* Destroy runqueue etc associated with a task group */
8275 void sched_destroy_group(struct task_group
*tg
)
8277 unsigned long flags
;
8280 spin_lock_irqsave(&task_group_lock
, flags
);
8281 for_each_possible_cpu(i
) {
8282 unregister_fair_sched_group(tg
, i
);
8283 unregister_rt_sched_group(tg
, i
);
8285 list_del_rcu(&tg
->list
);
8286 list_del_rcu(&tg
->siblings
);
8287 spin_unlock_irqrestore(&task_group_lock
, flags
);
8289 /* wait for possible concurrent references to cfs_rqs complete */
8290 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8293 /* change task's runqueue when it moves between groups.
8294 * The caller of this function should have put the task in its new group
8295 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8296 * reflect its new group.
8298 void sched_move_task(struct task_struct
*tsk
)
8301 unsigned long flags
;
8304 rq
= task_rq_lock(tsk
, &flags
);
8306 update_rq_clock(rq
);
8308 running
= task_current(rq
, tsk
);
8309 on_rq
= tsk
->se
.on_rq
;
8312 dequeue_task(rq
, tsk
, 0);
8313 if (unlikely(running
))
8314 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8316 set_task_rq(tsk
, task_cpu(tsk
));
8318 #ifdef CONFIG_FAIR_GROUP_SCHED
8319 if (tsk
->sched_class
->moved_group
)
8320 tsk
->sched_class
->moved_group(tsk
);
8323 if (unlikely(running
))
8324 tsk
->sched_class
->set_curr_task(rq
);
8326 enqueue_task(rq
, tsk
, 0);
8328 task_rq_unlock(rq
, &flags
);
8332 #ifdef CONFIG_FAIR_GROUP_SCHED
8333 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8335 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8336 struct rq
*rq
= cfs_rq
->rq
;
8339 spin_lock_irq(&rq
->lock
);
8343 dequeue_entity(cfs_rq
, se
, 0);
8345 se
->load
.weight
= shares
;
8346 se
->load
.inv_weight
= 0;
8349 enqueue_entity(cfs_rq
, se
, 0);
8351 spin_unlock_irq(&rq
->lock
);
8354 static DEFINE_MUTEX(shares_mutex
);
8356 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8359 unsigned long flags
;
8362 * We can't change the weight of the root cgroup.
8367 if (shares
< MIN_SHARES
)
8368 shares
= MIN_SHARES
;
8369 else if (shares
> MAX_SHARES
)
8370 shares
= MAX_SHARES
;
8372 mutex_lock(&shares_mutex
);
8373 if (tg
->shares
== shares
)
8376 spin_lock_irqsave(&task_group_lock
, flags
);
8377 for_each_possible_cpu(i
)
8378 unregister_fair_sched_group(tg
, i
);
8379 list_del_rcu(&tg
->siblings
);
8380 spin_unlock_irqrestore(&task_group_lock
, flags
);
8382 /* wait for any ongoing reference to this group to finish */
8383 synchronize_sched();
8386 * Now we are free to modify the group's share on each cpu
8387 * w/o tripping rebalance_share or load_balance_fair.
8389 tg
->shares
= shares
;
8390 for_each_possible_cpu(i
)
8391 set_se_shares(tg
->se
[i
], shares
);
8394 * Enable load balance activity on this group, by inserting it back on
8395 * each cpu's rq->leaf_cfs_rq_list.
8397 spin_lock_irqsave(&task_group_lock
, flags
);
8398 for_each_possible_cpu(i
)
8399 register_fair_sched_group(tg
, i
);
8400 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8401 spin_unlock_irqrestore(&task_group_lock
, flags
);
8403 mutex_unlock(&shares_mutex
);
8407 unsigned long sched_group_shares(struct task_group
*tg
)
8413 #ifdef CONFIG_RT_GROUP_SCHED
8415 * Ensure that the real time constraints are schedulable.
8417 static DEFINE_MUTEX(rt_constraints_mutex
);
8419 static unsigned long to_ratio(u64 period
, u64 runtime
)
8421 if (runtime
== RUNTIME_INF
)
8424 return div64_u64(runtime
<< 16, period
);
8427 #ifdef CONFIG_CGROUP_SCHED
8428 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8430 struct task_group
*tgi
, *parent
= tg
? tg
->parent
: NULL
;
8431 unsigned long total
= 0;
8434 if (global_rt_period() < period
)
8437 return to_ratio(period
, runtime
) <
8438 to_ratio(global_rt_period(), global_rt_runtime());
8441 if (ktime_to_ns(parent
->rt_bandwidth
.rt_period
) < period
)
8445 list_for_each_entry_rcu(tgi
, &parent
->children
, siblings
) {
8449 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8450 tgi
->rt_bandwidth
.rt_runtime
);
8454 return total
+ to_ratio(period
, runtime
) <
8455 to_ratio(ktime_to_ns(parent
->rt_bandwidth
.rt_period
),
8456 parent
->rt_bandwidth
.rt_runtime
);
8458 #elif defined CONFIG_USER_SCHED
8459 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8461 struct task_group
*tgi
;
8462 unsigned long total
= 0;
8463 unsigned long global_ratio
=
8464 to_ratio(global_rt_period(), global_rt_runtime());
8467 list_for_each_entry_rcu(tgi
, &task_groups
, list
) {
8471 total
+= to_ratio(ktime_to_ns(tgi
->rt_bandwidth
.rt_period
),
8472 tgi
->rt_bandwidth
.rt_runtime
);
8476 return total
+ to_ratio(period
, runtime
) < global_ratio
;
8480 /* Must be called with tasklist_lock held */
8481 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8483 struct task_struct
*g
, *p
;
8484 do_each_thread(g
, p
) {
8485 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8487 } while_each_thread(g
, p
);
8491 static int tg_set_bandwidth(struct task_group
*tg
,
8492 u64 rt_period
, u64 rt_runtime
)
8496 mutex_lock(&rt_constraints_mutex
);
8497 read_lock(&tasklist_lock
);
8498 if (rt_runtime
== 0 && tg_has_rt_tasks(tg
)) {
8502 if (!__rt_schedulable(tg
, rt_period
, rt_runtime
)) {
8507 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8508 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8509 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8511 for_each_possible_cpu(i
) {
8512 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8514 spin_lock(&rt_rq
->rt_runtime_lock
);
8515 rt_rq
->rt_runtime
= rt_runtime
;
8516 spin_unlock(&rt_rq
->rt_runtime_lock
);
8518 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8520 read_unlock(&tasklist_lock
);
8521 mutex_unlock(&rt_constraints_mutex
);
8526 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
8528 u64 rt_runtime
, rt_period
;
8530 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8531 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
8532 if (rt_runtime_us
< 0)
8533 rt_runtime
= RUNTIME_INF
;
8535 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8538 long sched_group_rt_runtime(struct task_group
*tg
)
8542 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
8545 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
8546 do_div(rt_runtime_us
, NSEC_PER_USEC
);
8547 return rt_runtime_us
;
8550 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
8552 u64 rt_runtime
, rt_period
;
8554 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
8555 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8557 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
8560 long sched_group_rt_period(struct task_group
*tg
)
8564 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8565 do_div(rt_period_us
, NSEC_PER_USEC
);
8566 return rt_period_us
;
8569 static int sched_rt_global_constraints(void)
8573 mutex_lock(&rt_constraints_mutex
);
8574 if (!__rt_schedulable(NULL
, 1, 0))
8576 mutex_unlock(&rt_constraints_mutex
);
8581 static int sched_rt_global_constraints(void)
8583 unsigned long flags
;
8586 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8587 for_each_possible_cpu(i
) {
8588 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
8590 spin_lock(&rt_rq
->rt_runtime_lock
);
8591 rt_rq
->rt_runtime
= global_rt_runtime();
8592 spin_unlock(&rt_rq
->rt_runtime_lock
);
8594 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
8600 int sched_rt_handler(struct ctl_table
*table
, int write
,
8601 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
8605 int old_period
, old_runtime
;
8606 static DEFINE_MUTEX(mutex
);
8609 old_period
= sysctl_sched_rt_period
;
8610 old_runtime
= sysctl_sched_rt_runtime
;
8612 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
8614 if (!ret
&& write
) {
8615 ret
= sched_rt_global_constraints();
8617 sysctl_sched_rt_period
= old_period
;
8618 sysctl_sched_rt_runtime
= old_runtime
;
8620 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
8621 def_rt_bandwidth
.rt_period
=
8622 ns_to_ktime(global_rt_period());
8625 mutex_unlock(&mutex
);
8630 #ifdef CONFIG_CGROUP_SCHED
8632 /* return corresponding task_group object of a cgroup */
8633 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
8635 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
8636 struct task_group
, css
);
8639 static struct cgroup_subsys_state
*
8640 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8642 struct task_group
*tg
, *parent
;
8644 if (!cgrp
->parent
) {
8645 /* This is early initialization for the top cgroup */
8646 init_task_group
.css
.cgroup
= cgrp
;
8647 return &init_task_group
.css
;
8650 parent
= cgroup_tg(cgrp
->parent
);
8651 tg
= sched_create_group(parent
);
8653 return ERR_PTR(-ENOMEM
);
8655 /* Bind the cgroup to task_group object we just created */
8656 tg
->css
.cgroup
= cgrp
;
8662 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8664 struct task_group
*tg
= cgroup_tg(cgrp
);
8666 sched_destroy_group(tg
);
8670 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8671 struct task_struct
*tsk
)
8673 #ifdef CONFIG_RT_GROUP_SCHED
8674 /* Don't accept realtime tasks when there is no way for them to run */
8675 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
8678 /* We don't support RT-tasks being in separate groups */
8679 if (tsk
->sched_class
!= &fair_sched_class
)
8687 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
8688 struct cgroup
*old_cont
, struct task_struct
*tsk
)
8690 sched_move_task(tsk
);
8693 #ifdef CONFIG_FAIR_GROUP_SCHED
8694 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
8697 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
8700 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
8702 struct task_group
*tg
= cgroup_tg(cgrp
);
8704 return (u64
) tg
->shares
;
8708 #ifdef CONFIG_RT_GROUP_SCHED
8709 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
8712 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
8715 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8717 return sched_group_rt_runtime(cgroup_tg(cgrp
));
8720 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
8723 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
8726 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
8728 return sched_group_rt_period(cgroup_tg(cgrp
));
8732 static struct cftype cpu_files
[] = {
8733 #ifdef CONFIG_FAIR_GROUP_SCHED
8736 .read_u64
= cpu_shares_read_u64
,
8737 .write_u64
= cpu_shares_write_u64
,
8740 #ifdef CONFIG_RT_GROUP_SCHED
8742 .name
= "rt_runtime_us",
8743 .read_s64
= cpu_rt_runtime_read
,
8744 .write_s64
= cpu_rt_runtime_write
,
8747 .name
= "rt_period_us",
8748 .read_u64
= cpu_rt_period_read_uint
,
8749 .write_u64
= cpu_rt_period_write_uint
,
8754 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
8756 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
8759 struct cgroup_subsys cpu_cgroup_subsys
= {
8761 .create
= cpu_cgroup_create
,
8762 .destroy
= cpu_cgroup_destroy
,
8763 .can_attach
= cpu_cgroup_can_attach
,
8764 .attach
= cpu_cgroup_attach
,
8765 .populate
= cpu_cgroup_populate
,
8766 .subsys_id
= cpu_cgroup_subsys_id
,
8770 #endif /* CONFIG_CGROUP_SCHED */
8772 #ifdef CONFIG_CGROUP_CPUACCT
8775 * CPU accounting code for task groups.
8777 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
8778 * (balbir@in.ibm.com).
8781 /* track cpu usage of a group of tasks */
8783 struct cgroup_subsys_state css
;
8784 /* cpuusage holds pointer to a u64-type object on every cpu */
8788 struct cgroup_subsys cpuacct_subsys
;
8790 /* return cpu accounting group corresponding to this container */
8791 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
8793 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
8794 struct cpuacct
, css
);
8797 /* return cpu accounting group to which this task belongs */
8798 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
8800 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
8801 struct cpuacct
, css
);
8804 /* create a new cpu accounting group */
8805 static struct cgroup_subsys_state
*cpuacct_create(
8806 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8808 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
8811 return ERR_PTR(-ENOMEM
);
8813 ca
->cpuusage
= alloc_percpu(u64
);
8814 if (!ca
->cpuusage
) {
8816 return ERR_PTR(-ENOMEM
);
8822 /* destroy an existing cpu accounting group */
8824 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8826 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8828 free_percpu(ca
->cpuusage
);
8832 /* return total cpu usage (in nanoseconds) of a group */
8833 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
8835 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8836 u64 totalcpuusage
= 0;
8839 for_each_possible_cpu(i
) {
8840 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8843 * Take rq->lock to make 64-bit addition safe on 32-bit
8846 spin_lock_irq(&cpu_rq(i
)->lock
);
8847 totalcpuusage
+= *cpuusage
;
8848 spin_unlock_irq(&cpu_rq(i
)->lock
);
8851 return totalcpuusage
;
8854 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
8857 struct cpuacct
*ca
= cgroup_ca(cgrp
);
8866 for_each_possible_cpu(i
) {
8867 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
8869 spin_lock_irq(&cpu_rq(i
)->lock
);
8871 spin_unlock_irq(&cpu_rq(i
)->lock
);
8877 static struct cftype files
[] = {
8880 .read_u64
= cpuusage_read
,
8881 .write_u64
= cpuusage_write
,
8885 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
8887 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
8891 * charge this task's execution time to its accounting group.
8893 * called with rq->lock held.
8895 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
8899 if (!cpuacct_subsys
.active
)
8904 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
8906 *cpuusage
+= cputime
;
8910 struct cgroup_subsys cpuacct_subsys
= {
8912 .create
= cpuacct_create
,
8913 .destroy
= cpuacct_destroy
,
8914 .populate
= cpuacct_populate
,
8915 .subsys_id
= cpuacct_subsys_id
,
8917 #endif /* CONFIG_CGROUP_CPUACCT */