4 * Kernel scheduler and related syscalls
6 * Copyright (C) 1991-2002 Linus Torvalds
8 * 1996-12-23 Modified by Dave Grothe to fix bugs in semaphores and
9 * make semaphores SMP safe
10 * 1998-11-19 Implemented schedule_timeout() and related stuff
12 * 2002-01-04 New ultra-scalable O(1) scheduler by Ingo Molnar:
13 * hybrid priority-list and round-robin design with
14 * an array-switch method of distributing timeslices
15 * and per-CPU runqueues. Cleanups and useful suggestions
16 * by Davide Libenzi, preemptible kernel bits by Robert Love.
17 * 2003-09-03 Interactivity tuning by Con Kolivas.
18 * 2004-04-02 Scheduler domains code by Nick Piggin
19 * 2007-04-15 Work begun on replacing all interactivity tuning with a
20 * fair scheduling design by Con Kolivas.
21 * 2007-05-05 Load balancing (smp-nice) and other improvements
23 * 2007-05-06 Interactivity improvements to CFS by Mike Galbraith
24 * 2007-07-01 Group scheduling enhancements by Srivatsa Vaddagiri
25 * 2007-11-29 RT balancing improvements by Steven Rostedt, Gregory Haskins,
26 * Thomas Gleixner, Mike Kravetz
30 #include <linux/module.h>
31 #include <linux/nmi.h>
32 #include <linux/init.h>
33 #include <linux/uaccess.h>
34 #include <linux/highmem.h>
35 #include <linux/smp_lock.h>
36 #include <asm/mmu_context.h>
37 #include <linux/interrupt.h>
38 #include <linux/capability.h>
39 #include <linux/completion.h>
40 #include <linux/kernel_stat.h>
41 #include <linux/debug_locks.h>
42 #include <linux/security.h>
43 #include <linux/notifier.h>
44 #include <linux/profile.h>
45 #include <linux/freezer.h>
46 #include <linux/vmalloc.h>
47 #include <linux/blkdev.h>
48 #include <linux/delay.h>
49 #include <linux/pid_namespace.h>
50 #include <linux/smp.h>
51 #include <linux/threads.h>
52 #include <linux/timer.h>
53 #include <linux/rcupdate.h>
54 #include <linux/cpu.h>
55 #include <linux/cpuset.h>
56 #include <linux/percpu.h>
57 #include <linux/kthread.h>
58 #include <linux/proc_fs.h>
59 #include <linux/seq_file.h>
60 #include <linux/sysctl.h>
61 #include <linux/syscalls.h>
62 #include <linux/times.h>
63 #include <linux/tsacct_kern.h>
64 #include <linux/kprobes.h>
65 #include <linux/delayacct.h>
66 #include <linux/reciprocal_div.h>
67 #include <linux/unistd.h>
68 #include <linux/pagemap.h>
69 #include <linux/hrtimer.h>
70 #include <linux/tick.h>
71 #include <linux/bootmem.h>
72 #include <linux/debugfs.h>
73 #include <linux/ctype.h>
74 #include <linux/ftrace.h>
75 #include <trace/sched.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
83 * Convert user-nice values [ -20 ... 0 ... 19 ]
84 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
87 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
88 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
89 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
92 * 'User priority' is the nice value converted to something we
93 * can work with better when scaling various scheduler parameters,
94 * it's a [ 0 ... 39 ] range.
96 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
97 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
98 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
101 * Helpers for converting nanosecond timing to jiffy resolution
103 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
105 #define NICE_0_LOAD SCHED_LOAD_SCALE
106 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
109 * These are the 'tuning knobs' of the scheduler:
111 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
112 * Timeslices get refilled after they expire.
114 #define DEF_TIMESLICE (100 * HZ / 1000)
117 * single value that denotes runtime == period, ie unlimited time.
119 #define RUNTIME_INF ((u64)~0ULL)
123 * Divide a load by a sched group cpu_power : (load / sg->__cpu_power)
124 * Since cpu_power is a 'constant', we can use a reciprocal divide.
126 static inline u32
sg_div_cpu_power(const struct sched_group
*sg
, u32 load
)
128 return reciprocal_divide(load
, sg
->reciprocal_cpu_power
);
132 * Each time a sched group cpu_power is changed,
133 * we must compute its reciprocal value
135 static inline void sg_inc_cpu_power(struct sched_group
*sg
, u32 val
)
137 sg
->__cpu_power
+= val
;
138 sg
->reciprocal_cpu_power
= reciprocal_value(sg
->__cpu_power
);
142 static inline int rt_policy(int policy
)
144 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
149 static inline int task_has_rt_policy(struct task_struct
*p
)
151 return rt_policy(p
->policy
);
155 * This is the priority-queue data structure of the RT scheduling class:
157 struct rt_prio_array
{
158 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
159 struct list_head queue
[MAX_RT_PRIO
];
162 struct rt_bandwidth
{
163 /* nests inside the rq lock: */
164 spinlock_t rt_runtime_lock
;
167 struct hrtimer rt_period_timer
;
170 static struct rt_bandwidth def_rt_bandwidth
;
172 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
174 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
176 struct rt_bandwidth
*rt_b
=
177 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
183 now
= hrtimer_cb_get_time(timer
);
184 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
189 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
192 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
196 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
198 rt_b
->rt_period
= ns_to_ktime(period
);
199 rt_b
->rt_runtime
= runtime
;
201 spin_lock_init(&rt_b
->rt_runtime_lock
);
203 hrtimer_init(&rt_b
->rt_period_timer
,
204 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
205 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
206 rt_b
->rt_period_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_UNLOCKED
;
209 static inline int rt_bandwidth_enabled(void)
211 return sysctl_sched_rt_runtime
>= 0;
214 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
218 if (rt_bandwidth_enabled() && rt_b
->rt_runtime
== RUNTIME_INF
)
221 if (hrtimer_active(&rt_b
->rt_period_timer
))
224 spin_lock(&rt_b
->rt_runtime_lock
);
226 if (hrtimer_active(&rt_b
->rt_period_timer
))
229 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
230 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
231 hrtimer_start_expires(&rt_b
->rt_period_timer
,
234 spin_unlock(&rt_b
->rt_runtime_lock
);
237 #ifdef CONFIG_RT_GROUP_SCHED
238 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
240 hrtimer_cancel(&rt_b
->rt_period_timer
);
245 * sched_domains_mutex serializes calls to arch_init_sched_domains,
246 * detach_destroy_domains and partition_sched_domains.
248 static DEFINE_MUTEX(sched_domains_mutex
);
250 #ifdef CONFIG_GROUP_SCHED
252 #include <linux/cgroup.h>
256 static LIST_HEAD(task_groups
);
258 /* task group related information */
260 #ifdef CONFIG_CGROUP_SCHED
261 struct cgroup_subsys_state css
;
264 #ifdef CONFIG_FAIR_GROUP_SCHED
265 /* schedulable entities of this group on each cpu */
266 struct sched_entity
**se
;
267 /* runqueue "owned" by this group on each cpu */
268 struct cfs_rq
**cfs_rq
;
269 unsigned long shares
;
272 #ifdef CONFIG_RT_GROUP_SCHED
273 struct sched_rt_entity
**rt_se
;
274 struct rt_rq
**rt_rq
;
276 struct rt_bandwidth rt_bandwidth
;
280 struct list_head list
;
282 struct task_group
*parent
;
283 struct list_head siblings
;
284 struct list_head children
;
287 #ifdef CONFIG_USER_SCHED
291 * Every UID task group (including init_task_group aka UID-0) will
292 * be a child to this group.
294 struct task_group root_task_group
;
296 #ifdef CONFIG_FAIR_GROUP_SCHED
297 /* Default task group's sched entity on each cpu */
298 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
299 /* Default task group's cfs_rq on each cpu */
300 static DEFINE_PER_CPU(struct cfs_rq
, init_cfs_rq
) ____cacheline_aligned_in_smp
;
301 #endif /* CONFIG_FAIR_GROUP_SCHED */
303 #ifdef CONFIG_RT_GROUP_SCHED
304 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
305 static DEFINE_PER_CPU(struct rt_rq
, init_rt_rq
) ____cacheline_aligned_in_smp
;
306 #endif /* CONFIG_RT_GROUP_SCHED */
307 #else /* !CONFIG_USER_SCHED */
308 #define root_task_group init_task_group
309 #endif /* CONFIG_USER_SCHED */
311 /* task_group_lock serializes add/remove of task groups and also changes to
312 * a task group's cpu shares.
314 static DEFINE_SPINLOCK(task_group_lock
);
316 #ifdef CONFIG_FAIR_GROUP_SCHED
317 #ifdef CONFIG_USER_SCHED
318 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
319 #else /* !CONFIG_USER_SCHED */
320 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
321 #endif /* CONFIG_USER_SCHED */
324 * A weight of 0 or 1 can cause arithmetics problems.
325 * A weight of a cfs_rq is the sum of weights of which entities
326 * are queued on this cfs_rq, so a weight of a entity should not be
327 * too large, so as the shares value of a task group.
328 * (The default weight is 1024 - so there's no practical
329 * limitation from this.)
332 #define MAX_SHARES (1UL << 18)
334 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
337 /* Default task group.
338 * Every task in system belong to this group at bootup.
340 struct task_group init_task_group
;
342 /* return group to which a task belongs */
343 static inline struct task_group
*task_group(struct task_struct
*p
)
345 struct task_group
*tg
;
347 #ifdef CONFIG_USER_SCHED
349 tg
= __task_cred(p
)->user
->tg
;
351 #elif defined(CONFIG_CGROUP_SCHED)
352 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
353 struct task_group
, css
);
355 tg
= &init_task_group
;
360 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
361 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
363 #ifdef CONFIG_FAIR_GROUP_SCHED
364 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
365 p
->se
.parent
= task_group(p
)->se
[cpu
];
368 #ifdef CONFIG_RT_GROUP_SCHED
369 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
370 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
376 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
377 static inline struct task_group
*task_group(struct task_struct
*p
)
382 #endif /* CONFIG_GROUP_SCHED */
384 /* CFS-related fields in a runqueue */
386 struct load_weight load
;
387 unsigned long nr_running
;
392 struct rb_root tasks_timeline
;
393 struct rb_node
*rb_leftmost
;
395 struct list_head tasks
;
396 struct list_head
*balance_iterator
;
399 * 'curr' points to currently running entity on this cfs_rq.
400 * It is set to NULL otherwise (i.e when none are currently running).
402 struct sched_entity
*curr
, *next
, *last
;
404 unsigned int nr_spread_over
;
406 #ifdef CONFIG_FAIR_GROUP_SCHED
407 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
410 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
411 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
412 * (like users, containers etc.)
414 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
415 * list is used during load balance.
417 struct list_head leaf_cfs_rq_list
;
418 struct task_group
*tg
; /* group that "owns" this runqueue */
422 * the part of load.weight contributed by tasks
424 unsigned long task_weight
;
427 * h_load = weight * f(tg)
429 * Where f(tg) is the recursive weight fraction assigned to
432 unsigned long h_load
;
435 * this cpu's part of tg->shares
437 unsigned long shares
;
440 * load.weight at the time we set shares
442 unsigned long rq_weight
;
447 /* Real-Time classes' related field in a runqueue: */
449 struct rt_prio_array active
;
450 unsigned long rt_nr_running
;
451 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
452 int highest_prio
; /* highest queued rt task prio */
455 unsigned long rt_nr_migratory
;
461 /* Nests inside the rq lock: */
462 spinlock_t rt_runtime_lock
;
464 #ifdef CONFIG_RT_GROUP_SCHED
465 unsigned long rt_nr_boosted
;
468 struct list_head leaf_rt_rq_list
;
469 struct task_group
*tg
;
470 struct sched_rt_entity
*rt_se
;
477 * We add the notion of a root-domain which will be used to define per-domain
478 * variables. Each exclusive cpuset essentially defines an island domain by
479 * fully partitioning the member cpus from any other cpuset. Whenever a new
480 * exclusive cpuset is created, we also create and attach a new root-domain
490 * The "RT overload" flag: it gets set if a CPU has more than
491 * one runnable RT task.
496 struct cpupri cpupri
;
501 * By default the system creates a single root-domain with all cpus as
502 * members (mimicking the global state we have today).
504 static struct root_domain def_root_domain
;
509 * This is the main, per-CPU runqueue data structure.
511 * Locking rule: those places that want to lock multiple runqueues
512 * (such as the load balancing or the thread migration code), lock
513 * acquire operations must be ordered by ascending &runqueue.
520 * nr_running and cpu_load should be in the same cacheline because
521 * remote CPUs use both these fields when doing load calculation.
523 unsigned long nr_running
;
524 #define CPU_LOAD_IDX_MAX 5
525 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
526 unsigned char idle_at_tick
;
528 unsigned long last_tick_seen
;
529 unsigned char in_nohz_recently
;
531 /* capture load from *all* tasks on this cpu: */
532 struct load_weight load
;
533 unsigned long nr_load_updates
;
539 #ifdef CONFIG_FAIR_GROUP_SCHED
540 /* list of leaf cfs_rq on this cpu: */
541 struct list_head leaf_cfs_rq_list
;
543 #ifdef CONFIG_RT_GROUP_SCHED
544 struct list_head leaf_rt_rq_list
;
548 * This is part of a global counter where only the total sum
549 * over all CPUs matters. A task can increase this counter on
550 * one CPU and if it got migrated afterwards it may decrease
551 * it on another CPU. Always updated under the runqueue lock:
553 unsigned long nr_uninterruptible
;
555 struct task_struct
*curr
, *idle
;
556 unsigned long next_balance
;
557 struct mm_struct
*prev_mm
;
564 struct root_domain
*rd
;
565 struct sched_domain
*sd
;
567 /* For active balancing */
570 /* cpu of this runqueue: */
574 unsigned long avg_load_per_task
;
576 struct task_struct
*migration_thread
;
577 struct list_head migration_queue
;
580 #ifdef CONFIG_SCHED_HRTICK
582 int hrtick_csd_pending
;
583 struct call_single_data hrtick_csd
;
585 struct hrtimer hrtick_timer
;
588 #ifdef CONFIG_SCHEDSTATS
590 struct sched_info rq_sched_info
;
592 /* sys_sched_yield() stats */
593 unsigned int yld_exp_empty
;
594 unsigned int yld_act_empty
;
595 unsigned int yld_both_empty
;
596 unsigned int yld_count
;
598 /* schedule() stats */
599 unsigned int sched_switch
;
600 unsigned int sched_count
;
601 unsigned int sched_goidle
;
603 /* try_to_wake_up() stats */
604 unsigned int ttwu_count
;
605 unsigned int ttwu_local
;
608 unsigned int bkl_count
;
612 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
614 static inline void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int sync
)
616 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, sync
);
619 static inline int cpu_of(struct rq
*rq
)
629 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
630 * See detach_destroy_domains: synchronize_sched for details.
632 * The domain tree of any CPU may only be accessed from within
633 * preempt-disabled sections.
635 #define for_each_domain(cpu, __sd) \
636 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
638 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
639 #define this_rq() (&__get_cpu_var(runqueues))
640 #define task_rq(p) cpu_rq(task_cpu(p))
641 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
643 static inline void update_rq_clock(struct rq
*rq
)
645 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
649 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
651 #ifdef CONFIG_SCHED_DEBUG
652 # define const_debug __read_mostly
654 # define const_debug static const
660 * Returns true if the current cpu runqueue is locked.
661 * This interface allows printk to be called with the runqueue lock
662 * held and know whether or not it is OK to wake up the klogd.
664 int runqueue_is_locked(void)
667 struct rq
*rq
= cpu_rq(cpu
);
670 ret
= spin_is_locked(&rq
->lock
);
676 * Debugging: various feature bits
679 #define SCHED_FEAT(name, enabled) \
680 __SCHED_FEAT_##name ,
683 #include "sched_features.h"
688 #define SCHED_FEAT(name, enabled) \
689 (1UL << __SCHED_FEAT_##name) * enabled |
691 const_debug
unsigned int sysctl_sched_features
=
692 #include "sched_features.h"
697 #ifdef CONFIG_SCHED_DEBUG
698 #define SCHED_FEAT(name, enabled) \
701 static __read_mostly
char *sched_feat_names
[] = {
702 #include "sched_features.h"
708 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
710 filp
->private_data
= inode
->i_private
;
715 sched_feat_read(struct file
*filp
, char __user
*ubuf
,
716 size_t cnt
, loff_t
*ppos
)
723 for (i
= 0; sched_feat_names
[i
]; i
++) {
724 len
+= strlen(sched_feat_names
[i
]);
728 buf
= kmalloc(len
+ 2, GFP_KERNEL
);
732 for (i
= 0; sched_feat_names
[i
]; i
++) {
733 if (sysctl_sched_features
& (1UL << i
))
734 r
+= sprintf(buf
+ r
, "%s ", sched_feat_names
[i
]);
736 r
+= sprintf(buf
+ r
, "NO_%s ", sched_feat_names
[i
]);
739 r
+= sprintf(buf
+ r
, "\n");
740 WARN_ON(r
>= len
+ 2);
742 r
= simple_read_from_buffer(ubuf
, cnt
, ppos
, buf
, r
);
750 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
751 size_t cnt
, loff_t
*ppos
)
761 if (copy_from_user(&buf
, ubuf
, cnt
))
766 if (strncmp(buf
, "NO_", 3) == 0) {
771 for (i
= 0; sched_feat_names
[i
]; i
++) {
772 int len
= strlen(sched_feat_names
[i
]);
774 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
776 sysctl_sched_features
&= ~(1UL << i
);
778 sysctl_sched_features
|= (1UL << i
);
783 if (!sched_feat_names
[i
])
791 static struct file_operations sched_feat_fops
= {
792 .open
= sched_feat_open
,
793 .read
= sched_feat_read
,
794 .write
= sched_feat_write
,
797 static __init
int sched_init_debug(void)
799 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
804 late_initcall(sched_init_debug
);
808 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
811 * Number of tasks to iterate in a single balance run.
812 * Limited because this is done with IRQs disabled.
814 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
817 * ratelimit for updating the group shares.
820 unsigned int sysctl_sched_shares_ratelimit
= 250000;
823 * Inject some fuzzyness into changing the per-cpu group shares
824 * this avoids remote rq-locks at the expense of fairness.
827 unsigned int sysctl_sched_shares_thresh
= 4;
830 * period over which we measure -rt task cpu usage in us.
833 unsigned int sysctl_sched_rt_period
= 1000000;
835 static __read_mostly
int scheduler_running
;
838 * part of the period that we allow rt tasks to run in us.
841 int sysctl_sched_rt_runtime
= 950000;
843 static inline u64
global_rt_period(void)
845 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
848 static inline u64
global_rt_runtime(void)
850 if (sysctl_sched_rt_runtime
< 0)
853 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
856 #ifndef prepare_arch_switch
857 # define prepare_arch_switch(next) do { } while (0)
859 #ifndef finish_arch_switch
860 # define finish_arch_switch(prev) do { } while (0)
863 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
865 return rq
->curr
== p
;
868 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
869 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
871 return task_current(rq
, p
);
874 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
878 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
880 #ifdef CONFIG_DEBUG_SPINLOCK
881 /* this is a valid case when another task releases the spinlock */
882 rq
->lock
.owner
= current
;
885 * If we are tracking spinlock dependencies then we have to
886 * fix up the runqueue lock - which gets 'carried over' from
889 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
891 spin_unlock_irq(&rq
->lock
);
894 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
895 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
900 return task_current(rq
, p
);
904 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
908 * We can optimise this out completely for !SMP, because the
909 * SMP rebalancing from interrupt is the only thing that cares
914 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
915 spin_unlock_irq(&rq
->lock
);
917 spin_unlock(&rq
->lock
);
921 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
925 * After ->oncpu is cleared, the task can be moved to a different CPU.
926 * We must ensure this doesn't happen until the switch is completely
932 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
936 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
939 * __task_rq_lock - lock the runqueue a given task resides on.
940 * Must be called interrupts disabled.
942 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
946 struct rq
*rq
= task_rq(p
);
947 spin_lock(&rq
->lock
);
948 if (likely(rq
== task_rq(p
)))
950 spin_unlock(&rq
->lock
);
955 * task_rq_lock - lock the runqueue a given task resides on and disable
956 * interrupts. Note the ordering: we can safely lookup the task_rq without
957 * explicitly disabling preemption.
959 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
965 local_irq_save(*flags
);
967 spin_lock(&rq
->lock
);
968 if (likely(rq
== task_rq(p
)))
970 spin_unlock_irqrestore(&rq
->lock
, *flags
);
974 void task_rq_unlock_wait(struct task_struct
*p
)
976 struct rq
*rq
= task_rq(p
);
978 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
979 spin_unlock_wait(&rq
->lock
);
982 static void __task_rq_unlock(struct rq
*rq
)
985 spin_unlock(&rq
->lock
);
988 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
991 spin_unlock_irqrestore(&rq
->lock
, *flags
);
995 * this_rq_lock - lock this runqueue and disable interrupts.
997 static struct rq
*this_rq_lock(void)
1002 local_irq_disable();
1004 spin_lock(&rq
->lock
);
1009 #ifdef CONFIG_SCHED_HRTICK
1011 * Use HR-timers to deliver accurate preemption points.
1013 * Its all a bit involved since we cannot program an hrt while holding the
1014 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1017 * When we get rescheduled we reprogram the hrtick_timer outside of the
1023 * - enabled by features
1024 * - hrtimer is actually high res
1026 static inline int hrtick_enabled(struct rq
*rq
)
1028 if (!sched_feat(HRTICK
))
1030 if (!cpu_active(cpu_of(rq
)))
1032 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1035 static void hrtick_clear(struct rq
*rq
)
1037 if (hrtimer_active(&rq
->hrtick_timer
))
1038 hrtimer_cancel(&rq
->hrtick_timer
);
1042 * High-resolution timer tick.
1043 * Runs from hardirq context with interrupts disabled.
1045 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1047 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1049 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1051 spin_lock(&rq
->lock
);
1052 update_rq_clock(rq
);
1053 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1054 spin_unlock(&rq
->lock
);
1056 return HRTIMER_NORESTART
;
1061 * called from hardirq (IPI) context
1063 static void __hrtick_start(void *arg
)
1065 struct rq
*rq
= arg
;
1067 spin_lock(&rq
->lock
);
1068 hrtimer_restart(&rq
->hrtick_timer
);
1069 rq
->hrtick_csd_pending
= 0;
1070 spin_unlock(&rq
->lock
);
1074 * Called to set the hrtick timer state.
1076 * called with rq->lock held and irqs disabled
1078 static void hrtick_start(struct rq
*rq
, u64 delay
)
1080 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1081 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1083 hrtimer_set_expires(timer
, time
);
1085 if (rq
== this_rq()) {
1086 hrtimer_restart(timer
);
1087 } else if (!rq
->hrtick_csd_pending
) {
1088 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
);
1089 rq
->hrtick_csd_pending
= 1;
1094 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1096 int cpu
= (int)(long)hcpu
;
1099 case CPU_UP_CANCELED
:
1100 case CPU_UP_CANCELED_FROZEN
:
1101 case CPU_DOWN_PREPARE
:
1102 case CPU_DOWN_PREPARE_FROZEN
:
1104 case CPU_DEAD_FROZEN
:
1105 hrtick_clear(cpu_rq(cpu
));
1112 static __init
void init_hrtick(void)
1114 hotcpu_notifier(hotplug_hrtick
, 0);
1118 * Called to set the hrtick timer state.
1120 * called with rq->lock held and irqs disabled
1122 static void hrtick_start(struct rq
*rq
, u64 delay
)
1124 hrtimer_start(&rq
->hrtick_timer
, ns_to_ktime(delay
), HRTIMER_MODE_REL
);
1127 static inline void init_hrtick(void)
1130 #endif /* CONFIG_SMP */
1132 static void init_rq_hrtick(struct rq
*rq
)
1135 rq
->hrtick_csd_pending
= 0;
1137 rq
->hrtick_csd
.flags
= 0;
1138 rq
->hrtick_csd
.func
= __hrtick_start
;
1139 rq
->hrtick_csd
.info
= rq
;
1142 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1143 rq
->hrtick_timer
.function
= hrtick
;
1144 rq
->hrtick_timer
.cb_mode
= HRTIMER_CB_IRQSAFE_PERCPU
;
1146 #else /* CONFIG_SCHED_HRTICK */
1147 static inline void hrtick_clear(struct rq
*rq
)
1151 static inline void init_rq_hrtick(struct rq
*rq
)
1155 static inline void init_hrtick(void)
1158 #endif /* CONFIG_SCHED_HRTICK */
1161 * resched_task - mark a task 'to be rescheduled now'.
1163 * On UP this means the setting of the need_resched flag, on SMP it
1164 * might also involve a cross-CPU call to trigger the scheduler on
1169 #ifndef tsk_is_polling
1170 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1173 static void resched_task(struct task_struct
*p
)
1177 assert_spin_locked(&task_rq(p
)->lock
);
1179 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1182 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1185 if (cpu
== smp_processor_id())
1188 /* NEED_RESCHED must be visible before we test polling */
1190 if (!tsk_is_polling(p
))
1191 smp_send_reschedule(cpu
);
1194 static void resched_cpu(int cpu
)
1196 struct rq
*rq
= cpu_rq(cpu
);
1197 unsigned long flags
;
1199 if (!spin_trylock_irqsave(&rq
->lock
, flags
))
1201 resched_task(cpu_curr(cpu
));
1202 spin_unlock_irqrestore(&rq
->lock
, flags
);
1207 * When add_timer_on() enqueues a timer into the timer wheel of an
1208 * idle CPU then this timer might expire before the next timer event
1209 * which is scheduled to wake up that CPU. In case of a completely
1210 * idle system the next event might even be infinite time into the
1211 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1212 * leaves the inner idle loop so the newly added timer is taken into
1213 * account when the CPU goes back to idle and evaluates the timer
1214 * wheel for the next timer event.
1216 void wake_up_idle_cpu(int cpu
)
1218 struct rq
*rq
= cpu_rq(cpu
);
1220 if (cpu
== smp_processor_id())
1224 * This is safe, as this function is called with the timer
1225 * wheel base lock of (cpu) held. When the CPU is on the way
1226 * to idle and has not yet set rq->curr to idle then it will
1227 * be serialized on the timer wheel base lock and take the new
1228 * timer into account automatically.
1230 if (rq
->curr
!= rq
->idle
)
1234 * We can set TIF_RESCHED on the idle task of the other CPU
1235 * lockless. The worst case is that the other CPU runs the
1236 * idle task through an additional NOOP schedule()
1238 set_tsk_thread_flag(rq
->idle
, TIF_NEED_RESCHED
);
1240 /* NEED_RESCHED must be visible before we test polling */
1242 if (!tsk_is_polling(rq
->idle
))
1243 smp_send_reschedule(cpu
);
1245 #endif /* CONFIG_NO_HZ */
1247 #else /* !CONFIG_SMP */
1248 static void resched_task(struct task_struct
*p
)
1250 assert_spin_locked(&task_rq(p
)->lock
);
1251 set_tsk_need_resched(p
);
1253 #endif /* CONFIG_SMP */
1255 #if BITS_PER_LONG == 32
1256 # define WMULT_CONST (~0UL)
1258 # define WMULT_CONST (1UL << 32)
1261 #define WMULT_SHIFT 32
1264 * Shift right and round:
1266 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1269 * delta *= weight / lw
1271 static unsigned long
1272 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1273 struct load_weight
*lw
)
1277 if (!lw
->inv_weight
) {
1278 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1281 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1285 tmp
= (u64
)delta_exec
* weight
;
1287 * Check whether we'd overflow the 64-bit multiplication:
1289 if (unlikely(tmp
> WMULT_CONST
))
1290 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1293 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1295 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1298 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1304 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1311 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1312 * of tasks with abnormal "nice" values across CPUs the contribution that
1313 * each task makes to its run queue's load is weighted according to its
1314 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1315 * scaled version of the new time slice allocation that they receive on time
1319 #define WEIGHT_IDLEPRIO 2
1320 #define WMULT_IDLEPRIO (1 << 31)
1323 * Nice levels are multiplicative, with a gentle 10% change for every
1324 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1325 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1326 * that remained on nice 0.
1328 * The "10% effect" is relative and cumulative: from _any_ nice level,
1329 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1330 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1331 * If a task goes up by ~10% and another task goes down by ~10% then
1332 * the relative distance between them is ~25%.)
1334 static const int prio_to_weight
[40] = {
1335 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1336 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1337 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1338 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1339 /* 0 */ 1024, 820, 655, 526, 423,
1340 /* 5 */ 335, 272, 215, 172, 137,
1341 /* 10 */ 110, 87, 70, 56, 45,
1342 /* 15 */ 36, 29, 23, 18, 15,
1346 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1348 * In cases where the weight does not change often, we can use the
1349 * precalculated inverse to speed up arithmetics by turning divisions
1350 * into multiplications:
1352 static const u32 prio_to_wmult
[40] = {
1353 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1354 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1355 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1356 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1357 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1358 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1359 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1360 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1363 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1366 * runqueue iterator, to support SMP load-balancing between different
1367 * scheduling classes, without having to expose their internal data
1368 * structures to the load-balancing proper:
1370 struct rq_iterator
{
1372 struct task_struct
*(*start
)(void *);
1373 struct task_struct
*(*next
)(void *);
1377 static unsigned long
1378 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1379 unsigned long max_load_move
, struct sched_domain
*sd
,
1380 enum cpu_idle_type idle
, int *all_pinned
,
1381 int *this_best_prio
, struct rq_iterator
*iterator
);
1384 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1385 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1386 struct rq_iterator
*iterator
);
1389 #ifdef CONFIG_CGROUP_CPUACCT
1390 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1392 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1395 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1397 update_load_add(&rq
->load
, load
);
1400 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1402 update_load_sub(&rq
->load
, load
);
1405 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1406 typedef int (*tg_visitor
)(struct task_group
*, void *);
1409 * Iterate the full tree, calling @down when first entering a node and @up when
1410 * leaving it for the final time.
1412 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1414 struct task_group
*parent
, *child
;
1418 parent
= &root_task_group
;
1420 ret
= (*down
)(parent
, data
);
1423 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1430 ret
= (*up
)(parent
, data
);
1435 parent
= parent
->parent
;
1444 static int tg_nop(struct task_group
*tg
, void *data
)
1451 static unsigned long source_load(int cpu
, int type
);
1452 static unsigned long target_load(int cpu
, int type
);
1453 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1455 static unsigned long cpu_avg_load_per_task(int cpu
)
1457 struct rq
*rq
= cpu_rq(cpu
);
1458 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1461 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1463 rq
->avg_load_per_task
= 0;
1465 return rq
->avg_load_per_task
;
1468 #ifdef CONFIG_FAIR_GROUP_SCHED
1470 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1473 * Calculate and set the cpu's group shares.
1476 update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1477 unsigned long sd_shares
, unsigned long sd_rq_weight
)
1480 unsigned long shares
;
1481 unsigned long rq_weight
;
1486 rq_weight
= tg
->cfs_rq
[cpu
]->load
.weight
;
1489 * If there are currently no tasks on the cpu pretend there is one of
1490 * average load so that when a new task gets to run here it will not
1491 * get delayed by group starvation.
1495 rq_weight
= NICE_0_LOAD
;
1498 if (unlikely(rq_weight
> sd_rq_weight
))
1499 rq_weight
= sd_rq_weight
;
1502 * \Sum shares * rq_weight
1503 * shares = -----------------------
1507 shares
= (sd_shares
* rq_weight
) / (sd_rq_weight
+ 1);
1508 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1510 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1511 sysctl_sched_shares_thresh
) {
1512 struct rq
*rq
= cpu_rq(cpu
);
1513 unsigned long flags
;
1515 spin_lock_irqsave(&rq
->lock
, flags
);
1517 * record the actual number of shares, not the boosted amount.
1519 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1520 tg
->cfs_rq
[cpu
]->rq_weight
= rq_weight
;
1522 __set_se_shares(tg
->se
[cpu
], shares
);
1523 spin_unlock_irqrestore(&rq
->lock
, flags
);
1528 * Re-compute the task group their per cpu shares over the given domain.
1529 * This needs to be done in a bottom-up fashion because the rq weight of a
1530 * parent group depends on the shares of its child groups.
1532 static int tg_shares_up(struct task_group
*tg
, void *data
)
1534 unsigned long rq_weight
= 0;
1535 unsigned long shares
= 0;
1536 struct sched_domain
*sd
= data
;
1539 for_each_cpu_mask(i
, sd
->span
) {
1540 rq_weight
+= tg
->cfs_rq
[i
]->load
.weight
;
1541 shares
+= tg
->cfs_rq
[i
]->shares
;
1544 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1545 shares
= tg
->shares
;
1547 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1548 shares
= tg
->shares
;
1551 rq_weight
= cpus_weight(sd
->span
) * NICE_0_LOAD
;
1553 for_each_cpu_mask(i
, sd
->span
)
1554 update_group_shares_cpu(tg
, i
, shares
, rq_weight
);
1560 * Compute the cpu's hierarchical load factor for each task group.
1561 * This needs to be done in a top-down fashion because the load of a child
1562 * group is a fraction of its parents load.
1564 static int tg_load_down(struct task_group
*tg
, void *data
)
1567 long cpu
= (long)data
;
1570 load
= cpu_rq(cpu
)->load
.weight
;
1572 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1573 load
*= tg
->cfs_rq
[cpu
]->shares
;
1574 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1577 tg
->cfs_rq
[cpu
]->h_load
= load
;
1582 static void update_shares(struct sched_domain
*sd
)
1584 u64 now
= cpu_clock(raw_smp_processor_id());
1585 s64 elapsed
= now
- sd
->last_update
;
1587 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1588 sd
->last_update
= now
;
1589 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1593 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1595 spin_unlock(&rq
->lock
);
1597 spin_lock(&rq
->lock
);
1600 static void update_h_load(long cpu
)
1602 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1607 static inline void update_shares(struct sched_domain
*sd
)
1611 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1619 #ifdef CONFIG_FAIR_GROUP_SCHED
1620 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1623 cfs_rq
->shares
= shares
;
1628 #include "sched_stats.h"
1629 #include "sched_idletask.c"
1630 #include "sched_fair.c"
1631 #include "sched_rt.c"
1632 #ifdef CONFIG_SCHED_DEBUG
1633 # include "sched_debug.c"
1636 #define sched_class_highest (&rt_sched_class)
1637 #define for_each_class(class) \
1638 for (class = sched_class_highest; class; class = class->next)
1640 static void inc_nr_running(struct rq
*rq
)
1645 static void dec_nr_running(struct rq
*rq
)
1650 static void set_load_weight(struct task_struct
*p
)
1652 if (task_has_rt_policy(p
)) {
1653 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1654 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1659 * SCHED_IDLE tasks get minimal weight:
1661 if (p
->policy
== SCHED_IDLE
) {
1662 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1663 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1667 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1668 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1671 static void update_avg(u64
*avg
, u64 sample
)
1673 s64 diff
= sample
- *avg
;
1677 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1679 sched_info_queued(p
);
1680 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1684 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1686 if (sleep
&& p
->se
.last_wakeup
) {
1687 update_avg(&p
->se
.avg_overlap
,
1688 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1689 p
->se
.last_wakeup
= 0;
1692 sched_info_dequeued(p
);
1693 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1698 * __normal_prio - return the priority that is based on the static prio
1700 static inline int __normal_prio(struct task_struct
*p
)
1702 return p
->static_prio
;
1706 * Calculate the expected normal priority: i.e. priority
1707 * without taking RT-inheritance into account. Might be
1708 * boosted by interactivity modifiers. Changes upon fork,
1709 * setprio syscalls, and whenever the interactivity
1710 * estimator recalculates.
1712 static inline int normal_prio(struct task_struct
*p
)
1716 if (task_has_rt_policy(p
))
1717 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1719 prio
= __normal_prio(p
);
1724 * Calculate the current priority, i.e. the priority
1725 * taken into account by the scheduler. This value might
1726 * be boosted by RT tasks, or might be boosted by
1727 * interactivity modifiers. Will be RT if the task got
1728 * RT-boosted. If not then it returns p->normal_prio.
1730 static int effective_prio(struct task_struct
*p
)
1732 p
->normal_prio
= normal_prio(p
);
1734 * If we are RT tasks or we were boosted to RT priority,
1735 * keep the priority unchanged. Otherwise, update priority
1736 * to the normal priority:
1738 if (!rt_prio(p
->prio
))
1739 return p
->normal_prio
;
1744 * activate_task - move a task to the runqueue.
1746 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1748 if (task_contributes_to_load(p
))
1749 rq
->nr_uninterruptible
--;
1751 enqueue_task(rq
, p
, wakeup
);
1756 * deactivate_task - remove a task from the runqueue.
1758 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1760 if (task_contributes_to_load(p
))
1761 rq
->nr_uninterruptible
++;
1763 dequeue_task(rq
, p
, sleep
);
1768 * task_curr - is this task currently executing on a CPU?
1769 * @p: the task in question.
1771 inline int task_curr(const struct task_struct
*p
)
1773 return cpu_curr(task_cpu(p
)) == p
;
1776 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1778 set_task_rq(p
, cpu
);
1781 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1782 * successfuly executed on another CPU. We must ensure that updates of
1783 * per-task data have been completed by this moment.
1786 task_thread_info(p
)->cpu
= cpu
;
1790 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1791 const struct sched_class
*prev_class
,
1792 int oldprio
, int running
)
1794 if (prev_class
!= p
->sched_class
) {
1795 if (prev_class
->switched_from
)
1796 prev_class
->switched_from(rq
, p
, running
);
1797 p
->sched_class
->switched_to(rq
, p
, running
);
1799 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
1804 /* Used instead of source_load when we know the type == 0 */
1805 static unsigned long weighted_cpuload(const int cpu
)
1807 return cpu_rq(cpu
)->load
.weight
;
1811 * Is this task likely cache-hot:
1814 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
1819 * Buddy candidates are cache hot:
1821 if (sched_feat(CACHE_HOT_BUDDY
) &&
1822 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
1823 &p
->se
== cfs_rq_of(&p
->se
)->last
))
1826 if (p
->sched_class
!= &fair_sched_class
)
1829 if (sysctl_sched_migration_cost
== -1)
1831 if (sysctl_sched_migration_cost
== 0)
1834 delta
= now
- p
->se
.exec_start
;
1836 return delta
< (s64
)sysctl_sched_migration_cost
;
1840 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
1842 int old_cpu
= task_cpu(p
);
1843 struct rq
*old_rq
= cpu_rq(old_cpu
), *new_rq
= cpu_rq(new_cpu
);
1844 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
1845 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
1848 clock_offset
= old_rq
->clock
- new_rq
->clock
;
1850 #ifdef CONFIG_SCHEDSTATS
1851 if (p
->se
.wait_start
)
1852 p
->se
.wait_start
-= clock_offset
;
1853 if (p
->se
.sleep_start
)
1854 p
->se
.sleep_start
-= clock_offset
;
1855 if (p
->se
.block_start
)
1856 p
->se
.block_start
-= clock_offset
;
1857 if (old_cpu
!= new_cpu
) {
1858 schedstat_inc(p
, se
.nr_migrations
);
1859 if (task_hot(p
, old_rq
->clock
, NULL
))
1860 schedstat_inc(p
, se
.nr_forced2_migrations
);
1863 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
1864 new_cfsrq
->min_vruntime
;
1866 __set_task_cpu(p
, new_cpu
);
1869 struct migration_req
{
1870 struct list_head list
;
1872 struct task_struct
*task
;
1875 struct completion done
;
1879 * The task's runqueue lock must be held.
1880 * Returns true if you have to wait for migration thread.
1883 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1885 struct rq
*rq
= task_rq(p
);
1888 * If the task is not on a runqueue (and not running), then
1889 * it is sufficient to simply update the task's cpu field.
1891 if (!p
->se
.on_rq
&& !task_running(rq
, p
)) {
1892 set_task_cpu(p
, dest_cpu
);
1896 init_completion(&req
->done
);
1898 req
->dest_cpu
= dest_cpu
;
1899 list_add(&req
->list
, &rq
->migration_queue
);
1905 * wait_task_inactive - wait for a thread to unschedule.
1907 * If @match_state is nonzero, it's the @p->state value just checked and
1908 * not expected to change. If it changes, i.e. @p might have woken up,
1909 * then return zero. When we succeed in waiting for @p to be off its CPU,
1910 * we return a positive number (its total switch count). If a second call
1911 * a short while later returns the same number, the caller can be sure that
1912 * @p has remained unscheduled the whole time.
1914 * The caller must ensure that the task *will* unschedule sometime soon,
1915 * else this function might spin for a *long* time. This function can't
1916 * be called with interrupts off, or it may introduce deadlock with
1917 * smp_call_function() if an IPI is sent by the same process we are
1918 * waiting to become inactive.
1920 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
1922 unsigned long flags
;
1929 * We do the initial early heuristics without holding
1930 * any task-queue locks at all. We'll only try to get
1931 * the runqueue lock when things look like they will
1937 * If the task is actively running on another CPU
1938 * still, just relax and busy-wait without holding
1941 * NOTE! Since we don't hold any locks, it's not
1942 * even sure that "rq" stays as the right runqueue!
1943 * But we don't care, since "task_running()" will
1944 * return false if the runqueue has changed and p
1945 * is actually now running somewhere else!
1947 while (task_running(rq
, p
)) {
1948 if (match_state
&& unlikely(p
->state
!= match_state
))
1954 * Ok, time to look more closely! We need the rq
1955 * lock now, to be *sure*. If we're wrong, we'll
1956 * just go back and repeat.
1958 rq
= task_rq_lock(p
, &flags
);
1959 trace_sched_wait_task(rq
, p
);
1960 running
= task_running(rq
, p
);
1961 on_rq
= p
->se
.on_rq
;
1963 if (!match_state
|| p
->state
== match_state
)
1964 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
1965 task_rq_unlock(rq
, &flags
);
1968 * If it changed from the expected state, bail out now.
1970 if (unlikely(!ncsw
))
1974 * Was it really running after all now that we
1975 * checked with the proper locks actually held?
1977 * Oops. Go back and try again..
1979 if (unlikely(running
)) {
1985 * It's not enough that it's not actively running,
1986 * it must be off the runqueue _entirely_, and not
1989 * So if it wa still runnable (but just not actively
1990 * running right now), it's preempted, and we should
1991 * yield - it could be a while.
1993 if (unlikely(on_rq
)) {
1994 schedule_timeout_uninterruptible(1);
1999 * Ahh, all good. It wasn't running, and it wasn't
2000 * runnable, which means that it will never become
2001 * running in the future either. We're all done!
2010 * kick_process - kick a running thread to enter/exit the kernel
2011 * @p: the to-be-kicked thread
2013 * Cause a process which is running on another CPU to enter
2014 * kernel-mode, without any delay. (to get signals handled.)
2016 * NOTE: this function doesnt have to take the runqueue lock,
2017 * because all it wants to ensure is that the remote task enters
2018 * the kernel. If the IPI races and the task has been migrated
2019 * to another CPU then no harm is done and the purpose has been
2022 void kick_process(struct task_struct
*p
)
2028 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2029 smp_send_reschedule(cpu
);
2034 * Return a low guess at the load of a migration-source cpu weighted
2035 * according to the scheduling class and "nice" value.
2037 * We want to under-estimate the load of migration sources, to
2038 * balance conservatively.
2040 static unsigned long source_load(int cpu
, int type
)
2042 struct rq
*rq
= cpu_rq(cpu
);
2043 unsigned long total
= weighted_cpuload(cpu
);
2045 if (type
== 0 || !sched_feat(LB_BIAS
))
2048 return min(rq
->cpu_load
[type
-1], total
);
2052 * Return a high guess at the load of a migration-target cpu weighted
2053 * according to the scheduling class and "nice" value.
2055 static unsigned long target_load(int cpu
, int type
)
2057 struct rq
*rq
= cpu_rq(cpu
);
2058 unsigned long total
= weighted_cpuload(cpu
);
2060 if (type
== 0 || !sched_feat(LB_BIAS
))
2063 return max(rq
->cpu_load
[type
-1], total
);
2067 * find_idlest_group finds and returns the least busy CPU group within the
2070 static struct sched_group
*
2071 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
2073 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2074 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
2075 int load_idx
= sd
->forkexec_idx
;
2076 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
2079 unsigned long load
, avg_load
;
2083 /* Skip over this group if it has no CPUs allowed */
2084 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
2087 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2089 /* Tally up the load of all CPUs in the group */
2092 for_each_cpu_mask_nr(i
, group
->cpumask
) {
2093 /* Bias balancing toward cpus of our domain */
2095 load
= source_load(i
, load_idx
);
2097 load
= target_load(i
, load_idx
);
2102 /* Adjust by relative CPU power of the group */
2103 avg_load
= sg_div_cpu_power(group
,
2104 avg_load
* SCHED_LOAD_SCALE
);
2107 this_load
= avg_load
;
2109 } else if (avg_load
< min_load
) {
2110 min_load
= avg_load
;
2113 } while (group
= group
->next
, group
!= sd
->groups
);
2115 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
2121 * find_idlest_cpu - find the idlest cpu among the cpus in group.
2124 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
,
2127 unsigned long load
, min_load
= ULONG_MAX
;
2131 /* Traverse only the allowed CPUs */
2132 cpus_and(*tmp
, group
->cpumask
, p
->cpus_allowed
);
2134 for_each_cpu_mask_nr(i
, *tmp
) {
2135 load
= weighted_cpuload(i
);
2137 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
2147 * sched_balance_self: balance the current task (running on cpu) in domains
2148 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
2151 * Balance, ie. select the least loaded group.
2153 * Returns the target CPU number, or the same CPU if no balancing is needed.
2155 * preempt must be disabled.
2157 static int sched_balance_self(int cpu
, int flag
)
2159 struct task_struct
*t
= current
;
2160 struct sched_domain
*tmp
, *sd
= NULL
;
2162 for_each_domain(cpu
, tmp
) {
2164 * If power savings logic is enabled for a domain, stop there.
2166 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
2168 if (tmp
->flags
& flag
)
2176 cpumask_t span
, tmpmask
;
2177 struct sched_group
*group
;
2178 int new_cpu
, weight
;
2180 if (!(sd
->flags
& flag
)) {
2186 group
= find_idlest_group(sd
, t
, cpu
);
2192 new_cpu
= find_idlest_cpu(group
, t
, cpu
, &tmpmask
);
2193 if (new_cpu
== -1 || new_cpu
== cpu
) {
2194 /* Now try balancing at a lower domain level of cpu */
2199 /* Now try balancing at a lower domain level of new_cpu */
2202 weight
= cpus_weight(span
);
2203 for_each_domain(cpu
, tmp
) {
2204 if (weight
<= cpus_weight(tmp
->span
))
2206 if (tmp
->flags
& flag
)
2209 /* while loop will break here if sd == NULL */
2215 #endif /* CONFIG_SMP */
2218 * try_to_wake_up - wake up a thread
2219 * @p: the to-be-woken-up thread
2220 * @state: the mask of task states that can be woken
2221 * @sync: do a synchronous wakeup?
2223 * Put it on the run-queue if it's not already there. The "current"
2224 * thread is always on the run-queue (except when the actual
2225 * re-schedule is in progress), and as such you're allowed to do
2226 * the simpler "current->state = TASK_RUNNING" to mark yourself
2227 * runnable without the overhead of this.
2229 * returns failure only if the task is already active.
2231 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
2233 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2234 unsigned long flags
;
2238 if (!sched_feat(SYNC_WAKEUPS
))
2242 if (sched_feat(LB_WAKEUP_UPDATE
)) {
2243 struct sched_domain
*sd
;
2245 this_cpu
= raw_smp_processor_id();
2248 for_each_domain(this_cpu
, sd
) {
2249 if (cpu_isset(cpu
, sd
->span
)) {
2258 rq
= task_rq_lock(p
, &flags
);
2259 old_state
= p
->state
;
2260 if (!(old_state
& state
))
2268 this_cpu
= smp_processor_id();
2271 if (unlikely(task_running(rq
, p
)))
2274 cpu
= p
->sched_class
->select_task_rq(p
, sync
);
2275 if (cpu
!= orig_cpu
) {
2276 set_task_cpu(p
, cpu
);
2277 task_rq_unlock(rq
, &flags
);
2278 /* might preempt at this point */
2279 rq
= task_rq_lock(p
, &flags
);
2280 old_state
= p
->state
;
2281 if (!(old_state
& state
))
2286 this_cpu
= smp_processor_id();
2290 #ifdef CONFIG_SCHEDSTATS
2291 schedstat_inc(rq
, ttwu_count
);
2292 if (cpu
== this_cpu
)
2293 schedstat_inc(rq
, ttwu_local
);
2295 struct sched_domain
*sd
;
2296 for_each_domain(this_cpu
, sd
) {
2297 if (cpu_isset(cpu
, sd
->span
)) {
2298 schedstat_inc(sd
, ttwu_wake_remote
);
2303 #endif /* CONFIG_SCHEDSTATS */
2306 #endif /* CONFIG_SMP */
2307 schedstat_inc(p
, se
.nr_wakeups
);
2309 schedstat_inc(p
, se
.nr_wakeups_sync
);
2310 if (orig_cpu
!= cpu
)
2311 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2312 if (cpu
== this_cpu
)
2313 schedstat_inc(p
, se
.nr_wakeups_local
);
2315 schedstat_inc(p
, se
.nr_wakeups_remote
);
2316 update_rq_clock(rq
);
2317 activate_task(rq
, p
, 1);
2321 trace_sched_wakeup(rq
, p
);
2322 check_preempt_curr(rq
, p
, sync
);
2324 p
->state
= TASK_RUNNING
;
2326 if (p
->sched_class
->task_wake_up
)
2327 p
->sched_class
->task_wake_up(rq
, p
);
2330 current
->se
.last_wakeup
= current
->se
.sum_exec_runtime
;
2332 task_rq_unlock(rq
, &flags
);
2337 int wake_up_process(struct task_struct
*p
)
2339 return try_to_wake_up(p
, TASK_ALL
, 0);
2341 EXPORT_SYMBOL(wake_up_process
);
2343 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2345 return try_to_wake_up(p
, state
, 0);
2349 * Perform scheduler related setup for a newly forked process p.
2350 * p is forked by current.
2352 * __sched_fork() is basic setup used by init_idle() too:
2354 static void __sched_fork(struct task_struct
*p
)
2356 p
->se
.exec_start
= 0;
2357 p
->se
.sum_exec_runtime
= 0;
2358 p
->se
.prev_sum_exec_runtime
= 0;
2359 p
->se
.last_wakeup
= 0;
2360 p
->se
.avg_overlap
= 0;
2362 #ifdef CONFIG_SCHEDSTATS
2363 p
->se
.wait_start
= 0;
2364 p
->se
.sum_sleep_runtime
= 0;
2365 p
->se
.sleep_start
= 0;
2366 p
->se
.block_start
= 0;
2367 p
->se
.sleep_max
= 0;
2368 p
->se
.block_max
= 0;
2370 p
->se
.slice_max
= 0;
2374 INIT_LIST_HEAD(&p
->rt
.run_list
);
2376 INIT_LIST_HEAD(&p
->se
.group_node
);
2378 #ifdef CONFIG_PREEMPT_NOTIFIERS
2379 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2383 * We mark the process as running here, but have not actually
2384 * inserted it onto the runqueue yet. This guarantees that
2385 * nobody will actually run it, and a signal or other external
2386 * event cannot wake it up and insert it on the runqueue either.
2388 p
->state
= TASK_RUNNING
;
2392 * fork()/clone()-time setup:
2394 void sched_fork(struct task_struct
*p
, int clone_flags
)
2396 int cpu
= get_cpu();
2401 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
2403 set_task_cpu(p
, cpu
);
2406 * Make sure we do not leak PI boosting priority to the child:
2408 p
->prio
= current
->normal_prio
;
2409 if (!rt_prio(p
->prio
))
2410 p
->sched_class
= &fair_sched_class
;
2412 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2413 if (likely(sched_info_on()))
2414 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2416 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2419 #ifdef CONFIG_PREEMPT
2420 /* Want to start with kernel preemption disabled. */
2421 task_thread_info(p
)->preempt_count
= 1;
2427 * wake_up_new_task - wake up a newly created task for the first time.
2429 * This function will do some initial scheduler statistics housekeeping
2430 * that must be done for every newly created context, then puts the task
2431 * on the runqueue and wakes it.
2433 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2435 unsigned long flags
;
2438 rq
= task_rq_lock(p
, &flags
);
2439 BUG_ON(p
->state
!= TASK_RUNNING
);
2440 update_rq_clock(rq
);
2442 p
->prio
= effective_prio(p
);
2444 if (!p
->sched_class
->task_new
|| !current
->se
.on_rq
) {
2445 activate_task(rq
, p
, 0);
2448 * Let the scheduling class do new task startup
2449 * management (if any):
2451 p
->sched_class
->task_new(rq
, p
);
2454 trace_sched_wakeup_new(rq
, p
);
2455 check_preempt_curr(rq
, p
, 0);
2457 if (p
->sched_class
->task_wake_up
)
2458 p
->sched_class
->task_wake_up(rq
, p
);
2460 task_rq_unlock(rq
, &flags
);
2463 #ifdef CONFIG_PREEMPT_NOTIFIERS
2466 * preempt_notifier_register - tell me when current is being being preempted & rescheduled
2467 * @notifier: notifier struct to register
2469 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2471 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2473 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2476 * preempt_notifier_unregister - no longer interested in preemption notifications
2477 * @notifier: notifier struct to unregister
2479 * This is safe to call from within a preemption notifier.
2481 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2483 hlist_del(¬ifier
->link
);
2485 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2487 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2489 struct preempt_notifier
*notifier
;
2490 struct hlist_node
*node
;
2492 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2493 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2497 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2498 struct task_struct
*next
)
2500 struct preempt_notifier
*notifier
;
2501 struct hlist_node
*node
;
2503 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2504 notifier
->ops
->sched_out(notifier
, next
);
2507 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2509 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2514 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2515 struct task_struct
*next
)
2519 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2522 * prepare_task_switch - prepare to switch tasks
2523 * @rq: the runqueue preparing to switch
2524 * @prev: the current task that is being switched out
2525 * @next: the task we are going to switch to.
2527 * This is called with the rq lock held and interrupts off. It must
2528 * be paired with a subsequent finish_task_switch after the context
2531 * prepare_task_switch sets up locking and calls architecture specific
2535 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2536 struct task_struct
*next
)
2538 fire_sched_out_preempt_notifiers(prev
, next
);
2539 prepare_lock_switch(rq
, next
);
2540 prepare_arch_switch(next
);
2544 * finish_task_switch - clean up after a task-switch
2545 * @rq: runqueue associated with task-switch
2546 * @prev: the thread we just switched away from.
2548 * finish_task_switch must be called after the context switch, paired
2549 * with a prepare_task_switch call before the context switch.
2550 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2551 * and do any other architecture-specific cleanup actions.
2553 * Note that we may have delayed dropping an mm in context_switch(). If
2554 * so, we finish that here outside of the runqueue lock. (Doing it
2555 * with the lock held can cause deadlocks; see schedule() for
2558 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2559 __releases(rq
->lock
)
2561 struct mm_struct
*mm
= rq
->prev_mm
;
2567 * A task struct has one reference for the use as "current".
2568 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2569 * schedule one last time. The schedule call will never return, and
2570 * the scheduled task must drop that reference.
2571 * The test for TASK_DEAD must occur while the runqueue locks are
2572 * still held, otherwise prev could be scheduled on another cpu, die
2573 * there before we look at prev->state, and then the reference would
2575 * Manfred Spraul <manfred@colorfullife.com>
2577 prev_state
= prev
->state
;
2578 finish_arch_switch(prev
);
2579 finish_lock_switch(rq
, prev
);
2581 if (current
->sched_class
->post_schedule
)
2582 current
->sched_class
->post_schedule(rq
);
2585 fire_sched_in_preempt_notifiers(current
);
2588 if (unlikely(prev_state
== TASK_DEAD
)) {
2590 * Remove function-return probe instances associated with this
2591 * task and put them back on the free list.
2593 kprobe_flush_task(prev
);
2594 put_task_struct(prev
);
2599 * schedule_tail - first thing a freshly forked thread must call.
2600 * @prev: the thread we just switched away from.
2602 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2603 __releases(rq
->lock
)
2605 struct rq
*rq
= this_rq();
2607 finish_task_switch(rq
, prev
);
2608 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2609 /* In this case, finish_task_switch does not reenable preemption */
2612 if (current
->set_child_tid
)
2613 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2617 * context_switch - switch to the new MM and the new
2618 * thread's register state.
2621 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2622 struct task_struct
*next
)
2624 struct mm_struct
*mm
, *oldmm
;
2626 prepare_task_switch(rq
, prev
, next
);
2627 trace_sched_switch(rq
, prev
, next
);
2629 oldmm
= prev
->active_mm
;
2631 * For paravirt, this is coupled with an exit in switch_to to
2632 * combine the page table reload and the switch backend into
2635 arch_enter_lazy_cpu_mode();
2637 if (unlikely(!mm
)) {
2638 next
->active_mm
= oldmm
;
2639 atomic_inc(&oldmm
->mm_count
);
2640 enter_lazy_tlb(oldmm
, next
);
2642 switch_mm(oldmm
, mm
, next
);
2644 if (unlikely(!prev
->mm
)) {
2645 prev
->active_mm
= NULL
;
2646 rq
->prev_mm
= oldmm
;
2649 * Since the runqueue lock will be released by the next
2650 * task (which is an invalid locking op but in the case
2651 * of the scheduler it's an obvious special-case), so we
2652 * do an early lockdep release here:
2654 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2655 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2658 /* Here we just switch the register state and the stack. */
2659 switch_to(prev
, next
, prev
);
2663 * this_rq must be evaluated again because prev may have moved
2664 * CPUs since it called schedule(), thus the 'rq' on its stack
2665 * frame will be invalid.
2667 finish_task_switch(this_rq(), prev
);
2671 * nr_running, nr_uninterruptible and nr_context_switches:
2673 * externally visible scheduler statistics: current number of runnable
2674 * threads, current number of uninterruptible-sleeping threads, total
2675 * number of context switches performed since bootup.
2677 unsigned long nr_running(void)
2679 unsigned long i
, sum
= 0;
2681 for_each_online_cpu(i
)
2682 sum
+= cpu_rq(i
)->nr_running
;
2687 unsigned long nr_uninterruptible(void)
2689 unsigned long i
, sum
= 0;
2691 for_each_possible_cpu(i
)
2692 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2695 * Since we read the counters lockless, it might be slightly
2696 * inaccurate. Do not allow it to go below zero though:
2698 if (unlikely((long)sum
< 0))
2704 unsigned long long nr_context_switches(void)
2707 unsigned long long sum
= 0;
2709 for_each_possible_cpu(i
)
2710 sum
+= cpu_rq(i
)->nr_switches
;
2715 unsigned long nr_iowait(void)
2717 unsigned long i
, sum
= 0;
2719 for_each_possible_cpu(i
)
2720 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2725 unsigned long nr_active(void)
2727 unsigned long i
, running
= 0, uninterruptible
= 0;
2729 for_each_online_cpu(i
) {
2730 running
+= cpu_rq(i
)->nr_running
;
2731 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
2734 if (unlikely((long)uninterruptible
< 0))
2735 uninterruptible
= 0;
2737 return running
+ uninterruptible
;
2741 * Update rq->cpu_load[] statistics. This function is usually called every
2742 * scheduler tick (TICK_NSEC).
2744 static void update_cpu_load(struct rq
*this_rq
)
2746 unsigned long this_load
= this_rq
->load
.weight
;
2749 this_rq
->nr_load_updates
++;
2751 /* Update our load: */
2752 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
2753 unsigned long old_load
, new_load
;
2755 /* scale is effectively 1 << i now, and >> i divides by scale */
2757 old_load
= this_rq
->cpu_load
[i
];
2758 new_load
= this_load
;
2760 * Round up the averaging division if load is increasing. This
2761 * prevents us from getting stuck on 9 if the load is 10, for
2764 if (new_load
> old_load
)
2765 new_load
+= scale
-1;
2766 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
2773 * double_rq_lock - safely lock two runqueues
2775 * Note this does not disable interrupts like task_rq_lock,
2776 * you need to do so manually before calling.
2778 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
2779 __acquires(rq1
->lock
)
2780 __acquires(rq2
->lock
)
2782 BUG_ON(!irqs_disabled());
2784 spin_lock(&rq1
->lock
);
2785 __acquire(rq2
->lock
); /* Fake it out ;) */
2788 spin_lock(&rq1
->lock
);
2789 spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
2791 spin_lock(&rq2
->lock
);
2792 spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
2795 update_rq_clock(rq1
);
2796 update_rq_clock(rq2
);
2800 * double_rq_unlock - safely unlock two runqueues
2802 * Note this does not restore interrupts like task_rq_unlock,
2803 * you need to do so manually after calling.
2805 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
2806 __releases(rq1
->lock
)
2807 __releases(rq2
->lock
)
2809 spin_unlock(&rq1
->lock
);
2811 spin_unlock(&rq2
->lock
);
2813 __release(rq2
->lock
);
2817 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2819 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2820 __releases(this_rq
->lock
)
2821 __acquires(busiest
->lock
)
2822 __acquires(this_rq
->lock
)
2826 if (unlikely(!irqs_disabled())) {
2827 /* printk() doesn't work good under rq->lock */
2828 spin_unlock(&this_rq
->lock
);
2831 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2832 if (busiest
< this_rq
) {
2833 spin_unlock(&this_rq
->lock
);
2834 spin_lock(&busiest
->lock
);
2835 spin_lock_nested(&this_rq
->lock
, SINGLE_DEPTH_NESTING
);
2838 spin_lock_nested(&busiest
->lock
, SINGLE_DEPTH_NESTING
);
2843 static void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2844 __releases(busiest
->lock
)
2846 spin_unlock(&busiest
->lock
);
2847 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
2851 * If dest_cpu is allowed for this process, migrate the task to it.
2852 * This is accomplished by forcing the cpu_allowed mask to only
2853 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2854 * the cpu_allowed mask is restored.
2856 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2858 struct migration_req req
;
2859 unsigned long flags
;
2862 rq
= task_rq_lock(p
, &flags
);
2863 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2864 || unlikely(!cpu_active(dest_cpu
)))
2867 trace_sched_migrate_task(rq
, p
, dest_cpu
);
2868 /* force the process onto the specified CPU */
2869 if (migrate_task(p
, dest_cpu
, &req
)) {
2870 /* Need to wait for migration thread (might exit: take ref). */
2871 struct task_struct
*mt
= rq
->migration_thread
;
2873 get_task_struct(mt
);
2874 task_rq_unlock(rq
, &flags
);
2875 wake_up_process(mt
);
2876 put_task_struct(mt
);
2877 wait_for_completion(&req
.done
);
2882 task_rq_unlock(rq
, &flags
);
2886 * sched_exec - execve() is a valuable balancing opportunity, because at
2887 * this point the task has the smallest effective memory and cache footprint.
2889 void sched_exec(void)
2891 int new_cpu
, this_cpu
= get_cpu();
2892 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2894 if (new_cpu
!= this_cpu
)
2895 sched_migrate_task(current
, new_cpu
);
2899 * pull_task - move a task from a remote runqueue to the local runqueue.
2900 * Both runqueues must be locked.
2902 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
2903 struct rq
*this_rq
, int this_cpu
)
2905 deactivate_task(src_rq
, p
, 0);
2906 set_task_cpu(p
, this_cpu
);
2907 activate_task(this_rq
, p
, 0);
2909 * Note that idle threads have a prio of MAX_PRIO, for this test
2910 * to be always true for them.
2912 check_preempt_curr(this_rq
, p
, 0);
2916 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2919 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2920 struct sched_domain
*sd
, enum cpu_idle_type idle
,
2924 * We do not migrate tasks that are:
2925 * 1) running (obviously), or
2926 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2927 * 3) are cache-hot on their current CPU.
2929 if (!cpu_isset(this_cpu
, p
->cpus_allowed
)) {
2930 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
2935 if (task_running(rq
, p
)) {
2936 schedstat_inc(p
, se
.nr_failed_migrations_running
);
2941 * Aggressive migration if:
2942 * 1) task is cache cold, or
2943 * 2) too many balance attempts have failed.
2946 if (!task_hot(p
, rq
->clock
, sd
) ||
2947 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2948 #ifdef CONFIG_SCHEDSTATS
2949 if (task_hot(p
, rq
->clock
, sd
)) {
2950 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2951 schedstat_inc(p
, se
.nr_forced_migrations
);
2957 if (task_hot(p
, rq
->clock
, sd
)) {
2958 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
2964 static unsigned long
2965 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2966 unsigned long max_load_move
, struct sched_domain
*sd
,
2967 enum cpu_idle_type idle
, int *all_pinned
,
2968 int *this_best_prio
, struct rq_iterator
*iterator
)
2970 int loops
= 0, pulled
= 0, pinned
= 0;
2971 struct task_struct
*p
;
2972 long rem_load_move
= max_load_move
;
2974 if (max_load_move
== 0)
2980 * Start the load-balancing iterator:
2982 p
= iterator
->start(iterator
->arg
);
2984 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
2987 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
2988 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2989 p
= iterator
->next(iterator
->arg
);
2993 pull_task(busiest
, p
, this_rq
, this_cpu
);
2995 rem_load_move
-= p
->se
.load
.weight
;
2998 * We only want to steal up to the prescribed amount of weighted load.
3000 if (rem_load_move
> 0) {
3001 if (p
->prio
< *this_best_prio
)
3002 *this_best_prio
= p
->prio
;
3003 p
= iterator
->next(iterator
->arg
);
3008 * Right now, this is one of only two places pull_task() is called,
3009 * so we can safely collect pull_task() stats here rather than
3010 * inside pull_task().
3012 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3015 *all_pinned
= pinned
;
3017 return max_load_move
- rem_load_move
;
3021 * move_tasks tries to move up to max_load_move weighted load from busiest to
3022 * this_rq, as part of a balancing operation within domain "sd".
3023 * Returns 1 if successful and 0 otherwise.
3025 * Called with both runqueues locked.
3027 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3028 unsigned long max_load_move
,
3029 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3032 const struct sched_class
*class = sched_class_highest
;
3033 unsigned long total_load_moved
= 0;
3034 int this_best_prio
= this_rq
->curr
->prio
;
3038 class->load_balance(this_rq
, this_cpu
, busiest
,
3039 max_load_move
- total_load_moved
,
3040 sd
, idle
, all_pinned
, &this_best_prio
);
3041 class = class->next
;
3043 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3046 } while (class && max_load_move
> total_load_moved
);
3048 return total_load_moved
> 0;
3052 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3053 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3054 struct rq_iterator
*iterator
)
3056 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3060 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3061 pull_task(busiest
, p
, this_rq
, this_cpu
);
3063 * Right now, this is only the second place pull_task()
3064 * is called, so we can safely collect pull_task()
3065 * stats here rather than inside pull_task().
3067 schedstat_inc(sd
, lb_gained
[idle
]);
3071 p
= iterator
->next(iterator
->arg
);
3078 * move_one_task tries to move exactly one task from busiest to this_rq, as
3079 * part of active balancing operations within "domain".
3080 * Returns 1 if successful and 0 otherwise.
3082 * Called with both runqueues locked.
3084 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3085 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3087 const struct sched_class
*class;
3089 for (class = sched_class_highest
; class; class = class->next
)
3090 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3097 * find_busiest_group finds and returns the busiest CPU group within the
3098 * domain. It calculates and returns the amount of weighted load which
3099 * should be moved to restore balance via the imbalance parameter.
3101 static struct sched_group
*
3102 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
3103 unsigned long *imbalance
, enum cpu_idle_type idle
,
3104 int *sd_idle
, const cpumask_t
*cpus
, int *balance
)
3106 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
3107 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
3108 unsigned long max_pull
;
3109 unsigned long busiest_load_per_task
, busiest_nr_running
;
3110 unsigned long this_load_per_task
, this_nr_running
;
3111 int load_idx
, group_imb
= 0;
3112 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3113 int power_savings_balance
= 1;
3114 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
3115 unsigned long min_nr_running
= ULONG_MAX
;
3116 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
3119 max_load
= this_load
= total_load
= total_pwr
= 0;
3120 busiest_load_per_task
= busiest_nr_running
= 0;
3121 this_load_per_task
= this_nr_running
= 0;
3123 if (idle
== CPU_NOT_IDLE
)
3124 load_idx
= sd
->busy_idx
;
3125 else if (idle
== CPU_NEWLY_IDLE
)
3126 load_idx
= sd
->newidle_idx
;
3128 load_idx
= sd
->idle_idx
;
3131 unsigned long load
, group_capacity
, max_cpu_load
, min_cpu_load
;
3134 int __group_imb
= 0;
3135 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3136 unsigned long sum_nr_running
, sum_weighted_load
;
3137 unsigned long sum_avg_load_per_task
;
3138 unsigned long avg_load_per_task
;
3140 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
3143 balance_cpu
= first_cpu(group
->cpumask
);
3145 /* Tally up the load of all CPUs in the group */
3146 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
3147 sum_avg_load_per_task
= avg_load_per_task
= 0;
3150 min_cpu_load
= ~0UL;
3152 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3155 if (!cpu_isset(i
, *cpus
))
3160 if (*sd_idle
&& rq
->nr_running
)
3163 /* Bias balancing toward cpus of our domain */
3165 if (idle_cpu(i
) && !first_idle_cpu
) {
3170 load
= target_load(i
, load_idx
);
3172 load
= source_load(i
, load_idx
);
3173 if (load
> max_cpu_load
)
3174 max_cpu_load
= load
;
3175 if (min_cpu_load
> load
)
3176 min_cpu_load
= load
;
3180 sum_nr_running
+= rq
->nr_running
;
3181 sum_weighted_load
+= weighted_cpuload(i
);
3183 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3187 * First idle cpu or the first cpu(busiest) in this sched group
3188 * is eligible for doing load balancing at this and above
3189 * domains. In the newly idle case, we will allow all the cpu's
3190 * to do the newly idle load balance.
3192 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3193 balance_cpu
!= this_cpu
&& balance
) {
3198 total_load
+= avg_load
;
3199 total_pwr
+= group
->__cpu_power
;
3201 /* Adjust by relative CPU power of the group */
3202 avg_load
= sg_div_cpu_power(group
,
3203 avg_load
* SCHED_LOAD_SCALE
);
3207 * Consider the group unbalanced when the imbalance is larger
3208 * than the average weight of two tasks.
3210 * APZ: with cgroup the avg task weight can vary wildly and
3211 * might not be a suitable number - should we keep a
3212 * normalized nr_running number somewhere that negates
3215 avg_load_per_task
= sg_div_cpu_power(group
,
3216 sum_avg_load_per_task
* SCHED_LOAD_SCALE
);
3218 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3221 group_capacity
= group
->__cpu_power
/ SCHED_LOAD_SCALE
;
3224 this_load
= avg_load
;
3226 this_nr_running
= sum_nr_running
;
3227 this_load_per_task
= sum_weighted_load
;
3228 } else if (avg_load
> max_load
&&
3229 (sum_nr_running
> group_capacity
|| __group_imb
)) {
3230 max_load
= avg_load
;
3232 busiest_nr_running
= sum_nr_running
;
3233 busiest_load_per_task
= sum_weighted_load
;
3234 group_imb
= __group_imb
;
3237 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3239 * Busy processors will not participate in power savings
3242 if (idle
== CPU_NOT_IDLE
||
3243 !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3247 * If the local group is idle or completely loaded
3248 * no need to do power savings balance at this domain
3250 if (local_group
&& (this_nr_running
>= group_capacity
||
3252 power_savings_balance
= 0;
3255 * If a group is already running at full capacity or idle,
3256 * don't include that group in power savings calculations
3258 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
3263 * Calculate the group which has the least non-idle load.
3264 * This is the group from where we need to pick up the load
3267 if ((sum_nr_running
< min_nr_running
) ||
3268 (sum_nr_running
== min_nr_running
&&
3269 first_cpu(group
->cpumask
) <
3270 first_cpu(group_min
->cpumask
))) {
3272 min_nr_running
= sum_nr_running
;
3273 min_load_per_task
= sum_weighted_load
/
3278 * Calculate the group which is almost near its
3279 * capacity but still has some space to pick up some load
3280 * from other group and save more power
3282 if (sum_nr_running
<= group_capacity
- 1) {
3283 if (sum_nr_running
> leader_nr_running
||
3284 (sum_nr_running
== leader_nr_running
&&
3285 first_cpu(group
->cpumask
) >
3286 first_cpu(group_leader
->cpumask
))) {
3287 group_leader
= group
;
3288 leader_nr_running
= sum_nr_running
;
3293 group
= group
->next
;
3294 } while (group
!= sd
->groups
);
3296 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
3299 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
3301 if (this_load
>= avg_load
||
3302 100*max_load
<= sd
->imbalance_pct
*this_load
)
3305 busiest_load_per_task
/= busiest_nr_running
;
3307 busiest_load_per_task
= min(busiest_load_per_task
, avg_load
);
3310 * We're trying to get all the cpus to the average_load, so we don't
3311 * want to push ourselves above the average load, nor do we wish to
3312 * reduce the max loaded cpu below the average load, as either of these
3313 * actions would just result in more rebalancing later, and ping-pong
3314 * tasks around. Thus we look for the minimum possible imbalance.
3315 * Negative imbalances (*we* are more loaded than anyone else) will
3316 * be counted as no imbalance for these purposes -- we can't fix that
3317 * by pulling tasks to us. Be careful of negative numbers as they'll
3318 * appear as very large values with unsigned longs.
3320 if (max_load
<= busiest_load_per_task
)
3324 * In the presence of smp nice balancing, certain scenarios can have
3325 * max load less than avg load(as we skip the groups at or below
3326 * its cpu_power, while calculating max_load..)
3328 if (max_load
< avg_load
) {
3330 goto small_imbalance
;
3333 /* Don't want to pull so many tasks that a group would go idle */
3334 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
3336 /* How much load to actually move to equalise the imbalance */
3337 *imbalance
= min(max_pull
* busiest
->__cpu_power
,
3338 (avg_load
- this_load
) * this->__cpu_power
)
3342 * if *imbalance is less than the average load per runnable task
3343 * there is no gaurantee that any tasks will be moved so we'll have
3344 * a think about bumping its value to force at least one task to be
3347 if (*imbalance
< busiest_load_per_task
) {
3348 unsigned long tmp
, pwr_now
, pwr_move
;
3352 pwr_move
= pwr_now
= 0;
3354 if (this_nr_running
) {
3355 this_load_per_task
/= this_nr_running
;
3356 if (busiest_load_per_task
> this_load_per_task
)
3359 this_load_per_task
= cpu_avg_load_per_task(this_cpu
);
3361 if (max_load
- this_load
+ busiest_load_per_task
>=
3362 busiest_load_per_task
* imbn
) {
3363 *imbalance
= busiest_load_per_task
;
3368 * OK, we don't have enough imbalance to justify moving tasks,
3369 * however we may be able to increase total CPU power used by
3373 pwr_now
+= busiest
->__cpu_power
*
3374 min(busiest_load_per_task
, max_load
);
3375 pwr_now
+= this->__cpu_power
*
3376 min(this_load_per_task
, this_load
);
3377 pwr_now
/= SCHED_LOAD_SCALE
;
3379 /* Amount of load we'd subtract */
3380 tmp
= sg_div_cpu_power(busiest
,
3381 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3383 pwr_move
+= busiest
->__cpu_power
*
3384 min(busiest_load_per_task
, max_load
- tmp
);
3386 /* Amount of load we'd add */
3387 if (max_load
* busiest
->__cpu_power
<
3388 busiest_load_per_task
* SCHED_LOAD_SCALE
)
3389 tmp
= sg_div_cpu_power(this,
3390 max_load
* busiest
->__cpu_power
);
3392 tmp
= sg_div_cpu_power(this,
3393 busiest_load_per_task
* SCHED_LOAD_SCALE
);
3394 pwr_move
+= this->__cpu_power
*
3395 min(this_load_per_task
, this_load
+ tmp
);
3396 pwr_move
/= SCHED_LOAD_SCALE
;
3398 /* Move if we gain throughput */
3399 if (pwr_move
> pwr_now
)
3400 *imbalance
= busiest_load_per_task
;
3406 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3407 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3410 if (this == group_leader
&& group_leader
!= group_min
) {
3411 *imbalance
= min_load_per_task
;
3421 * find_busiest_queue - find the busiest runqueue among the cpus in group.
3424 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
3425 unsigned long imbalance
, const cpumask_t
*cpus
)
3427 struct rq
*busiest
= NULL
, *rq
;
3428 unsigned long max_load
= 0;
3431 for_each_cpu_mask_nr(i
, group
->cpumask
) {
3434 if (!cpu_isset(i
, *cpus
))
3438 wl
= weighted_cpuload(i
);
3440 if (rq
->nr_running
== 1 && wl
> imbalance
)
3443 if (wl
> max_load
) {
3453 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
3454 * so long as it is large enough.
3456 #define MAX_PINNED_INTERVAL 512
3459 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3460 * tasks if there is an imbalance.
3462 static int load_balance(int this_cpu
, struct rq
*this_rq
,
3463 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3464 int *balance
, cpumask_t
*cpus
)
3466 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
3467 struct sched_group
*group
;
3468 unsigned long imbalance
;
3470 unsigned long flags
;
3475 * When power savings policy is enabled for the parent domain, idle
3476 * sibling can pick up load irrespective of busy siblings. In this case,
3477 * let the state of idle sibling percolate up as CPU_IDLE, instead of
3478 * portraying it as CPU_NOT_IDLE.
3480 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3481 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3484 schedstat_inc(sd
, lb_count
[idle
]);
3488 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
3495 schedstat_inc(sd
, lb_nobusyg
[idle
]);
3499 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
3501 schedstat_inc(sd
, lb_nobusyq
[idle
]);
3505 BUG_ON(busiest
== this_rq
);
3507 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
3510 if (busiest
->nr_running
> 1) {
3512 * Attempt to move tasks. If find_busiest_group has found
3513 * an imbalance but busiest->nr_running <= 1, the group is
3514 * still unbalanced. ld_moved simply stays zero, so it is
3515 * correctly treated as an imbalance.
3517 local_irq_save(flags
);
3518 double_rq_lock(this_rq
, busiest
);
3519 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3520 imbalance
, sd
, idle
, &all_pinned
);
3521 double_rq_unlock(this_rq
, busiest
);
3522 local_irq_restore(flags
);
3525 * some other cpu did the load balance for us.
3527 if (ld_moved
&& this_cpu
!= smp_processor_id())
3528 resched_cpu(this_cpu
);
3530 /* All tasks on this runqueue were pinned by CPU affinity */
3531 if (unlikely(all_pinned
)) {
3532 cpu_clear(cpu_of(busiest
), *cpus
);
3533 if (!cpus_empty(*cpus
))
3540 schedstat_inc(sd
, lb_failed
[idle
]);
3541 sd
->nr_balance_failed
++;
3543 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
3545 spin_lock_irqsave(&busiest
->lock
, flags
);
3547 /* don't kick the migration_thread, if the curr
3548 * task on busiest cpu can't be moved to this_cpu
3550 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
3551 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3553 goto out_one_pinned
;
3556 if (!busiest
->active_balance
) {
3557 busiest
->active_balance
= 1;
3558 busiest
->push_cpu
= this_cpu
;
3561 spin_unlock_irqrestore(&busiest
->lock
, flags
);
3563 wake_up_process(busiest
->migration_thread
);
3566 * We've kicked active balancing, reset the failure
3569 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
3572 sd
->nr_balance_failed
= 0;
3574 if (likely(!active_balance
)) {
3575 /* We were unbalanced, so reset the balancing interval */
3576 sd
->balance_interval
= sd
->min_interval
;
3579 * If we've begun active balancing, start to back off. This
3580 * case may not be covered by the all_pinned logic if there
3581 * is only 1 task on the busy runqueue (because we don't call
3584 if (sd
->balance_interval
< sd
->max_interval
)
3585 sd
->balance_interval
*= 2;
3588 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3589 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3595 schedstat_inc(sd
, lb_balanced
[idle
]);
3597 sd
->nr_balance_failed
= 0;
3600 /* tune up the balancing interval */
3601 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
3602 (sd
->balance_interval
< sd
->max_interval
))
3603 sd
->balance_interval
*= 2;
3605 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3606 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3617 * Check this_cpu to ensure it is balanced within domain. Attempt to move
3618 * tasks if there is an imbalance.
3620 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
3621 * this_rq is locked.
3624 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
,
3627 struct sched_group
*group
;
3628 struct rq
*busiest
= NULL
;
3629 unsigned long imbalance
;
3637 * When power savings policy is enabled for the parent domain, idle
3638 * sibling can pick up load irrespective of busy siblings. In this case,
3639 * let the state of idle sibling percolate up as IDLE, instead of
3640 * portraying it as CPU_NOT_IDLE.
3642 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
3643 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3646 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
3648 update_shares_locked(this_rq
, sd
);
3649 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
3650 &sd_idle
, cpus
, NULL
);
3652 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
3656 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
3658 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
3662 BUG_ON(busiest
== this_rq
);
3664 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
3667 if (busiest
->nr_running
> 1) {
3668 /* Attempt to move tasks */
3669 double_lock_balance(this_rq
, busiest
);
3670 /* this_rq->clock is already updated */
3671 update_rq_clock(busiest
);
3672 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
3673 imbalance
, sd
, CPU_NEWLY_IDLE
,
3675 double_unlock_balance(this_rq
, busiest
);
3677 if (unlikely(all_pinned
)) {
3678 cpu_clear(cpu_of(busiest
), *cpus
);
3679 if (!cpus_empty(*cpus
))
3685 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
3686 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3687 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3690 sd
->nr_balance_failed
= 0;
3692 update_shares_locked(this_rq
, sd
);
3696 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
3697 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
3698 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
3700 sd
->nr_balance_failed
= 0;
3706 * idle_balance is called by schedule() if this_cpu is about to become
3707 * idle. Attempts to pull tasks from other CPUs.
3709 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
3711 struct sched_domain
*sd
;
3712 int pulled_task
= -1;
3713 unsigned long next_balance
= jiffies
+ HZ
;
3716 for_each_domain(this_cpu
, sd
) {
3717 unsigned long interval
;
3719 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3722 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
3723 /* If we've pulled tasks over stop searching: */
3724 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
3727 interval
= msecs_to_jiffies(sd
->balance_interval
);
3728 if (time_after(next_balance
, sd
->last_balance
+ interval
))
3729 next_balance
= sd
->last_balance
+ interval
;
3733 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
3735 * We are going idle. next_balance may be set based on
3736 * a busy processor. So reset next_balance.
3738 this_rq
->next_balance
= next_balance
;
3743 * active_load_balance is run by migration threads. It pushes running tasks
3744 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
3745 * running on each physical CPU where possible, and avoids physical /
3746 * logical imbalances.
3748 * Called with busiest_rq locked.
3750 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
3752 int target_cpu
= busiest_rq
->push_cpu
;
3753 struct sched_domain
*sd
;
3754 struct rq
*target_rq
;
3756 /* Is there any task to move? */
3757 if (busiest_rq
->nr_running
<= 1)
3760 target_rq
= cpu_rq(target_cpu
);
3763 * This condition is "impossible", if it occurs
3764 * we need to fix it. Originally reported by
3765 * Bjorn Helgaas on a 128-cpu setup.
3767 BUG_ON(busiest_rq
== target_rq
);
3769 /* move a task from busiest_rq to target_rq */
3770 double_lock_balance(busiest_rq
, target_rq
);
3771 update_rq_clock(busiest_rq
);
3772 update_rq_clock(target_rq
);
3774 /* Search for an sd spanning us and the target CPU. */
3775 for_each_domain(target_cpu
, sd
) {
3776 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
3777 cpu_isset(busiest_cpu
, sd
->span
))
3782 schedstat_inc(sd
, alb_count
);
3784 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
3786 schedstat_inc(sd
, alb_pushed
);
3788 schedstat_inc(sd
, alb_failed
);
3790 double_unlock_balance(busiest_rq
, target_rq
);
3795 atomic_t load_balancer
;
3797 } nohz ____cacheline_aligned
= {
3798 .load_balancer
= ATOMIC_INIT(-1),
3799 .cpu_mask
= CPU_MASK_NONE
,
3803 * This routine will try to nominate the ilb (idle load balancing)
3804 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
3805 * load balancing on behalf of all those cpus. If all the cpus in the system
3806 * go into this tickless mode, then there will be no ilb owner (as there is
3807 * no need for one) and all the cpus will sleep till the next wakeup event
3810 * For the ilb owner, tick is not stopped. And this tick will be used
3811 * for idle load balancing. ilb owner will still be part of
3814 * While stopping the tick, this cpu will become the ilb owner if there
3815 * is no other owner. And will be the owner till that cpu becomes busy
3816 * or if all cpus in the system stop their ticks at which point
3817 * there is no need for ilb owner.
3819 * When the ilb owner becomes busy, it nominates another owner, during the
3820 * next busy scheduler_tick()
3822 int select_nohz_load_balancer(int stop_tick
)
3824 int cpu
= smp_processor_id();
3827 cpu_set(cpu
, nohz
.cpu_mask
);
3828 cpu_rq(cpu
)->in_nohz_recently
= 1;
3831 * If we are going offline and still the leader, give up!
3833 if (!cpu_active(cpu
) &&
3834 atomic_read(&nohz
.load_balancer
) == cpu
) {
3835 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3840 /* time for ilb owner also to sleep */
3841 if (cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
3842 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3843 atomic_set(&nohz
.load_balancer
, -1);
3847 if (atomic_read(&nohz
.load_balancer
) == -1) {
3848 /* make me the ilb owner */
3849 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
3851 } else if (atomic_read(&nohz
.load_balancer
) == cpu
)
3854 if (!cpu_isset(cpu
, nohz
.cpu_mask
))
3857 cpu_clear(cpu
, nohz
.cpu_mask
);
3859 if (atomic_read(&nohz
.load_balancer
) == cpu
)
3860 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
3867 static DEFINE_SPINLOCK(balancing
);
3870 * It checks each scheduling domain to see if it is due to be balanced,
3871 * and initiates a balancing operation if so.
3873 * Balancing parameters are set up in arch_init_sched_domains.
3875 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
3878 struct rq
*rq
= cpu_rq(cpu
);
3879 unsigned long interval
;
3880 struct sched_domain
*sd
;
3881 /* Earliest time when we have to do rebalance again */
3882 unsigned long next_balance
= jiffies
+ 60*HZ
;
3883 int update_next_balance
= 0;
3887 for_each_domain(cpu
, sd
) {
3888 if (!(sd
->flags
& SD_LOAD_BALANCE
))
3891 interval
= sd
->balance_interval
;
3892 if (idle
!= CPU_IDLE
)
3893 interval
*= sd
->busy_factor
;
3895 /* scale ms to jiffies */
3896 interval
= msecs_to_jiffies(interval
);
3897 if (unlikely(!interval
))
3899 if (interval
> HZ
*NR_CPUS
/10)
3900 interval
= HZ
*NR_CPUS
/10;
3902 need_serialize
= sd
->flags
& SD_SERIALIZE
;
3904 if (need_serialize
) {
3905 if (!spin_trylock(&balancing
))
3909 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
3910 if (load_balance(cpu
, rq
, sd
, idle
, &balance
, &tmp
)) {
3912 * We've pulled tasks over so either we're no
3913 * longer idle, or one of our SMT siblings is
3916 idle
= CPU_NOT_IDLE
;
3918 sd
->last_balance
= jiffies
;
3921 spin_unlock(&balancing
);
3923 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
3924 next_balance
= sd
->last_balance
+ interval
;
3925 update_next_balance
= 1;
3929 * Stop the load balance at this level. There is another
3930 * CPU in our sched group which is doing load balancing more
3938 * next_balance will be updated only when there is a need.
3939 * When the cpu is attached to null domain for ex, it will not be
3942 if (likely(update_next_balance
))
3943 rq
->next_balance
= next_balance
;
3947 * run_rebalance_domains is triggered when needed from the scheduler tick.
3948 * In CONFIG_NO_HZ case, the idle load balance owner will do the
3949 * rebalancing for all the cpus for whom scheduler ticks are stopped.
3951 static void run_rebalance_domains(struct softirq_action
*h
)
3953 int this_cpu
= smp_processor_id();
3954 struct rq
*this_rq
= cpu_rq(this_cpu
);
3955 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
3956 CPU_IDLE
: CPU_NOT_IDLE
;
3958 rebalance_domains(this_cpu
, idle
);
3962 * If this cpu is the owner for idle load balancing, then do the
3963 * balancing on behalf of the other idle cpus whose ticks are
3966 if (this_rq
->idle_at_tick
&&
3967 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
3968 cpumask_t cpus
= nohz
.cpu_mask
;
3972 cpu_clear(this_cpu
, cpus
);
3973 for_each_cpu_mask_nr(balance_cpu
, cpus
) {
3975 * If this cpu gets work to do, stop the load balancing
3976 * work being done for other cpus. Next load
3977 * balancing owner will pick it up.
3982 rebalance_domains(balance_cpu
, CPU_IDLE
);
3984 rq
= cpu_rq(balance_cpu
);
3985 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
3986 this_rq
->next_balance
= rq
->next_balance
;
3993 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
3995 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
3996 * idle load balancing owner or decide to stop the periodic load balancing,
3997 * if the whole system is idle.
3999 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4003 * If we were in the nohz mode recently and busy at the current
4004 * scheduler tick, then check if we need to nominate new idle
4007 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4008 rq
->in_nohz_recently
= 0;
4010 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4011 cpu_clear(cpu
, nohz
.cpu_mask
);
4012 atomic_set(&nohz
.load_balancer
, -1);
4015 if (atomic_read(&nohz
.load_balancer
) == -1) {
4017 * simple selection for now: Nominate the
4018 * first cpu in the nohz list to be the next
4021 * TBD: Traverse the sched domains and nominate
4022 * the nearest cpu in the nohz.cpu_mask.
4024 int ilb
= first_cpu(nohz
.cpu_mask
);
4026 if (ilb
< nr_cpu_ids
)
4032 * If this cpu is idle and doing idle load balancing for all the
4033 * cpus with ticks stopped, is it time for that to stop?
4035 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4036 cpus_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4042 * If this cpu is idle and the idle load balancing is done by
4043 * someone else, then no need raise the SCHED_SOFTIRQ
4045 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4046 cpu_isset(cpu
, nohz
.cpu_mask
))
4049 if (time_after_eq(jiffies
, rq
->next_balance
))
4050 raise_softirq(SCHED_SOFTIRQ
);
4053 #else /* CONFIG_SMP */
4056 * on UP we do not need to balance between CPUs:
4058 static inline void idle_balance(int cpu
, struct rq
*rq
)
4064 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4066 EXPORT_PER_CPU_SYMBOL(kstat
);
4069 * Return any ns on the sched_clock that have not yet been banked in
4070 * @p in case that task is currently running.
4072 unsigned long long task_delta_exec(struct task_struct
*p
)
4074 unsigned long flags
;
4078 rq
= task_rq_lock(p
, &flags
);
4080 if (task_current(rq
, p
)) {
4083 update_rq_clock(rq
);
4084 delta_exec
= rq
->clock
- p
->se
.exec_start
;
4085 if ((s64
)delta_exec
> 0)
4089 task_rq_unlock(rq
, &flags
);
4095 * Account user cpu time to a process.
4096 * @p: the process that the cpu time gets accounted to
4097 * @cputime: the cpu time spent in user space since the last update
4099 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
4101 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4104 p
->utime
= cputime_add(p
->utime
, cputime
);
4105 account_group_user_time(p
, cputime
);
4107 /* Add user time to cpustat. */
4108 tmp
= cputime_to_cputime64(cputime
);
4109 if (TASK_NICE(p
) > 0)
4110 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
4112 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4113 /* Account for user time used */
4114 acct_update_integrals(p
);
4118 * Account guest cpu time to a process.
4119 * @p: the process that the cpu time gets accounted to
4120 * @cputime: the cpu time spent in virtual machine since the last update
4122 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
)
4125 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4127 tmp
= cputime_to_cputime64(cputime
);
4129 p
->utime
= cputime_add(p
->utime
, cputime
);
4130 account_group_user_time(p
, cputime
);
4131 p
->gtime
= cputime_add(p
->gtime
, cputime
);
4133 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
4134 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
4138 * Account scaled user cpu time to a process.
4139 * @p: the process that the cpu time gets accounted to
4140 * @cputime: the cpu time spent in user space since the last update
4142 void account_user_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4144 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime
);
4148 * Account system cpu time to a process.
4149 * @p: the process that the cpu time gets accounted to
4150 * @hardirq_offset: the offset to subtract from hardirq_count()
4151 * @cputime: the cpu time spent in kernel space since the last update
4153 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
4156 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4157 struct rq
*rq
= this_rq();
4160 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
4161 account_guest_time(p
, cputime
);
4165 p
->stime
= cputime_add(p
->stime
, cputime
);
4166 account_group_system_time(p
, cputime
);
4168 /* Add system time to cpustat. */
4169 tmp
= cputime_to_cputime64(cputime
);
4170 if (hardirq_count() - hardirq_offset
)
4171 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
4172 else if (softirq_count())
4173 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
4174 else if (p
!= rq
->idle
)
4175 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
4176 else if (atomic_read(&rq
->nr_iowait
) > 0)
4177 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4179 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4180 /* Account for system time used */
4181 acct_update_integrals(p
);
4185 * Account scaled system cpu time to a process.
4186 * @p: the process that the cpu time gets accounted to
4187 * @hardirq_offset: the offset to subtract from hardirq_count()
4188 * @cputime: the cpu time spent in kernel space since the last update
4190 void account_system_time_scaled(struct task_struct
*p
, cputime_t cputime
)
4192 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime
);
4196 * Account for involuntary wait time.
4197 * @p: the process from which the cpu time has been stolen
4198 * @steal: the cpu time spent in involuntary wait
4200 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
4202 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
4203 cputime64_t tmp
= cputime_to_cputime64(steal
);
4204 struct rq
*rq
= this_rq();
4206 if (p
== rq
->idle
) {
4207 p
->stime
= cputime_add(p
->stime
, steal
);
4208 account_group_system_time(p
, steal
);
4209 if (atomic_read(&rq
->nr_iowait
) > 0)
4210 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
4212 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
4214 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
4218 * Use precise platform statistics if available:
4220 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
4221 cputime_t
task_utime(struct task_struct
*p
)
4226 cputime_t
task_stime(struct task_struct
*p
)
4231 cputime_t
task_utime(struct task_struct
*p
)
4233 clock_t utime
= cputime_to_clock_t(p
->utime
),
4234 total
= utime
+ cputime_to_clock_t(p
->stime
);
4238 * Use CFS's precise accounting:
4240 temp
= (u64
)nsec_to_clock_t(p
->se
.sum_exec_runtime
);
4244 do_div(temp
, total
);
4246 utime
= (clock_t)temp
;
4248 p
->prev_utime
= max(p
->prev_utime
, clock_t_to_cputime(utime
));
4249 return p
->prev_utime
;
4252 cputime_t
task_stime(struct task_struct
*p
)
4257 * Use CFS's precise accounting. (we subtract utime from
4258 * the total, to make sure the total observed by userspace
4259 * grows monotonically - apps rely on that):
4261 stime
= nsec_to_clock_t(p
->se
.sum_exec_runtime
) -
4262 cputime_to_clock_t(task_utime(p
));
4265 p
->prev_stime
= max(p
->prev_stime
, clock_t_to_cputime(stime
));
4267 return p
->prev_stime
;
4271 inline cputime_t
task_gtime(struct task_struct
*p
)
4277 * This function gets called by the timer code, with HZ frequency.
4278 * We call it with interrupts disabled.
4280 * It also gets called by the fork code, when changing the parent's
4283 void scheduler_tick(void)
4285 int cpu
= smp_processor_id();
4286 struct rq
*rq
= cpu_rq(cpu
);
4287 struct task_struct
*curr
= rq
->curr
;
4291 spin_lock(&rq
->lock
);
4292 update_rq_clock(rq
);
4293 update_cpu_load(rq
);
4294 curr
->sched_class
->task_tick(rq
, curr
, 0);
4295 spin_unlock(&rq
->lock
);
4298 rq
->idle_at_tick
= idle_cpu(cpu
);
4299 trigger_load_balance(rq
, cpu
);
4303 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
4304 defined(CONFIG_PREEMPT_TRACER))
4306 static inline unsigned long get_parent_ip(unsigned long addr
)
4308 if (in_lock_functions(addr
)) {
4309 addr
= CALLER_ADDR2
;
4310 if (in_lock_functions(addr
))
4311 addr
= CALLER_ADDR3
;
4316 void __kprobes
add_preempt_count(int val
)
4318 #ifdef CONFIG_DEBUG_PREEMPT
4322 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
4325 preempt_count() += val
;
4326 #ifdef CONFIG_DEBUG_PREEMPT
4328 * Spinlock count overflowing soon?
4330 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
4333 if (preempt_count() == val
)
4334 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4336 EXPORT_SYMBOL(add_preempt_count
);
4338 void __kprobes
sub_preempt_count(int val
)
4340 #ifdef CONFIG_DEBUG_PREEMPT
4344 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
4347 * Is the spinlock portion underflowing?
4349 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
4350 !(preempt_count() & PREEMPT_MASK
)))
4354 if (preempt_count() == val
)
4355 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
4356 preempt_count() -= val
;
4358 EXPORT_SYMBOL(sub_preempt_count
);
4363 * Print scheduling while atomic bug:
4365 static noinline
void __schedule_bug(struct task_struct
*prev
)
4367 struct pt_regs
*regs
= get_irq_regs();
4369 printk(KERN_ERR
"BUG: scheduling while atomic: %s/%d/0x%08x\n",
4370 prev
->comm
, prev
->pid
, preempt_count());
4372 debug_show_held_locks(prev
);
4374 if (irqs_disabled())
4375 print_irqtrace_events(prev
);
4384 * Various schedule()-time debugging checks and statistics:
4386 static inline void schedule_debug(struct task_struct
*prev
)
4389 * Test if we are atomic. Since do_exit() needs to call into
4390 * schedule() atomically, we ignore that path for now.
4391 * Otherwise, whine if we are scheduling when we should not be.
4393 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
4394 __schedule_bug(prev
);
4396 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
4398 schedstat_inc(this_rq(), sched_count
);
4399 #ifdef CONFIG_SCHEDSTATS
4400 if (unlikely(prev
->lock_depth
>= 0)) {
4401 schedstat_inc(this_rq(), bkl_count
);
4402 schedstat_inc(prev
, sched_info
.bkl_count
);
4408 * Pick up the highest-prio task:
4410 static inline struct task_struct
*
4411 pick_next_task(struct rq
*rq
, struct task_struct
*prev
)
4413 const struct sched_class
*class;
4414 struct task_struct
*p
;
4417 * Optimization: we know that if all tasks are in
4418 * the fair class we can call that function directly:
4420 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
4421 p
= fair_sched_class
.pick_next_task(rq
);
4426 class = sched_class_highest
;
4428 p
= class->pick_next_task(rq
);
4432 * Will never be NULL as the idle class always
4433 * returns a non-NULL p:
4435 class = class->next
;
4440 * schedule() is the main scheduler function.
4442 asmlinkage
void __sched
schedule(void)
4444 struct task_struct
*prev
, *next
;
4445 unsigned long *switch_count
;
4451 cpu
= smp_processor_id();
4455 switch_count
= &prev
->nivcsw
;
4457 release_kernel_lock(prev
);
4458 need_resched_nonpreemptible
:
4460 schedule_debug(prev
);
4462 if (sched_feat(HRTICK
))
4465 spin_lock_irq(&rq
->lock
);
4466 update_rq_clock(rq
);
4467 clear_tsk_need_resched(prev
);
4469 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
4470 if (unlikely(signal_pending_state(prev
->state
, prev
)))
4471 prev
->state
= TASK_RUNNING
;
4473 deactivate_task(rq
, prev
, 1);
4474 switch_count
= &prev
->nvcsw
;
4478 if (prev
->sched_class
->pre_schedule
)
4479 prev
->sched_class
->pre_schedule(rq
, prev
);
4482 if (unlikely(!rq
->nr_running
))
4483 idle_balance(cpu
, rq
);
4485 prev
->sched_class
->put_prev_task(rq
, prev
);
4486 next
= pick_next_task(rq
, prev
);
4488 if (likely(prev
!= next
)) {
4489 sched_info_switch(prev
, next
);
4495 context_switch(rq
, prev
, next
); /* unlocks the rq */
4497 * the context switch might have flipped the stack from under
4498 * us, hence refresh the local variables.
4500 cpu
= smp_processor_id();
4503 spin_unlock_irq(&rq
->lock
);
4505 if (unlikely(reacquire_kernel_lock(current
) < 0))
4506 goto need_resched_nonpreemptible
;
4508 preempt_enable_no_resched();
4509 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
4512 EXPORT_SYMBOL(schedule
);
4514 #ifdef CONFIG_PREEMPT
4516 * this is the entry point to schedule() from in-kernel preemption
4517 * off of preempt_enable. Kernel preemptions off return from interrupt
4518 * occur there and call schedule directly.
4520 asmlinkage
void __sched
preempt_schedule(void)
4522 struct thread_info
*ti
= current_thread_info();
4525 * If there is a non-zero preempt_count or interrupts are disabled,
4526 * we do not want to preempt the current task. Just return..
4528 if (likely(ti
->preempt_count
|| irqs_disabled()))
4532 add_preempt_count(PREEMPT_ACTIVE
);
4534 sub_preempt_count(PREEMPT_ACTIVE
);
4537 * Check again in case we missed a preemption opportunity
4538 * between schedule and now.
4541 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4543 EXPORT_SYMBOL(preempt_schedule
);
4546 * this is the entry point to schedule() from kernel preemption
4547 * off of irq context.
4548 * Note, that this is called and return with irqs disabled. This will
4549 * protect us against recursive calling from irq.
4551 asmlinkage
void __sched
preempt_schedule_irq(void)
4553 struct thread_info
*ti
= current_thread_info();
4555 /* Catch callers which need to be fixed */
4556 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
4559 add_preempt_count(PREEMPT_ACTIVE
);
4562 local_irq_disable();
4563 sub_preempt_count(PREEMPT_ACTIVE
);
4566 * Check again in case we missed a preemption opportunity
4567 * between schedule and now.
4570 } while (unlikely(test_thread_flag(TIF_NEED_RESCHED
)));
4573 #endif /* CONFIG_PREEMPT */
4575 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
4578 return try_to_wake_up(curr
->private, mode
, sync
);
4580 EXPORT_SYMBOL(default_wake_function
);
4583 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
4584 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
4585 * number) then we wake all the non-exclusive tasks and one exclusive task.
4587 * There are circumstances in which we can try to wake a task which has already
4588 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
4589 * zero in this (rare) case, and we handle it by continuing to scan the queue.
4591 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
4592 int nr_exclusive
, int sync
, void *key
)
4594 wait_queue_t
*curr
, *next
;
4596 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
4597 unsigned flags
= curr
->flags
;
4599 if (curr
->func(curr
, mode
, sync
, key
) &&
4600 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
4606 * __wake_up - wake up threads blocked on a waitqueue.
4608 * @mode: which threads
4609 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4610 * @key: is directly passed to the wakeup function
4612 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
4613 int nr_exclusive
, void *key
)
4615 unsigned long flags
;
4617 spin_lock_irqsave(&q
->lock
, flags
);
4618 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
4619 spin_unlock_irqrestore(&q
->lock
, flags
);
4621 EXPORT_SYMBOL(__wake_up
);
4624 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
4626 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
4628 __wake_up_common(q
, mode
, 1, 0, NULL
);
4632 * __wake_up_sync - wake up threads blocked on a waitqueue.
4634 * @mode: which threads
4635 * @nr_exclusive: how many wake-one or wake-many threads to wake up
4637 * The sync wakeup differs that the waker knows that it will schedule
4638 * away soon, so while the target thread will be woken up, it will not
4639 * be migrated to another CPU - ie. the two threads are 'synchronized'
4640 * with each other. This can prevent needless bouncing between CPUs.
4642 * On UP it can prevent extra preemption.
4645 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
4647 unsigned long flags
;
4653 if (unlikely(!nr_exclusive
))
4656 spin_lock_irqsave(&q
->lock
, flags
);
4657 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
4658 spin_unlock_irqrestore(&q
->lock
, flags
);
4660 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
4663 * complete: - signals a single thread waiting on this completion
4664 * @x: holds the state of this particular completion
4666 * This will wake up a single thread waiting on this completion. Threads will be
4667 * awakened in the same order in which they were queued.
4669 * See also complete_all(), wait_for_completion() and related routines.
4671 void complete(struct completion
*x
)
4673 unsigned long flags
;
4675 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4677 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
4678 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4680 EXPORT_SYMBOL(complete
);
4683 * complete_all: - signals all threads waiting on this completion
4684 * @x: holds the state of this particular completion
4686 * This will wake up all threads waiting on this particular completion event.
4688 void complete_all(struct completion
*x
)
4690 unsigned long flags
;
4692 spin_lock_irqsave(&x
->wait
.lock
, flags
);
4693 x
->done
+= UINT_MAX
/2;
4694 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
4695 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
4697 EXPORT_SYMBOL(complete_all
);
4699 static inline long __sched
4700 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
4703 DECLARE_WAITQUEUE(wait
, current
);
4705 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
4706 __add_wait_queue_tail(&x
->wait
, &wait
);
4708 if (signal_pending_state(state
, current
)) {
4709 timeout
= -ERESTARTSYS
;
4712 __set_current_state(state
);
4713 spin_unlock_irq(&x
->wait
.lock
);
4714 timeout
= schedule_timeout(timeout
);
4715 spin_lock_irq(&x
->wait
.lock
);
4716 } while (!x
->done
&& timeout
);
4717 __remove_wait_queue(&x
->wait
, &wait
);
4722 return timeout
?: 1;
4726 wait_for_common(struct completion
*x
, long timeout
, int state
)
4730 spin_lock_irq(&x
->wait
.lock
);
4731 timeout
= do_wait_for_common(x
, timeout
, state
);
4732 spin_unlock_irq(&x
->wait
.lock
);
4737 * wait_for_completion: - waits for completion of a task
4738 * @x: holds the state of this particular completion
4740 * This waits to be signaled for completion of a specific task. It is NOT
4741 * interruptible and there is no timeout.
4743 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
4744 * and interrupt capability. Also see complete().
4746 void __sched
wait_for_completion(struct completion
*x
)
4748 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
4750 EXPORT_SYMBOL(wait_for_completion
);
4753 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
4754 * @x: holds the state of this particular completion
4755 * @timeout: timeout value in jiffies
4757 * This waits for either a completion of a specific task to be signaled or for a
4758 * specified timeout to expire. The timeout is in jiffies. It is not
4761 unsigned long __sched
4762 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
4764 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
4766 EXPORT_SYMBOL(wait_for_completion_timeout
);
4769 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
4770 * @x: holds the state of this particular completion
4772 * This waits for completion of a specific task to be signaled. It is
4775 int __sched
wait_for_completion_interruptible(struct completion
*x
)
4777 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
4778 if (t
== -ERESTARTSYS
)
4782 EXPORT_SYMBOL(wait_for_completion_interruptible
);
4785 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
4786 * @x: holds the state of this particular completion
4787 * @timeout: timeout value in jiffies
4789 * This waits for either a completion of a specific task to be signaled or for a
4790 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
4792 unsigned long __sched
4793 wait_for_completion_interruptible_timeout(struct completion
*x
,
4794 unsigned long timeout
)
4796 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
4798 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
4801 * wait_for_completion_killable: - waits for completion of a task (killable)
4802 * @x: holds the state of this particular completion
4804 * This waits to be signaled for completion of a specific task. It can be
4805 * interrupted by a kill signal.
4807 int __sched
wait_for_completion_killable(struct completion
*x
)
4809 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
4810 if (t
== -ERESTARTSYS
)
4814 EXPORT_SYMBOL(wait_for_completion_killable
);
4817 * try_wait_for_completion - try to decrement a completion without blocking
4818 * @x: completion structure
4820 * Returns: 0 if a decrement cannot be done without blocking
4821 * 1 if a decrement succeeded.
4823 * If a completion is being used as a counting completion,
4824 * attempt to decrement the counter without blocking. This
4825 * enables us to avoid waiting if the resource the completion
4826 * is protecting is not available.
4828 bool try_wait_for_completion(struct completion
*x
)
4832 spin_lock_irq(&x
->wait
.lock
);
4837 spin_unlock_irq(&x
->wait
.lock
);
4840 EXPORT_SYMBOL(try_wait_for_completion
);
4843 * completion_done - Test to see if a completion has any waiters
4844 * @x: completion structure
4846 * Returns: 0 if there are waiters (wait_for_completion() in progress)
4847 * 1 if there are no waiters.
4850 bool completion_done(struct completion
*x
)
4854 spin_lock_irq(&x
->wait
.lock
);
4857 spin_unlock_irq(&x
->wait
.lock
);
4860 EXPORT_SYMBOL(completion_done
);
4863 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
4865 unsigned long flags
;
4868 init_waitqueue_entry(&wait
, current
);
4870 __set_current_state(state
);
4872 spin_lock_irqsave(&q
->lock
, flags
);
4873 __add_wait_queue(q
, &wait
);
4874 spin_unlock(&q
->lock
);
4875 timeout
= schedule_timeout(timeout
);
4876 spin_lock_irq(&q
->lock
);
4877 __remove_wait_queue(q
, &wait
);
4878 spin_unlock_irqrestore(&q
->lock
, flags
);
4883 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
4885 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4887 EXPORT_SYMBOL(interruptible_sleep_on
);
4890 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4892 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
4894 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
4896 void __sched
sleep_on(wait_queue_head_t
*q
)
4898 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
4900 EXPORT_SYMBOL(sleep_on
);
4902 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
4904 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
4906 EXPORT_SYMBOL(sleep_on_timeout
);
4908 #ifdef CONFIG_RT_MUTEXES
4911 * rt_mutex_setprio - set the current priority of a task
4913 * @prio: prio value (kernel-internal form)
4915 * This function changes the 'effective' priority of a task. It does
4916 * not touch ->normal_prio like __setscheduler().
4918 * Used by the rt_mutex code to implement priority inheritance logic.
4920 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
4922 unsigned long flags
;
4923 int oldprio
, on_rq
, running
;
4925 const struct sched_class
*prev_class
= p
->sched_class
;
4927 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
4929 rq
= task_rq_lock(p
, &flags
);
4930 update_rq_clock(rq
);
4933 on_rq
= p
->se
.on_rq
;
4934 running
= task_current(rq
, p
);
4936 dequeue_task(rq
, p
, 0);
4938 p
->sched_class
->put_prev_task(rq
, p
);
4941 p
->sched_class
= &rt_sched_class
;
4943 p
->sched_class
= &fair_sched_class
;
4948 p
->sched_class
->set_curr_task(rq
);
4950 enqueue_task(rq
, p
, 0);
4952 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
4954 task_rq_unlock(rq
, &flags
);
4959 void set_user_nice(struct task_struct
*p
, long nice
)
4961 int old_prio
, delta
, on_rq
;
4962 unsigned long flags
;
4965 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4968 * We have to be careful, if called from sys_setpriority(),
4969 * the task might be in the middle of scheduling on another CPU.
4971 rq
= task_rq_lock(p
, &flags
);
4972 update_rq_clock(rq
);
4974 * The RT priorities are set via sched_setscheduler(), but we still
4975 * allow the 'normal' nice value to be set - but as expected
4976 * it wont have any effect on scheduling until the task is
4977 * SCHED_FIFO/SCHED_RR:
4979 if (task_has_rt_policy(p
)) {
4980 p
->static_prio
= NICE_TO_PRIO(nice
);
4983 on_rq
= p
->se
.on_rq
;
4985 dequeue_task(rq
, p
, 0);
4987 p
->static_prio
= NICE_TO_PRIO(nice
);
4990 p
->prio
= effective_prio(p
);
4991 delta
= p
->prio
- old_prio
;
4994 enqueue_task(rq
, p
, 0);
4996 * If the task increased its priority or is running and
4997 * lowered its priority, then reschedule its CPU:
4999 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
5000 resched_task(rq
->curr
);
5003 task_rq_unlock(rq
, &flags
);
5005 EXPORT_SYMBOL(set_user_nice
);
5008 * can_nice - check if a task can reduce its nice value
5012 int can_nice(const struct task_struct
*p
, const int nice
)
5014 /* convert nice value [19,-20] to rlimit style value [1,40] */
5015 int nice_rlim
= 20 - nice
;
5017 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
5018 capable(CAP_SYS_NICE
));
5021 #ifdef __ARCH_WANT_SYS_NICE
5024 * sys_nice - change the priority of the current process.
5025 * @increment: priority increment
5027 * sys_setpriority is a more generic, but much slower function that
5028 * does similar things.
5030 asmlinkage
long sys_nice(int increment
)
5035 * Setpriority might change our priority at the same moment.
5036 * We don't have to worry. Conceptually one call occurs first
5037 * and we have a single winner.
5039 if (increment
< -40)
5044 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
5050 if (increment
< 0 && !can_nice(current
, nice
))
5053 retval
= security_task_setnice(current
, nice
);
5057 set_user_nice(current
, nice
);
5064 * task_prio - return the priority value of a given task.
5065 * @p: the task in question.
5067 * This is the priority value as seen by users in /proc.
5068 * RT tasks are offset by -200. Normal tasks are centered
5069 * around 0, value goes from -16 to +15.
5071 int task_prio(const struct task_struct
*p
)
5073 return p
->prio
- MAX_RT_PRIO
;
5077 * task_nice - return the nice value of a given task.
5078 * @p: the task in question.
5080 int task_nice(const struct task_struct
*p
)
5082 return TASK_NICE(p
);
5084 EXPORT_SYMBOL(task_nice
);
5087 * idle_cpu - is a given cpu idle currently?
5088 * @cpu: the processor in question.
5090 int idle_cpu(int cpu
)
5092 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
5096 * idle_task - return the idle task for a given cpu.
5097 * @cpu: the processor in question.
5099 struct task_struct
*idle_task(int cpu
)
5101 return cpu_rq(cpu
)->idle
;
5105 * find_process_by_pid - find a process with a matching PID value.
5106 * @pid: the pid in question.
5108 static struct task_struct
*find_process_by_pid(pid_t pid
)
5110 return pid
? find_task_by_vpid(pid
) : current
;
5113 /* Actually do priority change: must hold rq lock. */
5115 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
5117 BUG_ON(p
->se
.on_rq
);
5120 switch (p
->policy
) {
5124 p
->sched_class
= &fair_sched_class
;
5128 p
->sched_class
= &rt_sched_class
;
5132 p
->rt_priority
= prio
;
5133 p
->normal_prio
= normal_prio(p
);
5134 /* we are holding p->pi_lock already */
5135 p
->prio
= rt_mutex_getprio(p
);
5140 * check the target process has a UID that matches the current process's
5142 static bool check_same_owner(struct task_struct
*p
)
5144 const struct cred
*cred
= current_cred(), *pcred
;
5148 pcred
= __task_cred(p
);
5149 match
= (cred
->euid
== pcred
->euid
||
5150 cred
->euid
== pcred
->uid
);
5155 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
5156 struct sched_param
*param
, bool user
)
5158 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
5159 unsigned long flags
;
5160 const struct sched_class
*prev_class
= p
->sched_class
;
5163 /* may grab non-irq protected spin_locks */
5164 BUG_ON(in_interrupt());
5166 /* double check policy once rq lock held */
5168 policy
= oldpolicy
= p
->policy
;
5169 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
5170 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
5171 policy
!= SCHED_IDLE
)
5174 * Valid priorities for SCHED_FIFO and SCHED_RR are
5175 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
5176 * SCHED_BATCH and SCHED_IDLE is 0.
5178 if (param
->sched_priority
< 0 ||
5179 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
5180 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
5182 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
5186 * Allow unprivileged RT tasks to decrease priority:
5188 if (user
&& !capable(CAP_SYS_NICE
)) {
5189 if (rt_policy(policy
)) {
5190 unsigned long rlim_rtprio
;
5192 if (!lock_task_sighand(p
, &flags
))
5194 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
5195 unlock_task_sighand(p
, &flags
);
5197 /* can't set/change the rt policy */
5198 if (policy
!= p
->policy
&& !rlim_rtprio
)
5201 /* can't increase priority */
5202 if (param
->sched_priority
> p
->rt_priority
&&
5203 param
->sched_priority
> rlim_rtprio
)
5207 * Like positive nice levels, dont allow tasks to
5208 * move out of SCHED_IDLE either:
5210 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
5213 /* can't change other user's priorities */
5214 if (!check_same_owner(p
))
5219 #ifdef CONFIG_RT_GROUP_SCHED
5221 * Do not allow realtime tasks into groups that have no runtime
5224 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
5225 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
5229 retval
= security_task_setscheduler(p
, policy
, param
);
5235 * make sure no PI-waiters arrive (or leave) while we are
5236 * changing the priority of the task:
5238 spin_lock_irqsave(&p
->pi_lock
, flags
);
5240 * To be able to change p->policy safely, the apropriate
5241 * runqueue lock must be held.
5243 rq
= __task_rq_lock(p
);
5244 /* recheck policy now with rq lock held */
5245 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
5246 policy
= oldpolicy
= -1;
5247 __task_rq_unlock(rq
);
5248 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5251 update_rq_clock(rq
);
5252 on_rq
= p
->se
.on_rq
;
5253 running
= task_current(rq
, p
);
5255 deactivate_task(rq
, p
, 0);
5257 p
->sched_class
->put_prev_task(rq
, p
);
5260 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
5263 p
->sched_class
->set_curr_task(rq
);
5265 activate_task(rq
, p
, 0);
5267 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
5269 __task_rq_unlock(rq
);
5270 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
5272 rt_mutex_adjust_pi(p
);
5278 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
5279 * @p: the task in question.
5280 * @policy: new policy.
5281 * @param: structure containing the new RT priority.
5283 * NOTE that the task may be already dead.
5285 int sched_setscheduler(struct task_struct
*p
, int policy
,
5286 struct sched_param
*param
)
5288 return __sched_setscheduler(p
, policy
, param
, true);
5290 EXPORT_SYMBOL_GPL(sched_setscheduler
);
5293 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
5294 * @p: the task in question.
5295 * @policy: new policy.
5296 * @param: structure containing the new RT priority.
5298 * Just like sched_setscheduler, only don't bother checking if the
5299 * current context has permission. For example, this is needed in
5300 * stop_machine(): we create temporary high priority worker threads,
5301 * but our caller might not have that capability.
5303 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
5304 struct sched_param
*param
)
5306 return __sched_setscheduler(p
, policy
, param
, false);
5310 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5312 struct sched_param lparam
;
5313 struct task_struct
*p
;
5316 if (!param
|| pid
< 0)
5318 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
5323 p
= find_process_by_pid(pid
);
5325 retval
= sched_setscheduler(p
, policy
, &lparam
);
5332 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
5333 * @pid: the pid in question.
5334 * @policy: new policy.
5335 * @param: structure containing the new RT priority.
5338 sys_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
5340 /* negative values for policy are not valid */
5344 return do_sched_setscheduler(pid
, policy
, param
);
5348 * sys_sched_setparam - set/change the RT priority of a thread
5349 * @pid: the pid in question.
5350 * @param: structure containing the new RT priority.
5352 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
5354 return do_sched_setscheduler(pid
, -1, param
);
5358 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
5359 * @pid: the pid in question.
5361 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
5363 struct task_struct
*p
;
5370 read_lock(&tasklist_lock
);
5371 p
= find_process_by_pid(pid
);
5373 retval
= security_task_getscheduler(p
);
5377 read_unlock(&tasklist_lock
);
5382 * sys_sched_getscheduler - get the RT priority of a thread
5383 * @pid: the pid in question.
5384 * @param: structure containing the RT priority.
5386 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
5388 struct sched_param lp
;
5389 struct task_struct
*p
;
5392 if (!param
|| pid
< 0)
5395 read_lock(&tasklist_lock
);
5396 p
= find_process_by_pid(pid
);
5401 retval
= security_task_getscheduler(p
);
5405 lp
.sched_priority
= p
->rt_priority
;
5406 read_unlock(&tasklist_lock
);
5409 * This one might sleep, we cannot do it with a spinlock held ...
5411 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
5416 read_unlock(&tasklist_lock
);
5420 long sched_setaffinity(pid_t pid
, const cpumask_t
*in_mask
)
5422 cpumask_t cpus_allowed
;
5423 cpumask_t new_mask
= *in_mask
;
5424 struct task_struct
*p
;
5428 read_lock(&tasklist_lock
);
5430 p
= find_process_by_pid(pid
);
5432 read_unlock(&tasklist_lock
);
5438 * It is not safe to call set_cpus_allowed with the
5439 * tasklist_lock held. We will bump the task_struct's
5440 * usage count and then drop tasklist_lock.
5443 read_unlock(&tasklist_lock
);
5446 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
5449 retval
= security_task_setscheduler(p
, 0, NULL
);
5453 cpuset_cpus_allowed(p
, &cpus_allowed
);
5454 cpus_and(new_mask
, new_mask
, cpus_allowed
);
5456 retval
= set_cpus_allowed_ptr(p
, &new_mask
);
5459 cpuset_cpus_allowed(p
, &cpus_allowed
);
5460 if (!cpus_subset(new_mask
, cpus_allowed
)) {
5462 * We must have raced with a concurrent cpuset
5463 * update. Just reset the cpus_allowed to the
5464 * cpuset's cpus_allowed
5466 new_mask
= cpus_allowed
;
5476 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
5477 cpumask_t
*new_mask
)
5479 if (len
< sizeof(cpumask_t
)) {
5480 memset(new_mask
, 0, sizeof(cpumask_t
));
5481 } else if (len
> sizeof(cpumask_t
)) {
5482 len
= sizeof(cpumask_t
);
5484 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
5488 * sys_sched_setaffinity - set the cpu affinity of a process
5489 * @pid: pid of the process
5490 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5491 * @user_mask_ptr: user-space pointer to the new cpu mask
5493 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
5494 unsigned long __user
*user_mask_ptr
)
5499 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
5503 return sched_setaffinity(pid
, &new_mask
);
5506 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
5508 struct task_struct
*p
;
5512 read_lock(&tasklist_lock
);
5515 p
= find_process_by_pid(pid
);
5519 retval
= security_task_getscheduler(p
);
5523 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
5526 read_unlock(&tasklist_lock
);
5533 * sys_sched_getaffinity - get the cpu affinity of a process
5534 * @pid: pid of the process
5535 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
5536 * @user_mask_ptr: user-space pointer to hold the current cpu mask
5538 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
5539 unsigned long __user
*user_mask_ptr
)
5544 if (len
< sizeof(cpumask_t
))
5547 ret
= sched_getaffinity(pid
, &mask
);
5551 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
5554 return sizeof(cpumask_t
);
5558 * sys_sched_yield - yield the current processor to other threads.
5560 * This function yields the current CPU to other tasks. If there are no
5561 * other threads running on this CPU then this function will return.
5563 asmlinkage
long sys_sched_yield(void)
5565 struct rq
*rq
= this_rq_lock();
5567 schedstat_inc(rq
, yld_count
);
5568 current
->sched_class
->yield_task(rq
);
5571 * Since we are going to call schedule() anyway, there's
5572 * no need to preempt or enable interrupts:
5574 __release(rq
->lock
);
5575 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
5576 _raw_spin_unlock(&rq
->lock
);
5577 preempt_enable_no_resched();
5584 static void __cond_resched(void)
5586 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
5587 __might_sleep(__FILE__
, __LINE__
);
5590 * The BKS might be reacquired before we have dropped
5591 * PREEMPT_ACTIVE, which could trigger a second
5592 * cond_resched() call.
5595 add_preempt_count(PREEMPT_ACTIVE
);
5597 sub_preempt_count(PREEMPT_ACTIVE
);
5598 } while (need_resched());
5601 int __sched
_cond_resched(void)
5603 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
5604 system_state
== SYSTEM_RUNNING
) {
5610 EXPORT_SYMBOL(_cond_resched
);
5613 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
5614 * call schedule, and on return reacquire the lock.
5616 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
5617 * operations here to prevent schedule() from being called twice (once via
5618 * spin_unlock(), once by hand).
5620 int cond_resched_lock(spinlock_t
*lock
)
5622 int resched
= need_resched() && system_state
== SYSTEM_RUNNING
;
5625 if (spin_needbreak(lock
) || resched
) {
5627 if (resched
&& need_resched())
5636 EXPORT_SYMBOL(cond_resched_lock
);
5638 int __sched
cond_resched_softirq(void)
5640 BUG_ON(!in_softirq());
5642 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
5650 EXPORT_SYMBOL(cond_resched_softirq
);
5653 * yield - yield the current processor to other threads.
5655 * This is a shortcut for kernel-space yielding - it marks the
5656 * thread runnable and calls sys_sched_yield().
5658 void __sched
yield(void)
5660 set_current_state(TASK_RUNNING
);
5663 EXPORT_SYMBOL(yield
);
5666 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
5667 * that process accounting knows that this is a task in IO wait state.
5669 * But don't do that if it is a deliberate, throttling IO wait (this task
5670 * has set its backing_dev_info: the queue against which it should throttle)
5672 void __sched
io_schedule(void)
5674 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5676 delayacct_blkio_start();
5677 atomic_inc(&rq
->nr_iowait
);
5679 atomic_dec(&rq
->nr_iowait
);
5680 delayacct_blkio_end();
5682 EXPORT_SYMBOL(io_schedule
);
5684 long __sched
io_schedule_timeout(long timeout
)
5686 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
5689 delayacct_blkio_start();
5690 atomic_inc(&rq
->nr_iowait
);
5691 ret
= schedule_timeout(timeout
);
5692 atomic_dec(&rq
->nr_iowait
);
5693 delayacct_blkio_end();
5698 * sys_sched_get_priority_max - return maximum RT priority.
5699 * @policy: scheduling class.
5701 * this syscall returns the maximum rt_priority that can be used
5702 * by a given scheduling class.
5704 asmlinkage
long sys_sched_get_priority_max(int policy
)
5711 ret
= MAX_USER_RT_PRIO
-1;
5723 * sys_sched_get_priority_min - return minimum RT priority.
5724 * @policy: scheduling class.
5726 * this syscall returns the minimum rt_priority that can be used
5727 * by a given scheduling class.
5729 asmlinkage
long sys_sched_get_priority_min(int policy
)
5747 * sys_sched_rr_get_interval - return the default timeslice of a process.
5748 * @pid: pid of the process.
5749 * @interval: userspace pointer to the timeslice value.
5751 * this syscall writes the default timeslice value of a given process
5752 * into the user-space timespec buffer. A value of '0' means infinity.
5755 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
5757 struct task_struct
*p
;
5758 unsigned int time_slice
;
5766 read_lock(&tasklist_lock
);
5767 p
= find_process_by_pid(pid
);
5771 retval
= security_task_getscheduler(p
);
5776 * Time slice is 0 for SCHED_FIFO tasks and for SCHED_OTHER
5777 * tasks that are on an otherwise idle runqueue:
5780 if (p
->policy
== SCHED_RR
) {
5781 time_slice
= DEF_TIMESLICE
;
5782 } else if (p
->policy
!= SCHED_FIFO
) {
5783 struct sched_entity
*se
= &p
->se
;
5784 unsigned long flags
;
5787 rq
= task_rq_lock(p
, &flags
);
5788 if (rq
->cfs
.load
.weight
)
5789 time_slice
= NS_TO_JIFFIES(sched_slice(&rq
->cfs
, se
));
5790 task_rq_unlock(rq
, &flags
);
5792 read_unlock(&tasklist_lock
);
5793 jiffies_to_timespec(time_slice
, &t
);
5794 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
5798 read_unlock(&tasklist_lock
);
5802 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
5804 void sched_show_task(struct task_struct
*p
)
5806 unsigned long free
= 0;
5809 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
5810 printk(KERN_INFO
"%-13.13s %c", p
->comm
,
5811 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
5812 #if BITS_PER_LONG == 32
5813 if (state
== TASK_RUNNING
)
5814 printk(KERN_CONT
" running ");
5816 printk(KERN_CONT
" %08lx ", thread_saved_pc(p
));
5818 if (state
== TASK_RUNNING
)
5819 printk(KERN_CONT
" running task ");
5821 printk(KERN_CONT
" %016lx ", thread_saved_pc(p
));
5823 #ifdef CONFIG_DEBUG_STACK_USAGE
5825 unsigned long *n
= end_of_stack(p
);
5828 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
5831 printk(KERN_CONT
"%5lu %5d %6d\n", free
,
5832 task_pid_nr(p
), task_pid_nr(p
->real_parent
));
5834 show_stack(p
, NULL
);
5837 void show_state_filter(unsigned long state_filter
)
5839 struct task_struct
*g
, *p
;
5841 #if BITS_PER_LONG == 32
5843 " task PC stack pid father\n");
5846 " task PC stack pid father\n");
5848 read_lock(&tasklist_lock
);
5849 do_each_thread(g
, p
) {
5851 * reset the NMI-timeout, listing all files on a slow
5852 * console might take alot of time:
5854 touch_nmi_watchdog();
5855 if (!state_filter
|| (p
->state
& state_filter
))
5857 } while_each_thread(g
, p
);
5859 touch_all_softlockup_watchdogs();
5861 #ifdef CONFIG_SCHED_DEBUG
5862 sysrq_sched_debug_show();
5864 read_unlock(&tasklist_lock
);
5866 * Only show locks if all tasks are dumped:
5868 if (state_filter
== -1)
5869 debug_show_all_locks();
5872 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
5874 idle
->sched_class
= &idle_sched_class
;
5878 * init_idle - set up an idle thread for a given CPU
5879 * @idle: task in question
5880 * @cpu: cpu the idle task belongs to
5882 * NOTE: this function does not set the idle thread's NEED_RESCHED
5883 * flag, to make booting more robust.
5885 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
5887 struct rq
*rq
= cpu_rq(cpu
);
5888 unsigned long flags
;
5890 spin_lock_irqsave(&rq
->lock
, flags
);
5893 idle
->se
.exec_start
= sched_clock();
5895 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
5896 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
5897 __set_task_cpu(idle
, cpu
);
5899 rq
->curr
= rq
->idle
= idle
;
5900 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
5903 spin_unlock_irqrestore(&rq
->lock
, flags
);
5905 /* Set the preempt count _outside_ the spinlocks! */
5906 #if defined(CONFIG_PREEMPT)
5907 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
5909 task_thread_info(idle
)->preempt_count
= 0;
5912 * The idle tasks have their own, simple scheduling class:
5914 idle
->sched_class
= &idle_sched_class
;
5918 * In a system that switches off the HZ timer nohz_cpu_mask
5919 * indicates which cpus entered this state. This is used
5920 * in the rcu update to wait only for active cpus. For system
5921 * which do not switch off the HZ timer nohz_cpu_mask should
5922 * always be CPU_MASK_NONE.
5924 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
5927 * Increase the granularity value when there are more CPUs,
5928 * because with more CPUs the 'effective latency' as visible
5929 * to users decreases. But the relationship is not linear,
5930 * so pick a second-best guess by going with the log2 of the
5933 * This idea comes from the SD scheduler of Con Kolivas:
5935 static inline void sched_init_granularity(void)
5937 unsigned int factor
= 1 + ilog2(num_online_cpus());
5938 const unsigned long limit
= 200000000;
5940 sysctl_sched_min_granularity
*= factor
;
5941 if (sysctl_sched_min_granularity
> limit
)
5942 sysctl_sched_min_granularity
= limit
;
5944 sysctl_sched_latency
*= factor
;
5945 if (sysctl_sched_latency
> limit
)
5946 sysctl_sched_latency
= limit
;
5948 sysctl_sched_wakeup_granularity
*= factor
;
5950 sysctl_sched_shares_ratelimit
*= factor
;
5955 * This is how migration works:
5957 * 1) we queue a struct migration_req structure in the source CPU's
5958 * runqueue and wake up that CPU's migration thread.
5959 * 2) we down() the locked semaphore => thread blocks.
5960 * 3) migration thread wakes up (implicitly it forces the migrated
5961 * thread off the CPU)
5962 * 4) it gets the migration request and checks whether the migrated
5963 * task is still in the wrong runqueue.
5964 * 5) if it's in the wrong runqueue then the migration thread removes
5965 * it and puts it into the right queue.
5966 * 6) migration thread up()s the semaphore.
5967 * 7) we wake up and the migration is done.
5971 * Change a given task's CPU affinity. Migrate the thread to a
5972 * proper CPU and schedule it away if the CPU it's executing on
5973 * is removed from the allowed bitmask.
5975 * NOTE: the caller must have a valid reference to the task, the
5976 * task must not exit() & deallocate itself prematurely. The
5977 * call is not atomic; no spinlocks may be held.
5979 int set_cpus_allowed_ptr(struct task_struct
*p
, const cpumask_t
*new_mask
)
5981 struct migration_req req
;
5982 unsigned long flags
;
5986 rq
= task_rq_lock(p
, &flags
);
5987 if (!cpus_intersects(*new_mask
, cpu_online_map
)) {
5992 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
5993 !cpus_equal(p
->cpus_allowed
, *new_mask
))) {
5998 if (p
->sched_class
->set_cpus_allowed
)
5999 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
6001 p
->cpus_allowed
= *new_mask
;
6002 p
->rt
.nr_cpus_allowed
= cpus_weight(*new_mask
);
6005 /* Can the task run on the task's current CPU? If so, we're done */
6006 if (cpu_isset(task_cpu(p
), *new_mask
))
6009 if (migrate_task(p
, any_online_cpu(*new_mask
), &req
)) {
6010 /* Need help from migration thread: drop lock and wait. */
6011 task_rq_unlock(rq
, &flags
);
6012 wake_up_process(rq
->migration_thread
);
6013 wait_for_completion(&req
.done
);
6014 tlb_migrate_finish(p
->mm
);
6018 task_rq_unlock(rq
, &flags
);
6022 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
6025 * Move (not current) task off this cpu, onto dest cpu. We're doing
6026 * this because either it can't run here any more (set_cpus_allowed()
6027 * away from this CPU, or CPU going down), or because we're
6028 * attempting to rebalance this task on exec (sched_exec).
6030 * So we race with normal scheduler movements, but that's OK, as long
6031 * as the task is no longer on this CPU.
6033 * Returns non-zero if task was successfully migrated.
6035 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6037 struct rq
*rq_dest
, *rq_src
;
6040 if (unlikely(!cpu_active(dest_cpu
)))
6043 rq_src
= cpu_rq(src_cpu
);
6044 rq_dest
= cpu_rq(dest_cpu
);
6046 double_rq_lock(rq_src
, rq_dest
);
6047 /* Already moved. */
6048 if (task_cpu(p
) != src_cpu
)
6050 /* Affinity changed (again). */
6051 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
6054 on_rq
= p
->se
.on_rq
;
6056 deactivate_task(rq_src
, p
, 0);
6058 set_task_cpu(p
, dest_cpu
);
6060 activate_task(rq_dest
, p
, 0);
6061 check_preempt_curr(rq_dest
, p
, 0);
6066 double_rq_unlock(rq_src
, rq_dest
);
6071 * migration_thread - this is a highprio system thread that performs
6072 * thread migration by bumping thread off CPU then 'pushing' onto
6075 static int migration_thread(void *data
)
6077 int cpu
= (long)data
;
6081 BUG_ON(rq
->migration_thread
!= current
);
6083 set_current_state(TASK_INTERRUPTIBLE
);
6084 while (!kthread_should_stop()) {
6085 struct migration_req
*req
;
6086 struct list_head
*head
;
6088 spin_lock_irq(&rq
->lock
);
6090 if (cpu_is_offline(cpu
)) {
6091 spin_unlock_irq(&rq
->lock
);
6095 if (rq
->active_balance
) {
6096 active_load_balance(rq
, cpu
);
6097 rq
->active_balance
= 0;
6100 head
= &rq
->migration_queue
;
6102 if (list_empty(head
)) {
6103 spin_unlock_irq(&rq
->lock
);
6105 set_current_state(TASK_INTERRUPTIBLE
);
6108 req
= list_entry(head
->next
, struct migration_req
, list
);
6109 list_del_init(head
->next
);
6111 spin_unlock(&rq
->lock
);
6112 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
6115 complete(&req
->done
);
6117 __set_current_state(TASK_RUNNING
);
6121 /* Wait for kthread_stop */
6122 set_current_state(TASK_INTERRUPTIBLE
);
6123 while (!kthread_should_stop()) {
6125 set_current_state(TASK_INTERRUPTIBLE
);
6127 __set_current_state(TASK_RUNNING
);
6131 #ifdef CONFIG_HOTPLUG_CPU
6133 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
6137 local_irq_disable();
6138 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
6144 * Figure out where task on dead CPU should go, use force if necessary.
6145 * NOTE: interrupts should be disabled by the caller
6147 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
6149 unsigned long flags
;
6156 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
6157 cpus_and(mask
, mask
, p
->cpus_allowed
);
6158 dest_cpu
= any_online_cpu(mask
);
6160 /* On any allowed CPU? */
6161 if (dest_cpu
>= nr_cpu_ids
)
6162 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6164 /* No more Mr. Nice Guy. */
6165 if (dest_cpu
>= nr_cpu_ids
) {
6166 cpumask_t cpus_allowed
;
6168 cpuset_cpus_allowed_locked(p
, &cpus_allowed
);
6170 * Try to stay on the same cpuset, where the
6171 * current cpuset may be a subset of all cpus.
6172 * The cpuset_cpus_allowed_locked() variant of
6173 * cpuset_cpus_allowed() will not block. It must be
6174 * called within calls to cpuset_lock/cpuset_unlock.
6176 rq
= task_rq_lock(p
, &flags
);
6177 p
->cpus_allowed
= cpus_allowed
;
6178 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
6179 task_rq_unlock(rq
, &flags
);
6182 * Don't tell them about moving exiting tasks or
6183 * kernel threads (both mm NULL), since they never
6186 if (p
->mm
&& printk_ratelimit()) {
6187 printk(KERN_INFO
"process %d (%s) no "
6188 "longer affine to cpu%d\n",
6189 task_pid_nr(p
), p
->comm
, dead_cpu
);
6192 } while (!__migrate_task_irq(p
, dead_cpu
, dest_cpu
));
6196 * While a dead CPU has no uninterruptible tasks queued at this point,
6197 * it might still have a nonzero ->nr_uninterruptible counter, because
6198 * for performance reasons the counter is not stricly tracking tasks to
6199 * their home CPUs. So we just add the counter to another CPU's counter,
6200 * to keep the global sum constant after CPU-down:
6202 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
6204 struct rq
*rq_dest
= cpu_rq(any_online_cpu(*CPU_MASK_ALL_PTR
));
6205 unsigned long flags
;
6207 local_irq_save(flags
);
6208 double_rq_lock(rq_src
, rq_dest
);
6209 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
6210 rq_src
->nr_uninterruptible
= 0;
6211 double_rq_unlock(rq_src
, rq_dest
);
6212 local_irq_restore(flags
);
6215 /* Run through task list and migrate tasks from the dead cpu. */
6216 static void migrate_live_tasks(int src_cpu
)
6218 struct task_struct
*p
, *t
;
6220 read_lock(&tasklist_lock
);
6222 do_each_thread(t
, p
) {
6226 if (task_cpu(p
) == src_cpu
)
6227 move_task_off_dead_cpu(src_cpu
, p
);
6228 } while_each_thread(t
, p
);
6230 read_unlock(&tasklist_lock
);
6234 * Schedules idle task to be the next runnable task on current CPU.
6235 * It does so by boosting its priority to highest possible.
6236 * Used by CPU offline code.
6238 void sched_idle_next(void)
6240 int this_cpu
= smp_processor_id();
6241 struct rq
*rq
= cpu_rq(this_cpu
);
6242 struct task_struct
*p
= rq
->idle
;
6243 unsigned long flags
;
6245 /* cpu has to be offline */
6246 BUG_ON(cpu_online(this_cpu
));
6249 * Strictly not necessary since rest of the CPUs are stopped by now
6250 * and interrupts disabled on the current cpu.
6252 spin_lock_irqsave(&rq
->lock
, flags
);
6254 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6256 update_rq_clock(rq
);
6257 activate_task(rq
, p
, 0);
6259 spin_unlock_irqrestore(&rq
->lock
, flags
);
6263 * Ensures that the idle task is using init_mm right before its cpu goes
6266 void idle_task_exit(void)
6268 struct mm_struct
*mm
= current
->active_mm
;
6270 BUG_ON(cpu_online(smp_processor_id()));
6273 switch_mm(mm
, &init_mm
, current
);
6277 /* called under rq->lock with disabled interrupts */
6278 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
6280 struct rq
*rq
= cpu_rq(dead_cpu
);
6282 /* Must be exiting, otherwise would be on tasklist. */
6283 BUG_ON(!p
->exit_state
);
6285 /* Cannot have done final schedule yet: would have vanished. */
6286 BUG_ON(p
->state
== TASK_DEAD
);
6291 * Drop lock around migration; if someone else moves it,
6292 * that's OK. No task can be added to this CPU, so iteration is
6295 spin_unlock_irq(&rq
->lock
);
6296 move_task_off_dead_cpu(dead_cpu
, p
);
6297 spin_lock_irq(&rq
->lock
);
6302 /* release_task() removes task from tasklist, so we won't find dead tasks. */
6303 static void migrate_dead_tasks(unsigned int dead_cpu
)
6305 struct rq
*rq
= cpu_rq(dead_cpu
);
6306 struct task_struct
*next
;
6309 if (!rq
->nr_running
)
6311 update_rq_clock(rq
);
6312 next
= pick_next_task(rq
, rq
->curr
);
6315 next
->sched_class
->put_prev_task(rq
, next
);
6316 migrate_dead(dead_cpu
, next
);
6320 #endif /* CONFIG_HOTPLUG_CPU */
6322 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
6324 static struct ctl_table sd_ctl_dir
[] = {
6326 .procname
= "sched_domain",
6332 static struct ctl_table sd_ctl_root
[] = {
6334 .ctl_name
= CTL_KERN
,
6335 .procname
= "kernel",
6337 .child
= sd_ctl_dir
,
6342 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
6344 struct ctl_table
*entry
=
6345 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
6350 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
6352 struct ctl_table
*entry
;
6355 * In the intermediate directories, both the child directory and
6356 * procname are dynamically allocated and could fail but the mode
6357 * will always be set. In the lowest directory the names are
6358 * static strings and all have proc handlers.
6360 for (entry
= *tablep
; entry
->mode
; entry
++) {
6362 sd_free_ctl_entry(&entry
->child
);
6363 if (entry
->proc_handler
== NULL
)
6364 kfree(entry
->procname
);
6372 set_table_entry(struct ctl_table
*entry
,
6373 const char *procname
, void *data
, int maxlen
,
6374 mode_t mode
, proc_handler
*proc_handler
)
6376 entry
->procname
= procname
;
6378 entry
->maxlen
= maxlen
;
6380 entry
->proc_handler
= proc_handler
;
6383 static struct ctl_table
*
6384 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
6386 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
6391 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
6392 sizeof(long), 0644, proc_doulongvec_minmax
);
6393 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
6394 sizeof(long), 0644, proc_doulongvec_minmax
);
6395 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
6396 sizeof(int), 0644, proc_dointvec_minmax
);
6397 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
6398 sizeof(int), 0644, proc_dointvec_minmax
);
6399 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
6400 sizeof(int), 0644, proc_dointvec_minmax
);
6401 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
6402 sizeof(int), 0644, proc_dointvec_minmax
);
6403 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
6404 sizeof(int), 0644, proc_dointvec_minmax
);
6405 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
6406 sizeof(int), 0644, proc_dointvec_minmax
);
6407 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
6408 sizeof(int), 0644, proc_dointvec_minmax
);
6409 set_table_entry(&table
[9], "cache_nice_tries",
6410 &sd
->cache_nice_tries
,
6411 sizeof(int), 0644, proc_dointvec_minmax
);
6412 set_table_entry(&table
[10], "flags", &sd
->flags
,
6413 sizeof(int), 0644, proc_dointvec_minmax
);
6414 set_table_entry(&table
[11], "name", sd
->name
,
6415 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
6416 /* &table[12] is terminator */
6421 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
6423 struct ctl_table
*entry
, *table
;
6424 struct sched_domain
*sd
;
6425 int domain_num
= 0, i
;
6428 for_each_domain(cpu
, sd
)
6430 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
6435 for_each_domain(cpu
, sd
) {
6436 snprintf(buf
, 32, "domain%d", i
);
6437 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6439 entry
->child
= sd_alloc_ctl_domain_table(sd
);
6446 static struct ctl_table_header
*sd_sysctl_header
;
6447 static void register_sched_domain_sysctl(void)
6449 int i
, cpu_num
= num_online_cpus();
6450 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
6453 WARN_ON(sd_ctl_dir
[0].child
);
6454 sd_ctl_dir
[0].child
= entry
;
6459 for_each_online_cpu(i
) {
6460 snprintf(buf
, 32, "cpu%d", i
);
6461 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
6463 entry
->child
= sd_alloc_ctl_cpu_table(i
);
6467 WARN_ON(sd_sysctl_header
);
6468 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
6471 /* may be called multiple times per register */
6472 static void unregister_sched_domain_sysctl(void)
6474 if (sd_sysctl_header
)
6475 unregister_sysctl_table(sd_sysctl_header
);
6476 sd_sysctl_header
= NULL
;
6477 if (sd_ctl_dir
[0].child
)
6478 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
6481 static void register_sched_domain_sysctl(void)
6484 static void unregister_sched_domain_sysctl(void)
6489 static void set_rq_online(struct rq
*rq
)
6492 const struct sched_class
*class;
6494 cpu_set(rq
->cpu
, rq
->rd
->online
);
6497 for_each_class(class) {
6498 if (class->rq_online
)
6499 class->rq_online(rq
);
6504 static void set_rq_offline(struct rq
*rq
)
6507 const struct sched_class
*class;
6509 for_each_class(class) {
6510 if (class->rq_offline
)
6511 class->rq_offline(rq
);
6514 cpu_clear(rq
->cpu
, rq
->rd
->online
);
6520 * migration_call - callback that gets triggered when a CPU is added.
6521 * Here we can start up the necessary migration thread for the new CPU.
6523 static int __cpuinit
6524 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
6526 struct task_struct
*p
;
6527 int cpu
= (long)hcpu
;
6528 unsigned long flags
;
6533 case CPU_UP_PREPARE
:
6534 case CPU_UP_PREPARE_FROZEN
:
6535 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
6538 kthread_bind(p
, cpu
);
6539 /* Must be high prio: stop_machine expects to yield to it. */
6540 rq
= task_rq_lock(p
, &flags
);
6541 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
6542 task_rq_unlock(rq
, &flags
);
6543 cpu_rq(cpu
)->migration_thread
= p
;
6547 case CPU_ONLINE_FROZEN
:
6548 /* Strictly unnecessary, as first user will wake it. */
6549 wake_up_process(cpu_rq(cpu
)->migration_thread
);
6551 /* Update our root-domain */
6553 spin_lock_irqsave(&rq
->lock
, flags
);
6555 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6559 spin_unlock_irqrestore(&rq
->lock
, flags
);
6562 #ifdef CONFIG_HOTPLUG_CPU
6563 case CPU_UP_CANCELED
:
6564 case CPU_UP_CANCELED_FROZEN
:
6565 if (!cpu_rq(cpu
)->migration_thread
)
6567 /* Unbind it from offline cpu so it can run. Fall thru. */
6568 kthread_bind(cpu_rq(cpu
)->migration_thread
,
6569 any_online_cpu(cpu_online_map
));
6570 kthread_stop(cpu_rq(cpu
)->migration_thread
);
6571 cpu_rq(cpu
)->migration_thread
= NULL
;
6575 case CPU_DEAD_FROZEN
:
6576 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
6577 migrate_live_tasks(cpu
);
6579 kthread_stop(rq
->migration_thread
);
6580 rq
->migration_thread
= NULL
;
6581 /* Idle task back to normal (off runqueue, low prio) */
6582 spin_lock_irq(&rq
->lock
);
6583 update_rq_clock(rq
);
6584 deactivate_task(rq
, rq
->idle
, 0);
6585 rq
->idle
->static_prio
= MAX_PRIO
;
6586 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
6587 rq
->idle
->sched_class
= &idle_sched_class
;
6588 migrate_dead_tasks(cpu
);
6589 spin_unlock_irq(&rq
->lock
);
6591 migrate_nr_uninterruptible(rq
);
6592 BUG_ON(rq
->nr_running
!= 0);
6595 * No need to migrate the tasks: it was best-effort if
6596 * they didn't take sched_hotcpu_mutex. Just wake up
6599 spin_lock_irq(&rq
->lock
);
6600 while (!list_empty(&rq
->migration_queue
)) {
6601 struct migration_req
*req
;
6603 req
= list_entry(rq
->migration_queue
.next
,
6604 struct migration_req
, list
);
6605 list_del_init(&req
->list
);
6606 spin_unlock_irq(&rq
->lock
);
6607 complete(&req
->done
);
6608 spin_lock_irq(&rq
->lock
);
6610 spin_unlock_irq(&rq
->lock
);
6614 case CPU_DYING_FROZEN
:
6615 /* Update our root-domain */
6617 spin_lock_irqsave(&rq
->lock
, flags
);
6619 BUG_ON(!cpu_isset(cpu
, rq
->rd
->span
));
6622 spin_unlock_irqrestore(&rq
->lock
, flags
);
6629 /* Register at highest priority so that task migration (migrate_all_tasks)
6630 * happens before everything else.
6632 static struct notifier_block __cpuinitdata migration_notifier
= {
6633 .notifier_call
= migration_call
,
6637 static int __init
migration_init(void)
6639 void *cpu
= (void *)(long)smp_processor_id();
6642 /* Start one for the boot CPU: */
6643 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
6644 BUG_ON(err
== NOTIFY_BAD
);
6645 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
6646 register_cpu_notifier(&migration_notifier
);
6650 early_initcall(migration_init
);
6655 #ifdef CONFIG_SCHED_DEBUG
6657 static inline const char *sd_level_to_string(enum sched_domain_level lvl
)
6670 case SD_LV_ALLNODES
:
6679 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
6680 cpumask_t
*groupmask
)
6682 struct sched_group
*group
= sd
->groups
;
6685 cpulist_scnprintf(str
, sizeof(str
), sd
->span
);
6686 cpus_clear(*groupmask
);
6688 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
6690 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
6691 printk("does not load-balance\n");
6693 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
6698 printk(KERN_CONT
"span %s level %s\n",
6699 str
, sd_level_to_string(sd
->level
));
6701 if (!cpu_isset(cpu
, sd
->span
)) {
6702 printk(KERN_ERR
"ERROR: domain->span does not contain "
6705 if (!cpu_isset(cpu
, group
->cpumask
)) {
6706 printk(KERN_ERR
"ERROR: domain->groups does not contain"
6710 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
6714 printk(KERN_ERR
"ERROR: group is NULL\n");
6718 if (!group
->__cpu_power
) {
6719 printk(KERN_CONT
"\n");
6720 printk(KERN_ERR
"ERROR: domain->cpu_power not "
6725 if (!cpus_weight(group
->cpumask
)) {
6726 printk(KERN_CONT
"\n");
6727 printk(KERN_ERR
"ERROR: empty group\n");
6731 if (cpus_intersects(*groupmask
, group
->cpumask
)) {
6732 printk(KERN_CONT
"\n");
6733 printk(KERN_ERR
"ERROR: repeated CPUs\n");
6737 cpus_or(*groupmask
, *groupmask
, group
->cpumask
);
6739 cpulist_scnprintf(str
, sizeof(str
), group
->cpumask
);
6740 printk(KERN_CONT
" %s", str
);
6742 group
= group
->next
;
6743 } while (group
!= sd
->groups
);
6744 printk(KERN_CONT
"\n");
6746 if (!cpus_equal(sd
->span
, *groupmask
))
6747 printk(KERN_ERR
"ERROR: groups don't span domain->span\n");
6749 if (sd
->parent
&& !cpus_subset(*groupmask
, sd
->parent
->span
))
6750 printk(KERN_ERR
"ERROR: parent span is not a superset "
6751 "of domain->span\n");
6755 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
6757 cpumask_t
*groupmask
;
6761 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
6765 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
6767 groupmask
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
6769 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
6774 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
6783 #else /* !CONFIG_SCHED_DEBUG */
6784 # define sched_domain_debug(sd, cpu) do { } while (0)
6785 #endif /* CONFIG_SCHED_DEBUG */
6787 static int sd_degenerate(struct sched_domain
*sd
)
6789 if (cpus_weight(sd
->span
) == 1)
6792 /* Following flags need at least 2 groups */
6793 if (sd
->flags
& (SD_LOAD_BALANCE
|
6794 SD_BALANCE_NEWIDLE
|
6798 SD_SHARE_PKG_RESOURCES
)) {
6799 if (sd
->groups
!= sd
->groups
->next
)
6803 /* Following flags don't use groups */
6804 if (sd
->flags
& (SD_WAKE_IDLE
|
6813 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
6815 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
6817 if (sd_degenerate(parent
))
6820 if (!cpus_equal(sd
->span
, parent
->span
))
6823 /* Does parent contain flags not in child? */
6824 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
6825 if (cflags
& SD_WAKE_AFFINE
)
6826 pflags
&= ~SD_WAKE_BALANCE
;
6827 /* Flags needing groups don't count if only 1 group in parent */
6828 if (parent
->groups
== parent
->groups
->next
) {
6829 pflags
&= ~(SD_LOAD_BALANCE
|
6830 SD_BALANCE_NEWIDLE
|
6834 SD_SHARE_PKG_RESOURCES
);
6836 if (~cflags
& pflags
)
6842 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
6844 unsigned long flags
;
6846 spin_lock_irqsave(&rq
->lock
, flags
);
6849 struct root_domain
*old_rd
= rq
->rd
;
6851 if (cpu_isset(rq
->cpu
, old_rd
->online
))
6854 cpu_clear(rq
->cpu
, old_rd
->span
);
6856 if (atomic_dec_and_test(&old_rd
->refcount
))
6860 atomic_inc(&rd
->refcount
);
6863 cpu_set(rq
->cpu
, rd
->span
);
6864 if (cpu_isset(rq
->cpu
, cpu_online_map
))
6867 spin_unlock_irqrestore(&rq
->lock
, flags
);
6870 static void init_rootdomain(struct root_domain
*rd
)
6872 memset(rd
, 0, sizeof(*rd
));
6874 cpus_clear(rd
->span
);
6875 cpus_clear(rd
->online
);
6877 cpupri_init(&rd
->cpupri
);
6880 static void init_defrootdomain(void)
6882 init_rootdomain(&def_root_domain
);
6883 atomic_set(&def_root_domain
.refcount
, 1);
6886 static struct root_domain
*alloc_rootdomain(void)
6888 struct root_domain
*rd
;
6890 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
6894 init_rootdomain(rd
);
6900 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
6901 * hold the hotplug lock.
6904 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
6906 struct rq
*rq
= cpu_rq(cpu
);
6907 struct sched_domain
*tmp
;
6909 /* Remove the sched domains which do not contribute to scheduling. */
6910 for (tmp
= sd
; tmp
; ) {
6911 struct sched_domain
*parent
= tmp
->parent
;
6915 if (sd_parent_degenerate(tmp
, parent
)) {
6916 tmp
->parent
= parent
->parent
;
6918 parent
->parent
->child
= tmp
;
6923 if (sd
&& sd_degenerate(sd
)) {
6929 sched_domain_debug(sd
, cpu
);
6931 rq_attach_root(rq
, rd
);
6932 rcu_assign_pointer(rq
->sd
, sd
);
6935 /* cpus with isolated domains */
6936 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
6938 /* Setup the mask of cpus configured for isolated domains */
6939 static int __init
isolated_cpu_setup(char *str
)
6941 static int __initdata ints
[NR_CPUS
];
6944 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
6945 cpus_clear(cpu_isolated_map
);
6946 for (i
= 1; i
<= ints
[0]; i
++)
6947 if (ints
[i
] < NR_CPUS
)
6948 cpu_set(ints
[i
], cpu_isolated_map
);
6952 __setup("isolcpus=", isolated_cpu_setup
);
6955 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
6956 * to a function which identifies what group(along with sched group) a CPU
6957 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
6958 * (due to the fact that we keep track of groups covered with a cpumask_t).
6960 * init_sched_build_groups will build a circular linked list of the groups
6961 * covered by the given span, and will set each group's ->cpumask correctly,
6962 * and ->cpu_power to 0.
6965 init_sched_build_groups(const cpumask_t
*span
, const cpumask_t
*cpu_map
,
6966 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
6967 struct sched_group
**sg
,
6968 cpumask_t
*tmpmask
),
6969 cpumask_t
*covered
, cpumask_t
*tmpmask
)
6971 struct sched_group
*first
= NULL
, *last
= NULL
;
6974 cpus_clear(*covered
);
6976 for_each_cpu_mask_nr(i
, *span
) {
6977 struct sched_group
*sg
;
6978 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
6981 if (cpu_isset(i
, *covered
))
6984 cpus_clear(sg
->cpumask
);
6985 sg
->__cpu_power
= 0;
6987 for_each_cpu_mask_nr(j
, *span
) {
6988 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
6991 cpu_set(j
, *covered
);
6992 cpu_set(j
, sg
->cpumask
);
7003 #define SD_NODES_PER_DOMAIN 16
7008 * find_next_best_node - find the next node to include in a sched_domain
7009 * @node: node whose sched_domain we're building
7010 * @used_nodes: nodes already in the sched_domain
7012 * Find the next node to include in a given scheduling domain. Simply
7013 * finds the closest node not already in the @used_nodes map.
7015 * Should use nodemask_t.
7017 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
7019 int i
, n
, val
, min_val
, best_node
= 0;
7023 for (i
= 0; i
< nr_node_ids
; i
++) {
7024 /* Start at @node */
7025 n
= (node
+ i
) % nr_node_ids
;
7027 if (!nr_cpus_node(n
))
7030 /* Skip already used nodes */
7031 if (node_isset(n
, *used_nodes
))
7034 /* Simple min distance search */
7035 val
= node_distance(node
, n
);
7037 if (val
< min_val
) {
7043 node_set(best_node
, *used_nodes
);
7048 * sched_domain_node_span - get a cpumask for a node's sched_domain
7049 * @node: node whose cpumask we're constructing
7050 * @span: resulting cpumask
7052 * Given a node, construct a good cpumask for its sched_domain to span. It
7053 * should be one that prevents unnecessary balancing, but also spreads tasks
7056 static void sched_domain_node_span(int node
, cpumask_t
*span
)
7058 nodemask_t used_nodes
;
7059 node_to_cpumask_ptr(nodemask
, node
);
7063 nodes_clear(used_nodes
);
7065 cpus_or(*span
, *span
, *nodemask
);
7066 node_set(node
, used_nodes
);
7068 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
7069 int next_node
= find_next_best_node(node
, &used_nodes
);
7071 node_to_cpumask_ptr_next(nodemask
, next_node
);
7072 cpus_or(*span
, *span
, *nodemask
);
7075 #endif /* CONFIG_NUMA */
7077 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
7080 * SMT sched-domains:
7082 #ifdef CONFIG_SCHED_SMT
7083 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
7084 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
7087 cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7091 *sg
= &per_cpu(sched_group_cpus
, cpu
);
7094 #endif /* CONFIG_SCHED_SMT */
7097 * multi-core sched-domains:
7099 #ifdef CONFIG_SCHED_MC
7100 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
7101 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
7102 #endif /* CONFIG_SCHED_MC */
7104 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
7106 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7111 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7112 cpus_and(*mask
, *mask
, *cpu_map
);
7113 group
= first_cpu(*mask
);
7115 *sg
= &per_cpu(sched_group_core
, group
);
7118 #elif defined(CONFIG_SCHED_MC)
7120 cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7124 *sg
= &per_cpu(sched_group_core
, cpu
);
7129 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
7130 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
7133 cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
, struct sched_group
**sg
,
7137 #ifdef CONFIG_SCHED_MC
7138 *mask
= cpu_coregroup_map(cpu
);
7139 cpus_and(*mask
, *mask
, *cpu_map
);
7140 group
= first_cpu(*mask
);
7141 #elif defined(CONFIG_SCHED_SMT)
7142 *mask
= per_cpu(cpu_sibling_map
, cpu
);
7143 cpus_and(*mask
, *mask
, *cpu_map
);
7144 group
= first_cpu(*mask
);
7149 *sg
= &per_cpu(sched_group_phys
, group
);
7155 * The init_sched_build_groups can't handle what we want to do with node
7156 * groups, so roll our own. Now each node has its own list of groups which
7157 * gets dynamically allocated.
7159 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
7160 static struct sched_group
***sched_group_nodes_bycpu
;
7162 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
7163 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
7165 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
7166 struct sched_group
**sg
, cpumask_t
*nodemask
)
7170 *nodemask
= node_to_cpumask(cpu_to_node(cpu
));
7171 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7172 group
= first_cpu(*nodemask
);
7175 *sg
= &per_cpu(sched_group_allnodes
, group
);
7179 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
7181 struct sched_group
*sg
= group_head
;
7187 for_each_cpu_mask_nr(j
, sg
->cpumask
) {
7188 struct sched_domain
*sd
;
7190 sd
= &per_cpu(phys_domains
, j
);
7191 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
7193 * Only add "power" once for each
7199 sg_inc_cpu_power(sg
, sd
->groups
->__cpu_power
);
7202 } while (sg
!= group_head
);
7204 #endif /* CONFIG_NUMA */
7207 /* Free memory allocated for various sched_group structures */
7208 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7212 for_each_cpu_mask_nr(cpu
, *cpu_map
) {
7213 struct sched_group
**sched_group_nodes
7214 = sched_group_nodes_bycpu
[cpu
];
7216 if (!sched_group_nodes
)
7219 for (i
= 0; i
< nr_node_ids
; i
++) {
7220 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
7222 *nodemask
= node_to_cpumask(i
);
7223 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7224 if (cpus_empty(*nodemask
))
7234 if (oldsg
!= sched_group_nodes
[i
])
7237 kfree(sched_group_nodes
);
7238 sched_group_nodes_bycpu
[cpu
] = NULL
;
7241 #else /* !CONFIG_NUMA */
7242 static void free_sched_groups(const cpumask_t
*cpu_map
, cpumask_t
*nodemask
)
7245 #endif /* CONFIG_NUMA */
7248 * Initialize sched groups cpu_power.
7250 * cpu_power indicates the capacity of sched group, which is used while
7251 * distributing the load between different sched groups in a sched domain.
7252 * Typically cpu_power for all the groups in a sched domain will be same unless
7253 * there are asymmetries in the topology. If there are asymmetries, group
7254 * having more cpu_power will pickup more load compared to the group having
7257 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
7258 * the maximum number of tasks a group can handle in the presence of other idle
7259 * or lightly loaded groups in the same sched domain.
7261 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
7263 struct sched_domain
*child
;
7264 struct sched_group
*group
;
7266 WARN_ON(!sd
|| !sd
->groups
);
7268 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
7273 sd
->groups
->__cpu_power
= 0;
7276 * For perf policy, if the groups in child domain share resources
7277 * (for example cores sharing some portions of the cache hierarchy
7278 * or SMT), then set this domain groups cpu_power such that each group
7279 * can handle only one task, when there are other idle groups in the
7280 * same sched domain.
7282 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
7284 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
7285 sg_inc_cpu_power(sd
->groups
, SCHED_LOAD_SCALE
);
7290 * add cpu_power of each child group to this groups cpu_power
7292 group
= child
->groups
;
7294 sg_inc_cpu_power(sd
->groups
, group
->__cpu_power
);
7295 group
= group
->next
;
7296 } while (group
!= child
->groups
);
7300 * Initializers for schedule domains
7301 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
7304 #ifdef CONFIG_SCHED_DEBUG
7305 # define SD_INIT_NAME(sd, type) sd->name = #type
7307 # define SD_INIT_NAME(sd, type) do { } while (0)
7310 #define SD_INIT(sd, type) sd_init_##type(sd)
7312 #define SD_INIT_FUNC(type) \
7313 static noinline void sd_init_##type(struct sched_domain *sd) \
7315 memset(sd, 0, sizeof(*sd)); \
7316 *sd = SD_##type##_INIT; \
7317 sd->level = SD_LV_##type; \
7318 SD_INIT_NAME(sd, type); \
7323 SD_INIT_FUNC(ALLNODES
)
7326 #ifdef CONFIG_SCHED_SMT
7327 SD_INIT_FUNC(SIBLING
)
7329 #ifdef CONFIG_SCHED_MC
7334 * To minimize stack usage kmalloc room for cpumasks and share the
7335 * space as the usage in build_sched_domains() dictates. Used only
7336 * if the amount of space is significant.
7339 cpumask_t tmpmask
; /* make this one first */
7342 cpumask_t this_sibling_map
;
7343 cpumask_t this_core_map
;
7345 cpumask_t send_covered
;
7348 cpumask_t domainspan
;
7350 cpumask_t notcovered
;
7355 #define SCHED_CPUMASK_ALLOC 1
7356 #define SCHED_CPUMASK_FREE(v) kfree(v)
7357 #define SCHED_CPUMASK_DECLARE(v) struct allmasks *v
7359 #define SCHED_CPUMASK_ALLOC 0
7360 #define SCHED_CPUMASK_FREE(v)
7361 #define SCHED_CPUMASK_DECLARE(v) struct allmasks _v, *v = &_v
7364 #define SCHED_CPUMASK_VAR(v, a) cpumask_t *v = (cpumask_t *) \
7365 ((unsigned long)(a) + offsetof(struct allmasks, v))
7367 static int default_relax_domain_level
= -1;
7369 static int __init
setup_relax_domain_level(char *str
)
7373 val
= simple_strtoul(str
, NULL
, 0);
7374 if (val
< SD_LV_MAX
)
7375 default_relax_domain_level
= val
;
7379 __setup("relax_domain_level=", setup_relax_domain_level
);
7381 static void set_domain_attribute(struct sched_domain
*sd
,
7382 struct sched_domain_attr
*attr
)
7386 if (!attr
|| attr
->relax_domain_level
< 0) {
7387 if (default_relax_domain_level
< 0)
7390 request
= default_relax_domain_level
;
7392 request
= attr
->relax_domain_level
;
7393 if (request
< sd
->level
) {
7394 /* turn off idle balance on this domain */
7395 sd
->flags
&= ~(SD_WAKE_IDLE
|SD_BALANCE_NEWIDLE
);
7397 /* turn on idle balance on this domain */
7398 sd
->flags
|= (SD_WAKE_IDLE_FAR
|SD_BALANCE_NEWIDLE
);
7403 * Build sched domains for a given set of cpus and attach the sched domains
7404 * to the individual cpus
7406 static int __build_sched_domains(const cpumask_t
*cpu_map
,
7407 struct sched_domain_attr
*attr
)
7410 struct root_domain
*rd
;
7411 SCHED_CPUMASK_DECLARE(allmasks
);
7414 struct sched_group
**sched_group_nodes
= NULL
;
7415 int sd_allnodes
= 0;
7418 * Allocate the per-node list of sched groups
7420 sched_group_nodes
= kcalloc(nr_node_ids
, sizeof(struct sched_group
*),
7422 if (!sched_group_nodes
) {
7423 printk(KERN_WARNING
"Can not alloc sched group node list\n");
7428 rd
= alloc_rootdomain();
7430 printk(KERN_WARNING
"Cannot alloc root domain\n");
7432 kfree(sched_group_nodes
);
7437 #if SCHED_CPUMASK_ALLOC
7438 /* get space for all scratch cpumask variables */
7439 allmasks
= kmalloc(sizeof(*allmasks
), GFP_KERNEL
);
7441 printk(KERN_WARNING
"Cannot alloc cpumask array\n");
7444 kfree(sched_group_nodes
);
7449 tmpmask
= (cpumask_t
*)allmasks
;
7453 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
7457 * Set up domains for cpus specified by the cpu_map.
7459 for_each_cpu_mask_nr(i
, *cpu_map
) {
7460 struct sched_domain
*sd
= NULL
, *p
;
7461 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7463 *nodemask
= node_to_cpumask(cpu_to_node(i
));
7464 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7467 if (cpus_weight(*cpu_map
) >
7468 SD_NODES_PER_DOMAIN
*cpus_weight(*nodemask
)) {
7469 sd
= &per_cpu(allnodes_domains
, i
);
7470 SD_INIT(sd
, ALLNODES
);
7471 set_domain_attribute(sd
, attr
);
7472 sd
->span
= *cpu_map
;
7473 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7479 sd
= &per_cpu(node_domains
, i
);
7481 set_domain_attribute(sd
, attr
);
7482 sched_domain_node_span(cpu_to_node(i
), &sd
->span
);
7486 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7490 sd
= &per_cpu(phys_domains
, i
);
7492 set_domain_attribute(sd
, attr
);
7493 sd
->span
= *nodemask
;
7497 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7499 #ifdef CONFIG_SCHED_MC
7501 sd
= &per_cpu(core_domains
, i
);
7503 set_domain_attribute(sd
, attr
);
7504 sd
->span
= cpu_coregroup_map(i
);
7505 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7508 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7511 #ifdef CONFIG_SCHED_SMT
7513 sd
= &per_cpu(cpu_domains
, i
);
7514 SD_INIT(sd
, SIBLING
);
7515 set_domain_attribute(sd
, attr
);
7516 sd
->span
= per_cpu(cpu_sibling_map
, i
);
7517 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
7520 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, tmpmask
);
7524 #ifdef CONFIG_SCHED_SMT
7525 /* Set up CPU (sibling) groups */
7526 for_each_cpu_mask_nr(i
, *cpu_map
) {
7527 SCHED_CPUMASK_VAR(this_sibling_map
, allmasks
);
7528 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7530 *this_sibling_map
= per_cpu(cpu_sibling_map
, i
);
7531 cpus_and(*this_sibling_map
, *this_sibling_map
, *cpu_map
);
7532 if (i
!= first_cpu(*this_sibling_map
))
7535 init_sched_build_groups(this_sibling_map
, cpu_map
,
7537 send_covered
, tmpmask
);
7541 #ifdef CONFIG_SCHED_MC
7542 /* Set up multi-core groups */
7543 for_each_cpu_mask_nr(i
, *cpu_map
) {
7544 SCHED_CPUMASK_VAR(this_core_map
, allmasks
);
7545 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7547 *this_core_map
= cpu_coregroup_map(i
);
7548 cpus_and(*this_core_map
, *this_core_map
, *cpu_map
);
7549 if (i
!= first_cpu(*this_core_map
))
7552 init_sched_build_groups(this_core_map
, cpu_map
,
7554 send_covered
, tmpmask
);
7558 /* Set up physical groups */
7559 for (i
= 0; i
< nr_node_ids
; i
++) {
7560 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7561 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7563 *nodemask
= node_to_cpumask(i
);
7564 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7565 if (cpus_empty(*nodemask
))
7568 init_sched_build_groups(nodemask
, cpu_map
,
7570 send_covered
, tmpmask
);
7574 /* Set up node groups */
7576 SCHED_CPUMASK_VAR(send_covered
, allmasks
);
7578 init_sched_build_groups(cpu_map
, cpu_map
,
7579 &cpu_to_allnodes_group
,
7580 send_covered
, tmpmask
);
7583 for (i
= 0; i
< nr_node_ids
; i
++) {
7584 /* Set up node groups */
7585 struct sched_group
*sg
, *prev
;
7586 SCHED_CPUMASK_VAR(nodemask
, allmasks
);
7587 SCHED_CPUMASK_VAR(domainspan
, allmasks
);
7588 SCHED_CPUMASK_VAR(covered
, allmasks
);
7591 *nodemask
= node_to_cpumask(i
);
7592 cpus_clear(*covered
);
7594 cpus_and(*nodemask
, *nodemask
, *cpu_map
);
7595 if (cpus_empty(*nodemask
)) {
7596 sched_group_nodes
[i
] = NULL
;
7600 sched_domain_node_span(i
, domainspan
);
7601 cpus_and(*domainspan
, *domainspan
, *cpu_map
);
7603 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
7605 printk(KERN_WARNING
"Can not alloc domain group for "
7609 sched_group_nodes
[i
] = sg
;
7610 for_each_cpu_mask_nr(j
, *nodemask
) {
7611 struct sched_domain
*sd
;
7613 sd
= &per_cpu(node_domains
, j
);
7616 sg
->__cpu_power
= 0;
7617 sg
->cpumask
= *nodemask
;
7619 cpus_or(*covered
, *covered
, *nodemask
);
7622 for (j
= 0; j
< nr_node_ids
; j
++) {
7623 SCHED_CPUMASK_VAR(notcovered
, allmasks
);
7624 int n
= (i
+ j
) % nr_node_ids
;
7625 node_to_cpumask_ptr(pnodemask
, n
);
7627 cpus_complement(*notcovered
, *covered
);
7628 cpus_and(*tmpmask
, *notcovered
, *cpu_map
);
7629 cpus_and(*tmpmask
, *tmpmask
, *domainspan
);
7630 if (cpus_empty(*tmpmask
))
7633 cpus_and(*tmpmask
, *tmpmask
, *pnodemask
);
7634 if (cpus_empty(*tmpmask
))
7637 sg
= kmalloc_node(sizeof(struct sched_group
),
7641 "Can not alloc domain group for node %d\n", j
);
7644 sg
->__cpu_power
= 0;
7645 sg
->cpumask
= *tmpmask
;
7646 sg
->next
= prev
->next
;
7647 cpus_or(*covered
, *covered
, *tmpmask
);
7654 /* Calculate CPU power for physical packages and nodes */
7655 #ifdef CONFIG_SCHED_SMT
7656 for_each_cpu_mask_nr(i
, *cpu_map
) {
7657 struct sched_domain
*sd
= &per_cpu(cpu_domains
, i
);
7659 init_sched_groups_power(i
, sd
);
7662 #ifdef CONFIG_SCHED_MC
7663 for_each_cpu_mask_nr(i
, *cpu_map
) {
7664 struct sched_domain
*sd
= &per_cpu(core_domains
, i
);
7666 init_sched_groups_power(i
, sd
);
7670 for_each_cpu_mask_nr(i
, *cpu_map
) {
7671 struct sched_domain
*sd
= &per_cpu(phys_domains
, i
);
7673 init_sched_groups_power(i
, sd
);
7677 for (i
= 0; i
< nr_node_ids
; i
++)
7678 init_numa_sched_groups_power(sched_group_nodes
[i
]);
7681 struct sched_group
*sg
;
7683 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
,
7685 init_numa_sched_groups_power(sg
);
7689 /* Attach the domains */
7690 for_each_cpu_mask_nr(i
, *cpu_map
) {
7691 struct sched_domain
*sd
;
7692 #ifdef CONFIG_SCHED_SMT
7693 sd
= &per_cpu(cpu_domains
, i
);
7694 #elif defined(CONFIG_SCHED_MC)
7695 sd
= &per_cpu(core_domains
, i
);
7697 sd
= &per_cpu(phys_domains
, i
);
7699 cpu_attach_domain(sd
, rd
, i
);
7702 SCHED_CPUMASK_FREE((void *)allmasks
);
7707 free_sched_groups(cpu_map
, tmpmask
);
7708 SCHED_CPUMASK_FREE((void *)allmasks
);
7714 static int build_sched_domains(const cpumask_t
*cpu_map
)
7716 return __build_sched_domains(cpu_map
, NULL
);
7719 static cpumask_t
*doms_cur
; /* current sched domains */
7720 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
7721 static struct sched_domain_attr
*dattr_cur
;
7722 /* attribues of custom domains in 'doms_cur' */
7725 * Special case: If a kmalloc of a doms_cur partition (array of
7726 * cpumask_t) fails, then fallback to a single sched domain,
7727 * as determined by the single cpumask_t fallback_doms.
7729 static cpumask_t fallback_doms
;
7731 void __attribute__((weak
)) arch_update_cpu_topology(void)
7736 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
7737 * For now this just excludes isolated cpus, but could be used to
7738 * exclude other special cases in the future.
7740 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
7744 arch_update_cpu_topology();
7746 doms_cur
= kmalloc(sizeof(cpumask_t
), GFP_KERNEL
);
7748 doms_cur
= &fallback_doms
;
7749 cpus_andnot(*doms_cur
, *cpu_map
, cpu_isolated_map
);
7751 err
= build_sched_domains(doms_cur
);
7752 register_sched_domain_sysctl();
7757 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
,
7760 free_sched_groups(cpu_map
, tmpmask
);
7764 * Detach sched domains from a group of cpus specified in cpu_map
7765 * These cpus will now be attached to the NULL domain
7767 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
7772 unregister_sched_domain_sysctl();
7774 for_each_cpu_mask_nr(i
, *cpu_map
)
7775 cpu_attach_domain(NULL
, &def_root_domain
, i
);
7776 synchronize_sched();
7777 arch_destroy_sched_domains(cpu_map
, &tmpmask
);
7780 /* handle null as "default" */
7781 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
7782 struct sched_domain_attr
*new, int idx_new
)
7784 struct sched_domain_attr tmp
;
7791 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
7792 new ? (new + idx_new
) : &tmp
,
7793 sizeof(struct sched_domain_attr
));
7797 * Partition sched domains as specified by the 'ndoms_new'
7798 * cpumasks in the array doms_new[] of cpumasks. This compares
7799 * doms_new[] to the current sched domain partitioning, doms_cur[].
7800 * It destroys each deleted domain and builds each new domain.
7802 * 'doms_new' is an array of cpumask_t's of length 'ndoms_new'.
7803 * The masks don't intersect (don't overlap.) We should setup one
7804 * sched domain for each mask. CPUs not in any of the cpumasks will
7805 * not be load balanced. If the same cpumask appears both in the
7806 * current 'doms_cur' domains and in the new 'doms_new', we can leave
7809 * The passed in 'doms_new' should be kmalloc'd. This routine takes
7810 * ownership of it and will kfree it when done with it. If the caller
7811 * failed the kmalloc call, then it can pass in doms_new == NULL &&
7812 * ndoms_new == 1, and partition_sched_domains() will fallback to
7813 * the single partition 'fallback_doms', it also forces the domains
7816 * If doms_new == NULL it will be replaced with cpu_online_map.
7817 * ndoms_new == 0 is a special case for destroying existing domains,
7818 * and it will not create the default domain.
7820 * Call with hotplug lock held
7822 void partition_sched_domains(int ndoms_new
, cpumask_t
*doms_new
,
7823 struct sched_domain_attr
*dattr_new
)
7827 mutex_lock(&sched_domains_mutex
);
7829 /* always unregister in case we don't destroy any domains */
7830 unregister_sched_domain_sysctl();
7832 n
= doms_new
? ndoms_new
: 0;
7834 /* Destroy deleted domains */
7835 for (i
= 0; i
< ndoms_cur
; i
++) {
7836 for (j
= 0; j
< n
; j
++) {
7837 if (cpus_equal(doms_cur
[i
], doms_new
[j
])
7838 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
7841 /* no match - a current sched domain not in new doms_new[] */
7842 detach_destroy_domains(doms_cur
+ i
);
7847 if (doms_new
== NULL
) {
7849 doms_new
= &fallback_doms
;
7850 cpus_andnot(doms_new
[0], cpu_online_map
, cpu_isolated_map
);
7854 /* Build new domains */
7855 for (i
= 0; i
< ndoms_new
; i
++) {
7856 for (j
= 0; j
< ndoms_cur
; j
++) {
7857 if (cpus_equal(doms_new
[i
], doms_cur
[j
])
7858 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
7861 /* no match - add a new doms_new */
7862 __build_sched_domains(doms_new
+ i
,
7863 dattr_new
? dattr_new
+ i
: NULL
);
7868 /* Remember the new sched domains */
7869 if (doms_cur
!= &fallback_doms
)
7871 kfree(dattr_cur
); /* kfree(NULL) is safe */
7872 doms_cur
= doms_new
;
7873 dattr_cur
= dattr_new
;
7874 ndoms_cur
= ndoms_new
;
7876 register_sched_domain_sysctl();
7878 mutex_unlock(&sched_domains_mutex
);
7881 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
7882 int arch_reinit_sched_domains(void)
7886 /* Destroy domains first to force the rebuild */
7887 partition_sched_domains(0, NULL
, NULL
);
7889 rebuild_sched_domains();
7895 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
7899 if (buf
[0] != '0' && buf
[0] != '1')
7903 sched_smt_power_savings
= (buf
[0] == '1');
7905 sched_mc_power_savings
= (buf
[0] == '1');
7907 ret
= arch_reinit_sched_domains();
7909 return ret
? ret
: count
;
7912 #ifdef CONFIG_SCHED_MC
7913 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
7916 return sprintf(page
, "%u\n", sched_mc_power_savings
);
7918 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
7919 const char *buf
, size_t count
)
7921 return sched_power_savings_store(buf
, count
, 0);
7923 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
7924 sched_mc_power_savings_show
,
7925 sched_mc_power_savings_store
);
7928 #ifdef CONFIG_SCHED_SMT
7929 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
7932 return sprintf(page
, "%u\n", sched_smt_power_savings
);
7934 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
7935 const char *buf
, size_t count
)
7937 return sched_power_savings_store(buf
, count
, 1);
7939 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
7940 sched_smt_power_savings_show
,
7941 sched_smt_power_savings_store
);
7944 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
7948 #ifdef CONFIG_SCHED_SMT
7950 err
= sysfs_create_file(&cls
->kset
.kobj
,
7951 &attr_sched_smt_power_savings
.attr
);
7953 #ifdef CONFIG_SCHED_MC
7954 if (!err
&& mc_capable())
7955 err
= sysfs_create_file(&cls
->kset
.kobj
,
7956 &attr_sched_mc_power_savings
.attr
);
7960 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
7962 #ifndef CONFIG_CPUSETS
7964 * Add online and remove offline CPUs from the scheduler domains.
7965 * When cpusets are enabled they take over this function.
7967 static int update_sched_domains(struct notifier_block
*nfb
,
7968 unsigned long action
, void *hcpu
)
7972 case CPU_ONLINE_FROZEN
:
7974 case CPU_DEAD_FROZEN
:
7975 partition_sched_domains(1, NULL
, NULL
);
7984 static int update_runtime(struct notifier_block
*nfb
,
7985 unsigned long action
, void *hcpu
)
7987 int cpu
= (int)(long)hcpu
;
7990 case CPU_DOWN_PREPARE
:
7991 case CPU_DOWN_PREPARE_FROZEN
:
7992 disable_runtime(cpu_rq(cpu
));
7995 case CPU_DOWN_FAILED
:
7996 case CPU_DOWN_FAILED_FROZEN
:
7998 case CPU_ONLINE_FROZEN
:
7999 enable_runtime(cpu_rq(cpu
));
8007 void __init
sched_init_smp(void)
8009 cpumask_t non_isolated_cpus
;
8011 #if defined(CONFIG_NUMA)
8012 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
8014 BUG_ON(sched_group_nodes_bycpu
== NULL
);
8017 mutex_lock(&sched_domains_mutex
);
8018 arch_init_sched_domains(&cpu_online_map
);
8019 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
8020 if (cpus_empty(non_isolated_cpus
))
8021 cpu_set(smp_processor_id(), non_isolated_cpus
);
8022 mutex_unlock(&sched_domains_mutex
);
8025 #ifndef CONFIG_CPUSETS
8026 /* XXX: Theoretical race here - CPU may be hotplugged now */
8027 hotcpu_notifier(update_sched_domains
, 0);
8030 /* RT runtime code needs to handle some hotplug events */
8031 hotcpu_notifier(update_runtime
, 0);
8035 /* Move init over to a non-isolated CPU */
8036 if (set_cpus_allowed_ptr(current
, &non_isolated_cpus
) < 0)
8038 sched_init_granularity();
8041 void __init
sched_init_smp(void)
8043 sched_init_granularity();
8045 #endif /* CONFIG_SMP */
8047 int in_sched_functions(unsigned long addr
)
8049 return in_lock_functions(addr
) ||
8050 (addr
>= (unsigned long)__sched_text_start
8051 && addr
< (unsigned long)__sched_text_end
);
8054 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
8056 cfs_rq
->tasks_timeline
= RB_ROOT
;
8057 INIT_LIST_HEAD(&cfs_rq
->tasks
);
8058 #ifdef CONFIG_FAIR_GROUP_SCHED
8061 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
8064 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
8066 struct rt_prio_array
*array
;
8069 array
= &rt_rq
->active
;
8070 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
8071 INIT_LIST_HEAD(array
->queue
+ i
);
8072 __clear_bit(i
, array
->bitmap
);
8074 /* delimiter for bitsearch: */
8075 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
8077 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
8078 rt_rq
->highest_prio
= MAX_RT_PRIO
;
8081 rt_rq
->rt_nr_migratory
= 0;
8082 rt_rq
->overloaded
= 0;
8086 rt_rq
->rt_throttled
= 0;
8087 rt_rq
->rt_runtime
= 0;
8088 spin_lock_init(&rt_rq
->rt_runtime_lock
);
8090 #ifdef CONFIG_RT_GROUP_SCHED
8091 rt_rq
->rt_nr_boosted
= 0;
8096 #ifdef CONFIG_FAIR_GROUP_SCHED
8097 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
8098 struct sched_entity
*se
, int cpu
, int add
,
8099 struct sched_entity
*parent
)
8101 struct rq
*rq
= cpu_rq(cpu
);
8102 tg
->cfs_rq
[cpu
] = cfs_rq
;
8103 init_cfs_rq(cfs_rq
, rq
);
8106 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
8109 /* se could be NULL for init_task_group */
8114 se
->cfs_rq
= &rq
->cfs
;
8116 se
->cfs_rq
= parent
->my_q
;
8119 se
->load
.weight
= tg
->shares
;
8120 se
->load
.inv_weight
= 0;
8121 se
->parent
= parent
;
8125 #ifdef CONFIG_RT_GROUP_SCHED
8126 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
8127 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
8128 struct sched_rt_entity
*parent
)
8130 struct rq
*rq
= cpu_rq(cpu
);
8132 tg
->rt_rq
[cpu
] = rt_rq
;
8133 init_rt_rq(rt_rq
, rq
);
8135 rt_rq
->rt_se
= rt_se
;
8136 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
8138 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
8140 tg
->rt_se
[cpu
] = rt_se
;
8145 rt_se
->rt_rq
= &rq
->rt
;
8147 rt_se
->rt_rq
= parent
->my_q
;
8149 rt_se
->my_q
= rt_rq
;
8150 rt_se
->parent
= parent
;
8151 INIT_LIST_HEAD(&rt_se
->run_list
);
8155 void __init
sched_init(void)
8158 unsigned long alloc_size
= 0, ptr
;
8160 #ifdef CONFIG_FAIR_GROUP_SCHED
8161 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8163 #ifdef CONFIG_RT_GROUP_SCHED
8164 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
8166 #ifdef CONFIG_USER_SCHED
8170 * As sched_init() is called before page_alloc is setup,
8171 * we use alloc_bootmem().
8174 ptr
= (unsigned long)alloc_bootmem(alloc_size
);
8176 #ifdef CONFIG_FAIR_GROUP_SCHED
8177 init_task_group
.se
= (struct sched_entity
**)ptr
;
8178 ptr
+= nr_cpu_ids
* sizeof(void **);
8180 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8181 ptr
+= nr_cpu_ids
* sizeof(void **);
8183 #ifdef CONFIG_USER_SCHED
8184 root_task_group
.se
= (struct sched_entity
**)ptr
;
8185 ptr
+= nr_cpu_ids
* sizeof(void **);
8187 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
8188 ptr
+= nr_cpu_ids
* sizeof(void **);
8189 #endif /* CONFIG_USER_SCHED */
8190 #endif /* CONFIG_FAIR_GROUP_SCHED */
8191 #ifdef CONFIG_RT_GROUP_SCHED
8192 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8193 ptr
+= nr_cpu_ids
* sizeof(void **);
8195 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8196 ptr
+= nr_cpu_ids
* sizeof(void **);
8198 #ifdef CONFIG_USER_SCHED
8199 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
8200 ptr
+= nr_cpu_ids
* sizeof(void **);
8202 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
8203 ptr
+= nr_cpu_ids
* sizeof(void **);
8204 #endif /* CONFIG_USER_SCHED */
8205 #endif /* CONFIG_RT_GROUP_SCHED */
8209 init_defrootdomain();
8212 init_rt_bandwidth(&def_rt_bandwidth
,
8213 global_rt_period(), global_rt_runtime());
8215 #ifdef CONFIG_RT_GROUP_SCHED
8216 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
8217 global_rt_period(), global_rt_runtime());
8218 #ifdef CONFIG_USER_SCHED
8219 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
8220 global_rt_period(), RUNTIME_INF
);
8221 #endif /* CONFIG_USER_SCHED */
8222 #endif /* CONFIG_RT_GROUP_SCHED */
8224 #ifdef CONFIG_GROUP_SCHED
8225 list_add(&init_task_group
.list
, &task_groups
);
8226 INIT_LIST_HEAD(&init_task_group
.children
);
8228 #ifdef CONFIG_USER_SCHED
8229 INIT_LIST_HEAD(&root_task_group
.children
);
8230 init_task_group
.parent
= &root_task_group
;
8231 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
8232 #endif /* CONFIG_USER_SCHED */
8233 #endif /* CONFIG_GROUP_SCHED */
8235 for_each_possible_cpu(i
) {
8239 spin_lock_init(&rq
->lock
);
8241 init_cfs_rq(&rq
->cfs
, rq
);
8242 init_rt_rq(&rq
->rt
, rq
);
8243 #ifdef CONFIG_FAIR_GROUP_SCHED
8244 init_task_group
.shares
= init_task_group_load
;
8245 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
8246 #ifdef CONFIG_CGROUP_SCHED
8248 * How much cpu bandwidth does init_task_group get?
8250 * In case of task-groups formed thr' the cgroup filesystem, it
8251 * gets 100% of the cpu resources in the system. This overall
8252 * system cpu resource is divided among the tasks of
8253 * init_task_group and its child task-groups in a fair manner,
8254 * based on each entity's (task or task-group's) weight
8255 * (se->load.weight).
8257 * In other words, if init_task_group has 10 tasks of weight
8258 * 1024) and two child groups A0 and A1 (of weight 1024 each),
8259 * then A0's share of the cpu resource is:
8261 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
8263 * We achieve this by letting init_task_group's tasks sit
8264 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
8266 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
8267 #elif defined CONFIG_USER_SCHED
8268 root_task_group
.shares
= NICE_0_LOAD
;
8269 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
8271 * In case of task-groups formed thr' the user id of tasks,
8272 * init_task_group represents tasks belonging to root user.
8273 * Hence it forms a sibling of all subsequent groups formed.
8274 * In this case, init_task_group gets only a fraction of overall
8275 * system cpu resource, based on the weight assigned to root
8276 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
8277 * by letting tasks of init_task_group sit in a separate cfs_rq
8278 * (init_cfs_rq) and having one entity represent this group of
8279 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
8281 init_tg_cfs_entry(&init_task_group
,
8282 &per_cpu(init_cfs_rq
, i
),
8283 &per_cpu(init_sched_entity
, i
), i
, 1,
8284 root_task_group
.se
[i
]);
8287 #endif /* CONFIG_FAIR_GROUP_SCHED */
8289 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
8290 #ifdef CONFIG_RT_GROUP_SCHED
8291 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
8292 #ifdef CONFIG_CGROUP_SCHED
8293 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
8294 #elif defined CONFIG_USER_SCHED
8295 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
8296 init_tg_rt_entry(&init_task_group
,
8297 &per_cpu(init_rt_rq
, i
),
8298 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
8299 root_task_group
.rt_se
[i
]);
8303 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
8304 rq
->cpu_load
[j
] = 0;
8308 rq
->active_balance
= 0;
8309 rq
->next_balance
= jiffies
;
8313 rq
->migration_thread
= NULL
;
8314 INIT_LIST_HEAD(&rq
->migration_queue
);
8315 rq_attach_root(rq
, &def_root_domain
);
8318 atomic_set(&rq
->nr_iowait
, 0);
8321 set_load_weight(&init_task
);
8323 #ifdef CONFIG_PREEMPT_NOTIFIERS
8324 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
8328 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
8331 #ifdef CONFIG_RT_MUTEXES
8332 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
8336 * The boot idle thread does lazy MMU switching as well:
8338 atomic_inc(&init_mm
.mm_count
);
8339 enter_lazy_tlb(&init_mm
, current
);
8342 * Make us the idle thread. Technically, schedule() should not be
8343 * called from this thread, however somewhere below it might be,
8344 * but because we are the idle thread, we just pick up running again
8345 * when this runqueue becomes "idle".
8347 init_idle(current
, smp_processor_id());
8349 * During early bootup we pretend to be a normal task:
8351 current
->sched_class
= &fair_sched_class
;
8353 scheduler_running
= 1;
8356 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
8357 void __might_sleep(char *file
, int line
)
8360 static unsigned long prev_jiffy
; /* ratelimiting */
8362 if ((!in_atomic() && !irqs_disabled()) ||
8363 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
8365 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
8367 prev_jiffy
= jiffies
;
8370 "BUG: sleeping function called from invalid context at %s:%d\n",
8373 "in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
8374 in_atomic(), irqs_disabled(),
8375 current
->pid
, current
->comm
);
8377 debug_show_held_locks(current
);
8378 if (irqs_disabled())
8379 print_irqtrace_events(current
);
8383 EXPORT_SYMBOL(__might_sleep
);
8386 #ifdef CONFIG_MAGIC_SYSRQ
8387 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
8391 update_rq_clock(rq
);
8392 on_rq
= p
->se
.on_rq
;
8394 deactivate_task(rq
, p
, 0);
8395 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
8397 activate_task(rq
, p
, 0);
8398 resched_task(rq
->curr
);
8402 void normalize_rt_tasks(void)
8404 struct task_struct
*g
, *p
;
8405 unsigned long flags
;
8408 read_lock_irqsave(&tasklist_lock
, flags
);
8409 do_each_thread(g
, p
) {
8411 * Only normalize user tasks:
8416 p
->se
.exec_start
= 0;
8417 #ifdef CONFIG_SCHEDSTATS
8418 p
->se
.wait_start
= 0;
8419 p
->se
.sleep_start
= 0;
8420 p
->se
.block_start
= 0;
8425 * Renice negative nice level userspace
8428 if (TASK_NICE(p
) < 0 && p
->mm
)
8429 set_user_nice(p
, 0);
8433 spin_lock(&p
->pi_lock
);
8434 rq
= __task_rq_lock(p
);
8436 normalize_task(rq
, p
);
8438 __task_rq_unlock(rq
);
8439 spin_unlock(&p
->pi_lock
);
8440 } while_each_thread(g
, p
);
8442 read_unlock_irqrestore(&tasklist_lock
, flags
);
8445 #endif /* CONFIG_MAGIC_SYSRQ */
8449 * These functions are only useful for the IA64 MCA handling.
8451 * They can only be called when the whole system has been
8452 * stopped - every CPU needs to be quiescent, and no scheduling
8453 * activity can take place. Using them for anything else would
8454 * be a serious bug, and as a result, they aren't even visible
8455 * under any other configuration.
8459 * curr_task - return the current task for a given cpu.
8460 * @cpu: the processor in question.
8462 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8464 struct task_struct
*curr_task(int cpu
)
8466 return cpu_curr(cpu
);
8470 * set_curr_task - set the current task for a given cpu.
8471 * @cpu: the processor in question.
8472 * @p: the task pointer to set.
8474 * Description: This function must only be used when non-maskable interrupts
8475 * are serviced on a separate stack. It allows the architecture to switch the
8476 * notion of the current task on a cpu in a non-blocking manner. This function
8477 * must be called with all CPU's synchronized, and interrupts disabled, the
8478 * and caller must save the original value of the current task (see
8479 * curr_task() above) and restore that value before reenabling interrupts and
8480 * re-starting the system.
8482 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
8484 void set_curr_task(int cpu
, struct task_struct
*p
)
8491 #ifdef CONFIG_FAIR_GROUP_SCHED
8492 static void free_fair_sched_group(struct task_group
*tg
)
8496 for_each_possible_cpu(i
) {
8498 kfree(tg
->cfs_rq
[i
]);
8508 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8510 struct cfs_rq
*cfs_rq
;
8511 struct sched_entity
*se
, *parent_se
;
8515 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8518 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
8522 tg
->shares
= NICE_0_LOAD
;
8524 for_each_possible_cpu(i
) {
8527 cfs_rq
= kmalloc_node(sizeof(struct cfs_rq
),
8528 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8532 se
= kmalloc_node(sizeof(struct sched_entity
),
8533 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8537 parent_se
= parent
? parent
->se
[i
] : NULL
;
8538 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent_se
);
8547 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8549 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
8550 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
8553 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8555 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
8557 #else /* !CONFG_FAIR_GROUP_SCHED */
8558 static inline void free_fair_sched_group(struct task_group
*tg
)
8563 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8568 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
8572 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
8575 #endif /* CONFIG_FAIR_GROUP_SCHED */
8577 #ifdef CONFIG_RT_GROUP_SCHED
8578 static void free_rt_sched_group(struct task_group
*tg
)
8582 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
8584 for_each_possible_cpu(i
) {
8586 kfree(tg
->rt_rq
[i
]);
8588 kfree(tg
->rt_se
[i
]);
8596 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8598 struct rt_rq
*rt_rq
;
8599 struct sched_rt_entity
*rt_se
, *parent_se
;
8603 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
8606 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
8610 init_rt_bandwidth(&tg
->rt_bandwidth
,
8611 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
8613 for_each_possible_cpu(i
) {
8616 rt_rq
= kmalloc_node(sizeof(struct rt_rq
),
8617 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8621 rt_se
= kmalloc_node(sizeof(struct sched_rt_entity
),
8622 GFP_KERNEL
|__GFP_ZERO
, cpu_to_node(i
));
8626 parent_se
= parent
? parent
->rt_se
[i
] : NULL
;
8627 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent_se
);
8636 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8638 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
8639 &cpu_rq(cpu
)->leaf_rt_rq_list
);
8642 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8644 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
8646 #else /* !CONFIG_RT_GROUP_SCHED */
8647 static inline void free_rt_sched_group(struct task_group
*tg
)
8652 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
8657 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
8661 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
8664 #endif /* CONFIG_RT_GROUP_SCHED */
8666 #ifdef CONFIG_GROUP_SCHED
8667 static void free_sched_group(struct task_group
*tg
)
8669 free_fair_sched_group(tg
);
8670 free_rt_sched_group(tg
);
8674 /* allocate runqueue etc for a new task group */
8675 struct task_group
*sched_create_group(struct task_group
*parent
)
8677 struct task_group
*tg
;
8678 unsigned long flags
;
8681 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
8683 return ERR_PTR(-ENOMEM
);
8685 if (!alloc_fair_sched_group(tg
, parent
))
8688 if (!alloc_rt_sched_group(tg
, parent
))
8691 spin_lock_irqsave(&task_group_lock
, flags
);
8692 for_each_possible_cpu(i
) {
8693 register_fair_sched_group(tg
, i
);
8694 register_rt_sched_group(tg
, i
);
8696 list_add_rcu(&tg
->list
, &task_groups
);
8698 WARN_ON(!parent
); /* root should already exist */
8700 tg
->parent
= parent
;
8701 INIT_LIST_HEAD(&tg
->children
);
8702 list_add_rcu(&tg
->siblings
, &parent
->children
);
8703 spin_unlock_irqrestore(&task_group_lock
, flags
);
8708 free_sched_group(tg
);
8709 return ERR_PTR(-ENOMEM
);
8712 /* rcu callback to free various structures associated with a task group */
8713 static void free_sched_group_rcu(struct rcu_head
*rhp
)
8715 /* now it should be safe to free those cfs_rqs */
8716 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
8719 /* Destroy runqueue etc associated with a task group */
8720 void sched_destroy_group(struct task_group
*tg
)
8722 unsigned long flags
;
8725 spin_lock_irqsave(&task_group_lock
, flags
);
8726 for_each_possible_cpu(i
) {
8727 unregister_fair_sched_group(tg
, i
);
8728 unregister_rt_sched_group(tg
, i
);
8730 list_del_rcu(&tg
->list
);
8731 list_del_rcu(&tg
->siblings
);
8732 spin_unlock_irqrestore(&task_group_lock
, flags
);
8734 /* wait for possible concurrent references to cfs_rqs complete */
8735 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
8738 /* change task's runqueue when it moves between groups.
8739 * The caller of this function should have put the task in its new group
8740 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
8741 * reflect its new group.
8743 void sched_move_task(struct task_struct
*tsk
)
8746 unsigned long flags
;
8749 rq
= task_rq_lock(tsk
, &flags
);
8751 update_rq_clock(rq
);
8753 running
= task_current(rq
, tsk
);
8754 on_rq
= tsk
->se
.on_rq
;
8757 dequeue_task(rq
, tsk
, 0);
8758 if (unlikely(running
))
8759 tsk
->sched_class
->put_prev_task(rq
, tsk
);
8761 set_task_rq(tsk
, task_cpu(tsk
));
8763 #ifdef CONFIG_FAIR_GROUP_SCHED
8764 if (tsk
->sched_class
->moved_group
)
8765 tsk
->sched_class
->moved_group(tsk
);
8768 if (unlikely(running
))
8769 tsk
->sched_class
->set_curr_task(rq
);
8771 enqueue_task(rq
, tsk
, 0);
8773 task_rq_unlock(rq
, &flags
);
8775 #endif /* CONFIG_GROUP_SCHED */
8777 #ifdef CONFIG_FAIR_GROUP_SCHED
8778 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8780 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8785 dequeue_entity(cfs_rq
, se
, 0);
8787 se
->load
.weight
= shares
;
8788 se
->load
.inv_weight
= 0;
8791 enqueue_entity(cfs_rq
, se
, 0);
8794 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
8796 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
8797 struct rq
*rq
= cfs_rq
->rq
;
8798 unsigned long flags
;
8800 spin_lock_irqsave(&rq
->lock
, flags
);
8801 __set_se_shares(se
, shares
);
8802 spin_unlock_irqrestore(&rq
->lock
, flags
);
8805 static DEFINE_MUTEX(shares_mutex
);
8807 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
8810 unsigned long flags
;
8813 * We can't change the weight of the root cgroup.
8818 if (shares
< MIN_SHARES
)
8819 shares
= MIN_SHARES
;
8820 else if (shares
> MAX_SHARES
)
8821 shares
= MAX_SHARES
;
8823 mutex_lock(&shares_mutex
);
8824 if (tg
->shares
== shares
)
8827 spin_lock_irqsave(&task_group_lock
, flags
);
8828 for_each_possible_cpu(i
)
8829 unregister_fair_sched_group(tg
, i
);
8830 list_del_rcu(&tg
->siblings
);
8831 spin_unlock_irqrestore(&task_group_lock
, flags
);
8833 /* wait for any ongoing reference to this group to finish */
8834 synchronize_sched();
8837 * Now we are free to modify the group's share on each cpu
8838 * w/o tripping rebalance_share or load_balance_fair.
8840 tg
->shares
= shares
;
8841 for_each_possible_cpu(i
) {
8845 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
8846 set_se_shares(tg
->se
[i
], shares
);
8850 * Enable load balance activity on this group, by inserting it back on
8851 * each cpu's rq->leaf_cfs_rq_list.
8853 spin_lock_irqsave(&task_group_lock
, flags
);
8854 for_each_possible_cpu(i
)
8855 register_fair_sched_group(tg
, i
);
8856 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
8857 spin_unlock_irqrestore(&task_group_lock
, flags
);
8859 mutex_unlock(&shares_mutex
);
8863 unsigned long sched_group_shares(struct task_group
*tg
)
8869 #ifdef CONFIG_RT_GROUP_SCHED
8871 * Ensure that the real time constraints are schedulable.
8873 static DEFINE_MUTEX(rt_constraints_mutex
);
8875 static unsigned long to_ratio(u64 period
, u64 runtime
)
8877 if (runtime
== RUNTIME_INF
)
8880 return div64_u64(runtime
<< 20, period
);
8883 /* Must be called with tasklist_lock held */
8884 static inline int tg_has_rt_tasks(struct task_group
*tg
)
8886 struct task_struct
*g
, *p
;
8888 do_each_thread(g
, p
) {
8889 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
8891 } while_each_thread(g
, p
);
8896 struct rt_schedulable_data
{
8897 struct task_group
*tg
;
8902 static int tg_schedulable(struct task_group
*tg
, void *data
)
8904 struct rt_schedulable_data
*d
= data
;
8905 struct task_group
*child
;
8906 unsigned long total
, sum
= 0;
8907 u64 period
, runtime
;
8909 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
8910 runtime
= tg
->rt_bandwidth
.rt_runtime
;
8913 period
= d
->rt_period
;
8914 runtime
= d
->rt_runtime
;
8918 * Cannot have more runtime than the period.
8920 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
8924 * Ensure we don't starve existing RT tasks.
8926 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
8929 total
= to_ratio(period
, runtime
);
8932 * Nobody can have more than the global setting allows.
8934 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
8938 * The sum of our children's runtime should not exceed our own.
8940 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
8941 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
8942 runtime
= child
->rt_bandwidth
.rt_runtime
;
8944 if (child
== d
->tg
) {
8945 period
= d
->rt_period
;
8946 runtime
= d
->rt_runtime
;
8949 sum
+= to_ratio(period
, runtime
);
8958 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
8960 struct rt_schedulable_data data
= {
8962 .rt_period
= period
,
8963 .rt_runtime
= runtime
,
8966 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
8969 static int tg_set_bandwidth(struct task_group
*tg
,
8970 u64 rt_period
, u64 rt_runtime
)
8974 mutex_lock(&rt_constraints_mutex
);
8975 read_lock(&tasklist_lock
);
8976 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
8980 spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8981 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
8982 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
8984 for_each_possible_cpu(i
) {
8985 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
8987 spin_lock(&rt_rq
->rt_runtime_lock
);
8988 rt_rq
->rt_runtime
= rt_runtime
;
8989 spin_unlock(&rt_rq
->rt_runtime_lock
);
8991 spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
8993 read_unlock(&tasklist_lock
);
8994 mutex_unlock(&rt_constraints_mutex
);
8999 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
9001 u64 rt_runtime
, rt_period
;
9003 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9004 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
9005 if (rt_runtime_us
< 0)
9006 rt_runtime
= RUNTIME_INF
;
9008 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9011 long sched_group_rt_runtime(struct task_group
*tg
)
9015 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
9018 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
9019 do_div(rt_runtime_us
, NSEC_PER_USEC
);
9020 return rt_runtime_us
;
9023 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
9025 u64 rt_runtime
, rt_period
;
9027 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
9028 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9033 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
9036 long sched_group_rt_period(struct task_group
*tg
)
9040 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
9041 do_div(rt_period_us
, NSEC_PER_USEC
);
9042 return rt_period_us
;
9045 static int sched_rt_global_constraints(void)
9047 u64 runtime
, period
;
9050 if (sysctl_sched_rt_period
<= 0)
9053 runtime
= global_rt_runtime();
9054 period
= global_rt_period();
9057 * Sanity check on the sysctl variables.
9059 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
9062 mutex_lock(&rt_constraints_mutex
);
9063 read_lock(&tasklist_lock
);
9064 ret
= __rt_schedulable(NULL
, 0, 0);
9065 read_unlock(&tasklist_lock
);
9066 mutex_unlock(&rt_constraints_mutex
);
9070 #else /* !CONFIG_RT_GROUP_SCHED */
9071 static int sched_rt_global_constraints(void)
9073 unsigned long flags
;
9076 if (sysctl_sched_rt_period
<= 0)
9079 spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9080 for_each_possible_cpu(i
) {
9081 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
9083 spin_lock(&rt_rq
->rt_runtime_lock
);
9084 rt_rq
->rt_runtime
= global_rt_runtime();
9085 spin_unlock(&rt_rq
->rt_runtime_lock
);
9087 spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
9091 #endif /* CONFIG_RT_GROUP_SCHED */
9093 int sched_rt_handler(struct ctl_table
*table
, int write
,
9094 struct file
*filp
, void __user
*buffer
, size_t *lenp
,
9098 int old_period
, old_runtime
;
9099 static DEFINE_MUTEX(mutex
);
9102 old_period
= sysctl_sched_rt_period
;
9103 old_runtime
= sysctl_sched_rt_runtime
;
9105 ret
= proc_dointvec(table
, write
, filp
, buffer
, lenp
, ppos
);
9107 if (!ret
&& write
) {
9108 ret
= sched_rt_global_constraints();
9110 sysctl_sched_rt_period
= old_period
;
9111 sysctl_sched_rt_runtime
= old_runtime
;
9113 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
9114 def_rt_bandwidth
.rt_period
=
9115 ns_to_ktime(global_rt_period());
9118 mutex_unlock(&mutex
);
9123 #ifdef CONFIG_CGROUP_SCHED
9125 /* return corresponding task_group object of a cgroup */
9126 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
9128 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
9129 struct task_group
, css
);
9132 static struct cgroup_subsys_state
*
9133 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9135 struct task_group
*tg
, *parent
;
9137 if (!cgrp
->parent
) {
9138 /* This is early initialization for the top cgroup */
9139 return &init_task_group
.css
;
9142 parent
= cgroup_tg(cgrp
->parent
);
9143 tg
= sched_create_group(parent
);
9145 return ERR_PTR(-ENOMEM
);
9151 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9153 struct task_group
*tg
= cgroup_tg(cgrp
);
9155 sched_destroy_group(tg
);
9159 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9160 struct task_struct
*tsk
)
9162 #ifdef CONFIG_RT_GROUP_SCHED
9163 /* Don't accept realtime tasks when there is no way for them to run */
9164 if (rt_task(tsk
) && cgroup_tg(cgrp
)->rt_bandwidth
.rt_runtime
== 0)
9167 /* We don't support RT-tasks being in separate groups */
9168 if (tsk
->sched_class
!= &fair_sched_class
)
9176 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
9177 struct cgroup
*old_cont
, struct task_struct
*tsk
)
9179 sched_move_task(tsk
);
9182 #ifdef CONFIG_FAIR_GROUP_SCHED
9183 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
9186 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
9189 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
9191 struct task_group
*tg
= cgroup_tg(cgrp
);
9193 return (u64
) tg
->shares
;
9195 #endif /* CONFIG_FAIR_GROUP_SCHED */
9197 #ifdef CONFIG_RT_GROUP_SCHED
9198 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
9201 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
9204 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9206 return sched_group_rt_runtime(cgroup_tg(cgrp
));
9209 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
9212 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
9215 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
9217 return sched_group_rt_period(cgroup_tg(cgrp
));
9219 #endif /* CONFIG_RT_GROUP_SCHED */
9221 static struct cftype cpu_files
[] = {
9222 #ifdef CONFIG_FAIR_GROUP_SCHED
9225 .read_u64
= cpu_shares_read_u64
,
9226 .write_u64
= cpu_shares_write_u64
,
9229 #ifdef CONFIG_RT_GROUP_SCHED
9231 .name
= "rt_runtime_us",
9232 .read_s64
= cpu_rt_runtime_read
,
9233 .write_s64
= cpu_rt_runtime_write
,
9236 .name
= "rt_period_us",
9237 .read_u64
= cpu_rt_period_read_uint
,
9238 .write_u64
= cpu_rt_period_write_uint
,
9243 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
9245 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
9248 struct cgroup_subsys cpu_cgroup_subsys
= {
9250 .create
= cpu_cgroup_create
,
9251 .destroy
= cpu_cgroup_destroy
,
9252 .can_attach
= cpu_cgroup_can_attach
,
9253 .attach
= cpu_cgroup_attach
,
9254 .populate
= cpu_cgroup_populate
,
9255 .subsys_id
= cpu_cgroup_subsys_id
,
9259 #endif /* CONFIG_CGROUP_SCHED */
9261 #ifdef CONFIG_CGROUP_CPUACCT
9264 * CPU accounting code for task groups.
9266 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
9267 * (balbir@in.ibm.com).
9270 /* track cpu usage of a group of tasks */
9272 struct cgroup_subsys_state css
;
9273 /* cpuusage holds pointer to a u64-type object on every cpu */
9277 struct cgroup_subsys cpuacct_subsys
;
9279 /* return cpu accounting group corresponding to this container */
9280 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
9282 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
9283 struct cpuacct
, css
);
9286 /* return cpu accounting group to which this task belongs */
9287 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
9289 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
9290 struct cpuacct
, css
);
9293 /* create a new cpu accounting group */
9294 static struct cgroup_subsys_state
*cpuacct_create(
9295 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9297 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
9300 return ERR_PTR(-ENOMEM
);
9302 ca
->cpuusage
= alloc_percpu(u64
);
9303 if (!ca
->cpuusage
) {
9305 return ERR_PTR(-ENOMEM
);
9311 /* destroy an existing cpu accounting group */
9313 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9315 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9317 free_percpu(ca
->cpuusage
);
9321 /* return total cpu usage (in nanoseconds) of a group */
9322 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
9324 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9325 u64 totalcpuusage
= 0;
9328 for_each_possible_cpu(i
) {
9329 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9332 * Take rq->lock to make 64-bit addition safe on 32-bit
9335 spin_lock_irq(&cpu_rq(i
)->lock
);
9336 totalcpuusage
+= *cpuusage
;
9337 spin_unlock_irq(&cpu_rq(i
)->lock
);
9340 return totalcpuusage
;
9343 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
9346 struct cpuacct
*ca
= cgroup_ca(cgrp
);
9355 for_each_possible_cpu(i
) {
9356 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, i
);
9358 spin_lock_irq(&cpu_rq(i
)->lock
);
9360 spin_unlock_irq(&cpu_rq(i
)->lock
);
9366 static struct cftype files
[] = {
9369 .read_u64
= cpuusage_read
,
9370 .write_u64
= cpuusage_write
,
9374 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
9376 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
9380 * charge this task's execution time to its accounting group.
9382 * called with rq->lock held.
9384 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
9388 if (!cpuacct_subsys
.active
)
9393 u64
*cpuusage
= percpu_ptr(ca
->cpuusage
, task_cpu(tsk
));
9395 *cpuusage
+= cputime
;
9399 struct cgroup_subsys cpuacct_subsys
= {
9401 .create
= cpuacct_create
,
9402 .destroy
= cpuacct_destroy
,
9403 .populate
= cpuacct_populate
,
9404 .subsys_id
= cpuacct_subsys_id
,
9406 #endif /* CONFIG_CGROUP_CPUACCT */