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
29 #define pr_fmt(fmt) KBUILD_MODNAME ": " fmt
32 #include <linux/module.h>
33 #include <linux/nmi.h>
34 #include <linux/init.h>
35 #include <linux/uaccess.h>
36 #include <linux/highmem.h>
37 #include <linux/smp_lock.h>
38 #include <asm/mmu_context.h>
39 #include <linux/interrupt.h>
40 #include <linux/capability.h>
41 #include <linux/completion.h>
42 #include <linux/kernel_stat.h>
43 #include <linux/debug_locks.h>
44 #include <linux/perf_event.h>
45 #include <linux/security.h>
46 #include <linux/notifier.h>
47 #include <linux/profile.h>
48 #include <linux/freezer.h>
49 #include <linux/vmalloc.h>
50 #include <linux/blkdev.h>
51 #include <linux/delay.h>
52 #include <linux/pid_namespace.h>
53 #include <linux/smp.h>
54 #include <linux/threads.h>
55 #include <linux/timer.h>
56 #include <linux/rcupdate.h>
57 #include <linux/cpu.h>
58 #include <linux/cpuset.h>
59 #include <linux/percpu.h>
60 #include <linux/kthread.h>
61 #include <linux/proc_fs.h>
62 #include <linux/seq_file.h>
63 #include <linux/sysctl.h>
64 #include <linux/syscalls.h>
65 #include <linux/times.h>
66 #include <linux/tsacct_kern.h>
67 #include <linux/kprobes.h>
68 #include <linux/delayacct.h>
69 #include <linux/unistd.h>
70 #include <linux/pagemap.h>
71 #include <linux/hrtimer.h>
72 #include <linux/tick.h>
73 #include <linux/debugfs.h>
74 #include <linux/ctype.h>
75 #include <linux/ftrace.h>
78 #include <asm/irq_regs.h>
80 #include "sched_cpupri.h"
82 #define CREATE_TRACE_POINTS
83 #include <trace/events/sched.h>
86 * Convert user-nice values [ -20 ... 0 ... 19 ]
87 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
90 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
91 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
92 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
95 * 'User priority' is the nice value converted to something we
96 * can work with better when scaling various scheduler parameters,
97 * it's a [ 0 ... 39 ] range.
99 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
100 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
101 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
104 * Helpers for converting nanosecond timing to jiffy resolution
106 #define NS_TO_JIFFIES(TIME) ((unsigned long)(TIME) / (NSEC_PER_SEC / HZ))
108 #define NICE_0_LOAD SCHED_LOAD_SCALE
109 #define NICE_0_SHIFT SCHED_LOAD_SHIFT
112 * These are the 'tuning knobs' of the scheduler:
114 * default timeslice is 100 msecs (used only for SCHED_RR tasks).
115 * Timeslices get refilled after they expire.
117 #define DEF_TIMESLICE (100 * HZ / 1000)
120 * single value that denotes runtime == period, ie unlimited time.
122 #define RUNTIME_INF ((u64)~0ULL)
124 static inline int rt_policy(int policy
)
126 if (unlikely(policy
== SCHED_FIFO
|| policy
== SCHED_RR
))
131 static inline int task_has_rt_policy(struct task_struct
*p
)
133 return rt_policy(p
->policy
);
137 * This is the priority-queue data structure of the RT scheduling class:
139 struct rt_prio_array
{
140 DECLARE_BITMAP(bitmap
, MAX_RT_PRIO
+1); /* include 1 bit for delimiter */
141 struct list_head queue
[MAX_RT_PRIO
];
144 struct rt_bandwidth
{
145 /* nests inside the rq lock: */
146 raw_spinlock_t rt_runtime_lock
;
149 struct hrtimer rt_period_timer
;
152 static struct rt_bandwidth def_rt_bandwidth
;
154 static int do_sched_rt_period_timer(struct rt_bandwidth
*rt_b
, int overrun
);
156 static enum hrtimer_restart
sched_rt_period_timer(struct hrtimer
*timer
)
158 struct rt_bandwidth
*rt_b
=
159 container_of(timer
, struct rt_bandwidth
, rt_period_timer
);
165 now
= hrtimer_cb_get_time(timer
);
166 overrun
= hrtimer_forward(timer
, now
, rt_b
->rt_period
);
171 idle
= do_sched_rt_period_timer(rt_b
, overrun
);
174 return idle
? HRTIMER_NORESTART
: HRTIMER_RESTART
;
178 void init_rt_bandwidth(struct rt_bandwidth
*rt_b
, u64 period
, u64 runtime
)
180 rt_b
->rt_period
= ns_to_ktime(period
);
181 rt_b
->rt_runtime
= runtime
;
183 raw_spin_lock_init(&rt_b
->rt_runtime_lock
);
185 hrtimer_init(&rt_b
->rt_period_timer
,
186 CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
187 rt_b
->rt_period_timer
.function
= sched_rt_period_timer
;
190 static inline int rt_bandwidth_enabled(void)
192 return sysctl_sched_rt_runtime
>= 0;
195 static void start_rt_bandwidth(struct rt_bandwidth
*rt_b
)
199 if (!rt_bandwidth_enabled() || rt_b
->rt_runtime
== RUNTIME_INF
)
202 if (hrtimer_active(&rt_b
->rt_period_timer
))
205 raw_spin_lock(&rt_b
->rt_runtime_lock
);
210 if (hrtimer_active(&rt_b
->rt_period_timer
))
213 now
= hrtimer_cb_get_time(&rt_b
->rt_period_timer
);
214 hrtimer_forward(&rt_b
->rt_period_timer
, now
, rt_b
->rt_period
);
216 soft
= hrtimer_get_softexpires(&rt_b
->rt_period_timer
);
217 hard
= hrtimer_get_expires(&rt_b
->rt_period_timer
);
218 delta
= ktime_to_ns(ktime_sub(hard
, soft
));
219 __hrtimer_start_range_ns(&rt_b
->rt_period_timer
, soft
, delta
,
220 HRTIMER_MODE_ABS_PINNED
, 0);
222 raw_spin_unlock(&rt_b
->rt_runtime_lock
);
225 #ifdef CONFIG_RT_GROUP_SCHED
226 static void destroy_rt_bandwidth(struct rt_bandwidth
*rt_b
)
228 hrtimer_cancel(&rt_b
->rt_period_timer
);
233 * sched_domains_mutex serializes calls to arch_init_sched_domains,
234 * detach_destroy_domains and partition_sched_domains.
236 static DEFINE_MUTEX(sched_domains_mutex
);
238 #ifdef CONFIG_GROUP_SCHED
240 #include <linux/cgroup.h>
244 static LIST_HEAD(task_groups
);
246 /* task group related information */
248 #ifdef CONFIG_CGROUP_SCHED
249 struct cgroup_subsys_state css
;
252 #ifdef CONFIG_USER_SCHED
256 #ifdef CONFIG_FAIR_GROUP_SCHED
257 /* schedulable entities of this group on each cpu */
258 struct sched_entity
**se
;
259 /* runqueue "owned" by this group on each cpu */
260 struct cfs_rq
**cfs_rq
;
261 unsigned long shares
;
264 #ifdef CONFIG_RT_GROUP_SCHED
265 struct sched_rt_entity
**rt_se
;
266 struct rt_rq
**rt_rq
;
268 struct rt_bandwidth rt_bandwidth
;
272 struct list_head list
;
274 struct task_group
*parent
;
275 struct list_head siblings
;
276 struct list_head children
;
279 #ifdef CONFIG_USER_SCHED
281 /* Helper function to pass uid information to create_sched_user() */
282 void set_tg_uid(struct user_struct
*user
)
284 user
->tg
->uid
= user
->uid
;
289 * Every UID task group (including init_task_group aka UID-0) will
290 * be a child to this group.
292 struct task_group root_task_group
;
294 #ifdef CONFIG_FAIR_GROUP_SCHED
295 /* Default task group's sched entity on each cpu */
296 static DEFINE_PER_CPU(struct sched_entity
, init_sched_entity
);
297 /* Default task group's cfs_rq on each cpu */
298 static DEFINE_PER_CPU_SHARED_ALIGNED(struct cfs_rq
, init_tg_cfs_rq
);
299 #endif /* CONFIG_FAIR_GROUP_SCHED */
301 #ifdef CONFIG_RT_GROUP_SCHED
302 static DEFINE_PER_CPU(struct sched_rt_entity
, init_sched_rt_entity
);
303 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rt_rq
, init_rt_rq_var
);
304 #endif /* CONFIG_RT_GROUP_SCHED */
305 #else /* !CONFIG_USER_SCHED */
306 #define root_task_group init_task_group
307 #endif /* CONFIG_USER_SCHED */
309 /* task_group_lock serializes add/remove of task groups and also changes to
310 * a task group's cpu shares.
312 static DEFINE_SPINLOCK(task_group_lock
);
314 #ifdef CONFIG_FAIR_GROUP_SCHED
317 static int root_task_group_empty(void)
319 return list_empty(&root_task_group
.children
);
323 #ifdef CONFIG_USER_SCHED
324 # define INIT_TASK_GROUP_LOAD (2*NICE_0_LOAD)
325 #else /* !CONFIG_USER_SCHED */
326 # define INIT_TASK_GROUP_LOAD NICE_0_LOAD
327 #endif /* CONFIG_USER_SCHED */
330 * A weight of 0 or 1 can cause arithmetics problems.
331 * A weight of a cfs_rq is the sum of weights of which entities
332 * are queued on this cfs_rq, so a weight of a entity should not be
333 * too large, so as the shares value of a task group.
334 * (The default weight is 1024 - so there's no practical
335 * limitation from this.)
338 #define MAX_SHARES (1UL << 18)
340 static int init_task_group_load
= INIT_TASK_GROUP_LOAD
;
343 /* Default task group.
344 * Every task in system belong to this group at bootup.
346 struct task_group init_task_group
;
348 /* return group to which a task belongs */
349 static inline struct task_group
*task_group(struct task_struct
*p
)
351 struct task_group
*tg
;
353 #ifdef CONFIG_USER_SCHED
355 tg
= __task_cred(p
)->user
->tg
;
357 #elif defined(CONFIG_CGROUP_SCHED)
358 tg
= container_of(task_subsys_state(p
, cpu_cgroup_subsys_id
),
359 struct task_group
, css
);
361 tg
= &init_task_group
;
366 /* Change a task's cfs_rq and parent entity if it moves across CPUs/groups */
367 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
)
369 #ifdef CONFIG_FAIR_GROUP_SCHED
370 p
->se
.cfs_rq
= task_group(p
)->cfs_rq
[cpu
];
371 p
->se
.parent
= task_group(p
)->se
[cpu
];
374 #ifdef CONFIG_RT_GROUP_SCHED
375 p
->rt
.rt_rq
= task_group(p
)->rt_rq
[cpu
];
376 p
->rt
.parent
= task_group(p
)->rt_se
[cpu
];
382 static inline void set_task_rq(struct task_struct
*p
, unsigned int cpu
) { }
383 static inline struct task_group
*task_group(struct task_struct
*p
)
388 #endif /* CONFIG_GROUP_SCHED */
390 /* CFS-related fields in a runqueue */
392 struct load_weight load
;
393 unsigned long nr_running
;
398 struct rb_root tasks_timeline
;
399 struct rb_node
*rb_leftmost
;
401 struct list_head tasks
;
402 struct list_head
*balance_iterator
;
405 * 'curr' points to currently running entity on this cfs_rq.
406 * It is set to NULL otherwise (i.e when none are currently running).
408 struct sched_entity
*curr
, *next
, *last
;
410 unsigned int nr_spread_over
;
412 #ifdef CONFIG_FAIR_GROUP_SCHED
413 struct rq
*rq
; /* cpu runqueue to which this cfs_rq is attached */
416 * leaf cfs_rqs are those that hold tasks (lowest schedulable entity in
417 * a hierarchy). Non-leaf lrqs hold other higher schedulable entities
418 * (like users, containers etc.)
420 * leaf_cfs_rq_list ties together list of leaf cfs_rq's in a cpu. This
421 * list is used during load balance.
423 struct list_head leaf_cfs_rq_list
;
424 struct task_group
*tg
; /* group that "owns" this runqueue */
428 * the part of load.weight contributed by tasks
430 unsigned long task_weight
;
433 * h_load = weight * f(tg)
435 * Where f(tg) is the recursive weight fraction assigned to
438 unsigned long h_load
;
441 * this cpu's part of tg->shares
443 unsigned long shares
;
446 * load.weight at the time we set shares
448 unsigned long rq_weight
;
453 /* Real-Time classes' related field in a runqueue: */
455 struct rt_prio_array active
;
456 unsigned long rt_nr_running
;
457 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
459 int curr
; /* highest queued rt task prio */
461 int next
; /* next highest */
466 unsigned long rt_nr_migratory
;
467 unsigned long rt_nr_total
;
469 struct plist_head pushable_tasks
;
474 /* Nests inside the rq lock: */
475 raw_spinlock_t rt_runtime_lock
;
477 #ifdef CONFIG_RT_GROUP_SCHED
478 unsigned long rt_nr_boosted
;
481 struct list_head leaf_rt_rq_list
;
482 struct task_group
*tg
;
483 struct sched_rt_entity
*rt_se
;
490 * We add the notion of a root-domain which will be used to define per-domain
491 * variables. Each exclusive cpuset essentially defines an island domain by
492 * fully partitioning the member cpus from any other cpuset. Whenever a new
493 * exclusive cpuset is created, we also create and attach a new root-domain
500 cpumask_var_t online
;
503 * The "RT overload" flag: it gets set if a CPU has more than
504 * one runnable RT task.
506 cpumask_var_t rto_mask
;
509 struct cpupri cpupri
;
514 * By default the system creates a single root-domain with all cpus as
515 * members (mimicking the global state we have today).
517 static struct root_domain def_root_domain
;
522 * This is the main, per-CPU runqueue data structure.
524 * Locking rule: those places that want to lock multiple runqueues
525 * (such as the load balancing or the thread migration code), lock
526 * acquire operations must be ordered by ascending &runqueue.
533 * nr_running and cpu_load should be in the same cacheline because
534 * remote CPUs use both these fields when doing load calculation.
536 unsigned long nr_running
;
537 #define CPU_LOAD_IDX_MAX 5
538 unsigned long cpu_load
[CPU_LOAD_IDX_MAX
];
540 unsigned char in_nohz_recently
;
542 /* capture load from *all* tasks on this cpu: */
543 struct load_weight load
;
544 unsigned long nr_load_updates
;
550 #ifdef CONFIG_FAIR_GROUP_SCHED
551 /* list of leaf cfs_rq on this cpu: */
552 struct list_head leaf_cfs_rq_list
;
554 #ifdef CONFIG_RT_GROUP_SCHED
555 struct list_head leaf_rt_rq_list
;
559 * This is part of a global counter where only the total sum
560 * over all CPUs matters. A task can increase this counter on
561 * one CPU and if it got migrated afterwards it may decrease
562 * it on another CPU. Always updated under the runqueue lock:
564 unsigned long nr_uninterruptible
;
566 struct task_struct
*curr
, *idle
;
567 unsigned long next_balance
;
568 struct mm_struct
*prev_mm
;
575 struct root_domain
*rd
;
576 struct sched_domain
*sd
;
578 unsigned char idle_at_tick
;
579 /* For active balancing */
583 /* cpu of this runqueue: */
587 unsigned long avg_load_per_task
;
589 struct task_struct
*migration_thread
;
590 struct list_head migration_queue
;
598 /* calc_load related fields */
599 unsigned long calc_load_update
;
600 long calc_load_active
;
602 #ifdef CONFIG_SCHED_HRTICK
604 int hrtick_csd_pending
;
605 struct call_single_data hrtick_csd
;
607 struct hrtimer hrtick_timer
;
610 #ifdef CONFIG_SCHEDSTATS
612 struct sched_info rq_sched_info
;
613 unsigned long long rq_cpu_time
;
614 /* could above be rq->cfs_rq.exec_clock + rq->rt_rq.rt_runtime ? */
616 /* sys_sched_yield() stats */
617 unsigned int yld_count
;
619 /* schedule() stats */
620 unsigned int sched_switch
;
621 unsigned int sched_count
;
622 unsigned int sched_goidle
;
624 /* try_to_wake_up() stats */
625 unsigned int ttwu_count
;
626 unsigned int ttwu_local
;
629 unsigned int bkl_count
;
633 static DEFINE_PER_CPU_SHARED_ALIGNED(struct rq
, runqueues
);
636 void check_preempt_curr(struct rq
*rq
, struct task_struct
*p
, int flags
)
638 rq
->curr
->sched_class
->check_preempt_curr(rq
, p
, flags
);
641 static inline int cpu_of(struct rq
*rq
)
651 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
652 * See detach_destroy_domains: synchronize_sched for details.
654 * The domain tree of any CPU may only be accessed from within
655 * preempt-disabled sections.
657 #define for_each_domain(cpu, __sd) \
658 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
660 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
661 #define this_rq() (&__get_cpu_var(runqueues))
662 #define task_rq(p) cpu_rq(task_cpu(p))
663 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
664 #define raw_rq() (&__raw_get_cpu_var(runqueues))
666 inline void update_rq_clock(struct rq
*rq
)
668 rq
->clock
= sched_clock_cpu(cpu_of(rq
));
672 * Tunables that become constants when CONFIG_SCHED_DEBUG is off:
674 #ifdef CONFIG_SCHED_DEBUG
675 # define const_debug __read_mostly
677 # define const_debug static const
682 * @cpu: the processor in question.
684 * Returns true if the current cpu runqueue is locked.
685 * This interface allows printk to be called with the runqueue lock
686 * held and know whether or not it is OK to wake up the klogd.
688 int runqueue_is_locked(int cpu
)
690 return raw_spin_is_locked(&cpu_rq(cpu
)->lock
);
694 * Debugging: various feature bits
697 #define SCHED_FEAT(name, enabled) \
698 __SCHED_FEAT_##name ,
701 #include "sched_features.h"
706 #define SCHED_FEAT(name, enabled) \
707 (1UL << __SCHED_FEAT_##name) * enabled |
709 const_debug
unsigned int sysctl_sched_features
=
710 #include "sched_features.h"
715 #ifdef CONFIG_SCHED_DEBUG
716 #define SCHED_FEAT(name, enabled) \
719 static __read_mostly
char *sched_feat_names
[] = {
720 #include "sched_features.h"
726 static int sched_feat_show(struct seq_file
*m
, void *v
)
730 for (i
= 0; sched_feat_names
[i
]; i
++) {
731 if (!(sysctl_sched_features
& (1UL << i
)))
733 seq_printf(m
, "%s ", sched_feat_names
[i
]);
741 sched_feat_write(struct file
*filp
, const char __user
*ubuf
,
742 size_t cnt
, loff_t
*ppos
)
752 if (copy_from_user(&buf
, ubuf
, cnt
))
757 if (strncmp(buf
, "NO_", 3) == 0) {
762 for (i
= 0; sched_feat_names
[i
]; i
++) {
763 int len
= strlen(sched_feat_names
[i
]);
765 if (strncmp(cmp
, sched_feat_names
[i
], len
) == 0) {
767 sysctl_sched_features
&= ~(1UL << i
);
769 sysctl_sched_features
|= (1UL << i
);
774 if (!sched_feat_names
[i
])
782 static int sched_feat_open(struct inode
*inode
, struct file
*filp
)
784 return single_open(filp
, sched_feat_show
, NULL
);
787 static const struct file_operations sched_feat_fops
= {
788 .open
= sched_feat_open
,
789 .write
= sched_feat_write
,
792 .release
= single_release
,
795 static __init
int sched_init_debug(void)
797 debugfs_create_file("sched_features", 0644, NULL
, NULL
,
802 late_initcall(sched_init_debug
);
806 #define sched_feat(x) (sysctl_sched_features & (1UL << __SCHED_FEAT_##x))
809 * Number of tasks to iterate in a single balance run.
810 * Limited because this is done with IRQs disabled.
812 const_debug
unsigned int sysctl_sched_nr_migrate
= 32;
815 * ratelimit for updating the group shares.
818 unsigned int sysctl_sched_shares_ratelimit
= 250000;
819 unsigned int normalized_sysctl_sched_shares_ratelimit
= 250000;
822 * Inject some fuzzyness into changing the per-cpu group shares
823 * this avoids remote rq-locks at the expense of fairness.
826 unsigned int sysctl_sched_shares_thresh
= 4;
829 * period over which we average the RT time consumption, measured
834 const_debug
unsigned int sysctl_sched_time_avg
= MSEC_PER_SEC
;
837 * period over which we measure -rt task cpu usage in us.
840 unsigned int sysctl_sched_rt_period
= 1000000;
842 static __read_mostly
int scheduler_running
;
845 * part of the period that we allow rt tasks to run in us.
848 int sysctl_sched_rt_runtime
= 950000;
850 static inline u64
global_rt_period(void)
852 return (u64
)sysctl_sched_rt_period
* NSEC_PER_USEC
;
855 static inline u64
global_rt_runtime(void)
857 if (sysctl_sched_rt_runtime
< 0)
860 return (u64
)sysctl_sched_rt_runtime
* NSEC_PER_USEC
;
863 #ifndef prepare_arch_switch
864 # define prepare_arch_switch(next) do { } while (0)
866 #ifndef finish_arch_switch
867 # define finish_arch_switch(prev) do { } while (0)
870 static inline int task_current(struct rq
*rq
, struct task_struct
*p
)
872 return rq
->curr
== p
;
875 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
876 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
878 return task_current(rq
, p
);
881 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
885 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
887 #ifdef CONFIG_DEBUG_SPINLOCK
888 /* this is a valid case when another task releases the spinlock */
889 rq
->lock
.owner
= current
;
892 * If we are tracking spinlock dependencies then we have to
893 * fix up the runqueue lock - which gets 'carried over' from
896 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
898 raw_spin_unlock_irq(&rq
->lock
);
901 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
902 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
907 return task_current(rq
, p
);
911 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
915 * We can optimise this out completely for !SMP, because the
916 * SMP rebalancing from interrupt is the only thing that cares
921 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
922 raw_spin_unlock_irq(&rq
->lock
);
924 raw_spin_unlock(&rq
->lock
);
928 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
932 * After ->oncpu is cleared, the task can be moved to a different CPU.
933 * We must ensure this doesn't happen until the switch is completely
939 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
943 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
946 * __task_rq_lock - lock the runqueue a given task resides on.
947 * Must be called interrupts disabled.
949 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
953 struct rq
*rq
= task_rq(p
);
954 raw_spin_lock(&rq
->lock
);
955 if (likely(rq
== task_rq(p
)))
957 raw_spin_unlock(&rq
->lock
);
962 * task_rq_lock - lock the runqueue a given task resides on and disable
963 * interrupts. Note the ordering: we can safely lookup the task_rq without
964 * explicitly disabling preemption.
966 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
972 local_irq_save(*flags
);
974 raw_spin_lock(&rq
->lock
);
975 if (likely(rq
== task_rq(p
)))
977 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
981 void task_rq_unlock_wait(struct task_struct
*p
)
983 struct rq
*rq
= task_rq(p
);
985 smp_mb(); /* spin-unlock-wait is not a full memory barrier */
986 raw_spin_unlock_wait(&rq
->lock
);
989 static void __task_rq_unlock(struct rq
*rq
)
992 raw_spin_unlock(&rq
->lock
);
995 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
998 raw_spin_unlock_irqrestore(&rq
->lock
, *flags
);
1002 * this_rq_lock - lock this runqueue and disable interrupts.
1004 static struct rq
*this_rq_lock(void)
1005 __acquires(rq
->lock
)
1009 local_irq_disable();
1011 raw_spin_lock(&rq
->lock
);
1016 #ifdef CONFIG_SCHED_HRTICK
1018 * Use HR-timers to deliver accurate preemption points.
1020 * Its all a bit involved since we cannot program an hrt while holding the
1021 * rq->lock. So what we do is store a state in in rq->hrtick_* and ask for a
1024 * When we get rescheduled we reprogram the hrtick_timer outside of the
1030 * - enabled by features
1031 * - hrtimer is actually high res
1033 static inline int hrtick_enabled(struct rq
*rq
)
1035 if (!sched_feat(HRTICK
))
1037 if (!cpu_active(cpu_of(rq
)))
1039 return hrtimer_is_hres_active(&rq
->hrtick_timer
);
1042 static void hrtick_clear(struct rq
*rq
)
1044 if (hrtimer_active(&rq
->hrtick_timer
))
1045 hrtimer_cancel(&rq
->hrtick_timer
);
1049 * High-resolution timer tick.
1050 * Runs from hardirq context with interrupts disabled.
1052 static enum hrtimer_restart
hrtick(struct hrtimer
*timer
)
1054 struct rq
*rq
= container_of(timer
, struct rq
, hrtick_timer
);
1056 WARN_ON_ONCE(cpu_of(rq
) != smp_processor_id());
1058 raw_spin_lock(&rq
->lock
);
1059 update_rq_clock(rq
);
1060 rq
->curr
->sched_class
->task_tick(rq
, rq
->curr
, 1);
1061 raw_spin_unlock(&rq
->lock
);
1063 return HRTIMER_NORESTART
;
1068 * called from hardirq (IPI) context
1070 static void __hrtick_start(void *arg
)
1072 struct rq
*rq
= arg
;
1074 raw_spin_lock(&rq
->lock
);
1075 hrtimer_restart(&rq
->hrtick_timer
);
1076 rq
->hrtick_csd_pending
= 0;
1077 raw_spin_unlock(&rq
->lock
);
1081 * Called to set the hrtick timer state.
1083 * called with rq->lock held and irqs disabled
1085 static void hrtick_start(struct rq
*rq
, u64 delay
)
1087 struct hrtimer
*timer
= &rq
->hrtick_timer
;
1088 ktime_t time
= ktime_add_ns(timer
->base
->get_time(), delay
);
1090 hrtimer_set_expires(timer
, time
);
1092 if (rq
== this_rq()) {
1093 hrtimer_restart(timer
);
1094 } else if (!rq
->hrtick_csd_pending
) {
1095 __smp_call_function_single(cpu_of(rq
), &rq
->hrtick_csd
, 0);
1096 rq
->hrtick_csd_pending
= 1;
1101 hotplug_hrtick(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
1103 int cpu
= (int)(long)hcpu
;
1106 case CPU_UP_CANCELED
:
1107 case CPU_UP_CANCELED_FROZEN
:
1108 case CPU_DOWN_PREPARE
:
1109 case CPU_DOWN_PREPARE_FROZEN
:
1111 case CPU_DEAD_FROZEN
:
1112 hrtick_clear(cpu_rq(cpu
));
1119 static __init
void init_hrtick(void)
1121 hotcpu_notifier(hotplug_hrtick
, 0);
1125 * Called to set the hrtick timer state.
1127 * called with rq->lock held and irqs disabled
1129 static void hrtick_start(struct rq
*rq
, u64 delay
)
1131 __hrtimer_start_range_ns(&rq
->hrtick_timer
, ns_to_ktime(delay
), 0,
1132 HRTIMER_MODE_REL_PINNED
, 0);
1135 static inline void init_hrtick(void)
1138 #endif /* CONFIG_SMP */
1140 static void init_rq_hrtick(struct rq
*rq
)
1143 rq
->hrtick_csd_pending
= 0;
1145 rq
->hrtick_csd
.flags
= 0;
1146 rq
->hrtick_csd
.func
= __hrtick_start
;
1147 rq
->hrtick_csd
.info
= rq
;
1150 hrtimer_init(&rq
->hrtick_timer
, CLOCK_MONOTONIC
, HRTIMER_MODE_REL
);
1151 rq
->hrtick_timer
.function
= hrtick
;
1153 #else /* CONFIG_SCHED_HRTICK */
1154 static inline void hrtick_clear(struct rq
*rq
)
1158 static inline void init_rq_hrtick(struct rq
*rq
)
1162 static inline void init_hrtick(void)
1165 #endif /* CONFIG_SCHED_HRTICK */
1168 * resched_task - mark a task 'to be rescheduled now'.
1170 * On UP this means the setting of the need_resched flag, on SMP it
1171 * might also involve a cross-CPU call to trigger the scheduler on
1176 #ifndef tsk_is_polling
1177 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1180 static void resched_task(struct task_struct
*p
)
1184 assert_raw_spin_locked(&task_rq(p
)->lock
);
1186 if (test_tsk_need_resched(p
))
1189 set_tsk_need_resched(p
);
1192 if (cpu
== smp_processor_id())
1195 /* NEED_RESCHED must be visible before we test polling */
1197 if (!tsk_is_polling(p
))
1198 smp_send_reschedule(cpu
);
1201 static void resched_cpu(int cpu
)
1203 struct rq
*rq
= cpu_rq(cpu
);
1204 unsigned long flags
;
1206 if (!raw_spin_trylock_irqsave(&rq
->lock
, flags
))
1208 resched_task(cpu_curr(cpu
));
1209 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1214 * When add_timer_on() enqueues a timer into the timer wheel of an
1215 * idle CPU then this timer might expire before the next timer event
1216 * which is scheduled to wake up that CPU. In case of a completely
1217 * idle system the next event might even be infinite time into the
1218 * future. wake_up_idle_cpu() ensures that the CPU is woken up and
1219 * leaves the inner idle loop so the newly added timer is taken into
1220 * account when the CPU goes back to idle and evaluates the timer
1221 * wheel for the next timer event.
1223 void wake_up_idle_cpu(int cpu
)
1225 struct rq
*rq
= cpu_rq(cpu
);
1227 if (cpu
== smp_processor_id())
1231 * This is safe, as this function is called with the timer
1232 * wheel base lock of (cpu) held. When the CPU is on the way
1233 * to idle and has not yet set rq->curr to idle then it will
1234 * be serialized on the timer wheel base lock and take the new
1235 * timer into account automatically.
1237 if (rq
->curr
!= rq
->idle
)
1241 * We can set TIF_RESCHED on the idle task of the other CPU
1242 * lockless. The worst case is that the other CPU runs the
1243 * idle task through an additional NOOP schedule()
1245 set_tsk_need_resched(rq
->idle
);
1247 /* NEED_RESCHED must be visible before we test polling */
1249 if (!tsk_is_polling(rq
->idle
))
1250 smp_send_reschedule(cpu
);
1252 #endif /* CONFIG_NO_HZ */
1254 static u64
sched_avg_period(void)
1256 return (u64
)sysctl_sched_time_avg
* NSEC_PER_MSEC
/ 2;
1259 static void sched_avg_update(struct rq
*rq
)
1261 s64 period
= sched_avg_period();
1263 while ((s64
)(rq
->clock
- rq
->age_stamp
) > period
) {
1264 rq
->age_stamp
+= period
;
1269 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1271 rq
->rt_avg
+= rt_delta
;
1272 sched_avg_update(rq
);
1275 #else /* !CONFIG_SMP */
1276 static void resched_task(struct task_struct
*p
)
1278 assert_raw_spin_locked(&task_rq(p
)->lock
);
1279 set_tsk_need_resched(p
);
1282 static void sched_rt_avg_update(struct rq
*rq
, u64 rt_delta
)
1285 #endif /* CONFIG_SMP */
1287 #if BITS_PER_LONG == 32
1288 # define WMULT_CONST (~0UL)
1290 # define WMULT_CONST (1UL << 32)
1293 #define WMULT_SHIFT 32
1296 * Shift right and round:
1298 #define SRR(x, y) (((x) + (1UL << ((y) - 1))) >> (y))
1301 * delta *= weight / lw
1303 static unsigned long
1304 calc_delta_mine(unsigned long delta_exec
, unsigned long weight
,
1305 struct load_weight
*lw
)
1309 if (!lw
->inv_weight
) {
1310 if (BITS_PER_LONG
> 32 && unlikely(lw
->weight
>= WMULT_CONST
))
1313 lw
->inv_weight
= 1 + (WMULT_CONST
-lw
->weight
/2)
1317 tmp
= (u64
)delta_exec
* weight
;
1319 * Check whether we'd overflow the 64-bit multiplication:
1321 if (unlikely(tmp
> WMULT_CONST
))
1322 tmp
= SRR(SRR(tmp
, WMULT_SHIFT
/2) * lw
->inv_weight
,
1325 tmp
= SRR(tmp
* lw
->inv_weight
, WMULT_SHIFT
);
1327 return (unsigned long)min(tmp
, (u64
)(unsigned long)LONG_MAX
);
1330 static inline void update_load_add(struct load_weight
*lw
, unsigned long inc
)
1336 static inline void update_load_sub(struct load_weight
*lw
, unsigned long dec
)
1343 * To aid in avoiding the subversion of "niceness" due to uneven distribution
1344 * of tasks with abnormal "nice" values across CPUs the contribution that
1345 * each task makes to its run queue's load is weighted according to its
1346 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
1347 * scaled version of the new time slice allocation that they receive on time
1351 #define WEIGHT_IDLEPRIO 3
1352 #define WMULT_IDLEPRIO 1431655765
1355 * Nice levels are multiplicative, with a gentle 10% change for every
1356 * nice level changed. I.e. when a CPU-bound task goes from nice 0 to
1357 * nice 1, it will get ~10% less CPU time than another CPU-bound task
1358 * that remained on nice 0.
1360 * The "10% effect" is relative and cumulative: from _any_ nice level,
1361 * if you go up 1 level, it's -10% CPU usage, if you go down 1 level
1362 * it's +10% CPU usage. (to achieve that we use a multiplier of 1.25.
1363 * If a task goes up by ~10% and another task goes down by ~10% then
1364 * the relative distance between them is ~25%.)
1366 static const int prio_to_weight
[40] = {
1367 /* -20 */ 88761, 71755, 56483, 46273, 36291,
1368 /* -15 */ 29154, 23254, 18705, 14949, 11916,
1369 /* -10 */ 9548, 7620, 6100, 4904, 3906,
1370 /* -5 */ 3121, 2501, 1991, 1586, 1277,
1371 /* 0 */ 1024, 820, 655, 526, 423,
1372 /* 5 */ 335, 272, 215, 172, 137,
1373 /* 10 */ 110, 87, 70, 56, 45,
1374 /* 15 */ 36, 29, 23, 18, 15,
1378 * Inverse (2^32/x) values of the prio_to_weight[] array, precalculated.
1380 * In cases where the weight does not change often, we can use the
1381 * precalculated inverse to speed up arithmetics by turning divisions
1382 * into multiplications:
1384 static const u32 prio_to_wmult
[40] = {
1385 /* -20 */ 48388, 59856, 76040, 92818, 118348,
1386 /* -15 */ 147320, 184698, 229616, 287308, 360437,
1387 /* -10 */ 449829, 563644, 704093, 875809, 1099582,
1388 /* -5 */ 1376151, 1717300, 2157191, 2708050, 3363326,
1389 /* 0 */ 4194304, 5237765, 6557202, 8165337, 10153587,
1390 /* 5 */ 12820798, 15790321, 19976592, 24970740, 31350126,
1391 /* 10 */ 39045157, 49367440, 61356676, 76695844, 95443717,
1392 /* 15 */ 119304647, 148102320, 186737708, 238609294, 286331153,
1395 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
);
1398 * runqueue iterator, to support SMP load-balancing between different
1399 * scheduling classes, without having to expose their internal data
1400 * structures to the load-balancing proper:
1402 struct rq_iterator
{
1404 struct task_struct
*(*start
)(void *);
1405 struct task_struct
*(*next
)(void *);
1409 static unsigned long
1410 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1411 unsigned long max_load_move
, struct sched_domain
*sd
,
1412 enum cpu_idle_type idle
, int *all_pinned
,
1413 int *this_best_prio
, struct rq_iterator
*iterator
);
1416 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
1417 struct sched_domain
*sd
, enum cpu_idle_type idle
,
1418 struct rq_iterator
*iterator
);
1421 /* Time spent by the tasks of the cpu accounting group executing in ... */
1422 enum cpuacct_stat_index
{
1423 CPUACCT_STAT_USER
, /* ... user mode */
1424 CPUACCT_STAT_SYSTEM
, /* ... kernel mode */
1426 CPUACCT_STAT_NSTATS
,
1429 #ifdef CONFIG_CGROUP_CPUACCT
1430 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
);
1431 static void cpuacct_update_stats(struct task_struct
*tsk
,
1432 enum cpuacct_stat_index idx
, cputime_t val
);
1434 static inline void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
) {}
1435 static inline void cpuacct_update_stats(struct task_struct
*tsk
,
1436 enum cpuacct_stat_index idx
, cputime_t val
) {}
1439 static inline void inc_cpu_load(struct rq
*rq
, unsigned long load
)
1441 update_load_add(&rq
->load
, load
);
1444 static inline void dec_cpu_load(struct rq
*rq
, unsigned long load
)
1446 update_load_sub(&rq
->load
, load
);
1449 #if (defined(CONFIG_SMP) && defined(CONFIG_FAIR_GROUP_SCHED)) || defined(CONFIG_RT_GROUP_SCHED)
1450 typedef int (*tg_visitor
)(struct task_group
*, void *);
1453 * Iterate the full tree, calling @down when first entering a node and @up when
1454 * leaving it for the final time.
1456 static int walk_tg_tree(tg_visitor down
, tg_visitor up
, void *data
)
1458 struct task_group
*parent
, *child
;
1462 parent
= &root_task_group
;
1464 ret
= (*down
)(parent
, data
);
1467 list_for_each_entry_rcu(child
, &parent
->children
, siblings
) {
1474 ret
= (*up
)(parent
, data
);
1479 parent
= parent
->parent
;
1488 static int tg_nop(struct task_group
*tg
, void *data
)
1495 /* Used instead of source_load when we know the type == 0 */
1496 static unsigned long weighted_cpuload(const int cpu
)
1498 return cpu_rq(cpu
)->load
.weight
;
1502 * Return a low guess at the load of a migration-source cpu weighted
1503 * according to the scheduling class and "nice" value.
1505 * We want to under-estimate the load of migration sources, to
1506 * balance conservatively.
1508 static unsigned long source_load(int cpu
, int type
)
1510 struct rq
*rq
= cpu_rq(cpu
);
1511 unsigned long total
= weighted_cpuload(cpu
);
1513 if (type
== 0 || !sched_feat(LB_BIAS
))
1516 return min(rq
->cpu_load
[type
-1], total
);
1520 * Return a high guess at the load of a migration-target cpu weighted
1521 * according to the scheduling class and "nice" value.
1523 static unsigned long target_load(int cpu
, int type
)
1525 struct rq
*rq
= cpu_rq(cpu
);
1526 unsigned long total
= weighted_cpuload(cpu
);
1528 if (type
== 0 || !sched_feat(LB_BIAS
))
1531 return max(rq
->cpu_load
[type
-1], total
);
1534 static struct sched_group
*group_of(int cpu
)
1536 struct sched_domain
*sd
= rcu_dereference(cpu_rq(cpu
)->sd
);
1544 static unsigned long power_of(int cpu
)
1546 struct sched_group
*group
= group_of(cpu
);
1549 return SCHED_LOAD_SCALE
;
1551 return group
->cpu_power
;
1554 static int task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
);
1556 static unsigned long cpu_avg_load_per_task(int cpu
)
1558 struct rq
*rq
= cpu_rq(cpu
);
1559 unsigned long nr_running
= ACCESS_ONCE(rq
->nr_running
);
1562 rq
->avg_load_per_task
= rq
->load
.weight
/ nr_running
;
1564 rq
->avg_load_per_task
= 0;
1566 return rq
->avg_load_per_task
;
1569 #ifdef CONFIG_FAIR_GROUP_SCHED
1571 static __read_mostly
unsigned long *update_shares_data
;
1573 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
);
1576 * Calculate and set the cpu's group shares.
1578 static void update_group_shares_cpu(struct task_group
*tg
, int cpu
,
1579 unsigned long sd_shares
,
1580 unsigned long sd_rq_weight
,
1581 unsigned long *usd_rq_weight
)
1583 unsigned long shares
, rq_weight
;
1586 rq_weight
= usd_rq_weight
[cpu
];
1589 rq_weight
= NICE_0_LOAD
;
1593 * \Sum_j shares_j * rq_weight_i
1594 * shares_i = -----------------------------
1595 * \Sum_j rq_weight_j
1597 shares
= (sd_shares
* rq_weight
) / sd_rq_weight
;
1598 shares
= clamp_t(unsigned long, shares
, MIN_SHARES
, MAX_SHARES
);
1600 if (abs(shares
- tg
->se
[cpu
]->load
.weight
) >
1601 sysctl_sched_shares_thresh
) {
1602 struct rq
*rq
= cpu_rq(cpu
);
1603 unsigned long flags
;
1605 raw_spin_lock_irqsave(&rq
->lock
, flags
);
1606 tg
->cfs_rq
[cpu
]->rq_weight
= boost
? 0 : rq_weight
;
1607 tg
->cfs_rq
[cpu
]->shares
= boost
? 0 : shares
;
1608 __set_se_shares(tg
->se
[cpu
], shares
);
1609 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
1614 * Re-compute the task group their per cpu shares over the given domain.
1615 * This needs to be done in a bottom-up fashion because the rq weight of a
1616 * parent group depends on the shares of its child groups.
1618 static int tg_shares_up(struct task_group
*tg
, void *data
)
1620 unsigned long weight
, rq_weight
= 0, sum_weight
= 0, shares
= 0;
1621 unsigned long *usd_rq_weight
;
1622 struct sched_domain
*sd
= data
;
1623 unsigned long flags
;
1629 local_irq_save(flags
);
1630 usd_rq_weight
= per_cpu_ptr(update_shares_data
, smp_processor_id());
1632 for_each_cpu(i
, sched_domain_span(sd
)) {
1633 weight
= tg
->cfs_rq
[i
]->load
.weight
;
1634 usd_rq_weight
[i
] = weight
;
1636 rq_weight
+= weight
;
1638 * If there are currently no tasks on the cpu pretend there
1639 * is one of average load so that when a new task gets to
1640 * run here it will not get delayed by group starvation.
1643 weight
= NICE_0_LOAD
;
1645 sum_weight
+= weight
;
1646 shares
+= tg
->cfs_rq
[i
]->shares
;
1650 rq_weight
= sum_weight
;
1652 if ((!shares
&& rq_weight
) || shares
> tg
->shares
)
1653 shares
= tg
->shares
;
1655 if (!sd
->parent
|| !(sd
->parent
->flags
& SD_LOAD_BALANCE
))
1656 shares
= tg
->shares
;
1658 for_each_cpu(i
, sched_domain_span(sd
))
1659 update_group_shares_cpu(tg
, i
, shares
, rq_weight
, usd_rq_weight
);
1661 local_irq_restore(flags
);
1667 * Compute the cpu's hierarchical load factor for each task group.
1668 * This needs to be done in a top-down fashion because the load of a child
1669 * group is a fraction of its parents load.
1671 static int tg_load_down(struct task_group
*tg
, void *data
)
1674 long cpu
= (long)data
;
1677 load
= cpu_rq(cpu
)->load
.weight
;
1679 load
= tg
->parent
->cfs_rq
[cpu
]->h_load
;
1680 load
*= tg
->cfs_rq
[cpu
]->shares
;
1681 load
/= tg
->parent
->cfs_rq
[cpu
]->load
.weight
+ 1;
1684 tg
->cfs_rq
[cpu
]->h_load
= load
;
1689 static void update_shares(struct sched_domain
*sd
)
1694 if (root_task_group_empty())
1697 now
= cpu_clock(raw_smp_processor_id());
1698 elapsed
= now
- sd
->last_update
;
1700 if (elapsed
>= (s64
)(u64
)sysctl_sched_shares_ratelimit
) {
1701 sd
->last_update
= now
;
1702 walk_tg_tree(tg_nop
, tg_shares_up
, sd
);
1706 static void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1708 if (root_task_group_empty())
1711 raw_spin_unlock(&rq
->lock
);
1713 raw_spin_lock(&rq
->lock
);
1716 static void update_h_load(long cpu
)
1718 if (root_task_group_empty())
1721 walk_tg_tree(tg_load_down
, tg_nop
, (void *)cpu
);
1726 static inline void update_shares(struct sched_domain
*sd
)
1730 static inline void update_shares_locked(struct rq
*rq
, struct sched_domain
*sd
)
1736 #ifdef CONFIG_PREEMPT
1738 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
);
1741 * fair double_lock_balance: Safely acquires both rq->locks in a fair
1742 * way at the expense of forcing extra atomic operations in all
1743 * invocations. This assures that the double_lock is acquired using the
1744 * same underlying policy as the spinlock_t on this architecture, which
1745 * reduces latency compared to the unfair variant below. However, it
1746 * also adds more overhead and therefore may reduce throughput.
1748 static inline int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1749 __releases(this_rq
->lock
)
1750 __acquires(busiest
->lock
)
1751 __acquires(this_rq
->lock
)
1753 raw_spin_unlock(&this_rq
->lock
);
1754 double_rq_lock(this_rq
, busiest
);
1761 * Unfair double_lock_balance: Optimizes throughput at the expense of
1762 * latency by eliminating extra atomic operations when the locks are
1763 * already in proper order on entry. This favors lower cpu-ids and will
1764 * grant the double lock to lower cpus over higher ids under contention,
1765 * regardless of entry order into the function.
1767 static int _double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1768 __releases(this_rq
->lock
)
1769 __acquires(busiest
->lock
)
1770 __acquires(this_rq
->lock
)
1774 if (unlikely(!raw_spin_trylock(&busiest
->lock
))) {
1775 if (busiest
< this_rq
) {
1776 raw_spin_unlock(&this_rq
->lock
);
1777 raw_spin_lock(&busiest
->lock
);
1778 raw_spin_lock_nested(&this_rq
->lock
,
1779 SINGLE_DEPTH_NESTING
);
1782 raw_spin_lock_nested(&busiest
->lock
,
1783 SINGLE_DEPTH_NESTING
);
1788 #endif /* CONFIG_PREEMPT */
1791 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
1793 static int double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1795 if (unlikely(!irqs_disabled())) {
1796 /* printk() doesn't work good under rq->lock */
1797 raw_spin_unlock(&this_rq
->lock
);
1801 return _double_lock_balance(this_rq
, busiest
);
1804 static inline void double_unlock_balance(struct rq
*this_rq
, struct rq
*busiest
)
1805 __releases(busiest
->lock
)
1807 raw_spin_unlock(&busiest
->lock
);
1808 lock_set_subclass(&this_rq
->lock
.dep_map
, 0, _RET_IP_
);
1812 #ifdef CONFIG_FAIR_GROUP_SCHED
1813 static void cfs_rq_set_shares(struct cfs_rq
*cfs_rq
, unsigned long shares
)
1816 cfs_rq
->shares
= shares
;
1821 static void calc_load_account_active(struct rq
*this_rq
);
1822 static void update_sysctl(void);
1823 static int get_update_sysctl_factor(void);
1825 static inline void __set_task_cpu(struct task_struct
*p
, unsigned int cpu
)
1827 set_task_rq(p
, cpu
);
1830 * After ->cpu is set up to a new value, task_rq_lock(p, ...) can be
1831 * successfuly executed on another CPU. We must ensure that updates of
1832 * per-task data have been completed by this moment.
1835 task_thread_info(p
)->cpu
= cpu
;
1839 #include "sched_stats.h"
1840 #include "sched_idletask.c"
1841 #include "sched_fair.c"
1842 #include "sched_rt.c"
1843 #ifdef CONFIG_SCHED_DEBUG
1844 # include "sched_debug.c"
1847 #define sched_class_highest (&rt_sched_class)
1848 #define for_each_class(class) \
1849 for (class = sched_class_highest; class; class = class->next)
1851 static void inc_nr_running(struct rq
*rq
)
1856 static void dec_nr_running(struct rq
*rq
)
1861 static void set_load_weight(struct task_struct
*p
)
1863 if (task_has_rt_policy(p
)) {
1864 p
->se
.load
.weight
= prio_to_weight
[0] * 2;
1865 p
->se
.load
.inv_weight
= prio_to_wmult
[0] >> 1;
1870 * SCHED_IDLE tasks get minimal weight:
1872 if (p
->policy
== SCHED_IDLE
) {
1873 p
->se
.load
.weight
= WEIGHT_IDLEPRIO
;
1874 p
->se
.load
.inv_weight
= WMULT_IDLEPRIO
;
1878 p
->se
.load
.weight
= prio_to_weight
[p
->static_prio
- MAX_RT_PRIO
];
1879 p
->se
.load
.inv_weight
= prio_to_wmult
[p
->static_prio
- MAX_RT_PRIO
];
1882 static void update_avg(u64
*avg
, u64 sample
)
1884 s64 diff
= sample
- *avg
;
1888 static void enqueue_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1891 p
->se
.start_runtime
= p
->se
.sum_exec_runtime
;
1893 sched_info_queued(p
);
1894 p
->sched_class
->enqueue_task(rq
, p
, wakeup
);
1898 static void dequeue_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1901 if (p
->se
.last_wakeup
) {
1902 update_avg(&p
->se
.avg_overlap
,
1903 p
->se
.sum_exec_runtime
- p
->se
.last_wakeup
);
1904 p
->se
.last_wakeup
= 0;
1906 update_avg(&p
->se
.avg_wakeup
,
1907 sysctl_sched_wakeup_granularity
);
1911 sched_info_dequeued(p
);
1912 p
->sched_class
->dequeue_task(rq
, p
, sleep
);
1917 * __normal_prio - return the priority that is based on the static prio
1919 static inline int __normal_prio(struct task_struct
*p
)
1921 return p
->static_prio
;
1925 * Calculate the expected normal priority: i.e. priority
1926 * without taking RT-inheritance into account. Might be
1927 * boosted by interactivity modifiers. Changes upon fork,
1928 * setprio syscalls, and whenever the interactivity
1929 * estimator recalculates.
1931 static inline int normal_prio(struct task_struct
*p
)
1935 if (task_has_rt_policy(p
))
1936 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
1938 prio
= __normal_prio(p
);
1943 * Calculate the current priority, i.e. the priority
1944 * taken into account by the scheduler. This value might
1945 * be boosted by RT tasks, or might be boosted by
1946 * interactivity modifiers. Will be RT if the task got
1947 * RT-boosted. If not then it returns p->normal_prio.
1949 static int effective_prio(struct task_struct
*p
)
1951 p
->normal_prio
= normal_prio(p
);
1953 * If we are RT tasks or we were boosted to RT priority,
1954 * keep the priority unchanged. Otherwise, update priority
1955 * to the normal priority:
1957 if (!rt_prio(p
->prio
))
1958 return p
->normal_prio
;
1963 * activate_task - move a task to the runqueue.
1965 static void activate_task(struct rq
*rq
, struct task_struct
*p
, int wakeup
)
1967 if (task_contributes_to_load(p
))
1968 rq
->nr_uninterruptible
--;
1970 enqueue_task(rq
, p
, wakeup
);
1975 * deactivate_task - remove a task from the runqueue.
1977 static void deactivate_task(struct rq
*rq
, struct task_struct
*p
, int sleep
)
1979 if (task_contributes_to_load(p
))
1980 rq
->nr_uninterruptible
++;
1982 dequeue_task(rq
, p
, sleep
);
1987 * task_curr - is this task currently executing on a CPU?
1988 * @p: the task in question.
1990 inline int task_curr(const struct task_struct
*p
)
1992 return cpu_curr(task_cpu(p
)) == p
;
1995 static inline void check_class_changed(struct rq
*rq
, struct task_struct
*p
,
1996 const struct sched_class
*prev_class
,
1997 int oldprio
, int running
)
1999 if (prev_class
!= p
->sched_class
) {
2000 if (prev_class
->switched_from
)
2001 prev_class
->switched_from(rq
, p
, running
);
2002 p
->sched_class
->switched_to(rq
, p
, running
);
2004 p
->sched_class
->prio_changed(rq
, p
, oldprio
, running
);
2008 * kthread_bind - bind a just-created kthread to a cpu.
2009 * @p: thread created by kthread_create().
2010 * @cpu: cpu (might not be online, must be possible) for @k to run on.
2012 * Description: This function is equivalent to set_cpus_allowed(),
2013 * except that @cpu doesn't need to be online, and the thread must be
2014 * stopped (i.e., just returned from kthread_create()).
2016 * Function lives here instead of kthread.c because it messes with
2017 * scheduler internals which require locking.
2019 void kthread_bind(struct task_struct
*p
, unsigned int cpu
)
2021 /* Must have done schedule() in kthread() before we set_task_cpu */
2022 if (!wait_task_inactive(p
, TASK_UNINTERRUPTIBLE
)) {
2027 p
->cpus_allowed
= cpumask_of_cpu(cpu
);
2028 p
->rt
.nr_cpus_allowed
= 1;
2029 p
->flags
|= PF_THREAD_BOUND
;
2031 EXPORT_SYMBOL(kthread_bind
);
2035 * Is this task likely cache-hot:
2038 task_hot(struct task_struct
*p
, u64 now
, struct sched_domain
*sd
)
2042 if (p
->sched_class
!= &fair_sched_class
)
2046 * Buddy candidates are cache hot:
2048 if (sched_feat(CACHE_HOT_BUDDY
) && this_rq()->nr_running
&&
2049 (&p
->se
== cfs_rq_of(&p
->se
)->next
||
2050 &p
->se
== cfs_rq_of(&p
->se
)->last
))
2053 if (sysctl_sched_migration_cost
== -1)
2055 if (sysctl_sched_migration_cost
== 0)
2058 delta
= now
- p
->se
.exec_start
;
2060 return delta
< (s64
)sysctl_sched_migration_cost
;
2064 void set_task_cpu(struct task_struct
*p
, unsigned int new_cpu
)
2066 int old_cpu
= task_cpu(p
);
2067 struct cfs_rq
*old_cfsrq
= task_cfs_rq(p
),
2068 *new_cfsrq
= cpu_cfs_rq(old_cfsrq
, new_cpu
);
2070 #ifdef CONFIG_SCHED_DEBUG
2072 * We should never call set_task_cpu() on a blocked task,
2073 * ttwu() will sort out the placement.
2075 WARN_ON(p
->state
!= TASK_RUNNING
&& p
->state
!= TASK_WAKING
);
2078 trace_sched_migrate_task(p
, new_cpu
);
2080 if (old_cpu
!= new_cpu
) {
2081 p
->se
.nr_migrations
++;
2082 perf_sw_event(PERF_COUNT_SW_CPU_MIGRATIONS
,
2085 p
->se
.vruntime
-= old_cfsrq
->min_vruntime
-
2086 new_cfsrq
->min_vruntime
;
2088 __set_task_cpu(p
, new_cpu
);
2091 struct migration_req
{
2092 struct list_head list
;
2094 struct task_struct
*task
;
2097 struct completion done
;
2101 * The task's runqueue lock must be held.
2102 * Returns true if you have to wait for migration thread.
2105 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
2107 struct rq
*rq
= task_rq(p
);
2110 * If the task is not on a runqueue (and not running), then
2111 * the next wake-up will properly place the task.
2113 if (!p
->se
.on_rq
&& !task_running(rq
, p
))
2116 init_completion(&req
->done
);
2118 req
->dest_cpu
= dest_cpu
;
2119 list_add(&req
->list
, &rq
->migration_queue
);
2125 * wait_task_context_switch - wait for a thread to complete at least one
2128 * @p must not be current.
2130 void wait_task_context_switch(struct task_struct
*p
)
2132 unsigned long nvcsw
, nivcsw
, flags
;
2140 * The runqueue is assigned before the actual context
2141 * switch. We need to take the runqueue lock.
2143 * We could check initially without the lock but it is
2144 * very likely that we need to take the lock in every
2147 rq
= task_rq_lock(p
, &flags
);
2148 running
= task_running(rq
, p
);
2149 task_rq_unlock(rq
, &flags
);
2151 if (likely(!running
))
2154 * The switch count is incremented before the actual
2155 * context switch. We thus wait for two switches to be
2156 * sure at least one completed.
2158 if ((p
->nvcsw
- nvcsw
) > 1)
2160 if ((p
->nivcsw
- nivcsw
) > 1)
2168 * wait_task_inactive - wait for a thread to unschedule.
2170 * If @match_state is nonzero, it's the @p->state value just checked and
2171 * not expected to change. If it changes, i.e. @p might have woken up,
2172 * then return zero. When we succeed in waiting for @p to be off its CPU,
2173 * we return a positive number (its total switch count). If a second call
2174 * a short while later returns the same number, the caller can be sure that
2175 * @p has remained unscheduled the whole time.
2177 * The caller must ensure that the task *will* unschedule sometime soon,
2178 * else this function might spin for a *long* time. This function can't
2179 * be called with interrupts off, or it may introduce deadlock with
2180 * smp_call_function() if an IPI is sent by the same process we are
2181 * waiting to become inactive.
2183 unsigned long wait_task_inactive(struct task_struct
*p
, long match_state
)
2185 unsigned long flags
;
2192 * We do the initial early heuristics without holding
2193 * any task-queue locks at all. We'll only try to get
2194 * the runqueue lock when things look like they will
2200 * If the task is actively running on another CPU
2201 * still, just relax and busy-wait without holding
2204 * NOTE! Since we don't hold any locks, it's not
2205 * even sure that "rq" stays as the right runqueue!
2206 * But we don't care, since "task_running()" will
2207 * return false if the runqueue has changed and p
2208 * is actually now running somewhere else!
2210 while (task_running(rq
, p
)) {
2211 if (match_state
&& unlikely(p
->state
!= match_state
))
2217 * Ok, time to look more closely! We need the rq
2218 * lock now, to be *sure*. If we're wrong, we'll
2219 * just go back and repeat.
2221 rq
= task_rq_lock(p
, &flags
);
2222 trace_sched_wait_task(rq
, p
);
2223 running
= task_running(rq
, p
);
2224 on_rq
= p
->se
.on_rq
;
2226 if (!match_state
|| p
->state
== match_state
)
2227 ncsw
= p
->nvcsw
| LONG_MIN
; /* sets MSB */
2228 task_rq_unlock(rq
, &flags
);
2231 * If it changed from the expected state, bail out now.
2233 if (unlikely(!ncsw
))
2237 * Was it really running after all now that we
2238 * checked with the proper locks actually held?
2240 * Oops. Go back and try again..
2242 if (unlikely(running
)) {
2248 * It's not enough that it's not actively running,
2249 * it must be off the runqueue _entirely_, and not
2252 * So if it was still runnable (but just not actively
2253 * running right now), it's preempted, and we should
2254 * yield - it could be a while.
2256 if (unlikely(on_rq
)) {
2257 schedule_timeout_uninterruptible(1);
2262 * Ahh, all good. It wasn't running, and it wasn't
2263 * runnable, which means that it will never become
2264 * running in the future either. We're all done!
2273 * kick_process - kick a running thread to enter/exit the kernel
2274 * @p: the to-be-kicked thread
2276 * Cause a process which is running on another CPU to enter
2277 * kernel-mode, without any delay. (to get signals handled.)
2279 * NOTE: this function doesnt have to take the runqueue lock,
2280 * because all it wants to ensure is that the remote task enters
2281 * the kernel. If the IPI races and the task has been migrated
2282 * to another CPU then no harm is done and the purpose has been
2285 void kick_process(struct task_struct
*p
)
2291 if ((cpu
!= smp_processor_id()) && task_curr(p
))
2292 smp_send_reschedule(cpu
);
2295 EXPORT_SYMBOL_GPL(kick_process
);
2296 #endif /* CONFIG_SMP */
2299 * task_oncpu_function_call - call a function on the cpu on which a task runs
2300 * @p: the task to evaluate
2301 * @func: the function to be called
2302 * @info: the function call argument
2304 * Calls the function @func when the task is currently running. This might
2305 * be on the current CPU, which just calls the function directly
2307 void task_oncpu_function_call(struct task_struct
*p
,
2308 void (*func
) (void *info
), void *info
)
2315 smp_call_function_single(cpu
, func
, info
, 1);
2323 * - fork, @p is stable because it isn't on the tasklist yet
2325 * - exec, @p is unstable XXX
2327 * - wake-up, we serialize ->cpus_allowed against TASK_WAKING so
2328 * we should be good.
2331 int select_task_rq(struct task_struct
*p
, int sd_flags
, int wake_flags
)
2333 int cpu
= p
->sched_class
->select_task_rq(p
, sd_flags
, wake_flags
);
2336 * In order not to call set_task_cpu() on a blocking task we need
2337 * to rely on ttwu() to place the task on a valid ->cpus_allowed
2340 * Since this is common to all placement strategies, this lives here.
2342 * [ this allows ->select_task() to simply return task_cpu(p) and
2343 * not worry about this generic constraint ]
2345 if (unlikely(!cpumask_test_cpu(cpu
, &p
->cpus_allowed
) ||
2346 !cpu_active(cpu
))) {
2348 cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
2350 * XXX: race against hot-plug modifying cpu_active_mask
2352 BUG_ON(cpu
>= nr_cpu_ids
);
2360 * try_to_wake_up - wake up a thread
2361 * @p: the to-be-woken-up thread
2362 * @state: the mask of task states that can be woken
2363 * @sync: do a synchronous wakeup?
2365 * Put it on the run-queue if it's not already there. The "current"
2366 * thread is always on the run-queue (except when the actual
2367 * re-schedule is in progress), and as such you're allowed to do
2368 * the simpler "current->state = TASK_RUNNING" to mark yourself
2369 * runnable without the overhead of this.
2371 * returns failure only if the task is already active.
2373 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
,
2376 int cpu
, orig_cpu
, this_cpu
, success
= 0;
2377 unsigned long flags
;
2378 struct rq
*rq
, *orig_rq
;
2380 if (!sched_feat(SYNC_WAKEUPS
))
2381 wake_flags
&= ~WF_SYNC
;
2383 this_cpu
= get_cpu();
2386 rq
= orig_rq
= task_rq_lock(p
, &flags
);
2387 update_rq_clock(rq
);
2388 if (!(p
->state
& state
))
2398 if (unlikely(task_running(rq
, p
)))
2402 * In order to handle concurrent wakeups and release the rq->lock
2403 * we put the task in TASK_WAKING state.
2405 * First fix up the nr_uninterruptible count:
2407 if (task_contributes_to_load(p
))
2408 rq
->nr_uninterruptible
--;
2409 p
->state
= TASK_WAKING
;
2410 __task_rq_unlock(rq
);
2412 cpu
= select_task_rq(p
, SD_BALANCE_WAKE
, wake_flags
);
2413 if (cpu
!= orig_cpu
)
2414 set_task_cpu(p
, cpu
);
2416 rq
= __task_rq_lock(p
);
2417 update_rq_clock(rq
);
2419 WARN_ON(p
->state
!= TASK_WAKING
);
2422 #ifdef CONFIG_SCHEDSTATS
2423 schedstat_inc(rq
, ttwu_count
);
2424 if (cpu
== this_cpu
)
2425 schedstat_inc(rq
, ttwu_local
);
2427 struct sched_domain
*sd
;
2428 for_each_domain(this_cpu
, sd
) {
2429 if (cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
2430 schedstat_inc(sd
, ttwu_wake_remote
);
2435 #endif /* CONFIG_SCHEDSTATS */
2438 #endif /* CONFIG_SMP */
2439 schedstat_inc(p
, se
.nr_wakeups
);
2440 if (wake_flags
& WF_SYNC
)
2441 schedstat_inc(p
, se
.nr_wakeups_sync
);
2442 if (orig_cpu
!= cpu
)
2443 schedstat_inc(p
, se
.nr_wakeups_migrate
);
2444 if (cpu
== this_cpu
)
2445 schedstat_inc(p
, se
.nr_wakeups_local
);
2447 schedstat_inc(p
, se
.nr_wakeups_remote
);
2448 activate_task(rq
, p
, 1);
2452 * Only attribute actual wakeups done by this task.
2454 if (!in_interrupt()) {
2455 struct sched_entity
*se
= ¤t
->se
;
2456 u64 sample
= se
->sum_exec_runtime
;
2458 if (se
->last_wakeup
)
2459 sample
-= se
->last_wakeup
;
2461 sample
-= se
->start_runtime
;
2462 update_avg(&se
->avg_wakeup
, sample
);
2464 se
->last_wakeup
= se
->sum_exec_runtime
;
2468 trace_sched_wakeup(rq
, p
, success
);
2469 check_preempt_curr(rq
, p
, wake_flags
);
2471 p
->state
= TASK_RUNNING
;
2473 if (p
->sched_class
->task_wake_up
)
2474 p
->sched_class
->task_wake_up(rq
, p
);
2476 if (unlikely(rq
->idle_stamp
)) {
2477 u64 delta
= rq
->clock
- rq
->idle_stamp
;
2478 u64 max
= 2*sysctl_sched_migration_cost
;
2483 update_avg(&rq
->avg_idle
, delta
);
2488 task_rq_unlock(rq
, &flags
);
2495 * wake_up_process - Wake up a specific process
2496 * @p: The process to be woken up.
2498 * Attempt to wake up the nominated process and move it to the set of runnable
2499 * processes. Returns 1 if the process was woken up, 0 if it was already
2502 * It may be assumed that this function implies a write memory barrier before
2503 * changing the task state if and only if any tasks are woken up.
2505 int wake_up_process(struct task_struct
*p
)
2507 return try_to_wake_up(p
, TASK_ALL
, 0);
2509 EXPORT_SYMBOL(wake_up_process
);
2511 int wake_up_state(struct task_struct
*p
, unsigned int state
)
2513 return try_to_wake_up(p
, state
, 0);
2517 * Perform scheduler related setup for a newly forked process p.
2518 * p is forked by current.
2520 * __sched_fork() is basic setup used by init_idle() too:
2522 static void __sched_fork(struct task_struct
*p
)
2524 p
->se
.exec_start
= 0;
2525 p
->se
.sum_exec_runtime
= 0;
2526 p
->se
.prev_sum_exec_runtime
= 0;
2527 p
->se
.nr_migrations
= 0;
2528 p
->se
.last_wakeup
= 0;
2529 p
->se
.avg_overlap
= 0;
2530 p
->se
.start_runtime
= 0;
2531 p
->se
.avg_wakeup
= sysctl_sched_wakeup_granularity
;
2533 #ifdef CONFIG_SCHEDSTATS
2534 p
->se
.wait_start
= 0;
2536 p
->se
.wait_count
= 0;
2539 p
->se
.sleep_start
= 0;
2540 p
->se
.sleep_max
= 0;
2541 p
->se
.sum_sleep_runtime
= 0;
2543 p
->se
.block_start
= 0;
2544 p
->se
.block_max
= 0;
2546 p
->se
.slice_max
= 0;
2548 p
->se
.nr_migrations_cold
= 0;
2549 p
->se
.nr_failed_migrations_affine
= 0;
2550 p
->se
.nr_failed_migrations_running
= 0;
2551 p
->se
.nr_failed_migrations_hot
= 0;
2552 p
->se
.nr_forced_migrations
= 0;
2554 p
->se
.nr_wakeups
= 0;
2555 p
->se
.nr_wakeups_sync
= 0;
2556 p
->se
.nr_wakeups_migrate
= 0;
2557 p
->se
.nr_wakeups_local
= 0;
2558 p
->se
.nr_wakeups_remote
= 0;
2559 p
->se
.nr_wakeups_affine
= 0;
2560 p
->se
.nr_wakeups_affine_attempts
= 0;
2561 p
->se
.nr_wakeups_passive
= 0;
2562 p
->se
.nr_wakeups_idle
= 0;
2566 INIT_LIST_HEAD(&p
->rt
.run_list
);
2568 INIT_LIST_HEAD(&p
->se
.group_node
);
2570 #ifdef CONFIG_PREEMPT_NOTIFIERS
2571 INIT_HLIST_HEAD(&p
->preempt_notifiers
);
2576 * fork()/clone()-time setup:
2578 void sched_fork(struct task_struct
*p
, int clone_flags
)
2580 int cpu
= get_cpu();
2584 * We mark the process as waking here. This guarantees that
2585 * nobody will actually run it, and a signal or other external
2586 * event cannot wake it up and insert it on the runqueue either.
2588 p
->state
= TASK_WAKING
;
2591 * Revert to default priority/policy on fork if requested.
2593 if (unlikely(p
->sched_reset_on_fork
)) {
2594 if (p
->policy
== SCHED_FIFO
|| p
->policy
== SCHED_RR
) {
2595 p
->policy
= SCHED_NORMAL
;
2596 p
->normal_prio
= p
->static_prio
;
2599 if (PRIO_TO_NICE(p
->static_prio
) < 0) {
2600 p
->static_prio
= NICE_TO_PRIO(0);
2601 p
->normal_prio
= p
->static_prio
;
2606 * We don't need the reset flag anymore after the fork. It has
2607 * fulfilled its duty:
2609 p
->sched_reset_on_fork
= 0;
2613 * Make sure we do not leak PI boosting priority to the child.
2615 p
->prio
= current
->normal_prio
;
2617 if (!rt_prio(p
->prio
))
2618 p
->sched_class
= &fair_sched_class
;
2620 if (p
->sched_class
->task_fork
)
2621 p
->sched_class
->task_fork(p
);
2624 cpu
= select_task_rq(p
, SD_BALANCE_FORK
, 0);
2626 set_task_cpu(p
, cpu
);
2628 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
2629 if (likely(sched_info_on()))
2630 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
2632 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
2635 #ifdef CONFIG_PREEMPT
2636 /* Want to start with kernel preemption disabled. */
2637 task_thread_info(p
)->preempt_count
= 1;
2639 plist_node_init(&p
->pushable_tasks
, MAX_PRIO
);
2645 * wake_up_new_task - wake up a newly created task for the first time.
2647 * This function will do some initial scheduler statistics housekeeping
2648 * that must be done for every newly created context, then puts the task
2649 * on the runqueue and wakes it.
2651 void wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
2653 unsigned long flags
;
2656 rq
= task_rq_lock(p
, &flags
);
2657 BUG_ON(p
->state
!= TASK_WAKING
);
2658 p
->state
= TASK_RUNNING
;
2659 update_rq_clock(rq
);
2660 activate_task(rq
, p
, 0);
2661 trace_sched_wakeup_new(rq
, p
, 1);
2662 check_preempt_curr(rq
, p
, WF_FORK
);
2664 if (p
->sched_class
->task_wake_up
)
2665 p
->sched_class
->task_wake_up(rq
, p
);
2667 task_rq_unlock(rq
, &flags
);
2670 #ifdef CONFIG_PREEMPT_NOTIFIERS
2673 * preempt_notifier_register - tell me when current is being preempted & rescheduled
2674 * @notifier: notifier struct to register
2676 void preempt_notifier_register(struct preempt_notifier
*notifier
)
2678 hlist_add_head(¬ifier
->link
, ¤t
->preempt_notifiers
);
2680 EXPORT_SYMBOL_GPL(preempt_notifier_register
);
2683 * preempt_notifier_unregister - no longer interested in preemption notifications
2684 * @notifier: notifier struct to unregister
2686 * This is safe to call from within a preemption notifier.
2688 void preempt_notifier_unregister(struct preempt_notifier
*notifier
)
2690 hlist_del(¬ifier
->link
);
2692 EXPORT_SYMBOL_GPL(preempt_notifier_unregister
);
2694 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2696 struct preempt_notifier
*notifier
;
2697 struct hlist_node
*node
;
2699 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2700 notifier
->ops
->sched_in(notifier
, raw_smp_processor_id());
2704 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2705 struct task_struct
*next
)
2707 struct preempt_notifier
*notifier
;
2708 struct hlist_node
*node
;
2710 hlist_for_each_entry(notifier
, node
, &curr
->preempt_notifiers
, link
)
2711 notifier
->ops
->sched_out(notifier
, next
);
2714 #else /* !CONFIG_PREEMPT_NOTIFIERS */
2716 static void fire_sched_in_preempt_notifiers(struct task_struct
*curr
)
2721 fire_sched_out_preempt_notifiers(struct task_struct
*curr
,
2722 struct task_struct
*next
)
2726 #endif /* CONFIG_PREEMPT_NOTIFIERS */
2729 * prepare_task_switch - prepare to switch tasks
2730 * @rq: the runqueue preparing to switch
2731 * @prev: the current task that is being switched out
2732 * @next: the task we are going to switch to.
2734 * This is called with the rq lock held and interrupts off. It must
2735 * be paired with a subsequent finish_task_switch after the context
2738 * prepare_task_switch sets up locking and calls architecture specific
2742 prepare_task_switch(struct rq
*rq
, struct task_struct
*prev
,
2743 struct task_struct
*next
)
2745 fire_sched_out_preempt_notifiers(prev
, next
);
2746 prepare_lock_switch(rq
, next
);
2747 prepare_arch_switch(next
);
2751 * finish_task_switch - clean up after a task-switch
2752 * @rq: runqueue associated with task-switch
2753 * @prev: the thread we just switched away from.
2755 * finish_task_switch must be called after the context switch, paired
2756 * with a prepare_task_switch call before the context switch.
2757 * finish_task_switch will reconcile locking set up by prepare_task_switch,
2758 * and do any other architecture-specific cleanup actions.
2760 * Note that we may have delayed dropping an mm in context_switch(). If
2761 * so, we finish that here outside of the runqueue lock. (Doing it
2762 * with the lock held can cause deadlocks; see schedule() for
2765 static void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
2766 __releases(rq
->lock
)
2768 struct mm_struct
*mm
= rq
->prev_mm
;
2774 * A task struct has one reference for the use as "current".
2775 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
2776 * schedule one last time. The schedule call will never return, and
2777 * the scheduled task must drop that reference.
2778 * The test for TASK_DEAD must occur while the runqueue locks are
2779 * still held, otherwise prev could be scheduled on another cpu, die
2780 * there before we look at prev->state, and then the reference would
2782 * Manfred Spraul <manfred@colorfullife.com>
2784 prev_state
= prev
->state
;
2785 finish_arch_switch(prev
);
2786 perf_event_task_sched_in(current
, cpu_of(rq
));
2787 finish_lock_switch(rq
, prev
);
2789 fire_sched_in_preempt_notifiers(current
);
2792 if (unlikely(prev_state
== TASK_DEAD
)) {
2794 * Remove function-return probe instances associated with this
2795 * task and put them back on the free list.
2797 kprobe_flush_task(prev
);
2798 put_task_struct(prev
);
2804 /* assumes rq->lock is held */
2805 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*prev
)
2807 if (prev
->sched_class
->pre_schedule
)
2808 prev
->sched_class
->pre_schedule(rq
, prev
);
2811 /* rq->lock is NOT held, but preemption is disabled */
2812 static inline void post_schedule(struct rq
*rq
)
2814 if (rq
->post_schedule
) {
2815 unsigned long flags
;
2817 raw_spin_lock_irqsave(&rq
->lock
, flags
);
2818 if (rq
->curr
->sched_class
->post_schedule
)
2819 rq
->curr
->sched_class
->post_schedule(rq
);
2820 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
2822 rq
->post_schedule
= 0;
2828 static inline void pre_schedule(struct rq
*rq
, struct task_struct
*p
)
2832 static inline void post_schedule(struct rq
*rq
)
2839 * schedule_tail - first thing a freshly forked thread must call.
2840 * @prev: the thread we just switched away from.
2842 asmlinkage
void schedule_tail(struct task_struct
*prev
)
2843 __releases(rq
->lock
)
2845 struct rq
*rq
= this_rq();
2847 finish_task_switch(rq
, prev
);
2850 * FIXME: do we need to worry about rq being invalidated by the
2855 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
2856 /* In this case, finish_task_switch does not reenable preemption */
2859 if (current
->set_child_tid
)
2860 put_user(task_pid_vnr(current
), current
->set_child_tid
);
2864 * context_switch - switch to the new MM and the new
2865 * thread's register state.
2868 context_switch(struct rq
*rq
, struct task_struct
*prev
,
2869 struct task_struct
*next
)
2871 struct mm_struct
*mm
, *oldmm
;
2873 prepare_task_switch(rq
, prev
, next
);
2874 trace_sched_switch(rq
, prev
, next
);
2876 oldmm
= prev
->active_mm
;
2878 * For paravirt, this is coupled with an exit in switch_to to
2879 * combine the page table reload and the switch backend into
2882 arch_start_context_switch(prev
);
2885 next
->active_mm
= oldmm
;
2886 atomic_inc(&oldmm
->mm_count
);
2887 enter_lazy_tlb(oldmm
, next
);
2889 switch_mm(oldmm
, mm
, next
);
2891 if (likely(!prev
->mm
)) {
2892 prev
->active_mm
= NULL
;
2893 rq
->prev_mm
= oldmm
;
2896 * Since the runqueue lock will be released by the next
2897 * task (which is an invalid locking op but in the case
2898 * of the scheduler it's an obvious special-case), so we
2899 * do an early lockdep release here:
2901 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
2902 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
2905 /* Here we just switch the register state and the stack. */
2906 switch_to(prev
, next
, prev
);
2910 * this_rq must be evaluated again because prev may have moved
2911 * CPUs since it called schedule(), thus the 'rq' on its stack
2912 * frame will be invalid.
2914 finish_task_switch(this_rq(), prev
);
2918 * nr_running, nr_uninterruptible and nr_context_switches:
2920 * externally visible scheduler statistics: current number of runnable
2921 * threads, current number of uninterruptible-sleeping threads, total
2922 * number of context switches performed since bootup.
2924 unsigned long nr_running(void)
2926 unsigned long i
, sum
= 0;
2928 for_each_online_cpu(i
)
2929 sum
+= cpu_rq(i
)->nr_running
;
2934 unsigned long nr_uninterruptible(void)
2936 unsigned long i
, sum
= 0;
2938 for_each_possible_cpu(i
)
2939 sum
+= cpu_rq(i
)->nr_uninterruptible
;
2942 * Since we read the counters lockless, it might be slightly
2943 * inaccurate. Do not allow it to go below zero though:
2945 if (unlikely((long)sum
< 0))
2951 unsigned long long nr_context_switches(void)
2954 unsigned long long sum
= 0;
2956 for_each_possible_cpu(i
)
2957 sum
+= cpu_rq(i
)->nr_switches
;
2962 unsigned long nr_iowait(void)
2964 unsigned long i
, sum
= 0;
2966 for_each_possible_cpu(i
)
2967 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
2972 unsigned long nr_iowait_cpu(void)
2974 struct rq
*this = this_rq();
2975 return atomic_read(&this->nr_iowait
);
2978 unsigned long this_cpu_load(void)
2980 struct rq
*this = this_rq();
2981 return this->cpu_load
[0];
2985 /* Variables and functions for calc_load */
2986 static atomic_long_t calc_load_tasks
;
2987 static unsigned long calc_load_update
;
2988 unsigned long avenrun
[3];
2989 EXPORT_SYMBOL(avenrun
);
2992 * get_avenrun - get the load average array
2993 * @loads: pointer to dest load array
2994 * @offset: offset to add
2995 * @shift: shift count to shift the result left
2997 * These values are estimates at best, so no need for locking.
2999 void get_avenrun(unsigned long *loads
, unsigned long offset
, int shift
)
3001 loads
[0] = (avenrun
[0] + offset
) << shift
;
3002 loads
[1] = (avenrun
[1] + offset
) << shift
;
3003 loads
[2] = (avenrun
[2] + offset
) << shift
;
3006 static unsigned long
3007 calc_load(unsigned long load
, unsigned long exp
, unsigned long active
)
3010 load
+= active
* (FIXED_1
- exp
);
3011 return load
>> FSHIFT
;
3015 * calc_load - update the avenrun load estimates 10 ticks after the
3016 * CPUs have updated calc_load_tasks.
3018 void calc_global_load(void)
3020 unsigned long upd
= calc_load_update
+ 10;
3023 if (time_before(jiffies
, upd
))
3026 active
= atomic_long_read(&calc_load_tasks
);
3027 active
= active
> 0 ? active
* FIXED_1
: 0;
3029 avenrun
[0] = calc_load(avenrun
[0], EXP_1
, active
);
3030 avenrun
[1] = calc_load(avenrun
[1], EXP_5
, active
);
3031 avenrun
[2] = calc_load(avenrun
[2], EXP_15
, active
);
3033 calc_load_update
+= LOAD_FREQ
;
3037 * Either called from update_cpu_load() or from a cpu going idle
3039 static void calc_load_account_active(struct rq
*this_rq
)
3041 long nr_active
, delta
;
3043 nr_active
= this_rq
->nr_running
;
3044 nr_active
+= (long) this_rq
->nr_uninterruptible
;
3046 if (nr_active
!= this_rq
->calc_load_active
) {
3047 delta
= nr_active
- this_rq
->calc_load_active
;
3048 this_rq
->calc_load_active
= nr_active
;
3049 atomic_long_add(delta
, &calc_load_tasks
);
3054 * Update rq->cpu_load[] statistics. This function is usually called every
3055 * scheduler tick (TICK_NSEC).
3057 static void update_cpu_load(struct rq
*this_rq
)
3059 unsigned long this_load
= this_rq
->load
.weight
;
3062 this_rq
->nr_load_updates
++;
3064 /* Update our load: */
3065 for (i
= 0, scale
= 1; i
< CPU_LOAD_IDX_MAX
; i
++, scale
+= scale
) {
3066 unsigned long old_load
, new_load
;
3068 /* scale is effectively 1 << i now, and >> i divides by scale */
3070 old_load
= this_rq
->cpu_load
[i
];
3071 new_load
= this_load
;
3073 * Round up the averaging division if load is increasing. This
3074 * prevents us from getting stuck on 9 if the load is 10, for
3077 if (new_load
> old_load
)
3078 new_load
+= scale
-1;
3079 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) >> i
;
3082 if (time_after_eq(jiffies
, this_rq
->calc_load_update
)) {
3083 this_rq
->calc_load_update
+= LOAD_FREQ
;
3084 calc_load_account_active(this_rq
);
3091 * double_rq_lock - safely lock two runqueues
3093 * Note this does not disable interrupts like task_rq_lock,
3094 * you need to do so manually before calling.
3096 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
3097 __acquires(rq1
->lock
)
3098 __acquires(rq2
->lock
)
3100 BUG_ON(!irqs_disabled());
3102 raw_spin_lock(&rq1
->lock
);
3103 __acquire(rq2
->lock
); /* Fake it out ;) */
3106 raw_spin_lock(&rq1
->lock
);
3107 raw_spin_lock_nested(&rq2
->lock
, SINGLE_DEPTH_NESTING
);
3109 raw_spin_lock(&rq2
->lock
);
3110 raw_spin_lock_nested(&rq1
->lock
, SINGLE_DEPTH_NESTING
);
3113 update_rq_clock(rq1
);
3114 update_rq_clock(rq2
);
3118 * double_rq_unlock - safely unlock two runqueues
3120 * Note this does not restore interrupts like task_rq_unlock,
3121 * you need to do so manually after calling.
3123 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
3124 __releases(rq1
->lock
)
3125 __releases(rq2
->lock
)
3127 raw_spin_unlock(&rq1
->lock
);
3129 raw_spin_unlock(&rq2
->lock
);
3131 __release(rq2
->lock
);
3135 * If dest_cpu is allowed for this process, migrate the task to it.
3136 * This is accomplished by forcing the cpu_allowed mask to only
3137 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
3138 * the cpu_allowed mask is restored.
3140 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
3142 struct migration_req req
;
3143 unsigned long flags
;
3146 rq
= task_rq_lock(p
, &flags
);
3147 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
)
3148 || unlikely(!cpu_active(dest_cpu
)))
3151 /* force the process onto the specified CPU */
3152 if (migrate_task(p
, dest_cpu
, &req
)) {
3153 /* Need to wait for migration thread (might exit: take ref). */
3154 struct task_struct
*mt
= rq
->migration_thread
;
3156 get_task_struct(mt
);
3157 task_rq_unlock(rq
, &flags
);
3158 wake_up_process(mt
);
3159 put_task_struct(mt
);
3160 wait_for_completion(&req
.done
);
3165 task_rq_unlock(rq
, &flags
);
3169 * sched_exec - execve() is a valuable balancing opportunity, because at
3170 * this point the task has the smallest effective memory and cache footprint.
3172 void sched_exec(void)
3174 int new_cpu
, this_cpu
= get_cpu();
3175 new_cpu
= select_task_rq(current
, SD_BALANCE_EXEC
, 0);
3177 if (new_cpu
!= this_cpu
)
3178 sched_migrate_task(current
, new_cpu
);
3182 * pull_task - move a task from a remote runqueue to the local runqueue.
3183 * Both runqueues must be locked.
3185 static void pull_task(struct rq
*src_rq
, struct task_struct
*p
,
3186 struct rq
*this_rq
, int this_cpu
)
3188 deactivate_task(src_rq
, p
, 0);
3189 set_task_cpu(p
, this_cpu
);
3190 activate_task(this_rq
, p
, 0);
3191 check_preempt_curr(this_rq
, p
, 0);
3195 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
3198 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
3199 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3202 int tsk_cache_hot
= 0;
3204 * We do not migrate tasks that are:
3205 * 1) running (obviously), or
3206 * 2) cannot be migrated to this CPU due to cpus_allowed, or
3207 * 3) are cache-hot on their current CPU.
3209 if (!cpumask_test_cpu(this_cpu
, &p
->cpus_allowed
)) {
3210 schedstat_inc(p
, se
.nr_failed_migrations_affine
);
3215 if (task_running(rq
, p
)) {
3216 schedstat_inc(p
, se
.nr_failed_migrations_running
);
3221 * Aggressive migration if:
3222 * 1) task is cache cold, or
3223 * 2) too many balance attempts have failed.
3226 tsk_cache_hot
= task_hot(p
, rq
->clock
, sd
);
3227 if (!tsk_cache_hot
||
3228 sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
3229 #ifdef CONFIG_SCHEDSTATS
3230 if (tsk_cache_hot
) {
3231 schedstat_inc(sd
, lb_hot_gained
[idle
]);
3232 schedstat_inc(p
, se
.nr_forced_migrations
);
3238 if (tsk_cache_hot
) {
3239 schedstat_inc(p
, se
.nr_failed_migrations_hot
);
3245 static unsigned long
3246 balance_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3247 unsigned long max_load_move
, struct sched_domain
*sd
,
3248 enum cpu_idle_type idle
, int *all_pinned
,
3249 int *this_best_prio
, struct rq_iterator
*iterator
)
3251 int loops
= 0, pulled
= 0, pinned
= 0;
3252 struct task_struct
*p
;
3253 long rem_load_move
= max_load_move
;
3255 if (max_load_move
== 0)
3261 * Start the load-balancing iterator:
3263 p
= iterator
->start(iterator
->arg
);
3265 if (!p
|| loops
++ > sysctl_sched_nr_migrate
)
3268 if ((p
->se
.load
.weight
>> 1) > rem_load_move
||
3269 !can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3270 p
= iterator
->next(iterator
->arg
);
3274 pull_task(busiest
, p
, this_rq
, this_cpu
);
3276 rem_load_move
-= p
->se
.load
.weight
;
3278 #ifdef CONFIG_PREEMPT
3280 * NEWIDLE balancing is a source of latency, so preemptible kernels
3281 * will stop after the first task is pulled to minimize the critical
3284 if (idle
== CPU_NEWLY_IDLE
)
3289 * We only want to steal up to the prescribed amount of weighted load.
3291 if (rem_load_move
> 0) {
3292 if (p
->prio
< *this_best_prio
)
3293 *this_best_prio
= p
->prio
;
3294 p
= iterator
->next(iterator
->arg
);
3299 * Right now, this is one of only two places pull_task() is called,
3300 * so we can safely collect pull_task() stats here rather than
3301 * inside pull_task().
3303 schedstat_add(sd
, lb_gained
[idle
], pulled
);
3306 *all_pinned
= pinned
;
3308 return max_load_move
- rem_load_move
;
3312 * move_tasks tries to move up to max_load_move weighted load from busiest to
3313 * this_rq, as part of a balancing operation within domain "sd".
3314 * Returns 1 if successful and 0 otherwise.
3316 * Called with both runqueues locked.
3318 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3319 unsigned long max_load_move
,
3320 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3323 const struct sched_class
*class = sched_class_highest
;
3324 unsigned long total_load_moved
= 0;
3325 int this_best_prio
= this_rq
->curr
->prio
;
3329 class->load_balance(this_rq
, this_cpu
, busiest
,
3330 max_load_move
- total_load_moved
,
3331 sd
, idle
, all_pinned
, &this_best_prio
);
3332 class = class->next
;
3334 #ifdef CONFIG_PREEMPT
3336 * NEWIDLE balancing is a source of latency, so preemptible
3337 * kernels will stop after the first task is pulled to minimize
3338 * the critical section.
3340 if (idle
== CPU_NEWLY_IDLE
&& this_rq
->nr_running
)
3343 } while (class && max_load_move
> total_load_moved
);
3345 return total_load_moved
> 0;
3349 iter_move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3350 struct sched_domain
*sd
, enum cpu_idle_type idle
,
3351 struct rq_iterator
*iterator
)
3353 struct task_struct
*p
= iterator
->start(iterator
->arg
);
3357 if (can_migrate_task(p
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
3358 pull_task(busiest
, p
, this_rq
, this_cpu
);
3360 * Right now, this is only the second place pull_task()
3361 * is called, so we can safely collect pull_task()
3362 * stats here rather than inside pull_task().
3364 schedstat_inc(sd
, lb_gained
[idle
]);
3368 p
= iterator
->next(iterator
->arg
);
3375 * move_one_task tries to move exactly one task from busiest to this_rq, as
3376 * part of active balancing operations within "domain".
3377 * Returns 1 if successful and 0 otherwise.
3379 * Called with both runqueues locked.
3381 static int move_one_task(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
3382 struct sched_domain
*sd
, enum cpu_idle_type idle
)
3384 const struct sched_class
*class;
3386 for_each_class(class) {
3387 if (class->move_one_task(this_rq
, this_cpu
, busiest
, sd
, idle
))
3393 /********** Helpers for find_busiest_group ************************/
3395 * sd_lb_stats - Structure to store the statistics of a sched_domain
3396 * during load balancing.
3398 struct sd_lb_stats
{
3399 struct sched_group
*busiest
; /* Busiest group in this sd */
3400 struct sched_group
*this; /* Local group in this sd */
3401 unsigned long total_load
; /* Total load of all groups in sd */
3402 unsigned long total_pwr
; /* Total power of all groups in sd */
3403 unsigned long avg_load
; /* Average load across all groups in sd */
3405 /** Statistics of this group */
3406 unsigned long this_load
;
3407 unsigned long this_load_per_task
;
3408 unsigned long this_nr_running
;
3410 /* Statistics of the busiest group */
3411 unsigned long max_load
;
3412 unsigned long busiest_load_per_task
;
3413 unsigned long busiest_nr_running
;
3415 int group_imb
; /* Is there imbalance in this sd */
3416 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3417 int power_savings_balance
; /* Is powersave balance needed for this sd */
3418 struct sched_group
*group_min
; /* Least loaded group in sd */
3419 struct sched_group
*group_leader
; /* Group which relieves group_min */
3420 unsigned long min_load_per_task
; /* load_per_task in group_min */
3421 unsigned long leader_nr_running
; /* Nr running of group_leader */
3422 unsigned long min_nr_running
; /* Nr running of group_min */
3427 * sg_lb_stats - stats of a sched_group required for load_balancing
3429 struct sg_lb_stats
{
3430 unsigned long avg_load
; /*Avg load across the CPUs of the group */
3431 unsigned long group_load
; /* Total load over the CPUs of the group */
3432 unsigned long sum_nr_running
; /* Nr tasks running in the group */
3433 unsigned long sum_weighted_load
; /* Weighted load of group's tasks */
3434 unsigned long group_capacity
;
3435 int group_imb
; /* Is there an imbalance in the group ? */
3439 * group_first_cpu - Returns the first cpu in the cpumask of a sched_group.
3440 * @group: The group whose first cpu is to be returned.
3442 static inline unsigned int group_first_cpu(struct sched_group
*group
)
3444 return cpumask_first(sched_group_cpus(group
));
3448 * get_sd_load_idx - Obtain the load index for a given sched domain.
3449 * @sd: The sched_domain whose load_idx is to be obtained.
3450 * @idle: The Idle status of the CPU for whose sd load_icx is obtained.
3452 static inline int get_sd_load_idx(struct sched_domain
*sd
,
3453 enum cpu_idle_type idle
)
3459 load_idx
= sd
->busy_idx
;
3462 case CPU_NEWLY_IDLE
:
3463 load_idx
= sd
->newidle_idx
;
3466 load_idx
= sd
->idle_idx
;
3474 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
3476 * init_sd_power_savings_stats - Initialize power savings statistics for
3477 * the given sched_domain, during load balancing.
3479 * @sd: Sched domain whose power-savings statistics are to be initialized.
3480 * @sds: Variable containing the statistics for sd.
3481 * @idle: Idle status of the CPU at which we're performing load-balancing.
3483 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3484 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3487 * Busy processors will not participate in power savings
3490 if (idle
== CPU_NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
3491 sds
->power_savings_balance
= 0;
3493 sds
->power_savings_balance
= 1;
3494 sds
->min_nr_running
= ULONG_MAX
;
3495 sds
->leader_nr_running
= 0;
3500 * update_sd_power_savings_stats - Update the power saving stats for a
3501 * sched_domain while performing load balancing.
3503 * @group: sched_group belonging to the sched_domain under consideration.
3504 * @sds: Variable containing the statistics of the sched_domain
3505 * @local_group: Does group contain the CPU for which we're performing
3507 * @sgs: Variable containing the statistics of the group.
3509 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3510 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3513 if (!sds
->power_savings_balance
)
3517 * If the local group is idle or completely loaded
3518 * no need to do power savings balance at this domain
3520 if (local_group
&& (sds
->this_nr_running
>= sgs
->group_capacity
||
3521 !sds
->this_nr_running
))
3522 sds
->power_savings_balance
= 0;
3525 * If a group is already running at full capacity or idle,
3526 * don't include that group in power savings calculations
3528 if (!sds
->power_savings_balance
||
3529 sgs
->sum_nr_running
>= sgs
->group_capacity
||
3530 !sgs
->sum_nr_running
)
3534 * Calculate the group which has the least non-idle load.
3535 * This is the group from where we need to pick up the load
3538 if ((sgs
->sum_nr_running
< sds
->min_nr_running
) ||
3539 (sgs
->sum_nr_running
== sds
->min_nr_running
&&
3540 group_first_cpu(group
) > group_first_cpu(sds
->group_min
))) {
3541 sds
->group_min
= group
;
3542 sds
->min_nr_running
= sgs
->sum_nr_running
;
3543 sds
->min_load_per_task
= sgs
->sum_weighted_load
/
3544 sgs
->sum_nr_running
;
3548 * Calculate the group which is almost near its
3549 * capacity but still has some space to pick up some load
3550 * from other group and save more power
3552 if (sgs
->sum_nr_running
+ 1 > sgs
->group_capacity
)
3555 if (sgs
->sum_nr_running
> sds
->leader_nr_running
||
3556 (sgs
->sum_nr_running
== sds
->leader_nr_running
&&
3557 group_first_cpu(group
) < group_first_cpu(sds
->group_leader
))) {
3558 sds
->group_leader
= group
;
3559 sds
->leader_nr_running
= sgs
->sum_nr_running
;
3564 * check_power_save_busiest_group - see if there is potential for some power-savings balance
3565 * @sds: Variable containing the statistics of the sched_domain
3566 * under consideration.
3567 * @this_cpu: Cpu at which we're currently performing load-balancing.
3568 * @imbalance: Variable to store the imbalance.
3571 * Check if we have potential to perform some power-savings balance.
3572 * If yes, set the busiest group to be the least loaded group in the
3573 * sched_domain, so that it's CPUs can be put to idle.
3575 * Returns 1 if there is potential to perform power-savings balance.
3578 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3579 int this_cpu
, unsigned long *imbalance
)
3581 if (!sds
->power_savings_balance
)
3584 if (sds
->this != sds
->group_leader
||
3585 sds
->group_leader
== sds
->group_min
)
3588 *imbalance
= sds
->min_load_per_task
;
3589 sds
->busiest
= sds
->group_min
;
3594 #else /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3595 static inline void init_sd_power_savings_stats(struct sched_domain
*sd
,
3596 struct sd_lb_stats
*sds
, enum cpu_idle_type idle
)
3601 static inline void update_sd_power_savings_stats(struct sched_group
*group
,
3602 struct sd_lb_stats
*sds
, int local_group
, struct sg_lb_stats
*sgs
)
3607 static inline int check_power_save_busiest_group(struct sd_lb_stats
*sds
,
3608 int this_cpu
, unsigned long *imbalance
)
3612 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
3615 unsigned long default_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3617 return SCHED_LOAD_SCALE
;
3620 unsigned long __weak
arch_scale_freq_power(struct sched_domain
*sd
, int cpu
)
3622 return default_scale_freq_power(sd
, cpu
);
3625 unsigned long default_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3627 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3628 unsigned long smt_gain
= sd
->smt_gain
;
3635 unsigned long __weak
arch_scale_smt_power(struct sched_domain
*sd
, int cpu
)
3637 return default_scale_smt_power(sd
, cpu
);
3640 unsigned long scale_rt_power(int cpu
)
3642 struct rq
*rq
= cpu_rq(cpu
);
3643 u64 total
, available
;
3645 sched_avg_update(rq
);
3647 total
= sched_avg_period() + (rq
->clock
- rq
->age_stamp
);
3648 available
= total
- rq
->rt_avg
;
3650 if (unlikely((s64
)total
< SCHED_LOAD_SCALE
))
3651 total
= SCHED_LOAD_SCALE
;
3653 total
>>= SCHED_LOAD_SHIFT
;
3655 return div_u64(available
, total
);
3658 static void update_cpu_power(struct sched_domain
*sd
, int cpu
)
3660 unsigned long weight
= cpumask_weight(sched_domain_span(sd
));
3661 unsigned long power
= SCHED_LOAD_SCALE
;
3662 struct sched_group
*sdg
= sd
->groups
;
3664 if (sched_feat(ARCH_POWER
))
3665 power
*= arch_scale_freq_power(sd
, cpu
);
3667 power
*= default_scale_freq_power(sd
, cpu
);
3669 power
>>= SCHED_LOAD_SHIFT
;
3671 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
3672 if (sched_feat(ARCH_POWER
))
3673 power
*= arch_scale_smt_power(sd
, cpu
);
3675 power
*= default_scale_smt_power(sd
, cpu
);
3677 power
>>= SCHED_LOAD_SHIFT
;
3680 power
*= scale_rt_power(cpu
);
3681 power
>>= SCHED_LOAD_SHIFT
;
3686 sdg
->cpu_power
= power
;
3689 static void update_group_power(struct sched_domain
*sd
, int cpu
)
3691 struct sched_domain
*child
= sd
->child
;
3692 struct sched_group
*group
, *sdg
= sd
->groups
;
3693 unsigned long power
;
3696 update_cpu_power(sd
, cpu
);
3702 group
= child
->groups
;
3704 power
+= group
->cpu_power
;
3705 group
= group
->next
;
3706 } while (group
!= child
->groups
);
3708 sdg
->cpu_power
= power
;
3712 * update_sg_lb_stats - Update sched_group's statistics for load balancing.
3713 * @sd: The sched_domain whose statistics are to be updated.
3714 * @group: sched_group whose statistics are to be updated.
3715 * @this_cpu: Cpu for which load balance is currently performed.
3716 * @idle: Idle status of this_cpu
3717 * @load_idx: Load index of sched_domain of this_cpu for load calc.
3718 * @sd_idle: Idle status of the sched_domain containing group.
3719 * @local_group: Does group contain this_cpu.
3720 * @cpus: Set of cpus considered for load balancing.
3721 * @balance: Should we balance.
3722 * @sgs: variable to hold the statistics for this group.
3724 static inline void update_sg_lb_stats(struct sched_domain
*sd
,
3725 struct sched_group
*group
, int this_cpu
,
3726 enum cpu_idle_type idle
, int load_idx
, int *sd_idle
,
3727 int local_group
, const struct cpumask
*cpus
,
3728 int *balance
, struct sg_lb_stats
*sgs
)
3730 unsigned long load
, max_cpu_load
, min_cpu_load
;
3732 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
3733 unsigned long sum_avg_load_per_task
;
3734 unsigned long avg_load_per_task
;
3737 balance_cpu
= group_first_cpu(group
);
3738 if (balance_cpu
== this_cpu
)
3739 update_group_power(sd
, this_cpu
);
3742 /* Tally up the load of all CPUs in the group */
3743 sum_avg_load_per_task
= avg_load_per_task
= 0;
3745 min_cpu_load
= ~0UL;
3747 for_each_cpu_and(i
, sched_group_cpus(group
), cpus
) {
3748 struct rq
*rq
= cpu_rq(i
);
3750 if (*sd_idle
&& rq
->nr_running
)
3753 /* Bias balancing toward cpus of our domain */
3755 if (idle_cpu(i
) && !first_idle_cpu
) {
3760 load
= target_load(i
, load_idx
);
3762 load
= source_load(i
, load_idx
);
3763 if (load
> max_cpu_load
)
3764 max_cpu_load
= load
;
3765 if (min_cpu_load
> load
)
3766 min_cpu_load
= load
;
3769 sgs
->group_load
+= load
;
3770 sgs
->sum_nr_running
+= rq
->nr_running
;
3771 sgs
->sum_weighted_load
+= weighted_cpuload(i
);
3773 sum_avg_load_per_task
+= cpu_avg_load_per_task(i
);
3777 * First idle cpu or the first cpu(busiest) in this sched group
3778 * is eligible for doing load balancing at this and above
3779 * domains. In the newly idle case, we will allow all the cpu's
3780 * to do the newly idle load balance.
3782 if (idle
!= CPU_NEWLY_IDLE
&& local_group
&&
3783 balance_cpu
!= this_cpu
&& balance
) {
3788 /* Adjust by relative CPU power of the group */
3789 sgs
->avg_load
= (sgs
->group_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
3793 * Consider the group unbalanced when the imbalance is larger
3794 * than the average weight of two tasks.
3796 * APZ: with cgroup the avg task weight can vary wildly and
3797 * might not be a suitable number - should we keep a
3798 * normalized nr_running number somewhere that negates
3801 avg_load_per_task
= (sum_avg_load_per_task
* SCHED_LOAD_SCALE
) /
3804 if ((max_cpu_load
- min_cpu_load
) > 2*avg_load_per_task
)
3807 sgs
->group_capacity
=
3808 DIV_ROUND_CLOSEST(group
->cpu_power
, SCHED_LOAD_SCALE
);
3812 * update_sd_lb_stats - Update sched_group's statistics for load balancing.
3813 * @sd: sched_domain whose statistics are to be updated.
3814 * @this_cpu: Cpu for which load balance is currently performed.
3815 * @idle: Idle status of this_cpu
3816 * @sd_idle: Idle status of the sched_domain containing group.
3817 * @cpus: Set of cpus considered for load balancing.
3818 * @balance: Should we balance.
3819 * @sds: variable to hold the statistics for this sched_domain.
3821 static inline void update_sd_lb_stats(struct sched_domain
*sd
, int this_cpu
,
3822 enum cpu_idle_type idle
, int *sd_idle
,
3823 const struct cpumask
*cpus
, int *balance
,
3824 struct sd_lb_stats
*sds
)
3826 struct sched_domain
*child
= sd
->child
;
3827 struct sched_group
*group
= sd
->groups
;
3828 struct sg_lb_stats sgs
;
3829 int load_idx
, prefer_sibling
= 0;
3831 if (child
&& child
->flags
& SD_PREFER_SIBLING
)
3834 init_sd_power_savings_stats(sd
, sds
, idle
);
3835 load_idx
= get_sd_load_idx(sd
, idle
);
3840 local_group
= cpumask_test_cpu(this_cpu
,
3841 sched_group_cpus(group
));
3842 memset(&sgs
, 0, sizeof(sgs
));
3843 update_sg_lb_stats(sd
, group
, this_cpu
, idle
, load_idx
, sd_idle
,
3844 local_group
, cpus
, balance
, &sgs
);
3846 if (local_group
&& balance
&& !(*balance
))
3849 sds
->total_load
+= sgs
.group_load
;
3850 sds
->total_pwr
+= group
->cpu_power
;
3853 * In case the child domain prefers tasks go to siblings
3854 * first, lower the group capacity to one so that we'll try
3855 * and move all the excess tasks away.
3858 sgs
.group_capacity
= min(sgs
.group_capacity
, 1UL);
3861 sds
->this_load
= sgs
.avg_load
;
3863 sds
->this_nr_running
= sgs
.sum_nr_running
;
3864 sds
->this_load_per_task
= sgs
.sum_weighted_load
;
3865 } else if (sgs
.avg_load
> sds
->max_load
&&
3866 (sgs
.sum_nr_running
> sgs
.group_capacity
||
3868 sds
->max_load
= sgs
.avg_load
;
3869 sds
->busiest
= group
;
3870 sds
->busiest_nr_running
= sgs
.sum_nr_running
;
3871 sds
->busiest_load_per_task
= sgs
.sum_weighted_load
;
3872 sds
->group_imb
= sgs
.group_imb
;
3875 update_sd_power_savings_stats(group
, sds
, local_group
, &sgs
);
3876 group
= group
->next
;
3877 } while (group
!= sd
->groups
);
3881 * fix_small_imbalance - Calculate the minor imbalance that exists
3882 * amongst the groups of a sched_domain, during
3884 * @sds: Statistics of the sched_domain whose imbalance is to be calculated.
3885 * @this_cpu: The cpu at whose sched_domain we're performing load-balance.
3886 * @imbalance: Variable to store the imbalance.
3888 static inline void fix_small_imbalance(struct sd_lb_stats
*sds
,
3889 int this_cpu
, unsigned long *imbalance
)
3891 unsigned long tmp
, pwr_now
= 0, pwr_move
= 0;
3892 unsigned int imbn
= 2;
3894 if (sds
->this_nr_running
) {
3895 sds
->this_load_per_task
/= sds
->this_nr_running
;
3896 if (sds
->busiest_load_per_task
>
3897 sds
->this_load_per_task
)
3900 sds
->this_load_per_task
=
3901 cpu_avg_load_per_task(this_cpu
);
3903 if (sds
->max_load
- sds
->this_load
+ sds
->busiest_load_per_task
>=
3904 sds
->busiest_load_per_task
* imbn
) {
3905 *imbalance
= sds
->busiest_load_per_task
;
3910 * OK, we don't have enough imbalance to justify moving tasks,
3911 * however we may be able to increase total CPU power used by
3915 pwr_now
+= sds
->busiest
->cpu_power
*
3916 min(sds
->busiest_load_per_task
, sds
->max_load
);
3917 pwr_now
+= sds
->this->cpu_power
*
3918 min(sds
->this_load_per_task
, sds
->this_load
);
3919 pwr_now
/= SCHED_LOAD_SCALE
;
3921 /* Amount of load we'd subtract */
3922 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3923 sds
->busiest
->cpu_power
;
3924 if (sds
->max_load
> tmp
)
3925 pwr_move
+= sds
->busiest
->cpu_power
*
3926 min(sds
->busiest_load_per_task
, sds
->max_load
- tmp
);
3928 /* Amount of load we'd add */
3929 if (sds
->max_load
* sds
->busiest
->cpu_power
<
3930 sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
)
3931 tmp
= (sds
->max_load
* sds
->busiest
->cpu_power
) /
3932 sds
->this->cpu_power
;
3934 tmp
= (sds
->busiest_load_per_task
* SCHED_LOAD_SCALE
) /
3935 sds
->this->cpu_power
;
3936 pwr_move
+= sds
->this->cpu_power
*
3937 min(sds
->this_load_per_task
, sds
->this_load
+ tmp
);
3938 pwr_move
/= SCHED_LOAD_SCALE
;
3940 /* Move if we gain throughput */
3941 if (pwr_move
> pwr_now
)
3942 *imbalance
= sds
->busiest_load_per_task
;
3946 * calculate_imbalance - Calculate the amount of imbalance present within the
3947 * groups of a given sched_domain during load balance.
3948 * @sds: statistics of the sched_domain whose imbalance is to be calculated.
3949 * @this_cpu: Cpu for which currently load balance is being performed.
3950 * @imbalance: The variable to store the imbalance.
3952 static inline void calculate_imbalance(struct sd_lb_stats
*sds
, int this_cpu
,
3953 unsigned long *imbalance
)
3955 unsigned long max_pull
;
3957 * In the presence of smp nice balancing, certain scenarios can have
3958 * max load less than avg load(as we skip the groups at or below
3959 * its cpu_power, while calculating max_load..)
3961 if (sds
->max_load
< sds
->avg_load
) {
3963 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3966 /* Don't want to pull so many tasks that a group would go idle */
3967 max_pull
= min(sds
->max_load
- sds
->avg_load
,
3968 sds
->max_load
- sds
->busiest_load_per_task
);
3970 /* How much load to actually move to equalise the imbalance */
3971 *imbalance
= min(max_pull
* sds
->busiest
->cpu_power
,
3972 (sds
->avg_load
- sds
->this_load
) * sds
->this->cpu_power
)
3976 * if *imbalance is less than the average load per runnable task
3977 * there is no gaurantee that any tasks will be moved so we'll have
3978 * a think about bumping its value to force at least one task to be
3981 if (*imbalance
< sds
->busiest_load_per_task
)
3982 return fix_small_imbalance(sds
, this_cpu
, imbalance
);
3985 /******* find_busiest_group() helpers end here *********************/
3988 * find_busiest_group - Returns the busiest group within the sched_domain
3989 * if there is an imbalance. If there isn't an imbalance, and
3990 * the user has opted for power-savings, it returns a group whose
3991 * CPUs can be put to idle by rebalancing those tasks elsewhere, if
3992 * such a group exists.
3994 * Also calculates the amount of weighted load which should be moved
3995 * to restore balance.
3997 * @sd: The sched_domain whose busiest group is to be returned.
3998 * @this_cpu: The cpu for which load balancing is currently being performed.
3999 * @imbalance: Variable which stores amount of weighted load which should
4000 * be moved to restore balance/put a group to idle.
4001 * @idle: The idle status of this_cpu.
4002 * @sd_idle: The idleness of sd
4003 * @cpus: The set of CPUs under consideration for load-balancing.
4004 * @balance: Pointer to a variable indicating if this_cpu
4005 * is the appropriate cpu to perform load balancing at this_level.
4007 * Returns: - the busiest group if imbalance exists.
4008 * - If no imbalance and user has opted for power-savings balance,
4009 * return the least loaded group whose CPUs can be
4010 * put to idle by rebalancing its tasks onto our group.
4012 static struct sched_group
*
4013 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
4014 unsigned long *imbalance
, enum cpu_idle_type idle
,
4015 int *sd_idle
, const struct cpumask
*cpus
, int *balance
)
4017 struct sd_lb_stats sds
;
4019 memset(&sds
, 0, sizeof(sds
));
4022 * Compute the various statistics relavent for load balancing at
4025 update_sd_lb_stats(sd
, this_cpu
, idle
, sd_idle
, cpus
,
4028 /* Cases where imbalance does not exist from POV of this_cpu */
4029 /* 1) this_cpu is not the appropriate cpu to perform load balancing
4031 * 2) There is no busy sibling group to pull from.
4032 * 3) This group is the busiest group.
4033 * 4) This group is more busy than the avg busieness at this
4035 * 5) The imbalance is within the specified limit.
4036 * 6) Any rebalance would lead to ping-pong
4038 if (balance
&& !(*balance
))
4041 if (!sds
.busiest
|| sds
.busiest_nr_running
== 0)
4044 if (sds
.this_load
>= sds
.max_load
)
4047 sds
.avg_load
= (SCHED_LOAD_SCALE
* sds
.total_load
) / sds
.total_pwr
;
4049 if (sds
.this_load
>= sds
.avg_load
)
4052 if (100 * sds
.max_load
<= sd
->imbalance_pct
* sds
.this_load
)
4055 sds
.busiest_load_per_task
/= sds
.busiest_nr_running
;
4057 sds
.busiest_load_per_task
=
4058 min(sds
.busiest_load_per_task
, sds
.avg_load
);
4061 * We're trying to get all the cpus to the average_load, so we don't
4062 * want to push ourselves above the average load, nor do we wish to
4063 * reduce the max loaded cpu below the average load, as either of these
4064 * actions would just result in more rebalancing later, and ping-pong
4065 * tasks around. Thus we look for the minimum possible imbalance.
4066 * Negative imbalances (*we* are more loaded than anyone else) will
4067 * be counted as no imbalance for these purposes -- we can't fix that
4068 * by pulling tasks to us. Be careful of negative numbers as they'll
4069 * appear as very large values with unsigned longs.
4071 if (sds
.max_load
<= sds
.busiest_load_per_task
)
4074 /* Looks like there is an imbalance. Compute it */
4075 calculate_imbalance(&sds
, this_cpu
, imbalance
);
4080 * There is no obvious imbalance. But check if we can do some balancing
4083 if (check_power_save_busiest_group(&sds
, this_cpu
, imbalance
))
4091 * find_busiest_queue - find the busiest runqueue among the cpus in group.
4094 find_busiest_queue(struct sched_group
*group
, enum cpu_idle_type idle
,
4095 unsigned long imbalance
, const struct cpumask
*cpus
)
4097 struct rq
*busiest
= NULL
, *rq
;
4098 unsigned long max_load
= 0;
4101 for_each_cpu(i
, sched_group_cpus(group
)) {
4102 unsigned long power
= power_of(i
);
4103 unsigned long capacity
= DIV_ROUND_CLOSEST(power
, SCHED_LOAD_SCALE
);
4106 if (!cpumask_test_cpu(i
, cpus
))
4110 wl
= weighted_cpuload(i
) * SCHED_LOAD_SCALE
;
4113 if (capacity
&& rq
->nr_running
== 1 && wl
> imbalance
)
4116 if (wl
> max_load
) {
4126 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
4127 * so long as it is large enough.
4129 #define MAX_PINNED_INTERVAL 512
4131 /* Working cpumask for load_balance and load_balance_newidle. */
4132 static DEFINE_PER_CPU(cpumask_var_t
, load_balance_tmpmask
);
4135 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4136 * tasks if there is an imbalance.
4138 static int load_balance(int this_cpu
, struct rq
*this_rq
,
4139 struct sched_domain
*sd
, enum cpu_idle_type idle
,
4142 int ld_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
4143 struct sched_group
*group
;
4144 unsigned long imbalance
;
4146 unsigned long flags
;
4147 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4149 cpumask_copy(cpus
, cpu_active_mask
);
4152 * When power savings policy is enabled for the parent domain, idle
4153 * sibling can pick up load irrespective of busy siblings. In this case,
4154 * let the state of idle sibling percolate up as CPU_IDLE, instead of
4155 * portraying it as CPU_NOT_IDLE.
4157 if (idle
!= CPU_NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4158 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4161 schedstat_inc(sd
, lb_count
[idle
]);
4165 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
4172 schedstat_inc(sd
, lb_nobusyg
[idle
]);
4176 busiest
= find_busiest_queue(group
, idle
, imbalance
, cpus
);
4178 schedstat_inc(sd
, lb_nobusyq
[idle
]);
4182 BUG_ON(busiest
== this_rq
);
4184 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
4187 if (busiest
->nr_running
> 1) {
4189 * Attempt to move tasks. If find_busiest_group has found
4190 * an imbalance but busiest->nr_running <= 1, the group is
4191 * still unbalanced. ld_moved simply stays zero, so it is
4192 * correctly treated as an imbalance.
4194 local_irq_save(flags
);
4195 double_rq_lock(this_rq
, busiest
);
4196 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4197 imbalance
, sd
, idle
, &all_pinned
);
4198 double_rq_unlock(this_rq
, busiest
);
4199 local_irq_restore(flags
);
4202 * some other cpu did the load balance for us.
4204 if (ld_moved
&& this_cpu
!= smp_processor_id())
4205 resched_cpu(this_cpu
);
4207 /* All tasks on this runqueue were pinned by CPU affinity */
4208 if (unlikely(all_pinned
)) {
4209 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4210 if (!cpumask_empty(cpus
))
4217 schedstat_inc(sd
, lb_failed
[idle
]);
4218 sd
->nr_balance_failed
++;
4220 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
4222 raw_spin_lock_irqsave(&busiest
->lock
, flags
);
4224 /* don't kick the migration_thread, if the curr
4225 * task on busiest cpu can't be moved to this_cpu
4227 if (!cpumask_test_cpu(this_cpu
,
4228 &busiest
->curr
->cpus_allowed
)) {
4229 raw_spin_unlock_irqrestore(&busiest
->lock
,
4232 goto out_one_pinned
;
4235 if (!busiest
->active_balance
) {
4236 busiest
->active_balance
= 1;
4237 busiest
->push_cpu
= this_cpu
;
4240 raw_spin_unlock_irqrestore(&busiest
->lock
, flags
);
4242 wake_up_process(busiest
->migration_thread
);
4245 * We've kicked active balancing, reset the failure
4248 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
4251 sd
->nr_balance_failed
= 0;
4253 if (likely(!active_balance
)) {
4254 /* We were unbalanced, so reset the balancing interval */
4255 sd
->balance_interval
= sd
->min_interval
;
4258 * If we've begun active balancing, start to back off. This
4259 * case may not be covered by the all_pinned logic if there
4260 * is only 1 task on the busy runqueue (because we don't call
4263 if (sd
->balance_interval
< sd
->max_interval
)
4264 sd
->balance_interval
*= 2;
4267 if (!ld_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4268 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4274 schedstat_inc(sd
, lb_balanced
[idle
]);
4276 sd
->nr_balance_failed
= 0;
4279 /* tune up the balancing interval */
4280 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
4281 (sd
->balance_interval
< sd
->max_interval
))
4282 sd
->balance_interval
*= 2;
4284 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4285 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4296 * Check this_cpu to ensure it is balanced within domain. Attempt to move
4297 * tasks if there is an imbalance.
4299 * Called from schedule when this_rq is about to become idle (CPU_NEWLY_IDLE).
4300 * this_rq is locked.
4303 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
4305 struct sched_group
*group
;
4306 struct rq
*busiest
= NULL
;
4307 unsigned long imbalance
;
4311 struct cpumask
*cpus
= __get_cpu_var(load_balance_tmpmask
);
4313 cpumask_copy(cpus
, cpu_active_mask
);
4316 * When power savings policy is enabled for the parent domain, idle
4317 * sibling can pick up load irrespective of busy siblings. In this case,
4318 * let the state of idle sibling percolate up as IDLE, instead of
4319 * portraying it as CPU_NOT_IDLE.
4321 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
4322 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4325 schedstat_inc(sd
, lb_count
[CPU_NEWLY_IDLE
]);
4327 update_shares_locked(this_rq
, sd
);
4328 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, CPU_NEWLY_IDLE
,
4329 &sd_idle
, cpus
, NULL
);
4331 schedstat_inc(sd
, lb_nobusyg
[CPU_NEWLY_IDLE
]);
4335 busiest
= find_busiest_queue(group
, CPU_NEWLY_IDLE
, imbalance
, cpus
);
4337 schedstat_inc(sd
, lb_nobusyq
[CPU_NEWLY_IDLE
]);
4341 BUG_ON(busiest
== this_rq
);
4343 schedstat_add(sd
, lb_imbalance
[CPU_NEWLY_IDLE
], imbalance
);
4346 if (busiest
->nr_running
> 1) {
4347 /* Attempt to move tasks */
4348 double_lock_balance(this_rq
, busiest
);
4349 /* this_rq->clock is already updated */
4350 update_rq_clock(busiest
);
4351 ld_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
4352 imbalance
, sd
, CPU_NEWLY_IDLE
,
4354 double_unlock_balance(this_rq
, busiest
);
4356 if (unlikely(all_pinned
)) {
4357 cpumask_clear_cpu(cpu_of(busiest
), cpus
);
4358 if (!cpumask_empty(cpus
))
4364 int active_balance
= 0;
4366 schedstat_inc(sd
, lb_failed
[CPU_NEWLY_IDLE
]);
4367 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4368 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4371 if (sched_mc_power_savings
< POWERSAVINGS_BALANCE_WAKEUP
)
4374 if (sd
->nr_balance_failed
++ < 2)
4378 * The only task running in a non-idle cpu can be moved to this
4379 * cpu in an attempt to completely freeup the other CPU
4380 * package. The same method used to move task in load_balance()
4381 * have been extended for load_balance_newidle() to speedup
4382 * consolidation at sched_mc=POWERSAVINGS_BALANCE_WAKEUP (2)
4384 * The package power saving logic comes from
4385 * find_busiest_group(). If there are no imbalance, then
4386 * f_b_g() will return NULL. However when sched_mc={1,2} then
4387 * f_b_g() will select a group from which a running task may be
4388 * pulled to this cpu in order to make the other package idle.
4389 * If there is no opportunity to make a package idle and if
4390 * there are no imbalance, then f_b_g() will return NULL and no
4391 * action will be taken in load_balance_newidle().
4393 * Under normal task pull operation due to imbalance, there
4394 * will be more than one task in the source run queue and
4395 * move_tasks() will succeed. ld_moved will be true and this
4396 * active balance code will not be triggered.
4399 /* Lock busiest in correct order while this_rq is held */
4400 double_lock_balance(this_rq
, busiest
);
4403 * don't kick the migration_thread, if the curr
4404 * task on busiest cpu can't be moved to this_cpu
4406 if (!cpumask_test_cpu(this_cpu
, &busiest
->curr
->cpus_allowed
)) {
4407 double_unlock_balance(this_rq
, busiest
);
4412 if (!busiest
->active_balance
) {
4413 busiest
->active_balance
= 1;
4414 busiest
->push_cpu
= this_cpu
;
4418 double_unlock_balance(this_rq
, busiest
);
4420 * Should not call ttwu while holding a rq->lock
4422 raw_spin_unlock(&this_rq
->lock
);
4424 wake_up_process(busiest
->migration_thread
);
4425 raw_spin_lock(&this_rq
->lock
);
4428 sd
->nr_balance_failed
= 0;
4430 update_shares_locked(this_rq
, sd
);
4434 schedstat_inc(sd
, lb_balanced
[CPU_NEWLY_IDLE
]);
4435 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
4436 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
4438 sd
->nr_balance_failed
= 0;
4444 * idle_balance is called by schedule() if this_cpu is about to become
4445 * idle. Attempts to pull tasks from other CPUs.
4447 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
4449 struct sched_domain
*sd
;
4450 int pulled_task
= 0;
4451 unsigned long next_balance
= jiffies
+ HZ
;
4453 this_rq
->idle_stamp
= this_rq
->clock
;
4455 if (this_rq
->avg_idle
< sysctl_sched_migration_cost
)
4458 for_each_domain(this_cpu
, sd
) {
4459 unsigned long interval
;
4461 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4464 if (sd
->flags
& SD_BALANCE_NEWIDLE
)
4465 /* If we've pulled tasks over stop searching: */
4466 pulled_task
= load_balance_newidle(this_cpu
, this_rq
,
4469 interval
= msecs_to_jiffies(sd
->balance_interval
);
4470 if (time_after(next_balance
, sd
->last_balance
+ interval
))
4471 next_balance
= sd
->last_balance
+ interval
;
4473 this_rq
->idle_stamp
= 0;
4477 if (pulled_task
|| time_after(jiffies
, this_rq
->next_balance
)) {
4479 * We are going idle. next_balance may be set based on
4480 * a busy processor. So reset next_balance.
4482 this_rq
->next_balance
= next_balance
;
4487 * active_load_balance is run by migration threads. It pushes running tasks
4488 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
4489 * running on each physical CPU where possible, and avoids physical /
4490 * logical imbalances.
4492 * Called with busiest_rq locked.
4494 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
4496 int target_cpu
= busiest_rq
->push_cpu
;
4497 struct sched_domain
*sd
;
4498 struct rq
*target_rq
;
4500 /* Is there any task to move? */
4501 if (busiest_rq
->nr_running
<= 1)
4504 target_rq
= cpu_rq(target_cpu
);
4507 * This condition is "impossible", if it occurs
4508 * we need to fix it. Originally reported by
4509 * Bjorn Helgaas on a 128-cpu setup.
4511 BUG_ON(busiest_rq
== target_rq
);
4513 /* move a task from busiest_rq to target_rq */
4514 double_lock_balance(busiest_rq
, target_rq
);
4515 update_rq_clock(busiest_rq
);
4516 update_rq_clock(target_rq
);
4518 /* Search for an sd spanning us and the target CPU. */
4519 for_each_domain(target_cpu
, sd
) {
4520 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
4521 cpumask_test_cpu(busiest_cpu
, sched_domain_span(sd
)))
4526 schedstat_inc(sd
, alb_count
);
4528 if (move_one_task(target_rq
, target_cpu
, busiest_rq
,
4530 schedstat_inc(sd
, alb_pushed
);
4532 schedstat_inc(sd
, alb_failed
);
4534 double_unlock_balance(busiest_rq
, target_rq
);
4539 atomic_t load_balancer
;
4540 cpumask_var_t cpu_mask
;
4541 cpumask_var_t ilb_grp_nohz_mask
;
4542 } nohz ____cacheline_aligned
= {
4543 .load_balancer
= ATOMIC_INIT(-1),
4546 int get_nohz_load_balancer(void)
4548 return atomic_read(&nohz
.load_balancer
);
4551 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
4553 * lowest_flag_domain - Return lowest sched_domain containing flag.
4554 * @cpu: The cpu whose lowest level of sched domain is to
4556 * @flag: The flag to check for the lowest sched_domain
4557 * for the given cpu.
4559 * Returns the lowest sched_domain of a cpu which contains the given flag.
4561 static inline struct sched_domain
*lowest_flag_domain(int cpu
, int flag
)
4563 struct sched_domain
*sd
;
4565 for_each_domain(cpu
, sd
)
4566 if (sd
&& (sd
->flags
& flag
))
4573 * for_each_flag_domain - Iterates over sched_domains containing the flag.
4574 * @cpu: The cpu whose domains we're iterating over.
4575 * @sd: variable holding the value of the power_savings_sd
4577 * @flag: The flag to filter the sched_domains to be iterated.
4579 * Iterates over all the scheduler domains for a given cpu that has the 'flag'
4580 * set, starting from the lowest sched_domain to the highest.
4582 #define for_each_flag_domain(cpu, sd, flag) \
4583 for (sd = lowest_flag_domain(cpu, flag); \
4584 (sd && (sd->flags & flag)); sd = sd->parent)
4587 * is_semi_idle_group - Checks if the given sched_group is semi-idle.
4588 * @ilb_group: group to be checked for semi-idleness
4590 * Returns: 1 if the group is semi-idle. 0 otherwise.
4592 * We define a sched_group to be semi idle if it has atleast one idle-CPU
4593 * and atleast one non-idle CPU. This helper function checks if the given
4594 * sched_group is semi-idle or not.
4596 static inline int is_semi_idle_group(struct sched_group
*ilb_group
)
4598 cpumask_and(nohz
.ilb_grp_nohz_mask
, nohz
.cpu_mask
,
4599 sched_group_cpus(ilb_group
));
4602 * A sched_group is semi-idle when it has atleast one busy cpu
4603 * and atleast one idle cpu.
4605 if (cpumask_empty(nohz
.ilb_grp_nohz_mask
))
4608 if (cpumask_equal(nohz
.ilb_grp_nohz_mask
, sched_group_cpus(ilb_group
)))
4614 * find_new_ilb - Finds the optimum idle load balancer for nomination.
4615 * @cpu: The cpu which is nominating a new idle_load_balancer.
4617 * Returns: Returns the id of the idle load balancer if it exists,
4618 * Else, returns >= nr_cpu_ids.
4620 * This algorithm picks the idle load balancer such that it belongs to a
4621 * semi-idle powersavings sched_domain. The idea is to try and avoid
4622 * completely idle packages/cores just for the purpose of idle load balancing
4623 * when there are other idle cpu's which are better suited for that job.
4625 static int find_new_ilb(int cpu
)
4627 struct sched_domain
*sd
;
4628 struct sched_group
*ilb_group
;
4631 * Have idle load balancer selection from semi-idle packages only
4632 * when power-aware load balancing is enabled
4634 if (!(sched_smt_power_savings
|| sched_mc_power_savings
))
4638 * Optimize for the case when we have no idle CPUs or only one
4639 * idle CPU. Don't walk the sched_domain hierarchy in such cases
4641 if (cpumask_weight(nohz
.cpu_mask
) < 2)
4644 for_each_flag_domain(cpu
, sd
, SD_POWERSAVINGS_BALANCE
) {
4645 ilb_group
= sd
->groups
;
4648 if (is_semi_idle_group(ilb_group
))
4649 return cpumask_first(nohz
.ilb_grp_nohz_mask
);
4651 ilb_group
= ilb_group
->next
;
4653 } while (ilb_group
!= sd
->groups
);
4657 return cpumask_first(nohz
.cpu_mask
);
4659 #else /* (CONFIG_SCHED_MC || CONFIG_SCHED_SMT) */
4660 static inline int find_new_ilb(int call_cpu
)
4662 return cpumask_first(nohz
.cpu_mask
);
4667 * This routine will try to nominate the ilb (idle load balancing)
4668 * owner among the cpus whose ticks are stopped. ilb owner will do the idle
4669 * load balancing on behalf of all those cpus. If all the cpus in the system
4670 * go into this tickless mode, then there will be no ilb owner (as there is
4671 * no need for one) and all the cpus will sleep till the next wakeup event
4674 * For the ilb owner, tick is not stopped. And this tick will be used
4675 * for idle load balancing. ilb owner will still be part of
4678 * While stopping the tick, this cpu will become the ilb owner if there
4679 * is no other owner. And will be the owner till that cpu becomes busy
4680 * or if all cpus in the system stop their ticks at which point
4681 * there is no need for ilb owner.
4683 * When the ilb owner becomes busy, it nominates another owner, during the
4684 * next busy scheduler_tick()
4686 int select_nohz_load_balancer(int stop_tick
)
4688 int cpu
= smp_processor_id();
4691 cpu_rq(cpu
)->in_nohz_recently
= 1;
4693 if (!cpu_active(cpu
)) {
4694 if (atomic_read(&nohz
.load_balancer
) != cpu
)
4698 * If we are going offline and still the leader,
4701 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4707 cpumask_set_cpu(cpu
, nohz
.cpu_mask
);
4709 /* time for ilb owner also to sleep */
4710 if (cpumask_weight(nohz
.cpu_mask
) == num_active_cpus()) {
4711 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4712 atomic_set(&nohz
.load_balancer
, -1);
4716 if (atomic_read(&nohz
.load_balancer
) == -1) {
4717 /* make me the ilb owner */
4718 if (atomic_cmpxchg(&nohz
.load_balancer
, -1, cpu
) == -1)
4720 } else if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4723 if (!(sched_smt_power_savings
||
4724 sched_mc_power_savings
))
4727 * Check to see if there is a more power-efficient
4730 new_ilb
= find_new_ilb(cpu
);
4731 if (new_ilb
< nr_cpu_ids
&& new_ilb
!= cpu
) {
4732 atomic_set(&nohz
.load_balancer
, -1);
4733 resched_cpu(new_ilb
);
4739 if (!cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4742 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4744 if (atomic_read(&nohz
.load_balancer
) == cpu
)
4745 if (atomic_cmpxchg(&nohz
.load_balancer
, cpu
, -1) != cpu
)
4752 static DEFINE_SPINLOCK(balancing
);
4755 * It checks each scheduling domain to see if it is due to be balanced,
4756 * and initiates a balancing operation if so.
4758 * Balancing parameters are set up in arch_init_sched_domains.
4760 static void rebalance_domains(int cpu
, enum cpu_idle_type idle
)
4763 struct rq
*rq
= cpu_rq(cpu
);
4764 unsigned long interval
;
4765 struct sched_domain
*sd
;
4766 /* Earliest time when we have to do rebalance again */
4767 unsigned long next_balance
= jiffies
+ 60*HZ
;
4768 int update_next_balance
= 0;
4771 for_each_domain(cpu
, sd
) {
4772 if (!(sd
->flags
& SD_LOAD_BALANCE
))
4775 interval
= sd
->balance_interval
;
4776 if (idle
!= CPU_IDLE
)
4777 interval
*= sd
->busy_factor
;
4779 /* scale ms to jiffies */
4780 interval
= msecs_to_jiffies(interval
);
4781 if (unlikely(!interval
))
4783 if (interval
> HZ
*NR_CPUS
/10)
4784 interval
= HZ
*NR_CPUS
/10;
4786 need_serialize
= sd
->flags
& SD_SERIALIZE
;
4788 if (need_serialize
) {
4789 if (!spin_trylock(&balancing
))
4793 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
4794 if (load_balance(cpu
, rq
, sd
, idle
, &balance
)) {
4796 * We've pulled tasks over so either we're no
4797 * longer idle, or one of our SMT siblings is
4800 idle
= CPU_NOT_IDLE
;
4802 sd
->last_balance
= jiffies
;
4805 spin_unlock(&balancing
);
4807 if (time_after(next_balance
, sd
->last_balance
+ interval
)) {
4808 next_balance
= sd
->last_balance
+ interval
;
4809 update_next_balance
= 1;
4813 * Stop the load balance at this level. There is another
4814 * CPU in our sched group which is doing load balancing more
4822 * next_balance will be updated only when there is a need.
4823 * When the cpu is attached to null domain for ex, it will not be
4826 if (likely(update_next_balance
))
4827 rq
->next_balance
= next_balance
;
4831 * run_rebalance_domains is triggered when needed from the scheduler tick.
4832 * In CONFIG_NO_HZ case, the idle load balance owner will do the
4833 * rebalancing for all the cpus for whom scheduler ticks are stopped.
4835 static void run_rebalance_domains(struct softirq_action
*h
)
4837 int this_cpu
= smp_processor_id();
4838 struct rq
*this_rq
= cpu_rq(this_cpu
);
4839 enum cpu_idle_type idle
= this_rq
->idle_at_tick
?
4840 CPU_IDLE
: CPU_NOT_IDLE
;
4842 rebalance_domains(this_cpu
, idle
);
4846 * If this cpu is the owner for idle load balancing, then do the
4847 * balancing on behalf of the other idle cpus whose ticks are
4850 if (this_rq
->idle_at_tick
&&
4851 atomic_read(&nohz
.load_balancer
) == this_cpu
) {
4855 for_each_cpu(balance_cpu
, nohz
.cpu_mask
) {
4856 if (balance_cpu
== this_cpu
)
4860 * If this cpu gets work to do, stop the load balancing
4861 * work being done for other cpus. Next load
4862 * balancing owner will pick it up.
4867 rebalance_domains(balance_cpu
, CPU_IDLE
);
4869 rq
= cpu_rq(balance_cpu
);
4870 if (time_after(this_rq
->next_balance
, rq
->next_balance
))
4871 this_rq
->next_balance
= rq
->next_balance
;
4877 static inline int on_null_domain(int cpu
)
4879 return !rcu_dereference(cpu_rq(cpu
)->sd
);
4883 * Trigger the SCHED_SOFTIRQ if it is time to do periodic load balancing.
4885 * In case of CONFIG_NO_HZ, this is the place where we nominate a new
4886 * idle load balancing owner or decide to stop the periodic load balancing,
4887 * if the whole system is idle.
4889 static inline void trigger_load_balance(struct rq
*rq
, int cpu
)
4893 * If we were in the nohz mode recently and busy at the current
4894 * scheduler tick, then check if we need to nominate new idle
4897 if (rq
->in_nohz_recently
&& !rq
->idle_at_tick
) {
4898 rq
->in_nohz_recently
= 0;
4900 if (atomic_read(&nohz
.load_balancer
) == cpu
) {
4901 cpumask_clear_cpu(cpu
, nohz
.cpu_mask
);
4902 atomic_set(&nohz
.load_balancer
, -1);
4905 if (atomic_read(&nohz
.load_balancer
) == -1) {
4906 int ilb
= find_new_ilb(cpu
);
4908 if (ilb
< nr_cpu_ids
)
4914 * If this cpu is idle and doing idle load balancing for all the
4915 * cpus with ticks stopped, is it time for that to stop?
4917 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) == cpu
&&
4918 cpumask_weight(nohz
.cpu_mask
) == num_online_cpus()) {
4924 * If this cpu is idle and the idle load balancing is done by
4925 * someone else, then no need raise the SCHED_SOFTIRQ
4927 if (rq
->idle_at_tick
&& atomic_read(&nohz
.load_balancer
) != cpu
&&
4928 cpumask_test_cpu(cpu
, nohz
.cpu_mask
))
4931 /* Don't need to rebalance while attached to NULL domain */
4932 if (time_after_eq(jiffies
, rq
->next_balance
) &&
4933 likely(!on_null_domain(cpu
)))
4934 raise_softirq(SCHED_SOFTIRQ
);
4937 #else /* CONFIG_SMP */
4940 * on UP we do not need to balance between CPUs:
4942 static inline void idle_balance(int cpu
, struct rq
*rq
)
4948 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
4950 EXPORT_PER_CPU_SYMBOL(kstat
);
4953 * Return any ns on the sched_clock that have not yet been accounted in
4954 * @p in case that task is currently running.
4956 * Called with task_rq_lock() held on @rq.
4958 static u64
do_task_delta_exec(struct task_struct
*p
, struct rq
*rq
)
4962 if (task_current(rq
, p
)) {
4963 update_rq_clock(rq
);
4964 ns
= rq
->clock
- p
->se
.exec_start
;
4972 unsigned long long task_delta_exec(struct task_struct
*p
)
4974 unsigned long flags
;
4978 rq
= task_rq_lock(p
, &flags
);
4979 ns
= do_task_delta_exec(p
, rq
);
4980 task_rq_unlock(rq
, &flags
);
4986 * Return accounted runtime for the task.
4987 * In case the task is currently running, return the runtime plus current's
4988 * pending runtime that have not been accounted yet.
4990 unsigned long long task_sched_runtime(struct task_struct
*p
)
4992 unsigned long flags
;
4996 rq
= task_rq_lock(p
, &flags
);
4997 ns
= p
->se
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
4998 task_rq_unlock(rq
, &flags
);
5004 * Return sum_exec_runtime for the thread group.
5005 * In case the task is currently running, return the sum plus current's
5006 * pending runtime that have not been accounted yet.
5008 * Note that the thread group might have other running tasks as well,
5009 * so the return value not includes other pending runtime that other
5010 * running tasks might have.
5012 unsigned long long thread_group_sched_runtime(struct task_struct
*p
)
5014 struct task_cputime totals
;
5015 unsigned long flags
;
5019 rq
= task_rq_lock(p
, &flags
);
5020 thread_group_cputime(p
, &totals
);
5021 ns
= totals
.sum_exec_runtime
+ do_task_delta_exec(p
, rq
);
5022 task_rq_unlock(rq
, &flags
);
5028 * Account user cpu time to a process.
5029 * @p: the process that the cpu time gets accounted to
5030 * @cputime: the cpu time spent in user space since the last update
5031 * @cputime_scaled: cputime scaled by cpu frequency
5033 void account_user_time(struct task_struct
*p
, cputime_t cputime
,
5034 cputime_t cputime_scaled
)
5036 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5039 /* Add user time to process. */
5040 p
->utime
= cputime_add(p
->utime
, cputime
);
5041 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5042 account_group_user_time(p
, cputime
);
5044 /* Add user time to cpustat. */
5045 tmp
= cputime_to_cputime64(cputime
);
5046 if (TASK_NICE(p
) > 0)
5047 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5049 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5051 cpuacct_update_stats(p
, CPUACCT_STAT_USER
, cputime
);
5052 /* Account for user time used */
5053 acct_update_integrals(p
);
5057 * Account guest cpu time to a process.
5058 * @p: the process that the cpu time gets accounted to
5059 * @cputime: the cpu time spent in virtual machine since the last update
5060 * @cputime_scaled: cputime scaled by cpu frequency
5062 static void account_guest_time(struct task_struct
*p
, cputime_t cputime
,
5063 cputime_t cputime_scaled
)
5066 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5068 tmp
= cputime_to_cputime64(cputime
);
5070 /* Add guest time to process. */
5071 p
->utime
= cputime_add(p
->utime
, cputime
);
5072 p
->utimescaled
= cputime_add(p
->utimescaled
, cputime_scaled
);
5073 account_group_user_time(p
, cputime
);
5074 p
->gtime
= cputime_add(p
->gtime
, cputime
);
5076 /* Add guest time to cpustat. */
5077 if (TASK_NICE(p
) > 0) {
5078 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
5079 cpustat
->guest_nice
= cputime64_add(cpustat
->guest_nice
, tmp
);
5081 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
5082 cpustat
->guest
= cputime64_add(cpustat
->guest
, tmp
);
5087 * Account system cpu time to a process.
5088 * @p: the process that the cpu time gets accounted to
5089 * @hardirq_offset: the offset to subtract from hardirq_count()
5090 * @cputime: the cpu time spent in kernel space since the last update
5091 * @cputime_scaled: cputime scaled by cpu frequency
5093 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
5094 cputime_t cputime
, cputime_t cputime_scaled
)
5096 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5099 if ((p
->flags
& PF_VCPU
) && (irq_count() - hardirq_offset
== 0)) {
5100 account_guest_time(p
, cputime
, cputime_scaled
);
5104 /* Add system time to process. */
5105 p
->stime
= cputime_add(p
->stime
, cputime
);
5106 p
->stimescaled
= cputime_add(p
->stimescaled
, cputime_scaled
);
5107 account_group_system_time(p
, cputime
);
5109 /* Add system time to cpustat. */
5110 tmp
= cputime_to_cputime64(cputime
);
5111 if (hardirq_count() - hardirq_offset
)
5112 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
5113 else if (softirq_count())
5114 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
5116 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
5118 cpuacct_update_stats(p
, CPUACCT_STAT_SYSTEM
, cputime
);
5120 /* Account for system time used */
5121 acct_update_integrals(p
);
5125 * Account for involuntary wait time.
5126 * @steal: the cpu time spent in involuntary wait
5128 void account_steal_time(cputime_t cputime
)
5130 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5131 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5133 cpustat
->steal
= cputime64_add(cpustat
->steal
, cputime64
);
5137 * Account for idle time.
5138 * @cputime: the cpu time spent in idle wait
5140 void account_idle_time(cputime_t cputime
)
5142 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
5143 cputime64_t cputime64
= cputime_to_cputime64(cputime
);
5144 struct rq
*rq
= this_rq();
5146 if (atomic_read(&rq
->nr_iowait
) > 0)
5147 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, cputime64
);
5149 cpustat
->idle
= cputime64_add(cpustat
->idle
, cputime64
);
5152 #ifndef CONFIG_VIRT_CPU_ACCOUNTING
5155 * Account a single tick of cpu time.
5156 * @p: the process that the cpu time gets accounted to
5157 * @user_tick: indicates if the tick is a user or a system tick
5159 void account_process_tick(struct task_struct
*p
, int user_tick
)
5161 cputime_t one_jiffy_scaled
= cputime_to_scaled(cputime_one_jiffy
);
5162 struct rq
*rq
= this_rq();
5165 account_user_time(p
, cputime_one_jiffy
, one_jiffy_scaled
);
5166 else if ((p
!= rq
->idle
) || (irq_count() != HARDIRQ_OFFSET
))
5167 account_system_time(p
, HARDIRQ_OFFSET
, cputime_one_jiffy
,
5170 account_idle_time(cputime_one_jiffy
);
5174 * Account multiple ticks of steal time.
5175 * @p: the process from which the cpu time has been stolen
5176 * @ticks: number of stolen ticks
5178 void account_steal_ticks(unsigned long ticks
)
5180 account_steal_time(jiffies_to_cputime(ticks
));
5184 * Account multiple ticks of idle time.
5185 * @ticks: number of stolen ticks
5187 void account_idle_ticks(unsigned long ticks
)
5189 account_idle_time(jiffies_to_cputime(ticks
));
5195 * Use precise platform statistics if available:
5197 #ifdef CONFIG_VIRT_CPU_ACCOUNTING
5198 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5204 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5206 struct task_cputime cputime
;
5208 thread_group_cputime(p
, &cputime
);
5210 *ut
= cputime
.utime
;
5211 *st
= cputime
.stime
;
5215 #ifndef nsecs_to_cputime
5216 # define nsecs_to_cputime(__nsecs) nsecs_to_jiffies(__nsecs)
5219 void task_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5221 cputime_t rtime
, utime
= p
->utime
, total
= cputime_add(utime
, p
->stime
);
5224 * Use CFS's precise accounting:
5226 rtime
= nsecs_to_cputime(p
->se
.sum_exec_runtime
);
5231 temp
= (u64
)(rtime
* utime
);
5232 do_div(temp
, total
);
5233 utime
= (cputime_t
)temp
;
5238 * Compare with previous values, to keep monotonicity:
5240 p
->prev_utime
= max(p
->prev_utime
, utime
);
5241 p
->prev_stime
= max(p
->prev_stime
, cputime_sub(rtime
, p
->prev_utime
));
5243 *ut
= p
->prev_utime
;
5244 *st
= p
->prev_stime
;
5248 * Must be called with siglock held.
5250 void thread_group_times(struct task_struct
*p
, cputime_t
*ut
, cputime_t
*st
)
5252 struct signal_struct
*sig
= p
->signal
;
5253 struct task_cputime cputime
;
5254 cputime_t rtime
, utime
, total
;
5256 thread_group_cputime(p
, &cputime
);
5258 total
= cputime_add(cputime
.utime
, cputime
.stime
);
5259 rtime
= nsecs_to_cputime(cputime
.sum_exec_runtime
);
5264 temp
= (u64
)(rtime
* cputime
.utime
);
5265 do_div(temp
, total
);
5266 utime
= (cputime_t
)temp
;
5270 sig
->prev_utime
= max(sig
->prev_utime
, utime
);
5271 sig
->prev_stime
= max(sig
->prev_stime
,
5272 cputime_sub(rtime
, sig
->prev_utime
));
5274 *ut
= sig
->prev_utime
;
5275 *st
= sig
->prev_stime
;
5280 * This function gets called by the timer code, with HZ frequency.
5281 * We call it with interrupts disabled.
5283 * It also gets called by the fork code, when changing the parent's
5286 void scheduler_tick(void)
5288 int cpu
= smp_processor_id();
5289 struct rq
*rq
= cpu_rq(cpu
);
5290 struct task_struct
*curr
= rq
->curr
;
5294 raw_spin_lock(&rq
->lock
);
5295 update_rq_clock(rq
);
5296 update_cpu_load(rq
);
5297 curr
->sched_class
->task_tick(rq
, curr
, 0);
5298 raw_spin_unlock(&rq
->lock
);
5300 perf_event_task_tick(curr
, cpu
);
5303 rq
->idle_at_tick
= idle_cpu(cpu
);
5304 trigger_load_balance(rq
, cpu
);
5308 notrace
unsigned long get_parent_ip(unsigned long addr
)
5310 if (in_lock_functions(addr
)) {
5311 addr
= CALLER_ADDR2
;
5312 if (in_lock_functions(addr
))
5313 addr
= CALLER_ADDR3
;
5318 #if defined(CONFIG_PREEMPT) && (defined(CONFIG_DEBUG_PREEMPT) || \
5319 defined(CONFIG_PREEMPT_TRACER))
5321 void __kprobes
add_preempt_count(int val
)
5323 #ifdef CONFIG_DEBUG_PREEMPT
5327 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
5330 preempt_count() += val
;
5331 #ifdef CONFIG_DEBUG_PREEMPT
5333 * Spinlock count overflowing soon?
5335 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
5338 if (preempt_count() == val
)
5339 trace_preempt_off(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5341 EXPORT_SYMBOL(add_preempt_count
);
5343 void __kprobes
sub_preempt_count(int val
)
5345 #ifdef CONFIG_DEBUG_PREEMPT
5349 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
5352 * Is the spinlock portion underflowing?
5354 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
5355 !(preempt_count() & PREEMPT_MASK
)))
5359 if (preempt_count() == val
)
5360 trace_preempt_on(CALLER_ADDR0
, get_parent_ip(CALLER_ADDR1
));
5361 preempt_count() -= val
;
5363 EXPORT_SYMBOL(sub_preempt_count
);
5368 * Print scheduling while atomic bug:
5370 static noinline
void __schedule_bug(struct task_struct
*prev
)
5372 struct pt_regs
*regs
= get_irq_regs();
5374 pr_err("BUG: scheduling while atomic: %s/%d/0x%08x\n",
5375 prev
->comm
, prev
->pid
, preempt_count());
5377 debug_show_held_locks(prev
);
5379 if (irqs_disabled())
5380 print_irqtrace_events(prev
);
5389 * Various schedule()-time debugging checks and statistics:
5391 static inline void schedule_debug(struct task_struct
*prev
)
5394 * Test if we are atomic. Since do_exit() needs to call into
5395 * schedule() atomically, we ignore that path for now.
5396 * Otherwise, whine if we are scheduling when we should not be.
5398 if (unlikely(in_atomic_preempt_off() && !prev
->exit_state
))
5399 __schedule_bug(prev
);
5401 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
5403 schedstat_inc(this_rq(), sched_count
);
5404 #ifdef CONFIG_SCHEDSTATS
5405 if (unlikely(prev
->lock_depth
>= 0)) {
5406 schedstat_inc(this_rq(), bkl_count
);
5407 schedstat_inc(prev
, sched_info
.bkl_count
);
5412 static void put_prev_task(struct rq
*rq
, struct task_struct
*prev
)
5414 if (prev
->state
== TASK_RUNNING
) {
5415 u64 runtime
= prev
->se
.sum_exec_runtime
;
5417 runtime
-= prev
->se
.prev_sum_exec_runtime
;
5418 runtime
= min_t(u64
, runtime
, 2*sysctl_sched_migration_cost
);
5421 * In order to avoid avg_overlap growing stale when we are
5422 * indeed overlapping and hence not getting put to sleep, grow
5423 * the avg_overlap on preemption.
5425 * We use the average preemption runtime because that
5426 * correlates to the amount of cache footprint a task can
5429 update_avg(&prev
->se
.avg_overlap
, runtime
);
5431 prev
->sched_class
->put_prev_task(rq
, prev
);
5435 * Pick up the highest-prio task:
5437 static inline struct task_struct
*
5438 pick_next_task(struct rq
*rq
)
5440 const struct sched_class
*class;
5441 struct task_struct
*p
;
5444 * Optimization: we know that if all tasks are in
5445 * the fair class we can call that function directly:
5447 if (likely(rq
->nr_running
== rq
->cfs
.nr_running
)) {
5448 p
= fair_sched_class
.pick_next_task(rq
);
5453 class = sched_class_highest
;
5455 p
= class->pick_next_task(rq
);
5459 * Will never be NULL as the idle class always
5460 * returns a non-NULL p:
5462 class = class->next
;
5467 * schedule() is the main scheduler function.
5469 asmlinkage
void __sched
schedule(void)
5471 struct task_struct
*prev
, *next
;
5472 unsigned long *switch_count
;
5478 cpu
= smp_processor_id();
5482 switch_count
= &prev
->nivcsw
;
5484 release_kernel_lock(prev
);
5485 need_resched_nonpreemptible
:
5487 schedule_debug(prev
);
5489 if (sched_feat(HRTICK
))
5492 raw_spin_lock_irq(&rq
->lock
);
5493 update_rq_clock(rq
);
5494 clear_tsk_need_resched(prev
);
5496 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
5497 if (unlikely(signal_pending_state(prev
->state
, prev
)))
5498 prev
->state
= TASK_RUNNING
;
5500 deactivate_task(rq
, prev
, 1);
5501 switch_count
= &prev
->nvcsw
;
5504 pre_schedule(rq
, prev
);
5506 if (unlikely(!rq
->nr_running
))
5507 idle_balance(cpu
, rq
);
5509 put_prev_task(rq
, prev
);
5510 next
= pick_next_task(rq
);
5512 if (likely(prev
!= next
)) {
5513 sched_info_switch(prev
, next
);
5514 perf_event_task_sched_out(prev
, next
, cpu
);
5520 context_switch(rq
, prev
, next
); /* unlocks the rq */
5522 * the context switch might have flipped the stack from under
5523 * us, hence refresh the local variables.
5525 cpu
= smp_processor_id();
5528 raw_spin_unlock_irq(&rq
->lock
);
5532 if (unlikely(reacquire_kernel_lock(current
) < 0))
5533 goto need_resched_nonpreemptible
;
5535 preempt_enable_no_resched();
5539 EXPORT_SYMBOL(schedule
);
5541 #ifdef CONFIG_MUTEX_SPIN_ON_OWNER
5543 * Look out! "owner" is an entirely speculative pointer
5544 * access and not reliable.
5546 int mutex_spin_on_owner(struct mutex
*lock
, struct thread_info
*owner
)
5551 if (!sched_feat(OWNER_SPIN
))
5554 #ifdef CONFIG_DEBUG_PAGEALLOC
5556 * Need to access the cpu field knowing that
5557 * DEBUG_PAGEALLOC could have unmapped it if
5558 * the mutex owner just released it and exited.
5560 if (probe_kernel_address(&owner
->cpu
, cpu
))
5567 * Even if the access succeeded (likely case),
5568 * the cpu field may no longer be valid.
5570 if (cpu
>= nr_cpumask_bits
)
5574 * We need to validate that we can do a
5575 * get_cpu() and that we have the percpu area.
5577 if (!cpu_online(cpu
))
5584 * Owner changed, break to re-assess state.
5586 if (lock
->owner
!= owner
)
5590 * Is that owner really running on that cpu?
5592 if (task_thread_info(rq
->curr
) != owner
|| need_resched())
5602 #ifdef CONFIG_PREEMPT
5604 * this is the entry point to schedule() from in-kernel preemption
5605 * off of preempt_enable. Kernel preemptions off return from interrupt
5606 * occur there and call schedule directly.
5608 asmlinkage
void __sched
preempt_schedule(void)
5610 struct thread_info
*ti
= current_thread_info();
5613 * If there is a non-zero preempt_count or interrupts are disabled,
5614 * we do not want to preempt the current task. Just return..
5616 if (likely(ti
->preempt_count
|| irqs_disabled()))
5620 add_preempt_count(PREEMPT_ACTIVE
);
5622 sub_preempt_count(PREEMPT_ACTIVE
);
5625 * Check again in case we missed a preemption opportunity
5626 * between schedule and now.
5629 } while (need_resched());
5631 EXPORT_SYMBOL(preempt_schedule
);
5634 * this is the entry point to schedule() from kernel preemption
5635 * off of irq context.
5636 * Note, that this is called and return with irqs disabled. This will
5637 * protect us against recursive calling from irq.
5639 asmlinkage
void __sched
preempt_schedule_irq(void)
5641 struct thread_info
*ti
= current_thread_info();
5643 /* Catch callers which need to be fixed */
5644 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
5647 add_preempt_count(PREEMPT_ACTIVE
);
5650 local_irq_disable();
5651 sub_preempt_count(PREEMPT_ACTIVE
);
5654 * Check again in case we missed a preemption opportunity
5655 * between schedule and now.
5658 } while (need_resched());
5661 #endif /* CONFIG_PREEMPT */
5663 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int wake_flags
,
5666 return try_to_wake_up(curr
->private, mode
, wake_flags
);
5668 EXPORT_SYMBOL(default_wake_function
);
5671 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
5672 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
5673 * number) then we wake all the non-exclusive tasks and one exclusive task.
5675 * There are circumstances in which we can try to wake a task which has already
5676 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
5677 * zero in this (rare) case, and we handle it by continuing to scan the queue.
5679 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
5680 int nr_exclusive
, int wake_flags
, void *key
)
5682 wait_queue_t
*curr
, *next
;
5684 list_for_each_entry_safe(curr
, next
, &q
->task_list
, task_list
) {
5685 unsigned flags
= curr
->flags
;
5687 if (curr
->func(curr
, mode
, wake_flags
, key
) &&
5688 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
5694 * __wake_up - wake up threads blocked on a waitqueue.
5696 * @mode: which threads
5697 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5698 * @key: is directly passed to the wakeup function
5700 * It may be assumed that this function implies a write memory barrier before
5701 * changing the task state if and only if any tasks are woken up.
5703 void __wake_up(wait_queue_head_t
*q
, unsigned int mode
,
5704 int nr_exclusive
, void *key
)
5706 unsigned long flags
;
5708 spin_lock_irqsave(&q
->lock
, flags
);
5709 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
5710 spin_unlock_irqrestore(&q
->lock
, flags
);
5712 EXPORT_SYMBOL(__wake_up
);
5715 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
5717 void __wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
5719 __wake_up_common(q
, mode
, 1, 0, NULL
);
5722 void __wake_up_locked_key(wait_queue_head_t
*q
, unsigned int mode
, void *key
)
5724 __wake_up_common(q
, mode
, 1, 0, key
);
5728 * __wake_up_sync_key - wake up threads blocked on a waitqueue.
5730 * @mode: which threads
5731 * @nr_exclusive: how many wake-one or wake-many threads to wake up
5732 * @key: opaque value to be passed to wakeup targets
5734 * The sync wakeup differs that the waker knows that it will schedule
5735 * away soon, so while the target thread will be woken up, it will not
5736 * be migrated to another CPU - ie. the two threads are 'synchronized'
5737 * with each other. This can prevent needless bouncing between CPUs.
5739 * On UP it can prevent extra preemption.
5741 * It may be assumed that this function implies a write memory barrier before
5742 * changing the task state if and only if any tasks are woken up.
5744 void __wake_up_sync_key(wait_queue_head_t
*q
, unsigned int mode
,
5745 int nr_exclusive
, void *key
)
5747 unsigned long flags
;
5748 int wake_flags
= WF_SYNC
;
5753 if (unlikely(!nr_exclusive
))
5756 spin_lock_irqsave(&q
->lock
, flags
);
5757 __wake_up_common(q
, mode
, nr_exclusive
, wake_flags
, key
);
5758 spin_unlock_irqrestore(&q
->lock
, flags
);
5760 EXPORT_SYMBOL_GPL(__wake_up_sync_key
);
5763 * __wake_up_sync - see __wake_up_sync_key()
5765 void __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
5767 __wake_up_sync_key(q
, mode
, nr_exclusive
, NULL
);
5769 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
5772 * complete: - signals a single thread waiting on this completion
5773 * @x: holds the state of this particular completion
5775 * This will wake up a single thread waiting on this completion. Threads will be
5776 * awakened in the same order in which they were queued.
5778 * See also complete_all(), wait_for_completion() and related routines.
5780 * It may be assumed that this function implies a write memory barrier before
5781 * changing the task state if and only if any tasks are woken up.
5783 void complete(struct completion
*x
)
5785 unsigned long flags
;
5787 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5789 __wake_up_common(&x
->wait
, TASK_NORMAL
, 1, 0, NULL
);
5790 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5792 EXPORT_SYMBOL(complete
);
5795 * complete_all: - signals all threads waiting on this completion
5796 * @x: holds the state of this particular completion
5798 * This will wake up all threads waiting on this particular completion event.
5800 * It may be assumed that this function implies a write memory barrier before
5801 * changing the task state if and only if any tasks are woken up.
5803 void complete_all(struct completion
*x
)
5805 unsigned long flags
;
5807 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5808 x
->done
+= UINT_MAX
/2;
5809 __wake_up_common(&x
->wait
, TASK_NORMAL
, 0, 0, NULL
);
5810 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5812 EXPORT_SYMBOL(complete_all
);
5814 static inline long __sched
5815 do_wait_for_common(struct completion
*x
, long timeout
, int state
)
5818 DECLARE_WAITQUEUE(wait
, current
);
5820 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
5821 __add_wait_queue_tail(&x
->wait
, &wait
);
5823 if (signal_pending_state(state
, current
)) {
5824 timeout
= -ERESTARTSYS
;
5827 __set_current_state(state
);
5828 spin_unlock_irq(&x
->wait
.lock
);
5829 timeout
= schedule_timeout(timeout
);
5830 spin_lock_irq(&x
->wait
.lock
);
5831 } while (!x
->done
&& timeout
);
5832 __remove_wait_queue(&x
->wait
, &wait
);
5837 return timeout
?: 1;
5841 wait_for_common(struct completion
*x
, long timeout
, int state
)
5845 spin_lock_irq(&x
->wait
.lock
);
5846 timeout
= do_wait_for_common(x
, timeout
, state
);
5847 spin_unlock_irq(&x
->wait
.lock
);
5852 * wait_for_completion: - waits for completion of a task
5853 * @x: holds the state of this particular completion
5855 * This waits to be signaled for completion of a specific task. It is NOT
5856 * interruptible and there is no timeout.
5858 * See also similar routines (i.e. wait_for_completion_timeout()) with timeout
5859 * and interrupt capability. Also see complete().
5861 void __sched
wait_for_completion(struct completion
*x
)
5863 wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_UNINTERRUPTIBLE
);
5865 EXPORT_SYMBOL(wait_for_completion
);
5868 * wait_for_completion_timeout: - waits for completion of a task (w/timeout)
5869 * @x: holds the state of this particular completion
5870 * @timeout: timeout value in jiffies
5872 * This waits for either a completion of a specific task to be signaled or for a
5873 * specified timeout to expire. The timeout is in jiffies. It is not
5876 unsigned long __sched
5877 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
5879 return wait_for_common(x
, timeout
, TASK_UNINTERRUPTIBLE
);
5881 EXPORT_SYMBOL(wait_for_completion_timeout
);
5884 * wait_for_completion_interruptible: - waits for completion of a task (w/intr)
5885 * @x: holds the state of this particular completion
5887 * This waits for completion of a specific task to be signaled. It is
5890 int __sched
wait_for_completion_interruptible(struct completion
*x
)
5892 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_INTERRUPTIBLE
);
5893 if (t
== -ERESTARTSYS
)
5897 EXPORT_SYMBOL(wait_for_completion_interruptible
);
5900 * wait_for_completion_interruptible_timeout: - waits for completion (w/(to,intr))
5901 * @x: holds the state of this particular completion
5902 * @timeout: timeout value in jiffies
5904 * This waits for either a completion of a specific task to be signaled or for a
5905 * specified timeout to expire. It is interruptible. The timeout is in jiffies.
5907 unsigned long __sched
5908 wait_for_completion_interruptible_timeout(struct completion
*x
,
5909 unsigned long timeout
)
5911 return wait_for_common(x
, timeout
, TASK_INTERRUPTIBLE
);
5913 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
5916 * wait_for_completion_killable: - waits for completion of a task (killable)
5917 * @x: holds the state of this particular completion
5919 * This waits to be signaled for completion of a specific task. It can be
5920 * interrupted by a kill signal.
5922 int __sched
wait_for_completion_killable(struct completion
*x
)
5924 long t
= wait_for_common(x
, MAX_SCHEDULE_TIMEOUT
, TASK_KILLABLE
);
5925 if (t
== -ERESTARTSYS
)
5929 EXPORT_SYMBOL(wait_for_completion_killable
);
5932 * try_wait_for_completion - try to decrement a completion without blocking
5933 * @x: completion structure
5935 * Returns: 0 if a decrement cannot be done without blocking
5936 * 1 if a decrement succeeded.
5938 * If a completion is being used as a counting completion,
5939 * attempt to decrement the counter without blocking. This
5940 * enables us to avoid waiting if the resource the completion
5941 * is protecting is not available.
5943 bool try_wait_for_completion(struct completion
*x
)
5945 unsigned long flags
;
5948 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5953 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5956 EXPORT_SYMBOL(try_wait_for_completion
);
5959 * completion_done - Test to see if a completion has any waiters
5960 * @x: completion structure
5962 * Returns: 0 if there are waiters (wait_for_completion() in progress)
5963 * 1 if there are no waiters.
5966 bool completion_done(struct completion
*x
)
5968 unsigned long flags
;
5971 spin_lock_irqsave(&x
->wait
.lock
, flags
);
5974 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
5977 EXPORT_SYMBOL(completion_done
);
5980 sleep_on_common(wait_queue_head_t
*q
, int state
, long timeout
)
5982 unsigned long flags
;
5985 init_waitqueue_entry(&wait
, current
);
5987 __set_current_state(state
);
5989 spin_lock_irqsave(&q
->lock
, flags
);
5990 __add_wait_queue(q
, &wait
);
5991 spin_unlock(&q
->lock
);
5992 timeout
= schedule_timeout(timeout
);
5993 spin_lock_irq(&q
->lock
);
5994 __remove_wait_queue(q
, &wait
);
5995 spin_unlock_irqrestore(&q
->lock
, flags
);
6000 void __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
6002 sleep_on_common(q
, TASK_INTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6004 EXPORT_SYMBOL(interruptible_sleep_on
);
6007 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6009 return sleep_on_common(q
, TASK_INTERRUPTIBLE
, timeout
);
6011 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
6013 void __sched
sleep_on(wait_queue_head_t
*q
)
6015 sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, MAX_SCHEDULE_TIMEOUT
);
6017 EXPORT_SYMBOL(sleep_on
);
6019 long __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
6021 return sleep_on_common(q
, TASK_UNINTERRUPTIBLE
, timeout
);
6023 EXPORT_SYMBOL(sleep_on_timeout
);
6025 #ifdef CONFIG_RT_MUTEXES
6028 * rt_mutex_setprio - set the current priority of a task
6030 * @prio: prio value (kernel-internal form)
6032 * This function changes the 'effective' priority of a task. It does
6033 * not touch ->normal_prio like __setscheduler().
6035 * Used by the rt_mutex code to implement priority inheritance logic.
6037 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
6039 unsigned long flags
;
6040 int oldprio
, on_rq
, running
;
6042 const struct sched_class
*prev_class
= p
->sched_class
;
6044 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
6046 rq
= task_rq_lock(p
, &flags
);
6047 update_rq_clock(rq
);
6050 on_rq
= p
->se
.on_rq
;
6051 running
= task_current(rq
, p
);
6053 dequeue_task(rq
, p
, 0);
6055 p
->sched_class
->put_prev_task(rq
, p
);
6058 p
->sched_class
= &rt_sched_class
;
6060 p
->sched_class
= &fair_sched_class
;
6065 p
->sched_class
->set_curr_task(rq
);
6067 enqueue_task(rq
, p
, 0);
6069 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6071 task_rq_unlock(rq
, &flags
);
6076 void set_user_nice(struct task_struct
*p
, long nice
)
6078 int old_prio
, delta
, on_rq
;
6079 unsigned long flags
;
6082 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
6085 * We have to be careful, if called from sys_setpriority(),
6086 * the task might be in the middle of scheduling on another CPU.
6088 rq
= task_rq_lock(p
, &flags
);
6089 update_rq_clock(rq
);
6091 * The RT priorities are set via sched_setscheduler(), but we still
6092 * allow the 'normal' nice value to be set - but as expected
6093 * it wont have any effect on scheduling until the task is
6094 * SCHED_FIFO/SCHED_RR:
6096 if (task_has_rt_policy(p
)) {
6097 p
->static_prio
= NICE_TO_PRIO(nice
);
6100 on_rq
= p
->se
.on_rq
;
6102 dequeue_task(rq
, p
, 0);
6104 p
->static_prio
= NICE_TO_PRIO(nice
);
6107 p
->prio
= effective_prio(p
);
6108 delta
= p
->prio
- old_prio
;
6111 enqueue_task(rq
, p
, 0);
6113 * If the task increased its priority or is running and
6114 * lowered its priority, then reschedule its CPU:
6116 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
6117 resched_task(rq
->curr
);
6120 task_rq_unlock(rq
, &flags
);
6122 EXPORT_SYMBOL(set_user_nice
);
6125 * can_nice - check if a task can reduce its nice value
6129 int can_nice(const struct task_struct
*p
, const int nice
)
6131 /* convert nice value [19,-20] to rlimit style value [1,40] */
6132 int nice_rlim
= 20 - nice
;
6134 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
6135 capable(CAP_SYS_NICE
));
6138 #ifdef __ARCH_WANT_SYS_NICE
6141 * sys_nice - change the priority of the current process.
6142 * @increment: priority increment
6144 * sys_setpriority is a more generic, but much slower function that
6145 * does similar things.
6147 SYSCALL_DEFINE1(nice
, int, increment
)
6152 * Setpriority might change our priority at the same moment.
6153 * We don't have to worry. Conceptually one call occurs first
6154 * and we have a single winner.
6156 if (increment
< -40)
6161 nice
= TASK_NICE(current
) + increment
;
6167 if (increment
< 0 && !can_nice(current
, nice
))
6170 retval
= security_task_setnice(current
, nice
);
6174 set_user_nice(current
, nice
);
6181 * task_prio - return the priority value of a given task.
6182 * @p: the task in question.
6184 * This is the priority value as seen by users in /proc.
6185 * RT tasks are offset by -200. Normal tasks are centered
6186 * around 0, value goes from -16 to +15.
6188 int task_prio(const struct task_struct
*p
)
6190 return p
->prio
- MAX_RT_PRIO
;
6194 * task_nice - return the nice value of a given task.
6195 * @p: the task in question.
6197 int task_nice(const struct task_struct
*p
)
6199 return TASK_NICE(p
);
6201 EXPORT_SYMBOL(task_nice
);
6204 * idle_cpu - is a given cpu idle currently?
6205 * @cpu: the processor in question.
6207 int idle_cpu(int cpu
)
6209 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
6213 * idle_task - return the idle task for a given cpu.
6214 * @cpu: the processor in question.
6216 struct task_struct
*idle_task(int cpu
)
6218 return cpu_rq(cpu
)->idle
;
6222 * find_process_by_pid - find a process with a matching PID value.
6223 * @pid: the pid in question.
6225 static struct task_struct
*find_process_by_pid(pid_t pid
)
6227 return pid
? find_task_by_vpid(pid
) : current
;
6230 /* Actually do priority change: must hold rq lock. */
6232 __setscheduler(struct rq
*rq
, struct task_struct
*p
, int policy
, int prio
)
6234 BUG_ON(p
->se
.on_rq
);
6237 p
->rt_priority
= prio
;
6238 p
->normal_prio
= normal_prio(p
);
6239 /* we are holding p->pi_lock already */
6240 p
->prio
= rt_mutex_getprio(p
);
6241 if (rt_prio(p
->prio
))
6242 p
->sched_class
= &rt_sched_class
;
6244 p
->sched_class
= &fair_sched_class
;
6249 * check the target process has a UID that matches the current process's
6251 static bool check_same_owner(struct task_struct
*p
)
6253 const struct cred
*cred
= current_cred(), *pcred
;
6257 pcred
= __task_cred(p
);
6258 match
= (cred
->euid
== pcred
->euid
||
6259 cred
->euid
== pcred
->uid
);
6264 static int __sched_setscheduler(struct task_struct
*p
, int policy
,
6265 struct sched_param
*param
, bool user
)
6267 int retval
, oldprio
, oldpolicy
= -1, on_rq
, running
;
6268 unsigned long flags
;
6269 const struct sched_class
*prev_class
= p
->sched_class
;
6273 /* may grab non-irq protected spin_locks */
6274 BUG_ON(in_interrupt());
6276 /* double check policy once rq lock held */
6278 reset_on_fork
= p
->sched_reset_on_fork
;
6279 policy
= oldpolicy
= p
->policy
;
6281 reset_on_fork
= !!(policy
& SCHED_RESET_ON_FORK
);
6282 policy
&= ~SCHED_RESET_ON_FORK
;
6284 if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
6285 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
&&
6286 policy
!= SCHED_IDLE
)
6291 * Valid priorities for SCHED_FIFO and SCHED_RR are
6292 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL,
6293 * SCHED_BATCH and SCHED_IDLE is 0.
6295 if (param
->sched_priority
< 0 ||
6296 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
6297 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
6299 if (rt_policy(policy
) != (param
->sched_priority
!= 0))
6303 * Allow unprivileged RT tasks to decrease priority:
6305 if (user
&& !capable(CAP_SYS_NICE
)) {
6306 if (rt_policy(policy
)) {
6307 unsigned long rlim_rtprio
;
6309 if (!lock_task_sighand(p
, &flags
))
6311 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
6312 unlock_task_sighand(p
, &flags
);
6314 /* can't set/change the rt policy */
6315 if (policy
!= p
->policy
&& !rlim_rtprio
)
6318 /* can't increase priority */
6319 if (param
->sched_priority
> p
->rt_priority
&&
6320 param
->sched_priority
> rlim_rtprio
)
6324 * Like positive nice levels, dont allow tasks to
6325 * move out of SCHED_IDLE either:
6327 if (p
->policy
== SCHED_IDLE
&& policy
!= SCHED_IDLE
)
6330 /* can't change other user's priorities */
6331 if (!check_same_owner(p
))
6334 /* Normal users shall not reset the sched_reset_on_fork flag */
6335 if (p
->sched_reset_on_fork
&& !reset_on_fork
)
6340 #ifdef CONFIG_RT_GROUP_SCHED
6342 * Do not allow realtime tasks into groups that have no runtime
6345 if (rt_bandwidth_enabled() && rt_policy(policy
) &&
6346 task_group(p
)->rt_bandwidth
.rt_runtime
== 0)
6350 retval
= security_task_setscheduler(p
, policy
, param
);
6356 * make sure no PI-waiters arrive (or leave) while we are
6357 * changing the priority of the task:
6359 raw_spin_lock_irqsave(&p
->pi_lock
, flags
);
6361 * To be able to change p->policy safely, the apropriate
6362 * runqueue lock must be held.
6364 rq
= __task_rq_lock(p
);
6365 /* recheck policy now with rq lock held */
6366 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
6367 policy
= oldpolicy
= -1;
6368 __task_rq_unlock(rq
);
6369 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6372 update_rq_clock(rq
);
6373 on_rq
= p
->se
.on_rq
;
6374 running
= task_current(rq
, p
);
6376 deactivate_task(rq
, p
, 0);
6378 p
->sched_class
->put_prev_task(rq
, p
);
6380 p
->sched_reset_on_fork
= reset_on_fork
;
6383 __setscheduler(rq
, p
, policy
, param
->sched_priority
);
6386 p
->sched_class
->set_curr_task(rq
);
6388 activate_task(rq
, p
, 0);
6390 check_class_changed(rq
, p
, prev_class
, oldprio
, running
);
6392 __task_rq_unlock(rq
);
6393 raw_spin_unlock_irqrestore(&p
->pi_lock
, flags
);
6395 rt_mutex_adjust_pi(p
);
6401 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
6402 * @p: the task in question.
6403 * @policy: new policy.
6404 * @param: structure containing the new RT priority.
6406 * NOTE that the task may be already dead.
6408 int sched_setscheduler(struct task_struct
*p
, int policy
,
6409 struct sched_param
*param
)
6411 return __sched_setscheduler(p
, policy
, param
, true);
6413 EXPORT_SYMBOL_GPL(sched_setscheduler
);
6416 * sched_setscheduler_nocheck - change the scheduling policy and/or RT priority of a thread from kernelspace.
6417 * @p: the task in question.
6418 * @policy: new policy.
6419 * @param: structure containing the new RT priority.
6421 * Just like sched_setscheduler, only don't bother checking if the
6422 * current context has permission. For example, this is needed in
6423 * stop_machine(): we create temporary high priority worker threads,
6424 * but our caller might not have that capability.
6426 int sched_setscheduler_nocheck(struct task_struct
*p
, int policy
,
6427 struct sched_param
*param
)
6429 return __sched_setscheduler(p
, policy
, param
, false);
6433 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
6435 struct sched_param lparam
;
6436 struct task_struct
*p
;
6439 if (!param
|| pid
< 0)
6441 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
6446 p
= find_process_by_pid(pid
);
6448 retval
= sched_setscheduler(p
, policy
, &lparam
);
6455 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
6456 * @pid: the pid in question.
6457 * @policy: new policy.
6458 * @param: structure containing the new RT priority.
6460 SYSCALL_DEFINE3(sched_setscheduler
, pid_t
, pid
, int, policy
,
6461 struct sched_param __user
*, param
)
6463 /* negative values for policy are not valid */
6467 return do_sched_setscheduler(pid
, policy
, param
);
6471 * sys_sched_setparam - set/change the RT priority of a thread
6472 * @pid: the pid in question.
6473 * @param: structure containing the new RT priority.
6475 SYSCALL_DEFINE2(sched_setparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6477 return do_sched_setscheduler(pid
, -1, param
);
6481 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
6482 * @pid: the pid in question.
6484 SYSCALL_DEFINE1(sched_getscheduler
, pid_t
, pid
)
6486 struct task_struct
*p
;
6494 p
= find_process_by_pid(pid
);
6496 retval
= security_task_getscheduler(p
);
6499 | (p
->sched_reset_on_fork
? SCHED_RESET_ON_FORK
: 0);
6506 * sys_sched_getparam - get the RT priority of a thread
6507 * @pid: the pid in question.
6508 * @param: structure containing the RT priority.
6510 SYSCALL_DEFINE2(sched_getparam
, pid_t
, pid
, struct sched_param __user
*, param
)
6512 struct sched_param lp
;
6513 struct task_struct
*p
;
6516 if (!param
|| pid
< 0)
6520 p
= find_process_by_pid(pid
);
6525 retval
= security_task_getscheduler(p
);
6529 lp
.sched_priority
= p
->rt_priority
;
6533 * This one might sleep, we cannot do it with a spinlock held ...
6535 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
6544 long sched_setaffinity(pid_t pid
, const struct cpumask
*in_mask
)
6546 cpumask_var_t cpus_allowed
, new_mask
;
6547 struct task_struct
*p
;
6553 p
= find_process_by_pid(pid
);
6560 /* Prevent p going away */
6564 if (!alloc_cpumask_var(&cpus_allowed
, GFP_KERNEL
)) {
6568 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
)) {
6570 goto out_free_cpus_allowed
;
6573 if (!check_same_owner(p
) && !capable(CAP_SYS_NICE
))
6576 retval
= security_task_setscheduler(p
, 0, NULL
);
6580 cpuset_cpus_allowed(p
, cpus_allowed
);
6581 cpumask_and(new_mask
, in_mask
, cpus_allowed
);
6583 retval
= set_cpus_allowed_ptr(p
, new_mask
);
6586 cpuset_cpus_allowed(p
, cpus_allowed
);
6587 if (!cpumask_subset(new_mask
, cpus_allowed
)) {
6589 * We must have raced with a concurrent cpuset
6590 * update. Just reset the cpus_allowed to the
6591 * cpuset's cpus_allowed
6593 cpumask_copy(new_mask
, cpus_allowed
);
6598 free_cpumask_var(new_mask
);
6599 out_free_cpus_allowed
:
6600 free_cpumask_var(cpus_allowed
);
6607 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
6608 struct cpumask
*new_mask
)
6610 if (len
< cpumask_size())
6611 cpumask_clear(new_mask
);
6612 else if (len
> cpumask_size())
6613 len
= cpumask_size();
6615 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
6619 * sys_sched_setaffinity - set the cpu affinity of a process
6620 * @pid: pid of the process
6621 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6622 * @user_mask_ptr: user-space pointer to the new cpu mask
6624 SYSCALL_DEFINE3(sched_setaffinity
, pid_t
, pid
, unsigned int, len
,
6625 unsigned long __user
*, user_mask_ptr
)
6627 cpumask_var_t new_mask
;
6630 if (!alloc_cpumask_var(&new_mask
, GFP_KERNEL
))
6633 retval
= get_user_cpu_mask(user_mask_ptr
, len
, new_mask
);
6635 retval
= sched_setaffinity(pid
, new_mask
);
6636 free_cpumask_var(new_mask
);
6640 long sched_getaffinity(pid_t pid
, struct cpumask
*mask
)
6642 struct task_struct
*p
;
6643 unsigned long flags
;
6651 p
= find_process_by_pid(pid
);
6655 retval
= security_task_getscheduler(p
);
6659 rq
= task_rq_lock(p
, &flags
);
6660 cpumask_and(mask
, &p
->cpus_allowed
, cpu_online_mask
);
6661 task_rq_unlock(rq
, &flags
);
6671 * sys_sched_getaffinity - get the cpu affinity of a process
6672 * @pid: pid of the process
6673 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
6674 * @user_mask_ptr: user-space pointer to hold the current cpu mask
6676 SYSCALL_DEFINE3(sched_getaffinity
, pid_t
, pid
, unsigned int, len
,
6677 unsigned long __user
*, user_mask_ptr
)
6682 if (len
< cpumask_size())
6685 if (!alloc_cpumask_var(&mask
, GFP_KERNEL
))
6688 ret
= sched_getaffinity(pid
, mask
);
6690 if (copy_to_user(user_mask_ptr
, mask
, cpumask_size()))
6693 ret
= cpumask_size();
6695 free_cpumask_var(mask
);
6701 * sys_sched_yield - yield the current processor to other threads.
6703 * This function yields the current CPU to other tasks. If there are no
6704 * other threads running on this CPU then this function will return.
6706 SYSCALL_DEFINE0(sched_yield
)
6708 struct rq
*rq
= this_rq_lock();
6710 schedstat_inc(rq
, yld_count
);
6711 current
->sched_class
->yield_task(rq
);
6714 * Since we are going to call schedule() anyway, there's
6715 * no need to preempt or enable interrupts:
6717 __release(rq
->lock
);
6718 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
6719 do_raw_spin_unlock(&rq
->lock
);
6720 preempt_enable_no_resched();
6727 static inline int should_resched(void)
6729 return need_resched() && !(preempt_count() & PREEMPT_ACTIVE
);
6732 static void __cond_resched(void)
6734 add_preempt_count(PREEMPT_ACTIVE
);
6736 sub_preempt_count(PREEMPT_ACTIVE
);
6739 int __sched
_cond_resched(void)
6741 if (should_resched()) {
6747 EXPORT_SYMBOL(_cond_resched
);
6750 * __cond_resched_lock() - if a reschedule is pending, drop the given lock,
6751 * call schedule, and on return reacquire the lock.
6753 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
6754 * operations here to prevent schedule() from being called twice (once via
6755 * spin_unlock(), once by hand).
6757 int __cond_resched_lock(spinlock_t
*lock
)
6759 int resched
= should_resched();
6762 lockdep_assert_held(lock
);
6764 if (spin_needbreak(lock
) || resched
) {
6775 EXPORT_SYMBOL(__cond_resched_lock
);
6777 int __sched
__cond_resched_softirq(void)
6779 BUG_ON(!in_softirq());
6781 if (should_resched()) {
6789 EXPORT_SYMBOL(__cond_resched_softirq
);
6792 * yield - yield the current processor to other threads.
6794 * This is a shortcut for kernel-space yielding - it marks the
6795 * thread runnable and calls sys_sched_yield().
6797 void __sched
yield(void)
6799 set_current_state(TASK_RUNNING
);
6802 EXPORT_SYMBOL(yield
);
6805 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
6806 * that process accounting knows that this is a task in IO wait state.
6808 void __sched
io_schedule(void)
6810 struct rq
*rq
= raw_rq();
6812 delayacct_blkio_start();
6813 atomic_inc(&rq
->nr_iowait
);
6814 current
->in_iowait
= 1;
6816 current
->in_iowait
= 0;
6817 atomic_dec(&rq
->nr_iowait
);
6818 delayacct_blkio_end();
6820 EXPORT_SYMBOL(io_schedule
);
6822 long __sched
io_schedule_timeout(long timeout
)
6824 struct rq
*rq
= raw_rq();
6827 delayacct_blkio_start();
6828 atomic_inc(&rq
->nr_iowait
);
6829 current
->in_iowait
= 1;
6830 ret
= schedule_timeout(timeout
);
6831 current
->in_iowait
= 0;
6832 atomic_dec(&rq
->nr_iowait
);
6833 delayacct_blkio_end();
6838 * sys_sched_get_priority_max - return maximum RT priority.
6839 * @policy: scheduling class.
6841 * this syscall returns the maximum rt_priority that can be used
6842 * by a given scheduling class.
6844 SYSCALL_DEFINE1(sched_get_priority_max
, int, policy
)
6851 ret
= MAX_USER_RT_PRIO
-1;
6863 * sys_sched_get_priority_min - return minimum RT priority.
6864 * @policy: scheduling class.
6866 * this syscall returns the minimum rt_priority that can be used
6867 * by a given scheduling class.
6869 SYSCALL_DEFINE1(sched_get_priority_min
, int, policy
)
6887 * sys_sched_rr_get_interval - return the default timeslice of a process.
6888 * @pid: pid of the process.
6889 * @interval: userspace pointer to the timeslice value.
6891 * this syscall writes the default timeslice value of a given process
6892 * into the user-space timespec buffer. A value of '0' means infinity.
6894 SYSCALL_DEFINE2(sched_rr_get_interval
, pid_t
, pid
,
6895 struct timespec __user
*, interval
)
6897 struct task_struct
*p
;
6898 unsigned int time_slice
;
6899 unsigned long flags
;
6909 p
= find_process_by_pid(pid
);
6913 retval
= security_task_getscheduler(p
);
6917 rq
= task_rq_lock(p
, &flags
);
6918 time_slice
= p
->sched_class
->get_rr_interval(rq
, p
);
6919 task_rq_unlock(rq
, &flags
);
6922 jiffies_to_timespec(time_slice
, &t
);
6923 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
6931 static const char stat_nam
[] = TASK_STATE_TO_CHAR_STR
;
6933 void sched_show_task(struct task_struct
*p
)
6935 unsigned long free
= 0;
6938 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
6939 pr_info("%-13.13s %c", p
->comm
,
6940 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
6941 #if BITS_PER_LONG == 32
6942 if (state
== TASK_RUNNING
)
6943 pr_cont(" running ");
6945 pr_cont(" %08lx ", thread_saved_pc(p
));
6947 if (state
== TASK_RUNNING
)
6948 pr_cont(" running task ");
6950 pr_cont(" %016lx ", thread_saved_pc(p
));
6952 #ifdef CONFIG_DEBUG_STACK_USAGE
6953 free
= stack_not_used(p
);
6955 pr_cont("%5lu %5d %6d 0x%08lx\n", free
,
6956 task_pid_nr(p
), task_pid_nr(p
->real_parent
),
6957 (unsigned long)task_thread_info(p
)->flags
);
6959 show_stack(p
, NULL
);
6962 void show_state_filter(unsigned long state_filter
)
6964 struct task_struct
*g
, *p
;
6966 #if BITS_PER_LONG == 32
6967 pr_info(" task PC stack pid father\n");
6969 pr_info(" task PC stack pid father\n");
6971 read_lock(&tasklist_lock
);
6972 do_each_thread(g
, p
) {
6974 * reset the NMI-timeout, listing all files on a slow
6975 * console might take alot of time:
6977 touch_nmi_watchdog();
6978 if (!state_filter
|| (p
->state
& state_filter
))
6980 } while_each_thread(g
, p
);
6982 touch_all_softlockup_watchdogs();
6984 #ifdef CONFIG_SCHED_DEBUG
6985 sysrq_sched_debug_show();
6987 read_unlock(&tasklist_lock
);
6989 * Only show locks if all tasks are dumped:
6992 debug_show_all_locks();
6995 void __cpuinit
init_idle_bootup_task(struct task_struct
*idle
)
6997 idle
->sched_class
= &idle_sched_class
;
7001 * init_idle - set up an idle thread for a given CPU
7002 * @idle: task in question
7003 * @cpu: cpu the idle task belongs to
7005 * NOTE: this function does not set the idle thread's NEED_RESCHED
7006 * flag, to make booting more robust.
7008 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
7010 struct rq
*rq
= cpu_rq(cpu
);
7011 unsigned long flags
;
7013 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7016 idle
->state
= TASK_RUNNING
;
7017 idle
->se
.exec_start
= sched_clock();
7019 cpumask_copy(&idle
->cpus_allowed
, cpumask_of(cpu
));
7020 __set_task_cpu(idle
, cpu
);
7022 rq
->curr
= rq
->idle
= idle
;
7023 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
7026 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7028 /* Set the preempt count _outside_ the spinlocks! */
7029 #if defined(CONFIG_PREEMPT)
7030 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
7032 task_thread_info(idle
)->preempt_count
= 0;
7035 * The idle tasks have their own, simple scheduling class:
7037 idle
->sched_class
= &idle_sched_class
;
7038 ftrace_graph_init_task(idle
);
7042 * In a system that switches off the HZ timer nohz_cpu_mask
7043 * indicates which cpus entered this state. This is used
7044 * in the rcu update to wait only for active cpus. For system
7045 * which do not switch off the HZ timer nohz_cpu_mask should
7046 * always be CPU_BITS_NONE.
7048 cpumask_var_t nohz_cpu_mask
;
7051 * Increase the granularity value when there are more CPUs,
7052 * because with more CPUs the 'effective latency' as visible
7053 * to users decreases. But the relationship is not linear,
7054 * so pick a second-best guess by going with the log2 of the
7057 * This idea comes from the SD scheduler of Con Kolivas:
7059 static int get_update_sysctl_factor(void)
7061 unsigned int cpus
= min_t(int, num_online_cpus(), 8);
7062 unsigned int factor
;
7064 switch (sysctl_sched_tunable_scaling
) {
7065 case SCHED_TUNABLESCALING_NONE
:
7068 case SCHED_TUNABLESCALING_LINEAR
:
7071 case SCHED_TUNABLESCALING_LOG
:
7073 factor
= 1 + ilog2(cpus
);
7080 static void update_sysctl(void)
7082 unsigned int factor
= get_update_sysctl_factor();
7084 #define SET_SYSCTL(name) \
7085 (sysctl_##name = (factor) * normalized_sysctl_##name)
7086 SET_SYSCTL(sched_min_granularity
);
7087 SET_SYSCTL(sched_latency
);
7088 SET_SYSCTL(sched_wakeup_granularity
);
7089 SET_SYSCTL(sched_shares_ratelimit
);
7093 static inline void sched_init_granularity(void)
7100 * This is how migration works:
7102 * 1) we queue a struct migration_req structure in the source CPU's
7103 * runqueue and wake up that CPU's migration thread.
7104 * 2) we down() the locked semaphore => thread blocks.
7105 * 3) migration thread wakes up (implicitly it forces the migrated
7106 * thread off the CPU)
7107 * 4) it gets the migration request and checks whether the migrated
7108 * task is still in the wrong runqueue.
7109 * 5) if it's in the wrong runqueue then the migration thread removes
7110 * it and puts it into the right queue.
7111 * 6) migration thread up()s the semaphore.
7112 * 7) we wake up and the migration is done.
7116 * Change a given task's CPU affinity. Migrate the thread to a
7117 * proper CPU and schedule it away if the CPU it's executing on
7118 * is removed from the allowed bitmask.
7120 * NOTE: the caller must have a valid reference to the task, the
7121 * task must not exit() & deallocate itself prematurely. The
7122 * call is not atomic; no spinlocks may be held.
7124 int set_cpus_allowed_ptr(struct task_struct
*p
, const struct cpumask
*new_mask
)
7126 struct migration_req req
;
7127 unsigned long flags
;
7132 * Since we rely on wake-ups to migrate sleeping tasks, don't change
7133 * the ->cpus_allowed mask from under waking tasks, which would be
7134 * possible when we change rq->lock in ttwu(), so synchronize against
7135 * TASK_WAKING to avoid that.
7138 while (p
->state
== TASK_WAKING
)
7141 rq
= task_rq_lock(p
, &flags
);
7143 if (p
->state
== TASK_WAKING
) {
7144 task_rq_unlock(rq
, &flags
);
7148 if (!cpumask_intersects(new_mask
, cpu_active_mask
)) {
7153 if (unlikely((p
->flags
& PF_THREAD_BOUND
) && p
!= current
&&
7154 !cpumask_equal(&p
->cpus_allowed
, new_mask
))) {
7159 if (p
->sched_class
->set_cpus_allowed
)
7160 p
->sched_class
->set_cpus_allowed(p
, new_mask
);
7162 cpumask_copy(&p
->cpus_allowed
, new_mask
);
7163 p
->rt
.nr_cpus_allowed
= cpumask_weight(new_mask
);
7166 /* Can the task run on the task's current CPU? If so, we're done */
7167 if (cpumask_test_cpu(task_cpu(p
), new_mask
))
7170 if (migrate_task(p
, cpumask_any_and(cpu_active_mask
, new_mask
), &req
)) {
7171 /* Need help from migration thread: drop lock and wait. */
7172 struct task_struct
*mt
= rq
->migration_thread
;
7174 get_task_struct(mt
);
7175 task_rq_unlock(rq
, &flags
);
7176 wake_up_process(rq
->migration_thread
);
7177 put_task_struct(mt
);
7178 wait_for_completion(&req
.done
);
7179 tlb_migrate_finish(p
->mm
);
7183 task_rq_unlock(rq
, &flags
);
7187 EXPORT_SYMBOL_GPL(set_cpus_allowed_ptr
);
7190 * Move (not current) task off this cpu, onto dest cpu. We're doing
7191 * this because either it can't run here any more (set_cpus_allowed()
7192 * away from this CPU, or CPU going down), or because we're
7193 * attempting to rebalance this task on exec (sched_exec).
7195 * So we race with normal scheduler movements, but that's OK, as long
7196 * as the task is no longer on this CPU.
7198 * Returns non-zero if task was successfully migrated.
7200 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7202 struct rq
*rq_dest
, *rq_src
;
7205 if (unlikely(!cpu_active(dest_cpu
)))
7208 rq_src
= cpu_rq(src_cpu
);
7209 rq_dest
= cpu_rq(dest_cpu
);
7211 double_rq_lock(rq_src
, rq_dest
);
7212 /* Already moved. */
7213 if (task_cpu(p
) != src_cpu
)
7215 /* Affinity changed (again). */
7216 if (!cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7220 * If we're not on a rq, the next wake-up will ensure we're
7224 deactivate_task(rq_src
, p
, 0);
7225 set_task_cpu(p
, dest_cpu
);
7226 activate_task(rq_dest
, p
, 0);
7227 check_preempt_curr(rq_dest
, p
, 0);
7232 double_rq_unlock(rq_src
, rq_dest
);
7236 #define RCU_MIGRATION_IDLE 0
7237 #define RCU_MIGRATION_NEED_QS 1
7238 #define RCU_MIGRATION_GOT_QS 2
7239 #define RCU_MIGRATION_MUST_SYNC 3
7242 * migration_thread - this is a highprio system thread that performs
7243 * thread migration by bumping thread off CPU then 'pushing' onto
7246 static int migration_thread(void *data
)
7249 int cpu
= (long)data
;
7253 BUG_ON(rq
->migration_thread
!= current
);
7255 set_current_state(TASK_INTERRUPTIBLE
);
7256 while (!kthread_should_stop()) {
7257 struct migration_req
*req
;
7258 struct list_head
*head
;
7260 raw_spin_lock_irq(&rq
->lock
);
7262 if (cpu_is_offline(cpu
)) {
7263 raw_spin_unlock_irq(&rq
->lock
);
7267 if (rq
->active_balance
) {
7268 active_load_balance(rq
, cpu
);
7269 rq
->active_balance
= 0;
7272 head
= &rq
->migration_queue
;
7274 if (list_empty(head
)) {
7275 raw_spin_unlock_irq(&rq
->lock
);
7277 set_current_state(TASK_INTERRUPTIBLE
);
7280 req
= list_entry(head
->next
, struct migration_req
, list
);
7281 list_del_init(head
->next
);
7283 if (req
->task
!= NULL
) {
7284 raw_spin_unlock(&rq
->lock
);
7285 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
7286 } else if (likely(cpu
== (badcpu
= smp_processor_id()))) {
7287 req
->dest_cpu
= RCU_MIGRATION_GOT_QS
;
7288 raw_spin_unlock(&rq
->lock
);
7290 req
->dest_cpu
= RCU_MIGRATION_MUST_SYNC
;
7291 raw_spin_unlock(&rq
->lock
);
7292 WARN_ONCE(1, "migration_thread() on CPU %d, expected %d\n", badcpu
, cpu
);
7296 complete(&req
->done
);
7298 __set_current_state(TASK_RUNNING
);
7303 #ifdef CONFIG_HOTPLUG_CPU
7305 static int __migrate_task_irq(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
7309 local_irq_disable();
7310 ret
= __migrate_task(p
, src_cpu
, dest_cpu
);
7316 * Figure out where task on dead CPU should go, use force if necessary.
7318 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
7321 const struct cpumask
*nodemask
= cpumask_of_node(cpu_to_node(dead_cpu
));
7324 /* Look for allowed, online CPU in same node. */
7325 for_each_cpu_and(dest_cpu
, nodemask
, cpu_active_mask
)
7326 if (cpumask_test_cpu(dest_cpu
, &p
->cpus_allowed
))
7329 /* Any allowed, online CPU? */
7330 dest_cpu
= cpumask_any_and(&p
->cpus_allowed
, cpu_active_mask
);
7331 if (dest_cpu
< nr_cpu_ids
)
7334 /* No more Mr. Nice Guy. */
7335 if (dest_cpu
>= nr_cpu_ids
) {
7336 cpuset_cpus_allowed_locked(p
, &p
->cpus_allowed
);
7337 dest_cpu
= cpumask_any_and(cpu_active_mask
, &p
->cpus_allowed
);
7340 * Don't tell them about moving exiting tasks or
7341 * kernel threads (both mm NULL), since they never
7344 if (p
->mm
&& printk_ratelimit()) {
7345 pr_info("process %d (%s) no longer affine to cpu%d\n",
7346 task_pid_nr(p
), p
->comm
, dead_cpu
);
7351 /* It can have affinity changed while we were choosing. */
7352 if (unlikely(!__migrate_task_irq(p
, dead_cpu
, dest_cpu
)))
7357 * While a dead CPU has no uninterruptible tasks queued at this point,
7358 * it might still have a nonzero ->nr_uninterruptible counter, because
7359 * for performance reasons the counter is not stricly tracking tasks to
7360 * their home CPUs. So we just add the counter to another CPU's counter,
7361 * to keep the global sum constant after CPU-down:
7363 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
7365 struct rq
*rq_dest
= cpu_rq(cpumask_any(cpu_active_mask
));
7366 unsigned long flags
;
7368 local_irq_save(flags
);
7369 double_rq_lock(rq_src
, rq_dest
);
7370 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
7371 rq_src
->nr_uninterruptible
= 0;
7372 double_rq_unlock(rq_src
, rq_dest
);
7373 local_irq_restore(flags
);
7376 /* Run through task list and migrate tasks from the dead cpu. */
7377 static void migrate_live_tasks(int src_cpu
)
7379 struct task_struct
*p
, *t
;
7381 read_lock(&tasklist_lock
);
7383 do_each_thread(t
, p
) {
7387 if (task_cpu(p
) == src_cpu
)
7388 move_task_off_dead_cpu(src_cpu
, p
);
7389 } while_each_thread(t
, p
);
7391 read_unlock(&tasklist_lock
);
7395 * Schedules idle task to be the next runnable task on current CPU.
7396 * It does so by boosting its priority to highest possible.
7397 * Used by CPU offline code.
7399 void sched_idle_next(void)
7401 int this_cpu
= smp_processor_id();
7402 struct rq
*rq
= cpu_rq(this_cpu
);
7403 struct task_struct
*p
= rq
->idle
;
7404 unsigned long flags
;
7406 /* cpu has to be offline */
7407 BUG_ON(cpu_online(this_cpu
));
7410 * Strictly not necessary since rest of the CPUs are stopped by now
7411 * and interrupts disabled on the current cpu.
7413 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7415 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7417 update_rq_clock(rq
);
7418 activate_task(rq
, p
, 0);
7420 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7424 * Ensures that the idle task is using init_mm right before its cpu goes
7427 void idle_task_exit(void)
7429 struct mm_struct
*mm
= current
->active_mm
;
7431 BUG_ON(cpu_online(smp_processor_id()));
7434 switch_mm(mm
, &init_mm
, current
);
7438 /* called under rq->lock with disabled interrupts */
7439 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
7441 struct rq
*rq
= cpu_rq(dead_cpu
);
7443 /* Must be exiting, otherwise would be on tasklist. */
7444 BUG_ON(!p
->exit_state
);
7446 /* Cannot have done final schedule yet: would have vanished. */
7447 BUG_ON(p
->state
== TASK_DEAD
);
7452 * Drop lock around migration; if someone else moves it,
7453 * that's OK. No task can be added to this CPU, so iteration is
7456 raw_spin_unlock_irq(&rq
->lock
);
7457 move_task_off_dead_cpu(dead_cpu
, p
);
7458 raw_spin_lock_irq(&rq
->lock
);
7463 /* release_task() removes task from tasklist, so we won't find dead tasks. */
7464 static void migrate_dead_tasks(unsigned int dead_cpu
)
7466 struct rq
*rq
= cpu_rq(dead_cpu
);
7467 struct task_struct
*next
;
7470 if (!rq
->nr_running
)
7472 update_rq_clock(rq
);
7473 next
= pick_next_task(rq
);
7476 next
->sched_class
->put_prev_task(rq
, next
);
7477 migrate_dead(dead_cpu
, next
);
7483 * remove the tasks which were accounted by rq from calc_load_tasks.
7485 static void calc_global_load_remove(struct rq
*rq
)
7487 atomic_long_sub(rq
->calc_load_active
, &calc_load_tasks
);
7488 rq
->calc_load_active
= 0;
7490 #endif /* CONFIG_HOTPLUG_CPU */
7492 #if defined(CONFIG_SCHED_DEBUG) && defined(CONFIG_SYSCTL)
7494 static struct ctl_table sd_ctl_dir
[] = {
7496 .procname
= "sched_domain",
7502 static struct ctl_table sd_ctl_root
[] = {
7504 .procname
= "kernel",
7506 .child
= sd_ctl_dir
,
7511 static struct ctl_table
*sd_alloc_ctl_entry(int n
)
7513 struct ctl_table
*entry
=
7514 kcalloc(n
, sizeof(struct ctl_table
), GFP_KERNEL
);
7519 static void sd_free_ctl_entry(struct ctl_table
**tablep
)
7521 struct ctl_table
*entry
;
7524 * In the intermediate directories, both the child directory and
7525 * procname are dynamically allocated and could fail but the mode
7526 * will always be set. In the lowest directory the names are
7527 * static strings and all have proc handlers.
7529 for (entry
= *tablep
; entry
->mode
; entry
++) {
7531 sd_free_ctl_entry(&entry
->child
);
7532 if (entry
->proc_handler
== NULL
)
7533 kfree(entry
->procname
);
7541 set_table_entry(struct ctl_table
*entry
,
7542 const char *procname
, void *data
, int maxlen
,
7543 mode_t mode
, proc_handler
*proc_handler
)
7545 entry
->procname
= procname
;
7547 entry
->maxlen
= maxlen
;
7549 entry
->proc_handler
= proc_handler
;
7552 static struct ctl_table
*
7553 sd_alloc_ctl_domain_table(struct sched_domain
*sd
)
7555 struct ctl_table
*table
= sd_alloc_ctl_entry(13);
7560 set_table_entry(&table
[0], "min_interval", &sd
->min_interval
,
7561 sizeof(long), 0644, proc_doulongvec_minmax
);
7562 set_table_entry(&table
[1], "max_interval", &sd
->max_interval
,
7563 sizeof(long), 0644, proc_doulongvec_minmax
);
7564 set_table_entry(&table
[2], "busy_idx", &sd
->busy_idx
,
7565 sizeof(int), 0644, proc_dointvec_minmax
);
7566 set_table_entry(&table
[3], "idle_idx", &sd
->idle_idx
,
7567 sizeof(int), 0644, proc_dointvec_minmax
);
7568 set_table_entry(&table
[4], "newidle_idx", &sd
->newidle_idx
,
7569 sizeof(int), 0644, proc_dointvec_minmax
);
7570 set_table_entry(&table
[5], "wake_idx", &sd
->wake_idx
,
7571 sizeof(int), 0644, proc_dointvec_minmax
);
7572 set_table_entry(&table
[6], "forkexec_idx", &sd
->forkexec_idx
,
7573 sizeof(int), 0644, proc_dointvec_minmax
);
7574 set_table_entry(&table
[7], "busy_factor", &sd
->busy_factor
,
7575 sizeof(int), 0644, proc_dointvec_minmax
);
7576 set_table_entry(&table
[8], "imbalance_pct", &sd
->imbalance_pct
,
7577 sizeof(int), 0644, proc_dointvec_minmax
);
7578 set_table_entry(&table
[9], "cache_nice_tries",
7579 &sd
->cache_nice_tries
,
7580 sizeof(int), 0644, proc_dointvec_minmax
);
7581 set_table_entry(&table
[10], "flags", &sd
->flags
,
7582 sizeof(int), 0644, proc_dointvec_minmax
);
7583 set_table_entry(&table
[11], "name", sd
->name
,
7584 CORENAME_MAX_SIZE
, 0444, proc_dostring
);
7585 /* &table[12] is terminator */
7590 static ctl_table
*sd_alloc_ctl_cpu_table(int cpu
)
7592 struct ctl_table
*entry
, *table
;
7593 struct sched_domain
*sd
;
7594 int domain_num
= 0, i
;
7597 for_each_domain(cpu
, sd
)
7599 entry
= table
= sd_alloc_ctl_entry(domain_num
+ 1);
7604 for_each_domain(cpu
, sd
) {
7605 snprintf(buf
, 32, "domain%d", i
);
7606 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7608 entry
->child
= sd_alloc_ctl_domain_table(sd
);
7615 static struct ctl_table_header
*sd_sysctl_header
;
7616 static void register_sched_domain_sysctl(void)
7618 int i
, cpu_num
= num_possible_cpus();
7619 struct ctl_table
*entry
= sd_alloc_ctl_entry(cpu_num
+ 1);
7622 WARN_ON(sd_ctl_dir
[0].child
);
7623 sd_ctl_dir
[0].child
= entry
;
7628 for_each_possible_cpu(i
) {
7629 snprintf(buf
, 32, "cpu%d", i
);
7630 entry
->procname
= kstrdup(buf
, GFP_KERNEL
);
7632 entry
->child
= sd_alloc_ctl_cpu_table(i
);
7636 WARN_ON(sd_sysctl_header
);
7637 sd_sysctl_header
= register_sysctl_table(sd_ctl_root
);
7640 /* may be called multiple times per register */
7641 static void unregister_sched_domain_sysctl(void)
7643 if (sd_sysctl_header
)
7644 unregister_sysctl_table(sd_sysctl_header
);
7645 sd_sysctl_header
= NULL
;
7646 if (sd_ctl_dir
[0].child
)
7647 sd_free_ctl_entry(&sd_ctl_dir
[0].child
);
7650 static void register_sched_domain_sysctl(void)
7653 static void unregister_sched_domain_sysctl(void)
7658 static void set_rq_online(struct rq
*rq
)
7661 const struct sched_class
*class;
7663 cpumask_set_cpu(rq
->cpu
, rq
->rd
->online
);
7666 for_each_class(class) {
7667 if (class->rq_online
)
7668 class->rq_online(rq
);
7673 static void set_rq_offline(struct rq
*rq
)
7676 const struct sched_class
*class;
7678 for_each_class(class) {
7679 if (class->rq_offline
)
7680 class->rq_offline(rq
);
7683 cpumask_clear_cpu(rq
->cpu
, rq
->rd
->online
);
7689 * migration_call - callback that gets triggered when a CPU is added.
7690 * Here we can start up the necessary migration thread for the new CPU.
7692 static int __cpuinit
7693 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
7695 struct task_struct
*p
;
7696 int cpu
= (long)hcpu
;
7697 unsigned long flags
;
7702 case CPU_UP_PREPARE
:
7703 case CPU_UP_PREPARE_FROZEN
:
7704 p
= kthread_create(migration_thread
, hcpu
, "migration/%d", cpu
);
7707 kthread_bind(p
, cpu
);
7708 /* Must be high prio: stop_machine expects to yield to it. */
7709 rq
= task_rq_lock(p
, &flags
);
7710 __setscheduler(rq
, p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
7711 task_rq_unlock(rq
, &flags
);
7713 cpu_rq(cpu
)->migration_thread
= p
;
7714 rq
->calc_load_update
= calc_load_update
;
7718 case CPU_ONLINE_FROZEN
:
7719 /* Strictly unnecessary, as first user will wake it. */
7720 wake_up_process(cpu_rq(cpu
)->migration_thread
);
7722 /* Update our root-domain */
7724 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7726 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7730 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7733 #ifdef CONFIG_HOTPLUG_CPU
7734 case CPU_UP_CANCELED
:
7735 case CPU_UP_CANCELED_FROZEN
:
7736 if (!cpu_rq(cpu
)->migration_thread
)
7738 /* Unbind it from offline cpu so it can run. Fall thru. */
7739 kthread_bind(cpu_rq(cpu
)->migration_thread
,
7740 cpumask_any(cpu_online_mask
));
7741 kthread_stop(cpu_rq(cpu
)->migration_thread
);
7742 put_task_struct(cpu_rq(cpu
)->migration_thread
);
7743 cpu_rq(cpu
)->migration_thread
= NULL
;
7747 case CPU_DEAD_FROZEN
:
7748 cpuset_lock(); /* around calls to cpuset_cpus_allowed_lock() */
7749 migrate_live_tasks(cpu
);
7751 kthread_stop(rq
->migration_thread
);
7752 put_task_struct(rq
->migration_thread
);
7753 rq
->migration_thread
= NULL
;
7754 /* Idle task back to normal (off runqueue, low prio) */
7755 raw_spin_lock_irq(&rq
->lock
);
7756 update_rq_clock(rq
);
7757 deactivate_task(rq
, rq
->idle
, 0);
7758 __setscheduler(rq
, rq
->idle
, SCHED_NORMAL
, 0);
7759 rq
->idle
->sched_class
= &idle_sched_class
;
7760 migrate_dead_tasks(cpu
);
7761 raw_spin_unlock_irq(&rq
->lock
);
7763 migrate_nr_uninterruptible(rq
);
7764 BUG_ON(rq
->nr_running
!= 0);
7765 calc_global_load_remove(rq
);
7767 * No need to migrate the tasks: it was best-effort if
7768 * they didn't take sched_hotcpu_mutex. Just wake up
7771 raw_spin_lock_irq(&rq
->lock
);
7772 while (!list_empty(&rq
->migration_queue
)) {
7773 struct migration_req
*req
;
7775 req
= list_entry(rq
->migration_queue
.next
,
7776 struct migration_req
, list
);
7777 list_del_init(&req
->list
);
7778 raw_spin_unlock_irq(&rq
->lock
);
7779 complete(&req
->done
);
7780 raw_spin_lock_irq(&rq
->lock
);
7782 raw_spin_unlock_irq(&rq
->lock
);
7786 case CPU_DYING_FROZEN
:
7787 /* Update our root-domain */
7789 raw_spin_lock_irqsave(&rq
->lock
, flags
);
7791 BUG_ON(!cpumask_test_cpu(cpu
, rq
->rd
->span
));
7794 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
7802 * Register at high priority so that task migration (migrate_all_tasks)
7803 * happens before everything else. This has to be lower priority than
7804 * the notifier in the perf_event subsystem, though.
7806 static struct notifier_block __cpuinitdata migration_notifier
= {
7807 .notifier_call
= migration_call
,
7811 static int __init
migration_init(void)
7813 void *cpu
= (void *)(long)smp_processor_id();
7816 /* Start one for the boot CPU: */
7817 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
7818 BUG_ON(err
== NOTIFY_BAD
);
7819 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
7820 register_cpu_notifier(&migration_notifier
);
7824 early_initcall(migration_init
);
7829 #ifdef CONFIG_SCHED_DEBUG
7831 static __read_mostly
int sched_domain_debug_enabled
;
7833 static int __init
sched_domain_debug_setup(char *str
)
7835 sched_domain_debug_enabled
= 1;
7839 early_param("sched_debug", sched_domain_debug_setup
);
7841 static int sched_domain_debug_one(struct sched_domain
*sd
, int cpu
, int level
,
7842 struct cpumask
*groupmask
)
7844 struct sched_group
*group
= sd
->groups
;
7847 cpulist_scnprintf(str
, sizeof(str
), sched_domain_span(sd
));
7848 cpumask_clear(groupmask
);
7850 printk(KERN_DEBUG
"%*s domain %d: ", level
, "", level
);
7852 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
7853 pr_cont("does not load-balance\n");
7855 pr_err("ERROR: !SD_LOAD_BALANCE domain has parent\n");
7859 pr_cont("span %s level %s\n", str
, sd
->name
);
7861 if (!cpumask_test_cpu(cpu
, sched_domain_span(sd
))) {
7862 pr_err("ERROR: domain->span does not contain CPU%d\n", cpu
);
7864 if (!cpumask_test_cpu(cpu
, sched_group_cpus(group
))) {
7865 pr_err("ERROR: domain->groups does not contain CPU%d\n", cpu
);
7868 printk(KERN_DEBUG
"%*s groups:", level
+ 1, "");
7872 pr_err("ERROR: group is NULL\n");
7876 if (!group
->cpu_power
) {
7878 pr_err("ERROR: domain->cpu_power not set\n");
7882 if (!cpumask_weight(sched_group_cpus(group
))) {
7884 pr_err("ERROR: empty group\n");
7888 if (cpumask_intersects(groupmask
, sched_group_cpus(group
))) {
7890 pr_err("ERROR: repeated CPUs\n");
7894 cpumask_or(groupmask
, groupmask
, sched_group_cpus(group
));
7896 cpulist_scnprintf(str
, sizeof(str
), sched_group_cpus(group
));
7898 pr_cont(" %s", str
);
7899 if (group
->cpu_power
!= SCHED_LOAD_SCALE
) {
7900 pr_cont(" (cpu_power = %d)", group
->cpu_power
);
7903 group
= group
->next
;
7904 } while (group
!= sd
->groups
);
7907 if (!cpumask_equal(sched_domain_span(sd
), groupmask
))
7908 pr_err("ERROR: groups don't span domain->span\n");
7911 !cpumask_subset(groupmask
, sched_domain_span(sd
->parent
)))
7912 pr_err("ERROR: parent span is not a superset of domain->span\n");
7916 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
7918 cpumask_var_t groupmask
;
7921 if (!sched_domain_debug_enabled
)
7925 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
7929 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
7931 if (!alloc_cpumask_var(&groupmask
, GFP_KERNEL
)) {
7932 printk(KERN_DEBUG
"Cannot load-balance (out of memory)\n");
7937 if (sched_domain_debug_one(sd
, cpu
, level
, groupmask
))
7944 free_cpumask_var(groupmask
);
7946 #else /* !CONFIG_SCHED_DEBUG */
7947 # define sched_domain_debug(sd, cpu) do { } while (0)
7948 #endif /* CONFIG_SCHED_DEBUG */
7950 static int sd_degenerate(struct sched_domain
*sd
)
7952 if (cpumask_weight(sched_domain_span(sd
)) == 1)
7955 /* Following flags need at least 2 groups */
7956 if (sd
->flags
& (SD_LOAD_BALANCE
|
7957 SD_BALANCE_NEWIDLE
|
7961 SD_SHARE_PKG_RESOURCES
)) {
7962 if (sd
->groups
!= sd
->groups
->next
)
7966 /* Following flags don't use groups */
7967 if (sd
->flags
& (SD_WAKE_AFFINE
))
7974 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
7976 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
7978 if (sd_degenerate(parent
))
7981 if (!cpumask_equal(sched_domain_span(sd
), sched_domain_span(parent
)))
7984 /* Flags needing groups don't count if only 1 group in parent */
7985 if (parent
->groups
== parent
->groups
->next
) {
7986 pflags
&= ~(SD_LOAD_BALANCE
|
7987 SD_BALANCE_NEWIDLE
|
7991 SD_SHARE_PKG_RESOURCES
);
7992 if (nr_node_ids
== 1)
7993 pflags
&= ~SD_SERIALIZE
;
7995 if (~cflags
& pflags
)
8001 static void free_rootdomain(struct root_domain
*rd
)
8003 synchronize_sched();
8005 cpupri_cleanup(&rd
->cpupri
);
8007 free_cpumask_var(rd
->rto_mask
);
8008 free_cpumask_var(rd
->online
);
8009 free_cpumask_var(rd
->span
);
8013 static void rq_attach_root(struct rq
*rq
, struct root_domain
*rd
)
8015 struct root_domain
*old_rd
= NULL
;
8016 unsigned long flags
;
8018 raw_spin_lock_irqsave(&rq
->lock
, flags
);
8023 if (cpumask_test_cpu(rq
->cpu
, old_rd
->online
))
8026 cpumask_clear_cpu(rq
->cpu
, old_rd
->span
);
8029 * If we dont want to free the old_rt yet then
8030 * set old_rd to NULL to skip the freeing later
8033 if (!atomic_dec_and_test(&old_rd
->refcount
))
8037 atomic_inc(&rd
->refcount
);
8040 cpumask_set_cpu(rq
->cpu
, rd
->span
);
8041 if (cpumask_test_cpu(rq
->cpu
, cpu_active_mask
))
8044 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
8047 free_rootdomain(old_rd
);
8050 static int init_rootdomain(struct root_domain
*rd
, bool bootmem
)
8052 gfp_t gfp
= GFP_KERNEL
;
8054 memset(rd
, 0, sizeof(*rd
));
8059 if (!alloc_cpumask_var(&rd
->span
, gfp
))
8061 if (!alloc_cpumask_var(&rd
->online
, gfp
))
8063 if (!alloc_cpumask_var(&rd
->rto_mask
, gfp
))
8066 if (cpupri_init(&rd
->cpupri
, bootmem
) != 0)
8071 free_cpumask_var(rd
->rto_mask
);
8073 free_cpumask_var(rd
->online
);
8075 free_cpumask_var(rd
->span
);
8080 static void init_defrootdomain(void)
8082 init_rootdomain(&def_root_domain
, true);
8084 atomic_set(&def_root_domain
.refcount
, 1);
8087 static struct root_domain
*alloc_rootdomain(void)
8089 struct root_domain
*rd
;
8091 rd
= kmalloc(sizeof(*rd
), GFP_KERNEL
);
8095 if (init_rootdomain(rd
, false) != 0) {
8104 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
8105 * hold the hotplug lock.
8108 cpu_attach_domain(struct sched_domain
*sd
, struct root_domain
*rd
, int cpu
)
8110 struct rq
*rq
= cpu_rq(cpu
);
8111 struct sched_domain
*tmp
;
8113 /* Remove the sched domains which do not contribute to scheduling. */
8114 for (tmp
= sd
; tmp
; ) {
8115 struct sched_domain
*parent
= tmp
->parent
;
8119 if (sd_parent_degenerate(tmp
, parent
)) {
8120 tmp
->parent
= parent
->parent
;
8122 parent
->parent
->child
= tmp
;
8127 if (sd
&& sd_degenerate(sd
)) {
8133 sched_domain_debug(sd
, cpu
);
8135 rq_attach_root(rq
, rd
);
8136 rcu_assign_pointer(rq
->sd
, sd
);
8139 /* cpus with isolated domains */
8140 static cpumask_var_t cpu_isolated_map
;
8142 /* Setup the mask of cpus configured for isolated domains */
8143 static int __init
isolated_cpu_setup(char *str
)
8145 alloc_bootmem_cpumask_var(&cpu_isolated_map
);
8146 cpulist_parse(str
, cpu_isolated_map
);
8150 __setup("isolcpus=", isolated_cpu_setup
);
8153 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
8154 * to a function which identifies what group(along with sched group) a CPU
8155 * belongs to. The return value of group_fn must be a >= 0 and < nr_cpu_ids
8156 * (due to the fact that we keep track of groups covered with a struct cpumask).
8158 * init_sched_build_groups will build a circular linked list of the groups
8159 * covered by the given span, and will set each group's ->cpumask correctly,
8160 * and ->cpu_power to 0.
8163 init_sched_build_groups(const struct cpumask
*span
,
8164 const struct cpumask
*cpu_map
,
8165 int (*group_fn
)(int cpu
, const struct cpumask
*cpu_map
,
8166 struct sched_group
**sg
,
8167 struct cpumask
*tmpmask
),
8168 struct cpumask
*covered
, struct cpumask
*tmpmask
)
8170 struct sched_group
*first
= NULL
, *last
= NULL
;
8173 cpumask_clear(covered
);
8175 for_each_cpu(i
, span
) {
8176 struct sched_group
*sg
;
8177 int group
= group_fn(i
, cpu_map
, &sg
, tmpmask
);
8180 if (cpumask_test_cpu(i
, covered
))
8183 cpumask_clear(sched_group_cpus(sg
));
8186 for_each_cpu(j
, span
) {
8187 if (group_fn(j
, cpu_map
, NULL
, tmpmask
) != group
)
8190 cpumask_set_cpu(j
, covered
);
8191 cpumask_set_cpu(j
, sched_group_cpus(sg
));
8202 #define SD_NODES_PER_DOMAIN 16
8207 * find_next_best_node - find the next node to include in a sched_domain
8208 * @node: node whose sched_domain we're building
8209 * @used_nodes: nodes already in the sched_domain
8211 * Find the next node to include in a given scheduling domain. Simply
8212 * finds the closest node not already in the @used_nodes map.
8214 * Should use nodemask_t.
8216 static int find_next_best_node(int node
, nodemask_t
*used_nodes
)
8218 int i
, n
, val
, min_val
, best_node
= 0;
8222 for (i
= 0; i
< nr_node_ids
; i
++) {
8223 /* Start at @node */
8224 n
= (node
+ i
) % nr_node_ids
;
8226 if (!nr_cpus_node(n
))
8229 /* Skip already used nodes */
8230 if (node_isset(n
, *used_nodes
))
8233 /* Simple min distance search */
8234 val
= node_distance(node
, n
);
8236 if (val
< min_val
) {
8242 node_set(best_node
, *used_nodes
);
8247 * sched_domain_node_span - get a cpumask for a node's sched_domain
8248 * @node: node whose cpumask we're constructing
8249 * @span: resulting cpumask
8251 * Given a node, construct a good cpumask for its sched_domain to span. It
8252 * should be one that prevents unnecessary balancing, but also spreads tasks
8255 static void sched_domain_node_span(int node
, struct cpumask
*span
)
8257 nodemask_t used_nodes
;
8260 cpumask_clear(span
);
8261 nodes_clear(used_nodes
);
8263 cpumask_or(span
, span
, cpumask_of_node(node
));
8264 node_set(node
, used_nodes
);
8266 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
8267 int next_node
= find_next_best_node(node
, &used_nodes
);
8269 cpumask_or(span
, span
, cpumask_of_node(next_node
));
8272 #endif /* CONFIG_NUMA */
8274 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
8277 * The cpus mask in sched_group and sched_domain hangs off the end.
8279 * ( See the the comments in include/linux/sched.h:struct sched_group
8280 * and struct sched_domain. )
8282 struct static_sched_group
{
8283 struct sched_group sg
;
8284 DECLARE_BITMAP(cpus
, CONFIG_NR_CPUS
);
8287 struct static_sched_domain
{
8288 struct sched_domain sd
;
8289 DECLARE_BITMAP(span
, CONFIG_NR_CPUS
);
8295 cpumask_var_t domainspan
;
8296 cpumask_var_t covered
;
8297 cpumask_var_t notcovered
;
8299 cpumask_var_t nodemask
;
8300 cpumask_var_t this_sibling_map
;
8301 cpumask_var_t this_core_map
;
8302 cpumask_var_t send_covered
;
8303 cpumask_var_t tmpmask
;
8304 struct sched_group
**sched_group_nodes
;
8305 struct root_domain
*rd
;
8309 sa_sched_groups
= 0,
8314 sa_this_sibling_map
,
8316 sa_sched_group_nodes
,
8326 * SMT sched-domains:
8328 #ifdef CONFIG_SCHED_SMT
8329 static DEFINE_PER_CPU(struct static_sched_domain
, cpu_domains
);
8330 static DEFINE_PER_CPU(struct static_sched_group
, sched_groups
);
8333 cpu_to_cpu_group(int cpu
, const struct cpumask
*cpu_map
,
8334 struct sched_group
**sg
, struct cpumask
*unused
)
8337 *sg
= &per_cpu(sched_groups
, cpu
).sg
;
8340 #endif /* CONFIG_SCHED_SMT */
8343 * multi-core sched-domains:
8345 #ifdef CONFIG_SCHED_MC
8346 static DEFINE_PER_CPU(struct static_sched_domain
, core_domains
);
8347 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_core
);
8348 #endif /* CONFIG_SCHED_MC */
8350 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
8352 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8353 struct sched_group
**sg
, struct cpumask
*mask
)
8357 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8358 group
= cpumask_first(mask
);
8360 *sg
= &per_cpu(sched_group_core
, group
).sg
;
8363 #elif defined(CONFIG_SCHED_MC)
8365 cpu_to_core_group(int cpu
, const struct cpumask
*cpu_map
,
8366 struct sched_group
**sg
, struct cpumask
*unused
)
8369 *sg
= &per_cpu(sched_group_core
, cpu
).sg
;
8374 static DEFINE_PER_CPU(struct static_sched_domain
, phys_domains
);
8375 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_phys
);
8378 cpu_to_phys_group(int cpu
, const struct cpumask
*cpu_map
,
8379 struct sched_group
**sg
, struct cpumask
*mask
)
8382 #ifdef CONFIG_SCHED_MC
8383 cpumask_and(mask
, cpu_coregroup_mask(cpu
), cpu_map
);
8384 group
= cpumask_first(mask
);
8385 #elif defined(CONFIG_SCHED_SMT)
8386 cpumask_and(mask
, topology_thread_cpumask(cpu
), cpu_map
);
8387 group
= cpumask_first(mask
);
8392 *sg
= &per_cpu(sched_group_phys
, group
).sg
;
8398 * The init_sched_build_groups can't handle what we want to do with node
8399 * groups, so roll our own. Now each node has its own list of groups which
8400 * gets dynamically allocated.
8402 static DEFINE_PER_CPU(struct static_sched_domain
, node_domains
);
8403 static struct sched_group
***sched_group_nodes_bycpu
;
8405 static DEFINE_PER_CPU(struct static_sched_domain
, allnodes_domains
);
8406 static DEFINE_PER_CPU(struct static_sched_group
, sched_group_allnodes
);
8408 static int cpu_to_allnodes_group(int cpu
, const struct cpumask
*cpu_map
,
8409 struct sched_group
**sg
,
8410 struct cpumask
*nodemask
)
8414 cpumask_and(nodemask
, cpumask_of_node(cpu_to_node(cpu
)), cpu_map
);
8415 group
= cpumask_first(nodemask
);
8418 *sg
= &per_cpu(sched_group_allnodes
, group
).sg
;
8422 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
8424 struct sched_group
*sg
= group_head
;
8430 for_each_cpu(j
, sched_group_cpus(sg
)) {
8431 struct sched_domain
*sd
;
8433 sd
= &per_cpu(phys_domains
, j
).sd
;
8434 if (j
!= group_first_cpu(sd
->groups
)) {
8436 * Only add "power" once for each
8442 sg
->cpu_power
+= sd
->groups
->cpu_power
;
8445 } while (sg
!= group_head
);
8448 static int build_numa_sched_groups(struct s_data
*d
,
8449 const struct cpumask
*cpu_map
, int num
)
8451 struct sched_domain
*sd
;
8452 struct sched_group
*sg
, *prev
;
8455 cpumask_clear(d
->covered
);
8456 cpumask_and(d
->nodemask
, cpumask_of_node(num
), cpu_map
);
8457 if (cpumask_empty(d
->nodemask
)) {
8458 d
->sched_group_nodes
[num
] = NULL
;
8462 sched_domain_node_span(num
, d
->domainspan
);
8463 cpumask_and(d
->domainspan
, d
->domainspan
, cpu_map
);
8465 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8468 pr_warning("Can not alloc domain group for node %d\n", num
);
8471 d
->sched_group_nodes
[num
] = sg
;
8473 for_each_cpu(j
, d
->nodemask
) {
8474 sd
= &per_cpu(node_domains
, j
).sd
;
8479 cpumask_copy(sched_group_cpus(sg
), d
->nodemask
);
8481 cpumask_or(d
->covered
, d
->covered
, d
->nodemask
);
8484 for (j
= 0; j
< nr_node_ids
; j
++) {
8485 n
= (num
+ j
) % nr_node_ids
;
8486 cpumask_complement(d
->notcovered
, d
->covered
);
8487 cpumask_and(d
->tmpmask
, d
->notcovered
, cpu_map
);
8488 cpumask_and(d
->tmpmask
, d
->tmpmask
, d
->domainspan
);
8489 if (cpumask_empty(d
->tmpmask
))
8491 cpumask_and(d
->tmpmask
, d
->tmpmask
, cpumask_of_node(n
));
8492 if (cpumask_empty(d
->tmpmask
))
8494 sg
= kmalloc_node(sizeof(struct sched_group
) + cpumask_size(),
8497 pr_warning("Can not alloc domain group for node %d\n",
8502 cpumask_copy(sched_group_cpus(sg
), d
->tmpmask
);
8503 sg
->next
= prev
->next
;
8504 cpumask_or(d
->covered
, d
->covered
, d
->tmpmask
);
8511 #endif /* CONFIG_NUMA */
8514 /* Free memory allocated for various sched_group structures */
8515 static void free_sched_groups(const struct cpumask
*cpu_map
,
8516 struct cpumask
*nodemask
)
8520 for_each_cpu(cpu
, cpu_map
) {
8521 struct sched_group
**sched_group_nodes
8522 = sched_group_nodes_bycpu
[cpu
];
8524 if (!sched_group_nodes
)
8527 for (i
= 0; i
< nr_node_ids
; i
++) {
8528 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
8530 cpumask_and(nodemask
, cpumask_of_node(i
), cpu_map
);
8531 if (cpumask_empty(nodemask
))
8541 if (oldsg
!= sched_group_nodes
[i
])
8544 kfree(sched_group_nodes
);
8545 sched_group_nodes_bycpu
[cpu
] = NULL
;
8548 #else /* !CONFIG_NUMA */
8549 static void free_sched_groups(const struct cpumask
*cpu_map
,
8550 struct cpumask
*nodemask
)
8553 #endif /* CONFIG_NUMA */
8556 * Initialize sched groups cpu_power.
8558 * cpu_power indicates the capacity of sched group, which is used while
8559 * distributing the load between different sched groups in a sched domain.
8560 * Typically cpu_power for all the groups in a sched domain will be same unless
8561 * there are asymmetries in the topology. If there are asymmetries, group
8562 * having more cpu_power will pickup more load compared to the group having
8565 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
8567 struct sched_domain
*child
;
8568 struct sched_group
*group
;
8572 WARN_ON(!sd
|| !sd
->groups
);
8574 if (cpu
!= group_first_cpu(sd
->groups
))
8579 sd
->groups
->cpu_power
= 0;
8582 power
= SCHED_LOAD_SCALE
;
8583 weight
= cpumask_weight(sched_domain_span(sd
));
8585 * SMT siblings share the power of a single core.
8586 * Usually multiple threads get a better yield out of
8587 * that one core than a single thread would have,
8588 * reflect that in sd->smt_gain.
8590 if ((sd
->flags
& SD_SHARE_CPUPOWER
) && weight
> 1) {
8591 power
*= sd
->smt_gain
;
8593 power
>>= SCHED_LOAD_SHIFT
;
8595 sd
->groups
->cpu_power
+= power
;
8600 * Add cpu_power of each child group to this groups cpu_power.
8602 group
= child
->groups
;
8604 sd
->groups
->cpu_power
+= group
->cpu_power
;
8605 group
= group
->next
;
8606 } while (group
!= child
->groups
);
8610 * Initializers for schedule domains
8611 * Non-inlined to reduce accumulated stack pressure in build_sched_domains()
8614 #ifdef CONFIG_SCHED_DEBUG
8615 # define SD_INIT_NAME(sd, type) sd->name = #type
8617 # define SD_INIT_NAME(sd, type) do { } while (0)
8620 #define SD_INIT(sd, type) sd_init_##type(sd)
8622 #define SD_INIT_FUNC(type) \
8623 static noinline void sd_init_##type(struct sched_domain *sd) \
8625 memset(sd, 0, sizeof(*sd)); \
8626 *sd = SD_##type##_INIT; \
8627 sd->level = SD_LV_##type; \
8628 SD_INIT_NAME(sd, type); \
8633 SD_INIT_FUNC(ALLNODES
)
8636 #ifdef CONFIG_SCHED_SMT
8637 SD_INIT_FUNC(SIBLING
)
8639 #ifdef CONFIG_SCHED_MC
8643 static int default_relax_domain_level
= -1;
8645 static int __init
setup_relax_domain_level(char *str
)
8649 val
= simple_strtoul(str
, NULL
, 0);
8650 if (val
< SD_LV_MAX
)
8651 default_relax_domain_level
= val
;
8655 __setup("relax_domain_level=", setup_relax_domain_level
);
8657 static void set_domain_attribute(struct sched_domain
*sd
,
8658 struct sched_domain_attr
*attr
)
8662 if (!attr
|| attr
->relax_domain_level
< 0) {
8663 if (default_relax_domain_level
< 0)
8666 request
= default_relax_domain_level
;
8668 request
= attr
->relax_domain_level
;
8669 if (request
< sd
->level
) {
8670 /* turn off idle balance on this domain */
8671 sd
->flags
&= ~(SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8673 /* turn on idle balance on this domain */
8674 sd
->flags
|= (SD_BALANCE_WAKE
|SD_BALANCE_NEWIDLE
);
8678 static void __free_domain_allocs(struct s_data
*d
, enum s_alloc what
,
8679 const struct cpumask
*cpu_map
)
8682 case sa_sched_groups
:
8683 free_sched_groups(cpu_map
, d
->tmpmask
); /* fall through */
8684 d
->sched_group_nodes
= NULL
;
8686 free_rootdomain(d
->rd
); /* fall through */
8688 free_cpumask_var(d
->tmpmask
); /* fall through */
8689 case sa_send_covered
:
8690 free_cpumask_var(d
->send_covered
); /* fall through */
8691 case sa_this_core_map
:
8692 free_cpumask_var(d
->this_core_map
); /* fall through */
8693 case sa_this_sibling_map
:
8694 free_cpumask_var(d
->this_sibling_map
); /* fall through */
8696 free_cpumask_var(d
->nodemask
); /* fall through */
8697 case sa_sched_group_nodes
:
8699 kfree(d
->sched_group_nodes
); /* fall through */
8701 free_cpumask_var(d
->notcovered
); /* fall through */
8703 free_cpumask_var(d
->covered
); /* fall through */
8705 free_cpumask_var(d
->domainspan
); /* fall through */
8712 static enum s_alloc
__visit_domain_allocation_hell(struct s_data
*d
,
8713 const struct cpumask
*cpu_map
)
8716 if (!alloc_cpumask_var(&d
->domainspan
, GFP_KERNEL
))
8718 if (!alloc_cpumask_var(&d
->covered
, GFP_KERNEL
))
8719 return sa_domainspan
;
8720 if (!alloc_cpumask_var(&d
->notcovered
, GFP_KERNEL
))
8722 /* Allocate the per-node list of sched groups */
8723 d
->sched_group_nodes
= kcalloc(nr_node_ids
,
8724 sizeof(struct sched_group
*), GFP_KERNEL
);
8725 if (!d
->sched_group_nodes
) {
8726 pr_warning("Can not alloc sched group node list\n");
8727 return sa_notcovered
;
8729 sched_group_nodes_bycpu
[cpumask_first(cpu_map
)] = d
->sched_group_nodes
;
8731 if (!alloc_cpumask_var(&d
->nodemask
, GFP_KERNEL
))
8732 return sa_sched_group_nodes
;
8733 if (!alloc_cpumask_var(&d
->this_sibling_map
, GFP_KERNEL
))
8735 if (!alloc_cpumask_var(&d
->this_core_map
, GFP_KERNEL
))
8736 return sa_this_sibling_map
;
8737 if (!alloc_cpumask_var(&d
->send_covered
, GFP_KERNEL
))
8738 return sa_this_core_map
;
8739 if (!alloc_cpumask_var(&d
->tmpmask
, GFP_KERNEL
))
8740 return sa_send_covered
;
8741 d
->rd
= alloc_rootdomain();
8743 pr_warning("Cannot alloc root domain\n");
8746 return sa_rootdomain
;
8749 static struct sched_domain
*__build_numa_sched_domains(struct s_data
*d
,
8750 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
, int i
)
8752 struct sched_domain
*sd
= NULL
;
8754 struct sched_domain
*parent
;
8757 if (cpumask_weight(cpu_map
) >
8758 SD_NODES_PER_DOMAIN
* cpumask_weight(d
->nodemask
)) {
8759 sd
= &per_cpu(allnodes_domains
, i
).sd
;
8760 SD_INIT(sd
, ALLNODES
);
8761 set_domain_attribute(sd
, attr
);
8762 cpumask_copy(sched_domain_span(sd
), cpu_map
);
8763 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8768 sd
= &per_cpu(node_domains
, i
).sd
;
8770 set_domain_attribute(sd
, attr
);
8771 sched_domain_node_span(cpu_to_node(i
), sched_domain_span(sd
));
8772 sd
->parent
= parent
;
8775 cpumask_and(sched_domain_span(sd
), sched_domain_span(sd
), cpu_map
);
8780 static struct sched_domain
*__build_cpu_sched_domain(struct s_data
*d
,
8781 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8782 struct sched_domain
*parent
, int i
)
8784 struct sched_domain
*sd
;
8785 sd
= &per_cpu(phys_domains
, i
).sd
;
8787 set_domain_attribute(sd
, attr
);
8788 cpumask_copy(sched_domain_span(sd
), d
->nodemask
);
8789 sd
->parent
= parent
;
8792 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8796 static struct sched_domain
*__build_mc_sched_domain(struct s_data
*d
,
8797 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8798 struct sched_domain
*parent
, int i
)
8800 struct sched_domain
*sd
= parent
;
8801 #ifdef CONFIG_SCHED_MC
8802 sd
= &per_cpu(core_domains
, i
).sd
;
8804 set_domain_attribute(sd
, attr
);
8805 cpumask_and(sched_domain_span(sd
), cpu_map
, cpu_coregroup_mask(i
));
8806 sd
->parent
= parent
;
8808 cpu_to_core_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8813 static struct sched_domain
*__build_smt_sched_domain(struct s_data
*d
,
8814 const struct cpumask
*cpu_map
, struct sched_domain_attr
*attr
,
8815 struct sched_domain
*parent
, int i
)
8817 struct sched_domain
*sd
= parent
;
8818 #ifdef CONFIG_SCHED_SMT
8819 sd
= &per_cpu(cpu_domains
, i
).sd
;
8820 SD_INIT(sd
, SIBLING
);
8821 set_domain_attribute(sd
, attr
);
8822 cpumask_and(sched_domain_span(sd
), cpu_map
, topology_thread_cpumask(i
));
8823 sd
->parent
= parent
;
8825 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
, d
->tmpmask
);
8830 static void build_sched_groups(struct s_data
*d
, enum sched_domain_level l
,
8831 const struct cpumask
*cpu_map
, int cpu
)
8834 #ifdef CONFIG_SCHED_SMT
8835 case SD_LV_SIBLING
: /* set up CPU (sibling) groups */
8836 cpumask_and(d
->this_sibling_map
, cpu_map
,
8837 topology_thread_cpumask(cpu
));
8838 if (cpu
== cpumask_first(d
->this_sibling_map
))
8839 init_sched_build_groups(d
->this_sibling_map
, cpu_map
,
8841 d
->send_covered
, d
->tmpmask
);
8844 #ifdef CONFIG_SCHED_MC
8845 case SD_LV_MC
: /* set up multi-core groups */
8846 cpumask_and(d
->this_core_map
, cpu_map
, cpu_coregroup_mask(cpu
));
8847 if (cpu
== cpumask_first(d
->this_core_map
))
8848 init_sched_build_groups(d
->this_core_map
, cpu_map
,
8850 d
->send_covered
, d
->tmpmask
);
8853 case SD_LV_CPU
: /* set up physical groups */
8854 cpumask_and(d
->nodemask
, cpumask_of_node(cpu
), cpu_map
);
8855 if (!cpumask_empty(d
->nodemask
))
8856 init_sched_build_groups(d
->nodemask
, cpu_map
,
8858 d
->send_covered
, d
->tmpmask
);
8861 case SD_LV_ALLNODES
:
8862 init_sched_build_groups(cpu_map
, cpu_map
, &cpu_to_allnodes_group
,
8863 d
->send_covered
, d
->tmpmask
);
8872 * Build sched domains for a given set of cpus and attach the sched domains
8873 * to the individual cpus
8875 static int __build_sched_domains(const struct cpumask
*cpu_map
,
8876 struct sched_domain_attr
*attr
)
8878 enum s_alloc alloc_state
= sa_none
;
8880 struct sched_domain
*sd
;
8886 alloc_state
= __visit_domain_allocation_hell(&d
, cpu_map
);
8887 if (alloc_state
!= sa_rootdomain
)
8889 alloc_state
= sa_sched_groups
;
8892 * Set up domains for cpus specified by the cpu_map.
8894 for_each_cpu(i
, cpu_map
) {
8895 cpumask_and(d
.nodemask
, cpumask_of_node(cpu_to_node(i
)),
8898 sd
= __build_numa_sched_domains(&d
, cpu_map
, attr
, i
);
8899 sd
= __build_cpu_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8900 sd
= __build_mc_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8901 sd
= __build_smt_sched_domain(&d
, cpu_map
, attr
, sd
, i
);
8904 for_each_cpu(i
, cpu_map
) {
8905 build_sched_groups(&d
, SD_LV_SIBLING
, cpu_map
, i
);
8906 build_sched_groups(&d
, SD_LV_MC
, cpu_map
, i
);
8909 /* Set up physical groups */
8910 for (i
= 0; i
< nr_node_ids
; i
++)
8911 build_sched_groups(&d
, SD_LV_CPU
, cpu_map
, i
);
8914 /* Set up node groups */
8916 build_sched_groups(&d
, SD_LV_ALLNODES
, cpu_map
, 0);
8918 for (i
= 0; i
< nr_node_ids
; i
++)
8919 if (build_numa_sched_groups(&d
, cpu_map
, i
))
8923 /* Calculate CPU power for physical packages and nodes */
8924 #ifdef CONFIG_SCHED_SMT
8925 for_each_cpu(i
, cpu_map
) {
8926 sd
= &per_cpu(cpu_domains
, i
).sd
;
8927 init_sched_groups_power(i
, sd
);
8930 #ifdef CONFIG_SCHED_MC
8931 for_each_cpu(i
, cpu_map
) {
8932 sd
= &per_cpu(core_domains
, i
).sd
;
8933 init_sched_groups_power(i
, sd
);
8937 for_each_cpu(i
, cpu_map
) {
8938 sd
= &per_cpu(phys_domains
, i
).sd
;
8939 init_sched_groups_power(i
, sd
);
8943 for (i
= 0; i
< nr_node_ids
; i
++)
8944 init_numa_sched_groups_power(d
.sched_group_nodes
[i
]);
8946 if (d
.sd_allnodes
) {
8947 struct sched_group
*sg
;
8949 cpu_to_allnodes_group(cpumask_first(cpu_map
), cpu_map
, &sg
,
8951 init_numa_sched_groups_power(sg
);
8955 /* Attach the domains */
8956 for_each_cpu(i
, cpu_map
) {
8957 #ifdef CONFIG_SCHED_SMT
8958 sd
= &per_cpu(cpu_domains
, i
).sd
;
8959 #elif defined(CONFIG_SCHED_MC)
8960 sd
= &per_cpu(core_domains
, i
).sd
;
8962 sd
= &per_cpu(phys_domains
, i
).sd
;
8964 cpu_attach_domain(sd
, d
.rd
, i
);
8967 d
.sched_group_nodes
= NULL
; /* don't free this we still need it */
8968 __free_domain_allocs(&d
, sa_tmpmask
, cpu_map
);
8972 __free_domain_allocs(&d
, alloc_state
, cpu_map
);
8976 static int build_sched_domains(const struct cpumask
*cpu_map
)
8978 return __build_sched_domains(cpu_map
, NULL
);
8981 static cpumask_var_t
*doms_cur
; /* current sched domains */
8982 static int ndoms_cur
; /* number of sched domains in 'doms_cur' */
8983 static struct sched_domain_attr
*dattr_cur
;
8984 /* attribues of custom domains in 'doms_cur' */
8987 * Special case: If a kmalloc of a doms_cur partition (array of
8988 * cpumask) fails, then fallback to a single sched domain,
8989 * as determined by the single cpumask fallback_doms.
8991 static cpumask_var_t fallback_doms
;
8994 * arch_update_cpu_topology lets virtualized architectures update the
8995 * cpu core maps. It is supposed to return 1 if the topology changed
8996 * or 0 if it stayed the same.
8998 int __attribute__((weak
)) arch_update_cpu_topology(void)
9003 cpumask_var_t
*alloc_sched_domains(unsigned int ndoms
)
9006 cpumask_var_t
*doms
;
9008 doms
= kmalloc(sizeof(*doms
) * ndoms
, GFP_KERNEL
);
9011 for (i
= 0; i
< ndoms
; i
++) {
9012 if (!alloc_cpumask_var(&doms
[i
], GFP_KERNEL
)) {
9013 free_sched_domains(doms
, i
);
9020 void free_sched_domains(cpumask_var_t doms
[], unsigned int ndoms
)
9023 for (i
= 0; i
< ndoms
; i
++)
9024 free_cpumask_var(doms
[i
]);
9029 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
9030 * For now this just excludes isolated cpus, but could be used to
9031 * exclude other special cases in the future.
9033 static int arch_init_sched_domains(const struct cpumask
*cpu_map
)
9037 arch_update_cpu_topology();
9039 doms_cur
= alloc_sched_domains(ndoms_cur
);
9041 doms_cur
= &fallback_doms
;
9042 cpumask_andnot(doms_cur
[0], cpu_map
, cpu_isolated_map
);
9044 err
= build_sched_domains(doms_cur
[0]);
9045 register_sched_domain_sysctl();
9050 static void arch_destroy_sched_domains(const struct cpumask
*cpu_map
,
9051 struct cpumask
*tmpmask
)
9053 free_sched_groups(cpu_map
, tmpmask
);
9057 * Detach sched domains from a group of cpus specified in cpu_map
9058 * These cpus will now be attached to the NULL domain
9060 static void detach_destroy_domains(const struct cpumask
*cpu_map
)
9062 /* Save because hotplug lock held. */
9063 static DECLARE_BITMAP(tmpmask
, CONFIG_NR_CPUS
);
9066 for_each_cpu(i
, cpu_map
)
9067 cpu_attach_domain(NULL
, &def_root_domain
, i
);
9068 synchronize_sched();
9069 arch_destroy_sched_domains(cpu_map
, to_cpumask(tmpmask
));
9072 /* handle null as "default" */
9073 static int dattrs_equal(struct sched_domain_attr
*cur
, int idx_cur
,
9074 struct sched_domain_attr
*new, int idx_new
)
9076 struct sched_domain_attr tmp
;
9083 return !memcmp(cur
? (cur
+ idx_cur
) : &tmp
,
9084 new ? (new + idx_new
) : &tmp
,
9085 sizeof(struct sched_domain_attr
));
9089 * Partition sched domains as specified by the 'ndoms_new'
9090 * cpumasks in the array doms_new[] of cpumasks. This compares
9091 * doms_new[] to the current sched domain partitioning, doms_cur[].
9092 * It destroys each deleted domain and builds each new domain.
9094 * 'doms_new' is an array of cpumask_var_t's of length 'ndoms_new'.
9095 * The masks don't intersect (don't overlap.) We should setup one
9096 * sched domain for each mask. CPUs not in any of the cpumasks will
9097 * not be load balanced. If the same cpumask appears both in the
9098 * current 'doms_cur' domains and in the new 'doms_new', we can leave
9101 * The passed in 'doms_new' should be allocated using
9102 * alloc_sched_domains. This routine takes ownership of it and will
9103 * free_sched_domains it when done with it. If the caller failed the
9104 * alloc call, then it can pass in doms_new == NULL && ndoms_new == 1,
9105 * and partition_sched_domains() will fallback to the single partition
9106 * 'fallback_doms', it also forces the domains to be rebuilt.
9108 * If doms_new == NULL it will be replaced with cpu_online_mask.
9109 * ndoms_new == 0 is a special case for destroying existing domains,
9110 * and it will not create the default domain.
9112 * Call with hotplug lock held
9114 void partition_sched_domains(int ndoms_new
, cpumask_var_t doms_new
[],
9115 struct sched_domain_attr
*dattr_new
)
9120 mutex_lock(&sched_domains_mutex
);
9122 /* always unregister in case we don't destroy any domains */
9123 unregister_sched_domain_sysctl();
9125 /* Let architecture update cpu core mappings. */
9126 new_topology
= arch_update_cpu_topology();
9128 n
= doms_new
? ndoms_new
: 0;
9130 /* Destroy deleted domains */
9131 for (i
= 0; i
< ndoms_cur
; i
++) {
9132 for (j
= 0; j
< n
&& !new_topology
; j
++) {
9133 if (cpumask_equal(doms_cur
[i
], doms_new
[j
])
9134 && dattrs_equal(dattr_cur
, i
, dattr_new
, j
))
9137 /* no match - a current sched domain not in new doms_new[] */
9138 detach_destroy_domains(doms_cur
[i
]);
9143 if (doms_new
== NULL
) {
9145 doms_new
= &fallback_doms
;
9146 cpumask_andnot(doms_new
[0], cpu_active_mask
, cpu_isolated_map
);
9147 WARN_ON_ONCE(dattr_new
);
9150 /* Build new domains */
9151 for (i
= 0; i
< ndoms_new
; i
++) {
9152 for (j
= 0; j
< ndoms_cur
&& !new_topology
; j
++) {
9153 if (cpumask_equal(doms_new
[i
], doms_cur
[j
])
9154 && dattrs_equal(dattr_new
, i
, dattr_cur
, j
))
9157 /* no match - add a new doms_new */
9158 __build_sched_domains(doms_new
[i
],
9159 dattr_new
? dattr_new
+ i
: NULL
);
9164 /* Remember the new sched domains */
9165 if (doms_cur
!= &fallback_doms
)
9166 free_sched_domains(doms_cur
, ndoms_cur
);
9167 kfree(dattr_cur
); /* kfree(NULL) is safe */
9168 doms_cur
= doms_new
;
9169 dattr_cur
= dattr_new
;
9170 ndoms_cur
= ndoms_new
;
9172 register_sched_domain_sysctl();
9174 mutex_unlock(&sched_domains_mutex
);
9177 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
9178 static void arch_reinit_sched_domains(void)
9182 /* Destroy domains first to force the rebuild */
9183 partition_sched_domains(0, NULL
, NULL
);
9185 rebuild_sched_domains();
9189 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
9191 unsigned int level
= 0;
9193 if (sscanf(buf
, "%u", &level
) != 1)
9197 * level is always be positive so don't check for
9198 * level < POWERSAVINGS_BALANCE_NONE which is 0
9199 * What happens on 0 or 1 byte write,
9200 * need to check for count as well?
9203 if (level
>= MAX_POWERSAVINGS_BALANCE_LEVELS
)
9207 sched_smt_power_savings
= level
;
9209 sched_mc_power_savings
= level
;
9211 arch_reinit_sched_domains();
9216 #ifdef CONFIG_SCHED_MC
9217 static ssize_t
sched_mc_power_savings_show(struct sysdev_class
*class,
9220 return sprintf(page
, "%u\n", sched_mc_power_savings
);
9222 static ssize_t
sched_mc_power_savings_store(struct sysdev_class
*class,
9223 const char *buf
, size_t count
)
9225 return sched_power_savings_store(buf
, count
, 0);
9227 static SYSDEV_CLASS_ATTR(sched_mc_power_savings
, 0644,
9228 sched_mc_power_savings_show
,
9229 sched_mc_power_savings_store
);
9232 #ifdef CONFIG_SCHED_SMT
9233 static ssize_t
sched_smt_power_savings_show(struct sysdev_class
*dev
,
9236 return sprintf(page
, "%u\n", sched_smt_power_savings
);
9238 static ssize_t
sched_smt_power_savings_store(struct sysdev_class
*dev
,
9239 const char *buf
, size_t count
)
9241 return sched_power_savings_store(buf
, count
, 1);
9243 static SYSDEV_CLASS_ATTR(sched_smt_power_savings
, 0644,
9244 sched_smt_power_savings_show
,
9245 sched_smt_power_savings_store
);
9248 int __init
sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
9252 #ifdef CONFIG_SCHED_SMT
9254 err
= sysfs_create_file(&cls
->kset
.kobj
,
9255 &attr_sched_smt_power_savings
.attr
);
9257 #ifdef CONFIG_SCHED_MC
9258 if (!err
&& mc_capable())
9259 err
= sysfs_create_file(&cls
->kset
.kobj
,
9260 &attr_sched_mc_power_savings
.attr
);
9264 #endif /* CONFIG_SCHED_MC || CONFIG_SCHED_SMT */
9266 #ifndef CONFIG_CPUSETS
9268 * Add online and remove offline CPUs from the scheduler domains.
9269 * When cpusets are enabled they take over this function.
9271 static int update_sched_domains(struct notifier_block
*nfb
,
9272 unsigned long action
, void *hcpu
)
9276 case CPU_ONLINE_FROZEN
:
9277 case CPU_DOWN_PREPARE
:
9278 case CPU_DOWN_PREPARE_FROZEN
:
9279 case CPU_DOWN_FAILED
:
9280 case CPU_DOWN_FAILED_FROZEN
:
9281 partition_sched_domains(1, NULL
, NULL
);
9290 static int update_runtime(struct notifier_block
*nfb
,
9291 unsigned long action
, void *hcpu
)
9293 int cpu
= (int)(long)hcpu
;
9296 case CPU_DOWN_PREPARE
:
9297 case CPU_DOWN_PREPARE_FROZEN
:
9298 disable_runtime(cpu_rq(cpu
));
9301 case CPU_DOWN_FAILED
:
9302 case CPU_DOWN_FAILED_FROZEN
:
9304 case CPU_ONLINE_FROZEN
:
9305 enable_runtime(cpu_rq(cpu
));
9313 void __init
sched_init_smp(void)
9315 cpumask_var_t non_isolated_cpus
;
9317 alloc_cpumask_var(&non_isolated_cpus
, GFP_KERNEL
);
9318 alloc_cpumask_var(&fallback_doms
, GFP_KERNEL
);
9320 #if defined(CONFIG_NUMA)
9321 sched_group_nodes_bycpu
= kzalloc(nr_cpu_ids
* sizeof(void **),
9323 BUG_ON(sched_group_nodes_bycpu
== NULL
);
9326 mutex_lock(&sched_domains_mutex
);
9327 arch_init_sched_domains(cpu_active_mask
);
9328 cpumask_andnot(non_isolated_cpus
, cpu_possible_mask
, cpu_isolated_map
);
9329 if (cpumask_empty(non_isolated_cpus
))
9330 cpumask_set_cpu(smp_processor_id(), non_isolated_cpus
);
9331 mutex_unlock(&sched_domains_mutex
);
9334 #ifndef CONFIG_CPUSETS
9335 /* XXX: Theoretical race here - CPU may be hotplugged now */
9336 hotcpu_notifier(update_sched_domains
, 0);
9339 /* RT runtime code needs to handle some hotplug events */
9340 hotcpu_notifier(update_runtime
, 0);
9344 /* Move init over to a non-isolated CPU */
9345 if (set_cpus_allowed_ptr(current
, non_isolated_cpus
) < 0)
9347 sched_init_granularity();
9348 free_cpumask_var(non_isolated_cpus
);
9350 init_sched_rt_class();
9353 void __init
sched_init_smp(void)
9355 sched_init_granularity();
9357 #endif /* CONFIG_SMP */
9359 const_debug
unsigned int sysctl_timer_migration
= 1;
9361 int in_sched_functions(unsigned long addr
)
9363 return in_lock_functions(addr
) ||
9364 (addr
>= (unsigned long)__sched_text_start
9365 && addr
< (unsigned long)__sched_text_end
);
9368 static void init_cfs_rq(struct cfs_rq
*cfs_rq
, struct rq
*rq
)
9370 cfs_rq
->tasks_timeline
= RB_ROOT
;
9371 INIT_LIST_HEAD(&cfs_rq
->tasks
);
9372 #ifdef CONFIG_FAIR_GROUP_SCHED
9375 cfs_rq
->min_vruntime
= (u64
)(-(1LL << 20));
9378 static void init_rt_rq(struct rt_rq
*rt_rq
, struct rq
*rq
)
9380 struct rt_prio_array
*array
;
9383 array
= &rt_rq
->active
;
9384 for (i
= 0; i
< MAX_RT_PRIO
; i
++) {
9385 INIT_LIST_HEAD(array
->queue
+ i
);
9386 __clear_bit(i
, array
->bitmap
);
9388 /* delimiter for bitsearch: */
9389 __set_bit(MAX_RT_PRIO
, array
->bitmap
);
9391 #if defined CONFIG_SMP || defined CONFIG_RT_GROUP_SCHED
9392 rt_rq
->highest_prio
.curr
= MAX_RT_PRIO
;
9394 rt_rq
->highest_prio
.next
= MAX_RT_PRIO
;
9398 rt_rq
->rt_nr_migratory
= 0;
9399 rt_rq
->overloaded
= 0;
9400 plist_head_init_raw(&rt_rq
->pushable_tasks
, &rq
->lock
);
9404 rt_rq
->rt_throttled
= 0;
9405 rt_rq
->rt_runtime
= 0;
9406 raw_spin_lock_init(&rt_rq
->rt_runtime_lock
);
9408 #ifdef CONFIG_RT_GROUP_SCHED
9409 rt_rq
->rt_nr_boosted
= 0;
9414 #ifdef CONFIG_FAIR_GROUP_SCHED
9415 static void init_tg_cfs_entry(struct task_group
*tg
, struct cfs_rq
*cfs_rq
,
9416 struct sched_entity
*se
, int cpu
, int add
,
9417 struct sched_entity
*parent
)
9419 struct rq
*rq
= cpu_rq(cpu
);
9420 tg
->cfs_rq
[cpu
] = cfs_rq
;
9421 init_cfs_rq(cfs_rq
, rq
);
9424 list_add(&cfs_rq
->leaf_cfs_rq_list
, &rq
->leaf_cfs_rq_list
);
9427 /* se could be NULL for init_task_group */
9432 se
->cfs_rq
= &rq
->cfs
;
9434 se
->cfs_rq
= parent
->my_q
;
9437 se
->load
.weight
= tg
->shares
;
9438 se
->load
.inv_weight
= 0;
9439 se
->parent
= parent
;
9443 #ifdef CONFIG_RT_GROUP_SCHED
9444 static void init_tg_rt_entry(struct task_group
*tg
, struct rt_rq
*rt_rq
,
9445 struct sched_rt_entity
*rt_se
, int cpu
, int add
,
9446 struct sched_rt_entity
*parent
)
9448 struct rq
*rq
= cpu_rq(cpu
);
9450 tg
->rt_rq
[cpu
] = rt_rq
;
9451 init_rt_rq(rt_rq
, rq
);
9453 rt_rq
->rt_se
= rt_se
;
9454 rt_rq
->rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
9456 list_add(&rt_rq
->leaf_rt_rq_list
, &rq
->leaf_rt_rq_list
);
9458 tg
->rt_se
[cpu
] = rt_se
;
9463 rt_se
->rt_rq
= &rq
->rt
;
9465 rt_se
->rt_rq
= parent
->my_q
;
9467 rt_se
->my_q
= rt_rq
;
9468 rt_se
->parent
= parent
;
9469 INIT_LIST_HEAD(&rt_se
->run_list
);
9473 void __init
sched_init(void)
9476 unsigned long alloc_size
= 0, ptr
;
9478 #ifdef CONFIG_FAIR_GROUP_SCHED
9479 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9481 #ifdef CONFIG_RT_GROUP_SCHED
9482 alloc_size
+= 2 * nr_cpu_ids
* sizeof(void **);
9484 #ifdef CONFIG_USER_SCHED
9487 #ifdef CONFIG_CPUMASK_OFFSTACK
9488 alloc_size
+= num_possible_cpus() * cpumask_size();
9491 ptr
= (unsigned long)kzalloc(alloc_size
, GFP_NOWAIT
);
9493 #ifdef CONFIG_FAIR_GROUP_SCHED
9494 init_task_group
.se
= (struct sched_entity
**)ptr
;
9495 ptr
+= nr_cpu_ids
* sizeof(void **);
9497 init_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9498 ptr
+= nr_cpu_ids
* sizeof(void **);
9500 #ifdef CONFIG_USER_SCHED
9501 root_task_group
.se
= (struct sched_entity
**)ptr
;
9502 ptr
+= nr_cpu_ids
* sizeof(void **);
9504 root_task_group
.cfs_rq
= (struct cfs_rq
**)ptr
;
9505 ptr
+= nr_cpu_ids
* sizeof(void **);
9506 #endif /* CONFIG_USER_SCHED */
9507 #endif /* CONFIG_FAIR_GROUP_SCHED */
9508 #ifdef CONFIG_RT_GROUP_SCHED
9509 init_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9510 ptr
+= nr_cpu_ids
* sizeof(void **);
9512 init_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9513 ptr
+= nr_cpu_ids
* sizeof(void **);
9515 #ifdef CONFIG_USER_SCHED
9516 root_task_group
.rt_se
= (struct sched_rt_entity
**)ptr
;
9517 ptr
+= nr_cpu_ids
* sizeof(void **);
9519 root_task_group
.rt_rq
= (struct rt_rq
**)ptr
;
9520 ptr
+= nr_cpu_ids
* sizeof(void **);
9521 #endif /* CONFIG_USER_SCHED */
9522 #endif /* CONFIG_RT_GROUP_SCHED */
9523 #ifdef CONFIG_CPUMASK_OFFSTACK
9524 for_each_possible_cpu(i
) {
9525 per_cpu(load_balance_tmpmask
, i
) = (void *)ptr
;
9526 ptr
+= cpumask_size();
9528 #endif /* CONFIG_CPUMASK_OFFSTACK */
9532 init_defrootdomain();
9535 init_rt_bandwidth(&def_rt_bandwidth
,
9536 global_rt_period(), global_rt_runtime());
9538 #ifdef CONFIG_RT_GROUP_SCHED
9539 init_rt_bandwidth(&init_task_group
.rt_bandwidth
,
9540 global_rt_period(), global_rt_runtime());
9541 #ifdef CONFIG_USER_SCHED
9542 init_rt_bandwidth(&root_task_group
.rt_bandwidth
,
9543 global_rt_period(), RUNTIME_INF
);
9544 #endif /* CONFIG_USER_SCHED */
9545 #endif /* CONFIG_RT_GROUP_SCHED */
9547 #ifdef CONFIG_GROUP_SCHED
9548 list_add(&init_task_group
.list
, &task_groups
);
9549 INIT_LIST_HEAD(&init_task_group
.children
);
9551 #ifdef CONFIG_USER_SCHED
9552 INIT_LIST_HEAD(&root_task_group
.children
);
9553 init_task_group
.parent
= &root_task_group
;
9554 list_add(&init_task_group
.siblings
, &root_task_group
.children
);
9555 #endif /* CONFIG_USER_SCHED */
9556 #endif /* CONFIG_GROUP_SCHED */
9558 #if defined CONFIG_FAIR_GROUP_SCHED && defined CONFIG_SMP
9559 update_shares_data
= __alloc_percpu(nr_cpu_ids
* sizeof(unsigned long),
9560 __alignof__(unsigned long));
9562 for_each_possible_cpu(i
) {
9566 raw_spin_lock_init(&rq
->lock
);
9568 rq
->calc_load_active
= 0;
9569 rq
->calc_load_update
= jiffies
+ LOAD_FREQ
;
9570 init_cfs_rq(&rq
->cfs
, rq
);
9571 init_rt_rq(&rq
->rt
, rq
);
9572 #ifdef CONFIG_FAIR_GROUP_SCHED
9573 init_task_group
.shares
= init_task_group_load
;
9574 INIT_LIST_HEAD(&rq
->leaf_cfs_rq_list
);
9575 #ifdef CONFIG_CGROUP_SCHED
9577 * How much cpu bandwidth does init_task_group get?
9579 * In case of task-groups formed thr' the cgroup filesystem, it
9580 * gets 100% of the cpu resources in the system. This overall
9581 * system cpu resource is divided among the tasks of
9582 * init_task_group and its child task-groups in a fair manner,
9583 * based on each entity's (task or task-group's) weight
9584 * (se->load.weight).
9586 * In other words, if init_task_group has 10 tasks of weight
9587 * 1024) and two child groups A0 and A1 (of weight 1024 each),
9588 * then A0's share of the cpu resource is:
9590 * A0's bandwidth = 1024 / (10*1024 + 1024 + 1024) = 8.33%
9592 * We achieve this by letting init_task_group's tasks sit
9593 * directly in rq->cfs (i.e init_task_group->se[] = NULL).
9595 init_tg_cfs_entry(&init_task_group
, &rq
->cfs
, NULL
, i
, 1, NULL
);
9596 #elif defined CONFIG_USER_SCHED
9597 root_task_group
.shares
= NICE_0_LOAD
;
9598 init_tg_cfs_entry(&root_task_group
, &rq
->cfs
, NULL
, i
, 0, NULL
);
9600 * In case of task-groups formed thr' the user id of tasks,
9601 * init_task_group represents tasks belonging to root user.
9602 * Hence it forms a sibling of all subsequent groups formed.
9603 * In this case, init_task_group gets only a fraction of overall
9604 * system cpu resource, based on the weight assigned to root
9605 * user's cpu share (INIT_TASK_GROUP_LOAD). This is accomplished
9606 * by letting tasks of init_task_group sit in a separate cfs_rq
9607 * (init_tg_cfs_rq) and having one entity represent this group of
9608 * tasks in rq->cfs (i.e init_task_group->se[] != NULL).
9610 init_tg_cfs_entry(&init_task_group
,
9611 &per_cpu(init_tg_cfs_rq
, i
),
9612 &per_cpu(init_sched_entity
, i
), i
, 1,
9613 root_task_group
.se
[i
]);
9616 #endif /* CONFIG_FAIR_GROUP_SCHED */
9618 rq
->rt
.rt_runtime
= def_rt_bandwidth
.rt_runtime
;
9619 #ifdef CONFIG_RT_GROUP_SCHED
9620 INIT_LIST_HEAD(&rq
->leaf_rt_rq_list
);
9621 #ifdef CONFIG_CGROUP_SCHED
9622 init_tg_rt_entry(&init_task_group
, &rq
->rt
, NULL
, i
, 1, NULL
);
9623 #elif defined CONFIG_USER_SCHED
9624 init_tg_rt_entry(&root_task_group
, &rq
->rt
, NULL
, i
, 0, NULL
);
9625 init_tg_rt_entry(&init_task_group
,
9626 &per_cpu(init_rt_rq_var
, i
),
9627 &per_cpu(init_sched_rt_entity
, i
), i
, 1,
9628 root_task_group
.rt_se
[i
]);
9632 for (j
= 0; j
< CPU_LOAD_IDX_MAX
; j
++)
9633 rq
->cpu_load
[j
] = 0;
9637 rq
->post_schedule
= 0;
9638 rq
->active_balance
= 0;
9639 rq
->next_balance
= jiffies
;
9643 rq
->migration_thread
= NULL
;
9645 rq
->avg_idle
= 2*sysctl_sched_migration_cost
;
9646 INIT_LIST_HEAD(&rq
->migration_queue
);
9647 rq_attach_root(rq
, &def_root_domain
);
9650 atomic_set(&rq
->nr_iowait
, 0);
9653 set_load_weight(&init_task
);
9655 #ifdef CONFIG_PREEMPT_NOTIFIERS
9656 INIT_HLIST_HEAD(&init_task
.preempt_notifiers
);
9660 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
);
9663 #ifdef CONFIG_RT_MUTEXES
9664 plist_head_init_raw(&init_task
.pi_waiters
, &init_task
.pi_lock
);
9668 * The boot idle thread does lazy MMU switching as well:
9670 atomic_inc(&init_mm
.mm_count
);
9671 enter_lazy_tlb(&init_mm
, current
);
9674 * Make us the idle thread. Technically, schedule() should not be
9675 * called from this thread, however somewhere below it might be,
9676 * but because we are the idle thread, we just pick up running again
9677 * when this runqueue becomes "idle".
9679 init_idle(current
, smp_processor_id());
9681 calc_load_update
= jiffies
+ LOAD_FREQ
;
9684 * During early bootup we pretend to be a normal task:
9686 current
->sched_class
= &fair_sched_class
;
9688 /* Allocate the nohz_cpu_mask if CONFIG_CPUMASK_OFFSTACK */
9689 zalloc_cpumask_var(&nohz_cpu_mask
, GFP_NOWAIT
);
9692 zalloc_cpumask_var(&nohz
.cpu_mask
, GFP_NOWAIT
);
9693 alloc_cpumask_var(&nohz
.ilb_grp_nohz_mask
, GFP_NOWAIT
);
9695 /* May be allocated at isolcpus cmdline parse time */
9696 if (cpu_isolated_map
== NULL
)
9697 zalloc_cpumask_var(&cpu_isolated_map
, GFP_NOWAIT
);
9702 scheduler_running
= 1;
9705 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
9706 static inline int preempt_count_equals(int preempt_offset
)
9708 int nested
= preempt_count() & ~PREEMPT_ACTIVE
;
9710 return (nested
== PREEMPT_INATOMIC_BASE
+ preempt_offset
);
9713 void __might_sleep(char *file
, int line
, int preempt_offset
)
9716 static unsigned long prev_jiffy
; /* ratelimiting */
9718 if ((preempt_count_equals(preempt_offset
) && !irqs_disabled()) ||
9719 system_state
!= SYSTEM_RUNNING
|| oops_in_progress
)
9721 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
9723 prev_jiffy
= jiffies
;
9725 pr_err("BUG: sleeping function called from invalid context at %s:%d\n",
9727 pr_err("in_atomic(): %d, irqs_disabled(): %d, pid: %d, name: %s\n",
9728 in_atomic(), irqs_disabled(),
9729 current
->pid
, current
->comm
);
9731 debug_show_held_locks(current
);
9732 if (irqs_disabled())
9733 print_irqtrace_events(current
);
9737 EXPORT_SYMBOL(__might_sleep
);
9740 #ifdef CONFIG_MAGIC_SYSRQ
9741 static void normalize_task(struct rq
*rq
, struct task_struct
*p
)
9745 update_rq_clock(rq
);
9746 on_rq
= p
->se
.on_rq
;
9748 deactivate_task(rq
, p
, 0);
9749 __setscheduler(rq
, p
, SCHED_NORMAL
, 0);
9751 activate_task(rq
, p
, 0);
9752 resched_task(rq
->curr
);
9756 void normalize_rt_tasks(void)
9758 struct task_struct
*g
, *p
;
9759 unsigned long flags
;
9762 read_lock_irqsave(&tasklist_lock
, flags
);
9763 do_each_thread(g
, p
) {
9765 * Only normalize user tasks:
9770 p
->se
.exec_start
= 0;
9771 #ifdef CONFIG_SCHEDSTATS
9772 p
->se
.wait_start
= 0;
9773 p
->se
.sleep_start
= 0;
9774 p
->se
.block_start
= 0;
9779 * Renice negative nice level userspace
9782 if (TASK_NICE(p
) < 0 && p
->mm
)
9783 set_user_nice(p
, 0);
9787 raw_spin_lock(&p
->pi_lock
);
9788 rq
= __task_rq_lock(p
);
9790 normalize_task(rq
, p
);
9792 __task_rq_unlock(rq
);
9793 raw_spin_unlock(&p
->pi_lock
);
9794 } while_each_thread(g
, p
);
9796 read_unlock_irqrestore(&tasklist_lock
, flags
);
9799 #endif /* CONFIG_MAGIC_SYSRQ */
9803 * These functions are only useful for the IA64 MCA handling.
9805 * They can only be called when the whole system has been
9806 * stopped - every CPU needs to be quiescent, and no scheduling
9807 * activity can take place. Using them for anything else would
9808 * be a serious bug, and as a result, they aren't even visible
9809 * under any other configuration.
9813 * curr_task - return the current task for a given cpu.
9814 * @cpu: the processor in question.
9816 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9818 struct task_struct
*curr_task(int cpu
)
9820 return cpu_curr(cpu
);
9824 * set_curr_task - set the current task for a given cpu.
9825 * @cpu: the processor in question.
9826 * @p: the task pointer to set.
9828 * Description: This function must only be used when non-maskable interrupts
9829 * are serviced on a separate stack. It allows the architecture to switch the
9830 * notion of the current task on a cpu in a non-blocking manner. This function
9831 * must be called with all CPU's synchronized, and interrupts disabled, the
9832 * and caller must save the original value of the current task (see
9833 * curr_task() above) and restore that value before reenabling interrupts and
9834 * re-starting the system.
9836 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
9838 void set_curr_task(int cpu
, struct task_struct
*p
)
9845 #ifdef CONFIG_FAIR_GROUP_SCHED
9846 static void free_fair_sched_group(struct task_group
*tg
)
9850 for_each_possible_cpu(i
) {
9852 kfree(tg
->cfs_rq
[i
]);
9862 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9864 struct cfs_rq
*cfs_rq
;
9865 struct sched_entity
*se
;
9869 tg
->cfs_rq
= kzalloc(sizeof(cfs_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9872 tg
->se
= kzalloc(sizeof(se
) * nr_cpu_ids
, GFP_KERNEL
);
9876 tg
->shares
= NICE_0_LOAD
;
9878 for_each_possible_cpu(i
) {
9881 cfs_rq
= kzalloc_node(sizeof(struct cfs_rq
),
9882 GFP_KERNEL
, cpu_to_node(i
));
9886 se
= kzalloc_node(sizeof(struct sched_entity
),
9887 GFP_KERNEL
, cpu_to_node(i
));
9891 init_tg_cfs_entry(tg
, cfs_rq
, se
, i
, 0, parent
->se
[i
]);
9902 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9904 list_add_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
,
9905 &cpu_rq(cpu
)->leaf_cfs_rq_list
);
9908 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9910 list_del_rcu(&tg
->cfs_rq
[cpu
]->leaf_cfs_rq_list
);
9912 #else /* !CONFG_FAIR_GROUP_SCHED */
9913 static inline void free_fair_sched_group(struct task_group
*tg
)
9918 int alloc_fair_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9923 static inline void register_fair_sched_group(struct task_group
*tg
, int cpu
)
9927 static inline void unregister_fair_sched_group(struct task_group
*tg
, int cpu
)
9930 #endif /* CONFIG_FAIR_GROUP_SCHED */
9932 #ifdef CONFIG_RT_GROUP_SCHED
9933 static void free_rt_sched_group(struct task_group
*tg
)
9937 destroy_rt_bandwidth(&tg
->rt_bandwidth
);
9939 for_each_possible_cpu(i
) {
9941 kfree(tg
->rt_rq
[i
]);
9943 kfree(tg
->rt_se
[i
]);
9951 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
9953 struct rt_rq
*rt_rq
;
9954 struct sched_rt_entity
*rt_se
;
9958 tg
->rt_rq
= kzalloc(sizeof(rt_rq
) * nr_cpu_ids
, GFP_KERNEL
);
9961 tg
->rt_se
= kzalloc(sizeof(rt_se
) * nr_cpu_ids
, GFP_KERNEL
);
9965 init_rt_bandwidth(&tg
->rt_bandwidth
,
9966 ktime_to_ns(def_rt_bandwidth
.rt_period
), 0);
9968 for_each_possible_cpu(i
) {
9971 rt_rq
= kzalloc_node(sizeof(struct rt_rq
),
9972 GFP_KERNEL
, cpu_to_node(i
));
9976 rt_se
= kzalloc_node(sizeof(struct sched_rt_entity
),
9977 GFP_KERNEL
, cpu_to_node(i
));
9981 init_tg_rt_entry(tg
, rt_rq
, rt_se
, i
, 0, parent
->rt_se
[i
]);
9992 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
9994 list_add_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
,
9995 &cpu_rq(cpu
)->leaf_rt_rq_list
);
9998 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10000 list_del_rcu(&tg
->rt_rq
[cpu
]->leaf_rt_rq_list
);
10002 #else /* !CONFIG_RT_GROUP_SCHED */
10003 static inline void free_rt_sched_group(struct task_group
*tg
)
10008 int alloc_rt_sched_group(struct task_group
*tg
, struct task_group
*parent
)
10013 static inline void register_rt_sched_group(struct task_group
*tg
, int cpu
)
10017 static inline void unregister_rt_sched_group(struct task_group
*tg
, int cpu
)
10020 #endif /* CONFIG_RT_GROUP_SCHED */
10022 #ifdef CONFIG_GROUP_SCHED
10023 static void free_sched_group(struct task_group
*tg
)
10025 free_fair_sched_group(tg
);
10026 free_rt_sched_group(tg
);
10030 /* allocate runqueue etc for a new task group */
10031 struct task_group
*sched_create_group(struct task_group
*parent
)
10033 struct task_group
*tg
;
10034 unsigned long flags
;
10037 tg
= kzalloc(sizeof(*tg
), GFP_KERNEL
);
10039 return ERR_PTR(-ENOMEM
);
10041 if (!alloc_fair_sched_group(tg
, parent
))
10044 if (!alloc_rt_sched_group(tg
, parent
))
10047 spin_lock_irqsave(&task_group_lock
, flags
);
10048 for_each_possible_cpu(i
) {
10049 register_fair_sched_group(tg
, i
);
10050 register_rt_sched_group(tg
, i
);
10052 list_add_rcu(&tg
->list
, &task_groups
);
10054 WARN_ON(!parent
); /* root should already exist */
10056 tg
->parent
= parent
;
10057 INIT_LIST_HEAD(&tg
->children
);
10058 list_add_rcu(&tg
->siblings
, &parent
->children
);
10059 spin_unlock_irqrestore(&task_group_lock
, flags
);
10064 free_sched_group(tg
);
10065 return ERR_PTR(-ENOMEM
);
10068 /* rcu callback to free various structures associated with a task group */
10069 static void free_sched_group_rcu(struct rcu_head
*rhp
)
10071 /* now it should be safe to free those cfs_rqs */
10072 free_sched_group(container_of(rhp
, struct task_group
, rcu
));
10075 /* Destroy runqueue etc associated with a task group */
10076 void sched_destroy_group(struct task_group
*tg
)
10078 unsigned long flags
;
10081 spin_lock_irqsave(&task_group_lock
, flags
);
10082 for_each_possible_cpu(i
) {
10083 unregister_fair_sched_group(tg
, i
);
10084 unregister_rt_sched_group(tg
, i
);
10086 list_del_rcu(&tg
->list
);
10087 list_del_rcu(&tg
->siblings
);
10088 spin_unlock_irqrestore(&task_group_lock
, flags
);
10090 /* wait for possible concurrent references to cfs_rqs complete */
10091 call_rcu(&tg
->rcu
, free_sched_group_rcu
);
10094 /* change task's runqueue when it moves between groups.
10095 * The caller of this function should have put the task in its new group
10096 * by now. This function just updates tsk->se.cfs_rq and tsk->se.parent to
10097 * reflect its new group.
10099 void sched_move_task(struct task_struct
*tsk
)
10101 int on_rq
, running
;
10102 unsigned long flags
;
10105 rq
= task_rq_lock(tsk
, &flags
);
10107 update_rq_clock(rq
);
10109 running
= task_current(rq
, tsk
);
10110 on_rq
= tsk
->se
.on_rq
;
10113 dequeue_task(rq
, tsk
, 0);
10114 if (unlikely(running
))
10115 tsk
->sched_class
->put_prev_task(rq
, tsk
);
10117 set_task_rq(tsk
, task_cpu(tsk
));
10119 #ifdef CONFIG_FAIR_GROUP_SCHED
10120 if (tsk
->sched_class
->moved_group
)
10121 tsk
->sched_class
->moved_group(tsk
);
10124 if (unlikely(running
))
10125 tsk
->sched_class
->set_curr_task(rq
);
10127 enqueue_task(rq
, tsk
, 0);
10129 task_rq_unlock(rq
, &flags
);
10131 #endif /* CONFIG_GROUP_SCHED */
10133 #ifdef CONFIG_FAIR_GROUP_SCHED
10134 static void __set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10136 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10141 dequeue_entity(cfs_rq
, se
, 0);
10143 se
->load
.weight
= shares
;
10144 se
->load
.inv_weight
= 0;
10147 enqueue_entity(cfs_rq
, se
, 0);
10150 static void set_se_shares(struct sched_entity
*se
, unsigned long shares
)
10152 struct cfs_rq
*cfs_rq
= se
->cfs_rq
;
10153 struct rq
*rq
= cfs_rq
->rq
;
10154 unsigned long flags
;
10156 raw_spin_lock_irqsave(&rq
->lock
, flags
);
10157 __set_se_shares(se
, shares
);
10158 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
10161 static DEFINE_MUTEX(shares_mutex
);
10163 int sched_group_set_shares(struct task_group
*tg
, unsigned long shares
)
10166 unsigned long flags
;
10169 * We can't change the weight of the root cgroup.
10174 if (shares
< MIN_SHARES
)
10175 shares
= MIN_SHARES
;
10176 else if (shares
> MAX_SHARES
)
10177 shares
= MAX_SHARES
;
10179 mutex_lock(&shares_mutex
);
10180 if (tg
->shares
== shares
)
10183 spin_lock_irqsave(&task_group_lock
, flags
);
10184 for_each_possible_cpu(i
)
10185 unregister_fair_sched_group(tg
, i
);
10186 list_del_rcu(&tg
->siblings
);
10187 spin_unlock_irqrestore(&task_group_lock
, flags
);
10189 /* wait for any ongoing reference to this group to finish */
10190 synchronize_sched();
10193 * Now we are free to modify the group's share on each cpu
10194 * w/o tripping rebalance_share or load_balance_fair.
10196 tg
->shares
= shares
;
10197 for_each_possible_cpu(i
) {
10199 * force a rebalance
10201 cfs_rq_set_shares(tg
->cfs_rq
[i
], 0);
10202 set_se_shares(tg
->se
[i
], shares
);
10206 * Enable load balance activity on this group, by inserting it back on
10207 * each cpu's rq->leaf_cfs_rq_list.
10209 spin_lock_irqsave(&task_group_lock
, flags
);
10210 for_each_possible_cpu(i
)
10211 register_fair_sched_group(tg
, i
);
10212 list_add_rcu(&tg
->siblings
, &tg
->parent
->children
);
10213 spin_unlock_irqrestore(&task_group_lock
, flags
);
10215 mutex_unlock(&shares_mutex
);
10219 unsigned long sched_group_shares(struct task_group
*tg
)
10225 #ifdef CONFIG_RT_GROUP_SCHED
10227 * Ensure that the real time constraints are schedulable.
10229 static DEFINE_MUTEX(rt_constraints_mutex
);
10231 static unsigned long to_ratio(u64 period
, u64 runtime
)
10233 if (runtime
== RUNTIME_INF
)
10236 return div64_u64(runtime
<< 20, period
);
10239 /* Must be called with tasklist_lock held */
10240 static inline int tg_has_rt_tasks(struct task_group
*tg
)
10242 struct task_struct
*g
, *p
;
10244 do_each_thread(g
, p
) {
10245 if (rt_task(p
) && rt_rq_of_se(&p
->rt
)->tg
== tg
)
10247 } while_each_thread(g
, p
);
10252 struct rt_schedulable_data
{
10253 struct task_group
*tg
;
10258 static int tg_schedulable(struct task_group
*tg
, void *data
)
10260 struct rt_schedulable_data
*d
= data
;
10261 struct task_group
*child
;
10262 unsigned long total
, sum
= 0;
10263 u64 period
, runtime
;
10265 period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10266 runtime
= tg
->rt_bandwidth
.rt_runtime
;
10269 period
= d
->rt_period
;
10270 runtime
= d
->rt_runtime
;
10273 #ifdef CONFIG_USER_SCHED
10274 if (tg
== &root_task_group
) {
10275 period
= global_rt_period();
10276 runtime
= global_rt_runtime();
10281 * Cannot have more runtime than the period.
10283 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10287 * Ensure we don't starve existing RT tasks.
10289 if (rt_bandwidth_enabled() && !runtime
&& tg_has_rt_tasks(tg
))
10292 total
= to_ratio(period
, runtime
);
10295 * Nobody can have more than the global setting allows.
10297 if (total
> to_ratio(global_rt_period(), global_rt_runtime()))
10301 * The sum of our children's runtime should not exceed our own.
10303 list_for_each_entry_rcu(child
, &tg
->children
, siblings
) {
10304 period
= ktime_to_ns(child
->rt_bandwidth
.rt_period
);
10305 runtime
= child
->rt_bandwidth
.rt_runtime
;
10307 if (child
== d
->tg
) {
10308 period
= d
->rt_period
;
10309 runtime
= d
->rt_runtime
;
10312 sum
+= to_ratio(period
, runtime
);
10321 static int __rt_schedulable(struct task_group
*tg
, u64 period
, u64 runtime
)
10323 struct rt_schedulable_data data
= {
10325 .rt_period
= period
,
10326 .rt_runtime
= runtime
,
10329 return walk_tg_tree(tg_schedulable
, tg_nop
, &data
);
10332 static int tg_set_bandwidth(struct task_group
*tg
,
10333 u64 rt_period
, u64 rt_runtime
)
10337 mutex_lock(&rt_constraints_mutex
);
10338 read_lock(&tasklist_lock
);
10339 err
= __rt_schedulable(tg
, rt_period
, rt_runtime
);
10343 raw_spin_lock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10344 tg
->rt_bandwidth
.rt_period
= ns_to_ktime(rt_period
);
10345 tg
->rt_bandwidth
.rt_runtime
= rt_runtime
;
10347 for_each_possible_cpu(i
) {
10348 struct rt_rq
*rt_rq
= tg
->rt_rq
[i
];
10350 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
10351 rt_rq
->rt_runtime
= rt_runtime
;
10352 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
10354 raw_spin_unlock_irq(&tg
->rt_bandwidth
.rt_runtime_lock
);
10356 read_unlock(&tasklist_lock
);
10357 mutex_unlock(&rt_constraints_mutex
);
10362 int sched_group_set_rt_runtime(struct task_group
*tg
, long rt_runtime_us
)
10364 u64 rt_runtime
, rt_period
;
10366 rt_period
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10367 rt_runtime
= (u64
)rt_runtime_us
* NSEC_PER_USEC
;
10368 if (rt_runtime_us
< 0)
10369 rt_runtime
= RUNTIME_INF
;
10371 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10374 long sched_group_rt_runtime(struct task_group
*tg
)
10378 if (tg
->rt_bandwidth
.rt_runtime
== RUNTIME_INF
)
10381 rt_runtime_us
= tg
->rt_bandwidth
.rt_runtime
;
10382 do_div(rt_runtime_us
, NSEC_PER_USEC
);
10383 return rt_runtime_us
;
10386 int sched_group_set_rt_period(struct task_group
*tg
, long rt_period_us
)
10388 u64 rt_runtime
, rt_period
;
10390 rt_period
= (u64
)rt_period_us
* NSEC_PER_USEC
;
10391 rt_runtime
= tg
->rt_bandwidth
.rt_runtime
;
10393 if (rt_period
== 0)
10396 return tg_set_bandwidth(tg
, rt_period
, rt_runtime
);
10399 long sched_group_rt_period(struct task_group
*tg
)
10403 rt_period_us
= ktime_to_ns(tg
->rt_bandwidth
.rt_period
);
10404 do_div(rt_period_us
, NSEC_PER_USEC
);
10405 return rt_period_us
;
10408 static int sched_rt_global_constraints(void)
10410 u64 runtime
, period
;
10413 if (sysctl_sched_rt_period
<= 0)
10416 runtime
= global_rt_runtime();
10417 period
= global_rt_period();
10420 * Sanity check on the sysctl variables.
10422 if (runtime
> period
&& runtime
!= RUNTIME_INF
)
10425 mutex_lock(&rt_constraints_mutex
);
10426 read_lock(&tasklist_lock
);
10427 ret
= __rt_schedulable(NULL
, 0, 0);
10428 read_unlock(&tasklist_lock
);
10429 mutex_unlock(&rt_constraints_mutex
);
10434 int sched_rt_can_attach(struct task_group
*tg
, struct task_struct
*tsk
)
10436 /* Don't accept realtime tasks when there is no way for them to run */
10437 if (rt_task(tsk
) && tg
->rt_bandwidth
.rt_runtime
== 0)
10443 #else /* !CONFIG_RT_GROUP_SCHED */
10444 static int sched_rt_global_constraints(void)
10446 unsigned long flags
;
10449 if (sysctl_sched_rt_period
<= 0)
10453 * There's always some RT tasks in the root group
10454 * -- migration, kstopmachine etc..
10456 if (sysctl_sched_rt_runtime
== 0)
10459 raw_spin_lock_irqsave(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10460 for_each_possible_cpu(i
) {
10461 struct rt_rq
*rt_rq
= &cpu_rq(i
)->rt
;
10463 raw_spin_lock(&rt_rq
->rt_runtime_lock
);
10464 rt_rq
->rt_runtime
= global_rt_runtime();
10465 raw_spin_unlock(&rt_rq
->rt_runtime_lock
);
10467 raw_spin_unlock_irqrestore(&def_rt_bandwidth
.rt_runtime_lock
, flags
);
10471 #endif /* CONFIG_RT_GROUP_SCHED */
10473 int sched_rt_handler(struct ctl_table
*table
, int write
,
10474 void __user
*buffer
, size_t *lenp
,
10478 int old_period
, old_runtime
;
10479 static DEFINE_MUTEX(mutex
);
10481 mutex_lock(&mutex
);
10482 old_period
= sysctl_sched_rt_period
;
10483 old_runtime
= sysctl_sched_rt_runtime
;
10485 ret
= proc_dointvec(table
, write
, buffer
, lenp
, ppos
);
10487 if (!ret
&& write
) {
10488 ret
= sched_rt_global_constraints();
10490 sysctl_sched_rt_period
= old_period
;
10491 sysctl_sched_rt_runtime
= old_runtime
;
10493 def_rt_bandwidth
.rt_runtime
= global_rt_runtime();
10494 def_rt_bandwidth
.rt_period
=
10495 ns_to_ktime(global_rt_period());
10498 mutex_unlock(&mutex
);
10503 #ifdef CONFIG_CGROUP_SCHED
10505 /* return corresponding task_group object of a cgroup */
10506 static inline struct task_group
*cgroup_tg(struct cgroup
*cgrp
)
10508 return container_of(cgroup_subsys_state(cgrp
, cpu_cgroup_subsys_id
),
10509 struct task_group
, css
);
10512 static struct cgroup_subsys_state
*
10513 cpu_cgroup_create(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10515 struct task_group
*tg
, *parent
;
10517 if (!cgrp
->parent
) {
10518 /* This is early initialization for the top cgroup */
10519 return &init_task_group
.css
;
10522 parent
= cgroup_tg(cgrp
->parent
);
10523 tg
= sched_create_group(parent
);
10525 return ERR_PTR(-ENOMEM
);
10531 cpu_cgroup_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10533 struct task_group
*tg
= cgroup_tg(cgrp
);
10535 sched_destroy_group(tg
);
10539 cpu_cgroup_can_attach_task(struct cgroup
*cgrp
, struct task_struct
*tsk
)
10541 #ifdef CONFIG_RT_GROUP_SCHED
10542 if (!sched_rt_can_attach(cgroup_tg(cgrp
), tsk
))
10545 /* We don't support RT-tasks being in separate groups */
10546 if (tsk
->sched_class
!= &fair_sched_class
)
10553 cpu_cgroup_can_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10554 struct task_struct
*tsk
, bool threadgroup
)
10556 int retval
= cpu_cgroup_can_attach_task(cgrp
, tsk
);
10560 struct task_struct
*c
;
10562 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10563 retval
= cpu_cgroup_can_attach_task(cgrp
, c
);
10575 cpu_cgroup_attach(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
,
10576 struct cgroup
*old_cont
, struct task_struct
*tsk
,
10579 sched_move_task(tsk
);
10581 struct task_struct
*c
;
10583 list_for_each_entry_rcu(c
, &tsk
->thread_group
, thread_group
) {
10584 sched_move_task(c
);
10590 #ifdef CONFIG_FAIR_GROUP_SCHED
10591 static int cpu_shares_write_u64(struct cgroup
*cgrp
, struct cftype
*cftype
,
10594 return sched_group_set_shares(cgroup_tg(cgrp
), shareval
);
10597 static u64
cpu_shares_read_u64(struct cgroup
*cgrp
, struct cftype
*cft
)
10599 struct task_group
*tg
= cgroup_tg(cgrp
);
10601 return (u64
) tg
->shares
;
10603 #endif /* CONFIG_FAIR_GROUP_SCHED */
10605 #ifdef CONFIG_RT_GROUP_SCHED
10606 static int cpu_rt_runtime_write(struct cgroup
*cgrp
, struct cftype
*cft
,
10609 return sched_group_set_rt_runtime(cgroup_tg(cgrp
), val
);
10612 static s64
cpu_rt_runtime_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10614 return sched_group_rt_runtime(cgroup_tg(cgrp
));
10617 static int cpu_rt_period_write_uint(struct cgroup
*cgrp
, struct cftype
*cftype
,
10620 return sched_group_set_rt_period(cgroup_tg(cgrp
), rt_period_us
);
10623 static u64
cpu_rt_period_read_uint(struct cgroup
*cgrp
, struct cftype
*cft
)
10625 return sched_group_rt_period(cgroup_tg(cgrp
));
10627 #endif /* CONFIG_RT_GROUP_SCHED */
10629 static struct cftype cpu_files
[] = {
10630 #ifdef CONFIG_FAIR_GROUP_SCHED
10633 .read_u64
= cpu_shares_read_u64
,
10634 .write_u64
= cpu_shares_write_u64
,
10637 #ifdef CONFIG_RT_GROUP_SCHED
10639 .name
= "rt_runtime_us",
10640 .read_s64
= cpu_rt_runtime_read
,
10641 .write_s64
= cpu_rt_runtime_write
,
10644 .name
= "rt_period_us",
10645 .read_u64
= cpu_rt_period_read_uint
,
10646 .write_u64
= cpu_rt_period_write_uint
,
10651 static int cpu_cgroup_populate(struct cgroup_subsys
*ss
, struct cgroup
*cont
)
10653 return cgroup_add_files(cont
, ss
, cpu_files
, ARRAY_SIZE(cpu_files
));
10656 struct cgroup_subsys cpu_cgroup_subsys
= {
10658 .create
= cpu_cgroup_create
,
10659 .destroy
= cpu_cgroup_destroy
,
10660 .can_attach
= cpu_cgroup_can_attach
,
10661 .attach
= cpu_cgroup_attach
,
10662 .populate
= cpu_cgroup_populate
,
10663 .subsys_id
= cpu_cgroup_subsys_id
,
10667 #endif /* CONFIG_CGROUP_SCHED */
10669 #ifdef CONFIG_CGROUP_CPUACCT
10672 * CPU accounting code for task groups.
10674 * Based on the work by Paul Menage (menage@google.com) and Balbir Singh
10675 * (balbir@in.ibm.com).
10678 /* track cpu usage of a group of tasks and its child groups */
10680 struct cgroup_subsys_state css
;
10681 /* cpuusage holds pointer to a u64-type object on every cpu */
10683 struct percpu_counter cpustat
[CPUACCT_STAT_NSTATS
];
10684 struct cpuacct
*parent
;
10687 struct cgroup_subsys cpuacct_subsys
;
10689 /* return cpu accounting group corresponding to this container */
10690 static inline struct cpuacct
*cgroup_ca(struct cgroup
*cgrp
)
10692 return container_of(cgroup_subsys_state(cgrp
, cpuacct_subsys_id
),
10693 struct cpuacct
, css
);
10696 /* return cpu accounting group to which this task belongs */
10697 static inline struct cpuacct
*task_ca(struct task_struct
*tsk
)
10699 return container_of(task_subsys_state(tsk
, cpuacct_subsys_id
),
10700 struct cpuacct
, css
);
10703 /* create a new cpu accounting group */
10704 static struct cgroup_subsys_state
*cpuacct_create(
10705 struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10707 struct cpuacct
*ca
= kzalloc(sizeof(*ca
), GFP_KERNEL
);
10713 ca
->cpuusage
= alloc_percpu(u64
);
10717 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10718 if (percpu_counter_init(&ca
->cpustat
[i
], 0))
10719 goto out_free_counters
;
10722 ca
->parent
= cgroup_ca(cgrp
->parent
);
10728 percpu_counter_destroy(&ca
->cpustat
[i
]);
10729 free_percpu(ca
->cpuusage
);
10733 return ERR_PTR(-ENOMEM
);
10736 /* destroy an existing cpu accounting group */
10738 cpuacct_destroy(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10740 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10743 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++)
10744 percpu_counter_destroy(&ca
->cpustat
[i
]);
10745 free_percpu(ca
->cpuusage
);
10749 static u64
cpuacct_cpuusage_read(struct cpuacct
*ca
, int cpu
)
10751 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10754 #ifndef CONFIG_64BIT
10756 * Take rq->lock to make 64-bit read safe on 32-bit platforms.
10758 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
10760 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10768 static void cpuacct_cpuusage_write(struct cpuacct
*ca
, int cpu
, u64 val
)
10770 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10772 #ifndef CONFIG_64BIT
10774 * Take rq->lock to make 64-bit write safe on 32-bit platforms.
10776 raw_spin_lock_irq(&cpu_rq(cpu
)->lock
);
10778 raw_spin_unlock_irq(&cpu_rq(cpu
)->lock
);
10784 /* return total cpu usage (in nanoseconds) of a group */
10785 static u64
cpuusage_read(struct cgroup
*cgrp
, struct cftype
*cft
)
10787 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10788 u64 totalcpuusage
= 0;
10791 for_each_present_cpu(i
)
10792 totalcpuusage
+= cpuacct_cpuusage_read(ca
, i
);
10794 return totalcpuusage
;
10797 static int cpuusage_write(struct cgroup
*cgrp
, struct cftype
*cftype
,
10800 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10809 for_each_present_cpu(i
)
10810 cpuacct_cpuusage_write(ca
, i
, 0);
10816 static int cpuacct_percpu_seq_read(struct cgroup
*cgroup
, struct cftype
*cft
,
10817 struct seq_file
*m
)
10819 struct cpuacct
*ca
= cgroup_ca(cgroup
);
10823 for_each_present_cpu(i
) {
10824 percpu
= cpuacct_cpuusage_read(ca
, i
);
10825 seq_printf(m
, "%llu ", (unsigned long long) percpu
);
10827 seq_printf(m
, "\n");
10831 static const char *cpuacct_stat_desc
[] = {
10832 [CPUACCT_STAT_USER
] = "user",
10833 [CPUACCT_STAT_SYSTEM
] = "system",
10836 static int cpuacct_stats_show(struct cgroup
*cgrp
, struct cftype
*cft
,
10837 struct cgroup_map_cb
*cb
)
10839 struct cpuacct
*ca
= cgroup_ca(cgrp
);
10842 for (i
= 0; i
< CPUACCT_STAT_NSTATS
; i
++) {
10843 s64 val
= percpu_counter_read(&ca
->cpustat
[i
]);
10844 val
= cputime64_to_clock_t(val
);
10845 cb
->fill(cb
, cpuacct_stat_desc
[i
], val
);
10850 static struct cftype files
[] = {
10853 .read_u64
= cpuusage_read
,
10854 .write_u64
= cpuusage_write
,
10857 .name
= "usage_percpu",
10858 .read_seq_string
= cpuacct_percpu_seq_read
,
10862 .read_map
= cpuacct_stats_show
,
10866 static int cpuacct_populate(struct cgroup_subsys
*ss
, struct cgroup
*cgrp
)
10868 return cgroup_add_files(cgrp
, ss
, files
, ARRAY_SIZE(files
));
10872 * charge this task's execution time to its accounting group.
10874 * called with rq->lock held.
10876 static void cpuacct_charge(struct task_struct
*tsk
, u64 cputime
)
10878 struct cpuacct
*ca
;
10881 if (unlikely(!cpuacct_subsys
.active
))
10884 cpu
= task_cpu(tsk
);
10890 for (; ca
; ca
= ca
->parent
) {
10891 u64
*cpuusage
= per_cpu_ptr(ca
->cpuusage
, cpu
);
10892 *cpuusage
+= cputime
;
10899 * Charge the system/user time to the task's accounting group.
10901 static void cpuacct_update_stats(struct task_struct
*tsk
,
10902 enum cpuacct_stat_index idx
, cputime_t val
)
10904 struct cpuacct
*ca
;
10906 if (unlikely(!cpuacct_subsys
.active
))
10913 percpu_counter_add(&ca
->cpustat
[idx
], val
);
10919 struct cgroup_subsys cpuacct_subsys
= {
10921 .create
= cpuacct_create
,
10922 .destroy
= cpuacct_destroy
,
10923 .populate
= cpuacct_populate
,
10924 .subsys_id
= cpuacct_subsys_id
,
10926 #endif /* CONFIG_CGROUP_CPUACCT */
10930 int rcu_expedited_torture_stats(char *page
)
10934 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10936 void synchronize_sched_expedited(void)
10939 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
10941 #else /* #ifndef CONFIG_SMP */
10943 static DEFINE_PER_CPU(struct migration_req
, rcu_migration_req
);
10944 static DEFINE_MUTEX(rcu_sched_expedited_mutex
);
10946 #define RCU_EXPEDITED_STATE_POST -2
10947 #define RCU_EXPEDITED_STATE_IDLE -1
10949 static int rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
10951 int rcu_expedited_torture_stats(char *page
)
10956 cnt
+= sprintf(&page
[cnt
], "state: %d /", rcu_expedited_state
);
10957 for_each_online_cpu(cpu
) {
10958 cnt
+= sprintf(&page
[cnt
], " %d:%d",
10959 cpu
, per_cpu(rcu_migration_req
, cpu
).dest_cpu
);
10961 cnt
+= sprintf(&page
[cnt
], "\n");
10964 EXPORT_SYMBOL_GPL(rcu_expedited_torture_stats
);
10966 static long synchronize_sched_expedited_count
;
10969 * Wait for an rcu-sched grace period to elapse, but use "big hammer"
10970 * approach to force grace period to end quickly. This consumes
10971 * significant time on all CPUs, and is thus not recommended for
10972 * any sort of common-case code.
10974 * Note that it is illegal to call this function while holding any
10975 * lock that is acquired by a CPU-hotplug notifier. Failing to
10976 * observe this restriction will result in deadlock.
10978 void synchronize_sched_expedited(void)
10981 unsigned long flags
;
10982 bool need_full_sync
= 0;
10984 struct migration_req
*req
;
10988 smp_mb(); /* ensure prior mod happens before capturing snap. */
10989 snap
= ACCESS_ONCE(synchronize_sched_expedited_count
) + 1;
10991 while (!mutex_trylock(&rcu_sched_expedited_mutex
)) {
10993 if (trycount
++ < 10)
10994 udelay(trycount
* num_online_cpus());
10996 synchronize_sched();
10999 if (ACCESS_ONCE(synchronize_sched_expedited_count
) - snap
> 0) {
11000 smp_mb(); /* ensure test happens before caller kfree */
11005 rcu_expedited_state
= RCU_EXPEDITED_STATE_POST
;
11006 for_each_online_cpu(cpu
) {
11008 req
= &per_cpu(rcu_migration_req
, cpu
);
11009 init_completion(&req
->done
);
11011 req
->dest_cpu
= RCU_MIGRATION_NEED_QS
;
11012 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11013 list_add(&req
->list
, &rq
->migration_queue
);
11014 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11015 wake_up_process(rq
->migration_thread
);
11017 for_each_online_cpu(cpu
) {
11018 rcu_expedited_state
= cpu
;
11019 req
= &per_cpu(rcu_migration_req
, cpu
);
11021 wait_for_completion(&req
->done
);
11022 raw_spin_lock_irqsave(&rq
->lock
, flags
);
11023 if (unlikely(req
->dest_cpu
== RCU_MIGRATION_MUST_SYNC
))
11024 need_full_sync
= 1;
11025 req
->dest_cpu
= RCU_MIGRATION_IDLE
;
11026 raw_spin_unlock_irqrestore(&rq
->lock
, flags
);
11028 rcu_expedited_state
= RCU_EXPEDITED_STATE_IDLE
;
11029 synchronize_sched_expedited_count
++;
11030 mutex_unlock(&rcu_sched_expedited_mutex
);
11032 if (need_full_sync
)
11033 synchronize_sched();
11035 EXPORT_SYMBOL_GPL(synchronize_sched_expedited
);
11037 #endif /* #else #ifndef CONFIG_SMP */