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
22 #include <linux/module.h>
23 #include <linux/nmi.h>
24 #include <linux/init.h>
25 #include <asm/uaccess.h>
26 #include <linux/highmem.h>
27 #include <linux/smp_lock.h>
28 #include <asm/mmu_context.h>
29 #include <linux/interrupt.h>
30 #include <linux/capability.h>
31 #include <linux/completion.h>
32 #include <linux/kernel_stat.h>
33 #include <linux/debug_locks.h>
34 #include <linux/security.h>
35 #include <linux/notifier.h>
36 #include <linux/profile.h>
37 #include <linux/freezer.h>
38 #include <linux/vmalloc.h>
39 #include <linux/blkdev.h>
40 #include <linux/delay.h>
41 #include <linux/smp.h>
42 #include <linux/threads.h>
43 #include <linux/timer.h>
44 #include <linux/rcupdate.h>
45 #include <linux/cpu.h>
46 #include <linux/cpuset.h>
47 #include <linux/percpu.h>
48 #include <linux/kthread.h>
49 #include <linux/seq_file.h>
50 #include <linux/syscalls.h>
51 #include <linux/times.h>
52 #include <linux/tsacct_kern.h>
53 #include <linux/kprobes.h>
54 #include <linux/delayacct.h>
57 #include <asm/unistd.h>
60 * Scheduler clock - returns current time in nanosec units.
61 * This is default implementation.
62 * Architectures and sub-architectures can override this.
64 unsigned long long __attribute__((weak
)) sched_clock(void)
66 return (unsigned long long)jiffies
* (1000000000 / HZ
);
70 * Convert user-nice values [ -20 ... 0 ... 19 ]
71 * to static priority [ MAX_RT_PRIO..MAX_PRIO-1 ],
74 #define NICE_TO_PRIO(nice) (MAX_RT_PRIO + (nice) + 20)
75 #define PRIO_TO_NICE(prio) ((prio) - MAX_RT_PRIO - 20)
76 #define TASK_NICE(p) PRIO_TO_NICE((p)->static_prio)
79 * 'User priority' is the nice value converted to something we
80 * can work with better when scaling various scheduler parameters,
81 * it's a [ 0 ... 39 ] range.
83 #define USER_PRIO(p) ((p)-MAX_RT_PRIO)
84 #define TASK_USER_PRIO(p) USER_PRIO((p)->static_prio)
85 #define MAX_USER_PRIO (USER_PRIO(MAX_PRIO))
88 * Some helpers for converting nanosecond timing to jiffy resolution
90 #define NS_TO_JIFFIES(TIME) ((TIME) / (1000000000 / HZ))
91 #define JIFFIES_TO_NS(TIME) ((TIME) * (1000000000 / HZ))
94 * These are the 'tuning knobs' of the scheduler:
96 * Minimum timeslice is 5 msecs (or 1 jiffy, whichever is larger),
97 * default timeslice is 100 msecs, maximum timeslice is 800 msecs.
98 * Timeslices get refilled after they expire.
100 #define MIN_TIMESLICE max(5 * HZ / 1000, 1)
101 #define DEF_TIMESLICE (100 * HZ / 1000)
102 #define ON_RUNQUEUE_WEIGHT 30
103 #define CHILD_PENALTY 95
104 #define PARENT_PENALTY 100
105 #define EXIT_WEIGHT 3
106 #define PRIO_BONUS_RATIO 25
107 #define MAX_BONUS (MAX_USER_PRIO * PRIO_BONUS_RATIO / 100)
108 #define INTERACTIVE_DELTA 2
109 #define MAX_SLEEP_AVG (DEF_TIMESLICE * MAX_BONUS)
110 #define STARVATION_LIMIT (MAX_SLEEP_AVG)
111 #define NS_MAX_SLEEP_AVG (JIFFIES_TO_NS(MAX_SLEEP_AVG))
114 * If a task is 'interactive' then we reinsert it in the active
115 * array after it has expired its current timeslice. (it will not
116 * continue to run immediately, it will still roundrobin with
117 * other interactive tasks.)
119 * This part scales the interactivity limit depending on niceness.
121 * We scale it linearly, offset by the INTERACTIVE_DELTA delta.
122 * Here are a few examples of different nice levels:
124 * TASK_INTERACTIVE(-20): [1,1,1,1,1,1,1,1,1,0,0]
125 * TASK_INTERACTIVE(-10): [1,1,1,1,1,1,1,0,0,0,0]
126 * TASK_INTERACTIVE( 0): [1,1,1,1,0,0,0,0,0,0,0]
127 * TASK_INTERACTIVE( 10): [1,1,0,0,0,0,0,0,0,0,0]
128 * TASK_INTERACTIVE( 19): [0,0,0,0,0,0,0,0,0,0,0]
130 * (the X axis represents the possible -5 ... 0 ... +5 dynamic
131 * priority range a task can explore, a value of '1' means the
132 * task is rated interactive.)
134 * Ie. nice +19 tasks can never get 'interactive' enough to be
135 * reinserted into the active array. And only heavily CPU-hog nice -20
136 * tasks will be expired. Default nice 0 tasks are somewhere between,
137 * it takes some effort for them to get interactive, but it's not
141 #define CURRENT_BONUS(p) \
142 (NS_TO_JIFFIES((p)->sleep_avg) * MAX_BONUS / \
145 #define GRANULARITY (10 * HZ / 1000 ? : 1)
148 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
149 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)) * \
152 #define TIMESLICE_GRANULARITY(p) (GRANULARITY * \
153 (1 << (((MAX_BONUS - CURRENT_BONUS(p)) ? : 1) - 1)))
156 #define SCALE(v1,v1_max,v2_max) \
157 (v1) * (v2_max) / (v1_max)
160 (SCALE(TASK_NICE(p) + 20, 40, MAX_BONUS) - 20 * MAX_BONUS / 40 + \
163 #define TASK_INTERACTIVE(p) \
164 ((p)->prio <= (p)->static_prio - DELTA(p))
166 #define INTERACTIVE_SLEEP(p) \
167 (JIFFIES_TO_NS(MAX_SLEEP_AVG * \
168 (MAX_BONUS / 2 + DELTA((p)) + 1) / MAX_BONUS - 1))
170 #define TASK_PREEMPTS_CURR(p, rq) \
171 ((p)->prio < (rq)->curr->prio)
173 #define SCALE_PRIO(x, prio) \
174 max(x * (MAX_PRIO - prio) / (MAX_USER_PRIO / 2), MIN_TIMESLICE)
176 static unsigned int static_prio_timeslice(int static_prio
)
178 if (static_prio
< NICE_TO_PRIO(0))
179 return SCALE_PRIO(DEF_TIMESLICE
* 4, static_prio
);
181 return SCALE_PRIO(DEF_TIMESLICE
, static_prio
);
185 * task_timeslice() scales user-nice values [ -20 ... 0 ... 19 ]
186 * to time slice values: [800ms ... 100ms ... 5ms]
188 * The higher a thread's priority, the bigger timeslices
189 * it gets during one round of execution. But even the lowest
190 * priority thread gets MIN_TIMESLICE worth of execution time.
193 static inline unsigned int task_timeslice(struct task_struct
*p
)
195 return static_prio_timeslice(p
->static_prio
);
199 * These are the runqueue data structures:
203 unsigned int nr_active
;
204 DECLARE_BITMAP(bitmap
, MAX_PRIO
+1); /* include 1 bit for delimiter */
205 struct list_head queue
[MAX_PRIO
];
209 * This is the main, per-CPU runqueue data structure.
211 * Locking rule: those places that want to lock multiple runqueues
212 * (such as the load balancing or the thread migration code), lock
213 * acquire operations must be ordered by ascending &runqueue.
219 * nr_running and cpu_load should be in the same cacheline because
220 * remote CPUs use both these fields when doing load calculation.
222 unsigned long nr_running
;
223 unsigned long raw_weighted_load
;
225 unsigned long cpu_load
[3];
227 unsigned long long nr_switches
;
230 * This is part of a global counter where only the total sum
231 * over all CPUs matters. A task can increase this counter on
232 * one CPU and if it got migrated afterwards it may decrease
233 * it on another CPU. Always updated under the runqueue lock:
235 unsigned long nr_uninterruptible
;
237 unsigned long expired_timestamp
;
238 /* Cached timestamp set by update_cpu_clock() */
239 unsigned long long most_recent_timestamp
;
240 struct task_struct
*curr
, *idle
;
241 unsigned long next_balance
;
242 struct mm_struct
*prev_mm
;
243 struct prio_array
*active
, *expired
, arrays
[2];
244 int best_expired_prio
;
248 struct sched_domain
*sd
;
250 /* For active balancing */
253 int cpu
; /* cpu of this runqueue */
255 struct task_struct
*migration_thread
;
256 struct list_head migration_queue
;
259 #ifdef CONFIG_SCHEDSTATS
261 struct sched_info rq_sched_info
;
263 /* sys_sched_yield() stats */
264 unsigned long yld_exp_empty
;
265 unsigned long yld_act_empty
;
266 unsigned long yld_both_empty
;
267 unsigned long yld_cnt
;
269 /* schedule() stats */
270 unsigned long sched_switch
;
271 unsigned long sched_cnt
;
272 unsigned long sched_goidle
;
274 /* try_to_wake_up() stats */
275 unsigned long ttwu_cnt
;
276 unsigned long ttwu_local
;
278 struct lock_class_key rq_lock_key
;
281 static DEFINE_PER_CPU(struct rq
, runqueues
);
283 static inline int cpu_of(struct rq
*rq
)
293 * The domain tree (rq->sd) is protected by RCU's quiescent state transition.
294 * See detach_destroy_domains: synchronize_sched for details.
296 * The domain tree of any CPU may only be accessed from within
297 * preempt-disabled sections.
299 #define for_each_domain(cpu, __sd) \
300 for (__sd = rcu_dereference(cpu_rq(cpu)->sd); __sd; __sd = __sd->parent)
302 #define cpu_rq(cpu) (&per_cpu(runqueues, (cpu)))
303 #define this_rq() (&__get_cpu_var(runqueues))
304 #define task_rq(p) cpu_rq(task_cpu(p))
305 #define cpu_curr(cpu) (cpu_rq(cpu)->curr)
307 #ifndef prepare_arch_switch
308 # define prepare_arch_switch(next) do { } while (0)
310 #ifndef finish_arch_switch
311 # define finish_arch_switch(prev) do { } while (0)
314 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
315 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
317 return rq
->curr
== p
;
320 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
324 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
326 #ifdef CONFIG_DEBUG_SPINLOCK
327 /* this is a valid case when another task releases the spinlock */
328 rq
->lock
.owner
= current
;
331 * If we are tracking spinlock dependencies then we have to
332 * fix up the runqueue lock - which gets 'carried over' from
335 spin_acquire(&rq
->lock
.dep_map
, 0, 0, _THIS_IP_
);
337 spin_unlock_irq(&rq
->lock
);
340 #else /* __ARCH_WANT_UNLOCKED_CTXSW */
341 static inline int task_running(struct rq
*rq
, struct task_struct
*p
)
346 return rq
->curr
== p
;
350 static inline void prepare_lock_switch(struct rq
*rq
, struct task_struct
*next
)
354 * We can optimise this out completely for !SMP, because the
355 * SMP rebalancing from interrupt is the only thing that cares
360 #ifdef __ARCH_WANT_INTERRUPTS_ON_CTXSW
361 spin_unlock_irq(&rq
->lock
);
363 spin_unlock(&rq
->lock
);
367 static inline void finish_lock_switch(struct rq
*rq
, struct task_struct
*prev
)
371 * After ->oncpu is cleared, the task can be moved to a different CPU.
372 * We must ensure this doesn't happen until the switch is completely
378 #ifndef __ARCH_WANT_INTERRUPTS_ON_CTXSW
382 #endif /* __ARCH_WANT_UNLOCKED_CTXSW */
385 * __task_rq_lock - lock the runqueue a given task resides on.
386 * Must be called interrupts disabled.
388 static inline struct rq
*__task_rq_lock(struct task_struct
*p
)
395 spin_lock(&rq
->lock
);
396 if (unlikely(rq
!= task_rq(p
))) {
397 spin_unlock(&rq
->lock
);
398 goto repeat_lock_task
;
404 * task_rq_lock - lock the runqueue a given task resides on and disable
405 * interrupts. Note the ordering: we can safely lookup the task_rq without
406 * explicitly disabling preemption.
408 static struct rq
*task_rq_lock(struct task_struct
*p
, unsigned long *flags
)
414 local_irq_save(*flags
);
416 spin_lock(&rq
->lock
);
417 if (unlikely(rq
!= task_rq(p
))) {
418 spin_unlock_irqrestore(&rq
->lock
, *flags
);
419 goto repeat_lock_task
;
424 static inline void __task_rq_unlock(struct rq
*rq
)
427 spin_unlock(&rq
->lock
);
430 static inline void task_rq_unlock(struct rq
*rq
, unsigned long *flags
)
433 spin_unlock_irqrestore(&rq
->lock
, *flags
);
436 #ifdef CONFIG_SCHEDSTATS
438 * bump this up when changing the output format or the meaning of an existing
439 * format, so that tools can adapt (or abort)
441 #define SCHEDSTAT_VERSION 14
443 static int show_schedstat(struct seq_file
*seq
, void *v
)
447 seq_printf(seq
, "version %d\n", SCHEDSTAT_VERSION
);
448 seq_printf(seq
, "timestamp %lu\n", jiffies
);
449 for_each_online_cpu(cpu
) {
450 struct rq
*rq
= cpu_rq(cpu
);
452 struct sched_domain
*sd
;
456 /* runqueue-specific stats */
458 "cpu%d %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu %lu",
459 cpu
, rq
->yld_both_empty
,
460 rq
->yld_act_empty
, rq
->yld_exp_empty
, rq
->yld_cnt
,
461 rq
->sched_switch
, rq
->sched_cnt
, rq
->sched_goidle
,
462 rq
->ttwu_cnt
, rq
->ttwu_local
,
463 rq
->rq_sched_info
.cpu_time
,
464 rq
->rq_sched_info
.run_delay
, rq
->rq_sched_info
.pcnt
);
466 seq_printf(seq
, "\n");
469 /* domain-specific stats */
471 for_each_domain(cpu
, sd
) {
472 enum idle_type itype
;
473 char mask_str
[NR_CPUS
];
475 cpumask_scnprintf(mask_str
, NR_CPUS
, sd
->span
);
476 seq_printf(seq
, "domain%d %s", dcnt
++, mask_str
);
477 for (itype
= SCHED_IDLE
; itype
< MAX_IDLE_TYPES
;
479 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu "
482 sd
->lb_balanced
[itype
],
483 sd
->lb_failed
[itype
],
484 sd
->lb_imbalance
[itype
],
485 sd
->lb_gained
[itype
],
486 sd
->lb_hot_gained
[itype
],
487 sd
->lb_nobusyq
[itype
],
488 sd
->lb_nobusyg
[itype
]);
490 seq_printf(seq
, " %lu %lu %lu %lu %lu %lu %lu %lu %lu"
492 sd
->alb_cnt
, sd
->alb_failed
, sd
->alb_pushed
,
493 sd
->sbe_cnt
, sd
->sbe_balanced
, sd
->sbe_pushed
,
494 sd
->sbf_cnt
, sd
->sbf_balanced
, sd
->sbf_pushed
,
495 sd
->ttwu_wake_remote
, sd
->ttwu_move_affine
,
496 sd
->ttwu_move_balance
);
504 static int schedstat_open(struct inode
*inode
, struct file
*file
)
506 unsigned int size
= PAGE_SIZE
* (1 + num_online_cpus() / 32);
507 char *buf
= kmalloc(size
, GFP_KERNEL
);
513 res
= single_open(file
, show_schedstat
, NULL
);
515 m
= file
->private_data
;
523 const struct file_operations proc_schedstat_operations
= {
524 .open
= schedstat_open
,
527 .release
= single_release
,
531 * Expects runqueue lock to be held for atomicity of update
534 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
537 rq
->rq_sched_info
.run_delay
+= delta_jiffies
;
538 rq
->rq_sched_info
.pcnt
++;
543 * Expects runqueue lock to be held for atomicity of update
546 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
549 rq
->rq_sched_info
.cpu_time
+= delta_jiffies
;
551 # define schedstat_inc(rq, field) do { (rq)->field++; } while (0)
552 # define schedstat_add(rq, field, amt) do { (rq)->field += (amt); } while (0)
553 #else /* !CONFIG_SCHEDSTATS */
555 rq_sched_info_arrive(struct rq
*rq
, unsigned long delta_jiffies
)
558 rq_sched_info_depart(struct rq
*rq
, unsigned long delta_jiffies
)
560 # define schedstat_inc(rq, field) do { } while (0)
561 # define schedstat_add(rq, field, amt) do { } while (0)
565 * this_rq_lock - lock this runqueue and disable interrupts.
567 static inline struct rq
*this_rq_lock(void)
574 spin_lock(&rq
->lock
);
579 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
581 * Called when a process is dequeued from the active array and given
582 * the cpu. We should note that with the exception of interactive
583 * tasks, the expired queue will become the active queue after the active
584 * queue is empty, without explicitly dequeuing and requeuing tasks in the
585 * expired queue. (Interactive tasks may be requeued directly to the
586 * active queue, thus delaying tasks in the expired queue from running;
587 * see scheduler_tick()).
589 * This function is only called from sched_info_arrive(), rather than
590 * dequeue_task(). Even though a task may be queued and dequeued multiple
591 * times as it is shuffled about, we're really interested in knowing how
592 * long it was from the *first* time it was queued to the time that it
595 static inline void sched_info_dequeued(struct task_struct
*t
)
597 t
->sched_info
.last_queued
= 0;
601 * Called when a task finally hits the cpu. We can now calculate how
602 * long it was waiting to run. We also note when it began so that we
603 * can keep stats on how long its timeslice is.
605 static void sched_info_arrive(struct task_struct
*t
)
607 unsigned long now
= jiffies
, delta_jiffies
= 0;
609 if (t
->sched_info
.last_queued
)
610 delta_jiffies
= now
- t
->sched_info
.last_queued
;
611 sched_info_dequeued(t
);
612 t
->sched_info
.run_delay
+= delta_jiffies
;
613 t
->sched_info
.last_arrival
= now
;
614 t
->sched_info
.pcnt
++;
616 rq_sched_info_arrive(task_rq(t
), delta_jiffies
);
620 * Called when a process is queued into either the active or expired
621 * array. The time is noted and later used to determine how long we
622 * had to wait for us to reach the cpu. Since the expired queue will
623 * become the active queue after active queue is empty, without dequeuing
624 * and requeuing any tasks, we are interested in queuing to either. It
625 * is unusual but not impossible for tasks to be dequeued and immediately
626 * requeued in the same or another array: this can happen in sched_yield(),
627 * set_user_nice(), and even load_balance() as it moves tasks from runqueue
630 * This function is only called from enqueue_task(), but also only updates
631 * the timestamp if it is already not set. It's assumed that
632 * sched_info_dequeued() will clear that stamp when appropriate.
634 static inline void sched_info_queued(struct task_struct
*t
)
636 if (unlikely(sched_info_on()))
637 if (!t
->sched_info
.last_queued
)
638 t
->sched_info
.last_queued
= jiffies
;
642 * Called when a process ceases being the active-running process, either
643 * voluntarily or involuntarily. Now we can calculate how long we ran.
645 static inline void sched_info_depart(struct task_struct
*t
)
647 unsigned long delta_jiffies
= jiffies
- t
->sched_info
.last_arrival
;
649 t
->sched_info
.cpu_time
+= delta_jiffies
;
650 rq_sched_info_depart(task_rq(t
), delta_jiffies
);
654 * Called when tasks are switched involuntarily due, typically, to expiring
655 * their time slice. (This may also be called when switching to or from
656 * the idle task.) We are only called when prev != next.
659 __sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
661 struct rq
*rq
= task_rq(prev
);
664 * prev now departs the cpu. It's not interesting to record
665 * stats about how efficient we were at scheduling the idle
668 if (prev
!= rq
->idle
)
669 sched_info_depart(prev
);
671 if (next
!= rq
->idle
)
672 sched_info_arrive(next
);
675 sched_info_switch(struct task_struct
*prev
, struct task_struct
*next
)
677 if (unlikely(sched_info_on()))
678 __sched_info_switch(prev
, next
);
681 #define sched_info_queued(t) do { } while (0)
682 #define sched_info_switch(t, next) do { } while (0)
683 #endif /* CONFIG_SCHEDSTATS || CONFIG_TASK_DELAY_ACCT */
686 * Adding/removing a task to/from a priority array:
688 static void dequeue_task(struct task_struct
*p
, struct prio_array
*array
)
691 list_del(&p
->run_list
);
692 if (list_empty(array
->queue
+ p
->prio
))
693 __clear_bit(p
->prio
, array
->bitmap
);
696 static void enqueue_task(struct task_struct
*p
, struct prio_array
*array
)
698 sched_info_queued(p
);
699 list_add_tail(&p
->run_list
, array
->queue
+ p
->prio
);
700 __set_bit(p
->prio
, array
->bitmap
);
706 * Put task to the end of the run list without the overhead of dequeue
707 * followed by enqueue.
709 static void requeue_task(struct task_struct
*p
, struct prio_array
*array
)
711 list_move_tail(&p
->run_list
, array
->queue
+ p
->prio
);
715 enqueue_task_head(struct task_struct
*p
, struct prio_array
*array
)
717 list_add(&p
->run_list
, array
->queue
+ p
->prio
);
718 __set_bit(p
->prio
, array
->bitmap
);
724 * __normal_prio - return the priority that is based on the static
725 * priority but is modified by bonuses/penalties.
727 * We scale the actual sleep average [0 .... MAX_SLEEP_AVG]
728 * into the -5 ... 0 ... +5 bonus/penalty range.
730 * We use 25% of the full 0...39 priority range so that:
732 * 1) nice +19 interactive tasks do not preempt nice 0 CPU hogs.
733 * 2) nice -20 CPU hogs do not get preempted by nice 0 tasks.
735 * Both properties are important to certain workloads.
738 static inline int __normal_prio(struct task_struct
*p
)
742 bonus
= CURRENT_BONUS(p
) - MAX_BONUS
/ 2;
744 prio
= p
->static_prio
- bonus
;
745 if (prio
< MAX_RT_PRIO
)
747 if (prio
> MAX_PRIO
-1)
753 * To aid in avoiding the subversion of "niceness" due to uneven distribution
754 * of tasks with abnormal "nice" values across CPUs the contribution that
755 * each task makes to its run queue's load is weighted according to its
756 * scheduling class and "nice" value. For SCHED_NORMAL tasks this is just a
757 * scaled version of the new time slice allocation that they receive on time
762 * Assume: static_prio_timeslice(NICE_TO_PRIO(0)) == DEF_TIMESLICE
763 * If static_prio_timeslice() is ever changed to break this assumption then
764 * this code will need modification
766 #define TIME_SLICE_NICE_ZERO DEF_TIMESLICE
767 #define LOAD_WEIGHT(lp) \
768 (((lp) * SCHED_LOAD_SCALE) / TIME_SLICE_NICE_ZERO)
769 #define PRIO_TO_LOAD_WEIGHT(prio) \
770 LOAD_WEIGHT(static_prio_timeslice(prio))
771 #define RTPRIO_TO_LOAD_WEIGHT(rp) \
772 (PRIO_TO_LOAD_WEIGHT(MAX_RT_PRIO) + LOAD_WEIGHT(rp))
774 static void set_load_weight(struct task_struct
*p
)
776 if (has_rt_policy(p
)) {
778 if (p
== task_rq(p
)->migration_thread
)
780 * The migration thread does the actual balancing.
781 * Giving its load any weight will skew balancing
787 p
->load_weight
= RTPRIO_TO_LOAD_WEIGHT(p
->rt_priority
);
789 p
->load_weight
= PRIO_TO_LOAD_WEIGHT(p
->static_prio
);
793 inc_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
795 rq
->raw_weighted_load
+= p
->load_weight
;
799 dec_raw_weighted_load(struct rq
*rq
, const struct task_struct
*p
)
801 rq
->raw_weighted_load
-= p
->load_weight
;
804 static inline void inc_nr_running(struct task_struct
*p
, struct rq
*rq
)
807 inc_raw_weighted_load(rq
, p
);
810 static inline void dec_nr_running(struct task_struct
*p
, struct rq
*rq
)
813 dec_raw_weighted_load(rq
, p
);
817 * Calculate the expected normal priority: i.e. priority
818 * without taking RT-inheritance into account. Might be
819 * boosted by interactivity modifiers. Changes upon fork,
820 * setprio syscalls, and whenever the interactivity
821 * estimator recalculates.
823 static inline int normal_prio(struct task_struct
*p
)
827 if (has_rt_policy(p
))
828 prio
= MAX_RT_PRIO
-1 - p
->rt_priority
;
830 prio
= __normal_prio(p
);
835 * Calculate the current priority, i.e. the priority
836 * taken into account by the scheduler. This value might
837 * be boosted by RT tasks, or might be boosted by
838 * interactivity modifiers. Will be RT if the task got
839 * RT-boosted. If not then it returns p->normal_prio.
841 static int effective_prio(struct task_struct
*p
)
843 p
->normal_prio
= normal_prio(p
);
845 * If we are RT tasks or we were boosted to RT priority,
846 * keep the priority unchanged. Otherwise, update priority
847 * to the normal priority:
849 if (!rt_prio(p
->prio
))
850 return p
->normal_prio
;
855 * __activate_task - move a task to the runqueue.
857 static void __activate_task(struct task_struct
*p
, struct rq
*rq
)
859 struct prio_array
*target
= rq
->active
;
862 target
= rq
->expired
;
863 enqueue_task(p
, target
);
864 inc_nr_running(p
, rq
);
868 * __activate_idle_task - move idle task to the _front_ of runqueue.
870 static inline void __activate_idle_task(struct task_struct
*p
, struct rq
*rq
)
872 enqueue_task_head(p
, rq
->active
);
873 inc_nr_running(p
, rq
);
877 * Recalculate p->normal_prio and p->prio after having slept,
878 * updating the sleep-average too:
880 static int recalc_task_prio(struct task_struct
*p
, unsigned long long now
)
882 /* Caller must always ensure 'now >= p->timestamp' */
883 unsigned long sleep_time
= now
- p
->timestamp
;
888 if (likely(sleep_time
> 0)) {
890 * This ceiling is set to the lowest priority that would allow
891 * a task to be reinserted into the active array on timeslice
894 unsigned long ceiling
= INTERACTIVE_SLEEP(p
);
896 if (p
->mm
&& sleep_time
> ceiling
&& p
->sleep_avg
< ceiling
) {
898 * Prevents user tasks from achieving best priority
899 * with one single large enough sleep.
901 p
->sleep_avg
= ceiling
;
903 * Using INTERACTIVE_SLEEP() as a ceiling places a
904 * nice(0) task 1ms sleep away from promotion, and
905 * gives it 700ms to round-robin with no chance of
906 * being demoted. This is more than generous, so
907 * mark this sleep as non-interactive to prevent the
908 * on-runqueue bonus logic from intervening should
909 * this task not receive cpu immediately.
911 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
914 * Tasks waking from uninterruptible sleep are
915 * limited in their sleep_avg rise as they
916 * are likely to be waiting on I/O
918 if (p
->sleep_type
== SLEEP_NONINTERACTIVE
&& p
->mm
) {
919 if (p
->sleep_avg
>= ceiling
)
921 else if (p
->sleep_avg
+ sleep_time
>=
923 p
->sleep_avg
= ceiling
;
929 * This code gives a bonus to interactive tasks.
931 * The boost works by updating the 'average sleep time'
932 * value here, based on ->timestamp. The more time a
933 * task spends sleeping, the higher the average gets -
934 * and the higher the priority boost gets as well.
936 p
->sleep_avg
+= sleep_time
;
939 if (p
->sleep_avg
> NS_MAX_SLEEP_AVG
)
940 p
->sleep_avg
= NS_MAX_SLEEP_AVG
;
943 return effective_prio(p
);
947 * activate_task - move a task to the runqueue and do priority recalculation
949 * Update all the scheduling statistics stuff. (sleep average
950 * calculation, priority modifiers, etc.)
952 static void activate_task(struct task_struct
*p
, struct rq
*rq
, int local
)
954 unsigned long long now
;
962 /* Compensate for drifting sched_clock */
963 struct rq
*this_rq
= this_rq();
964 now
= (now
- this_rq
->most_recent_timestamp
)
965 + rq
->most_recent_timestamp
;
970 * Sleep time is in units of nanosecs, so shift by 20 to get a
971 * milliseconds-range estimation of the amount of time that the task
974 if (unlikely(prof_on
== SLEEP_PROFILING
)) {
975 if (p
->state
== TASK_UNINTERRUPTIBLE
)
976 profile_hits(SLEEP_PROFILING
, (void *)get_wchan(p
),
977 (now
- p
->timestamp
) >> 20);
980 p
->prio
= recalc_task_prio(p
, now
);
983 * This checks to make sure it's not an uninterruptible task
984 * that is now waking up.
986 if (p
->sleep_type
== SLEEP_NORMAL
) {
988 * Tasks which were woken up by interrupts (ie. hw events)
989 * are most likely of interactive nature. So we give them
990 * the credit of extending their sleep time to the period
991 * of time they spend on the runqueue, waiting for execution
992 * on a CPU, first time around:
995 p
->sleep_type
= SLEEP_INTERRUPTED
;
998 * Normal first-time wakeups get a credit too for
999 * on-runqueue time, but it will be weighted down:
1001 p
->sleep_type
= SLEEP_INTERACTIVE
;
1006 __activate_task(p
, rq
);
1010 * deactivate_task - remove a task from the runqueue.
1012 static void deactivate_task(struct task_struct
*p
, struct rq
*rq
)
1014 dec_nr_running(p
, rq
);
1015 dequeue_task(p
, p
->array
);
1020 * resched_task - mark a task 'to be rescheduled now'.
1022 * On UP this means the setting of the need_resched flag, on SMP it
1023 * might also involve a cross-CPU call to trigger the scheduler on
1028 #ifndef tsk_is_polling
1029 #define tsk_is_polling(t) test_tsk_thread_flag(t, TIF_POLLING_NRFLAG)
1032 static void resched_task(struct task_struct
*p
)
1036 assert_spin_locked(&task_rq(p
)->lock
);
1038 if (unlikely(test_tsk_thread_flag(p
, TIF_NEED_RESCHED
)))
1041 set_tsk_thread_flag(p
, TIF_NEED_RESCHED
);
1044 if (cpu
== smp_processor_id())
1047 /* NEED_RESCHED must be visible before we test polling */
1049 if (!tsk_is_polling(p
))
1050 smp_send_reschedule(cpu
);
1053 static inline void resched_task(struct task_struct
*p
)
1055 assert_spin_locked(&task_rq(p
)->lock
);
1056 set_tsk_need_resched(p
);
1061 * task_curr - is this task currently executing on a CPU?
1062 * @p: the task in question.
1064 inline int task_curr(const struct task_struct
*p
)
1066 return cpu_curr(task_cpu(p
)) == p
;
1069 /* Used instead of source_load when we know the type == 0 */
1070 unsigned long weighted_cpuload(const int cpu
)
1072 return cpu_rq(cpu
)->raw_weighted_load
;
1076 struct migration_req
{
1077 struct list_head list
;
1079 struct task_struct
*task
;
1082 struct completion done
;
1086 * The task's runqueue lock must be held.
1087 * Returns true if you have to wait for migration thread.
1090 migrate_task(struct task_struct
*p
, int dest_cpu
, struct migration_req
*req
)
1092 struct rq
*rq
= task_rq(p
);
1095 * If the task is not on a runqueue (and not running), then
1096 * it is sufficient to simply update the task's cpu field.
1098 if (!p
->array
&& !task_running(rq
, p
)) {
1099 set_task_cpu(p
, dest_cpu
);
1103 init_completion(&req
->done
);
1105 req
->dest_cpu
= dest_cpu
;
1106 list_add(&req
->list
, &rq
->migration_queue
);
1112 * wait_task_inactive - wait for a thread to unschedule.
1114 * The caller must ensure that the task *will* unschedule sometime soon,
1115 * else this function might spin for a *long* time. This function can't
1116 * be called with interrupts off, or it may introduce deadlock with
1117 * smp_call_function() if an IPI is sent by the same process we are
1118 * waiting to become inactive.
1120 void wait_task_inactive(struct task_struct
*p
)
1122 unsigned long flags
;
1127 rq
= task_rq_lock(p
, &flags
);
1128 /* Must be off runqueue entirely, not preempted. */
1129 if (unlikely(p
->array
|| task_running(rq
, p
))) {
1130 /* If it's preempted, we yield. It could be a while. */
1131 preempted
= !task_running(rq
, p
);
1132 task_rq_unlock(rq
, &flags
);
1138 task_rq_unlock(rq
, &flags
);
1142 * kick_process - kick a running thread to enter/exit the kernel
1143 * @p: the to-be-kicked thread
1145 * Cause a process which is running on another CPU to enter
1146 * kernel-mode, without any delay. (to get signals handled.)
1148 * NOTE: this function doesnt have to take the runqueue lock,
1149 * because all it wants to ensure is that the remote task enters
1150 * the kernel. If the IPI races and the task has been migrated
1151 * to another CPU then no harm is done and the purpose has been
1154 void kick_process(struct task_struct
*p
)
1160 if ((cpu
!= smp_processor_id()) && task_curr(p
))
1161 smp_send_reschedule(cpu
);
1166 * Return a low guess at the load of a migration-source cpu weighted
1167 * according to the scheduling class and "nice" value.
1169 * We want to under-estimate the load of migration sources, to
1170 * balance conservatively.
1172 static inline unsigned long source_load(int cpu
, int type
)
1174 struct rq
*rq
= cpu_rq(cpu
);
1177 return rq
->raw_weighted_load
;
1179 return min(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1183 * Return a high guess at the load of a migration-target cpu weighted
1184 * according to the scheduling class and "nice" value.
1186 static inline unsigned long target_load(int cpu
, int type
)
1188 struct rq
*rq
= cpu_rq(cpu
);
1191 return rq
->raw_weighted_load
;
1193 return max(rq
->cpu_load
[type
-1], rq
->raw_weighted_load
);
1197 * Return the average load per task on the cpu's run queue
1199 static inline unsigned long cpu_avg_load_per_task(int cpu
)
1201 struct rq
*rq
= cpu_rq(cpu
);
1202 unsigned long n
= rq
->nr_running
;
1204 return n
? rq
->raw_weighted_load
/ n
: SCHED_LOAD_SCALE
;
1208 * find_idlest_group finds and returns the least busy CPU group within the
1211 static struct sched_group
*
1212 find_idlest_group(struct sched_domain
*sd
, struct task_struct
*p
, int this_cpu
)
1214 struct sched_group
*idlest
= NULL
, *this = NULL
, *group
= sd
->groups
;
1215 unsigned long min_load
= ULONG_MAX
, this_load
= 0;
1216 int load_idx
= sd
->forkexec_idx
;
1217 int imbalance
= 100 + (sd
->imbalance_pct
-100)/2;
1220 unsigned long load
, avg_load
;
1224 /* Skip over this group if it has no CPUs allowed */
1225 if (!cpus_intersects(group
->cpumask
, p
->cpus_allowed
))
1228 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
1230 /* Tally up the load of all CPUs in the group */
1233 for_each_cpu_mask(i
, group
->cpumask
) {
1234 /* Bias balancing toward cpus of our domain */
1236 load
= source_load(i
, load_idx
);
1238 load
= target_load(i
, load_idx
);
1243 /* Adjust by relative CPU power of the group */
1244 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
1247 this_load
= avg_load
;
1249 } else if (avg_load
< min_load
) {
1250 min_load
= avg_load
;
1254 group
= group
->next
;
1255 } while (group
!= sd
->groups
);
1257 if (!idlest
|| 100*this_load
< imbalance
*min_load
)
1263 * find_idlest_cpu - find the idlest cpu among the cpus in group.
1266 find_idlest_cpu(struct sched_group
*group
, struct task_struct
*p
, int this_cpu
)
1269 unsigned long load
, min_load
= ULONG_MAX
;
1273 /* Traverse only the allowed CPUs */
1274 cpus_and(tmp
, group
->cpumask
, p
->cpus_allowed
);
1276 for_each_cpu_mask(i
, tmp
) {
1277 load
= weighted_cpuload(i
);
1279 if (load
< min_load
|| (load
== min_load
&& i
== this_cpu
)) {
1289 * sched_balance_self: balance the current task (running on cpu) in domains
1290 * that have the 'flag' flag set. In practice, this is SD_BALANCE_FORK and
1293 * Balance, ie. select the least loaded group.
1295 * Returns the target CPU number, or the same CPU if no balancing is needed.
1297 * preempt must be disabled.
1299 static int sched_balance_self(int cpu
, int flag
)
1301 struct task_struct
*t
= current
;
1302 struct sched_domain
*tmp
, *sd
= NULL
;
1304 for_each_domain(cpu
, tmp
) {
1306 * If power savings logic is enabled for a domain, stop there.
1308 if (tmp
->flags
& SD_POWERSAVINGS_BALANCE
)
1310 if (tmp
->flags
& flag
)
1316 struct sched_group
*group
;
1317 int new_cpu
, weight
;
1319 if (!(sd
->flags
& flag
)) {
1325 group
= find_idlest_group(sd
, t
, cpu
);
1331 new_cpu
= find_idlest_cpu(group
, t
, cpu
);
1332 if (new_cpu
== -1 || new_cpu
== cpu
) {
1333 /* Now try balancing at a lower domain level of cpu */
1338 /* Now try balancing at a lower domain level of new_cpu */
1341 weight
= cpus_weight(span
);
1342 for_each_domain(cpu
, tmp
) {
1343 if (weight
<= cpus_weight(tmp
->span
))
1345 if (tmp
->flags
& flag
)
1348 /* while loop will break here if sd == NULL */
1354 #endif /* CONFIG_SMP */
1357 * wake_idle() will wake a task on an idle cpu if task->cpu is
1358 * not idle and an idle cpu is available. The span of cpus to
1359 * search starts with cpus closest then further out as needed,
1360 * so we always favor a closer, idle cpu.
1362 * Returns the CPU we should wake onto.
1364 #if defined(ARCH_HAS_SCHED_WAKE_IDLE)
1365 static int wake_idle(int cpu
, struct task_struct
*p
)
1368 struct sched_domain
*sd
;
1374 for_each_domain(cpu
, sd
) {
1375 if (sd
->flags
& SD_WAKE_IDLE
) {
1376 cpus_and(tmp
, sd
->span
, p
->cpus_allowed
);
1377 for_each_cpu_mask(i
, tmp
) {
1388 static inline int wake_idle(int cpu
, struct task_struct
*p
)
1395 * try_to_wake_up - wake up a thread
1396 * @p: the to-be-woken-up thread
1397 * @state: the mask of task states that can be woken
1398 * @sync: do a synchronous wakeup?
1400 * Put it on the run-queue if it's not already there. The "current"
1401 * thread is always on the run-queue (except when the actual
1402 * re-schedule is in progress), and as such you're allowed to do
1403 * the simpler "current->state = TASK_RUNNING" to mark yourself
1404 * runnable without the overhead of this.
1406 * returns failure only if the task is already active.
1408 static int try_to_wake_up(struct task_struct
*p
, unsigned int state
, int sync
)
1410 int cpu
, this_cpu
, success
= 0;
1411 unsigned long flags
;
1415 struct sched_domain
*sd
, *this_sd
= NULL
;
1416 unsigned long load
, this_load
;
1420 rq
= task_rq_lock(p
, &flags
);
1421 old_state
= p
->state
;
1422 if (!(old_state
& state
))
1429 this_cpu
= smp_processor_id();
1432 if (unlikely(task_running(rq
, p
)))
1437 schedstat_inc(rq
, ttwu_cnt
);
1438 if (cpu
== this_cpu
) {
1439 schedstat_inc(rq
, ttwu_local
);
1443 for_each_domain(this_cpu
, sd
) {
1444 if (cpu_isset(cpu
, sd
->span
)) {
1445 schedstat_inc(sd
, ttwu_wake_remote
);
1451 if (unlikely(!cpu_isset(this_cpu
, p
->cpus_allowed
)))
1455 * Check for affine wakeup and passive balancing possibilities.
1458 int idx
= this_sd
->wake_idx
;
1459 unsigned int imbalance
;
1461 imbalance
= 100 + (this_sd
->imbalance_pct
- 100) / 2;
1463 load
= source_load(cpu
, idx
);
1464 this_load
= target_load(this_cpu
, idx
);
1466 new_cpu
= this_cpu
; /* Wake to this CPU if we can */
1468 if (this_sd
->flags
& SD_WAKE_AFFINE
) {
1469 unsigned long tl
= this_load
;
1470 unsigned long tl_per_task
;
1472 tl_per_task
= cpu_avg_load_per_task(this_cpu
);
1475 * If sync wakeup then subtract the (maximum possible)
1476 * effect of the currently running task from the load
1477 * of the current CPU:
1480 tl
-= current
->load_weight
;
1483 tl
+ target_load(cpu
, idx
) <= tl_per_task
) ||
1484 100*(tl
+ p
->load_weight
) <= imbalance
*load
) {
1486 * This domain has SD_WAKE_AFFINE and
1487 * p is cache cold in this domain, and
1488 * there is no bad imbalance.
1490 schedstat_inc(this_sd
, ttwu_move_affine
);
1496 * Start passive balancing when half the imbalance_pct
1499 if (this_sd
->flags
& SD_WAKE_BALANCE
) {
1500 if (imbalance
*this_load
<= 100*load
) {
1501 schedstat_inc(this_sd
, ttwu_move_balance
);
1507 new_cpu
= cpu
; /* Could not wake to this_cpu. Wake to cpu instead */
1509 new_cpu
= wake_idle(new_cpu
, p
);
1510 if (new_cpu
!= cpu
) {
1511 set_task_cpu(p
, new_cpu
);
1512 task_rq_unlock(rq
, &flags
);
1513 /* might preempt at this point */
1514 rq
= task_rq_lock(p
, &flags
);
1515 old_state
= p
->state
;
1516 if (!(old_state
& state
))
1521 this_cpu
= smp_processor_id();
1526 #endif /* CONFIG_SMP */
1527 if (old_state
== TASK_UNINTERRUPTIBLE
) {
1528 rq
->nr_uninterruptible
--;
1530 * Tasks on involuntary sleep don't earn
1531 * sleep_avg beyond just interactive state.
1533 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1537 * Tasks that have marked their sleep as noninteractive get
1538 * woken up with their sleep average not weighted in an
1541 if (old_state
& TASK_NONINTERACTIVE
)
1542 p
->sleep_type
= SLEEP_NONINTERACTIVE
;
1545 activate_task(p
, rq
, cpu
== this_cpu
);
1547 * Sync wakeups (i.e. those types of wakeups where the waker
1548 * has indicated that it will leave the CPU in short order)
1549 * don't trigger a preemption, if the woken up task will run on
1550 * this cpu. (in this case the 'I will reschedule' promise of
1551 * the waker guarantees that the freshly woken up task is going
1552 * to be considered on this CPU.)
1554 if (!sync
|| cpu
!= this_cpu
) {
1555 if (TASK_PREEMPTS_CURR(p
, rq
))
1556 resched_task(rq
->curr
);
1561 p
->state
= TASK_RUNNING
;
1563 task_rq_unlock(rq
, &flags
);
1568 int fastcall
wake_up_process(struct task_struct
*p
)
1570 return try_to_wake_up(p
, TASK_STOPPED
| TASK_TRACED
|
1571 TASK_INTERRUPTIBLE
| TASK_UNINTERRUPTIBLE
, 0);
1573 EXPORT_SYMBOL(wake_up_process
);
1575 int fastcall
wake_up_state(struct task_struct
*p
, unsigned int state
)
1577 return try_to_wake_up(p
, state
, 0);
1580 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
);
1582 * Perform scheduler related setup for a newly forked process p.
1583 * p is forked by current.
1585 void fastcall
sched_fork(struct task_struct
*p
, int clone_flags
)
1587 int cpu
= get_cpu();
1590 cpu
= sched_balance_self(cpu
, SD_BALANCE_FORK
);
1592 set_task_cpu(p
, cpu
);
1595 * We mark the process as running here, but have not actually
1596 * inserted it onto the runqueue yet. This guarantees that
1597 * nobody will actually run it, and a signal or other external
1598 * event cannot wake it up and insert it on the runqueue either.
1600 p
->state
= TASK_RUNNING
;
1603 * Make sure we do not leak PI boosting priority to the child:
1605 p
->prio
= current
->normal_prio
;
1607 INIT_LIST_HEAD(&p
->run_list
);
1609 #if defined(CONFIG_SCHEDSTATS) || defined(CONFIG_TASK_DELAY_ACCT)
1610 if (unlikely(sched_info_on()))
1611 memset(&p
->sched_info
, 0, sizeof(p
->sched_info
));
1613 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
1616 #ifdef CONFIG_PREEMPT
1617 /* Want to start with kernel preemption disabled. */
1618 task_thread_info(p
)->preempt_count
= 1;
1621 * Share the timeslice between parent and child, thus the
1622 * total amount of pending timeslices in the system doesn't change,
1623 * resulting in more scheduling fairness.
1625 local_irq_disable();
1626 p
->time_slice
= (current
->time_slice
+ 1) >> 1;
1628 * The remainder of the first timeslice might be recovered by
1629 * the parent if the child exits early enough.
1631 p
->first_time_slice
= 1;
1632 current
->time_slice
>>= 1;
1633 p
->timestamp
= sched_clock();
1634 if (unlikely(!current
->time_slice
)) {
1636 * This case is rare, it happens when the parent has only
1637 * a single jiffy left from its timeslice. Taking the
1638 * runqueue lock is not a problem.
1640 current
->time_slice
= 1;
1641 task_running_tick(cpu_rq(cpu
), current
);
1648 * wake_up_new_task - wake up a newly created task for the first time.
1650 * This function will do some initial scheduler statistics housekeeping
1651 * that must be done for every newly created context, then puts the task
1652 * on the runqueue and wakes it.
1654 void fastcall
wake_up_new_task(struct task_struct
*p
, unsigned long clone_flags
)
1656 struct rq
*rq
, *this_rq
;
1657 unsigned long flags
;
1660 rq
= task_rq_lock(p
, &flags
);
1661 BUG_ON(p
->state
!= TASK_RUNNING
);
1662 this_cpu
= smp_processor_id();
1666 * We decrease the sleep average of forking parents
1667 * and children as well, to keep max-interactive tasks
1668 * from forking tasks that are max-interactive. The parent
1669 * (current) is done further down, under its lock.
1671 p
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(p
) *
1672 CHILD_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1674 p
->prio
= effective_prio(p
);
1676 if (likely(cpu
== this_cpu
)) {
1677 if (!(clone_flags
& CLONE_VM
)) {
1679 * The VM isn't cloned, so we're in a good position to
1680 * do child-runs-first in anticipation of an exec. This
1681 * usually avoids a lot of COW overhead.
1683 if (unlikely(!current
->array
))
1684 __activate_task(p
, rq
);
1686 p
->prio
= current
->prio
;
1687 p
->normal_prio
= current
->normal_prio
;
1688 list_add_tail(&p
->run_list
, ¤t
->run_list
);
1689 p
->array
= current
->array
;
1690 p
->array
->nr_active
++;
1691 inc_nr_running(p
, rq
);
1695 /* Run child last */
1696 __activate_task(p
, rq
);
1698 * We skip the following code due to cpu == this_cpu
1700 * task_rq_unlock(rq, &flags);
1701 * this_rq = task_rq_lock(current, &flags);
1705 this_rq
= cpu_rq(this_cpu
);
1708 * Not the local CPU - must adjust timestamp. This should
1709 * get optimised away in the !CONFIG_SMP case.
1711 p
->timestamp
= (p
->timestamp
- this_rq
->most_recent_timestamp
)
1712 + rq
->most_recent_timestamp
;
1713 __activate_task(p
, rq
);
1714 if (TASK_PREEMPTS_CURR(p
, rq
))
1715 resched_task(rq
->curr
);
1718 * Parent and child are on different CPUs, now get the
1719 * parent runqueue to update the parent's ->sleep_avg:
1721 task_rq_unlock(rq
, &flags
);
1722 this_rq
= task_rq_lock(current
, &flags
);
1724 current
->sleep_avg
= JIFFIES_TO_NS(CURRENT_BONUS(current
) *
1725 PARENT_PENALTY
/ 100 * MAX_SLEEP_AVG
/ MAX_BONUS
);
1726 task_rq_unlock(this_rq
, &flags
);
1730 * Potentially available exiting-child timeslices are
1731 * retrieved here - this way the parent does not get
1732 * penalized for creating too many threads.
1734 * (this cannot be used to 'generate' timeslices
1735 * artificially, because any timeslice recovered here
1736 * was given away by the parent in the first place.)
1738 void fastcall
sched_exit(struct task_struct
*p
)
1740 unsigned long flags
;
1744 * If the child was a (relative-) CPU hog then decrease
1745 * the sleep_avg of the parent as well.
1747 rq
= task_rq_lock(p
->parent
, &flags
);
1748 if (p
->first_time_slice
&& task_cpu(p
) == task_cpu(p
->parent
)) {
1749 p
->parent
->time_slice
+= p
->time_slice
;
1750 if (unlikely(p
->parent
->time_slice
> task_timeslice(p
)))
1751 p
->parent
->time_slice
= task_timeslice(p
);
1753 if (p
->sleep_avg
< p
->parent
->sleep_avg
)
1754 p
->parent
->sleep_avg
= p
->parent
->sleep_avg
/
1755 (EXIT_WEIGHT
+ 1) * EXIT_WEIGHT
+ p
->sleep_avg
/
1757 task_rq_unlock(rq
, &flags
);
1761 * prepare_task_switch - prepare to switch tasks
1762 * @rq: the runqueue preparing to switch
1763 * @next: the task we are going to switch to.
1765 * This is called with the rq lock held and interrupts off. It must
1766 * be paired with a subsequent finish_task_switch after the context
1769 * prepare_task_switch sets up locking and calls architecture specific
1772 static inline void prepare_task_switch(struct rq
*rq
, struct task_struct
*next
)
1774 prepare_lock_switch(rq
, next
);
1775 prepare_arch_switch(next
);
1779 * finish_task_switch - clean up after a task-switch
1780 * @rq: runqueue associated with task-switch
1781 * @prev: the thread we just switched away from.
1783 * finish_task_switch must be called after the context switch, paired
1784 * with a prepare_task_switch call before the context switch.
1785 * finish_task_switch will reconcile locking set up by prepare_task_switch,
1786 * and do any other architecture-specific cleanup actions.
1788 * Note that we may have delayed dropping an mm in context_switch(). If
1789 * so, we finish that here outside of the runqueue lock. (Doing it
1790 * with the lock held can cause deadlocks; see schedule() for
1793 static inline void finish_task_switch(struct rq
*rq
, struct task_struct
*prev
)
1794 __releases(rq
->lock
)
1796 struct mm_struct
*mm
= rq
->prev_mm
;
1802 * A task struct has one reference for the use as "current".
1803 * If a task dies, then it sets TASK_DEAD in tsk->state and calls
1804 * schedule one last time. The schedule call will never return, and
1805 * the scheduled task must drop that reference.
1806 * The test for TASK_DEAD must occur while the runqueue locks are
1807 * still held, otherwise prev could be scheduled on another cpu, die
1808 * there before we look at prev->state, and then the reference would
1810 * Manfred Spraul <manfred@colorfullife.com>
1812 prev_state
= prev
->state
;
1813 finish_arch_switch(prev
);
1814 finish_lock_switch(rq
, prev
);
1817 if (unlikely(prev_state
== TASK_DEAD
)) {
1819 * Remove function-return probe instances associated with this
1820 * task and put them back on the free list.
1822 kprobe_flush_task(prev
);
1823 put_task_struct(prev
);
1828 * schedule_tail - first thing a freshly forked thread must call.
1829 * @prev: the thread we just switched away from.
1831 asmlinkage
void schedule_tail(struct task_struct
*prev
)
1832 __releases(rq
->lock
)
1834 struct rq
*rq
= this_rq();
1836 finish_task_switch(rq
, prev
);
1837 #ifdef __ARCH_WANT_UNLOCKED_CTXSW
1838 /* In this case, finish_task_switch does not reenable preemption */
1841 if (current
->set_child_tid
)
1842 put_user(current
->pid
, current
->set_child_tid
);
1846 * context_switch - switch to the new MM and the new
1847 * thread's register state.
1849 static inline struct task_struct
*
1850 context_switch(struct rq
*rq
, struct task_struct
*prev
,
1851 struct task_struct
*next
)
1853 struct mm_struct
*mm
= next
->mm
;
1854 struct mm_struct
*oldmm
= prev
->active_mm
;
1857 next
->active_mm
= oldmm
;
1858 atomic_inc(&oldmm
->mm_count
);
1859 enter_lazy_tlb(oldmm
, next
);
1861 switch_mm(oldmm
, mm
, next
);
1864 prev
->active_mm
= NULL
;
1865 WARN_ON(rq
->prev_mm
);
1866 rq
->prev_mm
= oldmm
;
1869 * Since the runqueue lock will be released by the next
1870 * task (which is an invalid locking op but in the case
1871 * of the scheduler it's an obvious special-case), so we
1872 * do an early lockdep release here:
1874 #ifndef __ARCH_WANT_UNLOCKED_CTXSW
1875 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
1878 /* Here we just switch the register state and the stack. */
1879 switch_to(prev
, next
, prev
);
1885 * nr_running, nr_uninterruptible and nr_context_switches:
1887 * externally visible scheduler statistics: current number of runnable
1888 * threads, current number of uninterruptible-sleeping threads, total
1889 * number of context switches performed since bootup.
1891 unsigned long nr_running(void)
1893 unsigned long i
, sum
= 0;
1895 for_each_online_cpu(i
)
1896 sum
+= cpu_rq(i
)->nr_running
;
1901 unsigned long nr_uninterruptible(void)
1903 unsigned long i
, sum
= 0;
1905 for_each_possible_cpu(i
)
1906 sum
+= cpu_rq(i
)->nr_uninterruptible
;
1909 * Since we read the counters lockless, it might be slightly
1910 * inaccurate. Do not allow it to go below zero though:
1912 if (unlikely((long)sum
< 0))
1918 unsigned long long nr_context_switches(void)
1921 unsigned long long sum
= 0;
1923 for_each_possible_cpu(i
)
1924 sum
+= cpu_rq(i
)->nr_switches
;
1929 unsigned long nr_iowait(void)
1931 unsigned long i
, sum
= 0;
1933 for_each_possible_cpu(i
)
1934 sum
+= atomic_read(&cpu_rq(i
)->nr_iowait
);
1939 unsigned long nr_active(void)
1941 unsigned long i
, running
= 0, uninterruptible
= 0;
1943 for_each_online_cpu(i
) {
1944 running
+= cpu_rq(i
)->nr_running
;
1945 uninterruptible
+= cpu_rq(i
)->nr_uninterruptible
;
1948 if (unlikely((long)uninterruptible
< 0))
1949 uninterruptible
= 0;
1951 return running
+ uninterruptible
;
1957 * Is this task likely cache-hot:
1960 task_hot(struct task_struct
*p
, unsigned long long now
, struct sched_domain
*sd
)
1962 return (long long)(now
- p
->last_ran
) < (long long)sd
->cache_hot_time
;
1966 * double_rq_lock - safely lock two runqueues
1968 * Note this does not disable interrupts like task_rq_lock,
1969 * you need to do so manually before calling.
1971 static void double_rq_lock(struct rq
*rq1
, struct rq
*rq2
)
1972 __acquires(rq1
->lock
)
1973 __acquires(rq2
->lock
)
1975 BUG_ON(!irqs_disabled());
1977 spin_lock(&rq1
->lock
);
1978 __acquire(rq2
->lock
); /* Fake it out ;) */
1981 spin_lock(&rq1
->lock
);
1982 spin_lock(&rq2
->lock
);
1984 spin_lock(&rq2
->lock
);
1985 spin_lock(&rq1
->lock
);
1991 * double_rq_unlock - safely unlock two runqueues
1993 * Note this does not restore interrupts like task_rq_unlock,
1994 * you need to do so manually after calling.
1996 static void double_rq_unlock(struct rq
*rq1
, struct rq
*rq2
)
1997 __releases(rq1
->lock
)
1998 __releases(rq2
->lock
)
2000 spin_unlock(&rq1
->lock
);
2002 spin_unlock(&rq2
->lock
);
2004 __release(rq2
->lock
);
2008 * double_lock_balance - lock the busiest runqueue, this_rq is locked already.
2010 static void double_lock_balance(struct rq
*this_rq
, struct rq
*busiest
)
2011 __releases(this_rq
->lock
)
2012 __acquires(busiest
->lock
)
2013 __acquires(this_rq
->lock
)
2015 if (unlikely(!irqs_disabled())) {
2016 /* printk() doesn't work good under rq->lock */
2017 spin_unlock(&this_rq
->lock
);
2020 if (unlikely(!spin_trylock(&busiest
->lock
))) {
2021 if (busiest
< this_rq
) {
2022 spin_unlock(&this_rq
->lock
);
2023 spin_lock(&busiest
->lock
);
2024 spin_lock(&this_rq
->lock
);
2026 spin_lock(&busiest
->lock
);
2031 * If dest_cpu is allowed for this process, migrate the task to it.
2032 * This is accomplished by forcing the cpu_allowed mask to only
2033 * allow dest_cpu, which will force the cpu onto dest_cpu. Then
2034 * the cpu_allowed mask is restored.
2036 static void sched_migrate_task(struct task_struct
*p
, int dest_cpu
)
2038 struct migration_req req
;
2039 unsigned long flags
;
2042 rq
= task_rq_lock(p
, &flags
);
2043 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
)
2044 || unlikely(cpu_is_offline(dest_cpu
)))
2047 /* force the process onto the specified CPU */
2048 if (migrate_task(p
, dest_cpu
, &req
)) {
2049 /* Need to wait for migration thread (might exit: take ref). */
2050 struct task_struct
*mt
= rq
->migration_thread
;
2052 get_task_struct(mt
);
2053 task_rq_unlock(rq
, &flags
);
2054 wake_up_process(mt
);
2055 put_task_struct(mt
);
2056 wait_for_completion(&req
.done
);
2061 task_rq_unlock(rq
, &flags
);
2065 * sched_exec - execve() is a valuable balancing opportunity, because at
2066 * this point the task has the smallest effective memory and cache footprint.
2068 void sched_exec(void)
2070 int new_cpu
, this_cpu
= get_cpu();
2071 new_cpu
= sched_balance_self(this_cpu
, SD_BALANCE_EXEC
);
2073 if (new_cpu
!= this_cpu
)
2074 sched_migrate_task(current
, new_cpu
);
2078 * pull_task - move a task from a remote runqueue to the local runqueue.
2079 * Both runqueues must be locked.
2081 static void pull_task(struct rq
*src_rq
, struct prio_array
*src_array
,
2082 struct task_struct
*p
, struct rq
*this_rq
,
2083 struct prio_array
*this_array
, int this_cpu
)
2085 dequeue_task(p
, src_array
);
2086 dec_nr_running(p
, src_rq
);
2087 set_task_cpu(p
, this_cpu
);
2088 inc_nr_running(p
, this_rq
);
2089 enqueue_task(p
, this_array
);
2090 p
->timestamp
= (p
->timestamp
- src_rq
->most_recent_timestamp
)
2091 + this_rq
->most_recent_timestamp
;
2093 * Note that idle threads have a prio of MAX_PRIO, for this test
2094 * to be always true for them.
2096 if (TASK_PREEMPTS_CURR(p
, this_rq
))
2097 resched_task(this_rq
->curr
);
2101 * can_migrate_task - may task p from runqueue rq be migrated to this_cpu?
2104 int can_migrate_task(struct task_struct
*p
, struct rq
*rq
, int this_cpu
,
2105 struct sched_domain
*sd
, enum idle_type idle
,
2109 * We do not migrate tasks that are:
2110 * 1) running (obviously), or
2111 * 2) cannot be migrated to this CPU due to cpus_allowed, or
2112 * 3) are cache-hot on their current CPU.
2114 if (!cpu_isset(this_cpu
, p
->cpus_allowed
))
2118 if (task_running(rq
, p
))
2122 * Aggressive migration if:
2123 * 1) task is cache cold, or
2124 * 2) too many balance attempts have failed.
2127 if (sd
->nr_balance_failed
> sd
->cache_nice_tries
) {
2128 #ifdef CONFIG_SCHEDSTATS
2129 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2130 schedstat_inc(sd
, lb_hot_gained
[idle
]);
2135 if (task_hot(p
, rq
->most_recent_timestamp
, sd
))
2140 #define rq_best_prio(rq) min((rq)->curr->prio, (rq)->best_expired_prio)
2143 * move_tasks tries to move up to max_nr_move tasks and max_load_move weighted
2144 * load from busiest to this_rq, as part of a balancing operation within
2145 * "domain". Returns the number of tasks moved.
2147 * Called with both runqueues locked.
2149 static int move_tasks(struct rq
*this_rq
, int this_cpu
, struct rq
*busiest
,
2150 unsigned long max_nr_move
, unsigned long max_load_move
,
2151 struct sched_domain
*sd
, enum idle_type idle
,
2154 int idx
, pulled
= 0, pinned
= 0, this_best_prio
, best_prio
,
2155 best_prio_seen
, skip_for_load
;
2156 struct prio_array
*array
, *dst_array
;
2157 struct list_head
*head
, *curr
;
2158 struct task_struct
*tmp
;
2161 if (max_nr_move
== 0 || max_load_move
== 0)
2164 rem_load_move
= max_load_move
;
2166 this_best_prio
= rq_best_prio(this_rq
);
2167 best_prio
= rq_best_prio(busiest
);
2169 * Enable handling of the case where there is more than one task
2170 * with the best priority. If the current running task is one
2171 * of those with prio==best_prio we know it won't be moved
2172 * and therefore it's safe to override the skip (based on load) of
2173 * any task we find with that prio.
2175 best_prio_seen
= best_prio
== busiest
->curr
->prio
;
2178 * We first consider expired tasks. Those will likely not be
2179 * executed in the near future, and they are most likely to
2180 * be cache-cold, thus switching CPUs has the least effect
2183 if (busiest
->expired
->nr_active
) {
2184 array
= busiest
->expired
;
2185 dst_array
= this_rq
->expired
;
2187 array
= busiest
->active
;
2188 dst_array
= this_rq
->active
;
2192 /* Start searching at priority 0: */
2196 idx
= sched_find_first_bit(array
->bitmap
);
2198 idx
= find_next_bit(array
->bitmap
, MAX_PRIO
, idx
);
2199 if (idx
>= MAX_PRIO
) {
2200 if (array
== busiest
->expired
&& busiest
->active
->nr_active
) {
2201 array
= busiest
->active
;
2202 dst_array
= this_rq
->active
;
2208 head
= array
->queue
+ idx
;
2211 tmp
= list_entry(curr
, struct task_struct
, run_list
);
2216 * To help distribute high priority tasks accross CPUs we don't
2217 * skip a task if it will be the highest priority task (i.e. smallest
2218 * prio value) on its new queue regardless of its load weight
2220 skip_for_load
= tmp
->load_weight
> rem_load_move
;
2221 if (skip_for_load
&& idx
< this_best_prio
)
2222 skip_for_load
= !best_prio_seen
&& idx
== best_prio
;
2223 if (skip_for_load
||
2224 !can_migrate_task(tmp
, busiest
, this_cpu
, sd
, idle
, &pinned
)) {
2226 best_prio_seen
|= idx
== best_prio
;
2233 pull_task(busiest
, array
, tmp
, this_rq
, dst_array
, this_cpu
);
2235 rem_load_move
-= tmp
->load_weight
;
2238 * We only want to steal up to the prescribed number of tasks
2239 * and the prescribed amount of weighted load.
2241 if (pulled
< max_nr_move
&& rem_load_move
> 0) {
2242 if (idx
< this_best_prio
)
2243 this_best_prio
= idx
;
2251 * Right now, this is the only place pull_task() is called,
2252 * so we can safely collect pull_task() stats here rather than
2253 * inside pull_task().
2255 schedstat_add(sd
, lb_gained
[idle
], pulled
);
2258 *all_pinned
= pinned
;
2263 * find_busiest_group finds and returns the busiest CPU group within the
2264 * domain. It calculates and returns the amount of weighted load which
2265 * should be moved to restore balance via the imbalance parameter.
2267 static struct sched_group
*
2268 find_busiest_group(struct sched_domain
*sd
, int this_cpu
,
2269 unsigned long *imbalance
, enum idle_type idle
, int *sd_idle
,
2270 cpumask_t
*cpus
, int *balance
)
2272 struct sched_group
*busiest
= NULL
, *this = NULL
, *group
= sd
->groups
;
2273 unsigned long max_load
, avg_load
, total_load
, this_load
, total_pwr
;
2274 unsigned long max_pull
;
2275 unsigned long busiest_load_per_task
, busiest_nr_running
;
2276 unsigned long this_load_per_task
, this_nr_running
;
2278 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2279 int power_savings_balance
= 1;
2280 unsigned long leader_nr_running
= 0, min_load_per_task
= 0;
2281 unsigned long min_nr_running
= ULONG_MAX
;
2282 struct sched_group
*group_min
= NULL
, *group_leader
= NULL
;
2285 max_load
= this_load
= total_load
= total_pwr
= 0;
2286 busiest_load_per_task
= busiest_nr_running
= 0;
2287 this_load_per_task
= this_nr_running
= 0;
2288 if (idle
== NOT_IDLE
)
2289 load_idx
= sd
->busy_idx
;
2290 else if (idle
== NEWLY_IDLE
)
2291 load_idx
= sd
->newidle_idx
;
2293 load_idx
= sd
->idle_idx
;
2296 unsigned long load
, group_capacity
;
2299 unsigned int balance_cpu
= -1, first_idle_cpu
= 0;
2300 unsigned long sum_nr_running
, sum_weighted_load
;
2302 local_group
= cpu_isset(this_cpu
, group
->cpumask
);
2305 balance_cpu
= first_cpu(group
->cpumask
);
2307 /* Tally up the load of all CPUs in the group */
2308 sum_weighted_load
= sum_nr_running
= avg_load
= 0;
2310 for_each_cpu_mask(i
, group
->cpumask
) {
2313 if (!cpu_isset(i
, *cpus
))
2318 if (*sd_idle
&& !idle_cpu(i
))
2321 /* Bias balancing toward cpus of our domain */
2323 if (idle_cpu(i
) && !first_idle_cpu
) {
2328 load
= target_load(i
, load_idx
);
2330 load
= source_load(i
, load_idx
);
2333 sum_nr_running
+= rq
->nr_running
;
2334 sum_weighted_load
+= rq
->raw_weighted_load
;
2338 * First idle cpu or the first cpu(busiest) in this sched group
2339 * is eligible for doing load balancing at this and above
2342 if (local_group
&& balance_cpu
!= this_cpu
&& balance
) {
2347 total_load
+= avg_load
;
2348 total_pwr
+= group
->cpu_power
;
2350 /* Adjust by relative CPU power of the group */
2351 avg_load
= (avg_load
* SCHED_LOAD_SCALE
) / group
->cpu_power
;
2353 group_capacity
= group
->cpu_power
/ SCHED_LOAD_SCALE
;
2356 this_load
= avg_load
;
2358 this_nr_running
= sum_nr_running
;
2359 this_load_per_task
= sum_weighted_load
;
2360 } else if (avg_load
> max_load
&&
2361 sum_nr_running
> group_capacity
) {
2362 max_load
= avg_load
;
2364 busiest_nr_running
= sum_nr_running
;
2365 busiest_load_per_task
= sum_weighted_load
;
2368 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2370 * Busy processors will not participate in power savings
2373 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2377 * If the local group is idle or completely loaded
2378 * no need to do power savings balance at this domain
2380 if (local_group
&& (this_nr_running
>= group_capacity
||
2382 power_savings_balance
= 0;
2385 * If a group is already running at full capacity or idle,
2386 * don't include that group in power savings calculations
2388 if (!power_savings_balance
|| sum_nr_running
>= group_capacity
2393 * Calculate the group which has the least non-idle load.
2394 * This is the group from where we need to pick up the load
2397 if ((sum_nr_running
< min_nr_running
) ||
2398 (sum_nr_running
== min_nr_running
&&
2399 first_cpu(group
->cpumask
) <
2400 first_cpu(group_min
->cpumask
))) {
2402 min_nr_running
= sum_nr_running
;
2403 min_load_per_task
= sum_weighted_load
/
2408 * Calculate the group which is almost near its
2409 * capacity but still has some space to pick up some load
2410 * from other group and save more power
2412 if (sum_nr_running
<= group_capacity
- 1) {
2413 if (sum_nr_running
> leader_nr_running
||
2414 (sum_nr_running
== leader_nr_running
&&
2415 first_cpu(group
->cpumask
) >
2416 first_cpu(group_leader
->cpumask
))) {
2417 group_leader
= group
;
2418 leader_nr_running
= sum_nr_running
;
2423 group
= group
->next
;
2424 } while (group
!= sd
->groups
);
2426 if (!busiest
|| this_load
>= max_load
|| busiest_nr_running
== 0)
2429 avg_load
= (SCHED_LOAD_SCALE
* total_load
) / total_pwr
;
2431 if (this_load
>= avg_load
||
2432 100*max_load
<= sd
->imbalance_pct
*this_load
)
2435 busiest_load_per_task
/= busiest_nr_running
;
2437 * We're trying to get all the cpus to the average_load, so we don't
2438 * want to push ourselves above the average load, nor do we wish to
2439 * reduce the max loaded cpu below the average load, as either of these
2440 * actions would just result in more rebalancing later, and ping-pong
2441 * tasks around. Thus we look for the minimum possible imbalance.
2442 * Negative imbalances (*we* are more loaded than anyone else) will
2443 * be counted as no imbalance for these purposes -- we can't fix that
2444 * by pulling tasks to us. Be careful of negative numbers as they'll
2445 * appear as very large values with unsigned longs.
2447 if (max_load
<= busiest_load_per_task
)
2451 * In the presence of smp nice balancing, certain scenarios can have
2452 * max load less than avg load(as we skip the groups at or below
2453 * its cpu_power, while calculating max_load..)
2455 if (max_load
< avg_load
) {
2457 goto small_imbalance
;
2460 /* Don't want to pull so many tasks that a group would go idle */
2461 max_pull
= min(max_load
- avg_load
, max_load
- busiest_load_per_task
);
2463 /* How much load to actually move to equalise the imbalance */
2464 *imbalance
= min(max_pull
* busiest
->cpu_power
,
2465 (avg_load
- this_load
) * this->cpu_power
)
2469 * if *imbalance is less than the average load per runnable task
2470 * there is no gaurantee that any tasks will be moved so we'll have
2471 * a think about bumping its value to force at least one task to be
2474 if (*imbalance
< busiest_load_per_task
) {
2475 unsigned long tmp
, pwr_now
, pwr_move
;
2479 pwr_move
= pwr_now
= 0;
2481 if (this_nr_running
) {
2482 this_load_per_task
/= this_nr_running
;
2483 if (busiest_load_per_task
> this_load_per_task
)
2486 this_load_per_task
= SCHED_LOAD_SCALE
;
2488 if (max_load
- this_load
>= busiest_load_per_task
* imbn
) {
2489 *imbalance
= busiest_load_per_task
;
2494 * OK, we don't have enough imbalance to justify moving tasks,
2495 * however we may be able to increase total CPU power used by
2499 pwr_now
+= busiest
->cpu_power
*
2500 min(busiest_load_per_task
, max_load
);
2501 pwr_now
+= this->cpu_power
*
2502 min(this_load_per_task
, this_load
);
2503 pwr_now
/= SCHED_LOAD_SCALE
;
2505 /* Amount of load we'd subtract */
2506 tmp
= busiest_load_per_task
* SCHED_LOAD_SCALE
/
2509 pwr_move
+= busiest
->cpu_power
*
2510 min(busiest_load_per_task
, max_load
- tmp
);
2512 /* Amount of load we'd add */
2513 if (max_load
* busiest
->cpu_power
<
2514 busiest_load_per_task
* SCHED_LOAD_SCALE
)
2515 tmp
= max_load
* busiest
->cpu_power
/ this->cpu_power
;
2517 tmp
= busiest_load_per_task
* SCHED_LOAD_SCALE
/
2519 pwr_move
+= this->cpu_power
*
2520 min(this_load_per_task
, this_load
+ tmp
);
2521 pwr_move
/= SCHED_LOAD_SCALE
;
2523 /* Move if we gain throughput */
2524 if (pwr_move
<= pwr_now
)
2527 *imbalance
= busiest_load_per_task
;
2533 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
2534 if (idle
== NOT_IDLE
|| !(sd
->flags
& SD_POWERSAVINGS_BALANCE
))
2537 if (this == group_leader
&& group_leader
!= group_min
) {
2538 *imbalance
= min_load_per_task
;
2548 * find_busiest_queue - find the busiest runqueue among the cpus in group.
2551 find_busiest_queue(struct sched_group
*group
, enum idle_type idle
,
2552 unsigned long imbalance
, cpumask_t
*cpus
)
2554 struct rq
*busiest
= NULL
, *rq
;
2555 unsigned long max_load
= 0;
2558 for_each_cpu_mask(i
, group
->cpumask
) {
2560 if (!cpu_isset(i
, *cpus
))
2565 if (rq
->nr_running
== 1 && rq
->raw_weighted_load
> imbalance
)
2568 if (rq
->raw_weighted_load
> max_load
) {
2569 max_load
= rq
->raw_weighted_load
;
2578 * Max backoff if we encounter pinned tasks. Pretty arbitrary value, but
2579 * so long as it is large enough.
2581 #define MAX_PINNED_INTERVAL 512
2583 static inline unsigned long minus_1_or_zero(unsigned long n
)
2585 return n
> 0 ? n
- 1 : 0;
2589 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2590 * tasks if there is an imbalance.
2592 static int load_balance(int this_cpu
, struct rq
*this_rq
,
2593 struct sched_domain
*sd
, enum idle_type idle
,
2596 int nr_moved
, all_pinned
= 0, active_balance
= 0, sd_idle
= 0;
2597 struct sched_group
*group
;
2598 unsigned long imbalance
;
2600 cpumask_t cpus
= CPU_MASK_ALL
;
2601 unsigned long flags
;
2604 * When power savings policy is enabled for the parent domain, idle
2605 * sibling can pick up load irrespective of busy siblings. In this case,
2606 * let the state of idle sibling percolate up as IDLE, instead of
2607 * portraying it as NOT_IDLE.
2609 if (idle
!= NOT_IDLE
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2610 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2613 schedstat_inc(sd
, lb_cnt
[idle
]);
2616 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, idle
, &sd_idle
,
2623 schedstat_inc(sd
, lb_nobusyg
[idle
]);
2627 busiest
= find_busiest_queue(group
, idle
, imbalance
, &cpus
);
2629 schedstat_inc(sd
, lb_nobusyq
[idle
]);
2633 BUG_ON(busiest
== this_rq
);
2635 schedstat_add(sd
, lb_imbalance
[idle
], imbalance
);
2638 if (busiest
->nr_running
> 1) {
2640 * Attempt to move tasks. If find_busiest_group has found
2641 * an imbalance but busiest->nr_running <= 1, the group is
2642 * still unbalanced. nr_moved simply stays zero, so it is
2643 * correctly treated as an imbalance.
2645 local_irq_save(flags
);
2646 double_rq_lock(this_rq
, busiest
);
2647 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2648 minus_1_or_zero(busiest
->nr_running
),
2649 imbalance
, sd
, idle
, &all_pinned
);
2650 double_rq_unlock(this_rq
, busiest
);
2651 local_irq_restore(flags
);
2653 /* All tasks on this runqueue were pinned by CPU affinity */
2654 if (unlikely(all_pinned
)) {
2655 cpu_clear(cpu_of(busiest
), cpus
);
2656 if (!cpus_empty(cpus
))
2663 schedstat_inc(sd
, lb_failed
[idle
]);
2664 sd
->nr_balance_failed
++;
2666 if (unlikely(sd
->nr_balance_failed
> sd
->cache_nice_tries
+2)) {
2668 spin_lock_irqsave(&busiest
->lock
, flags
);
2670 /* don't kick the migration_thread, if the curr
2671 * task on busiest cpu can't be moved to this_cpu
2673 if (!cpu_isset(this_cpu
, busiest
->curr
->cpus_allowed
)) {
2674 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2676 goto out_one_pinned
;
2679 if (!busiest
->active_balance
) {
2680 busiest
->active_balance
= 1;
2681 busiest
->push_cpu
= this_cpu
;
2684 spin_unlock_irqrestore(&busiest
->lock
, flags
);
2686 wake_up_process(busiest
->migration_thread
);
2689 * We've kicked active balancing, reset the failure
2692 sd
->nr_balance_failed
= sd
->cache_nice_tries
+1;
2695 sd
->nr_balance_failed
= 0;
2697 if (likely(!active_balance
)) {
2698 /* We were unbalanced, so reset the balancing interval */
2699 sd
->balance_interval
= sd
->min_interval
;
2702 * If we've begun active balancing, start to back off. This
2703 * case may not be covered by the all_pinned logic if there
2704 * is only 1 task on the busy runqueue (because we don't call
2707 if (sd
->balance_interval
< sd
->max_interval
)
2708 sd
->balance_interval
*= 2;
2711 if (!nr_moved
&& !sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2712 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2717 schedstat_inc(sd
, lb_balanced
[idle
]);
2719 sd
->nr_balance_failed
= 0;
2722 /* tune up the balancing interval */
2723 if ((all_pinned
&& sd
->balance_interval
< MAX_PINNED_INTERVAL
) ||
2724 (sd
->balance_interval
< sd
->max_interval
))
2725 sd
->balance_interval
*= 2;
2727 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2728 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2734 * Check this_cpu to ensure it is balanced within domain. Attempt to move
2735 * tasks if there is an imbalance.
2737 * Called from schedule when this_rq is about to become idle (NEWLY_IDLE).
2738 * this_rq is locked.
2741 load_balance_newidle(int this_cpu
, struct rq
*this_rq
, struct sched_domain
*sd
)
2743 struct sched_group
*group
;
2744 struct rq
*busiest
= NULL
;
2745 unsigned long imbalance
;
2748 cpumask_t cpus
= CPU_MASK_ALL
;
2751 * When power savings policy is enabled for the parent domain, idle
2752 * sibling can pick up load irrespective of busy siblings. In this case,
2753 * let the state of idle sibling percolate up as IDLE, instead of
2754 * portraying it as NOT_IDLE.
2756 if (sd
->flags
& SD_SHARE_CPUPOWER
&&
2757 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2760 schedstat_inc(sd
, lb_cnt
[NEWLY_IDLE
]);
2762 group
= find_busiest_group(sd
, this_cpu
, &imbalance
, NEWLY_IDLE
,
2763 &sd_idle
, &cpus
, NULL
);
2765 schedstat_inc(sd
, lb_nobusyg
[NEWLY_IDLE
]);
2769 busiest
= find_busiest_queue(group
, NEWLY_IDLE
, imbalance
,
2772 schedstat_inc(sd
, lb_nobusyq
[NEWLY_IDLE
]);
2776 BUG_ON(busiest
== this_rq
);
2778 schedstat_add(sd
, lb_imbalance
[NEWLY_IDLE
], imbalance
);
2781 if (busiest
->nr_running
> 1) {
2782 /* Attempt to move tasks */
2783 double_lock_balance(this_rq
, busiest
);
2784 nr_moved
= move_tasks(this_rq
, this_cpu
, busiest
,
2785 minus_1_or_zero(busiest
->nr_running
),
2786 imbalance
, sd
, NEWLY_IDLE
, NULL
);
2787 spin_unlock(&busiest
->lock
);
2790 cpu_clear(cpu_of(busiest
), cpus
);
2791 if (!cpus_empty(cpus
))
2797 schedstat_inc(sd
, lb_failed
[NEWLY_IDLE
]);
2798 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2799 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2802 sd
->nr_balance_failed
= 0;
2807 schedstat_inc(sd
, lb_balanced
[NEWLY_IDLE
]);
2808 if (!sd_idle
&& sd
->flags
& SD_SHARE_CPUPOWER
&&
2809 !test_sd_parent(sd
, SD_POWERSAVINGS_BALANCE
))
2811 sd
->nr_balance_failed
= 0;
2817 * idle_balance is called by schedule() if this_cpu is about to become
2818 * idle. Attempts to pull tasks from other CPUs.
2820 static void idle_balance(int this_cpu
, struct rq
*this_rq
)
2822 struct sched_domain
*sd
;
2823 int pulled_task
= 0;
2824 unsigned long next_balance
= jiffies
+ 60 * HZ
;
2826 for_each_domain(this_cpu
, sd
) {
2827 if (sd
->flags
& SD_BALANCE_NEWIDLE
) {
2828 /* If we've pulled tasks over stop searching: */
2829 pulled_task
= load_balance_newidle(this_cpu
,
2831 if (time_after(next_balance
,
2832 sd
->last_balance
+ sd
->balance_interval
))
2833 next_balance
= sd
->last_balance
2834 + sd
->balance_interval
;
2841 * We are going idle. next_balance may be set based on
2842 * a busy processor. So reset next_balance.
2844 this_rq
->next_balance
= next_balance
;
2848 * active_load_balance is run by migration threads. It pushes running tasks
2849 * off the busiest CPU onto idle CPUs. It requires at least 1 task to be
2850 * running on each physical CPU where possible, and avoids physical /
2851 * logical imbalances.
2853 * Called with busiest_rq locked.
2855 static void active_load_balance(struct rq
*busiest_rq
, int busiest_cpu
)
2857 int target_cpu
= busiest_rq
->push_cpu
;
2858 struct sched_domain
*sd
;
2859 struct rq
*target_rq
;
2861 /* Is there any task to move? */
2862 if (busiest_rq
->nr_running
<= 1)
2865 target_rq
= cpu_rq(target_cpu
);
2868 * This condition is "impossible", if it occurs
2869 * we need to fix it. Originally reported by
2870 * Bjorn Helgaas on a 128-cpu setup.
2872 BUG_ON(busiest_rq
== target_rq
);
2874 /* move a task from busiest_rq to target_rq */
2875 double_lock_balance(busiest_rq
, target_rq
);
2877 /* Search for an sd spanning us and the target CPU. */
2878 for_each_domain(target_cpu
, sd
) {
2879 if ((sd
->flags
& SD_LOAD_BALANCE
) &&
2880 cpu_isset(busiest_cpu
, sd
->span
))
2885 schedstat_inc(sd
, alb_cnt
);
2887 if (move_tasks(target_rq
, target_cpu
, busiest_rq
, 1,
2888 RTPRIO_TO_LOAD_WEIGHT(100), sd
, SCHED_IDLE
,
2890 schedstat_inc(sd
, alb_pushed
);
2892 schedstat_inc(sd
, alb_failed
);
2894 spin_unlock(&target_rq
->lock
);
2897 static void update_load(struct rq
*this_rq
)
2899 unsigned long this_load
;
2902 this_load
= this_rq
->raw_weighted_load
;
2904 /* Update our load: */
2905 for (i
= 0, scale
= 1; i
< 3; i
++, scale
<<= 1) {
2906 unsigned long old_load
, new_load
;
2908 old_load
= this_rq
->cpu_load
[i
];
2909 new_load
= this_load
;
2911 * Round up the averaging division if load is increasing. This
2912 * prevents us from getting stuck on 9 if the load is 10, for
2915 if (new_load
> old_load
)
2916 new_load
+= scale
-1;
2917 this_rq
->cpu_load
[i
] = (old_load
*(scale
-1) + new_load
) / scale
;
2922 * run_rebalance_domains is triggered when needed from the scheduler tick.
2924 * It checks each scheduling domain to see if it is due to be balanced,
2925 * and initiates a balancing operation if so.
2927 * Balancing parameters are set up in arch_init_sched_domains.
2929 static DEFINE_SPINLOCK(balancing
);
2931 static void run_rebalance_domains(struct softirq_action
*h
)
2933 int this_cpu
= smp_processor_id(), balance
= 1;
2934 struct rq
*this_rq
= cpu_rq(this_cpu
);
2935 unsigned long interval
;
2936 struct sched_domain
*sd
;
2938 * We are idle if there are no processes running. This
2939 * is valid even if we are the idle process (SMT).
2941 enum idle_type idle
= !this_rq
->nr_running
?
2942 SCHED_IDLE
: NOT_IDLE
;
2943 /* Earliest time when we have to call run_rebalance_domains again */
2944 unsigned long next_balance
= jiffies
+ 60*HZ
;
2946 for_each_domain(this_cpu
, sd
) {
2947 if (!(sd
->flags
& SD_LOAD_BALANCE
))
2950 interval
= sd
->balance_interval
;
2951 if (idle
!= SCHED_IDLE
)
2952 interval
*= sd
->busy_factor
;
2954 /* scale ms to jiffies */
2955 interval
= msecs_to_jiffies(interval
);
2956 if (unlikely(!interval
))
2959 if (sd
->flags
& SD_SERIALIZE
) {
2960 if (!spin_trylock(&balancing
))
2964 if (time_after_eq(jiffies
, sd
->last_balance
+ interval
)) {
2965 if (load_balance(this_cpu
, this_rq
, sd
, idle
, &balance
)) {
2967 * We've pulled tasks over so either we're no
2968 * longer idle, or one of our SMT siblings is
2973 sd
->last_balance
= jiffies
;
2975 if (sd
->flags
& SD_SERIALIZE
)
2976 spin_unlock(&balancing
);
2978 if (time_after(next_balance
, sd
->last_balance
+ interval
))
2979 next_balance
= sd
->last_balance
+ interval
;
2982 * Stop the load balance at this level. There is another
2983 * CPU in our sched group which is doing load balancing more
2989 this_rq
->next_balance
= next_balance
;
2993 * on UP we do not need to balance between CPUs:
2995 static inline void idle_balance(int cpu
, struct rq
*rq
)
3000 static inline void wake_priority_sleeper(struct rq
*rq
)
3002 #ifdef CONFIG_SCHED_SMT
3003 if (!rq
->nr_running
)
3006 spin_lock(&rq
->lock
);
3008 * If an SMT sibling task has been put to sleep for priority
3009 * reasons reschedule the idle task to see if it can now run.
3012 resched_task(rq
->idle
);
3013 spin_unlock(&rq
->lock
);
3017 DEFINE_PER_CPU(struct kernel_stat
, kstat
);
3019 EXPORT_PER_CPU_SYMBOL(kstat
);
3022 * This is called on clock ticks and on context switches.
3023 * Bank in p->sched_time the ns elapsed since the last tick or switch.
3026 update_cpu_clock(struct task_struct
*p
, struct rq
*rq
, unsigned long long now
)
3028 p
->sched_time
+= now
- p
->last_ran
;
3029 p
->last_ran
= rq
->most_recent_timestamp
= now
;
3033 * Return current->sched_time plus any more ns on the sched_clock
3034 * that have not yet been banked.
3036 unsigned long long current_sched_time(const struct task_struct
*p
)
3038 unsigned long long ns
;
3039 unsigned long flags
;
3041 local_irq_save(flags
);
3042 ns
= p
->sched_time
+ sched_clock() - p
->last_ran
;
3043 local_irq_restore(flags
);
3049 * We place interactive tasks back into the active array, if possible.
3051 * To guarantee that this does not starve expired tasks we ignore the
3052 * interactivity of a task if the first expired task had to wait more
3053 * than a 'reasonable' amount of time. This deadline timeout is
3054 * load-dependent, as the frequency of array switched decreases with
3055 * increasing number of running tasks. We also ignore the interactivity
3056 * if a better static_prio task has expired:
3058 static inline int expired_starving(struct rq
*rq
)
3060 if (rq
->curr
->static_prio
> rq
->best_expired_prio
)
3062 if (!STARVATION_LIMIT
|| !rq
->expired_timestamp
)
3064 if (jiffies
- rq
->expired_timestamp
> STARVATION_LIMIT
* rq
->nr_running
)
3070 * Account user cpu time to a process.
3071 * @p: the process that the cpu time gets accounted to
3072 * @hardirq_offset: the offset to subtract from hardirq_count()
3073 * @cputime: the cpu time spent in user space since the last update
3075 void account_user_time(struct task_struct
*p
, cputime_t cputime
)
3077 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3080 p
->utime
= cputime_add(p
->utime
, cputime
);
3082 /* Add user time to cpustat. */
3083 tmp
= cputime_to_cputime64(cputime
);
3084 if (TASK_NICE(p
) > 0)
3085 cpustat
->nice
= cputime64_add(cpustat
->nice
, tmp
);
3087 cpustat
->user
= cputime64_add(cpustat
->user
, tmp
);
3091 * Account system cpu time to a process.
3092 * @p: the process that the cpu time gets accounted to
3093 * @hardirq_offset: the offset to subtract from hardirq_count()
3094 * @cputime: the cpu time spent in kernel space since the last update
3096 void account_system_time(struct task_struct
*p
, int hardirq_offset
,
3099 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3100 struct rq
*rq
= this_rq();
3103 p
->stime
= cputime_add(p
->stime
, cputime
);
3105 /* Add system time to cpustat. */
3106 tmp
= cputime_to_cputime64(cputime
);
3107 if (hardirq_count() - hardirq_offset
)
3108 cpustat
->irq
= cputime64_add(cpustat
->irq
, tmp
);
3109 else if (softirq_count())
3110 cpustat
->softirq
= cputime64_add(cpustat
->softirq
, tmp
);
3111 else if (p
!= rq
->idle
)
3112 cpustat
->system
= cputime64_add(cpustat
->system
, tmp
);
3113 else if (atomic_read(&rq
->nr_iowait
) > 0)
3114 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3116 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3117 /* Account for system time used */
3118 acct_update_integrals(p
);
3122 * Account for involuntary wait time.
3123 * @p: the process from which the cpu time has been stolen
3124 * @steal: the cpu time spent in involuntary wait
3126 void account_steal_time(struct task_struct
*p
, cputime_t steal
)
3128 struct cpu_usage_stat
*cpustat
= &kstat_this_cpu
.cpustat
;
3129 cputime64_t tmp
= cputime_to_cputime64(steal
);
3130 struct rq
*rq
= this_rq();
3132 if (p
== rq
->idle
) {
3133 p
->stime
= cputime_add(p
->stime
, steal
);
3134 if (atomic_read(&rq
->nr_iowait
) > 0)
3135 cpustat
->iowait
= cputime64_add(cpustat
->iowait
, tmp
);
3137 cpustat
->idle
= cputime64_add(cpustat
->idle
, tmp
);
3139 cpustat
->steal
= cputime64_add(cpustat
->steal
, tmp
);
3142 static void task_running_tick(struct rq
*rq
, struct task_struct
*p
)
3144 if (p
->array
!= rq
->active
) {
3145 /* Task has expired but was not scheduled yet */
3146 set_tsk_need_resched(p
);
3149 spin_lock(&rq
->lock
);
3151 * The task was running during this tick - update the
3152 * time slice counter. Note: we do not update a thread's
3153 * priority until it either goes to sleep or uses up its
3154 * timeslice. This makes it possible for interactive tasks
3155 * to use up their timeslices at their highest priority levels.
3159 * RR tasks need a special form of timeslice management.
3160 * FIFO tasks have no timeslices.
3162 if ((p
->policy
== SCHED_RR
) && !--p
->time_slice
) {
3163 p
->time_slice
= task_timeslice(p
);
3164 p
->first_time_slice
= 0;
3165 set_tsk_need_resched(p
);
3167 /* put it at the end of the queue: */
3168 requeue_task(p
, rq
->active
);
3172 if (!--p
->time_slice
) {
3173 dequeue_task(p
, rq
->active
);
3174 set_tsk_need_resched(p
);
3175 p
->prio
= effective_prio(p
);
3176 p
->time_slice
= task_timeslice(p
);
3177 p
->first_time_slice
= 0;
3179 if (!rq
->expired_timestamp
)
3180 rq
->expired_timestamp
= jiffies
;
3181 if (!TASK_INTERACTIVE(p
) || expired_starving(rq
)) {
3182 enqueue_task(p
, rq
->expired
);
3183 if (p
->static_prio
< rq
->best_expired_prio
)
3184 rq
->best_expired_prio
= p
->static_prio
;
3186 enqueue_task(p
, rq
->active
);
3189 * Prevent a too long timeslice allowing a task to monopolize
3190 * the CPU. We do this by splitting up the timeslice into
3193 * Note: this does not mean the task's timeslices expire or
3194 * get lost in any way, they just might be preempted by
3195 * another task of equal priority. (one with higher
3196 * priority would have preempted this task already.) We
3197 * requeue this task to the end of the list on this priority
3198 * level, which is in essence a round-robin of tasks with
3201 * This only applies to tasks in the interactive
3202 * delta range with at least TIMESLICE_GRANULARITY to requeue.
3204 if (TASK_INTERACTIVE(p
) && !((task_timeslice(p
) -
3205 p
->time_slice
) % TIMESLICE_GRANULARITY(p
)) &&
3206 (p
->time_slice
>= TIMESLICE_GRANULARITY(p
)) &&
3207 (p
->array
== rq
->active
)) {
3209 requeue_task(p
, rq
->active
);
3210 set_tsk_need_resched(p
);
3214 spin_unlock(&rq
->lock
);
3218 * This function gets called by the timer code, with HZ frequency.
3219 * We call it with interrupts disabled.
3221 * It also gets called by the fork code, when changing the parent's
3224 void scheduler_tick(void)
3226 unsigned long long now
= sched_clock();
3227 struct task_struct
*p
= current
;
3228 int cpu
= smp_processor_id();
3229 struct rq
*rq
= cpu_rq(cpu
);
3231 update_cpu_clock(p
, rq
, now
);
3234 /* Task on the idle queue */
3235 wake_priority_sleeper(rq
);
3237 task_running_tick(rq
, p
);
3240 if (time_after_eq(jiffies
, rq
->next_balance
))
3241 raise_softirq(SCHED_SOFTIRQ
);
3245 #ifdef CONFIG_SCHED_SMT
3246 static inline void wakeup_busy_runqueue(struct rq
*rq
)
3248 /* If an SMT runqueue is sleeping due to priority reasons wake it up */
3249 if (rq
->curr
== rq
->idle
&& rq
->nr_running
)
3250 resched_task(rq
->idle
);
3254 * Called with interrupt disabled and this_rq's runqueue locked.
3256 static void wake_sleeping_dependent(int this_cpu
)
3258 struct sched_domain
*tmp
, *sd
= NULL
;
3261 for_each_domain(this_cpu
, tmp
) {
3262 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3271 for_each_cpu_mask(i
, sd
->span
) {
3272 struct rq
*smt_rq
= cpu_rq(i
);
3276 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3279 wakeup_busy_runqueue(smt_rq
);
3280 spin_unlock(&smt_rq
->lock
);
3285 * number of 'lost' timeslices this task wont be able to fully
3286 * utilize, if another task runs on a sibling. This models the
3287 * slowdown effect of other tasks running on siblings:
3289 static inline unsigned long
3290 smt_slice(struct task_struct
*p
, struct sched_domain
*sd
)
3292 return p
->time_slice
* (100 - sd
->per_cpu_gain
) / 100;
3296 * To minimise lock contention and not have to drop this_rq's runlock we only
3297 * trylock the sibling runqueues and bypass those runqueues if we fail to
3298 * acquire their lock. As we only trylock the normal locking order does not
3299 * need to be obeyed.
3302 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3304 struct sched_domain
*tmp
, *sd
= NULL
;
3307 /* kernel/rt threads do not participate in dependent sleeping */
3308 if (!p
->mm
|| rt_task(p
))
3311 for_each_domain(this_cpu
, tmp
) {
3312 if (tmp
->flags
& SD_SHARE_CPUPOWER
) {
3321 for_each_cpu_mask(i
, sd
->span
) {
3322 struct task_struct
*smt_curr
;
3329 if (unlikely(!spin_trylock(&smt_rq
->lock
)))
3332 smt_curr
= smt_rq
->curr
;
3338 * If a user task with lower static priority than the
3339 * running task on the SMT sibling is trying to schedule,
3340 * delay it till there is proportionately less timeslice
3341 * left of the sibling task to prevent a lower priority
3342 * task from using an unfair proportion of the
3343 * physical cpu's resources. -ck
3345 if (rt_task(smt_curr
)) {
3347 * With real time tasks we run non-rt tasks only
3348 * per_cpu_gain% of the time.
3350 if ((jiffies
% DEF_TIMESLICE
) >
3351 (sd
->per_cpu_gain
* DEF_TIMESLICE
/ 100))
3354 if (smt_curr
->static_prio
< p
->static_prio
&&
3355 !TASK_PREEMPTS_CURR(p
, smt_rq
) &&
3356 smt_slice(smt_curr
, sd
) > task_timeslice(p
))
3360 spin_unlock(&smt_rq
->lock
);
3365 static inline void wake_sleeping_dependent(int this_cpu
)
3369 dependent_sleeper(int this_cpu
, struct rq
*this_rq
, struct task_struct
*p
)
3375 #if defined(CONFIG_PREEMPT) && defined(CONFIG_DEBUG_PREEMPT)
3377 void fastcall
add_preempt_count(int val
)
3382 if (DEBUG_LOCKS_WARN_ON((preempt_count() < 0)))
3384 preempt_count() += val
;
3386 * Spinlock count overflowing soon?
3388 DEBUG_LOCKS_WARN_ON((preempt_count() & PREEMPT_MASK
) >=
3391 EXPORT_SYMBOL(add_preempt_count
);
3393 void fastcall
sub_preempt_count(int val
)
3398 if (DEBUG_LOCKS_WARN_ON(val
> preempt_count()))
3401 * Is the spinlock portion underflowing?
3403 if (DEBUG_LOCKS_WARN_ON((val
< PREEMPT_MASK
) &&
3404 !(preempt_count() & PREEMPT_MASK
)))
3407 preempt_count() -= val
;
3409 EXPORT_SYMBOL(sub_preempt_count
);
3413 static inline int interactive_sleep(enum sleep_type sleep_type
)
3415 return (sleep_type
== SLEEP_INTERACTIVE
||
3416 sleep_type
== SLEEP_INTERRUPTED
);
3420 * schedule() is the main scheduler function.
3422 asmlinkage
void __sched
schedule(void)
3424 struct task_struct
*prev
, *next
;
3425 struct prio_array
*array
;
3426 struct list_head
*queue
;
3427 unsigned long long now
;
3428 unsigned long run_time
;
3429 int cpu
, idx
, new_prio
;
3434 * Test if we are atomic. Since do_exit() needs to call into
3435 * schedule() atomically, we ignore that path for now.
3436 * Otherwise, whine if we are scheduling when we should not be.
3438 if (unlikely(in_atomic() && !current
->exit_state
)) {
3439 printk(KERN_ERR
"BUG: scheduling while atomic: "
3441 current
->comm
, preempt_count(), current
->pid
);
3442 debug_show_held_locks(current
);
3443 if (irqs_disabled())
3444 print_irqtrace_events(current
);
3447 profile_hit(SCHED_PROFILING
, __builtin_return_address(0));
3452 release_kernel_lock(prev
);
3453 need_resched_nonpreemptible
:
3457 * The idle thread is not allowed to schedule!
3458 * Remove this check after it has been exercised a bit.
3460 if (unlikely(prev
== rq
->idle
) && prev
->state
!= TASK_RUNNING
) {
3461 printk(KERN_ERR
"bad: scheduling from the idle thread!\n");
3465 schedstat_inc(rq
, sched_cnt
);
3466 now
= sched_clock();
3467 if (likely((long long)(now
- prev
->timestamp
) < NS_MAX_SLEEP_AVG
)) {
3468 run_time
= now
- prev
->timestamp
;
3469 if (unlikely((long long)(now
- prev
->timestamp
) < 0))
3472 run_time
= NS_MAX_SLEEP_AVG
;
3475 * Tasks charged proportionately less run_time at high sleep_avg to
3476 * delay them losing their interactive status
3478 run_time
/= (CURRENT_BONUS(prev
) ? : 1);
3480 spin_lock_irq(&rq
->lock
);
3482 switch_count
= &prev
->nivcsw
;
3483 if (prev
->state
&& !(preempt_count() & PREEMPT_ACTIVE
)) {
3484 switch_count
= &prev
->nvcsw
;
3485 if (unlikely((prev
->state
& TASK_INTERRUPTIBLE
) &&
3486 unlikely(signal_pending(prev
))))
3487 prev
->state
= TASK_RUNNING
;
3489 if (prev
->state
== TASK_UNINTERRUPTIBLE
)
3490 rq
->nr_uninterruptible
++;
3491 deactivate_task(prev
, rq
);
3495 cpu
= smp_processor_id();
3496 if (unlikely(!rq
->nr_running
)) {
3497 idle_balance(cpu
, rq
);
3498 if (!rq
->nr_running
) {
3500 rq
->expired_timestamp
= 0;
3501 wake_sleeping_dependent(cpu
);
3507 if (unlikely(!array
->nr_active
)) {
3509 * Switch the active and expired arrays.
3511 schedstat_inc(rq
, sched_switch
);
3512 rq
->active
= rq
->expired
;
3513 rq
->expired
= array
;
3515 rq
->expired_timestamp
= 0;
3516 rq
->best_expired_prio
= MAX_PRIO
;
3519 idx
= sched_find_first_bit(array
->bitmap
);
3520 queue
= array
->queue
+ idx
;
3521 next
= list_entry(queue
->next
, struct task_struct
, run_list
);
3523 if (!rt_task(next
) && interactive_sleep(next
->sleep_type
)) {
3524 unsigned long long delta
= now
- next
->timestamp
;
3525 if (unlikely((long long)(now
- next
->timestamp
) < 0))
3528 if (next
->sleep_type
== SLEEP_INTERACTIVE
)
3529 delta
= delta
* (ON_RUNQUEUE_WEIGHT
* 128 / 100) / 128;
3531 array
= next
->array
;
3532 new_prio
= recalc_task_prio(next
, next
->timestamp
+ delta
);
3534 if (unlikely(next
->prio
!= new_prio
)) {
3535 dequeue_task(next
, array
);
3536 next
->prio
= new_prio
;
3537 enqueue_task(next
, array
);
3540 next
->sleep_type
= SLEEP_NORMAL
;
3541 if (dependent_sleeper(cpu
, rq
, next
))
3544 if (next
== rq
->idle
)
3545 schedstat_inc(rq
, sched_goidle
);
3547 prefetch_stack(next
);
3548 clear_tsk_need_resched(prev
);
3549 rcu_qsctr_inc(task_cpu(prev
));
3551 update_cpu_clock(prev
, rq
, now
);
3553 prev
->sleep_avg
-= run_time
;
3554 if ((long)prev
->sleep_avg
<= 0)
3555 prev
->sleep_avg
= 0;
3556 prev
->timestamp
= prev
->last_ran
= now
;
3558 sched_info_switch(prev
, next
);
3559 if (likely(prev
!= next
)) {
3560 next
->timestamp
= now
;
3565 prepare_task_switch(rq
, next
);
3566 prev
= context_switch(rq
, prev
, next
);
3569 * this_rq must be evaluated again because prev may have moved
3570 * CPUs since it called schedule(), thus the 'rq' on its stack
3571 * frame will be invalid.
3573 finish_task_switch(this_rq(), prev
);
3575 spin_unlock_irq(&rq
->lock
);
3578 if (unlikely(reacquire_kernel_lock(prev
) < 0))
3579 goto need_resched_nonpreemptible
;
3580 preempt_enable_no_resched();
3581 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3584 EXPORT_SYMBOL(schedule
);
3586 #ifdef CONFIG_PREEMPT
3588 * this is the entry point to schedule() from in-kernel preemption
3589 * off of preempt_enable. Kernel preemptions off return from interrupt
3590 * occur there and call schedule directly.
3592 asmlinkage
void __sched
preempt_schedule(void)
3594 struct thread_info
*ti
= current_thread_info();
3595 #ifdef CONFIG_PREEMPT_BKL
3596 struct task_struct
*task
= current
;
3597 int saved_lock_depth
;
3600 * If there is a non-zero preempt_count or interrupts are disabled,
3601 * we do not want to preempt the current task. Just return..
3603 if (likely(ti
->preempt_count
|| irqs_disabled()))
3607 add_preempt_count(PREEMPT_ACTIVE
);
3609 * We keep the big kernel semaphore locked, but we
3610 * clear ->lock_depth so that schedule() doesnt
3611 * auto-release the semaphore:
3613 #ifdef CONFIG_PREEMPT_BKL
3614 saved_lock_depth
= task
->lock_depth
;
3615 task
->lock_depth
= -1;
3618 #ifdef CONFIG_PREEMPT_BKL
3619 task
->lock_depth
= saved_lock_depth
;
3621 sub_preempt_count(PREEMPT_ACTIVE
);
3623 /* we could miss a preemption opportunity between schedule and now */
3625 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3628 EXPORT_SYMBOL(preempt_schedule
);
3631 * this is the entry point to schedule() from kernel preemption
3632 * off of irq context.
3633 * Note, that this is called and return with irqs disabled. This will
3634 * protect us against recursive calling from irq.
3636 asmlinkage
void __sched
preempt_schedule_irq(void)
3638 struct thread_info
*ti
= current_thread_info();
3639 #ifdef CONFIG_PREEMPT_BKL
3640 struct task_struct
*task
= current
;
3641 int saved_lock_depth
;
3643 /* Catch callers which need to be fixed */
3644 BUG_ON(ti
->preempt_count
|| !irqs_disabled());
3647 add_preempt_count(PREEMPT_ACTIVE
);
3649 * We keep the big kernel semaphore locked, but we
3650 * clear ->lock_depth so that schedule() doesnt
3651 * auto-release the semaphore:
3653 #ifdef CONFIG_PREEMPT_BKL
3654 saved_lock_depth
= task
->lock_depth
;
3655 task
->lock_depth
= -1;
3659 local_irq_disable();
3660 #ifdef CONFIG_PREEMPT_BKL
3661 task
->lock_depth
= saved_lock_depth
;
3663 sub_preempt_count(PREEMPT_ACTIVE
);
3665 /* we could miss a preemption opportunity between schedule and now */
3667 if (unlikely(test_thread_flag(TIF_NEED_RESCHED
)))
3671 #endif /* CONFIG_PREEMPT */
3673 int default_wake_function(wait_queue_t
*curr
, unsigned mode
, int sync
,
3676 return try_to_wake_up(curr
->private, mode
, sync
);
3678 EXPORT_SYMBOL(default_wake_function
);
3681 * The core wakeup function. Non-exclusive wakeups (nr_exclusive == 0) just
3682 * wake everything up. If it's an exclusive wakeup (nr_exclusive == small +ve
3683 * number) then we wake all the non-exclusive tasks and one exclusive task.
3685 * There are circumstances in which we can try to wake a task which has already
3686 * started to run but is not in state TASK_RUNNING. try_to_wake_up() returns
3687 * zero in this (rare) case, and we handle it by continuing to scan the queue.
3689 static void __wake_up_common(wait_queue_head_t
*q
, unsigned int mode
,
3690 int nr_exclusive
, int sync
, void *key
)
3692 struct list_head
*tmp
, *next
;
3694 list_for_each_safe(tmp
, next
, &q
->task_list
) {
3695 wait_queue_t
*curr
= list_entry(tmp
, wait_queue_t
, task_list
);
3696 unsigned flags
= curr
->flags
;
3698 if (curr
->func(curr
, mode
, sync
, key
) &&
3699 (flags
& WQ_FLAG_EXCLUSIVE
) && !--nr_exclusive
)
3705 * __wake_up - wake up threads blocked on a waitqueue.
3707 * @mode: which threads
3708 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3709 * @key: is directly passed to the wakeup function
3711 void fastcall
__wake_up(wait_queue_head_t
*q
, unsigned int mode
,
3712 int nr_exclusive
, void *key
)
3714 unsigned long flags
;
3716 spin_lock_irqsave(&q
->lock
, flags
);
3717 __wake_up_common(q
, mode
, nr_exclusive
, 0, key
);
3718 spin_unlock_irqrestore(&q
->lock
, flags
);
3720 EXPORT_SYMBOL(__wake_up
);
3723 * Same as __wake_up but called with the spinlock in wait_queue_head_t held.
3725 void fastcall
__wake_up_locked(wait_queue_head_t
*q
, unsigned int mode
)
3727 __wake_up_common(q
, mode
, 1, 0, NULL
);
3731 * __wake_up_sync - wake up threads blocked on a waitqueue.
3733 * @mode: which threads
3734 * @nr_exclusive: how many wake-one or wake-many threads to wake up
3736 * The sync wakeup differs that the waker knows that it will schedule
3737 * away soon, so while the target thread will be woken up, it will not
3738 * be migrated to another CPU - ie. the two threads are 'synchronized'
3739 * with each other. This can prevent needless bouncing between CPUs.
3741 * On UP it can prevent extra preemption.
3744 __wake_up_sync(wait_queue_head_t
*q
, unsigned int mode
, int nr_exclusive
)
3746 unsigned long flags
;
3752 if (unlikely(!nr_exclusive
))
3755 spin_lock_irqsave(&q
->lock
, flags
);
3756 __wake_up_common(q
, mode
, nr_exclusive
, sync
, NULL
);
3757 spin_unlock_irqrestore(&q
->lock
, flags
);
3759 EXPORT_SYMBOL_GPL(__wake_up_sync
); /* For internal use only */
3761 void fastcall
complete(struct completion
*x
)
3763 unsigned long flags
;
3765 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3767 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3769 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3771 EXPORT_SYMBOL(complete
);
3773 void fastcall
complete_all(struct completion
*x
)
3775 unsigned long flags
;
3777 spin_lock_irqsave(&x
->wait
.lock
, flags
);
3778 x
->done
+= UINT_MAX
/2;
3779 __wake_up_common(&x
->wait
, TASK_UNINTERRUPTIBLE
| TASK_INTERRUPTIBLE
,
3781 spin_unlock_irqrestore(&x
->wait
.lock
, flags
);
3783 EXPORT_SYMBOL(complete_all
);
3785 void fastcall __sched
wait_for_completion(struct completion
*x
)
3789 spin_lock_irq(&x
->wait
.lock
);
3791 DECLARE_WAITQUEUE(wait
, current
);
3793 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3794 __add_wait_queue_tail(&x
->wait
, &wait
);
3796 __set_current_state(TASK_UNINTERRUPTIBLE
);
3797 spin_unlock_irq(&x
->wait
.lock
);
3799 spin_lock_irq(&x
->wait
.lock
);
3801 __remove_wait_queue(&x
->wait
, &wait
);
3804 spin_unlock_irq(&x
->wait
.lock
);
3806 EXPORT_SYMBOL(wait_for_completion
);
3808 unsigned long fastcall __sched
3809 wait_for_completion_timeout(struct completion
*x
, unsigned long timeout
)
3813 spin_lock_irq(&x
->wait
.lock
);
3815 DECLARE_WAITQUEUE(wait
, current
);
3817 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3818 __add_wait_queue_tail(&x
->wait
, &wait
);
3820 __set_current_state(TASK_UNINTERRUPTIBLE
);
3821 spin_unlock_irq(&x
->wait
.lock
);
3822 timeout
= schedule_timeout(timeout
);
3823 spin_lock_irq(&x
->wait
.lock
);
3825 __remove_wait_queue(&x
->wait
, &wait
);
3829 __remove_wait_queue(&x
->wait
, &wait
);
3833 spin_unlock_irq(&x
->wait
.lock
);
3836 EXPORT_SYMBOL(wait_for_completion_timeout
);
3838 int fastcall __sched
wait_for_completion_interruptible(struct completion
*x
)
3844 spin_lock_irq(&x
->wait
.lock
);
3846 DECLARE_WAITQUEUE(wait
, current
);
3848 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3849 __add_wait_queue_tail(&x
->wait
, &wait
);
3851 if (signal_pending(current
)) {
3853 __remove_wait_queue(&x
->wait
, &wait
);
3856 __set_current_state(TASK_INTERRUPTIBLE
);
3857 spin_unlock_irq(&x
->wait
.lock
);
3859 spin_lock_irq(&x
->wait
.lock
);
3861 __remove_wait_queue(&x
->wait
, &wait
);
3865 spin_unlock_irq(&x
->wait
.lock
);
3869 EXPORT_SYMBOL(wait_for_completion_interruptible
);
3871 unsigned long fastcall __sched
3872 wait_for_completion_interruptible_timeout(struct completion
*x
,
3873 unsigned long timeout
)
3877 spin_lock_irq(&x
->wait
.lock
);
3879 DECLARE_WAITQUEUE(wait
, current
);
3881 wait
.flags
|= WQ_FLAG_EXCLUSIVE
;
3882 __add_wait_queue_tail(&x
->wait
, &wait
);
3884 if (signal_pending(current
)) {
3885 timeout
= -ERESTARTSYS
;
3886 __remove_wait_queue(&x
->wait
, &wait
);
3889 __set_current_state(TASK_INTERRUPTIBLE
);
3890 spin_unlock_irq(&x
->wait
.lock
);
3891 timeout
= schedule_timeout(timeout
);
3892 spin_lock_irq(&x
->wait
.lock
);
3894 __remove_wait_queue(&x
->wait
, &wait
);
3898 __remove_wait_queue(&x
->wait
, &wait
);
3902 spin_unlock_irq(&x
->wait
.lock
);
3905 EXPORT_SYMBOL(wait_for_completion_interruptible_timeout
);
3908 #define SLEEP_ON_VAR \
3909 unsigned long flags; \
3910 wait_queue_t wait; \
3911 init_waitqueue_entry(&wait, current);
3913 #define SLEEP_ON_HEAD \
3914 spin_lock_irqsave(&q->lock,flags); \
3915 __add_wait_queue(q, &wait); \
3916 spin_unlock(&q->lock);
3918 #define SLEEP_ON_TAIL \
3919 spin_lock_irq(&q->lock); \
3920 __remove_wait_queue(q, &wait); \
3921 spin_unlock_irqrestore(&q->lock, flags);
3923 void fastcall __sched
interruptible_sleep_on(wait_queue_head_t
*q
)
3927 current
->state
= TASK_INTERRUPTIBLE
;
3933 EXPORT_SYMBOL(interruptible_sleep_on
);
3935 long fastcall __sched
3936 interruptible_sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3940 current
->state
= TASK_INTERRUPTIBLE
;
3943 timeout
= schedule_timeout(timeout
);
3948 EXPORT_SYMBOL(interruptible_sleep_on_timeout
);
3950 void fastcall __sched
sleep_on(wait_queue_head_t
*q
)
3954 current
->state
= TASK_UNINTERRUPTIBLE
;
3960 EXPORT_SYMBOL(sleep_on
);
3962 long fastcall __sched
sleep_on_timeout(wait_queue_head_t
*q
, long timeout
)
3966 current
->state
= TASK_UNINTERRUPTIBLE
;
3969 timeout
= schedule_timeout(timeout
);
3975 EXPORT_SYMBOL(sleep_on_timeout
);
3977 #ifdef CONFIG_RT_MUTEXES
3980 * rt_mutex_setprio - set the current priority of a task
3982 * @prio: prio value (kernel-internal form)
3984 * This function changes the 'effective' priority of a task. It does
3985 * not touch ->normal_prio like __setscheduler().
3987 * Used by the rt_mutex code to implement priority inheritance logic.
3989 void rt_mutex_setprio(struct task_struct
*p
, int prio
)
3991 struct prio_array
*array
;
3992 unsigned long flags
;
3996 BUG_ON(prio
< 0 || prio
> MAX_PRIO
);
3998 rq
= task_rq_lock(p
, &flags
);
4003 dequeue_task(p
, array
);
4008 * If changing to an RT priority then queue it
4009 * in the active array!
4013 enqueue_task(p
, array
);
4015 * Reschedule if we are currently running on this runqueue and
4016 * our priority decreased, or if we are not currently running on
4017 * this runqueue and our priority is higher than the current's
4019 if (task_running(rq
, p
)) {
4020 if (p
->prio
> oldprio
)
4021 resched_task(rq
->curr
);
4022 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4023 resched_task(rq
->curr
);
4025 task_rq_unlock(rq
, &flags
);
4030 void set_user_nice(struct task_struct
*p
, long nice
)
4032 struct prio_array
*array
;
4033 int old_prio
, delta
;
4034 unsigned long flags
;
4037 if (TASK_NICE(p
) == nice
|| nice
< -20 || nice
> 19)
4040 * We have to be careful, if called from sys_setpriority(),
4041 * the task might be in the middle of scheduling on another CPU.
4043 rq
= task_rq_lock(p
, &flags
);
4045 * The RT priorities are set via sched_setscheduler(), but we still
4046 * allow the 'normal' nice value to be set - but as expected
4047 * it wont have any effect on scheduling until the task is
4048 * not SCHED_NORMAL/SCHED_BATCH:
4050 if (has_rt_policy(p
)) {
4051 p
->static_prio
= NICE_TO_PRIO(nice
);
4056 dequeue_task(p
, array
);
4057 dec_raw_weighted_load(rq
, p
);
4060 p
->static_prio
= NICE_TO_PRIO(nice
);
4063 p
->prio
= effective_prio(p
);
4064 delta
= p
->prio
- old_prio
;
4067 enqueue_task(p
, array
);
4068 inc_raw_weighted_load(rq
, p
);
4070 * If the task increased its priority or is running and
4071 * lowered its priority, then reschedule its CPU:
4073 if (delta
< 0 || (delta
> 0 && task_running(rq
, p
)))
4074 resched_task(rq
->curr
);
4077 task_rq_unlock(rq
, &flags
);
4079 EXPORT_SYMBOL(set_user_nice
);
4082 * can_nice - check if a task can reduce its nice value
4086 int can_nice(const struct task_struct
*p
, const int nice
)
4088 /* convert nice value [19,-20] to rlimit style value [1,40] */
4089 int nice_rlim
= 20 - nice
;
4091 return (nice_rlim
<= p
->signal
->rlim
[RLIMIT_NICE
].rlim_cur
||
4092 capable(CAP_SYS_NICE
));
4095 #ifdef __ARCH_WANT_SYS_NICE
4098 * sys_nice - change the priority of the current process.
4099 * @increment: priority increment
4101 * sys_setpriority is a more generic, but much slower function that
4102 * does similar things.
4104 asmlinkage
long sys_nice(int increment
)
4109 * Setpriority might change our priority at the same moment.
4110 * We don't have to worry. Conceptually one call occurs first
4111 * and we have a single winner.
4113 if (increment
< -40)
4118 nice
= PRIO_TO_NICE(current
->static_prio
) + increment
;
4124 if (increment
< 0 && !can_nice(current
, nice
))
4127 retval
= security_task_setnice(current
, nice
);
4131 set_user_nice(current
, nice
);
4138 * task_prio - return the priority value of a given task.
4139 * @p: the task in question.
4141 * This is the priority value as seen by users in /proc.
4142 * RT tasks are offset by -200. Normal tasks are centered
4143 * around 0, value goes from -16 to +15.
4145 int task_prio(const struct task_struct
*p
)
4147 return p
->prio
- MAX_RT_PRIO
;
4151 * task_nice - return the nice value of a given task.
4152 * @p: the task in question.
4154 int task_nice(const struct task_struct
*p
)
4156 return TASK_NICE(p
);
4158 EXPORT_SYMBOL_GPL(task_nice
);
4161 * idle_cpu - is a given cpu idle currently?
4162 * @cpu: the processor in question.
4164 int idle_cpu(int cpu
)
4166 return cpu_curr(cpu
) == cpu_rq(cpu
)->idle
;
4170 * idle_task - return the idle task for a given cpu.
4171 * @cpu: the processor in question.
4173 struct task_struct
*idle_task(int cpu
)
4175 return cpu_rq(cpu
)->idle
;
4179 * find_process_by_pid - find a process with a matching PID value.
4180 * @pid: the pid in question.
4182 static inline struct task_struct
*find_process_by_pid(pid_t pid
)
4184 return pid
? find_task_by_pid(pid
) : current
;
4187 /* Actually do priority change: must hold rq lock. */
4188 static void __setscheduler(struct task_struct
*p
, int policy
, int prio
)
4193 p
->rt_priority
= prio
;
4194 p
->normal_prio
= normal_prio(p
);
4195 /* we are holding p->pi_lock already */
4196 p
->prio
= rt_mutex_getprio(p
);
4198 * SCHED_BATCH tasks are treated as perpetual CPU hogs:
4200 if (policy
== SCHED_BATCH
)
4206 * sched_setscheduler - change the scheduling policy and/or RT priority of a thread.
4207 * @p: the task in question.
4208 * @policy: new policy.
4209 * @param: structure containing the new RT priority.
4211 * NOTE that the task may be already dead.
4213 int sched_setscheduler(struct task_struct
*p
, int policy
,
4214 struct sched_param
*param
)
4216 int retval
, oldprio
, oldpolicy
= -1;
4217 struct prio_array
*array
;
4218 unsigned long flags
;
4221 /* may grab non-irq protected spin_locks */
4222 BUG_ON(in_interrupt());
4224 /* double check policy once rq lock held */
4226 policy
= oldpolicy
= p
->policy
;
4227 else if (policy
!= SCHED_FIFO
&& policy
!= SCHED_RR
&&
4228 policy
!= SCHED_NORMAL
&& policy
!= SCHED_BATCH
)
4231 * Valid priorities for SCHED_FIFO and SCHED_RR are
4232 * 1..MAX_USER_RT_PRIO-1, valid priority for SCHED_NORMAL and
4235 if (param
->sched_priority
< 0 ||
4236 (p
->mm
&& param
->sched_priority
> MAX_USER_RT_PRIO
-1) ||
4237 (!p
->mm
&& param
->sched_priority
> MAX_RT_PRIO
-1))
4239 if (is_rt_policy(policy
) != (param
->sched_priority
!= 0))
4243 * Allow unprivileged RT tasks to decrease priority:
4245 if (!capable(CAP_SYS_NICE
)) {
4246 if (is_rt_policy(policy
)) {
4247 unsigned long rlim_rtprio
;
4248 unsigned long flags
;
4250 if (!lock_task_sighand(p
, &flags
))
4252 rlim_rtprio
= p
->signal
->rlim
[RLIMIT_RTPRIO
].rlim_cur
;
4253 unlock_task_sighand(p
, &flags
);
4255 /* can't set/change the rt policy */
4256 if (policy
!= p
->policy
&& !rlim_rtprio
)
4259 /* can't increase priority */
4260 if (param
->sched_priority
> p
->rt_priority
&&
4261 param
->sched_priority
> rlim_rtprio
)
4265 /* can't change other user's priorities */
4266 if ((current
->euid
!= p
->euid
) &&
4267 (current
->euid
!= p
->uid
))
4271 retval
= security_task_setscheduler(p
, policy
, param
);
4275 * make sure no PI-waiters arrive (or leave) while we are
4276 * changing the priority of the task:
4278 spin_lock_irqsave(&p
->pi_lock
, flags
);
4280 * To be able to change p->policy safely, the apropriate
4281 * runqueue lock must be held.
4283 rq
= __task_rq_lock(p
);
4284 /* recheck policy now with rq lock held */
4285 if (unlikely(oldpolicy
!= -1 && oldpolicy
!= p
->policy
)) {
4286 policy
= oldpolicy
= -1;
4287 __task_rq_unlock(rq
);
4288 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4293 deactivate_task(p
, rq
);
4295 __setscheduler(p
, policy
, param
->sched_priority
);
4297 __activate_task(p
, rq
);
4299 * Reschedule if we are currently running on this runqueue and
4300 * our priority decreased, or if we are not currently running on
4301 * this runqueue and our priority is higher than the current's
4303 if (task_running(rq
, p
)) {
4304 if (p
->prio
> oldprio
)
4305 resched_task(rq
->curr
);
4306 } else if (TASK_PREEMPTS_CURR(p
, rq
))
4307 resched_task(rq
->curr
);
4309 __task_rq_unlock(rq
);
4310 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
4312 rt_mutex_adjust_pi(p
);
4316 EXPORT_SYMBOL_GPL(sched_setscheduler
);
4319 do_sched_setscheduler(pid_t pid
, int policy
, struct sched_param __user
*param
)
4321 struct sched_param lparam
;
4322 struct task_struct
*p
;
4325 if (!param
|| pid
< 0)
4327 if (copy_from_user(&lparam
, param
, sizeof(struct sched_param
)))
4332 p
= find_process_by_pid(pid
);
4334 retval
= sched_setscheduler(p
, policy
, &lparam
);
4341 * sys_sched_setscheduler - set/change the scheduler policy and RT priority
4342 * @pid: the pid in question.
4343 * @policy: new policy.
4344 * @param: structure containing the new RT priority.
4346 asmlinkage
long sys_sched_setscheduler(pid_t pid
, int policy
,
4347 struct sched_param __user
*param
)
4349 /* negative values for policy are not valid */
4353 return do_sched_setscheduler(pid
, policy
, param
);
4357 * sys_sched_setparam - set/change the RT priority of a thread
4358 * @pid: the pid in question.
4359 * @param: structure containing the new RT priority.
4361 asmlinkage
long sys_sched_setparam(pid_t pid
, struct sched_param __user
*param
)
4363 return do_sched_setscheduler(pid
, -1, param
);
4367 * sys_sched_getscheduler - get the policy (scheduling class) of a thread
4368 * @pid: the pid in question.
4370 asmlinkage
long sys_sched_getscheduler(pid_t pid
)
4372 struct task_struct
*p
;
4373 int retval
= -EINVAL
;
4379 read_lock(&tasklist_lock
);
4380 p
= find_process_by_pid(pid
);
4382 retval
= security_task_getscheduler(p
);
4386 read_unlock(&tasklist_lock
);
4393 * sys_sched_getscheduler - get the RT priority of a thread
4394 * @pid: the pid in question.
4395 * @param: structure containing the RT priority.
4397 asmlinkage
long sys_sched_getparam(pid_t pid
, struct sched_param __user
*param
)
4399 struct sched_param lp
;
4400 struct task_struct
*p
;
4401 int retval
= -EINVAL
;
4403 if (!param
|| pid
< 0)
4406 read_lock(&tasklist_lock
);
4407 p
= find_process_by_pid(pid
);
4412 retval
= security_task_getscheduler(p
);
4416 lp
.sched_priority
= p
->rt_priority
;
4417 read_unlock(&tasklist_lock
);
4420 * This one might sleep, we cannot do it with a spinlock held ...
4422 retval
= copy_to_user(param
, &lp
, sizeof(*param
)) ? -EFAULT
: 0;
4428 read_unlock(&tasklist_lock
);
4432 long sched_setaffinity(pid_t pid
, cpumask_t new_mask
)
4434 cpumask_t cpus_allowed
;
4435 struct task_struct
*p
;
4439 read_lock(&tasklist_lock
);
4441 p
= find_process_by_pid(pid
);
4443 read_unlock(&tasklist_lock
);
4444 unlock_cpu_hotplug();
4449 * It is not safe to call set_cpus_allowed with the
4450 * tasklist_lock held. We will bump the task_struct's
4451 * usage count and then drop tasklist_lock.
4454 read_unlock(&tasklist_lock
);
4457 if ((current
->euid
!= p
->euid
) && (current
->euid
!= p
->uid
) &&
4458 !capable(CAP_SYS_NICE
))
4461 retval
= security_task_setscheduler(p
, 0, NULL
);
4465 cpus_allowed
= cpuset_cpus_allowed(p
);
4466 cpus_and(new_mask
, new_mask
, cpus_allowed
);
4467 retval
= set_cpus_allowed(p
, new_mask
);
4471 unlock_cpu_hotplug();
4475 static int get_user_cpu_mask(unsigned long __user
*user_mask_ptr
, unsigned len
,
4476 cpumask_t
*new_mask
)
4478 if (len
< sizeof(cpumask_t
)) {
4479 memset(new_mask
, 0, sizeof(cpumask_t
));
4480 } else if (len
> sizeof(cpumask_t
)) {
4481 len
= sizeof(cpumask_t
);
4483 return copy_from_user(new_mask
, user_mask_ptr
, len
) ? -EFAULT
: 0;
4487 * sys_sched_setaffinity - set the cpu affinity of a process
4488 * @pid: pid of the process
4489 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4490 * @user_mask_ptr: user-space pointer to the new cpu mask
4492 asmlinkage
long sys_sched_setaffinity(pid_t pid
, unsigned int len
,
4493 unsigned long __user
*user_mask_ptr
)
4498 retval
= get_user_cpu_mask(user_mask_ptr
, len
, &new_mask
);
4502 return sched_setaffinity(pid
, new_mask
);
4506 * Represents all cpu's present in the system
4507 * In systems capable of hotplug, this map could dynamically grow
4508 * as new cpu's are detected in the system via any platform specific
4509 * method, such as ACPI for e.g.
4512 cpumask_t cpu_present_map __read_mostly
;
4513 EXPORT_SYMBOL(cpu_present_map
);
4516 cpumask_t cpu_online_map __read_mostly
= CPU_MASK_ALL
;
4517 EXPORT_SYMBOL(cpu_online_map
);
4519 cpumask_t cpu_possible_map __read_mostly
= CPU_MASK_ALL
;
4520 EXPORT_SYMBOL(cpu_possible_map
);
4523 long sched_getaffinity(pid_t pid
, cpumask_t
*mask
)
4525 struct task_struct
*p
;
4529 read_lock(&tasklist_lock
);
4532 p
= find_process_by_pid(pid
);
4536 retval
= security_task_getscheduler(p
);
4540 cpus_and(*mask
, p
->cpus_allowed
, cpu_online_map
);
4543 read_unlock(&tasklist_lock
);
4544 unlock_cpu_hotplug();
4552 * sys_sched_getaffinity - get the cpu affinity of a process
4553 * @pid: pid of the process
4554 * @len: length in bytes of the bitmask pointed to by user_mask_ptr
4555 * @user_mask_ptr: user-space pointer to hold the current cpu mask
4557 asmlinkage
long sys_sched_getaffinity(pid_t pid
, unsigned int len
,
4558 unsigned long __user
*user_mask_ptr
)
4563 if (len
< sizeof(cpumask_t
))
4566 ret
= sched_getaffinity(pid
, &mask
);
4570 if (copy_to_user(user_mask_ptr
, &mask
, sizeof(cpumask_t
)))
4573 return sizeof(cpumask_t
);
4577 * sys_sched_yield - yield the current processor to other threads.
4579 * This function yields the current CPU by moving the calling thread
4580 * to the expired array. If there are no other threads running on this
4581 * CPU then this function will return.
4583 asmlinkage
long sys_sched_yield(void)
4585 struct rq
*rq
= this_rq_lock();
4586 struct prio_array
*array
= current
->array
, *target
= rq
->expired
;
4588 schedstat_inc(rq
, yld_cnt
);
4590 * We implement yielding by moving the task into the expired
4593 * (special rule: RT tasks will just roundrobin in the active
4596 if (rt_task(current
))
4597 target
= rq
->active
;
4599 if (array
->nr_active
== 1) {
4600 schedstat_inc(rq
, yld_act_empty
);
4601 if (!rq
->expired
->nr_active
)
4602 schedstat_inc(rq
, yld_both_empty
);
4603 } else if (!rq
->expired
->nr_active
)
4604 schedstat_inc(rq
, yld_exp_empty
);
4606 if (array
!= target
) {
4607 dequeue_task(current
, array
);
4608 enqueue_task(current
, target
);
4611 * requeue_task is cheaper so perform that if possible.
4613 requeue_task(current
, array
);
4616 * Since we are going to call schedule() anyway, there's
4617 * no need to preempt or enable interrupts:
4619 __release(rq
->lock
);
4620 spin_release(&rq
->lock
.dep_map
, 1, _THIS_IP_
);
4621 _raw_spin_unlock(&rq
->lock
);
4622 preempt_enable_no_resched();
4629 static void __cond_resched(void)
4631 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
4632 __might_sleep(__FILE__
, __LINE__
);
4635 * The BKS might be reacquired before we have dropped
4636 * PREEMPT_ACTIVE, which could trigger a second
4637 * cond_resched() call.
4640 add_preempt_count(PREEMPT_ACTIVE
);
4642 sub_preempt_count(PREEMPT_ACTIVE
);
4643 } while (need_resched());
4646 int __sched
cond_resched(void)
4648 if (need_resched() && !(preempt_count() & PREEMPT_ACTIVE
) &&
4649 system_state
== SYSTEM_RUNNING
) {
4655 EXPORT_SYMBOL(cond_resched
);
4658 * cond_resched_lock() - if a reschedule is pending, drop the given lock,
4659 * call schedule, and on return reacquire the lock.
4661 * This works OK both with and without CONFIG_PREEMPT. We do strange low-level
4662 * operations here to prevent schedule() from being called twice (once via
4663 * spin_unlock(), once by hand).
4665 int cond_resched_lock(spinlock_t
*lock
)
4669 if (need_lockbreak(lock
)) {
4675 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4676 spin_release(&lock
->dep_map
, 1, _THIS_IP_
);
4677 _raw_spin_unlock(lock
);
4678 preempt_enable_no_resched();
4685 EXPORT_SYMBOL(cond_resched_lock
);
4687 int __sched
cond_resched_softirq(void)
4689 BUG_ON(!in_softirq());
4691 if (need_resched() && system_state
== SYSTEM_RUNNING
) {
4692 raw_local_irq_disable();
4694 raw_local_irq_enable();
4701 EXPORT_SYMBOL(cond_resched_softirq
);
4704 * yield - yield the current processor to other threads.
4706 * This is a shortcut for kernel-space yielding - it marks the
4707 * thread runnable and calls sys_sched_yield().
4709 void __sched
yield(void)
4711 set_current_state(TASK_RUNNING
);
4714 EXPORT_SYMBOL(yield
);
4717 * This task is about to go to sleep on IO. Increment rq->nr_iowait so
4718 * that process accounting knows that this is a task in IO wait state.
4720 * But don't do that if it is a deliberate, throttling IO wait (this task
4721 * has set its backing_dev_info: the queue against which it should throttle)
4723 void __sched
io_schedule(void)
4725 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4727 delayacct_blkio_start();
4728 atomic_inc(&rq
->nr_iowait
);
4730 atomic_dec(&rq
->nr_iowait
);
4731 delayacct_blkio_end();
4733 EXPORT_SYMBOL(io_schedule
);
4735 long __sched
io_schedule_timeout(long timeout
)
4737 struct rq
*rq
= &__raw_get_cpu_var(runqueues
);
4740 delayacct_blkio_start();
4741 atomic_inc(&rq
->nr_iowait
);
4742 ret
= schedule_timeout(timeout
);
4743 atomic_dec(&rq
->nr_iowait
);
4744 delayacct_blkio_end();
4749 * sys_sched_get_priority_max - return maximum RT priority.
4750 * @policy: scheduling class.
4752 * this syscall returns the maximum rt_priority that can be used
4753 * by a given scheduling class.
4755 asmlinkage
long sys_sched_get_priority_max(int policy
)
4762 ret
= MAX_USER_RT_PRIO
-1;
4773 * sys_sched_get_priority_min - return minimum RT priority.
4774 * @policy: scheduling class.
4776 * this syscall returns the minimum rt_priority that can be used
4777 * by a given scheduling class.
4779 asmlinkage
long sys_sched_get_priority_min(int policy
)
4796 * sys_sched_rr_get_interval - return the default timeslice of a process.
4797 * @pid: pid of the process.
4798 * @interval: userspace pointer to the timeslice value.
4800 * this syscall writes the default timeslice value of a given process
4801 * into the user-space timespec buffer. A value of '0' means infinity.
4804 long sys_sched_rr_get_interval(pid_t pid
, struct timespec __user
*interval
)
4806 struct task_struct
*p
;
4807 int retval
= -EINVAL
;
4814 read_lock(&tasklist_lock
);
4815 p
= find_process_by_pid(pid
);
4819 retval
= security_task_getscheduler(p
);
4823 jiffies_to_timespec(p
->policy
== SCHED_FIFO
?
4824 0 : task_timeslice(p
), &t
);
4825 read_unlock(&tasklist_lock
);
4826 retval
= copy_to_user(interval
, &t
, sizeof(t
)) ? -EFAULT
: 0;
4830 read_unlock(&tasklist_lock
);
4834 static inline struct task_struct
*eldest_child(struct task_struct
*p
)
4836 if (list_empty(&p
->children
))
4838 return list_entry(p
->children
.next
,struct task_struct
,sibling
);
4841 static inline struct task_struct
*older_sibling(struct task_struct
*p
)
4843 if (p
->sibling
.prev
==&p
->parent
->children
)
4845 return list_entry(p
->sibling
.prev
,struct task_struct
,sibling
);
4848 static inline struct task_struct
*younger_sibling(struct task_struct
*p
)
4850 if (p
->sibling
.next
==&p
->parent
->children
)
4852 return list_entry(p
->sibling
.next
,struct task_struct
,sibling
);
4855 static const char stat_nam
[] = "RSDTtZX";
4857 static void show_task(struct task_struct
*p
)
4859 struct task_struct
*relative
;
4860 unsigned long free
= 0;
4863 state
= p
->state
? __ffs(p
->state
) + 1 : 0;
4864 printk("%-13.13s %c", p
->comm
,
4865 state
< sizeof(stat_nam
) - 1 ? stat_nam
[state
] : '?');
4866 #if (BITS_PER_LONG == 32)
4867 if (state
== TASK_RUNNING
)
4868 printk(" running ");
4870 printk(" %08lX ", thread_saved_pc(p
));
4872 if (state
== TASK_RUNNING
)
4873 printk(" running task ");
4875 printk(" %016lx ", thread_saved_pc(p
));
4877 #ifdef CONFIG_DEBUG_STACK_USAGE
4879 unsigned long *n
= end_of_stack(p
);
4882 free
= (unsigned long)n
- (unsigned long)end_of_stack(p
);
4885 printk("%5lu %5d %6d ", free
, p
->pid
, p
->parent
->pid
);
4886 if ((relative
= eldest_child(p
)))
4887 printk("%5d ", relative
->pid
);
4890 if ((relative
= younger_sibling(p
)))
4891 printk("%7d", relative
->pid
);
4894 if ((relative
= older_sibling(p
)))
4895 printk(" %5d", relative
->pid
);
4899 printk(" (L-TLB)\n");
4901 printk(" (NOTLB)\n");
4903 if (state
!= TASK_RUNNING
)
4904 show_stack(p
, NULL
);
4907 void show_state_filter(unsigned long state_filter
)
4909 struct task_struct
*g
, *p
;
4911 #if (BITS_PER_LONG == 32)
4914 printk(" task PC stack pid father child younger older\n");
4918 printk(" task PC stack pid father child younger older\n");
4920 read_lock(&tasklist_lock
);
4921 do_each_thread(g
, p
) {
4923 * reset the NMI-timeout, listing all files on a slow
4924 * console might take alot of time:
4926 touch_nmi_watchdog();
4927 if (p
->state
& state_filter
)
4929 } while_each_thread(g
, p
);
4931 read_unlock(&tasklist_lock
);
4933 * Only show locks if all tasks are dumped:
4935 if (state_filter
== -1)
4936 debug_show_all_locks();
4940 * init_idle - set up an idle thread for a given CPU
4941 * @idle: task in question
4942 * @cpu: cpu the idle task belongs to
4944 * NOTE: this function does not set the idle thread's NEED_RESCHED
4945 * flag, to make booting more robust.
4947 void __cpuinit
init_idle(struct task_struct
*idle
, int cpu
)
4949 struct rq
*rq
= cpu_rq(cpu
);
4950 unsigned long flags
;
4952 idle
->timestamp
= sched_clock();
4953 idle
->sleep_avg
= 0;
4955 idle
->prio
= idle
->normal_prio
= MAX_PRIO
;
4956 idle
->state
= TASK_RUNNING
;
4957 idle
->cpus_allowed
= cpumask_of_cpu(cpu
);
4958 set_task_cpu(idle
, cpu
);
4960 spin_lock_irqsave(&rq
->lock
, flags
);
4961 rq
->curr
= rq
->idle
= idle
;
4962 #if defined(CONFIG_SMP) && defined(__ARCH_WANT_UNLOCKED_CTXSW)
4965 spin_unlock_irqrestore(&rq
->lock
, flags
);
4967 /* Set the preempt count _outside_ the spinlocks! */
4968 #if defined(CONFIG_PREEMPT) && !defined(CONFIG_PREEMPT_BKL)
4969 task_thread_info(idle
)->preempt_count
= (idle
->lock_depth
>= 0);
4971 task_thread_info(idle
)->preempt_count
= 0;
4976 * In a system that switches off the HZ timer nohz_cpu_mask
4977 * indicates which cpus entered this state. This is used
4978 * in the rcu update to wait only for active cpus. For system
4979 * which do not switch off the HZ timer nohz_cpu_mask should
4980 * always be CPU_MASK_NONE.
4982 cpumask_t nohz_cpu_mask
= CPU_MASK_NONE
;
4986 * This is how migration works:
4988 * 1) we queue a struct migration_req structure in the source CPU's
4989 * runqueue and wake up that CPU's migration thread.
4990 * 2) we down() the locked semaphore => thread blocks.
4991 * 3) migration thread wakes up (implicitly it forces the migrated
4992 * thread off the CPU)
4993 * 4) it gets the migration request and checks whether the migrated
4994 * task is still in the wrong runqueue.
4995 * 5) if it's in the wrong runqueue then the migration thread removes
4996 * it and puts it into the right queue.
4997 * 6) migration thread up()s the semaphore.
4998 * 7) we wake up and the migration is done.
5002 * Change a given task's CPU affinity. Migrate the thread to a
5003 * proper CPU and schedule it away if the CPU it's executing on
5004 * is removed from the allowed bitmask.
5006 * NOTE: the caller must have a valid reference to the task, the
5007 * task must not exit() & deallocate itself prematurely. The
5008 * call is not atomic; no spinlocks may be held.
5010 int set_cpus_allowed(struct task_struct
*p
, cpumask_t new_mask
)
5012 struct migration_req req
;
5013 unsigned long flags
;
5017 rq
= task_rq_lock(p
, &flags
);
5018 if (!cpus_intersects(new_mask
, cpu_online_map
)) {
5023 p
->cpus_allowed
= new_mask
;
5024 /* Can the task run on the task's current CPU? If so, we're done */
5025 if (cpu_isset(task_cpu(p
), new_mask
))
5028 if (migrate_task(p
, any_online_cpu(new_mask
), &req
)) {
5029 /* Need help from migration thread: drop lock and wait. */
5030 task_rq_unlock(rq
, &flags
);
5031 wake_up_process(rq
->migration_thread
);
5032 wait_for_completion(&req
.done
);
5033 tlb_migrate_finish(p
->mm
);
5037 task_rq_unlock(rq
, &flags
);
5041 EXPORT_SYMBOL_GPL(set_cpus_allowed
);
5044 * Move (not current) task off this cpu, onto dest cpu. We're doing
5045 * this because either it can't run here any more (set_cpus_allowed()
5046 * away from this CPU, or CPU going down), or because we're
5047 * attempting to rebalance this task on exec (sched_exec).
5049 * So we race with normal scheduler movements, but that's OK, as long
5050 * as the task is no longer on this CPU.
5052 * Returns non-zero if task was successfully migrated.
5054 static int __migrate_task(struct task_struct
*p
, int src_cpu
, int dest_cpu
)
5056 struct rq
*rq_dest
, *rq_src
;
5059 if (unlikely(cpu_is_offline(dest_cpu
)))
5062 rq_src
= cpu_rq(src_cpu
);
5063 rq_dest
= cpu_rq(dest_cpu
);
5065 double_rq_lock(rq_src
, rq_dest
);
5066 /* Already moved. */
5067 if (task_cpu(p
) != src_cpu
)
5069 /* Affinity changed (again). */
5070 if (!cpu_isset(dest_cpu
, p
->cpus_allowed
))
5073 set_task_cpu(p
, dest_cpu
);
5076 * Sync timestamp with rq_dest's before activating.
5077 * The same thing could be achieved by doing this step
5078 * afterwards, and pretending it was a local activate.
5079 * This way is cleaner and logically correct.
5081 p
->timestamp
= p
->timestamp
- rq_src
->most_recent_timestamp
5082 + rq_dest
->most_recent_timestamp
;
5083 deactivate_task(p
, rq_src
);
5084 __activate_task(p
, rq_dest
);
5085 if (TASK_PREEMPTS_CURR(p
, rq_dest
))
5086 resched_task(rq_dest
->curr
);
5090 double_rq_unlock(rq_src
, rq_dest
);
5095 * migration_thread - this is a highprio system thread that performs
5096 * thread migration by bumping thread off CPU then 'pushing' onto
5099 static int migration_thread(void *data
)
5101 int cpu
= (long)data
;
5105 BUG_ON(rq
->migration_thread
!= current
);
5107 set_current_state(TASK_INTERRUPTIBLE
);
5108 while (!kthread_should_stop()) {
5109 struct migration_req
*req
;
5110 struct list_head
*head
;
5114 spin_lock_irq(&rq
->lock
);
5116 if (cpu_is_offline(cpu
)) {
5117 spin_unlock_irq(&rq
->lock
);
5121 if (rq
->active_balance
) {
5122 active_load_balance(rq
, cpu
);
5123 rq
->active_balance
= 0;
5126 head
= &rq
->migration_queue
;
5128 if (list_empty(head
)) {
5129 spin_unlock_irq(&rq
->lock
);
5131 set_current_state(TASK_INTERRUPTIBLE
);
5134 req
= list_entry(head
->next
, struct migration_req
, list
);
5135 list_del_init(head
->next
);
5137 spin_unlock(&rq
->lock
);
5138 __migrate_task(req
->task
, cpu
, req
->dest_cpu
);
5141 complete(&req
->done
);
5143 __set_current_state(TASK_RUNNING
);
5147 /* Wait for kthread_stop */
5148 set_current_state(TASK_INTERRUPTIBLE
);
5149 while (!kthread_should_stop()) {
5151 set_current_state(TASK_INTERRUPTIBLE
);
5153 __set_current_state(TASK_RUNNING
);
5157 #ifdef CONFIG_HOTPLUG_CPU
5159 * Figure out where task on dead CPU should go, use force if neccessary.
5160 * NOTE: interrupts should be disabled by the caller
5162 static void move_task_off_dead_cpu(int dead_cpu
, struct task_struct
*p
)
5164 unsigned long flags
;
5171 mask
= node_to_cpumask(cpu_to_node(dead_cpu
));
5172 cpus_and(mask
, mask
, p
->cpus_allowed
);
5173 dest_cpu
= any_online_cpu(mask
);
5175 /* On any allowed CPU? */
5176 if (dest_cpu
== NR_CPUS
)
5177 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5179 /* No more Mr. Nice Guy. */
5180 if (dest_cpu
== NR_CPUS
) {
5181 rq
= task_rq_lock(p
, &flags
);
5182 cpus_setall(p
->cpus_allowed
);
5183 dest_cpu
= any_online_cpu(p
->cpus_allowed
);
5184 task_rq_unlock(rq
, &flags
);
5187 * Don't tell them about moving exiting tasks or
5188 * kernel threads (both mm NULL), since they never
5191 if (p
->mm
&& printk_ratelimit())
5192 printk(KERN_INFO
"process %d (%s) no "
5193 "longer affine to cpu%d\n",
5194 p
->pid
, p
->comm
, dead_cpu
);
5196 if (!__migrate_task(p
, dead_cpu
, dest_cpu
))
5201 * While a dead CPU has no uninterruptible tasks queued at this point,
5202 * it might still have a nonzero ->nr_uninterruptible counter, because
5203 * for performance reasons the counter is not stricly tracking tasks to
5204 * their home CPUs. So we just add the counter to another CPU's counter,
5205 * to keep the global sum constant after CPU-down:
5207 static void migrate_nr_uninterruptible(struct rq
*rq_src
)
5209 struct rq
*rq_dest
= cpu_rq(any_online_cpu(CPU_MASK_ALL
));
5210 unsigned long flags
;
5212 local_irq_save(flags
);
5213 double_rq_lock(rq_src
, rq_dest
);
5214 rq_dest
->nr_uninterruptible
+= rq_src
->nr_uninterruptible
;
5215 rq_src
->nr_uninterruptible
= 0;
5216 double_rq_unlock(rq_src
, rq_dest
);
5217 local_irq_restore(flags
);
5220 /* Run through task list and migrate tasks from the dead cpu. */
5221 static void migrate_live_tasks(int src_cpu
)
5223 struct task_struct
*p
, *t
;
5225 write_lock_irq(&tasklist_lock
);
5227 do_each_thread(t
, p
) {
5231 if (task_cpu(p
) == src_cpu
)
5232 move_task_off_dead_cpu(src_cpu
, p
);
5233 } while_each_thread(t
, p
);
5235 write_unlock_irq(&tasklist_lock
);
5238 /* Schedules idle task to be the next runnable task on current CPU.
5239 * It does so by boosting its priority to highest possible and adding it to
5240 * the _front_ of the runqueue. Used by CPU offline code.
5242 void sched_idle_next(void)
5244 int this_cpu
= smp_processor_id();
5245 struct rq
*rq
= cpu_rq(this_cpu
);
5246 struct task_struct
*p
= rq
->idle
;
5247 unsigned long flags
;
5249 /* cpu has to be offline */
5250 BUG_ON(cpu_online(this_cpu
));
5253 * Strictly not necessary since rest of the CPUs are stopped by now
5254 * and interrupts disabled on the current cpu.
5256 spin_lock_irqsave(&rq
->lock
, flags
);
5258 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5260 /* Add idle task to the _front_ of its priority queue: */
5261 __activate_idle_task(p
, rq
);
5263 spin_unlock_irqrestore(&rq
->lock
, flags
);
5267 * Ensures that the idle task is using init_mm right before its cpu goes
5270 void idle_task_exit(void)
5272 struct mm_struct
*mm
= current
->active_mm
;
5274 BUG_ON(cpu_online(smp_processor_id()));
5277 switch_mm(mm
, &init_mm
, current
);
5281 /* called under rq->lock with disabled interrupts */
5282 static void migrate_dead(unsigned int dead_cpu
, struct task_struct
*p
)
5284 struct rq
*rq
= cpu_rq(dead_cpu
);
5286 /* Must be exiting, otherwise would be on tasklist. */
5287 BUG_ON(p
->exit_state
!= EXIT_ZOMBIE
&& p
->exit_state
!= EXIT_DEAD
);
5289 /* Cannot have done final schedule yet: would have vanished. */
5290 BUG_ON(p
->state
== TASK_DEAD
);
5295 * Drop lock around migration; if someone else moves it,
5296 * that's OK. No task can be added to this CPU, so iteration is
5298 * NOTE: interrupts should be left disabled --dev@
5300 spin_unlock(&rq
->lock
);
5301 move_task_off_dead_cpu(dead_cpu
, p
);
5302 spin_lock(&rq
->lock
);
5307 /* release_task() removes task from tasklist, so we won't find dead tasks. */
5308 static void migrate_dead_tasks(unsigned int dead_cpu
)
5310 struct rq
*rq
= cpu_rq(dead_cpu
);
5311 unsigned int arr
, i
;
5313 for (arr
= 0; arr
< 2; arr
++) {
5314 for (i
= 0; i
< MAX_PRIO
; i
++) {
5315 struct list_head
*list
= &rq
->arrays
[arr
].queue
[i
];
5317 while (!list_empty(list
))
5318 migrate_dead(dead_cpu
, list_entry(list
->next
,
5319 struct task_struct
, run_list
));
5323 #endif /* CONFIG_HOTPLUG_CPU */
5326 * migration_call - callback that gets triggered when a CPU is added.
5327 * Here we can start up the necessary migration thread for the new CPU.
5329 static int __cpuinit
5330 migration_call(struct notifier_block
*nfb
, unsigned long action
, void *hcpu
)
5332 struct task_struct
*p
;
5333 int cpu
= (long)hcpu
;
5334 unsigned long flags
;
5338 case CPU_UP_PREPARE
:
5339 p
= kthread_create(migration_thread
, hcpu
, "migration/%d",cpu
);
5342 p
->flags
|= PF_NOFREEZE
;
5343 kthread_bind(p
, cpu
);
5344 /* Must be high prio: stop_machine expects to yield to it. */
5345 rq
= task_rq_lock(p
, &flags
);
5346 __setscheduler(p
, SCHED_FIFO
, MAX_RT_PRIO
-1);
5347 task_rq_unlock(rq
, &flags
);
5348 cpu_rq(cpu
)->migration_thread
= p
;
5352 /* Strictly unneccessary, as first user will wake it. */
5353 wake_up_process(cpu_rq(cpu
)->migration_thread
);
5356 #ifdef CONFIG_HOTPLUG_CPU
5357 case CPU_UP_CANCELED
:
5358 if (!cpu_rq(cpu
)->migration_thread
)
5360 /* Unbind it from offline cpu so it can run. Fall thru. */
5361 kthread_bind(cpu_rq(cpu
)->migration_thread
,
5362 any_online_cpu(cpu_online_map
));
5363 kthread_stop(cpu_rq(cpu
)->migration_thread
);
5364 cpu_rq(cpu
)->migration_thread
= NULL
;
5368 migrate_live_tasks(cpu
);
5370 kthread_stop(rq
->migration_thread
);
5371 rq
->migration_thread
= NULL
;
5372 /* Idle task back to normal (off runqueue, low prio) */
5373 rq
= task_rq_lock(rq
->idle
, &flags
);
5374 deactivate_task(rq
->idle
, rq
);
5375 rq
->idle
->static_prio
= MAX_PRIO
;
5376 __setscheduler(rq
->idle
, SCHED_NORMAL
, 0);
5377 migrate_dead_tasks(cpu
);
5378 task_rq_unlock(rq
, &flags
);
5379 migrate_nr_uninterruptible(rq
);
5380 BUG_ON(rq
->nr_running
!= 0);
5382 /* No need to migrate the tasks: it was best-effort if
5383 * they didn't do lock_cpu_hotplug(). Just wake up
5384 * the requestors. */
5385 spin_lock_irq(&rq
->lock
);
5386 while (!list_empty(&rq
->migration_queue
)) {
5387 struct migration_req
*req
;
5389 req
= list_entry(rq
->migration_queue
.next
,
5390 struct migration_req
, list
);
5391 list_del_init(&req
->list
);
5392 complete(&req
->done
);
5394 spin_unlock_irq(&rq
->lock
);
5401 /* Register at highest priority so that task migration (migrate_all_tasks)
5402 * happens before everything else.
5404 static struct notifier_block __cpuinitdata migration_notifier
= {
5405 .notifier_call
= migration_call
,
5409 int __init
migration_init(void)
5411 void *cpu
= (void *)(long)smp_processor_id();
5414 /* Start one for the boot CPU: */
5415 err
= migration_call(&migration_notifier
, CPU_UP_PREPARE
, cpu
);
5416 BUG_ON(err
== NOTIFY_BAD
);
5417 migration_call(&migration_notifier
, CPU_ONLINE
, cpu
);
5418 register_cpu_notifier(&migration_notifier
);
5425 #undef SCHED_DOMAIN_DEBUG
5426 #ifdef SCHED_DOMAIN_DEBUG
5427 static void sched_domain_debug(struct sched_domain
*sd
, int cpu
)
5432 printk(KERN_DEBUG
"CPU%d attaching NULL sched-domain.\n", cpu
);
5436 printk(KERN_DEBUG
"CPU%d attaching sched-domain:\n", cpu
);
5441 struct sched_group
*group
= sd
->groups
;
5442 cpumask_t groupmask
;
5444 cpumask_scnprintf(str
, NR_CPUS
, sd
->span
);
5445 cpus_clear(groupmask
);
5448 for (i
= 0; i
< level
+ 1; i
++)
5450 printk("domain %d: ", level
);
5452 if (!(sd
->flags
& SD_LOAD_BALANCE
)) {
5453 printk("does not load-balance\n");
5455 printk(KERN_ERR
"ERROR: !SD_LOAD_BALANCE domain"
5460 printk("span %s\n", str
);
5462 if (!cpu_isset(cpu
, sd
->span
))
5463 printk(KERN_ERR
"ERROR: domain->span does not contain "
5465 if (!cpu_isset(cpu
, group
->cpumask
))
5466 printk(KERN_ERR
"ERROR: domain->groups does not contain"
5470 for (i
= 0; i
< level
+ 2; i
++)
5476 printk(KERN_ERR
"ERROR: group is NULL\n");
5480 if (!group
->cpu_power
) {
5482 printk(KERN_ERR
"ERROR: domain->cpu_power not "
5486 if (!cpus_weight(group
->cpumask
)) {
5488 printk(KERN_ERR
"ERROR: empty group\n");
5491 if (cpus_intersects(groupmask
, group
->cpumask
)) {
5493 printk(KERN_ERR
"ERROR: repeated CPUs\n");
5496 cpus_or(groupmask
, groupmask
, group
->cpumask
);
5498 cpumask_scnprintf(str
, NR_CPUS
, group
->cpumask
);
5501 group
= group
->next
;
5502 } while (group
!= sd
->groups
);
5505 if (!cpus_equal(sd
->span
, groupmask
))
5506 printk(KERN_ERR
"ERROR: groups don't span "
5514 if (!cpus_subset(groupmask
, sd
->span
))
5515 printk(KERN_ERR
"ERROR: parent span is not a superset "
5516 "of domain->span\n");
5521 # define sched_domain_debug(sd, cpu) do { } while (0)
5524 static int sd_degenerate(struct sched_domain
*sd
)
5526 if (cpus_weight(sd
->span
) == 1)
5529 /* Following flags need at least 2 groups */
5530 if (sd
->flags
& (SD_LOAD_BALANCE
|
5531 SD_BALANCE_NEWIDLE
|
5535 SD_SHARE_PKG_RESOURCES
)) {
5536 if (sd
->groups
!= sd
->groups
->next
)
5540 /* Following flags don't use groups */
5541 if (sd
->flags
& (SD_WAKE_IDLE
|
5550 sd_parent_degenerate(struct sched_domain
*sd
, struct sched_domain
*parent
)
5552 unsigned long cflags
= sd
->flags
, pflags
= parent
->flags
;
5554 if (sd_degenerate(parent
))
5557 if (!cpus_equal(sd
->span
, parent
->span
))
5560 /* Does parent contain flags not in child? */
5561 /* WAKE_BALANCE is a subset of WAKE_AFFINE */
5562 if (cflags
& SD_WAKE_AFFINE
)
5563 pflags
&= ~SD_WAKE_BALANCE
;
5564 /* Flags needing groups don't count if only 1 group in parent */
5565 if (parent
->groups
== parent
->groups
->next
) {
5566 pflags
&= ~(SD_LOAD_BALANCE
|
5567 SD_BALANCE_NEWIDLE
|
5571 SD_SHARE_PKG_RESOURCES
);
5573 if (~cflags
& pflags
)
5580 * Attach the domain 'sd' to 'cpu' as its base domain. Callers must
5581 * hold the hotplug lock.
5583 static void cpu_attach_domain(struct sched_domain
*sd
, int cpu
)
5585 struct rq
*rq
= cpu_rq(cpu
);
5586 struct sched_domain
*tmp
;
5588 /* Remove the sched domains which do not contribute to scheduling. */
5589 for (tmp
= sd
; tmp
; tmp
= tmp
->parent
) {
5590 struct sched_domain
*parent
= tmp
->parent
;
5593 if (sd_parent_degenerate(tmp
, parent
)) {
5594 tmp
->parent
= parent
->parent
;
5596 parent
->parent
->child
= tmp
;
5600 if (sd
&& sd_degenerate(sd
)) {
5606 sched_domain_debug(sd
, cpu
);
5608 rcu_assign_pointer(rq
->sd
, sd
);
5611 /* cpus with isolated domains */
5612 static cpumask_t cpu_isolated_map
= CPU_MASK_NONE
;
5614 /* Setup the mask of cpus configured for isolated domains */
5615 static int __init
isolated_cpu_setup(char *str
)
5617 int ints
[NR_CPUS
], i
;
5619 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5620 cpus_clear(cpu_isolated_map
);
5621 for (i
= 1; i
<= ints
[0]; i
++)
5622 if (ints
[i
] < NR_CPUS
)
5623 cpu_set(ints
[i
], cpu_isolated_map
);
5627 __setup ("isolcpus=", isolated_cpu_setup
);
5630 * init_sched_build_groups takes the cpumask we wish to span, and a pointer
5631 * to a function which identifies what group(along with sched group) a CPU
5632 * belongs to. The return value of group_fn must be a >= 0 and < NR_CPUS
5633 * (due to the fact that we keep track of groups covered with a cpumask_t).
5635 * init_sched_build_groups will build a circular linked list of the groups
5636 * covered by the given span, and will set each group's ->cpumask correctly,
5637 * and ->cpu_power to 0.
5640 init_sched_build_groups(cpumask_t span
, const cpumask_t
*cpu_map
,
5641 int (*group_fn
)(int cpu
, const cpumask_t
*cpu_map
,
5642 struct sched_group
**sg
))
5644 struct sched_group
*first
= NULL
, *last
= NULL
;
5645 cpumask_t covered
= CPU_MASK_NONE
;
5648 for_each_cpu_mask(i
, span
) {
5649 struct sched_group
*sg
;
5650 int group
= group_fn(i
, cpu_map
, &sg
);
5653 if (cpu_isset(i
, covered
))
5656 sg
->cpumask
= CPU_MASK_NONE
;
5659 for_each_cpu_mask(j
, span
) {
5660 if (group_fn(j
, cpu_map
, NULL
) != group
)
5663 cpu_set(j
, covered
);
5664 cpu_set(j
, sg
->cpumask
);
5675 #define SD_NODES_PER_DOMAIN 16
5678 * Self-tuning task migration cost measurement between source and target CPUs.
5680 * This is done by measuring the cost of manipulating buffers of varying
5681 * sizes. For a given buffer-size here are the steps that are taken:
5683 * 1) the source CPU reads+dirties a shared buffer
5684 * 2) the target CPU reads+dirties the same shared buffer
5686 * We measure how long they take, in the following 4 scenarios:
5688 * - source: CPU1, target: CPU2 | cost1
5689 * - source: CPU2, target: CPU1 | cost2
5690 * - source: CPU1, target: CPU1 | cost3
5691 * - source: CPU2, target: CPU2 | cost4
5693 * We then calculate the cost3+cost4-cost1-cost2 difference - this is
5694 * the cost of migration.
5696 * We then start off from a small buffer-size and iterate up to larger
5697 * buffer sizes, in 5% steps - measuring each buffer-size separately, and
5698 * doing a maximum search for the cost. (The maximum cost for a migration
5699 * normally occurs when the working set size is around the effective cache
5702 #define SEARCH_SCOPE 2
5703 #define MIN_CACHE_SIZE (64*1024U)
5704 #define DEFAULT_CACHE_SIZE (5*1024*1024U)
5705 #define ITERATIONS 1
5706 #define SIZE_THRESH 130
5707 #define COST_THRESH 130
5710 * The migration cost is a function of 'domain distance'. Domain
5711 * distance is the number of steps a CPU has to iterate down its
5712 * domain tree to share a domain with the other CPU. The farther
5713 * two CPUs are from each other, the larger the distance gets.
5715 * Note that we use the distance only to cache measurement results,
5716 * the distance value is not used numerically otherwise. When two
5717 * CPUs have the same distance it is assumed that the migration
5718 * cost is the same. (this is a simplification but quite practical)
5720 #define MAX_DOMAIN_DISTANCE 32
5722 static unsigned long long migration_cost
[MAX_DOMAIN_DISTANCE
] =
5723 { [ 0 ... MAX_DOMAIN_DISTANCE
-1 ] =
5725 * Architectures may override the migration cost and thus avoid
5726 * boot-time calibration. Unit is nanoseconds. Mostly useful for
5727 * virtualized hardware:
5729 #ifdef CONFIG_DEFAULT_MIGRATION_COST
5730 CONFIG_DEFAULT_MIGRATION_COST
5737 * Allow override of migration cost - in units of microseconds.
5738 * E.g. migration_cost=1000,2000,3000 will set up a level-1 cost
5739 * of 1 msec, level-2 cost of 2 msecs and level3 cost of 3 msecs:
5741 static int __init
migration_cost_setup(char *str
)
5743 int ints
[MAX_DOMAIN_DISTANCE
+1], i
;
5745 str
= get_options(str
, ARRAY_SIZE(ints
), ints
);
5747 printk("#ints: %d\n", ints
[0]);
5748 for (i
= 1; i
<= ints
[0]; i
++) {
5749 migration_cost
[i
-1] = (unsigned long long)ints
[i
]*1000;
5750 printk("migration_cost[%d]: %Ld\n", i
-1, migration_cost
[i
-1]);
5755 __setup ("migration_cost=", migration_cost_setup
);
5758 * Global multiplier (divisor) for migration-cutoff values,
5759 * in percentiles. E.g. use a value of 150 to get 1.5 times
5760 * longer cache-hot cutoff times.
5762 * (We scale it from 100 to 128 to long long handling easier.)
5765 #define MIGRATION_FACTOR_SCALE 128
5767 static unsigned int migration_factor
= MIGRATION_FACTOR_SCALE
;
5769 static int __init
setup_migration_factor(char *str
)
5771 get_option(&str
, &migration_factor
);
5772 migration_factor
= migration_factor
* MIGRATION_FACTOR_SCALE
/ 100;
5776 __setup("migration_factor=", setup_migration_factor
);
5779 * Estimated distance of two CPUs, measured via the number of domains
5780 * we have to pass for the two CPUs to be in the same span:
5782 static unsigned long domain_distance(int cpu1
, int cpu2
)
5784 unsigned long distance
= 0;
5785 struct sched_domain
*sd
;
5787 for_each_domain(cpu1
, sd
) {
5788 WARN_ON(!cpu_isset(cpu1
, sd
->span
));
5789 if (cpu_isset(cpu2
, sd
->span
))
5793 if (distance
>= MAX_DOMAIN_DISTANCE
) {
5795 distance
= MAX_DOMAIN_DISTANCE
-1;
5801 static unsigned int migration_debug
;
5803 static int __init
setup_migration_debug(char *str
)
5805 get_option(&str
, &migration_debug
);
5809 __setup("migration_debug=", setup_migration_debug
);
5812 * Maximum cache-size that the scheduler should try to measure.
5813 * Architectures with larger caches should tune this up during
5814 * bootup. Gets used in the domain-setup code (i.e. during SMP
5817 unsigned int max_cache_size
;
5819 static int __init
setup_max_cache_size(char *str
)
5821 get_option(&str
, &max_cache_size
);
5825 __setup("max_cache_size=", setup_max_cache_size
);
5828 * Dirty a big buffer in a hard-to-predict (for the L2 cache) way. This
5829 * is the operation that is timed, so we try to generate unpredictable
5830 * cachemisses that still end up filling the L2 cache:
5832 static void touch_cache(void *__cache
, unsigned long __size
)
5834 unsigned long size
= __size
/ sizeof(long);
5835 unsigned long chunk1
= size
/ 3;
5836 unsigned long chunk2
= 2 * size
/ 3;
5837 unsigned long *cache
= __cache
;
5840 for (i
= 0; i
< size
/6; i
+= 8) {
5843 case 1: cache
[size
-1-i
]++;
5844 case 2: cache
[chunk1
-i
]++;
5845 case 3: cache
[chunk1
+i
]++;
5846 case 4: cache
[chunk2
-i
]++;
5847 case 5: cache
[chunk2
+i
]++;
5853 * Measure the cache-cost of one task migration. Returns in units of nsec.
5855 static unsigned long long
5856 measure_one(void *cache
, unsigned long size
, int source
, int target
)
5858 cpumask_t mask
, saved_mask
;
5859 unsigned long long t0
, t1
, t2
, t3
, cost
;
5861 saved_mask
= current
->cpus_allowed
;
5864 * Flush source caches to RAM and invalidate them:
5869 * Migrate to the source CPU:
5871 mask
= cpumask_of_cpu(source
);
5872 set_cpus_allowed(current
, mask
);
5873 WARN_ON(smp_processor_id() != source
);
5876 * Dirty the working set:
5879 touch_cache(cache
, size
);
5883 * Migrate to the target CPU, dirty the L2 cache and access
5884 * the shared buffer. (which represents the working set
5885 * of a migrated task.)
5887 mask
= cpumask_of_cpu(target
);
5888 set_cpus_allowed(current
, mask
);
5889 WARN_ON(smp_processor_id() != target
);
5892 touch_cache(cache
, size
);
5895 cost
= t1
-t0
+ t3
-t2
;
5897 if (migration_debug
>= 2)
5898 printk("[%d->%d]: %8Ld %8Ld %8Ld => %10Ld.\n",
5899 source
, target
, t1
-t0
, t1
-t0
, t3
-t2
, cost
);
5901 * Flush target caches to RAM and invalidate them:
5905 set_cpus_allowed(current
, saved_mask
);
5911 * Measure a series of task migrations and return the average
5912 * result. Since this code runs early during bootup the system
5913 * is 'undisturbed' and the average latency makes sense.
5915 * The algorithm in essence auto-detects the relevant cache-size,
5916 * so it will properly detect different cachesizes for different
5917 * cache-hierarchies, depending on how the CPUs are connected.
5919 * Architectures can prime the upper limit of the search range via
5920 * max_cache_size, otherwise the search range defaults to 20MB...64K.
5922 static unsigned long long
5923 measure_cost(int cpu1
, int cpu2
, void *cache
, unsigned int size
)
5925 unsigned long long cost1
, cost2
;
5929 * Measure the migration cost of 'size' bytes, over an
5930 * average of 10 runs:
5932 * (We perturb the cache size by a small (0..4k)
5933 * value to compensate size/alignment related artifacts.
5934 * We also subtract the cost of the operation done on
5940 * dry run, to make sure we start off cache-cold on cpu1,
5941 * and to get any vmalloc pagefaults in advance:
5943 measure_one(cache
, size
, cpu1
, cpu2
);
5944 for (i
= 0; i
< ITERATIONS
; i
++)
5945 cost1
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu2
);
5947 measure_one(cache
, size
, cpu2
, cpu1
);
5948 for (i
= 0; i
< ITERATIONS
; i
++)
5949 cost1
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu1
);
5952 * (We measure the non-migrating [cached] cost on both
5953 * cpu1 and cpu2, to handle CPUs with different speeds)
5957 measure_one(cache
, size
, cpu1
, cpu1
);
5958 for (i
= 0; i
< ITERATIONS
; i
++)
5959 cost2
+= measure_one(cache
, size
- i
* 1024, cpu1
, cpu1
);
5961 measure_one(cache
, size
, cpu2
, cpu2
);
5962 for (i
= 0; i
< ITERATIONS
; i
++)
5963 cost2
+= measure_one(cache
, size
- i
* 1024, cpu2
, cpu2
);
5966 * Get the per-iteration migration cost:
5968 do_div(cost1
, 2 * ITERATIONS
);
5969 do_div(cost2
, 2 * ITERATIONS
);
5971 return cost1
- cost2
;
5974 static unsigned long long measure_migration_cost(int cpu1
, int cpu2
)
5976 unsigned long long max_cost
= 0, fluct
= 0, avg_fluct
= 0;
5977 unsigned int max_size
, size
, size_found
= 0;
5978 long long cost
= 0, prev_cost
;
5982 * Search from max_cache_size*5 down to 64K - the real relevant
5983 * cachesize has to lie somewhere inbetween.
5985 if (max_cache_size
) {
5986 max_size
= max(max_cache_size
* SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5987 size
= max(max_cache_size
/ SEARCH_SCOPE
, MIN_CACHE_SIZE
);
5990 * Since we have no estimation about the relevant
5993 max_size
= DEFAULT_CACHE_SIZE
* SEARCH_SCOPE
;
5994 size
= MIN_CACHE_SIZE
;
5997 if (!cpu_online(cpu1
) || !cpu_online(cpu2
)) {
5998 printk("cpu %d and %d not both online!\n", cpu1
, cpu2
);
6003 * Allocate the working set:
6005 cache
= vmalloc(max_size
);
6007 printk("could not vmalloc %d bytes for cache!\n", 2 * max_size
);
6008 return 1000000; /* return 1 msec on very small boxen */
6011 while (size
<= max_size
) {
6013 cost
= measure_cost(cpu1
, cpu2
, cache
, size
);
6019 if (max_cost
< cost
) {
6025 * Calculate average fluctuation, we use this to prevent
6026 * noise from triggering an early break out of the loop:
6028 fluct
= abs(cost
- prev_cost
);
6029 avg_fluct
= (avg_fluct
+ fluct
)/2;
6031 if (migration_debug
)
6032 printk("-> [%d][%d][%7d] %3ld.%ld [%3ld.%ld] (%ld): "
6035 (long)cost
/ 1000000,
6036 ((long)cost
/ 100000) % 10,
6037 (long)max_cost
/ 1000000,
6038 ((long)max_cost
/ 100000) % 10,
6039 domain_distance(cpu1
, cpu2
),
6043 * If we iterated at least 20% past the previous maximum,
6044 * and the cost has dropped by more than 20% already,
6045 * (taking fluctuations into account) then we assume to
6046 * have found the maximum and break out of the loop early:
6048 if (size_found
&& (size
*100 > size_found
*SIZE_THRESH
))
6049 if (cost
+avg_fluct
<= 0 ||
6050 max_cost
*100 > (cost
+avg_fluct
)*COST_THRESH
) {
6052 if (migration_debug
)
6053 printk("-> found max.\n");
6057 * Increase the cachesize in 10% steps:
6059 size
= size
* 10 / 9;
6062 if (migration_debug
)
6063 printk("[%d][%d] working set size found: %d, cost: %Ld\n",
6064 cpu1
, cpu2
, size_found
, max_cost
);
6069 * A task is considered 'cache cold' if at least 2 times
6070 * the worst-case cost of migration has passed.
6072 * (this limit is only listened to if the load-balancing
6073 * situation is 'nice' - if there is a large imbalance we
6074 * ignore it for the sake of CPU utilization and
6075 * processing fairness.)
6077 return 2 * max_cost
* migration_factor
/ MIGRATION_FACTOR_SCALE
;
6080 static void calibrate_migration_costs(const cpumask_t
*cpu_map
)
6082 int cpu1
= -1, cpu2
= -1, cpu
, orig_cpu
= raw_smp_processor_id();
6083 unsigned long j0
, j1
, distance
, max_distance
= 0;
6084 struct sched_domain
*sd
;
6089 * First pass - calculate the cacheflush times:
6091 for_each_cpu_mask(cpu1
, *cpu_map
) {
6092 for_each_cpu_mask(cpu2
, *cpu_map
) {
6095 distance
= domain_distance(cpu1
, cpu2
);
6096 max_distance
= max(max_distance
, distance
);
6098 * No result cached yet?
6100 if (migration_cost
[distance
] == -1LL)
6101 migration_cost
[distance
] =
6102 measure_migration_cost(cpu1
, cpu2
);
6106 * Second pass - update the sched domain hierarchy with
6107 * the new cache-hot-time estimations:
6109 for_each_cpu_mask(cpu
, *cpu_map
) {
6111 for_each_domain(cpu
, sd
) {
6112 sd
->cache_hot_time
= migration_cost
[distance
];
6119 if (migration_debug
)
6120 printk("migration: max_cache_size: %d, cpu: %d MHz:\n",
6128 if (system_state
== SYSTEM_BOOTING
&& num_online_cpus() > 1) {
6129 printk("migration_cost=");
6130 for (distance
= 0; distance
<= max_distance
; distance
++) {
6133 printk("%ld", (long)migration_cost
[distance
] / 1000);
6138 if (migration_debug
)
6139 printk("migration: %ld seconds\n", (j1
-j0
) / HZ
);
6142 * Move back to the original CPU. NUMA-Q gets confused
6143 * if we migrate to another quad during bootup.
6145 if (raw_smp_processor_id() != orig_cpu
) {
6146 cpumask_t mask
= cpumask_of_cpu(orig_cpu
),
6147 saved_mask
= current
->cpus_allowed
;
6149 set_cpus_allowed(current
, mask
);
6150 set_cpus_allowed(current
, saved_mask
);
6157 * find_next_best_node - find the next node to include in a sched_domain
6158 * @node: node whose sched_domain we're building
6159 * @used_nodes: nodes already in the sched_domain
6161 * Find the next node to include in a given scheduling domain. Simply
6162 * finds the closest node not already in the @used_nodes map.
6164 * Should use nodemask_t.
6166 static int find_next_best_node(int node
, unsigned long *used_nodes
)
6168 int i
, n
, val
, min_val
, best_node
= 0;
6172 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6173 /* Start at @node */
6174 n
= (node
+ i
) % MAX_NUMNODES
;
6176 if (!nr_cpus_node(n
))
6179 /* Skip already used nodes */
6180 if (test_bit(n
, used_nodes
))
6183 /* Simple min distance search */
6184 val
= node_distance(node
, n
);
6186 if (val
< min_val
) {
6192 set_bit(best_node
, used_nodes
);
6197 * sched_domain_node_span - get a cpumask for a node's sched_domain
6198 * @node: node whose cpumask we're constructing
6199 * @size: number of nodes to include in this span
6201 * Given a node, construct a good cpumask for its sched_domain to span. It
6202 * should be one that prevents unnecessary balancing, but also spreads tasks
6205 static cpumask_t
sched_domain_node_span(int node
)
6207 DECLARE_BITMAP(used_nodes
, MAX_NUMNODES
);
6208 cpumask_t span
, nodemask
;
6212 bitmap_zero(used_nodes
, MAX_NUMNODES
);
6214 nodemask
= node_to_cpumask(node
);
6215 cpus_or(span
, span
, nodemask
);
6216 set_bit(node
, used_nodes
);
6218 for (i
= 1; i
< SD_NODES_PER_DOMAIN
; i
++) {
6219 int next_node
= find_next_best_node(node
, used_nodes
);
6221 nodemask
= node_to_cpumask(next_node
);
6222 cpus_or(span
, span
, nodemask
);
6229 int sched_smt_power_savings
= 0, sched_mc_power_savings
= 0;
6232 * SMT sched-domains:
6234 #ifdef CONFIG_SCHED_SMT
6235 static DEFINE_PER_CPU(struct sched_domain
, cpu_domains
);
6236 static DEFINE_PER_CPU(struct sched_group
, sched_group_cpus
);
6238 static int cpu_to_cpu_group(int cpu
, const cpumask_t
*cpu_map
,
6239 struct sched_group
**sg
)
6242 *sg
= &per_cpu(sched_group_cpus
, cpu
);
6248 * multi-core sched-domains:
6250 #ifdef CONFIG_SCHED_MC
6251 static DEFINE_PER_CPU(struct sched_domain
, core_domains
);
6252 static DEFINE_PER_CPU(struct sched_group
, sched_group_core
);
6255 #if defined(CONFIG_SCHED_MC) && defined(CONFIG_SCHED_SMT)
6256 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6257 struct sched_group
**sg
)
6260 cpumask_t mask
= cpu_sibling_map
[cpu
];
6261 cpus_and(mask
, mask
, *cpu_map
);
6262 group
= first_cpu(mask
);
6264 *sg
= &per_cpu(sched_group_core
, group
);
6267 #elif defined(CONFIG_SCHED_MC)
6268 static int cpu_to_core_group(int cpu
, const cpumask_t
*cpu_map
,
6269 struct sched_group
**sg
)
6272 *sg
= &per_cpu(sched_group_core
, cpu
);
6277 static DEFINE_PER_CPU(struct sched_domain
, phys_domains
);
6278 static DEFINE_PER_CPU(struct sched_group
, sched_group_phys
);
6280 static int cpu_to_phys_group(int cpu
, const cpumask_t
*cpu_map
,
6281 struct sched_group
**sg
)
6284 #ifdef CONFIG_SCHED_MC
6285 cpumask_t mask
= cpu_coregroup_map(cpu
);
6286 cpus_and(mask
, mask
, *cpu_map
);
6287 group
= first_cpu(mask
);
6288 #elif defined(CONFIG_SCHED_SMT)
6289 cpumask_t mask
= cpu_sibling_map
[cpu
];
6290 cpus_and(mask
, mask
, *cpu_map
);
6291 group
= first_cpu(mask
);
6296 *sg
= &per_cpu(sched_group_phys
, group
);
6302 * The init_sched_build_groups can't handle what we want to do with node
6303 * groups, so roll our own. Now each node has its own list of groups which
6304 * gets dynamically allocated.
6306 static DEFINE_PER_CPU(struct sched_domain
, node_domains
);
6307 static struct sched_group
**sched_group_nodes_bycpu
[NR_CPUS
];
6309 static DEFINE_PER_CPU(struct sched_domain
, allnodes_domains
);
6310 static DEFINE_PER_CPU(struct sched_group
, sched_group_allnodes
);
6312 static int cpu_to_allnodes_group(int cpu
, const cpumask_t
*cpu_map
,
6313 struct sched_group
**sg
)
6315 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(cpu
));
6318 cpus_and(nodemask
, nodemask
, *cpu_map
);
6319 group
= first_cpu(nodemask
);
6322 *sg
= &per_cpu(sched_group_allnodes
, group
);
6326 static void init_numa_sched_groups_power(struct sched_group
*group_head
)
6328 struct sched_group
*sg
= group_head
;
6334 for_each_cpu_mask(j
, sg
->cpumask
) {
6335 struct sched_domain
*sd
;
6337 sd
= &per_cpu(phys_domains
, j
);
6338 if (j
!= first_cpu(sd
->groups
->cpumask
)) {
6340 * Only add "power" once for each
6346 sg
->cpu_power
+= sd
->groups
->cpu_power
;
6349 if (sg
!= group_head
)
6355 /* Free memory allocated for various sched_group structures */
6356 static void free_sched_groups(const cpumask_t
*cpu_map
)
6360 for_each_cpu_mask(cpu
, *cpu_map
) {
6361 struct sched_group
**sched_group_nodes
6362 = sched_group_nodes_bycpu
[cpu
];
6364 if (!sched_group_nodes
)
6367 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6368 cpumask_t nodemask
= node_to_cpumask(i
);
6369 struct sched_group
*oldsg
, *sg
= sched_group_nodes
[i
];
6371 cpus_and(nodemask
, nodemask
, *cpu_map
);
6372 if (cpus_empty(nodemask
))
6382 if (oldsg
!= sched_group_nodes
[i
])
6385 kfree(sched_group_nodes
);
6386 sched_group_nodes_bycpu
[cpu
] = NULL
;
6390 static void free_sched_groups(const cpumask_t
*cpu_map
)
6396 * Initialize sched groups cpu_power.
6398 * cpu_power indicates the capacity of sched group, which is used while
6399 * distributing the load between different sched groups in a sched domain.
6400 * Typically cpu_power for all the groups in a sched domain will be same unless
6401 * there are asymmetries in the topology. If there are asymmetries, group
6402 * having more cpu_power will pickup more load compared to the group having
6405 * cpu_power will be a multiple of SCHED_LOAD_SCALE. This multiple represents
6406 * the maximum number of tasks a group can handle in the presence of other idle
6407 * or lightly loaded groups in the same sched domain.
6409 static void init_sched_groups_power(int cpu
, struct sched_domain
*sd
)
6411 struct sched_domain
*child
;
6412 struct sched_group
*group
;
6414 WARN_ON(!sd
|| !sd
->groups
);
6416 if (cpu
!= first_cpu(sd
->groups
->cpumask
))
6422 * For perf policy, if the groups in child domain share resources
6423 * (for example cores sharing some portions of the cache hierarchy
6424 * or SMT), then set this domain groups cpu_power such that each group
6425 * can handle only one task, when there are other idle groups in the
6426 * same sched domain.
6428 if (!child
|| (!(sd
->flags
& SD_POWERSAVINGS_BALANCE
) &&
6430 (SD_SHARE_CPUPOWER
| SD_SHARE_PKG_RESOURCES
)))) {
6431 sd
->groups
->cpu_power
= SCHED_LOAD_SCALE
;
6435 sd
->groups
->cpu_power
= 0;
6438 * add cpu_power of each child group to this groups cpu_power
6440 group
= child
->groups
;
6442 sd
->groups
->cpu_power
+= group
->cpu_power
;
6443 group
= group
->next
;
6444 } while (group
!= child
->groups
);
6448 * Build sched domains for a given set of cpus and attach the sched domains
6449 * to the individual cpus
6451 static int build_sched_domains(const cpumask_t
*cpu_map
)
6454 struct sched_domain
*sd
;
6456 struct sched_group
**sched_group_nodes
= NULL
;
6457 int sd_allnodes
= 0;
6460 * Allocate the per-node list of sched groups
6462 sched_group_nodes
= kzalloc(sizeof(struct sched_group
*)*MAX_NUMNODES
,
6464 if (!sched_group_nodes
) {
6465 printk(KERN_WARNING
"Can not alloc sched group node list\n");
6468 sched_group_nodes_bycpu
[first_cpu(*cpu_map
)] = sched_group_nodes
;
6472 * Set up domains for cpus specified by the cpu_map.
6474 for_each_cpu_mask(i
, *cpu_map
) {
6475 struct sched_domain
*sd
= NULL
, *p
;
6476 cpumask_t nodemask
= node_to_cpumask(cpu_to_node(i
));
6478 cpus_and(nodemask
, nodemask
, *cpu_map
);
6481 if (cpus_weight(*cpu_map
)
6482 > SD_NODES_PER_DOMAIN
*cpus_weight(nodemask
)) {
6483 sd
= &per_cpu(allnodes_domains
, i
);
6484 *sd
= SD_ALLNODES_INIT
;
6485 sd
->span
= *cpu_map
;
6486 cpu_to_allnodes_group(i
, cpu_map
, &sd
->groups
);
6492 sd
= &per_cpu(node_domains
, i
);
6494 sd
->span
= sched_domain_node_span(cpu_to_node(i
));
6498 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6502 sd
= &per_cpu(phys_domains
, i
);
6504 sd
->span
= nodemask
;
6508 cpu_to_phys_group(i
, cpu_map
, &sd
->groups
);
6510 #ifdef CONFIG_SCHED_MC
6512 sd
= &per_cpu(core_domains
, i
);
6514 sd
->span
= cpu_coregroup_map(i
);
6515 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6518 cpu_to_core_group(i
, cpu_map
, &sd
->groups
);
6521 #ifdef CONFIG_SCHED_SMT
6523 sd
= &per_cpu(cpu_domains
, i
);
6524 *sd
= SD_SIBLING_INIT
;
6525 sd
->span
= cpu_sibling_map
[i
];
6526 cpus_and(sd
->span
, sd
->span
, *cpu_map
);
6529 cpu_to_cpu_group(i
, cpu_map
, &sd
->groups
);
6533 #ifdef CONFIG_SCHED_SMT
6534 /* Set up CPU (sibling) groups */
6535 for_each_cpu_mask(i
, *cpu_map
) {
6536 cpumask_t this_sibling_map
= cpu_sibling_map
[i
];
6537 cpus_and(this_sibling_map
, this_sibling_map
, *cpu_map
);
6538 if (i
!= first_cpu(this_sibling_map
))
6541 init_sched_build_groups(this_sibling_map
, cpu_map
, &cpu_to_cpu_group
);
6545 #ifdef CONFIG_SCHED_MC
6546 /* Set up multi-core groups */
6547 for_each_cpu_mask(i
, *cpu_map
) {
6548 cpumask_t this_core_map
= cpu_coregroup_map(i
);
6549 cpus_and(this_core_map
, this_core_map
, *cpu_map
);
6550 if (i
!= first_cpu(this_core_map
))
6552 init_sched_build_groups(this_core_map
, cpu_map
, &cpu_to_core_group
);
6557 /* Set up physical groups */
6558 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6559 cpumask_t nodemask
= node_to_cpumask(i
);
6561 cpus_and(nodemask
, nodemask
, *cpu_map
);
6562 if (cpus_empty(nodemask
))
6565 init_sched_build_groups(nodemask
, cpu_map
, &cpu_to_phys_group
);
6569 /* Set up node groups */
6571 init_sched_build_groups(*cpu_map
, cpu_map
, &cpu_to_allnodes_group
);
6573 for (i
= 0; i
< MAX_NUMNODES
; i
++) {
6574 /* Set up node groups */
6575 struct sched_group
*sg
, *prev
;
6576 cpumask_t nodemask
= node_to_cpumask(i
);
6577 cpumask_t domainspan
;
6578 cpumask_t covered
= CPU_MASK_NONE
;
6581 cpus_and(nodemask
, nodemask
, *cpu_map
);
6582 if (cpus_empty(nodemask
)) {
6583 sched_group_nodes
[i
] = NULL
;
6587 domainspan
= sched_domain_node_span(i
);
6588 cpus_and(domainspan
, domainspan
, *cpu_map
);
6590 sg
= kmalloc_node(sizeof(struct sched_group
), GFP_KERNEL
, i
);
6592 printk(KERN_WARNING
"Can not alloc domain group for "
6596 sched_group_nodes
[i
] = sg
;
6597 for_each_cpu_mask(j
, nodemask
) {
6598 struct sched_domain
*sd
;
6599 sd
= &per_cpu(node_domains
, j
);
6603 sg
->cpumask
= nodemask
;
6605 cpus_or(covered
, covered
, nodemask
);
6608 for (j
= 0; j
< MAX_NUMNODES
; j
++) {
6609 cpumask_t tmp
, notcovered
;
6610 int n
= (i
+ j
) % MAX_NUMNODES
;
6612 cpus_complement(notcovered
, covered
);
6613 cpus_and(tmp
, notcovered
, *cpu_map
);
6614 cpus_and(tmp
, tmp
, domainspan
);
6615 if (cpus_empty(tmp
))
6618 nodemask
= node_to_cpumask(n
);
6619 cpus_and(tmp
, tmp
, nodemask
);
6620 if (cpus_empty(tmp
))
6623 sg
= kmalloc_node(sizeof(struct sched_group
),
6627 "Can not alloc domain group for node %d\n", j
);
6632 sg
->next
= prev
->next
;
6633 cpus_or(covered
, covered
, tmp
);
6640 /* Calculate CPU power for physical packages and nodes */
6641 #ifdef CONFIG_SCHED_SMT
6642 for_each_cpu_mask(i
, *cpu_map
) {
6643 sd
= &per_cpu(cpu_domains
, i
);
6644 init_sched_groups_power(i
, sd
);
6647 #ifdef CONFIG_SCHED_MC
6648 for_each_cpu_mask(i
, *cpu_map
) {
6649 sd
= &per_cpu(core_domains
, i
);
6650 init_sched_groups_power(i
, sd
);
6654 for_each_cpu_mask(i
, *cpu_map
) {
6655 sd
= &per_cpu(phys_domains
, i
);
6656 init_sched_groups_power(i
, sd
);
6660 for (i
= 0; i
< MAX_NUMNODES
; i
++)
6661 init_numa_sched_groups_power(sched_group_nodes
[i
]);
6664 struct sched_group
*sg
;
6666 cpu_to_allnodes_group(first_cpu(*cpu_map
), cpu_map
, &sg
);
6667 init_numa_sched_groups_power(sg
);
6671 /* Attach the domains */
6672 for_each_cpu_mask(i
, *cpu_map
) {
6673 struct sched_domain
*sd
;
6674 #ifdef CONFIG_SCHED_SMT
6675 sd
= &per_cpu(cpu_domains
, i
);
6676 #elif defined(CONFIG_SCHED_MC)
6677 sd
= &per_cpu(core_domains
, i
);
6679 sd
= &per_cpu(phys_domains
, i
);
6681 cpu_attach_domain(sd
, i
);
6684 * Tune cache-hot values:
6686 calibrate_migration_costs(cpu_map
);
6692 free_sched_groups(cpu_map
);
6697 * Set up scheduler domains and groups. Callers must hold the hotplug lock.
6699 static int arch_init_sched_domains(const cpumask_t
*cpu_map
)
6701 cpumask_t cpu_default_map
;
6705 * Setup mask for cpus without special case scheduling requirements.
6706 * For now this just excludes isolated cpus, but could be used to
6707 * exclude other special cases in the future.
6709 cpus_andnot(cpu_default_map
, *cpu_map
, cpu_isolated_map
);
6711 err
= build_sched_domains(&cpu_default_map
);
6716 static void arch_destroy_sched_domains(const cpumask_t
*cpu_map
)
6718 free_sched_groups(cpu_map
);
6722 * Detach sched domains from a group of cpus specified in cpu_map
6723 * These cpus will now be attached to the NULL domain
6725 static void detach_destroy_domains(const cpumask_t
*cpu_map
)
6729 for_each_cpu_mask(i
, *cpu_map
)
6730 cpu_attach_domain(NULL
, i
);
6731 synchronize_sched();
6732 arch_destroy_sched_domains(cpu_map
);
6736 * Partition sched domains as specified by the cpumasks below.
6737 * This attaches all cpus from the cpumasks to the NULL domain,
6738 * waits for a RCU quiescent period, recalculates sched
6739 * domain information and then attaches them back to the
6740 * correct sched domains
6741 * Call with hotplug lock held
6743 int partition_sched_domains(cpumask_t
*partition1
, cpumask_t
*partition2
)
6745 cpumask_t change_map
;
6748 cpus_and(*partition1
, *partition1
, cpu_online_map
);
6749 cpus_and(*partition2
, *partition2
, cpu_online_map
);
6750 cpus_or(change_map
, *partition1
, *partition2
);
6752 /* Detach sched domains from all of the affected cpus */
6753 detach_destroy_domains(&change_map
);
6754 if (!cpus_empty(*partition1
))
6755 err
= build_sched_domains(partition1
);
6756 if (!err
&& !cpus_empty(*partition2
))
6757 err
= build_sched_domains(partition2
);
6762 #if defined(CONFIG_SCHED_MC) || defined(CONFIG_SCHED_SMT)
6763 int arch_reinit_sched_domains(void)
6768 detach_destroy_domains(&cpu_online_map
);
6769 err
= arch_init_sched_domains(&cpu_online_map
);
6770 unlock_cpu_hotplug();
6775 static ssize_t
sched_power_savings_store(const char *buf
, size_t count
, int smt
)
6779 if (buf
[0] != '0' && buf
[0] != '1')
6783 sched_smt_power_savings
= (buf
[0] == '1');
6785 sched_mc_power_savings
= (buf
[0] == '1');
6787 ret
= arch_reinit_sched_domains();
6789 return ret
? ret
: count
;
6792 int sched_create_sysfs_power_savings_entries(struct sysdev_class
*cls
)
6796 #ifdef CONFIG_SCHED_SMT
6798 err
= sysfs_create_file(&cls
->kset
.kobj
,
6799 &attr_sched_smt_power_savings
.attr
);
6801 #ifdef CONFIG_SCHED_MC
6802 if (!err
&& mc_capable())
6803 err
= sysfs_create_file(&cls
->kset
.kobj
,
6804 &attr_sched_mc_power_savings
.attr
);
6810 #ifdef CONFIG_SCHED_MC
6811 static ssize_t
sched_mc_power_savings_show(struct sys_device
*dev
, char *page
)
6813 return sprintf(page
, "%u\n", sched_mc_power_savings
);
6815 static ssize_t
sched_mc_power_savings_store(struct sys_device
*dev
,
6816 const char *buf
, size_t count
)
6818 return sched_power_savings_store(buf
, count
, 0);
6820 SYSDEV_ATTR(sched_mc_power_savings
, 0644, sched_mc_power_savings_show
,
6821 sched_mc_power_savings_store
);
6824 #ifdef CONFIG_SCHED_SMT
6825 static ssize_t
sched_smt_power_savings_show(struct sys_device
*dev
, char *page
)
6827 return sprintf(page
, "%u\n", sched_smt_power_savings
);
6829 static ssize_t
sched_smt_power_savings_store(struct sys_device
*dev
,
6830 const char *buf
, size_t count
)
6832 return sched_power_savings_store(buf
, count
, 1);
6834 SYSDEV_ATTR(sched_smt_power_savings
, 0644, sched_smt_power_savings_show
,
6835 sched_smt_power_savings_store
);
6839 * Force a reinitialization of the sched domains hierarchy. The domains
6840 * and groups cannot be updated in place without racing with the balancing
6841 * code, so we temporarily attach all running cpus to the NULL domain
6842 * which will prevent rebalancing while the sched domains are recalculated.
6844 static int update_sched_domains(struct notifier_block
*nfb
,
6845 unsigned long action
, void *hcpu
)
6848 case CPU_UP_PREPARE
:
6849 case CPU_DOWN_PREPARE
:
6850 detach_destroy_domains(&cpu_online_map
);
6853 case CPU_UP_CANCELED
:
6854 case CPU_DOWN_FAILED
:
6858 * Fall through and re-initialise the domains.
6865 /* The hotplug lock is already held by cpu_up/cpu_down */
6866 arch_init_sched_domains(&cpu_online_map
);
6871 void __init
sched_init_smp(void)
6873 cpumask_t non_isolated_cpus
;
6876 arch_init_sched_domains(&cpu_online_map
);
6877 cpus_andnot(non_isolated_cpus
, cpu_possible_map
, cpu_isolated_map
);
6878 if (cpus_empty(non_isolated_cpus
))
6879 cpu_set(smp_processor_id(), non_isolated_cpus
);
6880 unlock_cpu_hotplug();
6881 /* XXX: Theoretical race here - CPU may be hotplugged now */
6882 hotcpu_notifier(update_sched_domains
, 0);
6884 /* Move init over to a non-isolated CPU */
6885 if (set_cpus_allowed(current
, non_isolated_cpus
) < 0)
6889 void __init
sched_init_smp(void)
6892 #endif /* CONFIG_SMP */
6894 int in_sched_functions(unsigned long addr
)
6896 /* Linker adds these: start and end of __sched functions */
6897 extern char __sched_text_start
[], __sched_text_end
[];
6899 return in_lock_functions(addr
) ||
6900 (addr
>= (unsigned long)__sched_text_start
6901 && addr
< (unsigned long)__sched_text_end
);
6904 void __init
sched_init(void)
6908 for_each_possible_cpu(i
) {
6909 struct prio_array
*array
;
6913 spin_lock_init(&rq
->lock
);
6914 lockdep_set_class(&rq
->lock
, &rq
->rq_lock_key
);
6916 rq
->active
= rq
->arrays
;
6917 rq
->expired
= rq
->arrays
+ 1;
6918 rq
->best_expired_prio
= MAX_PRIO
;
6922 for (j
= 1; j
< 3; j
++)
6923 rq
->cpu_load
[j
] = 0;
6924 rq
->active_balance
= 0;
6927 rq
->migration_thread
= NULL
;
6928 INIT_LIST_HEAD(&rq
->migration_queue
);
6930 atomic_set(&rq
->nr_iowait
, 0);
6932 for (j
= 0; j
< 2; j
++) {
6933 array
= rq
->arrays
+ j
;
6934 for (k
= 0; k
< MAX_PRIO
; k
++) {
6935 INIT_LIST_HEAD(array
->queue
+ k
);
6936 __clear_bit(k
, array
->bitmap
);
6938 // delimiter for bitsearch
6939 __set_bit(MAX_PRIO
, array
->bitmap
);
6943 set_load_weight(&init_task
);
6946 open_softirq(SCHED_SOFTIRQ
, run_rebalance_domains
, NULL
);
6949 #ifdef CONFIG_RT_MUTEXES
6950 plist_head_init(&init_task
.pi_waiters
, &init_task
.pi_lock
);
6954 * The boot idle thread does lazy MMU switching as well:
6956 atomic_inc(&init_mm
.mm_count
);
6957 enter_lazy_tlb(&init_mm
, current
);
6960 * Make us the idle thread. Technically, schedule() should not be
6961 * called from this thread, however somewhere below it might be,
6962 * but because we are the idle thread, we just pick up running again
6963 * when this runqueue becomes "idle".
6965 init_idle(current
, smp_processor_id());
6968 #ifdef CONFIG_DEBUG_SPINLOCK_SLEEP
6969 void __might_sleep(char *file
, int line
)
6972 static unsigned long prev_jiffy
; /* ratelimiting */
6974 if ((in_atomic() || irqs_disabled()) &&
6975 system_state
== SYSTEM_RUNNING
&& !oops_in_progress
) {
6976 if (time_before(jiffies
, prev_jiffy
+ HZ
) && prev_jiffy
)
6978 prev_jiffy
= jiffies
;
6979 printk(KERN_ERR
"BUG: sleeping function called from invalid"
6980 " context at %s:%d\n", file
, line
);
6981 printk("in_atomic():%d, irqs_disabled():%d\n",
6982 in_atomic(), irqs_disabled());
6983 debug_show_held_locks(current
);
6984 if (irqs_disabled())
6985 print_irqtrace_events(current
);
6990 EXPORT_SYMBOL(__might_sleep
);
6993 #ifdef CONFIG_MAGIC_SYSRQ
6994 void normalize_rt_tasks(void)
6996 struct prio_array
*array
;
6997 struct task_struct
*p
;
6998 unsigned long flags
;
7001 read_lock_irq(&tasklist_lock
);
7002 for_each_process(p
) {
7006 spin_lock_irqsave(&p
->pi_lock
, flags
);
7007 rq
= __task_rq_lock(p
);
7011 deactivate_task(p
, task_rq(p
));
7012 __setscheduler(p
, SCHED_NORMAL
, 0);
7014 __activate_task(p
, task_rq(p
));
7015 resched_task(rq
->curr
);
7018 __task_rq_unlock(rq
);
7019 spin_unlock_irqrestore(&p
->pi_lock
, flags
);
7021 read_unlock_irq(&tasklist_lock
);
7024 #endif /* CONFIG_MAGIC_SYSRQ */
7028 * These functions are only useful for the IA64 MCA handling.
7030 * They can only be called when the whole system has been
7031 * stopped - every CPU needs to be quiescent, and no scheduling
7032 * activity can take place. Using them for anything else would
7033 * be a serious bug, and as a result, they aren't even visible
7034 * under any other configuration.
7038 * curr_task - return the current task for a given cpu.
7039 * @cpu: the processor in question.
7041 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7043 struct task_struct
*curr_task(int cpu
)
7045 return cpu_curr(cpu
);
7049 * set_curr_task - set the current task for a given cpu.
7050 * @cpu: the processor in question.
7051 * @p: the task pointer to set.
7053 * Description: This function must only be used when non-maskable interrupts
7054 * are serviced on a separate stack. It allows the architecture to switch the
7055 * notion of the current task on a cpu in a non-blocking manner. This function
7056 * must be called with all CPU's synchronized, and interrupts disabled, the
7057 * and caller must save the original value of the current task (see
7058 * curr_task() above) and restore that value before reenabling interrupts and
7059 * re-starting the system.
7061 * ONLY VALID WHEN THE WHOLE SYSTEM IS STOPPED!
7063 void set_curr_task(int cpu
, struct task_struct
*p
)